Onsets: Suprasegmental and Prosodic Behaviour (Cambridge Studies in Linguistics)

O N S ETS

The concept of the ‘onset’, i.e. the consonant(s) before the vowel of a syllable, is critical within phonology. While phonologists have examined the segmental behaviour of onsets, their prosodic status has instead been largely overlooked. In fact, most previous accounts have stipulated that onsets are insignificant when it comes to the ‘heaviness’ of syllables. In this book Nina Topintzi presents a new theory of onsets, arguing for their fundamental role in the structure of language both in the underlying and surface representation, unlike previous assumptions. To capture the weight behaviour of onsets, a novel account is proposed that relates their interaction with voicing, tone and stress. Using numerous case-studies and data from a variety of languages and phenomena (including stress, compensatory lengthening, gemination and word minimality), the book introduces a model that reflects the true behaviour of onsets, demonstrating profound implications for syllable and weight theories. n i na t o p i n t zi is a Teaching Fellow in the School of English at Aristotle University of Thessaloniki.

In this series 80. 81. 82. 83. 84. 85. 86. 87. 88. 89. 90. 91. 92. 93. 94. 95. 96. 97. 98. 99. 100. 101. 102. 103. 104. 105. 106. 107. 108. 109. 110. 111. 112. 113. 114. 115. 116. 117. 118. 119. 120. 121. 122. 123. 124. 125.

an drea mor o : The raising of predicates: predicative noun phrases and the theory of clause structure roger lass : Historical linguistics and language change j oh n m. a n de r son : A notional theory of syntactic categories bernd hein e : Possession: cognitive sources, forces and grammaticalization n omi ertesc h i k - shi r : The dynamics of focus structure j oh n colem an : Phonological representations: their names, forms and powers christina y. b e t h in : Slavic prosody: language change and phonological theory ba rba ra d an cyg ie r : Conditionals and prediction claire lefe b vr e : Creole genesis and the acquisition of grammar: the case of Haitian creole h ein z giege r ic h : Lexical strata in English keren rice : Morpheme order and semantic scope april mc mah on : Lexical phonology and the history of English matthew y. ch e n : Tone Sandhi: patterns across Chinese dialects grego ry t. st u mp : Inflectional morphology: a theory of paradigm structure j oa n by bee : Phonology and language use laurie bau e r : Morphological productivity tho ma s ern s t : The syntax of adjuncts eliza beth c l oss t raug ot t and r i c ha r d b. da s her : Regularity in semantic change maya hick man n : Children’s discourse: person, space and time across languages d iane bla kem o r e : Relevance and linguistic meaning: the semantics and pragmatics of discourse markers ia n ro berts and an na r ou ssou : Syntactic change: a minimalist approach to grammaticalization do nka minkova : Alliteration and sound change in early English ma rk c. ba ke r : Lexical categories: verbs, nouns and adjectives carlo ta s . sm i th : Modes of discourse: the local structure of texts ro chelle li e b er : Morphology and lexical semantics h olger die s se l : The acquisition of complex sentences s ha ro n ink e l as and c h e r yl zol l : Reduplication: doubling in morphology s usa n edwa r ds : Fluent aphasia barbara da ncygi e r and e ve s we e t se r : Mental spaces in grammar: conditional constructions hew baerm an , d un stan b r ow n an d g r ev il le g . corbett : The syntax-morphology interface: a study of syncretism marcus tom al in : Linguistics and the formal sciences: the origins of generative grammar sa mu el d. ep st e i n and t. dan i e l se e ly : Derivations in minimalism paul de lacy : Markedness: reduction and preservation in phonology yehu da n. fal k : Subjects and their properties p. h . matth ew s : Syntactic relations: a critical survey mark c. bak e r : The syntax of agreement and concord gillian cat ri ona r am c h an d : Verb meaning and the lexicon: a first phase syntax pieter muy sk e n : Functional categories j ua n uriag e r e ka : Syntactic anchors: on semantic structuring d. robert l a dd : Intonational phonology second edition leonard h. b ab b y : The syntax of argument structure b. elan dre sh e r : The contrastive hierarchy in phonology dav id adg er , dan i e l h ar b ou r and l au r e l j. watk ins : Mirrors and microparameters: phrase structure beyond free word order n iina ning zh ang : Coordination in syntax n eil smith : Acquiring phonology n ina top in t zi : Onsets: suprasegmental and prosodic behaviour

Earlier issues not listed are also available

C AM B RIDGE ST UDI ES IN LING UIS TIC S G eneral E di to rs: p. austi n , j. bresna n , b. com rie , s . cr ai n , w. dressler , c. j. ew en , r. la ss, d. l ig ht foot , k. r ice , i . robe rts , s. ro m ain e , n. v. s mit h

Onsets: Suprasegmental and Prosodic Behaviour

O NS E TS S U P R A S E G M E N TA L A N D P R O S O D I C B E H AV I O U R

N I NA TO P I N T Z I Aristotle University, Thessaloniki

ca mbri d g e u ni v e r si t y p r e s s Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paulo, Delhi, Dubai, Tokyo Cambridge University Press The Edinburgh Building, Cambridge CB2 8RU, UK Published in the United States of America by Cambridge University Press, New York www.cambridge.org Information on this title: www.cambridge.org/9780521493352 © Nina Topintzi 2010 This publication is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published 2010 Printed in the United Kingdom at the University Press, Cambridge A catalogue record for this publication is available from the British Library Library of Congress Cataloguing in Publication data Topintzi, Nina. Onsets : suprasegmental and prosodic behaviour / Nina Topintzi. p. cm. – (Cambridge studies in linguistics ; 125) Includes bibliographical references and index. ISBN 978-0-521-49335-2 (hardback) 1. Prosodic analysis (Linguistics) 2. Grammar, Comparative and general–Phonology. 3. Syllabication. I. Title. II. Series. P224.T67 2010 414'.6–dc22 2010001517 ISBN 978-0-521-49335-2 Hardback Cambridge University Press has no responsibility for the persistence or accuracy of URLs for external or third-party internet websites referred to in this publication, and does not guarantee that any content on such websites is, or will remain, accurate or appropriate.

Contents

Preface Acknowledgements

page ix xii

1

Onsets and weight: the theory

1.1 1.2 1.3 1.4

Aims Why a book on moraic onsets? The theory of onset weight Organization of the book

1 2 11 33

2

Onsets and stress

37

2.1 2.2

37

2.5 2.6

Introduction Languages where the quality of onsets (QO) matters exclusively (❶) Languages where the presence of onsets (PO) matters exclusively (❷) PO and QO effects simultaneously in a language: the case of Pirahã (❸) Arabela, a case ambiguous between (❶) and (❸) Discussion and conclusion

3

Onsets and compensatory lengthening

100

3.1 3.2 3.3 3.4

Aims Samothraki Greek Extensions of the proposal Conclusion

100 102 122 134

4

Onsets and word minimality

137

4.1 4.2 4.3 4.4 4.5

Introduction Bella Coola background Word minimality Alternatives that dispense with onset moraicity Conclusion

137 138 146 159 163

2.3 2.4

1

39 57 75 83 97

vii

viii

Contents

5

Onsets and geminates

164

5.1 5.2 5.3 5.4 5.5

Introduction Word-initial moraic-onset geminates Word-medial moraic-onset geminates Typology, more theory and alternatives Concluding remarks

164 166 184 191 200

6

Other real and not so real onset-sensitive data: brief case-studies

202

6.1 6.2 6.3 6.4

Aims Additional phenomena and/or languages indicative of moraic onsets Misanalysed onset-sensitive languages Summary

202 203 222 228

7

Conclusion and discussion of alternatives

230

7.1 7.2 7.3

The book’s main themes Alternatives Conclusion

230 237 245

References Subject index Language index

247 264 266

Preface

Theme As the title communicates, this book is about onsets. In particular, it focuses on their suprasegmental and prosodic behaviour. This on its own is quite interesting, given that most current phonological theory assigns no such role to onsets. The core of the discussion is devoted to proving this assertion wrong. The pivotal proposal this book wishes to make is that onset weight exists. And like rimal weight, it participates in the same phenomena, that is, stress, compensatory lengthening, word minimality and gemination, among others. However, onset weight is not unrestricted. It may either be produced on the surface as a response to a weight-inducing requirement or it may be lexical, in which case its source is found in the underlying representation. The former type of onset weight is termed coerced, whereas the latter is dubbed distinctive, following Morén’s (2001) similar distinction for coda weight. The first type of weight – but not the latter – is subject to certain limitations and thus can be assigned to a subset of segments only. These restrictions represent a leading idea in the book. In particular, it is claimed that the optimal weightful onsets are those that lack the feature [+voice]. As a result, the prototypical moraic onsets are [−voice] ones, i.e. the voiceless obstruents. Justification for this account comes from the relationship between voicing, tone and stress. Examination of the tonogenesis facts reveals that the pitch perturbation due to (the lack of) voicing is commonly phonologized as tone. Extending this idea, I propose that in some languages such pitch perturbation is phonologized as stress, in terms of moras. The languages Karo, Pirahã and Arabela exhibit this pattern. Of paramount importance is a related claim made with regard to sonorants. These are argued to be marked on a language-specific basis as [+voi] or lack any [voi] feature whatsoever. This statement is supported by the behaviour of sonorants in the examination of tone and stress. It is postulated that they may pattern with the voiced obstruents when they bear [+voi]; but they may also ix

x

Preface

pattern alongside the voiceless obstruents, if they lack [voi] specification. As explained, the system built is such that it can treat [−voi] segments and those lacking [voi] uniformly to the exclusion of [+voi] ones. Still, as far as onset-sensitive stress is concerned, a crucial distinction is made. Languages may be sensitive to the presence of an onset (PO effect) or to its quality (QO effect), but also to both (PO&QO effect). Importantly, PO effects are down to alignment considerations, whereas QO effects depend on weight considerations. For instance, in Aranda, an onsetful syllable attracts stress more than an onsetless one (PO); in Karo, syllables with onsets of a certain quality attract stress more than others (QO); and finally, in Pirahã, the two effects are combined (PO&QO). The subsequent discussion centres on the interaction of onsets with other phenomena, namely compensatory lengthening, word minimality and gemination. The reasoning behind this is the following: onset-sensitive stress on its own is not sufficient to uphold the onset-weight hypothesis. This is because there are other ways to account for stress, e.g. the concept of prominence, which bypass reference to weight. Thus, providing evidence on the effects of onsets from uncontroversially weight-based phenomena surely offers solid grounding to the onset-weight theory. To this end, a large number of case-studies is examined and formalized using the framework of Optimality Theory (OT) (Prince and Smolensky 1993/2004; McCarthy and Prince 1995). Many of the analyses are very detailed, thus also supplying the technical implementation of the theoretical ideas proposed here. Finally, the theoretical proposals that are presented often extend beyond onsets. For instance, the chapters on compensatory lengthening and geminates provide full-fledged analyses (in terms of theoretical and empirical scope, as well as technical detail) that promise to account for a wide range of facts and cases. Audience and use This book revolves around the topic of onsets, but in doing so, it addresses several core phenomena in suprasegmental and prosodic phonology. It is thus of interest to anyone who works on syllables, geminates, weight theory, compensatory lengthening, word minimality, tone and reduplication. It will also prove useful to segmental phonologists, especially those interested in [voice] and its interaction with the prosodic phenomena of tone and stress.

Preface

xi

Typologists interested in phonology can certainly use this book as a resource on some typologically rare languages and find information on languages hardly discussed elsewhere. This study should be accessible to anyone with some background in phonology. While it is true that its technical aspects will be fully appreciated by OT-theorists (or anyone with some good knowledge of OT), the core insights, such as the weightfulness of onsets and the relationship between stress and voicing, are theory neutral. The book can be used in various ways: by the phonologist in general, as a new model of syllable and weight to apply in her/his work; by the OT-theorist who wishes to espouse (some of) the ideas proposed here for their technical merits; by the typologist, as a resource for typological studies; and by the student, as a useful review of past and recent theories on syllable weight, compensatory lengthening and gemination. A final word This book is an updated and revised version of my 2006 thesis titled ‘Moraic Onsets’, at University College London. Much of the material in here is significantly altered, although the core idea remains the same: ‘onsets can be moraic’! Most analyses have been changed (e.g. Arabela, Bella Coola, Samothraki Greek, etc.), while the discussion on [voice] as well as sonorants has been updated to reflect findings of more recent work, such as Tang (2008). Some material has been discarded in the interest of clarity, while certain other sections have been added (see the discussion on medial-onset geminates). The interested reader can consult Topintzi (2006b) for comparison. Finally, earlier versions of some of the material presented here have appeared in previously published work of mine. This includes: Topintzi (2006a) and Topintzi (2008a, b).

Acknowledgements

In December 2001, as an MA student at UCL, I listened to a colloquium talk by Dan Everett on the phonology of Pirahã. I thought Dan was an engaging speaker and the data he reported on fascinating. A year later, as a first-year PhD student, I thought about this talk again; Pirahã has been notorious for its onset-sensitive stress. But, was it the only such language? Could there be more cases like it, and in fact, could it be that the role of onsets in prosody had so far been underestimated? These questions persist within this book – and previously in my thesis – in the hope that by the end of it a satisfactory answer will have been given. During the years that I’ve been looking into this topic (2002–9), a number of people generously contributed to it with their suggestions, comments and help in gaining access to relevant material. First and foremost, my thanks go to Moira Yip, my thesis supervisor and academic role model, for teaching me good phonology (any ‘bad’ phonology here is purely my fault), helping me shape my own ideas, sharing her expertise and vast knowledge and for the most exquisite ‘Tosca’ I have and will possibly ever enjoy. Dan Everett has been a constant resource regarding Pirahã and provided a lot of encouragement when I first took up this project. Since 2005, Marc van Oostendorp has been my ‘life-vest’ whenever I seemed to get stuck on a piece of data, always offering insightful ways to tackle the problem in question. I thank both a lot! Many thanks also go to Matt Gordon, Iggy Roca and a third anonymous reviewer for this book. I thank them for their careful reading and their invaluable comments. I am especially grateful to Matt, who offered the most helpful feedback possible, disregarding his own views on the issue – expressed in Gordon (2005) – and generously offering remarks to make my own work genuinely better. Besides the above, I also gratefully acknowledge: Mary Baltazani, JeanMarc Beltzung, Juliette Blevins, Lev Blumenfeld, Emily Curtis, Paul de Lacy, Mark Donohue, Paloma Garcia-Bellido, Spike Gildea, Rob Goedemans, Carolina González, Beverley Goodman, Carlos Gussenhoven, John Hajek, xii

Acknowledgements

xiii

John Harris, Bruce Hayes, Ben Hermans, Larry Hyman, Wouter Jansen, Brian Jose, Dasha Kavitskaya, Michael Kenstowicz, Debbie Koop (SIL Peru), YenHwei Lin, Angeliki Malikouti-Drachman, Bruce Morén, Elliott Moreton, David Nash, Mary Pearce, Anthi Revithiadou, Curt Rice, Keren Rice, Roland Rich, Joe Salmons, Toshi Shiraishi, Jen Smith, Paul Smolensky, Donca Steriade, Mary Ruth Wise, Yi Xu and Hideki Zamma. Thanks to Eleftheria Giakoumaki and Angelos Afroudakis at the Κέντρον Ερεύνης των Νεοελληνικών ∆ιαλέκτων και Ιδιωμάτων (Academy of Athens Research Centre for the Greek Dialects and Idioms) – Ι.Λ.Ν.Ε for granting me access to older Samothraki Greek sources. Also to Marianna Ronga, Maria Tsolaki and Allia Foteinou, my informants on Samothraki Greek. Warm thanks also go to my fellow students at UCL (2002–6) and my colleagues at the Universities of Patras and Thessaloniki (AUTH), where I have been employed for the most part till now (2007–9). Many thanks (and congratulations!!) to colleagues and friends for standing by and for occasionally putting up with me during the last couple of years: Xara and Tasoula Balaska, Mary Baltazani, Yannis and Labros Giogios, Ioanna Kappa, Giouli Kommata, Marika Lekakou, Despoina Papadopoulou, Vina Tsakali and Eirini Tsiknidou. For wonderful co-operation and communication, I am grateful to Helen Barton, my editor at CUP. I kept my warmest thanks for last. This goes without saying to my family: my parents Betty and Takis and my sister Ermina (and to our newest, furrier, family members Astreha and Akrali). Their contribution is one that this work could not have existed without: unconditional love and unfailing support.

1 Onsets and weight: the theory

. . . no language has a rule stressing the penultimate syllable unless it begins with a voiced consonant, in which case one stresses the antepenultimate syllable Hyman 1985: 96 [Karo] stress can be predicted by the onset of the last syllable: if it is a voiced stop consonant, then the stress shifts one syllable to the left Gabas 1999: 39

1.1

Aims

This book explores the role of onsets in syllable-weight theory and consequently on prosody. In particular, I will be arguing that onsets, like nuclei and codas, can bear weight. Although this idea is one that has to be seriously considered by every phonologist interested in stress, weight and prosodic structure, it has nonetheless been overlooked. In fact, virtually all work on syllable structure and weight, the most prominent being Hyman (1985), Hayes (1989), Goedemans (1998), Morén (2001) and Gordon (2006, the published version of his 1999 thesis), maintains that onsets can never be weightful. However, in most cases, this claim is made purely by stipulation and for convenience, as e.g. Morén (2001: 8) also acknowledges: ‘Onsets are typically non-moraic. Although this is not the only logical possibility, it is convenient and I assume it here’ (emphasis added mine). The present book aims to fill this gap in the literature and challenge the standard assumption that sees onsets as weightless, offering fresh insights around this topic. This comes in support of the recent marginal literature – basically Hajek and Goedemans (2003) and Gordon (2005) – that hesitantly admits some type of weightful onsets. However, unlike those accounts that focus on specific aspects of the issue, e.g. stress (Gordon 2005) or geminates (Hajek and Goedemans 2003), this book offers the first comprehensive study in terms of argumentation, length and concreteness on the issue of onset weight.

1

2

Onsets

Using findings of the past literature as a starting point, it presents a novel theory whose ambition is to encompass phenomena from a host of languages (many of which had never been discussed in this light before) in a unifying, explanatory and restrictive way. It argues against the prosodic inertness of onsets and proposes a modification of the syllable-weight model where onsets, like codas, can sometimes be moraic and sometimes not. To this end, the book may also be used as a resource for other researchers, as it contains a collection of languages and data which have been argued (rightly or not) to exhibit onset sensitivity. This first chapter consists of two parts. The first (§1.2) is quite introductory. It aims at establishing the need for a book of this kind and sets the scene for the topic under consideration. The second, longer, part (§1.3) is more theoretical and technical in nature. It supplies the theory that will be implemented in the forthcoming chapters where several case-studies will be discussed and analysed. 1.2

Why a book on moraic onsets?

This part starts by briefly looking into syllable weight (§1.2.1), as well as some of the models that have been proposed to capture it. It will soon become evident that moraic theory (Hyman 1985; Hayes 1989) stands out as the most successful of all (§.1.2.2). In its standard conception, however, moraic theory proves empirically insufficient, since it explicitly excludes a range of cases and data that are actually attested (§1.2.3–4). Rather than replacing it, though, this book argues that with some modification – namely by allowing onsets to be weightful – moraic theory can incorporate these cases too, and thus emerge as a complete, accurate and yet restrictive theory of weight. Section 1.3 explains how this is possible. 1.2.1 Syllable weight Syllable weight refers to the idea that syllables of different structure behave in different ways in prosodic phenomena and processes such as stress, reduplication, tone, compensatory lengthening, word minimality and others. In the languages that make a distinction based on weight, syllables are either heavy or light. They are heavy if they contain a long vowel (VV) and light if they are simply made up of a short vowel (V). Syllables with a short vowel followed by a coda (VC) are heavy or light depending on the language. In many languages stress is attracted to heavy syllables. Hopi and Lenakel are languages of this sort. However, while in Hopi VC counts as heavy, in Lenakel it counts as light.

Onsets and weight: the theory (1)

3

Hopi: VV/VC=heavy; V=light a. qǿq.tø.som.pi sóː.ja b. qø.tǿ.som.pi ko.jó.ŋo

‘headbands’ ‘planting stick’ ‘headband’ ‘turkey’

In Hopi (Jeanne 1982, cited in Gussenhoven and Jacobs 2005: 145), the first syllable is stressed if it is heavy (C)VV or (C)VC (1a), but if it is light (C)V, then the second syllable gets stress (1b).1 In Lenakel (Lynch 1974, 1978 cited in Hayes 1995), on the other hand, (C)VCs are considered light and primary stress appears on the penult (2a). Simplifying a bit, secondary stress is (usually) assigned to the first syllable and to alternate syllables after that (2b). However, this pattern may be disrupted; (C)VCs do not get secondary stress (cf. unstressed mol in (2d)) unless they happen to be located in a position that would receive rhythmic stress anyway; in contrast, heavy (C)VVs get stress no matter what their position (cf. kìː in 2c). (2)

Lenakel: VV=heavy; VC/V=light a. éheŋ rɨmáwŋɨn b. lètupwàlukáluk c. nɨkìːnílar ̀ olkéykey d. rƚ m

‘to blow the nose’ ‘he ate’ ‘in the lungs’ ‘their (pl.) hearts’ ‘he liked it’

(Lynch 1978: 16) (Lynch 1978: 19) (Lynch 1974: 183) (Lynch 1974: 198) (Lynch 1978: 19)

Several other phenomena make reference to syllable weight. A by-no-meansexhaustive list includes: (3)

Phenomena involving syllable weight I.

Compensatory lengthening: the lengthening that occurs after deletion of a segment e.g. Turkish (Roca and Johnson 1999) tahsil → taːsil ‘education’

II.

Word minimality: the minimum word size some languages impose [commonly (C)VC or (C)VV] e.g. Dalabon (Capell 1962; Garrett 1999) words that are CVC bad ‘stone’ or CVV biː ‘man’ are allowed, CV words are banned

III. Poetic metre: the organization of syllables into feet in songs or poetry e.g. Greek and Latin dactylic hexameter in epic poetry. The verse consists of six metra, each of which is made up of one heavy and two short syllables ( ̶""); however, two short syllables can be replaced by one heavy in which case we have a spondee ( ̶ ̶). Boundaries of metra are marked by parentheses: 1

In this chapter, unless stated otherwise, the acute accent marks primary stress, the grave accent means secondary stress and underlining denotes the reduplicated portion. I will interchangeably use VV or Vː to refer to long bimoraic vowels.

4

Onsets (cārmǐnă) (quāe vūl)(tīs cōg)(nōscǐtě); (cārmǐnă) (vōbīs) (Vergil, Eclogues VI. 25) IV. Reduplication: the repetition of part of a word (commonly a heavy syllable) e.g. Mokilese progressive (Harrison 1976; McCarthy and Prince 1986) poki~pokpoki ‘beat’, kookɔ~kookookɔ ‘grind coconut’, but pa~paapa ‘weave’ V. (Prosodic) Truncation: shortening of forms, as in the production of nicknames e.g. among other patterns, acceptable Japanese nicknames are a single heavy syllable (Mester 1990; Benua 1995) Midori~Mii-čan or JuNko~JuN-čan (-čan is the diminutive suffix) VI. Gemination: the consonant doubling that often occurs after short stressed vowels, so that the syllable is rendered heavy e.g. Kukatj (Breen 1992) or in Swedish dialects (Kiparsky 2008b) such as viss.na ‘to wilt’, takk.sa ‘rate’, hall.va ‘half’ VII. Tone: the use of pitch to mark different morphemes e.g. in Hausa (Gordon 2006) contour tones are only tolerated in heavy, but not in light, CV syllables, i.e. lâːláː ‘indolence’, mântáː ‘forget’, râsːáː ‘branches’

The distinction between heavy and light syllables was recognized as early as Jakobson (1931) and Trubetzkoy (1939) and has since been formalized in three major ways: a) CV theory (McCarthy 1979; Clements and Keyser 1983), b) the X-slot model (Levin 1985), and c) the moraic model (Hyman 1985; Hayes 1989). All three theories assign abstract weight units to segments in the syllable. The difference lies in what kinds of units these are and exactly what syllable constituents are identified, which of course has repercussions on the predictions made. For example, the syllable tan would be represented in the first two models in the following way. (4)

CV theory a. σ

X-slot model σ b.

CVC

R

t a n

O NC XXX

R=rime O=onset N=nucleus C=coda

t a n

In CV theory (4a), segments are marked as C-ones and V-ones, whereas in the X-slot model (4b), the more generic tag X is used to refer to both consonants and vowels. The latter notation has a welcome result; there is evidence that the C and V labels can sometimes be far too specific. For instance,

Onsets and weight: the theory

5

in Ancient Greek, the form esmi ‘I am’ underwent s-deletion and subsequent compensatory lengthening. In some dialects, the resulting form was emmi with C-lengthening, while in others it was eemi with V-lengthening. CV theory can account for emmi, because the vacated C-position of s is filled by a consonant too, but it fails to do so in eemi where the position is held by a V. This is not a problem shared by the X-slot theory, since X slots, by being general enough, circumvent this problem of labelling. X slots too, however, prove inadequate. Consider the example of Japanese from (3V) above. Given hypocoristics such as Midori~Mii-čan or JuNko~JuNčan, one can simply state that hypocoristic formation involves heavy [XXX] templates, i.e. [CVV] or [CVC] syllables. This idea cannot be maintained once other possible nicknames are considered, as shown in (5). (5)

Japanese Hypocoristics (Benua 1995) Midori Hanako

Mido-čan, Mii-čan Hana-čan, Haa-čan, Hač-čan

In these examples, the nicknames – excluding the diminutive suffix čan – can either be monosyllabic [CVV] or [CVC] or bisyllabic [CVCV]. Obviously, this pattern cannot be captured by a template [XXX]. Data like these have led to a further improvement of the syllable-weight theory by utilizing the concept of moras, as proposed in Hyman (1985) and especially Hayes (1989). 1.2.2 Advantages of moraic theory Hyman (1985) proposes a model which consists of weight units (WUs) whose function is virtually identical to moras, which is why I will simplify and use moras for Hyman’s representations too (6a). For our purposes, the most important property of this model is that underlyingly all segments start off with at least one WU (6a.i). Crucially on the surface, onsets lose their WU (indicated by the crossed-out mora in (6a.ii)) due to the universal application of the Onset-Creation Rule (OCR). This rule applies whenever a [+cons] segment is followed by a [−cons] segment and its effect is to delete the WU of the [+cons] segment. Subsequently, the [+cons] feature matrix associates to the WU of the [−cons] segment on its right. In other words, the nucleic WU/mora dominates both the onset and the nucleus of the syllable (6a.ii). (6)

a. Hyman (1985) i. underlying form µ µµ t a n

ii. surface form σ

b. Hayes (1989) surface form σ

µ µµ

µµ

t a n

t a n

6

Onsets

This differs from Hayes (1989), who assumes that the nucleic mora is not shared between the onset and the nucleus, but only associates to the nucleus. The onset instead links directly to the syllable node as depicted in (6b). Note that although I have represented the coda consonant in (6b) as moraic, a singleton coda consonant may be non-moraic on the surface (compare Hopi (C)VμCμ with Lenakel (C)VμC previously). If it is moraic, this is the result of the application of the Weight-by-Position rule which assigns moras on codas. Thus, monomoraic (C)V syllables are light, while bimoraic (C)VVs are heavy; (C)VC can be light or heavy on a language-specific basis. Moras are also grouped into feet (McCarthy and Prince 1986; Hayes 1995), which are part of higher prosodic structure that includes prosodic words (Selkirk 1980, 1984a; Nespor and Vogel 1986; Itô and Mester 1992). Reference to feet and moras allows us to account for numerous data, many of which cannot be adequately accounted for in other timing models. A concrete example of this sort emerges in the consideration of the seemingly problematic data from Japanese hypocoristic formation in (5) above. To account for the attested patterns, the X-slot model needs to impose both [XXX] and [XXXX] templates; worse still, the CV model needs to utilize [CVV] or [CVC] or [CVCV] templates depending on the nickname considered each time. Evidently, none of these approaches is insightful. The same is not true for the moraic framework, which can propose a uniform and straightforward account, namely that Japanese nicknames have the size of a single bimoraic foot and consequently can emerge as either heavy CVV/CVC monosyllables or light CV disyllables. A similarly neat explanation is available for the Ancient Greek compensatory lengthening (CL) data above. Given that CL is just about the preservation of the mora after the deletion of the segment that hosted it through lengthening of a neighbouring segment (Hayes 1989, but see Ch. 3), either V- or C-lengthening will do. Consequently, the preservation of the mora of the deleted s in esmi can yield either eemi or emmi leading to the attested dialectal variation. In sum, moraic theory has proven more successful compared to its predecessors, and because of this, it will be the timing model assumed in this work. 1.2.3

The traditional stance of moraic theory towards onsets, and its problems Moraic theory à la Hayes (1989) claims that one of its strong points is that it does not allow moraic onsets. Hayes argues that onset consonant deletion

Onsets and weight: the theory

7

never causes CL, which is only natural if onsets never bear weight. Notably, this effect cannot be as easily derived in the previous frameworks, whereby a timing slot is assigned to the onset, implying, at least in principle, that this constituent too may be active in prosodic processes. One way that has been utilized to avoid this superficially implausible result was to introduce the rimal node (cf. (4b)) as the only one that could bear weight. Moraic theory is advantageous in that respect too. Given that it only assigns moras to nuclei and codas, it can dispense with the rimal node. Despite its virtues, moraic theory bases its argument against the existence of weightful onsets on the fallacious argument of rarity; as Hayes (1995: 51) puts it: ‘Onset segments are prosodically inert . . . While this claim is not fully valid at the observational level, it is so well supported across languages that it serves as the central observation for formal theories of syllable weight’. Similar statements are made by other researchers too, who admit that claiming that onsets never play any role in weight is not entirely accurate. For example, Gordon (2006: 3) observes that ‘in Latin, as in virtually all languages, the onset is ignored for purposes of calculating weight’ (emphasis added mine). The bottom line seems to be that because the overwhelming majority of languages ignore onsets for prosodic processes or – to put it another way – because the prosody of languages so rarely pays attention to onsets, syllable-weight theories have so far ignored onsets by stipulating that they are prosodically inert. The present book instead challenges the ‘convenience’ of the traditional assumptions and takes the position that although it is true that there is a very strong tendency for onsets not to matter for weight purposes, this is by no means universal. A more complete understanding of syllable structure and weight thus forces us to include onsets in the syllable-weight equation. This then suggests a more literal interpretation of the term ‘syllable weight’, which up to now has basically corresponded to rimal weight. In fact, admitting the participation of onsets in prosodic phenomena seems to be the null hypothesis, rather than excluding them as the traditional theory advocates. I will consequently attempt to show that certain stress and syllable-weight facts cannot be re-analysed in any way other than by admitting weightful onsets, thus moraic onsets do exist in some languages and are represented in the way shown in (7b). Their introduction does not undermine moraic theory, but aspires to improve the range of facts that the theory can account for. Importantly, I am not suggesting that the presence of moraic onsets is unrestricted; rather, it is regulated by certain patterns pertaining to voicing (§1.3.3.1–3, §1.3.3.5) or

8

Onsets

underlying moraic specification (§1.3.4). In this view, moraic theory remains in an advantageous position, because even after the introduction of moraic onsets, it can still distinguish between languages that have them (7b) and ones that do not (7a) by simply assigning a mora on the onset of the former but not the latter. (7)

a. Non-moraic onsets σ µ C V

b. Moraic onsets σ µ

µ

C V

While the presence or lack of moraicity in moraic theory is built in (cf. the case of codas), other timing models do not have this option, stating that onsets should either be consistently weightful or weightless across languages. Neither situation reflects reality, however, as we will see in due course. To conclude, in principle, there is nothing wrong with having a moraic onset. Consequently, within the current proposal, onsets come in two flavours: nonmoraic (7a) – as in most languages – and moraic (7b). The latter’s distribution is systematic and restricted, as discussed in §1.3. An onset can still be seen as the tautosyllabic prenucleic consonant. 1.2.4 Onset weight: a brief overview of the empirical data To be able to follow the argumentation and justification of the theory that will be presented in the following sections, it is at this point important to consider briefly some of the data that provide the basis upon which it will be built. The subsequent chapters of course develop detailed case-studies of these and many more data where the full range of onset-weight effects are analysed extensively. 1.2.4.1 Onsets and stress Our attention will first be drawn to stress. The stress algorithm of a handful of languages is sensitive to the presence and/or to the quality of an onset. These two factors are independent of one another, as I will be arguing, so it is possible that they interact or act separately (8). In languages like Karo (❶) only the onset’s quality matters (QO systems); in others, e.g. Aranda (❷), only its presence does (PO systems); while in Pirahã (❸) both the presence and the quality of the onset are instrumental to stress assignment. More commonly, of course, neither of the two factors exerts any influence on stress assignment (❹). The data in (9) exemplify.

Onsets and weight: the theory (8)

(9)

9

Presence and quality of onset interaction in stress QO

PO

Languages

✓ ✕ ✓ ✕

✕ ✓ ✓ ✕

Karo Aranda, Banawá, Dutch Pirahã Greek, Russian, etc.

Pattern Identifier ❶ ❷ ❸ ❹

Presence and/or quality effect of onsets on stress ❶ Karo (stressed syllables in bold; Gabas 1998: 22, 1999: 39–41) a. cigi ‘spot’ pibεʔ ‘foot’ b. pakːɔ ‘fish (sp.)’ ‘fontanel’ nahek c. maŋgɔt ‘again’ kɨɾɨwεp ‘butterfly’ d. pε.ɔdn ‘skin’ e.i ‘irara’2 ❷ Aranda (accents indicate primary (acute) and secondary (grave) stress; Strehlow 1944; diacritics ignored) a. tárama ‘to laugh’ kútuŋùla ‘ceremonial assistant’ b. ankáta ‘Jew lizard’ ulámbulàmba ‘water-fowl’ ❸ Pirahã (stressed syllables in bold, acute accent = H tone, no accent = L tone; Everett and Everett 1984; Everett 1988) a. ko.ʔo.pa ‘stomach’ b. poo.gáí.hi.aí ‘banana’ c. ʔí.bo.gi ‘milk’ d. biísai ~ miísai ‘red’ ❹ Greek a. pérazma perúka ekpébo ekpobí b. étimos eláfi eðáfi éðrano

‘way-through’ ‘wig’ ‘transmit’ ‘show’ ‘ready-masc-sg’ ‘deer’ ‘land-pl’ ‘bench, desk’

In Karo, default stress is word final, unless some requirement, i.e. tone, nasalization or onset voicing, causes shift from that position. In particular, final 2

Gabas (1998) is written in Portuguese. I have not translated the glosses into English.

10

Onsets

voiced obstruent onsets repel stress (9❶a), whereas voiceless obstruent (9❶b) and sonorant (9❶c) onsets do not. Onsetless syllables are allowed (Gabas 1999: 24), and they can carry stress too if they make the best available stress bearers (9❶d). This suggests that onsetless syllables in Karo are not treated in any special way. In Aranda, on the other hand, stress on onsetless syllables is avoided, so that actually the first onsetful syllable receives stress irrespective of its type (compare onsetful (9❷a) with onsetless (9❷b)). In Pirahã (Everett and Everett 1984; Everett 1988), the rightmost heaviest syllable of the final three in a word receives stress according to the following hierarchy: PVV > BVV > VV > PV > BV (P = voiceless onset, B = voiced onset, > = is heavier than). Thus, stress is final if all syllables are of the same type (9❸a), but may appear elsewhere when syllables are different. In particular, onsetful syllables attract stress more than onsetless ones (9❸b, gáí > aí), but also onsets of a certain type, i.e. voiceless obstruents, attract stress more than the voiced obstruents and sonorants (9❸c, ʔí > bo, gi; 9❸d, sai > bií, mií). Finally, in languages such as Greek, the presence or quality of onsets plays absolutely no role in the stress algorithm. Syllables with onsets of any type may receive stress (9❹a) and onsetless syllables may carry stress too (9❹b). 1.2.4.2 Onsets and geminates Another situation where onsets are prosodically active is in languages such as Trukese and Pattani Malay that exhibit initial geminates (represented as CiCi or as Cː in the sources). Evidence for the contribution of onsets to weight comes from word-minimality effects (Trukese) and stress (Pattani Malay). More specifically, minimal words in Trukese are either CVV or CiCiV. The latter include a geminate. A straightforward account of this pattern is that the minimal word is bimoraic, thus implying that geminates must contribute a mora; given that the geminate can be plausibly syllabified in an onset position, it is reasonable to propose that Trukese allows moraic onsets. (10)

Trukese initial geminates and minimal words (Dyen 1965; Goodenough and Sugita 1980; Hart 1991; Davis 1999b) a. CVV words b. CiCiV words

maa oo tto kka

‘behaviour’ ‘omen’ ‘clam sp.’ ‘taro sp.’

Similar results obtain in Pattani Malay, which also has initial geminates as shown in (11b). Comparison of the minimal pairs in (11a) and (11b) reveals

Onsets and weight: the theory

11

that geminates have to be weight-contributing since they attract stress (11b). Again, an analysis that represents initial geminates as moraic onsets seems in order. (11)

Pattani Malay initial geminates and stress (Yupho 1989) a. buwɔ́h ɟal5 b. bːúwɔh ɟːálε

1.3

‘fruit’ ‘street/path’ ‘to bear fruit’ ‘to walk’

The theory of onset weight

1.3.1 Types of moraic onsets Having provided empirical evidence on the existence of onset weight, the question that now arises is: what kind of moraic onsets are there? I would like to propose that there are two types of moraic onsets: a) distinctive and b) coerced, inspired by Morén’s (2001) discussion on weight. Distinctive weight refers to phonemic weight distinctions; for instance, the difference between Hungarian vicε ‘janitor’ and vic.cε ‘his joke’ is one that will be represented in the input as /vicε/ vs. /vicμε/. Coerced weight refers to weight acquired in the output as a result of a requirement such as Word Minimality, Stress-to-Weight, or Weight-by-Position, etc. and where an input /C/ turns in the output into [Cμ]. I argue that such a distinction is not only applicable to codas, as Morén claims,3 but to onsets too. In other words, this grouping distinguishes between: a) singleton non-moraic onsets vs. singleton moraic onsets (coerced weight) and b) singleton non-moraic onsets vs. geminate moraic ones (distinctive weight). At this point, it is important to clarify certain terms.4 As stated, distinctive weight refers to underlying weight, whereas coerced weight refers to derived weight on the surface. Moraic onsets like the ones appearing in Karo or Pirahã are clearly coerced (9). These appear moraic, but without any evidence for underlying moraicity and without any surface contrast with their non-moraic counterparts (i.e. we don’t get both [p] and [pμ], but only the latter). Similarly, other consonants clearly bear distinctive weight; these must therefore be underlyingly moraic, i.e. geminates, since on the surface we find contrasts between 3

4

As a matter of fact, Morén’s distinctive–coerced contrast applies to both vowels and consonants, but he explicitly stipulates the absence of onset weight. If this stipulation is wrong, as I claim it is, then he too would be bound to expect a similar contrast in onsets. I am indebted to Bruce Morén (personal communication (p.c.)) for relevant discussion.

12

Onsets

them and their singleton counterparts, e.g. Trukese [təə] ‘islet’ vs. [tto] ‘clam’ (Davis 1999b). These two cases so far are the prototypical coerced and distinctive moraic onsets, respectively. This grouping is not always as clear-cut, though. Another type of moraic onset seems to exist that I will call here ‘geminated’.5 These are consonants whose weight appears intermediate between distinctive and coerced, as they present properties of both. Consider, for instance, heavy consonants whose weight emerges as the product of reduplication or prefixation, e.g. Marshallese reduplication /korap/ > [yokkoraprap] (§6.2.4.2). These are similar to geminates (distinctive weight) in that they present a contrast between non-moraic and moraic consonants in the same position, e.g. [ko.rap] vs. [yo.kko.rap.rap]; but they also resemble coerced moraic onsets, since this contrast does not seem to be underlying.6 In the schema below (12), straight, solid links between distinctive and true geminates, and coerced and non-geminated ones denote the prototypical cases of each category, as explained above. For the reasons mentioned before, ‘geminated’ consonants are somewhere in between, but in the absence of negative evidence, have a closer link to distinctive weight (indicated by the thicker solid line), than to coerced weight (indicated by the dotted line). Notably, the difference between ‘geminates’ and ‘geminated’ consonants seems to be irrelevant to the forthcoming discussion (see Ch. 5 and Ch. 6), so I will no longer employ it. Instead, I will uniformly term both geminates. There are two reasons for this: first, as noted above and in Chapter 5, the distinction is often difficult to determine or even undetectable; second, I will show that moraic onsets that are unambiguously distinctive or coerced largely participate in the same weight phenomena thus rendering such simplification justifiable.7

5

6

7

Note that the related term ‘geminate’ is somewhat confusing, as it is used in the literature to refer to several things. For instance, some researchers call sequences of identical consonants ‘geminates’ although they lack weight (Selkirk 1990; Tranel 1991), as in e.g. Malayalam. I will claim that these are not geminates, but simply doubled consonants with two root nodes (cf. §5.4.2.2 for fuller discussion). The latter conclusion is, however, not uncontested (also see fn. 11 in Chapter 5). If, for instance, prefixation or reduplication involve the presence of an underlyingly unlinked mora that as a result of the constraint ranking ends up creating a moraic onset, we could still possibly claim that this moraic onset has distinctive weight, i.e. is a geminate, because it has a mora whose source is in the input. If this proves correct, then such consonants would need to be considered distinctive, patterning exactly like the geminates. I say ‘largely’ because unambiguous coerced onsets cannot be assumed in the cases of geminates, since, by definition, the latter are distinctive.

Onsets and weight: the theory (12)

13

Moraic Onsets

Distinctive True Geminates Surface contrast between [C] & [Cµ] in same position that comes from the input, i.e. /C/ → [C] vs. /Cµ/ → [Cµ]

Coerced Geminated Surface contrast between [C] & [Cµ] in same position, but debatable if this comes from the input

Non-geminated Surface weight appears which must be induced by a constraint, i.e. generally /C/ → [Cµ], but due to onset markedness considerations, only a subset of consonants may present this pattern. The rest remain /C/ → [C]. See below for the importance of voicing for onsetmoraicity and sonority for codamoraicity

While the difference between ‘geminate’ and ‘geminated’ consonants will be disregarded, the difference in the nature of weight in the examination of prototypical cases of weight has repercussions on what kinds of moraic consonants can be found (Morén 2001). In distinctive weight there are no expectations as to what type of segments will show contrasts between singletons versus geminates. Since weight is lexically specified in the input, this will be unpredictable and consequently perhaps arbitrary. High-ranking moraic faithfulness – which I assume throughout – ensures that weight contrast of this type is preserved on the surface. But the same principle does not hold in coerced weight, which is moulded on the surface, i.e. where markedness constraints are applicable. For nucleus and coda weight, sonority considerations become relevant (Zec 1988, 1995; Morén 2001). Focusing on codas in particular, hierarchies referring to moraicity show that more sonorous segments are more likely to be moraic as well as to attract stress, leading to the conclusion that the more sonorous a coda segment is, the more likely it is for it to be moraic. In this light, a case like Kwakwala (Zec 1988, 1995), where the moraicity of codas in CVC syllables is variable depending on whether the codas are sonorous, is an instance of coerced weight. In particular, non-glottalized sonorant codas are moraic, whereas glottalized sonorants and obstruents are not. The prediction, then, is that in languages with coerced coda weight, sonority is crucial and thus it should not be possible for non-sonorous codas to be moraic with more sonorous ones being non-moraic. In languages with distinctive weight, however, no similar restrictions apply. Lexical specification occurs and therefore sonority reversals with respect to

14

Onsets

moraicity are admitted. For instance, in languages that only allow one type of coda-onset geminate, this is commonly the less sonorous voiceless rather than the more sonorous voiced one, e.g. Lak, Nez Perce, Ocaina, Ojibwa, Totonac, Yakut and Japanese (Morén 2001). If this idea is on the right track, then we would expect similar effects in onset moraicity too. Languages like Trukese, Pattani Malay, Marshallese and Puluwat have geminates that can be analysed as distinctive moraic onsets. In most of these cases, all consonants are able to appear as singletons or geminates with the possible exception of Pattani Malay (§5.2.1) and Puluwat (§5.3.4), where a subset of the consonants cannot geminate, namely /ɣ/ in Pattani Malay and /w, y/ in Puluwat. But this is exactly what is anticipated if geminates are lexically specified as such. These languages will be dealt with in more detail in Chapters 5 and 6. As for coerced moraic onsets, we would expect them to adhere to certain markedness considerations, which I will argue relate to voicing rather than directly to sonority. Coerced moraic onsets emerge in the stress systems of Pirahã, Arabela and Karo (§1.2.4, Ch. 2) or the minimality facts of Bella Coola ((32) here and Ch. 4) and are enforced by the generic constraint Moraic Onset (cf. (30) §1.3.3.5), a handy tag for various constraints, such as Weightby-Position with regard to onsets, Word Minimality, etc. Summing up, then, I argue that the contrast between distinctive and coerced weight is applicable in onset weight too. The profile of coerced moraic onsets will be identified in §1.3.2 and §1.3.3.5 and further explored in Chapters 2 and 4 with reference to stress and word minimality. Discussion on distinctive weight follows in §1.3.4 and Chapter 5. 1.3.2 Coerced moraic onsets Previously we saw that the preferred moraic codas are those that are more sonorous. This is by no means a coincidence, since it relates to the issue of what optimal nuclei and margins look like. Prince and Smolensky (1993/2004) argue that optimal peaks are high in sonority, whereas optimal margins are low in sonority. This is the result of the harmonic peak and margin hierarchies in (13). (13)

Peaksyll

a > e,o > i,u > . . . > p,t,k

Marginsyll

p,t,k > . . . > i,u > e,o > a

Technically, these can be translated into the following constraints (for the harmonic alignment schema, see Prince and Smolensky 1993: 136 and McCarthy 2002: 20–2):

Onsets and weight: the theory (14)

15

Peak prominence: *Peak/p,t,k >> . . . >> *Peak /i,u >> *Peak /e,o >> *Peak /a Margin prominence: *Mar/a >> *Mar/e,o >> *Mar/i,u >> . . . >> *Mar/p,t,k

Note, however, that these hierarchies and fixed rankings are not exactly accurate, since codas actually prefer to be quite sonorous and at any rate much more sonorous than optimal onsets (Murray and Vennemann 1983; Vennemann 1988; Zec 1988, 1995; Clements 1990; Smith 2003). Bearing this in mind, the above hierarchy needs to be refined as below (15):8 (15)

Syllable-constituent well-formedness Nucleus prominence: *Peak/p,t,k >> . . . >> *Peak/i,u >> *Peak/e,o >> *Peak/a Coda prominence: *Coda/p,t,k >> . . . >> *Coda/i,u >> *Coda/e,o >> *Coda/a Onset prominence: *Ons/a >> *Ons/e,o >> *Ons/i,u >> . . . >> *Ons/p,t,k

This correctly generates the fact that optimal nuclei and codas are high in sonority, whereas optimal onsets are low in sonority. Moreover, observe that preferred moraic codas, and – of course – nuclei, are those that are high in sonority (Zec 1988, 1995; Yip 1992). One would perhaps expect, then, that the more sonorous a segment is, the more likely it is to be moraic. While this works fine for codas and nuclei, it does not account equally well for the newly introduced moraic onsets. The problem is that in onsets, as the empirical data suggest (see 9❸), voiceless segments are preferably moraic as opposed to their voiced counterparts, although the former are less sonorous than the latter. But this point only proves problematic if one maintains the connection between moraicity and sonority for onset weight too. I instead propose that while syllable well-formedness is down to sonority considerations for all constituents, the same does not hold for weight. For the latter, I suggest that a specific aspect or component of sonority, namely voicing, and not sonority itself, takes priority. For the time being, however, I will say that with regard to syllable margins, weight markedness relies on voicing for onsets and on sonority for codas, as indicated in the fixed rankings below. 8

Of course, the combined interaction of syllabification considerations, the requirements on what can constitute an onset or a coda as shown in (15) and the high-ranked constraint Nuc which demands that syllables have nuclei ensures that the more sonorous segments syllabify as nuclei and not as codas (e.g. usually only high vowels can be glides and syllabified in margins, but nothing more sonorous than that). In other words, although low vowels are both the best possible nuclei and codas according to (15), we do not predict that low vowels will be syllabified in a coda. This is simply because the coda and peak hierarchies interact with one another and the requirement to have a nucleus will give priority in picking out the low vowel as a nucleus and not as a coda (for a similar argument, see Baertsch and Davis 2003).

16 (16)

Onsets Moraic markedness for margins Onset moraicity: *μ/Ons/[+voi] >> *μ/Ons *μ/Ons: Moraic onsets of no type are allowed (Muller 1999; Mellander 2003a; §5.4.1) Coda moraicity: *μ/Coda/[−son] >> *μ/Coda *μ/Coda: Moraic codas are banned (adapted from Rosenthall and van der Hulst 1999: 503)

Three-predicate constraints similar to the ones here have been used in other works too, e.g. de Lacy (2000). Thus, a constraint such as *μ/Ons/[+voi] states that a [+voi] onset makes the worst (indicated by * and its top ranking) moraic (μ) onset (Ons) possible. Notice, however, that these fixed rankings merely tell us that certain onsets (or codas) make worse moraic constituents than others. They do not impose moraicity of any sort. Anticipating the discussion that follows in §1.3.3.5, I claim that the constraint that does that is Be Moraic, which interleaved among the three positions available, as shown below, generates an equal number of different patterns, i.e. where all onsets/codas are moraic (17.1), some are (17.2) and none are (17.3). (17)

Moraic markedness for margins and BE MORAIC Onset moraicity: (1) Coda moraicity:

*µ/ONS/[+voi] >> *µ/ONS (2) BE MORAIC

(3)

*µ/CODA/[–son] >> *µ/CODA

1.3.3 The effect of voicing In this section, I first justify the preponderance of voicing for weight by correlating it with pitch, stress and tone in §1.3.3.1 and §1.3.3.3. I then explain (§1.3.3.4) why voicing effects on weight ultimately appear on onsets only and not on other constituents (nuclei, codas), where sonority is instead chosen. 1.3.3.1 Voice and tone patterns It is a well-known fact that voiceless onsets usually raise the pitch of the following vowel, whereas voiced ones lower it (Haudricourt 1954; Hombert 1978; Kingston and Solnit 1988a,b; Peng 1992; Bradshaw 1999; Yip 2002; Tang 2008). In voiceless obstruents, the cricothyroid muscle stretches the vocal folds to obstruct vocal fold vibration resulting in vocal fold tensing, which in turn leads to higher F0. In voiced obstruents, the larynx and hyoid bone are lower and a lowered larynx results in lower F0 (Yip 2002: 6–7; Honda 2004).

Onsets and weight: the theory

17

In fact, depression of F0 after voiced stops is very likely universal as Kingston and Solnit (1988b) state. During the course of history, numerous languages have lost the voicing contrast and have reinterpreted it in terms of tones. A common scenario for such a language is to now only have voiceless onsets and a contrast between L and H tones. L tones emerge where there used to be a preceding voiced onset, whereas H tones occur where there was originally also a voiceless onset. Some languages present both effects at once: Bade, Jingpho, certain Chinese dialects, Masa, Yabem, etc. (see Tang 2008 and references therein). In fact, this ‘evolutionary’ path can be synchronically traced among dialects of the same language. Yip (2002: 35, citing Svantesson 1983) notes that in the southern dialects of Kammu there is still a voicing contrast in onset position for the obstruents, but this no longer exists in the northern dialects. There, an H tone appears in positions that correspond to voiceless obstruents and an L tone in positions that correspond to voiced obstruents in the southern dialects. This pattern is illustrated below. (18)

Kammu dialects South (voicing contrast–no tone) North (no voicing contrast–tone contrast) klaaŋ kláaŋ ‘eagle’ glaaŋ klàaŋ ‘stone’

Other languages still maintain the voicing contrast, but consistently show lower pitch after a voiced obstruent and higher pitch after a voiceless obstruent. One example of this type is in Songjiang, a Wu dialect of Chinese (examples again from Yip 2002: 7). In (19), 5 denotes the highest pitch, 1 the lowest. The first digit marks the pitch at the beginning of the syllable, the second marks the one at the end of the syllable. (19)

Songjiang tones ti ti ti

53 44 35

‘low’ ‘bottom’ ‘emperor’

di di di

31 22 13

‘lift’ ‘younger brother’ ‘field’

In her survey, Bradshaw (1999) attempts to explain facts of this sort by using a single unary feature that represents both [voi] and L tone. Crucially for her, [voi] is the only feature allowed to interact with tone, but this is problematic. First, there is evidence that aspiration and glottalization may also affect tone (Yip 2002; Tang 2008), which suggests that voice and tone cannot be identified as a single feature. Second, and more important for our purposes, unary [voi] fails to capture a fuller range of facts. More specifically, while voiced obstruents with unary [voi] are anticipated to interact with tone, voiceless ones are not, since they lack the feature altogether. Thus, finding cases where

18

Onsets

voiceless obstruents clearly interact with tone presents difficulties for a unary approach and highlights the need for binary [+/− voi]. Tang (2008: 3, based on Schuh 2002) presents data that fit this profile. Bade has a process of rightward high-tone spreading across word boundaries (20a–b), blocked by non-implosive voiced obstruents (20c–d). This is anticipated under unary [voi]. (20)

High-tone spreading in Bade and its blocking by voiced obstruents: /H#LH/ → [H#H!H] a. b. c. d.

/nә́ tә̀nkә̀kú/ /nә́ ɗùwàɬú/ /nә́ tә̀mbә̀lú/ /nә́ bàzàrtú/

[nә́ tә́nkә́kú] [nә́ ɗúwáɬú] [nә́ tә́mbә̀lú] [nә́ bàzàrtú]

‘I pressed’ ‘I got tired’ ‘I pushed’ ‘I shamed’

*tә́mbә́lú *bázártú

Surprisingly, though, voiceless obstruents also interact with tone (21c–d), because they block another process in the language, namely L-tone spreading (21a–b). L tone spreads following an H-toned syllable across a clitic boundary or within the same morpheme. The spreading occurs only if the target syllable also precedes another H tone (consider kóːrón and the verbs before it). This situation receives an obvious account in a binary [+/− voi] approach, where voiced obstruents and voiceless ones are equipped with a value of binary [voi] and consequently may interact with tone, but is less straightforward in a unary [voi] approach. (21)

Low-tone spreading in Bade and its blocking by voiceless obstruents a. b. c. d.

/dʒә̀ɗgә́ kóːrón/ /dʒә̀ kә̀rә́ kóːrón/ /dʒә̀ dә̀psә́ kóːrón/ /dʒә̀ gàfá kóːrón/

[dʒә̀ɗgә̀ kóːrón] [dʒә̀ kә̀rә̀ kóːrón] [dʒә̀ dә̀psә́ kóːrón] [dʒә̀ gàfá kóːrón]

‘we followed a donkey’ ‘we stole a donkey’ ‘we hid a donkey’ ‘we caught a donkey’

Shifting our attention to sonorants now, the picture gets more complicated. The empirical facts indicate that sonorants present variable behaviour. As Bradshaw (1999) reports, in some cases, they pattern with the voiced obstruents (Nupe, Ngizim), and in others with the voiceless obstruents (Suma, Siswati, Yaka, Miya, Digo, Bade). As an example, in Nupe the L tone of a nominal prefix or tense/aspect marker spreads onto the root vowel provided the first root consonant is not a voiceless obstruent (22c). Sonorants (22b) and voiced obstruents (22a) are transparent for L-tone spreading: (22)

Nupe L-spread: sonorants and voiced obstruents pattern together (George 1970) a. Voiced obstruents /è+dú/ → [èdǔ] /gí/ → [à gǐ]

‘taxes’ ‘will eat’

Onsets and weight: the theory b. Sonorants /è+lé/ → [èlě] /lá/ → [èl ]

19

‘past’ ‘is carrying’

c. Voiceless obstruents /è+tú/ → [ètú] ‘parasite’ /tí/ → [ètí] ‘is hooting’

On the other hand, in Bade, as seen before, H-tone spreading across words changes underlying /HLH/ to [HH!H]. Spreading occurs if the intervening consonant is a voiceless obstruent (23a) or a sonorant (23b), but not if it is a voiced obstruent (23c). (23)

Bade H-spread: sonorants and voiceless obstruents pattern together (Schuh 1978) a. nә́n kàtáw → nә́n ká!táw b. nә́n làwáw → nә́n lá!wáw c. nә́n gàfáw → nә́n gàfáw

* nә́n gá!fáw

‘I returned’ ‘I ran’ ‘I caught’

Such variability may have both a phonetic and a phonological explanation. Phonetically, unlike obstruents, sonorants do not automatically perturb the F0 of adjacent vowels, thus ‘there is no phonetic reason to expect that the fundamental frequency of the following vowel will be either elevated or depressed by the sonorant’s laryngeal articulation; instead the speaker may be free to choose either elevation or depression of fundamental frequency’ (Kingston and Solnit 1988a: 276). Phonologically, my claim goes as follows. The major class feature of sonorants is [+son]. This accounts for the spontaneous phonetic voicing (see Rice and Avery 1989) that is so characteristic of sonorants (Ohala 1983). This is, however, different from their phonological behaviour with respect to the binary [+/−voi] feature. Allowing sonorants to be phonologically either [+voi] or entirely unspecified for this feature, their dual behaviour presented above is entirely anticipated. In particular, consider Nupe above, where sonorants behave like the voiced obstruents in admitting L-tone spreading through them. This will happen regardless of whether they are [+voi] or unspecified. Assuming that this spreading is only blocked by [−voi], i.e. by voiceless obstruents, then the facts in (22) fall out naturally. In Bade, on the other hand, it must be the case that voiced obstruents are [+voi], their voiceless counterparts [−voi] and sonorants unspecified. H-tone spreading (20) occurs over voiceless obstruents, but is blocked by [+voi] obstruents. L-tone spreading (21), in contrast, occurs over voiced obstruents, but is blocked by [−voi] obstruents, indicating the necessity of using binary features, since both are actively involved in blocking spreading

20

Onsets

processes. Interestingly, sonorants pattern with the voiceless obstruents (23) in also admitting H-tone spreading over them. This suggests that sonorants do not bear [+voi] which otherwise blocks such a process (cf. 20). But if sonorants are unspecified for [voi], then they should also be transparent to L-tone spreading (21), since this is only blocked by [−voi] obstruents. Indeed, this is the case, as the examples below in (24) confirm. (24)

Bade sonorants do not block L-tone spreading (Schuh 2002) a. /dʒә̀ kә̀rә́ kóːrón/ b. /kàyáːn pә́m/

[dʒә̀ kә̀rә̀ kóːrón] [kàyàːn pә́m]

‘we stole a donkey’ [cf. (21b)] ‘not a squirrel’

At this point, the cautious reader may raise the following point: if Nupe works equally well whether sonorants are [+voi] or unspecified, and if in Bade they must be unspecified, then would it not perhaps be more economical to leave them unspecified, as Tang (2008) suggests? This is a valid point, but I will show that it is an over-simplification. The reason sonorants in Nupe are compatible with either a [+voi] or no voice specification analysis is because they appear as transparent to a pre-existing L-tone that spreads. To see their behaviour clearly, we need an instance whereby sonorants do not behave passively, but instead actively induce a process. A suitable case appears in Kotoko (Odden 2007), a language that presents three tones, High, Low and Mid (H = ,́ L = ,̀ Mid = no mark), and where sonorants – like voiced obstruents – function as depressor consonants in the absence of any pre-existing associated tone. One representative set of examples is offered by the future tense which segmentally takes the form [npfx–stem–àsfx]. This tense also comes with an underlying H tone. Importantly, if the stem begins with a voiceless obstruent, then the H tone surfaces (25a). But if it begins with a voiced obstruent (25b) or a sonorant (25c) then the tone lowers into M. Since sonorants here exhibit an overt lowering effect, it must be the case that they are [+voi]. Additional evidence for the sonorants’ [+voi] specification will be provided later in this book in the discussion on Pirahã (§2.4.2) and Arabela (§2.5.1). (25)

Kotoko sonorants must be [+voice] a. Voiceless obstruents take H tone on first syllable n-sáp-à m-páy-à

‘chase’ ‘bury’

n-sáb-à n-cә́nh-à

‘grow’ ‘be sated’

b. Voiced obstruents take M-tone on first syllable (depressor consonants) n-zəgl-à m-ban-à

‘carry’ ‘bathe’

n-dunkw-à ‘throw’ n-gulan-à ‘laugh’

c. Sonorants take M-tone on first syllable (depressor consonants) n-wac-à m-mar-à

‘leave’ ‘die’

n-yey-à n-law-à

‘call’ ‘fight’

Onsets and weight: the theory

21

Consequently, sonorants are of two types, as presented below (26). (26)

Types of sonorants a. [+son]

Lar | | [+voi]

b.

[+son]

Lar

a: sonorants as in Kotoko (and possibly Nupe, Ngizim) b: sonorants as in Bade, Suma, Siswati, Yaka, Miya, Digo

1.3.3.2 Justifying binary features Before moving on, I need to justify the use of a binary-feature approach over a unary one. Some of the argumentation has been mentioned above in passing, but as this is an important feature of the analysis, I will briefly repeat it for clarity. Given the specification of sonorants as [+voi] or nothing and the specification of voiced obstruents as [+voi], one might perhaps be tempted to employ a unary approach by getting rid of [−voi] and considering voiceless obstruents as lacking a voice-feature specification altogether. This would then transform the moraic markedness hierarchy of onsets in (16) to: *μ/Ons/[voi] >> *μ/ Ons. But excising [−voi] seems impossible; recall that in Bade (see (21)), it was imperative to block L-tone spreading. Nonetheless, Marc van Oostendorp (personal communication (p.c.)) proposes a possible way out, assuming the following constraints: a) Align-T, which induces tone spreading; b) L ⊃ [voice] (Hansson 2004), which says that L tone implies [voice]; c) Ident−[voice]; and d) NoGap, which militates against non-local spreading. Suppose further that the latter three constraints dominate the first one, so that: L ⊃ [voice], Ident−[voice], NoGap >> Align-T. Since L tones cannot occur on voiceless consonants (L ⊃ [voice]), which also cannot change into voiced (because of Ident−[voice]), and since consonants cannot be skipped (due to NoGap), the only alternative is for spreading to stop at voiceless consonants. Thus, technically, a unary-based solution is available for the problem in Bade. Besides likely difficulties in extending this proposal to other similar cases and the repercussions of L ⊃ [voice] more generally, a more serious problem is the following: if we adopt monovalent features for [voice] we would plausibly want to do that for [sonorant] too. But then, how could we express the coda moraic markedness hierarchy (16), since in codas, it is the [−son] sounds that tend to be non-moraic? A unary approach would restate the coda markedness hierarchy as: *μ/Coda/[son] >> *μ/Coda, but this would preferentially ban moras from sonorous codas, instead of the non-sonorous ones as desired.

22

Onsets

Of course, it could be argued that rather than [son], we could use [obstruent], i.e. *μ/Coda/[obstr] >> *μ/Coda, but this complicates the analysis unnecessarily. Briefly, we know independently that [son] is, among others, indispensable for syllable well-formedness, tone and stress, hence use of [obstruent] does not make our system more restricted either technically or empirically. Consider, for instance, Somali and Dakota voiceless stops which become voiced in codas, a fact that de Lacy (2006) convincingly attributes to a sonority increase that is preferable in that position. Given that voiced stops are more sonorous than their voiceless counterparts (see (15) and Topintzi 2006b, §1.4.1.1 for discussion), this pattern can be accounted for by a [son]-based account, but not by an [obstr]-based one. From now on, I will thus assume use of binary [+/−voi] and [+/−son]. 1.3.3.3 Interaction between voicing and stress The present section develops a fundamental idea in the book, namely that pitch perturbation due to voicing is not systematically phonologized as tone across languages. Instead, in some languages, it is phonologized as stress by means of weight. To begin with, consider the connection between tone and stress, as this proves instrumental for my purposes. More specifically, in some languages, H-toned moras attract stress (de Lacy 1999, 2002). One example of this sort is Golin, where stress falls on the rightmost H-toned syllable. In the absence of H-toned syllables, stress is word final. (27)

Golin H tones and stress (Bunn and Bunn 1970; stress marked in bold, accents indicate tone) LLL HLL LHL LLH HHL HLH

kàw.lì.gì á.kò.là gò.má.gì ò.nì.bá sí.bá.gì én.dè.rín

‘post’ ‘wild fig tree’ ‘type of sweet potato’ ‘snake’ ‘sweet potato type’ ‘fire’

These facts lead us to the following reasoning: we have seen that H-toned moras, i.e. moras with high pitch perturbation, may attract stress. We have also seen that one source of high pitch perturbation is voicing – in fact, lack thereof. And we also know that pitch is one of the phonetic correlates of stress. What we predict, then, is that in some languages pitch perturbation due to voicing is phonologized as stress (rather than tone) in terms of moras.9 In other words, 9

Of course, other laryngeal properties may also be relevant for stress purposes. These include, for instance, stronger aspiration in stressed syllables as in English, German, Farsi and Maori (González

Onsets and weight: the theory

23

it should be possible to have languages whose high pitch perturbation [−voi] onsets contribute to weight, as opposed to their [+voi] counterparts. To put it even more plainly, the claim is that voiceless obstruent onsets are more likely to add a mora to their syllable than the voiced counterparts, so that PV > BV [P= −voi, B= +voi]. The most straightforward area to test this is through the examination of stress patterns. Given that heavy syllables are frequently the preferred stress bearers (according to the Weight-to-Stress Principle, WSP), we would expect to see systems where PV syllables are preferentially stressed over BV ones. This obviously fits the profile of the systems ❶ and ❸ mentioned previously in (9). This task will be taken up in detail in Chapter 2. The relevant facts are repeated here for ease of reference. (28)

PV syllables attract stress more than BV ones (heavy syllables in bold) ❶ Karo stress is final unless [. . . PV.BV#] sequences are involved. In this case, it is penult: cigi ‘spot’ pibεʔ ‘foot’ ❸ Pirahã’s stress algorithm imposes stress on the rightmost heaviest syllable. Assuming PV > BV, we get facts like the one below: ʔí.bo.gi ‘milk’

1.3.3.4 Why voice and not sonority? One question to consider at this point is why voicing should be used for the onset-weight contrast rather than the more familiar and more general sonority that is used in nuclei and coda weight contrasts. Posing this question in such a way is of course rather biased, in the sense that the underlying assumption is that the sonority-based weight contrast is the default, while the voicing-based one is an abnormality that requires explanation. But this is only one way of seeing things. Another is to start with the reverse assumption – a view I am presently endorsing – namely that a voicing-based weight contrast is the default, but when the conditions for its application are weakened or even disappear, we turn to the more general sonority-based weight contrast. Voicing and sonority are after all related. Although sonority scales are notoriously varied (see Parker 2002 for a good overview) and there have been 2003) and glottalization (usually post-glottalization) before stressed syllables in languages like Coast Tshimshian, Gitksan, Saanich and Lilloet (González 2003). Notably, however, these patterns only exhibit laryngeal manipulations when found in stressed syllables. What we would like to see for our purposes is cases where aspiration or glottalization attracts stress in analogy to the voicing effect found in Pirahã or Karo. I have been unable to find examples of this sort.

24

Onsets

suggestions that voicing should not even be included in them (see Clements 1990), there is evidence to suggest that voicing is a factor in the sonority hierarchy. For instance, de Lacy (2006: 123, 171–2 and references therein) presents data from Somali and Dakota, where he argues that a voicing process occurs to increase sonority (also see above), while Parker (2002) makes a similar point by making reference to Imdlawn Tashlhiyt Berber and Kolokuma Ijo (Williamson 1965). In fact, lack of consensus is less about whether voicing should be included in the sonority hierarchy and more about whether any further subdivision exists based on the continuancy or not of consonants; in particular, while everyone agrees that among obstruents, voiced fricatives are the most sonorous and voiceless stops the least sonorous, there is debate as to whether voiced stops are more, less or equally sonorous to voiceless fricatives. All three possibilities have been considered in the literature, as Parker (2002: 68–9) observes. The instrumental data he presents, however, support the proposal that the whole voiced obstruent series is more sonorous than the whole voiceless counterpart, a conclusion that is in full accordance with the data from Pirahã, Karo and Arabela. But the relationship between sonority and voicing does not explain why the latter is prioritized over the former with regard to weight contrasts. The main reason, I speculate, is economy. It is long established that pitch perturbation due to voicing is used for tonal purposes. The point to make here is that this phonetic cue is constantly available and universal (see Kingston and Solnit 1988b on their discussion of pitch depression after voiced stops), so when it is not used for tonal purposes, it may be employed for another phenomenon it can associate with, namely stress, by means of moras. Sonority cannot offer such an effect – recall that sonorants do not cause any inherent perturbation – so its relationship with weight (and consequently stress) in onsets would be random and coincidental. So far so good, but now another issue arises. If voicing is rather the default factor for weight contrasts, then why does it not extend to other syllable constituents? Why is it that nucleus and coda weight relies on general sonority and not on specific voicing? If economy was the answer previously, now it is unreliability. What I mean by that is that when considering nuclei, voicing is rarely contrastive. As Gordon’s (1998b) typological survey reveals, although there are several languages that present voiced and voiceless vowels phonetically, in less than a handful is the contrast phonologically underlying. He mentions Sedang, Gujarati and Jalapa Mazatec. While in my view rare cases are often extremely insightful, it is doubtful whether a weight-based contrast would rely on a practically phonologically non-contrastive property, i.e. nucleus voicing.

Onsets and weight: the theory

25

A somewhat similar reasoning applies to codas. Unlike onsets, where the voicing contrast tends to be enhanced, in codas, the same contrast tends to be neutralized, as the numerous cases of final devoicing attest, e.g. German, Dutch, Catalan, Russian and others (Lombardi 1999; Wetzels and Mascaró 2001). Again then, using voicing as a cue to coda weight would be possible, but frequently non-existent, and thus on the whole unreliable. The discussion in this section has so far focused on justifying the superiority of a voice-based onset-weight analysis over a sonority-based one on the grounds of economy and reliability of contrast. But this can also be demonstrated by the insufficiency of a sonority-based account in terms of its predictions or erroneous conclusions. First, consider the sonority hierarchy again, this time in relation to manner distinctions. According to many theorists (see also above and Parker 2002), within the category of obstruents, fricatives are more sonorous than stops. If this is indeed the case, then we would expect to have sonority-increasing processes in positions where sonority must be greater. Indeed, all kinds of lenition processes (e.g. voicing, spirantization, l-vocalization, etc.) achieve this (Lavoie 2001), but the one which is of interest to us here is spirantization. The reasoning is the following. Continuancy distinctions are important for sonority purposes. If it were true that the onset-weight contrast were due to sonority considerations (rather than voicing), we could perhaps expect languages where e.g. stops contribute onset moras, whereas fricatives do not. To my knowledge, cases of this sort do not exist. On the other hand, in coda position – where sonority undoubtedly matters for weight purposes – one can find instances where fricatives contribute weight, but stops do not. For instance, in Oowekyala (Howe 2001), non-moraic stops spirantize in coda position in order to become moraic, e.g. [ləka] ‘to play the stone throwing game’ vs. [ləxsut] ‘to peck a hole through stone’. Importantly, fricatives also occur in onsets, e.g. [wəxa] ‘to split’, [casa] ‘to pour water on’ (Howe 2000). Note that the lack of attention to manner distinctions of consonants in the onset position (as far as stress and weight are concerned) also finds a parallel with the voicing and tone relationship, where manner again seems to be irrelevant. A second, and perhaps more problematic, aspect of a sonority-based account is that it is not able to capture the duality in the behaviour of sonorants. This is because if [+/−son] is the relevant feature, then sonorants should be treated uniformly. They would either need to behave like the voiced or like the voiceless obstruents, but only systematically so across languages. The empirical facts, however, falsify this, as we have seen previously (cf. Pirahã, Kotoko

26

Onsets

vs. Karo, Bade). Utilizing [+/−voice] and treating sonorants as either [+voi] or unspecified produces the correct result. All in all, a voicing-based onset-weight account is more successful than a sonority-based one in that it is both conceptually and empirically accurate, plus technically simple and viable. To express all that has been mentioned above in Optimality Theoretic terms, we only need to utilize the ranking in (29), which explicitly states that [+voi] onsets are always the worst moraic onsets, and implicitly [−voi] ones are the best ones.10 (29)

Onset moraic markedness hierarchy *μ/Ons/[+voi] >> *μ/Ons

1.3.3.5 Patterns of coerced moraic-onset weight The ranking in (29) informs us about the preferable moraic onsets, i.e. the [−voi] ones, but on its own has nothing to say about cases where onsets are indeed moraic, simply because it bans onset moraicity altogether (cf. discussion with regard to (16)). For this reason, we need another markedness constraint that imposes moraicity. The use of markedness rather than faithfulness bears on the issue that faithfulness would need to make crucial reference to the moraicity of the input. This would be undesirable for two reasons. First, we would need to restrict inputs in particular ways against ‘Richness of the Base’ (see Prince and Smolensky 1993/2004), but this could also potentially predict different outputs for the same input depending on whether underlying moraicity is assumed or not (an illustration of this problem appears when considering compensatory lengthening in §3.2.2.2). Moreover, here we are talking about coerced weight, i.e. weight that is imposed in the output, therefore recourse to faithfulness would not help us achieve the right results. For a similar argument, see Morén (2001). Morén argues that coerced weight is achieved through the force of a higherranked constraint that requires moraicity of segments on the surface, such as F(oo)tBin(arity), W(or)d Min(imality) or W(eight)byP(osition). Such constraints are part of the cover constraint Be Moraic that Morén informally uses to assign moraicity on codas. Looking at WbyP in particular, this is a constraint that is generally well accepted, despite the lack of good justification for its existence (other than its practical convenience). In effect, it imposes 10

Note that I have made use here of Positional Markedness (e.g. Steriade 1997; Zoll 1998; Walker 2001), i.e. of a context-specific markedness constraint, namely *μ/Ons/[+voi], as well as of general markedness. Such constraints are justified in Zoll (1998). An alternative formulation such as *μ/Ons/[+voi] >> *μ/Ons/[−voi] is non-viable, since it would have nothing to say about sonorants in languages where they are unspecified for voice.

Onsets and weight: the theory

27

moraicity on codas and is really an alias for Moraic Coda which has been used in some works such as Broselow, Chen and Huffman (1997). In principle, however, there is no reason why a similar constraint could not be applicable for onsets. In fact, it is merely a stipulation that it applies on codas only. It is thus reasonable to extend such a constraint to onset segments too by means of (30). (30)

Moraic Onset: Onsets are moraic (see Crosswhite 2001)

Following Morén, I too will use the cover constraint Be Moraic as the constraint which assigns moraicity on segments (unless it is required to be more precise). Now, if this constraint is interleaved among the constraints in (29), i.e. *μ/Ons/[+voi] >> *μ/Ons, we generate three possible coerced onset-weight patterns. (31)

Patterns of coerced onset weight a. *μ/Ons/[+voi] >> *μ/Ons >> Be Moraic b. *μ/Ons/[+voi] >> Be Moraic >> *μ/Ons c. Be Moraic >> *μ/Ons/[+voi] >> *μ/Ons

(31a) describes the great majority of languages where no moraic onsets whatsoever are allowed. (31b) refers to instances where only onsets lacking [+voi] act as moraic. These will consistently be the [−voi] ones. Depending on the sonorants’ [voi] specification, this also captures their dual behaviour: when they are [+voi], we get languages such as Pirahã, where only voiceless obstruents are moraic; when they lack a voice specification, as is the case in Karo, sonorants pattern with the [−voi] obstruents in being moraic, since they escape being penalized by *μ/Ons/[+voi] and consequently are allowed to be moraic too. Abstracting away from sonorants, the ranking in (31b) suggests that the preferred heavier onsets are the voiceless rather than the voiced ones. This is entirely compatible with independent empirical support. First, there is a tendency of voiceless obstruents to be longer than voiced ones (Ohala 1983, 1997; Maddieson 1997).11 This property is phonologized in Swiss German (Ham 2001 for Bernese in particular; Kraehenmann 2001) and appears as an underlying phonemic contrast. More generally, it is the case that in languages that allow only one type of geminate, these are commonly voiceless rather than voiced, e.g. Lak, Nez Perce, Ocaina, Ojibwa, Totonac, Yakut and Japanese (Morén 2001). 11

As far as Pirahã goes, experimental work by K. Everett (1998) shows that onset consonants in stressed syllables are longer than those in stressless syllables, but there is no discussion about the voicing properties involved, and whether these have any effect.

28

Onsets

Consonant gradation in Finnish and related languages can be understood along these lines too (Anttila 1994; Gordon 1998a). In Finnish, double stops degeminate and simple stops weaken before a branching rime, so that we get alternations of the type tt > t and t > d, e.g. takka > takan ‘fireplace-gen’, but takkana / *takana ‘fireplace-ess’ as there is no branching rime there, and similarly suti > sudin ‘brush-gen’, but sutina / *sudina ‘brush-ess’. Also in Yolngu Djapu, Morphy (1983) mentions that the contrast between voiceless and voiced stops has also been described as one between geminate and simple stops, respectively, so /b/ actually corresponds to p and /p/ to pp. In the Cuna ‘talking backwards’ language game (Sherzer 1970), the first syllable of the word moves to the end. The inversion is straightforward when the consonants involved include voiced stops, e.g. dage → geda, obsa → saob, but when voiceless stops are encountered, then the inversion takes a different shape. Thus, for the regular form sapan we do not get *pansa as anticipated, but instead bansab, and similarly sate → desad and not *tesa. This is very similar to Djapu above. If voiced stops actually act as singletons, while the voiceless ones act as geminates, then this is exactly the pattern we would expect to find. The last pattern (31c) refers to languages where all onsets are moraic independently of their type. Crucially here, voiced and voiceless obstruents do behave in the same way, but only because sonorants show the same behaviour too. Bella Coola Word Minimality offers a potential case for such a pattern. (32)

Bella Coola Word Minimality (Bagemihl 1998) Word Shape a. V b. VV c. VC d. CV

Examples * ya ‘good’ n̩ ƛ’ ‘dark, night’ ƛ’i ‘fast’

As we will see in Chapter 4, my claim is that in this language, moraic onsets arise only in [CV] words, so that the bimoraic word minimum is satisfied. However, this may be equally satisfied by all types of onsets, e.g. [ƛ’i] ‘fast’, [xwl̩] ‘to pull’, [wl̩] ‘to spill’, [mu] ‘fishtrap’ (Nater 1979: 170, 171, 175), suggesting that sonorant onsets too are moraic. Given the claim that sonorants are either unspecified or specified as [+voi], to be able to argue that Bella Coola offers evidence for (31c), we would need to see what happens with voiced obstruents which are consistently [+voi]. Had Bella Coola allowed voiced obstruents occurring in [CV] words, then we would be bound to argue that they are moraic, in the same way all other onsets are, supporting (31c); on the other hand, if the language had voiced obstruents, but these failed to occur

Onsets and weight: the theory

29

in [CV] words, then presumably voiced obstruents would be unable to be moraic onsets, thus only [−voi] segments could be moraic (evidence for (31b)). Unfortunately, since in Bella Coola voiced obstruents may appear allophonically in intervocalic position, but never word initially (see Nater 1979: 171), both options seem possible. One final point is worth mentioning here. As is evident, the cases in (31a) and (31c) make no internal distinction in how onsets are treated: they are either all non-moraic or all moraic, respectively. Their specific quality only matters in (31b), which is, however, the most interesting pattern for our purposes and is discussed extensively in forthcoming chapters. For this reason, (31a) and (31c) can be further simplified, by doing away with the contextual onset moraic markedness constraints and just keeping the general *μ/Onset. The updated patterns are as follows: (33)

Alternative way to state patterns of coerced onset weight a. *μ/Ons >> Be Moraic b. *μ/Ons/[+voi] >> Be Moraic >> *μ/Ons c. Be Moraic >> *μ/Ons

1.3.3.6 Interim summary This section has presented several key ideas of the book, including: i) the interaction among voice-induced pitch perturbation, tone, stress and onsets, ii) the dual nature of sonorants and iii) the patterns of coerced onset weight based on the moraic-onset markedness constraints of (29) in conjunction with a moraicity inducing constraint (30). Our attention now shifts to underlyingly moraic consonants, i.e. geminates, which instead of straddling the syllable boundary (Hayes 1989; Morén 2001; Davis 1999a) are wholly syllabified in an onset position. By virtue of their input mora specification, such consonants are not bound to the moraic markedness constraints mentioned before, simply because faithfulness may require them to keep their underlying moras. 1.3.4 Moraic-onset geminates Distinctive weight refers to the weight contrast found between short and long vowels and singleton and geminate consonants. Our focus is on the latter. Traditionally (after Hayes 1989), geminates have been assumed to be consonants which are underlyingly specified as moraic. Thus, the distinction between singleton and geminate is one between input non-moraic consonant and input moraic consonant as shown in (34).

30 (34)

Onsets Singletons vs. geminates in the input (Hayes 1989) µ C

C

Since then, various works have questioned the validity of the claim that geminates are always moraic, including, most prominently, Tranel (1991) and more recently Muller (2001). Others, however, have shown that instances supporting the absence of moraicity in geminates may be refuted or at the very least weakened when data are examined more closely, concluding that geminates are consistently moraic (Curtis 2003). This latter approach is also adopted here. While geminates may emerge in any position within the word, it is by far most common that they arise word medially. In that position, their surface representation is standardly taken to be one whereby the geminate straddles the syllable boundary in a so-called ‘flopped’ structure (Hayes 1989). Evidently, the first part of the geminate syllabifies in a coda (and carries a mora), and the second part directly associates to the onset of the following syllable. This structure achieves two goals: i) it represents the increased length of geminates compared to singletons and ii) it allows the coda, rather than the onset, to host a mora. (35)

‘Flopped’ structure of geminates word medially σ σ µ Cː

On the other hand, such a configuration faces an insurmountable problem, namely the inability to represent initial geminates in languages like Pattani Malay ((36) and §5.2.1) and Trukese (§5.2.2) among others. (36)

Initial singletons vs. geminates in Pattani Malay (Abramson 1999, 2003) Singletons pagi tido labɔ ɣatɔ

Geminates ‘morning’ ‘to sleep’ ‘to profit’ ‘comprehensive’

pːagi tːido lːabɔ ɣːatɔ

‘early morning’ ‘put to sleep’ ‘cause to be late’ ‘to spread out’

This is because initially no coda is available that can simultaneously host the mora and the first part of the geminate. Consequently, an alternative must be employed. Curtis (2003) addresses this issue and presents a number of possible solutions.

Onsets and weight: the theory (37)

31

Possible moraic representations of initial geminates c. σ b. PrWd a. σ … σ µ µ µ µ pː o

µ

µ

pː o

pː o

(37a) – proposed in Davis (1999b) – suggests that the geminate’s mora has no linking to higher prosodic structure, and thus should contribute no weight. But this is incorrect; for instance, as we will see in Trukese, geminates count for word-minimality purposes (§5.2.2). (37b) is favoured by Curtis, but partly identifies geminates with unsyllabified consonants like the ones surfacing in Bella Coola (Bagemihl 1991; §4.2.1), Piro (Lin 1997; §3.3.2) or Arabic (Kiparsky 2002). As there is no argument against the syllabification of these consonants, such a representation for initial geminates also has to be discarded. Curtis finally entertains the possibility of the structure in (37c) – i.e. an onset geminate – but does not investigate it further. This, I claim, is actually the right representation of initial moraic geminates. Moreover, I will be proposing that this representation is not only possible word initially, but also word medially as several languages testify (§5.3: Marshallese, Trique, Djeebbana, Puluwat, Eastern Popoloca). There are various advantages to (37c): first, the geminate gets to realize its mora unobstructed; second, the mora is within the onset, i.e. a syllabified position, and as such can naturally render its syllable heavy; and third, the geminate links to an onset, that is, a syllable constituent that is well received cross-linguistically. The only possible objection to this configuration is the absence of double linking that may perhaps reflect the increased length of geminates. However, as I will show in §5.4.2.2, double linking may cause more problems than it solves, since the increased length of geminates is actually a phonetic rather than a phonological reflex. Single linking and the absence of ‘flopped’ structure is consequently the default configuration for initial, but also final, geminates as shown below. (38)

a. Word initially (proposed here) σ µ µ Wd[Cː V

b. Word finally (Ham 2001) σ µ µ V Cː]Wd

But if moraic onsets are the right representation of geminates word initially and if they can also arise word medially, then what is the justification for the

32

Onsets

existence of ‘flopped’ geminates too in that position? In other words, isn’t it more economical to have a single representation for geminates, rather than two, depending on their position in the word? To understand why both structures are necessary, one needs to take into account the ‘flopped’ structure which is required in instances of languages that ban moraic onsets – which so happen to form the majority. Recall that a geminate is simply an underlyingly moraic consonant. If moraic onsets are not allowed, then the only thing that can be done is to have the geminate straddle the syllable boundary, by having a coda (39b) as the host of the geminate’s mora and an onset that renders the next syllable onsetful, a universal preference over onsetless syllables (39a). If, however, moraic onsets are permitted and if codas are avoided – at least whenever possible – then the structure in (39c) is preferred. (39)

Geminates word medially a. Avoidance of onsetless σs

*

σ µ

σ

b. Structure if moraic onsets are banned σ σ

c. Structure if moraic onsets are permitted σ σ

µ

µ

µ

µ

µ

µ

µ

µ

V Cː

V

V Cː

V

V

Cː

V

Chapter 5 will extensively discuss and analyse a number of languages with initial and medial geminates, arguing that use of moraic-onset geminates is imperative. We will also examine their interaction with stress (Pattani Malay, §5.2.1; Marshallese, §5.3.1), compensatory lengthening (Pattani Malay, §5.2.1; Trukese, §5.2.2; Trique, §5.3.2), word minimality (Trukese, §5.2.2) and reduplication (Bellonese, §6.2.4.1; Marshallese, §6.2.4.2). Finally, another category of consonants will be considered (§5.4.2), namely those that lack moras, but are nonetheless long. These have also been tagged geminates in some of the literature (e.g. Tranel 1991), but I will claim (following Ham 2001) that such long consonants are not actually geminates. Instead, they are doubled consonants with two root nodes (see Selkirk 1990 and §5.4.2). I will argue that Berawan (García-Bellido and Clayre 1997; §6.3.1) is one of the languages having such consonants and that it has wrongly been considered in the past to possess onset geminates. As I show, there is no convincing evidence in support of this claim. Importantly, the distinction between geminates and doubled consonants as separate entities makes a significant prediction, namely that they may co-occur in the same language. Ham (2001) presents Bernese as a case of this sort.

Onsets and weight: the theory 1.4

33

Organization of the book

This chapter has set the scene for what will follow. Having established the inadequacy of standard moraic theory (§1.2), it has been proposed that a modified version of it is in order, where onsets are allowed to be moraic on a language-specific basis (§1.3). This updated moraic theory is undoubtedly more advantageous, for three main reasons (accompanied by representative specific examples), as shown below. (40)

Core advantages of modified moraic theory that admits moraic onsets I.

It accounts for a wide range of data and weight-based phenomena involving onsets that had till now resisted satisfactory analysis, such as: i) onset-sensitive stress (Ch. 2) ii) certain cases of word minimality (Chs. 4 and 6) iii) onsets as a trigger or target of compensatory lengthening (CL; Chs. 3, 5 and 6) iv) various instances of gemination (Chs. 5 and 6)

II. It facilitates or offers theoretical solutions to certain problematic cases for phonological theory, including: i)

a theoretical explanation and proposal about the dual nature of sonorants (in the examination of tone and stress; §1.3.3) ii) an output-based analysis of CL with special focus on cases where onsets serve as the trigger or target of CL (Chs. 3, 5 and 6) iii) an account of initial geminates (Chs. 5 and 6) iv) a full exploration of word-medial geminates (Ch. 5)

III. It is highly restrictive, since: i)

it claims that moraic onsets are of two types only: a) distinctive, b) coerced ii) moreover in the case of b): their distribution is regulated by the wellestablished connection between onsets-voicing-tone, as well as the connection between H tone and stress iii) interleaving of Be Moraic in the fixed ranking *μ/Ons/[+voi] >> *μ/Ons produces all cases of coerced onset weight iv) it makes specific predictions on the patterns of moraic onsets that can be found

All these issues and more are explored in the remaining chapters, which are structured as follows: Chapter 2 focuses on onset-sensitive stress, where it is argued that ‘quality of onset’ effects (QO Pattern ❶) are attributed to weight, whereas ‘presence of onset’ effects (PO Patterns ❷) are due to alignment considerations, rather than weight. Given the independence of QO and PO effects, it is predicted that these may also interact. Pirahã confirms such a prediction (Pattern ❸ in §2.4).

34

Onsets

From Chapter 3 onward, attention largely shifts towards weight-sensitive phenomena beyond stress. This is especially targeted towards readers who endorse the idea that various aspects of stress can be reinterpreted in terms of prominence (see §7.2.1), thus by-passing mention of weight. My aim here is to show that many other weight-based phenomena crucially refer to onsets, thus alluding to their moraicity is unavoidable. Chapter 3 is devoted to compensatory lengthening (CL). For reasons to be explained later, CL is not particularly enlightening about the coerced vs. distinctive weight issue, but is nonetheless a highly common weight-based phenomenon and as such could not be missing from a book of this type. Using data from Samothraki Greek CL after onset r-loss (§3.2), a novel account of CL is offered, which is both empirically as well as technically grounded. Notably, extensions of the proposed approach capture a host of other CL instances, such as CL after deletion of unsyllabified material, in a unifying way (§3.3.2). Most important for our purposes, however, is the finding that onsets not only trigger CL (through their loss), but also are its targets (§3.3.4, §5.2.1, §5.2.2.1 and §6.2.1) when neighbouring material deletes. This can only be accounted for if onsets are allowed to be weight-carriers. Chapter 4 investigates word minimality in Bella Coola and shows that [CV] words behave like other minimal words in the language which happen to be bimoraic. An elaborate analysis is built, whose purpose is to persuade the reader that only a moraic-onset approach can capture the full range of facts. Chapter 5 yet again shifts its orientation somewhat, this time, by focusing on distinctive onset weight, i.e. geminates (and ‘geminated’ onsets, §1.3.1). The behaviour of such consonants in respect of various phenomena, including stress, CL, word minimality, etc., is examined. I argue that onset geminates not only appear word initially (Davis 1994; Curtis 2003), but also word medially (§5.3). This is entirely anticipated if such geminates syllabify as weightbearing onsets. For completeness, I consider numerous typological issues (§5.4.1, §5.4.2.3) and confront possible objections to the proposed model. Chapter 6 complements the previous chapters, less in theory and more in empirical facts. It reports on a number of languages which could be taken to support onset moraicity through a range of processes and phenomena (some of which are previously undiscussed, e.g. reduplication, metrics), but where the data are not robust enough to do so conclusively. Its second aim is equally important; it investigates languages which in previous accounts have implicitly or explicitly been argued to support onset moraicity, but where I argue that this is false (see Bislama and Nankina in Gordon 2005; Berawan in GarcíaBellido and Clayre 1997). This chapter then can serve as a starting point for

Onsets and weight: the theory

35

investigations of what kinds of patterns we can expect and what we should be aware of in our search for onset-weight languages. Finally, Chapter 7 summarizes the most important points of the book, examines alternatives and then ventures out on questions for future research. The two major rivals of the present theory are briefly reviewed and are ultimately declared unsuccessful. These involve what I dub as: i) the weight-as-a-functionof-prominence account (Gordon 2005); ii) the prominence-driven-onset-stress account (with Smith 2005 being one of its major proponents). Some tentative discussion follows revolving around issues such as: a) the weightfulness of CCV onsets, b) the few data that seem less well behaved according to the theory laid out in this book and c) the possibility of phonetic experimentation and testing of these phonological claims. All these obviously merit further investigation. By the end of this study, it is hoped for that readers will have been convinced that syllable and weight theory should be modified in such a way so that both non-moraic and moraic onsets (41a, 41b) are allowed. This would entail a model more symmetrical in nature, where both codas and onsets are able to emerge in weightless and weightful versions on a language-specific basis. In fact, this should be the null hypothesis. Rather than stipulating that onset weight does not exist – based on its relative rarity – what previous theories should have been doing (but for the most part haven’t) is to justify the lack thereof. (41)

a. Non-moraic onsets σ

C

b. Moraic onsets σ

µ

µ

µ

V

C

V

In the present book, I have offered a different perspective. By denying the arbitrary stipulation that onset weight is lacking, I have searched for languages that support the existence of onset weight. As will be seen, there seem to be several of these, and at any rate, many more than previously thought. Onsets and codas are not as different, then, at least with regard to their prosodic behaviour. If we dig through more closely, we will perhaps find even more.12 To the extent that there is some asymmetry, I believe that it relates more to the fact that moraic onsets are much rarer than moraic codas, rather than to their phonological representation per se. Although I am basically interested 12

A recent example of this sort is Djeebbana mentioned in Wolf (2008), who understands geminates there as moraic onsets based on Topintzi (2008a). Notably, I was made aware of Djeebbana only recently and include it here through Wolf’s work.

36

Onsets

in the latter in this book, I offer concrete proposals about this asymmetry later on (see §3.3.3, §5.4.1, §6.2.3). However, my hunch is that such explanations can be complemented by more functional accounts (e.g. Smith 2005; Gordon 2005) that in a sense extend beyond purely phonological considerations and move on to the realm of general cognitive abilities (e.g. sensitivity to auditory stimuli, Delgutte 1982; Viemeister 1980).

2 Onsets and stress

2.1

Introduction

I begin the exploration of weightful onsets by examining onset-sensitive stress. Recall from the previous chapter that onset-sensitive stress depends on two independent dimensions: a) the quality of the onset (QO); b) the presence of the onset (PO). I claim that only the former is attributed to weight (§2.2, §2.4, §2.5), whereas the latter relates to alignment considerations that require stressed syllables to be onsetful (§2.3). (1)

Presence and quality of onset interaction in stress QO

PO

✓ ✕ ✓ ✕

✕ ✓ ✓ ✕

Languages

Pattern Identifier 1

Karo (Arabela) Aranda, Banawá, Dutch Pirahã (Arabela) Greek, Russian, etc.

❶ ❷ ❸ ❹

Such independence is supported by the fact that in some languages only QO or PO effects emerge, but in others they interact, as in Pirahã. The following sections discuss each of these patterns with detailed case-studies. For convenience, patterns are numbered so that they can easily be identified in the text. Each of the languages examined exhibits the following pattern of coerced onset weight, discussed in §1.3.3.5: (2)

*μ/Ons/[+voi] >> Be Moraic >> *μ/Ons

Given that the ranking between *μ/Ons/[+voi] >> *μ/Ons is fixed, as argued in §1.3.3.4, the possible rankings generated by these constraints are three, depending on the position of Be Moraic. The least interesting case for our purposes is 1

Arabela most likely lacks onsetless syllables, therefore it cannot serve as a testing ground with respect to the presence of the onset issue. This is why I position it in both cells.

37

38

Onsets

when it is at the bottom, because this entails the absence of onset weight. Most languages work this way. The other pattern, i.e. when Be Moraic is at the top, implies that all onsets are moraic. While this is an interesting case, I have found no example of this sort, at least not with regard to stress. Bella Coola word minimality, however, is a possible candidate (§1.3.3.5 and Ch. 4). This leaves us with the case schematized in (2). Karo (§2.2), Pirahã (§2.4) and Arabela (§2.5) each exhibit this ranking, so that voiceless onsets are moraic and voiced ones are not. What they all share is that the former attract stress on their host syllable due to the action of the Weight-to-Stress Principle which demands that heavy syllables are stressed. Each of the languages also demonstrates an additional twist that makes it even more interesting. In Karo it is the co-existence of onset and coda weight, in Pirahã it is the primacy of nucleic weight over general, i.e. onset-induced, weight, and in Arabela it is the contextual stress shift that only applies with regard to primary stress. Besides discussing and analysing the relevant data, my aim in this chapter is to argue that the only peculiar thing about onset weight is exactly that. Otherwise, it is pretty regular. For instance, the primacy of nucleic weight over onset weight is analogous to the primacy of nucleic over coda weight, as many studies report (Hammond 1999 for English; Morén 2000 for Kashmiri; Gordon 2006); similarly, the contextual application of stress and weight finds other parallels too (see Rosenthall and van der Hulst 1999 for Goroa; McGarrity 2003 for Huariapano). The nature of onset weight is consequently parallel to other types of syllable weight, a fact that renders it much less exotic than usually considered. The chapter is structured as follows: Section 2.2 deals with languages where only QO effects are traceable (❶). A detailed case-study of Karo follows (§2.2.1) that involves discussion of the several aspects important for Karo stress (§2.2.1.1) with special focus on onset voicing and its analysis (§2.2.1.2). An alternative account offered by Blumenfeld (2006) is shown to be inadequate (§2.2.1.3). Section 2.3 next addresses languages where only PO effects emerge (❷). These are further sub-divided into two patterns: onset-on-stress effects (§2.3.1) and stress-on-onset effects (§2.3.2). While both refer to the requirement that stressed syllables are onsetful, the former describes languages where the normal stress pattern is manipulated in such a way that stress shifts and falls on an onsetful syllable, as in Iowa-Oto (§2.3.1.1), while in the latter, stress does not change, but if it were to dock onto an onsetless syllable, then this acquires an onset, as in Dutch (§2.3.2.1). This part of the chapter closes

Onsets and stress

39

with a discussion of alternatives and the advantages of the present account (§2.3.3). Section 2.4 explores Pirahã, a language that simultaneously exhibits QO and PO effects (❸). Finally, §2.5 focuses on Arabela, a language that again shows QO effects, but it is unclear whether it allows onsetless syllables, so it cannot be tested for PO effects. Section 2.6 offers a brief summary of the points made in this chapter. 2.2

Languages where the quality of onsets (QO) matters exclusively (❶)

The presence of QO effects is evident in at least three languages: Pirahã, Arabela and Karo,2 but it is only the latter that also clearly lacks PO effects. In Pirahã, as we will see (§2.4), PO effects are also at play, while no certain conclusion can be reached for Arabela, given that it most likely lacks onsetless syllables. Karo and Pirahã also diverge in another respect. As mentioned before (§1.3.3), in QO languages, onset consonants attract stress provided they lack [+voi]. Voiced obstruents always bear [+voi] and voiceless obstruents always [−voi]. Sonorants, on the other hand (§1.3.3.1), obligatorily bear [+son] and may also present [+voi] or be unspecified for [voi] on a language-specific basis. In Karo, I claim, sonorants are underspecified, thus pattern with the voiceless obstruents (§2.2.1.2), whereas in Pirahã, they bear [+voi] and consequently pattern with the voiced obstruents (§2.4.2). Finally, Arabela onset sonorants are similar to those in Pirahã in not attracting stress and thus are considered non-moraic. However, the facts with regard to voiced obstruents are not entirely clear, hence no direct comparison with them can be done, as in Pirahã. I start the discussion with data from Karo. 2.2.1 Karo Karo is a Tupi language spoken in the Rondônia State, Brazil, by approximately 150 Arara Indians. The data presented here are those that Nilson Gabas Jr. has collected during field work and published in Gabas (1998) and Gabas (1999), henceforth referred to as G983 and G99, respectively. I begin by presenting the phonemic consonantal inventory first.

2

3

Davis (1985, 1988) argues that in Madimadi (Mathimathi) coronal onsets render their syllables heavy. However, Gahl (1996) has shown that Mathimathi’s stress is morphologically regulated. G98 is written in Portuguese. As I have not attempted any translation of the glosses into English, these are given in the original language.

40 (3)

Onsets Karo phonemic consonants

Stops Nasal Fricative Approximant

bilabial

alveolar

palatal

velar

glottal

p b m

t r n

c

k g ŋ

ʔ h

w

y

Importantly, the language seems to treat /r/ (which is phonetically a flap [ɾ]) as a /d/ phonologically. Gabas shows that [ɾ] behaves phonologically and morphologically as the voiced stops [b] and [g] do and thus is included among them. Alternations across words like the ones in (4) provide evidence for this fact. Vowel length is not distinctive in Karo, but optional. See Gabas (1998: 22) for details. (4)

Data from G98: 22 pә̃t + ʔaʔ → pә̃ːɾ aʔ (/t/ turns into [ɾ]) up + ʔaʔ → uːb ʔaʔ (/p/ turns into [b])4 tak + ʔɨp → taːg ɨp (/k/ turns into [g])

This property is not unprecedented; other Tupi languages treat /b r g/ as the voiced counterparts of /p t k/ (Rodrigues p.c. to Gabas 1999: 12), while in Gadsup, Papua New Guinea (Frantz and Frantz 1966: 3), underlying voiced alveolar /d/ surfaces as either [d] or [ɾ] intervocalically. Voicing basically contrasts in stressed syllables. Elsewhere, its distribution is regulated by various lenition and fortition processes (discussed in more detail in §2.2.1.3). Following Gabas (1998, 1999), I will be arguing that the voicing contrast is nonetheless lexical and that it conditions stress placement. (5)

Karo voicing facts (as summarized in Blumenfeld 2006) Voicing only contrasts in stressed syllables, e.g. matːet ‘yesterday’ vs. koɾét ‘bird (sp.)’ [G99: 17] II Elsewhere, voiceless stops occur except in intervocalic onsets of stressless σs a. V_ ́ V (in this position voiceless stops are lengthened), e.g. matːet b. ʔ_V e.g. piʔti *piʔɾi ‘heavy’ [G99: 18] c. #_V e.g. pewít *bewít ‘honey’ [G99: 45] d. V_# e.g. naʔcәk *naʔcәg ‘hole’ [G99: 14] III Voiced stops occur intervocalically e. V_V e.g. pibeʔ *pipeʔ ‘foot’ [G99: 39] I

4

Unlike the other examples, the initial glottal of /ʔaʔ/ here remains. Gabas does not discuss why this is the case.

Onsets and stress

41

Another point of interest is that the only fricative in the language is /h/, which is also less frequent than other phonemes (G99: 12). Phonetically, the following consonants surface. (6)

Karo phonetic consonants

Stops Flap Nasal Fricative Approximant

bilabial

alveolar

palatal

velar

glottal

p pː p’ b

t tː t’

c cː

k kː k’ g

ʔ

m mb bm β w wP

ɾɾP n nd dn ç y yP

ŋ ŋg gŋ ɣ

h

Some remarks that will also facilitate the presentation of the data below are in order. Long voiceless stops appear as onsets of stressed syllables (see §2.2.1.3). Unreleased voiceless stops occur as final codas. In all other contexts, we get the plain alternants. Nasal sonorants appear as pre-stopped nasals, e.g. bm, in word-final codas of stressed syllables with an oral vowel, but as post-stopped nasals, e.g. mb, in onsets of stressed syllables that include an oral vowel. Plain nasals occur everywhere else. [b]~[β] and [g]~[ɣ] are in variation as onsets of unstressed syllables preceded by an open syllable, while [c]~[ç] are in entirely free variation. Finally, the nasalization of the approximants and the flap is predictable based on the nasalization of surrounding vowels. The phonemic vowels of Karo include: i) the oral i, e, ɨ, ә, a, u and o, and ii) the nasal L, N, a̰ and o̰. Phonetically, two points are important. The first is that the oral mid vowel /e/ surfaces as [e] in H-toned syllables, and [ε] in unstressed or L-tone stressed syllables. Somewhat similarly, /o/ surfaces as [o] in H-toned syllables and in unstressed syllables, but as [ɔ] in L-tone stressed syllables. The second point is that the nasalized /a̰/ surfaces as [ә̰]. 2.2.1.1 Karo tone, nasalization and stress With this much segmental information at hand, we may proceed in looking at Karo in more detail. Syllables are of the (C)V(C) type. Long vowels are not phonemic and generally do not arise. Sequences of vowels are syllabified in different syllables. Most words do not exceed three syllables in length. Karo distinguishes between H and L tone, e.g. [pḛ́n] ‘to open’ vs. [pNn] ‘to step’, [cádn] ‘to wash’ vs. [cadn] ‘to pluck’ (G99: 44). H tone is marked with the acute accent, and L tone with the absence of any mark. Apart from having phonemically nasal segments, the language also possesses rules that assign nasalization to neighbouring segments (see G98: 63).

42

Onsets

Stress is generally final, but may be pushed back to an earlier position if one of the following conditions is applicable, in the order specified below. The overwhelming majority of words present stress in a disyllabic window at the right edge. Only rarely, antepenult stress is found, but the conditions where this occurs are not clear (see fn. 6 for discussion). Boldface indicates the stressed syllable. (7)

Priorities for Karo stress assignment i)

If a syllable has H tone, then it gets stress yogá ‘egg’ (G99: 43) koɾétQ ‘guan (sp.)’ (G99: 43) mandógodn ‘rabbit (sp.)’ (G99: 40)

ii) If there is no H-toned syllable, but there is one with a nasal vowel, then this receives stress maʔo̰ pә̰ŋa moɾ ya piɾo̰n caɾ k

‘ant (sp.)’ ‘dar’ ‘miçanga’ ‘redondo’ ‘slow’

(G99: 42) (G98: 17) (G98: 39) (G98: 30) (G99: 23)

iii) If none of the above is applicable, and if the final syllable has a voiced stop onset, then the penult gets stressed5

Although my focus will be on the final condition, I will briefly discuss the other two for completeness.6 First, a phonological contrast between high and low tones exists and is independent of stress. Some minimal pairs illustrating this are presented below. (8)

5 6

Karo contrastive tones (Gabas 1999) a. cádn H tone ‘to wash’ vs. cadn L tone ‘to pluck’ b. tóy H tone ‘to disappear’ vs. tɔy L tone ‘to see’ c. pḛ́n H tone ‘to open’ vs. pNn L tone ‘to step’

I will discuss the case of a voiced stop onset in the penult too at a later point. H tone and nasalization spreading occur in Karo under certain circumstances. Certain data seem to suggest that stress assignment precedes both. For instance, the underlying form for [nayúá] is /nayúa/ ‘ant’s house’ and for [wíúpQ] it is /wíup/ ‘native, non-domesticated’ (G99: 43). Since default stress is final, we would expect stress on the final syllable instead, given that this also has H tone. This is not the case though. The same point obtains for nasalized vowels too. In addition, an optional rule turns oral vowels into their nasalized counterparts when they are between nasals (G98: 63), e.g /mani/ → [mani] ~ [mә̰ni] ‘macaxeira’; /anana/ → [ananda] ~ [anә̰nda] ‘abacaxi’. Again, stress seems to precede nasalization, otherwise we should get *[mә̰ni] or *[anә̰nda].

Onsets and stress

43

Second, notice that stressing a syllable with H tone or – in the absence of that – nasalization is top priority, so much so that it occurs even at the expense of stressing a final syllable with a voiced stop onset,7 e.g. koɾét or piɾo̰n above. Third, this strong effect of tone and nasalization with regard to stress assignment should not come as a surprise. De Lacy (2002) has explored the relationship between tone and stress and has shown that H-toned syllables commonly attract stress. As far as nasalization is concerned, John Hajek (p.c.) suggests that in fact nasalized vowels may be simply long, in which case their primacy for stress falls out easily (although some reference to the Weight-toStress Principle (N) (see §2.4.1) would be needed to assign more importance to weight coming from vowels, compared to other moraic segments). The next sections provide the relevant data that illustrate the onset voicing effect on stress. The analysis follows. 2.2.1.2 Stress and onset voicing Setting tone and nasalization aside, we can focus on the cases where neither is present in the word, thus stress is normally word final, unless that syllable has a voiced obstruent onset, in which case stress retracts to the penult. Both patterns are illustrated below. (9)

Karo stress and onset voicing (G99: 14, 39–41) a. maʔpε kɔyɔ yaʔmbɔ pakːɔ b. kɨɾɨwεp kuɾuʔcu c. yaba pibεʔ wεɾε kaɾo mәga iʔcɔgɔ

7

‘gourd’ ‘crab’ ‘yam (sp.)’ ‘fontanel’ ‘butterfly’ ‘saliva’ ‘rodent (sp.)’ ‘foot’ ‘frog’ ‘macaw’ ‘mouse’ ‘quati (sp.)’

Some other data need to be treated with caution. Gabas seems to suggest that there is a disyllabic window at the right edge of the word, which may receive stress. I have, however, found three examples where this is violated by allowing antepenultimate stress on [iyánoʔndɔnd] ‘machado’ (G98: 16), /agóaʔpәt/ → [agóáʔpәtQ] ‘pajé’ and /có̰abeʔ/ → [có̰abeʔ] ‘bow’ (G98: 73). One could perhaps attribute this to the fact that stressing an input H tone is of such priority that it can override the window. Some other examples that exhibit nasalization seem problematic in the light of Gabas’ claim (G99: 42) that there can be only one underlying nasal per word, and yet here we have two in words such as: /pa̰pa̰/ → pә̰pːә̰ ‘yam (sp.)’ (G99: 23), /mә̰yga̰ɾa/ → mә̰y̰gә̰ɾa ‘snake’ (G99: 17). Furthermore, G99: 43 argues that ‘high pitch never occurs in a

44

Onsets

(9a) presents cases with final stress. The examples here involve all combinations of sonorants and voiceless obstruents in the two final syllables of the word. Stress is consistently final. This is also the case in (9b). This time, the penult includes a voiced stop onset, whereas the final has either a sonorant or a voiceless stop. Things are different in (9c). Here the final has a voiced stop onset, and the penult has either a voiceless stop or a sonorant. Stress is now systematically penultimate. 2.2.1.2.1 Moraic onsets To analyse this set of data, I will make reference to onset moraicity and claim that in Karo voiceless obstruents and sonorants are moraic onsets, whereas voiced obstruents are not, due to the ranking (cf. (17) in §1.3.2): (10)

*μ/Ons/[+voi] >> Be Moraic >> *μ/Ons

Recall that sonorants can behave either like the voiced obstruents or like their voiceless counterparts (§1.3.3.1 and §1.3.3.4). In Karo, we see sonorants behaving like the voiceless obstruents (cf. (9)), thus suggesting that they are unspecified for [voi]. As a result, they are subject to Be Moraic >> *μ/Ons, meaning that they, alongside voiceless obstruents, act as moraic. However, Gabas (1998: 80–4) claims that sonorants and voiced obstruents pattern together as they allow H-tone spreading across them, whereas voiceless obstruents block such spreading. This finding is unsatisfactory for two reasons: first, as Gabas himself acknowledges, this is cross-linguistically an unlikely pattern since one would expect voiced obstruents as blockers of H-tone spreading given that these are standardly depressors (e.g. Bade in §1.3.3.1 Example (20)). But even if this were possible, a more serious problem is that the examples with voiceless obstruents Gabas gives are really not comparable with their sonorant and voiced obstruent counterparts. For voiced obstruents and sonorants, stress is penultimate and there is an input H tone on the same syllable. In the output, the H tone spreads onto the final, e.g. /o/+ /pitégat/ → [owitːégátQ] and /o/+/káwan/ → [okːáwán]. On the other hand, the examples with voiceless obstruents involve final stress and no lexical H tone whatsoever, making spreading inherently impossible, e.g. /caʔpe/ → [caʔpε] or /moca/ → [mocːa].

syllable of a word which also contains another syllable with an underlying nasal vowel’. Several counterexamples arise, though: /có̰abeʔ/, có̰abeʔ ‘bow’ (G98: 73); /pá̰ɾam/, pә̰́ɾ̰ә̰́m ‘wood (sp.)’ (G99: 46); /ká̰ɾam/, kә̰́ɾ̰ә̰́m ‘hummingbird’ (G99: 46); /pá̰ŋan/, pә̰́ŋә̰́n ‘to give’ (G99: 46); /kéra̰n/, kːéɾә̰́n ‘dormiram’ (G98: 80); éɾo̰m ‘lamber’ (G98: 24).

Onsets and stress

45

I thus believe it would be hazardous to draw any conclusions from such dubious data, particularly since the stress facts highlight a robust patterning between sonorants and voiceless obstruents. Crucially, Karo sonorants are not simply inert in stress assignment. They actively attract stress away from syllables with voiced obstruents, as evidenced by a word like [mәga] ‘mouse’. If they were inert, then we would only see stress retraction when the final syllable had a voiced onset and the penult had a voiceless one, but not when the penult had a sonorant one. To illustrate, let us see how moraicity would be assigned to a word like /pibεʔ/. Note that I assign moras onto input vowels following Hayes (1989), who considers short vowels as underlyingly monomoraic and long vowels as underlyingly bimoraic. This is a common assumption in most analyses. Even Rosenthall (1997), who argues that vowels acquire moras through the constraint V-mora,8 claims that long vowels are lexically specified in terms of two moras and are not subject to V-mora. I take a more simplified approach in assuming that both short and long vowels are specified as such in the input by means of moras. However, I do not present the vocalic moras in the Karo tableaux, since all vowels are essentially monomoraic. As a result, a heavy syllable in Karo is one where the onset (or coda, see §2.2.1.2.4) contributes a mora. The inclusion of moras on vowels will prove necessary in languages like Pirahã that make reference to them. Observe that superscripted moras after an onset indicate that it is a moraic onset. (11)

*μ/Ons/[+voi] >> Be Moraic >> *μ/Ons for /pibeʔ/ pibeʔ

Be Moraic

*μ/Ons

pμi.beʔ

*

*

b.

pi.beʔ

**!

c.

pμi.bμeʔ

☞ a.

*μ/Ons/[+voi]

*!

**

The last candidate leaves the competition early on, since by having moras on all onsets, it violates top-ranked *μ/Ons/[+voi] due to bµeʔ. The second candidate avoids this problem by having no mora on the onset of beʔ, but also fails to assign a mora on the voiceless onset of pi. As a result, it violates Be Moraic more severely than (11a), which manages to assign moras on the voiceless obstruent onset only. Identical results would be generated if the first onset was a sonorant. 8

The V-mora constraint states that for every vocalic root node rt1, there is a mora μ1 (Rosenthall 1997: 26 in ROA-126 version). Effectively, this says that every vowel is (mono-)moraic.

46

Onsets

2.2.1.2.2 The core analysis The stress pattern we need to capture is that stress is generally placed on the final syllable (12i), unless that syllable includes a voiced onset and the penult either a voiceless obstruent or a sonorant, in which case it shifts to the penult (12ii). In the representations below, [+voi] marks voiced obstruents, while ≠[+voi] refers to unspecified sonorants and [−voi] obstruents. (12)

Karo stress i. ii.

General pattern: Stress shift:

σ σ ≠[+voi]

σ≠[+voi] σ[+voi]

The way this can be accounted for is that stress is restricted within a disyllabic window at the right edge, in which the Weight-to-Stress Principle (WSP) is active and dominates the preference for final stress captured by AlignHd-R. (13)

WSP: Heavy syllables are stressed (Prince 1990) Align-Head-Right: Align the head syllable of a prosodic word to the right edge of the prosodic word (McCarthy and Prince 1993)

(14)

WSP >> Align-Hd-R

WSP is violated if syllables with moraic onsets receive no stress. In particular, when the ultima has a non-moraic onset and the penult a moraic one, then penult stress appears as a better candidate, since WSP is satisfied. This point is schematically presented below (15). (15)

Penult stress as a result of stress shift (syllables with moraic onsets have an indexed mora) σ≠ [+voi]

σ [+voi]

☞ a.

σμ ≠ [+voi]

σ [+voi]

b.

σμ ≠ [+voi]

σ [+voi]

WSP

Align-Hd-R *

*!

When the ultima has any other type of onset, such stress shift does not occur, as there is no trigger for its application. In particular, if the penult has a voiced onset (cf. (16)), then both WSP and Align-Hd-R are satisfied more satisfactorily in final stress (16a) than in penult stress (16b). (16)

Final stress as a result of simultaneous WSP and ALIGN-HD-R satisfaction σ [+voi]

σ≠ [+voi]

☞ a.

σ [+voi]

σμ ≠ [+voi]

b.

σ [+voi]

σμ ≠ [+voi]

WSP

Align-Hd-R

*!

*

Onsets and stress

47

In the other cases which involve all the combinations of voiceless obstruents and sonorants in the final two onsets, WSP is equally satisfied by both candidates. Consequently, Align-Hd-R picks out the candidate with final default stress (cf. (17)). (17)

Final stress as a result of equal WSP satisfaction. ALIGN-HD-R decides σ≠ [+voi]

σ≠ [+voi]

WSP

☞ a.

σμ ≠ +voi]

σμ ≠[+voi]

*

b.

σμ ≠ [+voi]

σμ ≠[+voi]

*

Align-Hd-R

*!

This ranking also makes a further prediction. Recall from earlier that in Karo penultimate stress occurs when the final onset is a voiced obstruent. This is actually the claim that G99: 40 makes. It thus implies that penult stress should occur even when the penult itself has a voiced obstruent onset. Gabas nonetheless provides no example of this sort. What is more, he presents two examples which he treats as exceptions to this generalization. These are: (18)

σ[+voi] σ[+voi] words: exceptions for Gabas, norm for current account (G99: 41) a. kiɾibɔp b. mɨɾɨɾɨy

*kiɾibɔpQ *mɨɾɨɾɨy

‘frog (sp.)’ ‘toad (sp.)’

Notably, both penult and ultima have voiced onsets. And yet, final stress arises. While this is an exception to Gabas’ statement, it is in fact anticipated given the analysis sketched here. The reason stress shift does not occur is because both candidates vacuously satisfy WSP, therefore Align-Hd-R is again the determining factor in favour of final stress (19a). (19)

Voiced onsets: Final stress because ALIGN-HD-R decides σ [+voi]

σ [+voi]

☞ a.

σ [+voi]

σ [+voi]

b.

σ [+voi]

σ [+voi]

WSP

Align-Hd-R

*!

In sum, then, the part of the Karo stress algorithm that bears on the onset quality issue reveals that sonorants pattern alongside voiceless obstruents in being moraic onsets. The ranking of *μ/Ons/[+voi] >> Be Moraic >> *μ/Ons (cf. (10)) combined with the assumption that Karo sonorants lack the [voi] feature produces this result. The second major part of the analysis involves the ranking WSP >> Align-Hd-R (cf. (14)), which requires that heavy syllables get stress. When the ultima has a voiced onset and the penult another type of onset, then

48

Onsets

the penultimate syllable receives stress as it is heavy. When the ultima has a ≠ [+voi] onset, then final stress simultaneously satisfies WSP (by stressing a heavy syllable) and right-edge alignment of stress. Similarly, if both ultima and penult have voiced onsets, stress is again word final by vacuous satisfaction of WSP and perfect stress alignment. 2.2.1.2.3 Onsetless syllables At this point, we need to justify the exclusiveness of the QO effects in Karo in the absence of PO effects. To show this, consider the following facts that indicate that Karo allows onsetless syllables, both word initially and word medially. (20)

Onsetless syllables in Karo a. i.y̰ә̰ i.cːɨ a.mε.kːɔ b. wí.upQ có̰.a.beʔ c. é.ɾo̰m ḭ.y̰a d. i.ɔt e.i

‘Brazil nut’ ‘water’ ‘jaguar’ ‘native, non-domesticated’ ‘bow’ ‘lamber’ ‘bird’ ‘pescoço’ ‘irara’

(G99: 18) (G99: 19) (G99: 21) (G99: 43) (G98: 73/G99: 25) (G98: 24) (G99: 42) (G98: 28) (G98: 22)

(20a) exemplifies onsetless syllables word initially, and (20b) word medially. Stress is assigned as previously, so that syllables with H tone or nasalization attract stress. (20c) illustrates that onsetless syllables may receive stress if they are the best possible stress bearers given that they also carry tone or nasalization. But more importantly, (20d) shows that even in the absence of these, onsetless syllables can still act as stress bearers. Onsetless syllables then in Karo do not push stress away, nor do they require the epenthesis of an onset to be able to maintain stress (cf. Dutch in §2.3.2.1), verifying the claim made earlier (see §1.2.4.1 and §2.1) that the QO and PO effects are independent of one another. In Karo, the quality of the onset matters, but its presence is unimportant. Onsetless syllables may carry stress in the same way onsetful ones do. 2.2.1.2.4 Exceptions and the role of codas In Gabas (1999: 41), ten examples are listed as exceptions to the anticipated stress shift. The first two have already been dealt with in (19), where it was shown that when both the penult and the ultima contain a voiced obstruent, there is no reason for stress to shift. The next three examples indeed emerge as genuine exceptions given what we have said so far. Although nothing conclusive can be stated at this point, I speculate that

Onsets and stress

49

perhaps final syllables here possess H tone or nasalization that has not been identified (see fn. 6 for some additional dubious data) and consequently attract stress on the final syllable. (21)

Truly exceptional cases (G99: 41)9 /acibe/ /pobo/ /cagɔp/

[acibε] [pɔbɔ] [cagɔp]

*[acibε] *[pɔbɔ] *[cagɔpQ]

‘root’ ‘owl’ ‘dish’

The final five examples are below.10 We will deal with these more extensively. (22)

Exceptions that involve final sonorant codas (G99: 41) /korem/ /koran/ /pirun/ /pagon/ /yogoy/

[koɾεbm] [koɾadn] [piɾudn] [pagɔdn] [yogɔy]

*[kɔɾεbm] *[kɔɾadn] *[piɾudn] *[pagɔdn] *[yɔgɔy]

‘also’ ‘fish (sp.)’ ‘round’ ‘friend’ ‘breath’

Since the final onset in all cases here is [+voi] and is preceded by a non-voiced penultimate onset, we would expect penult stress (15), but this is not what happens. Note, however, that final syllables also possess a coda which comprises a sonorant consonant. It is thus possible to propose that these codas are moraic. If they are, then the moraic composition of a word like [yogɔy] is the following: [yμoμ.gɔμyμ]. Since both syllables are bimoraic, they are equally eligible for stress. Align-Hd-R can now play a role in selecting final stress.11 That this line of thought is promising is supported by the following observations. First, word medially, one only finds the coda /ʔ/ and much less frequently /h/ (G99: 12, fn. 6), e.g. ihy̰әy ̰ ̰ ‘piranha’ (G99: 18), which cannot be moraic because in a word like [piʔ.ti], the final syllable receives stress (23). If the penult coda had a mora, then that syllable would be trimoraic and the final one bimoraic. Due to the ranking WSP >> Align-Hd-R, and the gradient computation of WSP (for justification see §2.4.1), we would thus expect penult stress contrary to facts. (23)

Non-moraic codas word medially maʔpε naʔtɔ piʔti

9

10 11

‘gourd’ ‘tapir’ ‘heavy’

(G99: 14) (G99: 14) (G99: 18)

Two more examples are exceptional given the current proposal: [aɔɾɔ] instead of [*aɔɾɔ] ‘parrot’ (G99: 39) and [yaɨ] rather than [*yaɨ] ‘macaco guariba’(G98: 25). Thanks to Moira Yip (p.c.) for suggesting closer examination of these data. If sonorant codas are indeed moraic, then one could perhaps wonder why [é.ɾo̰m] does not get stress on the final syllable. The answer lies in the fact that moraicity in Karo only becomes important in cases where tone and nasalization do not regulate stress. Here, the word includes

Onsets

50

Second, word finally one finds a larger set of codas that includes the unreleased stops, pewít ‘honey, sweet’ (G99: 13), the glottal stop – primarily as a coda of monosyllabic functional words, as in ʔaʔ ‘classifier’ (G98: 13) and the sonorants (discussed below). The non-moraicity of non-sonorants again can also be seen word finally. (24)

Non-moraic non-sonorants word finally cɨɾɨpQ pibeʔ

‘bird (sp.)’ ‘foot’

(G99: 39) (G99: 39)

Since default stress is final and since voiceless obstruent onsets contribute a mora to their syllable, the first syllable of these words is bimoraic. Had the non-sonorant coda added a mora, then the final syllable should be bimoraic too and should receive stress due to better stress alignment. The fact that it does not shows that the final coda adds no mora. On the other hand, sonorant final codas receive stress – as in the alleged exceptions of (22) – because they bear moras, a proposal which is in accordance with Zec’s (1988, 1995) system that predicts languages where only sonorant codas are moraic, whereas non-sonorant ones are not (e.g. Kwakwala).12 Karo, however, is unique in that it simultaneously presents the relationship between sonority & moraicity for codas as well as that between voicing & moraicity for onsets. The pattern in (22) can thus be easily accounted for with the existing constraints and under the assumption that sonorous codas are moraic.13 (25)

Moraic sonorant codas: WSP >> ALIGN-HD-R yogɔy ☞ a. b.

WSP

μ μ

μ μ

*

μ μ

μ μ

*

y o gɔ y

y o gɔ y

Align-Hd-R

*!

The partial moraicity of codas also explains why these two words receive final stress.

12

13

both tone and nasalization, rendering moraicity unimportant. Also, as was noted before (cf. fn. 6), stress seems to precede nasalization, thus in the presence of tone too it will be attracted onto an H-toned syllable, rather than the final nasalized (and heavy) one. Of course, stress will appear on the penult if this is a bearer of H tone or nasalization, both of which override weight considerations: /yogo̰yom/ yogo̰y̰o̰m ‘beard, moustache’ (G99: 42) /ko̰nam/ ko̰nә̰m ‘crazy’ (G99: 42) /káwan/ káwán ‘be fat’ (G99: 46) I have been able to find one exception to this pattern, namely the word [εcːigudn] ‘ant-eater (sp.)’ (G99: 40) instead of the anticipated *εcːigudn.

Onsets and stress (26)

pεɔdn pεnaɔy

*pεɔdn *pεnaɔy

‘skin’ ‘dançar’

51

(G99: 40) (G98: 16)

Without final moraic sonorant codas, we would expect penult stress, because the final syllable would have just one mora. The penult would have two due to the onset and nucleus moras, therefore stress should be dragged onto that syllable. The fact that it is not is difficult to account for, but not if sonorant codas contribute moras. In that case, the final syllable is bimoraic too, consequently stress remains put as alignment is perfect. 2.2.1.3 Blumenfeld (2006) While the present account has followed Gabas (1998, 1999) in arguing that Karo stress is predictable and voicing is contrastive, there is an obvious alternative, namely one which treats stress as lexically determined and the voicing of segments as predictable. Such an approach has been adopted by Blumenfeld (2006) and is based on the following voicing facts for Karo (repeated for convenience from (5)). (27)

Karo voicing facts (as summarized in Blumenfeld 2006) I

Voicing only contrasts in stressed syllables, e.g. matːet ‘yesterday’ vs. koɾét ‘bird (sp.)’ [G99: 17]

II Elsewhere, voiceless stops occur except in intervocalic onsets of stressless σs a. b. c. d.

V_ ́ V ʔ_V #_V V_#

(in this position voiceless stops are lengthened), e.g. matːet e.g. piʔti *piʔɾi ‘heavy’ [G99: 18] e.g. pewít *bewít ‘honey’ [G99: 45] e.g. naʔcәk *naʔcәg ‘hole’ [G99: 14]

III Voiced stops occur intervocalically e. V_V

e.g. pibeʔ *pipeʔ ‘foot’ [G99: 39]

Building on the fact that voiced and voiceless stops only contrast in stressed syllables, Blumenfeld argues that elsewhere voiceless stops emerge, except in intervocalic onsets of stressless syllables. There, the voiced counterparts arise instead as a product of intervocalic lenition which generally applies. Other morphophonemic processes in which voiceless stops become voiced provide support for lenition, e.g. as in the context [V]_[V][−accent] where the vowels belong to different morphemes, e.g. /e-tati/ → [eɾati] ‘te trouxe’. However, to account for the voiced–voiceless contrast in stressed positions, an extra ingredient needs to be added. It is argued that voiced stops are allowed in the input, but only if the syllables that contain them carry stress. Stress is preserved in the output due to top-ranked Faith Stress, while feature identity ensures that voicing will also survive through the use of high-ranking Max-[voi]. A few

52

Onsets

more constraints are added to complete the picture. Effectively, then, the proposal maintains that it is not voicing that is unpredictable in Karo, but rather it is stress. This analysis offers an accurate account of the distribution of voiced stops (less so of voiceless stops) and implicitly suggests that the location of stress is completely accidental due to its unpredictability. In fairness, this view’s major advantage is that it insightfully accounts for the lenition processes that appear in the language. Nonetheless, while the ‘stress conditions voicing’ approach may seem at a first glance plausible, I will show that there are several points where it proves inadequate or makes wrong predictions. First, recall from (27) that Blumenfeld’s goal is to explain the distribution of voicing through the absence of contrast between voiced and voiceless stops, unless underlying stress is involved, in which case both patterns may emerge. But while this analysis superficially seems to discard underlying distinctions of voicing by rendering it predictable, it actually makes crucial use of them with reference to stressed syllables. In fact, it ties voicing with stress (but not the other way round) in a manner which seems rather arbitrary or at least without any justification. The burden thus moves to the input stress specification (rather than contrastive voicing), by posing an underlying limitation, namely that voicing should go hand in hand with stress. As a result of the Richness of the Base (RotB) (Prince and Smolensky 1993/2004) and the alleged predictability of voicing we would need to consider inputs with voiced consonants in stressless positions too, e.g. as in the first consonant in hypothetical /baki/ (in line with Blumenfeld, I assume for the moment that lexical stress specification occurs in the input). Under Blumenfeld’s account and without further modifications, this would be bound to emerge as [baki], because of high-ranked Max-[voi] which bans the loss of [voi] in consonants. But this is wrong, since as we have already seen in (27c), word-initial stops are always voiceless. This problem, however, is resolved if we allow underlying voicing contrasts – so as to be consistent with RotB – and permit separate cross-linguistically justified processes to account for the given distribution. In particular, (27c) could be attributed to word-initial fortition (Lavoie 2001), while strengthening of onset stops in stressed positions (27a) is also a possibility (Lavoie 2001; González 2003), although, admittedly, a less common process. (27d) looks like a familiar case of final devoicing, while (27e) also illustrates a widespread process of intervocalic lenition. Note that if Blumenfeld were to consider a fuller range of inputs, as hinted above, processes like the aforementioned ones would be needed one way or another to account for the fortition of /baki/ to [paki].

Onsets and stress

53

An additional important point which is merely mentioned but not discussed in Blumenfeld is also that voiceless stops in front of stressed syllables (27a) get lengthened [Gabas 1998: 10 states that only voiceless stops have long allophones]. This is significant in two respects; first it supports the idea of fortition, since gemination is a familiar case of strengthening (Lavoie 2001; González 2003). Moreover, it seems to support the idea that such consonants bear moras. Given that the effect of stress attraction appears on the final syllable as in e.g. pakːɔ, it seems reasonable to suggest that this geminate is actually wholly included in the final syllable, i.e. in the onset (a position argued for in §5.3), instead of having the more familiar ‘flopped’ representation of geminates (Hayes 1989) where one part of the geminate is in the coda position of the preceding syllable and the other in the onset of the following one. If this is on the right track, then it just seems that abstract weight is enhanced by lengthening. The gemination pattern is totally elusive in Blumenfeld.14 It is additionally unclear whether words with – default – final stress should have any underlying specification of stress or not (although in /baki/ above I assumed that they do for illustration purposes). Moreover, claiming that Karo has a partially unpredictable stress system groups it among other languages with lexical stress systems such as Russian or Greek (Revithiadou 1999; Alderete 2001). In these languages, however, it has been convincingly shown that morphemes (roots and affixes) bear particular stress properties or requirements in the input. This occurs in a constrained and consistent manner, unlike in Karo, where the situation seems to be random. However, the most important defect of Blumenfeld’s analysis is one that he also recognizes. If it were the case that stress was lexically specified, then we should expect unpredictable stress placement sometimes on the penult and sometimes on the ultima, irrespective of whether the final syllable had a voiced onset or not. And yet, penult stress only occurs when the final onset is voiced (setting aside of course the cases where the prevalence of tone or nasalization commands otherwise). More concretely, apart from a handful of exceptions, there are no instances where the final onset has a sonorant or a voiceless obstruent and stress appears on the penult. These cases consistently present final stress. 14

However, in examples such as [naʔtup] or [piʔti] no lengthening applies. If the geminates are wholly syllabified in the ultima onset, as suggested, at first glance no straightforward explanation is available as to why *[piʔ.tti] fails to arise. On the other hand, if gemination had the usual ‘flopped’ representation, i.e. as in *[piʔt.ti], then the form would be easily eliminated given that complex codas are banned. However, it is more likely that this is more of a phonetic effect. Having a ʔCː sequence would be very difficult to produce as it would require sustaining stop closure for a very long time. Moreover, given that the coda that arises is a glottal stop, some amount of its glottal closure would end up with the following consonant, but even clusters of C’.C are commonly undesirable, e.g. in Tonkawa (Kisseberth 1970).

54

Onsets

Such a result is entirely unexpected in an ‘unpredictable stress’ account. Blumenfeld does not solve this problem, but merely states that it is one shared with the way Gabas would have to analyse the data. In particular, Gabas also acknowledges the existence of exceptions, e.g. kɔɾεbm and not *kɔɾεbm ‘also’ (G99: 41),15 so he would need to treat these as cases of underlyingly specified stress and use undominated Faith Stress to explain them. But once Faith Stress is used, then nothing precludes consideration of inputs such as /mani/ where the penult is stressed in the input although the ultima has a voiceless onset. The expectation is now that this should surface as [mani] even though empirically this does not occur. On the whole, then, Blumenfeld’s analysis presents numerous shortcomings and is less economical than the current proposal. To get the Karo data right, Blumenfeld needs to assume lexical specification of stress, underlying voicing (despite his claim that this is not needed) and a close tie between voicing and stress in the lexicon. On the other hand, the present proposal only makes use of underlying voicing distinctions and generates stress on the surface in a predictable manner. The present proposal, then, is more advantageous than the ‘unpredictable stress’ analysis. It also proves to be more successful in terms of how many correct patterns it predicts, as illustrated in (28). Although a couple of patterns are missed, these may either be superficial exceptions, or they may be attributed to various other reasons instead of being ascribed to the structure of the grammar (see below). Comparing the results for each of the analyses will make this clear. (28)

Comparison between Blumenfeld’s ‘lexical stress’ analysis and current analysis (a check mark denotes that the prediction is correct given the data; italics denote that the prediction is partly correct) Combination Attested

I.

15

a. b. c. d. e. f. g.

+voi +son +voi +son –voi +voi +son

+voi +son +son +voi +voi –voi –voi

kiɾibɔp, mɨɾɨɾɨy wayɔ, mani aɣaya, paɾamit waɾo, ndɔgat, yaβa cεgo, paɾat, kaɾo εɾεpːɔ naʔtup, mocːay, nahεk

In my analysis, these are not exceptions (cf. §2.2.1.2.4).

Predicted Blumenfeld

Current work

✓ YES ✓ YES ✓ YES ✓ YES ✓ YES ✓ YES ✓ YES

✓ YES (19a) ✓ YES (17a) ✓ YES (16a) ✓ YES (15a) ✓ YES (15a) ✓ YES (16a) ✓ YES (17a)

Onsets and stress Combination Attested

Predicted Blumenfeld

Current work

II. h. –voi i. +voi j. +son

–voi –voi –voi

NO NO NO

✓ NO ✓ NO ✓ NO

✓ NO (17b) ✓ NO (16b) ✓ NO (17b)

III. k. l. m. n. o.

+voi +son +son +voi +son

NO NO NO NO NO

YES YES YES YES YES

✓ NO (19b) ✓ NO (17b) ✓ NO (16b) ✓ NO (15b) ✓ NO (17b)

+voi +son +voi son –voi

IV. p. –voi q. –voi r.

–voi

55

+voi acibεʔ, pɔbɔ, cagɔpQ ✓ YES +son pεwit ✓ YES in [σσ]Wd; otherwise lenition –voi pakːɔ, caʔpε, piʔti ✓ YES in [σσ]Wd; otherwise lenition

NO (15b) YES generally (17a) YES generally (17a)

Source and glosses: a. kiɾibɔpQ ‘frog (sp.)’, mɨɾɨɾiy ‘toad (sp.) [G99: 41]; b. wayɔ ‘jacaré’ [G98: 40], mani ‘macaxeira’ [G98: 24]; c. aɣaya ‘cacau’ [G98: 20], paɾamitQ ‘aranha’ [G98: 24]; d. waɾo ‘caracol’ [G98: 26], ndɔgat ‘comer (intr)’ [G98: 16], yaβa ‘paca’ [G98: 18]; e. cεgo ‘macaco (esp.)’ [G98: 25], paɾatQ ‘curimba (peixe)’ [G98: 40], kaɾo ‘arara’ [G98: 49]; f. εɾεpːɔ ‘tamanduá’ [G98: 25]; g. naʔtupQ ‘final’ [G98: 27], mocːay ‘mucura’ [G98: 27], nahεkQ ‘moleira’ [G98: 38]; p. acibεʔ ‘raiz’ [G98: 76], pɔbɔ ‘owl’ [G99: 41], cagɔpQ ‘dish’ [G99: 41]; q. pεwitQ ‘honey’ [G99: 21];16 r. pakːɔ ‘pacu’ [G98: 40], caʔpε ‘escama’ [G98: 25], piʔti ‘pesado’ [G98: 24]

This table can be split into four parts. In (I), both analyses agree in correctly predicting certain examples as attested (28a–g); in (II), they do the same but for unattested cases (28h–j). Part (III), i.e. (28k–o), is particularly important, since it highlights those instances which only the present approach correctly predicts as unattested, while Blumenfeld fails to do so. Finally, (IV) presents two cases (28q, r) where the right predictions are made half of the time under the current account, and only one case (28p) which is correctly predicted in Blumenfeld’s account, but not in the present one. It is thus evident from the above that Blumenfeld’s approach captures 13/18 patterns, while the current approach captures 15/18 rising to possibly 17/18 patterns. The analysis that utilizes lexical stress specification predicts numerous cases for which no data exist. This is because it fails to see the connection between 16

Note that Gabas (1998: 51) transcribes this as pεwít instead, i.e. with a high tone which would already attract stress.

56

Onsets

the moraicity and voicing properties of the segments. However, a few cases are correctly ruled out (28h, i, j), not because of stress considerations – as in my account – but because they should be impossible forms. In particular, they involve voiceless obstruents in stressless non-initial positions,17 which, according to Blumenfeld, should be lenited and appear with voiced obstruents instead. Forms like (28q) and (28r) are predicted by Blumenfeld but, as I note, only in disyllabic words. This is because the first −voi segment will remain −voi word initially in a disyllabic word, but will turn into +voi due to lenition once it finds itself in a word-medial position, as would happen in a trisyllabic word, e.g. σ[−voi] σ[−voi] σ[−voi] → σ[−voi] σ[+voi] σ[−voi]. Examples of this form exist, e.g. [paɾamitQ] ‘aranha’ (G98: 11), [ciβεkːɔ́dn] ‘urubu’ (G98: 18). In the present approach, no similar distinction is made since a −voi followed by a stressed syllable with a −voi/son onset should occur independently of the length of the word. However, even though the voicing pattern in the two examples above is missed, the same is not true for their stress pattern which is correctly predicted. Sonorant or voiceless obstruent onsets attract stress more than the voiced obstruent ones, by virtue of their moraicity. The only potentially real counterexample for the present analysis, then, is the case presented by the words in (28p), which disrespect the anticipated stress shift pattern. Nonetheless, I contend that these are far too few to weaken the generalization. Moreover, instead of imposing lexical stress specification and faithfulness to underlying stress for exceptions, it is equally likely or even plausible to consider that some other factor may be involved here. For instance, as we have seen, high tone or nasalization outweigh conditions of stress placement based on onset quality, so it could perhaps be the case that such words actually have a H-tone final syllable and this has simply not been recognized (cf. the case of pεwitQ(28q and fn. 16)). On the other hand, several configurations, as indicated in (III), are predicted as possible in Blumenfeld’s account, but none of these arises in reality. All these patterns are generated under the hypothesis that stress is unpredictable, which consequently must be on the wrong track, given the empirical facts. In the current state of affairs and with the data available, the present approach is consequently superior to Blumenfeld’s alternative analysis both empirically and conceptually. This is because it offers the fullest possible examination of 17

Word initially, syllables with stressless voiceless obstruent onsets are protected, while word medially they appear voiced. This is a fact that Blumenfeld captures through lenition processes, whereas the current analysis fails to do so, but could be amended accordingly. This is why (28q, r) are only partially correctly predicted, i.e. there are no finally stressed trisyllables where σ2= [−voi]On and σ3= [−voi]On or where σ2=[−voi]On and σ3=[+son]On.

Onsets and stress

57

facts, it provides the most satisfactory account of default stress and stress shift and it manages to minimize the set of assumptions involved. Closing this section, we have seen that Karo manipulates the consonants’ voicing and lack thereof for stress assignment purposes. Sonorants and voiceless obstruents pattern together in attracting stress on the syllable that hosts them, unlike voiced obstruents which repel stress, unless there is no better stress contender, i.e. when the only other possible location for stress also bears a voiced obstruent. At the same time, §2.2.1.2.3 has demonstrated that onsetless syllables may also bear stress, thus supporting the claim that Karo is a QO-only language (❶). The next section explores languages that present PO-only (❷) effects. 2.3

Languages where the presence of onsets (PO) matters exclusively (❷)

Like Karo, there are other languages which are sensitive to onsets in their stress algorithms, but unlike Karo, only pay attention to the presence of an onset and not to its quality. Two representative examples of this sort are Alyawarra and Dutch. In Alyawarra, stress appears on the first syllable if the word has an initial consonant, e.g. kíra ‘meat’, párriːka ‘fence’, but on the second syllable if the word has a vowel at the start, e.g. athá ‘I (ergative)’, ilípa ‘axe’. In Dutch, on the other hand, stress will be assigned on a particular syllable, but if this happens to be onsetless, then it will – under certain conditions (see §2.3.2.1.2 on Dutch) – acquire a glottal stop, e.g. /aɔrta/ → [a.ʔɔ́r.ta] ‘aorta’ vs. /xaɔs/ → [xáː.ɔs] ‘chaos’ (Booij 1995: 65). I will be arguing that both patterns are induced by the constraint ALIGNσ́O as formulated below. (29)

ALIGNσ́O: Align-L (σ́, C), i.e. Align the left edge of every stressed syllable with a consonant (Topintzi 2006b)

ALIGNσ́O states that a stressed syllable must be onsetful, but does not impose any requirements on how this can be achieved. There are in fact at least two ways in which ALIGNσ́O can be satisfied. The first involves manipulation of the general stress system of the language so that stress falls on an onsetful syllable. In this instance, the presence of an onset affects stress placement by means of stress shift, rendering such shift an onset-on-stress effect. Alyawarra, as well as Aranda, Banawá and Iowa-Oto, are representative languages (§2.3.1). Another way to satisfy ALIGNσ́O is by manipulating the segmental structure. In other words, stress assignment applies as usual, but if it happens to be the case that it would fall on an onsetless syllable, then an onset is inserted.

58

Onsets

This constitutes a stress-on-onset effect because it is stress that triggers onset epenthesis. This is exactly the case in Dutch (§2.3.2.1), Ainu (§2.3.2.2) and marginally in English (§2.3.2). Discussion relevant to the use and necessity of ALIGNσ́O as well as alternatives is presented in §2.3.3. 2.3.1 Onset-on-stress (OS) effects The OS effect is illustrated by a handful of languages, which present the following, roughly stated, stress algorithm: (30)

Stress the first syllable if onsetful, otherwise stress the second one

Languages which have this algorithm include: Aranda (Strehlow 1944), Alyawarra (Yallop 1977) and other Australian languages, such as Lamalama, Mbabaram, Umbuygamu, Umbindhamu, Linngithig, Uradhi, Kuku-Thaypan, Kaytetj and Agwamin (most of these are Cape York and Arandic languages; see Davis 1985; Goedemans 1998 and Blevins 2001 for more details). That this is not an areal feature is supported by the presence of the phenomenon in unrelated languages of North and South America, namely Iowa-Oto (Robinson 1975), Banawá (Buller, Buller and Everett 1993) and Juma (Abrahamson and Abrahamson 1984). Relevant data illustrate. (31)

I. Aranda stress a. Consonant-initial words of three or more syllables ráːtama kútungùla lélantìnama

‘to emerge’ ‘ceremonial assistant’ ‘to walk along’

b. Vowel-initial words of three or more syllables ergúma arálkama ulámbulàmba

‘to seize’ ‘to yawn’ ‘water-fowl’

c. Words of two syllables (C- or V-initial) ílba áːtwa kála gúra

‘ear’ ‘man’ ‘already’ ‘bandicoot’

II. Alyawarra stress d. Consonant-initial words of two or more syllables kwíya kíra párriːka ngáyakwa

‘girl’ ‘meat’ ‘fence’ ‘hungry’

Onsets and stress

59

e. Vowel-initial or glide-initial words of two or more syllables athá iːlpá ilípa arrákirta walíːmparra

‘I (ergative)’ ‘ear’ ‘axe’ ‘mouth’ ‘pelican’

III. Banawá stress f. Consonant-initial words dísa bádi mákarì fúanà tátikùne tìnarífabùne

‘shoot with an arrow’ ‘name’ ‘cloth’ ‘lost’ ‘hair’ ‘you are going to work’

g. Vowel-initial words of more than two moras idía owárià ufábunè

‘to marry’ ‘one’ ‘I drink’

h. Disyllabic vowel-initial words of two moras ába íta áwi

‘fish’ ‘sit’ ‘tapir’

In Aranda, C-initial words receive stress on the first syllable (31a), but V-initial ones have stress on the second syllable (31b). There is, however, one exception: in disyllabic words, stress is word initial regardless of whether the word begins with a vowel or a consonant (31c). Presumably, this relates to the fact that Aranda avoids final stress due to NonFinality (or prefers to create binary feet according to Goedemans (1996)), thus words like *il(bá) or *a(rálka)(má) with final stress are impossible. The closely related Alyawarra is similar, with two exceptions: first, syllables beginning with a glide pattern with onsetless syllables (31e);18 second, disyllables beginning with an onsetless syllable show the same behaviour as their polysyllabic counterparts (31e), presumably 18

This is not as surprising as it seems. Smith (2003) argues that there are two types of glides: true onset ones, and nuclear ones. The latter are actually part of the nucleic node and thus exempt from statements that refer to onsets. Smith claims that among related Campidanian Sardinian dialects, the Iglesias dialect has nuclear onglides, whereas the Sestu dialect has true onset glides. In addition, there are cases where the distinction between the two glides is found in a single language as is the case in French, Spanish, Slovak and English. For details, the reader is referred to Smith (2003). The implication of this result is that, in Alyawarra, onset glides could be treated as nuclear onglides, which means that they are not true onsets. Consequently, the resemblance between glide-initial and onsetless syllables falls out naturally. Moreover, this

60

Onsets

because, in Alyawarra, the need to assign stress on onsetful syllables (cf. 31d, e) overrides any non-finality considerations. However, even when this is not an issue, final stress avoidance is still evident, as in e.g. arrákirta rather than *arrákirtá. In Banawá, feet are built on moras (not on syllables) and vowel sequences form tautosyllabic diphthongs. Long vowels do not occur unless for minimality purposes (Everett 1996; Hyde 2007). Banawá is quite similar to Aranda in that it too stresses the onsetless syllable in bimoraic disyllabic words (31h), because otherwise word minimality, which requires a bimoraic foot per word, would be violated. However, in words of increased moraicity, the avoidance of stressed onsetless syllables can freely operate, which is why stress shifts to the second syllable (31g). In consonant-initial words, the algorithm operates as normal, assigning stress on the first syllable, since it is onsetful (31f). Finally, in Juma, data are limited, but again, stress seems to be sensitive to the presence of an onset. Juma differs from all the languages above in that this effect emerges word finally, i.e. ɔ́kɨtɨ́ ‘esta cortando’ where the final syllable is onsetful, versus peyíkɔpía ‘espécie de passaro’ where it is not. However, a word of caution is in order. In the limited data available, it is possible to re-analyse cases like peyíkɔpía as containing a final diphthong, i.e. pe.yí.kɔ.pía rather than a sequence of heterosyllabic vowels, i.e. pe.yí.kɔ.pí.a, which is what Abrahamson and Abrahamson (1984) seem to suggest. If the diphthong analysis proves right, then Juma no longer seems to illustrate onset sensitivity of this type. It should be obvious that all these data share one thing in common: stress preferably docks onto the first onsetful syllable (almost always counting from the left edge of the word rightward). The next section is devoted to the analysis of this pattern based on the implementation of ALIGNσ́O. Given that Aranda and Alyawarra have received an account along similar lines in Goedemans (1996), while Banawá is discussed in Hyde (2007), I provide an analysis of the lesser-known, barely discussed, stress system of Iowa-Oto (Robinson 1975).19

19

theoretical assumption is compatible with Yallop’s statements that ‘glides are part of the phonetic realization of the vowels’ (1977: 19), as well as that ‘wa- and yu- word-initially represent phonologically simple vowels’ (1977: 20). Aranda and Alyawarra have been discussed in terms of Align-FtO (which is similar to ALIGNσ́O; see §2.3.3 for discussion) by Goedemans (1996). Other analyses of Aranda are Takahashi (1994), Downing (1998) and Smith (2005), while Alyawarra is also dealt with in Downing (1998). Banawá has not been dealt with in terms of ALIGNσ́O/ALIGN-FTO, but can be accommodated by it. However, since it presents some complications regarding the vowel combinations it allows, minimality considerations and foot parsing, providing a full account would take us too far afield. Other – albeit quite different from the present one – analyses of Banawá

Onsets and stress

61

2.3.1.1 Iowa-Oto Iowa-Oto is a virtually extinct language, part of the Chiwere sub-group of the Siouan language family. According to the Ethnologue (www.ethnologue.com), the last fluent speakers died in 1996. A few remain with some knowledge of the language. The stress system is virtually identical to that of Alyawarra, although glideinitial words seem to pattern with consonant-initial ones. Consonant-initial words receive primary stress on the first syllable (32a). In vowel-initial words, primary stress docks onto the second syllable (32b). The number of syllables in a word is irrelevant. (32)

Iowa-Oto stress (Robinson 1975) a. Consonant-initial words péce náwe hérota páxoče wíwaθòče

‘fire’ ‘leaf’ ‘morning’ ‘Iowa’ ‘machine’

b. Vowel-initial words aháta ithá

‘outside’ ‘there’

Obviously, ALIGNσ́O favours stress on onsetful syllables and causes shift away from the first syllable if it is onsetless, i.e. ALIGNσ́O >> Align-L (PrWd, Ft) (i.e. Align the left edge of the PrWd with the left edge of a Foot; abbreviated as Align-L). With regard to secondary stress, we can observe that while rhythmic stress is desirable as in wíwaθòče, it is restricted from docking onto the final syllable, e.g. hérota not *hérotà. This can be attributed to a relatively high-ranked NonFin or to FtBin which requires that feet are bisyllabic. While any of these would do, for current purposes, I will use FtBin so that I also draw attention to the rhythmic nature of stress assignment that prefers binary groupings of feet. What this entails is that syllables may remain unparsed to escape a FtBin violation, hence: FtBin >> PARSE-σ. Despite this, it is not the case that feet are never unary. They can be so to avoid stressing an onsetless syllable, e.g. ithá not *ítha; consequently, ALIGNσ́O must be very highly ranked and certainly dominating FtBin. The following cases illustrate. include: Buller, Buller and Everett (1993), Everett (1996), Downing (1998) and Gordon (2005). To my knowledge, Iowa-Oto is only discussed in passing by Gordon (2005), so the present analysis is much more detailed.

62 (33)

Onsets ALIGNσ́O, Dep-C >> Align-L nawe ☞ a. b.

c. d. ☞ e.

ALIGNσ́O

Dep-C

Align-L

náwe nawé

*!

ahata

ALIGNσ́O

áhata ʔáhata aháta

*!

Dep-C

Align-L

*! *

The example /ahata/ → [aháta] compared to the normal stress pattern as illustrated by /nawe/ → [náwe] reveals that ALIGNσ́O >> Align-L. In other words, the regular stress algorithm, which requires primary stress on the first syllable, is disrupted by the presence of ALIGNσ́O which forces shift of stress to the peninitial. ALIGNσ́O could also be satisfied by consonant insertion and stress on the first syllable as in (33d). This is ruled out, though, because Dep-C (i.e. that every C in the output must have a correspondent in the input, i.e. no consonant epenthesis) is top-ranked too. Consequently, ALIGNσ́O is satisfied by altering the preferred prosodic alignment of the stressed syllable. (34)

FtBin >> PARSE-σ herota ☞ a. b.

(héro)ta (héro)(tà) wiwaθoče

☞ c. d. e.

FtBin

PARSE-σ *

*! FtBin

PARSE-σ

(wíwa)(θòče) (wíwa)θoče (wíwa)(θò)(čè)

** **

(34a) and (34b) provide the next ranking argument, namely that FtBin >> PARSE-σ, since disyllabic feet are preferred even at the expense of leaving certain syllables unparsed. In wíwaθòče, perfect satisfaction of both constraints is possible, as shown in (34c). The final point arises when we consider a bisyllabic onsetless-initial word. Our grammar does not yet tell us what will happen, since there are two independent conflicting demands. On the one hand, stress needs to dock onto an

Onsets and stress

63

onsetful syllable due to high-ranking ALIGNσ́O, therefore for an input like /ita/ we would expect [i(thá)]. On the other hand, feet need to be disyllabic due to high-ranking FtBin; consequently, we would predict [(ítha)]. In reality, what we get is [i(thá)], thus we have a new ranking argument, namely that ALIGNσ́O >> FtBin, exemplified below. (35)

ALIGNσ́O >> FtBin /ita/ ☞ a. b.

ALIGNσ́O

h

i(t á) h

(ít a)

FtBin *

*!

All in all, then, the Iowa-Oto stress system can be described by the ranking: (36)

ALIGNσ́O, Dep-C >> Align-L (PrWd, Ft), FtBin >> PARSE-σ

2.3.2 Stress-on-onset (SO) effects The preceding discussion examined the OS effect, where an onsetful syllable attracts stress due to ALIGNσ́O. Satisfaction of ALIGNσ́O comes through prosodic misalignment of the normal stress algorithm of the language. ALIGNσ́O must therefore dominate the Align-Stress constraints, and also rank at least as highly as Dep-C. This is important, because had Dep-C been low ranked, it would be possible to satisfy ALIGNσ́O not by misalignment, but by the segmental fix of onset-insertion in front of a stressed syllable. This pattern marginally appears in British English where words like co-operate, geometry and reaction can be realized in careful speech as [kәƱʔ̲ɒ́pәreıt], [dʒiʔ̱ɒ́mәtrı] and [riːʔ̲ǽkʃәn] respectively (inserted ʔ is underlined; Cruttenden 2001: 169) or when contrasts like [dʒiʔ̲ɒ́grәfı] ‘geography’ vs. [dʒiːәƱgrǽfık] ‘geographic’ optionally emerge (Michael Ashby, p.c.). While this pattern is facilitated by the presence of stress on the second vowel, it is reported that it can also arise in other hiatus contexts in the absence of stress (Cruttenden 2001: 169). However, even in cases where stress is assigned on the second vowel, the pattern is not systematic. For this reason, I will instead use data from Dutch (§2.3.2.1), where the phenomenon is robustly documented, as well as Ainu (§2.3.2.2). 2.3.2.1 Dutch SO effects emerge in Dutch by means of ʔ-insertion in front of the stressed syllable under hiatus. The relevant data are presented below.

64 (37)

Onsets Dutch ʔ-insertion under hiatus when V1 = /a/ (Booij 1995: 65) a. ʔ-insertion /paεlja/ /aɔrta/ /kaunda/

[pa.ʔ5l.ja] [a.ʔɔ́r.ta] [ka.ʔún.da]

‘paella’ ‘aorta’ Kaunda [Zambia’s first president]

[xáː.ɔs] [fáː.ra.oː]20

‘chaos’ ‘Pharaoh’

b. no ʔ-insertion /xaɔs/ /farao/

The data above present a hiatus environment where the first vowel is /a/.21 In the first set of examples (37a), a glottal stop is inserted between /a/ and the following vowel on the surface. The second set illustrates a near-identical environment, where ʔ-insertion fails. The crucial difference is that the underlyingly onsetless syllable in (37a) is stressed in the output in contrast to the one in (37b), which is not. We can therefore argue that the reason the glottal is inserted in (37a), but not in (37b), is because ALIGNσ́O is in action, positing a requirement for a stressed syllable to begin with an onset. If ALIGNσ́O >> Dep-C then a consonant is going to be inserted to fulfil this need. No similar requirement, however, applies for unstressed onsetless syllables, therefore any such insertion would violate Dep-C pointlessly. Note that the inserted consonant, as we see, is the glottal stop /ʔ/, a consonant which is unmarked in terms of its place specification (Prince and Smolensky 1993/2004) under most theories of place markedness (Alderete et al. 1999; Lombardi 2002; de Lacy 2006). Moreover, by being voiceless, /ʔ/ is entirely compatible with the idea put forward in this book, namely that the lack of voicing increases the pitch of the following vowel providing a cue for stress. Here, the stress algorithm alone decides where stress will dock, but if it happens to be on an onsetless syllable, then the addition of /ʔ/ renders such a syllable more well formed by virtue of the presence of an onset. To simplify things slightly, I will use DEP-ʔ instead of Dep-C.

20

21

Carlos Gussenhoven (p.c.) observes that in a word like [fáː.ra.òː], there is also secondary stress on the final syllable. Insertion of /ʔ/ there is optional and at any rate seems different from the one affecting main stress, where insertion is obligatory. Booij does not offer the syllabification of these words, but this can be inferred by statements regarding syllabification earlier in his book. For instance, Booij (1995: 36) reports that branching onsets cannot contain two sonorant consonants, banning combinations of nasals with liquids or glides, or liquids with glides. Given the consonant inventory of Dutch too (p. 7), /lj/ cannot be interpreted as the palatal lateral /ʎ/ either. For this reason, I take it that /lj/ genuinely refers to this consonant sequence, which, given the above, has to be syllabified as a coda-onset one. Marc van Oostendorp (p.c.) also confirms that this is the right syllabification.

Onsets and stress

65

2.3.2.1.1 Background on Dutch stress The first step in understanding the phenomenon is to provide some background on Dutch stress based on Gussenhoven (2008). This will be quite general and brief, as numerous detailed accounts of Dutch stress are available (see van der Hulst 1984; Lahiri and Koreman 1988; Kager 1989, among others). Simplifying a bit, Dutch primary stress is assigned from right to left forming a quantity-sensitive trochee, thus Trochee (‘feet are trochaic’) must be ranked high. Secondary stress is assigned from left to right in a quantity-insensitive manner. Primary stress falls on the penult if it is heavy, i.e. CVV or CVC as in e.g. a.xáː.ta ‘Agatha’ or a.mán.da ‘Amanda’. Since CVCs are treated as heavy, Moraic Coda (or WbyP) should be active, although with a caveat: it does not apply beyond the left of primary stress. WSP must therefore be quite high ranking so that heavy syllables attract stress. The opposite also occurs, namely that stressed syllables are heavy, suggesting that Stress-to-Weight (S-to-W) is also high ranking. If the penult is light, stress goes on the antepenult, as in máː.ra.tɔn ‘marathon’. Final stress is generally avoided, because NonFin (‘main stress is not word final’) dominates Align-R (HdFt, PrWd) (‘the head foot is at the right edge of the word’). Also NonFin >> WSP holds, as the example (áːrɔn) rather than *a(rɔ́n) indicates. There is an exception to final stress avoidance, though: if the final syllable is super-heavy, i.e. contains a CVVC or CVCC, then stress docks onto that syllable, e.g. àt.mi.ráːl ‘admiral’ or lèː.di.kánt ‘bed’. According to Gussenhoven, this can be accounted for by imposing the constraint Super-Heavy-to-Stress-Principle (SHSP), an extended version of WSP that refers to trimoraic syllables (but see §2.4.1 for an alternative gradient evaluation of WSP). If this is ranked above NonFin, then we will correctly get final stress in kanáːl instead of penultimate *kánaːl. Finally, adjacent stresses are not allowed, therefore NoClash has to be dominant. In the subsequent analysis, I will follow Gussenhoven (2008) as faithfully as possible, although there are some weak – mostly technical – points in his analysis, such as the claim that WbyP does not work beyond the left of main stress or that FtBin can be satisfied either at the moraic or syllabic level, allowing, for instance, a quadrimoraic bisyllabic foot such as (áːrɔn), which would by most analyses be considered two feet. However, all these are tangential to the main point made here. Since I argue that ALIGNσ́O – my focus here – does not affect the stress algorithm, but merely turns onsetless syllables which have been designated stressed into onsetful ones, then the exact analysis of Dutch stress does not really matter, so long as it manages to produce the right stress. As far as I can see, what would most likely vary among different analyses is the footing assumed, but again this poses no threat to the ensuing analysis.

66

Onsets

2.3.2.1.2 The core analysis It should be obvious that the data in (37) are compatible with the facts mentioned above. It is not the case that ALIGNσ́O affects the general stress algorithm. Had it done so, then since feet are formed from rightto-left and assuming ALIGNσ́O dominates WSP and NonFin, we could perhaps expect to stress the word ‘aorta’ as aɔrtá instead of aʔɔ́rta since in the former the first onsetful syllable receives stress. This would be more or less what seems to happen in Juma mentioned in the discussion of the OS effects, where the first onsetful syllable from the right receives stress. Dutch is not like that, though. It will assign stress following the normal stress algorithm, but if it so happens that stress falls on an onsetless syllable in a hiatus environment, then it will insert a glottal stop. Thus, ALIGNσ́O must be dominated by the constraints mentioned before, which are responsible for Dutch stress. This is illustrated below. For convenience, I group the constraints above under the tag Stress Constraints. (38)

ʔ-insertion22

a. b. ☞ c.

/aɔrta/

Stress Constr

a.ɔr.(táː) a.(ɔ ́r.ta) a.(ʔɔ ́r.ta)

*! (NON-FIN)

ALIGNσ́O

DEP-ʔ

*! *

(38a) perfectly satisfies ALIGNσ́O – as does the winner – but violates at least one of the constraints that regulate the stress algorithm in the language, e.g. NonFinality. (38b) satisfies the stress constraints, but fails with respect to lowerranked ALIGNσ́O. Because of the ranking ALIGNσ́O >> DEP-ʔ, there is an additional candidate to be considered, namely one with glottal stop insertion, which actually proves to be the winner.23 This ranking can also neatly produce the absence of ʔ-epenthesis, when the onsetless syllable involved is stressless, as in xáɔs. (39)

Lack of ʔ-insertion Stress Constr

a. b. c. ☞ d. 22 23

/xaɔs/

NonFin

xa.(ɔ ́s) xa.(ʔɔ ́s) (xáː.ʔɔs) (xáː.ɔs)

*! *!

WSP

ALIGNσ́O

DEP-ʔ

* *(WSP) *(WSP)

* *!

I assume that the constraint Max dominates at least DEP-ʔ so that no input material is lost. Gussenhoven argues that feet in Dutch are either (HL) or (H). Examples (38b) and (38c) are footed as (HL) accordingly. The same holds for (41b) and (41c) below.

Onsets and stress

67

To understand why (39d) is chosen here, we need to make explicit reference to the constraints NonFin and WSP which are contained in the generic Stress Constr group. Their relative ranking is NonFin >> WSP as argued above. This comes in handy now. Candidates (39a) and (39b) stress the final syllable violating NonFin, which requires that final syllables are stressless. (39c) and (39d) avoid violation of NonFin by stressing the penult. But this comes at a cost; since the final closed syllable is argued to be heavy, then it incurs a WSP violation by remaining unstressed. One way or another, (39c) and (39d) tie at this point, and therefore the evaluation continues. The crucial bit comes here. Unlike in (38), ALIGNσ́O is vacuously satisfied in (39c) and (39d), because although they include an onsetless syllable, this is also unstressed. Since there is no requirement that stressless onsetless syllables acquire an onset, (39c) and (39d) escape violations of ALIGNσ́O. As a result, the addition of a ʔ-onset as in (39c) gratuitously violates DEP-ʔ and consequently loses. (39d) without ʔ-insertion is chosen. As it stands, this analysis has an undesirable result. It predicts that an initial onsetless syllable should acquire an onset if it is stressed. This is bound to happen, because of the ranking ALIGNσ́O >> DEP-ʔ, thus an example like [óli] ‘oil’ should actually surface as *[ʔóli] instead. How can this problem be resolved? Positional faithfulness (Beckman 1999; Casali 1996) offers a way out. Positional faithfulness has been designed to account for instances of positional neutralization and preservation of contrasts. It has been repeatedly found that a contrast may be retained or lost in all positions, but if it is to be lost in a specific environment, then this will be a weak position, not a strong one. The general schema that produces contrast maintenance in a strong position only is given by: (40)

Contrast in strong position only FaithPositional >> M >> Faith

This means that a contrast will be generally lost because Markedness dominates general Faithfulness, except in a strong position, since Faithfulness relativized to this position protects the contrast from being lost. We can now transfer this schema to the case at hand, replacing Faith with DEP-ʔ, M with ALIGNσ́O and FaithPositional with DEP-ʔinit-σ (see McCarthy 2002: 32–6). Empirically, this means that in Dutch we should be able to find both onsetful and onsetless initially stressed syllables, but word medially, stressed syllables should all be onsetful. The contrast between onsetful and onsetless stressed syllables is thus lost in the weak medial position, but maintained initially. The following tableau exemplifies.

68 (41)

Onsets No ʔ-insertion in stressed onsetless syllables word initially /oli/ a. o(líː)

Stress Constr

DEP-ʔinit-σ

ALIGNσ́O

DEP-ʔ

*! (Non-Fin)

b. (ʔóːli)

*!

☞ c. (óːli)

* *

2.3.2.1.3 Dutch summary The preceding discussion has illustrated how ALIGNσ́O may not only force stress shift to an onsetful syllable (§2.3.1), but also how it can force insertion of a consonant in a stressed syllable so that it becomes onsetful (§2.3.2). The analysis of Dutch, however, centred on the exploration of hiatus whereby the first vowel is /a/. This is because /ʔ/-insertion only applies there. Elsewhere, glide formation is preferred to resolve hiatus, regardless of stress considerations. For a detailed analysis that incorporates the complete range of facts, see Topintzi (2006b, §3.3.2.3). 2.3.2.2 Ainu Ainu (Shiraishi 1998, following Tamura 1996; Shiraishi, p.c.), a language isolate spoken in Japan, is another candidate for an SO effect rather similar to the Dutch one. In the Saru dialect that Shiraishi has studied, one may observe the avoidance of stress on onsetless high vowels. In this language, stress can only dock onto one of the first two syllables, the preference being the peninitial, e.g. a.pá ‘entrance’ unless the first hosts a heavy (C)VC syllable, in which case stress appears on it, e.g. áp.to ‘rain’. In addition, glide formation emerges from underlying /i/ and /u/ vowels, but only when these are within the stress window. To visualize, consider the following examples. (42)

a. Glide formation applies /ko-i-o-map/ [koyómap]

*ko.í.o.map

‘to show affection to, love other person’s child’ (Tamura 1996: 355) /ku-i-or-ot/ [kuyórot] *ku.í.o.rot ‘to join to’ (Tamura 1996: 263) /e-u-e-tusmak/ [ewétusmak] *e.ú.e.tus.mak ‘to compete with’ (Tamura 1996: 144)

b. Glide formation does not apply24 /i-ko-i-o-map/ [ikóiyomap] *ikoyomap

24

‘to love, show affection to another person’s children’ (Tamura 1996: 225)

I am grateful to Toshi Shiraishi (p.c.) for discussion on this point. He notes that the examples in (42b) indeed do not undergo the glide formation in (42a) – which as I claim is induced by stress

Onsets and stress /yay-ko-u-epeker/

[yáykou wepeker]

*yaykow epeker

69

‘to think over one’s troubles’ (Tamura 1996: 858)

Glide formation then serves to offer an onset to an otherwise onsetless syllable, but only if that syllable is to be stressed (42a). Like Dutch then, Ainu remains faithful to its normal stress algorithm, but when it can turn a (high) vowel into a glide, it will, as a response to ALIGNσ́O.25 Outside the stress window, glide formation is inapplicable (42b). This view also neatly accounts for another set of data that Shiraishi leaves undiscussed: forms like /i-omap/ ‘to show affection to children’ and /u-epeker/ ‘to tell a folktale/a folktale’ end up [iyómap] and [uwépeker], respectively. If ALIGNσ́O drives glide formation, then it understandably applies here. The glide can now constitute the onset of the second syllable, the preferred stress position in Ainu. Alternative forms such as *[yomáp] and [wepéker] are possibly ruled out because of the special properties of initial positions. As McCarthy and Prince (1993) note, it is not uncommon for languages to permit onsetless syllables word initially but nowhere else, as in e.g. Axininca Campa. Ainu is presumably like that, with the added proviso that onset creation (through glide formation) is only enforced within the initial disyllabic stress window. 2.3.3 Why ALIGNσ́O? This section justifies the use of ALIGNσ́O over other formulations of similar constraints and explains why the onset effects it covers cannot be captured by the onset-weight constraints. The definition of ALIGNσ́O is repeated below. (43)

ALIGNσ́O: Align-L (σ́, C), i.e. Align the left edge of every stressed syllable with a consonant

Direct reference to stressed syllables like in (43) is not uncommon; Nelson (1998) utilizes AnCHOR-σ́ for French truncation, while Madsen (2000) and Beckman (1999) make reference to MAX-σ́ in Spanish blend formation and Scots Gaelic, respectively. ALIGNσ́O is virtually identical to Hyde’s (2007)

25

requirements – but instead emerge with glide insertion beyond the initial stress window. This seems a different process that occurs independently to provide for a syllable onset. Ainu has a further complicating factor. In cases where the underlying high vowel is preconsonantal, stress appears on the initial onsetless syllable that now includes a diphthong, e.g. /e-u-ka-oma-p/ > [éu.ka.o.map] ‘a hut with a triangular roof’. Glide formation here cannot be induced by ALIGNσ́O, since turning /u/ into a glide would form no onset. Turning /e/ into a glide would be rendered an onset, but no glides from mid-vowels can occur. It must also be the case that [eu] is heavy – which is reasonable if it is a diphthong (Shiraishi does not comment on this point) – because otherwise we would get *[eu.ká.o.map].

70

Onsets

Align-L (FtHd, C). As far as I can see, these constraints are empirically the same, but the latter is more explicit in allowing alignment of a similar sort for other prosodic categories. As I will not tackle this possibility here, I will stick to the earlier ALIGNσ́O, but maintain that its replacement by Align-L (FtHd, C) does not affect the argument and analyses made here. ALIGNσ́O has been inspired by Goedemans’ (1996, 1998) earlier AlignFtO constraint, which can be seen as a positional markedness constraint that favours the alignment of feet with consonants (Zoll 1998). (44)

Align-FtO: Align-L (Ft, C), i.e. Align the left edge of every foot with a consonant

There are some reasons why (43) should be preferred over (44). What (44) says is that the left edge of the foot has to start with a consonant. Although Goedemans does not discuss this issue, this constraint essentially must refer to trochaic languages, and not to iambic ones, because it only targets the left edge of a foot irrespective of its type. To see the problem, consider a bisyllabic trochaic foot (σ́σ) and a bisyllabic iambic foot (σσ́). In the first case, the left edge of the foot coincides with the left edge of the foot-head, i.e. the stressed syllable. In the latter, though, the left edge of the foot happens to be the left edge of the foot-tail, since the foot-head lies on the second syllable. Therefore, such a constraint makes the prediction that in an iambic system the onset effect will not appear on the stressed syllable, but on the stressless one. This seems a highly unlikely and potentially problematic prediction. Perhaps, Juma – to the extent that the data described according to Abrahamson and Abrahamson (1984) are accurate – offers a concrete example. In Juma, the final syllable gets stressed unless it is a [.V], in which case stress moves to the penult. Whether the foot is iambic or trochaic cannot be determined, but this shift would only work if the foot was trochaic, i.e. in final stress we would have a degenerate foot as in …σ (σ́)# and then with shift, a binary …(σ́ σ)# with the foot-edge coinciding with an onsetful syllable. If the foot is iambic, however, default final stress would work fine – assuming of course that the penult is onsetful – i.e. …(σ σ́)#, but stress shift to the penult would produce …(σ σ́) σ#. The problem is now that Align-FtO cannot account for this shift, since the left edge of the foot does not coincide with the stressed syllable which has to be onsetful. ALIGNσ́O, however, does not face this problem, since it directly refers to the stressed syllable. Another peculiar prediction of Align-FtO is the following. Consider a language which forms iambs from left to right, e.g. (CV. ́CV).CV.CV.26 26

I ignore here the strong tendency for iambs to have a heavy syllable in the foot-head and make no reference to additional feet that could be built in these strings, as they are not relevant to the argument.

Onsets and stress

71

Suppose further that this language shows the onset-sensitivity effects described in this section. Now, take the string V.CV.CV.CV. How would this be parsed? According to Alignσ́O, the left edge of the stressed syllable must begin with a consonant, yielding the parsing: (V. ́CV).CV.CV. Align-FtO, though, will select the parsing where the first syllable is left unparsed, i.e. V.(CV. ́CV).CV. Essentially, under Align-FtO, the stress algorithm for this language would say ‘stress the second syllable if the first is onsetful, but stress the third if it is onsetless’. On the other hand, Alignσ́O would consistently build iambs without such a seemingly poorly grounded stress shift. However, none of the languages in question clearly presents an iambic system to prove one or other formulation right. One possible exception nonetheless is Pirahã, which, as we will see in §2.4.1, also utilizes Alignσ́O. According to Green (1995), Pirahã is a mixed system where both trochees (for antepenultimate and penultimate stress) and iambs (for final stress) are employed. If this is right, only Alignσ́O can be used to yield the correct results for both trochaic and iambic patterns. But even so, the fact that Alignσ́O does not in the first place refer to any footing can be advantageous in cases where analyses agree on where stress is located, but not quite on the exact footing, as in Dutch (§2.3.2.1.1) or Juma (earlier in this section). Note that in this view, Alignσ́O is virtually identical to Smith’s (2005) ONSET/σ́́ constraint, which requires that a prominent position, like a stressed syllable, has an onset. However, there are two reasons why I do not choose this formulation. First, by using ONSET/σ́́ as the relevant constraint, one might assume that I espouse Smith’s theory as a whole. As §7.2.1 discusses, there are some difficulties in this model, particularly when stress is sensitive to the quality of onsets, rendering it inadequate to explain this phenomenon satisfactorily. The second reason is the following: if alignment is the right way to think of sensitivity to the presence of an onset, then under an analysis along the lines of Alignσ́O (or, for that matter, Align-FtO too), we could expect to find a corresponding constraint at the right edge, so that the right edge of a stressed syllable is closed by a consonant. This is not possible in Smith (2005), however, since it would imply the existence of a constraint CODA/σ́́, which is not predicted by the positional augmentation model she puts forward. And yet a potential example of this kind seems to appear in the Brazilian language of Kaingang (Yip 1992), a language lacking long vowels and geminates, but allowing the nasal codas m, n, ɲ, ŋ with their allophones and, more controversially, the voiced continuants w, y, r. What is interesting for our purposes is that under certain conditions, stress-attracting penultimate open syllables acquire

72

Onsets

a default velar nasal in reduplication – and sometimes outside reduplication too – so that the syllable becomes heavy and thus can more easily accommodate stress. A constraint like Align-R (σ́, C) would straightforwardly account for this fact. It is also conceivable to think of Italian Raddoppiamento Sintattico in a similar way. Basically what happens in this is that in a sequence of two words W1 and W2 where the last syllable of W1 ends in a stressed vowel and W2 begins with a consonant, the initial consonant of W2 geminates, providing a coda for the final syllable of W1 (Borrelli 2002; McCrary 2002). A question that comes to mind is whether the facts above can be analysed in a way other than using ALIGNσ́O. One alternative is that offered by Downing (1998), who observes that in some languages onsetless syllables are treated as ill-formed and consequently are excluded from certain prosodic processes such as stress, tone and reduplication. In the default situation, she explains, morphological and prosodic constituents are co-extensive with one another. However, there are also constraints which require that prosodic constituents are aligned with optimal, i.e. onsetful, syllables. If these constraints are high ranked, misalignment between morphological and prosodic constituents is enforced, leaving the ill-formed onsetless syllables outside the prosodic domains, i.e. they are extraprosodic. As a result, onsetless syllables do not participate in the prosodic processes in question, hence their exceptional behaviour. While this account has good empirical coverage particularly with respect to tone and reduplication, it fails to account for some instances of onsetsensitive stress. In particular, since it relies heavily on matching morphological constituents with prosodic ones, it is designed to act at the edges of morphemes and prosodic constituents. Thus, although it can account for the cases presented in this section, since they refer to the (left) edges of words, it fails to capture the facts in Pirahã where the onsetful syllables attract stress more than the onsetless ones also word medially and intra-morphemically as in [pia.hao.gi.so.ai.pi] ‘cooking banana’. On the other hand, a constraint like ALIGNσ́O makes no similar claim, and therefore it can freely apply to the case of Pirahã as well. It should be noted, however, that the increased frequency of ALIGNσ́O effects at the left edge of the word relates to the fact that most of the languages that present this phenomenon are Australian (especially of the Cape York and Arandic families). These languages are renowned for the historical process of initial dropping, that is, the loss of the initial consonant of the word and stress shift from the first syllable to the second. In this way *[ ́CVCV…] >

Onsets and stress

73

[V ́CV…].27 This then gives us the ‘stress the first onsetful syllable’ effect. The sequence of the events above is under debate. Blevins (2001) claims that stress shift occurred first followed by consonant loss.28 Goedemans (1998) suggests that the order was the reverse, i.e. that consonant loss served as the trigger of stress shift. Had it been the other way round, he argues, we should find languages where stress shift has already applied, but consonant loss has not yet. Evidence one way or another is not conclusive, and at any rate is orthogonal to the point of interest for us here, namely that – synchronically – onsetful syllables attract stress. More importantly, the diachronic explanation is on its own insufficient, since no evidence is available for a similar process having occurred in the native American languages (Banawá and Iowa-Oto). The use of ALIGNσ́O is therefore indispensable. A remaining alternative needs to be considered: would it be possible to understand the presence of onset effects (OS and SO) in terms of onset weight, as Gordon (2005: §10.1.2) attempts? This would obviously fit the present account and would plausibly provide an even more uniform account. However, it is not possible for both empirical and theoretical reasons. First, consider languages with OS effects. Rather than claiming that ALIGNσ́O is responsible for stress attraction on onsetful syllables, we could perhaps argue that onsetless syllables are monomoraic, whereas onsetful ones are bimoraic and consequently attract stress on them. There is, however, evidence that shows that onsetful syllables are not bimoraic. For instance, in Banawá minimal words are either CVV, e.g. fáa ‘water’, búu ‘beat’, or (C)V.CV, e.g. áwi ‘tapir’, téme ‘foot’. CV words are impossible, e.g. *fa, *bi. This is exactly what an ALIGNσ́O analysis of Banawá predicts, but not what an analysis that assigns weight on such onsets expects. As for SO effects, we could assume that in Dutch glottal-stop onsets contribute to weight, so that e.g. [ʔV] attracts stress as much as a heavy CVC (where onset C is not /ʔ/). One could then argue that actually all onsets in Dutch are moraic, so that CVC is trimoraic, whereas [ʔV] is bimoraic. It then falls out that CVC is a better stress bearer. But if CVC is also trimoraic, then we no longer have an explanation for the fact that Dutch avoids stressing final syllables, unless they are super-heavy CVVC or CVCC, e.g. àt.mi.ráːl 27

28

Since most of these languages lack complex onsets, this process presumably happened in some, but not all, cases, otherwise all words would be vowel-initial! Blevins (2001) reports numerous other Australian languages where she claims the sequence of facts was stress shift → C-loss. These include Mpalitjahn, Yinwum, Linngithigh, Anguthimri, Ngkoth, Mbiywom, Rimanggudinhma, Umbuygamu, Lamalama, Kuku-Thaypan, Mbabaram, Wamin, Yanga, Mbara, and more questionably Ogunyjan, Olgol, Oykangand and Nganyaywana.

74

Onsets

‘admiral’ or lèː.di.kánt ‘bed’. We would expect CVC to also be able to bear stress word finally, but this is not the case. Given that all three CVVC, CVCC and CVC share an onset and at least one coda, the difference in their weight properties and behaviour needs to be attributed to syllable weight that disregards the onset. The only other possibility, of course, would be to argue that CVVC and CVCC are hyper-heavy quadrimoraic (!) whereas CVC is only super-heavy trimoraic. But this would stretch the argument too far. While trimoraic syllables are relatively rare, they nonetheless exist, e.g. Estonian, Hindi (Hayes 1995), Old English (Hayes 1989), Hungarian, Sinhalese (Cohn 2003 and references therein) or Finnish (Anttila 2006). On the other hand, to my knowledge, no evidence exists for quadrimoraic syllables. But it may now well be the case that this issue needs to be reconsidered in the light of the introduction of moraic onsets, since, in theory at least, a sequence like [CμVμμCμ] is possible. I leave this possibility open for future investigation. 2.3.4 Interim summary Section 2.3 focused on the discussion of pattern ❷, i.e. the onset-presence-only (PO) effects relevant to stress. I argued that these effects basically come in two forms, OS and SO effects, both of which can be attributed to the influence of ALIGNσ́O that requires stress on an onsetful syllable. These are summarized below. (45)

(I) OS effects In languages like Aranda, Alyawarra, Banawá or Iowa-Oto (§2.3.1), ALIGNσ́O determines the stress algorithm. Stress is normally placed on the first syllable, unless it is onsetless, in which case it shifts to the second (onsetful) syllable. Relevant ranking here: ALIGNσ́O >> Align-L (PrWd, Ft) (II) SO effects In languages like Dutch and Ainu (§2.3.2), ALIGNσ́O does not determine the stress algorithm which is regulated by other prosodic principles. However, if stress falls on a syllable which is otherwise onsetless, ALIGNσ́O forces it to acquire an onset, e.g. Dutch: xá.ɔs but a.ʔór.ta. Relevant ranking here: ALIGNσ́O >> Dep-C

Section 2.3.3 also demonstrated the existence of evidence in support of the independent nature of the onset quality (§2.2) as opposed to onset presence effects (§2.3). This relates to both the inability of an exclusively weight-based analysis to extend to pattern ❷, as well as the observation that QO effects (❶) (Karo) vs. PO effects (❷) (Banawá or Dutch) can emerge separately in languages. Given this result, the logical conclusion is that onset-quality and

Onsets and stress

75

onset-presence effects can also co-occur in a language (Pattern ❸). I claim that Pirahã is such a case, as discussed in detail next. 2.4

PO and QO effects simultaneously in a language: the case of Pirahã (❸)

Pirahã is spoken by about 200 Indians in north-western Brazil. Its phonemic inventory is presented in (46) and its syllable structure is (C)V(V), which means that codas are banned. Single monovocalic syllables, i.e. V, are also prohibited. (46)

Pirahã phonemes Pirahã consonants p

t

k

b

ʔ

g s

h Pirahã vowels

i o/u a

Like Karo, this language too has tone, which again does not interfere with stress (K. Everett 1998). The data which show the independence of tone from stress are presented below. All vowels are specified for tone ( ́ = H tone, no accent = Low tone). (47)

Pirahã: tone independent of stress (P=voiceless onset, B=voiced onset) i) PVBV a. HH: tígí LL: pigi b. LH: sabí HL: ʔábi

‘small parrot’ ‘swift’ ‘mean’ ‘to stay’

ii) PVPV a. LH: tiʔí HL: tíhi

‘honey bee’ ‘tobacco’

Evidently, given that both H and L tones are possible after voiced and voiceless consonants, tone cannot be dependent on the preceding consonant’s voicing. Moreover, we see that stress stays constant, whereas the tonal patterns may change, which clearly suggests that stress and tone are independent of each other.

76

Onsets

The stress algorithm basically requires that stress emerges on the rightmost heaviest syllable within the trisyllabic window at the right edge of the word. Heaviness is defined according to the scale in (48). Pirahã, like other languages in the world, e.g. Greek, Italian, Spanish, only treats one of the last three syllables as possible stress bearers. The stress facts that confirm each of the steps in the weight hierarchy are presented in (49). Data are from Everett and Everett (1984), Everett (1988) and Everett (p.c.) (E&E=Everett&Everett 1984, E=Everett 1988). (48)

Pirahã weight scale PVV > BVV > VV > PV > BV (P=voiceless obstr.; B=voiced obstr. and sonorants)

(49)

Pirahã examples a. PVV > BVV káo.bá.baí pa.hai.bií pii.bi.gái b. BVV > VV bii.oá.ii poo.gáí.hi.aí

‘almost fell’ ‘proper name’ ‘deep water’

(E: 239) (E&E: 708) (Everett p.c.)

‘tired [literally: being without blood]’ ‘banana’

(Everett p.c.) (E&E: 709)

c. VV > PV pia.hao.gi.so.ai.pi ‘cooking banana’

(E&E: 710)

d. PV > BV ʔa.ba.gi ti.po.gi

‘toucan’ ‘species of bird’

(E&E: 710) (E&E: 710)

e. rightmost heaviest stress ʔa.ba.pa ‘Amapá’ (city name) *ʔa.ba.pa ho.áo.íi ‘shotgun’ *ho.áo.íi ti.po.gi ‘species of bird’ *ti.po.gi paó.hoa.hai ‘anaconda’ *paó.hoa.hai/*paó.hoa.hai

(E&E: 710) (E&E: 710) (E&E: 710) (E&E: 707)

Example (49e) in particular illustrates that when syllables are of the same weight, then the rightmost one receives stress. For example, in tipogi, both the antepenult and the penult can carry stress by virtue of being the heaviest in the word. Yet, only po does so, since it is the rightmost of the two. Observation of these data leads to the conclusion that weight depends on three different parameters, laid out below. (50)

Pirahã weight depends on i) VV vs. V ii) CV vs. V iii) PV vs. BV

The first parameter recognizes the heaviness of long vowels over short ones. This is a very familiar and recurrent pattern cross-linguistically. The

Onsets and stress

77

second factor refers to the ability of an onsetful syllable to attract stress more than an onsetless one. This is attributed to ALIGNσ́O, as discussed in §2.3. Finally, the third factor refers to the sensitivity to onset quality, as explored in §2.2. Pirahã is interesting because it combines all these factors in its weight and stress systems. The resulting analysis is presented below. 2.4.1 Analysis In Pirahã, voiceless obstruent onsets attract stress more than those with sonorant or voiced stops (for justification on their common patterning, see §2.4.2). As discussed before (§1.3.3.1 & §1.3.3.3), this distribution is accounted for by the relation between voicing and stress in analogy to the effects that voicing has on pitch in tone languages. Formally, then, to account for the distribution of moraic onsets in Pirahã, we only need the ranking below. (51)

Pirahã moraic onsets *μ/Ons/[+voi] >> Be Moraic >> *μ/Ons

In §2.2.1.2.1 I introduced the assumption with regard to the underlying moraicity of vowels (both short and long) and the use of superscripted moras after an onset to indicate a moraic onset. The same applies here and throughout. As a reminder of how onset moraicity is assigned, briefly consider the following example from Pirahã. (52)

*μ/Ons/[+voi] >> Be Moraic >> *μ/Ons for /tipogi/ tiμpoμgiμ ☞ a. b. c.

μ μ

μ μ

μ μ

μ

μ μ

μ μ

*μ/Ons/[+voi] μ

t i .p o .gi μ

t i .po .gi

μ μ

t i .p o .g i

*!

Be Moraic

*μ/Ons

*

**

**!

* ***

The last candidate leaves the competition early on, since by having moras on all onsets, it violates top-ranked *μ/Ons/[+voi] due to gμiμ. The second candidate avoids this problem by having no mora on the onset of gi, but also fails to assign a mora on one of the other onsets. As a result, it violates Be Moraic more severely than (52a), which manages to assign moras on all voiceless obstruent onsets. Onset moraicity of course has repercussions on the moraic make-up of Pirahã syllables, which is illustrated in (53).

78 (53)

Onsets How many moras per syllable? Total number of moras PVV = BVV = VV = PV = BV =

3μ 2μ 2μ 2μ 1μ

Onset

Mora composition

YES YES NO YES YES

(2μ-nucleic, 1μ-onset) (2μ-nucleic) (2μ-nucleic) (1μ-nucleic, 1μ-onset) (1μ-nucleic)

Two things seem clear at first glance. PVV is unambiguously heavier than all other syllables, in the same way that BV is the lightest. The former is trimoraic, whereas the latter is only monomoraic. A potential problem now is that the remaining syllable types, BVV, VV and PV, are all considered bimoraic. Thus, they should behave as equally heavy, and yet a well-defined hierarchy exists among them, i.e. BVV > VV > PV. To see why this should be the case, we need to take into account that BVV differs from VV in having an onset that the latter lacks. Such a preference for stressing onsetful syllables is captured by ALIGNσ́O, repeated below. (54)

ALIGNσ́O: Align-L (σ́, C) (see also §2.3)

This leaves us with the VV > PV part of the scale. PV not only has an onset, which is missing from VV, but it also happens to be a voiceless one, hence moraic. Nonetheless, it acts as lighter than VV. Observe, however, that VV has two moras too, but both come from the nucleus, unlike PV where one is linked to the nucleus and the other to the onset. In other words, this pattern indicates that nucleic weight takes priority over overall weight. Translating this in terms of constraints, we need the constraint which requires that heavy syllables receive stress (i.e. the Weight-to-Stress Principle (WSP cf. (55)) to appear in a special version where heavy syllables due to nucleic moras receive stress. Such a constraint can be called WSP(N) where N stands for nucleus (56). (55)

WSP: Heavy syllables are stressed (NB: gradiently assessed, i.e. a stressless bimoraic syllable incurs one violation; a stressless trimoraic syllable incurs two violations)

(56)

WSP(nucleus): Heavy syllables due to nucleic moras are stressed

Similar constraints have been proposed in Hammond (1999) with respect to English, where it is claimed that under certain circumstances CVV behaves as heavier than CVC. In fact, in cases where codas count for weight purposes, there are many languages that treat CVV and CVC as equally heavy, e.g. Brahui, Latin, Votic (Gordon 2006), but numerous others which observe a CVV > CVC weight hierarchy, e.g. Kashmiri (Morén 2000, 2001), Klamath, Chickasaw (Gordon 2002).

Onsets and stress

79

It is thus the combination of onset moraicity (51), the alignment of a stressed syllable with an onsetful one (54) and the WSP facts (55–56) that produce the complex system of Pirahã. Two more things need to be accounted for. First, we have seen that among equally heavy/light syllables (49e), the rightmost one receives stress. The constraint Align-Head-Right (57a) is responsible for this bias. Second, Pirahã strictly observes the trisyllabic window at the right edge for stress assignment. Although this issue is interesting on its own merit (see Green 1995), for current purposes, I will set it aside and merely assume the constraint in (57b). To simplify things, in the tableaux that follow, *Extended Lapse Right is assumed but not included. (57)

a. Align-Head-Right: Align the head syllable of a prosodic word to the right edge of the prosodic word (McCarthy and Prince 1993) b. *Extended Lapse Right: No sequences of more than two consecutive stressless syllables at the right edge of the word (Elenbaas and Kager 1999; Gordon 2005)

Putting all this together, the following ranking obtains (58), which can be shown through several ranking arguments, beginning with (59) which illustrates the heaviness of voiceless obstruent onsets vs. their voiced counterparts. (58)

Proposed ranking for Pirahã: WSP(N), WSP >> ALIGNσ́O >> Align-Hd-R

(59)

WSP >> Align-Hd-R PVV.BVV: toi.bai ‘woman’s name’ (Everett p.c.) toμiμbaμiμ ☞ a. b.

WSP

Align-Hd-R

μ μ μ

μ μ

*

*

μ μ μ

μ μ

**!

t o i .ba i

t o i .ba i

In a nutshell, misalignment of stress is compelled as a result of better satisfaction of the WSP by the first candidate compared to the perfectly aligned second one. It can be observed that violations of WSP are gradiently assessed so that (59a) incurs one violation and (59b) two. Although gradience is not uncontested (see McCarthy 2003a), there are various arguments in favour of it (see Hyde 2007 for discussion). However, the standard interpretation of WSP only asks that heavy syllables receive stress, thus each of (59a) and (59b) should incur just one violation of WSP, leading to a tie. Align-Hd-R would thus determine the outcome, by wrongly selecting (59b) as the winner. But this is not the only possible way to understand WSP. Note that the syllable [.tμoμiμ.] has three moras, whereas [.baμiμ.] has only two. Given that the former is heavier than the latter, it intuitively makes sense that when comparing heavy syllables, WSP is more severely violated if the heavier one [.tμoμiμ.] remains unstressed.

80

Onsets

There are at least two ways to tackle this issue. The first is to make use of PkProm: a superheavy syllable makes a better stress-bearer than a heavy syllable, and a heavy syllable a better one than a light syllable, i.e. μμμ > μμ > μ (adapted from Prince and Smolensky 1993, 2004). (60)

PkProm: μμμ > μμ > μ

PkProm would correctly choose (59a), since its super-heavy trimoraic syllable has priority for stress over the heavy second syllable. But even if PkProm is used instead of WSP, it cannot subsume the effects of WSP(N), which is important for the analysis as shown in (61) below. Since WSP(N) is used, we expect WSP to be present in the ranking too, and therefore it would be preferable to try to solve the problem by making use of this constraint. This is possible if we assume that WSP is gradiently evaluated depending on the moraic make-up of syllables. In some sense, then, WSP subsumes PkProm since it can now favour stressing a super-heavy syllable over a heavy one. This is why (59a) only receives one violation of WSP, whereas (59b) receives two; the former leaves a heavy syllable unstressed, while the second leaves a super-heavy one unstressed, making this a fatal violation. It is well known that while standard WSP and PkProm overlap to a certain extent, originally only PkProm could select a super-heavy stressed syllable over a heavy stressed one. This is no longer the case if WSP is amended as suggested. Nonetheless, calculation of violations is still different. WSP receives violations for every heavy/super-heavy syllable that remains unstressed, whereas PkProm is satisfied as soon as the heaviest syllable in the word is stressed. This may be just one syllable, leaving others – which are equally heavy – unstressed. For PkProm this makes no difference. All that matters is to stress one – the heaviest – syllable in the word. This difference may prove important before deciding to discard PkProm after the present modification of WSP, but I will leave this issue open for further investigation. With this in mind, we can move on and see why PkProm could not replace WSP(N) too. (61)

WSP(N) >> ALIGNσ́O29 PV.VV.PV: ho.aí.pi ‘type of fish’ (E&E: 710)30

29

30

Words in Pirahã must begin with consonants (Everett p.c.), so no disyllabic word of this type is available. One could wonder what would happen if in this tableau the relevant constraints were WSP(N) and Align-Hd-R instead. While the correct result would be generated too, we know that ALIGNσ́O is independently needed in rankings like the one in (64), where it is in conflict with Align-Hd-R. Therefore, its relative ranking with respect to other constraints needs to be established. Had (61) been WSP(N), ALIGNσ́O >> Align-Hd-R, then (61b) would have wrongly won due to perfect alignment. WSP(N) >> ALIGNσ́O ensures that this is not the case.

Onsets and stress hoμaμiμpiμ ☞ a.

hμoμ.aμíμ.pμiμ

b.

hμoμ.aμíμ.pμiμ

WSP(N)

81

ALIGNσ O ́ *

*!

Candidate (61a) with a stressed VV syllable is preferred over a PV stressed one. Note that neither WSP nor PkProm alone would be able to select the right winner here. Every syllable in this word is bimoraic, therefore no matter which is stressed, the results are the same. If PkProm is used, then no candidate violates it, because the heaviest syllable in the word happens to be a bimoraic one. On the contrary, if WSP is used, then every candidate would violate it twice, since in each candidate two heavy bimoraic syllables stay unstressed. In both cases, then, candidates would tie, and ALIGNσ́O would wrongly favour candidate (61b), since it stresses an onsetful syllable (and manages to align the stressed syllable with the right word edge). The use of WSP(N), however, resolves the problem. WSP(N) focuses on nucleic moras only. As a result, leaving heavy – due to onset weight – ho and pi unstressed causes no violation of WSP(N). On the other hand, ai is heavy due to nucleic moras, therefore staying stressless incurs a WSP(N) violation. Ranking WSP(N) over ALIGNσ́O generates the right result. So far, then, we have achieved two rankings: (62)

WSP >> Align-Hd-R

(cf. (59))

WSP(N) >> ALIGNσ́O

(cf. (61))

Further rankings are possible. First, in accordance with Panini’s Theorem – also stated as the Elsewhere Condition (Kiparsky 1973) – we expect that the more specific WSP(N) should rank over the general WSP, producing: (63)

WSP(N) >> WSP >> ALIGNσ́O, ALign-Hd-R

What remains to be done is to determine the ranking of the two bottom-ranked constraints. This is possible too; among syllables of equal weight, the preference is to stress the one that is also onsetful, so it must be that ALIGNσ́O >> Align-Hd-R (64). (64)

ALIGNσ́O >> ALign-Hd-R BVV.VV: gao.ii ‘proper name’ (E&E: 709) gaμoμiμμ ☞ a. b.

ALIGNσ O ́

μ μ μμ

ga o .i

μ μ μμ

ga o .i

ALign-Hd-R *

*!

82

Onsets

In the light of the above, the final ranking for Pirahã is shaped as shown below. (65)

Final ranking: WSP(N) >> WSP >> ALIGNσ́O >> Align-Hd-R

One final point needs to be illustrated. Recall that among equal candidates for stress, the rightmost one actually gets it. (66) presents a representative example. All the syllables in [kμoμʔμoμpμaμ] are bimoraic, owing one of their moras to the nucleus and the other to the onset. They are thus identical in terms of their weight structure. In this environment, we can see that the rightmost syllable will receive stress due to perfect right alignment. (66)

Align-Hd-R seen in action PV.PV.PV: ko.ʔo.pa ‘stomach’ (E: 239) koμʔoμpaμ

WSP(N)

a. kμoμ.ʔμoμ.pμaμ

WSP

ALIGNσ O ́ Align-Hd-R

**

**!

b. k o .ʔ o .p a

**

*!

☞ c. k o .ʔ o .p a

**

μ μ μ μ

μ μ

μ μ

μ μ μ μ

The next tableau illustrates how the right output is chosen when all the constraints discussed are simultaneously considered. (67)

PVV.BVV.PV.VV: poo.gáí.hi.aí ‘species of fruit’ (E: 209) poμμgaμiμhiμaμiμ

ALIGNσ O ́ Align-Hd-R

WSP(N)

WSP

μ μ

μ μ

μ μ

*

**

**

μ μ

μ μ

μ μ

b. …ga i .h i .a i

* *!

**

*

c. …gaμiμ.hμiμ.aμ iμ

*

**

☞ a. …ga i .h i .a i

*!

In this four-syllable word, the heaviest syllable is actually poo, but this is beyond the stress-bearing trisyllabic window, and thus can never receive stress. Of the remaining options, (67b) loses because it fails to stress either of the two heavy syllables due to nucleic weight, unlike its contenders that fare better with respect to WSP(N). (67a) and (67c) at that point tie, until ALIGNσ́O discards the candidate with the stressed onsetless syllable (67c). (67a) wins despite its violations of Align-Hd-R, which is too low ranked to matter. To sum up, the Pirahã weight system boils down to three different factors: a) priority of nucleic weight, b) moraicity of voiceless onsets and c) preference for onsetful stressed syllables. I have proposed an analysis that integrates all these points and generates the right results.

Onsets and stress

83

2.4.2 Sonorants pattern with voiced stops The discussion up to now has claimed that Pirahã sonorants behave like voiced stops, but so far no relevant examples have been presented in support of that. Everett (p.c.) offers such evidence. He notes that while there are no input sonorants in Pirahã (see the inventory in (46)), they may nonetheless arise as allophones of the voiced stops. In particular, input /b/ has three surface realizations: [b], [m] and the bilabial trill [b], where the last two are sonorants. Input /g/ emerges as [g], sonorant [n] or as the sui generis voiced apico-laminal alveolar-labial double flap [ɾ”ɾ̼] only reported in Pirahã. Finally, while [n] seems an unlikely allophone for /g/, this seems more reasonable in the light of the fact that Proto-Mura – the language family Pirahã belongs to and exists as sole survivor of – *d shifted to /r/ in Mura and /g/ in Pirahã (Everett p.c.). Some rules that regulate the distribution of the allophones of the voiced stops are given below. (68)

b → m/ pause__

(69)

b varies freely with b/ i or a__o

(70)

g varies freely with [ɾ”ɾ̼]/ o__i

According to E&E (1984: 708, 710), the following words are stressed as shown: (71)

a. ʔíbogi b. biísai

‘milk’ ‘red’

Each of these fulfils the environment where at least one of the rules above may apply. (72)

a. (69)/(70) apply b. (68) applies

ʔíBogi / ʔíboɾ”ɾ̼i / ʔíBoɾ”ɾ̼i miísai

The crucial observation is that in each of these cases, where a sonorant is involved stress is identical to cases where voiced stops appear instead. Therefore, we can safely assume that sonorants apart from [+son] are also specified as [+voi] in Pirahã. 2.5

Arabela, a case ambiguous between (❶) and (❸)

The final language to be discussed in this chapter is Arabela, a Zaparoan language spoken in Peru by about 50 Indians (Rich 1963; Payne and Rich 1988). Arabela is very interesting because although it generally presents a normal rhythmic stress algorithm, this pattern is disrupted in a particular context that has to do with the voicing of the onsets involved. My stress examples all come

84

Onsets

from Payne and Rich (1988). A preliminary description of stress appears in Rich (1963), which is superseded by Payne and Rich where the effects of onsets were noted in a published form for the first time. I will start with the basics of the language by first presenting the syllable structure and working my way to the segmental inventory (§2.5.1), since the latter is going to prove relevant for the stress facts we examine here (§2.5.2). A thorough analysis follows (§2.5.3). 2.5.1 Arabela syllables and segments Syllable structure is discussed in Rich (1963), but several points remain dubious. It is claimed that onsetless syllables and codas are allowed, but examples are usually ambiguous. For instance, words such as /rupaa/ ‘mouth’ or /siinu/ ‘to raise a creature’ are syllabified according to Rich (1963: 201) as [ru.pa.a] and [si.i.ni]. It is also noted that in such V syllables, as they are referred to, the V is always the same as the preceding one. This makes it quite likely that what we are talking about is actually a long vowel. As for codas, to the extent that these exist, most of the time they are glides, as in [kway.ni.nyu] ‘to hunt’ or [su.wo.kwaw] ‘ground corn’, although other sequences that seem to involve sonorant codas also arise, e.g. /sapartu/ ‘shoulder blade’ or /mante/ ‘moth’. Other work, however, presents Arabela as representative of a language that only has CV and CCV syllables (Levelt and van de Vijver 2004; Gussenhoven and Jacobs 2005). This conflict is not resolved in the more recent of the Arabela papers, i.e. Payne and Rich (1988), where no example includes anything that can be construed as a coda or as a long vowel. Importantly, though, even if codas do exist in the language, they are irrelevant for weight and stress purposes, e.g. (mòkoʃ)(tyáka) ‘palm fruit’ instead of *(mò)(kòʃ)(tyáka), which shows that codas are not moraic. We can now move to the phonemic inventory, which is presented in (73) and followed by some examples in (74). (73)

Arabela phonemic inventory vowels: i, e, a, o, u consonants: stops p, t, k, fricatives s, ʃ, h, nasals m, n; liquid r and glides w, y

(74)

a. stops pinyu

‘to hit’

tinyu

b. fricatives siyokwa

‘tucuayo bird’

ʃiyokwa ‘grease’

c. nasals miyano

‘plaything’

niyano

d. liquid – glides riyano ‘he is breathing’ tawe

‘to fall’

kinyu

‘to stay’

‘he is coming’ ‘foreigner’

hayunu ‘pulling’

Onsets and stress

85

While the language seems to lack voiced obstruents phonemically, according to Rich (1963) they nonetheless arise as variants of the voiceless stops in certain contexts which are unfortunately not specified. Rich’s relevant data are presented in (75). Note especially that in the morpheme meaning ‘father’ there is alternation between k, x and g. Rich (1963) suggests that these alternants occur freely and progressively lenite as one moves further within the phrase. But this is not supported by the more recent description of Payne and Rich (1988), where the only obstruents presented are the voiceless stops. For this reason, the term ‘voiced consonants’ in Arabela in practice refers to sonorant onsets. As a reminder to the reader of the [+voi] feature that sonorants bear, sonorants will occasionally be characterized as voiced. (75)

/ke/ /nake/ /kanaake/ /saako/

[kiʔ] [nãxiʔ] [kanããgiʔ] [saaɣoʔ]

‘father’ ‘his father’ ‘our (excl.) father’ ‘corn’

2.5.2 Stress in Arabela Arabela stress is rhythmic and normally creates trochees from left to right (76a). The rightmost stress is the primary one. Degenerate monosyllabic feet are also admitted, e.g. (sàma)(rú). According to Payne and Rich (1988), there is, however, an interesting exceptional pattern that involves stress shift (76b); they state that ‘if a word-final syllable that would have received stress has a voiced onset, and the immediately preceding syllable has a voiceless onset, then the syllable with the voiceless onset is stressed’ (1988: 6).31 (76)

Arabela stress a. Normal pattern (tèna)(kári) (sàma)(rú) (hùwa)(hàni)(yá)

‘afternoon’ ‘spirit’ ‘peaceful’

b. Stress shift (nòwa)(ʃ ì)(ʃáno) (sàpo)(hò)(sáno) (mwèra)(tì)(tyénu) 31

*(nòwa)(ʃ ìʃa)(nó) *(sàpo)(hòsa)(nó) *(mwèra)(tìtye)(nú)

‘brightened’ ‘deceived’ ‘cause to be seen’

As explained in the main text, the rhythmic groupings here are partly determined by onset voicing. This implies that the pitch perturbation of voicing must be interpreted as stress, not as tone. Note that it has been suggested that apart from metrical feet, tonal feet also exist in languages such as Hausa, Northern Mande, Isixhosa (Leben 2001) or Kera (Pearce 2005). Leben (2001), however, observes that the characteristic of a metrical foot is that it contains one stress per foot, whereas constituency in tonal feet is not as clear. Arabela obviously belongs to the languages with metrical feet.

86

Onsets

In (76b) we would presumably anticipate stress on the first, third and fifth syllables, but instead stress docks onto the first, third and fourth syllables. This pattern can be accounted for by making use of moraic onsets in a manner similar to Pirahã. In Arabela, too, voiceless onsets are moraic, but voiced, i.e. sonorant, ones are not. Thus, the moraicity of Arabela onsets is identical to that of Pirahã. Consonants with [−voi] attract stress more by virtue of their moraicity. Note that although all the cited examples relevant to stress shift involve final onset sonorants, Payne and Rich (1988) talk about voicing and not sonorancy. Had they intended reference to sonorants only, I believe they would have been explicit about sonorancy instead of voicing since they must have been aware of the voiced obstruents mentioned in Rich (1963), given that Rich is one of the authors of Payne and Rich (1988). Use of [voice] then as the relevant feature rather than [son] is further corroborated. The by-now-familiar onset-moraicity ranking in (77) captures the Arabela situation provided sonorants are specified as [+voi]. (77)

Onset moraicity in Arabela *μ/Ons/[+voi] >> Be Moraic >> *μ/Ons

The hypothetical word in (78) – based on the actual rupohonu word – exemplifies. To simplify, Arabela tableaux only present superscripted moras which accompany the relevant onsets, since no example involves long vowels (whose existence is generally questionable). (78)

*μ/Ons/[+voi] >> Be Moraic >> *μ/Ons pohonu

*μ/Ons/[+voi]

Be Moraic

*μ/Ons

a.

pμohμonμu

b.

pohonu

***!

c.

μ

**!

*

*

**

☞ d.

poh onu μ

μ

p oh onu

*!

***

The ranking in (77), however, gives us no information as to the footing of the winning candidate; pμohμonu could reasonably be parsed as [(pμòhμo)(nú)] or [(pμò)(hμónu)] among other possibilities. Numerous other ingredients are required to achieve the right results. These will be explored in the next section. Summing up this section, it has been evident that onsets in Arabela influence stress assignment. In particular, it is the quality of onsets that

Onsets and stress

87

matters, implying that Arabela is a type ❶ or type ❸ language. The actual classification depends on the language’s behaviour with respect to onsetless syllables and stress. However, evidence for the existence of onsetless syllables is lacking, so it is reasonable to follow Levelt and van de Vijver (2004) and Gussenhoven and Jacobs (2005) in assuming that the language only allows CV and CCV syllables. As a result, Arabela’s behaviour with respect to onsetless syllables cannot be tested for and consequently could equally be trivially grouped among languages like Karo where the presence of the onset does not matter for stress (❶) or languages like Pirahã where it does (❸). 2.5.3 Stress analysis This section offers an analysis of Arabela’s stress system. First, an account of the normal rhythmic pattern is provided (§2.5.3.1) and then the stress shift pattern is discussed (§2.5.3.2). It is argued that stress shift only affects primary stress which is systematically located at the right edge of the word; secondary stresses, on the other hand, never exhibit stress shift. The shift is enforced by the requirement to assign main stress on heavy syllables, which in Arabela translates to syllables with voiceless onsets. The full range of facts described by Payne and Rich (1988) is captured, while an additional potential case of stress shift is predicted, which at present cannot be confirmed by the available data (§2.5.3.3). The section closes with a brief discussion around the contextual character of the Arabela stress shift pattern (§2.5.3.4). 2.5.3.1 Rhythmic pattern Recall from (76b) that the default stress system of Arabela involves trochees rhythmically formed from left to right. Rightmost stress is the primary one, therefore the following two constraints are undominated in Arabela. Since these are never violated, I will not consider them in the tableaux that follow unless required. (79)

Trochee: Feet have initial prominence Align-HdFt-R: Align-R (HdFt, PrWd)

(Kager 1999) (Hyde 2001)

The head foot is rightmost in the prosodic word

Additional prosodic constraints are needed to account for the parsing of all syllables into feet, the directionality of footing and the size of feet. These are listed next.

88 (80)

Onsets PARSE-σ: Syllables are parsed by feet

(Prince and Smolensky 1993/2004)

32

FtBin(Max): Feet are maximally bisyllabic33 (Everett 1996) All-Ft-R: Align-R (Ft, PrWd) (McCarthy and Prince 1993) Align the right edge of every foot with the right edge of the PrWd All-Ft-L: Align-L (Ft, PrWd) (McCarthy and Prince 1993) Align the left edge of every foot with the left edge of the PrWd

Rightward directionality of footing establishes that All-Ft-R >> All-Ft-L. (81)

All-Ft-R >> All-Ft-L samaru

All-Ft-R

All-Ft-L

☞ a.

(sàma)(rú)

*

**

b.

(sà)(máru)

**!

*

FtBin(Max) needs to dominate All-Ft-R as illustrated below. (82)

FtBin(Max) >> All-Ft-R (cf. [hùwahàniyá]) σσσσσ

FtBin(Max)

a.

(σσσσσ)

*!

b.

(σσσ)(σσ)

*!

**

c.

(σσ)(σσσ)

*!

***

☞ d.

32

33

(σσ)(σσ)(σ)

All-Ft-R

****

The more traditional FtBin runs into trouble. PARSE-σ would need to dominate FtBin since all syllables are parsed into feet. All-Ft-R is needed to account for the rightward directionality of footing. The problem arises with the interaction of FtBin and All-Ft-R as observed in large odd-syllable words where either too many syllables could be stuffed into a foot or ternary feet are preferred. None of these actually occurs. The problem is resolved once we use high-ranking FtBin(Max) to allow feet to be unary or bisyllabic. In this case, FtBin(Min), i.e. feet are minimally bisyllabic, must be assumed to be very low-ranked since degenerate feet are admitted. FtBin requires feet to be binary either on the syllabic or moraic level. In Arabela we claim that FtBin holds on the syllabic level, thus feet have to be bisyllabic. This is interesting since the language is claimed to be weight sensitive, but is analysed as quantity insensitive when it comes to foot formation. Perhaps a similar claim can be made for iambic feet in Paumari. Based on the fact that in this language long vowels can surface in the non-head position of an iamb, Everett (2003) argues that they are monomoraic. It could be possible, though, to claim that long vowels are bimoraic, but that the language forms feet on the syllabic rather than the moraic level. This is quite strange given that generally iambs favour durational differences (cf. the Iambic/ Trochaic Law in Hayes 1995). However, Everett shows convincingly that Paumari does not seem to adhere to the law for independent reasons, thus making the QI formation of feet less peculiar.

Onsets and stress

89

The first three candidates fail FtBin(Max) as they include either ternary or even larger feet, which exceed the bisyllabic maximum. The winning candidate is (82d), since its unary and bisyllabic feet perfectly satisfy FtBin(Max) even at the expense of All-Ft-R. We can also easily show that PARSE-σ >> All-Ft-R. (83)

PARSE-σ >> All-Ft-R PaRSE-σ

tenakari ☞ a. b.

(tèna)(kári)

All-Ft-R **

tena(kári)

**!

The constraints PARSE-σ and FtBin(Max) present no conflict with one another, so we can assume that they are placed next to each other (84). (84)

PARSE-σ, FtBin(Max) PARSE-σ

samaru ☞ a. b.

FtBin(Max)

(sàma)(rú) (sáma)ru

*!

The rankings we have thus far established are presented in (85). (85)

a. Moraic-onset ranking: *μ/Ons/[+voi] >> Be Moraic >> *μ/Ons b. Prosodic ranking: Trochee, Align-HdFt-R, PARSE-σ, FtBin(Max) >> All-Ft-R >> All-Ft-L

2.5.3.2 Stress shift Up to this point, the rankings in (85) appear unrelated. The link between them is provided once one considers the stress shift data, repeated in (86). (86)

Arabela stress shift nòwaʃ ìʃáno sàpohòsáno mwèratìtyénu

*nòwaʃ ìʃanó *sàpohòsanó *mwèratìtyenú

‘brightened’ ‘deceived’ ‘cause to be seen’

In these cases, it can be observed that while secondary stress is rhythmic, primary stress shifts from the anticipated ultima that hosts a sonorant onset to the penult with a voiceless onset. Presumably, it is the added weight of the penult – due to its moraic onset – that causes such shift. Obviously, the WSP can be called into action to achieve this result. Importantly, however, such stress shift

90

Onsets

only occurs with regard to primary stress, implying that there is an asymmetry between primary and secondary stress. Primary stress seems to be weight sensitive, whereas secondary stress is not. Asymmetries between primary and secondary stress are well attested. McGarrity (2003) discusses a range of patterns, one of which is exactly the type found in Arabela. Other languages with quantity-sensitive main stress, but with quantity-insensitive secondary stress, include: Huariapano, Inga, Seneca, Maung, Djabugay, Yukulta and Surinam Carib. A straightforward way to account for this asymmetry is to invoke a specific version of WSP that focuses on main stress, i.e. WSPPrWd (87). As before, due to the Elsewhere Condition (Kiparsky 1973), WSPPrWd should rank above WSP, since it is more specific. Note that constraints relativized to primary stress are argued for by several theorists, based on the fact that stressed positions are strong and as such able to exhibit special properties as opposed to unstressed positions (Casali 1996; Beckman 1999; Smith 2005; de Lacy 2006). (87)

WSP: Heavy syllables receive stress WSPPrWd: Heavy syllables receive primary stress (cf. McGarrity 2003)

We are now in a position to account for the contextual stress shift found in Arabela. This is due to the ranking of WSPPrWd over All-Ft-R, as illustrated by the word sàpohòsáno. This example also illustrates the importance of utilizing WSPPrWd. Had the more general WSP applied only, then we would expect multiple secondary stresses that we actually do not get (cf. 88b). (88)

voiceless–voiced WSPPrWd >> All-Ft-R >>WSP WSPPrWd

All-Ft-R

WSP **

a.

(sμàpμo)(hμòsμa)(nó)

****!

**** (4)

b.

(sμà)(pμò)(hμò)(sμáno)

***

******!*** (9)

(sμàpμo)(hμò)(sμáno)

***

***** (5)

☞ c.

*

The rhythmic stress pattern (88a) loses early on, because by stressing the final syllable with a non-moraic sonorant WSPPrWd is fatally violated. WSPPrWd asks that all heavy syllables receive primary stress, so (88b) and (88c) tie as they only assign main stress on one (in fact the rightmost one) of the heavy syllables. Consequently, the evaluation is passed on to All-Ft-R which favours (88c). Three points are worth noting at this stage: first, WSP alone is not sufficient to produce the correct result, because if it dominated All-Ft-R, it

Onsets and stress

91

would wrongly pronounce (88b) as the winner. Dominance of All-Ft-R over WSP would be equally undesirable, as it would pick the rhythmic candidate (88a). WSPPrWd resolves this problem. Second, we need to ensure that a candidate like (sá)(pó)(hó)(sáno) is not the winner. This is identical to (88b) with the exception that all its stresses are primary and thus completely obedient to WSPPrWd. We can reasonably assume, however, that each word assigns one primary stress, due to Culminativity (89), so such a form would be impossible. Ranking Culminativity above WSPPrWd is sufficient to rule out (sá)(pó)(hó)(sáno) and let the remaining constraints in (88) select the winner (88c/90c). (89)

Culminativity: There is only one primary stress per word (Liberman and Prince 1977: 262; Hayes 1995: 24)

(90)

Culminativity >> WSPPrWd >> All-Ft-R >>WSP Culminativity a. ☞ b. ☞ c.

(sμá)(pμó)(hμó)(sμáno) μ

μ

μ

***!

μ

(s à)(p ò)(h ò)(s áno) μ

μ

μ

WSPPrWd

***

μ

(s àp o)(h ò)(s áno)

***

While the preceding discussion has located WSPPrWd in the prosodic hierarchy of (85b), it has not yet positioned it in the moraic-onset hierarchy of (85a). Consider again sàpohòsáno. In order that this wins over the rhythmic sàpohòsanó, WSPPrWd needs to be employed, which of course is sensitive to the heaviness of syllables, which in its turn is dependent on the moraic-onset hierarchy. Let us consider four candidates in total: one that has rhythmic stress (91b); one that has stress retraction (91a); and the moraic counterparts for each of these two (91d, b). (91)

Be Moraic >> WSPPrWd >> All-Ft-R Be Moraic

WSPPrWd

All-Ft-R

a.

(sàpo)(hò)(sáno)

****!

***** (5)

b.

(sàpo)(hòsa)(nó)

****!

**** (4)

☞ c.

(sμàpμo)(hμò)(sμáno)

*

***

***** (5)

d.

(sμàpμo)(hμòsμa)(nó)

*

****!

**** (4)

To simplify the tableau, I have only included the most relevant constraints and only the candidates that have voiceless moraic onsets in accordance with (85a).

92

Onsets

Inclusion of Be Moraic is indispensable. Without it, WSPPrWd would be vacuously satisfied by candidates (91a, b), leaving All-Ft-R to pick the rhythmic candidate (91b) as the winner. The empirical data, however, suggest that stress shift occurs in this instance. To produce this effect, onset moraicity needs to be taken into account so that the non-moraic candidates (91a, b) are dispensed with. WSPPrWd can now correctly choose (91c). The resulting rankings that incorporate the prosodic and moraic onset aspects as well as the WSPPrWd appear in (92). (92)

Mini-grammar of Arabela *µ/ONS/[+voi]

CULMINATIVITY

TROCHEE, ALIGN-HDFT-R, PARSE-σ, FTBIN(MAX)

BE MORAIC *µ/ONS

WSPPrWd

ALL-FT-R WSP

ALL-FT-L

2.5.3.3 Stress and the range of voiceless–voiced sonorant combinations According to the description of the language (Payne and Rich 1988), in words with an odd number of syllables where the ultima has a voiced onset and the penult has a voiceless one, stress appears on the penult rather than the anticipated ultima. This section, then, explores two issues: first, whether the proposed analysis, offered in (92), accounts for all possible combinations of voiced sonorants and voiceless consonants in the last two syllables of the word; and second, whether the appearance of a voiceless onset elsewhere disrupts the normal algorithm. In practice, the following cases have to be investigated: (93)

Full range of sonorant-voiceless onsets in the final two syllables of oddsyllable words34 Penult a. Voiced (Sonorant) b. Voiced (Sonorant) c. Voiceless d. Voiceless

34

Ultima

Example

Anticipated Result

Voiced (Sonorant) Voiceless Voiceless Voiced (Sonorant)

hùwahàniyá

no shift to penult

sàkamànahá kòkotáka sàpohòsáno

no shift to penult no shift to penult shift to penult

Since *μ/Ons/[+voi] >> Be Moraic >> μ/Ons has been established and to simplify representations as much as possible, from now on I will no longer indicate the moras on voiceless onsets (unless required for illustration purposes), but will simply assume them.

Onsets and stress

93

As shown before, it is the case that only in (93d) is there shift of stress. It remains to be seen whether the other results are generated given the system proposed above. Let us first examine an example where both final syllables contain a sonorant onset. (94)

voiced–voiced Align-HdFt-R >> WSPPrWd >> All-Ft-R Align-HdFt-R

WSPPrWd

All-Ft-R

(hùwa)(hàni)(yá)

**

**** (4)

b. (hùwa)(hà)(níya)

**

*****! (5)

*

**** (4)

☞ a.

c.

(hùwa)(háni)(yà)

*!

Given this ranking, the rhythmic pattern is correctly identified as the winner. However, observe the importance of high-ranked Align-HdFt-R, which requires that the head-foot is aligned to the right edge of the prosodic word. Candidate (94c) presents main stress on the heavy syllable ha and in this way satisfies WSPPrWd more satisfactorily than its contenders. Despite that, it loses, as main stress is further away from the right edge of the prosodic word. Finally, candidate (94b) presents gratuitous stress shift and loses on All-Ft-R. (95)

voiced–voiceless Align-HdFt-R >> WSPPrWd >> All-Ft-R Align-HdFt-R ☞ a. (sàka)(màna)(há) b. (sàka)(mà)(náha) c. (sà)(kà)(màna)(há)

WSPPrWd

All-Ft-R

**

**** (4)

***!

***** (5)

**

*****!*** (8)

In (95) there is no trigger for stress shift, since the ultima, which contains a voiceless onset, would normally receive primary stress due to rhythmic stress. Thus, (95a) is unsurprisingly chosen. The final case we need to consider is one that involves voiceless onsets on both the final two syllables. Again, the rhythmic pattern wins. (96)

voiceless – voiceless35 Align-HdFt-R >> WSPPrWd >> All-Ft-R

35

I have unfortunately been unable to find an odd-syllable word that would illustrate this pattern (cf. (93c)). However, I tested the proposed grammar with a hypothetical word that would serve as an example and the correct – according to Payne’s and Rich’s (1988) account – results emerged, i.e. no stress shift.

94

Onsets Align-HdFt-R

WSPPrWd

All-Ft-R

(kòko)(táka)

***

** (2)

b.

(kò)(kò)(tà)(ká)

***

***!*** (6)

c.

(kòko)(tà)(ká)

***

***! (3)

☞ a.

The example above has an additional advantage: it is an instance where rightward stress shift (cf. (96c)) could perhaps be expected, given that the final syllable has a voiceless onset. This now brings us to the second issue investigated in this section. Do voiceless consonants in positions earlier or later than the penult cause stress shift? In (96) it is evident that such stress shift does not occur. However, it is conceivable that in other cases it does. Consider, for instance, a case with a voiceless onset in the ultima and a sonorant one in the penult, like the hypothetical /kokonaka/. The crucial difference with (93b) – where no stress shift occurs – is that here the word contains an even number of syllables. Our system actually predicts (kòko)(nà)(ká) with rightward stress shift, instead of the rhythmic pattern (kòko)(náka). This is because the drive to satisfy WSPPrWd compels shift of stress to the right as in (97a). The alternative [(kóko)(nàka)] (97e) is no option as it violates top-ranked Align-HdFt-R. At present, no empirical data are available that verify this prediction, but it should be held in mind for future investigation. (97)

even-σ-word: voiced–voiceless – rightward stress shift Align-HdFt-R >> WSPPrWd >> All-Ft-R Align-HdFt-R

WSPPrWd

All-Ft-R

☞ a.

(kòko)(nà)(ká)

**

*** (3)

b.

(kò)(kòna)(ká)

**

****! (4)

c.

(kòko)(náka)

***!

** (2)

d.

(kò)(kò)(náka)

***!

***** (5)

e.

(kóko)(nàka)

**

** (2)

*!

On the other hand, stressing syllables with voiceless onsets early in the word does not offer a similar WSPPrWd reward at the right edge; instead, multiplication of feet implies multiplication of right foot alignment violations. (98) exemplifies. (98)

voiceless onsets early in word – no rightward stress shift Align-HdFt-R >> WSPPrWd >> All-Ft-R

Onsets and stress Align-HdFt-R ☞ a. (sàka)(màna)(há) b. (sà)(kà)(màna)(há) c. (sà)(kàma)(náha) d. (sà)(kàma)(nà)(há)

95

WSPPrWd

All-Ft-R

**

**** (4)

**

*****!*** (8)

***!

****** (6)

**

*****!** (7)

Candidate (98c) is first ruled out because it fails to assign primary stress on any of the heavy syllables. The remaining candidates compete in terms of better satisfaction of All-Ft-R. The most successful proves to be the rhythmic candidate (98a) rather than the others which exhibit rightward stress shift. Summarizing the preceding analysis, it has been shown that the Arabela stress system is generally rhythmic, requiring all syllables to be parsed into binary and, where needed, unary feet at the right edge of the word satisfying All-Ft-R to the highest degree possible. This rhythmic pattern occurs in spite of the dominant Align-HdFt-R and WSPPrWd whose effects are usually concealed, as their influence can only be seen under certain circumstances. This is because their combined effect, i.e. stress shift, can only arise on the final head foot (Align-HdFt-R), whose foot-head is preferably a heavy syllable (WSPPrWd), i.e. one with a voiceless onset. Stress shift with regard to secondary stress earlier in the word does not satisfy WSPPrWd better – since this focuses on primary stress – but merely violates All-Ft-R fatally. Consequently, stress shift proves advantageous only in odd-syllable words that contain a sequence of voiceless–voiced sonorants in the penult and ultima, respectively. Rather than final stress, stress shift to the penult takes place (93d). In other odd-syllable words that lack the voiceless–voiced sonorant sequence in the final two syllables, such stress shift is gratuitous. No trigger exists for stress shift and as a result All-Ft-R violations are so many that the optimal candidates present the normal rhythmic pattern (cf. (94), (95), (96)). The analysis, however, makes an additional prediction that requires verification from further empirical data. The prediction is that in even-syllable words with voiced sonorant–voiceless onsets in the final two syllables, primary stress will present rightward stress to the ultima so that it better satisfies WSPPrWd while still observing Align-HdFt-R (97). 2.5.3.4 Contextual stress shift The Arabela stress system might at first seem surprising given its peculiar contextual character. Not only does it present stress shift relativized to primary stress only, but it also bases its pattern on the heaviness of onsets. Is this stress

96

Onsets

shift in Arabela as bizarre as one would think? I would like to suggest that it is not. Asymmetries between the two types of stress are in fact well attested, as McGarrity (2003) shows. For instance, there are languages with: a) sonoritysensitive primary stress, but sonority-insensitive secondary stress (Nganasan, Kara), b) trochaic foot for main stress, but iambic foot for secondary stress (Sentani), c) quantity-sensitive primary stress and quantity-insensitive secondary stress (Huariapano), among many other patterns. In Arabela, what we find is contextual stress shift; it only occurs with regard to main stress, because of the WSP relativized to primary stress. Furthermore, heavy syllables in Arabela are those that begin with a weightful onset. A somewhat analogous case emerges in Goroa (Seidel 1900; Rosenthall and van der Hulst 1999), this time, however, due to coda weight (99). Stress in Goroa falls either on the leftmost long vowel (99a–b) or on the penult (99e–f). A further complexity is added by closed syllables which are generally light (99b, f) except word finally (99c–d), where they are heavy and can receive stress to satisfy higher-ranking metrical constraints. (99)

Goroa stress a. dúːgnunoː b. girambóːda c. adúx d. axemís e. oromíla f. idirdána

‘thumb’ ‘short’ ‘heavy’ ‘hear’ ‘because’ ‘sweet’

In the light of such data, the only peculiar thing with Arabela is that stress shift is caused by onsets, and not by codas. Other than that, the emerging pattern is rather ordinary. Finally, it is largely an analytical issue to suggest that codas (or onsets) are contextually moraic – as Rosenthall and van der Hulst (1999) and Morén (2000) do – or that they are consistently moraic, but subject to different constraints making reference to their moraicity, e.g. the distinction between WSPPrWd and WSP or WSP(N) and WSP used in Arabela and Pirahã, respectively. Independently of the view taken, the intuition is the same, that is, the moraic behaviour of syllables may not be uniform across all the syllables in the word. 2.5.4 Arabela summary This section has dealt with Arabela, a very little-studied Peruvian language, whose trochaic stress system is particularly interesting as it is generally

Onsets and stress

97

rhythmic, unless disrupted by stress shift owing to the heaviness of the syllables in the head foot. Heaviness depends on onset moraicity, so that voiceless onsets are moraic, but voiced (sonorant) ones are not (§2.5.2). The proposed account (§2.5.3) has treated stress shift as the result of the interaction between the language’s prosodic constraints, onset moraicity and WSPPrWd, so that stress shift only occurs with respect to primary, but not secondary, stress. 2.6

Discussion and conclusion

This chapter has focused on onset-sensitive stress claiming that there are essentially two dimensions of onset-sensitive stress that are independent from one another and as such may interact. These dimensions involve sensitivity to: a) the quality of the onset (QO) and b) the presence of the onset (PO). The tableau below (100) summarizes the languages discussed and the patterns that each exhibits. (100)

Presence and quality of onset interaction in stress QO

PO

Languages

✓ ✕ ✓ ✕

✕ ✓ ✓ ✕

Karo (Arabela) Aranda, Banawá, Dutch Pirahã (Arabela) Greek, Russian, etc.

Pattern Identifier ❶ ❷ ❸ ❹

The independence of QO and PO from one another is supported by the fact that in some languages only QO (❶) or PO (❷) effects emerge, but in others they interact, as in Pirahã (❸). In languages with QO effects (❶&❸), it has been shown that the moraic onset hierarchy (101) that assigns onset weight on certain onsets is in action. (101)

*μ/Ons/[+voi] >> Be Moraic >> *μ/Ons

This entails that voiceless onsets will be systematically moraic, but those which bear [+voi], i.e. voiced obstruents, will not. Interestingly, sonorants may pattern along voiced or voiceless obstruents depending on whether they bear [+voi] or are unspecified for [voi]. As this is language specific, in languages like Arabela and Pirahã sonorants bear [+voi], but in Karo they lack the feature altogether. In all three languages, the Weight-to-Stress Principle (WSP) – that requires that heavy syllables are stressed – is responsible for generating the stress assignment or stress shift effects found.

98

Onsets

The ranking in (101) is founded on the idea that pitch perturbation due to voicing may also be used for stress, besides its primary use for tone. During the production of a voiceless consonant, the vocal folds are stiff, which results in pitch raising. While speakers usually utilize this type of F0 raising for tone, this phonetic cue is present and can be exploited for other purposes too. I have suggested that certain languages do manipulate it for stress. I have also explored how this is done phonologically and formalized it accordingly by means of moraic onsets. However, it is anticipated that this effect could also be tested experimentally. For instance, we could conduct a perception experiment in one of the languages that present QO effects (Karo, Pirahã, Arabela) and test whether listeners would tend to perceive some syllables as more stressed than others depending on the onset’s voicing. The hypothesis obviously would be that syllables with voiceless onsets would be perceived as more stressed than syllables with voiced onsets.36 Pirahã in particular is ideal for such testing, as it is a language where tone functions orthogonally to stress, and consequently any attraction of stress by high tone could not be attributed to native-language influence. Future experiments would also need to add sonorants into the equation and see whether the hypothesis about their ambiguous nature is verified. In languages like Karo, they should induce similar results with voiceless obstruents, but in others, such as Pirahã and Arabela, they should pattern with the voiced obstruents. As far as PO effects (patterns ❷&❸) are concerned, however, I have argued that the driving force behind them is not onset weight, but ALIGNσ́O, which 36

A tentative experiment along these lines has been conducted in Topintzi (2006b) for English. Briefly, the sequences pa and ba, placed within carrier phrases, were elicited from a male British English speaker. Manipulation of the F0 of pa produced a further token matching the F0 of ba (i.e. lower by 7.6 Hz). Three tokens were thus available: bā, pā (the so-called flat chain) and the original pá (with higher pitch than the other two). Any differences in amplitude, length and quality of vowel were eliminated by direct manipulation. The tokens were then combined to get syllables of alternating voicing. The resulting sixteen tokens ranged between two and five syllables. The sequences were then randomized and presented to eight subjects, whose task was to identify the syllables they perceived as stressed. The task was repeated four times in total, generating sixty-four stimuli per speaker. The statistical analysis confirmed that pa was perceived as stressed more frequently in the high rather than the flat conditions, a fact that verified the phonological hypothesis pursued throughout; pa was also perceived as stressed more frequently in pa-initial words and in words with an even number of syllables (2 and 4) compared with words with an odd number of syllables (3 and 5). As Matthew Gordon (p.c.) correctly observes, the results of the experiment are compatible with the present phonological approach. However, its conclusions cannot be secured, since it failed to consider certain factors that might prove vital in our understanding of what goes on. For instance, the experiment did not look at the effect of pitch on /ba/ which would allow for weighing the relative effect of pitch versus voicing. It is nonetheless a first step to phonetically evaluating the predictions of the current theory.

Onsets and stress

99

asks that stressed syllables are onsetful. This can be achieved in one of two ways: either the prosodic aspects of the normal stress system of the language will be manipulated so that misalignment of stress to an onsetful syllable will occur (e.g. Alyawarra), or the stress algorithm will remain unaffected, but an onset will be added to a syllable that has already been designated as the stressed one, if it is onsetless (e.g. Dutch). Of course, the majority of languages present neither QO nor PO effects, so these are classified as type ❹ languages. This concludes my investigation of onset-sensitive stress with regard to coerced onset weight. I will, however, return to stress in Chapter 5 when I discuss onset geminates, i.e. distinctive onset weight, and their significance for the stress algorithm of Pattani Malay (§5.2.1) and Marshallese (§5.3.1). Later, in Chapter 7 (§7.2), stress again crops up in the discussion of possible alternatives to the present account, particularly those which treat onset-sensitive stress as a prominence effect (Everett 1988; Hayes 1995; Smith 2005) or as function of prominence (Gordon 2005). Interestingly, most of the previous research devoted to onset effects on prosody has focused on stress (see references above, but also Davis 1988; Goedemans 1996, 1998; Gordon 2005) without much discussion beyond it (but see Hajek and Goedemans 2003). However, it is the investigation of onset-sensitive prosodic phenomena other than stress which is pivotal for the thesis made in this book. According to the vast majority of analysts, phenomena like compensatory lengthening, word minimality, gemination or reduplication can only be explained by means of weight. The implication of this is straightforward: if we were to discover that onsets actively participate in such phenomena, then we undoubtedly have robust argumentation in favour of onset weight. Unsurprisingly – given the aims and scope of this book – this is exactly what I will be arguing. By the end of this exploration, it should have become clear that onsets often play a leading rather than an insignificant role in a range of weight-based phenomena. Consequently, the remaining chapters of the book are devoted to detailed examination of the relevant phenomena and analysis of case-studies. For completeness and for further justification, however, alternative approaches will also be considered when appropriate. The phenomena investigated are, in order of appearance: compensatory lengthening (Chs. 3 and 6), word minimality (Chs. 4 and 6) and geminates (Chs. 5 and 6), as well as brief discussions related to reduplication, metrics and tone (Ch. 6).

3 Onsets and compensatory lengthening

3.1

Aims

Compensatory lengthening (CL) is a phenomenon commonly addressed for weight purposes. It refers to the lengthening of a segment that occurs on the surface as a response to a neighbouring segment’s deletion. Although, according to Morin (1992), both consonants and vowels may trigger and undergo CL (1), the ‘classical’ – in the sense of most common – case of CL involves the deletion of a coda consonant and the lengthening of the preceding vowel (1.I). For instance, in Turkish, coda-h deletes before a continuant or a nasal stop, e.g. fihrist ~ fiːrist ‘index’ (Sezer 1986: 230). (1)

CL triggers and undergoers

I II III IV

Lost segment

Lengthened segment

C C V V

V C V C

In his seminal paper, Hayes (1989) accounts for CL through the use of moras, arguing that segment loss does not necessarily imply mora loss too. Instead, the latter may survive by means of re-association to an adjacent melodic unit. Thus, in Turkish, h’s mora is left behind despite the disappearance of its original host, leading to the lengthening of the vowel before it. Consequently, the number of moras before segment loss and after lengthening remains stable. Central to this approach is the prediction – stated explicitly by Hayes (1989: 285) – that consonant loss from an onset position will never lead to CL, simply because onsets never bear any moras to start with, and as such are inherently incapable of causing lengthening. Crucially, such a prediction depends on the stipulation that onsets are never moraic. In a framework such as the present one, the opposite is anticipated. Indeed, this chapter aims at showing that there is good evidence that onsets not 100

Onsets and compensatory lengthening

101

only may cause CL, as in Samothraki Greek (§3.2.3), but they can also undergo CL, as in Trique (§3.3.4). The analysis I will be constructing works for a wide range of standard and more ‘exotic’ CL cases. Although it is designed to conform to important tenets of Optimality Theory (OT), such as parallelism and Richness of the Base (RotB), it is flexible enough that it can be incorporated in other, serial or rule-based frameworks. This distinguishes it from most other analyses that are somewhat ‘framework-particular’. However, this is also the reason why CL cannot tell us much with regard to the distinctive/coerced distinction. To conform to RotB, no moraic specifications should be imposed. But then this means that CL cannot be about mora preservation (since it should not depend on mora specification). Indeed, the analysis builds on the idea that CL is about position preservation through a mora, which may be underlyingly present but does not have to be. This latter property of the proposed system suggests that CL cannot be used as a criterion for distinctive/coerced weight. But this is only a peripheral issue; instead, my main aim in this chapter is to stress a point that remains invariable across models or frameworks, namely that CL – like other phenomena discussed in this book – establishes that onsets can be moraic on a language-specific basis. This conclusion is inescapable if one considers cases where onsets undergo CL, like in Trique (§3.3.4) or Pattani Malay (§5.2.1), and is compatible with onsets that trigger CL, as in Samothraki Greek (§3.2), Onondaga (§6.2.1.1) and others (see §6.2.1.2 for an overview). Independently of preferences on theoretical models or technical details, this is the insight that the present chapter wishes to contribute. The chapter is structured as follows: Section 3.2 presents the Samothraki Greek facts that form the basis of this chapter’s argument. In general, r deletes from an onset position triggering lengthening of the following vowel, while in a coda position it stays put. Justification for the r’s existence and a presentation of the facts is provided. The idea pursued is that CL is about position preservation through a mora, rather than mora preservation per se. The core analysis implements this idea and a more detailed analysis follows. Possible alternatives and their flaws are also considered for completeness. Section 3.3 provides further evidence for the current analysis of CL through its extensions to both rather standard cases of CL, such as prenasalization or glide formation accompanied by lengthening, and more challenging ones, such as CL caused by unsyllabified segments. Other interesting issues are also explored, namely the absence of determinism in CL-application that Hayes’ (1989) – perhaps most influential to date – account of CL predicts under certain

102

Onsets

circumstances, as well as an explanation as to why CL from onset loss is so exceedingly rare. Section 3.4 offers some concluding remarks. 3.2

Samothraki Greek

Samothraki Greek (henceforth: SamG) is a Northern Greek dialect, spoken on the island of Samothraki in north-eastern Greece. Like other Northern Greek dialects, it is typical in raising stressless e, o to i, u, respectively, e.g. péde > pédi ‘five’, potamós > putamós ‘river’, and losing underlying i and u (with some exceptions), e.g. tiɣáni > tɣaɲ ‘frying pan’, kufós > kfos ‘deaf’. Unlike other dialects, though, it is unique in exhibiting the phenomenon of r-deletion from an onset position, the loss of which is in most cases compensated for by the lengthening of the following vowel. The remainder of this section examines the relevant data, how we can understand and analyse them, as well as considering alternatives that – as we will see – prove inadequate. 3.2.1 The data The data reported below are largely due to Katsanis (1996; henceforth K). Some extra data come from additional sources and from personal communication with Maria Tsolaki and Allia Foteinou, who are native speakers of SamG, and Marianna Ronga at the Aristotle University of Thessaloniki, who also has knowledge of the data. These will be indicated appropriately. SamG presents a case of r-deletion accompanied by subsequent CL. What is fascinating, however, is that CL only occurs when the r deletes in an onset position affecting the following vowel. The forms on the left of ‘>’ show the pronunciation of the words in Standard Modern Greek (SMG) – with the added processes of raising and deletion discussed before that are applicable in Northern dialects – while those on its right indicate the corresponding SamG forms. In most instances, onset r deletes inducing lengthening of the following vowel. This is indeed the case for initial singleton onsets (2) and for r as the second (or third) component of initial and medial complex onsets as in (3). (2)

CL environment I: Deletion of r initially and lengthening (K: 50–1)

[here and throughout all glosses are mine] ra > aː ri > iː ru > uː re > eː ro > oː

rafts > áːfts riɣaɲ > íːɣaɲ rúxa > úːxa réma > éːma róɣa > óːɣa

‘tailor (masc.)’ ‘oregano’ ‘clothes’ ‘stream’ ‘nipple, berry (of a grape)’

Onsets and compensatory lengthening (3)

103

CL environment II: r in onset CrVC: deletion and lengthening a. Biconsonantal clusters (K: 54–6) prótus > póːtus pr+o > poː vr+i > viː vris’ > víːs’1 fr+e > feː fréna > féːna xr+o > xoː xróma > xóːma kr+a > kaː krató > kaːtó ɣr+a > ɣaː ɣráfo > ɣáːfu θr+o > θoː θrónos > θóːnus ∂ ´ r+a > ∂ ´ aː ∂ ´ rákos > ∂ ´ áːkus br+e > beː ɣabré > ɣabéː dr+u > duː ∂ ´ édru > ∂ ´ éduː tr+u > tuː metrún > mitúːn

‘first’ ‘tap’ ‘brakes’ ‘colour’ ‘I hold’ ‘I write’ ‘throne’ ‘dragon’ ‘bridegroom’ ‘tree’ ‘they count’

b. Triconsonantal clusters (K: 58–9) spr+a > spaː áspra > áspaː xtr+a > xtaː éxtra > éxtaː zdr+u > zduː sidrufçá > zduːfçá ftr+a > ftaː ráftra > áːftaː

‘white’ ‘hostility’ ‘company (of people)’ ‘tailor (fem.)’

The lengthening is clearly audible and visible, as the spectrograms from the speech of Allia Foteinou below show (despite the noise that was unavoidable at the time of recording). The SMG words she was asked to pronounce in SamG were [friði] ‘eye-brow’ and [fiði] ‘snake’, placed within the carrier phrase ‘Η λέξη __ είναι εύκολη’ The word __ is easy. The result was [fiːð] vs. [fið], creating a nice minimal pair, where the only difference is in the duration of the vowels, as indicated by the highlighted portions. The respective durations were: 0.270 ms and 0.117 ms.

Figure 3.1: Spectrogram of [fiːð] 1

The apostrophe expresses palatalization due to an underlying /i/ that has deleted on the surface.

104

Onsets

Figure 3.2: Spectrogram of [fið]

In certain environments, however, r either fails to delete altogether or disappears, but causes no lengthening. Surprisingly enough, lack of r-deletion is observed when the r is syllabified as a coda, the locus classicus for CL. This fact is stated in all the SamG descriptions that I have checked (Papadopoulos 1927; Krekoukias 1964; Afroudakis 1985; Katsanis 1996) with the exception of Andriotis (1926–1928) who suggests that r-deletion and lengthening occurs in a coda position too, e.g. várka > váːka ‘boat’. Given that the reports of most researchers also concur with the data that I have collected myself, it is safe to say that coda r indeed survives. (4)

No deletion and no lengthening: Coda r stays

fanár figár

‘lantern’ (K: 48) ‘moon’ (K: 58)

arpázu karpós

‘I grab’ (K: 48) ‘seed’ (K: 48)

On the other hand, an r that would originally syllabify as a medial onset deletes, but without any lengthening. (5)

Deletion and no CL (I): Deletion of r word medially and no lengthening (K: 52)2 aro > ao iru > iu are > ae iri > ii eru > eu uri > ui era > ia ara > aa ure > ue

2

θaró > θaó léftirus > léftius varéʎ > vaéʎ θirí∂ ´ a > θíi∂ ´a kséru > kséu lurí > luí méra > mía skára > skáa kurévu > kuévu

‘I reckon’ ‘free’ ‘barrel’ ‘pigeon-hole’ ‘I know’ ‘strap, strip’ ‘day’ ‘grill’ ‘I cut someone’s hair’

Raising in [mía] is unexpected, since the mid-vowel is stressed. For this idiosyncratic process, the reader should consult Katsanis (1996: 37–8). It is also unclear why [θíi∂ ´ a] receives stress on the first vowel instead of [θií∂ ´ a].

Onsets and compensatory lengthening

105

The final environment where r fails to delete is in complex clusters of the type CrVV. Unlike CrVC clusters (3), CrVV clusters maintain the r and consequently no lengthening appears. What we get is C+V+rjV, presumably with metathesis and glide formation (6). In all the relevant cases, the first vowel is front i or e, while words that in SMG include a VV sequence where the first vowel is back a, o, u, like fráula, krúo, práos, akróasi, are not used in the dialect (Ronga p.c.). We thus cannot currently test what happens in the sequences C+r+a/o/u+V, which is why the ensuing analysis only deals with C+r+i/e+V sequences. (6)

Deletion and no CL (II): C+r+i/e+V → C+i+rjV (Ronga p.c.) priakóni > pirjakóɲ alétria > alétirja tría > tirjá

‘jagged file used to sharpen knives’ ‘plough (plural)’ ‘three’

Notably, when the C in the sequences above is velar, then in addition to what happens in (6), the front vowels i/e get centralized to î (or ê occasionally) (7).3 Centralization also occurs in C[velar]+r+i/e+C sequences (8).4 A treatment of centralization would take us too far afield here; since it is not practically vital for the ensuing discussion, I refer the interested reader to Topintzi and van Oostendorp (2009b) for a detailed account. (7)

C[Velar]+r+i/e+V → C[Velar]+î+rjV (K: 71) áɣrios > áɣîrjus kréas > kîrjás

(8)

‘wild’ ‘meat’

́ axríastos > axî rjastus ɣriá > ɣîrjá

C[Velar]+r+i/e+C → C[Velar]+îː/êː+C ́ gr+i > gîː grízos > gî ːzus ‘grey’ ́ kr+i > kîː krínu > kî ːnu ‘I judge’ kr+e > kêː kremnós > kêːmnús ‘precipice’

‘unneeded’ ‘old woman’ (K: 56) (K: 56) (K: 72)

3.2.1.1 Summary of the patterns For ease of reference, the preceding patterns are summarized in the following tableau. (9)

SamG r patterns

Coda r Singleton onset r 3

4

word initially

Deletion

Lengthening

Ex.

NO

NO

(4)

YES

YES

(2)

Other comments

I maintain Katsanis’s representation of [î] and [ê]. According to Ronga (p.c.), though, [î] is likely IPA [ɨ] and [ê] is [ә]. She also maintains that [î] is the unstressed realization of [ê]. According to Maria Tsolaki (p.c.), some of the words above are instead pronounced as: axrjastus, kirjás’ (where the final s is palatalized) and gːmnus.

106

Onsets

Complex onset r

Deletion

Lengthening

Ex.

Other comments

word medially

YES

NO

(5)

Cr+V+C

YES

YES

(3)

(also when C=velar & V=front→ centralization, see (8))

Cr+V[i,e]+V

NO

NO

(6)

glide appearance +metathesis (+centralization as above, see (7))

3.2.1.2 Justifying r Before proceeding to the analysis, an important question has to be addressed.5 What is the evidence for the underlying r in the first place in SamG? Perhaps the strongest evidence comes from alternations between r in coda and onset positions. Such cases may emerge at morpheme boundaries and with morphologically related words. If r exists underlyingly, the expectation is that it will show up in coda position, but be absent if it were to syllabify as an onset. An example of that sort is found in the words σύρ-ω ‘I drag’ and σύρ-θηκ-α ‘I was dragged’. Hyphens indicate morphological boundaries. In SMG these are pronounced and syllabified as [sí.ro] and [sír.θi.ka], exhibiting an alternation between r in onset and coda positions, respectively. In SamG, however, as Allia Foteinou confirms (p.c.), these words are pronounced [sío] and [sírka] (with simplification of the -θik- suffix), as anticipated, given the facts of SamG above and under the condition that r indeed exists in the input. Kiparsky (2008a) provides additional evidence that r exists synchronically in SamG by considering the interaction between vowel changes in unstressed positions and r-loss. Conclusively, then, there are various sources in support of the presence of r still in the dialect. It needs to be mentioned, however, that most SamG speakers are bilingual between their dialect and the Standard (Athenian) one, especially due to exposure to it through the media. According to Katsanis (1996: 27), most speakers of SamG belong to older generations, while younger ones have little or no knowledge of the dialect. He also briefly refers to social and educational circumstances that affect the speakers’ knowledge of the dialect. It is thus unclear to what extent SMG influences idiolects in maintaining the r and, conversely, to what extent SamG r-loss surfaces in the Standard. My own – very brief – experience with two speakers of the dialect suggests 5

Thanks to Iggy Roca for pointing this out and for discussion on the matter.

Onsets and compensatory lengthening

107

that r-loss genuinely occurs, but not in an absolute manner, plus there is inter-speaker variation as to the words that present r-deletion and lengthening. For instance, the word [frási] ‘phrase’ above was pronounced by Maria Tsolaki as [faːs], and [fras] by Allia Foteinou – similar to the SMG [frási], but with the characteristic Northern Greek unstressed /i/ deletion – although in other instances, Allia would produce lengthened forms, e.g. [vaːka] for SMG [vraka] ‘breeches’. However, whenever lengthening applies there is consistency as to where it does, that is, it always involves forms that in Standard Greek include an onset r. The ability of speakers, like Allia above, to replace the r in words like [fras] with lengthening in other words like [vaːka] suggests that at some level of abstraction r is indeed present. While I believe that this level is indeed the underlying one, the die-hard supporter of the view that this process is no longer synchronic (despite Kiparsky’s 2008a counter-evidence) should bear in mind that even if that is the case, r-loss did occur historically, which serves equally well to claim that SamG – synchronically or diachronically – illustrates an instance where onset loss causes CL. Thus, while in the ensuing analysis I will be assuming the presence of r in the input, readers who feel that this is not well justified should think of this process as historical rather than synchronic. Crucially, this does not affect the main argument, namely that onsets may trigger CL. 3.2.2 The basic analysis Moving on to the analysis now, the account hinges on the claim that segmental positions in a string require a correspondent on the surface. There are basically two ways to achieve this: either by preserving the segment (with total or partial featural identity towards the input) that occupied that position underlyingly, or by losing the segment, but having a prosodic correspondent for it in the guise of a mora. That latter instance is what we call CL. In other words, CL is seen as an alternative strategy for a segment to show up in the output with lengthening acting as a ‘cue’ for the lost segment. This can be coded in the constraint termed Position Correspondence (PosCorr). (10)

PosCorr: An input segment must have an output correspondent either segmentally by means of a root node or prosodically by means of a mora

I will basically argue that CL – and consequently SamG CL too – involves a high-ranking markedness constraint that militates against the presence of an offending segment resulting in its deletion. The latter is compensated for by the lengthening of a neighbouring segment, the second mora of which gets identified with the position that hosted the original – now lost – segment.

108

Onsets

To see how this works, consider the following hypothetical example of synchronic CL. The root /kan/ may either surface on its own or be followed by vowel-initial /-a/ or consonant-initial /-ta/ suffixes. However, codas are not permitted, thus the /n/ in (11a, b) deletes and lengthening of the preceding vowel emerges, i.e. CL occurs. On the other hand, (11c) presents no lengthening, and the /n/ survives as it is no longer in syllable-final position. (11)

→ → →

a. /kan/ b. /kan-ta/ c. /kan-a/

[kaː] [kaːta] [kana]

These facts are accounted for in the next tableaux. First, note the use of indices as a means to identify corresponding positions. To avoid too many indices, moras of input segments will have no indices. Only the inserted mora (in bold) due to CL, e.g. the one numbered μ3 in (12e), will have an index to mark the mapping to the deleted segment, here /n3/. The index of /k/ is omitted as it is irrelevant for current purposes. (12)

/kan/ → [kaː] /k1a2μn3/

NoCoda

a.

ka2μn3

*!

b.

ka2μn3μ

*!

c.

ka2μ

d. ☞ e.

ka

μ 2,3

ka

μμ3 2

PosCorr Uniformity

Dep-μ

* *! *! *

Tableau (12) caters for the situation arising in (11a, b). The high-ranking markedness constraint NoCoda eliminates both (12a) and (12b) which include a coda, without paying any attention to the moraic specification of each. (12c) avoids the NoCoda violation, but at the cost of a – segmental or prosodic – correspondent for position 3, thus PosCorr is violated. A tempting candidate now is (12d) which merges the deleted segment with the previous vowel, offering a segmental correspondent for position 3 in the vowel. To exclude this, all we need to do is rank Uniformity (13) highly enough. (13)

Uniformity: No element of the output has more than one correspondent in the input, i.e. no coalescence (McCarthy and Prince 1995)

The winner (12e) resolves all problems above by deleting the coda and lengthening the previous vowel whose mora with the index 3 now provides a prosodic correspondent for the position originally occupied by /n/. The only violation it incurs is that of low-ranking DEP-μ. This ranking works smoothly for /kan-a/ → [kana] too.

Onsets and compensatory lengthening (14)

109

/kan-a/ → [kana] /k1a2μn3 -a4μ/

NoCoda

PosCorr Uniformity

Dep-μ

☞ a. ka 2μn3a4μ b. ka2μμ3n3a4μ

*!

μ μ 2 4

c. ka a

*!

Here, no trigger for /n/-deletion (14c) exists, so a segmental correspondent (14a) is preferred over a prosodic one (14b). While both satisfy PosCorr, (14b) is worse since it does so at the expense of DEP-μ. Thus, we can see that the right results obtain. More generally, the pattern for CL involves the following: (15)

Pattern for CL – preliminary version:

- a trigger for deletion of segments in particular syllabic positions must exist (markedness constraint M) - PosCorr >> DEP-μ ensures that CL occurs, i.e. insertion of a mora identified with the position of the deleted input segment - Uniformity >> DEP-μ: prosodic rather than segmental identification is preferred We will see that this system has to be slightly modified (§3.2.2.2), but before doing so, let us explore some of its typological consequences. 3.2.2.1 Lack of CL An appealing aspect of the proposed system is that it can also derive the absence of CL as in hypothetical /kan/ → [ka] *[kaː]. One way to do this is by ranking PosCorr low in the ranking (16), so that markedness is best satisfied by deleting the position wholly (17c). (16)

Deletion of a segment not followed by CL – preliminary version: - requires minimal re-ranking: DEP-μ >> PosCorr implying that no trace (segmental or prosodic) of the deleted segment is left behind

(17)

Lack of CL: /kan/ → [ka] /k1a2μn3/ a. b. ☞ c. d. e.

ka2μn3 μ μ 2 3

ka n ka

μ 2

ka

μμ3 2

ka

μ 2,3

No Coda

Dep-μ

Uniformity PosCorr

*! *!

* * *! *!

110

Onsets

Another way to achieve the lack of CL is by keeping PosCorr highly ranked and at the same time by ranking DEP-μ >> Uniformity. In this way, the winning output would satisfy PosCorr segmentally – rather than prosodically – by means of fusion (18e). (18)

Lack of CL: /kan/ → [ka] /k1a2μn3/

NoCoda

μ 2 3

*!

μ μ 2 3

*!

a. ka n b. ka n c. ka

μ 2

d. ka

μμ3 2

☞ e. ka

μ 2,3

PosCorr

DEP-μ

Uniformity

* *! *! *

Deciding between the two possibilities is not always straightforward. However, given that (18e) involves fusion, it should – at least sometimes – be possible to recognize whether the segment is the product of coalescence. For example, in the case of /kan/ becoming [ka], fusion would presumably have occurred if the resulting vowel presented nasalization, i.e. as in [kã]. The French data below support an analysis along these lines, therefore the winning candidate in French would be the product of the ranking in (18). (19)

French: [V̰] ~ [Vn] alternations in masculine-feminine (Tranel 1987: 70–1) bon fin certain

[bõ] [fε̃] [sεrtε̃]

vs. vs. vs.

bonne fine certaine

[bɔn] [fin] [sεrtεn]

‘good’ ‘thin’ ‘sure’

A similar example is found in Portuguese (Mateus and d’Andrade 2002), where it is argued that nasalized vowels are the product of an underlying sequence of a vowel and a nasal. For instance, the prefix /in-/ appears as [in-] before a vowel, but as a nasalized vowel [ĩ] before a consonant. In the latter case, nasal consonant deletion must have occurred, with the nasal feature being left behind and linked to the preceding vowel. (20)

Portuguese nasalized vowels (Mateus and d’Andrade 2002: 22) a. V-initial stems acabado ‘finished’ oportuno ‘opportune’

inacabado inoportuno

b. C-initial items capaz ‘able’ posto ‘put in place’

incapaz imposto

[inɐkɐbádu] [inopuɾtúnu] [ĩkɐpáʃ ] [ĩpóʃ tu]

‘unfinished’ ‘inopportune’ ‘unable’ ‘tax’

Onsets and compensatory lengthening

111

Whenever such direct evidence is lacking, the assumption will be that the language prefers to delete the position altogether leaving no relic of it behind, i.e. as in (17c). 3.2.2.2 Replacing DEP-μ with P-DEP-μ The careful reader will have noticed a seemingly problematic aspect of the analysis as it stands: in order that the correct outputs are generated in tableaux (17) and (18), it is vital that input /n/ lacks a mora, so that lengthening on the surface violates the fairly high-ranked DEP-μ which thus eliminates the candidate that has undergone CL. But what if the input contains a mora for /n/? The problem is non-trivial, as we are in fact obliged to consider such a possibility given Richness of the Base, which requires that inputs are unconstrained. Utilizing the same ranking and candidates, but for the input /k1a2μn3μ/ instead, predicts the candidate [ka2μμ3], i.e. with CL, as the winner. Obviously, this is not a desirable result, as it suggests that we can in principle get contradictory outputs with respect to the same phenomenon in the same language depending on the input assumed. What we need to solve this problem is to somehow ban lengthening in /kaμn(μ)/ → [kaμμ] independently of the moraic status of /n/. We can achieve this by modifying DEP-μ in such a way so that it militates against lengthening, but of course not against underlying (vowel) length. Positional μ-licensing, as proposed by Bermúdez-Otero, offers such a version of DEP-μ (see BermúdezOtero 2001; Campos-Astorkiza 2004). (21)

Positional μ-licensing: A segment α is positionally μ-licensed by a mora iff µ is the only prosodic unit directly dominating α

A positional μ-licenser is a mora that is the only prosodic unit dominating a given segment a. In contrast, a non-positional μ-licenser is a mora that is not the sole prosodic unit immediately dominating α. Under positional μ-licensing, the requirements of DEP-μ – now P(OSITIONAL)-DEP-μ – change as in (22). (22)

P-DEP-μ: A non-positional µ-licenser mora in the output has a correspondent in the input

P-DEP-μ only penalizes inserted moras that lead to lengthening, exactly as in candidates (17d–18d). An evaluation of this constraint and a comparison with classic DEP-μ follows. Only relevant structures are considered.

112 (23)

Onsets P-DEP-μ vs. DEP-μ

a.

b.

c.

d.

Input

Output

µ1 µ2

µ 1 µ2

V

V

µ1

µ 1 µ2

V

V

µ1

µ1 µ2

V C

V

µ1 µ2

µ1 µ2

V C

V

Positional µ-licenser

P-DEP-µ DEP-µ

No







(a) long vowel

No

*

*

(b) lengthened vowel

No

*

*

(c) non-moraic C-deletion and lengthening

No

*




(d) moraic Cdeletion and lengthening

(23a) presents an underlyingly long vowel. None of the output moras is a positional μ-licenser (since no single μ directly dominates the segment V), but P-DEP-μ is satisfied, since both moras have a correspondent in the input. The same reasoning applies to the evaluation of DEP-μ. The two constraints also agree in their evaluation with respect to the lengthening in (23b). This violates DEP-μ (because of the inserted μ2 in the output) and also P-DEP-μ (because the non-positionally μ-licensed μ2 has no correspondent in the input). The evaluation of (23c) is identical to that of (23b), the only difference being that the input also contains a non-moraic coda that gets deleted. Up to this point, no difference in the evaluation of both versions of DEP-μ emerges. This changes when we consider a /VμCμ/ sequence with a moraic C (23d). DEP-μ incurs no violation in [Vμμ], since no mora has been inserted. But for P-DEP-μ this is beside the point. Vowel lengthening again occurs and given that the non-positionally μ-licensed μ2 linked to V has no correspondent linked to the same segment in the input, a P-DEP-μ violation obtains. This difference will give us the right results for both non-moraic and moraic inputs, as well as for both the presence and absence of CL. These are summarized in the tableaux in (24) and (25). (24)

Deletion of segment followed by CL: PosCorr >> P-DEP-μ (i) non-moraic input /C1V2μC3/ a. ☞ b.

V2μ2 μ2μ3 2

V

PosCorr

P-DEP-μ

*! *

Onsets and compensatory lengthening

113

(ii) moraic input /C1V2μC3μ/ a. ☞ b. (25)

V2μ2

P-DEP-μ

PosCorr *!

V2μ2μ3

*

Deletion of segment not followed by CL: P-DEP-μ >> PosCorr (i) non-moraic input /C1V2μC3/ ☞ a. b.

P-DEP-μ

PosCorr

μ2 2

V

μ2μ3 2

V

* *!

(ii) moraic input /C1V2μC3μ/ ☞ a. b.

P-DEP-μ

PosCorr

V2μ2 μ2μ3 2

V

* *!

3.2.2.3 Interim summary The analysis of CL and lack thereof is thus accordingly modified as shown below. (26)

Deletion of a segment followed by CL – final version: M, Uniformity, PosCorr >> P-DEP-μ

(27)

Deletion of a segment not followed by CL – final version: i) no trace of deleted position: M, Uniformity, P-DEP-μ >> PosCorr (cf. (17)) ii) segmental correspondent: M, PosCorr, P-DEP-μ >> Uniformity (cf. (18))

We now have all the machinery in place to tackle the SamG CL data. 3.2.3 SamG analysis The essentials of the SamG account involve the CL ranking in (26), that is: M, Uniformity, PosCorr >> P-DEP-μ. The markedness constraint that triggers onset deletion in SamG is *ONSET/r, as imposed by the proposed placelessness of SamG r (§3.2.3.1). The absence of CL word medially relates to the crosslinguistic tendency of avoiding super-long vocalic transitions (Kavitskaya 2002), which would have been generated had CL applied in that position (§3.2.3.2.1). Word initially (§3.2.3.1.1) or in complex onsets followed by a

114

Onsets

V+C sequence (§3.2.3.1.2) no risk of that sort appears, which is why CL can apply freely. However, CL does not apply in two more cases: a) in complex clusters followed by V+V sequences, as it would create undesirable structures (§3.2.3.2.2), and b) in coda position, where *ONSET/r becomes irrelevant, so no trigger for r-deletion exists and hence no reason to lengthen the preceding vowel (§3.2.3.2.3). The following sub-sections present the detailed analysis. 3.2.3.1 r-deletion and CL The first part of the analysis deals with the two cases where onset r deletes and CL occurs, that is, in word-initial singleton onsets ((2) and §3.2.3.1.1) and complex onsets followed by a VC sequence ((3) and §3.2.3.1.2). The proposal hinges on an important assumption, namely that SamG r is placeless and that placeless segments are banned from onset positions. This is captured by the following constraint. (28)

*ONSET/r: r is disallowed in onset position

The constraint in (28) serves as the markedness constraint that triggers r-deletion and creates the circumstances for the application of CL. As an added bonus, it also provides the basis for a specific case of vowel spreading (§3.2.3.2.2). Support for the constraint in (28) comes from two observations: first, r-placelessness has been argued for in other languages too, such as Yoruba (Akinlabi 1993) or English (Rice 1992); second, placelessness is tolerated or even required in codas, e.g. Selayarese ʔ (Rice 1992) or Japanese ŋ (Yip 1991), but not in onsets, because placeful onsets are preferred. Replacing then M with *ONSET/r in the general schema of CL in (26) we get the core ranking for SamG CL as in (29). Some extra constraints will be introduced in the following sections to account for the whole phenomenon, and, in particular, to explain instances where CL fails to emerge. But first, let us consider cases where it does emerge. (29)

Core ranking for SamG CL *Onset/r, Uniformity, PosCorr >> P-Dep-μ

3.2.3.1.1 Word-initial singleton r As we know, word initially, r is lost and subsequent CL occurs. The ranking *ONSET/r, Uniformity, PosCorr >> P-DEP-μ eliminates each of the first three candidates, which present a placeless onset (30a), coalescence (30b) or lack of a correspondent for /r1/ (30c). Therefore, (30d) wins, because the added μ1 that produces lengthening only causes a violation of low-ranked P-DEP-μ.

Onsets and compensatory lengthening (30)

115

CL of singleton r word initially: /rasu/ → [áːsu] ‘priest’s cloth’ *ONSET/r, Uniformity, PosCorr >> P-DEP-μ /r1a2μs3u4μ/ *ONSET/r Uniformity PosCorr P-DEP-μ a. r1a2μ..

*!

μ

b. a1,2 ..

*!

c. a2μ..

*!

☞ d. a2μ1μ..

*

3.2.3.1.2 r in complex onsets followed by VC sequences The ranking employed above also generates the right results for CrVC > CVːC sequences. (31)

CL of r in Cr+V+C: /mitrun/ → [mitúːn] ‘they count’

*ONSET/r, Uniformity, PosCorr >> P-DEP-μ μ μ /m1i2 t 3r4u5 n6/ μ 3 4 5 6

a. …t r u n

b. …t3u4,5μn6

*ONSET/r

Uniformity PosCorr P-DEP-μ

*! *!

c. …t3u5μn6 ☞ d. …t3u5μ4μn6

*! *

3.2.3.2 Lack of CL Emergence of CL is straightforward given the ranking proposed. However, there are other instances where CL fails to occur. This is either because r deletes but other factors block lengthening from happening (§3.2.3.2.1), or because there is no trigger for the r to delete in the first place (§3.2.3.2.2–3). 3.2.3.2.1 Word-medial singleton r Unlike deletion of word-initial r (§3.2.3.1.1), word-medial r-deletion induces no lengthening. This makes sense if we consider what would happen had lengthening occurred to an input like /l1u2μr3i4μ/. The anticipated outcome would be a VːV (VVV) sequence that would take the shape of either luːi [l1u2μ2μ3i4μ] or luiː [l1u2μi4μ3μ4]. But such super-long sequences tend to be avoided in languages, as Kavitskaya states (2002: 43). She attributes the ban on VːV to perceptual reasons; in particular, since the vowel-to-vowel transitions are always long, it is unlikely that a two-vowel sequence will be interpreted as a three-vowel sequence. Encoding this observation by means of a highly ranked markedness constraint (32), we get the desired effect in SamG. (32)

*SUPER-LONG Vocalic Hiatus (*S-L VH) / *VːV

Note that the term ‘hiatus’ is used rather than Kavitskaya’s original ‘transitions’ to allow it also to refer to cases where the vowels flanking r are identical. As

116

Onsets

Moira Yip (p.c.) points out, the term ‘transitions’ would correctly predict lack of lengthening when the vowels surrounding r were different, e.g. luri > luí, *luːí, but wrongly entail lengthening when the vowels are the same, as in θirí∂́a > *θíːi∂́a, instead of the actual *θíi∂́a, that lacks lengthening (cf. (5)). For this reason, the more neutral term ‘hiatus’ is chosen which also bans *θíːi∂́a. Adding *S-L VH to the so-far established ranking produces the tableau in (33). (33)

no CL in singleton r word medially: /luri/ → [luí] *S-L VH, *ONSET/r >> PosCorr >> P-DEP-μ /l1u2μr3i4μ/

*S-L VH

μμ3 μ 2 4

*!

*

*!

*

a. ..u

i ..

μ μ3μ 2 4

b. ..u i

μ μ 2 3 4

..

c. ..u r i .. ☞ d. ..u2μi4μ..

*ONSET/r

PosCorr

P-DEP-μ

*! *

*S-L VH rules out the outputs with super-long vocalic sequences (33a–b) and *ONSET/r does the same for (33c), which fails to delete the onset r. (33d) is now the winner despite its PosCorr violation implying that there is no segmental or prosodic correspondent for /r3/. Importantly, we now need to slightly modify our ranking and propose that *ONSET/r >> PosCorr holds, as the ranking argument between (33c) and (33d) suggests. (34)

SamG CL (to be revised): Uniformity, *S-L VH, *ONSET/r >> PosCorr >> P-DEP-μ

3.2.3.2.2 r in complex onsets followed by VV sequences As mentioned before, when a complex onset cluster that includes r is followed by i/e + V, no r-deletion occurs, but also no lengthening takes place. The question posed, then, is why input /alétria/ leads to [alétirja] ‘plough (plural)’ instead of the anticipated *[alétiːja]. We will see how this is excluded in a moment. Katsanis’ (1996) speculative account involves a number of processes like epenthesis, deletion and coalescence that apply at various stages of the derivation, but he admits that the proposed solution is unsatisfactory (K: 57). Topintzi (2006a) provides additional reasons why this is so and ends up with a solution that is also adopted here. In particular, it is claimed that r does not delete because it has moved to a coda position due to metathesis, i.e. r + i/e → i/e + r. Recall that r is placeless (§3.2.3.1), but placeless consonants are penalized in onsets, not in codas. Thus, r may survive in that position. Moreover, exactly because r is placeless, it does not block the rightward spreading of its adjacent i/e on its left that provides an onset for the next syllable. This situation is pictured below.

Onsets and compensatory lengthening (35)

117

Metathesis and rightward-spreading over placeless /r/ for input: /a1l2e3t4r5i6a7/ σ

σ

σ

σ

µ

µ

µ

µ

a1

l2 e3 t4 i6

r5

a7

Root Root Place

Of course, there are several other candidates that need to be considered for input /a1l2e3t4r5i6a7/. Some of these can easily be excluded given the constraints utilized so far and thus will not be included in the ensuing tableaux. For instance, *[aletiːa] or *[aletiaː], which correspond to the following relevant moraic structures […t4i6μ5μ6a7] and […t4i6μ6a7μ5μ7], respectively, both violate high-ranking *S-L VH (32) as well as low-ranking P-DEP-μ. The fully faithful *[a1l2e3t4r5i6a7] also violates top-ranking *ONSET/r, while *[a1l2e3t4i5,6a7] provides a segmental correspondent for /r5/, but coalescence is not an approved method to apply in SamG due to high-ranking Uniformity. We are thus left with two reasonable candidates; obviously, one is the winner […t4i6r5j6a7] and the other is *[aletiːja], which can be represented as […t4i6μ5μ6 j6a7], as shown in (36). (36)

Multiple linking in *[aletiːja] σ

σ

σ

σ

µ

µ

µ 5 µ6

µ

a1

l2

e 3 t4

i6

a7

Root Place

Observe that in this structure, the place features of /i/ are distributed to three positions (two to moras μ5 and μ6, one as the onset of the final syllable). Following Topintzi and van Oostendorp (2009b), such a structure is banned because of a binarity constraint on feature association. (37)

BinSpan / BinAss(F): A feature span is (maximally) binary

Justification for this constraint comes from various phenomena, such as vowel harmony spreading in Cappadocian Greek dialects, Ekegusii tone spreading and SamG centralization (see Topintzi and van Oostendorp 2009a and 2009b for

118

Onsets

details) and Limburg Dutch diminutive umlaut (Hermans and van Oostendorp 2008; Topintzi and van Oostendorp 2009a). Other approaches along these lines, but less explicit, include McCarthy (2004) and Key (2007). Note that such a constraint is reminiscent of other constraints banning ternary structures, such as prohibitions against trimoraic syllables (*3μ) or against ternary feet (FtBin). One way to avoid the triple featural association would be to argue that the glide is a product of insertion rather than spreading. This would, however, violate Dep-Seg, which presumably is also ranked high in the language. All that is left now is to examine the constraints that get violated in the winning candidate (35) and incorporate them in the already established ranking. Metathesis of /tri/ in /aletria/ to [tir] and multiple correspondence of /i6/ to [i6] and the glide [j6] violate Linearity and Integrity, respectively, which consequently have to be low ranked. Linearity should not be too low ranked, however, because we do not want metathesis to be the general strategy of resolving complex onset clusters with r when a VC sequence follows. To see this, consider /krato/, which becomes [kaːto]. To get this result instead of the metathesized [karto] the following ranking is in order. (38)

Linearity >> P-DEP-μ: /krato/ → [kaːtó] *[kartó] /k1r2a3μt4o5μ/ ..a3μ r2..

a. ☞ b.

..a

P-DEP-μ

Linearity *!

μ2μ3 3

..

*

Metathesis, then, is the chosen solution to satisfy PosCorr only in CrVV sequences, but not more generally, in which case violation of bottom-ranked P-DEP-μ is preferred. The state of affairs with respect to CrVV input strings is summarized in the next tableau. (39)

no CL in Cr+V+V: /a1l2e3t4r5i6a7/ → [a1l2e3t4i6μr5j6a7μ] *ONSET/r, Uniformity, *S-L VH, BinSpan >> PosCorr >> Integrity, Linearity >> P-DEP-μ /a1l2e3t4r5i6a7/ *ONS uni- *S-L Bin Pos Integr Linear P-DEP-μ /r form VH Span Corr

a. …t4r5i6μa7μ μ μ 4 5,6 7

b. …t i

a

c. …t4i6

μ5μ6

d. …t4i6

μ5μ6 μ

μ7 7

a

j6a7

μ 7

e. …t4i6 a

☞ f. …t4i6 r j a μ

μ 5 6 7

*! *! *!

* *! *! *

*

Onsets and compensatory lengthening

119

This ranking is consistent and compatible with the previous results. Given (33) and the undominated nature of BinSpan, the part of the ranking *Onset/r, Uniformity, *S-L VH, BinSpan >> PosCorr is justified. Comparison between (39e) and (39f) offers the ranking argument PosCorr >> Integrity, Linearity, while (38) confirms that Linearity >> P-DEP-μ. The final ranking for SamG CL, capturing the cases where CL indeed applies and those where it does not, is given by the mini-grammar: (40)

Final ranking for SamG CL

*ONSET/r, Uniformity, *S-L VH, BinSpan >> PosCorr >> Integrity, Linearity >> P-DEP-μ 3.2.3.2.3 r in codas The final – and easiest – case we need to consider is one where the r would syllabify as a coda. Actually, as we have seen (4), coda r survives intact because it escapes the violation of *ONSET/r and hence causes no lengthening. The ranking in (40) also ensures this outcome. The entirely faithful candidate (41d) wins while providing correspondents for all the positions in the string, unlike (41b) which loses one and unlike (41a) and (41c) which incur other constraint violations in their effort to provide different segmental or prosodic position correspondents, respectively. (41)

no CL when r in coda: /karpos/ → [karpós] UNIFORMITY, *ONSET/r >> PosCorr >> P-DEP-μ /k1a2μr3p4o5μs6/ Uniformity *ONSET/r PosCorr a. k1a2, 3μp4.. μ 1 2 4

b. k a p .. μμ3 1 2

c. k a

..

P-DEP-μ

*! *! *!

☞ d. k a r .. μ 1 2 3

3.2.3.3 Summary of SamG analysis and alternatives As explained throughout §3.2.3, SamG CL exemplifies the general pattern that CL presents in the current theory, namely: Markedness, Uniformity >> PosCorr >> P-DEP-μ; the triggering force for onset r-deletion rests with the assumption that r in SamG is placeless and the idea that placeless onsets are not tolerated as represented by the Markedness constraint *ONSET/r. Deletion of r satisfies the latter constraint, while lengthening of the vowel satisfies PosCorr at the minimal cost of a P-DEP-μ violation. A number of additional constraints are introduced to capture the full range of facts and in

120

Onsets

particular the instances where CL fails to emerge. For completeness, the following sub-sections summarize how the major alternative proposals (would) account for the SamG data (for more detailed discussion, see Topintzi 2006a). As will become obvious, they prove either inadequate in certain respects or even totally unsuccessful. 3.2.3.3.1 CL as mora preservation The most famous and insightful pre-OT analysis of CL is due to Hayes (1989), who argues that CL is a process whereby segments delete, but the moras they host remain. As a result, the total number of moras remains constant. Importantly, in Hayes’ model, onsets never bear weight, and as a result are incapable of producing CL. Hayes is, however, aware of (some of) the SamG data and employs various mechanisms to explain them without making reference to onset weight. For instance, he sees the lengthening to [CVː] in /CrV/ sequences as a product of a number of processes, namely: a) V-epenthesis splitting the cluster to [CVrV], b) intervocalic r-loss and c) merging of [VV] to [Vː]. Hayes justifies this analysis based on unclear, or at best ambiguous, data found in sources of SamG, e.g. Heisenberg (1921), Newton (1972) and Katsanis (1996). The most important problem for him, however, is that he fails to consider a crucial aspect of the SamG facts that proves condemning for his proposal. In particular, his analysis misses the facts of word-initial singleton r-deletion that also produces lengthening. To maintain the proposal outlined above, Hayes would need to employ vowel epenthesis here too, but this is totally unmotivated as no complex onset cluster exists. As a result, r would not delete, since the appropriate, i.e. intervocalic, environment for doing so would not exist, hence lengthening would be impossible in the first place. The attempt, then, to divert interest to an alleged epenthetic vowel rather than the onset itself is futile. Reference to onsets is imperative given the behaviour of r word initially. 3.2.3.3.2 A phonetic approach to CL Kavitskaya (2002) builds an account of CL that treats it as phonologization of inherent phonetic length. Her argument is based on the fact that vowels in CVC syllables are phonetically longer when followed by certain consonants whose transitions can be misheard as part of the vowel (i.e. sonorants, approximants). Upon deletion of such consonants, the ‘excess’ length of vowels can be justified if phonologized. Thus, vowels are reinterpreted by listeners as phonemically longer. Given that the burden in Kavitskaya’s model is on the profile of consonants that can cause CL, she allows for onset deletion to induce CL. In fact, she

Onsets and compensatory lengthening

121

considers the SamG data (actually a fuller range than Hayes), although not the complete set. Her claim is that r is vocalic enough to be reinterpreted as additional vowel length when lost, unless blocked by a prohibition against super-long vowel sequences, as is the case with medial singleton r. However, Kavitskaya fails to explain the SamG coda r facts. Recall that r’s in codas survive, but this is puzzling in her model. If reinterpretation of the length of the r occurs in onset position, then it should surely also occur in the prototypical position for such reinterpretation, i.e. the coda, but it does not. Another prediction that proves erroneous given the SamG facts relates to the behaviour of glides. One would expect that if r is vocalic enough to lead to reinterpretation of extra vowel length, then similar effects should be caused by the more vocalic glides. However, no empirical evidence supports such an expectation. Thus, while Kavitskaya is right in permitting CL from onset loss, the implementation of this idea at least with regard to SamG is less successful. 3.2.3.3.3 OT analyses of CL Within OT numerous proposals have been put forward to deal with CL. Given that many have used Hayes’ (1989) theory of mora conservation as a starting point, a common issue they encounter is the following: in order that CL of the /CVC/ → [CVː] type works in a derivational model, the segment that gets to be deleted should have acquired a mora, so that when it deletes, its mora can survive and migrate to a neighbouring segment. The usual way to do this is by applying Weight-by-Position (WbyP). In classic OT, i.e. one that uses a single input–output mapping, there is no space for such a derivational step, so one would need to assume that the segment which eventually gets deleted should bear a mora underlyingly, e.g. /CVμCμ/, and let high-ranking MAX-μ force its realization on another segment. The problem with such a move is that it violates the Richness of the Base, which states that no restrictions should be placed on inputs, thus we are also obliged to consider the same input, but without any mora on the coda, i.e. /CVμC/. In a language in which CVC CL occurs, we should be able to generate the same [CVː] output regardless of this mora specification, but the problem is that emergence of CL hinges exactly on that specification, so that [CVμ] would always be preferred over [CVμμ] for input /CVμC/, because the latter also violates DEP-μ. Alternatively, we could indeed suggest that /CVμCμ/ and /CVμC/ map to different outputs, but this then implies an undesirable contrast between non-moraic and moraic singletons in the same language (see Bermúdez-Otero 2001; Campos-Astorkiza 2004 for discussion).

122

Onsets

In the current theory, this problem has been avoided by abandoning the idea that CL is about mora preservation; instead, it involves position preservation through a mora. Introduction of P-DEP-μ (§3.2.2.2) additionally ensures that no matter what the assumed moraic status of the input, lengthening (but not inherent length) will be penalized. Also, unlike many other OT accounts of CL, no need to introduce intermediate levels emerges. Some do this more covertly than others. For instance, Sprouse (1997) mimics the intermediate level by means of what he calls ‘enriched input’, McCarthy (2003b) by means of a ‘sympathetic candidate’ and Goldrick (2000) requires reference to covert structure and proliferation of (to a certain extent new) constraints to capture different types of CL. Others introduce intermediate levels explicitly. Stratal Optimality Theory (Kiparsky 2000; Bermúdez-Otero in prep.) introduces levels making a more principled use of morphological strata such as the stem- or word-level. This grounding in morphology is promising, but for the most part inapplicable to CL, which is often oblivious to morphological levels and consequently its adequate treatment remains elusive. Most recently, the Candidate-Chain Theory (OT-CC; McCarthy 2007; Shaw 2007) explicitly makes reference to intermediate stages by means of candidate chains, each member of which needs to be more unmarked than the previous one. In addition, a set of Prec (edence) constraints are imposed which function as extrinsic rule ordering. These two components of the framework suggest a full circle back to serial derivational approaches. It remains to be seen whether such a hugely costly move is justified. Finally, a different avenue that has been explored and which is somewhat reminiscent of the present proposal refers to segmental faithfulness that keeps track of the number of segments involved (indicated by means of indices). Accounts of this sort also fail as they suffer from various problems. For instance, Lee (1996) is inconsistent in sometimes using moras and sometimes segments to account for CL, Hermans (2001) fails to predict CL after onset loss in the first place and Sumner (1999), although accepting such a possibility, nevertheless predicts various unattested structures (see §6.1.3. in Topintzi 2006a for details). 3.3

Extensions of the proposal

Preservation of positions through moras offers a successful account of SamG CL. To further corroborate this idea, a number of additional advantages and extensions will be discussed next. First, I start by ensuring that besides uncommon instances of CL, e.g. that of SamG, this analysis can also account for more

Onsets and compensatory lengthening

123

straightforward cases, such as CL due to prenasalization and glide formation. This is tackled in §3.3.1. Then, I present the multitude of ways in which the present analysis is superior over others; in particular, we see that: a) it explains CL caused by the deletion of unsyllabified segments (§3.3.2), b) it explains why onsets, like other syllable constituents, may cause CL (§3.3.3), and further to that, c) it predicts that onsets may not only be the triggers of CL, but also the undergoers (§3.3.4). These results are particularly significant because, although they emerge as natural predictions of the current framework, they are also actually verified through empirical data, some of which have barely been observed, but others have already been discussed in the literature in different terms (see Piro in Lin 1997). Notably, though, until now no single framework has managed to unify all these cases under the same analysis, a task the present proposal aspires to achieve. 3.3.1 CL without segment loss The data discussed so far involved deletion of a segment and lengthening of a neighbouring one. But there are also other – rather typical – cases of CL, where lengthening occurs despite the absence of deletion of a segment. Standard examples include Bantu prenasalization and glide formation. More specifically, Luganda inputs like /ba-ntu/ surface as [baːntu] ‘people’ with prenasalization and /mu-iko/ as [mwiːko] ‘trowel’ with glide formation. Both processes are accompanied by lengthening. This is at first glance surprising, given that no segment loss occurs. However, a closer inspection of the definition of PosCorr reveals that those cases too are anticipated. (42)

PosCorr: An input segment must have an output correspondent either segmentally by means of a root node or prosodically by means of a mora

The root node here plays a key role. As work by Clements (1986), Sagey (1986) and Rosenthall (1997) indicates, in the output sequences mentioned above, complex segments emerge that consist of a single root node which contains the nasal features or those of the high vowel – in baːntu and mwiːko, respectively. Thus, no segmental correspondent for the nasal or the high vowel exists by means of a root node. Nonetheless, under the ranking PosCorr >> P-DEP-μ, some correspondent for them must exist. In the present circumstances, that can only be a prosodic correspondent by means of lengthening of the neighbouring vowel. Data from Kinyarwanda provide a nice illustration of those claims (Sagey 1986), as the language admits sequences of a consonant followed by a glide. Compelling evidence that this is a single root node containing a complex segment with multiple articulations relates to the fact that Kinyarwanda bans

124

Onsets

complex onsets. In particular, German loanwords that involve complex onsets are split by epenthesis in Kinyarwanda so that simplex onsets are created, e.g. Republik > repuburika, Präsident > perezida, etc. In contrast, complex segments are permitted, a fact that can easily be explained if these are made of a single root node. Adopting for illustration purposes Sagey (1986: 75), Kinyarwanda examples that contain complex segments (43a) receive the representation in (43b) with glide formation and subsequent CL. As is apparent, the first vowel creates a complex segment with the preceding consonant under a single timing slot (or root node). The first vowel’s timing slot re-associates with the second vowel leading to CL. (43)

a. /ku-i-Bon-a/ → [kwi:Bona]6 ‘to see oneself’ /ku-gu-ir-a/ → [kugwi:ra] ‘to fall on’ b.

C

V

V

x

x

x

C

V

V

x

x

x

⇒ O

N

C

V

V

x

x

x

⇒ O

N

σ

O

σ

N σ

In the current model, these representations can be modified as exemplified in (44) for glide formation and in (45) for prenasalization. Common to both is that the features of the nasal and the glide are not realized in separate independent root nodes, but as parts of the root nodes of their neighbouring segments. (44)

Glide formation and CL in current model C1

V2

V3

r.n

r.n

r.n

µ2 ⇒

…1 …1 …1

6

µ3

C1

V3

r.n

r.n

…2 …2 …2

The transcription [kwiːBona] which differs from Sagey’s original [kwiiBona] is chosen to draw attention to the fact that we are talking about complex segments with multiple articulations as well as long vowels (rather than a sequence of identical vowels).

Onsets and compensatory lengthening (45)

a. /ba-nde/ → [ba:nde] ‘who?’

125

/ku-ngana/ → [ku: gana] ‘to be equal’

b. Prenasalization and CL in current model V1

N2

C3

r.n

r.n

r.n

µ1 ⇒

µ2

V1

C3 r.n

r.n

[nas]2

…3 …3 …3

Turning now to an OT analysis, (46) depicts how prenasalization can be accounted for (a similar analysis holds for glide formation too). Prenasalization informally refers to the triggering force behind prenasalization (whatever that is). If high ranked, as here, it rules out non-prenasalized candidates early on (cf. (46b)). All other candidates involve prenasalization; coda-onset syllabification is not considered, simply because the language bans codas on the whole. Among the remaining candidates, then, (46a) violates Uniformity due to fusion and (46c) fails PosCorr under the assumption that there is no position correspondent for n3, based on the previous empirical evidence that nasals cannot appear with their own root node. Consequently, (46d) wins, because its violation of P-DEP-μ – due to the addition of a mora as an output prosodic correspondent for /n/ – is too low ranked. (46)

/ba-nde/ → [baːnde] Prenasa/b1a2μn3d4e5μ/ Uniformity lization PosCorr μ n μ 1 2 3,4 5

a. b a . d e μ 1 2

μ 3 4 5

b. b a .n d e c.

μ n3 μ 1 2 4 5

ba . d e

☞ d. b a

μ2μ3 n μ 1 2 4 5

.de

P-DEP-μ

*! *! *! *

3.3.2 CL after deletion of unsyllabified segments The approach advocated here claims that CL is intertwined with the preservation of phonological positions. Given that CL is not tied with moraic segments in certain syllabic positions, it is predicted that CL can also occur as a result of deletion of segments that are unsyllabified or non-moraic. Data from Piro seem to confirm such a prediction.

126

Onsets

Piro is an Arawakan language spoken in Eastern Peru (Lin 1997; Matteson 1965). While there is consensus that all syllables in Piro are onsetful and open, some controversy exists over how the various consonant clusters are syllabified. Matteson (1965) argues for a (C)(C)CV structure, whereas Lin (1997) suggests that syllables are simply CV and remaining consonants are unsyllabified. Lin observes that in all positions in the word, all possible clusters are allowed, ignoring sonority sequencing principles. Only some clusters are banned, since their co-occurrence leads to OCP violations by sharing similar place and/or manner of articulation. (47)

Sample of Piro clusters tpa mwenutu wyoptota ksu tmennu smota

‘curve’ ‘cheap’ ‘we receive’ ‘tube’ ‘flaw’ ‘blunt point’

pto wmahataya ywalitxa skota mtenotu msa

‘…’s group’ ‘we lack’ ‘hip’ ‘lower abdomen’ ‘short’ ‘empty corn cob’

Consonants which are not in a prevocalic position surface either as syllabic or are followed by a very short epenthetic vowel, whose role is to facilitate consonant–consonant transitions. This vowel is inactive phonologically speaking, so it ‘is best treated as a phonetic phenomenon’ (Lin 1997: 406). Piro’s interesting property for our purposes is the fact that it has a ‘boundary vowel deletion’ (BVD) rule which deletes the final vowel of each lexical root and derived stem at each stage of suffixation. The process applies cyclically from the innermost domain to the outermost one. There are two exceptions to BVD. First, certain suffixes are lexically marked as not triggering the deletion process (see (49)). Second, BVD fails to apply if it would lead to illicit or extra-long clusters (I am following Lin’s notation in marking the deleted vowel with the underscore and the non-deleted vowels with boldface). (48)

Boundary vowel deletion (BVD) nika + ya + waka + lu ‘to eat it there’ to eat-loc-place-it cycle 1 nika+ya → nik_ya cycle 2 nikya+waka → nikyawaka cycle 3 nikyawaka+lu → nikyawak_lu

On the first cycle the stem vowel deletes, but this does not repeat on the next cycle, since deletion would lead to the illicit [kyw] cluster. On the final cycle BVD can again apply, because no similar risk arises. On other occasions, BVD is accompanied by subsequent consonant deletion which optionally or obligatorily leads to CL of the previous vowel (49). Since it is not my aim to focus

Onsets and compensatory lengthening

127

on the exact conditions under which BVD is blocked or CL applies (see Lin 1997, 1998a), I will focus on the process itself as illustrated in the following example. For presentational purposes, I will assume that a simplified version of a constraint *CCC is in effect banning three-consonant clusters, whereas OCP forbids certain clusters that share place or manner of articulation. Both of these lead to consonant deletion and subsequent CL. Of course, a complete analysis of Piro requires consideration of the full range of facts. (49)

BVD, C-deletion and CL (non-BVD-triggering affixes are marked with capitals) i.

obligatory CL nika+ka → nikka → niːka hira+re-TA → hirreta → hiːreta xitxi+tši → xitxtši → xiːtši

ii. optional CL wane+YA+yi → waneyyi ~ waneːyi ruhi…TA+hima → ruhhimata ~ ruːhimata

‘he is eaten’ ‘to drink’ ‘foot’ ‘you have’ ‘it is said that he answers’

Significantly, Piro has no underlying weight contrasts. There are no geminates other than the ones which are the result of morphological concatenation. Moreover, the distinction between long and short vowels is not phonemic, but derived, since all long vowels are the outcome of CL. In addition, since consonants are not syllabified as codas, they are presumably syllabified as onsets and against all expectations – of standard moraic theory – seemingly trigger CL. Piro is thus an interesting case because it appears to present a counterargument to Hayes’ (1989) claims in more than one way and challenges most versions of moraic theory. In particular, it shows that pre-existing syllable weight contrast is not a prerequisite for CL and that – adopting Matteson (1965) – onsets may participate in CL, e.g. xi.txtši → xiː.tši (49). An analysis based on the preservation of phonological positions, on the other hand, faces no such problem. Since syllabification and moraification are beside the point, data like those presented in Piro are entirely expected, as explained below. In an attempt to avoid these problems, while sticking to the core of Hayes’ CL account, Lin makes two basic assumptions in her analysis of Piro. First, she claims that consonant clusters are not syllabified as complex onsets. The latter syllabification – apart from the CL problem – makes it difficult to account for certain co-occurrence restrictions. For instance, if consonants of any type make legitimate onsets, it is difficult to explain why *CCC should hold. Thus, Lin proposes that only the prevocalic consonant is syllabified in an onset. Other

128

Onsets

consonants before it remain unsyllabified. The absence of sonority restrictions and the possibility for virtually all combinations between consonants (other than the OCP-violating ones) is compatible with this proposal. This does away with the problem of CL after onset loss that Matteson (1965) faces. Second, Lin assumes that all segments, and consequently unsyllabified consonants too, are underlyingly moraic (see Bagemihl 1998), an issue I will return to later. This move has two gains. First, under the assumption that CL is about mora preservation, it accounts for the occurrence of CL after consonant deletion. Second, it allows Lin to modify Hayes’ claim about the necessity of input syllable weight contrasts as a prerequisite for CL to the following generalization: ‘a language that has the moraic and non-moraic contrast among consonants or the monomoraic and bimoraic contrast among vowels can exhibit CL’ (Lin 1997: 424). According to Lin, Piro falls under the first category, since it has a contrast between unsyllabified moraic consonants and syllabified moraless onset consonants. Notably, however, some of Lin’s proposals generate a host of other problems. In particular, the weight distinction she employs makes reference to output contrasts, which can be troublesome and circular. To illustrate, imagine a language that not only lacks unsyllabified consonants altogether but also any positive evidence about the moraicity of consonants, which could consequently be considered as non-moraic. Given Lin’s claims, the only other way for this language to have CL is by making reference to the long vs. short vowel contrast. But it is actually possible that a language fits this criterion by presenting no other distinction but a V vs. VV contrast due to CL itself. This is of course circular, since the validity of CL is based on the moraicity contrast, which in its turn hinges upon the CL facts. Additionally, the success of her analysis crucially depends on the assumption that all segments are underlyingly moraic. This is a severe restriction over the structure of the inputs, thereby violating the Richness of the Base – a dubious for some, yet important tenet of OT (Prince and Smolensky 1993/2004). These problems vanish in the alternative that treats CL as position preservation. The idea is that unsyllabified segments may lack syllable affiliation, but are like all other segments in requiring an output correspondent. As we have seen before, this may come in the guise of an entirely faithful output that preserves the segments in question or in the shape of moras which serve as the prosodic remnant of that position. When a consonant deletes, the first option is ruled out. The only solution, then, is CL due to PosCorr >> P-DEP-μ. Notably, assignment of input moras plays no role in this, contra Lin (1997), since independently of whether the consonant that deletes includes an underlying mora or

Onsets and compensatory lengthening

129

not, P-DEP-μ is identically violated when lengthening of the preceding vowel occurs. A sketch of this account follows. (50)

/mtasa+xe/ → [mtaːxe] – first attempt /m1t2a3s4a5+x6e7/ a. m1t2a3s4a5x6e7 b. m1t2a3s4x6e7 c. m1t2a3x6e7 ☞ d. m t a

μ3μ4 1 2 3 6 7

xe

BVD

OCP f

Pos Corr

*!

*

P-DEP-μ

*!

**! *

*

(50a) violates the constraint which forces the boundary vowel deletion process, informally stated here as BVD. OCPf – after Lin – is a constraint militating against adjacent fricatives. This is violated in (50b) where /s/ and /x/ are found next to each other. We are now left with (50c) and (50d). Both apply BVD (so deletion of a5 equally violates PosCorr) and delete the fricative /s/ to avoid an OCPf violation. The difference lies in the fact that (50d) ensures an output correspondent for /s4/ by inserting a mora and thus lengthening the vowel /a3/. In this way it satisfies PosCorr better. The lowest-ranked P-DEP-μ violation is thus unimportant. By failing to lengthen, (50c) lacks such a correspondent and therefore loses with respect to PosCorr’s evaluation. A challenging question that arises at this point relates to the absence of lengthening after the vowel’s deletion due to the application of BVD. More concretely, the winning candidate presents /a5/ deletion, but there seems to be no moraic correspondent for this segment. While it is quite unclear what the remedy should be in Piro, there is one possible way out of this. Perhaps, Piro provides evidence that PosCorr can be realized by means of its subparts PosCorr-C and PosCorr-V. The former requires a position correspondent for consonants and the latter for vowels. This should not come as a surprise since PosCorr’s related constraint Max-Seg has often been claimed to require separation into its two variants: Max-C and Max-V (used in several analyses, e.g. Kager 1999 on Southeastern Tepehuan, Hall 2000 on Zoque, Kiparsky 2002 on Arabic, etc.). For Piro, then, all we would need to do is to rank PosCorr-C above P-DEP-μ so that its deletion is followed by subsequent lengthening, but PosCorr-V below P-DEP-μ, so that no lengthening is caused after vowel deletion. This becomes obvious in the comparison of candidates (51d) vs. the winner (51e). The former fails to provide a position correspondent for both s4 and a5. (51e) is similar in lacking a position correspondent for a5, but preserves /s/’s position by lengthening the previous vowel.

130

Onsets

This prioritization of consonantal over vocalic positions is reflected in the ranking: PosCorr-C >> P-DEP-μ >> PosCorr-V. One more candidate deserves extra discussion, namely (51c). This is like the winner (51e), but in a way also better, since it has a prosodic correspondent for s4 by means of a mora. However, (51c) incurs a more serious violation of P-DEP-μ, given that the non-positionally μ-licensed moras (µ4 and µ5) linked to a3 have no correspondents linked to them in the input. Alternatively, (51c) can be ruled out because it includes a trimoraic syllable and thus violates *3μ. For illustration purposes, I assume the former. (51)

/mtasa+xe/ → [mtaːxe] – revised version /m1t2a 3s4a5+x6e7/ a. m1t2a3s4a 5x6e7 b. m1t2a3s4x6e7

BVD

OCPf

Pos P-DEP-μ Pos Corr-C Corr-V

*! *!

*

μ3μ4μ5 1 2 3 6 7

c. m t a

xe

d. m1t2a3x6e7 ☞ e. m t a

μ3μ4 1 2 3 6 7

xe

**! *!

* *

*

3.3.3 Onsets are like other CL-inducing constituents By this point, a concern, or even objection, will have emerged. If the ‘CL-asposition-preservation’ enterprise is on the right track and if onsets can indeed induce lengthening, then why is it so exceedingly rare? Importantly, this question pertains to other syllable constituents too, although it is virtually never asked. To begin answering this question, consider first coda deletion, which is commonly accompanied by lengthening, but this is by no means mandatory. As a result, a great number of potential CL cases actually never really emerge. For example, Diola-Fogny coda consonants delete in heterorganic coda-onset clusters (Sapir 1965, Kager 1999) as a means of satisfying the coda condition, but no lengthening of the preceding vowel occurs in spite of the existence of phonemic vowel length (Sapir 1965: 6). For instance, /let-ku-jaw/ surfaces as [lekujaw] and not *[leːkujaw] ‘they won’t go’ or /jaw-bu-ŋar/ becomes [jabuŋar] rather than *[jaːbuŋar] ‘voyager’. The fact that not all instances of coda deletion lead to CL is also acknowledged in Hayes (1989); in a rule-based derivational system like his, which ties moraicity and CL so closely, such variability can easily be attributed to the point at which WbyP applies: if it applies before deletion, then CL will occur;

Onsets and compensatory lengthening

131

if later than deletion, then CL cannot occur. Furthermore, in Hayes’ framework, onsets are never moraic, so their loss cannot cause CL. There is, however, a prediction that Hayes makes, namely that vowel deletion should always trigger CL, simply because nuclei are always moraic.7 But this is wrong. To see that, consider the Australian language Lardil (Kenstowicz 1994a and references cited therein). First, notice that final consonants delete if they are noncoronal but without inducing lengthening of the previous vowel, i.e. /ŋaluk/ > *[ŋaluː] or /thuraraŋ/ > *[thuraraː], although Lardil also admits long vowels, e.g. [kudmee] ‘happy’, [maari] ‘extract’, [waaka] ‘crow’ (Hale 1981a). To explain this lack of lengthening, perhaps one could assume that WbyP applies too late to be able to cause CL. The stem alternations below exemplify. (52)

Lardil stem alternations: final C-deletion Absolute ŋalu thurara

Inflected ŋaluk-in ‘story’ thuraraŋ-in ‘shark’

This, however, leaves unexplained the lack of lengthening in the Lardil data below where final vowel apocope applies (in stems of three or more syllables). This time a vowel – which is unquestionably moraic – deletes, but again no CL applies, against all expectations. (53)

Lardil stem alternations: final V-apocope8 Absolute yalul mayar

Inflected yalulu-n mayara-n

‘flame’ ‘rainbow’

A quick search in standard textbooks reveals how easy it is to find several similar examples where a vowel deletes in the absence of any lengthening. Some of the examples include: Chukchee vowel apocope (Kenstowicz 1994a: 105), e.g. nileq ~ nileqe-t (absolute sg. ~ absolute pl.) ‘match’; Tangale vowel elision (Kenstowicz 1994a: 96); Klamath first stem (short) vowel syncope between a CV-prefix and a following CV syllable, e.g. mačhaːt-ka ‘listens’ ~ sna-mčhaːt-iːla ‘causes to hear’ (Odden 2005: 121), and Icelandic i-deletion and vowel syncope (Odden 2005: 189–90). Of course, in most cases things are more complicated, so the interested reader is invited to check these works and the references cited therein for details. 7

8

This is overwhelmingly the case. However, there are instances where some vowels have been claimed to contribute no mora. See, for instance, Shaw (2004) who argues that schwa in Mohawk is non-moraic. V-apocope and final C-deletion may also co-occur when the stem ≥ 3σ and the consonant in question is labial or velar. Unsurprisingly, CL again does not apply, e.g. the inflected [putuka-n] and [tipitipi-n] emerge in the absolute form as [putu] ‘short’ or [tipiti] ‘rock-cod species’.

132

Onsets

In sum, segment deletion, even if the segment is uncontroversially moraic, i.e. a vowel, is not guaranteed to be counterbalanced by CL. Consequently, the proposal that attributes the absence of onset CL to the lack of onset weight cannot be maintained. Instead, it seems more likely that the absence of CL after onset loss simply relates to the lack of CL that is anyway observed in languages. While this observation gives us a general clue to the less-than-expected frequency of CL, it provides no specific information as to why onset CL in particular is not only fairly uncommon, but in fact increasingly rare. I would like to suggest that there are some additional reasons at play which pertain specifically to onsets. Most prominently, it is the case that cross-linguistically coda loss is overwhelmingly more common than onset loss. The latter is repeatedly observed in the beginning of the word in Australian languages, but virtually none exhibits CL. Nonetheless, it is reasonable to say that since onset loss is on its own quite rare, the chance for CL after such loss obviously dramatically decreases. An important exception seems to be Proto-Austronesian (PAN) initial consonant loss, which, according to Zewen (1977: 9–10, 36), has generally led to CL, relics of which can still be seen in Marshallese (MAR) today (54). (54)

Proto-Austronesian initial consonant loss and CL in Marshallese9 PAN

MAR

Gloss

bitun binih ɣaput’ pukәt puna hat’ap

iːju iːne äːut oːk uːn aːt

‘star’ ‘seed’ ‘to wrap around’ ‘drag net’ ‘origin’ ‘smoke’

An additional factor is that deletion of an onset and subsequent lengthening can often lead to super-long vocalic hiatus, which, as we have seen before (§3.2.3.2.1), is universally non-preferred (Kavitskaya 2002). This type of hiatus would arise in virtually all cases where there is a singleton intervocalic onset, i.e. VCV, that gets deleted. By ranking the constraint *S-L VH highly, we can ensure that lengthening of this type cannot occur. As a matter of fact, Samothraki Greek is an excellent example of this type. Here, r deletes from an onset position, and causes lengthening whenever this would not violate *S-L VH, namely word initially and in a complex onset cluster. Word medially, its 9

́ ‘obstructed’. Also Cluster simplification did not lead to CL, as in PAN *mpәnәd > MAR boɳ note that /b/́ stands for a rounded bilabial stop and /t’/ for an unreleased dental stop.

Onsets and compensatory lengthening

133

deletion would result in the prohibited VːV configuration, thus no lengthening takes place. This makes the strong prediction that there will be no language with compensatory lengthening after onset loss word medially, without equivalent lengthening word initially. In fact, this prediction is borne out in the Marshallese example above: PAN ɣaput’ > MAR äːut, where the MAR form presents deletion of both the initial as well as the medial onset. Lengthening, however, occurs only after the loss of the former. More generally, then, since *S-L VH seems to be cross-linguistically highly ranked, it falls out that it frequently destroys the environment for CL after onset loss. Putting all these factors together should give us a good idea as to why onset CL is so rare. 3.3.4 Onsets as a target of CL A further prediction that the present model makes is that onsets not only should be able to induce CL, but may also undergo it by means of bearing a mora. In such cases, loss of a position causes lengthening of an onset consonant which is now rendered a geminate (see Ch. 5 for examination of geminates as moraic onsets). Certain dialects of Trique (Oto-Manguean) provide a compelling example of this type, whereby a vowel is lost or shortened and the preceding consonant lengthens in response. Trique is a tonal language, which also exhibits stress consistently on the final syllable. This syllable is special in a number of ways: it may license numerous tonal, consonantal and vocalic contrasts (see Hollenbach 1977 for details). The relevant contrast for current purposes is the existence of fortis stops and fortis sibilants in the San Juan Copala dialect of Trique (henceforth SJC) that are crucially found only in the onset of the final syllable and nowhere else. The fortis consonants are voiceless and unaspirated (Hollenbach 1977: 36–7). The interesting effect arises in phrasal stress, which is expressed by means of final-syllable intensity and by lengthening of open long vowels or shortening of short vowels. ‘The latter is accompanied by compensatory lengthening of an immediately preceding fortis stop, fortis sibilant or resonant’ (Hollenbach 1977: 49; emphasis added mine). Incidentally, note that the consonants that undergo CL are exactly the ones predicted in the current model to be coerced moraic onsets under the ranking *μ/ONS/[+voi] >> Moraic >> *μ/ONS (cf. (31) in §1.3.3.5), assuming of course that resonants in SJC are unspecified for [voi]. Presumably, here the reduction of the short vowel is compensated for by lengthening of the preceding onset consonant. Taking into account the lack of codas – other than word-final ones, which are necessarily laryngeal (Hollenbach 1977: 36) – in SJC Trique, it must be the case that the newly created geminate syllabifies wholly in a moraic onset.

134

Onsets

A similar instance of CL appears in the San Andrés Chicahuaxtla (SAC) dialect of Trique. SAC differs from SJC in having both fortis resonants /mː nː lː yː wː/ which only appear in monosyllables and lenis sonorants /m n l y w/. The characteristic of fortis resonants is that they are quite long (Hollenbach 1977: 50). Interestingly, SAC fortis resonants developed mainly by a lengthening of simple resonants to compensate for the loss of a penult (Longacre 1957: 18). Possibly, the easiest way to achieve this change was by deleting the [+voi] feature of non-moraic-onset sonorants and turning them into unspecified moraic ones. Again, given *μ/ONS/[+voi] >> Moraic >> *μ/ONS, only the fortis resonants should be unspecified for [voi].10 (55)

SAC Trique fortis resonants (Hollenbach 1977: 58; a dot underneath a vowel indicates it is short) SJC SJC SJC

yum­ yan® yuwe

SAC SAC SAC

mːL nːa wːe

‘swdet potbto& ‘loft’ ‘century plant’

In an analysis that treats initial moraic geminates as weightful onsets (see Ch. 5 for data and elaboration), this is clearly an instance of CL that leads to a moraic onset. 3.4

Conclusion

This chapter has treated moraicity as the remedy for, rather than the trigger of, compensatory lengthening. I have proposed an analysis applicable to a wide range of CL cases, although by no means to all. Perhaps some further finetuning will be required to handle a more exhaustive list of CL cases. The general idea has been to introduce the notion of ‘position preservation’. Under the reasonable assumption that a segment requires an output correspondent, I have claimed that this can occur either segmentally as a root node, or prosodically via a mora. This is dubbed PosCorr. When the segment deletes, only the prosodic correspondent can maintain the position in question. In the languages where PosCorr is of high priority, lengthening through a mora occurs as a response to its satisfaction (Samothraki Greek, Piro, etc.). The net effect is CL. In the languages where such lengthening is banned, PosCorr is either sacrificed so that no output correspondent for the segment exists (Lardil, §3.3.3), or it is satisfied by having a segmental correspondent by means of fusion (Portuguese, §3.2.2.1). 10

It is thus possible that SAC Trique has two types of resonants: the lenis ones which are [+voi], and the fortis ones which are unspecified for [voi].

Onsets and compensatory lengthening

135

While my account has claimed that onsets can both trigger and undergo CL, this chapter has for the most part been structured in such a way that onsets could be assigned moras, but do not have to be. For instance, r in SamG could bear a mora (Topintzi 2007), but the proposed analysis by no means relies on that. This has been deliberate given the general framework of analysis assumed, i.e. OT (Prince and Smolensky 1993/2004), in which restrictions over inputs are banned (according to RoTB). We thus want to make sure that our analysis will work independently of the input moraicity of the segment that triggers CL. Previous accounts could achieve this either by assigning moraicity at a level intermediate between input and output, thus re-introducing serialism in parallelist OT (Sprouse 1997; Goldrick 2000; McCarthy 2003b, 2007; Shaw 2007), or by simply violating RoTB itself (Lee 1996; Lin 1997) – another important tenet of the model. With this in mind, the proposal put forward has been purposely devised in such a way as to side-step the issues above. The gain in doing so is significant. Apart from not compromising OT’s orientation to the output or running up against RoTB, the theory is, we could say, ‘framework-neutral’. This means that although it is designed to conform to the basic theoretical tenets of OT, it is also consistent with other frameworks, e.g. rule-based derivational ones that introduce moraicity at some level of representation through rule application. In other words, whether we assume that CL-triggering onsets can be moraic underlyingly or become moraic at some stage or bear no mora whatsoever, the point is the same: onsets do trigger CL. And this is the insight that the present chapter would like to offer independently of the framework assumed. Nonetheless, and faithful to this book’s aims, I claim that CL actually must be used as supporting evidence for onset moraicity. This is inevitably so, when one examines instances where onsets act as undergoers of, rather as triggers for, CL. Regardless of CL being about position preservation through a mora (this chapter) or – in a more standard view – about mora preservation (analyses à la Hayes 1989), data like those of Trique (§3.3.4; §5.3.2) reveal that onsets bear moras as a response to the deletion of a segment. We will see additional cases of that sort when discussing Pattani Malay and Trukese (§5.2.1 and §5.2.2, respectively). ‘CL as position preservation’ has further merits too. Exactly because it focuses on positions rather than on particular syllable constituents, it treats all segments as equal and thus as equally able to cause CL.11 Consequently, it 11

On closer inspection, some refinement may be required, since it is an oversimplification to say that all segments are equal. For instance, epenthetic segments in some languages are invisible for certain processes, e.g. the Marshallese epenthetic yV- prefix in reduplication (Hendricks 1999).

136

Onsets

provides elegant solutions to cases previously considered problematic, such as CL due to unsyllabified segments (§3.3.2) and, of course, CL after onset loss. As a result of its output orientation, it also indicates that reference to terms such as ‘CL after coda loss’ or ‘CL after onset loss’ is descriptively useful, but inaccurate. This is because the lost segments cannot be identified as codas or onsets, since in the input no syllabification applies, while in the output, the segments in question no longer surface. In such instances, then, it makes much more sense to simply talk about CL after consonant deletion.

4 Onsets and word minimality

4.1

Introduction

This chapter deals with word minimality, that is, the requirement that certain languages have for words to exhibit a minimum size. Highly common minimum size requirements are that words are at least bimoraic (C)VV as in Ket or Mocha (Gordon 2006) or (C)VV/(C)VC, e.g. English or Evenki (Gordon 2006), or bisyllabic, e.g. Pitta-Pitta (Hayes 1995: 201). Standard moraic theory (Hyman 1985; Hayes 1989) predicts that no language with a bimoraic wordminimum will exist that allows [CV] words alongside clearly bimoraic [VV] (and [VC]) words, but bans [V] ones. This is indeed impossible given a model that onsets never contribute weight, but entirely predicted by a model such as the present one. Indeed, the minimality data of Bella Coola (also known as Nuxálk) lend support to this approach. This chapter argues that Bella Coola onsets are moraic, although in a very marginal environment, namely in [CV] words only. This happens so that the overarching requirement of bimoraic word minimality can be satisfied. In this way, Bella Coola presents a case of coerced onset weight, where onset moraicity is enforced by a higher imperative in the language, namely word minimality (§1.3.3.5). That onsets are not generally moraic in the language is certified by the root maximality data, which impose a four-mora maximum for roots. Attested roots conform to this restriction with the provision that onsets do not contribute any moras, offering evidence that they bear no weight. An analysis framed in OT, such as the one that follows, can capture the contextual character of onset-moraicity enforcement. To the extent that Bella Coola allows moraic onsets, it is particularly interesting, as it potentially illustrates an instance of a language where all types of onsets can act as moraic independently of their type (§1.3.3.5), because of the ranking: (1)

BeMoraic >> *μ/Ons/[+voi] >> *μ/Ons

(cf. (35c) in Ch. 1)

In such a language, all voiced and voiceless obstruents, as well as sonorants, would be rendered moraic to satisfy word minimality, because of the top 137

138

Onsets

ranking Be Moraic, e.g. [ƛ’i] ‘fast’, [xwl̩] ‘to pull’, [wl̩] ‘to spill’, [mu] ‘fishtrap’ (Nater 1979: 170, 171, 175). Evidently, as the examples suggest, voiceless obstruents and sonorants must be able to bear moras. Because sonorants may be unspecified for [voi] or bear [+voi] (§1.3.3.1 & §1.3.3.4–5; Ch. 2), to be able to argue that Bella Coola offers evidence for (1), we would need to see what happens with voiced obstruents which are consistently [+voi]. Had Bella Coola allowed voiced obstruents occurring in [CV] words, then we would be bound to argue that they are moraic, in the same way all other onsets are, supporting (1); on the other hand, if the language had voiced obstruents, but these failed to occur in [CV] words, then presumably voiced obstruents would be unable to be moraic onsets, thus only segments lacking [+voi] could be moraic, providing evidence for the pattern *μ/Ons/[+voi] >> BeMoraic >> *μ/Ons. However, what actually goes on cannot be determined, given that Bella Coola lacks word-initial voiced obstruents (Nater 1979: 171). This chapter is structured as follows. Section 4.2 offers a general phonological background of Bella Coola focusing on the root-maximality data (§4.2.2), whose analysis involves the use of moras and crucially their lack in onsets. Section 4.3 deals with the main phenomenon here, that of word minimality providing a strong argument in favour of moraic onsets, since the words that satisfy it are [VV], [VC] and [CV] but not [V]. The core analysis of the minimal and maximal patterns follows (§4.3.1), supplemented by a lengthier analysis that covers the full range of facts and targeted to the OT theorist who wishes to ensure that basic OT tenets like RoTB are observed (§4.3.2). Section 4.4 considers the chief alternatives that avoid use of onset moraicity and aims at persuading the reader that only an onset-weight account is viable. Section 4.5 offers some concluding remarks. 4.2

Bella Coola background

4.2.1 Basic facts To be able to comprehend the word-minimality facts of Bella Coola, it is vital that a more general discussion of the phonology of the language is made. Bella Coola is a Salish language of British Columbia, notorious – like its sister languages – for long sequences of consonants (cf. (4)). Its inventory follows. (2)

Bella Coola inventory p p̕

t t̕

m m|

n n|

c c̕ s

ƛ̕ ɬ l l|

k k̕ x y

kw k̓w xw w

q q̕ χ

qw q̕w χw

ʔ (h)

[NB: /c/=alveolar affricate, /ɬ/=lateral fricative, /ƛ’/=lateral glottalic affricate]

Onsets and word minimality

139

The data and facts I will use are based on Bagemihl (1991 and 1998; henceforth B91 and B98, respectively; page numbers are also included for cross-reference with Bagemihl’s original sources which include Nater (1977, 1984), Newman (1947, 1971) and Saunders and Davis (1972)). Much of what follows in this section reviews B91’s arguments for the relatively normal syllabification of Bella Coola in maximal TRVVC syllables (where T=obstruent, R=sonorant; I use R to represent a sonorant consonant which is either the second consonant in an onset cluster or is syllabic; this should be obvious from the context), and the existence of unsyllabified obstruents. This is important because, as we will see, syllables and unsyllabified consonants play a key role in the overall analysis below. Given the TRVVC template, only TR complex onsets are allowed. TT and RR sequences occur, but are not complex onsets. Vowels and syllabic sonorants can serve as nuclei. Singleton obstruents syllabify either as onsets or codas depending on the environment. Codas receive moras, while onsets are claimed not to be moraic. The latter fact is generally true, but will be disputed later in a particular instance. Some examples follow. (3)

Syllables with vocalic nuclei a. b. c. d. e.

CV CVC CVVC TRVC TRVVC

ƛ’i kwit’ niiχw xnas c’wiiχw

‘fast’ ‘to pry loose’ ‘fire’ ‘woman’ ‘having grey hair’

(B98: 75) (B98: 75) (B91: 599) (B91: 601) (B91: 619)

Syllables with sonorant nuclei f. CR g. CRC

c’m̩ tl̩q̓w

‘index finger’ ‘to swallow something’

(B98: 79) (B98: 79)

In strings of obstruents – which can be as many as three or four root-internally and up to eight or nine across morphemes, e.g. [ɬx̓wtɬcxw] ‘you spat on me’: ɬx̓wt ‘to spit’, −ɬ past tense, −c lSg, −xw 2sg (B98: 74) – the one immediately before the nucleus syllabifies as an onset, the one immediately after a nucleus syllabifies as a coda and the remaining consonants are unsyllabified.1 (4)

Sequences with unsyllabified consonants a. b. c. d. e. f.

1

stn| st’xwm| sqwciɬ c’klakt cipsx pɬtkn|

‘tree’ ‘floor mat’ ‘ventral posterior fin’ ‘ten’ ‘fisher’ ‘bark of bitter-cherry tree’

(B91: 609) (B91: 609) (B91: 609) (B98: 78) (B98: 80) (B98: 79)

In this chapter, I will use boldface to represent unsyllabified consonants.

140

Onsets

This finding is established through the behaviour of obstruents; evidence from reduplication, vowel allophony and glottal stop distribution indicates that they cannot constitute nuclei or syllable margins, thus they must be unsyllabified. For instance, roots with a vocalic or sonorant nucleus get reduplicated with one of the lexically determined CV, CVC or V templates,2 e.g. qayt → qaqayt-i ‘hat → toadstool-diminutive’ with a vocalic nucleus (B91: 598) and tl̩k̕w → tl̩t l̩k w̕ ‘swallow → continuative’ with a sonorant one (B91: 599). Sonorants and vowels, then, can serve as nuclei, but this is not true for obstruents. Had they been able to, then we would expect similar reduplication facts, e.g. kɬ̩- → *kɬ |kɬ-| ‘fall’ (B91: 606), but these are absent from the Bella Coola reduplication system. In fact, most obstruent-only words either do not reduplicate at all or exhibit sonorant or vowel epenthesis. In the latter case, the newly supplied nucleus allows them to get syllabified and reduplicated. Additional arguments against the syllabification of obstruent clusters in syllable margins are available. First, within syllables, a strict sonority profile is observed where sonority rises from the onset to the nucleus and then falls towards the coda. This is why complex onsets can only be TR – presenting rising sonority – and not TT or RR. TT clusters, on the other hand, are indifferent towards sonority considerations (cf. sqwciɬ (4c) or pɬtkn̩ (4f)), therefore they cannot be part of the syllable. Second, reduplication of TR-initial bases leads to a reduplicant where the first consonant of the original cluster syllabifies as an onset and the second – sonorant one – becomes syllabic (5a). Had TT clusters been syllabified too, then we would expect them to pattern identically with TR clusters. But as we have seen already, words made of TT clusters can at best reduplicate if a vowel also epenthesizes (5b). (5)

Reduplication in TR sequences3 a. xwnaɬ → xwn̩xwnaaɬ-i4

2

3

4

‘spring of water → diminutive’ (B91: 615)

Reduplication patterns in Bella Coola are complex but, as my focus is not reduplication, most of the time I will use examples of CV (the simplest) reduplication. The copying of the obstruent and the sonorant (as a nucleus) in the reduplicant is merely a generalization, as data are complicated when one considers other templates of reduplication (i.e. CVC or V patterns) as well as other processes that can accompany them, e.g. vowel syncope, glottal insertion. In the cases presented here, we might perhaps have expected a reduplicated form like (i) TRV.TRV or (ii) TV.TRV for a base /TRV/. The reason TR.TRV emerges as optimal could be attributed to Emergence of the Unmarked issues (McCarthy and Prince 1994), namely that although complex onsets are allowed generally in the language, reduplicants present the unmarked case banning complex onsets. This would exclude (i). Option (ii) would need to be excluded by a constraint that would explain why the original vocalic nucleus of the base is not used as the reduplicant’s nucleus too; the reason is that in this case, the sonorant would have to be skipped, causing a Contiguity violation. Lengthening of the base vowel applies here (as in (7) too). This will not be of concern to us in the current work. For details, the interested reader is referred to Bagemihl (1991).

Onsets and word minimality

141

Reduplication with TT-only sequences (in the cases that this applies)5 b. ɬq’- (base) → ɬn̩q’ (n-insertion) → ɬn̩ɬn̩q’- (reduplication) ‘slap → continuative’ (B91: 607)

As expected, reduplication of words containing initial obstruent clusters followed by a nucleus is possible – exactly because there is already a nucleus available. Interestingly, however, a discrepancy arises when compared with reduplication of TR-initial sequences. In TR clusters, the reduplicant appears prefixed immediately before the base TR sequence and effectively – since the cluster is initial – the reduplicant appears in word-initial position. In TT clusters, though, the reduplicant is not positioned in the same place, i.e. word initially. Instead, it is located before the last member of the cluster. (6)

a. Reduplication in TR-initial sequences xwnaɬ → xwn̩xwnaaɬ-i

‘spring of water → diminutive’ (B91: 615)

b. Reduplication in TT-initial sequences p’ɬa- → p’ɬaɬa ‘wink, bat the eyes → continuative’ (B91: 609) tqn̩k- → tqn̩qn̩k ‘be under → underwear’ (B91: 609) c. Reduplication in a combination of clusters skma → skm̩ kma-y

‘moose → diminutive’ (B91: 615)

The question, then, is whether there is any way to unify these patterns. In fact there is. It simply suffices to say that the reduplicant gets prefixed before the first base syllable.6 In (6a) this coincides with the beginning of the word, since the TR clusters are part of the syllable. In (6b), the reduplicant appears before the last obstruent (in the cluster) of the base, thus failing to align with the left edge of the word. This comes as no surprise if we assume that the first consonant stays unsyllabified and remains outside the reduplication domain. (6c) illustrates the interaction of unsyllabified consonants and TR clusters. As is now anticipated, the reduplicant consists of a copy of the onset obstruent and the sonorant that becomes syllabic. This is prefixed to the syllable and preceded by the unsyllabified consonant s.

5

6

Underlining is used to represent the reduplicant, while normal typeface is used for the base. Boldface is reserved for the unsyllabified consonants. Based on examples such as milixw → milmilixw-ɬp ‘bear berry → plant of the bear berry’ (gloss mentioned on p. 603), Bagemihl (1991: 613) argues that the reduplicant should be prefixed to a foot, rather than a syllable, since mil spans two syllables. This is, however, not necessary in OT. All we need to say is that a CVC template needs to be filled from base segments without having to make specific reference to base syllabification. For present purposes at least, affixation to the first base syllable is sufficient.

142

Onsets

So far, we have seen that obstruents cannot be nuclei and that a maximum of one obstruent can syllabify as an onset. The same result obtains for codas too, as the reduplication facts again reveal. (7)

Reduplication in final TT clusters yaɬk- → yaɬyaaɬkχwalx → χwal χwalx

‘do too much → continuative’ (B91: 617) ‘to melt → solder’ (B91: 617)

In these examples the postulated unsyllabified consonant is at the end of the word. If this were to be incorporated in the preceding syllable, we would expect the reduplicated forms: *yaɬkyaaɬk or *χwalx χwalx,7 but again the reduplicant prefixes before the syllable and fails to copy the final obstruent of the base, as this is not part of the syllable.8 These data, then, confirm that the maximal syllable in Bella Coola is TRVVC. Only a single obstruent can syllabify in an onset and a coda position. Further obstruents are bound to stay unsyllabified. In order that they surface, and given that they cannot be licensed by syllables, Bagemihl claims that obstruents are licensed by moras. As we will see next, this proposal is corroborated by the Bella Coola root-maximality facts, the restrictions of which are computed on the basis of moras and take unsyllabified obstruents into consideration. 4.2.2 Root maximality While it has been shown that Bella Coola only involves a superficially complex syllable structure, its moraic structure is much richer. The major generalization that can be drawn is that Bella Coola roots consist of maximally four moras (i.e. two bimoraic feet),9 where a root is a monomorphemic base to which affixes are added (B98: 91, note 8). Roots may also occur as independent words without any affixes (B98: 93, note 18). Crucially, note that this is a root restriction and not a word one (on the distinction between morphological root and word and its repercussions in the analysis, see Topintzi 2006b, §4.5.1).10 7 8

9

10

Notice that in this instance, the reduplication pattern is a bimoraic one. Alternatively, one might suggest that the reason that forms like *yaɬkyaaɬk are excluded is because these complex codas are banned from reduplicants. In fact, we will see later that the language generally lacks complex codas. This interpretation is indeed possible, but by no means provides a counter-argument to the explanation given in the text. The latter is consistent with the independently established fact that consonants remain unsyllabified under certain circumstances and thus offers a more unified approach. Yoruba (Ola 1995) is another language that posits a two-feet maximum for prosodic words. Ola (1995: 282–3 and references cited therein) also mentions languages that make use of a two-feet prosodic template. These include: Japanese, Ponapean, English and of course Bella Coola. The overwhelming majority of Bella Coola roots conform to the four-mora limit (over 94% out of the 1,169 monomorphemic roots listed in Nater (1977). Since the exceptions are either personal

Onsets and word minimality

143

Polymorphemic words frequently exceed this limit, e.g. xɬp’χwɬtɬpɬɬs ‘he had had in his possession a bunchberry plant’ (B98: 74) [xɬ- ‘to have, possess’; p’χwɬt ‘bunchberry’; -ɬp ‘tree, plant’; -ɬɬ pluperfect; -s possessive]. Some representative examples of possible roots featuring actual data are presented next along with hypothetical roots that exceed the four-mora generalization. The latter are consistently ill formed (a fuller list can be found in B98: 77–81).11 (8)

Heavy–light syllable sequence (B98: 79) a. b. c. d.

CVCCV CVCCCV CVCCCCV CVCCCCCV

k’ucxi ƛ’iq’ɬkn̩ * *

‘maggots’ ‘(low) dwarf blueberry’

3μ [μμ][μ] 4μ [μμ]μ[μ] 5μ [μμ]μμ[μ] 6μ [μμ]μμμ[μ]

Here, the first syllable is closed, followed by an open syllable. Between the coda of the first syllable and the onset of the second, only one unsyllabified consonant may intervene (8b). Additional ones would lead to an increase in moras beyond the total of four per root that the language accepts. Had the sequence been one of two closed syllables, then no unsyllabified consonants would be permitted, since the four-mora maximum would have been reached already as illustrated in (9). (9)

Heavy–heavy syllable sequence (B98: 79) a. CVCCVC b. CVCCCVC c. CVCCCCVC

qacqiɬ * *

‘ant’

4μ [μμ][μμ] 5μ [μμ]μ[μμ] 6μ [μμ]μμ[μμ]

A generalization that becomes evident is that the more syllables a root has, the less space for clustering – and thus potentially unsyllabified consonants – is available. This can be shown when considering trisyllabic roots. (10)

Trisyllabic roots (B98: 81) a. CVCVCV b. CVCVCVC c. CVCCVCV d. CVCCVCVC

quluci k’anawiɬ t’ixɬala *

‘skunk’ ‘bow of boat’ ‘robin’

3μ [μ][μ][μ] 4μ [μ][μ][μμ] 4μ [μμ][μ][μ] 5μ [μμ][μ][μμ]

Here, no unsyllabified consonants are permitted. In fact, no more than one heavy syllable can arise. Only one instance of a trisyllabic root with an unsyllabified

11

or geographic names, loanwords or possibly morphologically complex forms, the generalization put forward seems well grounded (see Bagemihl 1998, Appendix B for the exceptions). Following Bagemihl’s notation, moras within brackets indicate syllabified moras, while those without any brackets flanking them are the moras contributed by unsyllabified segments.

144

Onsets

consonant is imaginable and this presupposes that all syllables have to be light, so that there is space for one more moraic unsyllabified consonant, namely CCVCVCV, e.g. stapiɬm̩ ‘bat (animal)’ (B98: 81). Conversely, increasing clustering is possible in monosyllabic roots. These can maximally be CCCCV when open, e.g. pɬtkn ̩ ‘bark of bitter-cherry tree’ (B98: 79), i.e. with three unsyllabified moraic consonants and one mora from the vowel, and CCCVC when closed, e.g. q’χtis ‘fish weir made of rocks’ (B98: 79), or CVCCC, e.g. cipsx ‘fisher’ (B98: 80), with final clustering. Thus, extensive clustering is permitted either initially or finally, but not simultaneously at both edges. (11)

Clustering at the edges in monosyllabic roots (B98: 80)12 a. CCVCC b. CCCVCC c. CCVCCC

kwpaɬɬ * *

‘liver’

4μ μ[μμ]μ 5μ μμ[μμ]μ 5μ μ[μμ]μμ

4.2.2.1 Untenability of other counting criteria Before concluding that the maximality criterion is founded on mora count, we would need to dispense with any argument based on other types of counting, e.g. segment, consonant or vowel counting. For instance, to argue for segment counting, the four-mora maximum has to be interpreted in terms of segments, e.g. a root is maximally four segments long (or something longer than that, but in any case always consistently so). This is impossible to do. In fact, no type of counting, i.e. segment-, C- or V-counting, apart from the moraic one, can account for the facts in a uniform manner. This is illustrated below. (12)

Possible counting criteria C-counting V-counting13 Seg-counting μ-counting a. b. c. d.

12

13

c’klakt p’χwɬt pɬtkn̩ ƛ’iq’ɬkn̩

5 4 4 4

1 0 1 2

6 4 5 6

4 4 4 4

Bear in mind that while a form like CCCVCC (11b) is bad since it exceeds the mora maximum by virtue of the far too many unsyllabified obstruents, the near identical CCCVCC form is well formed, e.g. c’klakt (12a). The difference here is that a complex onset that fits the sonority profile has been created. Since the cluster is syllabified in the onset, it no longer carries a mora, thus the total number of moras involved now reaches the maximum of four (i.e. μ[μμ]μ). This fact is an additional argument for the correct interpretation of the root-maximality facts and their relationship to unsyllabified consonants and onsets. I count syllabic sonorants as Vs by virtue of their function as nuclei. Perhaps nucleus counting might seem more appropriate, but I am simplifying here and keep the V-counting in line with Bagemihl (1998) who considers sonorants as Vs when discussing word minimality.

Onsets and word minimality C-counting V-counting e. f. g.

k’anawiɬ ƛ’aqwakila miank

4 4 3

145

Seg-counting μ-counting

3 4 2

7 8 5

4 4 4

The well-formed roots above present variable numbers in terms of consonants, vowels or segments, but their moras are consistently four (the alleged maximum). However, given the table above, the following maxima obtain for the rest of the categories: five consonants, four vowels or eight segments. It would thus be interesting to see whether these maxima generally produce well-formed roots as well. If any of them does, then we could claim that it is this criterion and not the mora one which is responsible for the observed root maximum. Assuming, for instance, that the language uses a segment-counting maximum (i.e. eight segments), then we would expect roots with eight segments to be consistently attested. The same should occur with the consonant (five consonants) or vowel maximum (four vowels). (13)

Unattested roots based on C-, V- or Seg- maxima C-counting V-counting Seg-counting μ-counting a. CVCCCCV (B98: 75) b. CCCCC (B98: 75) c. CCCVCVCV (B98: 81) d. CCVCCVC (B98: 80)

5 5

2 0

7 5

5 5

5

3

8

5

5

2

7

5

(13) shows that many roots that conform to these requirements are nevertheless unattested. All the roots in (13) have five consonants – which in (12a) is acceptable – and yet are unattested. So is the case in (13c) that has three vowels, although in (12e) this does not cause any problems. Under the segment-counting approach, eight segments do not ensure the well-formedness of the root, as (13a–d) illustrate. In fact, (13b) has only five segments, but it still does not occur. Thus, even though these examples contain a number of units which other extant roots exhibit, they do not appear as well-formed roots. Only one thing unites them: they all exceed four moras, establishing that this is the root maximum. A final alternative that needs to be considered for completeness is an extraprosodic account of unsyllabified consonants. This too fails, as B98 shows, because it predicts that unsyllabified consonants should be able to arise freely at both word edges. The problem arises once we consider the following: if

146

Onsets

there are roots which contain up to three such consonants at the left edge, e.g. pɬtkn̩ ‘bark of bitter-cherry tree’ (B98: 79), and up to two at the right edge, e.g. ʔastxw ‘(to be) inside’ (B98: 84), then why should these not appear simultaneously, forming e.g. *CCC.CVC.CC? This seems coincidental unless one considers that structures like the above exceed the four-mora maximum. 4.3

Word minimality

With this much background, it is now possible to tackle the word-minimality facts and then proceed to an analysis of all well-formed roots in Bella Coola. The data are presented below.14 (14)

Minimal word size in Bella Coola Word Shape a. V b. VV c. VC d. C e. CV / CR f. CC g. CCC h. CCCC i. CCCCC

Examples * ya15 ‘good’ n̩ƛ’ ‘dark, night’ * ƛ’i ‘fast’ / tk’ ‘sticky’ sχp ‘to tie a knot’ p’χwɬt ‘bunchberries’ *

c’m̩ ‘index finger’ (B98: 79)

The absence of [V] (14a) words and the simultaneous admission of [VV] (14b) and [VC] (14c) words is totally in line with the idea that the word minimum in Bella Coola is bimoraic. Somewhat more strangely, [CV] (14e) words are also permitted, but this is only surprising in standard moraic theories (see Hyman 1985; Hayes 1989 and indeed Bagemihl’s 1998 own analysis) that deny any contribution of onsets to weight. In the present model, this is entirely anticipated. The remaining facts also fall out naturally. Given §4.2.1–2, we know that unsyllabified obstruents bear moras and that roots cannot be longer than four moras. Consequently, [C] (14d) and [CCCCC] (14i) are unacceptable, as they 14

15

Notice that since roots are the only morphemes which can occur independently as words without any affixes, a minimal constraint on roots is equivalent to a minimal-word constraint (Bagemihl 1998: 93). Bagemihl (1998: 89) presents this as an example of VV and later on (1998: 93) mentions that there are no examples of roots consisting of only a long vowel, which may be an accidental gap, but such absence may more likely be attributed to the extreme rarity of long vowels in morpheme-final position (Nater 1984: 17). Although a bimoraic structure for ya is possible (cf. Smith 2003: example (8b.ii)), one might still argue that such an example constitutes a CV sequence and not a VV one. For details of an analysis along these lines, see §4.4.3.

Onsets and word minimality

147

are subminimal and supraminimal, respectively. In contrast, words (14f–h) are attested, as they conform to both word minimality and root maximality. The situation above is schematized in (15) with regard to the most important word templates in question. Here and throughout, moras within brackets indicate moras that may or may not be present underlyingly, so as to conform to Richness of the Base (Prince and Smolensky 1993/2004), which dictates that inputs should not be constrained. I assume, however, that vocalic moras are present underlyingly, because the language has phonemic vowel length. (15)

Word Shape Respects WdMin?

Prosodic Structure Underlying

a.

C

No

b.

CC

Yes

After Syllabification

(µ)

µ

(µ) (µ)

µ µ

C

C C

C

σ c.

V

No

µ

µ V σ

d.

CV

Yes

(µ) µ

µ

µ

C V

C

V

My suggestion is that onsets are moraic only in an extremely restricted environment, namely that of CV words, so as to satisfy word minimality. We can tell that in larger roots, this is not the case due to the maximality restrictions discussed in §4.2.2. In particular, roots such as CCVCC sq̕waɬkw ‘ashes’ (B98: 76) or CVCCC muχwɬt ‘to cry huyp (a dance cry)’ (B98: 75) or CVCCCV nuχwski ‘soapberries’ (B98: 75), among others, are admitted, because all of them are maximally quadrimoraic. Had the onsets counted moraically, all these forms would include either five or six moras. Given the maximality restrictions, they should thus fail to surface. The fact that they surface clearly suggests that onsets are not moraic with the exception of CV words.16 The following sections present a detailed analysis of the above. 16

Alternatively, one may propose that onsets are consistently moraic and the only thing that needs to be done is to revise the mora maximum to five or six moras so that moraic onsets are also taken into account. However, this is not correct. To illustrate, suppose we modified the limit to six moras, so that roots like CVCCCV above are admitted [i.e. with all segments moraic]. The question would then be why roots such as CCCCCV or CVCCCC or CCVCCC (B98: 79–80) etc. are not well formed, although these comprise six moras too. Such an amendment would ruin the generalization formerly established, which, as we have seen, is well grounded. Evidently, then, onsets are moraic only in CV roots and the moraic limit should not be modified.

148

Onsets

4.3.1 The core analysis The core analysis deals with the environment where moraic onsets appear; in particular, I argue that moraic onsets only emerge so that word minimality (16), a top priority in the language, is satisfied at the expense of allowing moraic onsets (17). This renders CV words minimal. (16)

WdMin: Words are minimally bimoraic

(17)

*Moraic Onset: Moraic onsets are banned (cf. (16) in §1.3.3, §5.4.1)

Larger forms already satisfy word minimality, thus use of moraic onsets is superfluous. As said, given that the language also presents a root-maximality condition, we can actually establish that onsets do not contribute any moras in other cases, because otherwise we would wrongly predict the non-existence of certain roots such as CVC.CVC ones. Had their onsets contributed weight, then they should be unattested, as they would exceed the four-mora maximum. In due course, numerous technical details will be added to complete the analysis. While these details might be more relevant to the OT theorist, careful technical development is crucial for empirical reasons too. For instance, I will devote special attention to why the output of a /CV/ word appears moraified and syllabified as [[CμVμ]σ]Wd. This is because there is independent evidence from reduplication that CV forms get syllabified (cf. §4.2.1). It is therefore important that all other reasonably conceivable candidates such as: [C[Vμ]σ]Wd, the null parse, [CμVμ]Wd and [Cμ[Vμ]σ]Wd examined in (19), (25), (34) and (36), respectively are eliminated (§4.3.2.2). Similarly, we also know that words consisting of a single C or V do not occur. Consequently, in this case too, alternative possibilities must be dispensed with (see also §4.4). I will thus start by laying out the basic proposal and then move on to an elaborate analysis. The form we need to examine first is /CV/. Given that WdMin (16) is a requirement that is never violated in the language, it is reasonable to assume that it is undominated. The proposed solution admits moraic onsets, hence WdMin >> *Moraic Onset. Due to Richness of the Base (Prince and Smolensky 1993/2004), it is plausible to consider both moraically specified and unspecified inputs, i.e. /CμVμ/ and /CVμ/, respectively17 (or more simply /C(μ)V(μ)/), and ensure that mora retention or insertion in the onset is the strategy chosen to satisfy WdMin. This means that WdMin must also dominate Dep-μ. I will in particular employ the P-Dep-μ version of that constraint, as argued 17

Since vowels are moraically specified either as monomoraic (short) or bimoraic (long), the terms ‘moraicity’ and ‘non-moraicity’ refer to consonants.

Onsets and word minimality

149

in §3.2.2.2, for reasons that will become clear soon. We therefore obtain the ranking in (18). Note that the position of P-Dep-μ is not yet justified. This is postponed till §4.3.2.1, example (30). I nonetheless include it, as it will prove an essential component of the full minimality analysis. (18)

WdMin, P-Dep-μ >> *Moraic Onset

Let us see how /CV/ fares in terms of these constraints. Note that variable moras are shown in brackets. Constraint violations, however, are shared in both moraic and non-moraic inputs. (19)

/C(µ)Vµ/ --- [[CµVµ]σ]PrWd: WDMIN, P-DEP-µ >> *MORAIC ONSET WDMIN a.

P-DEP-µ

*MORAIC ONSET

Wd σ *!

µ C ☞

V Wd

b.

σ µ

µ

C

V

*

An output like (19a) violates WdMin. The candidate in (19b) has a moraic onset, but this violation is less important than the one incurred by the first candidate. P-Dep-μ is also present, but yields no effect in this instance. To see why, recall from §3.2.2.2 the definitions of positional mora licensing and P-Dep-μ. (20)

Positional μ-licensing: A segment α is positionally μ-licensed by a mora iff μ is the only prosodic unit directly dominating α

(21)

P-Dep-μ: A non-positional μ-licenser mora in the output has a correspondent in the input

P-Dep-μ only refers to non-positional μ-licensers, but outputs such as (19a) and (19b) only involve positional μ-licensers, since each segment is uniquely dominated by one mora. Consequently, P-Dep-μ is vacuously satisfied in both.

150

Onsets

In a form larger than CV, WdMin can be satisfied by other means, so the claim is that no moraic onsets will appear in such a case, because although low ranked, *Moraic Onset violations are taken into account. Dominance of *Moraic Onset over Max-μ – the constraint responsible for mora preservation – ensures that no moras will appear on onsets. This ranking argument is illustrated if we consider a moraically specified input like the one in (22). (22)

/CµVµCµ/ --- [[CVµCµ]σ]PrWd: WDMIN >> P-DEP-µ, *MORAIC ONSET >> MAX-µ WDMIN

P-DEP-µ

*MORAIC ONSET

MAX-µ

Wd

a.

σ µ

µ

C V

C

µ

☞

b.

*!

Wd σ µ C

*

µ

V C

Taking root maximality into consideration – encoded in the undominated constraint in (23) – we can also see why in some cases no full-fledged output wins. Since RtMax sets a limit to the number of moras, the null parse is selected (24). (23)

RtMax: No root may exceed four moras (B98: 77)

(24)

MParse: Morphemes are parsed into morphological constituents (Prince and Smolensky 1993/2004)

This is briefly illustrated here (the choice of MParse is justified in §4.5.2, Topintzi 2006b). (25)

/C(µ)C(µ)C(µ)V(µ)C(µ)C(µ)/ ---∅: RTMAX >> MPARSE RTMAX

MPARSE

Wd

a.

σ µ µ

µ µ µ

*!

C C C V C C ☞

b.

∅

*

Onsets and word minimality

151

Let us abstract away from the exact input mora specification and focus on the two candidates presented here. Anticipating the relevant discussion in (40), let us also assume that unsyllabified consonants only survive if they bear a mora. Then, as (25a) shows, a [CCCVCC] root fails to emerge. This is because it includes three unsyllabified consonants, all of which need to be moraically licensed in order to survive. But this means that the root maximum must be exceeded. Candidate (25a) therefore is ruled out and the null parse is considered optimal. Notice, however, that a root such as CCVCC (26) with just one fewer unsyllabified consonant is well formed, e.g. kwpaɬɬ ‘liver’ (B98: 80). What this means is that this root adheres to RtMax and implies that the onset of the CVC syllable has to be non-moraic. (26)

/C(µ)C(µ)V(µ)C(µ)C(µ)/ --- [Cµ[CVµCµ]σCµ]PrWd: RTMAX, *MORAIC ONSET >> MPARSE RTMAX

*MORAIC ONSET

*!

*

MPARSE

Wd

a.

σ µ µ µ µ µ C C V CC ☞

b.

∅

c.

Wd

*!

σ µ

µ µ µ

C CV C C

Note that no (direct) ranking argument can be formed between RtMax and *Moraic Onset, since the winning candidate violates both. Soon, however, we will be able to indirectly infer that RtMax >> *Moraic Onset, through the interaction of other constraints. At this stage, the inclusion of this tableau merely highlights the fact that, apart from reasons specific to our analysis suggesting that supraminimal words contain no moraic onsets (cf. (22)), there is also empirical evidence leading to this conclusion. Roots of the type in (26) are acceptable, so clearly the upper limit of moras per root entails that moraic onsets do not appear in maximal roots. Summarizing the core part of this analysis, we have provided the ranking arguments given in (28) and have proposed a mini-grammar for Bella Coola which is compatible with the one linearly represented in (27). (27)

RtMax, WdMin >> MParse >> P-Dep-μ, *Moraic Onset >> Max-μ

152

Onsets P-DEP-µ

WDMIN

(28) (19)

RTMAX (25)

*MORAIC ONSET (22)

MPARSE

MAX-␮

The exact ranking of these constraints as indicated in (27) will be justified as the analysis progresses. Additional constraints will also be employed, but these will just be needed to polish the analysis so that various details are covered. This job is done next. 4.3.2 Analysis in detail This section provides fuller analysis of the facts. I will begin by examining the word-minimality facts. 4.3.2.1 Word minimality Let us first consider the case of monosegmental words. These fail to emerge (cf. (15a) and (15c)). Since they are subminimal forms, they violate WdMin. However, unlike /CV/ forms which circumvent the same problem by assigning a mora to their onset (if one is not underlyingly present, that is), /C/ or /V/ forms are not rectified in any analogous way. An obvious remedy for a monoconsonantal form would be the insertion of a vowel, but this is not allowed, therefore the null parse is preferred. This entails that Dep-Seg – which penalizes segment insertion – is highly ranked and dominates MParse. (29)

/C(µ)/ --- [∅]: WDMIN, DEP-SEG >> MPARSE WDMIN a.

DEP-SEG

MPARSE

Wd σ *!

µ C Wd

b.

σ µ

µ

C ☞

c.

*!

V ∅

*

Things are slightly more complicated when one considers monovocalic forms. Here, apart from the insertion of a segment, e.g. a consonantal onset,

Onsets and word minimality

153

which would already be excluded due to Dep-Seg, another option is feasible, namely lengthening of the vowel so that it becomes bimoraic. This is where use of P-Dep-μ becomes vital. (30)

/Vμ/ and /Vμμ/ roots: WdMin, P-Dep-μ >> MParse (i)

Vμ

WdMin

a.

Vμ

*!

b.

Vμμ

☞ c.

Ø

(ii)

Vμμ

☞ d.

Vμμ

e.

P-Dep-μ

MParse

*!

Ø

*

*!

In §3.2.2.2 we saw that moras of underlyingly long vowels (30d) are nonpositional μ-licensers, since they both dominate a single segment. However, since their input–output configuration is identical, they satisfy P-Dep-μ. This is not the case for a lengthened vowel (30b), whose moras – like those of the long one – are non-positional μ-licensers, but in addition suffer from a P-Dep-μ violation, exactly because the added mora in the lengthened vowel has no input correspondent. This explains the contrast between (30b) and (30d). Monomoraic candidates of course violate WdMin (30a), so are easily excluded. It is thus the ranking between P-Dep-μ and MParse that generates the relevant results. If the former dominates the latter, then the null parse (30c) is naturally the optimal candidate for /Vμ/, but not for /Vμμ/. In that latter case, the faithful output candidate (30d) avoids the violation of P-Dep-μ, thus violation of low-ranked MParse becomes decisive in eliminating the null parse as the winning candidate. So far we have seen that a series of constraints dominates MParse, i.e. RtMax >> MParse (25), Dep-Seg >> MParse (29), P-Dep-μ >> MParse (30) and WdMin >> MParse (29). WdMin also dominates *Moraic Onset (19). What this suggests, then, is that the relative ranking of MParse and *Moraic Onset may be somehow related, but we cannot yet determine exactly how. Consideration of CC sequences provides the missing link. We already know that when both consonants are obstruents (TT), no nucleus can be formed. This can be accounted for by the ranking: (31)

*Nuc/Obstr >> Nuc >> *Nuc/Son >> *Nuc/Vowel

This says that only sonorants and vowels can serve as nuclei, but obstruents cannot. Given that Dep-Seg and *Nuc/Obstr are very highly ranked, a /CC/

154

Onsets

form is destined to either arise as the null output or unsyllabified. The empirical data suggest the latter, thus MParse must dominate the constraint which penalizes unsyllabified structures. I claim that this is the μ/σ as illustrated below. (32)

μ/σ: Every mora is dominated by a syllable (adapted from Lin 1997)

This belongs to the group of prosodic licensing constraints (see Lin 1997 for other constraints of this type) which express that every prosodic constituent must be dominated by another one higher in the prosodic hierarchy (cf. Strict Layering Hypothesis: Selkirk 1984b; Weak Layering Hypothesis: Itô and Mester 1992). (33)

/C(µ)C(µ)/ --- [CµCµ]PrWd: MPARSE >> µ/σ MPARSE ☞

a.

Wd µ

µ

C b.

µ/σ

*

C ∅

*!

The constraint μ/σ is violated by (33a), but it is (33b) that loses as it presents the more serious violation of MParse. The question now is whether μ/σ can have any unwelcome repercussions in the consideration of our very first example, i.e. /CV/, which is clearly both a syllable, as the reduplication facts show (§4.2.1), and has a moraic onset to satisfy WdMin. The answer is negative, provided μ/σ >> *Moraic Onset. (34)

/C(µ)Vµ/ --- [[CµVµ]σ]PrWd: µ/σ >> *MORAIC ONSET µ/σ a.

☞

*MORAIC ONSET

Wd µ

µ

C

V

b.

*!

Wd σ µ

µ

C

V

*

Candidates (34a) and (34b) are very similar, the only difference being that (34b) is syllabified. μ/σ is decisive in favouring (34b) over (34a) by virtue of

Onsets and word minimality

155

μ/σ >> *Moraic Onset. Thus, μ/σ is not a high priority of the language, but whenever it can be satisfied, it will be, as it is here.18 An additional reasonable candidate is (35), where the mora is not linked to an onset, but to an unsyllabified consonant. Apparently, Onset is in operation here, requiring that whenever syllabification can apply, the prevocalic consonant must syllabify as an onset instead of staying unsyllabified. A further ranking is established, as shown in (36). (35)

Wd

*

σ

(36)

µ

µ

C

V

Onset >> *Moraic Onset Onset ☞

a.

[[CμVμ]σ]Wd (34b)

b.

[Cμ[Vμ]σ]Wd (35)

*Moraic Onset *

*!

Consideration of the representations above establishes that a /CV/ sequence can only be syllabified as [[CμVμ]σ]Wd, i.e. with a moraic onset. The only remaining word-minimality case that we have not yet discussed is that of /VC/. Since Bella Coola prefers syllabified codas to unsyllabified consonants, it must be that the constraint which militates against unsyllabified segments, namely Parse-Seg, dominates No Coda as shown below. (37)

Parse-Seg >> No Coda

To sum up, Bella Coola monomoraic words are banned as there is no way to simultaneously satisfy both WdMin and P-Dep-μ. The optimal output is thus found in the guise of the null parse. Bimoraic or larger words are, however, well formed. In the case of bimoraic words, we can find examples that end up unsyllabified, i.e. CC, with an obstruent cluster, or those which are syllabified by virtue of a sonorant segment that can act as a nucleus, e.g. CV or VC. The latter form presents a straightforward example of a heavy syllable where 18

The constraint Onset will be used next. One might wonder whether this could be used here instead of μ/σ. I believe the answer is no. Onset will impose certain requirements on prenucleic consonants of an already syllabified nucleus (as in (35)). In (34a), there is no syllabification at all, so Onset would be vacuously satisfied.

156

Onsets

the coda contributes to weight. Things are more surprising in the case of CV, which, as we know, has to be both syllabified and bimoraic. Since one mora comes from the vowel, the other is bound to come from the onset. This is the only example where we can see moraic onsets in Bella Coola in action. Thus, the major constraints pertaining to word minimality and their respective ranking is given below. (38)

WdMin, P-Dep-μ >> MParse >> μ/σ >> *Moraic Onset

So the picture we have shaped so far looks like: (39)

*NUC /OBSTR

WDMIN

P-DEP-µ

DEP-SEG (30)

(29)&(30) NUC

(25)

(33) ONSET

(36)

µ/σ (34)

*NUC/V

(29)

MPARSE

(19)

(31) *NUC/SON

RTMAX

*MORAIC ONSET

PARSE-SEG (37) .

NO CODA

(22) MAX-µ

4.3.2.2 Remaining roots I will briefly examine how the analysis built so far accounts for roots which are neither minimal nor maximal, but somewhere in between, with special focus on words including several unsyllabified obstruents. First, recall that moraic onsets only emerge in [CV] words as a response to WdMin requirements. In larger roots, WdMin can be satisfied by other means, thus violations of *Moraic Onset prove gratuitous, as we have seen in (22) and (26). Second, numerous unsyllabified consonants emerge which are not stray erased (§4.2). This can be explained using the notion of moraic licensing after Bagemihl (1991, 1998) and Lin (1997) for Piro. In these works, it has been proposed that unsyllabified consonants do not delete because they can be licensed by moras. This is an assumption shared here as well. (40)

Moraic Licensing (MLic): An unsyllabified consonant is realized, i.e. licensed, when dominated by a mora

We know that the maximal syllable in Bella Coola is TRVVC, which implies that complex codas are banned. Evidently, *Complex Coda is undominated

Onsets and word minimality

157

in the language. Given the maximal syllable, any remaining surface obstruents must be unsyllabified, but must also be moraic so that they satisfy MLic. At the same time, another constraint, Max-Seg, requires that as many segments as possible surface. The obvious solution to reconcile Max-Seg, which requires the output realization of all segments, and MLic, which deletes non-moraic unsyllabified consonants, is to attach a mora to all consonants. In this way, both constraints are satisfied. However, MLic does not act unrestrictedly. It is controlled by RtMax, which discards numerous roots which exceed four moras. As a result, it excludes the candidate *CμCμCVμCμCμ for an input like /CCCVCC/. RtMax also regulates the distribution of other reasonable candidates. Effectively, in these forms either a mora gets deleted to obtain the four-mora maximum or one of the underlying segments fails to be realized. The main contenders thus are: (41a) with mora sharing, (41b) with segment deletion and (41c) with unsyllabified non-moraic consonants. Exclusion of all those candidates renders the null parse the winner. (41)

a.

b.

Wd σ µ

C

C

C

c.

Wd

Wd σ

σ

µ

µ

µ

µ

V

C

C

C

C

µ

µ

µ

V

C

C

C

µ

µ

µ

µ

C C

V

C

C

In (41a) the onset is non-moraic, while the nucleus V and the coda C receive a mora. There are three unsyllabified consonants, but because only two more moras are available, two of the segments share a mora. Such a representation violates *Mora Sharing. In (41b) a consonant has deleted, violating MaxSeg.19 Finally, in (41c) one of the unsyllabified consonants is left without a mora. This causes a MLic violation since every unsyllabified segment must be dominated by a mora. The constraints that are violated each time are the following: (42) 19

*Mora Sharing: Segments do not share moras (Sprouse 1996: 398, 406; Broselow, Chen and Huffman 1997: 65)

This possibility cannot really be ruled out. Perhaps the input for the word sq̉waɬkw ‘ashes’ (B98: 76) might be the hypothetical /tsq̉waɬkw/. Then the initial consonant could delete and we would obtain the output above. However, by Lexicon Optimization (Prince and Smolensky 1993/2004), the input /sq̉waɬkw/ will be preferred. Moreover, since RtMax is a root constraint, all the forms it evaluates are underived. As a result, no alternations – occurring in processes such as deletion in derived environments – emerge that could offer us empirical evidence one way or the other.

158

Onsets Max-Seg: Do not delete segments (McCarthy and Prince 1995) MLic: An unsyllabified consonant is realized, i.e. licensed, when dominated by a mora (Bagemihl 1991, 1998; Lin 1997; cf. (40))

All these can be summarized in the following tableau: (43)

/C(µ)C(µ)C(µ)V(µ)C(µ)C(µ)/ --- ∅ RTMAX, MLIC, *MORA SHARE >> MAX-SEG >> MPARSE RT MAX a.

MLIC

*µ SHARE

MAXSEG

M PARSE

Wd σ µ µ

µ µ µ

*!

C C CV C C (41c) b. [CCµ[CVµCµ]σCµ]PrWd (41a) c. [Cµ1Cµ1[CVµCµ]σCµ]PrWd (41b) d. [Cµ[CVµCµ]σCµ]PrWd ☞

e.

*! *!

∅

*! *

In sum, whenever the root exceeds the four-mora maximum imposed in the language, the form cannot survive, since it violates RtMax. Other possible forms are excluded too, because these also violate other high-ranking markedness constraints. Consequently, the null parse proves to be the optimal output. 4.3.2.3 Summary of analysis The preceding sections have offered an analysis of Bella Coola roots with respect to both word minimality and root maximality. A major conclusion has been that the language allows moraic onsets in a very restricted environment, namely that of [CV] words only. The core phenomenon of word minimality is described by: (44)

WdMin, P-Dep-μ >> MParse >> μ/σ >> *Moraic Onset

WdMin is a very highly ranked constraint requiring that all words satisfy it by being bimoraic. This is obviously the case for [VV] and [(C)VC] words. In [CV] words, this requirement leads to the emergence of moraic onsets. At the same time, no segments can be added in the output, while vowel lengthening is banned. This means that for words made of a single vowel or consonant, only the null parse can be generated. In words of the [CC] type the word minimum is respected despite the lack of syllabification. Outside word minimality and more generally in the language, the preferred syllabification is quite typical in that CVC syllables are constructed where the

Onsets and word minimality

159

nucleus is either a vowel or a sonorant. Complex onsets of rising sonority are also allowed, but complex codas are banned. Obstruent consonant-clusters cannot syllabify, but do not delete either, because Max-Seg is high ranked. The only way for these consonants to surface is by being moraically licensed (highranking MLic). Indeed, this is what happens, provided that the total number of moras within a root does not exceed four (due to high-ranking RtMax). If the maximum is exceeded, the null parse wins. The resulting ranking for Bella Coola is: (45)

Final ranking in Bella Coola RTMAX, *µSHARE, MLIC (43) *NUC /OBSTR WDMIN P-DEP-µ DEP-SEG MAX-SEG (30) (29) (43) (29)&(30) MARSE (19) NUC (33) (31) *NUC/SON

ONSET

*NUC/V

(36)

µ/σ

(34) *MORAIC ONSET

PARSE-SEG (37) .

NO CODA

(22) MAX-µ

4.4

Alternatives that dispense with onset moraicity

At this point, it is useful to consider some alternatives that avoid use of onset moraicity and see how they fare with regard to the full range of empirical facts. It will become evident that onset moraicity cannot be dispensed with. 4.4.1 Bagemihl (1998) The major explicit alternative to the current analysis is the one offered by Bagemihl (1998). For convenience, the minimality data are repeated below. (46)

Minimal word size in Bella Coola a. b. c. d.

Word Shape V VV VC C

Examples * ya ‘good’ n̩ƛ’ ‘dark, night’ *

160

Onsets e. f. g. h. i.

CV / CR CC CCC CCCC CCCCC

ƛ’i tk’ sχp p’χwɬt *

‘fast’ / ‘sticky’ ‘to tie a knot’ ‘bunchberries’

c’m̩ ‘index finger’ (B98: 79)

Like the present analysis, Bagemihl puts forward an account that bases the minimality criterion on bimoraicity, correctly excluding words like [V] or [C] and permitting [VV] or [VC]. His analysis is, however, weakened when considering the [CV] words; given that he rejects the idea that onsets may bear moras, he needs some way to explain how [CV] words can satisfy minimality. To resolve the problem, Bagemihl assumes that all segments are underlyingly moraic so that /V/ syllables are underlyingly monomoraic, while /CV/ ones are underlyingly bimoraic. Similarly, /C/ is monomoraic and /CC/ is bimoraic. Word minimality is also assumed to be an input condition, thus /CV/ and /CC/ satisfy it, while /V/ and /C/ do not. In the output, however, [CV] and [V] are monomoraic – since the singleton C syllabifies as an onset, it cannot retain its mora – and syllabified, while [C] and [CC] both remain unsyllabified (although the former is monomoraic and the latter bimoraic). Bagemihl’s proposal is untenable in an output-oriented model. First, word minimality cannot be treated as a condition in the input, simply because restrictions of this sort must spring from constraint interaction. Second, Bagemihl’s proposal crucially hinges on underlying mora specification. Given Richness of the Base (Prince and Smolensky 1993/2004), this is undesirable, because no such restrictions should be placed on inputs. Both problems are avoided in the present, entirely output-oriented, account, which treats WdMin as a constraint on the output. It is its interplay with other constraints that generates the attested data. Obligatory underlying moraicity is also dispensed with, offering an analysis that conforms to RotB. Finally, due to the inherent violability of the OT constraints, it is possible to state that moraic onsets will only arise in a restricted environment, i.e. only as a means to satisfy WdMin. When this is not at stake, *Moraic Onset, which is still active, albeit low ranked, ensures that no superfluous weightful onsets will arise. 4.4.2 An alternative OT account Given OT’s success in offering a neat account of the Bella Coola facts, it is at this point worthwhile examining whether an OT alternative that eliminates moraic onsets is feasible. To this end, I will briefly examine one such possibility as suggested by Elliott Moreton (p.c.). Two assumptions are crucial: (i) the ranking Max-Seg >> WdMin holds in Bella Coola and (ii) the traditional notion of bimoraicity is adopted. Then all of CVC, VC and VV words would be well

Onsets and word minimality

161

formed by virtue of bimoraicity, while CV ones would be accepted due to the overriding importance of Max-Seg. However, C and V words, despite satisfying Max-Seg, would fail minimality, thus they would be unable to surface. The problem with this approach is that there is no reason why /CV/ would not lengthen to [CVV] or [CVC] so that it would simultaneously satisfy bimoraicity. One possibility is that Dep-Seg is also top ranked so that /CV/ cannot become [CVC]. This would not stop it from becoming [CVV], however, so again no good explanation obtains. An apparent remedy would be the use of high-ranked P-Dep-μ, so that lengthening is banned. Is, then, the ranking in (47) a real problem for my analysis? Consider the following tableau that illustrate the problem: (47)

P-Dep-μ, Max-Seg, Dep-Seg >> WdMin >> MParse P-Dep-μ

Max-Seg, Dep-Seg

WdMin

MParse

/Vμ/ a. Vμμ

*!

b. Vμ ☞ c.

*!

O

*

/Vμμ/ ☞ d. Vμμ e.

O

*!

/C/ f.

*!

CVμ

g. C

* *!

☞ h. O

*

/CV/ i.

CVμμ

j.

CVμ


k. O

*! *! *

This grammar yields the right results for /V/ and /C/ roots where the null parse is generated (47c, h), while it chooses a long-vowelled output for a /Vμμ/ input. The problem arises in a /CV/ input where the optimal candidate is wrongly the null parse (47k). The problem could be resolved if the ranking between WdMin and MParse was reversed. MParse >> WdMin would now select (47j) as the winner, but it would also entail that [V] and [C] outputs, (47b) and (47g) respectively,

162

Onsets

should survive too. This is clearly wrong as the empirical facts show. Thus, an alternative which makes no reference to onset moraicity is not viable. 4.4.3 A segmental account Early in the discussion (see fn. 15), it was mentioned that final long vowels are rare, but Bagemihl considers structures like ya as bimoraic VV ones. Although this is certainly a possibility (see Smith 2003: example (8b.ii)), one could nonetheless challenge this idea and argue that it instead constitutes a CV sequence. In this case and under a moraic approach, the lack of VV words is unexpected, but not so in an approach that bases minimality on a segmental criterion. There, VV can be treated as a single segment – if one segment means one root node – alongside words consisting of only V or C. With this in mind, it could be argued that the minimal word in Bella Coola merely has to consist of two segments. Since all of V, C and VV include just one segment, it would be correctly predicted that these fail the minimality criterion, and as such never surface as possible roots. In light of the above, could we then successfully pursue the claim that the language simply requires minimally bisegmental words? The answer will be negative. First, we have seen that apart from word minimality, the language also possesses a root-maximality restriction stating that roots consist of maximally four moras (§4.2.2). If a segment-based analysis is put forward for the minimality case, one would prefer to maintain a uniform analysis for the maximality effects too. But, as examined in §4.2.2.1, only a mora-based account can explain the root-maximality facts, thus the only option is to propose a uniformly moraic approach. Alternatively, one could utilize different approaches for the WdMin and the RtMax facts, but this would add an unnecessary extra component to the analysis. Another observation is that – to my knowledge – besides the case of Yakima Sahaptin (Curtis 2003: 193–207; Hargus and Beavert 2006), no principled account in favour of a segmental analysis for word minimality is available. This of course is not intended as a crucial argument against the segmental analysis, given that it rests on the fallacious argument that rare phenomena can be ignored in analyses precisely because they are rare. After all, this very book deals with a rare phenomenon! It should thus be merely taken as an observation, whose strength would be corroborated if a re-analysis of Yakima is in order and if no other languages seem suggestive of segmental word minimality. In Yakima Sahaptin, words seem to be of the type [CCV] or [CVC]. [CV] words do not occur, but surprisingly CVV syllables, although admitted in the language, do not satisfy the minimality criterion either. Curtis suggests

Onsets and word minimality

163

that minimality is not a prosodic phenomenon, but instead ‘is related to lexical recoverability and the overwhelming frequency of consonant clusters in Sahaptin’ (2003: 206). A possible analysis that ends up positing a disyllabic minimum for Yakima words is provided in Topintzi (2006b, §4.2.2.2, fn. 18). However, Hargus and Beavert (2006) claim that such a solution is not possible and that we should instead pose a biconsonantal minimum. I will not tackle this issue here, but simply make the reader aware of this view too. At any rate, a biconsonantal word minimum is not useful in the case of Bella Coola, which merely presents itself as an additional language that observes the bimoraic word minimum, provided onsets are allowed to be moraic on the surface. 4.5

Conclusion

The present chapter has focused on a clearly weight-based phenomenon, that of word minimality, and has shown that languages like Bella Coola calculate the onset for minimality purposes. In particular, it has been argued that onsets are moraic, albeit in a very marginal environment, namely that of a [CV] word. In this way, the overarching requirement of bimoraic word minimality can be satisfied. Elsewhere onsets are non-moraic, as the root-maximality facts verify (see §4.2.2). Bella Coola thus presents a case of coerced onset weight, where onset moraicity is enforced by a higher imperative in the language, that of word minimality. An elaborate analysis followed, aiming to provide a unifying account for the word-minimality and root-maximality data. Bearing in mind that word minimality is undoubtedly a weight-based phenomenon, use of moraic onsets in Bella Coola offers a particularly strong argument in favour of onset weight. Before concluding, it should be mentioned that the rankings presented in the preceding analysis are entirely consistent with the empirical facts, but some technical details might be somewhat different from what has been presented here. For example, extra-large roots could shorten by deleting segments and moras to fit the RtMax pattern rather than yield the null parse. By looking at the empirical facts we simply cannot tell what exactly happens, but even if it were the case that these alternative strategies were optionally or wholly adopted, then we would simply require some modification of the rankings suggested. The most important point to stress here is that irrespective of the specific details of the grammar, nothing changes with respect to the – up to now – controversial [CV] facts. No matter what the source of [CV] outputs is, e.g. /V/, /C/ or /CV/, reference to moraic onsets must be made.

5 Onsets and geminates

5.1

Introduction

The preceding chapters dealt with the behaviour of moraic onsets with regard to stress (Ch. 2), compensatory lengthening (Ch. 3) and word minimality (Ch. 4). All these phenomena provide evidence to a varying degree for the existence of coerced moraic onsets, i.e. onset weight enforced on the surface. The foremost of the three is of course stress. Recall that stress systems sensitive to the quality of the onset exhibit certain and predictable patterns according to which onsets may be moraic, as discussed extensively in §1.3.3, §2.2, §2.4–5. The word-minimality facts relating to onsets in Bella Coola seem to work as predicted by the coerced onset weight conditions (§1.3.3.5 and Ch. 4), although the data currently available are too scarce to state this with certainty. Finally, the analysis of compensatory lengthening was modelled in such a way so that it is framework neutral. More concretely, it is compatible with both the underlying presence and the absence of moras on constituents that eventually get to be syllabified as onsets on the surface, in other words, it does not care about input onset moraicity. Consequently, CL does not provide much evidence for either coerced or underlying onset weight (§3.1). This chapter clearly shifts its interest to consonants that are underlyingly moraic, i.e. /Cμ/, and that on the surface constitute geminates (for the difference between geminate and geminated consonants, see the discussion in §1.3.2 and fn. 11 of this chapter for completeness). Their moraicity is evident in the same phenomena that coerced weight is applicable, such as stress (§5.2.1; §5.3.1) or word minimality (§5.2.2). It is also present in CL (§5.2; §5.3.2), but, as said above, CL is not well equipped to identify this distinction. In most proposals within moraic theory (see Hayes 1989; Davis 1994; Morén 2001, among many others), geminates are considered to straddle the syllable boundary (the so-called ‘flopped’ structure), having one link to the coda of one syllable and another to the onset of the next. On most occasions, the coda is also moraic, thus rendering its syllable heavy. 164

Onsets and geminates (1)

165

‘Flopped’ structure of geminates word medially σ σ µ Cː

The reasoning behind splitting the geminate into two halves is two-fold. First, it reflects the geminate’s phonetically longer duration. For instance, Ladefoged and Maddieson (1996: 91–2) report that, depending on the language, geminate stops may be up to three times longer than their singleton counterparts. Second, it satisfies syllabification needs (Ham 2001), namely that besides a moraic coda, an onset is needed for the following syllable, a requirement served by the second link of the geminate. While this is one way to view things, it is not the only one possible and, as this chapter explains, is not even an adequate one for at least some cases. I will thus argue that certain empirical data discussed in the rest of this chapter call for a representation of geminates which treats them as moraic onsets (2) rather than coda-onset configurations. (2)

Geminates as moraic onsets σ µ

µ

Cː

V

Unlike (1), this representation contains no double linking. Instead, the geminate syllabifies wholly as a moraic onset within a single syllable. Consequently, it renders heavy the syllable that hosts it rather than the one before. This is particularly useful for geminates in word-initial positions, whose representation according to (1) is highly problematic, simply because initially, there is no coda for the first half of the geminate to link to.1 Undoubtedly, this is a significant advantage for the representation in (2), which also manages to remain compatible with syllabification needs. More challenging perhaps are the phonetic facts, since the absence of double linking seemingly misses the point that geminates are longer than singletons. However, as will be explained in more detail later (§5.4.2.2), it is the enforcement of double linking that generates concerns and may lead to a number of incorrect predictions rather than the 1

The same goes for word-final position, only this time there is no available onset for the geminate. Word-final geminates are not discussed here, but for a possible representation see Ham (2001).

166

Onsets

other way round. This is because the duration of the geminate across languages is not stable and thus it is incorrect to suggest that it is systematically double the duration of the singleton, as implied by the double linking in (1). Instead, all that is required in the phonology is that a geminate has a surface mora that started off in the input. It is the task of phonetics – and not phonology – to regulate how the geminate’s mora will be interpreted durationally on a languagespecific basis. The structure of this chapter is the following: Section 5.2 focuses on two case-studies, Pattani Malay (§5.2.1) and Trukese (§5.2.2), proposing that the most adequate analysis of their initial geminates is by means of moraic onsets. Such a conclusion leads to the theme of the next section, i.e. the possibility that moraic-onset geminates also emerge word medially. Section 5.3 centres around this topic using data from various languages and claims that a uniform analysis for onset geminates in initial and medial positions is in order. Section 5.4 deals with a number of pertinent issues. These include the preservation of contrast between moraic geminates and moraic singletons (§5.4.2.3), the relationship between phonetic length and phonological weight in geminates (§5.4.2.2), as well as discussion on various alternative phonological representations of onset geminates that fail compared to the present one (§5.4.3). Before moving on, let me clarify one notational point. Recall that the term geminate applies to consonants that are underlyingly moraic, i.e. /Cμ/, as opposed to singletons that lack input moras, i.e. /C/, but which can nonetheless appear moraic on the surface. This is what happens in moraic singletons. Thus, in the output, both are moraic [Cμ]. To differentiate between the two, surface geminates are denoted either as Cː or as CiCi and surface singletons as C (their mora indicated too whenever appropriate). Such double notation is justified for both practical and theoretical reasons. Practically speaking, comparison between the present text and the original data sources is facilitated. Theoretically speaking, differences in syllabification among geminates are more easily highlighted. For example, medial coda-onset geminates are best represented by means of CiCi, while tautosyllabic ones – like the ones proposed here – are easily indicated through the Cː notation. Note that length on the skeletal level merely serves to refer to geminates instead of simple moraic singleton consonants. 5.2

Word-initial moraic-onset geminates

As mentioned previously, in word-initial position, the representation of geminates is highly problematic, given that there is no material for the first part of the geminate to link to. But this is only problematic if one wishes to adhere strictly to standard moraic theory as proposed by Hayes (1989), who stipulates

Onsets and geminates

167

the prosodic inertness of the onset. If we dispense with this idea, as the present model does, then geminates in onset position come for free. This section presents case-studies of two languages of this type, namely Pattani Malay and Trukese, and shows how the moraic-onset analysis captures their facts more fully. The analysis built here has been inspired by previous work making similar claims as suggested by Hajek and Goedemans (2003; henceforth H&G) for Pattani Malay and Churchyard (1991) for Trukese. These approaches are, however, quite different from the current approach, as they do not extend to word-medial position nor do they establish links with other phenomena beyond geminates. Pattani Malay 5.2.1 Pattani Malay is a dialect of Malay spoken widely among the Muslim communities of the southern provinces of Thailand. The data used here are mainly due to Yupho (1989) and less so to Abramson (1991, 1998, 1999, 2003). Pattani’s phonemic inventory contains: the stops /p t c k ʔ b d J- g/, fricatives /s h z ɣ/, nasals /m n ɲ ŋ/, post-stopped nasals /mb nd ɲJ- ŋg/, liquids /r l/, the glides /w y/,2 and twelve vowels /i e ε ɨ a u o ɔ ε̰ a̰ µ ɔ̰/. With the exception of the schwa-like /ɨ/, which is relatively short even in open syllables, vowels are long in open syllables and short in closed ones. Thus, vowel length is non-phonemic. This is not the case for consonant length, which is contrastive, but only word initially, distinguishing between short consonants and long geminates, as in (3). (3)

Initial geminates vs. singletons in Pattani Malay (Abramson 1999, 2003) Singletons pagi tido makε labɔ ɣatɔ sεpaʔ cabε butɔ

Geminates ‘morning’ ‘to sleep’ ‘to eat’ ‘to profit’ ‘comprehensive’ ‘to kick’ ‘branch’ ‘blind’

pːagi tːido mːakε lːabɔ ɣːatɔ sːεpaʔ cːabε bːutɔ

‘early morning’ ‘put to sleep’ ‘to be eaten’ ‘cause to be late’ ‘to spread out’ ‘to be kicked’ ‘side road’ ‘kind of tree’

According to Abramson (1987), the primary phonetic correlate for initial geminates in Pattani Malay is the increased duration of the closure or constriction, 2

This is H&G’s description. Yupho excludes post-stopped nasals from the phonemes, as she considers them clusters of nasals and stops. However, the fact that – as she admits – the stop portion is almost inaudible and that such consonants syllabify in the onset position merely suggests secondary articulation of a single consonant. Note also that /ɣ/ is a rhotic-like velar (H&G: 83), so it is not clear whether it should be considered a fricative. For H&G (2003: 83), /ɣ/ does not geminate, but [ɣːatɔ] is a clear counterexample.

168

Onsets

which is about three times longer than that of their singleton counterparts (Abramson 1986). However, this is not the only cue, since, for instance, despite their silent occlusions, speakers detect the geminate–singleton distinction even in voiceless stops. Abramson found that greater amplitude (Abramson 1991) and higher fundamental frequency (Abramson 1999) contribute to the perception of geminates. In particular, amplitude differences are moderately significant for voiced stops and continuants (Abramson 2004), whereas the effect of fundamental frequency is also present in voiced stops, albeit smaller compared to that found in their voiceless counterparts. Hardly any effect is found in fricatives, whereas none at all is found in nasals. Geminates are primarily found as the product of initial syllable or morpheme reduction and occasionally in loanwords, e.g. tːa from Thai /taːn/ ‘police station’ (Yupho 1989). In the former case, the initial syllable of a word deletes, generating a free variant whose – original – second onset now geminates as in e.g. buwi ~ wːi ‘give’, sɨdadu ~ dːadu ‘police’, pɨmatɔ ~ mːatɔ ‘jewellery’ (Yupho 1989: 130). One way to view this would be that the loss of the initial syllable is followed by compensatory lengthening. Given that no reference to any sort of segment or slot preservation can be made – as the mora clearly moves from a vowel onto an onset – we are forced to accept that the onset has acquired a mora, thus becoming a moraic-onset geminate, as in [sɨμdaμduμ] ~ [dːμaμduμ] (Michael Kenstowicz, p.c.). The variant mentioned also emerges in some morphological environments: i) in words with the structure: derivational/verbal prefix /Cɨ-/ + stem; ii) in reduplicated forms where instead of the reduplicated form, one finds a geminate. (‘Yupho’s limited discussion on reduplication provides no further details about the process; it nonetheless seems compatible with the proposal made in §6.2.3.) iii) in cases where a functional word deletes. All in all, geminate formation interacts with morphology to a point, but the facts are too confusing to allow any sure conclusion to be made. On the other hand, there are cases, like loanwords and first-syllable deletion, where no morphology gets involved. I will thus assume that gemination is a phonological phenomenon which may be, but does not necessarily have to be, morphologically conditioned. Some representative examples of this sort using prefixed forms and their variants are given in (4). (4)

Geminates in prefixed forms (Yupho 1989) Unprefixed form

Prefixed form

Geminate variant

ɟalε buwɔh diɣi kaɟi

bɨɟalε bɨbuwɔh bɨdiɣi mɨŋaɟi

ɟːalε bːuwɔh dːiɣi ŋːaɟi

‘road, path’ ‘fruit’ ‘self’ ‘no gloss’

~ ~ ~ ~

‘to walk’ ‘to bear fruit’ ‘to stand’ ‘to study’

Onsets and geminates

169

Compensatory lengthening is the first phenomenon in Pattani which supports the phonological significance of geminates. The second is stress assignment, in which geminates play a predominant role. According to Yupho (1989: 134–5), primary stress is normally word final (5i), while all preceding syllables bear secondary stress (5i.a), unless they include the vowel /ɨ/, in which case they are stressless (5i.b). However, in words that begin with geminates, primary stress shifts and is placed on the initial syllable, with all remaining ones receiving secondary stress (5ii). Unlike H&G (2003), but like Yupho (1989), I do not mark vowel length for reasons that will be explained next. (5)

Stress in Pattani Malay i) Words lacking geminates a. ɟàlέ ‘road, path’ dàlέ ‘in, deep’ mà̰k¶nέ̰ ‘food’ b. sɨpɨnɔ́ ‘perfect, complete’ bɨlàkέ ‘back’ ii) Words with initial geminates mːátɔ̀ ‘jewellery’ ɟːál¶ ‘to walk’

(Yupho 1989: 133) (Yupho 1989: 134) (Yupho 1989: 135) (Yupho 1989: 128; stress here constructed given algorithm) (Yupho 1989: 134) (Yupho 1989: 135) (Yupho 1989: 133)

Interestingly, a number of minimal pairs are created, such as [bùwɔ́h] ‘fruit’, [ɟàl¶] ‘road’ vs. [bːúwɔ̀h] ‘to bear fruit’ and [ɟːál¶] ‘to walk’ (see Yupho 1989: 135). The versions with initial singletons present final stress, while those with initial geminates bear initial stress, even if the geminate is followed by /ɨ/ (6), which, as we have seen in (5i.b), otherwise repels stress. (6)

Geminates and /ɨ/

kɨdá ‘shop’ vs. kː-í dà ‘to the shop’ [from /kɨ+kɨda/] (Yupho 1989: 133) 5.2.1.1 The analysis Using the data from compensatory lengthening and stress above as evidence for weight-sensitive phenomena in Pattani, the current account claims that the language distinguishes between heavy and light syllables. Heavy syllables include a geminate moraic-onset consonant. Light syllables are all the remaining syllable types. This claim explains the compensatory lengthening facts, but also offers a full account of the stress facts. In particular, I argue that Pattani stress is partly weight and partly quality based. While default stress is final, it can shift to the first position if the syllable there is heavy, i.e. it contains a geminate, because

170

Onsets

of the WSP (Weight-to-Stress Principle), which is why we find initial stress whenever an initial geminate is present. As for the quality component of the stress system, one needs to compare the behaviour of the vowel /ɨ/ as opposed to all the other vowels of the language. Only /ɨ/ cannot receive stress at all (unless preceded by a geminate), whereas the other vowels can have primary or secondary stress. This asymmetrical behaviour will be attributed to the low sonority of /ɨ/ and to the well-known fact that, in certain languages, low-sonority vowels tend to or systematically remain unstressed (Kenstowicz 1994b, Crosswhite 1999, de Lacy 2007). Pattani Malay is an additional example of such a case. The full analysis follows. 5.2.1.1.1 The full-fledged account As explained, geminate-initial syllables in Pattani are bimoraic because of the mora the onset geminate (7) contributes. It is also common for heavy syllables (cf. several cases in Chapter 2, for instance) to attract stress because of the WSP. (7)

Initial geminates as moraic onsets σ µ

µ

# Cː

V

In the light of the lack of phonemic vowel-length contrast in Pattani, vowel length can be treated as completely phonetic. Consequently, there is only one type of heavy syllable in Pattani, i.e. those beginning with a geminate. All other types of syllable are monomoraic. This state of affairs is summarized in (8). (8)

Pattani Malay syllables and their moras in current approach i. CV(V)=1μ ii. CVC=1μ iii. CːV=2μ

It is worth mentioning that the Pattani schwa-like /ɨ/ is treated on a par with other vowels, although it never gets lengthened. This is unsurprising; the same happens in Yakima Sahaptin (Hargus 2001), while Yiddish (Albright 2002: 7) and Yupik (Bakovic 1996: 13) exhibit lack of long /ə/. Such resistance to lengthening may relate to the inherent weakness of /ɨ/ due to the lack of featural content (van Oostendorp 2000) or to the low sonority that is typical of

Onsets and geminates

171

central vowels which occasionally renders them incapable of bearing stress (Kenstowicz 1994b; de Lacy 2007). Either approach is workable in Pattani, although de Lacy’s generally seems more successful (see, for instance, his discussion on Nganasan). The claim, then, is that while /ɨ/ is quantitatively similar to the other vowels, it is qualitatively different from them, which explains why in (9) the primary stress data can be accounted for only if both the following factors are considered: i) whether the word begins with a geminate or not ((9a) vs. (9b)) and ii) whether the first syllable contains a central vowel or not ((9i) vs. (9ii)). It is also evident that quantity takes priority over quality, but this is an issue we will return to in due course. (9)

Summary of stress patterns in Pattani Malay a. i) Non-geminate, non-/ɨ/ words [bùwɔ́h] ii) Non-geminate, /ɨ/ words [pɨmàtɔ́] b. i) Geminate, non-/ɨ/ words [bːúwɔ̀h] ii) Geminate, /ɨ/ words [kːɨdà]

(σ̀)μ (σ́)μ σμ (σ̀)μ (σ́)μ (σ́)μμ (σ̀)μ (σ́)μμ (σ̀)μ

Let us first focus on the quantity issue alone, that is, a comparison between (9a.i) and (9b.i). In the former case, stress is final. Given that all syllables are monomoraic, we can reason that this is the default position of stress in Pattani accounted for by the alignment constraint Align-R (PrWd, HdFt). (10)

Align-R (PrWd, HdFt): The right edge of the prosodic word aligns with the right edge of the head foot (McCarthy and Prince 1993)

The stress shift on the first syllable in (9b.i) will be accounted for by weight considerations. More concretely, due to the initial geminate, the first syllable is now heavy by virtue of one mora for the vowel and one for the initial onset geminate. If heavy syllables attract stress, then WSP is in action and should dominate Align-R (PrWd, HdFt), in order to dictate word-initial primary stress. However, WSP only requires stress on heavy syllables. It does not require that this stress is primary. What we need, then, is a version of this constraint specific to primary stress, namely WSPPrWd (see McGarrity 2003). (11)

WSPPrWd: Heavy syllables receive primary stress

WSP can still be present in the ranking, but since it is WSPPrWd that regulates the outcome and not WSP, I will exclude the latter from tableaux examination.

172

Onsets

With this in mind, the only candidates we need to take into account are: i) (12a) with primary stress on the first syllable, and ii) (12b) with secondary stress on the first syllable. Due to the ranking WSPPrWd >> Align-R (PrWd, HdFt), the winner will be the former candidate, despite misaligning primary stress from the right edge. (12)

WSPPrWd >> Align-R (PrWd, HdFt) bːuwɔh ☞

a.

(bːú)(wɔ̀h)

b.

(bːù)(wɔ h ́ )

WSPPrWd

Align-R (PrWd, HdFt) *

*!

While this discussion accounts for (9b.i), it has nothing to say about why a candidate such as [(bːú)woh] is excluded. Or, more generally, it does not provide an explanation as to why secondary stress emerges, despite stress clashes or formation of non-binary feet. Answering this issue will offer a more complete analysis of (9b.i), but also of (9a.i). In effect, what the data suggest is the construction of monosyllabic feet. To produce these, we can make use of the constraints in (13) (Prince 1997, citing Bruce Tesar; Green 2002) and demand that they are simultaneously satisfied. As becomes evident in (15), this is only possible in the event of a monosyllabic foot. (13)

Align-L (Ft, FtHd): Align the L edge of every foot with the L edge of a foot-head Align-R (Ft, FtHd): Align the R edge of every foot with the R edge of a foot-head [NB: Following Prince (1997), I will abbreviate these as Ft-Hd-L and Ft-Hd-R, respectively]

If these foot alignment constraints are sufficiently high ranked, then we can produce languages that only allow monosyllabic feet (Prince 1997), such as Cantonese (Moira Yip, p.c.). The strictly monosyllabic feet of Pattani Malay call for a similar account. The ranking in (14) expresses exactly this fact: all Pattani syllables have to be parsed (PARSE-σ) into monosyllabic feet, even at the expense of creating several stress clashes (*Clash) or non-binary feet (FtBin). Tableau (15) exemplifies this result. (14)

Monosyllabic feet in Pattani Malay: Ft-Hd-L, Ft-Hd-R, PARSE-σ >> FtBin, *Clash

Onsets and geminates (15)

173

[bùwɔ́h]: Ft-Hd-L, Ft-Hd-R, PARSE-σ, Align-R (PrWd, HdFt) >> FtBin, *Clash buwɔh

FtHd-L

a. (buwɔ h ́ )

*!

b. (búwɔh) c. bu(wɔ h ́ ) d. (bú)(wɔ̀h) ☞ e. (bù)(wɔ h ́ )

Ft- Parse-σ Align-R FtBin *Clash Hd-R (PrWd, HdFt)

*! *!

* *!

**

*

**

*

To keep tableaux as manageable as possible, from now on I will use the cover constraint FtForm to represent the constraints Ft-Hd-L, Ft-Hd-R, PARSE-σ, each of which rules out one of (15a–c). (15d) is perhaps the most interesting rival to the winner (15e). The two are identical, the difference being in the order in which they place primary and secondary stress. However, recall that default stress in Pattani is final (unless WSPPrWd is involved), thus Align-R (PrWd, HdFt) (10) provides the relevant solution. Notice also that its position in the ranking, as presented here, cannot be established through a ranking argument. I just assume it for illustration purposes. The rankings in (12) and (15) have dealt with the quantity aspect of Pattani stress. However, inclusion of the central vowel in a word complicates things. In particular, why is it [pɨmàtɔ́] and not *[pɨmàtɔ́] (9a.ii), which we would expect since all syllables, save the primary-stressed one, receive secondary stress? Perhaps [ɨ] is not a good stress bearer as mentioned before. But, then, how can it get stress in something like [kːí-dà] (9b.ii)? The conundrum is resolved as soon as we consider some additional facts about central vowels and if we allow quantitative factors to rule over qualitative ones in Pattani. As said, central vowels like /ɨ/ are cross-linguistically weak and low in terms of sonority, often rendering them inappropriate stress bearers. Kenstowicz (1994b) in particular utilizes data from Kobon, Chukchee, Aljutor and Mari to make this point and to devise a hierarchy which reflects the sensitivity of stress to the relative sonority of the vowels involved. This hierarchy explicitly states that there will be languages which avoid stressing a schwa-like vowel. Importantly, this Peak Hierarchy is unrelated to that of Prince and Smolensky (1993/2004) which refers to the segments that make the best/worst syllable nuclei. Also note that *P/ә refers to central schwa-like vowels such as /ә/ and /ɨ/.

174 (16)

Onsets Kenstowicz’s (1994b) Peak Hierarchy:

*P/ә >> *P/i, u >> *P/e, o >> *P/a Additional evidence in support of this hierarchy is found in Chuvash and Javanese (Gordon 2002), Yil (de Lacy 2007) and Yakima Sahaptin (Hargus 2001), which most strikingly resembles Pattani’s /ɨ/ in most of its distributional properties. All we need, then, is to use a constraint that highlights /ɨ/’s marked nature and the fact that it cannot get stressed. This is effectively Kenstowicz’s *P/ә, which for our purposes will be simplified to *P/ɨ. Ranking *P/ɨ >> FtForm ensures that the syllable hosting /ɨ/ will be left unstressed. This is a welcome result as the example [pɨmàtɔ́] (9a.ii and below) illustrates. (17)

[pɨmàtɔ́]: *P/ɨ >> FtForm, Align-R (PrWd, HdFt) pɨmatɔ

*P/ɨ

FtForm

a.

(pɨ )(mà)(tɔ )́

b.

pɨ(matɔ )́

* (Ft-Hd-L), *! (PARSE-σ)

c.

pɨ(má)(tɔ)̀

* (PARSE-σ)

☞ d.

(pɨmà)(tɔ )́

* (Ft-Hd-L)

☞ e.

pɨ(mà)(tɔ )́

* (PARSE-σ)

Align-R (PrWd, HdFt)

*!

*!

Here, we get primary stress on the final syllable and secondary on the syllable before. The first syllable remains unstressed. This can be done by either parsing it in the foot-tail position (17d) or by leaving it unparsed (17e). Either of the two will do, as they are the same empirically. Things change dramatically if the central vowel is preceded by an onset ́ (9b.ii). Quality-based considerations ask that the initial geminate, as in [kːı-dà] syllable is stressless; from the point of view of quantity, though, this syllable must get stressed, because it is bimoraic. As the empirical facts show, quantity takes priority over quality, thus forming the ranking argument WSPPrWd >> *P/ɨ as illustrated below. (18)

WSPPrWd >> *P/ɨ kːɨda ☞

a.

́ (kːı-)(dà)

b.

kːɨ(dá)

WSPPrWd

*P/ɨ *

*!

With this modification, the full ranking for Pattani Malay stress is given below.

Onsets and geminates (19)

175

Pattani Malay stress (final version) WSPPrWd >> *P/ɨ >> Ft-Hd-L/R, PARSE-σ, Align-R (PrWd, HdFt) >> FtBin, *Clash

Summing up, we have seen that in words composed of monomoraic syllables only, primary stress is final (Align-R (PrWd, HdFt)) and all remaining syllables get footed receiving secondary stress (PARSE-σ). The resultant feet are monosyllabic because Ft-Hd-L/R >> FtBin. The only exception to this is with a syllable that comprises the schwa-like /ɨ/, which has to stay unstressed (highranked *P/ɨ). Nonetheless, this too is subject to the top-ranked WSPPrWd which places primary stress on heavy syllables. In Pattani, heavy syllables are only those starting with an initial geminate, which is syllabified as a moraic onset. Consequently, geminate-initial syllables always receive primary stress regardless of the subsequent vowel. Even if it is /ɨ/, the ranking WSPPrWd >> *P/ɨ ensures the priority of quantity over quality. 5.2.1.1.2 Against alternatives Several other accounts of the Pattani Malay data are possible, but all present flaws or serious disadvantages. These are thoroughly discussed in Topintzi (2008a) and briefly summarized here. To start with, consider the analysis of Hajek and Goedemans (2003, henceforth H&G), who first brought Yupho’s (1989) data to light. Their account shares with the present one the use of moraic onsets. Otherwise, though, they diverge in significant ways. More concretely, they put forward an exclusively weight-based analysis that interprets the phonetic vowel length phonologically too, thus claiming that there is a phonemic vowel-length contrast between monomoraic [Cɨ] syllables and bimoraic [CVː] ones. While this claim is not concretely justified, it is not entirely unsubstantiated either, and it could perhaps be adopted had it not generated a crucial problem for H&G. To be more precise, H&G inevitably have to conclude that both [CVː] and [Cːɨ] syllables are bimoraic. And yet, the latter have to be somehow heavier than the former because of examples like [kːí-da] – which for them is [kμɨμ.daμμ] – because otherwise, final default stress would be anticipated. H&G’s answer to this issue is that onset weight in Pattani takes priority over nucleic weight. But this is deeply flawed. Besides receiving no other external back-up, it opens the door for extensions to this idea so that coda weight could also take priority over nucleic weight. This would mean that we should be able to find languages where CVC syllables are heavier than CVV ones. Possible illustrations of exactly this pattern have been proposed in Dutch, Tiberian Hebrew and Kashaya Pomo, but on closer inspection,

176

Onsets

these may prove to be fake counterexamples (see Gussenhoven 2008 for Dutch). If margin-weight priority was on the right track, then we would additionally expect languages with extremely complicated systems such as CμVμCμ > CμVμμ, VμμCμ > CμVμ, VμCμ > Vμμ > Vμ. To my knowledge, these are unattested. Another, more interesting, possibility would be to dispense with geminates altogether and argue instead that these are syllabic consonants and that the general stress system of Pattani is based on iambs (Moira Yip, p.c.). Along with certain other assumptions, this account could be appealing, but lacks support concerning the presence of syllabic consonants in Pattani. More importantly, though, it encounters a significant empirical problem, namely the inability to deal with the stress contrast between examples like ‘non-geminated’ dìɣí and ‘geminated’ dːíɣì. If their only difference is the presence/absence of syllabic consonants (what I’ve called geminates in my analysis), then why is it that only words with ‘syllabic’ consonants systematically receive initial primary stress? No straightforward answer exists for this question. The most promising account, however, treats geminates as an early stage of tonogenesis, i.e. a gradual phonological shift towards tone. This idea appears tentatively in Abramson (2004) and is based on two facts about Pattani: first, that F0 is one of the perceptual features associated with the initial geminates, and second, that Pattani speakers are bilingual in Thai – a tonal language (Abramson 2004). However, this evidence is fragmentary and is weakened once one examines the full picture of the Pattani facts. To begin with, higher F0, i.e. fundamental frequency (Abramson 1999) – along with greater amplitude (Abramson 1999) – is just a secondary cue for the singleton versus geminate contrast in Pattani and cannot serve as a sufficient cue for this perceptual contrast (Abramson 1999, 2004). Instead, on the whole it is longer duration that serves as the primary and robust cue for the contrast (Abramson 1987, 1991, 1998, 1999, 2003). The combination of these three correlates actually seems to point in the direction of stress (see e.g. Hayes 1995; Hyman 2006), rather than tone, which corroborates the present proposal. With regard to Abramson’s second observation, it is very much true that a language may become tonal through contact, as is the case for Cham (Thurgood 1996) or Middle Korean (Ramsay 2001). Opinions are split, however, on whether exposure to a tonal second language is sufficient to turn the first language into a tonal one. Salmons (1992: 57), for instance, argues for ‘a relatively uniform path of development cross-linguistically from tone to pitch then to stress’ (emphasis added mine), rather than the other way round. Evans (2001: 81), on the other hand, suggests that the tones of Southern Qiang

Onsets and geminates

177

derived – among others – through contact with Mandarin. Proto-Qiang and Northern Qiang instead present lexical stress. Presumably, then, at the time of separation from the Northern dialects, Southern Qiang had stress too. For the sake of argument, let’s suppose that a stress-to-tone shift indeed occurred. Then what this would imply is that duration should be demoted to a secondary cue and F0 promoted to a primary one for the singleton–geminate contrast (Salmons, p.c.). But, as we have explained before, it is the other way round. In sum, the moraic-onset-geminate proposal captures the full range of facts. Unlike the syllabic consonants approach, it explains why stress shifts only in geminate-initial words. Unlike the tonogenesis approach, it explains why variants such as [buwi] ~ [wːi] ‘give’ or [sɨdadu] ~ [dːadu] ‘police’ exist. This is because onset weight is involved and compensatory lengthening occurs. Unlike H&G – which also captures the range of facts – it carries no unsubstantiated predictions. And finally, it is also compatible with all the phonetic facts. Indeed, syllables with initial geminates sound accented, as Abramson reports, but this salience is not tone. It is stress, correlates of which can be pitch, duration and amplitude. Pattani exhibits all three. On a more phonological level, if we merely analyse initial geminates as moraic onsets that render the carrier syllable-heavy, then the stress (and compensatory lengthening) data find natural explanation, since heavy syllables commonly attract stress. Trukese 5.2.2 The next language I will consider is Trukese, a Micronesian language, spoken on the islands of the Truk Atoll south of Guam. Initial geminates and their effects can be identified in word-minimality data as well as in a type of compensatory lengthening dubbed geminate throwback (Davis and Torretta 1998). I will argue that in Trukese too initial geminates are best represented and analysed as moraic onsets. 5.2.2.1 The data Trukese minimal words consist of either a long vowel, i.e. (C)VV, or a geminate plus a short vowel, i.e. CiCiV (in the notation here, I follow other sources of Trukese). CVC or CV words are not allowed (Davis 1999b),3 which means that singleton codas contribute no mora (Muller 1999). More generally, Trukese 3

Although Muller (1999: 395, fn. 3, and p. 394, example in (2)) observes that there are verbs which are monomoraic, such as ma ‘to be ashamed’ or kak ‘ring’, this is never the case for nouns.

178

Onsets

singleton non-moraic codas only occur word finally and geminate moraic codas only occur word medially (Muller 1999; Davis and Torretta 1998: 112, fn. 2). Relevant minimal words appear below. (20)

Trukese minimal words a. CVV words maa oo téé4 núú b. CCV words tto kka čča

‘behaviour’ ‘omen’ ‘islet’ ‘unripe coconut’ ‘clam sp.’ ‘taro sp.’ ‘blood’

These data suggest that a bimoraic word minimum is imposed in Trukese, which can either be satisfied by a long vowel or by a geminate followed by a short vowel. Geminates consequently contribute to weight. The same conclusion is reached when another pervasive process of Trukese is examined, namely the deletion of the root-final mora when it coincides with the end of the word. Under certain circumstances, this deletion may interact with Word Minimality. First, let us examine the case where it does not. Deletion of the final mora in a trimoraic word generates no other effect, since the resulting word still satisfies the bimoraic word minimum (21). Here, only unsuffixed words undergo final mora deletion; suffixed ones are protected by the non-moraic suffix /-n/ in a coda position. (21)

Final mora deletion in words of more than or equal to 3μ Suffixed form sawaa-n ‘taro of’ orosseti-n ‘shore of’

Unsuffixed form sawa ‘taro’ orosset ‘shore’

In smaller words, things change. If the original word is bimoraic, deletion produces a monomoraic word. Since this is not permitted, lengthening occurs, which can take the guise of either lengthening of a vowel (22) or consonant gemination (23). (22)

Final mora deletion causing V-lengthening Suffixed form a. fasa-n ‘nest of’ b. fæne-n ‘building of’ c. tappa-mw ‘your coconut’

4

Unsuffixed form faas ‘nest’ *fas, *fass fææn ‘building’ *fæn, *fænn taap ‘coconut’ *ttap, *tapp, *tap

é=[ʌ], ú=[ɨ] (Davis and Torretta 1998: 112, fn. 2).

Onsets and geminates (23)

179

Final mora deletion causing C-gemination Suffixed form a. fitta-mw ‘your package’ b. kikki-k ‘you move’ c. kunnu-n ‘its rotation’

Unsuffixed form ffit ‘package’ *fiit, *fitt kkik ‘move’ *kiik, *kikk kkun ‘rotate’ *kuun, *kunn

There are numerous points to observe here. First, something that all forms share: it is impossible to have or generate (after final mora deletion) a final geminate, e.g. *fass, *fitt, etc. Geminates are either medial or initial. With regard to the latter in particular, we have direct evidence that they must be moraic, because as the unsuffixed forms of (23) demonstrate, they alone can fulfil the bimoraicity requirement without any additional V-lengthening. Second, bimoraic suffixed forms that comprise no geminate can appear unsuffixed only with V-lengthening (22) and not C-gemination. Third, suffixed forms with a medial geminate, i.e. CVCiCiV, may contrast as to the way they appear unsuffixed. In some cases, vowel lengthening is preferred, as in [taap] (22c); still in others, consonant gemination wins over, e.g. [kkun] (23c). The latter is what is termed ‘gemination throwback’ (Davis and Torretta 1998). Our task, then, is to analyse these data, showing how minimality and compensatory lengthening interact and explaining under which circumstances V-lengthening and C-gemination apply. 5.2.2.2 The analysis To begin with, the most obvious thing we need is the constraint WdMin which permits bimoraic [CVμμ] and [CμVμ] words, but bans subminimal [CVμC] and [CVμ] ones. This must be high ranked, given that all words conform to it. (24)

WdMin: Words are minimally bimoraic

Next, consideration of (21–23) leads to the requirement of a markedness constraint that induces final mora deletion. For this purpose, we can use Free-μ (see Davis and Torretta 1998) or *μ# (Muller 1999). This deletion is not always compensated for, as the data in (21) illustrate. Essentially, then, no compensatory lengthening arises, e.g. /oμroμsμeμtiμ/ > [oμroμsμeμt] *[oμroμsμeμμt], a fact that is accounted for by the ranking P-Dep-μ >> PosCorr discussed previously (see §3.2.2.2 on the definitions and the way these constraints work in the system). At this point, just recall that P-Dep-μ applies to non-positionally licensed moras, that is, a mora that non-uniquely dominates a segment and violates P-Dep-μ in cases of lengthening. P-Dep-μ is by nature designed in such a way that it disregards positionally licensed moras, e.g. cases where a single mora dominates a C, as in a moraic coda. Compensatory lengthening takes place on other occasions, however, as justified by the data in (22–23); more concretely, if satisfaction of WdMin is at

180

Onsets

stake, then lengthening is imposed so that the resulting word is at least bimoraic, e.g. /faμsaμ/ > [faμμs] *[faμs]. WdMin also regulates the application of *μ#. In particular, a bimoraic word like /maμμ/ ‘behaviour’ will stay [maμμ] rather than shorten to the subminimal *[maμ], indicating that WdMin >> *μ#. This sketchy analysis is displayed in the ranking below. (25)

Initial ranking of Trukese WdMin >> *μ# >> P-Dep-μ >> PosCorr

Such a ranking, however, provides no insight that would allow us to predict in which cases V-lengthening occurs and in which C-gemination. In particular, at first sight, the most problematic examples seem to be those of (23) as opposed to (22c), since all share [CVCiCiV] suffixed forms, but diverge in the unsuffixed patterns. To account for this unpredictability, I follow Muller (1999) in arguing that this stems from a lexical contrast between forms that underlyingly include one geminate (26b), and those that include two (26a). This claim is based on the following alternations. (26)

Morphological alternations and geminates5 Prefixation a. initial geminate o-kkun ‘rotate it’ æ-ppεp ‘cause to skip’ b. initial singleton fita-tap

Suffixation

No affixation

medial geminate kunnu-n ‘its rotation’ pεppa-n ‘its skipping’ medial geminate tappa-n

initial geminate kkun ‘rotate’ ppεp ‘skip’ long vowel (nouns) taap

In (26a) we find singleton–geminate alternations between two consonants, e.g. [k]–[kk] and [n]–[nn]. In contrast, in (26b), only p alternates between [p]–[pp], but [t] merely appears as a singleton. We are now ready to provide a full-fledged analysis of the Trukese data. We can break down the relevant data into four patterns. (27)

Overview: The Trukese final deletion patterns Surface Surface μForm Struc i. ii. iii. iv.

5

[orosset] [faas] [taap] [ppep]

oμroμsμeμt faμμs taμμp p μ eμ p

Input μStruc

Lengthening

Ex.

/oμroμsμeμtiμ / /faμsaμ/ /taμpμaμ / /pμeμpμaμ/

NO YES: V-lengthening YES: V-lengthening YES: C-gemination

(21) (22a) (22c) (23)

To highlight the length alternations involved in (26), I mark corresponding consonants in boldface and italics.

Onsets and geminates

181

Two important proposals pursued in the current work are implied through (27) and will become evident in the ensuing tableaux. These are clearly stated in (28). (28)

Crucial representational issues for the current analysis: i) underlying consonantal moraicity indicates geminates ii) initial geminates are syllabified as moraic onsets

Let us start by considering (27i). For this, it is enough to merely consider the constraints in (25). Since word minimality is satisfied, the winner must be one that fails to have a positional correspondent for the final mora hosted by /i/ (29a). (29)

No lengthening: /oμroμsμeμtiμ/ → [oμroμsμeμt] /oμroμsμeμtiμ/ ☞

a. b. c.

μ

μ μ μ

μ

μ μ μμ

μ

μ μ μ μ

WdMin

*μ#

P-Dep-μ PosCorr

o ro s e t

*

o ro s e t

*!

o ro s e ti

*!

Next, comes the case of (27ii), that is, a bimoraic stem which lacks any underlying geminate and has to lengthen after final mora deletion to fulfil the minimality requirement. (30)

Incorrect prediction instead of V-lengthening: /faμsaμ/ → [faμμs] /faμsaμ/ μ

WdMin

*μ#

a.

fa s

b.

faμsaμ

*!

c.

faμsμ

*!


d.

fμaμs

☞

faμμs

e.

P-Dep-μ PosCorr

*!

*

*!

The role of WdMin becomes evident, as it excludes (30a). The very similar (30c) is also ruled out, since it has a mora word finally. An identical violation is incurred by faithful (30b). The remaining two candidates are more problematic. Instead of vowel lengthening (30e), this ranking incorrectly produces the initially geminated candidate (30d). Obviously, something needs to be added to the ranking. This crucial bit is the incorporation of the familiar *Moraic Onset / *μ/Onset (cf. (16) in §1.3.3), which bans moraic onsets of any type. In the view that initial geminates are represented as moraic onsets, the ranking of *Moraic Onset above P-Dep-μ suffices to yield the correct output.

182 (31)

Onsets Correct prediction of V-lengthening: /faμsaμ/ → [faμμs] /faμsaμ /

☞

*μ/Onset

μ μ

a.

fas

b.

μμ

P-Dep-μ

*!

fa s

*

This leaves us with the cases with geminates. When there is one underlying geminate (27iii), again vowel lengthening emerges after final mora deletion. To avoid repetition, I exclude from evaluation candidates that violate WdMin and *μ#, which, as we have seen in (30), never stand a chance, and only consider the two most promising candidates. As anticipated, (32b) is the winner. (32)

Correct prediction of V-lengthening: /taμpμaμ/ → [taμμp] /taμpμaμ/

WdMin

*μ#

a. tμaμp

*μ/Onset P-Dep-μ PosCorr *!

☞ b. ta p μμ

*!

The remaining pattern is the one that includes two underlying geminates. Unlike all previous instances, initial gemination is now preferred over vowel lengthening. Applying the ranking established so far, however, no longer captures the facts. (33)

Incorrect prediction of C-gemination: /pμeμpμaμ/ → [pμeμp] /pμeμpμaμ/

WdMin

*μ#

*μ/Onset P-Dep-μ PosCorr


a. peμμp

*

☞ b. p e p μ μ

*!

We now face a paradox. To get V-lengthening in (27iii), the ranking *Moraic Onset >> P-Dep-μ is required. To get C-gemination in (27iv), the opposite is seemingly in order. To resolve the conundrum we need to introduce yet another constraint, which, if ranked above *μ/Onset, produces the desirable effect. (34)

(35)

Ident-Link-μ: Corresponding mora-segment associations between input and output must be the same (see McCarthy’s 2000 NoDelink and NoSpread; Morén’s 2001 Max/Dep-Link-μ) Correct prediction of C-gemination: /pμeμpμaμ/ → [pμeμp] /pμeμpμaμ / WdMin μ μ μ

a. p e p

b. pμeμμp

*μ#

IdentLink-μ

*!

*μ /Ons

P-Dep-μ

* **!

*

*

Onsets and geminates

/pμeμpμaμ / WdMin μμ

c. pe p ☞ d. p e p μ μ

*μ#

IdentLink-μ

*μ /Ons

**!* *

183

P-Dep-μ *

*

This fuller tableau evaluates the most reasonable candidates. The first keeps most of the input moras, also without messing in with the underlying associations, but fails because it retains the final mora, thus violating *μ#. The second candidate suffers from all kinds of violations. It geminates the consonant, producing a violation of *μ/Ons; it lengthens the vowel, thus violating P-Dep-μ; and, most importantly, it deletes the mora link of the second p and inserts a new one in the vowel, leading to double violation of Ident-Link-μ. The incorrect winner before, i.e. (33a) and now (35c) entails a similar set of violations. By avoiding a moraic onset, though, it adds an Ident-Link-μ violation. Finally, (35d) wins, because it minimally alters the mora-segment associations at the expense of the rather low-ranked *μ/Onset. Adding Ident-Link-μ to the previous tableau does not alter the outcomes in any undesirable way, a fact interested readers can double-check for themselves. Consequently, the ranking the Trukese facts require is this one: (36)

Final ranking for Trukese WdMin >> *μ# >> Ident-Link-μ >> *μ/Ons >> P-Dep-μ >> PosCorr

One final thing needs to be examined. How about the cases in (26a) which exhibit singleton–geminate alternations in different morphological environments, e.g. prefixation: æ-ppεp ‘cause to skip’; suffixation: pεppa-n ‘its skipping’; no affixation: ppεp ‘skip’? We have an account for the bare root and prefixed forms, but we also need one for the suffixed form. As we know, underlyingly there are two geminates, so perhaps we would expect the output: *ppεppa-n. Its failure to arise can be attributed to banning the co-occurrence of multiple geminates within a single morpheme. In other words, this can be seen as an OCP effect, a position also adopted by both Davis and Torretta (1998) and Muller (1999). For ease of reference I will call this OCP Gem{morpheme},6 but more likely it is about a ban on consecutive moraic consonants within the same morpheme. 6

This is not an inter-morphemic restriction, cf. kučču-ŋŋaw ‘to fit badly’ (Muller 1999: 403) which is well formed. The OCP explanation is not particularly appealing since the OCP usually makes reference to tones or features, not to more abstract notions such as weight. In fact, it is difficult to imagine exactly how one would devise an OCP(weight) constraint. Perhaps, there is instead an upper bound on the moras a word can include, cf. similar restrictions in Bella Coola (Bagemihl 1998) or Yoruba (Ola 1995, 1997). For current purposes, however, I will make use of the OCP.

184

Onsets

This explanation rules out (37d) below and suggests that in cases of this sort, only one geminate may survive. But which? It can be shown that it is systematically the second one. Although ranked quite low, *Moraic Onset will rule out unneeded moraic onsets given the chance. But in principle – and in practice as we will see in §5.3 – the medial geminate too could also syllabify as a moraic onset. There is a way round this; keeping only one of the geminates – in line with the OCP – and syllabifying it not as a moraic onset, but as a coda-onset sequence will avoid violating *Moraic Onset, only at the expense of low-ranked NoCoda. This state of affairs is depicted in (37). (37)

Suffixed input forms with two geminates /pμεμpμaμ+n/ ppεppa+n

OCP Gem

Ident- *Moraic NoCoda Link-μ Onset

☞ a. pεμpμaμn

pεp.pan

*

b. pεμpμaμn

pε.ppan

*

*!

μ μ

ppε.pan

*

*!

μ

c. p ε pa n μ μ μ μ

d. p ε p a n

ppεp.pan

*!

*

*

5.2.2.3 Summary Trukese is similar to Pattani Malay in that it too presents initial moraic geminates, which are best represented as moraic onsets. Unlike Pattani, whose geminate effects become visible with regard to stress (and compensatory lengthening), in Trukese, geminates play a vital role in word minimality and in the process of final mora deletion which, depending on the circumstances, induces compensatory lengthening. Both case-studies demonstrated that underlying weight – like its coerced variant – participates in the same types of phenomena, e.g. stress, word minimality, CL, etc., with similar effects.

5.3

Word-medial moraic-onset geminates

Up to this point, the discussion has centred on initial geminates and their representation as moraic onsets. Attention now shifts to medial geminates. Although there is no absolute consensus, the majority of geminate accounts syllabify them medially in a moraic-coda–non-moraic-onset configuration, as shown below.

Onsets and geminates (38)

185

Geminates word medially: most common representation (see Hayes 1989) σ

σ µ

µ

µ

V

Cː

V

The geminate’s weight and increased length are represented respectively by means of a mora associated to the coda and by double linking to higher positions. The uniqueness of this representation for word-medial geminates, however, is merely an illusion, as Ham (2001) has convincingly explained. The only part of (38) that moraic theory really necessitates is that the geminate bears a mora, since in standard moraic theory (Hayes 1989) a geminate is by definition an underlyingly moraic consonant. The geminate’s double association, however, responds to syllabification issues. To see this, imagine a situation where the geminate was singly linked to a coda (39a). In this case, a [VCː.V] sequence would be created where the second syllable would systematically be onsetless. But this is empirically flawed as, given the opportunity, languages prefer onsetful syllables over onsetless ones. The structure in (39b) achieves exactly this through an additional link to the following syllable, a move that seemingly buys for free the extra length geminates commonly display. We will see later that this result actually puts (39b) also at a disadvantage (§5.4.2.1–2). (39)

Geminates word medially a. Avoidance of onsetless σs *

σ

b. Preferred structure

σ

σ

σ

µ

µ

µ

µ

µ

µ

V

Cː

V

V

Cː

V

According to standard moraic theory, then, (39b) is the only configuration that simultaneously fulfils the moraification and syllabification requirements of geminates. Or is it? For it to hold true, a vital hypothesis – but one that is only implicitly assumed – must be correct, namely that onsets can never bear moras. Recall that this is widely accepted, but in reality is merely a stipulation that the present work hopes by now to have dissolved. The alternative, then, would be to also allow a medial geminate to syllabify wholly within a syllable as a moraic onset.

186 (40)

Onsets Geminates as moraic onsets σ µ

µ

Cː

V

More accurately, (40) is not a mere alternative, but an obvious prediction of the current model. There is no reason for a geminate which syllabifies as a moraic onset word initially not to do the same word medially. In fact, this is entirely anticipated. The forthcoming sections discuss several cases of this sort. Notably, the discussion is not as extensive as was the case with Pattani Malay and Trukese. This is due to the more scattered information available on these languages. 5.3.1 Marshallese Marshallese is a Micronesian language in which the effects due to medial geminates become evident in the exploration of stress.7 The relevant data come from Zewen (1977: 40–1), which is the most specific description of stress to date. Most other works on Marshallese ignore stress (Abo et al. 1976; Bender 1976, 1991; Hendricks 1999). The remaining ones either discuss it briefly (Lynch 2000 based on Rehg 1993) or too generally (Bender et al. in prep), leaving many issues unanswered. Stress (marked in boldface below) is assigned within a trisyllabic window at the right edge. If the final three syllables are light, the antepenult gets stress (41a). Codas – at least final ones – do not render their syllable heavy, thus they do not count for stress purposes (e.g. ekajet and not *ekajet).8 If there is a heavy syllable, then this receives stress (41b). If the word is disyllabic and made of syllables that are either both heavy or both light, then penultimate stress occurs, i.e. the leftmost of the two syllables gets stressed (41c).

7

8

Marshallese also presents a relevant case of reduplication, but discussion about this is postponed till §6.2.4.2. Marshallese diphthongs are: ao, au, ai, ae, ei, öu. These act as long vowels. It is implied that the remaining vowel sequences are heterosyllabic. Spelling conventions used in sources and adopted here: i) from Zewen (1977): b ́ = rounded bilabial plosive, i.e. something like bw, ¸ = rounded bilabial nasal, i.e. something like mw, ¹ = ɳ, » = ɭ, r = alveolar-coronal trill, r̀ = dental-apical trill; ii) from Abo et al. (1976): ¹ = nw, ū = ɯ.

Onsets and geminates (41)

187

a. Trisyllabic words LLL: Antepenultimate stress ekajet nukileb jekaru lakatib

‘to judge’ ‘to have a big family’ ‘coconut syrup’ ‘to make angry’

b. Trisyllabic words LHL or (L)LH: Stress H je.uː.rur ‘commotion, excitement’ je.ro.aːn ‘to waste’ köräː ‘woman’ jeläː ‘to know’ kijeːk ‘fire’ c. Disyllabic words (either LL or HH): Penultimate stress neb́ar ‘to praise’ maːjaːj ‘to be clear of underwood’

Schematically, this stress pattern can be given by the ranking below. (43) reminds us of the relevant constraints. (42)

*Extended Lapse Right >> WSP >> Align-Head-Left

(43)

a. Align-Head-Left: Align the head syllable of a prosodic word to the left edge of the prosodic word (McCarthy and Prince 1993) b. *Extended Lapse Right: No sequences of more than two consecutive stressless syllables at the right edge of the word (Elenbaas and Kager 1999; Gordon 2005)

Stress tends to be at the left side of the word, maximally reaching the antepenult, because *Extended Lapse Right restricts its location within the last three syllables of the word. In the absence of heavy syllables, it will dock onto the leftmost stressable syllable; but if a heavy syllable is present and within the stress window, then stress is assigned on that syllable. Finally, codas contribute no weight, meaning that their host syllable remains light. Data with medial geminates fit this description too. Note that here only the stressed vowel is indicated (in bold). This is deliberately so, because I wish to remain neutral at this point with regard to the assumed geminate syllabification. (44)

Geminate stress (Zewen 1977: 27) jib̍b̍uŋ (y)em̍ m̍ an emmerˋ

‘morning’ ‘good’ ‘to be ripe’

188

Onsets

Only under tautosyllabic syllabification and contribution of a mora by the geminate9 can the facts fall out naturally. (45)

ji.b̍b̍uŋ w

*jib̍.b̍uŋ w

(y)e.m m an

*(y)emw.mwan

As shown in (45), the geminate renders its syllable heavy and thus stressattracting. The alternative coda-onset geminate syllabification fails conspicuously. Consider jib̍.b̍uŋ, for example. Since codas do not contribute any weight, then both syllables should be light, in which case we would expect jib̍.b̍uŋ, because this is the leftmost syllable that can bear stress (cf. 41c). However, given the presence of a geminate, such a medial coda could be moraic. Even in that case, though, jib̍.b̍uŋ is again predicted since it is the heaviest syllable (cf. 41b). Consequently, the geminate must be tautosyllabic, i.e. wholly syllabified in an onset position, and moraic, i.e. so as to attract stress. Unfortunately, the lack of further data does not allow us to shed more light on this issue, but one potential test for stress could be the following: Abo et al. (1976: xxxi) report that Marshallese commonly employs the prefix ri with its various alternants to form person nouns, e.g. ruwa ‘sailor’. Now, a number of person nouns have their first consonant doubled when this prefix is added, thus tarian̹ae > rūttarian̹ae ‘soldier’. If stress assignment indeed works as mentioned, then for a base like katta, we would expect katta if the geminate is tautosyllabic, but katta if it is heterosyllabic. Adding the prefix ri- should lead to consonant doubling, e.g. something like ri-kkatta.10 Given that the leftmost heaviest syllable receives stress, then depending on whether the geminate is tautosyllabic or heterosyllabic, we should expect rikkatta or rikkatta, respectively. The Marshallese facts point towards an analysis of geminates as medial moraic onsets, thus arguing against the unique representation of medial geminates. Depending on the language, these can be heterosyllabic in a coda-onset configuration, or tautosyllabic in a heavy syllable. 5.3.2 Trique Recall from the discussion in §3.3.4 that Trique is a language whereby deletion is expressed by means of CL, presumably in the onset position. Here, it is not 9

10

I have managed to spot only four examples with geminate consonants and stress clearly marked. Three of them appear in (44). The exception is ,ar̀r̀ik ‘boy’ (Zewen 1977: 27) with stress on the first syllable. This suggests a heterosyllabic syllabification. Unfortunately, investigation for additional data has not been fruitful. Two geminates within a word occur as in yibbiddikdik ‘many crumbs, grains, morsels’ (Abo et al. 1976: 32).

Onsets and geminates

189

that onset deletion triggers CL, but deletion of another segment (or syllable), e.g. yan® > nːa leads to CL through an onset. On more speculative ground now, and in the absence of other information, we can argue that the onset produced is a geminate exactly like the ones found in Pattani Malay. The only difference is that, in at least some dialects, e.g. the San Juan Copala one (Hollenbach 1977), it appears word medially, rather than word initially. In all likelihood, then, Trique is another language that seems to exhibit medial onset geminates. 5.3.3 Djeebbana The Australian language Djeebbana (also Ndjébbana) is also cited by Wolf (2008) as another language that utilizes medial onset geminates (but not initial ones). The full analysis can be found in that work. For our purposes, it suffices to say that in Djeebbana, a process of prefixation creates the right environment for gemination to occur. In doing so, the geminate surfaces as a heavy onset, rendering its syllable heavy and as such a better carrier of stress (due to the Stress-to-Weight Principle).11 5.3.4 Puluwat Puluwat is yet another Austronesian language – actually a relative of Trukese – that exhibits geminates in all three positions (Elbert 1974). It clearly exhibits a contrast between singletons and geminates (including some nice minimal pairs), as shown in the following examples. 11

Cases like those in Trique, Djeebbana, Bellonese (§6.2.4.1), Marshallese (§6.2.4.2), and perhaps – under some interpretations – Pattani Malay above, may not strictly include ‘geminates’ if the term solely refers to underlyingly moraic consonants. This is because in each of these cases, the geminate is more ‘derived’ (as a result of prefixation, reduplication, etc) rather than underlying. However, the matter is far from clear. If prefixation or reduplication (in Djeebbana, Bellonese, Marshallese and Pattani Malay) involves the presence of an unlinked mora in the input that due to the constraint ranking ends up creating a moraic onset, we could still possibly claim that this moraic onset has distinctive weight, i.e. is a geminate, because it has a mora whose source is in the input. Otherwise, these should be considered ‘derived’, i.e. coerced moraic onsets. Trique is even more perplexing. It involves CL, which, as we have seen in Ch. 3, is quite problematic assuming preservation of underlying moras in an OT framework. The present theory has bypassed this problem by suggesting that CL is about position preservation through a mora, whose source may be in the input or not (to satisfy RotB). Thus, CL is not a good testing ground to answer the matter of distinctive weight. That said, certain observations should be kept in mind. First, distinctive moraic onsets undoubtedly exist, as shown in Trukese (§5.2.2), Marshallese (§5.3.1) and Puluwat (§5.3.4), thus use of geminates for onsets too is entirely justified. Second, as we have seen, both distinctive and coerced weight participate in the same phenomena. It is just that we expect the former to exhibit more freedom as to which onsets may appear moraic (since they are underlyingly specified as such). And finally, to avoid confusion and given that, in many instances, it is unclear whether we should consider a case as having distinctive or coerced weight, I have decided to treat prototypical geminates and ‘geminated’ ones on a par (see §1.3.2 for discussion).

190 (46)

Onsets Puluwat initial singletons and geminates a. word initially par tefan b. word medially yapaar sinisiy c. word finally lap pwɨl

‘blind’ ‘his relationship’

ppar ttef

‘attached’ ‘to tear’

‘to reward dancers’ yappar ‘change’ kilissow

‘breadfruit sap’ ‘thanks’

‘big’ ‘burn’

‘to upset’ ‘to break’

rapp pwɨll

Unfortunately, though, Elbert’s poor discussion on stress – and no other weightsensitive phenomenon – does not provide us with further information about this language.12 However, a few remarks can be made through what is currently known about it. First, Puluwat seems to once more verify the distinction between distinctive and coerced weight. As mentioned, it is the former and not the latter that could allow for patterns that do not conform to the voicing moraic markedness of onsets, because by being lexical, distinctive weight allows for arbitrary contrasts. Indeed, all Puluwat consonants may appear as singletons or geminates with the exception of /w/ and /y/. This finding is irrelevant to voicing. Second, Puluwat speakers themselves call geminates ‘heavy’ and singletons ‘light’, an interesting anecdotal piece of information that perhaps highlights the difference in weight between the two. Finally, the data Elbert presents point to the direction that initial geminates syllabify as moraic onsets, medial ones as straddling the syllable boundary and final ones as moraic codas. This is because Puluwat allows for both geminate consonants and long vowels, admitting all of the following monosyllabic words: CV, CVC (1 mora); CːV, CVː, CːVC, CVːC, CVCː (2 moras); CːVː, CːVːC, CVːCː, CːVCː (3 moras). Such patterning is of course possible under the assumption that geminates contribute one mora, long vowels two, and singleton onsets and codas none. That this idea is on the right track is supported by the lack of *[CːVːCː] words. Had these geminates been non-moraic, i.e. fake geminates (see §5.4.2.2 for discussion), then this form should be perfectly well formed. But if geminates bear weight, then its absence is straightforward; such a form would be quadrimoraic, a pattern universally banned! 12

At some point, Elbert (1974: 5) states that ‘/h/ differs from all other consonants in that syllables beginning with /h/ are always stressed’ (emphasis added mine). Goedemans (1998: 134) takes this at face value. I am more sceptical, since Elbert soon enough (1974: 13) retreats from this, saying that ‘Syllables beginning with h- are usually stressed’ (emphasis added mine). In the absence of a clear discussion about Puluwat stress and given the inconsistency mentioned above, I have decided not to take this observation into account.

Onsets and geminates

191

Eastern Popoloca 5.3.5 In Eastern Popoloca (Oto-Manguean), spoken in San Juan Atzingo, Mexico (Kalstrom and Pike 1968), stressed syllables include a ‘component of length’ (1968: 16). This may either be a long vowel or a long consonant (in the present terminology, a geminate). We thus get pairs such as the following: (47)

Eastern Popoloca (tones are omitted) Stress Type 1 (includes geminate) nčisːé ‘clay pitcher’ thakːó ‘early in the evening’

Stress Type 2 (includes long V) nčíːse ‘mesh bag’ tháːko ‘is teaching’

Now the question is whether a heavy syllable attracts stress or whether stress can only dock onto a heavy syllable. In other words, does Eastern Popoloca utilize the Weight-to-Stress Principle (WSP) or the Stress-to-Weight Principle (STW)? In the light of this statement: ‘Every phonological word has either a long consonant or a long vowel, but not both’, I am inclined to suggest that STW is applicable here, because otherwise it gets more difficult to explain why we cannot have two or more long vowels or geminates per word (unless some kind of OCP is in action). This analysis is supported by other information about Popoloca. Kalstrom and Pike mention that Stress Type 1 usually occurs on the last syllable of a verb or noun stem (and on certain suffixes). Stress Type 2, on the other hand, is more mobile. It also always occurs on the next-to-last stem syllable. These facts, combined with the examples in (47), lead us to the following conclusion: i) Stress in E. Popoloca is lexical, since we can have contrastive stress in identical stems; ii) The stress-bearing syllable must be heavy. If the last stem syllable is to be stressed, then heaviness is achieved by means of a medial onset geminate. If it is the syllable before it, then a long vowel emerges instead.

5.4

Typology, more theory and alternatives

This section deals with typological predictions that arise from the current proposal, issues that potentially question its validity, as well as reasonable alternatives to it. First with typology; I begin by considering the factorial typology of medial geminates and show that the only two attested configurations are the ‘flopped’ structure and the onset geminate (§5.4.1). This task is subsequently complemented by looking into the representation of geminates in various positions within the word (§5.4.2.3). Next, the relationship between the phonetics and phonology of geminates is addressed. Given that I subscribe to the view that gemination only requires the

192

Onsets

presence of an underlying mora and consequently double linking of the geminate to higher prosodic structure is not obligatory, the relationship between weight and length in geminates is possibly at stake. In §5.4.2.2, however, I present recent phonetic work (after Ham 2001) that shows that the added duration of geminates observed in the phonetics is the product of an underlying mora and not of double linking per se. Moreover, I argue that the phonological contrast between geminates and singletons is preserved in all positions (see §5.4.2.3). The purpose of this discussion is more to show that the present account is consistent with various phonetic and phonological findings, than to offer a compelling argument in favour of it. Finally, §5.4.3 briefly examines alternatives and argues that the present account is superior as it captures a wider range of empirical facts, while being theoretically and technically workable. Generating medial geminates: a factorial typology 5.4.1 This chapter has pursued the idea that moraic-onset geminates exist. Word initially, they constitute the sole representation available, while medially they co-exist with the coda-onset geminates that straddle the syllable boundary, the so-called ‘flopped’ structure. Both representations are illustrated in (48). (48)

Geminates word medially a. Moraic onsets σ

b. ‘Flopped’ structure σ σ

µ

µ

µ

µ

µ

Cː

V

V Cː

V

The conditions under which each emerges are down to the particular constraint ranking the language at hand promotes. In particular, (48b) contains a coda, therefore violating NoCoda (Prince and Smolensky 1993/2004). (48a) lacks such a violation, but incurs that of *Moraic Onset, due to the weightful onset it introduces. A language with geminates of the type in (48a) should thus have NoCoda >> *Moraic Onset, so that codas will be minimized in favour of tautosyllabic moraic onsets. Such a ranking does not necessarily imply that the language will lack codas altogether. It just indicates that they will not be the preferred representation for underlyingly moraic consonants, but they could emerge on other occasions, e.g. in non-geminate CC clusters. The opposite ranking *Moraic Onset >> NoCoda expresses the fact that geminates are

Onsets and geminates

193

heterosyllabic. As we know, in the majority of languages, medial geminates present the ‘flopped’ structure. This should come as no surprise. Given that *Moraic Onset is so often cross-linguistically high ranked, it is natural that languages with medial moraic-onset geminates will be markedly few, because the most basic requirement for them, i.e. low-ranked *Moraic Onset, is unsatisfied. Of course, this relates to the reason moraic onsets are rare in the first place, a condition probably attributed to functional explanations, as discussed in Goedemans (1998), Smith (2005) and Gordon (2005). A factorial typology of the situation with medial geminates is depicted in the tableaux below. Although three constraints are employed (thus 3! = 6 grammars are possible), two tableaux suffice to draw the relevant point. As I will show, this is because in practice only two constraints conflict with each other. (49)

Medial geminate factorial typology /VCμV/

(i) ☞

a.

VCμi.CiV

b.

VCːμ.V

c.

V.CːμV μ

(ii)

☞

/VC V/ a.

VCμi.CiV

b.

VCːμ.V

c.

V.CːμV

Onset

*Moraic Onset

NoCoda *

*!

* *!

Onset

NoCoda

*Moraic Onset

*! *!

* *

Besides NoCoda and *Moraic Onset, which evaluate the configurations in (48), Onset is also used to evaluate a third possible representation for geminates examined earlier, i.e. [VCːμ.V] (39a). Previously, we discussed the implausibility of such a structure – despite being allowed by moraic theory – but now we can also exclude it on technical grounds too.13 OT neatly accounts for its absence. In particular, [VCːμ.V] is harmonically bounded by [VCμi.CiV]. While both violate NoCoda equally, the former also incurs a fatal violation of Onset, thus under no ranking of the proposed constraints, Onset, NoCoda and *Moraic Onset, can it be the designated winner. Effectively, then, (49a) and (49c) are the sole rivals, so the only conflict that matters is that between NoCoda and *Moraic Onset. The factorial typology for medial geminates is thus practically down to two rankings: *Moraic Onset >> NoCoda, and NoCoda >> *Moraic Onset, hence the two tableaux 13

For additional discussion on the implausibility of this candidate, see Topintzi (2008a, §3.1).

194

Onsets

above. In the first case, the ‘flopped’ structure is selected. In the second case, a moraic-onset geminate is admitted.14 This is the ranking needed for medial onset geminates like those of Marshallese and Trique previously examined. 5.4.2 What does it mean to be a geminate? As said, the most-often discussed geminates are the medial ones, which are usually taken to straddle the syllable boundary having associations to a moraic coda and a following non-moraic onset. An advantage of such a configuration is that it represents the increased length geminates demonstrate over the corresponding singletons. This is of course not the case in the present work, where onset geminates (in initial and medial positions) have just a single link to a moraic onset. Is double linking then imperative? If it is, then the present enterprise is on the wrong track. Obviously, this is not the route we wish to take. But if double linking is not essential, then do we perhaps run the risk of eliminating the contrast between geminates and singletons? This is tackled next. 5.4.2.1 Geminates in moraic theory Throughout this chapter, an assumption has been implicitly adopted, namely Hayes’ (1989) claim that in moraic theory geminates differ from singletons in that the former are underlyingly moraic consonants as opposed to the nonmoraic latter ones, i.e. /Cμ/ vs. /C/. In other words, the framework contrasts these two types of consonants with respect to their moraicity and not their linking structure. This is of utmost importance for our purposes, since it means that there is no real requirement for the present model – which explicitly uses standard moraic theory even if it slightly modifies it – to accommodate this property. As said, the double linking evident in the heterosyllabic structure of geminates is really the result of syllabification needs, while incidentally mirroring the increased length geminates exhibit. 5.4.2.2 Phonological weight and phonetic length in geminates According to Ladefoged and Maddieson (1996: 91–2), geminates are indeed longer than singletons. Depending on the language, geminate stops can be anywhere between one-and-a-half to three times longer than their singleton counterparts. Phonetically, then, the geminate’s duration varies, but its phonological weight remains constant, i.e. a mora. 14

For completeness, note that of the four remaining grammars in the factorial typology, the two with NoCoda top ranked choose (49c) as the winner and the two with *Moraic Onset top ranked choose (49a).

Onsets and geminates

195

This observation is consistent with the ideas of Hubbard (1994) and Ham (2001), who claim that moras are allocated a minimum target duration, whose implementation in terms of timing takes precedence over any other segmentspecific effects, such as place of articulation and voicing. By extension, if the main characteristic of geminates is to be moraic, then it is anticipated that the implementational priority is to achieve the mora’s minimum duration target, thus leaving less space for other durational segment-related differences. This view is called ‘the moraic primacy in the implementation of timing’. To test this hypothesis, Ham (2001) compared the segment durational effects of voicing and place of articulation between singletons and geminates in four languages: Bernese, Madurese, Levantine Arabic and Hungarian. With respect to place of articulation, a strong tendency is that closure duration decreases as the point of oral constriction becomes less anterior (Ham 2001: 215 and references therein). Ham found that the place effect on closure duration is substantially larger in singletons than in geminates. Similarly, although voiced stops are generally shorter than their voiceless counterparts due to the aerodynamic difficulties of sustaining vocal fold vibration in the presence of a complete oral seal (Ham 2001: 217 and references therein), singletons exhibited more voiceconditioned variation in closure duration than the geminates. The upshot of this is that geminates present smaller durational effects on the segmental level compared to singletons, exactly as predicted by ‘moraic primacy’. Such a conclusion is inconsistent with studies that view increased length as the primary characteristic of geminates rather than their moraicity (Selkirk 1990; Tranel 1991). As illustrated below, these accounts represent geminates phonologically via double linking in terms of two timing slots or two root nodes. We will see next that beyond the phonetic evidence, there are also phonological arguments against such accounts. (50)

Geminate representations based on length, i.e. ‘fake geminates’ in present context Selkirk (1990): as two root nodes

Tranel (1991): as two timing slots

σ µ t t e

x

x

t

[-cont] coronal

Motivation for such configurations comes from so-called ‘weightless geminates’, that is, consonants that are longer than singletons, but without bearing

196

Onsets

any weight as in Malayalam, Selkup and Tübatulabal (Tranel 1991) and possibly Tukang Besi (Donohue 1999; Topintzi 2006b, §7.2.5.3). In the present account, these cannot be real geminates, since geminates are by definition underlyingly moraic. Rather, they are simply doubled consonants with two root nodes (cf. (50a)), that is, ‘fake geminates’, which can receive a representation that involves a complex onset made of two identical consonantal root nodes. For instance, Malayalam’s ‘weightless geminates’ are tautosyllabic in the onset position, as evidenced by native-speaker intuitions, language games and word-level restrictions (K.P. Mohanan 1986: 73–4; T. Mohanan 1989). Such ‘fake geminates’ also appear under morphological concatenation (Hayes 1989) or across a syllable boundary when they involve duplicated segments (Schein and Steriade 1986). Actually, if geminates and doubled consonants are different species, as argued here, one may expect them to co-occur in the same language. This is what Ham (2001) argues for the case of Bernese, since he treats word-medial and word-final geminates as real geminates, but word-initial ones as doubled consonants resulting from morphological concatenation. An additional phonological argument in favour of geminates as underlyingly moraic rather than long is the following. According to Tranel (1991), singletons are singly linked and geminates are doubly linked to the segment. Since length is what distinguishes them and not weight, weight should be uniform within a language, i.e. all codas should be light or heavy regardless of whether they belong to a singleton or geminate coda. This is what he calls the Principle of Equal Weight for Codas. If right, this is a knock-out argument for the current account, because it proves that geminates may be long, but non-moraic, against expectations. What the Principle does not predict – and moraic theory does – is a language where geminates are heavy (because they bear an underlying mora), but singletons are not (because WbyP/Moraic Coda is inapplicable). Indeed, languages of this sort exist (Davis 1999a). In Koya, Sinhala and Hausa, CVV and CVG (G=geminate) syllables act as heavy, whereas CVC and CV ones behave as light. The claim of moraic theory that what makes a geminate is its underlying weight is therefore corroborated. Returning now to the issue of the segmental effects on duration discussed previously, the ‘length’ models above predict that these should either be greater in geminates, if the magnitude of the effects directly correlates to the number of timing slots, or they should be the same between geminates and singletons, under the assumption that the effects are distributed across both the timing slots (see Ham 2001: 223). Neither of these predictions

Onsets and geminates

197

corresponds to the empirical facts, whereby segmental effects actually minimize in geminates. Another problem these models face is failure to account for the variability of duration between geminates and singletons. Due to the double association to timing structure, geminates should be twice as long as singletons are. But, as we have seen above, they can be anything from one-and-a-half to three times longer than singletons (Ladefoged and Maddieson 1996). In the present framework that espouses moraic theory, all that is said is that phonologically geminates are underlyingly moraic and the added duration observed in the phonetics is the product of underlying moras, not of double linking per se. Thus, the phonetic implementation of phonological weight is regulated on a language-specific basis. Length, then, should not serve as a testing ground for the current proposal, since it is compatible with both single and double linking in geminates. In sum, then, the distinction between geminates and singletons is one of underlying phonological weight. Not only do a number of phonological arguments favour such a conclusion, but also several phonetic facts are consistent with it. While geminate length depends on language-specific phonetic factors, single or double linking is phonologically regulated in order that it satisfies syllabification requirements, such as the avoidance of codas (in the case of onset geminates) or the avoidance of moraic onsets (in the case of ‘flopped’ geminates). 5.4.2.3 Geminates in various positions The final issue to be discussed with respect to linking relates to the location of geminates. Word medially, there is enough phonological structure to syllabify the geminate either heterosyllabically [VCμi.CiV] or tautosyllabically [V.CːμV]. Word finally and word initially, though, only the single-linking option is available, since there is neither onset nor coda to host one half of the geminate. Thus, we find VCμː]# word finally in languages like Hungarian (Ham 2001) and #[CμːV word initially in Pattani Malay and Trukese (§5.2.1–5.2.2). This indicates once more that single vs. double linking cannot be the source of gemination. In turn, a new problem emerges. If phonetics and not phonology is responsible for the implementation of phonological weight, then how can we ensure that the surface contrast between singleton [CVCμ] and geminate [CVCːμ] is retained, given that on the surface both are phonologically moraic? I claim that word medially, the issue does not arise in the first place. The ‘flopped’ structure buys us the singleton–geminate contrast for free (51).

198 (51)

Onsets Word-medial distinction between singletons and geminates a. Singleton CVCµ

b. Geminate CVCːµ σ σ

σ

σ µ

µ

µ

µ

µ

µ

V

Ci Cj

V

V

Cː

V

Word finally, as Ham (2001) notes, given the sample of languages he examined (Bernese, Levantine Arabic and Hungarian), singletons were consistently light and geminates systematically heavy, a fact attributed to their inherent moraicity. The contrast is thus retained word finally too in the guise of: (52)

Word-final distinction between singletons and geminates a. Singleton CVC] σ

b. Geminate CVCːµ] σ

µ

µ

µ

V C]

V

Cː]

As a result, Ham predicts that no language will possess both heavy singleton CVCμ and geminate CVCːμ word finally, although word medially singletons can be moraic, since they differ from geminates in their syllabification (cf. (51)). It remains to be seen whether this prediction is borne out in a wider sample of languages. Finally, word initially, the contrast between singleton and geminate onsets is again maintained. For instance, in Pattani Malay, singleton [ɟàlέ] ‘road, path’ vs. geminate [ɟːál¶] ‘to walk’ is marked through the absence or presence of an onset mora respectively.15 (53)

Word-initial distinction between singletons and geminates a. Singleton CV σ

C 15

b. Geminate CµV σ

µ

µ

µ

V

Cː

V

The same contrast is also available word medially for cases like Marshallese with onset geminates.

Onsets and geminates

199

Further advantages of the current proposal 5.4.3 The view that sees initial geminates as moraic onsets simultaneously improves on or resolves problems that other accounts face without any additional mechanisms or assumptions. A fuller investigation of alternatives is found in Topintzi (2008a), while some additional discussion appears in §7.2. Recall from the previous discussion that one of the main problems that the double linking of the ‘flopped’ structure generated was its inability to extend to other positions, e.g. word initially. Answers to this problem include Davis (1999b) and Curtis (2003). (54)

a. Davis (1999b) σ µ

µ

Cː

V

b. Curtis (2003) PrWd … σ µ

µ

Cː

V

The first (54a) suggests that the first part of the geminate remains unlinked with the repercussion that the initial mora should not be able to participate in weight-related processes because to do so, linking to higher prosodic structure is required (Kiparsky 2002). Obviously, then, examples such as word minimality in Trukese demonstrate that such linking must exist. Curtis attempts to solve this problem by allowing the geminate to link to the PrWd, at the expense of a new problem; more concretely, (54b) seems to partially identify initial geminates with unsyllabified, yet moraic, consonants like the ones surfacing in Piro (Lin 1997 and §3.3.2) and Bella Coola (Bagemihl 1991, 1998 and Ch. 4), even though the status of geminates provides no evidence against syllabification. Treating initial geminates as moraic onsets instead, as in (55), bypasses all these problems at no extra cost; first, it fully prosodifies the geminate, permitting it to contribute to syllable weight without any complication (cf. (54a)); second, it is dominated by the syllable node, thus being entirely different from unsyllabified consonants (cf. (54b)); finally, by being in a moraic-onset position, it avoids the ‘flopped’ structure yet retains the mora that is characteristic of geminates. And recall that this final point was, after all, the problem that set off this chapter. (55)

Initial geminates as moraic onsets σ µ

µ

Cː

V

200 5.5

Onsets Concluding remarks

To conclude, Chapter 5 has focused on geminates, especially in initial and medial positions. Retaining the major insight of moraic theory, i.e. that geminates are underlyingly moraic consonants, it has extended the theory in a rather straightforward way. In particular, it has claimed that if realization of an underlying mora is all a geminate requires, then a syllabification whereby the geminate appears as a moraic onset is directly predicted by the model. The only way to stop this from happening is by explicitly and arbitrarily stating that such a configuration cannot exist. This is exactly what standard moraic theory does by stipulation (Hayes 1989). Despite its ‘convenience’ (see Morén 2001: 8) stemming from the fact that ‘in virtually all languages, the onset is ignored for purposes of calculating weight’ (Gordon 2006: 3), there is nonetheless a small number of languages that prove standard moraic theory wrong. This book has been dedicated to doing exactly that. Unlike the previous chapters that centred on phenomena that largely have to do with coerced onset weight, i.e. weight enforced on the surface, this chapter has shown that underlying onset weight also exists, as logic demands, and takes the shape of geminate consonants. Using empirical data from Pattani Malay and Trukese, the claim is that initial geminates are best analysed as moraic onsets. Detailed case-studies of both have been constructed to illustrate exactly how these work. Then, the discussion moved to its natural extension, namely that moraic-onset geminates may appear at the other possible position, that is, word medially. Due to the scarcity of data, the basis of the discussion was drawn on more speculative ground, using data from Marshallese, Trique, Djeebbana, Puluwat and Eastern Popoloca. The remainder of the chapter shifted onto more theoretical issues that involve the predictions the proposal makes. In particular, a factorial typology of onset geminates was constructed, while significant space was devoted to explaining why and how the geminate–singleton contrast bears on weight and not on length. This is non-trivial; although it is a basic assumption of moraic theory, it has to be justified – at least for the purposes of the current study – to answer any possible objections or alternatives that bear on the assumption that geminates are long rather than weightful. The simple ranking NoCoda >> *Moraic Onset generates onset geminates by demanding their syllabification wholly in the onset as a means to avoid the creation of closed syllables. The reverse ranking produces ‘flopped’

Onsets and geminates

201

geminates, which manage to avoid introduction of moraic onsets. Medial onset geminates may be rare, but this can be seen as a direct consequence of the general avoidance of moraic onsets that the overwhelming majority of languages exhibit. Their occurrence is thus scarce, but much commoner than previously thought. The empirical facts presented in this chapter testify to this.

6 Other real and not so real onset-sensitive data: brief case-studies 6.1

Aims

In the previous chapters, I have been arguing for the existence of onset moraicity by investigating the interaction between onsets and weight. Several casestudies have been explored to this end. In this chapter, this effort is continued, but takes a slightly different route. In particular, I examine a range of languages that fall within two broad categories. The first refers to languages whose data point to a moraic-onset analysis, but where the available evidence ranges from instances where arguing for onset weight is pretty straightforward, but the data are too sparse or too ambiguous to reach a safe conclusion, to instances that are quite controversial, but still indicative of some onset-weight effect (§6.2). The second category comprises languages which have in the past received analyses that made explicit or implicit use of moraic onsets, but, as I show, this is either misguided – in the sense that such an analysis is not strongly supported – or even entirely incorrect (§6.3). What, then, is the aim of this chapter? In fact, it is two-fold; first, it raises awareness of several additional languages which are suggestive of moraic onsets. Unsurprisingly, these are typically hardly documented. By addressing them and pointing out the questions they pose, it is hoped that future additional documentation and investigation should be able to clarify their status. At the same time, this chapter offers tentative discussion of other weight-based phenomena that potentially relate to onsets, the evidence of which, however, is not strong enough to merit analysis in their own separate chapter. These include tone (§6.2.3), reduplication (§6.2.4) and poetic metre (§6.2.5.2). The second aim may seem rather peculiar in a book that supports onset moraicity, because at first glance it seems to promote the opposite effect. More specifically, the action taken is to examine languages which have been falsely or less convincingly argued to present themselves as candidates for onset moraicity. I contend that this is equally important; besides attempting to clear the waters regarding real onset-moraicity cases, it aspires to present some of 202

Other real and not so real onset-sensitive data

203

the factors that can cause confusion and lead us to wrong conclusions with regard to the onset-moraicity status across languages. 6.2

Additional phenomena and/or languages indicative of moraic onsets

Discussion here begins with additional instances of compensatory lengthening interacting with onsets (Ch. 3), confirming that this is a less infrequent phenomenon than usually believed. More information is provided about Onondaga (§6.2.1.1) and less for other rather marginal cases (§6.2.1.2). Subsequently, very brief reference to an additional case of onset-sensitive stress is made (§6.2.2) before moving to the exploration of onset effects on tone. Based on the discussion throughout the book, some interesting predictions are made, plus justification is given as to why the interaction between the two is extremely rare (§6.2.3). Next, onset-weight effects with regard to reduplication are considered. As shown, in languages like Bellonese (§6.2.4.1) and Marshallese (§6.2.4.2), moraic onsets in the shape of geminates satisfy reduplication needs. The section closes by investigating three, this time more controversial, examples of onset-sensitive weight phenomena that involve word minimality (§6.2.5.1), some kind of CL (§6.2.5.3), and, interestingly, poetic metre (§6.2.5.2). Information on these cases is, however, insufficient, thus they should be treated with caution. 6.2.1 Compensatory lengthening Compensatory lengthening (CL) has been extensively discussed in Chapter 3. It has been shown that CL after onset loss is not only possible, but actually is the only way we can analyse data such as the phenomena in Samothraki Greek. Onsets, however, act not only as triggers for CL, but also as targets, in the sense that their gemination, i.e. their gain in moraicity, serves to compensate for the loss of neighbouring material. Languages that support this view have already been presented, namely Trique (§3.3.4), Pattani Malay (§5.2.1) and Trukese (§5.2.2). In this section we will look at a few more examples, where onsets cause CL, as in Onondaga (§6.2.1.1), and numerous other minor cases (§6.2.1.2). 6.2.1.1 Onondaga Onondaga (Michelson 1988) resembles Samothraki Greek (see Chapter 3), in that deletion of r results in lengthening. The facts are quite complex, so I will present a brief overview of the major points. Onondaga is one of the

204

Onsets

five Lake-Iroquoian languages, the other four being Seneca, Cayuga, Oneida and Mohawk. Onondaga and Seneca lost the phoneme r of the Proto-LakeIroquoian (PLI) in all environments, whereas Cayuga lost it intervocalically and after a postvocalic laryngeal. The discussion here centres only around Onondaga, unless reference to the other languages is needed. Intervocalically, r became w after round vowels (1a) and y after i (1b). In all other occasions r was lost (1c). The latter case resulted in a sequence of vowels, but no lengthening appeared.1 (1)

Intervocalic r-deletion or change to a glide without lengthening in Onondaga a. PLI *yothóːreʔ b. PLI *owíːraʔ c. PLI *waʔknóhareʔ

‘it is cold’ ‘baby’ ‘I wash’

Onond othóːweʔ Onond owiːyæʔ Onond waʔkoháeʔ

More interesting is the pre- and post-consonantal r loss. In both cases, r deleted and caused lengthening of the vowel before it (pre-consonantal r, e.g. PLI: *katóryeʔs ‘I’m breathing’, Onond: katóːyeʔs) or after it (post-consonantal). The latter is where we focus our attention. (2)

Post-consonantal r-deletion and lengthening in Onondaga PLI *yotshíʔkreʔ ‘clouds’ PLI *óʔkraʔ ‘snow flake’

Onond Onond

otshíʔkeːʔ óʔkæːʔ

Lengthening did not normally occur after *hr or *ʔr. Michelson states that the derived vowels are truly long and tautosyllabic since they are counted as one vowel for stress purposes. Seneca, on the other hand, presented r-deletion in the same environment without lengthening. Citing Woodbury (1981), Michelson (1988) proposes to explain the Onondaga pattern based on two facts. First, in dictionaries of the seventeenth and mid-eighteenth centuries, *Cr is frequently represented as Cer. Second, Mohawk regularly inserts an epenthetic e to break up *Cr clusters. One could then assume that Onondaga at some point inserted an epenthetic e too, which after r-deletion would coalesce with the following vowel, leading to a fused long vowel, i.e. *CrV > *CerV > *CeV > *CVV > *CVː (see Hayes 1989). While this is a possible explanation of the facts, it is not the only one. Onondaga’s r-deletion patterns seem to be consistent with what happens in SamG, i.e. lengthening of a vowel after cluster simplification, but not when the r deletes word medially.2 With regard to the information the dictionaries offer, numerous questions can be raised, the most pertinent one being why 1 2

If the vowel following r is back, then it is fronted, e.g. a → æ as in (1b). Onondaga differs from Samothraki Greek in also presenting deletion of r from the coda with subsequent lengthening. With respect to Onondaga initial r, Michelson (1988) reports that

Other real and not so real onset-sensitive data

205

it is that only some of the *Cr forms are represented as *Cer. Had this pattern been so pervasive, we would expect it to show up in all relevant forms. More importantly, beyond the information we receive from the dictionaries, no other conclusive empirical evidence exists to certify that indeed there was e-epenthesis in the *CrV context. Its occurrence in Mohawk does not imply that it also happened in Onondaga. Unlike Onondaga, Mohawk has extensive e-epenthesis. As a matter of fact, Michelson herself (1988: 146–7) describes just a single case of e-epenthesis in Onondaga, namely between an extrasyllabic non-laryngeal consonant and a following consonant (3). (3)

Onondaga e-epenthesis k-theʔt-haʔ > ke̲théʔthaʔ k-ʔny-a-ʔke-h > ke̲ʔnyáʔkeh

‘I’m pounding’ ‘my hand’

All these facts should at the very least make us sceptical towards the *Cer approach. A compensatory lengthening approach along the lines of Chapter 3 is also available, especially given the resemblance of the facts to the Samothraki Greek data. 6.2.1.2 Marginal cases Other minor cases of onset CL that usually occur with specific morphemes are mentioned in Davis (1985), Loporcaro (1991) and especially in Rialland (1993) and Beltzung (2007). In Maasai (Nilotic, Kenya; Davis 1985 and references therein), when intervocalic onset k deletes, lengthening of the following vowel occurs. In Romanesco Italian (Loporcaro 1991: 280), a highly morphologized case of non-synchronic CL occurs. Optional loss of the initial l of the definite article /lo la li le/ (masculine singular, feminine singular, masculine plural and feminine plural) and of the object clitic /lo la li le/ (him, her, them masculine, them feminine) when followed by an unstressed vowel leads to its lengthening, e.g. [lo stúpido] > [oː stúpido] ‘the stupid-masc.’ or [la brúːʃo] > [aː brúːʃo] ‘I burn her’. Beyond this environment, such CL does not appear. The examples reported next come from Beltzung (and references therein; glosses are not translated into English). The first example involves Ntcham, a Gur language of Togo, where deletion of the onset of the nominal suffix -ku leads to lengthening of the preceding vowel. (4)

Ntcham CL (Podi 1995: 345) (tones are omitted) /ku-sa-ku/ /ku-fe-ku/ /ku-mo-ku/

[kusaːu] [kufeːu] [kumoːu]

‘le champ’ ‘l’ecaille’ ‘l’herbe’

In Anuak/Anywa (Nilotic, Sudan), roots ending in c, k or glides systematically lose their final consonant when a vowel-initial suffix follows. If the vowel stem-initial r alternates with y/Ø, but, at least in the examples provided, this does not coincide with the word-initial position.

206

Onsets

preceding the – now lost – consonant is already long, then nothing else occurs (5a). But if that vowel is short, then it gets lengthened, apparently an instance of onset CL (5b). (5)

Anuak (Reh 1996: 90)

(tones are omitted)

a. Consonant deletion without lengthening /kiːc–ɛ/ [kiːe] ‘orphelins’ /a-paːk-ɛ/ [apaːe] ‘il l’a affilé’ /a-kiːw-a/ [akiːa] ‘je l’ai ramé’ /lʌːy-i/ [lʌːi] ‘animaux’ b. Consonant deletion with lengthening /kac-ɛ/ [kaːe] ‘moissons’ /a-bʌc-ɛ/ [abʌːe] ‘il l’a enveloppe’ /a-bʌk-ɛ/ [abʌːe] ‘il a bouilli’ /tuy-ɔ/ [tuːo] ‘germination’ /tɛw-ɛ/ [tiːɛ] ‘courants’

Other languages work similarly. In Turkana (Nilotic, Kenya; Dimmendaal 1983), there is a restriction against sequences of glides and [−ATR] vowels. To resolve structures of this type, deletion of the glide occurs making the following vowel long, e.g. [a-kɪ-yɛp] → [akɪɛːp] ‘couper’. In Gyore (Voltaic; Rialland 1993), the initial g vocalizes intervocalically or even gets deleted. When this happens, lengthening of the first vowel is observed, e.g. /ru-gɔ/ → [ruː(w)ɔ] ‘house’. In Lango (Nilotic, Uganda; Noonan 1992), the liquid of the agentive prefix /-là/ deletes intervocalically inducing lengthening of the previous vowel, e.g. /pé-là-bә̀ŋә́/ → [péːàbә̀ŋә́] ‘this is not a stupid person’ vs. /pé-á-bә̀ŋә́/ → [péábә̀ŋә́] ‘this is not stupidity’. These examples have demonstrated instances of CL where the onset is a trigger for lengthening and the latter only occurs with certain morphemes. Some caution, then, is required to investigate to what extent morphology plays any role. For instance, it may be the case that this phonological lengthening is employed to retain some morphological information. Even so, however, it is telling that the phonology actually chooses the strategy of lengthening after onset loss for that reason. If phonology were not to admit such a possibility – as suggested by standard moraic theory – then we would no longer have any explanation for data like these. 6.2.2 Stress Stress has been thoroughly discussed in Chapter 2. Here, I wish only to draw attention to a case that is consistent with the theory presented before, but for which I had no access to the original source at the time of writing. This is Tümpisa Shoshone (Dayley 1989, cited in Gordon 2005: 601), where

Other real and not so real onset-sensitive data

207

‘peninitial syllables with a voiceless onset only optionally attract stress away from an initial CV syllable, whereas a long vowel in the second syllable consistently attracts stress from initial CV’. 6.2.3 Tone Another so far undiscussed effect of onset weight is that on tone. This is in fact especially interesting due to a prediction the current model makes; in particular, since I have argued that moraic onsets exist, and since moras serve as tone-bearing units (TBUs) in several languages (see Yip 2002 for examples and p. 72–6 for specific discussion), a natural expectation would be to find moraic onsets that carry tone. I will argue that while this is possible, there are good reasons why it is vanishingly rare. It will become obvious that exploration of this issue largely relates to tone–consonant interactions, a topic that is complex and far from straightforward, as the vast literature on it suggests (e.g. Halle and Stevens 1971; Yip 1980, 1995; Bao 1990; Duanmu 1990; Peng 1992; Bradshaw 1999; Tang 2008, and many others). For this reason, at present I will only touch upon it and make some general statements, which should merely serve as a point of departure for further research. Data from Kpelle, a Mande language spoken in Liberia, form the empirical basis of the core discussion. (6)

Kpelle onsets as TBUs a. Hyman (1985: 44): i. initial obstruent Stem ‘my’ ‘his/her’ pólù ¸bólù b̀ólù ‘back’ túέ ńdúέ d̀úέ ‘front’ kɔ́ɔ́ Ãgɔ́ɔ́ g̀ɔ́ɔ́ ‘foot, leg’ fíí ¸víí v̀íí ‘hard breathing’ ii. initial sonorant Stem ‘my’ ‘his/her’ lēē ńēē 9ēē ‘mother’ yéé ɲ́éé ɲ̀éé ‘hand, arm’ mā̰lóŋ ¸ā̰lóŋ m̀ā̰lóŋ ‘misery’ ɲḭ́ŋ ɲ́ḭ́ŋ ɲ̀ḭ́ŋ ‘tooth’ b. Welmers (1962: 72): intervocalic obstruents kâpa ‘penny’ b̀ε̂bε ‘raffia purse’ g̀b̀εtε ‘fix it’ g̀b̀ôdo ‘leprosy’

In (6a), we see that for the production of the ‘my’ form, a high-toned nasal is prefixed to the stem. If the stem starts with an obstruent, then voicing assimilation

208

Onsets

also occurs (6a.i), but if it begins with a sonorant, then total assimilation and nasal simplification occurs (6a.ii). (NB: I am glossing over the specifics of these processes, as my focus here is on tone and onsets.) The ‘his/her’ forms appear with either an initial low-toned voiced obstruent or with a low-toned nasal. In light of the above, and due to the existence of minimal pairs such as (7), where a sonorant onset can appear toneless, L toned or H toned, Welmers proposes that in the stem – ‘his/her’ forms of (6a), the underlying distinction for [m]–[m̀] is one between /m/–/ ˋm/, while for word-initial [p]–[b], it is one between /p/–/ ˋp/. (7)

mare-kε̂i ‘a question’ m̀are kέ ‘ask him’

¸are kέ ‘ask me’

Moreover, Welmers proposes that the word-medial voicing contrast for obstruents shown in (6b) provides evidence that there is also an underlying /b/ that occurs word medially. His proposal can thus be summarized as: (8)

Phonemes and allophones according to Welmers a. /p/ /p/ / ̀p/ b. /b/ /b/

→ →

[p] word initially, e.g. pólù [b̀] word initially, e.g. b̀ólù

→

[b] intervocalically, e.g. b̀ε̂bε

However, I would like to reinterpret these contrasts in a slightly different way so that actually there is only one phonemic obstruent, and all others are derived allophonically. This can be supported by at least two pieces of evidence. First, Welmers himself notes that the initial [b] is heavily voiced and begins with low pitch, which is consistent with the fact that it derives from underlying / ̀p/ with a floating low tone. On the other hand, the intervocalic [b] is not heavily voiced and does not begin with a low pitch. This might suggest that in fact it is not underlyingly voiced, but it may simply derive from plain voiceless and toneless /p/, which shows up as slightly voiced due to ambient voicing. Second, Welmers does not fully discuss what happens with the underlying representations of the sonorants presented in (7). Following his reasoning in (8a), the obvious thing to say would be that [m] comes from /m/, [m̀] from / ̀m/ and [¸] from / ́ m/. Adding all these together, we get: (9)

Phonemes in Kpelle and their allophones – present proposal: a. Obstruents, e.g. /p/ / ́p/ → [p] word initially, e.g. pólù /ˋp/ → [b̀] word initially, e.g. b̀ólù /p/ → [p] intervocalically, e.g. b̀ε̂bε

Other real and not so real onset-sensitive data b. Sonorants, e.g. /m/ / ́m/ → [¸] / `m/ → [m̀] /m/ → [m]

209

word initially, e.g. ¸arε word initially, e.g. m̀arε word initially/medially, e.g. marε / d̀ámaa

The difference between this and Welmers’ suggestion is that while he imposes two phonemes for surface [b], I propose only one, namely /p/, which is manifested in various ways, depending on the presence of floating tone. The second point is that I also allow for an underlyingly H-toned obstruent – in analogy to sonorants – whose tone, however, fails to surface, because, as I will argue next, voiceless obstruents cannot be surface TBUs. Assuming this is right, then we would need to say that this high tone either deletes, remains afloat or shifts to a neighbouring segment. The final possibility, although appealing, is very difficult to test, since there is no agreement among researchers on some of the facts or the exact tonal association. For instance, while Welmers (1962: 75) transcribes the stem of the word meaning ‘back’ as [pôlu] with a contour tone, Hyman (1985: 44) gives it as [pólù]. As Hyman also acknowledges, the forms in (6) have caused a ‘heated debate’ over the correct transcription of the facts and their analysis, thus it would take us too far afield to provide a fuller analysis at this stage. Notably, though, the fact that certain onsets in Kpelle appear as tone bearing is undisputable. This can naturally be accommodated in a framework that admits moraic onsets, which consequently can act as TBUs. Two theoretical issues now need to be addressed. First, why is the phenomenon of onsets as TBUs so remarkably rare? I will argue that this is because TBUs and moraic onsets pose conflicting demands. And, second, why have I claimed that voiceless obstruents cannot be surface TBUs? This is because voiceless obstruents do not satisfy the phonetic conditions required for TBUs. The remainder of this section elaborates on these ideas. I will start the discussion by considering moraic onsets – the presence of which is quite rare already – and what constitutes the best moraic onset. As I have argued in §1.3.3, the preferred moraic onsets are the voiceless (obstruents) as well as the sonorants lacking [+voi] due to the fixed ranking: *μ/Ons/[+voi] >> *μ/Ons. Thus, if a language only possesses a single type of moraic onset, then it will be the voiceless obstruents, assuming that sonorants are [+voice] (10①), as in Pirahã due to *μ/Ons/[+voi] >> Moraic >> *μ/Ons. The same ranking, but with the added proviso that sonorants lack [+voi], generates a language where both voiceless obstruents and sonorants are moraic onsets (10②), like in Karo. Finally, the ranking Moraic >> *μ/ Ons/[+voi] >> *μ/Ons entails that all onsets will be moraic, i.e. voiceless

210

Onsets

obstruents, sonorants and voiced obstruents (10③). There is thus an implicational relationship, depicted below. (10)

Implicational relationship for moraic onsets Voiced obstruents > Sonorants > Voiceless obstruents 1 2 3

Consider what the consequences of these patterns are for tone. Since the physical correlate of tone is fundamental frequency, only voiced segments should be able to carry it, i.e. vowels, sonorants and voiced obstruents. The harmonics of voiced obstruents are markedly low in energy compared to those of sonorants, and as Gordon (2006: 88) states, voiced obstruents are expected to ‘contribute extremely little to the ability of a syllable to carry a contour tone or not’. More generally, it is not that voiced obstruents are incapable of participating in tonal assignment; they may do so, but only to a very limited extent. It thus follows that while a language may be diagnosed through weight-based phenomena as treating all segments – including obstruents – as moraic, only a subset of these may actually be TBUs, an idea originating in Steriade (1990). To see this more clearly, let us consider languages where CVV=CVC=2μ, i.e. where the coda is moraic, and focus on what the TBU can be. In many languages (11①), only vowels in nuclei receive tones, as in Ancient Greek (Steriade 1990). In others, such as Danish (Zec 1988; Steriade 1990), Early Greek (roughly equivalent to Homeric Greek, Steriade 1990), Lithuanian (Zec 1988; Steriade 1990; Gordon 2006), Kiowa (Gordon 2006) and Kunama (Connell, Hayward and Ashkaba 2000), the coda in CVC syllables is moraic no matter what its quality, but only sonorant codas can bear tone alongside the vocalic TBUs of nuclei (11②). Finally, in Hausa, Musey and Luganda, Gordon (2006) finds that not only sonorants and vowels are TBUs, but also the much rarer (voiced) obstruents (11③).3 The picture thus shaped is presented in (11a). The situation described in (11b) is discussed next.

3

I discuss Musey in more detail below. As for the other two, Hausa admits contour tones on CVC syllables with obstruents, e.g. râsːáː ‘branches’ (Gordon 2006), which, like Musey, could suggest that voiceless obstruents phonologically act as TBUs. In Luganda (Ashton et al. 1954), voiced and voiceless geminates seem able to carry tone, but data are too sparse to reach any conclusion. Notably, these are characterized as syllabic, so perhaps an analysis along the lines of sesqui-syllabic languages is in order (see main text below).

Other real and not so real onset-sensitive data (11)

211

Implicational relationship for TBUs a. Voiced obstruents > Sonorants > Vowels 1 2 3

b. Voiceless obstruents are not (surface??) TBUs universally

Now, the idea is that, since voiceless obstruents cannot be TBUs, it should be impossible to find tone on moraic onsets in a language with pattern (10①), where only voiceless moraic onsets are admitted. Recall that since (10①) is the default instance of moraic onsets, by eliminating this pattern, the likelihood of moraic onsets as TBUs is automatically hugely reduced. We are thus left with patterns (10②) and (10③). In (10②), both sonorants and voiceless obstruents are moraic, but due to the inability of the latter to be TBUs, only the former can assume this role. Thus, we can predict that it should be possible for a language to have the moraic-onset pattern of (10②) and the TBU pattern of (11②), where only nuclei and moraic sonorant onsets receive tone. In (10③), where all segments can be moraic onsets, the prediction is that TBUs can either be of the type in (11②), i.e. nuclei and sonorant onsets, or of the type in (11③), where, as well as vowels in nuclei, sonorant and voiced obstruent onsets can also be TBUs. Kpelle above seems to be an instance of the latter pattern (Hyman 1985; Welmers 1962). However, recall that in Kpelle I had argued that voiceless obstruents underlyingly may carry tone (9a), although they might surface as toneless. This relates to (11b), where I question the universality of the ban on voiceless obstruents as TBUs. To this end, consider Musey (Shryock 1995), cited in Gordon (2006), where consonants are divided into Type A (or High consonants) and Type B (or Low consonants). In the absence of lexical tone, word-initial Type A consonants induce mid tone on the first vowel, whereas Type B trigger low tone. Roughly, Type A consonants include the sonorants and the obstruents that historically used to be voiceless. Type B correspond to the historically voiced consonants. Processes such as rightward displacement of lexical L tone (12) suggest that the contrast between Type A and B consonants is genuine and causes tonal differences. If the mid tone after Type A consonants is treated as being introduced by them, then under the assumption that tones attach to moras, we would need to say that voiceless obstruents were at least at some point phonologically TBUs, although phonetically they are not (since the tone surfaces on the neighbouring vowel). Alternatively, as Shryock suggests, the mid tone may be assigned by default. But in this view too, something special has to be said

212

Onsets

about why only Type A consonants cause tonal displacement, i.e. why we do not get *ɦūnà too. (12)

Rightward displacement of lexical L tone in Musey a. Cliticization of /-na/ Type A: sà → sanà → sānà ‘person’ Type B: ɦù → ɦùnà ‘goat’ b. Subjunctive Type A: tò ‘sweep’ Type B: dò ‘pick’

Subjunctive with affixation tōm̀ ‘sweep it’ dòm̀ ‘pick it’

This then raises the issue about the possibility of distinguishing phonological from phonetic TBUs. If this proves necessary, then perhaps voiceless obstruents can be phonological TBUs, albeit not phonetic ones. This latter property explains why Kpelle / ́p/ is underlyingly admitted, although it fails to surface as such. Things get more complicated in the so-called sesqui-syllabic languages (Sloan 1988; Gafos 1998; Lin 1998b; Hendricks 2001) like Kammu (Svantesson 1983). In those languages, many words consist of a full syllable and a ‘minor syllable’ that usually precedes it. A minor syllable typically consists of an onset and a coda. The latter serves as the syllabic element which, in some sesqui-syllabic languages, can also carry tone. Commonly, the coda includes a sonorant, e.g. [t¸.ràʔ] ‘stove’, [kr̀.làŋ] ‘bulge’, but it may also contain a voiceless obstruent, which, despite expectations, can nonetheless carry tone, e.g. [pk̀.ték] ‘to tell a riddle’, [kś.rìs] ‘to shake something out’. In the light of these facts, it can be argued that the universal ban on surface voiceless obstruent TBUs does not hold either. However, the status of minor syllables remains elusive (see Lin 1998b for a brief overview of proposed accounts), and their special behaviour is unquestionable, given that they: are prosodically defective, occur at word edges, cannot bear main stress, and can be insensitive to stress assignment, etc. In fact, under one interpretation of minor syllables, the second consonant can be considered the coda of a syllable with an empty moraic nucleus, i.e. [kś.rìs] from above is actually [kә́s.rìs] (Gafos 1998 on Temiar), or the coda of a reduced vowel, i.e. [kә́s.rìs] (Coleman 1996 on Tashlhiyt Berber). If any of these is on the right track, then the (empty) nucleus, instead of the – sometimes voiceless obstruent – coda, would carry the tone, thus eliminating the problem above. I will thus adopt a position along these lines, and maintain that voiceless obstruents are never surface TBUs universally. In sum, it is anticipated that moraic onsets as TBUs should be an extremely rare phenomenon, given that onset moraicity and tone essentially impose

Other real and not so real onset-sensitive data

213

contradictory requirements. The preferable moraic onsets are the voiceless obstruents which lack F0, whereas tone needs to dock onto segments with fundamental frequency. Thus, for onsets to act as moraic TBUs, several conditions must be satisfied at once. This is possible only in highly restricted environments. In this view, then, their extreme rarity falls out naturally. 6.2.4 Geminates in reduplication and elsewhere In our discussion of Pattani Malay, we briefly mentioned the possibility that reduplication is expressed by means of onset consonant gemination (§5.2.1). Some languages, however, utilize this strategy in a more systematic way, as presented next. 6.2.4.1 Bellonese In Bellonese (Melanesian, Solomon Islands; Elbert 1988), syllables are of the (C)V(V) type, thus no codas are permitted. Interestingly, whenever reduplication is involved, all consonants can geminate in fast speech (NB: the reduplicant is shown underlined). (13)

Bellonese long consonants in reduplication (Elbert 1988: 17–18) Slow speech

Fast speech

Gloss

babange bebete hahatu ka kata lolongi momoe nanamu ngangaha pipiki sasaka tatau

bbange bbete hhatu k kata llongi m moe nnamu ngngaha ppiki ssaka ttau

‘to play’ ‘to untie’ ‘to fold’ ‘to laugh’ ‘weak’ ‘to lie down’ ‘to smell badly’ ‘to measure’ ‘to keep’ ‘to beg’ ‘to tattoo’

Elbert seems to attribute this phenomenon to the slow vs. fast speech distinction and thus implies that the geminated form is somehow generated after the deletion of the vowel of the CV reduplicant. To corroborate that, he claims that similar losses occur word medially in some non-reduplicated words. Nonetheless, I contend that a more interesting explanation of the facts is the following: the template for reduplication is merely a mora. Speakers can achieve this in two ways. They can either produce a CV syllable or they can

214

Onsets

geminate the first base consonant. While a geminate consonant is ideal in the sense that it is really the most minimal form one can get for mora reduplication, it introduces geminates which are quite marked. A CV-reduplicant avoids this problem at the expense of constructing a slightly larger form. Two pressures are then in conflict producing variation: on the one hand, the preference for minimum reduplicant size and on the other the avoidance of geminates. Depending on which of the two takes precedence, the above results obtain. There is another reason why this analysis seems to be on the right track. Elbert notes that this kind of pattern systematically arises only in reduplication. A very limited number of similar cases emerge outside reduplication, e.g. ghaghaghaba ~ ghghaghaba ‘a plantain’.4 In fact, Elbert offers only two more examples of the same type. Had the pattern described in (13) been purely related to the slow–fast speech difference, then why would it regularly and systematically only arise in reduplication? And why would it always and only involve deletion of the first vowel, given that word-medial vowel deletion also occurs, e.g. ghaghaghasa ~ ghaghghasa ‘a limpet’? The truth is that vowel deletion happens extremely marginally outside reduplication and seems random. The reduplicated pattern is, however, suggestive of a generalization which can much more adequately be attributed to initial-consonant gemination that involves construction of a moraic onset – a configuration also supported by the prohibition against codas. Bellonese then renders itself a compelling case of reduplication by means of initial-consonant moraic gemination. 6.2.4.2 Marshallese In §5.3.1, we examined medial geminates in Marshallese and their effect on stress. However, one of the dialects of Marshallese – the Ralik one – provides additional support for the moraicity of onset geminates through reduplication (Abo et al. 1976). Discussion of these data was postponed till now for two reasons: first, the focus in the previous chapter was on medial geminates, and the stress data most clearly highlighted their effect, and second, it was considered best to group the reduplication data together here, rather than introduce them earlier. With this much in mind, we may consider the Ralik data, which present an interesting case of consonant doubling to form the distributive. The distributive is an important category in Marshallese, which is very productive and covers a constellation of semantic distinctions (for discussion, see Bender 1991). Most distributives are formed by consonant doubling and final syllable reduplication, e.g. koto → kkototo ‘windy → always be windy’. However, these 4

Note, however, that despite Elbert’s claim, this still looks like reduplication.

Other real and not so real onset-sensitive data

215

two processes seem to be independent of one another since, as Bender notes (1991: 15), words which already present initial consonant doubling can form the distributive by having final copying, and those which already have final reduplication may instead present gemination only. The current discussion will thus address only the facts relevant to gemination. (14)

Marshallese distributive – Ralik dialect5 Root

Ralik

Gloss

korap tumej panuk bale diylah jekapen mede nib lokjak reja

yokkoraprap yuttumejmej yeppanuknuk yebbalele yiddiylahlah yejjekapenpen yemmedede yinnibnib yollokjakjak yerrejaja

‘gecko’ ‘open eyes under water’ ‘pile up, gather’ ‘a type of fish’ ‘nail’ ‘less than half full’ ‘young coconut meat’ ‘preemptive’ ‘be busy with’ ‘shave (from Engl)’

Obviously, ‘the reduplicant … surfaces in onset position of a syllable’ as Hendricks (1999: 36) also observes. Interestingly, in the Ratak dialect the reduplicant appears instead as a simple CV-copy, e.g. betah → bebetahtah, diylah → didiylahlah, etc. The difference between the dialects with regard to the formation of distributive reduplication is exactly analogous to reduplication in Bellonese (§6.2.4.1), i.e. reduplication of a mora. Ratak resembles slow speech in Bellonese, whereas Ralik corresponds to Bellonese fast speech in realizing reduplication through the use of a moraic onset. One more thing needs to be addressed at this point. The Ralik distributive is preceded by an epenthetic yV- prefix, whose vowel quality is identical to that of the base vowel, unless this is /a/, in which case the epenthetic vowel is /e/. Note that without the prefix for a base like reja we would expect *rrejaja instead of attested yerrejaja. This can be straightforwardly accounted for if we claim that Marshallese bans initial geminates. Unsurprisingly, this seems sensible since most languages prefer medial to initial geminates. While such an assumption accounts for the presence of a vowel to offer the required distance of the geminate from the word-initial position, it offers no insight as to 5

Following the practice of Hendricks, I present the reduplicated data not with a phonemic transcription, but with their spelling as presented in Abo et al. (1976). Apart from being a harmless simplification for current purposes, it is also preferred, as Abo et al. (1976) do not always give the phonemic transcription, so one would sometimes need to guess what this would really be.

216

Onsets

why a glide must also precede it. Presumably, this relates to a requirement that syllables must have an onset. Moreover, the yV- epenthetic prefix seems somewhat external to the reduplicated form, corresponding to no input morphological material, thus yV- is not part of the morphological word of the reduplicated form (Hendricks 1999; van Oostendorp’s 2005 ‘Coloured Containment’). In this view, yerrejaja then corresponds to ye[r-rejaja]Word. The presence of the prefix is then justified by phonological reasons only, i.e. avoiding an initial geminate and ensuring an onsetful syllable. In sum, then, we have seen that both stress and reduplication plead for an analysis that treats Marshallese geminates as medial moraic onsets. 6.2.5 More controversial cases Data in other languages are more perplexing, as they provide some evidence which favours an analysis that utilizes moraic onsets, but at the same time they either require additional assumptions in order to be maintained (e.g. Damin or Luganda), or may conflict with other facts in the language (e.g. Supyire). Nevertheless, they are mentioned here for completeness. 6.2.5.1 Word minimality in Damin Damin is a now-extinct initiation language used by the Lardil (Mornington Island, Gulf of Carpentaria, Australia (Hale 1981b, henceforth H;6 Hale and Nash 1997, henceforth H&N). It is phonologically interesting for many reasons, one being that it is the only language outside Africa that possesses clicks and shows rearticulation of some consonants. For our purposes, Damin is interesting, because according to H&N, it possesses a word minimum of CVV or CCV words. On the other hand, CVC minimal words are banned, although Damin allows closed syllables with /n/ or /rr/ as a coda (/rr/ is an apico-alveolar flap). On the whole, there is preference for syllables to be open (H&N: 254). For expository reasons, I present the consonantal inventory that Hale and Nash offer, without attempting to justify it. (15)

Damin consonantal inventory

stops nasals flaps 6

bilabial

laminodental

apicoalveolar

b

th

d n rr

apicodomal

laminoalveolar

dorsovelar

j ny

k ng

Many thanks to David Nash and to ASEDA, AIATSIS for making document 0028 available to me.

Other real and not so real onset-sensitive data

bilabial glides fricative ejectives nasals

laminodental

w f pf p’ m!

apicodomal

laminoalveolar

dorsovelar

y j2 k’ nh!2

lateral trill

apicoalveolar

217

n! n!2 l*

(rn!) ng*

pr pr2

The interested reader is referred to H&N: 252 for the detailed IPA transcriptions of these sounds. For current purposes, it suffices to say that the symbol ‘2’ indicates rearticulation, ‘*’ that the sound is voiceless, and ‘!’ signifies a standard click at the place of articulation denoted by the previous consonantal symbol. All clicks are nasalized. Intervocalic clusters are limited, but there are some. Onsets include singleton consonants, rearticulated consonants or certain clusters. In particular, the latter involve: a) the bilabial /f/ or /p’/ followed by /ny/ or /ng/; b) the bilabial /f/ or /pr2/ followed by /y/; c) the cluster /thrr/. On a highly speculative note, Hale and Nash (1997), following a suggestion by Morris Halle, propose that in fact all complex onsets in Damin, including the rearticulated ones, have to be analysed as ‘the double association to onset of a single segment’. In other words, they are treated as geminate consonants in their traditional representation (see Ch. 5). This is consistent with the fact that minimal words either contain a long vowel or are of the form CCV,7 where CC corresponds to consonant clusters or rearticulated consonants. Both types of words are treated as bimoraic. Hence, the onset is allowed to contribute a mora, a possibility that Hale and Nash (1997) themselves acknowledge. Some examples follow. (16)

Damin minimal words (H&N and H)8 CCV: a) rearticulated consonants n!2a ‘grandmother, paternal’ (H) n!2u ‘liquid’ (H&N) nh!2u ‘become hot’ (H)

7

8

The examples of rearticulated consonants provided in the sources only include rearticulated clicks. While there are words starting with /pr2/ and /j2/, these are followed by VV instead, e.g. pr2yuu and j2uu (H). They were the only examples of this kind I could spot. Document 0028 (Hale 1981b) comprises a Lardil–Damin vocabulary and a Lardil–English one. The former provides no English glosses for words. I have provided the glosses indirectly by consulting the Lardil–English vocabulary. I usually provide a single gloss, since for some words multiple correspondents are applicable, e.g. n!2a also means ‘brother in law’.

218

Onsets b) complex onsets fngu ‘be bewitched’ (H) fnyi ‘calf of leg or tail’ (H) thrru ‘woman’ (H) CVV: jii ‘grandchild, daughter’s child’ (H) thii ‘shark (general term)’ (H) kaa ‘now’ (H&N)

Note that some – superficially – CV words are allowed too, but these always involve one of the following consonants: /l*/, /ng*/ or /k’/, such as l*i ‘good, well’ (H), ng*i ‘tobacco’ (H). Presumably, these consonants actually act as rearticulated. For instance, /l*/ has two sequentially opening airstream mechanisms, i.e. an ingressive voiceless lateral with eggressive glottalic release, while /k’/ and /ng*/ have ‘double effort’ in the airstream with /k’/ being an ejective and /ng*/ having an extra pulmonic pressure (H&N: 257–8). While the Damin data seem intriguing, we would need more information to understand the nature of the consonants involved, the possible clusters, as well as the properties of rearticulation. Unfortunately, the language is no longer extant, thus the discussion above will necessarily remain speculative. 6.2.5.2 Metrics: the case of Luganda court songs According to Fabb (1997), the Luganda court songs – part of the court music of the Buganda kings of Uganda – are metrical (for more details, see Fabb 1997: 100–2). In particular, the basic line consists of 36 pulses divided into six 6-pulse constituents matched by accompanying handclaps. The music pulses match with moras in the line. For our purposes, it is interesting to examine the mora-counting algorithm. All the following configurations count as two moras: i) a long vowel, ii) a vowel preceded by a glide, e.g. kwa, and iii) a short vowel followed by two consonants (usually a geminate or a nasal+obstruent sequence). The following line illustrates (NB: • = pulse, • = implicitly accented pulse accompanied by handclap). (17)

Line from the song of Ssematimba and Kikwabanga9 (Katamba and Cooke 1987)

Call part •

•• ••

Response part ••

•• • ••

•

••

•••• • ••

• • • ••

   

  







   

µ µµ µµ µµ µµ µ µµ

µ

µµ

µ

A-baa - lia - ban - gi# nsi- ga - dde# bw’o-mu# we were 9

many

I am left

alone

 

µ µµ

•

µ µ µµ µ

•• •

•• ••

•

    

µµ µ µµ µµ µ

La- ba# Sse-ma- ti-mba# ne# Ki-kwa-ba-nga# Look at Ssematimba

and

Kikwabanga

The syllable [ma] in Ssematimba is monomoraic but matched to two pulses, as a result of lengthening in that position. The monosyllabic word [ne] that follows is treated as bimoraic, because

Other real and not so real onset-sensitive data

219

It is evident that the line comprises two half-lines: the call part and the response part. A few pulses may remain unmatched if the half-line is short. Notably, unmatched pulses only occur at the end of a half-line, but never within it. The reader is referred to Fabb (1997) for more details on the analysis of this metrical template, but what matters for our purposes is that syllables with complex onsets like bw’o or kwa count as heavy for metrical purposes. Of course, one could argue that such syllables are truly bw’oː or kwaː, but I do not consider this likely, since Fabb is cautious in being explicit about all other vowels involved by marking them as long or short.10 I thus take this to really indicate a short vowel preceded by a glide, but being able to act as bimoraic. Nonetheless, note that it is also possible that the glides here are what Smith (2003) calls ‘nuclear onglides’, i.e. glides that are actually dominated by the nucleus node instead of the syllable node. As such, they could contribute to the weight of the nucleus and consequently not count for onset moraicity. To assess this possibility, further research would be required on the status of Luganda glides. However, even if this is so, such a proposal cannot be extended due to the following observation. Syllables with a short vowel followed by a nasal+obstruent stop or with a geminate, e.g. ti in ti-mba or ga in ga-dde, are considered bimoraic, perhaps due to resyllabification, i.e. tim-ba and gad-de, from the onset to the coda of the previous syllable. But this may not be necessary if we assume, as Fabb does above, that complex onsets may contribute moras for metrical purposes. Although the correspondence between pulses and moras may need to be slightly modified, the number of moras is not affected, and neither is the overall pattern. Luganda court songs, then, provide a potential case where onset moraicity may be recognized in metrical structure. Of course, this proposal really depends on the possibility that complex onsets bear weight (as opposed to simplex ones), an issue addressed later in §6.3.2–4. 6.2.5.3 Vowel elision and CL in Supyire In this section I describe a process of vowel elision in Supyire spoken in Mali (Carlson 1990) which, according to an unpublished manuscript by Hajek and Goedemans (in prep.), could be treated as involving onset weight. Throughout I will not mark tone.

10

the following consonant /k/ in Kikwabanga is pronounced [tʃ] and acts like the geminates or nasal+stop sequences (see text below), rendering the preceding syllable heavy. Note that Fabb’s interpretation of the facts conflicts with that of Clements (1986), where kwa is actually [kwaː] from underlying /ku-a/. This undergoes glide formation and lengthening of the following vowel. Similarly, timba is actually [tiːmba] from underlying /tim-ba/. Prenasalization follows with subsequent lengthening of the previous vowel.

220

Onsets

The facts of interest are the following (Carlson 1990: 67–8). Unstressed high vowels /i/ and /u/ delete when followed by a resonant consonant /l/ and /n/ (but apparently not when the resonant is a different nasal or an approximant /w/ or /y/) and preceded by either a non-coronal stop (/p/, /b/, /k/) or a non-back fricative (/f/, /v/, /s/, /z/).11 There are two options for the resulting surface form, either consonant gemination, i.e. C1V1C2V2 → C1C2ːV2 or lengthening of the following vowel, i.e. C1V1C2V2 → C1C2V2ː. Some examples follow. (18)

Supyire gemination/lengthening after V-loss [NB: R is a uvular tap (Carlson 1990: 21)] pilaga biliwe kile kinaga file silege

C-gemination plːaRa blːiwe klːe knːaRa flːe slːeRe

V-lengthening plaːRa bliːwe kleː knaːRa fleː sleːRe

Gloss ‘night’ ‘slave’ ‘sky, God’ ‘fruit bar’ ‘approach’ ‘be ashamed’

Hajek and Goedemans (in prep.) argue that both sorts of lengthening (consonantal or vocalic) are a result of the preservation of the mora left behind by the high vowel. Moreover, they interpret the consonant gemination as a moraiconset geminate, based on two statements of Carlson. First, Supyire has no codas: ‘There are no closed syllables in Supyire. Syllables are basically CV or CVV. A few grammatical words… begin with a V syllable’ (Carlson 1990: 16). As a result, the first part of C2 cannot be considered a coda. Moreover, Supyire has a single instance of syllabic nasals which arises after the deletion of an unstressed vowel in the environment NVC[−voi stop] (Carlson 1990: 65–6). If these statements are right, then the consonant in question cannot be syllabic or a coda-onset geminate. The only alternative then seems to be that the consonant is syllabified in the onset and bears a mora, as a result of position preservation through a mora. To show that this is indeed the case, evidence is presented relating to the 1Sg gender class suffix -wV which only attaches to bimoraic roots, such as CVCV sara-wa ‘bee’ or CVV tʃeː-we ‘woman’, but not to monomoraic ones. The fact that the suffix can appear after forms generated through elision, like blːi-we ~ bliː-we ‘slave’, corroborates the idea that the resulting root ends up bimoraic. But there are certain other details about Supyire’s phonology which might make us sceptical about the validity of the above proposal. While preservation 11

There are words which conform to this description and yet show no elision, see e.g. sululi or pinini (Carlson 1990: 231). Perhaps, the vowel in question is stressed and that is why elision does not apply.

Other real and not so real onset-sensitive data

221

through a mora indeed seems to be occurring, its characterization may not be accurate. First, note that this process only occurs when C2 is a sonorant /l/ or /n/, and not with any other consonants. Second, the sonorant retains the tone of the elided vowel. Third, given the examples Carlson provides, this process seems to occur only when the relevant sequence appears word initially. These three properties are consistent with either syllabic consonants or sesqui-syllables (also see §6.2.3 in relation to tone). The latter occur mainly in the Mon-Khmer languages, like Kammu, Temiar or Semai (Svantesson 1983; Sloan 1988; Gafos 1998; Lin 1998b; Hendricks 2001 and references cited therein), which possess a large number of words that include a sequence of consonants (usually two), dubbed a minor syllable (or semi-syllable) and a normal syllable, e.g. CC.CVC as in Temiar br.caaʔ ‘to feed’ (Gafos 1998: 238) or Semai dh.dnoh ‘appearance of nodding constantly’ (Sloan 1988: 320). The first CC part of these words is not really a syllable as it includes no nucleic part, but instead is usually considered as an onset–coda sequence with one mora (hence the name sesqui-syllabic, i.e. words of one-and-a-half syllables). Such degenerate semi-syllables are confined to word-initial position (Sloan 1988), and according to John Hajek (p.c.) this is reasonable given that these languages are very markedly stress final and all other syllables tend to be reduced. A final property of minor syllables relevant for our purposes is that commonly the second consonant carries tone, e.g. Kammu k9.níʔ ‘behind’, hŕ.màal ‘soul’ (Svantesson 1983: 32–3). Adding all these traits together makes minor syllables look suspiciously similar to the CC[son] sequences of Supyire. Consequently, we conclude that it is possible to view the geminated consonant being syllabified as C1C̩2.C2V2 so that C2 is either syllabic or a coda of a semi-syllable. The only reason we could not yet adopt either of these two ideas with any certainty is Carlson’s statements mentioned above. If the language lacks codas, semi-syllables should not be possible, at least not under the interpretation that views semi-syllables as onset–coda sequences. The other statement maintains that the language has syllabic sonorants, but only in a very specific environment which does not correspond to the one in (18). Here too then syllabic sonorants would not be an option either. To construct a more informed opinion about Supyire we would then need answers to the following questions: i) on what empirical grounds does Carlson establish that the only syllabic sonorants in the language are the nasals which arise after the deletion of an unstressed vowel in the environment NVC[−voi stop]? In other words, how can we tell with certainty that other sonorants cannot be syllabic (or part of a semi-syllable)? ii) are there absolutely no examples of the type: kapilaga → kapːlaga [i.e. with V-elision word medially]? Hajek

222

Onsets

and Goedemans seem to suggest that there are none, since onset gemination is today a historic process in the urban variety of Supyire; iii) are there any examples where C2≠l, n? And iv) are there any other instances of geminates in Supyire? Until we receive satisfactory answers to these issues, I contend that no decision can be made for one or the other approach. 6.3

Misanalysed onset-sensitive languages

Beyond the languages discussed so far, other researchers have claimed that onset-weight effects are attested elsewhere. This section addresses a few cases of that sort arguing that this conclusion is incorrect, simply because certain data and facts that prove instrumental have not been taken into account. I claim instead that these languages are overall either well behaved (in terms of stress: Bislama and Nankina) or slightly unusual (i.e. Berawan alleged onset geminates), but at any rate provide no evidence for onset moraicity. A brief overview is presented here. Detailed accounts can be found in Topintzi (2006b: §7.3). 6.3.1 Berawan Let us begin with Berawan (García-Bellido and Clayre 1997, henceforth GB&C), an Austronesian language spoken on the island of Borneo, in Sarawak, East Malaysia. The aspect of Berawan which is of interest to us is the existence of alleged geminate consonants, the presence of which is supported by acoustic studies. All consonants – plosives, affricates, nasals and liquids – apart from the fricative s, the approximants w and j and the glottals ʔ and h, can geminate.12 (19)

Geminates vs. singletons in Berawan (GB&C: 25) Geminate lƱtːɔʔ ʔɐtʃːɔh sɐnːɐɪʔ

Singleton ‘float’ ‘dog’ ‘heat of sun’

lƱtɔh ʔɪtʃoː sɐnɐɪʔ

‘soggy’ ‘remove’ ‘insect sp.’

Geminates, or ‘heavy’ consonants in GB&C’s description, have a very restricted distribution, namely they ‘can only occur as onsets of the nuclear syllable’ (GB&C 1997: 25), i.e. the final syllable. Such a syllable is called nuclear because of some of its special properties, e.g. it uniquely hosts the full 12

The authors represent the singleton vs. geminate rhotic as /ɾ/ and /r/, respectively.

Other real and not so real onset-sensitive data

223

inventory of vowels and consonants (long and short), it receives nuclear H* pitch and bears stress (GB&C 1997: 27). GB&C claim that the nuclear syllable must be heavy, thus it is no coincidence that it is the only one that can also host a geminate. Had the other syllables included a geminate, then they would become heavy, a fact banned in the language – admittedly a rather circular argument. Following GB&C’s line of thought nonetheless, it must be assumed that geminates in Berawan are wholly syllabified in the onset, because otherwise they would render the penultimate syllable heavy rather than the final one, as desired. In addition, the Berawan minimal word is bimoraic CVV or CVC13 (1997: 27–8), but surprisingly not CːV, as we would expect if geminates were heavy. We could nevertheless argue that this is because initial geminates are banned, and can only appear word medially, thus necessarily in larger words. But this offers no similar rescue to the absence of [CV.CːV] words, which seems entirely unexpected. A solution is nonetheless possible. Assuming – as GB&C do – that the final syllable in Berawan must be heavy (although no similar expectations apply to other syllables), but claiming that so-called geminates are in fact weightless doubled consonants (§5.4.2.2), then facts fall into place. The reason we do not get [CVCːV] is because this would actually be [CVμ.CːVμ], i.e. *[LL], which is forbidden since the last syllable needs to be heavy. At the same time, configurations such as CV.CːVː e.g. [dʒikːɑː] ‘rendezvous’ (GB&C: ex. 1, p. 18), or CV.CːVC, e.g. [lakːεh] ‘omen bird’ (GB&C: ex. 9, p. 25), would be the only ones possible, because they could render the final syllable heavy due to the bimoraicity of the rime. A further problem is the presence of CːVːC syllables as in e.g. [ŋәɲːɑːŋ]. Had the heavy consonant and the coda contributed a mora each, then this syllable should be quadrimoraic. This is totally unprecedented. Positing instead three moras as the maximum syllable weight, as in Estonian (Hayes 1995) or Kashmiri (Morén 2000), is consistent with the presence of CːVːC syllables (where the first consonant is an initial non-moraic doubled C) as well as with the lack of *CV1ːV2C and *CV1V2ːC that would exceed this maximum. Overlooking the facts above can easily allow one to reach the conclusion that Berawan permits onset geminates. On closer inspection, however, as we have seen, this cannot be maintained. Assuming instead that Berawan only possesses weightless doubled consonants immediately provides a neat account for the problematic cases above. 13

A minimal syllable contains an onset (GB&C: 36), which explains why VV and VC words are banned, although bimoraic.

224

Onsets

6.3.2 Bislama Bislama (Camden 1977; Gordon 2005; Lynch 1975; Crowley 2004), the national language of Vanuatu, is a creole in which a large part of the vocabulary is based on English and French, but which has significantly diverged from both. Bislama is relevant here since Gordon (2005) based on Camden (1977) argues that it presents a stress algorithm which is based on the following weight scale, whereby complex onsets, but not singleton ones, contribute to weight: (20)

Bislama weight scale (Gordon 2005) CCVC > CVC > CCV > CV

Topintzi (2006b: §7.3.2) argues that this conclusion is fairly unsubstantiated, given that Camden’s description causes numerous concerns; in particular, he offers a stress algorithm that assumes highly peculiar syllabification, e.g. [(C)CVC.V] instead of [(C)CV.CV] for /(C)CVCV/, and does not present any real-life examples throughout, only abstract C/V forms. Setting this matter aside and assuming for the sake of the argument that the approaches of both Camden and Gordon are on the right track, then their predictions still diverge in many respects. For instance, Camden says that final CCV takes stress, unless we have V. ´CVC.CCV. From that – and other statements – Gordon infers that CCV attracts stress due to its heaviness, but not as much as a CVC syllable. If then we interpret Camden as saying that all words (in the language, that is, and not in his data only) with final CCV get stress apart from this one exception, then this distinguishes him from Gordon who would expect final stress only when the penult is V or CV, as shown below. (21)

Camden vs. Gordon with respect to CCV final syllables (the cases where the predictions differ are in bold) Camden

Gordon

σ.V. ´CCV

σ.V. C ́ CV [CCV > V]

σ.CV. ´CCV

σ.CV. C ́ CV [CCV > CV]

σ. C ́ VC.CCV σ.CCV. ´CCV

Exception

σ. ´CVC.CCV [CVC > CCV] σ. ´CCV.CCV [default]

σ.CCVC. ´CCV

σ. ´CCVC.CCV [CCVC > CCV]

σ.CCCV. ´CCV

σ. ´CCCV.CCV [CCCV > CCV?]

σ.CCCVC. ´CCV

σ. ´CCCVC.CCV [CCCVC > CCV]

Similar divergence between the two approaches emerges on other occasions. In addition, other accounts of Bislama point to yet more different stress algorithms. For instance, Lynch (1975: 193–5) observes certain tendencies, e.g. that

Other real and not so real onset-sensitive data

225

in almost all disyllables, quadrisyllables and the majority of trisyllables the penult receives stress. Lynch’s description is more in accordance with the recent discussion of Bislama stress by Crowley (2004). Stress is claimed to be unpredictable, although some strong tendencies can be identified. Words of Melanesian origin overwhelmingly present stress on the penult, as happens to be the case in the source languages. Words of French descent often have stress on the final syllable, whereas those of English origin usually keep the stress they present in English. It is interesting, however, to mention that Crowley (p.c. 22 March 2004) notes that some representative words like epril, futbrek, divos, baot all present penult stress. Obviously, épril contradicts both Camden and Gordon, since the final syllable is in both accounts a heavy one and yet receives no stress. Similarly for fútbrek, dívos, báot. While these are loanwords and their stress could be attributed to effects from the source language, such issues do not arise in either Camden or Gordon, who propose algorithms that take no account of the loanword phonology. One thing is for sure: while there is no consensus on the Bislama data, recent accounts of the facts in Crowley (2004), but also earlier ones, like Lynch (1975), do not suggest any effect from the onsets. Until more robust evidence arises in favour of onset effects on Bislama stress, I believe it is safer to assume that this language has a quite common system with penult stress and partial lexical stress largely as a result of the heavy borrowing from English, French and other Melanesian languages. 6.3.3 Nankina The last language to be considered in this section is Nankina (Spaulding and Spaulding 1994; Al-Harbi 2005; Gordon 2005), a language of Papua New Guinea spoken in the region of the Madang Province, Saidor District, in the upper Nankina River valley. This is the other language, besides Bislama, that Gordon (2005) presents as an example of a stress system where CCV > (C)V.14 Nankina normally presents penult stress (22a), with the exception of words with a final complex onset as in (22b). (22)

Nankina stress a. Normal – penultimate stress tɔ́.wuŋ ‘egg’ wɔ́.rε kʌ̀.ndεp ‘wood’ jέ.ɥi b. CCV attracts stress – final stress ε.kwí ‘bad’ ε.tsjέŋ

14

‘a sore’ ‘cause’ ‘light weight’

The other analysis of Nankina stress, i.e. Al-Harbi (2005), accepts Gordon’s interpretation of the data, although it eventually rightly concludes that CCV syllables do not behave as heavier than (C)V ones. However, it does so for the wrong reasons. Its drawbacks and a more detailed discussion can be found in Topintzi (2006b: §7.3.3).

226

Onsets

Obviously, such an account is based on the idea that there are complex onsets in Nankina. Instead, I argue that Nankina does not lend itself as a testing ground for a CCV > (C)V hypothesis, simply because if one examines the data more closely, it becomes obvious that Nankina lacks CCV syllables in the first place. This view is corroborated by Spaulding and Spaulding (1994: 3–4), who state that word-initial clusters such as /pt/, /bt/ or /kt/ are split by what they call ‘a transitional vowel [ɨ]’ whose exact phonetic realization varies. Thus, one finds: /pta/ – [pɨra] ‘stand’, /ktagwɔk/ – [kɨrangwɔk] ‘small knife’, /bt/ – [bɨt] ‘pig’. The same epenthetic vowel also appears in pidgin words that clearly involve complex onsets (23). Nothing like that happens in cases like /kwit/ – [kwit]/*[kɨwit] ‘bird’ or /mwεk/ – [mwεk]/*[mɨwεk] ‘lizard’. (23)

Nankina

Pidgin

Gloss

gɨras pɨrεs sɨtɔri bɨrupεra

gras ples stori blupela

‘grass, hair’ ‘village’ ‘story’ ‘blue’

Potential true complex onsets are thus banned from Nankina; apparent ones are, however, admitted, but this is because they are not really complex onsets, rather they are complex segments with complex or secondary articulation (Clements 1986; Sagey 1986; Rosenthall 1997, among others). Such are the affricates, prenasalized segments and consonant–glide sequences (with the most popular being /kw/ and /gw/). With this in mind, then, the facts above are easily explained. (22a) informs us that in sequences with syllables of the same type, i.e. CV(C), stress is preferably penultimate, but in sequences where the penult is onsetless, stress moves to the final onsetful syllable (22b). If CCV in Nankina is, as we claim, truly CV, then the [V. ´CCV] pattern actually translates to [V. ´CV]. But of course this is the familiar PO pattern (❷) extensively discussed in §2.3, whereby onsetful syllables attract stress more than onsetless ones. There is an additional piece of evidence in favour of the proposed account. A word like takwʌn ‘curse’ is stressed [tákwʌn] (Spaulding and Spaulding 1994: 18) and not *[takwʌ́n], as one would expect given Gordon’s proposed CCV > (C)V weight hierarchy. But of course, if /kw/ is really a single segment – albeit a complex one – it is only natural that the word should be stressed like [tákwʌn], since this is really a CV.CV(C) sequence. In other words, the syllables are equally light, so initial stress is chosen, since this is the normal pattern the language chooses when no other pressing requirement is applicable.

Other real and not so real onset-sensitive data

227

Consequently, Nankina does not provide support for CCV > CV. Even if we accept that complex onsets are allowed – despite the evidence that they are not – there can be no explanation of why CV and CCV act in the same way, other than that the language does not distinguish between the two. We have thus offered evidence that both alleged cases of CCV > CV (Gordon 2005) as manifested in Nankina and Bislama cannot be justified. Neither language seems to distinguish between complex and simplex onsets in terms of weight contribution. This is an issue I briefly address again from a different perspective in the next section. 6.3.4 Moraic contrast between CV and CCV? Although Nankina and Bislama have provided no convincing evidence in favour of a difference in the moraic structure between CV and CCV syllables, this is a possibility that should nonetheless be contemplated. In Damin (§6.2.5.1), CCV syllables satisfied the minimality criterion, in contrast to CV ones. In Luganda (§6.2.5.2), certain CCV syllables counted as heavy for metrical purposes. Finally, in Supyire (§6.2.5.3), gemination after CL occurred within complex clusters. Cases like these suggest that CCV syllables might behave as heavy unlike the simpler (C)V ones. The problem with each of these cases, however, is that the data are always ambiguous. In Damin, the process of rearticulation seems to suggest that these CCV sequences are more likely CːV geminates, which are of course expected to behave as heavy (see Ch. 5 on geminate onsets). In Luganda, seemingly complex onsets with a glide in the second position may be treated as ‘nuclear onglides’ following Smith (2003) and consequently be resyllabified in the nucleus. In this way, there would be no CCV syllables that are heavy. And in Supyire, it is unclear whether gemination really takes place within clusters, rather than creation of syllabic consonants or sesqui-syllables. Either of the two latter options would abandon the use of heavy CCV syllables. In the light of these objections, it remains an open issue whether the contrast CCV > (C)V can ever arise. Notably, however, a similar contrast in codas does not arise, that is, I am not aware of any language where: (C)VCC > (C)VC, (C)V. Such a language would treat open syllables with a short vowel and those with a singleton coda as light, but once the coda acquired an extra consonant, then the syllable would become heavy. I believe that such a weight scale is absent, although the somewhat similar (C)VCC > (C)VC > (C)V is attested. For instance, in Hindi (Hayes 1995), (C)VCC syllables are super-heavy, (C)VC ones are heavy and (C)V ones light. Presumably, the absence of such patterns can be attributed to the fact that Moraic Onset/Coda is really the constraint responsible for moraicity

228

Onsets

on onsets and codas, respectively. Since this merely looks at syllable positions without distinguishing between singleton and complex margins, then it would be impossible for a complex onset/coda to become moraic without the singleton becoming moraic too. A new implication arises; in particular, it is anticipated that if a language has weightful complex syllable margins, then it should also have weightful singleton margins. In principle then, languages with the weight system CCV > CV > V are possible. I have not been able to find such a language, but future investigation may perhaps be able to verify this prediction empirically. It is not unimaginable, however, that the search will be fruitless. To see why, consider what happens in codas. While numerous languages treat CVCs as heavy, only a few languages possess complex codas, let alone moraic ones. Taking into consideration that the moraic contrast between CV and V is much rarer than the moraic coda contrast in the first place, a CCV > CV > V appears even less likely. 6.4

Summary

The current chapter has presented data from numerous languages. Most of them conform to the theory laid out in Chapters 1–5, as they exhibit evidence in support of onset weight. A few exceptions have appeared in §6.3. Importantly, these are not exceptions in that they dispute the theory; rather, it is the case that more careful investigation of their phonology reveals that they possess no moraic onsets in the first place. They have, however, been included here, since according to previous studies they offered evidence in support of weightful onsets. I have shown that this was wrong. This chapter’s main contribution has therefore been to complement the list of languages that possess weightful onsets. This is important for a book whose aspiration (among others) is to constitute a resource for future research on the issue of onset weight. In addition, this chapter has been the most appropriate place to discuss other weight-based phenomena, such as tone and reduplication, that also interact with onsets. As empirical facts are scantier in these cases, they would not be able to sustain separate chapters of their own and consequently had to be incorporated elsewhere. With this chapter, the core of the book has been completed. Starting from cases of coerced onset weight and stress (Ch. 2), where considerations of voicing were of utmost importance, we used the phenomena of compensatory lengthening (Ch. 3) and word minimality (Ch. 4) as a stopover, and finally shifted our attention to clear cases of distinctive onset weight, namely onset

Other real and not so real onset-sensitive data

229

gemination (Ch. 5). As just said, Chapter 6 offered complementary material with the intention of providing a complete picture of the prosodic behaviour of onsets. What remains to be done is to ponder upon alternative treatments of the phenomena and see where they fail. Meanwhile, the main points and arguments made in the book will be summarized to facilitate its overall evaluation.

7 Conclusion and discussion of alternatives

7.1

The book’s main themes

This book has been about onsets and their behaviour in suprasegmental phonology. Having discarded the stipulation that most other phonological accounts take for granted, namely that onsets are prosodically inert, it has been argued instead that onsets may be moraic or not on a language-specific basis. This brings them on a par with codas. And like codas, the moraicity of onsets is regulated by markedness considerations or is subject to underlying specifications. In the latter instance, moraicity is lexical, hence unpredictable, leading to onset geminates. In the former, though, the claim has been that moraicity is basically down to voicing, and that in this type of moraic onset, voiceless onsets always make better mora bearers than their voiced counterparts. However, in the case of nuclei and codas, the more general concept of sonority takes over, because in those positions, voicing is not contrastive or reliable enough. Such cases have been tackled elsewhere (see Zec 1988, 1995; Morén 2001). Throughout the book, these ideas have received empirical support from a wide range of data and phenomena. We have looked at stress, word minimality, compensatory lengthening, gemination, reduplication, tone, and all have highlighted the important role onsets play in suprasegmental phonology. In the process, various formal, technical and analytical issues have been dealt with in order to provide a comprehensive theory of onset weight. These key ideas and themes are summarized in (1) and discussed in more detail immediately afterwards. (1)

Key ideas and themes in the book I. All syllable constituents, including onsets, may be moraic (see (4) below) II. Moraic onsets come in two flavours, just like codas do (see Morén 2001), namely distinctive and coerced. Both refer to consonants that end up being syllabified in the output as onsets, but the difference is that the first started off with an underlying mora, whereas the latter acquired it on the surface III. The nature of coerced moraic onsets is predictable. Prototypical moraic onsets are [−voi] ones, due to the markedness ranking: *μ/Ons/[+voi] >> *μ/Ons. The constraint BeMoraic which assigns moraicity can be

230

Conclusion and discussion of alternatives

231

interleaved in any of the three positions available, generating the following patterns: i. *μ/Ons/[+voi] >> *μ/Ons >> BeMoraic: no onsets are moraic ii. *μ/Ons/[+voi] >> BeMoraic >> *μ/Ons: only [−voi] onsets and those that lack voice specification are moraic iii. BeMoraic >> *μ/Ons/[+voi] >> *μ/Ons: all onsets are moraic IV. The significance of [+/– voice] related to coerced onset moraicity is justified based on the following facts: i. H-toned moras, i.e. moras with high pitch perturbation, may attract stress ii. One source of high pitch perturbation is the lack of voicing iii. Pitch is one of the phonetic correlates of stress → Proposal: in some languages pitch perturbation due to voicing is phonologized as stress (rather than tone) in terms of moras V. The behaviour of sonorants is variable depending on whether they are phonologically unspecified for [voi] or marked as [+voi], predicting that under the ranking in III.ii, two patterns are possible: α. if sonorants are unspecified, they behave like voiceless obstruents in being moraic onsets; β. if sonorants are [+voi], they behave like voiced obstruents in being non-moraic onsets VI. Distinctive moraic onsets are the onset geminates. Because they are lexically specified as moraic, they do not have to adhere to the moraic-onset distribution patterns above VII. Both coerced and distinctive moraic onsets may participate in the same weight-related phenomena, e.g. stress, word minimality, reduplication, compensatory lengthening, etc.

As explained before (see also §1.2.2), the idea that onsets are weightless has been dominant in current phonological theory, but without any satisfactory justification. As Scheer and Szigetvári (2002) aptly state: ‘moraic theory … ha[s] got nothing to say about why onsets never count’. Its answer to why they do not count actually is: ‘because onsets are non-moraic. Why are they non-moraic? Because moraic theory has decided that there is no mora attached to them. Why did moraic theory decide so? Because we observe that onsets never count for stress’. This is, of course, ‘overtly circular’ (Scheer and Szigetvári 2002), but also, as I have tried to show throughout the book, blatantly false too. In particular, I have provided evidence from a host of weight-sensitive phenomena, where besides nucleic and coda weight, onset weight is also instrumental. Importantly, the cases under study have extended beyond stress. This is significant, because, as discussed in §7.2.1, the literature concerning the prosodic behaviour of onsets has largely been limited to the investigation of stress. The importance of weight for stress has been thoroughly investigated (see Hayes 1995 for an overview), but nonetheless, there is research claiming that in at least some cases, it could be reinterpreted in terms of prominence

232

Onsets

(§7.2.1). Possibly then, such reasoning could expand to onset-sensitive stress. In this view, stress alone is not sufficient to support the existence of onset weight. Instead, other phenomena, where weight is undoubtedly involved, provide better justification. Compensatory lengthening, geminates, word minimality and reduplication, among others, perfectly fit the bill. The preceding chapters have shown how all of them – as well as other phenomena – interact with onsets offering robust support for their moraicity. The table (2) presents an exhaustive list of the relevant phenomena studied in this book – including stress – accompanied by the languages that illustrate them. It also points to the section(s) where they were explored in the book. When applicable, additional comments are also provided. (2)

Onset-sensitive phenomena investigated in the book and corresponding languages

Phenomenon

Language

Comments

Compensatory lengthening

Samothraki Greek (§3.2) Marshallese (§3.3.3) Trique (§3.3.4) Pattani Malay (§5.2.1) Trukese (§5.2.2) Onondaga (§6.2.1.1) Maasai, Romanesco Italian Ntcham, Anuak / Anywa, Turkana, Gyore, Lango (§6.2.1.2) Marshallese (§3.3.3 & §5.3.1 & §6.2.4.2) Trique (§3.3.4)

onsets as CL-trigger onsets as CL-trigger onsets as CL-target onsets as CL-target onsets as CL-target onsets as CL-trigger onsets as CL-trigger

Onset geminates

Pattani Malay (§5.2.1) Trukese (§5.2.2) Djeebbana (§5.3.3) Puluwat (§5.3.4)

Stress

Eastern Popoloca (§5.3.5) Bellonese (§6.2.4.1) Karo (§2.2.1) Aranda, Alyawarra, Lamalama, Mbabaram, Umbuygamu, Umbindhamu, Linngithig, Uradhi, Kuku-Thaypan, Kaytetj, Agwamin, Banawá, Juma (§2.3.1), Iowa-Oto (§2.3.1.1) English (§2.3.2), Dutch (§2.3.2.1), Ainu (§2.3.2.2) Pirahã (§2.4) Arabela (§2.5)

word medially word initially (& medially?) word initially word initially word medially word initially (& medially?) word medially word initially QO effect only PO effect only (Onset-on-Stress)

PO effect only (Stress-on-Onset) PO & QO effect ambiguous between QO-only and PO&QO

Conclusion and discussion of alternatives Phenomenon

Minimal word

Language

Comments

Marshallese (§5.3.1) Pattani Malay (§5.2.1) Djeebbana (§5.3.3) Eastern Popoloca (§5.3.5) Tümpisa Shoshone (§6.2.2)

due to geminates due to geminates due to geminates due to geminates presumably QO or PO&QO CV, (C)VV, CC, (C)VC but *V CːV, CVV but *(C)V(C) CːV, CVV, CCV? but *CVC

Bella Coola (§4.3) Trukese (§5.2.2) Damin (§6.2.5.1)

Tone Reduplication

Kpelle (§6.2.3) Musey (§6.2.3) Pattani Malay (§5.2.1) Bellonese (§6.2.4.1) Marshallese (§6.2.4.2)

233

expressed through gemination expressed through gemination expressed through gemination

What the above establishes, then, is that onsets not only count for stress – contra moraic theory and for that matter contra Scheer and Szigetvári (2002) too – but also for a host of other phenomena, which are unambiguously weight based. Thinking about this on another dimension, we thus find that weight is much more symmetrical than previously thought, as all three syllable constituents may bear it. In some sense, this is actually the null hypothesis. From the standpoint of standard moraic theory where only nuclei and codas can be moraic, finding cases where onsets are non-moraic is only natural, whereas the few cases where this is disputed are usually treated as a quirky deviation that is at best mentioned in passing. What this approach never asks, though, is why it should be normal that only two out of the three syllable constituents are weight bearing. This in turn generates the question: ‘why should it be that onsets are weightless?’ In light of the absence of any persuasive answer to this issue, as well as the conundrum of some counterexamples that even standard weight theorists acknowledge, this book has adopted a different standpoint. Assuming that weight is symmetrical, then it should be the case that onsets too can be moraic. And as we find, indeed they are. This view both conforms to the null hypothesis and has much more accurate empirical coverage. Coming to the issue of representations now, recall from Chapter 1 the structures that the major previous weight theories employed. These are repeated below for ease of reference.

234

Onsets

(3)

Previous models of syllable and weight CV-theory

a.

X slot-model

σ

b.

Standard moraic theory

σ

CVC

R

t a n

ONC

c. R=rime O=onset N=nucleus C=coda

σ µµ t a n

XXX t a n

CV theory proposes a structure (McCarthy 1979; Clements and Keyser 1983) where, depending on their consonantal or vocalic quality, segments receive C and V tags that also serve as timing units. For weight purposes, though, special status is assigned to the constituent nucleus which can contain underneath it: a long vowel (Vː), a diphthong (VV) or a VC sequence. In other words, this system only superficially offers the flat structure of (3a), since it makes explicit reference to the nucleus. Moreover, it implicitly marks the timing units under the nucleus node as different from those of the other constituents. This is because allegedly only the former can act as weight units. Obviously, CV theory cannot accommodate weightful onsets. The X-slot model (Levin 1985) solves other problems that CV theory faces (see §1.2), but not the weightlessness of onsets, since it assumes that only rimal material can bear weight. The same insight is shared by standard moraic theory (Hayes 1989), which, however, does away with the rime by making use of moras. Crucially, as explained next, these are optional. For reasons laid out in §1.2.1, moraic theory has been considered the most appropriate syllable and weight model as a working framework for this book. By now, however, I hope to have convinced readers of its weakness with regard to onset weight and to have paved the way for the modified moraic theory model that I illustrate below. (4)

Modified moraic theory proposed in this book σ (µ) (µ) (µ) t

a

n

In this updated model, a flatter structure of the syllable is proposed – reminiscent of Davis’ (1985) model – where all constituents may be weightful. However, they do not necessarily have to be. This is entirely unremarkable for

Conclusion and discussion of alternatives

235

codas, where the admittance of moraic and non-moraic ones is well accepted (e.g. Latin, Delaware, English, Kiowa and Turkish vs. Wargamay, Lenakel, Eastern Ojibwa, Khalkha Mongolian, respectively; see Hayes 1995; Morén 2001; Zec 1995, 2007). For onsets, I have been arguing throughout the book that they too are moraic on a language-specific basis. More striking is perhaps the case of nuclei, but for them too, there is work suggesting that they are occasionally weightless. Such is the case in Malagasy (Erwin 1996), Chuvash and Mari (Hyman 1985), Kabardian (Peterson 2007), Alamblak (Mellander 2003b), but also for /e/ in Mohawk (Pigott 1998; Michelson 1989), /ə/ in Musqueam (Shaw 2002) and for /ʌ/ and /ɨ/ in Au (Scorza 1985; Crosswhite 1999). This state of affairs is represented in the table below. (5)

The symmetrical nature of weight and the optionality of moraicity Nuclei Non-Moraic YES (Musqueam) Moraic

Codas

Onsets

YES (Lenakel)

YES (majority of Ls)

YES (majority of Ls) YES (Latin)

YES (this book)

An additional point that becomes evident through (4) is that the behaviour of onsets with regard to weight and the proposal about the symmetrical nature of syllable weight argues against the onset–rime distinction. Since onsets – like the rimal constituents – may also bear weight, the most important justification for the rime is now invalidated. Davis (1985) also offers argumentation against the onset–rime distinction through the presentation of phonotactic constraints holding between the onset and nucleus, as well as between onset and coda, e.g. in Yindjibarndi r cannot co-occur in both onset and coda of the same syllable. In his view, this establishes that the strong affiliation between nucleus and coda represented by the rime cannot exist. On a different note, Yip (2003) presents facts from the behaviour of Mandarin pre-nuclear glides which sometimes behave as if they are in the nucleus and sometimes as if they are in the onset. Given that pre-nuclear glides lie at the alleged constituent boundary between onset and rime, they are predicted to behave as either onset or rime but not as both. Their ‘schizophrenic’ behaviour, as Yip calls it, can only make us doubtful about the existence of the onset–rime distinction in the first place. Perhaps, the argument made presently delivers an even firmer blow. At this point, a reasonable objection arises. If the system is as symmetrical as suggested, then how can asymmetries on the emergence of coda vs. onset weight exist? In other words, how can we capture the undisputable reality that onset weight is much rarer than coda weight?

236

Onsets

Notably, the proposed structure in (4) makes claims about symmetries in representation, but has nothing to say about frequency of patterns. The latter instead relates to other phonological conditions. In particular, in §3.3.3 I showed that it generally holds true that not all potentially CL-inducing situations actually lead to CL. Adding to the equation that onset deletion is rare anyway, CL as a response to it is anticipated to be highly uncommon. Later (§5.4.1), an account for the relative rarity of onset gemination was presented. To produce onset geminates, *Moraic Onset must be ranked really low. As this happens only occasionally, it is anticipated that onset geminates will not be particularly frequent either. Finally, with regard to tone, I have argued that tone is concomitant with the presence of fundamental frequency, which is, however, absent from the prototypical moraic onsets, i.e. the voiceless ones. Since tone and onset moraicity impose contradictory demands, tonal effects from moraic onsets are predicted to be highly marginal. I have thus attributed the asymmetry and rarity of onset weight to a constellation of phonological factors, rather than to any single property. Furthermore, other, additional factors inform us about the asymmetry between onset and coda weight, but provide us with no clue as to the rarity aspect. More specifically, consider the discussion in §1.3.3, where it was claimed that voicing regulates the weight behaviour of onsets, but sonority regulates the weight behaviour of nuclei and codas. Taking this for granted does not explain why we do not find voicing-based onset-weight systems more frequently than we do. Although matters such as these are left open for future research, I suspect that functional explanations of the sort that have appeared in Smith (2005) or Gordon (2005), may prove quite enlightening. To conclude, onsets are like their other fellow-syllable constituents, i.e. moraic on a language-specific (and even intra-language) basis. As a result, it falls out naturally that they too participate in the same weight-based phenomena that nuclei and codas are generally accepted to participate in. A synopsis of them is illustrated in the following table. (6)

A (non-exhaustive) list of weight-related phenomena and their association to nuclei, codas and onsets with sample languages1 Moras in:

1

Phenomenon:

Nuclei

Codas

Onsets

Stress

Wargamay, Lenakel

English, German

Pirahã (§2.4), Karo (§2.2), Arabela (§2.5), Pattani Malay (§5.2.1)

Sources with respect to nuclei and codas are as follows: Hayes (1995) for stress, Kenstowicz (1994a) for geminates, Hayes (1989) for compensatory lengthening, Gordon (2006) for minimal word, McCarthy and Prince (1986) for reduplication.

Conclusion and discussion of alternatives

237

Moras in:

7.2

Phenomenon:

Nuclei

Codas

Onsets

Geminates

N/A

Italian, Japanese

Pattani Malay (§5.2.1), Trukese (§5.2.2), Marshallese (§3.2)

Compensatory lengthening

Middle English, Slavic

Latin, Komi Ižma

Samothraki Greek (§4.2), Trique (§5.3.2), Pattani Malay (§5.2.1)

Minimal word

Cebuano, Fijian

Basque, Koasati

Bella Coola (§3.3), Trukese (§5.2.2)

Reduplication

Ponapean, Mokilese

Mokilese, Manam

Marshallese (§5.3.1; §6.2.3.2), Bellonese (§6.2.3.1)

Alternatives

While the present work has aspired to provide a thorough investigation of onsetsensitive prosodic phenomena, it is by no means the only study on this subject. However, it is the only book-length study on the topic. It is also unique in addressing a range of phenomena in a unifying way. Other accounts tackle one onset-sensitive phenomenon at a time and thus have narrow empirical coverage and theoretical foundation. For the most part, these deal with stress (Davis 1985, 1988; Everett 1988; Hayes 1995; Downing 1998; Goedemans 1998; Smith 2005; Gordon 2005) or with gemination (Hajek and Goedemans 2003). The problems Hajek and Goedemans’ account on gemination faces have been explored in §5.2.1.1.2, hence they will not be considered here again. The stress accounts, however, merit some discussion and comparison with the present approach. Broadly, these can be distinguished between the prominence-only accounts, foremost of which is Smith (2005), and the weight-as-a-function-ofphonetic-prominence account of Gordon (2005). These are discussed in turn. 7.2.1 What is wrong with prominence-based accounts Like syllable weight, prominence is a notion widely used in phonology. Unlike syllable weight, whose manifestation is more clearly identified, the definition of prominence is elusive. Jensen (2004) correctly observes that depending on the analyst, prominence has different meanings and is sometimes related to – although considered distinct from – akin notions like stress and accent. It is therefore generally agreed, citing Jones (1956: 141), that ‘Stress is not the same as “prominence” […]; stress is one of the factors that may cause or help to cause a sound or syllable to be “prominent”’. According to Jones (1956: 142), the other factors which can make a sound more prominent are ‘inherent sonority, length and intonation’. Despite the lack of consensus on what prominence entails exactly, a shared assumption in

238

Onsets

more recent work seems to be that prominence is equated with ‘perceptual salience’ (Hayes 1995; Jensen 2004; Smith 2005), which is still quite vague. In spite of this weakness, prominence has been widely used to account for several data. Pirahã in particular (§2.4) has received much attention. For instance, Everett (1988) and Hayes (1995) use a prominence grid that renders voiceless obstruents more prominent and thus more stress attracting than their voiced counterparts, but this is entirely arbitrary, e.g. there is no explanation why PV acts as heavier than BV and not the other way round. Goedemans (1998) attempts to explain this direction of facts in a more systematic fashion phonetically, but eventually does not propose any more full-fledged analysis. Lack of a principled – possibly phonetic – account that makes clear predictions as to the anticipated and attested patterns is obviously an important shortcoming of these proposals. However, it is not the only one. While it is unclear what exactly prominence encompasses, it is nonetheless clear that it is much more general a concept than weight. For this reason, it is possible for prominence to include a large number of phenomena and easily lead to over-generation. Indeed, this is what happens. For instance, under prominence it would be possible to find a language whose syllables with an onset s attract stress more than others as a result of s’s loudness. Since ‘louder’ can easily entail ‘more prominent’, this is a predictable system, and yet – to my knowledge – unattested. No similar expectation arises in the weight-based account laid out throughout the book. Consequently, weight makes more specific predictions than prominence. Weight is more restrictive in another way too. It can maximally reach two or three moras, but there is no similar upper bound for prominence. Virtually anything goes, therefore a considerable overgeneration of patterns is predicted. Perhaps, however, the finishing blow against prominence is not over-generation but under-generation. The reasoning here is entirely empirical. Despite the negative consequences of prominence, it can capture the onset-sensitive stress facts. The problem nonetheless appears once we look beyond stress and examine clearly weight-based phenomena like word minimality (Ch. 4), compensatory lengthening (Ch. 3) or reduplication (Ch. 6). One can see that for these phenomena, too, onsets can be relevant. Such a result is absolutely surprising and totally unaccounted for within a prominence-driven framework simply because the notion of prominence is inherently designed to tackle phenomena like stress, but not others where only weight can be implemented. More generally, then, the inability of prominence to extend beyond stress virtually renders it an untenable approach for the onset-sensitive facts. Despite these flaws, there is at least one account that strives to be quite specific in its predictions as well as its understanding of prominence. This is

Conclusion and discussion of alternatives

239

Smith (2005) and for that reason, it deserves special reference. It needs to be mentioned, however, that only a small portion of it is devoted to onset sensitivity, thus it is inherently unable to serve as a complete account of onset weight. This work originates in research that distinguishes positions into ‘strong’ and ‘weak’ (Casali 1996; Beckman 1999, among others). Strong positions are those which preserve contrasts, whereas neutralization is found in weak positions, a phenomenon termed ‘positional neutralization’. For instance, in English or Catalan stressed syllables (strong positions), the full vowel inventory is found, but in unstressed syllables (weak positions), only a subset of it emerges. Strong positions are further grouped into two categories: i) phonetically strong positions containing stressed syllables (σ́), onsets and long vowels, and ii) psycholinguistically strong positions including roots and initial syllables (σ1). Smith focuses only on strong positions and claims that strong positions need to be augmented, i.e. acquire properties that enhance their perceptual salience. In other words, prominent positions need to become more prominent, a phenomenon dubbed ‘positional augmentation’. Properties that enhance prominence are the presence of: weight (HEAVYσ), stress (Have stress), high tone (HTone), [Place] features in consonants (HaveCPlace), onsets (Onset) or onsets of low sonority (*Onset/X), and nuclei of high sonority (*Peak/X).2 Conjoining these properties with the strong positions above generates a number of constraints relativized to the position under consideration. For instance, ONSET/σ́, *[ONSET/X]/σ́ and HTONE/σ require that a stressed syllable has an onset, a low-sonority onset and high tone, respectively. Not all possible combinations of prominent properties relativized to strong positions are admitted. These are regulated by two filters Smith introduces (see original work for details). Besides some general concerns this work raises (see §1.3.2 in Topintzi 2006b), there are certain problematic features in the way it proposes to treat onset-sensitive stress. In particular, Smith discusses two constraints relating to onsets, namely Onset and *Onset/X, both of which can be relativized to initial and stressed syllables. Thus we get: (7)

2

a. Onset: Syllables have onsets Onset/σ1: Initial syllables have onsets ONSET/σ́: Stressed syllables have onsets

The ‘X’ in these constraints stands for any step on the segmental sonority hierarchy. The version of sonority Smith assumes is in (8). Notably, the *Peak/X hierarchy takes the form *Peak/ voiceless obstruents >> … >> *Peak/low V, thus preferring segments of the highest sonority as nuclei. The *Onset/X hierarchy is the reverse, i.e. *Onset/low V >> … >> *Onset/voiceless obstruents, favouring voiceless obstruents as best possible onsets (see Prince and Smolensky 1993/2004).

240

Onsets b. *Onset/X: Syllables have less sonorous onsets *[ONSET/X]/σ1: Initial syllables have less sonorous onsets *[ONSET/X]/σ́: Stressed syllables have less sonorous onsets

For Onset and *[Onset/X], Smith claims that there is neurophysiological evidence suggesting that the presence of an onset and its specific quality enhance prominence. Besides such evidence, there is not much empirical support for these constraints, especially for the version of *[Onset/X] pertinent to stressed syllables, i.e. *[ONSET/X]/σ́. With this constraint, Smith generates the patterns of Pirahã (§2.4), where her claim is that low-sonority onsets attract stress, as well as the patterns of Niuafo`ou (de Lacy 2000) where glide onsets are avoided in stressed syllables. Notably, these patterns make reference to the extreme edges of the sonority hierarchy (8). If Smith were right, though, and sonority is indeed the reason one finds the Pirahã and Niuafo`ou patterns, then we should be able to find languages which utilize any of the intermediate cut-off points for the *[ONSET/X]/ σ́ constraints as well. (8)

Sonority Hierarchy with respect to consonants (Smith 2005: 56) glides > rhotics > laterals > nasals > voiced obstruents > voiceless obstruents

For instance, in analogy to Niuafo`ou, we would predict that there is a language where glides, rhotics and laterals are avoided in stressed syllables or another where all sonorants attract stress, and so on. Similarly, a pattern reminiscent of Pirahã could arise where voiceless obstruents, voiced obstruents and nasals attract stress to the exclusion of all other consonants. As far as I know, none of these arises, posing a serious problem for Smith’s proposal that sonority regulates onset sensitivity. However, this still leaves us with the pattern of Niuafo`ou, which is not predicted by the present theory. I suspect that actually Smith herself (2003) offers a possible solution, when she makes a distinction between ‘true onset’ glides and ‘nuclear’ onglides. Simplifying a bit, the former really syllabify in an onset position, whereas the latter are actually underneath the nucleus node. As a result, only the former are subject to constraints that make reference to onset structure (*Onset/X), whereas the latter escape them. It could thus be conceivable to build an analysis which treats Niuafo`ou glides as nuclear onglides. As such, they would be subject to a different set of constraints, i.e. probably the [*PEAK/X]/σ́ constraints, and consequently would not interfere with the genuine onset-sensitive stress effects.

Conclusion and discussion of alternatives

241

All in all, then, Smith’s (2005) proposal, albeit interesting, fails to provide a satisfactory account for the onset-sensitive stress data, since it over-generates in its predictions (see above). Despite being more constrained compared to other prominence-based accounts in other respects, it also under-generates. This is because it fails to account for the dual behaviour of sonorants. While it can explain the common patterning of sonorants and voiced obstruents in Pirahã by virtue of their increased sonority in contrast to the voiceless obstruents, it also predicts that sonorants will always be the least likely onsets to attract stress due to *[ONSET/X]/σ́. In principle, though, sonorants can pattern with the voiceless obstruents, but only when their voiced counterparts do so too. This is a direct consequence of the hierarchy in (8). In other words, Smith’s approach falsely predicts that a language like Karo, where sonorants pattern with the voiceless obstruents to the exclusion of the voiced ones, should not exist. 7.2.2 Gordon (2005) Gordon (2005) is, to my knowledge, the only work – besides the present one – that addresses onset weight exclusively, and not as a side issue. Although this in itself provides sufficient reason to consider it here, it is also an interesting and thorough account that can reasonably stand as a contender to the present approach. Its rather detailed examination and critique, then, is fully justified. Gordon’s account of onset weight is in line with other work by him (predominantly Gordon 2002, which examines rimal weight) that considers weight to be projected from phonetic prominence, that is, weight is seen as a function of prominence. For Gordon, phonetic prominence refers to greater perceptual energy, where perceptual energy is calculated as the integration of energy over time (Gordon 2002: 60). Thus, in a given language, syllables with greater perceptual energy are heavier than those with less energy. In his account, languages adopt weight distinctions that offer the greatest separation of heavy and light syllables in the perceptual energy domain (Gordon 2005: 602). Gordon’s investigation of onset sensitivity is restricted to stress, focusing both on languages where stress is attracted to the first onsetful syllable (e.g. Alyawarra in §2.3.1 here) and on those where stress is attracted to syllables with onsets of particular quality (e.g. Pirahã in §2.4 here). To account for onset weight and relate it to perceptual energy, Gordon discusses the notions of adaptation and recovery (Delgutte 1982; Viemeister 1980). The former expresses the decline in auditory sensitivity to a stimulus, while the latter refers to the fact that sensitivity to a new stimulus can be regained by means of a period of silence or reduced acoustic energy. Onsets – and in particular less intense and

242

Onsets

thus less sonorous onsets – provide exactly this recovery phase and enhance the perceptual energy of the rime. Thus, (low sonority) onsets can contribute to weight through the perceptual boost they offer to the following vowel, rendering it stress attracting. However, onsets themselves do not contribute directly to the auditory prominence of the syllable, but do so only through their effect on the following rime. This is a ‘crucial feature’ in Gordon’s (2005: 604) analysis. The perceptual energy of the onset itself is thus not addressed, but it is likely that it too could have effects on stress. Applying the same definition of perceptual energy to onsets too, then we would expect it would be enhanced when sonorants or fricatives appear as onsets, but not if stops appear as onsets. Another expectation with respect to perceptual energy emerges. Suppose we consider a couple of CV sequences and for each we examine three windows. Window 1 is the perceptual energy of the onset itself, window 2 refers to the recovery period, while window 3 is the steady state of the vowel, which can be disregarded, since it is expected to be the same in all cases. As explained above, in window 1 the perceptual energy will be increased when the onset is made up of a nasal compared to a voiceless stop, for example. On the other hand, lowsonority onsets such as voiceless stops will contribute to recovery (window 2) more significantly than more sonorous onsets such as nasals will. (9)

Perceptual energy effect from the onset itself and during the recovery phase Window 1 perceptual energy of the onset itself

Window 2 recovery

NV

+

−

PV

−

+

Window 3 steady state of the vowel

N = nasals, P = voiceless stops For Window 1: + = high perceptual energy − = low perceptual energy For Window 2: + = high recovery − = low recovery

The problem that Gordon faces is that he does not take into account window 1. In his system it seems to be the case that the effects during the recovery period (window 2) take priority over those of the onset perceptual energy (window 1), predicting that voiceless stop onsets make better weight bearers than nasals, i.e. PV > NV. However, it could be possible that in other languages window 1 matters more than 2, resulting in a situation exactly opposite to the one discussed above. In this case, more intense, thus more sonorous, onsets like nasals would

Conclusion and discussion of alternatives

243

be predicted as more able to carry weight, NV > PV. This issue is left unexplored. Note that if one also considers similar simultaneous effects in codas, very complex and presumably unlikely systems are produced. Even if we disregard this omission, perceptual energy and auditory boost – as currently understood – lead to incorrect predictions. For example, Gordon (2005: 604) proposes that less intense onsets should provide a bigger auditory boost. Within this class of onsets, he also includes obstruents, but although stops indeed have low intensity, the same does not hold for fricatives, some of which are rather intense sounds. In this view it would be surprising to find languages where fricatives pattern with the weightful onsets and not the weightless ones. Against expectations, though, in Pirahã, onset s patterns with the voiceless stops in contributing weight to the syllable (see [sabí] ex. (47b) in Ch. 2). The idea of auditory boost is also used to explain why in a VC1.C2V sequence, the coda C1 does not contribute any weight to the immediately following syllable. Delgutte (1982) has found that a period of 0–40 ms of preceding silence offers the greatest auditory boost to the following vowel, whereas a silence that lasts 40–100 ms only has a minor effect. Given that most consonants are longer than 40 ms, it falls out naturally that only the presence of the first prevocalic consonant (C2) will offer a significant auditory boost to the following vowel, whereas C1’s contribution to the auditory effect will be insignificant. Applying this line of thought to geminates is problematic, since Gordon does not predict any significant difference between geminates and singletons, given that the latter are already long enough to offer the auditory boost required. This does not predict the pattern in which a singleton does not render its following vowel stress attracting, while a geminate does. But this is exactly what we find in languages with onset geminates like Pattani Malay (§5.2.1), Trukese (§5.2.2), Marshallese (§5.3.1; §6.2.4.1) or Bellonese (§6.2.4.2). A further problem for Gordon is that he attempts to bring cases like Alyawarra (§2.3.1), Banawá (§2.3.1) and Dutch (§2.3.2.1) on a par with languages like Pirahã (§2.4). Recall that in the present framework, the first two respond to the alignment constraint ALIGNσ́O, which requires that stressed syllables are onsetful, by shifting stress from the default position to an onsetful syllable, while Dutch achieves this by inserting an onset to an otherwise onsetless stressed syllable. Pirahã also employs this constraint to achieve stress on a BVV syllable instead of a VV one, but also makes a weight distinction based on the quality of the onset. But for Gordon, this difference in the nature of the two phenomena does not exist. They are simply considered different manifestations of the same phenomenon, namely weight. As Gordon states, the constraint *PROM[Ø[X]R]σ

244

Onsets

(where X is material within the rime R of a syllable σ) ‘treats syllables with an onset as heavy’ (2005: 642) and is used to account for the stress-attracting feature of onsetful syllables. But this is an undesirable result, because it would entail that we should expect onsetful syllables in languages like Banawá or Dutch to behave as if they were heavy in phenomena other than stress. To clarify, consider Banawá, which sets a bimoraic word minimum, satisfied by V.CV and CVV words, but not by monomoraic CV words, as Everett (1996: ex. 12) explicitly mentions, e.g. *bi, *ka. If Gordon were right, and Banawá onsetful syllables were heavy, then why couldn’t they themselves satisfy the minimality requirement? In other words, why is a [CV] word impossible? Of course, it could be argued that for some reason the minimality criterion is satisfied by a different dimension of weight (e.g. the mixed-weight system of Pirahã, where one dimension of weight refers to onsetful syllables, and another to onsetful syllables of a certain quality), and not by onsetful syllables. Given that Gordon uses X slots and not moras, it would be a challenging feat to define this suitably (to see why, consider the problem of Japanese hypocoristics in §1.2.1), let alone be able to do that for a wide range of languages presenting similar facts (but see Garrett 1999 for discussion). Closing this section, an important drawback of Gordon’s account, like Smith’s (2005), is the inability to account for the dual behaviour of sonorants. Recall that sonorants would either pattern alongside voiceless obstruents in being moraic (as in Karo, §2.2.1) or alongside the voiced obstruents in being non-moraic (as in Pirahã, §2.4). This is entirely predicted in the system proposed in the book, which associates the pitch perturbation due to voicing with either tone or stress (§1.3.3). Phonetically, sonorants do not automatically perturb the F0 of adjacent vowels, so elevation or depression of fundamental frequency is observed on a language-specific basis. Depending on what happens in a language, we then anticipate that sonorants will behave variably. But this is not the case in Gordon, who presumably would expect that the perceptual energy of sonorants and the auditory boost they offer to the following rime should be systematic across languages. Since sonorants are quite intense, they offer a lower auditory boost than, for example, stops. Consequently, in a language that bases onset weight on an onset-quality distinction, Gordon predicts that sonorants may be less weightful than voiced onsets (or at least equally weightful), but never as weightful as voiceless onsets. Evidently, given the situation in Karo, this is plain wrong. Gordon’s paper is undoubtedly an important contribution and a very much needed addition to the literature on onset weight, which, despite its interest and consequences for syllable and weight theory, has been largely dismissed rather

Conclusion and discussion of alternatives

245

than explored. It nonetheless has a number of shortcomings of varying significance, both in terms of theory as well as of empirical coverage, that eventually render it untenable as a theory of onset weight. 7.3

Conclusion

In the prevailing moraic theory à la Hayes (1989) – what I have called here standard moraic theory – it is taken for granted that onsets are never weight bearers. Moraic theory has never really questioned why this should be so. Given that, according to the null hypothesis, all syllable constituents should be able to bear weight, the burden falls to standard moraic theory to explain why there is an asymmetry between the onset and the nucleus&coda or – if you will – the rime. And yet, only sporadic work has dealt with this issue, with much of that not actually subscribing to the standard theory either (see Gordon 2005). Eventually, the conclusion of past research has been that onset-sensitive instances do exist, albeit highly marginally. The next step has been to demarcate the distributional boundaries of these onset-sensitive data. With the exception of Hajek and Goedemans (2003), who talked about weightful onset geminates, all other work has claimed that onset sensitivity only arises with respect to stress. This has for the most part been convenient, because stress is closely related to the notion of prominence and consequently onset-sensitive stress could be subsumed under the ‘umbrella’-concept of prominence, rather than – the problematic for standard moraic theory – weight. But even so, what about Hajek and Goedemans’ data? To account for gemination through prominence is too far-fetched. This has superficially led us to a stalemate. On the one hand, prominence may account for onset-sensitive stress in accordance with standard moraic theory, but has nothing to say about onset geminates. On the other, onset geminates (and stress) may be analysed by means of onset weight, but standard moraic theory explicitly prohibits it. But this deadlock is only insurmountable if one accepts that standard moraic theory is right in its assumptions. If, instead, one realizes that the prosodic inertness of onsets is merely a stipulation that has never been adequately vindicated – besides the justification the empirical facts supposedly provide – then we are free to re-think the possible moraicity of onsets. This has been the starting point of this book. Taking at face value the empirical data of Pattani Malay, as described in Yupho (1989; cited in Hajek and Goedemans 2003), and along with them, the similar facts of Trukese (Churchyard 1991; Muller 1999), a moraic-onset approach has begun to seem

246

Onsets

appropriate. Further painstaking exploration has revealed that numerous other cases command a similar account. These have all been incorporated in this book. Notably, among them, cases of onset-sensitive stress have also been included. As I have shown earlier (§7.2.1), closer examination of the relevant data has unveiled that prominence-based accounts do not actually succeed in accounting for these facts, although they are purported to do so. The present book has thus endeavoured to bring the role of the onset into the limelight. In doing so, it has shown a basic assertion of moraic theory to be wrong. According to it, onsets are weightless. In reality, though, they are not. Instead, they are weightful on a language-specific basis. Such a conclusion, backed up by a restrictive model of onset weight (as described in the preceding chapters), captures a fuller range of the empirical data, uniformly accounts for a cluster of phenomena (stress, compensatory lengthening, onset geminates and word minimality, among others) and automatically gives a sought-after answer to a burning question for standard moraic theory. More specifically, in the question of ‘why are onsets never weight bearing?’, the present theory argues that this is an invalid question, simply because onsets can be moraic. Consequently, alongside language-specific weightless [CVC] and weightful [CVCμ] codas that are theoretically well accepted, this book comes to propose the same language-specific distinction for onsets too, i.e. [CV] and [CμV]. This conclusion thus enforces a slightly modified representation for syllabification and weight within the moraic framework (see (10)), where onsets come in two flavours: a) non-moraic (as previously thought) and b) moraic (as presently argued). (10)

Non-moraic onsets σ

C

and

Moraic onsets σ

µ

µ

µ

V

C

V

References

In these references, ROA refers to the Rutgers Optimality Archive, a freely accessible resource for research within the Optimality Theory tradition. It is available at: http:// roa.rutgers.edu/index.php3. All articles in the ROA are identified through a number, e.g. ROA-nnn, which is also indicated in the relevant references. Abo, Takaji, Byron W. Bender, Alfred Capelle and Tony DeBrum (1976) MarshalleseEnglish Dictionary. Honolulu: University Press of Hawaii. Abrahamson, Arno and Joyce Abrahamson (1984) Os Fonemas da Língua Júma. In Estudos Sobre Línguas Tupí do Brasil [Série Lingüística 11], R. Dooley (ed.), pp. 157–74, Brasília: Summer Institute of Linguistics. Abramson, Arthur (1986) The Perception of Word-Initial Consonant Length: Pattani Malay. Journal of the International Phonetic Association 16: 8–16. (1987) Word-initial Consonant Length in Pattani Malay. Proceedings of the XIth International Congress of Phonetic Sciences, Tallinn, pp. 68–70. (1991) Amplitude as a Cue to Word-initial Consonant Length: Pattani Malay. Proceedings of the XIIth International Congress of Phonetic Sciences, Aixen-Provence, pp. 98–101. (1998) The Complex Acoustic Output of a Single Articulatory Gesture: Pattani Malay Word-initial Consonant Length. In Papers from the Fourth Annual Meeting of the Southeast Asian Linguistics Society 1994, U. Warotamasikkhadit and T. Panakul (eds.), pp. 1–20, Tempe, Arizona: Program for Southeast Asian Studies, Arizona State University. (1999) Fundamental Frequency as a Cue to Word-initial Length: Pattani Malay. Proceedings of the XIVth International Congress of Phonetic Sciences, San Francisco, pp. 591–94. (2003) Acoustic Cues to Word-Initial Stop Length in Pattani Malay. In Proceedings of the 15th International Congress of Phonetic Sciences, Barcelona, Spain, August 3–9, 2003. Universitat Autònoma de Barcelona, pp. 387–90. (2004) Toward Prosodic Contrast: Suai and Pattani Malay. In International Symposium on Tonal Aspects of Languages: Emphasis on Tone Languages, Bernard Bel and Isabelle Marlien (eds.), pp. 1–4, Beijing: The Institute of Linguistics, Chinese Academy of Social Sciences. Afroudakis, Angelos (1985) Collection of Language Material from the Island of Samothraki. Manuscript No. 1293, Research Centre of Modern Greek Dialects and Idioms, Academy of Athens [Αφρουδάκης, Άγγελος (1985) Συλλογή Γλωσσικού Υλικού 247

248

References

από τη Νήσο Σαµοθράκη. Αριθμός Χειρογράφου: 1293, Κέντρον Ερεύνης των Νεοελληνικών ∆ιαλέκτων και Ιδιωμάτων – Ι.Λ.Ν.Ε. Ακαδημία Αθηνών]. Akinlabi, Akinbiyi (1993) Underspecification and the Phonology of Yoruba /r/. Linguistic Inquiry 24(1): 139–60. Al-Harbi, Awwad (2005) Nankina Trochees. ms., Umm Al-Qura University. Available as ROA-724. Albright, Adam (2002) Base Selection in Analogical Change in Yiddish. In The Proceedings of the 28th Annual Meeting of the Berkeley Linguistics Society (BLS 28), pp. 1–13. Alderete, John (2001) Morphologically Governed Accent in Optimality Theory. New York and London: Routledge. Originally 1999 doctoral dissertation, University of Massachusetts, Amherst. Alderete, John, Jill Beckman, Laura Benua, Amalia Gnanadesikan, John McCarthy and Suzanne Urbanczyk (1999) Reduplication with Fixed Segmentism. Linguistic Inquiry 30: 327–64. Andriotis, Nikolaos (1926–1928) Collection of Language Material from Samothraki. Manuscript No: 449, Research Centre of Modern Greek Dialects and Idioms, Academy of Athens [Ανδριώτης, Νικόλαος (1926–1928) Συλλογή Γλωσσικού Υλικού εκ Σαµοθράκης. Αριθμός Χειρογράφου: 449, Κέντρον Ερεύνης των Νεοελληνικών ∆ιαλέκτων και Ιδιωμάτων – Ι.Λ.Ν.Ε. Ακαδημία Αθηνών]. Anttila, Arto (1994) Consonant Gradation is Epenthesis. Handout from paper presented at the Stanford Phonology Workshop (Feb 1994). (2006) Variation and Opacity. Natural Language and Linguistic Theory 24(4): 893–944. Ashton, E. O., E. M. K. Mulira, E. G. M. Ndawula and A. N. Tucker (1954) A Luganda Grammar. London: Longmans, Green and Co. Baertsch, Karen and Stuart Davis (2003) The Split Margin Approach to Syllable Structure. ZAS Working Papers in Lingustics 32: 1–14. Bagemihl, Bruce (1991) Syllable Structure in Bella Coola. Linguistic Inquiry 22: 589–646. (1998) Maximality in Bella Coola (Nuxalk). In Salish Languages and Linguistics, Theoretical and Descriptive Perspectives, Ewa Czaykowska-Higgins and M. Dale Kinkade (eds.), pp. 71–98, Berlin, New York: Mouton de Gruyter. Bakovic, Eric (1996) Foot Harmony and Quantitative Adjustments. ms., Rutgers University. Available as ROA-168. Bao, Zhiming (1990) On the Nature of Tone. Doctoral dissertation, MIT. Beckman, Jill (1999) Positional Faithfulness: an Optimality Theoretic Treatment of Phonological Asymmetries. New York, London: Garland. Originally 1998 doctoral dissertation, University of Massachusetts, Amherst. Beltzung, Jean-Marc (2007) Allongements Compensatoires: une Typologie. Presented at the 7th Biennial Meeting of the Association for Linguistic Typology (ALT7), 24–26 September 2007, Paris, France. Bender, Byron W. (1976) Spoken Marshallese. Honolulu: University Press of Hawaii. (1991) On the Category Distributive. In Currents in Pacific Linguistics: papers on Austronesian languages and ethnolinguistics in honour of George W. Grace, Robert Blust (ed.), pp. 11–26, Canberra: Department of Linguistics, Australian National University.

References

249

Bender, Byron W., Alfred Capelle and Louise Pagotto (in prep.) The Sound System of Marshallese. ms. Benua, Laura (1995) Identity Effects in Morphological Truncation. In Papers in Optimality Theory, Jill Beckman, Laura Walsh Dickey and Suzanne Urbanczyk (eds.), pp. 77–136, Amherst, MA: GLSA. Bermúdez-Otero, Ricardo (2001) Underlyingly Nonmoraic Coda Consonants, Faithfulness and Sympathy. ms., University of Manchester. (in prep.) Stratal Optimality Theory. Oxford Studies in Theoretical Linguistics. Oxford University Press. Blevins, Juliette (2001) Where Have All the Onsets Gone? Initial Consonant Loss in Australian Aboriginal Languages. In Forty years on: Ken Hale and Australian languages, Jane Simpson, David Nash, Mary Laughren, Peter Austin, Barry Alpher (eds.), pp. 481–92, Canberra: Pacific Linguistics. Blumenfeld, Lev (2006) Constraints on Phonological Interactions. Doctoral dissertation, Stanford University. Booij, Geert (1995) The Phonology of Dutch. Oxford: Clarendon Press. Borrelli, Doris (2002) Raddoppiamento Sintattico in Italian: A Synchronic and Diachronic Cross-Dialectical Study. New York and London: Routledge. Bradshaw, Mary (1999) A Crosslinguistic Study of Consonant-Tone Interaction. Doctoral dissertation, Ohio State University. Breen, Gavan (1992) Some Problems in Kukatj Phonology. Australian Journal of Linguistics 12(1): 1–43. Broselow, Ellen, Su-I Chen and Marie Huffman (1997) Syllable Weight: Convergence of Phonology and Phonetics. Phonology 14: 47–82. Buller, Barbara, Ernest Buller and Daniel Everett (1993) Stress Placement, Syllable Structure, and Minimality in Banawá. International Journal of American Linguistics 59: 280–93. Bunn, Gordon and Ruth Bunn (1970) Golin Phonology. In Pacific Linguistics A23: Papers in New Guinea Linguistics No.11, pp.1–7. Camden, Pastor Bill (1977) A Descriptive Dictionary: Bislama to English. Rosebury, Australia: Bridge Printery. Campos-Astorkiza, Rebeka (2004) Faith in Moras: A Revised Approach to Prosodic Faithfulness. In Proceedings of North East Linguistics Society 34, K. Moulton and M. Wolf (eds.), Amherst, MA: GLSA. Capell, Arthur (1962) Some Linguistic Types in Australia. Oceanic Linguistics Monographs, Number 7, Sydney: University of Sydney. Carlson, Robert J. (1990) A Grammar of Supyire: Kampwo Dialect. Doctoral dissertation, University of Oregon. Casali, Roderic (1996) Resolving Hiatus. Doctoral dissertation, UCLA. Churchyard, Henry (1991) Compensatory Lengthening and ‘Gemination Throwback’ in Trukese and Puluwat as Evidence for Rime in Moraic Phonology. ms., University of Texas. Clements, George N. (1986) Compensatory Lengthening and Consonant Gemination in LuGanda. In Studies in Compensatory Lengthening, Leo Wetzels and Engin Sezer (eds.), pp. 37–77, Dordrecht: Foris.

250

References

(1990) The Role of the Sonority Cycle in Core Syllabification. In Papers in Laboratory Phonology I: Between the Grammar and Physics of Speech, John Kingston and Mary E. Beckman (eds.), pp. 283–333, Cambridge University Press. Clements, George N. and Samuel Jay Keyser (1983) CV Phonology: A Generative Theory of the Syllable. Cambridge, MA: MIT Press. Cohn, Abigail (2003) Phonological Structure and Phonetic Duration: the Role of the Mora. Working Papers of the Cornell Phonetics Laboratory 15: 69–100. Coleman, John S. (1996) Declarative Syllabification in Tashlhiyt Berber. In Current Trends in Phonology: Models and Methods, Volume I, Jacques Durand and Bernard Laks (eds.), pp. 177–218, Salford: European Studies Research Institute. Connell, Bruce A., Richard A. Hayward and John Abraha Ashkaba (2000) Observations on Kunama Tone. Studies in African Linguistics 29: 1–42. Crosswhite, Katherine (1999) Vowel Reduction in Optimality Theory. Doctoral dissertation, UCLA. (2001) Predicting Syllabicity and Moraicity in Dihovo Macedonian. In University of Rochester Working Papers in the Language Sciences 2(1), J. S. Magnuson and K. M. Crosswhite (eds.), pp. 1–22. Crowley, Terry (2004) Bislama Reference Grammar. Oceanic Linguistics Special Publication No. 31, Honolulu: University of Hawaii Press. Cruttenden, Alan (2001) Gimson’s Pronunciation of English. 6th edition, London: Arnold. Curtis, Emily Kathryn Jean (2003) Geminate Weight: Case Studies and Formal Models. Doctoral dissertation, University of Washington. Davis, Stuart (1985) Topics in Syllable Geometry. Doctoral dissertation, University of Arizona, Tucson. (1988) Syllable Onsets as a Factor in Stress Rules. Phonology 5: 1–19. (1994) Geminate Consonants in Moraic Phonology. WCCFL 13: 32–45. (1999a) On the Moraic Representation of Underlying Geminates. In The ProsodyMorphology Interface, René Kager, Harry van der Hulst and Wim Zonneveld (eds.), pp. 39–61, Cambridge University Press. (1999b) On the Representation of Initial Geminates. Phonology 16: 93–104. Davis, Stuart and Gina Torretta (1998) An Optimality-Theoretic Account of Compensatory Lengthening and Geminate Throwback in Trukese. NELS 28: 111–25. Dayley, Jon P. (1989) Tümpisa (Panamint) Shoshone grammar [University of California Publications in Linguistics 115]. Berkeley, CA: University of California Press. de Lacy, Paul (1999) Tone and Prominence. ms., University of Massachusetts, Amherst. (2000) Markedness in Prominent Positions. In Proceedings of HUMIT, MIT Working Papers in Linguistics 40, Ora Matushansky et al. (eds.) [also available as ROA-432]. (2002) The Interaction of Tone and Stress in Optimality Theory. Phonology 19: 1–32. (2006) Markedness: Reduction and Preservation in Phonology. Cambridge University Press. (2007) The Interaction of Tone, Sonority and Prosody. In The Cambridge Handbook of Phonology, Paul de Lacy (ed.), pp. 281–307, Cambridge University Press.

References

251

Delgutte, Bertrand (1982) Some Correlates of Phonetic Distinctions at the Level of the Auditory Nerve. In The Representation of Speech in the Peripheral Auditory System, R. Carlson and B. Granström (eds.), pp. 131–50, Amsterdam: Elsevier. Dimmendaal, Gerrit J. (1983) The Turkana Language. Dordrecht: Foris. Donohue, Mark (1999) A Grammar of Tukang Besi. Grammar Library series No. 20. Berlin: Mouton de Gruyter. Downing, Laura (1998) On the Prosodic Misalignment of Onsetless Syllables. Natural Language and Linguistic Theory 16: 1–52. Duanmu, San (1990) A Formal Study of Syllable, Tone, Stress and Domain in Chinese Languages. Doctoral dissertation, MIT. Dyen, Isidore (1965) A Sketch of Trukese Grammar. New Haven, Connecticut: American Oriental Society. Elbert, Samuel H. (1974) Puluwat grammar. Pacific Linguistics, series B-29 Canberra: Australian National University. (1988) Echo of a Culture: A Grammar of Rennell and Bellona. Oceanic Linguistics Special Publication, No. 22, Honolulu: University of Hawaii Press. Elenbaas, Nine and René Kager (1999) Ternary Rhythm and the Lapse Constraint. Phonology 16: 273–329. Erwin, Sean (1996) Quantity and Moras: An Amicable Separation. In UCLA Occasional Papers in Linguistics 17: The structure of Malagasy, vol. I, Matthew Pearson and Ileana Paul (eds.), pp. 2–30, Los Angeles: University of California Press. Evans, Jonathan P. (2001) Contact-Induced Tonogenesis in Southern Qiang. Language and Linguistics 2(2): 63–110. Everett, Daniel (1988) On Metrical Constituent Structure in Pirahã Phonology. Natural Language and Linguistic Theory 6: 207–46. (1996) Prosodic Levels and Constraints in Banawá and Suruwahá. Available as ROA-496. (2003) Iambic Feet in Paumari and the Theory of Foot Structure. Linguistic Discovery 2: 22–44. Everett, Daniel and Keren Everett (1984) On the Relevance of Syllable Onsets to Stress Placement. Linguistic Inquiry 15: 705–11. Everett, Keren (1998) Acoustic Correlates of Stress in Pirahã. MA Thesis, University of Pittsburgh. Fabb, Nigel (1997) Linguistics and Literature: Language in the Verbal Arts of the World. Oxford: Blackwell. Frantz, Chester I. and Marjorie Frantz (1966) Gadsup Phoneme and Toneme Units. Linguistic Circle of Canberra Publications A 7: 1–11, Papers in New Guinea Linguistics 5, Canberra: Linguistic Circle of Canberra, Australian National University. Gabas Jr., Nilson (1998) Estudo Fonológico da Língua Karo. Munich and Newcastle: LINCOM Europa. (1999) A Grammar of Karo (Tupi). Doctoral dissertation, University of California, Santa Barbara. Gafos, Diamandis (1998) Eliminating Long-Distance Consonantal Spreading. Natural Language and Linguistic Theory 16: 223–78. Gahl, Susanne (1996) Syllable Onsets as a Factor in Stress Rules: the Case of Mathimathi Revisited. Phonology 13: 329–44.

252

References

García-Bellido, Paloma and Beatrice Clayre (1997) Prosodic Constraints and Representations in the Berawan Word. Oxford Working Papers in Linguistics, Philology, and Phonetics 2: 17–52. Garrett, Edward (1999) Minimal Words Aren’t Minimal Feet. In UCLA Working Papers in Linguistics, Vol. 1, Papers in Phonology 2, Matthew Gordon (ed.), pp. 68–105. George, Isaac (1970) Nupe Tonology. Studies in African Linguistics 1: 100–22. Goedemans, Rob (1996) An Optimality Account of Onset-Sensitive Stress in QuantityInsensitive Languages. Linguistic Review 13: 33–47. (1998) Weightless Segments. The Hague: Holland Academic Graphics. Goldrick, Matthew (2000) Turbid Output Representations and the Unity of Opacity. In Proceedings of the North East Linguistic Society, Vol. I, Masako Hirotani, Andries Coetzee, Nancy Hall and Ji-yung Kim (eds.) pp. 231–45, Amherst, MA: GLSA. González, Carolina (2003) The Effect of Stress and Foot Structure on Consonantal Processes. Doctoral dissertation, University of Southern California. Goodenough, Ward H. and Hiroshi Sugita (1980) Trukese-English Dictionary. Philadelphia: The American Philosophical Society. Gordon, Matthew (1998a) A Fortition-Based Approach to Balto-Fennic-Sámi Consonant Gradation. Folia Linguistica Historica 18: 49–79. (1998b) The Phonetics and Phonology of Non-Modal Vowels: a Cross-linguistic Perspective. Berkeley Linguistics Society 24: 93–105. (2002) A Phonetically-Driven Account of Syllable Weight. Language 78: 51–80. (2004) Syllable Weight. In Phonetically Based Phonology, Bruce Hayes, Robert Kirchner and Donca Steriade (eds.), pp. 277–312, Cambridge University Press. (2005) A Perceptually-Driven Account of Onset-Sensitive Stress. Natural Language and Linguistic Theory 23: 595–653. (2006) Syllable Weight: Phonetics, Phonology, and Typology. New York and London: Routledge. Originally 1999 doctoral dissertation, UCLA. Green, Antony Dubach (2002) Word, Foot, and Syllable Structure in Burmese. ms., University of Potsdam, available as ROA-551. Green, Thomas (1995) The Stress Window in Pirahã: a Reanalysis of Rhythm in Optimality Theory. ms., MIT, available as ROA-45. Gussenhoven, Carlos (2008) Vowel Duration, Syllable Quantity and Stress in Dutch. In The Nature of the Word. Essays in Honor of Paul Kiparsky, K. Hanson and S. Inkelas (eds.), pp. 181–98, Cambridge, MA: MIT Press. Gussenhoven, Carlos and Haike Jacobs (2005) Understanding Phonology. First edition 1998, London: Hodder Arnold. Hajek, John and Rob Goedemans (2003) Word-initial Geminates and Stress in Pattani Malay. The Linguistic Review 20: 79–94. (in prep.) On the Representation of Onset Weight. ms., University of Melbourne, Utrecht University and Universiteit Leiden. Hale, Ken (1981a) Part I: Remarks on Lardil Phonology and Morphology. In A Preliminary Dictionary of Lardil, Ken Hale et al., Document 0028 of the Aboriginal Studies Electronic Data Archive (ASEDA), Canberra: AIATSIS (Australian Institute of Aboriginal and Torres Strait Islander Studies). (1981b) Part III: Damin Vocabulary. In A Preliminary Dictionary of Lardil, Ken Hale et al., Document 0028 of the Aboriginal Studies Electronic Data Archive

References

253

(ASEDA), Canberra: AIATSIS (Australian Institute of Aboriginal and Torres Strait Islander Studies). Hale, Ken and David Nash (1997) Damin and Lardil Phonotactics. In Boundary Rider: Essays in Honour of Geoffrey O’Grady, Darrell Tryon and Michael Walsh (eds.), Pacific Linguistics C-136, pp. 247–59, Canberra: Research School of Pacific and Asian Studies, the Australia National University. Hall, Tracy A. (2000) Syllabically Conditioned Coalescence and Deletion in Zoque: an Optimality-Theoretic Approach. Linguistics 38 (4): 711–38. Halle, Morris and Kenneth Stevens (1971) A Note on Laryngeal Features. RLE Quarterly Progress Report 101: 198–312. Ham, William (2001) Phonetic and Phonological Aspects of Geminate Timing. New York and London: Routledge. Hammond, Michael (1999) The Phonology of English. Oxford University Press. Hansson, Gunnar Olafur (2004) Tone and Voicing Agreement in Yabem. In WCCFL 23: Proceedings of the 23rd West Coast Conference on Formal Linguistics, B. Schmeiser, V. Chand, A. Kelleher and A. Rodrίguez (eds.), pp. 318–31, Somerville, MA: Cascadilla Press. Hargus, Sharon (2001) Quality Sensitive Stress Reconsidered. University of Washington Working Papers in Linguistics 20: 25–56. Hargus, Sharon and Virginia Beavert (2006) Word-initial Clusters and Minimality in Yakima Sahaptin. Phonology 23: 21–58. Harrison, Sheldon P. (1976) Mokilese Reference Grammar. Honolulu: University of Hawaii Press. Hart, Michele E. (1991) The Moraic Status of Initial Geminates in Trukese. BLS 17: 107–20. Haudricourt, Andre-George (1954) De l’Origine des Tons en Vietnamien. Journal Asiatique 242: 69–82. Hayes, Bruce (1989) Compensatory Lengthening in Moraic Phonology. Linguistic Inquiry 20: 253–306. (1995) Metrical Stress Theory: Principles and Case Studies. Chicago: University of Chicago Press. Heisenberg, August (1921) Die Liquida P im Dialekt von Samothrake. In For G. N. Chatzidakis, ɪ, pp. 89–99, Athens: P.D. Sakellariou. Hendricks, Sean (1999) Reduplication Without Template Constraints: A Study in BareConsonant Reduplication. Doctoral dissertation, University of Arizona. (2001) Bare-Consonant Reduplication without Prosodic Templates: Expressive Reduplication in Semai. Journal of East Asian Linguistics 10: 287–306. Hermans, Ben (2001) CL in OT. Linguistics in the Netherlands 18: 139–50. Hermans, Ben and Marc van Oostendorp (2008) Umlaut is Phonological: Evidence from Ineffability. Paper presented at the 16th Manchester Phonology Meeting, May 2008, Manchester, UK. Hollenbach, Barbara (1977) Phonetic vs. Phonemic Correspondence in Two Trique Dialects. In Studies in Otomanguean Phonology, William R. Merrifield (ed.), pp. 35–67, Dallas: Summer Institute of Linguistics and University of Texas at Arlington. Hombert, Jean-Marie (1978) Consonant Types, Vowel Quality, and Tone. In Tone: A Linguistic Survey, V. Fromkin (ed.), pp. 77–111, New York: Academic Press.

254

References

Honda, Kiyoshi (2004) Physiological Factors Causing Tonal Characteristics of Speech: from Global to Local Prosody. Paper presented at Speech Prosody 2004, Nara, Japan. Retrieved 2 Sept 2008: www.isca-speech.org/archive/sp2004/ sp04_739.pdf. Howe, Darin (2000) Oowekyala Segmental Phonology. Doctoral dissertation, University of British Columbia. (2001) Syllabic Obstruents in Oowekyala. Paper presented at: Society for the Study of the Indigenous Languages of the Americas, Workshop on American Indian Languages, University of California, Santa Barbara. Hubbard, Kathleen (1994) Duration in Moraic Theory. Doctoral dissertation, University of California, Berkeley. Hulst, Harry van der (1984) Syllable Structure and Stress in Dutch. Dordrecht: Foris. Hyde, Brett (2001) Metrical and Prosodic Structure in Optimality Theory. Doctoral dissertation, Rutgers University. (2007) Issues in Banawá Prosody: Onset Sensitivity, Minimal Words, and Syllable Integrity. Linguistic Inquiry 38(2): 239–85. Hyman, Larry (1985) A Theory of Phonological Weight. Publications in Language Sciences 19, Dordrecht: Foris [reissued in 2003 by the University of Chicago Press]. (2006) World-Prosodic Typology. Phonology 23: 225–57. Itô, Junko and Armin Mester (1992) Weak Layering and Word Binarity. ms., University of California, Santa Cruz. Jakobson, Roman (1931) Die Betonung und ihre Rolle in der Wort- und Syntagmaphonologie. Travaux du Cercle Linguistique de Prague 4 . Reprinted in R. Jakobson (1962) Selected Writings I: Phonological Studies, pp. 117–136, Mouton: The Hague. Jeanne, Laverne M. (1982) Some Phonological Rules of Hopi. International Journal of American Linguistics 48: 245–70. Jensen, Christian (2004) Stress and Accent: Prominence Relations in Southern Standard British English. Doctoral dissertation, University of Copenhagen. Jones, Daniel (1956) The Pronunciation of English. 4th edition, Cambridge University Press. Kager, René (1989) A Metrical Theory of Stress and Destressing in English and Dutch. Dordrecht: Foris. (1999) Optimality Theory. Cambridge University Press. Kalstrom, Marjorie K. and Eunice V. Pike (1968) Stress in the Phonological System of Eastern Popoloca. Phonetica 18: 16–30. Katamba, Francis and Peter Cooke (1987) Ssematimba ne Kikwabanga: the Music and Poetry of a Ganda Historical Song. World of Music 29(2): 49–68. Katsanis, Nikolaos (1996) The Dialect of Samothraki Greek. Municipality of Samothraki [Κατσάνης, Νικόλαος (1996) Το Γλωσσικό Ιδίωµα της Σαµοθράκης. ∆ήμος Σαμοθράκης]. Kavitskaya, Darya (2002) Compensatory Lengthening: Phonetics, Phonology, Diachrony. New York and London: Routledge. Kenstowicz, Michael (1994a) Phonology in Generative Grammar. Oxford: Blackwell. (1994b) Sonority-Driven Stress. Available as ROA-33.

References

255

Key, Michael (2007) Headed Spans and Bantu Tonology. In University of Massachusetts Occasional Papers in Linguistics 32: Papers in Optimality Theory III, L. Bateman et al. (eds.), pp. 187–207 [also available as ROA-809]. Amherst, MA: GLSA. Kingston, John and David Solnit (1988a) The Inadequacy of Underspecification. In Proceedings of NELS 19: 264–278. Amherst, MA: GLSA. (1988b) The Tones of Consonants. ms., University of Massachusetts, Amherst and University of Michigan. Kiparsky, Paul (1973) ‘Elsewhere’ in Phonology. In A Festschrift for Morris Halle, S. Anderson and P. Kiparsky (eds.), pp. 93–106, New York, NY: Holt, Rinehart and Winston. (2000) Opacity and Cyclicity. The Linguistic Review 17: 351–67. (2002) Syllables and Moras in Arabic. In The Syllable in Optimality Theory, Caroline Féry and Ruben van de Vijver (eds.), pp. 147–82, Cambridge University Press. (2008a) Compensatory Lengthening. ms., Stanford University, available at: www. stanford.edu/~kiparsky/. (2008b) Fenno-Swedish Quantity: Contrast in Stratal OT. In Rules, Constraints and Phonological Phenomena, Bert Vaux (ed.), pp. 185–219, Oxford University Press. Kisseberth, Charles (1970) Vowel Elision in Tonkawa and Derivational Constraints. In Studies Presented to Robert B. Lees by his Students, Jerrold M. Sadock and Anthony L. Vanek (eds.), pp. 109–37, Edmonton, AB and Champaign, IL: Linguistic Research. Kraehenmann, Astrid (2001) Swiss German Stops: Geminates all over the Word. Phonology 18: 109–45. Krekoukias, Dimitrios A. (1964) Collection of Language Material from Samothraki. Manuscript No. 839, Research Centre of Modern Greek Dialects and Idioms, Academy of Athens [Κρεκούκιας, ∆ημήτριος Α. (1964) Συλλογή Γλωσσικού Υλικού εκ Σαµοθράκης. Αριθμός Χειρογράφου: 839, Κέντρον Ερεύνης των Νεοελληνικών ∆ιαλέκτων και Ιδιωμάτων – Ι.Λ.Ν.Ε. Ακαδημία Αθηνών]. Ladefoged, Peter and Ian Maddieson (1996) The Sounds of the World’s Languages. Oxford; Cambridge, MA: Blackwell. Lahiri, Aditi and Jacques Koreman (1988) Syllable Weight and Quantity in Dutch. In Proceedings of WCCFL 7, Hagit Borer (ed.), pp. 217–28, Stanford, CA: The Center for the Study of Language and Information. Lavoie, Lisa M. (2001) Consonant Strength: Phonological Patterns and Phonetic Manifestations. New York and London: Garland Publishing. Leben, William (2001) Tonal Feet. Paper presented in Typology of African Prosodic Systems Workshop, Bielefeld University, Germany, May 18–20, 2001. Lee, Byung-Gun (1996) The Emergence of the Faithful. Available as ROA-166. Levelt, Clara and Ruben van de Vijver (2004) Syllable Types in Cross-Linguistic and Developmental Grammars. In Constraints in Phonological Acquisition, René Kager, Joe Pater and Wim Zonneveld (eds.), pp. 204–18, Cambridge University Press [also available as ROA-265]. Levin, Juliette (1985) A Metrical Theory of Syllabicity. Doctoral dissertation, MIT.

256

References

Liberman, Mark and Alan Prince (1977) On Stress and Linguistic Rhythm. Linguistic Inquiry 8(2): 249–336. Lin, Yen-Hwei (1997) Syllabic and Moraic Structures in Piro. Phonology 14: 403–36. (1998a) Cyclic and Noncyclic Affixation in Piro. In Phonology in Progress – Progress in Phonology: HILP Papers III, pp. 167–88, The Hague: Holland Academic Press. (1998b) On Minor Syllables. In Proceedings of LP’98, O. Fujimura, B.D. Joseph and B. Palek (eds.), pp. 163–83, Charles University in Prague: The Karolinum Press. Lombardi, Linda (1999) Positional Faithfulness and Voicing Assimilation in Optimality Theory. Natural Language and Linguistic Theory 17: 267–302. (2002) Markedness and the Typology of Epenthetic Vowels, ms. University of Amherst, available as ROA–578. Longacre, Robert E. (1957) Proto-Mixtecan. International Journal of American Linguistics 23(4) Part 3 (Publication 5 of the Indiana University Research Center in Anthropology, Folklore, and Linguistics). Loporcaro, Michele (1991) Compensatory Lengthening in Romanesco. In Certamen phonologicum II, Papers from the 1990 Cortona Phonology Meeting, P. M. Bertinetto, M. Kenstowicz and M. Loporcaro (eds.), pp. 279–307, Turin: Rosenberg & Sellier. Lynch, John (1974) Lenakel Phonology. Doctoral dissertation, University of Hawaii. (1975) Bislama Phonology and Grammar: A Review Article. Kivung 8(2): 186–204. (1978) A Grammar of Lenakel. Pacific Linguistics B55, Canberra: Australian National University. (2000) Reconstructing Proto-Oceanic Stress. Oceanic Linguistics 39: 53–82. Maddieson, Ian (1997) Phonetic Universals. In Handbook of Phonetic Sciences, William J. Hardcastle and John Laver (eds.), pp. 619–39, Oxford: Blackwell. Madsen, Shannon (2000) Stress Patterns in Spanish Blends. In UCLA Working Papers in Linguistics, no.4: Papers in Phonology 4, Adam Albright and Taehong Cho (eds.). Mateus, Maria Helena and Ernesto d’Andrade (2002) The Phonology of Portuguese. Oxford University Press. Matteson, Esther (1965) The Piro (Arawakan) Language. Berkeley: University of California Press. McCarthy, John (1979) Formal Problems in Semitic Phonology and Morphology. Doctoral dissertation, MIT, Cambridge, MA. (2000) Faithfulness and Prosodic Circumscription. In Optimality Theory: Syntax, Phonology, and Acquisition, Joost Dekkers, Frank van der Leeuw and Jeroen van de Weijer (eds.), pp. 151–89, Oxford University Press. (2002) A Thematic Guide to Optimality Theory. Cambridge University Press. (2003a) OT Constraints are Categorical. Phonology 20: 75–138. (2003b) Sympathy, Cumulativity, and the Duke-of-York Gambit. In The Optimal Syllable, Caroline Féry and Ruben van de Vijver (eds.), pp. 23–76, Cambridge University Press. (2004) Headed Spans and Autosegmental Spreading. ms., University of Massachusetts, Amherst. (2007) Hidden Generalizations: Phonological Opacity in Optimality Theory. London: Equinox.

References

257

McCarthy, John and Alan Prince (1986) Prosodic Morphology 1986. Report no. RuCCS-TR-32. New Brunswick, NJ: Rutgers University Center for Cognitive Science; available online at: http://people.umass.edu/jjmccart/research.html. (1993) Generalized Alignment. In Yearbook of Morphology, Geert Booij and Jaap van Marle (eds.), pp. 79–153, Dordrecht: Kluwer. (1994) The Emergence of the Unmarked: Optimality in Prosodic Morphology. In Proceedings of the North East Linguistic Society 24, Mercè Gonzàlez (ed.), pp. 333–79, Amherst, MA: GLSA. (1995) Faithfulness and Reduplicative Identity. In University of Massachusetts Occasional Papers in Linguistics 18, Jill Beckman, Laura Walsh Dickey and Suzanne Urbanczyk (eds.), pp. 249–384, Amherst, MA: GLSA. McCrary, Cristie (2002) Vowel Length, Consonant Length and RaddoppiamentoSintattico: Segment Duration in Italian Revisited. Paper presented at the 10th Manchester Phonology Meeting, Manchester, UK. McGarrity, Laura W. (2003) Constraints on Patterns of Primary and Secondary Stress. Doctoral dissertation, Indiana University. Mellander, Evan (2003a) A Prosodic Theory of Prominence and Rhythm. Doctoral dissertation, McGill University. (2003b) Weightless Syllables in Alamblak and the Case Against Head Dependence. Poster presented at the CASTL Kick-Off Conference, 2–4 October 2003; University of Tromsø, Norway. Mester, Armin (1990) Patterns of Truncation. Linguistic Inquiry 20: 478–85. Michelson, Karin (1988) A Comparative Study of Lake-Iroquoian Accent. Dordrecht: Kluwer. (1989) Invisibility: Vowels Without a Timing Slot in Mohawk. In Theoretical Perspectives on Native American Languages, Donna Gerdts and Karin Michelson (eds.), pp. 38–69, Albany, NY: SUNY Press. Mohanan, K. P. (1986) The Theory of Lexical Phonology. Dordrecht: Reidel. Mohanan, Tara (1989) Syllable Structure in Malayalam. Linguistic Inquiry 20: 589–625. Morén, Bruce (2000) The Puzzle of Kashmiri Stress: Implications for Weight Theory. Phonology 17: 365–96. (2001) Distinctiveness, Coercion and Sonority: a Unified Theory of Weight. London: Routledge. Originally 1999 doctoral dissertation, University of Maryland at College Park. Morin, Yves-Charles (1992) Phonological Interpretations of Historical Lengthening. In Phonologica 1992: Proceedings of the 7th International Phonology Meeting, Wolfgang U. Dressler, Martin Prinzhorn and John Rennison, pp. 135–55, Turin: Rosenberg & Sellier. Morphy, Frances (1983) Djapu, a Yolngu Dialect. In Handbook of Australian Languages Vol. III, R. M. W Dixon and Barry J. Blake (eds.), pp. 1–188, Amsterdam: John Benjamins. Muller, Jennifer (1999) A Unified Mora Account of Chuukese. In Proceedings of WCCFL 18, S. Bird, A. Carnie, J. Haugen and P. Norquest (eds.) Somerville, MA: Cascadilla Press. (2001) The Phonology and Phonetics of Word-Initial Geminates. Doctoral dissertation, Ohio State University.

258

References

Murray, Robert and Theo Vennemann (1983) Sound Change and Syllable Structure in Germanic Phonology. Language 59: 514–28. Nater, Hank F. (1977) Stem list of the Bella Coola language. PdR Press Publications of Salish Languages 4, Lisse, Holland: Peter de Ridder Press. (1979) Bella Coola Phonology. Lingua 49: 169–87. (1984) The Bella Coola Language. Canada: National Museum of Manitoba, Mercury Series. Ethnology Service Papers 92. Ottawa. Nelson, Nicole (1998) Right Anchor, Aweigh. ms., Rutgers University [also available as ROA-284]. Nespor, Marina and Irene Vogel (1986) Prosodic Phonology. Dordrecht: Foris. Newman, Stanley (1947) Bella Coola I: Phonology. International Journal of American Linguistics 13: 129–34. (1971) Bella Coola Reduplication. International Journal of American Linguistics 37: 34–8. Newton, Brian (1972) Loss of /r/ in a Modern Greek dialect. Language 48: 566–72. Noonan, Michael (1992) A Grammar of Lango. Berlin, New York: Mouton de Gruyter. Odden, David (2005) Introducing Phonology. Cambridge University Press. (2007) The Unnatural Tonology of Zina Kotoko. In Tones and Tunes; v.1: Typological Studies in Word and Sentence Prosody, Tomas Riad and Carlos Gussenhoven (eds.), pp. 63–89, Berlin and New York: Mouton de Gruyter. Ohala, John J. (1983) The Origin of Sound Patterns in Vocal Tract Constraints. In The Production of Speech, J. MacNeilage (ed.), pp. 189–216, New York: Springer Verlag. (1997) The Relation Between Phonetics and Phonology. In The Handbook of Phonetic Sciences, W. Hardcastle and J. Laver (eds.), pp. 674–94, Oxford: Blackwell. Ola, Olanike (1995) Properheadedness and Binarity: Prosodic Words in Yoruba. In Theoretical Approaches to African Linguistics, Akinbiyi Akinlabi (ed.), pp. 273–93, New Jersey: Africa World Press. (1997) Benue-Congo Prosodic Phonology and Morphology. Lincom Studies in African Linguistics 15, München, Newcastle: Lincom Europa. Oostendorp, Marc van (2000) Back to the Base. On the Richness of the Base Hypothesis. ms, Meertens Institute, Amsterdam. (2005) Derived Environment Effects and Consistency of Exponence. Talk presented at University College London, December 2005. Papadopoulos, Anthimos (1927) Grammar of Modern Greek Northern Idioms. Athens: P. D. Sakellarios [Παπαδόπουλος, Άνθιμος (1927) Γραµµατική των Βορείων Ιδιωµάτων της Νέας Ελληνικής. Εν Αθήναις: ∏. ∆. Σακελλάριος]. Parker, Steven G. (2002) Quantifying the Sonority Hierarchy. Doctoral dissertation, University of Massachusetts, Amherst. Payne, David and Furne Rich (1988) Sensitivity to Onset in Arabela Stress. ms., Información de Campo 743. Yarinacocha, Pucallpa, Perú: Instituto Lingüístico de Verano. Pearce, Mary (2005) Iambicity Without Stress in Kera. In Proceedings of the University of Cambridge Third Postgraduate Conference in Language Research (CamLing 2005), G. Chatzidamianos et al. (eds.), pp. 271–78, Cambridge Institute of Language Research.

References

259

Peng, Long (1992) A Unified Theory of Tone-Voice. Doctoral dissertation, University of Arizona. Peterson, Tyler (2007) Issues of Homophony and the Minimal Word in the Adyghan Languages. Paper presented at: Conference on the Languages of the Caucasus, 7–9 December 2007; Max Planck Institute for Evolutionary Anthropology, Leipzig, Germany. Pigott, Glyne (1995) Epenthesis and Syllable Weight. Natural Language and Linguistic Theory 13: 283–326. (1998) Foot Form and the Parsing of Weightless Syllables. Paper presented at the 34th Annual Meeting of the Chicago Linguistics Society. Podi, Napo (1995) Esquisse Comparative de l’Akasilimi et du Basaal. Doctoral dissertation, Université de Grenoble 3. Prince, Alan (1990) Quantitative Consequences of Rhythmic Organization. In Papers from the Chicago Linguistic Society 26/2, M. Ziolkowski, M. Noske and K. Deaton (eds.), pp. 355–98. (1997) Topics in Optimality Theory: TETU, GTH and BC. Handout from LSA Class, retrieved on 17 December 2005 from: http://ling.rutgers.edu/gamma/talks/ insthdt4.pdf. Prince, Alan and Paul Smolensky (1993/2004) Optimality Theory: Constraint Interaction in Generative Grammar. Malden, MA; Oxford, UK; Victoria, Australia: Blackwell. Ramsay, Robert (2001) Tonogenesis in Korean. In Proceedings of the Symposium Crosslinguistic Studies of Tonal Phenomena: Historical Development, Phonetics of Tone, and Descriptive Studies, Shigeki Kaji et al. (eds.), pp. 3–17, Tokyo, Tokyo University of Foreign Studies, Institute for Languages and Cultures of Asia and Africa. Reh, Mechthild (1996) Anywa Language: Description and Internal Reconstructions. Cologne: Ruddiger Köppe Verlag. Rehg, Kenneth (1993) Proto-Micronesian Prosody. In Tonality in Micronesian Languages, J. A. Edmondson and K. J. Gregerson (eds.), pp. 25–46, Honolulu: University of Hawaii Press. Revithiadou, Anthi (1999) Headmost Accent Wins: Head Dominance and Ideal Prosodic Form in Lexical Accent Systems. The Hague: Holland Academic Graphics. Originally 1998 doctoral dissertation, University of Leiden. Rialland, Annie (1993) L’allongement Compensatoire: Nature et Modèles. In Architecture des représentations phonologiques, B. Laks and A. Rialland (eds.), pp. 59–92, collection ‘Sciences du Langage’, Editions du CNRS. Rice, Keren (1992) On Deriving Sonority: A Structural Account of Sonority Relationships. Phonology 9: 61–99. Rice, Keren and Peter Avery (1989) On the Interaction between Sonorancy and Voicing. Toronto Working Papers in Linguistics 10: 65–82. Rich, Furne (1963) Arabela Phonemes. In Studies in Peruvian Indian Languages I, Benjamin Elson (ed.), pp. 193–206, Summer Institute of Linguistics Publications in Linguistics and Related Fields 9, Norman, OK: Summer Institute of Linguistics of the University of Oklahoma.

260

References

Robinson, Lila W. (1975) Some Phonological Rules of Iowa-Oto (Siouan). In Proceedings of the Mid-America Linguistics Conference, pp. 439–48, Linguistics Department, University of Kansas, Lawrence. Roca, Iggy and Wyn Johnson (1999) A Course in Phonology. Oxford: Blackwell. Rosenthall, Sam (1997) Vowel/Glide Alternation in a Theory of Constraint Interaction. New York and London: Garland Publishing. Originally 1994 doctoral dissertation, University of Massachusetts, Amherst. Rosenthall, Sam and Harry van der Hulst (1999) Weight-By-Position By Position. Natural Language and Linguistic Theory 17: 499–540. Sagey, Elizabeth (1986) The Representation of Features and Relations in Nonlinear Phonology. Doctoral dissertation, MIT. Salmons, Joe (1992) Accentual Change and Language Contact: Comparative Survey and Case Study of Early Northern Europe. Stanford, CA: Stanford University Press and London: Routledge. Sapir, Edward (1965) A Grammar of Diola-Fogny. Cambridge University Press. Saunders, Ross and Philip Davis (1972) Verbal Categories in Bella Coola: Reduplication. Paper presented at the XI Conference on American Indian Languages, San Diego. Scheer, Tobias and Péter Szigetvári (2002) Unified Representations for the Syllable and Stress. Paper presented at the 10th Manchester Phonology Meeting, 23–25 May 2002, Manchester, UK. Schein, Barry and Donca Steriade (1986) On Geminates. Linguistic Inquiry 17: 691–744. Schuh, Russell (1978) Tone Rules. In Tone: A Linguistic Survey, V. Fromkin (ed.), pp. 221–56, New York: Academic Press. (2002) The Phonology and Morphology of Bade and Ngizim. Retrieved from: www. humnet.ucla.edu/humnet/aflang/Yobe/Papers/Bade_Ngizim_proseminar/00bade_ ngizim.html. Scorza, David (1985) A Sketch of Au Morphophonemics and Syntax. In Pacific Linguistics A63, pp. 215–73. Canberra: Australian National University. Seidel, August (1900) Die Sprache von Ufromi in Deutsch-Ostafrika. Zeitschrift f^r afrikanische und oceanische Sprachen 5: 165–75. Selkirk, Elisabeth (1980) The Role of Prosodic Categories in English Word Stress. Linguistic Inquiry 11: 563–606. (1984a) On the Major Class Features and Syllable Theory. In Language Sound Structure: Studies in Phonology Presented to Moris Halle by his Teacher and Students, Mark Aronoff and Richard T. Oehrle (eds.), pp. 107–36, Cambridge, MA: MIT Press. (1984b) Phonology and Syntax: The Relation between Sound and Structure. Cambridge, MA and London: MIT Press. (1990) A Two-Root Theory of Length. In University of Massachusetts Occasional Papers 14: 123–71, Elaine Dunlap and Jaye Padgett (eds.), Amherst, MA: GLSA. Sezer, Engin (1986) An Autosegmental Analysis of Compensatory Lengthening in Turkish. In Studies in Compensatory Lengthening, Leo Wetzels and Engin Sezer (eds.), pp. 227–50, Dordrecht: Foris.

References

261

Shaw, Jason (2007) Compensatory Lengthening via Mora Preservation in OT-CC: Theory and Predictions. Paper presented at NELS 38. Shaw, Patricia A. (2002) On the Edge: Obstruent Clusters in Salish. In Proceedings of the Workshop on the Structure and Constituency of the Languages of the Americas (WSCLA 7), L.Bar-el, L, Watt and I. Wilson (eds.). (2009) Inside Access: The Prosodic Role of Internal Morphological Constituency. In The Nature of Word: essays in honor of Paul Kiparsky, K. Hanson and S. Inkelas (eds.), pp. 241–72, Cambridge, MA: MIT Press. Sherzer, Joel (1970) Talking Backwards in Cuna: The Sociological Reality of Phonological Descriptions. Southwestern Journal of Anthropology 26: 343–53. Shiraishi, Hidetoshi (1998) Evidence for Prominence-based Accent Shifting in Ainu (the Saru Dialect). In Proceedings of the 117th Meeting of the Japanese Society of Linguists, pp. 266–71. Shryock, Aaron Michael (1995) WPP, No. 89: Investigating Laryngeal Contrasts: An Acoustic Study of the Consonants of Musey. US Los Angeles: Department of Linguistics, UCLA. Sloan, Kelly (1988) Bare-Consonant Reduplication: Implications for a Prosodic Theory of Reduplication. In Proceedings of WCCFL 7, Hagit Borer (ed.), pp. 319–30, Stanford, CA: The Center for the Study of Language and Information. Smith, Jennifer L. (2003) Onset Sonority Constraints and Subsyllabic Structure. Submitted to Phonologica 2002, John R. Rennison, Friedrich Neubarth and Markus A. Pöchtrager (eds.). (2005) Phonological Augmentation in Prominent Positions. New York and London: Routledge. Spaulding, Craig and Pat Spaulding (1994) Phonology and Grammar of Nankina [Data papers on Papua New Guinea languages 41]. Ukarumpa, New Guinea: Summer Institute of Linguistics. Sprouse, Ronald (1996) Vowels that Borrow Moras: Geminates and Weight in OT. In Proceedings of the North Eastern Linguistic Society (NELS) 26 (1995), Kiyomi Kusumoto (ed.), pp. 393–408, Amherst: GLSA. (1997) A Case for Enriched Inputs. Handout for paper presented at TREND VII, held 3 May 1997 at UCSC. Available as ROA-193. Steriade, Donca (1990) Moras and Other Slots. In Formal Linguistics Society of Midamerica 1, D. Meyer and S. Tomioka (eds.) pp. 254–80, Madison: Department of Linguistics, University of Wisconsin. (1997) Phonetics in Phonology: the Case of Laryngeal Neutralization. ms., UCLA. Strehlow, Theodor G. H. (1944) Aranda Phonetics and Grammar. Oceania Monographs 7, Sydney University Press. Sumner, Meghan (1999) Compensatory Lengthening as Coalescence: Analysis and Implications. In Proceedings of WCCFL 18, S. Bird, A. Carnie, J. Haugen and P. Norquest (eds.), pp. 532–44, Somerville, MA: Cascadilla Press. Svantesson, Jan-Olof (1983) Kammu Phonology and Morphology. Malmö: CWK Gleerup. Takahashi, Toyomi (1994) Constraint Interaction in Aranda Stress. In UCL Working Papers in Linguistics 6: 479–507.

262

References

Tamura, Suzuko (1996) The Ainu-Japanese Dictionary: Saru Dialect. Tokyo: Sofukan. Tang, Katrina Elizabeth (2008) The Phonology and Phonetics of Consonant-Tone Interaction. Doctoral dissertation, University of California, Los Angeles. Thurgood, Graham (1996) Language Contact and the Directionality of Internal Drift: the Development of Tones and Registers in Chamic. Language 71(1): 1–31. Topintzi, Nina (2006a) A (not so) Paradoxical Instance of Compensatory Lengthening: Samothraki Greek and Theoretical Implications. Journal of Greek Linguistics 7: 71–119. (2006b) Moraic Onsets. Doctoral dissertation, UCL. (2007) Solving the Samothraki Greek Compensatory Lengthening Puzzle. In Selected Papers on Theoretical and Applied Linguistics of the 17th International Symposium (2005) Vol. I, Eleni Agathopoulou, Maria Dimitrakopoulou and Despoina Papadopoulou (eds.), pp. 53–63, Thessaloniki: Monochromia. (2008a) On the Existence of Moraic Onset Geminates. Natural Language and Linguistic Theory 26(1): 147–84. (2008b) Word Minimality in Bella Coola as Evidence for Moraic Onsets. In the Proceedings of the 41st Annual Meeting of the Chicago Linguistic Society (Main Session), 7–9 April 2005, pp. 489–504, Chicago: CLS. Topintzi, Nina and Marc van Oostendorp (2009a) Resolving Some Paradoxes Through Binary Feature Spans. Paper presented at OCP6 (Old World Conference in Phonology), 22–24 January 2009, Edinburgh, UK. (2009b) Palatalization and Centralization in Samothraki Greek. In Proceedings of the 29th Annual Meeting of the Linguistics Department 2008, Aristotle University of Thessaloniki, pp. 385–99. Tranel, Bernard (1987) The Sounds of French: an Introduction. Cambridge University Press. (1991) CVC Light Syllables, Geminates and Moraic Theory. Phonology 8: 291– 302. Trubetzkoy, Nikolay Sergeevich (1939) Grundzüge der Phonologie. Göttingen: Vandenhoeck and Ruprecht. Vennemann, Theo (1988) Preference Laws for Syllable Structure and the Exploration of Sound Change. Berlin: Mouton. Viemeister, Neal (1980) Adaptation of Masking. In Psychophysical, Physiological and Behavioural Studies in Hearing, Proceedings of the 5th International Symposium on Hearing, G. van den Brink and F. A. Bilsen (eds.), pp. 190–98, Delft: Delft University Press. Walker, Rachel (2001) Positional Markedness in Vowel Harmony. In Proceedings of HILP 5, Linguistics in Potsdam 12, Caroline Féry, Antony Dubach Green and Ruben van de Vijver (eds.), pp. 212–32, Potsdam: University of Potsdam. Welmers, William (1962) The Phonology of Kpelle. Journal of African Languages 1: 69–93. Wetzels, W. Leo and Joan Mascaró (2001) The Typology of Voicing and Devoicing. Language 77: 207–44. Williamson, Kay (1965) A Grammar of the Kolokuma Dialect of Ijo. Second edition 1969, West African Language Monographs 2, Cambridge University Press.

References

263

Wolf, Matthew (2008) Optimal Interleaving: Serial Phonology-Morphology Interaction in a Constraint-based Model. Doctoral dissertation, University of Massachusetts, Amherst. Woodbury, Hanni (1981) The Loss of a Phoneme. International Journal of American Linguistics 47: 103–20. Yallop, Colin (1977) Alyawarra: An Aboriginal Language of Central Australia. Research and Regional Studies 10, Canberra: Australian Institute of Aboriginal Studies. Yip, Moira (1980) The Tonal Phonology of Chinese. Doctoral dissertation, MIT. (1991) Coronals, Consonant Clusters and the Coda Condition. In The Special Status of Coronals. Internal and External Evidence, Caroline Paradis and Jean-François Prunet (eds.), pp. 61–78, San Diego: Academic Press. (1992) Sonorant vs. Obstruent Codas: a Prosodic Distinction. ms., University of California, Irvine. (1995) Tone in East Asian Languages. In Handbook of Phonological Theory, J. Goldsmith (ed.), pp. 476–94, Oxford: Blackwell. (2002) Tone. Cambridge University Press. (2003) Casting Doubt on the Onset/Rime Distinction. Lingua 113(8): 779–816. Yupho, Nawanit (1989) Consonant Clusters and Stress Rules in Pattani Malay. MonKhmer Studies 15: 125–37. Zec, Draga (1988) Sonority Constraints on Prosodic Structure. Doctoral dissertation, Stanford University. (1995) Sonority Constraints on Syllable Structure. Phonology 12: 85–129. (2007) The Syllable. In The Cambridge Handbook of Phonology, Paul de Lacy (ed.), pp. 161–94, Cambridge University Press. Zewen, François-Xavier Nicolas (1977) The Marshallese Language: A Study of its Phonology, Morphology, and Syntax. Berlin: Verlag von Dietrich Reimer. Zoll, Cheryl (1998) Positional Asymmetries and Licensing. ms., MIT. Available as ROA-282.

Subject index

ALIGNσ́O, 57, 69–73 Be Moraic, 26, 33 BinSpan / BinAss(F), 118–19 Blumenfeld (2006), 51–57 coalescence, 108, 110, 116–17 compensatory lengthening, 3, 100–2, 113–15, 120–30, 169, 178 onsets as a target of, 133–34 onsets as a trigger of, 203–6 consonant length, 194–97 culminativity, 91

*μ/Ons, 16, 148–52, 153–58, 181–84, 192–94, 236 markedness moraic, 16 coda, 16, 21 onset, 16, 21, 26, 29, 190 positional, 70 metathesis, 105, 116, 118 minimal word, see word minimality mora licensing, 156–58 positional, 111, 149 *Moraic Onset, see *μ/Ons nasalization, 41–43, 49, 110

doubled consonants, 196 dual nature of sonorants, see voice, and sonorants economy, 24–25 epenthesis, 58, 66, 116, 120, 124, 140, 205 functional approaches, 36, 193, 236 fundamental frequency (F0), 19, 168, 210, 244 fusion, see coalescence geminate, 164, 166, 194–97 fake, see doubled consonants initial, 167–68, 179–80, 199 onset, 184–91, 199 typology, 192–94, 197–98 glide formation, 123–24 Gordon (2005), 241–45 gradience, 79 Hayes (1989), see under theory, standard moraic iambic, 70–71, 96, 176

264

onset and compensatory lengthening, see under compensatory lengthening and geminates, see geminate, onset and word minimality, see word minimality moraic coerced, 11–14, 26–29 distinctive, 11–12, 14, 29–32 geminated, 12–13 *Onset/r, 114–19 onset weight, 8–11, 175 coerced, see onset, moraic, coerced distinctive, see onset, moraic, distinctive rarity of, 130–33, 193, 213, 235–36 P-Dep-μ, 111–12, 149, 153, 179–80 peak hierarchy, 173 pitch perturbation, see voice, -induced pitch perturbation position correspondence (PosCorr), 107–9, 123–25, 128–30, 179–80 prenasalization, 123–25

Subject index prominence, see also weight, as a function of prominence 237–41

265

spreading, 18–21, 44, 118 tonogenesis, 176–77 trochaic, 65, 70–71, 85, 87, 96

quantity, see under weight unsyllabified obstruents, 140–44 rarity argument, 7 reduplication, 140–42 reliability of contrast, 24–25 Richness of the Base, 26, 52, 101, 111, 121, 147–48, 160 root maximality, 142–44 RotB, see Richness of the Base *S-L VH (*Super-Long Vocalic Hiatus), 115–19, 132–33 Smith (2005), 238–41 sonority hierarchy, 23, 25, 240 influence on moraicity, 13–16, 50 -sensitive-stress, 96, 173, 239–40 stress clash, 65, 172 shift, 9, 38, 46–48, 56–57, 60–61, 70–74, 85–87, 89–90, 93–97, 169, 171, 243 syllable constituent well-formedness, 15, 22 onsetful, see under syllable, onsetless onsetless, 10, 39, 48, 57, 59–62, 64–69, 71–75, 84, 185, 226, 243 super-heavy, 65, 73, 227 theory CV, 4–6, 234 modified moraic (as discussed in the book), 7, 33, 234 of onset weight, see theory, modified moraic (as discussed in the book) standard moraic, 5–8 X-slot, 4–6, 234 tone and onsets, 207–13 and voice, see voice, and tone

voice and obstruents, 16–18 and sonorants, 18–21, 27–28, 44–45, 83, 85–86, 137 and sonority, 23–25 and stress, 22–23 and tone, 16–21, 25 contrast, 17, 24–25, 28, 40, 51–52, 208, 211 -induced pitch perturbation, 16, 19, 22, 24, 98, 231 unary vs, binary feature, 17–18, 21–22 unspecified for, see voice, and sonorants voicing, see under voice weight as a function of prominence, 241–45 coda, 13 nucleic, 13, 175 onset, see onset weight syllable, 2–5, 7 symmetrical, 233, 235 weight-by-position, 26 weight-to-stress principle, 46 gradient, 78–80 nucleic, 78 prosodic word, 90, 171 word minimality, 137–38, 146–49, 152–56, 160–61, 178, 181 segmental account of, 162–63 WSP, 46–51, 65–67, 78–82, 89–90, 187, see also weight-to-stress principle WSPPrWd, 90–95, 171, 174–75, see also weight-to-stress principle, prosodic word WSP(N), 78, 80–82, see also weight-to-stress principle, nucleic

Language index

Agwamin, 58 Ainu, 68–69 Alamblak, 235 Aljutor, 173 Alyawarra, 57–60, 74 Anuak, 205 Arabela, 83–95 Arabic, 31, 129 Levantine, 195, 198 Aranda, 9–10, 58–59, 74 Bade, 17–21, 26, 44 Banawá, 58–60, 73, 244 Basque, 237 Bella Coola, 28–29, 138–52, 158–63, 199 Bellonese, 213–14 Berawan, 222–23 Berber Tashlhiyt, 24, 212 Bernese, 27, 195–96, 198 Bislama, 224–25 Brahui, 78 Campa Axininca, 69 Carib Surinam, 90 Catalan, 25, 239 Cayuga, 204 Cebuano, 237 Cham, 176 Chickasaw, 78 Chinese Songjiang, 17 Chukchee, 131, 173 Chuvash, 174, 235 Cuna, 28

266

Dakota, 22, 24 Dalabon, 3 Damin, 216–18 Danish, 210 Delaware, 235 Digo, 18, 21 Diola Fogny, 130 Djabugay, 90 Djapu Yolngu, 28 Djeebbana, 189 Dutch, 25, 57, 63–68, 73, 175 Ekegusii, 118 English, 38, 58, 63, 78, 114, 137, 224, 235–36, 239 Middle, 237 Old, 74 Estonian, 74, 223 Evenki, 137 Fijian, 237 Finnish, 28, 74 French, 69, 110, 224 Gadsup, 40 German, 25, 124, 236 Swiss, 27 Golin, 22 Goroa, 38, 96 Greek Ancient, 3, 5–6, 210 Cappadocian, 117 Homeric, 210 Samothraki, 102–7, 113–22, 203 Standard Modern, 9, 53 Gujarati, 24 Gyore, 206

Language index Hausa, 4, 196, 210 Hebrew Tiberian, 175 Hopi, 2, 6 Huariapano, 38, 90, 96 Hungarian, 11, 74, 195, 197–98 Icelandic, 131 Ijo Kolokuma, 24 Inga, 90 Iowa-Oto, 58, 61–63, 73–74 Italian, 72, 237 Romanesco, 205 Japanese, 4–6, 14, 27, 114, 237 Javanese, 174 Jingpho, 17 Juma, 58, 60, 70 Kabardian, 235 Kaingang, 71–72 Kammu, 17, 212, 221 Kara, 96 Karo, 9–11, 23–24, 26–27, 39–57 Kashaya Pomo, 175 Kashmiri, 38, 78, 223 Kaytetj, 58 Ket, 137 Kinyarwanda, 123 Kiowa, 210, 235 Klamath, 78, 131 Koasati, 237 Kobon, 173 Komi Ižma, 237 Korean Middle, 176 Kotoko, 20–21, 25 Koya, 196 Kpelle, 207–9, 211–12 Kukatj, 4 Kuku-Thaypan, 58 Kunama, 210 Kwakwala, 13, 50 Lak, 14, 27 Lamalama, 58 Lango, 206 Lardil, 131

267

Latin, 3, 7, 78, 235, 237 Lenakel, 2–3, 6, 235–36 Linngithig, 58 Lithuanian, 210 Luganda, 123, 210, 218–19, 227 Maasai, 205 Madurese, 195 Malagasy, 235 Malay Pattani, 10, 30, 167–77, 198 Malayalam, 196 Manam, 237 Mandarin, 177, 235 Mari, 173, 235 Marshallese, 132–33, 186–88, 194, 214–16 Masa, 17 Maung, 90 Mazatec Jalapa, 24 Mbabaram, 58 Miya, 18, 21 Mocha, 137 Mohawk, 204–5, 235 Mokilese, 4, 237 Mongolian Khalkha, 235 Musey, 210–12 Musqueam, 235 Nankina, 225–27 Nez Perce, 14, 27 Nganasan, 96, 171 Niuafo`ou, 240 Ntcham, 205 Ocaina, 14, 27 Ojibwa, 14, 27, 235 Oneida, 204 Onondaga, 203–5 Pirahã, 9–11, 23–25, 71, 75–83, 240–41, 243–44 Piro, 126–30, 156, 199 Pitta-Pitta, 137 Ponapean, 237 Popoloca Eastern, 191 Portuguese, 110

268

Language index

Proto-Austronesian, 132 Puluwat, 189–90 Qiang, 176 Russian, 25, 53 Sahaptin Yakima, 162–63, 170, 174 Sedang, 24 Selayarese, 114 Selkup, 196 Seneca, 90, 204 Sentani, 96 Sinhalese, 74 Siswati, 18, 21 Somali, 22, 24 Spanish, 69 Suma, 18, 21 Swedish, 4 Tangale, 131 Temiar, 212, 221 Tepehuan Southeastern, 129 Totonac, 14, 27

Trique, 133–34, 188, 194 Trukese, 10, 177–84, 199 Tübatulabal, 196 Tukang Besi, 196 Tümpisa Shoshone, 207 Turkana, 206 Turkish, 3, 100, 235 Umbindhamu, 58 Umbuygamu, 58 Uradhi, 58 Votic, 78 Wargamay, 235–36 Yabem, 17 Yaka, 18, 21 Yakut, 14, 27 Yiddish, 170 Yil, 174 Yoruba, 114 Yukulta, 90 Yupik, 170 Zoque, 129