Larry M. Hyman and Frans Plank (Eds.

)
Phonological Typology
Phonology and Phonetics

Edited by
Aditi Lahiri

Volume 23
ISBN 978-3-11-044970-9
e-ISBN (PDF) 978-3-11-045193-1
e-ISBN (EPUB) 978-3-11-044992-1
ISSN 1861-4191

Library of Congress Control Number: 2018934237

Bibliografische Information der Deutschen Nationalbibliothek

The Deutsche Nationalbibliothek lists this publication in the Deutschen
Nationalbibliografie; detailed bibliographic data are available on the internet http://d
nb.dnb.de.

© 2018 Walter de Gruyter GmbH, Berlin/Boston

www.degruyter.com
Contents

Preface
Contributors

Larry M. Hyman
What is phonological typology?

Frans Plank
An implicational universal to defy: typology ⊃ ¬ phonology ≡ ph
onology ⊃ ¬ typology ≡ ¬ (typology ᴧ phonology) ≡ ¬ typology v
¬ phonology

Paul Kiparsky
Formal and empirical issues in phonological typology

Ian Maddieson
Is phonological typology possible without (universal) categorie
s?

Jeffrey Heinz
The computational nature of phonological generalizations

Anthony Brohan and Jeff Mielke

Frequent segmental alternations in P-base 3

Aditi Lahiri
Predicting universal phonological contrasts

B. Elan Dresher, Christopher Harvey, and Will Oxford

Contrastive feature hierarchies as a new lens on typology

Ellen Broselow
Laryngeal contrasts in second language phonology

Tomas Riad
The phonological typology of North Germanic accent

Carlos Gussenhoven
Prosodic typology meets phonological representations

Subject Index

Language Index

Author Index
Preface

This volume seeks to bring together two separate enterprises, typology and
phonology, which have often gone their separate ways, even when
apparently addressing each other. Although there has always been
phonology in the many approaches to typology, there has always been far
less of it than there has been of morphology and syntax. Particularly in
recent times, during which phonological theory has flourished in many
colours, the phonology in typological circles has centered either around
segment inventories and basic phonotactics or limited itself to crudely
categorising labels such as “syllable- vs. stress-timed languages”,
something which would barely pass muster with phonological theorists,
whose concerns have been “deeper” and usually more formal. On the other
hand, despite the numerous and diverse languages that would typically
inform phonological theorising, typological questions as such were rarely of
major consequence, if raised at all.
If one sees value or at any rate promise in typology, the research
programme for discovering and accounting for order in diversity, and if you
see no principled reasons to doubt that linguistic diversity is about as
orderly in phonology as in syntax and morphology (or at any rate
inflection), this state of affairs is regrettable. Since this is how we see it, we
felt obliged to lend a hand towards improving relationships.
In organised typology, the success of our efforts was limited. Though
long involved in various capacities in the Association for Linguistic
Typology (ALT), it looks like we and a few fellow campaigners have not
been able to significantly raise the profile of phonology at ALT’s biennial
conferences or in its other activities. The two of us hit rock bottom when a
workshop to boost phonology in typology that we suggested as part of an
ALT conference a few years ago did not find favour with the programme
committee, for reasons we ourselves found unconvincing. Reassuringly, the
phonological record of Linguistic Typology, the journal we have helped to
run for two decades, is better, but still comparatively modest.
 Perhaps we had gotten hold of the wrong end of the stick, trying to
proselytise in typological circles. Changing tack, for the present volume we
sought out linguists who define themselves and are perceived first and
foremost as phonologists. Though rarely meeting at typological get-
togethers and not so far published in LT, we were still expecting typological
awareness rather than complete innocence, because our remit for them was
as follows: Do typology! Present a sample piece of work appropriate for a
workshop and subsequent publication as you think it can and ought to be
done in state-of-the-art phonology!
And here we go. After two scene-setting and background-providing
contributions from the editors (Hyman; Plank), the phonologists assembled
here, typologically aware or indeed expert, address metatheoretical issues of
what it means to do phonological typology (Kiparsky; Maddieson).
Through different methodologies, they explore the possibilities of and
limitations on segmental alternations (Heinz; Brohan & Mielke), seek
system, not in segment inventories as such, but in featural contrasts, and
find variability, though not randomness, of contrastive systems in language
change and language acquisition (Lahiri; Dresher, Harvey, & Oxford;
Broselow) and seek limitations on diversity and change concerning
microvariation in tonal accent systems in North Germanic (Riad). An effort
at conceptual clarification of what it is that can be typologised in prosodic
typology – segments, constituents, their alignments – concludes
(Gussenhoven).
Typology’s subject matter is vast: for EVERYTHING about language –
units, paradigmatic systems, rules for and constraints on forming prosodic
constituents (syllables, feet . . . anything syntagmatically complex without
or with meaning) – we need to ascertain whether it is variable or invariable,
and if it is variable we want to know whether it varies independently or co-
varies with anything else. Despite the extensive knowledge accumulated in
phonology over the past century, phonological typologists still have
considerable work cut out for them for an unforeseeable time to come. The
present collection will have served its purpose if it gives them
encouragement and guidance. Naturally, phonological awareness or indeed
expertise of none-too-basic a nature will be an asset for those planning to
join in this enterprise: in phonology no less than in syntax and morphology,
generalising is futile if the particulars over which one generalises are
inadequately analysed.
 This volume derives from a workshop in Somerville College, University
of Oxford, on 11–13 August 2013 (funded by a European Research Council
Advanced Investigator Grant to Aditi Lahiri). Subsequent to the preparation
of this volume, a survey monograph of the same title appeared, by Matthew
K. Gordon (Oxford University Press, 2016), adding another voice to the
small chorus of performers striving to bring the adjective into harmony with
the noun (or vice versa) of our shared title. A big thanks from all of us who
were present goes to Aditi Lahiri, for there is no better host. There also is
no better series editor for this sort of thing either, cruel but fair with her
editors and authors: A thank you to our dear friend and colleague Aditi
Lahiri in this capacity, too.
We might as well dedicate this collection to her on the occasion of a
round birthday that must not go unmarked . . .
Larry M. Hyman
Frans Plank
Contributors

Anthony Brohan
Google, Mountain View, CA 94043, USA

Ellen Broselow
Department of Linguistics, Stony Brook
University, Stony Brook, NY 11794-4376, USA
[email protected]

B. Elan Dresher
Department of Linguistics, University of
Toronto, Toronto, Ontario, Canada M5S 3G3
[email protected]

Carlos Gussenhoven
Department of Linguistics, Radboud
University, PO Box 9103, 6550 HD Nijmegen, The Netherlands
[email protected]

Christopher Harvey
Department of Linguistics, University of
Toronto, Toronto, Ontario, Canada M5S 3G3
[email protected]

Jeffrey Heinz
Department of Linguistics and Institute of
Advanced Computational Science, Stony Brook
University, Stony Brook, NY 11794-4376, USA
[email protected]

Larry M. Hyman
Department of Linguistics, University of
California at Berkeley, Berkeley, CA 94704, USA
[email protected]
Paul Kiparsky
Department of Linguistics, Stanford University,
Stanford, CA 94305-2150, USA
[email protected]

Aditi Lahiri
Faculty of Linguistics, Philology and Phonetics,
University of Oxford, Oxford OX1 2HG, UK
[email protected]

Ian Maddieson
Department of Linguistics, University of New
Mexico, Albuquerque NM
[email protected]

Jeff Mielke
Department of English, North Carolina State
University, Raleigh, NC 27695-8105, USA
[email protected]

Will Oxford
Department of Linguistics, University of
Manitoba, Winnipeg, Manitoba, Canada R3T 5V5
[email protected]

Frans Plank
Somerville College, University of Oxford,
Oxford OX2 6HD, UK
[email protected]

Tomas Riad
Institutionen för svenska och flerspråkighet,
Stockholms universitet, 106 91 Stockholm,
Sweden
[email protected]
Larry M. Hyman
What is phonological typology?
Whatever typology is, it is on a roll at the moment and likely to continue.
(Nichols 2007: 236)

Abstract: In this contribution I raise the question of what phonological

typology is, can, or should be. I start by asking what linguistic typology is
and then turn to the problem: despite the intellectual overlap, there is rare
cross-communication in the study of sound systems by phonologists vs.
typologists. Despite earlier contributions by Trubetzkoy, Jakobson,
Martinet, Greenberg and others, and its inclusion in even earlier efforts
towards “holistic” typology, phonological typology is often
underrepresented or even excluded in typology textbooks. At the same time,
many, if not most phonologists do not see a difference between
phonological typology and crosslinguistic (formal) phonology. As a result,
they often address issues of comparison without awareness of the field of
typology and with little involvement in the foundational and
methodological questions/ controversies peppering the pages of Linguistic
Typology, e.g., its concern with distributions, whether pre-established
categories exist etc. I argue for a properties-based approach to typology,
showing that phonology has always been – and should remain – basically
typological in it concerns.

1 What is typology? What is phonology?

The purpose of this paper is to address the question of what phonological
typology is, can, or should be. To do so, one has to consider its relationship
both to typology and to phonology in general. Such a task is complicated by
at least three factors. First, there is no agreement on what typology is, let
alone phonological typology. In an article entitled “What, if anything, is
typology?”, the then president of the Association for Linguistic Typology
wrote:
 Typology has the hallmarks of a mature discipline: a society, conferences, journals,
books, textbooks, classic works, a founding father [Joseph H. Greenberg], and people
who are called and call themselves typologists. (Nichols 2007: 231)

While most typologists would probably self-identify as studying the

similarities vs. differences among languages, Nichols goes on to say that
“despite these conspicuous identifying marks”, typology should not be
recognized as a subfield of linguistics, but rather as “framework-neutral
analysis and theory plus some common applications of such analysis (which
include crosslinguistic comparison, geographical mapping, cladistics, and
reconstruction)” (p. 236). On the other hand, linguists who work in specific
formal frameworks may engage in crosslinguistic comparison, but typically
self-identify as syntacticians, morphologists, phonologists, etc. as they have
less interest in issues of geography, language classification and history.
The second problem in characterizing phonological typology is that
phonology is no longer the unified subfield that it once was. The following
assessment appears in a recent review of the multivolume Blackwell
companion to phonology (van Oostendorp et al. 2011):
Phonology is changing rapidly [. . .] Some phonologists collect the evidence for their
theories using introspection, fieldwork and descriptive grammars, while other trust only
quantitatively robust experimentation or corpus data. Some test phonological theory
computationally [. . .] whereas others prefer to compare theories on conceptual grounds
[. . .] (Gouskova 2013: 173)

Gouskova goes on to observe that the diversification within phonology has

become so great that “it is becoming harder for phonologists to talk to each
other, for who can be a computer scientist, phonetician, neurolinguist and
expert in adjacent fields such as morphology and syntax at the same time as
having a command of the extensive literature on phonology-internal
argumentation and phonological typology?” (p. 173).
Finally, whether typology and phonology are coherent subfields or not,
there has been precious little interaction between the two groups of
scholars. Most typologists do not work on phonology per se and usually cite
phonological examples only en passant, if at all (there is for example no
phonology in Whaley’s (1997) Introduction to typology). For their part,
phonologists frequently invoke typology, but without participation in the
society, conferences, journals etc. referred to above. While typology is
currently centered around crosslinguistic morphosyntax, phonology has
been transitioning from a descriptive/analytical to experimental field.
Slightly oversimplifying, “traditional” phonology from the time of the
phoneme has been concerned with the underlying structures needed to
account for the properties of sound systems. The methodology has largely
consisted of phonological argumentation on how best to analyze a wide
range of crosslinguistic phenomena. Given that phonology is part of
grammar, this naturally includes the interfaces of phonology with both
morphology and syntax, where the connection to grammatical typology
should be even more clear. However, today’s phonologist is more likely to
be involved in laboratory techniques where the methodologies are
instrumental, experimental, statistical, and computational. To the extent that
the questions focus on how what is produced and how what is in the signal
relates to the speaker’s mind, ears, and vocal tract, the results may appear
even more removed from the morphosyntactic core of the typology
movement.
This non-intersection is highly atypical when compared with the
interests of the founders of both fields. Joseph Greenberg’s foundational
work on typology and universals touched on virtually all aspects of
phonology, e.g., syllable structure (Greenberg 1962, 1978), distinctive
features (Greenberg, Jenkins, & Foss 1967), vowel harmony (Greenberg
1963), nasalized vowels (Greenberg 1966), glottalized consonants
(Greenberg 1970), word-prosodic systems (Greenberg & Kaschube 1976),
and so forth. His historical work on African languages also included
phonological reconstruction, e.g., of tone in Proto-Bantu (Greenberg 1948)
and labial consonants in Proto-Afro-Asiatic (Greenberg 1958). It is thus
striking how few major morphosyntactic typologists show an active
engagement with phonology today (but see Evans 1995, Donohue 1997,
Haspelmath 2006, Plank 1998, among others).
On the other side, the non-involvement of phonologists with the field of
typology stands in stark contrast to the fact that phonology has been
typological from its very beginning. In fact, the very notion of the phoneme
is a typological one, as evidenced in the following oft-cited passage:
[. . .] it almost goes without saying that two languages, A and B, may have identical
sounds but utterly distinct phone[mic] patterns; or they may have mutually
incompatible phonetic systems, from the articulatory and acoustic standpoint, but
identical or similar [phonemic] patterns. (Sapir 1925: 43)
The frequent comparison of allophonic aspiration in English with phonemic
aspiration in Thai, Korean, etc. is inherently typological, a statement about
how different sound systems can “phonologize” the same or similar
phonetic substance. Ever since the introduction of the phoneme
phonologists have been unified in recognizing that phonological
representations are distinct from the observed phonetics. In the 1930s the
Prague School developed the phonetics-phonology distinction further,
emphasizing how phonological systems differ in their structural properties.
Trubetzkoy’s (1939) Grundzüge der Phonologie is both a highly theoretical
and a thoroughly typological work. As any textbook in phonology would
explain, a specific phonetic distinction may have a quite different status in
different languages. A difference in voicing as in [t] vs. [d] may have a
distinctive (paradigmatic) function in distinguishing between morphemes,
e.g., bit vs. bid in English. It may instead have a demarcative (syntagmatic)
function helping to determine where one is in the spoken chain. In Basaá
there is a single set of underlying stops / P, T, K/, which are realized [p, t, k]
in stem-initial position vs. [b, d, g] (~ [β, r, ɣ]) stem-internally (Hyman
2003: 259). As a result, the prefixed word /ɓa-Tâ/ ‘fathers’ is pronounced
[ɓatâ] while the suffixed word /ɓáT-â/ ‘gather’ is pronounced [ɓádâ] (~
[bárâ]). A third possibility is that the voicing difference is non-distinctive or
allophonic. A well-known case of this comes from Korean, where /t/ is
realized [d] intervocalically. Thus, when /su/ ‘water’ and /to/ ‘way’ are
compounded, the result is [sudo] ‘waterway, waterworks’. Among the other
possibilities are free variation, as when the final /t/ of English bit is either
released or not and what Trubetzkoy calls the expressive function, where
differences indicate such things as social identity or attitude of the speaker,
e.g. the “expressive” aspiration in the phrase je t[h]’aime (Martinet 1960).
Once the phonological contrasts are established, a major component of
Trubetzkoy’s Grundzüge was to provide a typology of the contrasts found
in one vs. another system. He classifies distinctive contrasts according to
three different factors.
First, their relationship to the entire system of contrasts. This refers to
the number of segments in the set. For example, the set of oral labial stops
can be bilateral (/p/ vs. /b/) or multilateral (/p/ vs. /ph/ vs. /b/), depending on
the language. The relationship to the system is said to be proportional, if
other segments exhibit a parallel relation, e.g., bilateral /t/ vs. /d/ or
multilateral /t/ vs. /th/ vs. /d/. On the other hand, a contrast such as /l/ vs. /r/
is said to be isolated, since there is no other pair of phonemes which
realizes a parallel contrast.
Second, the relationship between the contrasting segments, which can
be privative, gradual, or equipollent. In a privative contrast one member has
a “mark” which is lacking in the other. Thus in a /ph/ vs. /p/ contrast, /ph/
has aspiration, while /p/ lacks it. Gradual contrasts refer to scalar features
such as the vowel height differences between / i, e, ɛ, æ/ or the pitch height
differences between High, Mid and Low tone. In equipollent contrasts the
segments are considered “logically equivalent”. An example is labial /p/ vs.
alveolar /t/, where each has a logically equivalent but different upper and
lower articulator. Trubetzkoy is careful to distinguish “logically” vs.
“actually” privative, gradual and equipollent, since it will depend on the
system. While it makes no sense to think of /p/ vs. /t/ as differing on a
continuous scale (they involve different articulators), Trubetzkoy might
consider the relation to be privative if a language were to have only labial
and alveolar places of articulation. In this case /p/ could be said to have a
labial mark while /t/ lacks it.1
Third, the extent of the contrast. This refers to whether the contrast is
realized in all environments or whether there are contexts in which the
contrast is neutralized. A well-known example of this is German final
devoicing, whereby /rat/ ‘advice’ and /rad/ ‘wheel’ are both realized [rat] in
isolation. Another is flapping in American English, e.g. metal and medal,
both pronounced [mɛɾl̩ ] (cf. etymologically related metallic and medallion
with [tʰ] and [d]).
The above examples not only establish that early modern phonology
was heavily steeped in typology, but that the founders had two different
ideas of phonological typology, depending on whether the starting point is
substance vs. form. In the first approach one asks how different systems
exploit a particular phonetic property. In the examples cited, it was seen that
obstruent voicing can be distinctive, demarcative or allophonic. The
possibilities can be more extensive, as in the case of nasality. As
summarized in (1), there are at least five possibilities for how nasality may
be underlyingly contrastive in a language (cf. Cohn 1993; Clements & Osu
2003):
In addition to the above distinctions, languages may vary in whether they
contain voiceless nasals, prenasalized or nasally released consonants, as
well in whether the contrasts are found on all nasalizable consonants (e.g.,
including liquids and glides) and on all vowels. Similar substance-directed
typology can be done with virtually any phonetic feature or property, e.g.,
voicing, aspiration, rounding, and so forth (cf. (1)). Still being substance-
directed, a typologist will likely be interested in how one vs. another of
these properties is distributed in the languages of the world, whether by
genetic affiliation or by geography.
The second approach to phonological typology is form-directed: in this
case the analyst explores the logical properties of a specific model. The
above examples from Trubetzkoy fall into this category, as he was
interested in the logical differences in the nature of the contrasts that his
model of phonology recognized. It mattered less that /l/ and /r/ differed in
laterality or rhoticity than the fact that they constitute an isolated bilateral
contrast in any language which has only these two liquids. This second,
form-directed approach finds reincarnation in virtually every model, if not
every proposal in phonological theory. Since early generative phonology
proposed ordered rules (Chomsky & Halle 1968), it was only natural that a
form-directed typology should develop how these rules apply to forms and,
in so doing, how they affect each other: a phonological rule could apply to a
form left-to-right, right-to-left, simultaneously, and cyclically or non-
cyclically. Earlier applying rules could be in feeding, bleeding,
counterfeeding, and counterbleeding relationships, creating and/ or taking
away inputs to which later rules could apply (Kiparsky 1968; Kenstowicz &
Kisseberth 1977). More recently, within optimality theory (Prince &
Smolensky 1993), all of the possible rankings can be exhaustively
computed in a “factorial typology” (cf. Gordon 2007). In short, almost any
formal property can be “typologized” in terms of its logical parameters.
2 Phonology vs. typology
In both of the above approaches to phonological typology there has been a
deep commitment to the idea that phonetics and phonology are distinct from
each other. As Buckley (2000: 2) puts it, “becoming divorced from the
phonetics is the very essence of phonology”. The key goal of phonology has
been to determine what is a possible phonological system. This has meant
both determining the universal properties of sound patterns in languages as
well as what is going on in the heads of speakers with respect to these
sound patterns. While these goals are directed towards the quest for
universals, the traditional approach has been to seek universals through the
study of language particulars, which can be quite diverse. Determining how
languages can vary within such confines has been the central goal of
traditional typology, where there has been a distinction (confusion?)
between two views of what typology is about. The first is that it concerns
the classification of languages into “types”. Thus, Hagège (1992: 7) defines
typology as “a principled way of classifying the languages of the world by
the most significant properties which distinguish one from another”. While
it is harder to find explicit definitions of phonological typology, Vajda’s
(2001) posting coincides with this view: “it is possible to classify languages
according to the phonemes they contain [. . .] typology is the study of
structural features across languages. Phonological typology involves
comparing languages according to the number or type of sounds they
contain”.2 The other view, which I have termed property-driven typology
(Hyman 2009: 213, 2012: 371), is that typology is not about the
classification of languages but rather the characterization of linguistic
properties: “Typology, thus, is not so much about the classification of
languages as about the distributions of individual traits – units, categories,
constructions, rules of all kinds – across the linguistic universe; these
distributions, not languages as such, are the primary objects of comparison”
(Plank 2001: 1399). Although I will come back to the issue of distributions
as a crucial ingredient of typology, note for now that Greenberg (1974: 14)
also explicitly recognizes the above two views: “all synchronic typologies
have this Janus-like nature in that the same data can be utilized either for a
typology of linguistic properties or a typology of individual languages”.
One reason why there has been so little interaction between typologists
and other linguists has been common misconceptions. Nichols (2007: 233–
234) debunks the following four misunderstandings about typology,
presumably including phonological typology:
(i) typology deals with only superficial grammatical phenomena, while
formal grammar deals with deeper abstraction
(ii) typology usually or often uses large surveys of hundreds of languages
(iii) in typology, explanations or theory are usually functionalist
(iv) the main theoretical constructs of typology are the implicational
correlation and the implicational hierarchy

Concerning the first misconception, Nichols goes on to cite the following,

to which I would add her own head- vs. dependent marking typology
(Nichols 1986):
I see no difference in analytic or theoretical profundity or abstraction between
generative parameters and original contributions of typology such as direct object vs.
primary object (Dryer 1986), verb-framed vs. satellite-framed lexicalization patterns
(Talmy 1985, Slobin 2004), various aspects of alignment (e.g., Dixon 1994, Dixon &
Aikhenvald (eds.) 2000), differential object marking (Bossong 1998, Aissen 2003),
referential density (Bickel 2003), and others.

She concludes that most typologists do not exploit large databases, many
(including herself) are not functionalists, and finally, implicational
statements are “a convenient format for presenting and testing results [. . .]
[but not] the be-all and end-all of typology”.
In fact, typologists disagree on a number of issues, including these:

(i) whether typology is a field:

what we call typology is not properly a subfield of linguistics but is simply framework-
neutral analysis and theory plus some of the common applications of such analysis
(which include crosslinguistic comparison, geographical mapping, cladistics, and
reconstruction). (Nichols 2007: 236)

(ii) whether it has internal subfields:

Linguistic typology includes three subdisciplines: qualitative typology, which deals

with the issue of comparing languages and within-language variance; quantitative
typology, which deals with the distribution of structural patterns in the world’s
languages; and theoretical typology, which explains these distributions. (Wikipedia
“Linguistic Typology”)
(iii) whether typology necessarily involves the quest for universals (or is
about diversity):

the goal of typology is to uncover universals of language, most of which are universals
of grammatical variation. (Croft 2003: 200)

(iv) what the role of theory should be in typology:

The hypothesis that typology is of theoretical interest is essentially the hypothesis that
the ways in which languages differ from each other are not entirely random, but show
various types of dependencies (Greenberg 1974: 54)

A traditional typologist might embellish, but presumably not object to,

Evans & Levinson’s (2010: 2740) statement that “[. . .] the goal of
linguistics is [. . .] to explain why languages have the properties they do”
(vs. the goal of linguistics is to explain how a speaker with a finite and
limited exposure can produce an infinite number of new sentences, how a
child by the age of two can do such-and-such, etc.). Be this as it may, let me
return to the view that typology is something which phonologists do all the
time (Hyman 2007). As I pointed out above via the quote from Sapir
(1925), phonology has always been explicitly crosslinguistic. Thus, both
phonological theory and phonological typology are concerned with how
languages encode the same phonetic substance into structured sound
systems:
Phonological typology is a classification of linguistic systems based on phonological
properties. There are four basic kinds of typology: ‘areal’ or ‘genetic’ typologies;
typologies based on ‘surface phonological properties’; typologies based on some
‘underlying phonological property’; and ‘parametric’ typologies [. . .] In addition,
phonological typology can refer to the classification of the elements that make up a
phonological system. For example, articulatory descriptors like ‘velar’ and ‘labial’
form part of a typology of speech sounds. (Hammond 2006: 523)

The inseparability of phonology and typology continues unbroken right up

to current optimality theory:
One of the most compelling features of OT, in my view, is the way that it unites
description of individual languages with explanation of language typology. As a
phonologist, I have always been impressed and sometimes overwhelmed by how the
complexity and idiosyncrasy of each language’s phonology is juxtaposed with the
 clarity and abundance of solid typological generalizations. Even though this is arguably
the central research problem of phonology and of linguistic theory in general, progress
in consolidating description and explanation has at best been halting and occasionally
retrograde. (McCarthy 2002: 1)

The fundamental assumption of OT that constraint ranking varies from language to

language has provided fertile ground for typological research in phonology. (Gordon
2007: 750)

Concerning the relation to phonetics, phonological analysis has always been

concerned with levels of representation, specifically with establishing the
nature of underlying representations and how these are brought to the
surface (by rules, input/ output conditions, etc.). While some take a single-
level inventory approach to phonological typology, a meaningful
PHONOLOGICAL typology must also be concerned with input-output relations
and the notion of structural contrast. Typologies such as those found in
Trubetzkoy (1939) or Hockett (1955) could not otherwise be possible.
There is no clear division between phonological typology and phonological theory.
Given their shared concern with the nature of phonological systems, one can’t do
insightful typology without addressing the same analytical issues that confront
phonological theory. Throughout the history of phonology, the two have been
inseparable both in principle and in practice. (Hyman 2007: 265)

In (1) above I provided a typology of the underlying representations

nasality can have in different phonological systems. Similarly, (2) shows
how different languages underlyingly systematize or “structure” Front and
Round “color” contrasts:

The systems in (2a, b) have triangular vowel systems with underlying front
unrounded and back rounded vowels, while (2c) represents a vertical central
vowel system with front and round features restricted to consonants (to
which the centralized vowels typically assimilate). (2d) represents a vowel
harmony system where some vowels are specified, others unspecified for
Front and Round. Finally, as in the case of nasality, Front and Round can be
prosodies on whole morphemes or words. Recall from (1) that some
languages lack nasality entirely. The situation is different concerning Front
and Round: while two languages (Qawasar and Yessan-Mayo) out of the
451 languages in the UPSID database (Maddieson & Precoda 1990;
Maddieson 1991) lack a front vowel, both have the palatal glide /y/. Of the
four languages (Jaqaru, Alawa, Nunggubuyu, and Nimboran) which lack a
round vowel, only Nimboran also lacks the labiovelar glide /w/ and hence
does not exploit the feature Round at all. (It is likely that a language will
turn up that in parallel fashion does not exploit the feature Front.) No
language has thus far been cited which fails to phonologize both Front and
Round.
This does not necessarily mean that there will be a total lack of nasality,
palatality, or rounding in phonetic outputs. Examples such as (1) and (2)
illustrate that phonological typology cannot be about surface outputs alone
(for which we might distinguish PHONETIC typology). One has to make a
choice of level, which is particularly problematic in the case of tone
systems. For example, Ik (Heine 1993) and Kom (Hyman 2005) both have
underlying / H, L/ but a third [M] (mid tone) on the surface which they
derive by the following rules:

Since the trigger H may drop out after conditioning L tone raising in Ik, and
similarly, the trigger L can drop out after triggering H tone lowering in
Kom, these languages have two underlying-contrastive tone heights / H, L/,
but three surface-contrastive tone heights [H, M, L]. Are these 2- or 3-
height systems? The only adequate approach is to typologize on the basis of
the relation between underlying and surface contrastive elements, i.e., both
Ik and Kom have a 2→3 tone-height system.

3 Property-driven phonological typology

In this section I want to present the arguments in favor of basing
phonological typology on properties rather than (whole) languages. There
are at least four reasons to resist the temptation to taxonomize languages
into “types” (Hyman 2012, 2015). First, this gives the impression that the
the labels are mutually exclusive. A good example is the stress- vs. tone
language distinction, about which van der Hulst (2011: 12) writes: “Hyman
[2009] [. . .] reduc[es] the typology of word prosodic systems to tone
languages and stress languages”. Although the work in question recognizes
two independent properties Tone and Stress-Accent, which produce four
situations, as in (4), what van der Hulst really meant to say is that I do not
recognize a third prosodic property called “pitch-accent”.

A second reason to avoid labeling language types is that this gives the
impression that there is a unique taxonomy. Consider the following
hypothetical exchange over whether German should be classified with
English vs. French on the basis of its vowel system. To illustrate, consider
the hypothetical exchange in (5):

(5) Typologist #1: German should be classified with English as a “tense-

lax vowel language”, since both contrast / i, u/ vs. / ɪ, ʊ/ (etc.), as
opposed to French.
Typologist #2: No! German should be classified with French as a
“front-rounded vowel language”, since both have /ü, ø/, as opposed
to English.
Typologist #3 (e.g. me): No! You’re both wrong. A property-driven
typology would look like the following table, which allows us to
also add Spanish:
An example of such an unproductive controversy arises in Beckman &
Venditti who ask “Is typology needed?” (2010: 641) and argue against
typologizing prosodic systems solely by function (e.g., tone vs. stress-
accent) because Mandarin tonal L+H is allegedly like English intonational
L+H*:
[That one is a toneme and the other intonational] does not change the fact that these
two languages are far more like each other in many other respects than either is to a
language such as Japanese. (Beckman & Venditti 2011: 531)

While Beckman & Venditti find the Mandarin and English L+H similarities
significant, compare the more usual view of Gussenhoven’s (2007: 256)
concerning the similar H+L in Japanese and English:
While phonologically comparable, the pitch accents of Japanese and English have very
different morphological statuses. In Japanese, they form part of the underlying
phonological specification of morphemes, along with the vowels and consonants.
Intonational pitch accents are morphemically independent of the words they come with,
and are chiefly used to express the information status of the expression. The fact that
the English example [. . .] seems to have an accentuation similar to the Japanese
example [. . .] is ENTIRELY ACCIDENTAL. (My emphasis; cf. Hyman 2012)

Related to this is the third argument: assigning a name to a system can

give the false impression that something has been accomplished. On
numerous occasions I have been approached with the comment, “I think my
language may have pitch-accent, not tone”. Upon probing such
pronouncements further I find that this often means nothing more than the
feeling that the tonal contrasts are more sparse in this language than in
certain other languages which contrast tone on every syllable.
This brings us to the fourth reason to avoid whole taxonomies: the
labels are often unclear. An “X language” can mean at least the following:
(i) a language that has X, e.g., a “tone language” has tone, a “click
language” has clicks; (ii) a language that lacks X, e.g., an “open syllable
language” lacks closed syllables, an “intonation language” lacks tone or
stress:
 intonation language n. A language which is neither a tone language nor a pitch
language; a language in which the universally present intonation constitutes the only
linguistic use of pitch. (Trask 1996: 184)

(iii) a language that marks X more than certain other languages, e.g., “tone
language” vs. “pitch-accent language”, “syllable language” vs. “word
language”:
A pitch-accent system is one in which pitch is the primary correlate of prominence and
there are significant constraints on the pitch patterns for words [. . .] (Bybee et al
1998:277)

A syllable language is one which dominantly refers to the syllable, a word language is
one which dominantly refers to the phonological word in its phonological make-up.
(Auer 1993: 91)

(iv) a language which combines a specific set of linked properties into a

“holistic” typology (see especially Plank 1998):
there are obvious links between phonology and morphology; for example, it has been
argued – most probably correctly – that vowel harmony is a phenomenon of
agglutinating languages, or that fusional languages have more morphophonological
rules than isolating ones. There may also be links between phonology and syntax, e.g.
between head/ modifier (operator/operand) serialization and the location of (sentence or
word) stress. (Auer 1993: 1–2)

Vowel harmony is a phonological process relating to the morphological word in

syllable-timed languages, whereas vowel reduction is a phonological process relating to
the phonological word in stress-timed languages. (Auer 1993: 8; cf. Donegan &
Stampe 1983)

Such multi-property typologies invariably run into exceptions, and hence

proposals of prototypes. A potentially useful deductive strategy is the
canonical approach to typology:
The canonical approach means that I take definitions to their logical end point, enabling
me to build theoretical spaces of possibilities. Unlike classical typology, only then does
one ask how this space is populated with real instances. The canonical instances, that
is, THE BEST, CLEAREST, INDISPUTABLE (the ones closely matching the canon) are
unlikely to be frequent [. . .] Nevertheless, the convergence of criteria fixes a canonical
point from which the phenomena actually found can be calibrated, following which
 there can be illuminating investigation of frequency distributions. (Corbett 2007: 9; my
emphasis – LMH]

In prosody, canonical systems combine properties to meet a basic function

(Hyman 2012). In Prague School terms, the definitional function of stress-
accent is SYNTAGMATIC: it should unambiguously identify and mark off
major category words within utterances. To best do this, canonical stress-
accent therefore should be:

In other words, stress should be “biunique”: One should be able to predict

the stress from the word boundaries and the word boundaries from the
stress. Stress is thus highly syntagmatic.
This contrasts with the definitional function of tone which, like
segmental features, is to distinguish morphemes. Thus, for a two-height [H,
L] system to best realize this function to distinguish the most morphemes,
the properties of the canonical system should include:
In contrast with the above, there is no canonical function for so-called
“pitch-accent” systems. Each of the following possibilities either fails to
provide a distinct function from that of stress-accent or represents an
arbitrary criterion:

(8) a. a language which has an obligatory (but not necesarily

culminative) H tone per word?
b. a language which has a culminative (but not necessarily
obligatory) H tone? (Hualde, 2006)
c. a language which has either a culminative or an obligatory H tone?
(van der Hulst 2011)
d. a language which has privative H tones (/H/ vs. Ø)? (Clark 1988)
e. a language which limits tonal contrasts to the stressed syllable?
f. a language which restricts its tones in whatever way?
g. a language which has only two tone heights (H, L)?

[. . .] if we push the use of accents to its limits (at the expense of using tones), this
implies allowing unaccented words (violating obligatoriness) and multiple accents
(violating culminativity). In this liberal view on acccent, only languages that have more
than a binary pitch contrast are necessarily tonal [. . .] (van der Hulst 2011: 13)

If systems can be as “liberally” typologized as van der Hulst entertains, as

just quoted, then something is clearly wrong. I suggest it is the misguided
notion that the goal of phonological typology is to taxonomize languages
into pre-determined named “types”. If we instead focus on the properties,
rather than classifying languages or their subsystems, we will better be able
to appreciate the richness of the variation found in the world’s languages.
4 Where do phonology and typology part
company?
So why should we distinguish phonological typology from phonology
proper? After all, phonology has always been typological, developing its
models on the basis of extensive crosslinguistic data (Chomsky & Halle
1968 cite over 100 languages, for instance). However, there are aspects of
typology in which most phonologists have expressed little interest, e.g.,
mapping out phonological properties by geography, language family, or
historical contact. (Some have little interest in linguistic reconstruction and
language history as well.) Diverging from the traditional view of typology
that I have been discussing is the distributional typology perspective
“What’s where why?”:
In the past century, typology was mostly used as an alternative method of pursuing one
of the same goals as generative grammar: to determine the limits of possible human
languages and, thereby, to contribute to a universal theory of grammar [. . .] that would
rule out as linguistically impossible what would seem logically imaginable, e.g., a
language with a gender distinction exclusively in the 1st person singular. Over the past
decade, typology has begun to emancipate itself from this goal and to turn from a
method into a full-fledged discipline, with its own research agenda, its own theories, its
own problems. What has reached centerstage is a fresh appreciation of linguistic
diversity in its own right, and the new goal of typology is the development of theories
that explain why linguistic diversity is the way it is – a goal first made explicit by
Nichols’s (1992) call for a science of population typology, parallel to population
biology. Instead of asking “what’s possible?”, more and more typologists ask “what’s
where why?”. (Bickel 2007: 239)

To the theoretical phonologist it matters little that retroflex or ejective

consonants cluster geographically in certain areas or occur only in certain
language families. Instead, phonologists, like other formal linguists, have
mostly been interested in the question of what is a possible phonology:
Most theoretical linguists, from whatever camp, consider that it is a central goal of
theoretical work on grammar to distinguish possible grammatical processes from
impossible ones and – for the former – to explain why some possible processes seem
more common [probable] than others. (Newmeyer 2005: 27)

Concerning this growing conception of typology, my impression is that

traditional phonology has been less concerned with the “where” than the
“how” (as in “how should we analyze this system?”). In this connection,
what is the difference between a phonological typologist and a formal
phonologist who works on languages? Is it a matter of goals (“research
agenda”), emphasis, or initial assumptions? The following characterizations
of Croft’s (2007: 87) are reminiscent of the distinction I made between
substance vs. form as the starting point in comparing phonological systems:
[. . .] the structuralist and generative method assumes the same formal theoretical
entities to exist across languages, and then looks for constructions with distribution
patterns that appear to distinguish those formal theoretical entities in the language.

Typological analysis proceeds very differently. A typologist uses a functional definition

of a situation type, such as the Keenan-Comrie functional definition of relative clauses,
and compares the different grammatical constructions used for that function across
languages, and seeks relationships among the constructions (or grammatical properties
of the constructions).

While such a distinction may be recognizable to many linguists, particularly

non-phonologists, structural and generative phonologists who have done
crosslinguistic studies and surveys differ in the degree to which they are
concerned about geographic and genetic distributions. Thus, comparing the
various crosslinguistic studies of stress-accent, compare the different
weighting given to the “what” vs. “where” in Hyman (1977) and van der
Hulst et al. (2010) vs. Halle & Vergnaud (1987) and Hayes (1995). These
studies may even differ in how they answer the “why”. (For an explicitly
distributional typology of phonological properties conducted by two
structuralist-generative phonologists, see Clements & Rialland 2008.)
This brings us to the role of historical explanation and the question of
how to reconcile universals vs. diversity in phonological systems, which
Kiparsky (2008: 52) addresses as follows:
An increasingly popular research program seeks the causes of typological
generalizations in recurrent historical processes, or even claims that all principled
explanations for universals reside in diachrony. Structural and generative grammar has
more commonly pursued the reverse direction of explanation, which grounds the way
language changes in its structural properties. The two programs can coexist without
contradiction or circularity as long as we can make a principled separation between true
universals, which constrain both synchronic grammars and language change, and
typological generalizations, which are simply the results of typical paths of change.
I think this sums up the non-contradiction in the fact that most phonologists
both seek to determine what is universal AND at the same time appreciate
the diversity that we find in the sound systems of the world’s languages. In
a rare article reflecting on the nature of phonological typology, Dressler
(1979) applies Seiler’s (1979) inductive vs. deductive typology to
phonology:
Work in the typology of process phonology is usually inductive. [. . .] The usual
method of research is the sampling of similar phonological processes in different
languages, the enumeration of frequent, general or exceptionless properties, of their
clustering, of probable hierarchies and implications, and attempts at explanation by
reference to phonetic data [. . .] Much less frequent are deductive process phonological
typologies, although they are of primary importance, if typology should be based on
language universals research [. . .]. (Dressler 1979: 261)

He goes on to point out the following apparently contradictory observations

concerning phonological typology:
Deductive research is easier in phonology than in grammar, since we simply know
more about the phonologies of the languages of the world than about their grammars;
on the other hand less deductive typology has been done in phonology than in
grammar. (Dressler 1979: 262)

Of course this all depends on what one counts as “phonological typology”.

The original title of our workshop was “What is phonological typology –
and why does it matter?” As a brief answer: We need to do phonological
typology for the same reason we do general phonology, namely in order to
understand why phonologies are the way they are. However, in the ever
expanding, diverse field of phonology, we have the opportunity to
incorporate the “What, where, why?” in a way that is harder in other
subfields of linguistics. Phonologists can and should be involved in (i)
looking at phenomena both in breadth (quantitatively) and in depth
(qualitatively), (ii) identifying the geographical and genetic distributions of
the phenomena, and (iii) considering a wide range of potential explanatory
sources in addressing the “why?” It is only in so doing that we will attain a
complete picture of what phonology can vs. cannot do and why.

References
Aissen, Judith. 2003. Differential object marking: Iconicity vs. economy. Natural
Language & Linguistic Theory 21. 435–483.
Auer, Peter. 1993. Is a rhythm-based typology possible? A study of the role of
prosody in phonological typology. KontRI Working Paper No. 21. University of
Konstanz.
Beckman, Mary E. & Jennifer J. Venditti. 2010. Tone and intonation. In William J.
Hardcastle & John Laver (eds.), The handbook of phonetic sciences, 603–
652. Oxford: Blackwell.
Beckman, Mary E. & Jennifer J. Venditti. 2011. Intonation. In John Goldsmith,
Jason Riggle, & Alan C. L. Yu (eds.), The handbook of phonological theory,
485–532. Oxford: Blackwell.
Bickel, Balthasar. 2003. Referential density in discourse and syntactic typology.
Language 79. 708–736.
Bickel, Balthasar. 2007. Typology in the 21st century: Major current developments.
Linguistic Typology 11. 239–251.
Bossong, Georg. 1998. Le marquage différential de l’objet dans les langues
d’Europe. In Jack Feuillet (ed.), Actance et valence dans les langues de
l’Europe, 193–258. Berlin: Mouton de Gruyter.
Buckley, Eugene. 2000. On the naturalness of unnatural rules. Proceedings from
the Second Workshop on American Indigenous Languages. UCSB Working
Papers in Linguistics 9. 1–14.
Bybee, Joan L., Paromita Chakraborti, Dagmar Jung, & Joanne Scheibman. 1998.
Prosody and segmental effect: Some paths of evolution for word stress.
Studies in Language 22. 267–314.
Chomsky, Noam & Morris Halle. 1968. The sound pattern of English. New York:
Harper & Row.
Clark, Mary. 1988. An accentual analysis of Zulu. In van der Hulst & Smith (eds.),
51–79.
Clements, G. N. & Sylvester Osu. 2003. Ikwere nasal harmony in typological
perspective. In Patrick Sauzet & Anne Zribi-Hertz (eds.), Typologie des
langues d’Afrique et universaux de la grammaire, vol. 2, 70–95. Paris:
L’Harmattan.
Clements, G. N. & Annie Rialland. 2008. Africa as a phonological area. In Bernd
Heine & Derek Nurse (eds.), A linguistic geography of Africa, 36–85.
Cambridge: Cambridge University Press.
Cohn, Abigail. 1993. A survey of the phonology of the feature [+nasal]. Working
Papers of the Cornell Phonetics Laboratory 8. 141–203.
Corbett, Greville G. 2007. Canonical typology, suppletion, and possible words.
Language 83. 8–42.
Croft, William. 2003. Typology and universals. Second edition. Cambridge:
Cambridge University Press.
Croft, William. 2007. Typology and linguistic theory in the past decade: A personal
view. Linguistic Typology 11. 79–91.
Dixon, R. M. W. 1994. Ergativity. Cambridge: Cambridge University Press.
Dixon, R. M. W. & Alexandra Y. Aikhenvald (eds.). 2000. Changing valency: Case
studies in transitivity. Cambridge: Cambridge University Press.
Donohue, Mark. 1997. Tone in New Guinea languages. Linguistic Typology 1.
347–386.
Donegan, Patricia & David Stampe. 1983. Rhythm and the holistic organization of
language structure. In F. Richardson (ed.), CLS 19, Parasession on the
interplay of phonology, morphology & syntax, 337–353. Chicago: Chicago
Linguistic Society.
Dressler, Wolfgang U. 1979. Reflections on phonological typology. Acta Linguistica
Academiae Scientiarum Hungaricae 29. 259–273.
Dryer, Matthew. 1986. Primary objects, secondary objects, and antidative.
Language 62. 808–845.
Evans, Nicholas & Stephen C. Levinson. 2010. Time for a sea-change in
linguistics: Reponses to comments on “The myth of language universals”.
Lingua 120. 2733–2758.
Gordon, Matthew. 2007. Typology in optimality theory. Language and Linguistics
Compass 1(6). 750–769.
Gordon, Matthew K. 2016. Phonological typology. Oxford University Press.
Gouskova, Maria. 2013. Review of Marc van Oostendorp, Colin J. Ewen, Elizabeth
Hume, & Keren Rice (eds.). 2011. The Blackwell companion to phonology
(Malden: Mass.: Wiley-Blackwell). Phonology 30. 173–179.
Greenberg, Joseph H. 1948. The tonal system of Proto-Bantu. Word 4. 196–208.
Greenberg. Joseph H. 1958. The labial consonants of Proto-Afro-Asiatic. Word 14.
295–302.
Greenberg, Joseph H. 1962. Is the vowel-consonant dichotomy universal? Word
18. 73–81.
Greenberg, Joseph H. 1963. Vowel harmony in African languages. Actes du
Second Colloque Internationale de Linguistique Negro-Africaine, 33–38.
Dakar: Université de Dakar, West African Languages Survey.
Greenberg, Joseph H. 1966. Synchronic and diachronic universals in phonology.
Language 42. 508–517.
Greenberg, Joseph H. 1970. Some generalizations concerning glottalic
consonants, especially implosives. International Journal of American
Linguistics 36. 123–145.
Greenberg, Joseph H. 1974. Language typology: A historical and analytic
overview. The Hague: Mouton.
Greenberg, Joseph H. 1978. Some generalizations concerning initial and final
consonant clusters. Linguistics 18. 5–34.
Greenberg, Joseph H., James J. Jenkins, & Donald J. Foss. 1967. Phonological
distinctive features as cues in learning. Journal of Experimental Psychology
77. 200–205.
Greenberg, Joseph H. & Dorothea Kaschube. 1976. Word prosodic systems: A
preliminary report. Working Papers in Language Universals 20. 1–18.
Gussenhoven, Carlos. 2007. The phonology of intonation. In Paul de Lacy (ed.),
The Cambridge handbook of phonology, 253–280. Cambridge: Cambridge
University Press.
Hagège, Claude. 1992. Morphological typology. In Oxford international
encyclopedia of linguistics, vol. 3, 7–8. Oxford: Oxford University Press.
Halle, Morris & Jean-Roger Vergnaud. 1987. An essay on stress. Cambridge, MA:
MIT Press.
Haspelmath, Martin. 2006. Against markedness (and what to replace it with).
Journal of Linguistics 42. 25–70.
Hammond, Michael 2006. Phonological typology. Encyclopedia of Language &
Linguistics. Online. http://www.sciencedirect.com/science/article/pii/B0080448
542000468
Hayes, Bruce. 1995. Metrical stress theory: Principles and case studies. Chicago:
University of Chicago Press.
Heine, Bernd. 1993. Ik dictionary. Köln: Rüdiger Köppe Verlag.
Hockett, Charles F. 1955. A manual of phonology. Memoir 11, International Journal
of American Linguistics.
Hualde, José Ignacio. 2006. Remarks on word-prosodic typology. Proceedings of
the 32nd Annual Meeting of the Berkeley Linguistics Society, 157–174.
Hulst, Harry van der. 2011. Pitch accent systems. In van Oostendorp, Ewen,
Hume, & Rice (eds.), vol. 2, #45.
Hulst, Harry van der, Rob Goedemans, & Ellen van Zanten (eds). 2010. A survey
of word accentual patterns in the languages of the world. Berlin: De Gruyter
Mouton.
Hyman, Larry M. 1977. On the nature of linguistic stress. In Larry M. Hyman (ed.),
Studies in stress and accent, 37–82. Southern California Occasional Papers
in Linguistics 4. Department of Linguistics, University of Southern California.
Hyman, Larry M. 2005. Initial vowel and prefix tone in Kom: Related to the Bantu
Augment? In Koen Bostoen & Jacky Maniacky (eds.), Studies in African
comparative linguistics, 313–341. Köln: Rüdiger Köppe Verlag.
Hyman, Larry M. 2007. Where’s phonology in typology? Linguistic Typology 11.
265–271.
Hyman, Larry M. 2009. How (not) to do phonological typology: The case of pitch-
accent. Language Sciences 31. 213–238.
Hyman, Larry M. 2011. Tone: Is it different? In John Goldsmith, Jason Riggle, &
Alan Yu (eds.), The handbook of phonological theory, 2nd edition, 197–239.
Oxford: Blackwell.
Hyman, Larry M. 2012. In defense of prosodic typology: A response to Beckman &
Venditti. Linguistic Typology 16. 341–385.
Hyman, Larry M. 2015. Towards a canonical typology of prosodic systems. In
Esther Herrera Zendejas (ed.), Tono, acento y estructuras métricas en
 lenguas mexicanas, 13–38. México: El Colegio de México.
Kenstowicz, Michael & Charles Kisseberth. 1977. Topics in phonological theory.
New York: Academic Press.
Kiparsky, Paul. 1968. Linguistic universals and linguistic change. In Emmon Bach
& Robert T. Harms (eds.), Universals in linguistic theory, 171–202. New York:
Holt, Rinehart & Winston.
Kiparsky, Paul. 2008. Universals constrain change; change results in typological
generalizations. In Jeff Good (ed.), Linguistic universals and language
change, 23–53. Oxford: Oxford University Press.
Maddieson, Ian. 1991. Testing the universality of phonological generalizations with
a phonetically specified segment database: Results and limitations. Phonetica
48. 193–206.
Maddieson, Ian & Kristin Precoda. 1990. Updating UPSID. UCLA Working Papers
in Phonetics 74. 104–111.
McCarthy, John J. 2002. A thematic guide to optimality theory. Cambridge:
Cambridge University Press.
Newmeyer, Frederick J. 2005. Possible and probable languages: A generative
perspective on linguistic typology. Oxford: Oxford University Press.
Nichols, Johanna. 1986. Head-marking and dependent-marking grammar.
Language 66. 56–119.
Nichols, Johanna. 1992. Language diversity in space and time. Chicago:
University of Chicago Press.
Nichols, Johanna. 2007. What, if anything, is typology? Linguistic Typology 11.
231–238.
Oostendorp, Marc van, Colin J. Ewen, Elizabeth Hume, & Keren Rice (eds.). 2011.
The Blackwell companion to phonology. Malden, MA: Wiley-Blackwell.
Plank, Frans. 1998. The co-variation of phonology with morphology and syntax: A
hopeful history. Linguistic Typology 2. 195–230.
Plank, Frans. 2001. Typology by the end of the 18th century. In Sylvain Auroux et
al. (eds.), History of the Language Sciences: An International Handbook on
the Evolution of the Study of Language from the Beginnings to the Present,
vol. 2, 1399–1414. Berlin: Walter de Gruyter.
Prince, Alan & Paul Smolensky. 1993/2004. Optimality theory: Constraint
interaction in generative grammar. Malden, MA: Blackwell.
Samuels, Bridget D. 2011. Phonological architecture: A biolinguistic perspective.
Oxford: Oxford University Press.
Sapir, Edward. 1925. Sound patterns in language. Language 1. 37–51.
Seiler, Hansjakob. 1979. Language universals research, questions, objectives, and
prospects. Acta Linguistica Academiae Scientiarum Hungaricae 29. 353–367.
Slobin, Dan I. 2004. The many ways to search for a frog: Linguistic typology and
the expression of motion events. In Sven Strömqvist & Ludo Verhoeven
(eds.), Relating events in narrative, volume 2: Typological and contextual
perspectives, 219–257. Mahwah, N.J.: Erlbaum.
Talmy, Leonard. 1985. Lexicalization patterns: Semantic structure in lexical forms.
In Timothy Shopen (ed.), Language typology and linguistic description, volume
3: Grammatical categories and the lexicon, 57–149. Cambridge: Cambridge
University Press.
Trask, R. L. 1996. A dictionary of phonetics and phonology. London: Routledge.
Trubetzkoy, Nikolai (1969 [1939]). Principles of phonology. Translated by
Christiane A. M. Baltaxe. Berkeley and Los Angeles: University of California
Press.
Vajda, Edward. 2001. Test materials dated August 17, 2001. http://pandora.cii.ww
u.edu/ vajda/ ling201/test2materials/Phonology3.htm.
Whaley, Lindsay J. 1997. Introduction to typology. Thousand Oaks, California:
Sage Publications.
Frans Plank
An implicational universal to defy:
typology ⊃ ¬ phonology ≡ phonology
⊃ ¬ typology ≡ ¬ (typology ᴧ
phonology) ≡ ¬ typology v ¬
phonology

Abstract: The purpose of this chapter is twofold: first, to assess how

typology has been dealing with phonology, from early days to the present;
second, focusing on phonology, to ask about asymmetries between
phonology and syntax-inflection in general and about typological concerns
in phonology itself. Looked at from both angles, the phonology–typology
relationship is seen to be special, and the impression is confirmed that, in
comparison especially with syntax, phonological typology as well as
typological phonology are behindhand in the quest for system in linguistic
diversity. Explanations are suggested in terms of the substance of subject
matters and of the attitudes to description and theory in different
subcommunities in linguistics.

1 Introduction
The purpose of this chapter is twofold: first, to assess how typology,
unceremoniously introduced in §2, has been dealing with phonology (§3),
from early days (§3.1) to the present (§3.2); second, focusing on phonology
(§4), to ask about an imbalance of phonology and syntax-inflection in
general (§4.1) and about typological concerns in phonology itself (§4.2).
Looked at from both angles, the phonology–typology relationship is seen to
be special, and the impression is confirmed that, in comparison especially
with syntax, phonological typology as well as typological phonology are
behindhand in the quest for system in linguistic diversity. (Though not all is
well about the syntax–typology relationship, either.) Explanations are
suggested in terms of the substance of subject matters and of the attitudes to
description and theory in different subcommunities in linguistics.

2 The typological programme and where it is in

arrears
In linguistics, typology is a research programme, not a subfield or a theory,
and its remit is (i) to chart linguistic diversity, (ii) to discover order or
indeed unity in diversity, and (iii) to make sense of what has been charted
and discovered. Regrettably but perhaps understandably, given how this
enterprise has usually been named after Gabelentz (1894, 1901) and given
the meanings of this term in other fields, typology has sometimes been
taken to be about the classification of the discipline’s cherished cardinal
individuals, namely languages, into types; but for linguistic typology “type”
in this sense of “class” is really a secondary and indeed expendable concept.
First and foremost typology’s objective is to identify elementary variables
where languages (or, to avoid this moot concept, mental lexicons and
grammars) can differ, and then to examine whether these variables vary
independently or vary together: when variables are found to co-vary, this
means crosslinguistic variation is in this respect limited. It is not to be
seriously expected that literally TOUT se tient, other than in a Saussurian
logical sense of all sign values in a sign system being interdependent.
Since lexicons and grammars are as diverse and uniform as they have
become over time – during the lifespan of individual speakers; across
generations of language acquirers; through contacts between speech
communities; in the evolution of our species – typology’s closest
association is with developmental linguistics. To the extent that linguistic
diversity is limited and orderly rather than being without limit and random,
and to the extent that any order that we are able to discover is not the result
of non-linguistic contingencies of the histories of populations (itself a
fascinating research field), patterns of diversity could be shaped (i) by
timeless typological laws constraining states of mental lexicons and
grammars and/ or (ii) by laws of historical change and stability,
constraining transitions between states. The evidence points to both being
effective, although constraints are not always easily recognised as clearly
being of one kind or the other.
By its very nature the typological research programme is all-inclusive:
elementary variables from ALL structural domains should be examined for
covariation. It is an empirical issue, not one to be resolved a priori, whether
some domains are in fact less tightly interconnected than others, showing
independent variation rather than co-variation.
This is the lofty idea. Here we are asking about an imbalance in
workaday practice: Within typology, on its own or in association with
developmental linguistics, is phonology different? Contextualising the other
way round: Within phonology, synchronic and diachronic, is typology
different? And these are not quite the same question, notwithstanding the
logical equivalence of the first two implications of the title when flipping
antecedent and consequent in contraposition under negation. Is there too
little phonological typology and/ or too little typological phonology?
Should, and could, there be more of either or both?
When diagnosing typology as phonologically deficient or healthy, or
phonology as typologically challenged or up to the mark, what is phonology
being compared with? Obviously the other structural domains into which
grammar is compartmentalised: syntax and morphology, the latter with its
subdomains of inflection and word formation. There is
semantics/pragmatics, too, with the construction of complex meaning as the
job of morphology and syntax. And there is the lexicon, storing the basic
building blocks for all constructing, the material for grammar to work with
(plus everything else that is non-compositional and therefore not taken care
of by the grammar). Word formation as the lexical part of morphology,
semantics/ pragmatics, and the lexicon hardly compare any less
unfavourably than phonology in their typological involvement. It is really
only in explicit or tacit comparison with syntax and with the syntactic part
of morphology, inflection, that phonology can possibly have grounds for
complaint. To assess the situation correctly, phonology, as the grammar of
sound, should perhaps be kept apart from phonetics; but then, typological
contexts seem especially conducive to a blurring of this distinction, thereby
strengthening the typological presence of the amalgamated domains.3
Now, ask around among experts – as I have informally done for years,
working with an editorial board of a typological journal and helping with
the programming of many a typological event, but sometimes also mingling
with phonologists with or without typological sympathies – and they will
very likely agree, especially when active in both phonology and typology,
that, yes, the relationship between typology and phonology is special,
namely less intimate or at any rate different, in comparison with syntax and
inflection. Is this a misperception?4

3 The evidence

3.1 Early typology

In the past, as the typological programme was gaining momentum,
inflection and syntax were distinctly in the limelight. Phonology was on the
stage, too, but was more of a sideshow. Historically speaking, anybody
complaining about an imbalance would have a point, then. Here is a bird’s-
eye view of the record.5

3.1.1 The pioneer years

The first typologist in my history was Tommaso Campanella (1568–1639).

A Dominican monk, philosopher, astrologer, and utopian social theorist, he
was the first to methodically do, among all his other activities, what today’s
typologists are doing, too – namely to take stock of linguistic diversity and
to determine whether it is limited insofar as the values of certain
grammatical variables, although logically independent of one another, co-
vary rather than varying independently. Campanella’s results, for the
variables that he studied, were positive (or else they would presumably
have gone unreported). His attention was attracted by parts of speech and
“accidence”, central to linguistic theory and descriptive grammars of the
day. Universal or Philosophical Grammar took for granted that parts-of-
speech grammar was in essence universally uniform; this was one of the
dogmas Campanella doubted, and his doubt was not global and vague. He
had evidence, coming from languages hitherto inaccessible (such as
Vietnamese, as described by missionaries of his acquaintance) or neglected
(such as the Romance vernaculars), that there was diversity where
contemporary linguistic theory decreed uniformity; but in light of this
evidence he could also identify specific constraints on diversity.
Campanella discovered that nouns were not words which universally
inflected for case and number, nor were verbs words which universally
inflected for some such categories as tense and personnumber. His
crosslinguistic evidence suggested to him that although such inflection was
an option, though not a necessity for either nouns or verbs, variation was
yet not random. He was aware of languages where verbs inflected, but
nouns did not (at any rate not for case: witness the Romance vernaculars),
but not of languages where nouns inflected (for case), but verbs did not
inflect. Thus, for all he knew, inflectional systems could not differ in all
conceivable ways, given the elementary variables of word classes and
inflectional categories: it looked like noun inflection implied verb
inflection, but not vice versa. Further, turning from a connection between
parts of speech to one between inflectional categories, nouns did not
universally have to inflect for both case and number. They could (as in the
Classical languages or also in Turkish); but they could also inflect only for
number (as in the Romance vernaculars) or for neither (as in Vietnamese);
but there was no language in Campanella’s smallish, yet diagnostically
instructive sample where nouns only inflected for case. Hence, by inductive
generalisation, diversity was limited here too, insofar as case inflection
implied number inflection, but not vice versa.
There was no continuity of typological research into inflection after
Campanella – which is not surprising because it would take a while before
you could be a morphologist or indeed linguist by profession. Nonetheless,
a wide range of individual inflectional categories as well as properties of
inflectional systems were frequently on the agenda when diversity and its
possible limits were examined, in whatever wider intellectual context. By
the end of the nineteenth century the typological distinction between
agglutination and flexion/fusion was well entrenched: a rich network of
implications was assumed, on deductive as well as inductive grounds, to be
connecting individual inflectional categories within and across word classes
with regard to variables such as separation vs. cumulation of categories,
invariance vs. variance of exponents, looseness vs. tightness of bonding
between stems and morphological markers, distinction vs. non-distinction
(syncretism) of term oppositions.
Words had always been principal units for linguistic theory and
descriptive grammar, and the structure of constructions most obviously
consisted in how words were ordered one after another. One did not have to
be a serious polyglot to be able to observe that they were ordered differently
in different languages, or also at different historical stages of one and the
same language. Familiarity with the right kind of languages would soon
suggest that there was indeed system to such differences. A traveller and
diplomat with profound Oriental expertise, François (de) Mesgnien (or,
Polish-style, Meninski, ca. 1620–98), apart from noting diversity in many
grammatical particulars where current linguistic theory had decreed
universal uniformity, discovered that different syntactic constructions did
not have their constituent parts ordered randomly: across all kinds of
constructions where one member was a head and the other a dependent
(verb–object, noun–genitive, noun–attributive adjective, adposition–noun
phrase, etc.), dependents would uniformly either precede or follow their
heads. There were as many variables as their were kinds of head–dependent
constructions; instead of varying independently, these variables all tended
to co-vary.
Word order would remain at centre stage in the eighteenth century,
although with a focus now on its relationship to inflection and periphrasis:
rich inflection, notably for case and agreement, was hypothesised to license
inversions, while impoverished inflection was hypothesised to necessitate
rigid syntactic ordering, especially of dependents (or determiners) after
their heads. This was how Abbé Gabriel Girard (ca. 1677–1748) and, a rare
grammarian by profession,6 Nicolas Beauzée (1717–89) had been
elaborating the old Scholastic theory of a universal ordo naturalis of ideas
and words expressing them. Publicised through the Encyclopédie française,
such typological schemes (like also those involving inflectional systems,
circulated by the Encyclopedia Britannica) were gaining currency, while
Mesgnien’s more discerning discovery that rigid word order itself permitted
of orderly variation – head before dependent throughout, as per ordo
naturalis, but also dependent before head throughout – was temporarily
falling into oblivion.
Beginning with occasional early sightings in Greenlandic, Basque, and
elsewhere of a special case (eventually called “ergative”) which marked
only transitive subjects, as opposed to intransitive subjects, which would
receive the same case as direct objects of transitive verbs, the universal
uniformity of such central syntactic concepts as subject and object came
into question. But relational alignment would acquire typological
significance only slowly, when it was realised that the potential for diversity
was indeed vast here, because there were numerous patterns of relational
identification across transitive and intransitive clauses, and alignments
could in principle differ from verb to verb, from nominal to nominal, from
tense/ aspect to tense/ aspect, and between different rules making reference
to syntactic relations (e.g., case marking, verb agreement, constituent
ordering, clause combining). At any rate, when Georg von der Gabelentz
(1840–93), a professor of Oriental languages and general linguistics but
originally trained in law and administration, set out a methodology for
getting ahead with the typological programme (posthumously published in
1894), he used alignment and word order for illustration. His two variables
were ergative vs. accusative alignment of case marking and the ordering
(uniform or divergent) of genitives and adjectives relative to their heads. Of
the four possible combinations of values, he found one underrepresented:
ergative alignment and uniform ordering of genitives and adjectives; and for
one he could presently cite no single attestation: accusative alignment and
divergent ordering of genitives and adjectives. (In future, committees of
experts, drawing up comprehensive checklists of variables and calculating
the statistics of co-variation, would have to confirm such gaps and
strikingly unequal value distributions.) The “conjunctures” to be inductively
inferred thus were these: if a language has ergative-absolutive alignment for
case marking, it will, with more than chance probability, also have
divergent genitive and adjective ordering; if a language has divergent
genitive and adjective ordering, it will, with considerably higher probability
(if not certainty), also have ergative-absolutive alignment.
Gabelentz was optimistic that it would eventually be established, rather
than continue to be merely decreed, that tout se tient – including even das
Lautwesen (Die Sprachwissenschaft, posthumous 2nd edition of 1901: 481).
But he did not live to flesh out how phonology was supposed to be
internally and externally interconnected and to duly downscale his
grandiose hope, for surely, unless you were thinking of Saussure’s “values”
of signs within a system, not EVERYTHING would turn out to cohere with
EVERYTHING else.

3.1.2 Sounds different

Actually, languages had long been observed, if often dimly, to differ in their
Lautwesen, and some such differences had begun to be presumed, if not
methodically established, to be systematically interconnected.
Updating Biblical descent stories, languages were distinguished (most
influentially by Bishop Isidore of Seville, ca. 560–636) depending on which
classes of sounds or “letters” they featured as somehow the most prominent:
guttural (the Semitic East), palatal (Greek and other Eastern
Mediterranean), or dental (the Romance West). It likewise betrayed some
awareness of sound inventories as harmonious and symmetrical systems of
contrasts that sounds felt to be indispensable were occasionally deplored to
be missing, such as the labial nasal in Iroquois; while, conversely,
uncommon kinds of seemingly difficult-to-produce sounds would
sometimes be reported from far-away regions, such as clicks from
southernmost Africa.
Words were heard, and described, to be stressed partly similarly and
partly differently, and sentences to be intoned and rhythmically organised
partly similarly and partly differently, in different languages; but not much
seems to have been made of this in early typology, when prosodic analysis
was still in its infancy and hard enough to practise on even the most
familiar languages closest to home. Once the awareness spread that pitch
could not only be employed for intonation, but also lexically, “tone
languages” would earn themselves a prominent place in galleries of
linguistic curios. But these exhibits long remained something like
monolithic erratic rocks of enigmatic provenance, obviously instantiating
crosslinguistic diversity, but of little apparent significance for seekers of
patterns of co-variation.
The length of words as measured in segments or syllables, and how
sounds could and could not be combined to articulate them, were variables
whose early typological standing was more assured. Although the words of
a language would differ as to their length and composition, there seemed to
prevail some consistency on both counts within languages, but not across
them. Further, there seemed a prospect especially of word length correlating
with something else, namely the richness or poverty (i) of sound inventory
(including prosody) and (ii) of inflectional morphology. It was at something
along these lines that such an early and inscrutable distinction as that
between naturales linguæ and grammaticæ linguæ as championed by
Guillaume Postel (1510–81), a Renaissance polymath with firsthand
experience of languages of the East, appeared to be driving.
Going beyond inventories and phonotactics, it was noticed in early
grammars of languages such as Turkish that certain similarity requirements
were imposed upon the vowels of the constituent parts of words. That such
vowel harmony conspicuously distinguished such languages from most
others where words could ostensibly make free use of the entire inventories
of sounds available to them was commented on at least as early as by
Mesgnien-Meninski; but it was not suspected to be implicated in constraints
on diversity. Elementary variables such as the dimensions for harmony
(e.g., front–central–back, rounded–unrounded), its progressive or regressive
direction, or vowels allowed to escape it, were not highlighted in early
comparisons. Arguably this precluded the recognition that such
phonological processes, regardless of the morphological and syntactic
environments in which they were embedded, themselves gave ample room
for variation – and that the empirical question was whether or not it was
actually exploited by different languages.7
An early apogee of holistic typology was reached with the
developmental scenario of James Burnett (1714–99, who as judge at the
High Court of Scotland took the title Lord Monboddo). For him, as for
other “conjectural historians” of the Scottish Enlightenment, language
development from its first origins consisted in increasing “articulation”, and
“material” and “formal” articulation were proceeding in tandem: the extent
of material articulation, pertaining in particular to (a) the elaboration of
sound inventories, (b) the complexity of syllable structures, (c) word length,
(d) accentual differentiation (as opposed to not-so-articulated tonal
modulation), would therefore correlate with the extent of formal
articulation, pertaining in particular to (a) the differentiation of parts of
speech, (b) the elaboration of inflectional and derivational systems, (c)
analytic syntax (as opposed to polysynthesis, where sentences are not yet
articulated into words). In Burnett’s comprehensive scheme little remained
of Latin-inspired Universal Grammar: just about all of morphosyntax and
phonology/phonetics had become themes with variations. But for him, too,
variation was not random but harmonious, and the reasoning was
developmental: co-variation was the result of the co-evolution of all
individual manifestations of ultimately one single fundamental human
capacity, that of separating wholes into parts and imposing ever finer
structure on the unstructured. Burnett’s immensely influential Scottish
contemporary Adam Smith (1723–90) had, in his occasional linguistic
writings, preferred the term “analysis” to Burnett’s “articulation” – whence
the typological triad of analytic, synthetic, and polysynthetic (or
“incorporating”) languages to emerge soon after, finding equal favour
among speculative and comparative grammarians. Burnett ventured further
into the articulation of sound matter than others at his time, and he was also
uncommonly well informed about contemporary languages, especially of
North America, which were taken for representatives of different
developmental stages of the human capacity for articulation. His inductive
generalising about co-variation/co-evolution was not hastier than was
customary; but in his case there were just too many languages, often ill
described, and too many variables, material as well as formal, to keep under
control.

3.1.3 Where sounds matter

Although the early protagonists of typology, from Campanella to Gabelentz,

were doing, within their means, what typologists are still doing today, they
were not academically trained professional linguists. Comparative
linguistics as an academic discipline emerged over the nineteenth century,
and what was being professionally compared then were sounds, words, and
inflections among the languages of primarily one family, Indo-European.8
The remit there was to work out the history of these sounds, words,
inflections with the aim of reconstructing ancestral languages and the
genealogical relationships between their descendents, preferably visualised
through family trees. Sound matters couldn’t have mattered more: sound
laws (so-called) after all were the crowning achievements in this
comparative enterprise, with the study of inflection and especially syntax
taking a back seat. Despite its intellectual triumphs, academic historical-
comparative linguistics remained precariously poised vis-à-vis the
philologies; but it at least managed to eke out a niche for itself – and in the
process marginalised typological comparison. Universal co-variations
between variables, often morphosyntactic, rarely phonological, continued to
be suggested, usually elaborating on old themes rather than breaking new
ground; but they were no match for sound laws, too far were they lagging
behind in the methodological rigour of their continuingly amateurish
investigation. Obviously, detecting order in variation is more demanding the
larger and the more diverse the set of languages to be compared – all
languages ever spoken (well, a hopefully representative subset of them all)
or only one family. Still, one can see why, at the end of the century,
Gabelentz would issue his rallying cry for typologists to get their act
together, at long last.
To do what the likes of Rasmus Rask, Jacob Grimm, Karl Verner, and
the Neogrammarians were famously doing, likewise building on
seventeenth and eighteenth century amateur predecessors, required
phonological expertise. Morphology, syntax, even semantics were not
wholly outside the scope of the new professionals; but practising the
Comparative Method, the tool that gained historical-comparative linguistics
the scholarly respectability that as yet eluded typology, at heart meant being
a phonologist. The languages under comparison, for the purpose of
ascertaining whether or not they had “sprung from some common source”,
had to be searched for words or morphemes that might be cognates;9 then
regular sound correspondences (more often differences rather than
identities) had to be identified; then plausible stories had to be constructed
for how a state of systematic difference, as between the daughter languages,
could have resulted from a postulated state of uniformity, as in the proto-
language, with regular sound change effectuating the transition. Latterly,
reconstructed phonological systems and hypothesised phonological changes
would sometimes be considered suspect if they were crosslinguistically
uncommon or indeed unique: but how could typology convincingly serve as
a control, when its own results were not beyond methodological doubt and
what it delivered on the phonological front was so little to begin with?

3.1.4 Interim summary

At the heart of doing typology, as an amateur or eventually with an

academic license, always lay the identification of elementary linguistic
variables and their examination for co-variation or independent variation.
There is no inherent reason why this enterprise should have disfavoured
phonology, or, as the Comparative Method did in the case of historical-
comparative linguistics, have favoured phonology.
If anything it should ceteris paribus have been easier to do phonological
than morphosyntactic typology, because the methodological issue of the
tertium comparationis is generally felt to be less of a problem here.
Assuming that “genitives”, “adjectives”, or “direct objects”, for example,
are the same kinds of thing in all languages under comparison just because
they bear the same name is riskier than taking crosslinguistic equivalence
for granted for many descriptive terms in phonology, where they are often
phonetically grounded, hence have a more solid claim to be universal.
Still, for the pre-academic times of typology as covered in the précis
above, phonology was indisputably a runner-up to inflection and syntax. We
have more than a century of phonology-in-typology and typology-in-
phonology yet to size up to bring us up to date: Has the balance shown
signs of shifting as typology was coming of age?

3.2 Typology these days

Easing the task of the historian, it was really only some sixty or seventy
years into the twentieth century when the tide for typology turned rather
dramatically from ebb to full flow. Still, instead of accompanying
typology’s march into modernity, with the typological programme
advancing through wider and deeper knowledge of languages, through more
penetrating linguistic analysis, and through refined typological
methodology, let’s zoom in on what we know best: ourselves. There are
several kinds of indicators that, with us today, typology and phonology have
remained the uneasy bedfellows that they had always been.10

3.2.1 Centers and projects

Typology’s overdue rise to global prominence was heralded by several local

research cooperations, usually gathering around a senior figure, focusing on
selected structural domains, holding workshops, and publishing working
papers and collective volumes.
In Prague the Linguistics Circle had a typological section, with Vilém
Mathesius’s pupil Vladimír Skalička (active from the 1930s to 60s) as the
eventual mastermind. From the 1960s onwards, the St. Petersburg/
Leningrad school of typology produced a long series of collective
monographs on grammatical categories, overseen by Aleksandr A.
Xolodovič, Viktor S. Xrakovskij, and Vladimir P. Nedjalkov. Roughly
concurrently, though with only modest interaction, there were the Language
Universals Project at Stanford, directed by Joseph H. Greenberg (1967–76),
and UNITYP (“Sprachliche Universalienforschung und Typologie unter
besonderer Berücksichtigung funktionaler Aspekte”) in Cologne under
Hansjakob Seiler (1972–92). In Paris, a little later (1984), RIVALC got
going, lead by Gilbert Lazard, and its remit was rather more specific:
“Recherche interlinguistique sur les variations d’actance et leur corrélats”.11
No phonology appears to have been on the typological agendas in St.
Petersburg/Leningrad, Cologne, and Paris. In Prague, phonological traits,
from segment inventories and phonotactics to rhythmic patterns, would be
conjectured to be implicated in Skalička’s “ideal” types, whose most
conspicuous hallmarks were syntactic and morphological, but which were
after all meant as holistic. However, the phonological typologising of
Nikolaj S. Trubetzkoy and Roman O. Jakobson, earlier members of the
Prague Cercle, would echo more resoundingly in Stanford, largely owing to
Joseph Greenberg himself, where phonology was equally represented with
syntax and morphology, each yielding one volume of the resulting book
series, Universals of human language (4 vols., Stanford: Stanford
University Press, 1978).
Eventually, funding was secured from the European Science Foundation
for a large-scale international typological research programme.12 From
1990–95, the nine theme groups of EUROTYP brought together over a
hundred collaborators from Europe in the EU sense and beyond, eventually
producing eight tomes, published in Mouton de Gruyter’s series Empirical
Approaches to Language Typology (forming its collective volume 20). One
somewhat isolated group and one volume of EUROTYP were on a
phonological theme, Word prosodic systems in the languages of Europe
(coordinated/edited by Harry van der Hulst); the rest was syntax and
inflection.

3.2.2 Results on record

The net results of the typological programme are instances of co-variation

among particular variables: How many of those on record,13 however
confidence-inspiring or dubious, are phonological and how many syntactic
and inflectional? Although incomplete and not updated over the last decade,
the UNIVERSALS ARCHIVE at Konstanz (http://typo.uni-konstanz.de/archive/in
tro/) still gives an impression that should not be far off. Of the over 2,000
universals documented, some age-old, others novel, the domain Syntax
accounts for 1129, Inflection and Morphology for 789 + 157 (with the
intended domain distinction somewhat unclear), Phonology (including
phonetics) and Prosodic Phonology for 543 + 62. (For completeness:
Lexicon 158, Word Formation 51, Semantics, Pragmatics, and Discourse
142 + 14 + 11.) Without disentangling assignments to multiple domains,
there are more than three times more syntactic and inflectional than
phonological universals deposited in this archive. Discarding the possibility
that far more phonological than morphosyntactic universals have
inadvertently escaped archiving, this can mean two things: (i)
crosslinguistic diversity is far more copiously constrained in syntax and
inflection than in phonology; or (ii) typologists have strongly preferred
syntax and inflection to phonology when prospecting for universals.

3.2.3 Conferences and journals

In typology, like in other academic enterprises, it is at conferences and in

journals that new ground is broken. (Amateurs used to work in isolation or
corresponded.) How does phonology stand its ground on these occasions?
The learned society devoted to the advancement of the scientific study
of typology, the Association for Linguistic Typology (ALT; http://www.ling
uistic-typology.org/index.html), has held biennial conferences since its
foundation in 1994. At ALT 1, in Vitoria-Gasteiz in 1995, 44 papers were
given, and a mere two and a half (or two halves) of these were on
phonology. (One by the present writer, eventually to mature into Plank
1998: not original research, but a historiographic piece meant as
encouragement.) This was certainly not an auspicious start of phonology in
organised typology. Subsequently phonology was to pick up slightly, but a
modest 5–10% of presentations used to be the limit from ALT 2–10. Only
most recently, at ALT 11, in Albuquerque in 2015, did phonology (and
phonetics), while remaining a distant second to syntax and inflection,
account for as many as 16% of the papers given (13 out of 81, including
theme and poster sessions).14
This last figure is getting close to the routine proportion of phonology
papers at the International Conferences on Historical Linguistics (ICHL),
biennially organised by the International Society of Historical Linguistics.
ICHL 22, convened in Napoli in 2015, had 29 phonology papers out of 246
papers accepted for general session as well as workshops; but 12% for
phonology is relatively low for ICHL’s, where the average over the years
has been closer to 20%, a proportion that is also confirmed by the selections
of papers published in the ICHL proceedings. Thus, the conference circuit
does not see mass migrations of phonologists to either historical or
typological fora of the kind of ICHL and ALT meetings. You are likely to
meet quite a few typologists you know from ALT at ICHL, though, but
what they have to say about the relationship between typology and
diachrony typically concerns morphosyntax rather than phonology, even
when their topic is grammaticalisation.
The figures for Linguistic Typology (LT), ALT’s journal, show an
imbalance, too, although less marked than at ALT conferences. As revealed
in the five-yearly editorial reports published in LT, the period of 2006–11
saw 30 submissions on phonology and phonetics and (ca.) 80 on syntax and
inflection, of which 15 and (ca.) 30 respectively were accepted; 2001–06
saw 20 submissions on phonology and phonetics and 64 on syntax and
inflection, of which 11 and 23 respectively were accepted. Like at ALT
conferences, phonology had a really poor start in this environment: 8
phonology and phonetics papers were submitted in 1995–2001 as against
100 for syntax and inflection, of which 2 and 30 respectively were
accepted. Interestingly, these figures also show that phonology submissions
had higher acceptance rates, reflecting superior quality or at any rate quality
that it was easier to reach consensus on.
Spotchecks suggest that the proportion of phonology to syntax and
inflection papers is somewhat, though not dramatically lower in LT than in
historical linguistics journals such as Diachronica, Journal of Historical
Linguistics, Folia Linguistica Historica, or Transactions of the Philological
Society. (In traditional-style historical-comparative journals such as
Indogermanische Forschungen, Historische Sprachforschung, or Journal of
Indo-European Studies, on the other hand, phonology easily holds its own
as of old.)

3.2.4 Databases
Since data collections are no longer jealously guarded as the collector’s
private property, an increasingly popular research tool in typology are
online databases. Among the thirty or so world-wide typological databases I
am aware of as recently active, the majority cover domains from syntax and
inflection. But it is not as overwhelming a majority as one might have
expected, since about a dozen are on phonology or include substantial
phonological data:
– UPSID: UCLA Phonological Segment Inventory Database
http://www.linguistics.ucla.edu/faciliti/sales/software.htm;
http://web.phonetik.uni-frankfurt.de/upsid_info.html
– LAPSyd: Lyon-Albuquerque Phonological Systems Database
http://www.lapsyd.ddl.ish-lyon.cnrs.fr/lapsyd/
– PHOIBLE Online: Phonetics Information Base and Lexicon
http://phoible.org/
– P-base
http://pbase.phon.chass.ncsu.edu/
– World Phonotactics Database
http://phonotactics.anu.edu.au/
– StressTyp2
http://st2.ullet.net/?
– XTone: Cross-Linguistic Tonal Database
http://xtone.linguistics.berkeley.edu/index.php
– Metathesis in Language
http://metathesisinlanguage.osu.edu/database.cfm
– Language Typology Database
http://www.unicaen.fr/typo_langues/index.php?malang=gb
– tds: Typological Database System
http://languagelink.let.uu.nl/tds/main.html
– WALS Online: World Atlas of Language Structures
http://wals.info/
– SignPhon: A Phonological Database for Sign Languages
http://www.ru.nl/sign-lang/projects/completed-projects/signphon/

Currently the most popular database is the World atlas of language

structures online (WALS, http://wals.info/), created under the auspices of the
now-defunct Linguistics Department of the Max Planck Institute for
Evolutionary Anthropology at Leipzig. WALS examines numerous lexical
and grammatical variables (“features”) across large numbers of languages,
with core and extended core samples of 100 and 200 languages and with
more than 2,500 languages figuring in one survey or another. This database
has been widely used to discover as well as to refute universals; but what
you get from WALS directly are geographical patterns of value distributions,
and these are of interest to historical linguists for all kinds of reasons. WALS
is divided up into 192 chapters, which mostly cover one feature, but some
cover several (last accessed 10 March 2016). Breaking down this number
by “areas”, there are 20 chapters devoted to phonological (or also phonetic)
features – or 22 if “Writing systems” and “Para-linguistic usages of clicks”
are added. Inflection (Morphology 12, Nominal Categories 29, Verbal
Categories 17) claims 58 chapters, Syntax (Nominal Syntax 8, Word Order
56, Simple Clauses 26, Complex Sentences 7) 97. (The Lexicon gets 13
chapters, of which four are about the sound shape of pronouns, and one is
about the sound shapes of the word for ‘tea’.) In sum, almost eight times
more morphosyntactic than phonological features were deemed worthy of
(or also amenable to) WALS-style treatment, which is once again consistent
with the underrepresentation of phonology in typology.

3.2.5 Teaching and texts

Unlike in the days of Campanella, Mesgnien-Meninski, and Gabelentz and

indeed up to less than fifty years before present, most modern typologists
will have been taught typology at some stage of their university training in
linguistics. On the teaching side, practising typologists often design their
own courses, with “learning by doing” as the chief didactic element, but
there has also been a proliferation of typology textbooks. Concerning the
specialist knowledge and know-how that textbook authors want to pass on
to their readers, there are differences among them in selection and
emphasis; but what virtually all share, with only one exception I am aware
of (Moravcsik 2013), is that phonology is not emphasised, if selected at all:
– Haarmann, Harald. 1976. Grundzüge der Sprachtypologie: Methodik,
Empirie und Systematik der Sprachen Europas. Stuttgart: Kohlhammer.
– Some phonology, mostly phoneme inventories; but not really a
textbook.
– Ineichen, Gustav. 1979. Allgemeine Sprachtypologie: Ansätze und
Methoden. Darmstadt: Wissenschaftliche Buchgesellschaft; 2nd edn.
 1991. – 3 pages “Zum Stellenwert der Phonologie”, but it has no
Stellenwert for the rest of the book.
– Mallinson, Graham & Barry J. Blake. 1981. Language typology: Cross-
linguistic studies in syntax. Amsterdam: North-Holland. – No
phonology: honest titling.
– Comrie, Bernard. 1981. Language universals and linguistic typology:
Syntax and morphology. Oxford: Blackwell; 2nd edn. 1989. – No
phonology: honest titling.
– Croft, William. 1990. Typology and universals. Cambridge: Cambridge
University Press; 2nd edn. 2003. – Almost no phonology.
– Whaley, Lindsay J. 1997. Introduction to typology: The unity and
diversity of language. Newbury Park: Sage. – No phonology.
– Cristofaro, Sonia & Paolo Ramat (eds.) 1999. Introduzione alla
tipologia linguistica. Roma: Carocci. – 8 classic papers selected for
didactic exposition: one, by Joseph Greenberg, half phonological.
– Song, Jae Sung. 2000. Linguistic typology: Morphology and syntax.
Harlow: Pearson Education. – No phonology: honest subtitling.
– Moure, Teresa. 2001. Universales del lenguaje y linguo-diversidad.
Barcelona: Ariel. – No phonology.
– Feuillet, Jack. 2006. Introduction à la typologie linguistique. Paris:
Honoré Champion. – No phonology.
– Velupillai, Viveka. 2012. An introduction to linguistic typology.
Amsterdam: Benjamins. – Ch. 4: Phonology, on segment inventories,
syllable structures, tone systems: 27 of 517 pages, all exclusively based
on WALS.
– Moravcsik, Edith A. 2013. Introducing language typology. Cambridge:
Cambridge University Press. – Ch. 5: Phonological typology, on a par
with chapters on lexical, syntactic, and morphological typology!
– Kahl, Thede & Michael Metzeltin. 2015. Sprachtypologie: Ein
Methoden- und Arbeitsbuch für Balkanologen, Romanisten und
allgemeine Sprachwissenschaftler. Wiesbaden: Harrassowitz. – No
phonology.

For some time now, prospective typologists have also been able to benefit
from summer (or autumn or winter) schools. Among the earliest I am aware
of as exclusively devoted to this subject were the typology schools of the
Deutsche Gesellschaft für Sprachwissenschaft at Mainz/Germany in 1998,
of the MOSCOW Typological Circle in or near MOSCOW in 1998, 2000, 2002,
and 2005, of ALT in Cagliari/Sardinia in 2003; the most recent, run by the
Fédération Typologie et Universaux Linguistique of the CNRS, will be at
the Ile de Porquerolles/ France in the autumn of 2016. There have always
been one or even two phonology courses at these schools, but one or two
dozen were on offer for those typology students keener on other matters,
such as inflection and syntax, methodology, and language/family surveys.15
Seeking guidance beyond the textbook level, apprentice typologists, and
whoever else is in need of ready reference about this field, can now also
consult specialised handbooks – currently these two:
– Haspelmath, Martin, Ekkehard König, Wulf Oestereicher, & Wolfgang
Raible (eds.). 2001/02. Language typology and language universals. 2
vols. Berlin: Walter de Gruyter.
– Song, Jae Sung (ed.). 2010. The Oxford handbook of linguistic typology.
Oxford: Oxford University Press.

With its two weighty tomes, the first is almost a compendium of linguistics
in its entirety; its section on “Phonology-based typology” (5 chapters), a
chapter on syllable/accent-counting as one “salient typological parameter”,
and occasional passing references to sound matters add up to some 90 pages
of phonology, out of 1,800. The second has one out of 30 chapters devoted
to segment/phoneme inventories, which yields an even worse proportion of
phonology (if this is what this chapter is, and not phonetics) to non-
phonological typology.
A handbook of sorts, too, is this set of three volumes – with a great deal
of morphology subsumed under “syntax”, but with no companion set
Language typology and phonological description:
– Shopen, Timothy (ed.). 1985. Language typology and syntactic
description. Vol. 1: Clause structure. Vol. 2: Complex constructions.
Vol. 3: Grammatical categories and the lexicon. Cambridge: Cambridge
University Press. (2nd edn., co-edited by Matthew S. Dryer, 2007.)

3.2.6 Specialisation

Few linguists who see themselves and are seen by others as typologists,
whatever further categorisations they might invite, are genuine all-rounders:
some wholly devote themselves to methodology (and might in fact be
statisticians), but most specialise in one structural domain or another – and
inflection and/or syntax specialists far outnumber phonology (and
phonetics) specialists. (Specialisation is not entirely novel in typology: the
chief expertise of a pioneer such as Gabelentz, unlike that of most of his
Neogrammarian colleagues at Leipzig, lay in syntax.) As crown witnesses I
call the five previous presidents of ALT, Bernard Comrie, Marianne
Mithun, Nicholas Evans, Anna Siewierska, Johanna Nichols: as typologists
all are primarily known for their work in syntax and inflection, although
most have a sound component to their work, too. Of the one editor and 27
associate editors who have so far overseen ALT’s journal, LT, one was a
phonologist (Larry Hyman, though also with morphosyntactic work to his
typological credit), one a phonetician (Ian Maddieson), and one divided her
time between phonology and morphosyntax (Joan Bybee); syntax and
inflection were the main expertise of the rest, with one or the other on rare
occasions moonlighting as phonologists (William Croft, Nicholas Evans,
Frans Plank, Martine Vanhove). Further evidence pointing in the same
direction is conveniently gathered from the ALT membership directory
(http://ling-asv.ling.su.se/alt_filer/membership.html): as their “special
interests” members do mention phonology or phonetics as such as well as
particular phonological/ phonetic topics such as tone, nasalisation and other
phonological processes, phonotactics, prosody, sound change, speech
perception; but these figures cannot compete with mentions of syntax and
morphology and particular morphosyntactic topics. Phonologists might of
course be doing their typology elsewhere – a question which we need to
return to (§3.2) before we can conclude that among today’s linguists with
typological interests phonologists are comparatively rare.

4 Reasons why

4.1 Is typology special?

4.1.1 Phonology outside typology

Undeniably, then, however you look at it, typology has been, and continues
to be, about co-variation and co-evolution in syntax and inflection much
more so than in phonology. Now, what are the reasons for this imbalance?
And is it desirable, and possible, to redress it in future?
A first step towards an answer is to raise a further question: Is typology
special?
Above (§2.2.4) I compared typology to historical linguistics with regard
to the amount of phonology one finds at specialised conferences and in
specialised journals, concluding, if tentatively, that it is less than
morphosyntax, too. It is probably only among those historical linguists
active in comparative (and internal) reconstruction that phonological
expertise will be at a premium. Making further comparisons with subfields
within linguistics where languages are being studied from some sort of a
selective perspective and with some special ulterior motives, the likelihood
is that a similar imbalance will be encountered. Take psycho- and
neurolinguistics, or sociolinguistics and anthropological linguistics, or
computational linguistics, and syntax (but not necessarily inflection) will
receive more attention than phonology, although the subject matter
supposedly is languages as such and there would not seem to be inherent
reasons for some structural domains being prioritised over others. Perhaps
dialectology is a rare subfield where the preferences among syntax and
phonology(-cum-phonetics) are reversed, with inflection possibly on a par
with phonology (and with the lexicon ahead of both).
If phonology is pitched against syntax in linguistics as a whole, it will
again emerge as the loser, although it will probably win second place before
morphology. Relevant evidence are the contents of general linguistics
journals (you name them) or the membership lists of learned societies
catering for the discipline as a whole (to name some where I made
spotchecks: Societas Linguistica Europaea, Linguistic Society of America,
Linguistics Association of Great Britain, Philological Society, Deutsche
Gesellschaft für Sprachwissenschaft, Société Linguistique de Paris, Società
di Linguistica Italiana, Società Italiana di Glottologia, Australian Linguistic
Society, Linguistic Society of India): more linguists specialise in, and
publish on, syntax than phonology or also morphology.16 As a subset of
linguists, typologists thus are not ESPECIALLY averse to phonology, then:
they are boringly average the way they like and dislike to specialise. If they
are special, it is probably in their idiosyncratic partiality to inflectional
morphology.
4.1.2 Quantum sufficit

The question is: Why? Is phonology felt by trainee linguists to be

prohibitively difficult? Or too easy, too little of an intellectual challenge? Is
the way phonology is being taught too forbidding, with the numbers of
initiates and potential future teachers thus dwindling from generation to
generation? Does phonology lack the allure of theoretical promise or
practical usefulness or do phonologists lack charisma, and does syntax and
do syntacticians have it? Whether or not these are plausible considerations,
there is also a simpler answer, which is that there just is less phonology to
be studied in comparison with syntax and morphology.
As an approximation to what languages are about, take descriptive
grammars (and dictionaries, but lexical typology is not our concern here).
Framed by an introductory presentation of the language to be described and
perhaps a core vocabulary and texts as appendices, they typically have three
parts: phonology, morphology (or in fact only inflection, with word
formation often set aside for separate treatment), and syntax, in this or more
rarely reverse order. Here are a few specimens:
– Sweet, Henry. 1892/98. A new English grammar: Logical and
historical. 2 volumes. Oxford: Clarendon Press. – Phonology 75 pages,
out of over 600.
– Jespersen, Otto. 1909–49. A Modern English grammar on historical
principles. London: Allen & Unwin. – One volume on “sounds and
spelling”, six on syntax and morphology.
– Sapir, Edward. 1922. The Takelma language of southwestern Oregon. In
Franz Boas (ed.), Handbook of American Indian languages, vol. 2, 1–
296. Washington: Bureau of American Ethnology. – Phonology 43
pages, morphology 247 pages (with no separate syntax); Sapir’s other
grammar, of Southern Paiute (1930), almost doubles the share of
phonology to one third.
– Bloomfield, Leonard. 1962. The Menomini language. Ed.
[posthumously] by Charles F. Hockett. New Haven: Yale University
Press. – “Sounds” and “morphophonemics” 46 pages out of 507 (plus
22 pages on “morphologic processes and constructions”, many filled
with attendant morphonological “modifications”, to be distinguished
from “morpholexical variation”).
– Kibrik, A. E., S. V. Kodzasov, I. P. Olovjannikova, & D. S. Samedov.
1977. Opyt strukturnogo opisanija arčinskogo jazyka. 4 volumes.
Moskva: Izdatel’stvo moskovskogo universiteta. – One volume lexicon,
word formation, and phonetics, with the latter (the responsibility of
Kodzasov) 160 pages out of 350, two volumes syntax, one volume texts
and vocabulary.
– Heidolph, Karl Erich, Walter Flämig, & Wolfgang Motsch (eds.). 1980.
Grundzüge einer deutschen Grammatik. Berlin: Akademie-Verlag. –
Phonology (the responsibility of Wolfgang U. Wurzel) 95 pages out of
over 1,000.
– Rice, Keren. 1989. A grammar of Slave (Mouton Grammar Library 5).
Berlin: Mouton de Gruyter. – Seven chapters out of 48 (including an
introduction to a typological character sketch of Slave as well as texts)
on phonology.
– Rischel, Jørgen. 1995. Minor Mlabri: A hunter-gatherer language of
Northern Indochina. Copenhagen: Museum Tusculanum Press. – One
chapter on phonology (called a “sketch”), morphology, syntax each, of
18, 50, and 65 pages respectively.
– Rennison, John R. 1997. Koromfe (Lingua/Croom Helm/Routledge
Descriptive Grammars). London: Routledge. – Syntax 137 pages,
morphology 227 pages, phonology 97 pages, lexicon 57 pages (with the
grammar structured according to the original LDS Questionnaire,
devised by Bernard Comrie and Norval Smith, the latter responsible for
the phonological questions: https://www.eva.mpg.de/lingua/tools-at-ling
board/questionnaire/linguaQ.php).17
– Evans, Nicholas, 2003. Bininj Gun-Wok: A pan-dialectal grammar of
Mayali, Kunwinjku and Kune. 2 volumes. Canberra: Pacific Linguistics.
– The two chapters on phonology and morphophonemics add up to 46
pages, out of over 700.
– Hualde, José Ignacio & Jon Ortiz de Urbina (eds.). 2011. A grammar of
Basque (Mouton Grammar Library 26). Berlin: Mouton de Gruyter. –
Phonology (the responsibility of Hualde and Gorka Elordieta) 97 pages,
morphology 249 pages, syntax 529 pages.
– Crane, Thera M., Larry M. Hyman, & Simon Nsielanga Tukumu et al.
2011. A grammar of Nzadi [B.865]: A Bantu language of the
Democratic Republic of the Congo. Berkeley: University of California
Publications in Linguistics – Two chapters out of ten on phonology
 (sound system, tone), plus one appendix on phonological reconstruction,
a similar proportion as in a similar earlier collective grammar by Hyman
on Aghem.

This sample is not entirely random: though differing in many respects, what
my dozen grammars have in common is that they were written or co-written
by recognised or in fact eminent phonologists.18 They are a rather select
group, because grammar-writing is not a common activity of phonologists,1
9 and one could therefore suspect my sample to be biased in favour of

phonology: grammar writers with no special phonological expertise could

be expected to do worse on this count. But COULD one do worse, purely in
terms of numbers of pages or chapters – if phonology experts themselves
standardly get along with about a fifth or less of the space that they need for
syntax and inflection to set out the phonology of the language they are
describing?
Sometimes the contrast is even starker in abbreviation. Ten-or-so-page
sketches or typological profiles of languages, as often prefixed to full
grammars or also published separately as basic sources of information for a
general audience, typically reduce phonology to segment inventory tables
or omit it altogether, but rarely are quite as laconic in highlighting what is
special (or not special, but typologically expected) about the syntax and
inflection of the languages concerned. For example, unusual for an
introduction to linguistics but true to its title, How languages work (edited
by Carol Genetti and essentially the work of her department at UC Santa
Barbara, Cambridge: Cambridge University Press, 2013) is accompanied by
13 “language profiles”; however, although the body of the book includes
three expert chapters on phonetics, phonology, and prosody, twelve portraits
give the impression that languages work without phonology, while one, on
Kabardian (by the author of the phonetics and phonology chapters, Matthew
Gordon), exclusively expands on the phonology of this language.
If such comparisons are something to go by, and if there are no
languages reversing these proportions whether fully described or aptly
sketched,20 we can conclude that, given comprehensive grammars, purged
of excessive verbiage and reduced to bare statements of units, paradigmatic
systems, and rules for and constraints on constructions, the “Minimum
Description Length” or “Kolmogorov Complexity” of any language will be
substantially less for their phonology than their syntax and inflection.
 Owing to this universal, as yet undisconfirmed, it will only be natural,
then, if – like elsewhere in linguistics – more research into linguistic
diversity and unity is about what there is more of to compare across
languages: syntax and inflection.
The inferiority of phonology may not only be quantitative, but
qualitative, too. Phonological patterns have been claimed to be less
complex than syntactic patterns in terms of the Chomsky Hierarchy of
formal objects: as pointed out by Heinz & Idsardi (2013),21 syntax can be
finite, regular, context-free, mildly context-sensitive, and context-sensitive,
covering almost the entire spectrum from lowest to highest complexity
(with computably-enumerable patterns the only ones unattested), while the
grammar of sound is limited to its lower region, namely to finite and regular
patterns. As such, this imbalance is unlikely to be behind the poorer
performance of phonology in typology: context-free and context-sensitive
patterns are not really what syntactic typology has primarily been thriving
on, with instances of such higher complexity in short crosslinguistic supply
even in syntax. However, Heinz & Idsardi’s explanatory speculation that
syntax and phonology draw on different learning mechanisms could have
typological relevance. Though learnability has rarely been an issue in
typology, any constraints on structural diversity clearly facilitate learning,
just like constraints on formal complexity do, and the relative importance of
co-variation detection in acquisition might be greater in one domain than in
another. If it is at a premium in syntax, it could be the manner of its learning
that ultimately accounts for the typological pre-eminence of syntax.

4.2 Is phonology different?

While the conclusion seems credible that there is less phonology to be
investigated especially in comparison with syntax, or less complexity to be
grappled with in phonology than in syntax, and hence there will be fewer
phonologists to investigate it, in linguistics in general and therefore also,
with a view to co-variation/co-evolution, in typology, it is not a wholly
satisfying diagnosis. There are further symptoms of a deeper disparity that
can hardly be accounted for through such quantitative or qualitative
differences between subject matters as just mentioned. They become
apparent when we turn around the question, from How much phonology is
there in typology? to How much typology is there in phonology? Is there
proportionally less than there is in syntax and inflection? While there are
quantitative differences here too, they are not the full story.

4.2.1 Typology in phonology

Probably, although concrete figures are impossible to obtain, linguists who

perceive themselves and are perceived by others as phonologists, of all
persuasions combined, are proportionally less likely than syntacticians and
morphologists to be members of typological associations such as ALT, to
attend conferences advertised as typological, and to publish in journals of a
typological profile (such as LT, Studies in Language, Sprachtypologie und
Universalienforschung, Linguistic Discovery). But then, there is no
monopoly on typology: journals and conferences dedicated to phonology
itself (e.g., the Manchester Phonology Meeting, Laboratory Phonology, the
journal Phonology) surely do not discourage or ostracise typological
contributions. But do they get fewer of those than journals and conferences
dedicated to syntax get (e.g., meetings like Syntax of the World’s
Languages, Wiener Morphologietagung, the journals Syntax, Morphology,
Word Structure)? Categorising work as typological whenever its theme is
the dialectics of diversity and unity and it specifically focuses on co-
variation/co-evolution among structural variables, whether or not it has
“typology” in its title, I would be surprised if differences turned out to be
significant: my impression is that there is relatively little on the recent
record that qualifies as typology for all three domains, allowing for certain
variations depending on theoretical frameworks.22 By contrast, fora for
general linguistics, or indeed general science (such as Science, Nature,
PLOS ONE), have seen a very tangible increase in typological studies over
the last four or five decades, but again it does not look like phonology is
conspicuously lagging behind morphology or syntax.

4.2.2 Teaching and texts

In introductions to phonology, to judge by teaching materials that are

available on the internet, phonological typology has become an element at
many institutions, which may or may not teach introductions to typology as
such; morphology and syntax introductions also often have typology
components, but perhaps more frequently this is taken care of in separate
typology courses. The textbook market would seem to reflect the situation
well: for example, picking out two popular series, in the Cambridge
Introductions to Language and Linguistics both the phonology and the
morphology texts, by Odden and Lieber, have chapters on typology, while
the Understanding Language texts lack such separate chapters for
phonology (Gussenhoven & Jacobs) as well as for morphology
(Haspelmath (& Sims)) and syntax (Tallerman), but address typological
issues in subsections and en passant, and more extensively for syntax and
morphology than for phonology.
– Lieber, Rochelle. 2009. Introducing morphology. Cambridge:
Cambridge University Press. (2nd edn. 2015.)
– Odden, David. 2005. Introducing phonology. Cambridge: Cambridge
University Press. (2nd edn. 2013.)
– Gussenhoven, Carlos & Haike Jacobs. 1998. Understanding phonology
London: Arnold. (3rd edn. 2011.)
– Haspelmath, Martin. 2002. Understanding morphology. London:
Arnold. (2nd edn. 2010, with co-author Andrea D. Sims.)
– Tallerman, Maggie. 1998. Understanding syntax. London: Arnold. (4th
edn. Routledge, 2015.)

4.2.3 Languages and linguistic theory

Over and above such similarities there is a curious and seemingly trivial
difference between introductions to phonology on the one hand and those to
morphology and syntax on the other: the former consistently serve up more
languages. According to their indices, Odden’s and Gussenhoven &
Jacobs’s phonology texts in one way or another make reference to some 150
languages, comparing to a little over 100 in Tallerman’s syntax and in
Haspelmath (& Sims)’s as well as Lieber’s morphology texts. An early text
such as Larry Hyman’s Phonology: Theory and analysis (New York: Holt,
1975) had examples from and analyses for over 80 languages, at a time
when introductions to syntax would make ends meet with one (often the
author’s own) and Peter Matthews’ Morphology, the first Cambridge
Textbook in Linguistics (Cambridge University Press, 1974), got along with
a modest 20.
 And such an imbalance is not encountered in textbooks alone. The
language index, for example, of the Blackwell Handbook of phonological
theory (edited by John A. Goldsmith, 1995) has 422 entries, while its
companion Handbook of morphology (edited by Andrew Spencer & Arnold
M. Zwicky, 1998), apart from missing the tag “theory” in the title, only has
159 entries for languages and families, incorporated in the subject index.
Major theoretical works in phonology are routinely brimming with
languages, too: to choose almost randomly, only think of Trubetzkoy’s
Grundzüge der Phonologie (Travaux du Cercle Linguistique de Prague 7,
Prague 1939), Wolfgang Dressler’s Morphophonology (Ann Arbor:
Karoma, 1985), John Goldsmith’s Autosegmental and metrical phonology
(Oxford: Blackwell, 1990), Bruce Hayes’ Metrical stress theory (Chicago:
University of Chicago Press, 1995), or Robert Ladd’s Intonational
phonology (Cambridge: Cambridge University Press, 1996). Devoted to a
single language, even Noam Chomsky & Moris Halle’s The sound pattern
of English (New York: Harper & Row, 1968) has a separate language index,
with as many as 101 entries. Comparable language coverage in landmark
monographs is hard to find in syntax: perhaps Guglielmo Cinque, with titles
such as Adverbs and functional heads: A cross-linguistic perspective
(Oxford: Oxford University Press, 1999) or The syntax of adjectives: A
comparative study (Cambridge, MA: MIT Press, 2010), or Mark Baker,
with The polysynthesis parameter (Oxford: Oxford University Press, 1996),
Lexical categories (Cambridge: Cambridge University Press, 2005), or The
syntax of agreement and concord (Cambridge: Cambridge University Press,
2008), come closest, but they are exceptions. Morphology is in between, as
is suggested by the language counts for some works from the morphological
Renaissance of the 1980s: Frans Plank, Morphologische (Ir-)Regularitäten
(Tübingen: Narr, 1981) has 20+; Wolfgang Wurzel, Flexionsmorphologie
und Natürlichkeit (Berlin: Akademie-Verlag, 1984) 75; Joan Bybee,
Morphology (Amsterdam: Benjamins, 1985) 50+; Andrew Carstairs,
Allomorphy in inflexion (London: Croom Helm, 1987) 48. The
morphological counterpart to SPE, Mark Aronoff’s Word formation in
Generative Grammar (Cambridge, MA: MIT Press, 1976), is mostly about
one language, English, but is livened up with tangential references to ten
others.
What could seem an idle ranking of publications by language density
arguably bears witness to a divergence of research traditions, and of an
estrangement of professional sub-communities, between syntax and
phonology.23
In syntax, and similarly in morphology, a tradition had developed of
elaborating theories and frameworks on a narrow basis: at the expense of
confronting crosslinguistic diversity, theorising would be informed by in-
depth looks at selected structural phenomena in one or a few particular
languages, and not just by utterances and texts, but also by native
judgments about them. This was not only the policy in Generative
Grammar: a collection such as Syntactic theory 1: Structuralist (edited by
Fred Householder, Harmondsworth: Penguin, 1972), assembling 23 classic
readings, accumulates a little over 100 languages and families, but most
individual chapters make their theoretical points on the basis of individual
languages, mostly English, with Ilocano (L. Bloomfield), Bilaan (K. L.
Pike), Teleéfoól (P. Healey), Sundanese (R. H. Robins), Vietnamese (P. J.
Honey), and Eskimo-Aleut (K. Bergsland) as sporadic co-stars and the rest
as bit-part players. (The only multi-language exceptions in this reader are
W. S. Allen, Transitivity and possession, and B. L. Whorf, Grammatical
categories.) A pre-structuralist classic, Wilhelm Havers’ Handbuch der
erklärenden Syntax (Heidelberg: Winter, 1931), had limited itself to a subset
of Indo- European, although with illustrations from spoken modern
languages and with occasional comparisons of these “Kultursprachen” to
none-too-specific “Natursprachen”.
Eventually, from the 1960s and 70s onwards, as typology was
beginning, through individual efforts like Joseph Greenberg’s, to attract
wider attention than ever before, languages in the plural would re-assert
their right to be heard not only for their phonology, but also their syntax.
With the Generative paradigm continuing to dominate, a misperception
arose of syntax being done in two ways: “theoretically”, engaging with
single languages against the backdrop of Universal Grammar (largely taken
for granted), vs. “descriptively”, dealing with multiple languages and
inductively inferring crosslinguistic generalisations about co-variation/co-
evolution. When “theorists” were withholding the honorary epithet
“theoretical” from the latter line, where theorising was primarily about
finding and explaining inductive generalisations, they were probably
encouraged by an occasional lack of subtlety in conceptualising syntactic
structures and processes and a reluctance to countenance abstract
representations. Though far better informed crosslinguistically, syntactic
typology as part and parcel of the “descriptive” approach remained
theoretically indeed sometimes a bit basic.24 It seemed like grammar was in
bare essence to be conceived of as a checklist of variables, possibly with
only two values, plus or minus – OV or VO? Adposition before/ after NP?
Genitive before/ after head noun? Ergative or accusative or other
alignment? A definite article, yes or no? Zero copula? Dual? Inclusive-
exclusive? Gender, and where applicable: how many? Doing typology then
typically meant searching for co-variation among such variables whose
values the typologist could easily glean at a glance from lots of descriptive
grammars.
In phonology, theory and framework development had never been
divorced from crosslinguistic awareness to a similarly alarming extent.
There were thus no grounds for a multilingual “descriptive” phonology to
split off from a monolingual “theoretical” phonology à la syntax.
Languages in the plural remained at the core of phonological theorising. In
principle this meant that typology could have been done as part of
theoretical phonology, rather than in a separate community where members
defined themselves as typologists and where “non-theoretical” syntacticians
were setting the agenda. And to some extent it was – namely to the extent
that phonological grammar could be conceived of as a checklist and values
could conveniently be checked for co-variation/ co-evolution:25 Does the
language have this segment and that? Does it have quantity contrasts for
vowels/ consonants? Does it permit onset clusters, and if so which? How do
its syllables go beyond CV? Does it enforce final devoicing? Does it have
vowel harmony? Is it tonal? Level or contour tones, and how many? Which
syllable of the word does it stress? However, this sort of thing – like listing
segment inventories and phonotactic templates – was never considered all
there is for phonological theory to address, and, when it was a point of
departure, it did not represent a theoretical issue or conclusion. Hence,
much of what was at the heart of phonological theorising, in variable
frameworks but invariably richly informed by diverse languages, has never
translated into phonological typology of the checklist-based variety, a style
so increasingly popular for syntactic and morphological typology and
accounting for the latter-day bulk of it.
When phonology is seen as phonologists see it, aiming at adequate
description and at the same time at making theoretical sense of what is
being described, the grammar of sound is substantially growing in sheer
volume. When syntax chapters in weightier descriptive grammars are
compared with monographs on the syntax of the same languages, like those
of the Cambridge Syntax Guides, there is no dramatic mismatch. The remit
for authors of these Guides, of which a dozen have so far been published
(mostly for European languages), is to be both descriptive and theoretical,
while the editorial team itself (Peter Austin, Bernard Comrie, Joan Bresnan,
David Lightfoot, Ian Roberts, Neil Smith), like the intended audience, is
patently divided between “descriptive” and “theoretical” allegiances. The
authors of the Phonology of the World’s Languages series of Clarendon/
Oxford University Press are likewise instructed, although by a single editor
(Jacques Durand), to attend to both description and explanation, to the
benefit of a single undivided body of intended readers, which should not be
put off by differences between theoretical frameworks – and the resultant
monographs, as yet 19, far exceed what would be found in even the most
comprehensive of descriptive grammars. Intriguingly, several languages
have both a Cambridge syntax guide and an Oxford phonology portrait
devoted to them, and the Kolmogorov Complexity or at any rate book
length is not necessarily less for the phonology: Arabic phonology (and
morphology; author Janet Watson) 336 pages, syntax (authors Joseph Aoun
et al.) 260 pages; Catalan phonology (Max Wheeler) 400, Spanish syntax
(Karen Zagona) 300; Welsh phonology (S. J. Hannahs) 198, Welsh syntax
(Robert Borsley & Maggie Tallerman) 412; Icelandic (& Faroese)
phonology (Kristján Árnason) 368, syntax (Höskuldur Thráinsson) 580;
Dutch phonology (Geert Booij) 218, syntax (Jan-Wouter Zwart) 418;
German phonology (Richard Wiese) 368, syntax (Hubert Haider) 368;
Hungarian phonology (Péter Siptár & Miklós Törkenczy) 336, syntax
(Katalin É. Kiss) 292; Chinese phonology (San Duanmu) 382, syntax
(James Huang et al.) 404.26

4.2.4 Languages and typology

Now, when phonology is being done with wide crosslinguistic awareness

and with equal descriptive and theoretical poise, will the inevitable outcome
be phonological typology – homemade, but like syntactic and
morphological typology aspiring to chart diversity and to discover order or
indeed unity? Is theoretical phonology, inseparable from descriptive
phonology, methodically investigating whether variables co-vary or co-
evolve, just like syntactic and morphological typology do, although not
normally in communion with “theoretical” syntax and morphology?
Having many languages at one’s analytic fingertips does not make one a
typologist. Charting diversity is one thing, seeking order is another – and
typology’s remit includes both. In multilingual theoretical phonology it is
often only the former which is on the agenda, and success consists in one’s
theoretical framework being able to insightfully deal with whatever can be
found across languages, however diverse. The chief interest is in what there
is and in its theoretical accommodation, rather than in what there isn’t – in
how diversity is constrained through variables systematically co-varying or
co-evolving. Optimality Theory is an obvious framework to mention here,
because it is sometimes considered “inherently typological”, if only at the
expense of redefining and narrowing typology’s remit. But to also
exemplify my point through a particular work: in metrical phonology a
typology of feet was developed (prominently in Hayes’ Metrical stress
theory, a book hard to beat on language density), which indeed aims to
constrain crosslinguistic diversity insofar as no language is supposed to
employ further conceivable kinds of feet beyond these original types
(essentially only iambs and trochees with variations). However, little
comparative effort went into asking, further, whether single languages are
free to employ different types of feet for different purposes (e.g., stress, all
kinds of segmental processes affecting syllable structure, poetic meter) and
for different lexical domains, or whether the choice of one foot type is a
determinant for all foot-sensitive patterns of a language. Foot typology has
only been taken in the direction of real co-variation/ co-evolution typology
by Elan Dresher & Aditi Lahiri (The Germanic foot: Metrical coherence in
Old English. Linguistic Inquiry 22. 251–286, 1991), when they
demonstrated that languages are “metrically coherent” – or at any rate the
Old Germanic languages that they examined were, doing all their relevant
business on the basis of just the resolved moraic trochee and no other foot
type. Prosodic typology, where Hyman has long been arguing for co-
variation among variables, rather than types of languages (“tone
languages”, “stress languages”, “pitch-accent languages”), as the object of
inquiry, is also moving in this direction, on such evidence as Prosodic
typology: The phonology of intonation and phrasing (2 volumes,
Cambridge: Cambridge University Press, 2005/2014), whose editor, Sun-
Ah Jun, concludes by extracting, from two dozen language-particular
accounts (all done in the same descriptive framework, ToBI), implicational
generalisations about prosodic prominence and rhythmic or prosodic units
such as these: “In stress languages (such as English, Arabic, Farsi, Swedish,
Chickasaw), the prominence of a word is always marked by postlexical
pitch accent (marking the head of the word), but not often by marking the
edge of the word. [...] Languages that do not have any feature of lexical
prosody (such as French, Bengali, Korean) mark the prominence of the
word demarcatively at the postlexical level” (updated in the second volume
through introducing the concept of “macrorhythm”).
These are phonology-internal developments aligning phonology with
morphosyntactic typology, but they have not found much resonance in
typological circles. Other than a subject matter and theoretical angles
unfamiliar, even impenetrable to many syntacticians and morphologists,
there remains a difference in emphasis on methodology. Morphologists and
syntacticians, often maligned as “non-theoretical”, have made huge efforts
since Georg von der Gabelentz to hone typological methodology: inductive
generalisations are no longer arrived at as naively as in days of old, when
assembling samples and determining statistical significance, if done at all,
was amateurs’ work. In phonological typology, theoretical accommodation,
far above the level of checklist grammar, continues to be prized more highly
than methodological sophistication, and cherry-picking continues to be
preferred over sampling.
There are lessons to be learnt either way.

4.2.5 Is phonology hopeless?

I don’t think typological phonology is hopeless, but I am less confident

about phonological typology.
Intonation and prosody in general have variously been conjectured to be
subject to intra-language variation of an extent unparalleled by syntax and
morphology. However, as shown by the recent Prosodic typology collection
which does recognise such variation, this does not seem an insurmountable
obstacle for typological system-seeking. There is a methodological
challenge here, insofar as the sampling points in (non-phonological)
typology have typically been languages:27 to cope with diversity of
prosodic dimensions, it is dialects and idiolects – individual mental
grammars – that ought to be sampled, once typological phonology proceeds
from cherry-picking to sampling.
The grammar of sound has been conjectured to be radically different
from other grammar: “[phonology] is very complex; it does not seem to
have any of the nice computational properties of the rest of the system”
(Chomsky 2012: 40). If it is these “nice computational properties” of syntax
and inflection which are responsible not only for unity, but also for
orderliness of diversity, constrained by co-variation or co-evolution, then
typologising in phonology would indeed be doomed. However, against such
speculation, informed by a perhaps somewhat minimalistic outlook on
syntax, stand arguments for rich parallels, like those collected in
Honeybone & Bermúdez-Otero (2006): such structural and architectural
parallels have all variously figured in both syntactic and phonological
typology. Also, as touched on above (§4.1.2), the exact opposite case to
Chomsky’s has been made by Heinz & Idsardi (2013) for phonology being
formally LESS complex than syntax, and different domain-specific learning
strategies might have typological repercussions. If you are unconvinced by
such theoretical reasoning and the body of work in phonology that is
typologically minded, read on because the present collection is the best
argument that typological phonology is not a dead loss.
Unlike typological phonology, it is phonology in typology that might be
hopeless, albeit for extraneous reasons. It is to be feared that typologists,
who as we saw are mostly syntacticians and inflectional morphologists by
professional specialisation, will continue to lack the serious grounding that
would enable them to appreciate what is going on in this (for them)
hermetic field of phonology, once it rises above the checklist level of
grammar and the going gets heavier than the tabulation of segment
inventories. For the time being, there will probably continue to be two
communities far apart and indifferent towards one another, phonologically
challenged organised typologists here and phonologists as do-it-yourself
typologists over there.

Acknowledgements: I can’t seriously pretend to be telling an inside story

of phonology, typological or otherwise; but without the many years of
casual instruction from Aditi Lahiri and Larry Hyman, this work would
have been pure science fiction (and I’m not sure whether utopian or
dystopian). For contributing to my phonological education on frequent
convivial occasions I’m also grateful to Elan Dresher, Carlos Gussenhoven,
Paul Kiparsky, and Tomas Riad. Comments from Larry Hyman, Aditi
Lahiri, and Edith Moravcsik have helped to improve the present chapter.
It took us ages to get our act together on this volume – which
incidentally began with a failed attempt by my co-editor and myself to have
phonology and typology discussed at an ALT conference. Thanks
everybody for their patience: It’s all my fault.

References
Note: Works dealt with as historiographical data are referenced in the body of the chapter.

Anderson, Stephen R. 1985. Phonology in the twentieth century: Theories of rules

and theories of representations. Chicago: University of Chicago Press.
Chomsky, Noam. 2012. The science of language: Interviews with James
McGilvray. Cambridge: Cambridge University Press.
Gabelentz, Georg von der. 1894. Hypologie der Sprachen, eine neue Aufgabe der
Sprachwissenschaft. Indogermanische Forschungen 4. 1–7.
Gabelentz, Georg von der. 1901. Die Sprachwissenschaft, ihre Aufgaben,
Methoden und bisherigen Ergebnisse. 2nd edn. Leipzig: Tauchnitz.
Heinz, Jeffrey & William Idsardi. 2013. What complexity differences reveal about
domains in language. Topics in Cognitive Science 5. 111–131.
Honeybone, Patrick & Ricardo Bermúdez-Otero (eds.). 2006. Linguistic
knowledge: Perspectives from phonology and from syntax. Thematic issue of
Lingua 116(5).
Hyman, Larry M. 2007. Where’s phonology in typology? Linguistic Typology 11.
265–271.
Morpurgo Davies, Anna. 1975. Language classification in the nineteenth century.
In Thomas A. Sebeok (ed.), Current trends in linguistics, vol. 13:
Historiography of linguistics, 717–827. The Hague: Mouton.
Morpurgo Davies, Anna. 1997. History of linguistics, vol. 4: Nineteenth century
linguistics. London: Longman.
Plank, Frans. 1991. Hypology, typology: The Gabelentz puzzle. Folia Linguistica
25. 421–458.
Plank, Frans. 1992. Adam Smith: Grammatical economist. In Peter Jones &
Andrew Skinner (eds.), Adam Smith reviewed, 21–55. Edinburgh: Edinburgh
University Press.
Plank, Frans. 1993. Des Lord Monboddo Ansichten von Ursprung und Entwicklung
der Sprache. Linguistische Berichte 144. 154–166.
Plank, Frans. 1995. Professor Pott und die Lehre der Allgemeinen
Sprachwissenschaft. Zeitschrift der Deutschen Morgenländischen
Gesellschaft 145. 328–364.
Plank, Frans. 1998. The co-variation of phonology with morphology and syntax: A
hopeful history. Linguistic Typology 2. 195–230.
Plank, Frans. 2001. Typology by the end of the 18th century. In Sylvain Auroux et
al. (eds.), History of the language sciences: An international handbook on the
evolution of the study of language, 1399–1414. Berlin: Walter de Gruyter.
Shibatani, Masayoshi & Theodora Bynon (eds.). 1995. Approaches to language
typology. Oxford: Clarendon.
Paul Kiparsky
Formal and empirical issues in
phonological typology

Abstract: The word level in the sense of Lexical Phonology and Stratal
OT, here referred to as the l-phonemic level, is a linguistically significant
level of representation, which captures what was right about the structural
phonemic level without inheriting its well-known problems. It does so in
virtue of encoding non-contrastive but distinctive as well as contrastive but
non-distinctive phonological properties. I show that phonological systems
which appear marginal or aberrant from the perspective of structural
phonemics and generative phonological underlying representations are
normalized at the l-phonemic level, and that certain phonological universals
become exceptionless only at this level. Dramatic instances include putative
vertical and one-vowel systems such as those of Arrernte and Kabardian,
and apparently syllable-less languages such as Gokana. I further argue that
“external evidence” from change, dispersion, poetic conventions, and
language games supports l-phonemic representations rather than classical
phonemic representations.
The larger methodological point is that there are no theory-neutral
grammars, and consequently no theory-neutral typology. In terms of
Hyman’s (2008) distinction, there are no “descriptive” universals of
language. All universals are “analytic”, and their validity often turns on a
set of critical cases where different solutions can be and have been
entertained. Therefore the search for better linguistic descriptions, more
illuminating typologies, and stronger cross-linguistic generalizations and
universals must go hand in hand.

1 Lexical representations
1.1 Problems with phonemes
Typological generalizations and universals are explicanda for linguistic
theory, but they are themselves theory-dependent, for in order to be
intelligible and falsifiable they must adhere to some explicit descriptive
framework. This mutual dependency comes to a head at the margins of
typological space, where reconciling typologies with descriptive
frameworks and the analyses dictated by them can involve a labyrinth of
choices. I explore a few of the tangled paths through it in the realm of
syllable structure and vowel systems.
Phonological typology has been based on three distinct levels of
representation: phonemic, phonetic, and morphophonemic (underlying,
“systematic phonemic”). Most work on segment inventories is framed in
terms of phonemic systems in the tradition of Trubetzkoy (1929, 1939),
Jakobson (1958), and Greenberg (1978). A major resource is the UPSID
collection of phonemic systems (Maddieson 1984, 2013, Maddieson &
Precoda 1990), which has the virtue of being genetically balanced (to the
extent possible), carefully vetted, and to some extent normalized to conform
to a standard set of analytic principles.28 The same resource has also been
used by phoneticians to investigate the typology of speech sound
inventories (Schwartz et al. 1997). Proponents of Dispersion Theory have
attempted to model the UPSID vowel systems, even though the theory is
strictly speaking about the phonetic realization of phonemes (maximization
of perceptual distance and minimization of articulatory effort). A growing
body of typology crucially relies on underlying representations (phonemes
in the classical generative phonological sense), such as Dresher (2009) and
Casali (2014). The analysis of Arrernte syllable structure that Evans &
Levinson (2009: 434) cite as part of their argument that universals are
“myths” is based on abstract underlying representations (Section 2.1
below).
Throwing abstract morphophonemic, phonemic, and phonetic
inventories in the same bin is unlikely to produce coherent typologies and
universals. So what kinds of categories and representations should
typologists look at? At least two criteria follow from the nature of typology
itself. We want typological categories that correlate with each other and
show some historical stability. And we want the categories to be based on
independently justified linguistically significant representations.
 It is not obvious that the phonemic level satisfies either of these criteria.
There is persuasive evidence for some level between abstract underlying
representations and phonetics at which phonology is accessed in language
use, including the classic “psychological reality” or “external evidence”
diagnostics such as versification and language games, as well as language
change, including sound change and phonologization, analogy, and
borrowing. But phonemic theories do not converge on this level. Depending
on how such fundamental issues as biuniqueness, invariance, linearity, and
morphological conditioning (“grammatical prerequisites”, junctures) are
resolved, phonemic analyses diverge for all but the simplest textbook cases,
and quite drastically for typologically challenging outlier systems of the
sort I’ll focus on here. For example, if we require linearity (a phoneme
cannot correspond to a sequence of sounds) Kabardian has seven vowel
phonemes. If we don’t require linearity, but do require biuniqueness, it has
three vowel phonemes; otherwise it has two. Each of these phonemic
analyses is currently advocated by researchers on Kabardian (Section 3.2).
I shall argue that what language users actually access, and what
language change reveals, is not exactly the classical phonemic level, but the
level of representations that emerges from the lexical phonology (in the
sense of Lexical Phonology and Stratal OT). I’ll refer to this as the level of
LEXICAL REPRESENTATIONS and to its elements as L-PHONEMES. I will argue
first that the classic diagnostics fit lexical representations rather than
phonemic representations, where they differ, and then that the typology of
syllable structure and phonological systems is best served by lexical
representations. At this level phonological systems converge on significant
common properties, and some important phonological near-universals turn
exceptionless. The global factors of dispersion, symmetry, and naturalness,
to the extent that they shape phonological systems, appear to take effect at
this level.
Although I concentrate on unusual syllabification and vowel systems,
these just highlight some inherent tensions between phonemics and
typology that arise less conspicuously in most languages. They are due to
the sparseness and segmentalism of phonemic representations.
The point of phonemic representations is that they should be stripped of
all predictable information. In Jakobson’s words (1958 [1962]: 525): “A
typology of either grammatical or phonological systems cannot be achieved
without subjecting them to a logical restatement which gives the maximum
economy by a strict extraction of redundancies.” I’ll defend the opposite
view, that a specific class of redundant information is phonologically
relevant,29 and that its omission can lead typology astray – namely just that
increment of information which accrues from the phonological computation
in the lexical module. In particular, lexical representations include word
stress, if the language has it, and word-level syllable structure, regardless of
whether these things are predictable in the language or relevant to any
phonological processes in it.30 Lexical representations include this
information for principled reasons, as we’ll see directly. But they exclude
postlexical feature specifications from sandhi processes and phonetic
implementation rules. In this respect lexical representations are thus more
like Praguian Wortphonologie than like Satzphonologie.
Besides sparseness, a second source of trouble for structural phonemics
is its segmentalist commitment (criticized in Scobbie & Stuart-Smith 2008).
It requires that a multiply associated feature be associated with exactly one
contrastive segment in its span. Segmentalism is implied by such concepts
as minimal pairs, the commutation test, and the view of a phonemic system
as an inventory of abstract contrastive segments. Structural phonemics has
no place for Harris’ long components, Firthian prosodies, or Goldsmith’s
autosegments, not even those versions of it that take distinctive features as
the basic units of phonology, such as Jakobson’s. OT phonology has
inherited segmentalism in its descriptive practice, and formalized it in
correspondence theory, but nothing about OT inherently requires it. OT is a
theory of constraint interaction, not a theory of representations. Lexical
representations differ from phonemic representations in that they record the
full cumulative effect of the stem-level and word-level phonological
computation, including any redundant features assigned in those two lexical
submodules, with one-to-many association of prosodies to segmental slots
where appropriate, while still excluding allophones introduced in the
postlexical module and phonetic implementation. This additional
information turns out to be important for phonological typology and
significant universals can be formulated over representations that
incorporate it.
An example will help make these points clear. Gravina (2014: 90–94)
describes Moloko (Central Chadic) as having a single underlying vowel /a/,
and a second vowel /ǝ/ which does not appear in underlying forms and is
predictably inserted where syllable structure requires.31 In addition, a word
may have one of two prosodies, palatalization and labialization (notated as
y, w), which color its vowels to yield six surface vowels altogether:

The prosodies spread leftward across a word, from suffix to stem and stem
to prefix, but they do not cross word boundaries (e.g. (2g)).32

There are three special cases. The vowel of a pre-pausal syllable is

neutralized to /a/, realized as [a], [ɛ], [ͻ] depending on the prosody, e.g.
(2g)). Labialized consonants and /w/ color adjacent [a] to [ͻ], [ǝ] to [u], and
[ɛ] to [œ] (the only source of [oe]), and /j/ colors adjacent [ǝ] to [i]. Finally,
a word-initial vowel is always [a], regardless of the prosody, e.g. (2h).33
So how many vowel phonemes does Moloko have? Just /a/, for the
epenthetic vowel is predictable, and the color prosodies are
suprasegmental? Or six, since there are in principle six-way surface
contrasts like CaCaC : CɛCɛC : CͻCͻC : CǝCǝC : CiCiC : CuCuC? The
puzzle is that the epenthetic vowel is non-contrastive, therefore allophonic,
but it is the bearer — in some cases the only bearer — of the phonologically
unpredictable and contrastive prosodies, therefore phonemic.
From the Stratal OT point of view, both answers are right. In this theory
of grammar, three levels of representation naturally emerge from the
phonological computation: (A) The STEM LEVEL is the innermost layer of
morphology and phonology. Its constraint system characterizes the form of
simple and derived stems, and derivatively via lexicon optimization the
form of the roots and affixes from which they are built, so that there is no
structural difference between underlying and derived representations at this
level. (B) The WORD LEVEL constraint system generates words, and
comprises constraints that apply in the span of a word, which in Moloko
includes the spread of the palatalization and labialization autosegments to
eligible slots. (C) The POSTLEXICAL PHONOLOGY applies to phrases and
sentences syntactically generated by combining words, generating PHONETIC
REPRESENTATIONS, the interface to speech production and perception.
At the stem level (whose representations I’ll put between braces)
Moloko has one vowel segment { a } and the autosegments { y } and { w },
formally the feature bundles [–consonantal, –back] and [–consonantal,
+back] (hence mutually incompatible and compatible only with vowels).
Epenthesis and the spread of the prosodies at the word level produce a
system of six vowels /a/, /ɛ/, /ͻ/, /ǝ/, /i/, /u/. Assuming that the structure-
insensitive local assimilation processes that apply across word boundaries
are postlexical, they add a seventh vowel [oe] to the repertoire. This is also
where prepausal neutralization of height takes effect.
Now we have to settle a terminological matter. What shall we call the
elements at the stem and word levels? The term “phoneme” is too handy to
give up, but it becomes ambiguous now that we have two significant
abstract levels of representation above the phonetic level, not just one as in
classical generative phonology. Should we use it for input segments (as in
Gravina 2014) or for lexical output segments, such as the six vowels of
Moloko (as Gravina 2014: 153 does for the similar Mbuko language)? To
avoid confusion I’ll refer to the underlying elements as M-PHONEMES
(morphophonemes) and to elements in lexical representations (the output of
the word phonology) as L-PHONEMES, or LEXICAL PHONEMES. Phonemes in the
traditional structuralist sense, not part of the proposed setup, are then S-
PHONEMES.
L-phonemes are at once more concrete and more abstract than s-
phonemes. As noted above, they are more concrete in that they incorporate
redundant features and prosodic properties that are assigned in the lexical
phonology. Consider syllable structure. Jakobson (1958) pointed out that all
languages have CV syllables.34 But in most languages syllabification is
entirely predictable from the segmental chain, which means that it is not s-
phonemic. At the s-phonemic level, such languages therefore have no
syllables, and in particular no CV syllables. More pointedly, some
languages have completely predictable and unremarkable CV syllable
structure that cannot be assigned at the s-phonemic level, even redundantly,
because its segmental substrate is not present there. For example, Kalam
words can contain long sequences of consonant phonemes, of which all but
the word-final one are automatically syllabified with an epenthetic [ɨ] (for
details, see Section 3.1 below), e.g. /pttt/ [ɸɨ´ɾɨ´ɾɨ´ɾ] ‘quivering’ (Pawley &
Bulmer 2011). In reality this language has a very strict syllable structure;
indeed it is a textbook illustration of syllabic universals and preferences, but
it violates almost all of them at the s-phonemic level, and a fortiori at the m-
phonemic level.
In the face of such examples, one might consider framing syllabic
typology at the phonetic level instead, where the regular syllable structure
of a language such as Kalam is patent. But this is not a promising solution
in general, for syllables are in general not phonetically characterizable.
They are only definable over sequences of discrete phonological segments
(see Section 2.2 below).
Similarly, if stress falls predictably within words, it has no place in
redundancy-free phonemic representations unless it plays a demarcative
role at the level of the sentence. For example, if secondary stresses fall
predictably on alternating syllables, they are absent from phonemic
representations even if they play a role in morphophonology and selection
of affix allomorphs. A consequence is that the theory and typology of stress
(Hyde 2002, Alber 2005, Kager 2007, and many others) simply cannot be
defined on s-phonemic or m-phonemic representations.
Conversely, l-phonemes are more abstract than s-phonemes in other
respects. They do not register postlexical sentence-level processes, even
when these neutralize contrasts or introduce new derived ones. In particular,
the lexical form of a word is not necessarily the same as its isolation form,
as we have just seen in Moloko. A better-known example is French, where
final consonants, present in the output of the word level since they surface
in liaison contexts, are dropped prepausally. Postlexical operations can
mask word phonology with respect to syllable structure, lexical stress, and
tone, to the point of altering its typological character. Ancient Greek is
lexically a tone-to-stress system, while the post lexical component is a
stress-to-tone system (Blumenfeld 2004). In preclassical Sanskrit, accent is
culminative at the word level: a word has one and only one pitch accent.
Postlexically, as a result of deaccentuation and glide formation across word
boundaries, a word may be unaccented or have more than one accent. It is
the culminative accentuation in the lexical phonology that is the
phonologically important property because it is driven by constraints that
lie at the core of the accentual phonology. Its loss in the postlexical
phonology is the extraneous result of various processes that have no
intrinsic connection to each other.
How do we identify the lexical representations of a language? The
bottom line is that we have to work out its phonology and morphology. But
there are some diagnostic shortcuts. Let us say that a constraint is ACTIVE at
a given level if it ranked in such a way that it is visible in at least some
derivation, i.e., that the output would be different if it were removed
entirely. A constraint that is active at the word level has the following
properties: (A) Its domain includes the entire lexical word, not only stems.3
5 (B) It interacts transparently with all other word-level constraints. (C) It is
asymmetrically bled and fed by stem-level constraints, and hence can
render them opaque. (D) It can be rendered opaque by postlexical
operations, resulting in derived surface contrasts. (E) It is sensitive to word-
level morphology; when words are built recursively, it applies cyclically at
each stage, which can lead to derived surface contrasts. (F) It is not
sensitive to stem-level morphology. (G) Like all of lexical phonology, it
operates on binary feature values rather than gradient feature values, and its
phonological context is defined in terms of binary feature values. Naturally
not all these properties can be positively instantiated for every case, but
whichever ones can be checked in a language should yield mutually
consistent results.
The properties (A)-(G) are not stipulated arbitrarily. They are
consequences of the principles of Stratal OT (Kiparsky 2000, to appear,
Bermúdez-Otero 2012, 2015). (A) follows because the natural way of
restricting a process to the word domain is to restrict the constraints that
drive it to the word level (“level 2”). The interaction and non-interaction
patterns (B)-(F) are implied by the architecture of the theory and basic OT.
Binarity of feature specifications (G) is a property of the entire lexical
phonology. Enhancement by new features is possible because Stratal OT
has no principle of Structure Preservation, as originally adopted in Lexical
Phonology, but soon abandoned in the face of a barrage of counterevidence
(Harris 1987, 1990, MacMahon 1991, Borowsky 1993, Hall 1993,
Martínez-Gil 1993, Kim 2001, Roca 2005). Structure-preservation in the
lexical phonology was an attempt to reconstruct certain structuralist
assumptions, with tenuous empirical support and no connection to the rest
of the theory.36 Finally, OT phonology has no constraints that prohibit
syllabification. That is, building syllable structure does not incur
faithfulness violations.37 The empirical reason is that such faithfulness
constraints would expand the factorial typology to predict nonexisting
phonological systems, such as languages with arbitrary phonotactics. In
particular, it would defeat the derivation of Jakobson’s generalization that
CV syllables are universal (Prince & Smolensky 1993).

1.2 Contrastiveness and distinctiveness

The basic concept of classical phonemics is that of phonological contrast, or
opposition. A phoneme is defined as a class of non-contrasting sounds.
Contrastiveness has been understood in two different ways. In American
structuralism, it had to do with CONTRASTIVE DISTRIBUTION. Two sounds
were held to contrast in a given phonological environment if they are
neither in complementary distribution nor in free variation in that
environment, i.e., if the occurrence of either of them in a given context
neither excludes nor implies the occurrence of the other in that context
(Bloch 1953). Functional approaches (Trubetzkoy 1939, Martinet 1964)
equated contrast with phonological RELEVANCE, or DISTINCTIVENESS, the
potential of distinguishing utterances as revealed by minimal or near-
minimal pairs and the commutation test. In Martinet’s formulation (1964:
53), the function of “phonic elements of a language […] is distinctive or
oppositional when they contribute to the identification, at one point of the
spoken chain, of one sign as opposed to all the other signs which could have
figured at that point if the message has been a different one”. The semiotic
grounding of this view of the phoneme is apparent.
It has been clear for some time that the distributional and semiotic
concepts of contrast don’t converge. Two empirical insights led to this
conclusion. One is that sounds which contrast in the distributional sense are
sometimes perceptually indistinguishable. Labov (1994) documented the
phenomenon of NEAR-MERGER, where speakers produce an instrumentally
measurable contrast that they cannot perceive, either in the speech of other
such speakers or when their own speech is played back to them. An
example is the source : sauce opposition in some U.S. dialects (Labov
1994, Ch. 12).38 Independently of this work, it was found that contextual
neutralization can be incomplete (Port & O’Dell 1985, Port & Crawford
1989, Dinnsen 1985, Piroth & Janker 2004, Kleber, John, & Harrington
2010). For example, some German speakers pronounce underlying voiced
and voiceless obstruents differently in word-final position, but not
differently enough to enable hearers to distinguish them reliably. Port and
his collaborators showed that German speakers can guess correctly whether
a given instance of German [bunt] corresponds to /bund/ ‘league’ or /bunt/
‘colorful’, with more than chance accuracy though far less well than a
normal distinctive minimal pair. Berber speakers consistently articulate
initial and final geminate voiceless stops longer than singletons, as in ttut
‘forget him’ : ttutt ‘forget her’ : tut ‘she hit’, even though this articulatory
difference has no audible effect (Ridouane 2007). These are non-distinctive
contrasts.
The second reason for separating contrastiveness and distinctiveness is
that distributionally non-contrastive, redundant features can contribute to
signaling phonemic distinctions (“the identification of signs” in Martinet’s
words) and are in that sense phonologically distinctive and relevant.
Russian front /i/ and back /ɨ/ are allophones respectively triggered by
palatalized and non-palatalized consonants. But Jakobson, Fant, & Halle
(1952) observed that the phonemic opposition between the consonants is
more effectively cued by the vowel allophones that they condition than by
the consonants themselves; in a noisy channel only the vocalic cues may be
perceived. Moreover, the backness distinction between [i] and [ɨ] is
phonologically relevant also in the sense that it actively participates in
phonological processes. In the lexical phonology, velars are palatalized
before /i/, but postlexically the direction of assimilation is reversed and
velars back a following /i/ to [ɨ] like all other [+back] consonants (Rubach
2000, Padgett 2010). Russian /ɨ/ is a non-contrastive but distinctive
segment.
One way of thinking of non-contrastive distinctive segments is that they
ENHANCE lexical feature contrasts by redundant features, beyond what
would result just from coarticulation (Stevens & Keyser 1989, Keyser &
Stevens 2006). The enhancements can appear on the contrastive segments
themselves or — what is more relevant here — on neighboring segments.
They can be more saliently distinctive than the structurally contrastive
segments they supplement, and historically more stable, often being
precursors of new contrasts that arise by phonologization. Just as the
realization of Russian /i/ as [i] or [ɨ] is a cue to the palatalization or
velarization of the preceding consonant, English vowel quantity and its
diphthongal reflexes (such as “Canadian Raising”) are cues to the voicing
distinction of coda consonants, Arabic vowel backing and lowering are cues
to the pharyngealization (“emphasis”) of consonants, etc. Such allophones
have been called QUASI-PHONEMES (Ebeling 1960, Korhonen 1969,
Liberman 1991, Janda 2003, Scobbie & Stuart-Smith 2008) or QUASI-
CONTRASTIVE (Ladd 2006).39
Stratal OT separates the distributional property of contrastiveness from
the perceptual/ functional property of distinctiveness in the following way.
Contrastiveness can be characterized at the input. Phonological derivations
cannot differentiate identical inputs: they can enhance, neutralize, and
displace contrasts, or translate prosodic or morphological differences into
segmental oppositions, but they cannot create distinctness from identity.
Featural contrastiveness is characterized by the dominant faithfulness
constraints at the stem level, the innermost layer of the lexicon. These
determine the availability of contrasts in lexical entries, and the extent to
which markedness constraints affect the shapes of morphemes and of
morpheme combinations at the stem level. Complexity and redundancy are
minimized at this level. The sparseness of the input representations and the
invariant underlying form of morphemes presumably serve to facilitate the
recognition, acquisition, and retrieval of the lexicon by making it easier to
recognize morphological relationships among words.
The distinctiveness of a contrast, on the other hand, is dependent on the
derivational level. At the word level, distinctiveness is characterized by the
available stem-level inputs (no “freedom of analysis” here anymore!) and
by the ranking of faithfulness constraints with respect to the markedness
constraints at that level. The promotion of markedness constraints can
optimize phonology for perception by enhancing lexical contrasts with
redundant features that help the hearer identify them, or neutralize stem-
level contrasts, or transpose them by a combination of contextual
enhancement and neutralization of the original source of the contrast.
Cyclic application or morphological conditioning can create new derived
feature contrasts which are not present in lexical representations, e.g.,
Belfast winter vs. printer (Wells 1982: 431, Borowsky 1993), British holey
[hɒʊ.li:] vs. holy [hʌʊ.li:] (Harris 1990, MacMahon 1991, Borowsky 1993).
The interface between the word-level and postlexical phonology is in
principle analogous. Reduction and coarticulation in phonetic
implementation minimizes effort in speech production at the cost of
distinctiveness. Cases of near contrasts and incomplete neutralization
probably arise here.

1.3 The argument from dispersion

We have seen that Moloko has one m-phonemic vowel and six l-phonemic
vowels. ALL very sparse vowel systems appear to have a larger,
typologically unremarkable symmetrical system of l-phonemic vowels due
to the spread of vocalic features from prosodies or from adjacent
consonants. The evidence will be laid out in Section 3. The l-phonemic
systems usually resemble the reconstructed vowel systems of the languages’
earlier stages or the current ones of related or neighboring languages, which
suggests that they are more stable than the m-phonemic systems. Although
the additional l-phonemic vowels are predictable, they are usually at least as
perceptually salient as the site of the contrastive features that trigger them,
if not more so, and can be as “psychologically real” to the speakers. In such
cases, the segmentalist extraction of redundancies arguably obscures the
way the language works.
Dispersion Theory holds that phonological systems maximize the
perceptual distinctiveness of contrasts. At the s-phonemic level, so-called
vertical vowel systems don’t have this property. Lindblom (1986, 1990) and
Flemming (1995) propose to reconcile vertical vowel systems with
Dispersion Theory by appealing to a conflicting pressure for minimally
complex articulations, which is best satisfied by central vowels such as /i/
and /e/ (see also Padgett & Ní Chiosáin 2001, Kaplan 2011 et al., among
others). Factorial typology then predicts that there should be vocalic
systems in which this markedness constraint dominates dispersion
(MINDIST) constraints, consisting of only “minimally complex” central
vowels. Such vowel inventories are however not attested either phonetically
or in the output of the lexical phonology of any language.40 In reality, all
languages exploit the major dimensions of the vowel space, including the
F2 dimension by distinguishing front and back vocalic l-phonemes
(including Kabardian and Marshallese, see Section 4 below). What is
special about so-called vertical vowel systems is that their peripheral
vowels are derived in the word phonology rather than underlying; they are
l-phonemes rather than m-phonemes.41
A more interesting class of apparent exceptions are certain Papua New
Guinea languages, whose high vowels have been phonemically analyzed as
underlying consonantal non-syllabic /y/ and /w/. I argue in Section 3.1 that
they are underlyingly indifferent with respect to syllabicity and appear in
both nuclear and marginal positions as required by the language’s stringent
syllable structure. When syllabification is taken into account, all these
languages have /i/ and /u/.
For lexical representations, we may therefore remove the if-clause from
Hyman’s (2008) Vocalic Universal #5, quoted in (3).

(3) A vowel system may be contrastive only for aperture only if its vowels
acquire vowel color from neighboring consonants.

We replace (3) with the categorical (4).

(4) All vowel systems have distinctive color and aperture.

This means that at the level of lexical representations there are no one-
dimensional vowel systems, whether vertical or horizontal. Minimal vowel
systems are triangular, making use of both the front/back dimension and the
high/ low dimension. To this we can now add the more specific substantive
generalization (5).

(5) All vowel systems have at least a low vowel and two non-low vowels.
One of the non-low vowels is a front unrounded high vowel, the other
is back.

An apparent counterexample to (4) and (5) is Qawasqar (a Fuegian

language), which UPSID cites as having the phonemes /ǝ, o, a/. But Clairis
(1977), UPSID’s source, represents them as /e, o, a/, and says that /e/ has
the allophones [i], [x] (devoicing?), [ɯ], [ǝ], [e], [æ], [a], of which [ǝ] is the
most frequent (hence UPSID’s /ǝ/). Clairis moreover says that [i] tends to
be stable in certain words, and proposes that /i/ is a distinct phoneme in
those words. If this is correct, Qawasqar has a non-minimal four-vowel
system /i, e, o, a/ of a common type (as in Campa, Klamath, Malagasy,
Mazatec, Nahuatl, Tacana).

1.4 The argument from symmetry and from the

loss of generalization
When phonemics is integrated into a derivational morphophonemics, the
divergence between contrastiveness and distinctiveness gives rise to a well-
known formal problem. One of the first mentions of it in the literature
occurs in Bloomfield’s description of Menomini height assimilation:
If postconsonantal y, w, or any one of the high vowels, i, ı¯, u, u¯, follows anywhere in
the word, the vowels e¯ and o¯ are raised to ı ¯ and u¯ […] Since u¯ occurs only in this
alternation, it is not a full phoneme. (Bloomfield 1939: ¶35).

So, e¯ → ı ¯ in (6a) and o¯ → u¯ in (6b) are parallel, but the former is

morphophonemic and the latter allophonic, since u¯ is in complementary
distribution with o¯.

(6) a. /maye¯cˇek-waɁ/ → mayı¯cˇekwaɁ ‘that which they eat’

b. /a¯tɛɁhno¯hk-uwɛw/ → a¯tɛɁnu¯hkuwɛw ‘he tells him a sacred
story’.

Bloomfield called Menomini u¯ a SEMI-PHONEME; other writers have used

the term QUASI-PHONEME. We can now define these as l-phonemes which are
not m-phonemes.
The phonemic level must mark the contrastive distinction between ı¯
and e¯ and exclude the redundant, predictable one between u¯ and o¯. But
if we want the derivation from morphophonemics to phonetics to pass
through a phonemic level, we must split raising into a morphophonemic
rule e¯ → ı ¯ (solid arrow in (7)) and an allophonic rule o¯ → u ¯ (dashed
arrow), although they are obviously the same process.

(7)
A formally identical English case was noted by Bloch (1941).
Although this duplication problem did not attract attention at the time, it
led to a crisis in phonemic theory when it was raised by Halle (1959: 22)
and Chomsky (1964) as an objection to any intermediate phonemic level
(Anderson 2000). Crucially, THIS PROBLEM DOES NOT ARISE IN STRATAL OT. At
any given level, the available contrasts are defined by the ranking of the
relevant faithfulness and markedness constraints. Schematically, the
asymmetric underlying vowel system of Menomini comes from a constraint
— call it *u¯ — that dominates IDENT(High) in the stem phonology, thereby
suppressing the height contrast between u¯ and o¯. In the word phonology,
both *u¯ and IDENT(High) are dominated by height assimilation, whose
activation brings in the new l-phoneme. In this way the grammar formally
characterizes both the neutralization of the contrast between u¯ and o¯ in
the stem phonology (which makes the height specification irrelevant in
input representations), and the derived distinction between them in the word
phonology.
The activation of context-sensitive markedness constraints not only
enhances feature distinctions and maximizes dispersion, but creates more
symmetric inventories, and maximizes feature economy in the sense of
Clements (2001, 2003). Like dispersion, symmetry and feature economy are
tendencies of phonological systems, not absolute requirements, but they are
quantifiable and statistically verifiable, as shown for feature economy by
Clements (2003), and legitimate criteria for adjudicating between different
phonemic solutions. I venture the following conjectures:

(8) a. L-phoneme inventories are never less symmetrical than m-

phoneme inventories.
b. L-phoneme inventories are never less dispersed than m-phoneme
inventories.

Jimi (another Central Chadic language) makes an instructive

comparison with Moloko. It has three basic underlying vowels, {i}, {ǝ},
{a}, plus long {i:} and {a:} (Gravina 2010: 134–139). Unlike Moloko, it
has no general vowel harmony, and its vowels are not normally affected by
adjacent consonants, e.g. (9a,b,c), except that {ǝ} becomes [i], [u] after {j},
{w} or next to {Ɂw}, {Ɂj}, as in (9d-g), and {a} becomes [e] after {rj}, {lj},
as in (9h,i).

Since {i} is an independent phoneme, whereas [u] and [e] are allophones
occurring only in the contexts just mentioned, a split derivation would again
be required to reconstruct an s-phonemic level. As in (7), the processes
marked by dashed lines introduce new l-phonemes, increasing both
symmetry and dispersion.

(10)

Unlike Menomini, Jimi does not achieve perfect symmetry. {a} is never
raised to {o} due to a gap in the consonant inventory. Jimi has no labialized
alveolars, so that the process that produces [e] after {rj}, {lj} has no
corresponding labial triggers *{rw}, *{lw}.

1.5 The argument from diachrony: Sound

change, analogy, borrowing
The concept of a quasi-phoneme originated in the literature on sound
change and phonologization, as part of the effort to solve the problem why
allophonic distinctions sometimes remain unaffected when their
conditioning environments disappear, and become phonemic instead
(Ebeling 1960, Korhonen 1969, Liberman 1991, Janda 2003). The idea was
that this happens when they for some reason have acquired “quasi-
phonemic” status before the environment changes — in Stratal OT terms,
when they have become l-phonemes by the criteria (A)-(G) in Section 1.1.
The allophones are already distinctive, but not yet contrastive. If sound
changes are initiated postlexically (as must be assumed for many
independent reasons) they do not affect lexical representations (Bermúdez-
Otero 2015, Kiparsky 2015). This follows from the feed-forward relation
between the strata.
When a phonological opposition first becomes distinctive at the word
level, there is no contrastive input from the stem level to realize it. If it is
instantiated at all, it is not by inputs from the stem level, but by types of
lexemes that don’t go through the stem phonology. This is where marginal
phonemes can enter the word phonology. These are of three types: (A) non-
lexical categories, including function words and interjections; (B) loan
vocabulary, in so far as it is unassimilated, as is independently diagnosable
by the unavailability of stem-level morphology for it; and (C) word-level
derivatives, where opacity can arise through cyclic application of
phonology (“analogy”). And these are indeed exactly the contexts where
new distinctions in a language first appear. The Menomini quasi-phoneme
[u:] appears out of its triggering context in interjections, loanwords, and for
some innovating speakers by analogical generalization from raising
contexts (Bloomfield 1962: ¶1.16). In English, phonological properties of
simple words are almost always retained in their word-level derivatives;
e.g., the lengthened vocalic nucleus of monosyllabic words is inherited by
their polysyllabic word-level derivatives (cart, carter vs. Carter). Contrasts
such as bible (short diphthong) vs. libel (long diphthong) have emerged in
some varieties of English (Scobbie & Stuart-Smith 2008).42
I conjecture that all phonemes arise as l-phonemes. They remain when a
sound change makes their conditioning environment opaque (whereas
postlexical allophones disappear), and become contrastive (“phonologized”,
in Praguian terms) when their distribution can no longer be predicted. This
is essentially equivalent to Hyman’s (1976, 2008) proposal of an automatic
→ extrinsic → phonemic trajectory of phonological alternations.
Correspondingly, we may suppose that all mergers pass through a near-
contrast stage. The phenomenon of DISPLACED CONTRAST is a combination of
merger and phonologization.
Adapting loanwords involves not only approximating their
pronunciation with native phonetic resources, but rendering the donor
language’s contrasts as best you can. In this task l-phonemes can be seen in
action. Hsieh, Kenstowicz, & Mou (2009) (see also Kenstowicz & Louriz
2009) show that Mandarin Chinese borrowers, using “reverse engineering”,
privilege the salient allophonic (in my terms l-phonemic) vowel distinction
over the less salient phonemic distinction between the nasal codas. With
only one s-phonemic low vowel /a/, pronounced [æ] before /n/ and [ɑ]
before /ŋ/, Mandarin cannot express the four-way English distinction ran :
rang : Ron : wrong. It chooses to map it into the available two-way
distinction as in (11a), giving up the /n/ : /ŋ/ contrast in order to maintain
the [æ] : [ɑ] contrast.

(11)

Thus Dan : Don would be rendered as [dæn] : [dɑŋ], while ban : bang
would both be [bæn]. By the criteria laid out in Section 1.1, /æ/ and /ɑ/ are
distinct l-phonemes in Mandarin, present in lexical representations just as
/n/ and /ŋ/ are. The loan phonology privileges the front/back feature on
vowels over the corresponding consonantal feature on nasal codas.
Presumably the vocalic distinction is perceptually more salient than the
consonantal distinction, as in the case of the Russian. Postlexical allophones
are not used for such “reverse engineering” because they are not
represented in the lexical phonology and unavailable for manipulation by
speakers. For example, English borrowings from Chinese don’t render tones
by consonant voicing, although this might well produce approximations of
at least some Chinese tonal contrasts.

1.6 The argument from poetic convention and

language games
The artistic use of language, not just versification but also text-setting and
language games, involves superimposing a second layer of constraints on
already structured linguistic representations. In so far as it relates to
phonology, the relevant level is obviously neither underlying
representations and phonetics but somewhere in between.43 The identity is
not between s-phonemes, but l-phonemes provide the right representations,
at least in many interesting cases. Here briefly are two telling examples.
The classical rules of French versification, which remained normative
into the twentieth century, stipulate that consonants that are deleted in
word-final position count for purposes of rhyme, except that homorganic
final voiced and voiceless obstruents are treated as equivalent. For example,
long and tronc rhyme, but neither of them rhymes with rond or pont, which
however rhyme with each other; none of them rhyme with son. Phonetically
all five words end the same way, and traditional phonemics would reflect
that: /tʁõ/ [tʁõ], /lõ/ [lõ], /mõ/ [mõ], /ʁõ/ [ʁõ], /sõ/ [sõ].
Morphophonologically they all end differently: {tronk}, {long}, {pont},
{rond}, {son} — the consonant shows up before suffixes, as in tronquer,
longue, ponter, ronde, sonner. It is only in lexical representations that the
two rhyming pairs match correctly: because of final devoicing at the word
level, long and tronc both end in /-k/, and dont and rond both end in /-t/.
The evidence that final devoicing of stops takes effect at the word level
whereas final deletion is postlexical is that, in the classical liaison system
(now as old-fashioned as the rhyming convention that reflects it) final
voiced stops appear in devoiced form before a following vocalic word in
close contact, e.g. long hiver [lͻ˜.ki.vɛʁ] ‘long winter’, grand homme
[grã.tͻm] ‘great man’. Therefore they must enter the postlexical phonology
with the final consonant present but in devoiced form. In sum, traditional
French versification conventions crucially refer to the lexical representation
that is computed by the word phonology and forms the input to the sentence
phonology:

(12) – Underlying (morphophonemic representation): {tʁonk}, {long},

{pont}, {ʁond}, {son}
– Lexical representation: /tʁõk/, /lõk/, /põt/, /ʁõt/, /son/
– Phonetic (and structuralist phonemic) representation: [tʁõ], [lõ],
[põ], [ʁõ], [sõ]
That the rhyming conventions could outlive the sixteenth century
pronunciation they reflect for centuries is presumably due to the fact that
the living morphophonology of the language kept them intelligible.
Guimarães & Nevins (2013) used invented language games to probe
whether Brazilian Portuguese nasal vowels are synchronically derived from
vowel+nasal sequences, or underlying. Their experiments showed that the
four nonlow nasal vowels [ı˜, e˜, õ, u˜] are derived, whereas the low nasal
vowel [ɐ˜] is an underlying segment. What accounts for this phonological
difference? The phonetic difference between the plain and nasalized low
vowel that G&N appeal to can hardly justify it. But there are some
phonological processes, not mentioned by G&N, which show the low nasal
vowel behaving differently from the non-low nasal vowels in a way which
suggests that /ã/, unlike the other nasal vowels, is an l-phoneme. Stems in {-
an-} regularly contract with a following ending {-a}, e.g. sã [sɐ˜] ‘sane’
(fem.), from /san-a/ (cf. masc. são [sɐ˜w˜ ], from /san-u/). {-Vn-}stems
where V is some other vowel than /a/ keep the stem form under these
circumstances, e.g. dona [ˈdonɐ] ‘lady’, from /don-a/ (masc. don [dõ], from
/don-u/), or in exceptional cases delete the nasal, e.g. boa [boɐ] ‘good’
/bon-a/ (masc. bom [bõ] /bon-u/). This distribution falls out if the low nasal
vowel is formed at the stem level, whereas the other nasal vowels are
formed postlexically.

(13) a. {san-} → /sɐ˜/ → /sɐ˜-a/ → [sɐ˜]

b. {don-} → /don-a/ → [ˈdonɐ]

2 Syllabification

2.1 Arrernte
Arrernte (an Arandic language of Australia) has been claimed to have only
VC(C) syllables (Breen & Pensalfini 1999 [B&P], Pensalfini 1998, Tabain
et al. 2004). Evans & Levinson (2009: 434) cite B&P’s work as “a clear
demonstration that Arrernte organizes its syllables around a VC(C)
structure and does not permit consonantal onsets […] An initially plausible
pattern turns out not to be universal after all, once the range of induction is
sufficiently extended”. VC(C) is indeed the most marked syllable type since
it violates ONSET, NOCODA, and *COMPLEX, and contradicts the following
generalizations (and Jakobson’s CV universal):

(14) a. All languages have syllables with onsets.

b. All languages have open syllables.
c. All languages that have syllables with complex codas have
syllables with simple codas.
d. All languages have syllables with simple onsets.

B&P’s claim that all Arrernte syllables lack onsets is about underlying
representations. About 25% of words as actually pronounced begin with a
consonant. Their analysis posits that they have an underlying initial /e-/,
which is then deleted. Their claim that all Arrernte syllables are closed is
likewise about UNDERLYING representations. According to Henderson &
Dobson (1994: 23) “nearly all Arrernte phonological words end in a central
vowel, though this vowel need not be pronounced, and is often absent in
sandhi when another vowel follows”. H&D’s transcription implies a
phonemicization that is consistent with all four universals in (14). It has no
underlying unpronounced initial /e-/, and posits final /-e/ where it is
pronounced. (15) shows B&P’s analysis in the first column, and H&D’s in
the second, with the actual pronunciation in the third.

So Arrernte’s syllable structure violates (14) at most in underlying

representations, and then only under B&P’s analysis which posits
underlying forms with initial vowels and final consonants that undergo
aphaeresis and paragoge, not under H&D’s analysis.44 Let us consider the
evidence for B&P’s analysis.
B&P’s first argument for their morphophonological analysis is based on
the generalization that the first syllable of a word is stressed, except if it is
onsetless, in which case the second syllable is stressed. For if all words are
assumed to begin with underlying V-, as in the first column of (15), stress
can just be assigned to the second underlying syllable.45 But an equally
simple stress rule exploits the well-documented weakness of onsetless
syllables, which causes them to be unstressable in some languages (Burzio
1994: 158, Downing 1998, Ryan 2014). Unambiguous evidence for this
treatment of the Arrernte-type pattern comes from Iowa-Ota stress. In this
language words are stressed on the first syllable, except if it has no onset, in
which case they are stressed on the second syllable (Topintzi 2010: 58 ff.,
with other examples of this pattern).

Iowa-Ota has the same stress pattern as Arrernte, but it cannot be reduced to
second-syllable stress by positing deleted initial vowels. Finnish secondary
stress exhibits the same pattern. Four-syllable words normally get a stress
on the third syllable, except if it is onsetless, e.g. á.te.ri.a ‘meal’,
kómp.pa.ni.a ‘(military) company’ (Karvonen 2005). KiKerewe
demonstrates the prosodic defectiveness of onsetless syllables in several
different ways: they are light, tonally defective, and do not induce
compensatory lengthening when desyllabified (Odden 1995). Regardless of
how the unstressability of onsetless syllables is modeled,46 it undermines
the argument for abstract /V-/ in Arrernte.
The second argument adduced by B&P for VC(C) syllabification in
Arrernte is based on the plural/reciprocal suffix. After a stem with an odd
number of syllables, the suffix is -err or -errirr. After a stem with an even
number of syllables, the suffix is -irr. Stems of more than one syllable can
also have the optional allomorph -ewarr. The syllable count comes out right
if an initial vowel is posited in words that begin with consonants, so that
(17a) begins in /et ̪-/ and (17c) begins in /ekwern/.47
The obvious alternative is that the allomorphy is stress-conditioned: the
allomorph -err- must head a foot, the allomorphs -irr and -ewarr cannot.48
B&P’s third argument is that the reduplication pattern of the
frequentative indicates VC(C) syllabification.

For B&P, the frequentative suffix consists of a disyllabic foot, the first
syllable pre-specified as -ep-, the second a copy of the final VC(C) syllable
of the root. But this is a weak argument because prosodic morphology
normally does not involve copying prosodic constituents of the base.
Rather, affixes are prosodic templates (defined by constraints) that get their
unspecified segmental content from the base (McCarthy & Prince 1986). If
the syllable structure of the reduplicant is fixed by the reduplication
morpheme itself, then it can’t tell us anything about the syllabification of
the base. The argument is further undermined by Pensalfini’s (1998)
observation that the same type of reduplication exists in Jingulu, which
uncontroversially has CV syllabification, and therefore in any case requires
some such alternative analysis. A straightforward formulation consistent
with the theory of Prosodic Morphology is that the suffix is /-epVC/, with
VC filled by the closest part of the stem melody, e.g. /empwarr/ → empwarr-
epVC → empwarr-eparr.
The fourth argument, from the play language Rabbit Talk, is especially
intriguing:49
It looks like the initial syllable of the word, VC(C) in B&P’s analysis, is
moved to the end. But an unproblematic alternative is that the word rhyme
(the portion of the word that includes the stressed vowel and everything that
follows it, boldfaced in (19)) is flipped with the residue (prosodic
circumscription), viz. (amp) (áŋkem) → aŋkem-amp.50
There is substantial positive evidence that Arrernte words do exhibit the
universal preference for CV. One indication comes from the rendering of
English loanwords. They insert a vowel after a final consonant, not before
an initial consonant as the VC(C) syllable canon would predict.

Arrernte songs categorically prefer CV. “In the Arandic [song] tradition,
quite generally, the consonant of a line-final suffix [...] is transferred to the
beginning of the line following, so that each line begins with a consonant,
even if the actual Arandic word heading the line is vowel initial [...]” (Hale
1984). Turpin (2012) moreover observes: “All sung syllables have an onset.
[...] creating a poetic line involves either deleting the line-initial vowel ([ɐ
ˈɳtǝpǝ] → [ˈɳtǝpǝ] ‘pigeon’) or inserting a consonant ([ɐˈlǝmǝ] →
[ˈwɐlǝmǝ] ‘stomach’).”
Postlexical syllabification shows CV preference as well. At the sentence
level, ONSET and NOCODA are maximized:

Moreover, epenthesis is obligatory in phrase-final position and e- never

occurs phrase-initially. This led Pensalfini (1998) to partially retract the
B&P analysis at least for the phrasal level. However the arguments
presented here seem to me to invalidate it for the word level too, and for
any level of representation. Not only do Arrernte’s lexical and postlexical
phonological processes actively FAVOR CV syllables, but the claim that its
output inventory of syllable types is derived from underlying
representations that have exclusively VC(C) syllables is difficult to sustain.5
1

2.2 The universality of syllables

Hyman (1985, 2011, 2015) argues that Gokana has no phonological rules or
constraints that must refer to syllable structure: its phonotactics and
phonological alternations can all be stated in terms of segments and moras.
Not all of them have to be stated that way, and the assumption that Gokana
has syllables would entail no complications, lost generalizations, or
violations of typological expectations, such as anti-syllabic phonotactics.
There might be cases where the syllabification is indeterminate, as Hyman
notes, in which case the theory would dictate the least marked syllable
structure.
Hyman’s point, then, is not that Gokana cannot have syllables but that
nothing is gained by positing syllable structure in it; all the action is in the
moras. Gokana is not the only language where the syllable plays second
fiddle to the mora: Japanese is another one, and the existence of syllables in
Japanese has likewise been doubted (Labrune 2012, but see Kubozono
2003, Kawahara 2012).
As Hyman is careful to note, a language can have syllables even if they
play no role in its phonology. Not every phonological property of a
language need be involved in its phonotactics and phonological
alternations. It would be no more surprising for a language to lack
phonological processes involving syllable structure than for it to lack
phonological processes involving vowel height.
Like feet, syllables meet more scepticism and outright rejection than
larger prosodic groupings such as phrases, both from linguists (Steriade
1999) and from phoneticians (Kohler 1966, Ohala 1990, Ohala &
Kawasaki-Fukumori 1997). The main reason is probably that in many
languages they are not phonetically demarcated, at least in casual speech,
nor directly identifiable by some crosslinguistically invariant property. (The
same is true of segments, as Hyman reminds me.) The syllable is not the
sole domain of coarticulation, nor a “puff of air”, nor a “chest pulse”.52 As
Anderson (1982: 546) states, “the facts of acoustics and of articulatory co-
articulation make it quite impossible to segment and identify the speech
stream directly in terms of such units”. Rather, syllables, their constituents,
and the boundaries between them are manifested in multiple ways by their
phonological effects, and it is the explanatory connections that can be based
on this construct that justify it.
It is true that many generalizations involving syllables can be restated
without reference to syllables, and doubters have sought to show that such
restatements are in some cases as insightful or even more so. Certainly the
early arguments for syllable structure in generative phonology, first as a
feature (Chomsky & Halle 1968), then as a boundary marker (Vennemann
1972), and finally a constituent (Kahn 1976), were not conclusive. For
example, the argument that syllable structure eliminates disjunctive
environments such as “before a consonant or word boundary” is not
sufficient because the generalization could be be equally well and possibly
better addressed by sophisticated competitors to syllabic formulations.
Some syllable effects can be expressed in terms of moras (Hyman 1985) or
feet.
Even in languages such as Chinese, where syllables are well demarcated
in speech, restrictions on codas and onsets could be reformulated in terms
of postvocalic and prevocalic position.53
The most persuasive argument for the syllable as a constituent is that it
ties together a number of phenomena that competing syllable-less theories
can only deal with separately in a disconnected way (Vaux & Wolfe 2009).
1. Distinctive syllabification, e.g. Finnish hau.is.sa ‘in pikes’ vs.
ha.uis.sa ‘in searches’, English ˌant.ˈac.id vs. fan.ˈtas.tic, respectively
with unaspirated coda t and aspirated onset t. These are predictable
from the morphology, but syllabification can also be contrastive, as
marginally in Sanskrit, where some words have inherent hiatus, e.g.
trisyllabic ti.ta.u- ‘sieve’ (vs. disyllabic da.dau ‘gave’). These pairs do
not differ as to mora count.
2. Differences between onset and rhyme position, e.g. British aspiration
or light vs. dark l, and ambisyllabicity in the corresponding American
cases (Gussenhoven 1986).
 3. Stressability. Polish ˈkrvi is monosyllabic. Polish has penult word
stress, so if the word were disyllabic, we’d expect *ˈkrˌvi. Serbo-
Croatian krˌˈvi is disyllabic; in Serbo-Croatian, syllabic r can bear
stress and pitch accent like vowels do.
4. Syllable counting, e.g., alternating stress, allomorphy dependent on
even/ odd parity.
5. Sonority sequencing: syllable boundaries as sonority troughs, nuclei as
sonority peaks.
6. Restrictions on syllable size.
7. Differences between open and closed syllables (not reducible to
syllable weight because heavy syllables can be open and light syllables
can be closed).
8. Prosodic morphology: reduplication, infixation, truncation.
9. Language use: text-setting in song and chanting, language games,
speech errors.

Obviously not all of these manifestations of syllable structure will exist in a

given language. But when they do, they converge.54
Suppose that Gokana has none of these things — no distinctive or
contrastive syllabification, no audible syllable boundaries, no processes or
constraints that distinguish onsets from codas, or count syllables, or are
sensitive to syllable boundaries, or constrain the size of syllables, no
evidence for syllables in versification or slogan-chanting, at least nothing
that could not also be stated in terms of moras. Then is Gokana a language
without syllables, or does it just not wear its syllable structure on its sleeve?
In the absence of instrumental evidence, which might one day help decide
the question (Fn. 25), we have to rely on phonological arguments. Let’s
consider the pros and cons.
Hyman’s idea that Gokana does not have syllables, only moras, is based
on an ex silentio argument advanced in a more general context by Clements
(2001: 72). The idea is that languages have syllable structure only when it is
“activated”, which is to say when it is “needed in order to express
generalizations about the phonological system”, such as phonotactic
restrictions or phonological alternations (Hyman 2011). But this actually
leads to an argument for syllables in Gokana: they explain many of its
phonological properties that would otherwise remain arbitrary. Gokana
obeys all cross-linguistic generalizations about syllable structure, including
the ones in (14) and Jakobson’s universal. Also, Gokana syllables are
maximally bimoraic, and *COMPLEX is unviolated. Vowel-initial words get a
predictable glottal stop, indicating that ONSET is active. It tolerates codas,
but so do most languages. Gokana syllable structure is about as vanilla as
can be. What makes it look odd is the frequency of hiatus, reflected in its
long sequences of vowels, including identical vowels, apparently without
rearticulation or other clues to syllable boundaries. Such examples don’t per
se argue against syllables, because they also occur in languages that
unquestionable have syllables. For example, Finnish has hiatus, e.g. vaa’an
[va:.an] ‘scale’ (gen.) vs. vaan ‘but, however’, disyllabic häät [hæ.æt]
‘carbon monoxides’ vs. monosyllabic häät [hæ:t] ‘wedding(s)’ (certainly
contrastive, but probably not distinctive in normal speech), and long
sequences of vowels such as hääyöaie [hæ:.yø.ai.e] ‘wedding night
intention’, but that doesn’t warrant the conclusion that Finnish has no
syllable structure — on the contrary, syllables are hugely important in
Finnish phonology and allomorphy.
There are also a number of Gokana-specific facts that syllables help
make sense of:

(22) a. Roots have the shapes CV, CVV, CVC, CVCV, but not *CVVV.
Analysis: they are minimally a syllable and maximally a bimoraic
foot, satisfying ONSET.
b. Derivational suffixes can have the shape -V or -CV. Analysis:
they are minimal (light) syllables.
c. Prosodic stems may be of the form CV, CVC, CVV, CVCV,
CVVCV, CVVCVV, CVVVV, but not *CVVVCV, *CVCVVV.
They are maximally disyllabic (disyllabic trochees), as Hyman
himself notes. Since Gokana syllables are maximally bimoraic,
the restrictions follow.55
d. Gokana has CV-reduplication. Analysis: the Gokana reduplicant
is a minimal (light) syllable, a very common type of reduplication
as predicted by Prosodic Morphology (McCarthy & Prince 1986,
1993).

Such constraints are obviously helpful to hearers and learners in parsing the
morphological structure of word. These data undermine even the weaker
claim that Gokana CAN be analyzed adequately without syllables.
 According to Hyman (2008), “imposing an arbitrary syllabification [on
the word adds nothing to our understanding of Gokana”. I find this
argument unconvincing for two reasons. First, the syllabification would not
be arbitrary, for it would have to be compatible with the language’s
constraints, including the ones in (22). Secondly, it seems too much to ask
that the syllabification of EVERY Gokana word should add something to our
understanding of the whole LANGUAGE. We don’t ask that of any other
aspect of the phonological analysis of words. Rather, the analysis of the
entire language has to be compatible with all its words and yield as many
explanatory dividends as possible, within the language and across
languages. A theory lives by the totality of its consequences.
A theoretical argument for the same conclusion follows from basic
assumptions of OT. A constraint can be defeated only by a more highly
ranked constraint. Prohibiting syllabification would require constraints that
defeat syllable structure assignment. But syllabification per se violates
neither faithfulness constraints or markedness constraints (although specific
marked syllable structures violate such constraints as ONSET and *CODA,
which can be ranked to yield the familiar factorial typology, and
RESYLLABIFICATION does constitute a faithfulness violation). Such
constraints are unmotivated and their adoption would expand the factorial
typology in undesirable ways. For example, a language without syllables
would not violate any constraints such as ONSET, *CODA, and *COMPLEX,
and consequently not be subject to phonotactic constraints captured by
those constraints.
This seems to me enough reason to reject the claim that Gokana has no
syllables. Even if the symptoms of syllabicity in (22) are discounted, the
very fact that the language is syllabifiable in conformity with typologically
well-established constraints and preferences would be incomprehensible if
it did not in fact have syllables. All in all Gokana speaks for rather than
against the universality of syllables and CV syllables in particular, just as
Arrernte does.
Japanese is a broadly similar case. It has the same kind of funny vowel
sequences as Gokana, e.g., Bloch’s example oooóóo ‘let’s cover the tail’,
and perhaps no syllable-conditioned phonological processes. Yet there is
evidence for one-mora and two-mora syllables (McCawley 1968, Kubozono
1999, 2003, Itô & Mester 2003), possibly three-mora syllables, though
Kubozono argues that these are divided into two syllables as /CV.VN/.
 Labrune disputes the existence of syllables in Japanese, citing the three-
way contrast in (23):

(23) a. an.i (three moras) /aNi/ [ˈãNi] ‘ease’

b. ani (two moras) /ani/ [ˈani] ‘older brother’
c. anni (three moras) /aNni/ [ˈãnni] ‘implicitly’

The contrast cited by Labrune has a natural syllabic interpretation, however:

an.i : a.ni : an.ni, with coda nasals counting as moras and causing
nasalization of the tautosyllabic vowel.56 Yet Labrune cites it as evidence
against syllables in Japanese, on the grounds that (23a) violates the
generalization that a closed syllable cannot be followed by a syllable that
starts with a vowel. One might as well cite it as evidence against moras in
Japanese, since it also violates the generalization that a consonant followed
by a vowel is not moraic. Of course both generalizations are correct for the
initial syllabification of unsyllabified segments, where onsets are
universally favored over codas, but in morphologically derived words the
syllabification of the base may prevail if faithfulness outranks these
markedness constraints.57 The moraic nasal in an.i ‘ease’, an.itsu ‘idleness’
is inherited from the root an from which they are formed (Itô & Mester
2015a: 296). This root can also be seen in words like anraku ‘ease’, anshin
‘peace of mind’, anga ‘quiet rest’, heian ‘peace’. Therefore the data in (23)
don’t confute the syllabic analysis. In fact, they support it. For Labrune’s
alternative that moraic nasals are “special moras” which are “prosodic units
in their own right which possess greater autonomy than syllabic codas” has
nothing to say about this morphological connection. Labrune defines the
class of “special moras” by enumeration as moraic nasals, second parts of
long vowels, i after a vowel, first parts of obstruent geminates, which are
exactly the weak (non-head) moras of heavy syllables. The fact that these
fall readily under a unifying syllabic characterization constitutes further
evidence for the syllable (Itô & Mester 2015b: 371).
Phonological descriptions of Japanese have found syllables useful to
represent surface contrasts such as kóo.o ‘likes and dislikes’ vs. ko.oo
‘response’, and variation in the pronuciation of vowel sequences, e.g. o.ó.i
(three syllables) and óo.i (two syllables) ‘is much’, ‘are many’, or o.ó.u
(three syllables) and oo.u ‘covers’ (Martin 1975: 17). Syllable structure also
plays a role in rules of accentuation and word formation, clearly productive
since many of them involve the nativization of borrowed words (Kubozono
1999: 53). On balance, then, the universality of the syllable seems well
supported.
The question whether all languages have syllables has a telling parallel
in the syntactic question whether all languages are configurational. The
claim that Gokana and Japanese lack syllables because they lack
diagnostics for syllables is analogous to the claim that a language lacks a
VP constituent because it has no diagnostics for it. In the absence of
positive evidence for or against a VP, should we posit a minimal flat clause
structure (Mohanan 1982, Austin & Bresnan 1996), or a VP?58 The latter
seems a better bet in view of previously overlooked evidence for a VP
constituent, in some cases, interestingly, involving phonological phrasing
(for Tamil, see Nagarajan 1995). Contemporary syntax leans towards the
universality of VP.

3 Vowel systems

3.1 Kalam
The generalization that all languages have an /i/-type vowel is contradicted
by analyses of some Papua New Guinean and Chadic languages. In these
languages syllabic and non-syllabic semivowels (high vowels and glides)
are in complementary distribution, but phonemicized as underlying /y/ and
/w/ (Fast 1953, Laycock 1965, Barreteau 1988, Comrie 1991, Pawley &
Bulmer 2011, Smith 1999). This analysis reduces the phonemic vowel
inventory to /e/, /a/, /o/, or just to /a/ (in some of the languages with an
additional epenthetic /ǝ/ or /ɨ/).59 Since i, y and u, w have the same
segmental feature content, differing only in syllabicity,60 this analysis
amounts to specifying syllabicity in the phonemic inventory, despite its
predictability, and despite the complementary distribution between the high
vowels and glides. I believe that the need to specify semivowels as
underlyingly non-syllabic in these languages is an artifact of segmentalist
phonemics, and present a Stratal OT analysis in which all underlying
segments are indifferent as to syllabicity, and the semivowels are derived
from underspecified {I}, {U}. This yields exactly the same output as
positing underlying /y/, /w/ or /i/, /u/, because the actual realizations are
determined by the languages’ strict syllable structure. I demonstrate this
with a reconsideration of the exemplary analysis of Kalam by Blevins &
Pawley (2010) and Pawley & Bulmer (2011).
In addition to the semivowels at issue, Kalam has the vowels /e/, /a/, /o/,
plus an epenthetic vowel which is inserted predictably after unsyllabifiable
consonants, realized as high central short [ɨ], with a word-final [ǝ]
allophone. Underlying forms can have long sequences of consonants, and
some words have no vowels at all. Only /y/, /w/, and / s/ (the language’s
only fricative consonant) occur as word-internal codas. Word-finally any
type of consonant is allowed, including obstruents, nasals, liquids, and
glides. Underlying forms are accommodated to a CV syllabic template
where possible by inserting the nucleus /ɨ/, driven by the basic constraints
ONSET and NOCODA.61
With respect to syllabification, Kalam phonemes can be divided into three
classes:

(25) a. Consonants / p b m t s d n c j n͂ k g ŋ l/ – always non-syllabic.

b. Vowels / e a o/ – always syllabic.
c. Semivowels / i u/ ⁓ /j w/ – syllabic and non-syllabic.

The epenthetic vowel [ɨ] ([ǝ] word-finally) is not in the underlying

inventory, and is pronounced shorter than underlying vowels and vocalized
semivowels:

(26) a. Epenthetic [ɨ], [ǝ] are short: /kn/ [ˈkɨn], /m/ [ˈmǝ]
b. Phonemic vowels are half-long: /kay/ [ˈkaˑj]
c. Including the syllabic allophones of the semivowels: /kyn/ [ˈkiˑn]

B&P give four arguments that the semivowels are always underlying
consonants. Their first argument is that while /a e o/ are found word-
initially in native words, no native words begin with /i/, /u/ or any central
vowel. Instead, words may begin with [ji] or [wu]. B&P analyze them as
beginning phonemically with /j/, /w/.

P&B (2011: 31) describe the vowels in such words as predictably inserted
“release vowels” colored by the adjacent semivowel. However, as
reproduced in (27), B&P consistently transcribe them as half-long, like
regular vowels and like the vocalized glides in words like (24b) /kyn/
[ˈkiˑn], but unlike the short vowel predictably inserted between two
consonants in (24a) /kn/ [ˈkɨn].62 That means that they are not release
vowels, but vocalized glides, which for P&B’s analysis means that the
semivowels are vocalized as [ji-], [wu-] before consonants, since no words
begin with [i-], [u-]. P&B do not account for these data. A simpler
alternative is that Kalam words are syllabified to have onsets where
possible, so initial glides are [ji-], [wu-] before consonants and [j-], [w-]
before vowels. Since semivowels can be both syllabic nuclei and margins,
an initial semivowel followed by a consonant can satisfy the the CV
preference by being syllabified as /ji-/, with the same melodic element
serving as onset and nucleus. For example, underlying {Im} (or for that
matter underlying {im} or {ym}) is then syllabified as /jim/, rather than as
*/im/, */jm/, or as */jm/. Since vowels cannot be affiliated with onsets, it
also follows that underlying {Am} (or {am}) must be syllabified as /am/
rather than as */a̯am/. This derives the generalization that words can begin
only with a consonant or with a true vowel but not with a semivowel or
with an epenthetic /ɨ/.63 There is then no need to specify the semivowels as
underlyingly non-syllabic. This analysis is straightforwardly implementable
in OT, as we’ll see shortly in tableau (31) below. Before proceeding to that
formalization, let us review B&P’s remaining three arguments for
underlying /y/, /w/.
 B&P’s second argument for underlying /y/, /w/ is that a word can end in
a consonant or in a vocalized semivowel [-i] and [-u], but not in a vowel [-
a], [-e], or [-o], e.g. (24d) /amy/ [ˈaˑˈmiˑ], but */ama/ *[ˈaˑˈmaˑ]. On their
assumption that [-i] and [-u] are underlying consonants, this follows from
the constraint that words cannot end in vowels. I assume the weaker
constraint that words cannot end in non-high vowels.64
B&P’s third argument is based on the distribution of the two allomorphs
of the negative prefix or proclitic /ma-/ ⁓ /m-/ ‘not, not yet’. The choice of
allomorph is phonologically determined: /m-/ occurs before vowels: /m-ag-
p/ ‘he did not speak’, /m-ow-p/ ‘he has not come’, /m-o-ng-gab/ ‘he will not
come’. /ma-/ occurs before consonants and before semivowels: /ma-pkp/ ‘it
has not struck’, /ma-dan/ ‘don’t touch’, /ma-ynb/ ‘it is not cooked’, /ma-
wkp/ ‘it is not cracked’.65

B&P’s assumption that semivowels are underlying consonants explains this

distribution. But so does my assumption that initial semivowels before
consonants are syllabified as CV- to maximize onsets, as discussed in
connection with the first argument.
B&P’s fourth argument is that the only word-internal surface vowel
sequences in Kalam are [i:a:], [i:o:], [i:e:], [u:a:], [u:o:], [u:e:]. These words
have an alternative monosyllabic pronunciation in free variation (P&B
2011: 32), and P&B sometimes transcribe them with a bridging glide (see
(29c)).
B&P reason that these are the only permitted word-internal vowel
sequences because they are underlying consonant+vowel sequences, hence
exempt from the prohibition against adjacent vowels. They would have to
add some constraint to ensure that such hiatus sequences only arise after a
consonant, in order to rule out such words as */yon/ *[ˈiˑˈyoˑn] in Kalam.
On my account this follows automatically without any additional
constraints. Since Kalam permits no tautosyllabic consonant clusters, and
specifically no initial clusters, we know that *COMPLEX is undominated.
ONSET forces syllables to have onsets where possible, so initial glides are
[ji-], [wu-] before consonants and [j-], [w-] before vowels, e.g., /im/ = /ym/
= /Im/ is syllabified as [jiˑm]. Vowels must be syllabic, so /am/ = /a̯am/ =
/Am/ is syllabified as /am/. This derives the generalization that words can
begin only with a consonant or with a true vowel but not with a semivowel
or with an epenthetic /ɨ/. We also know from such data as (24c,d,e) that
ONSET is dominated. The ranking *COMPLEX >> ONSET forces the
syllabification [ki.on] rather than [kyon].
All the pieces of our OT analysis are now in place. The constraints are
listed in (30) and the derivations are illustrated in (31). In the
representations, “v” stands for an empty nucleus (a syllable head with no
affiliated phoneme), realized as [ɨ]. I give alternative inputs in the tableau in
order to show the non-contrastive status of syllabicity in underlying
representations.

(30) Constraints
a. *V]ω : A word can’t end in a vowel.
b. SON: Consonants are non-syllabic, vowels are syllabic (see (25)).
c. NUC: A syllable has a nucleus.
d. IDENT(F): Input [αF] does not correspond to output [–αF].
e. DEP: Don’t insert a segment.
f. MAX: Don’t delete a segment.
g. *COMPLEX: A syllable does not have a consonant cluster.
h. ONSET: A syllable has an onset.
i. *CODA: A syllable does not have a coda.
That monoconsonantal words such as (24f) /m/ [‘̍mǝ] undergo epenthesis
follows not from reranking the constraints for those derivations, as B&P
suggest, but from standard constraints that are plainly unviolated in Kalam:
lexical words are accented, accents fall on a syllable, and syllables contain a
vocalic nucleus.66
Underlying forms are accommodated to a CV syllabic template by
inserting a vocalic nucleus realized as [ɨ] where necessary, driven by the
basic constraints ONSET and NOCODA. The insertion of the predictable [ɨ]-
vowels can be regarded as part of the syllabification process. The evidence
that they are epenthesized phonologically, rather than intruding
phonetically, is that they are obligatory regardless of speech rate, provide a
nucleus for syllables that would otherwise lack one, and carry word stress.
Several strands of evidence show that they originate specifically in the word
phonology rather than in the postlexical phonology. They are grouped into
binary feet within the domain of a word: in a word consisting of such
syllables, odd-numbered ones can get stressed (Pawley & Bulmer 2011:
30). Corroborating the lexical status of epenthetic vowels is the partial
unpredictability, or perhaps morphological conditioning, of these stresses.
For example, the second stress is on the fifth syllable in (24k) and on the
third syllable in (24l), perhaps because of the different morphological
structure. Finally, each member of a compound counts as a separate word
for purposes of the syllable count and word-finality. The inserted nucleus [ɨ]
is for these reasons an l-phoneme, not part of the s-phonemic representation
because it is predictable. Hence the regular CV syllable structure of the
language cannot be represented at the s-phonemic level. P&B posit initial
syllabifications with consonantal syllables like /t.d.k.s.pm/, /m.d.n.k.nŋ/,
contrary to Jakobson’s CV universal. The simple CV syllabification is,
however, visible in lexical representations, where the l-phoneme /ɨ/ is
present.
B&P say that Kalam ɨ does not neatly fit into Currie-Hall’s (2007)
typology of phonologically epenthetic vs. phonetically intrusive
(excrescent) vowels, on the grounds that it has two properties of intrusive
vowels in addition to the standard properties of epenthetic vowels: it does
not repair illicit structures, and it is a central vowel ([ɨ], word-finally [ǝ]).
Neither of these arguments hit the mark. The first argument overlooks the
generalization that Kalam epenthetic vowels provide a nucleus for
consonants that would otherwise have to be syllabified as codas but are
prohibited in coda position.67 But providing a nucleus for unsyllabifiable
consonants is repairing syllable structure. The second argument is based on
the incorrect premise that inserted central vowels are always intrusive rather
than epenthetic. There are many well-documented instances of epenthetic
central vowels, for example in German (Wiese 1986, 2000: 245), Catalan
(Wheeler 2005, Ch. 8), Armenian (Vaux 1998a, 1998b, 2003, Delisi 2015),
Slovenian (Jurgec 2007), some dialects of Berber (Dell & Tangi 1993),
Salishan languages (Parker 2011), and Mongolian (Svantesson 1995,
Svantesson et al. 2005) — all demonstrably phonological cases of ǝ-
epenthesis, some of them cyclic or morphologically conditioned, hence
definitely lexical. The correct generalization is the converse: intrusive
vowels are always central (unless of course they acquire peripheral features
from their context). In other words, independent peripheral quality is a
diagnostic for epenthetic vowels, but central quality is not a diagnostic of
intrusive vowels. That being the case, Kalam epenthetic vowels fit perfectly
into Hall’s typology; they do not have “mixed properties”.
Thinking that Kalam epenthetic vowels have mixed properties, B&P
classify them as a third category which they call REMNANT VOWELS. They
propose that remnant vowels arise from the historical loss of reduced
unstressed vowels, followed by reanalysis of deletion as insertion in the
complementary contexts (rule inversion) and possibly generalization of the
new insertion process, and that this origin explains their mixed properties.
This is no doubt how the Kalam epenthetic vowels arose. However, even
supposing contrary to fact that Kalam vowels had mixed properties, B&P’s
historical account would not explain that mixture, for there are numerous
synchronic epenthesis processes that are extended inversions of original
syncope and apocope processes and do NOT have mixed properties (see
Andersen 1969 on the synchrony and diachrony of Ukrainian paragoge; on
some analyses even the English schwa in the plural, genitive, and reduced
copula is a case). In any case there are at present no known clear instances
of epenthesis with mixed properties. The two-way distinction between
phonologically epenthetic and phonetically intrusive vowels offers a
sufficient typology of vowel insertion.
Comrie (1991) argues that of the seven Haruai vowels in (32), only /ǝ/ is
phonemic.

(32)

[ɨ] is an anaptyctic vowel predictably inserted after the first consonant in a

sequence of consonants or in a word that consists only of consonants.
Comrie analyzes the high vowels i and u as /y/ and /w/, on the basis of
arguments that are similar to those of B&P for Kalam. He analyzes a as /
ǝǝ/, and tentatively derives the mid vowels e and o from /ǝyǝ/ and /ǝwǝ/,
respectively, with the reservation that some lexical items are exceptions to
these contractions, and others undergo them only optionally. By strict
structuralist methods this amounts to a contrastive distribution, leaving
Haruai with a horizontal vowel system of three phonemes /e/, /ǝ/, /o/. My
suspicion is that they may be reanalyzed as morphophonemically
unspecified for syllabicity along the lines of what I proposed above for
Kalam, emerging at the lexical level as vocalic and consonantal l-phonemes
depending on syllable structure.

3.2 Kabardian
In the course of their argument that universals are “myths”, Evans &
Levinson (2009: 438) claim that it is “contested” whether all spoken
languages have vowel phonemes at all, citing the Northwestern Caucasian
languages, where “the quality of the vowel segments was long maintained
by many linguists to be entirely predictable from the consonantal context
(see Colarusso 1982; Halle 1970; Kuipers 1960)” (E&L 2009: 438). It must
be said that there is no such debate about Northwestern Caucasian or any
other languages; Colarusso (1982) and Halle (1970) demonstrated a
minimum inventory of two contrastive vowels.68 And it is not accurate that
“although most scholars have now swung round to recognizing two
contrasting vowels, the evidence for this hangs on the thread of a few
minimal pairs, mostly loanwords from Turkish or Arabic” (ibid.). Actually
the majority of scholars recognize a THREE-VOWEL system, and to the extent
that some have “swung round” to the two-vowel analysis, it is not from the
vowel-less analysis but from the three-vowel analysis. Nor does the
evidence particularly depend on Turkish and Arabic loans, or on minimal
pairs. On the contrary, the strongest evidence comes from native words and
the core vocabulary (Colarusso 1992: 22).
The history of scholarship on Kabardian phonology is worth reviewing
as an example of the theory-dependence of phonemic analyses. Older
reference grammars of Kabardian set up seven vowel phonemes: two
variable short vowels /ǝ/ and /ɐ/, whose realization depends mostly on the
following consonant, and five stable long vowels /ɑ: i: u: e: o:/, phonetically
more peripheral than the variable ones (Jakovlev 1948, Turčaninov &
Tsagov 1940, Abitov et al. 1957, Šagirov 1967, Bagov et al. 1970, followed
by Maddieson 1984, 2013). Jakovlev (1923) discovered that the stable
vowels can be derived by fusion of the short vowels with a glide. Most s-
phonemic theories allow fusion (e.g., phonemicizing English [ɚ] as /ǝr/, or
French and Portuguese nasal vowels as V+N sequences). In these it is
straightforward to reduce Kabardian /i: u: e: o:/ to underlying /ǝy/, /ǝw/, /
ɐy/, /ɐw/ respectively. Historically /ɑ:/ undoubtedly comes from an
analogous fusion of /ɐh/, but synchronically it can’t quite be reduced to that
under the strictures of s-phonemics. So in addition to /ǝ/ and /a/, analyses
typically posit it as a third s-phoneme. Its representation has been the
subject of some debate. Currently favored is /a:/ (Choi 1991, Matasovic ´
2006, Wood 1994, Gordon & Applebaum 2006, Applebaum & Gordon
2013). An older theory posits a vertical three-vowel system /ǝ/ /ɐ/ /a/
(Trubetzkoy 1925, 1929, 1939, Catford 1942, 1984, Kumaxov 1973, 1984).
Abstract generative analyses, on the other hand, can easily derive the fifth
long vowel [ɑ:] from /ɐh/, in some cases ultimately from /hɐ/ and /ɐɣ/ by
other processes. The result of that further analytic step is the two-vowel
system of Kuipers (1960), Halle (1970), and Colarusso (1992, 2006).
It cannot be emphasized enough that the seven-vowel, three-vowel, and
two-vowel solutions with its two variants do not reflect any disagreement
about Kabardian, only the differences between phonological theories. The
data and phonological generalizations of Kabardian are not at stake. Each
analysis follows rigorously from exactly the same facts depending on the
principles that it assumes. Even the choice between the two variants of the
three-vowel phonemic system is a deep question of principle: what is at
stake is whether phonemics should privilege phonetic criteria, or
morphophonemic criteria and the overall simplicity of the grammar. The
former in this case favor a qualitative opposition /ɐ/ : /ɑ/, the latter point to
a quantitative opposition /a/ : /a:/. Far from being a dismaying free-for-all,
this spectrum of analyses is heartening because it means that our
understanding of Kabardian has reached a point where it can be advanced
by sharpening phonological theory and typology by empirical work on
other languages.
The upshot, then, is that Kabardian has at least three vocalic phonemes
(s-phonemes), reducible to two underlying m-phonemes. With that and the
failure of the refutation of the CV universal, E&L’s case against
phonological universals falls apart.
Even though Kabardian is not vowel-less, it remains, at the level of s-
phonemic representations, an exception to the proposed universals on vowel
systems in (14). A look at its phonology makes it likely that its lexical
representations do conform to them. Phonetically, Kabardian makes full use
of the vowel space, with unrounded and rounded front vowels and rounded
back vowels, in three heights, ten vowels in all according to Colarusso
(1992). In (33) I give examples of his phonetic and underlying forms, to
which have added the phonemic representation according to the three-vowel
analysis.
This ten-vowel repertoire arises by assimilation in height, backness, and
rounding to a following consonant, if there is one.69 Vowels are fronted
before [–high] coronals (alveolars, alveopalatals, palatoalveolars), fronted
and raised before [+high] coronals (palatals and palato-alveolars), backed
before plain uvulars and pharyngeals, backed and rounded before rounded
uvulars, and raised and rounded before labiovelars (there are no plain
velars). Onset consonants also color the following vowel, but in a variable
and gradient manner at the level of phonetic implementation, as Colarusso
(1992: 31) makes clear. Word-finally and before labials and the laryngeal /
Ɂ/, which lack a distinctive tongue position, the vowels are unraised and
front (Colarusso 1982: 96, 1992: 30).70 These assimilations, summarized in
(34), generate ten surface vowels.

The chart in (35) summarizes the assimilation patterns of backness and

roundness, and (dashed arrows) of height; note that the mid vowels each
have two sources.

(35)

Before /Ɂw/, which triggers rounding but not backing, /ǝ/ and /a/ are realized
as ö, ͻ¨ , not shown in the two-dimensional diagram. The long vowels /i:/
/u:/ /e:/ /o:/ originate by the same assimilation processes before /y/ and /w/,
which are then deleted with compensatory lengthening. /h/ neutralizes /ǝ/
and /a/ without any other coloring effect, and deletes like the other glides,
giving /a:/.
By the criteria (A)-(G) of Section 1.1, these assimilations are
phonological processes, not coarticulation processes, and they take effect in
the word-level phonology, everywhere within the word domain, but not
across phonological word boundaries (Gordon & Applebaum 2010: 51).
They are categorical and operate on discrete feature values. Note especially
that [–high] consonants trigger a chain shift, so that [o] and [ɛ] represent
either /ǝ/ and /ɐ/, depending on the following consonant.
The surface vowels of Kabardian are thus l-phonemes. The ten-vowel
system that emerges at the word level is symmetric and dispersed. It is
isomorphic to UPSID’s ten-vowel system for Korean (Maddieson 1984:
283). Perhaps significantly, the four-way combination of the values of
[round] and [back] that is its outstanding typological feature is also found in
coterritorial Turkish and its relatives, and elsewhere in Eurasia (Uyghur,
Selkup, Seto, Dagur, among others).

3.3 Marshallese
The other famous case of a vertical vowel system is Marshallese. Bender
(1963) had posited the phonemic vowel system (36) (I have replaced the
unrounded back vowels with their offical IPA symbols).

This is isomorphic to the Mofu-Gudur vowel system mentioned in Fn. 32.

Bender (1968) reduced it to four vowels, taking the center series ɯ, ɤ, ʌ, ɑ
as basic (p. 20),71 and deriving the other eight vowels by assimilation to
adjacent consonants:

(37) a. [–back,–round] next to “light” (palatal or palatalized) consonants

py,ty,my,ny,ly,j,y
b. [+back,–round] next to “heavy” (plain, or perhaps velarized)
consonants p,t,k,m,n,ŋ,l,r,h
c. [+back,+round] next to labialized consonants kw,nw,ŋw,lw,rw,w

When a vowel appears between consonants that differ in backness and

rounding, it assimilates to both and is accordingly pronounced as a
diphthong. This results in 24 diphthongs [iɯ], [iu], etc.
Bender further noted that the second row of vowels / i ɤ u/ can be
eliminated from the inventory by deriving it from two sources: lowering of /
ɯ/ before deleted /ʌ/, and raising of /ʌ/ before deleted /ɯ/:

(38) a. /bwʌkwɯ/ [bwukw] ‘bladder’, cf. /bwʌkwɯ-ny/ [bwukwɯny] ‘his

bladder’
 b. /wɯbʌ/ [ubw] ‘chest’, cf. /wɯbʌ-ny/ [ubwʌny] ‘his chest’

He rejects this further reduction for the phonemic level because it would
violate biuniqueness. Words like [bwuŋw] ‘night’ could be phonemicized
either as /bwʌŋwɯ/ or as /bwɯŋwʌ/, unless one had access to the
morphonological information about the underlying second syllable from
suffixed forms, which is not available at the phonemic level. At the
morphophonemic level, this objection falls away.
Nevertheless Bender’s remarkable solution does not fully adhere to the
principle of biuniqueness, and transcends structuralist procedures of
segmentation and classification, for the context that triggers the vowel
allophones is sometimes itself deleted, and the phoneme /h/, the “heavy”
counterpart of /y/ and /w/, never surfaces at all. For example, the three-way
contrasts in long vowel vowels seen in (39) are due to deleted intervocalic
glides.

(39) a. /mayar/ mɛɛr ‘to tell a lie’, /mahaj/ mɑɑj ‘open field’, /mawar/
mͻͻr ‘bait’
b. /mʌyʌj/ meej ‘dark colored’, /rʌhʌj/ rʌʌj ‘bright colored’, /tʌwʌj/
tooj ‘conspicuous’

The implication of these data for a Stratal OT analysis is that vowel

coloring must take effect in the lexical phonology, and then made opaque
by postlexical processes. So in the derivation /mayar/ → mɛyɛr → mɛɛr,
lexical assimilation is masked by the postlexical deletion of its trigger.72
Independent evidence that intervocalic glide deletion is postlexical is
that Marshallese does not allow superheavy syllables (*CVVC, *CVCC),
except for those that arise by just this deletion process. Thus postlexical
glide deletion makes the syllable structure constraint opaque.
However, there is evidence that assimilation also applies postlexically,
as a coarticulation process. In rapid speech, consonant assimilation across
word boundaries can feed vowel allophony, especially in sequences of
identical consonants. For example, in (40b) /bek/ is pronounced like /bekw/
[bokw] ‘sand’ through assimilation to the initial /kw/ of the following word
(Byron Bender, p.c.).
Phonetic implementation adds further color nuances to the vowels, which
Bender interprets as due to “competing consonantal influences on a less
fully specified vowel”: the vowel in /tyʌkw/ moves from front to back with
increasing rounding, all at mid height: [tye̯ o¯̯okw]. Similarly, /kwʌty/ is
pronounced [kwo̯ǝe̯ ty].73 It is not uncommon for constraints to be enforced at
more than one level, with the respective applications obeying the ground
rules at each level, as appears to be the case here.
I conclude that Marshallese vowel assimilation is a categorical word-
level process operating on binary feature specifications, overlain by a
gradient postlexical coarticulation effect.
Since the relation of levels in Stratal OT does not have to be biunique,
the underlying three-height vertical system of three vocalic m-phonemes /ɯ
ʌ ɑ/ (or / i e ɑ/) that Bender (1968) entertained can be maintained. The
fourth vowel height arises in the lexical phonology, where all vowels also
assimilate to palatal and labial consonants. This yields the system of twelve
vocalic l-phonemes (36), which is identical to the s-phonemic inventory of
Bender (1963).

4 Conclusion
All putative phonological universals are framed in terms of theory-
dependent categories, and defined on some theory-dependent level of
representation, most often the phonemic level. Therefore the linguistic
descriptions on which they are based cannot be theory-neutral or
atheoretical. The approach of “describing each language in its own terms”
is at best aspirational. With one exception, all grammars I am aware of draw
heavily on existing descriptive frameworks.74 Since there are no theory-
neutral grammars, there is no theory-neutral typology. In terms of Hyman’s
(2008) distinction, there are no “descriptive” universals of language. All
universals are analytic, and their validity often turns on a set of critical
cases where different solutions can be and have been entertained. The
choice between these is not a matter of taste or whimsy but of different
assumptions, each one with testable empirical consequences in a multitude
of other languages. It follows that the search for better linguistic
descriptions, more illuminating typologies, and stronger cross-linguistic
generalizations and universals should go hand in hand.
Stratal OT’s word level representations encode the typologically
significant phonological properties omitted in s-phonemic representations,
including syllabification regardless of whether it is contrastive or not, and
“quasi-phonemes”. They also encode typologically significant abstract
structural information that is missing in the phonetic record, such as
metrical and prosodic structure and feature sharing, while omitting
postlexical features and structurally irrelevant coarticulation phenomena.
This makes word phonology the sweet spot where typological
generalizations appear at their tidiest: it seems likely that it obeys all
phonological universals that phonemic representations do, and then some.
The difference is most dramatic where phonemic theory imposes extremely
abstract analyses, as in vertical and one-vowel systems. But the argument
that the word level should replace phonemics in typological research can
also be made in languages where lexical representations are fairly close to
classical phonemic representations.
Since the lexical level of representation is empirically supported and
formally anchored in Stratal OT and Lexical Phonology, it is a good
candidate for replacing the classical s-phonemic level. That would remove
an unmotivated residue of structuralism and replace it with a well motivated
level of representation that serves some of the same functions. Our analysis
of unusual syllabification and vowel systems shows it to be a Goldilocks
level that is just right for typology, in that it conforms to some important
generalizations that are obscured for technical reasons in structuralist
phonemic representations, therby leading to cleaner typologies and turning
near-universals into solid exceptionless universals.
Since lexical representations and l-phonemes were not defined with an
eye on typology, its positive typological implications are a nice bonus that
supports Stratal OT. In broader perspective, the outcome encourages the
joint pursuit of linguistic theory and typology, where universals are not just
inductive generalizations from putatively theory-neutral linguistic
descriptions but hypotheses that at once guide analysis and are informed by
it. It has the hallmark of a good theory, that it leads BOTH to better linguistic
descriptions and to stronger cross-linguistic generalizations and universals.
Going beyond the typology of segmental inventories and syllable structure,
the relevance of lexical representations is worth exploring further in
dispersion theory, language acquisition, language use, and sound change.

Acknowledgement: Special thanks to Byron Bender for sharing his

beautiful work on Marshallese, and to Larry Hyman for lively discussion
and comments on a draft.

References
Abitov, M. L. et al. (eds.). 1957. Grammatika kabardino-čerkesskogo literaturnogo
jazyka. Moskva: Izd-vo Akademii Nauk SSSR.
Alber, Birgit. 2005. Clash, lapse and directionality. Natural Language & Linguistic
Theory 23. 485–542.
Andersen, Henning. 1969. A study in diachronic morphophonemics: The Ukrainian
prefixes. Language 45. 807–830.
Anderson, Stephen R. 1982. The analysis of French shwa, or How to get
something for nothing. Language 58. 534–573.
Anderson, Stephen R. 2000. Reflections on “On the phonetic rules of Russian”.
Folia linguistica 34. 11–27.
Applebaum, Ayla & Matthew Gordon. 2013. A comparative phonetic study of the
Circassian languages. In Chundra Cathcart, Shinae Kang, & Clare S. Sandy
(eds.), Proceedings of the 37th Annual Meeting of the Berkeley Linguistics
Society, 3–17. Berkeley: BLS.
Austin, Peter & Joan Bresnan. 1996. Non-configurationality in Australian aboriginal
languages. Natural Language & Linguistic Theory 14. 215–268.
Avery, Peter, B. Elan Dresher & Keren Rice (eds.). 2008. Contrast in phonology.
Theory, perception, acquisition. Berlin: De Gruyter Mouton.
Bagov, P. M., B. X. Balkarov, T. X. Kuaševa, M. A. Kumaxov & G. B. Rogova
(eds.). 1970. Grammatika kabardino-čerkesskogo literaturnogo jazyka, 1.
Fonetika i morfologija. Moskva: Nauka.
Barreteau, Daniel. 1988. Description du mofu-gudur: Langue de la famille
tchadique parlée au Cameroun. Paris: Editions de l’ORSTOM.
Bedell, George. 1968. Kokugaku grammatical theory. Ph.D. dissertation, MIT,
Cambridge, MA.
Bender, Byron W. 1963. Marshallese phonemics: Labialization or palatalization?
Word 19. 335–341.
Bender, Byron W. 1968. Marshallese phonology. Oceanic Linguistics 7. 16–35.
Bender, Byron W. 1971. Micronesian languages. In Thomas A. Sebeok (ed.),
Current trends in linguistics, vol. 8: Linguistics in Oceania, 426–465. The
Hague: Mouton.
Bermúdez-Otero, Ricardo. 2012. The architecture of grammar and the division of
labour in exponence. In Jochen Trommer (ed.), The morphology and
phonology of exponence: The state of the art, 8–83. Oxford: Oxford University
Press.
Bermúdez-Otero, Ricardo. 2015. Amphichronic explanation and the life cycle of
phonological processes. In Patrick Honeybone & Joseph Salmons (eds.), The
Oxford handbook of historical phonology, 374–399. Oxford: Oxford University
Press.
Bloch, Bernard. 1941. Phonemic overlapping. American Speech 16. 278–284.
Bloch, Bernard. 1953. Contrast. Language 29. 59–61.
Bloomfield, Leonard. 1939. Menomini morphophonemics. In Etudes phonologiques
dédiées à la mémoire de M. le prince N. S. Trubetzkoy, 105–115. Travaux du
Cercle Linguistique de Prague 8. Prague.
Bloomfield, Leonard. 1962. The Menomini language. Edited by Charles F. Hockett.
New Haven, CT: Yale University Press.
Blumenfeld, Lev. 2004. Tone-to-stress and stress-to-tone: Ancient Greek accent
revisited. Proceedings of the 30th Annual Meeting of the Berkeley Linguistics
Society.
Borowsky, Toni. 1993. On the word level. In Sharon Hargus & Ellen Kaisse (eds.),
Studies in Lexical Phonology, 199–234. New York: Academic Press.
Breen, Gavan & Rob Pensalfini. 1999. Arrernte: A language with no syllable
onsets. Linguistic Inquiry 30. 1–25.
Burzio, Luigi. 1994. Principles of English stress. Cambridge: Cambridge University
Press.
Casali, Roderic F. 2014. Assimilation, markedness and inventory structure in
tongue root harmony systems. ROA 1319. http://roa.rutgers.edu/content/articl
e/files/1319_casali_1.pdf.
Catford. J. C. 1942. The Kabardian language. Maître phonétique, 3rd Series 78.
15–18.
Catford, J. C. 1984. Instrumental data and linguistic phonetics. In Jo-Ann W. Higgs
& Robin Thelwall (eds.), Topics in linguistic phonetics: In honour of E. T.
Uldall, 23–48. Coleraine, N. Ireland: New University of Ulster.
Choi, John, D. 1991. An acoustic study of Kabardian vowels. Journal of the
International Phonetic Association 21. 4–12.
Chomsky, Noam. 1964. Current issues in linguistic theory. The Hague: Mouton.
Chomsky, Noam, & Morris Halle. 1968. The sound pattern of English. New York:
Harper & Row.
Clairis, Christos. 1977. Première approche du qawasqar: Identification et
phonologie. La Linguistique 13. 145–152.
Clements, G. N. 2001. Representational economy in constraint-based phonology.
In T. Alan Hall (ed.), Distinctive feature theory, 71–146. Berlin: Mouton de
Gruyter.
Clements, G. N. 2003. Feature economy in sound systems. Phonology 20. 287–
333.
Colarusso, John. 1982. Western Circassian vocalism. Folia Slavica 5. 89–114.
Colarusso, John. 1992. A grammar of the Kabardian language. Calgary: University
of Calgary Press.
Colarusso, John. 2006. Kabardian (East Circassian). München: Lincom Europa.
Comrie, Bernard. 1991. On Haruai vowels. In Andrew Pawley (ed.), Man and a
half: Essays in pacific anthropology in honour of Ralph Bulmer, 393–397.
Auckland: The Polynesian Society.
Currie-Hall, Daniel. 2007. The role and representation of contrast in phonological
theory. Ph.D. dissertation, University of Toronto. Toronto Working Papers in
Linguistics.
Currie-Hall, Kathleen. 2013. A typology of intermediate phonological relationships.
The Linguistic Review 30. 215–275.
DeCamp, David. 1958. The pronunciation of English in San Francisco. Part 1.
Orbis 7. 372–391.
DeCamp, David. 1959. The pronunciation of English in San Francisco. Part 2.
Orbis 8. 54–77.
Delisi, Jessica L. 2015. Epenthesis and prosodic structure in Armenian: A
diachronic account. Ph.D. dissertation, UCLA Indo-European Studies.
Dell, François & Oufae Tangi. 1993. Syllabification and empty nuclei in Ath-Sidhar
Rifain Berber. Journal of African Languages and Linguistics 13. 125–162.
Dinnsen, Daniel A. 1985. A re-examination of phonological neutralization. Journal
of Linguistics 21. 265–279.
Downing, Laura J. 1998. On the prosodic misalignment of onsetless syllables.
Natural Language & Linguistic Theory 16. 1–52.
Dresher, B. Elan. 2009. The contrastive hierarchy in phonology. Cambridge:
Cambridge University Press.
Ebeling, C. L. 1960. Linguistic units. The Hague: Mouton.
Evans, Nicholas & Stephen C. Levinson. 2009. The myth of language universals:
Language diversity and its importance for cognitive science. Behavioral and
Brain Sciences 32. 429–492.
Fast, P. W. 1953. Amuesha (Arawak) phonemes. International Journal of American
Linguistics 19. 191–194.
Flemming, Edward. 1995. Auditory representations in phonology. Ph.D.
dissertation, UCLA. (Published New York: Garland Press, 2002.)
Flemming, Edward. 2016. Dispersion theory and phonology. Oxford Research
Encyclopedias: Linguistics. http://linguistics.oxfordre.com/view/10.1093/acrefo
re/9780199384655.001.0001/acrefore-9780199384655-e-110
Fougeron, Cécile & Rachid Ridouane. 2008. On the phonetic implementation of
syllabic consonants and vowel-less syllables in Tashlhiyt. Estudios de
Fonética Experimental 17. 139–175.
Gordon, Matthew & Ayla Applebaum. 2006. Phonetic structures in Turkish
Kabardian. Journal of the International Phonetic Association 36. 159–186.
Gravina, Richard. 2014. The phonology of Proto-Central Chadic: The
reconstruction of the phonology and lexicon of Proto-Central Chadic, and the
linguistic history of the Central Chadic languages. Ph.D. dissertation, Leiden
University. https://www.lotpublications.nl/Documents/375_fulltext.pdf
Green, Jenny. 1994. A learner’s guide to Eastern and Central Arrernte. Alice
Springs: IAD Press.
Greenberg, Joseph H., Charles A. Ferguson & Edith A. Moravcsik (eds.). 1978.
Universals of human language, volume 2. Stanford: Stanford University Press.
Guimarães, Maximiliano & Andrew Nevins. 2013. Probing the representation of
nasal vowels in Brazilian Portuguese with language games. Organon 28. 155–
178. ling.auf.net/lingbuzz/001693/current.pdf
Gussenhoven, Carlos. 1986. English plosive allophones and ambisyllabicity.
Gramma 10. 119–142.
Hale, Mark. 2000. Marshallese phonology, the phonetics-phonology interface and
historical linguistics. The Linguistic Review 17. 241–257.
Hall, Tracy Alan. 1993. The phonology of German /R/. Phonology 10. 43–82.
Halle, Morris. 1959. The sound pattern of Russian. The Hague: Mouton.
Halle, Morris. 1970. Is Kabardian a vowel-less language? Foundations of
Language 6. 95–103.
Harris, James W. & Ellen M. Kaisse. 1999. Palatal vowels, glides and obstruents in
Argentinean Spanish. Phonology 16. 117–190.
Harris, John. 1987. Non-structure-preserving rules in Lexical Phonology:
Southeastern Bantu harmony. Lingua 72. 255–292.
Harris, John. 1990. Derived phonological contrasts. In Susan Ramsaran (ed.),
Studies in the pronunciation of English: A commemorative volume in honour
of A. C. Gimson, 87–105. London: Routledge.
Henderson, John & Veronica Dobson. 1994. Eastern and Central Arrernte to
English dictionary. Alice Springs: Institute for Aboriginal Development.
Hsieh, Feng-Fan, Michael Kenstowicz & Xiaomin Mou. 2009. Mandarin
adaptations of coda nasals in English loanwords. In Andrea Calabrese & Leo
Wetzels (eds.), Loan phonology: Issues and controversies, 131–154.
Amsterdam: John Benjamins.
Hualde, José Ignacio. 2005. Quasi-phonemic contrasts in Spanish. In Benjamin
Schmeiser, Vineeta Chand, Ann Kelleher & Angelo Rodriguez (eds.), West
Coast Conference on Formal Linguistics 23. Somerville, MA: Cascadilla
Press.
Hyde, Brett. 2002. A restrictive theory of metrical stress. Phonology 19. 313–359.
Hyman, Larry M. 1976. Phonologization. In Alphonse G. Juilland, A. M. Devine &
Laurence D. Stephens (eds.), Linguistic studies offered to Joseph H.
Greenberg on the occasion of his sixtieth birthday, 407–418. Saratoga, CA:
Anma Libri.
Hyman, Larry M. 1983. Are there syllables in Gokana? In Jonathan Kaye, Hilda
Koopman, Dominique Sportiche & André Dugas (eds.), Current approaches to
African linguistics, vol. 2, 171–179. Dordrecht: Foris.
Hyman, Larry M. 1985. A theory of phonological weight. Dordecht: Foris.
Hyman, Larry M. 2008. Universals in phonology. The Linguistic Review 25. 83–
137.
Hyman, Larry M. 2011. Does Gokana really have no syllables? Or: what’s so great
about being universal? Phonology 28. 55–85.
Hyman, Larry M. 2015. Does Gokana really have syllables? A postscript.
Phonology 32. 303–306.
Itô, Junko & Armin Mester. 2003. Japanese morphophonemics: Markedness and
word structure (Linguistic Inquiry Monograph Series 41). Cambridge, MA: MIT
Press.
Itô, Junko & Armin Mester. 2015a. Sino-Japanese phonology. In Haruo Kubozono
(ed.), Handbook of Japanese phonetics and phonology, 289–312. Berlin:
Mouton de Gruyter.
Itô, Junko & Armin Mester. 2015b. Word formation and phonological processes. In
Haruo Kubozono (ed.), Handbook of Japanese phonetics and phonology,
363–395. Berlin: Mouton de Gruyter.
Jakobson, Roman. 1931. Die Betonung und ihre Rolle in der Wort- und
Syntagmaphonologie. In Réunion Phonologique Internationale tenue à
Prague: 18–21/XII 1930, 164–183. (Reprinted in Jakobson, Selected writings,
vol. 1: Phonological studies, 117–136. The Hague: Mouton, 1962.)
Jakobson, Roman. 1958. Typological studies and their contribution to historical
comparative linguistics. In Proceedings of the 8th International Congress of
Linguists, Oslo. (Reprinted in Jakobson, Selected writings, vol. 1:
Phonological studies, 523–532. The Hague: Mouton, 1962.)
Jakobson, Roman, Gunnar Fant & Morris Halle. 1952. Preliminaries to speech
analysis. Cambridge, MA: Acoustics Laboratory, Massachusetts Institute of
Technology.
Jakovlev, N. F. 1923. Tablitsy fonetiki kabardinskogo jazyka. In Trudy podrazriada
issledovaniia severokavkazskikh jazykov pri Institute Vostokovedeniia v
Moskve, I. Moskva.
Jakovlev, N. F. 1948. Grammatika literaturnogo kabardino-čerkesskogo jazyka.
Moskva: Izd-vo AN SSSR.
Janda, Richard D. 2003. “Phonologization” as the start of dephoneticization – or,
on sound change and its aftermath: Of extension, lexicalization, and
morphologization. In Brian D. Joseph & Richard D. Janda (eds.), The
handbook of historical linguistics, 401–422. Oxford: Blackwell.
Jurgec, Peter. 2007. Schwa in Slovenian is epenthetic. 2nd Congress of the Slavic
Linguistic Society. Berlin: ZAS. http://www.hum.uit.no/a/jurgec/schwa.pdf
(accessed August 24 2007).
Kager, René. 2007. Feet and metrical stress. In Paul de Lacy (ed.), The
Cambridge handbook of phonology, 195–227. Cambridge: Cambridge
University Press.
Kahn, Daniel. 1976. Syllable-based generalizations in English phonology. Ph.D.
dissertation, MIT.
Kaplan, Abby. 2011. Phonology shaped by phonetics: The case of intervocalic
lenition. ROA 1077. http://roa.rutgers.edu/article/view/1107
Karvonen, Dan. 2005. Word prosody in Finnish. Ph.D. dissertation, University of
California at Santa Cruz.
Kawahara, Shigeto. 2012. Review of Laurence Labrune, The phonology of
Japanese (2012). Phonology 29. 540–548.
Kenstowicz, Michael & Nabila Louriz. 2009. Reverse engineering: Emphatic
consonants and the adaptation of vowels in French loanwords into Moroccan
Arabic. Brill’s Annual of Afroasiatic Languages and Linguistics 1. 41–74.
Kessler, Brett. 1994. Sandhi and syllables in Classical Sanskrit. In Erin Duncan,
Donka Farkas & Philip Spaelti (eds.), The proceedings of the 12th West Coast
Conference on Formal Linguistics, 35–50. Stanford, CA: CSLI Publications.
Keyser, Samuel Jay & Kenneth N. Stevens. 2006. Enhancement and overlap in
the speech chain. Language 82. 33–63.
Kim, Susan. 2001. Lexical Phonology and the fricative voicing rule. Journal of
Linguistics 29. 149–161.
Kiparsky, Paul. 1996. Allomorphy or morphophonology? In Rajendra Singh (ed.),
Trubetzkoy’s orphan: Proceedings of the Montréal roundtable
“Morphophonology: Contemporary Responses”, 13–31. Amsterdam: John
Benjamins.
Kiparsky, Paul. 2003. Syllables and moras in Arabic. In Caroline Féry & Ruben van
de Vijver (eds.), The syllable in Optimality Theory, 147–182. Cambridge:
Cambridge University Press.
Kiparsky, Paul. 2006. Amphichronic linguistics vs. Evolutionary Phonology.
Theoretical Linguistics 32. 217–236.
Kiparsky, Paul. 2015. Phonologization. In Patrick Honeybone & Joseph Salmons
(eds.), The Oxford handbook of historical phonology, 563–582. Oxford: Oxford
University Press.
Kiparsky, Paul. To Appear. Paradigms and opacity. Stanford: CSLI Press.
Kleber, Felicitas, Tina John & Jonathan Harrington. 2010. The implications for
speech perception of incomplete neutralization of final devoicing in German.
Journal of Phonetics 38. 185–196.
Kohler, Klaus J. 1966. Is the syllable a phonological universal? Journal of
Linguistics 2. 207–208
Korhonen, Mikko. 1969. Die Entwicklung der morphologischen Methode im
Lappischen. Finnisch-Ugrische Forschungen 37. 203–262.
Kubozono, Haruo. 1999. Mora and syllable. In Natsuko Tsujumura (ed.), The
handbook of Japanese linguistics, 31–61. Oxford: Blackwell.
Kubozono, Haruo. 2003. The syllable as a unit of prosodic organization in
Japanese. In Caroline Féry & Ruben van de Vijver (eds.), The syllable in
Optimality Theory, 99–122. Cambridge: Cambridge University Press.
Kumaxov, M. A. 1973. Teorija monovokalizma i zapadnokavkazskie jazyki.
Voprosy jazykoznanija 4. 54–67.
Kumaxov, M. A. 1984. Očerki obščego i kavkazskogo jazykoznanija. Nal’cik:
Izdatel’stvo El’brus.
Kuipers, Aert H. 1960. Phoneme and morpheme in Kabardian. The Hague:
Mouton.
Labov, William. 1994. Principles of linguistic change. Vol. 1: Internal factors.
Oxford: Wiley-Blackwell.
Labrune, Laurence. 2012. Questioning the universality of the syllable: Evidence
from Japanese. Phonology 29. 113–152.
Ladd, D. Robert. 2006. “Distinctive phones” in surface representation. In Louis M.
Goldstein, D. H. Whalen & Catherine T. Best (eds.), Laboratory Phonology,
vol. 8, 3–26. Berlin: Mouton de Gruyter.
Laycock, D. C. 1965. The Ndu language family (Linguistic Circle of Canberra
Publications, Series C, 1). Canberra: Australian National University.
Liberman, Anatoly. 1991. Phonologization in Germanic: Umlauts and vowel shifts.
In Elmer H. Antonsen & Hans Henrich Hock (eds.), Stæfcræft: Studies in
Germanic linguistics, 125–137. Amsterdam: John Benjamins.
Lindblom, Björn. 1986. Phonetic universals in vowel systems. In John J. Ohala &
Jeri J. Jaeger (eds.), Experimental phonology, 13–44. Orlando: Academic
Press.
Lindblom, Björn. 1990. Explaining phonetic variation: A sketch of the H&H theory.
In W. J. Hardcastle & A. Marchal (eds.), Speech production and speech
modelling, 403–439. Dordrecht: Kluwer.
McCarthy, John & Alan Prince. 1986. Prosodic morphology 1986. http://scholarwor
ks.umass.edu/linguist_faculty_pubs/13
McCarthy, John & Alan Prince. 1993. Prosodic morphology I: Constraint interaction
and satisfaction. http://scholarworks.umass.edu/linguist_faculty_pubs/14
McCawley, James D. 1968. The phonological component of a grammar of
Japanese. The Hague: Mouton.
MacMahon, April M. S. 1991. Lexical Phonology and sound change: The case of
the Scottish vowel length rule. Journal of Linguistics 27. 29–53.
Maddieson, Ian. 1984. Patterns of sounds. Cambridge: Cambridge University
Press.
Maddieson, Ian. 2013. Chapters in Matthew S. Dryer & Martin Haspelmath (eds.),
The world atlas of language structures online. Leipzig: Max Planck Institute for
Evolutionary Anthropology. http://wals.info/
Maddieson, Ian & Kristin Precoda. 1990. Updating UPSID. UCLA Working Papers
in Phonetics 74. 104–111.
Manaster-Ramer, Alexis. 1994. On three East Slavic non-counterexamples to
Stieber’s Law. Journal of Slavic Linguistics 2. 164–170.
Martin, Samuel. 1975. A reference grammar of Japanese. New Haven, CT: Yale
University Press.
Martinet, André. 1964. Elements of general linguistics. Chicago: University of
Chicago Press.
Martínez-Gil, Fernando. 1993. Galician nasal velarization as a case against
Structure Preservation. In Proceedings of the 19th Annual Meeting of the
Berkeley Linguistics Society, 254–267.
Matasović, Ranko. 2006. A short grammar of East Circassian (Kabardian). Zagreb.
http://mudrac.ffzg.unizg.hr/~rmatasov/KabardianGrammar.pdf
Mohanan, K. P. 1982. Grammatical relations and clause structure in Malayalam. In
Joan Bresnan (ed.), The mental representation of grammatical relations, 504–
589. Cambridge, MA: MIT Press.
Nagarajan, Hemalatha. 1995. Gemination of stops in Tamil: Implications for the
phonology-syntax interface.
https://www.ucl.ac.uk/pals/research/linguistics/publications/wpl/95paper
Odden, David. 1995. The status of onsetless syllables in Kikerewe. OSU Working
Papers in Linguistics 47. 89–110.
Ó Siadhail, Mícheál. 1989. Modern Irish: Grammatical structure and dialectal
variation. Cambridge: Cambridge University Press.
Ohala, John J. 1990. Alternatives to the sonority hierarchy for explaining
segmental sequential constraints. In Michael Ziolkowski, Manuela Noske, &
Karen Deaton (eds.), Papers from the 26th Regional Meeting of the Chicago
Linguistic Society. Vol. 2: Parasession on the syllable in phonetics and
phonology, 319–338. Chicago: CLS.
Ohala, John J. & Haruko Kawasaki-Fukumori. 1997. Alternatives to the sonority
hierarchy for explaining segmental sequential constraints: In Stig Eliasson &
Ernst Håkon Jahr (eds.), Language and its ecology: Essays in memory of
Einar Haugen, 343–365. Berlin: Mouton de Gruyter.
Padgett, Jaye. 2010. Russian consonant-vowel interactions and derivational
opacity. In W. Brown, A. Cooper, A. Fisher, E. Kesici, N. Predolac, & D. Zec
(eds.), Proceedings of the 18th Formal Approaches to Slavic Linguistics
meeting, 353–382. Ann Arbor: Michigan Slavic Publications.
Padgett, Jaye & Máire Ní Chiosáin. 2011. Markedness, segment realization and
locality in spreading. In Linda Lombardi (ed.), Segmental phonology in
Optimality Theory: Constraints and representations, 118–156. Cambridge:
Cambridge University Press.
Parker, Aliana. 2011. It’s that schwa again! Towards a typology of Salish schwa.
Working Papers of the Linguistics Circle of the University of Victoria 21. 9–21.
Pawley, Andrew & Ralph Bulmer. 2011. A dictionary of Kalam with ethnographic
notes. Canberra, A.C.T.: Pacific Linguistics, School of Culture, History and
Language, College of Asia and the Pacific, The Australian National University.
Pensalfini, Robert. 1998. The development of (apparently) onsetless
syllabification: A constraint-based approach. In M. Catherine Gruber, Derrick
Higgins, Kenneth Olson & Tamra Wysocki (eds.), Papers from the 32nd
Regional Meeting of the Chicago Linguistic Society, 167–178. Chicago: CLS.
Piroth, Hans Georg & Peter M. Janker. 2004. Speaker-dependent differences in
voicing and devoicing of German obstruents. Journal of Phonetics 32. 81–
109.
Port, Robert F. & Michael O’Dell. 1985. Neutralization of syllable-final voicing in
German. Journal of Phonetics 13. 455–471.
Port, Robert F. & Penny Crawford. 1989. Incomplete neutralization and pragmatics
in German. Journal of Phonetics 17. 257–282.
Prince, Alan & Paul Smolensky. 1993. Optimality Theory: Constraint interaction in
generative grammar. RuCCS Technical Report 2, Rutgers University,
Piscateway, NJ: Rutgers University. Center for Cognitive Science. Revised
version published 2004 by Blackwell.
Ridouane, Rachid. 2007. Gemination in Tashlhiyt Berber: An acoustic and
articulatory study. Journal of the International Phonetic Association 37. 119–
142.
Roca, Iggy. 2005. Strata, yes, structure-preservation, no. In Twan Geerts, Ivo van
Ginneken, & Haike Jacobs (eds.), Romance languages and linguistic theory
2003, 197–218. Amsterdam: John Benjamins.
Rood, David S. 1975. Implications of Wichita phonology. Language 51. 315–337.
Rubach, Jerzy. 2000. Backness switch in Russian. Phonology 17. 39–64.
Ryan, Kevin. 2014. Onsets contribute to syllable weight: Statistical evidence from
stress and meter. Language 90. 309–341.
Šagirov, A. K. 1967. Kabardinskij jazyk. In V. V. Vinogradov (ed.), Jazyki narodov
SSSR, vol. 4: Iberijsko-kavkazskie jazyki, 165–183. Moskva: Nauka.
Schwartz, Geoffrey. 2013. A representational parameter for onsetless syllables.
Journal of Linguistics 49. 613–646.
Schwartz, Jean-Luc, Louis-Jean Boë, Nathalie Vallée & Christian Abry. 1997.
Major trends in vowel system inventories. Journal of Phonetics 25. 233–253.
Scobbie, James M. & Jane Stuart-Smith. 2008. Quasi-phonemic contrast and the
fuzzy inventory: Examples from Scottish English. In Avery, Dresher & Rice
(eds.) 2008, 87–113.
Smith, Tony. 1999. Muyang phonology. Yaoundé: SIL.
Sommer, Bruce A. 1970a. An Australian language without CV syllables.
International Journal of American Linguistics 36. 57–58.
Sommer, Bruce A. 1970b. The shape of Kunjen syllables. In Didier L. Goyvaerts
(ed.), Phonology in the 1980s, 231–244. Ghent: Story-Scientia.
Steriade, Donca. 1999. Alternatives in syllable-based accounts of consonantal
phonotactics. In Osamu Fujimura, Brian Joseph & Bohumil Palek (eds.),
Proceedings of LP 1998, vol. 1, 205–246. Prague: Charles University and
Karolinum Press.
Stevens, Kenneth N. & Samuel Jay Keyser. 1989. Primary features and their
enhancement in consonants. Language 65. 81–106.
Svantesson, Jan-Olof. 1995. Cyclic syllabification in Mongolian. Natural Language
& Linguistic Theory 13. 755–766.
Svantesson, Jan-Olof, Anna Tsendina, Anastasia M. Karlsson & Vivan Franzen.
2005. The phonology of Mongolian. Oxford: Oxford University Press.
Tabain, Marija, Gavan Breen & Andrew Butcher. 2004. VC vs. CV syllables: A
comparison of Aboriginal languages with English. Journal of the International
Phonetic Association 34. 175–200.
Topintzi, Nina. 2010. Onsets: Suprasegmental and prosodic behaviour.
Cambridge: Cambridge University Press.
Topintzi, Nina & Andrew Nevins. 2017. Moraic onsets in Arrernte. Phonology 34.
615–650.
Trubetzkoy, N. S. 1925. Review of Jakovlev 1923. Bulletin de la Société de
Linguistique de Paris 26. 277–286.
Trubetzkoy, N. S. 1929. Zur allgemeinen Theorie der phonologischen
Vokalsysteme. Travaux du Cercle Linguistique de Prague 1. 39–67.
Trubetzkoy, N. S. 1939. Grundzüge der Phonologie. Prague: Travaux du Cercle
Linguistique de Prague, No. 7.
Turčaninov, G. & M. Tsagov. 1940. Grammatika kabardinskogo jazyka. Moskva:
Izd-vo Akademii Nauk.
Turpin, Myfany. 2012. The metrics of Kaytetye rain songs, a ceremonial repertory
of Central Australia. http://linguistics.ucla.edu/event/icalrepeat.detail/2012/10/
03/325/-/pho
Vaux, Bert. 1998a. The laryngeal specifications of fricatives. Linguistic Inquiry 29.
497–511.
Vaux, Bert. 1998b. The phonology of Armenian. Oxford: Oxford University Press.
Vaux, Bert. 2003. Syllabification in Armenian, universal grammar, and the lexicon.
Linguistic Inquiry 34. 91–125.
Vaux, Bert & Andrew Wolfe. 2009. The appendix. In Eric Raimy & Charles E.
Cairns (eds.), Contemporary views on architecture and representations in
phonology, 101–143. Cambridge, MA: MIT Press.
Vaux, Bert & Bridget Samuels. 2015. Explaining vowel systems: Dispersion theory
vs. natural selection. The Linguistic Review 32. 573–599.
Vennemann, Theo. 1972. On the theory of syllabic phonology. Linguistische
Berichte 18. 1–18.
Versteegh, Kees. 1997. Landmarks in linguistic thought, vol. 3: The Arabic
linguistic tradition. London: Routledge.
Wells, John. 1982. Accents of English. Cambridge: Cambridge University Press.
Wiese, Richard. 1986. Schwa and the structure of words in German. Linguistics
24. 697–724.
Wiese, Richard. 2000. The phonology of German. Oxford: Oxford University Press.
Wood, Sidney. 1994. A spectrographic analysis of vowel allophones in Kabardian.
Working Papers 42. 241–250. Lund: Lund University Department of
Linguistics.
Ian Maddieson
Is phonological typology possible
without (universal) categories?

Abstract: Discussions of the sound structure of languages rely on a long

tradition of categorization. In this chapter the question of whether it is
possible to devise a continuous descriptive framework that does not rely on
categories is considered and rejected. Complete identity between utterances
does not occur and in any case cannot form the basis for generalization.
Certain frameworks devised to compare rhythm types, sonority, and overall
basic phonological complexity employ scalar variables, but these are in
practise founded on categorical assumptions. Doing typological work in
phonology without reliance on categories is considered unlikely to be
possible.

1 Introduction
Phoneticians and phonologists have generally relied on a basic descriptive
framework which presupposes a set of categories anchored in local and
dynamic aspects of the speech production mechanism and in the auditory
and perceptual systems and the mental processing capacities largely
common to all humans, as well as in the nature of the acoustic signal that
carries speech between the speaker and the listener. These include terms for
places of articulation such as bilabial, velar, or pharyngeal, labels for
categories of articulatory configurations and their auditory characteristics
such as plosive, fricative, or nasal, and categories for acoustic properties
such as burst, formant, and noise. Specific entities such as voiceless bilabial
plosive ([p]) or low central unrounded vowel ([a]) are also referenced. In
addition, higher-order categories such as consonant and vowel, liquid,
sonorant and obstruent, coronal and guttural are customary. Categories such
as the syllable and its component parts of onset, nucleus, rhyme, and coda,
and other larger units such as the intonational phrase are also familiar.
Analytical concepts such as tones, phonemes (or similar notions of
contrastive elements), and stress, and inventories of these elements as well
as categories that express relationships between variant forms, such as
assimilation, gemination, or lenition, also form part of this framework.
Comparison between the phonetic and phonological properties of languages
has mostly been based on such categorical properties: this or that language
has a similar vowel system to another but a distinct one from yet others;
these languages allow limited syllable structures but others allow a larger
range of structures; these languages require nasals before stops to assimilate
in place but these others don’t, and so on.
The most familiar body of work on phonetics and phonology from the
nineteenth, twentieth, and twenty-first centuries (including, for example,
Sweet 1877, Jespersen 1889, Trubetzkoy 1939, Hockett 1955, Catford
1977, Chomsky & Halle 1968, Maddieson 1984, Stevens 1998, Ladefoged
& Maddieson 1986, etc., etc.) for the most part assumes that the categories
established are more-or-less valid for any language without explicitly
arguing the point. And much of this familiar conceptual framework and
terminology has roots in considerably older traditions of scholarship in
Greek, Roman, Arabic, Indian, or Chinese cultures which also imply that –
even if the terms are used to describe properties of specific languages such
as Latin or Sanskrit – the descriptive framework itself is not language-
specific.
In other words, much of the terminology used in the phonetic sciences
and applied in phonological analysis refers to categories that are determined
outside the scope of an individual language. That is, they seem to fit the bill
of being “pre-determined categories” of the sort that Haspelmath (2007)
declared “do not exist”. Haspelmath (2010) argues that cross-language
comparison, and hence any form of linguistic typology, cannot be based on
“descriptive categories” but must instead be based on “comparative
concepts”. This seems like a distinction without a difference. The notion of
a category is of course widely discussed in philosophical literature and in
many specialized fields, and is open to divergent interpretations, but by-
and-large anything that can be called a concept can be interpreted as a
category. Harnad (2005) in a trenchant (and entertaining) article argues that
any cognitive act is necessarily an act of categorization. The very name
“typology” implies recognition of types, that is, categories. But purely
physical scales can be non-categorical. As an example, Harnad mentions
the categorical set of colors as opposed to the continuous property of
electromagnetic wavelength/ frequency which human perceptual and
cognitive systems divide up in colors.
In this chapter I consider whether it is conceivable (or useful) to discuss
within- and between-language similarities and differences without forming
categories, i.e., without appealing to any discrete variables (not necessarily
the familiar ones) that are taken to be language-independent. That is, can
we insightfully compare languages or their phonological attributes without
establishing types? In particular the foundation of various continuous-
seeming scales proposed in the literature will be discussed.

2 Commensurability
Any kind of linguistic analysis, most especially typology, depends on being
able to say that some tokens are exemplars of the same “entity” or can be
placed in a commensurable space: otherwise each speech act is sui generis
and no generalizations are possible.
“Sameness” could be physical identity; in which case it would not be
necessary to form any kind of over-arching category to subsume any
differences. But no two utterances, even by the same speaker of the same
lexical string in the same language, are ever identical. Hence, IDENTITY can
never provide a basis for grouping of phonetic/phonological samples.
Repetitions of the same utterance by the same speaker even under similar
conditions differ in many details. Consider the two spectrograms in Figure
4.1. These show two repetitions by a female English speaker of a string of
digits which form a familiar telephone number. The speaker is very
habituated to saying this string and the two repetitions are so similar that
listeners cannot reliably say if they have heard two playings of the same
recording or two different recordings.
These two utterances have almost identical overall timing and very
similar F0 contours – but nonetheless they differ in many details of timing,
amplitude, and spectral composition, some of which are indicated in the
annotations provided on the figure.
Note that for convenience the differences are mostly described here
using categorical labels, e.g., vowel, nasal, burst, formant, etc., since these
terms are familiar. However, in principle it is possible to largely avoid these
categorical labels by using circumlocutions referring only to continuous
variables, such as “the time interval between the first major increase in
signal amplitude and the following salient reduction in amplitude” instead
of “first vowel”.

Figure 4.1: Spectrograms of two repetitions of a digit string by the

same speaker. Differences between the two repetitions are noted on the
lower spectrogram.

A non-categorical alternative to identity is therefore to rely on such

continuous variables which can provide a space within which comparisons
of a scalar nature can be made. Work in the phonetic sciences, including in
the Laboratory Phonology paradigm, frequently measures properties of
utterances in terms of continuous variables in spatial, temporal, amplitude,
frequency and other dimensions. These are typically employed to quantify
differences between samples considered to exemplify different categories.
But can typology be done using continuous variables? That is, could we do
comparison between languages – the sorting out of similarities and
differences between them – without using “types”? Some work appears to
attempt to do this, for example, the extensive amount of work devoted to
quantitative studies of rhythm.
3 Scalar measures of rhythm
This research tradition is inspired by the rhythmic typology proposed by
Lloyd James (1940), and popularized by Pike (1946) and Abercrombie
(1967). This approach proposed a division into “syllable-timed” and
“stress-timed” languages based on whether syllables or stresses are are
believed to be closer to being isochronous. Later, “mora-timed” languages,
with Japanese as the prototype, were added as a third major class. The
clearly categorical nature of this typology was challenged by Dauer (1983)
who compared English, Spanish, Italian, and Greek. In all four of these
languages the inter-stress interval increases LINEARLY with an increase in
syllable count, despite a consensus that English and Spanish (at least)
belong to different rhythm categories. Moreover, the slope of the line fitting
interstress interval and syllable count is similar in all languages and mean
syllable duration is very similar, as shown in Figure 4.2 (data replotted after
Dauer 1967).
Dauer did not conclude that languages cannot be grouped by rhythmic
properties (though some readers interpreted her data that way). Instead, she
suggested that a variety of factors lead to judgments of rhythmic difference,
and that these place languages on a continuum rather than in discrete
categories. These factors concern (at least) syllable structure, the role of
stress, and vowel reduction. She argues that “stress-timed” languages have
more closed syllables than “syllable-timed” ones, calculating that closed
syllables constitute 56% in English, versus 30% in French and 26% in
Spanish. “Stress-timed” languages have more heavy syllables than
“syllable-timed” ones, and heavy syllables tend to attract stress and hence
have more stresses per unit time. “Stress-timed” languages have more
reduction of vowel quality in unstressed syllables than “syllable-timed”
ones and “stress-timed” languages have a more important contrastive
function for stress and therefore are more likely to allow alternative stress
patterns syntagmatically. Dauer illustrates a partial continuum as in Figure
4.3.
 Figure 4.2: Utterance duration by syllable count in four languages
(average of data from 2–4 speakers).

Figure 4.3: Dauer’s exemplification of a continuum of “stress-

timedness”.

Ramus et al. (1999), motivated, inter alia, by Dauer’s suggestions,

proposed that it would be possible to make continuous measures which
reflect the linguistic impressions of rhythm type. Specifically they
suggested measuring the following three parameters in a sample of
connected speech for any given language:
– %V proportion of speech that is vocalic (the inverse is %C)
– ΔV standard deviation of vocalic intervals
– ΔC standard deviation of consonantal intervals

Of these three measures, %V reflects how much of the speech is vocalic,

ΔV reflects the variability of vocalic intervals, and ΔC reflects the
variability of consonantal intervals.
 Ramus et al. (1999) made measurements of sample sentences in data
from eight languages: English, Polish, Dutch, French, Spanish, Italian,
Catalan, and Japanese. In the traditional rhythmic typology, the Germanic
languages are considered “stress-timed”, the Romance languages are
considered “syllable-timed”, and Japanese is considered “mora-timed”.
Polish has no agreed classification. Figure 4.4 plots the two-dimensional
grouping of these languages on the %V and ΔC variables. This pair of
variables groups languages into the traditional categories: English and
Dutch are grouped together, the four Romance languages are grouped
together, with Japanese distant from both these groups. Polish is classified
with “stress-timed” languages in the space defined by these two variables.
Figure 4.5 plots mean %V vs ΔV. This pair of variables also groups the
expected languages into the traditional categories, with the two Germanic
languages together, the four Romance languages together, and Japanese
apart. However, here Polish is in a group of its own.
In principle, thus, these languages can be placed in a continuous
multidimensional space, which supports – but does not depend on – a
classification based on intuitive impressions of rhythmic similarity/
difference that was originally based on assigning them to categories.
Substantial subsequent work inspired by the original study by Ramus et
al. has measured data on more languages, suggested modified indices,
tested if perceptual similarity estimates match measured distances,
examined the effects of variables such as speech rate and text type, as well
as other factors. For example, Grabe & Low (2002) provide data on 18
languages, as shown in Figure 4.6.
 Figure 4.4: Mean values of ΔC vs %V by language, after Figure 1 in
Ramus et al. (1999).

Figure 4.5: Mean values of ΔV vs %V by language, after Figure 2 in

Ramus et al. (1999).

At first glance it may appear that this work is a promising effort to do

phonological typology with continuous rather than categorical variables.
However, this is not really the case, for both pragmatic and principled
reasons. First of all it doesn’t work in practical terms: the proposed
measures turn out to be highly variable across individual speakers, speech
rates, language samples, etc. (see, for example, Dellwo et al. 2012, Arvaniti
2012). Very strikingly the languages in common between the Ramus et al.
(1999) and Grabe & Low (2002) studies are placed quite differently in ΔC /
%V space as can be seen by comparing the results in Figures 4.4 and 4.6. In
Figure 4.4 Catalan is close to the other Romance languages, whereas in Fig
ure 4.6, Catalan is distant from French and Spanish (and Rumanian) and
close to the Germanic languages in the sample. Polish is close to English
and German in Figure 4.4 but distant from all other languages in Figure 4.6.
Japanese is distant from all other languages in Figure 4.4 but falls with the
Germanic languages in Figure 4.6.

Figure 4.6: Mean ΔC vs %V over 18 languages after Figure 3 in Grabe

& Low (2002).

These inconsistencies could be due to various causes, in particular, the

fact that measurements of these parameters on individuals speaking the
same language vary considerably, as illustrated in Figure 4.7 showing
variation among eight speakers of Swiss German on the %V parameter. The
means for these different individuals range from about 38% to about 47%,
which is greater than the difference between English and French in Ramus
et al. (which is about 5%) or between French and German in Grabe & Low
(which is about 4%), and each individual also varies over a considerable
range within the speech samples used.
Another major source of within-language variability comes from the
fact that measurements on the same language at different speech rates vary.
Dellwo (2009) provided ΔC measurements from five languages from
speakers speaking a varying rates. The results are shown in Figure 4.8 with
the data divided into five speech-rate “bins”. The left panel shows the raw
data. In all five languages the variability of consonant intervals declines as
speech rate increases, but more sharply for those languages with higher
variability at slow rates. Hence fast-rate German overlaps with slow-rate
French (as highlighted by the dotted line). Applying a ratenormalization
which standardizes speech rate according to the local CV alternation rate, as
in the right panel, to some extent stabilizes the data from Czech, Italian, and
French. Normalized ΔC in these languages at the fastest rate is similar to
that at the slowest rate. However, English and Italian are now confounded at
faster rates.

Figure 4.7: Individual speaker means and variation of %V for 8 Swiss

German speakers from Dellwo et al. (2012).
 Figure 4.8: Raw and rate-normalized measures of ΔC in five languages
by speech rate (from Dellwo 2009).

These practical problems illustrating the instability of the proposed

rhythmic parameters are compounded by one of principle. Obviously, the
calculation of indices such ΔC or %V relies on a prior categorization of the
speech data into intervals labeled as Consonants and Vowels. These are
presumably precisely the kind of “pre-established categories” whose
existence was questioned in Haspelmath (2007). The resulting metrics are
continuous variables, which enable languages to be compared in a multi-
dimensional continuous space, but they depend on an assumption of
categoriality.

4 Phonological complexity scales

There are parallel problems for other seemingly continuous metrics which
have at times been put forward as ways to place languages along a
continuous scale or in a multi-dimensional non-categorical space. One
example is the measure labeled TPD (for “Total Phoneme Diversity”) by
Atkinson (2011). Atkinson proposed a single index designed to reflect
phonological complexity in order to test his hypothesis that phonological
complexity broadly declines with distance from humanity’s original African
homeland (analogically with the established decline in genetic diversity,
due to the founder effect). Figure 4.9 shows the fit between TPD and
distance from a hypothesized origin in Africa to the present location of a
large sample of languages. When languages are grouped into his choice of
six major geographical areas, Africa has the highest mean TPD value, and
South America and Oceania the lowest, as shown in Figure 4.10. Thus these
language groups can be compared along a single continuous variable.
However, the TPD score, itself continuous, is calculated from counts of
consonants, (basic) vowels and tones – founded on traditional categories,
and the assumption that such entities take part in categorical contrasts.
Improved (but still simple-minded) phonological complexity scores (e.g.,
Maddieson et al. 2011) can be proposed, but these remain derived from
categorical data.

Figure 4.9: TPD vs distance from a hypothesized origin in (West) Africa

(Figure 3 in Atkinson 2011).

5 Sonority scaling
Similarly, work by cultural anthropologists examining the relationship
between cultural/ environmental factors and phonological structures is
based ultimately on categorical data, even when a continuous scale is used.
This work seems to be little known among linguists, so will be summarized
in some detail here.

Figure 4.10: Mean and range of TPD scores for languages grouped
into six regional clusters (after Figure 4 in Anderson 2011).

Munroe et al. (1996) proposed that there is a major influence of climate on

phonological structure. Their hypothesis was that people in warm climates
spend more time outdoors than those in cold climates. This means that they
are often communicating over larger distances under poorer transmission
conditions (because they are outdoors). Hence, there is a greater need to
optimize the sound transmission characteristics of the language. They
interpreted this (originally) as predicting that warm climate languages will
show a preference for simple CV syllables.
Munroe et al. counted the proportion of CV syllables in up to 200 words
of the “basic” lexicon of 60 languages chosen to represent the 60 major
cultural areas recognized in the Human Relations Area Files (current
version at http://hraf.yale.edu). Their basic result was that languages
situated in warm climates did indeed generally have a higher proportion of
CV syllables in the wordlists. (Their dividing line between “warm” and
“cold” was that a cold climate has mean winter temperature of 10° C (50°
F) or lower for 5 or more months per year.) In a follow-up study Munroe &
Silander (1999) examined if the effect can be found within language
families (as a check on whether it is an artifact of where language families
with different inherited prototypical syllabic patterns are located).
The findings by Munroe and colleagues were challenged by Ember &
Ember (1999) who proposed that “baby-holding” rather than climate was a
better predictor of the frequency of simple CV structures across languages.
And this prompted Munroe’s team to counter that calculating a continuous
“sonority score” and making climate a multi-valued, rather than binary,
variable re-affirmed climate as a significant predictor of cross-language
structural differences in their phonology. Their result is shown in Figure 4.1
0, which plots a sonority score against the number of cold months per year.
The debate continued, with Ember & Ember (2007) pointing to the large
variation in sonority scores among languages in warm climates seen in Figu
re 4.11, and looking for further factors that might explain this. They
proposed three additional factors: (i) degree of sexual freedom (which
correlates with higher sonority); (ii) degree of plant cover (correlates with
higher sonority); (iii) degree of “mountain-ness” (correlates with lower
sonority). Their figure relating scores for extra-marital sex (taken from
Huber et al. 2004) to the sonority scores provided by Fought et al. (2004) is
shown in Figure 4.12. Very broadly, the higher sonority scores occur in
languages spoken by what are reported as more sexually free societies.
 Figure 4.11: Plot of “sonority score” vs number of cold months, after
Ember & Ember (2010), using data from Fought et al (2004). Linear fit
shown, R2 = .329.
 Figure 4.12: Boxplot of mean sonority score by sexual freedom (after
Ember & Ember 2010, based on data in Fought et al 2004, and Huber
et al 2004).

The correlations found in these studies are intriguing, and in all

probability some relationship between phonological characteristics and
climatic and ecological variables is real (e.g., Maddieson & Coupé 2015).
However, here the interest is primarily in the effort to place languages along
a continuous scale of sonority. But note that in this case also, as with the
work on rhythm types, the constructed continuous scale used to compare
the languages is one based on categorical values. The sonority score was
calculated by assigning a value between 100 and 2 to phonetic classes such
as low vowels (scored 100), nasals (scored 9), and stops (scored 2). A score
is constructed for each word in word-lists from each language by summing
the values for each segment in the word and dividing by the number of
segments counted (affricates, for example, count as two segments) and the
mean word-score per language is then calculated. This procedure evidently
depends entirely on the prior recognition of the categories of individual
segments to which sonority values are assigned.
6 Non-categorical sonority scaling?
The fact that efforts to place languages on a continuous sonority scale seem
to yield interesting results suggests that it could be worth considering the
construction of a purely continuous scale of sonority that does not depend
on a prior categorical labeling.
It is possible to imagine a scale that depended only on a measure of the
intensity of recorded speech samples, either of a list of words as in Fought
et al. (2004) or of continuous speech. The higher the proportion of speech
duration that consists of high-intensity sound the more sonorous it would be
considered. This could in principle be measured as the mean intensity found
in the samples being compared. (Note that intensity is a measure of the
integrated and rectified instantaneous amplitudes throughout a sample.)
There are, however, considerable problems in implementing such a
comparison, both of a practical and a theoretical nature. The major practical
problem revolves around the fact that amplitude values in a recording
depend not only on the speech power itself but on the conditions and
equipment used in the recording and on potential moment-to-moment
variations in the orientation of the speaker to the microphone. Also
individual speakers speak louder or more quietly than others. All these
factors mean that a crude measure of mean amplitude of individual samples
might not be representative of characteristics of a language. Possible
normalization schemes could be proposed, such as scaling amplitude values
in relation to the maximum observed in the sample, or to an average of the
maxima recorded in the tokens of the low central vowel /a/ which is present
in most languages and generally assumed to be the most sonorous segment.
Note that the latter strategy would depend crucially on identifying the
category of /a/ tokens, and would fail for languages that did not have a
comparable segment.
Amplitude also varies as a function of such factors as utterance type and
position in an utterance. In many languages, for example, amplitude
typically diminishes over the duration of an utterance. Consequently a given
word occurring late in an utterance would have lower amplitude than the
same word occurring early in the utterance. If a hypothesis of cross-
language relative sonority differences is based on the relative frequency of
more versus less sonorous segments in the LEXICON (rather than in running
speech), then this is a potentially serious confound. An unguided
comparison of mean speech intensity across samples is thus unlikely to be
informative.
Despite these problems it is certainly possible to conceive continuous
scales of language-level sonority. Rather than a purely algorithmic
procedure a preliminary test of a scoring using a minimally categorical
classification guided by local amplitude peaks and valleys has been
conducted. Speech samples are divided into sonorant and non-sonorant
intervals. Sonorance is defined by two properties: presence of a
(quasi-)periodic fundamental frequency and presence of a formant-like
structure in the mid-frequency range for speech (c. 300–3000 Hz). In
general there is on the order of 10 dB difference between peak amplitude in
a sonorous interval and the amplitude valley in a non-sonorous interval. A
language’s score is the percentage of a speech sample’s duration that is
sonorous according to these criteria.
The test data for this project is a subset of the language materials used
in Easterday et al. (2011). That study was designed to examine – regardless
of any attempt to assign languages to rhythm classes – whether the factors
suggested by Dauer (1983) as influencing rhythmic type judgments, such as
syllable structure and vowel reduction, actually correlate with the metrics
put forward by Ramus et al. (1999). The speech samples used are drawn
from those created by the Global Recordings Network (see http://globalreco
rdings.net/en/), an evangelical Christian organization that makes recordings
of proselytizing materials in a wide range of languages publicly available
online. Although it is not the intended purpose of these materials, they
provide a potentially valuable tool for cross-linguistic research, particularly
of the kind that simply examines acoustic properties without any need for a
transcription of the content. The recordings are well-matched in terms of
style and content since they are based on the same set of prepared didactic
texts and pictures.
Easterday et al. (2011) segmented the recordings used in their study into
Consonant and Vowel intervals, similar to the process in the work on
rhythmic typology discussed earlier. These analyses are being relabeled so
as to calculate the percentage of duration in running speech which is
periodic and broadly falls above a (locally-defined) intensity threshold.
Samples are of monologues of running speech spoken fluently but in a
moderately careful style. Pauses between sentences or other utterance units
are excluded from the duration calculation. The actual speech duration
measured in each sample varies between about one half and one minute in
length (range 27.3–52.3). Voiceless stop closures following a pause are
included as part of the pause duration. To date this re-analysis has been
completed for one speaker of each of 13 languages. The results are shown
in tabular form in Table 4.1 and graphed against distance from the Equator
in Figure 4.13.
The results of this analysis show that the range from highest to lowest
sonorant percentage is quite wide, but strikingly the majority of the
languages cluster around being about two-thirds composed of sonorous
material in running speech, at least in these samples. This suggests that this
is an approximate norm for languages. The results remain too meager so far
to draw strong conclusions, but they do provide a further hint in support of
the suggestion that some aspects of linguistic structure have links to climate
and environment. The distance of a language’s location from the Equator
can serve as a very rough proxy for warmer vs. cooler climate. Figure 4.13
plots the sonority score against the language’s point location as shown in
the LAPSyD database (Maddieson et al. 2013). There is indeed some
correlation between lower sonority and distance from the Equator. This
aligns with the proposal in Fought et al. (2004) and other suggestions that
sound patterns in languages are in part designed in response to
environmental factors (e.g. Maddieson et al. 2013, Maddieson & Coupé
2015).
Table 4.1: Sonorous percentage in sample of 13 languages.
% SONORANT
Qiang 62.4 %
Nez Perce 64.4 %
Thaayore 64.9 %
Tamazight 65.7 %
Chukchi 66.8 %
Sheko 67.0 %
Sgaw Karen 67.1 %
Mixtec 68.4 %
Pohnpeian 68.5 %
Maori 69.1 %
Ekagi 71.2 %
Towa 76.9 %
Maung 93.4 %
 Figure 4.13: Sonorous percentage against absolute distance in
degrees from the Equator.

However, once more, the basis for constructing this particular

continuous variable depends on a categorization of speech intervals, in this
case one that is functionally equivalent to establishing the binary categories
of sonorous and non-sonorous. The languages themselves are neither sorted
into categories, nor said to possess or lack particular categorical properties,
so the role of categorization in this exercise is minimal, but it is still there.

7 Final remarks
It has been shown that several proposals to characterize aspects of
phonological typology along continuous scales that are found in the
literature turn out to be based on a prior step involving categorical
classifications. A thought experiment to devise a sonority measure that was
entirely free of non-categorical assumptions seemed to founder on practical
and theoretical difficulties. However, one that makes only a minimal appeal
to prior established categories shows some promise of providing interesting
differentiation between languages, and a potentially intriguing connection
to hypotheses suggesting adaptation to aspects of the environment. This
perhaps shows that cross-language comparisons using scalar variables,
albeit derived from categorization, may nonetheless be of some interest.
 Is the failure to devise purely continuous scales for typological
properties simply a failure of imagination, due to over-familiarity with the
traditional way of observing phonetic and phonological characteristics
through the lens of established categories? Quite possibly. It seems also
possible that Harnad is right: “To cognize is to categorize”. Any attempt to
distribute languages along a parameter seems to entail defining some
property which is present or absent either in an absolute or a gradient
fashion in each language examined. A property is necessarily a categorical
entity. This does not mean that the categories that are familiar in the
established traditions in the phonetic sciences are necessarily the most
useful we can devise, or that they are applicable to all languages. But to this
phonological typologist, it does not seem practicable to compare languages
in the absence of categories.

References
Abercrombie, David. 1967. Elements of general phonetics. Edinburgh: Edinburgh
University Press.
Arvaniti, Amalia. 2012. The usefulness of metrics in the quantification of speech
rhythm. Journal of Phonetics 40. 351–373.
Atkinson, Quentin. 2011. Phonemic diversity supports serial founder effect model
of language expansion from Africa. Science 332. 346–349.
Catford, J. C. 1977. Fundamental problems in phonetics. London: Longman.
Chomsky, Noam & Morris Halle. 1968. The sound pattern of English. New York:
Harper & Row.
Dauer, Rebecca. 1967. Stress-timing and syllable-timing re-analyzed. Journal of
Phonetics 11. 51–62.
Dellwo, Volker. 2009. Choosing the right rate normalization methods for
measurements of speech rhythm. In Proceedings of AISV.
Dellwo, Volker, Adrian Leemann, & Marie-José Kolly. 2012. Speaker idiosyncratic
rhythmic features in the speech signal. In Proceedings of Interspeech 2012,
Portland OR.
Easterday, Shelece, Jason Timm, & Ian Maddieson. 2011. The effects of
phonological structure on the acoustic correlates of rhythm. ICPhS Hong
Kong.
Ember, Carol R. & Marvin Ember. 2007. Climate, econiche, and sexuality:
Influences on sonority in language. American Anthropologist, New Series 109.
180–185.
Ember, Carol R. & Marvin Ember. 2007. Rejoinder to Munroe and Fought‘s
commentary. American Anthropologist, New Series 109. 785.
Ember, Marvin & Carol R. Ember. 1999. Cross-language predictors of consonant-
vowel syllables. American Anthropologist, New Series 101. 730–742.
Fought, John G., Robert L. Munroe, Carmen R. Fought, & Erin M. Good. 2004.
Sonority and climate in a world sample of languages. Cross-Cultural Research
38. 27–51.
Grabe, Esther & E. L. Low. 2003. Durational variability in speech and the rhythm
class hypothesis. In Carlos Gussenhoven & Natasha Warner (eds.), Papers in
laboratory phonology 7, 515–546. Berlin: Mouton de Gruyter.
Harnad, Stevan. 2005. To cognize is to categorize: Cognition is categorization. In
Henri Cohen & Claire Lefebvre (eds.), Handbook of categorization in cognitive
science, 19–43. Amsterdam: Elsevier.
Haspelmath, Martin. 2007. Pre-established categories don’t exist: Consequences
for language description and typology. Linguistic Typology 11. 119–132.
Haspelmath, Martin. 2010. Comparative concepts and descriptive categories in
crosslinguistic studies. Language 86. 663–687.
Hockett, Charles. 1955. A manual of phonology (Indiana University Publications in
Anthropology and Linguistics, Memoir 11). Bloomington IN: Indiana University.
Huber, Brad R., Vendula Linhartova, & Dana Cope. 2004. Measuring paternal
certainty using cross-cultural data. World Cultures 15. 48–59.
Jespersen, Otto. 1889. The articulations of speech sounds represented by means
of analphabetic symbols. Marburg: N. G. Elwert.
Ladefoged, Peter & Ian Maddieson. Sounds of the world’s languages. Oxford:
Blackwell.
Lloyd James, Arthur. 1940. Speech signals in telephony. London: Pitman.
Maddieson, Ian. 1984. Patterns of sounds. Cambridge: Cambridge University
Press.
Maddieson, Ian, Tanmoy Bhattacharya, Eric D. Smith, & William Croft. 2011.
Geographical distribution of phonological complexity. Linguistic Typology 15.
267–279.
Maddieson, Ian & Christophe Coupé. 2015. Human language diversity and the
acoustic adaptation hypothesis. Proceedings of Meetings on Acoustics 25,
060005. http://dx.doi.org/10.1121/2.0000198
Maddieson, Ian, Sébastien Flavier, Christophe Coupé, Egidio Marsico, & François
Pellegrino. 2013. LAPSyD: Lyon-Albuquerque Phonological Systems
Database. Interspeech 2013, Lyon.
Munroe, Robert L., R. H. Munroe, & S. Winters. 1996. Cross-cultural correlates of
the CV syllable. Cross-Cultural Research 30. 60–83.
Munroe, Robert L. & Megan Silander. 1999. Climate and the consonant-vowel
syllable: A replication within language families. Cross-Cultural Research 33.
43–62.
Pike, Kenneth L. 1946. Intonation of American English. Ann Arbor: University of
Michigan.
Ramus, Franck, Marina Nespor, & Jacques Mehler. 1999. Correlates of linguistic
rhythm in the speech signal. Cognition 73. 265–292.
Stevens, Kenneth L. 1998. Acoustic phonetics. Cambridge, MA: MIT Press.
Sweet, Henry. 1877. A handbook of phonetics. Oxford: Clarendon Press.
Trubetzkoy, Nikolai S. 1939. Grundzüge der Phonologie. Travaux du Cercle
Linguistique de Prague 7.
Jeffrey Heinz
The computational nature of
phonological generalizations

Abstract: This chapter studies the nature of the typology of phonological

markedness constraints and the nature of the typology of the transformation
from underlying to surface forms from a computational perspective. It
argues that there are strong computational laws that constrain the form of
these constraints and transformations. These laws are currently stated most
clearly in terms of the so-called subregular hierarchies, which have been
established for stringsets (for modeling constraints) and are currently being
established for string-to-string maps (for modeling the transformations). It is
anticipated that future research will reveal equally powerful laws applicable
to non-string-based representations. Finally, this chapter argues that these
laws arise as a natural consequence of how humans generalize from data.

1 Illuminating the phonological component of

grammar
Wilhelm von Humboldt’s phrase “language makes infinite use of finite
means” (1836/1999) is oft-cited by Chomsky because not only does it
encapsulate an important characteristic of natural language, but it also
highlights why GENERATIVE grammars play an important role in
understanding this aspect of language. In brief, generative grammars are the
finite means, but the linguistic knowledge they represent can be applied to
unboundedly many linguistic forms. The psychological reality of generative
grammars is the powerful scientific hypothesis which underlies all work in
generative linguistics.
In this chapter, we will study generative grammars from both a
typological and computational perspective. The same Wilhelm von
Humboldt is reported (Frans Plank, p.c.) to have suggested that in order to
do linguistic typology, two encyclopedias are necessary. The first is “an
encyclopedia of categories” and the second is “an encyclopedia of
languages”. The encyclopedia of categories provides an ontology with which
the encyclopedia of languages — and the generalizations about them — can
be studied. The object of inquiry is natural language and the linguistic
generalizations. But the light we shine on them comes from the encyclopedia
of categories.
This chapter will argue that the theory of computation provides a
meaningful and insightful encyclopedia of categories, with which linguistic
generalizations ought to be studied. (My discussion is limited to
phonological generalizations, though much important work in a similar vein
exists for other kinds of linguistic generalizations (Chomsky 1956; Gazdar &
Pullum 1982; Shieber 1985; Rogers 1998; Kobele 2006; Graf 2013).) I will
endeavor to explain that when phonological generalizations are studied in
this light, there are computational laws which govern important aspects of
their nature. I will also argue that current phonological theory does not
account for these laws, and I will make suggestions as to how phonological
theory might be modified to do so.
In this way, the goals of this chapter are similar to the goals of Charles
Kisseberth in his 1970 paper “On the functional unity of phonological
rules”:
I will show [. . .] that a rather rich set of diverse phenomenon is related in a complex, but
quite coherent way. The theory of phonology has hitherto been blind to phenomena of
this sort [. . .] and I will attempt to make some suggestions about the kind of apparatus
the facts [. . .] seem to require that a theory of phonology contain. I am not, however,
principally interested in proposing detailed formalism; instead I would like to encourage
phonologists to look at the phonological component of a grammar in a particular way.
(Kisseberth 1970: 239, emphasis in original)

Kisseberth argued that important generalizations in languages were missed

by not paying attention to the functional unity of phonological rules. The
introduction of surface constraints into phonological theory followed, and
later became one of the cornerstones of Optimality Theory (OT; Prince &
Smolensky 1993, 2004).
Similarly, I am arguing that important generalizations in languages are
being missed by not paying attention to the computational nature of
phonological generalizations. Yes, I am talking about computational
generalizations of phonological generalizations. I argue that these meta-
generalizations are important because they too suggest a conspiracy of sorts:
phonological generalizations across languages are distinct, but they exhibit a
very strong tendency to exhibit particular computational properties.
I will argue that, when phonological generalizations are studied under
this light, the hypothesized computational laws are SUFFICIENTLY EXPRESSIVE
to account for the impressive range of cross-linguistic variation, and are
simultaneously VERY RESTRICTIVE in the sense that strong predictions are
made about which logically possible phonological generalizations are not
humanly possible ones. I will argue that in this respect these computational
laws better match the attested typology than what is predicted by classical
OT (and many of its variants), which, I will argue, is NEITHER sufficiently
expressive nor restrictive. I will also argue that the restrictive nature of the
computational laws help answer questions about how such phonological
generalizations can be learned.
The arguments that I am making in this chapter are not novel. They have
previously been published in articles and conference proceedings. This
chapter thus provides a roadmap for phonologists of this literature, and
attempts to present a unified, overarching perspective on both the
importance of this computational encyclopedia of categories and its
implications for a theory of phonology.75

2 What is phonology?
The fundamental insight in the twentieth century which shaped the
development of generative phonology is that the best explanation of the
systematic variation in the pronunciation of morphemes is to posit a single
underlying mental representation of the phonetic form of each morpheme
and to derive its pronounced variants with context-sensitive transformations.
This development, present in Chomsky (1951) and Halle (1959), was
perhaps stated most fully and completely with Chomsky & Halle (1968), and
persists in OT (Prince & Smolensky 2004) today.
Thus there is a point of agreement between different theories of
phonology, which is stated in (1).

(1) There exist underlying representations of morphemes which are

transformed to surface representations.
As a result of this fundamental insight, every particular theory of phonology
grapples with three fundamental questions:

(2) a. What is the nature of the abstract, underlying, lexical

representations?
b. What is the nature of the concrete, surface representations?
c. What is the nature of the transformation from underlying forms to
surface forms?

I would like to give some examples of how phonological theories aim to

answer these questions. It is not possible in this chapter to comprehensively
survey the range of answers that have been offered. Therefore, I only
highlight some answers (and only in very broad strokes).
Rule-based theories, as exemplified by Chomsky & Halle (1968), for
example, have argued that the abstract underlying representations are subject
to languagespecific morpheme structure constraints (MSCs). The
transformations from underlying forms to surface forms are due to language-
specific rules, which are applied in a language-specific order. Constraints on
surface representations were, generally speaking, not part of the ontology of
these theories, and therefore were not posited to have any psychological
reality. Such generalizations — the phonotactic generalizations — were
derivable from the interaction of the MSCs and the rules.
On the other hand, in classical OT (Prince & Smolensky 1993, 2004),
there are no constraints on underlying representations (richness of the base),
but there are psychologically real, universal constraints on surface forms
(markedness constraints). The transformation from underlying forms to
surface forms is formulated as an OPTIMIZATION over these markedness
constraints, in addition to constraints which penalize differences between
surface and underlying forms (so-called faithfulness constraints). While both
the markedness and faithfulness constraints are universal, their relative
importance is language-specific. So in every language the surface
pronunciation of an underlying representation is predicted to be the optimal
form (the one that violates the most important constraints the least), though
what is optimal can vary across languages because the relative importance of
the constraints can vary across languages.
These two theories are radically different in what they take to be
psychologically real. The ontologies of the theories are very different.
Perhaps this is most clear with respect to the concept of phonemes (Dresher
2011). Phonemes exist as a consequence of the ontology of rule-based
theories, but they do not as a consequence of the ontology of OT. This is
simply because phonemes are a kind of MSC; underlying representations of
morphemes must be constructed out of them, and nothing else. In OT, there
are no MSCs and hence there are no phonemes. Consequently,
generalizations regarding complementary distribution are explained in a very
different manner in the two theories, and they promote different views of the
notion of contrast. Despite these differences however, there is an important
point of agreement: in both theories, complementary distribution of speech
sounds in surface forms is the outcome of a transformation of underlying
forms to surface forms.
This is the point I wish to emphasize: neither theory abandons the
fundamental insight stated in (1).76 The theories offer radical different
answers to the questions in (2), but THEY AGREE ON THE QUESTIONS BEING
ASKED.
Like earlier research in generative grammar, research in computational
phonology agrees with the insight in (1) and the questions being asked in (2).
In this chapter we ask three derivative questions. In theories like SPE, which
posit morpheme structure constraints, what does the theory of computation
bring to the nature of these generalizations regarding underlying
representations? In theories like OT, which posit markedness constraints,
what does it bring to the nature of these phonotactic generalizations? And for
ALL theories of phonology, What does it bring to the study of the nature of
the transformations?
The theory of computation provides a way to answer these questions.
The encyclopedia of categories it provides allows these different
generalizations to be classified according to computational criteria. What
makes this approach valuable is that it is about as atheoretical as one can get.
This is because it explicitly separates the intensional descriptions of the
generalizations from their extensions. The intensional description of the
generalization is the one given by a phonologist in their grammatical
description of the generalization. It is the “finite means” in Humboldt’s
sense. The extension of this intensional description is one that typically
describes an infinite-sized object. It is the “infinite use” in Humboldt’s
sense. Mathematically, this infinitely-sized object exists. It is like a perfect
circle, a set of infinitely many points each exactly the same distance from a
center. But we can never see the object in its entirety. We cannot see an
infinity of points, even if we know they are there. The situation with
linguistic generalizations is similar. The extension is there, but they cannot
be written down in their entirety since they are not finite. But we can write
down a grammar which can be understood as generating the infinite set, in
the same way that a perfect circle can be generated by specifying a center
point and a distance, the radius.
The same perfect circle can be described in other ways as well. If we
employ the Cartesian plane, we could generate a circle with an equation of
the form (x − a)2 + (y − b)2 = r2 where the r is the radius of the circle and (a,
b) is its center. The equation is interpreted as follows: all and only points (x,
y) which satisfy the equation belong to the circle. The equation is an
intensional description and the set of points, the circle, is its extension.
We can also describe a circle on a plane with polar coordinates instead of
Cartesian ones. Recall that polar coordinates are of the form (r, θ) where r is
the radius and θ is an angle. The equation r = 2a cos(θ) + 2b sin(θ) provides
the general form of the circle with the radius given by √a +b 2
2

and the center by (a, b) (in Cartesian coordinates). The

polar equation is interpreted like the Cartesian one: all and only points (r, θ)
which satisfy the equation belong to the circle.
There are some interesting differences between these two coordinate
systems. Each point in the Cartesian system has a unique representation, but
each point in the polar system has infinitely many representations (since the
same angle can be described in infinitely many ways, e.g. 0° = 360° = 720°
= . . .). If the center of the circle is the origin, the polar equation simplifies to
r = a whereas the Cartesian equation remains more complicated x2 + y2 = r2.
Thus, the polar equation r = 4 and the Cartesian equation x2 + y2 = 16 are
different equations with different interpretations, but they describe the same
unique circle: one of radius four centered around the origin. The two
equations differ intensionally, but their extension is the same.
It seems strange to ask which of these two descriptions is the “right”
description of a circle. They are different descriptions of the same thing.
Some descriptions might be more useful than others for some purposes. It
also interesting to ask what properties the circles have irrespective of a
particular description. For instance the length of the perimeter and the area
of a circle are certainly relatable to these descriptions, but they are also in a
sense independent of the particulars. The perimeter and area depend on the
radius but not the center, though both appear in the equations. This suggests
that the radius is a more fundamental structure to a circle than its center,
though both certainly matter.
The analogy I wish to draw is that rule-based and OT-theoretic
formalisms are like the Cartesian and polar systems. The analogy is far from
perfect, but it is instructive. Both rule-based and OT analyses provide
descriptions of platonic, infinitely sized objects. In many cases, but not all,
the two formalisms describe the same object, insofar as the empirical
evidence allows.
What is this object? The transformations from underlying forms to
surface forms can be thought of as a FUNCTION, in the mathematical sense of
the word. Another word for function becoming prevalent in the phonological
literature is MAP (Tesar 2014). There are three parts to a function. One, there
is its domain, which is the set of objects the function applies to. Two, there is
its co-domain, which is the set of objects to which the elements of the
domain are mapped. Three, there is the map itself, which says which domain
elements are transformed to which co-domain elements. Thus to specify a
function, one needs to provide a description of its domain, its co-domain,
and a description of which domain elements become which co-domain
elements.
This lines up nearly perfectly with the fundamental questions of
phonological theory. The underlying representations correspond to the
domain. The surface representations are the co-domain. And the
transformation from underlying to surface forms is the map from domain
elements to co-domain elements. From this perspective, describing the
phonology of a language requires describing aspects of this function,
regardless of whether the function is described intensionally with SPE-style
or OT grammars.
Further, in linguistic typology we are actually interested in the CLASS of
such functions that correspond to possible human phonologies. If the
phonologies of languages are circles we would be interested in the universal
properties of circles and the extent of their variation. Circles are pretty
simple, so the answers are straightforward. All circles have a center and a
radius, but their centers can be different points and their radii can have
different lengths. What universal properties do phonological functions
share? What kind of variation does the human animal permit in this
function?
This is why computational approaches to language have much to offer.
Studying the extensions of constraints and transformations through the lens
of a computationally-grounded encyclopedia of categories helps us better
understand the nature of phonological component of grammar.

3 Representing constraints and transformations

Ultimately, phonological grammars represent the functions mentioned
earlier. However, unlike circles, phonologies are not described with single
equations; instead, phonological grammars contain multiple, interacting
parts. In OT grammars those parts are constraints. In rule-based grammars
those parts are rules. In this section, we put these intensions aside and
examine the extensions of phonotactic constraints and the extensions of
phonological transformations. Then in the next sections we examine the
computational nature of those extensions.

3.1 Phonotactic knowledge and markedness

constraints
Halle (1978) gives phonotactic knowledge as an example of knowledge that
is learned but not taught. He provides an experiment demonstrating this
knowledge, whose results are shown in Table 5.1. I have informally
conducted this experiment myself on dozens, if not hundreds of young
adults, who are native speakers of English. When presented visually with
orthographic representations of the words in Table 5.1 (but ungrouped),
student reliably and uniformly identify thole, plast and flitch as the English
words.77 Just as circles are fruitfully thought of as an infinite set of points,
phonotactic knowledge can likewise be thought of as an infinite set of
strings. All possible English words are in the set; all logically possible,
impossible words are out of the set. This is but one concrete way to see that
“language makes infinite use of finite means”; generative grammars allow us
to distinguish among infinitely many logically possible forms. One question
linguists address is: What is the nature of this infinite set?
Table 5.1: Words from Halle (1978).
possible English words impossible English words
thole ptak
plast hlad
flitch sram
mgla
vlas
dnom
rtut

Markedness constraints in OT express phonotactic knowledge.

Markedness constraints are said to prohibit marked structures so they
distinguish well-formed structures from ill-formed ones. We will consider
their extensions as follows: all surface forms with zero violations are in the
set; all surface forms with nonzero violations are out of the set (cf. McCarthy
2003). Therefore, the extensions of these constraints can be interpreted as all
and only those strings which are well-formed according to the constraint;
they are those structures which DO NOT CONTAIN the marked structure as a
sub-structure.
For example, consider the constraint ∗NC˳ (Pater 2001), which states that
nasals followed by voiceless consonants are marked sequences. The
extension of this constraint can be conceived as the set of strings not
containing marked structure, some of which are explicitly shown in (3).

(3) {a, b, aba, anda, anba, . . . }.

In fact every logically possible string which does not contain this marked
substructure is in the extension of the ∗NC˳ constraint. As will be discussed
in greater detail in Section 5.1, substrings like n‘t are sub-structures of
strings.
Another example comes from syllable structure. It is widely held that
codas are marked. Words with codas are said to violate the constraint
NOCODA. Thus the well-formed structures picked out by this constraint are
all and only those strings which do not contain codas as indicated by (4).

(4) {a, a.ba, pa.pa, . . . }

The representations in (4) differ from those in (3) because they include a
symbol for the syllable boundary. The available symbols, and the choice of
representation more generally, is an important issue, to which I will return at
the end of this chapter.

3.2 Transformations
Extensions of transformations can also be described as infinite sets. In this
case the elements of the set are pairs: the first element of the pair represents
the INPUT and the second element the OUTPUT. Such extensions have been
called MAPS by Tesar (2014) and others.
As an example, consider the SPE-style rule shown in (5), which
epenthesizes [ɨ] between stridents.

The extension of this rule can be interpreted as every pair of strings (i, o)
such that if i is the input to the rule o would be the output. The extension of
(5) is shown in (6).

(6) { (wiʃz, wiʃɨz), (dagz, dagz), . . . }

Here is another example. Consider the rule in (7), which devoices word-final
obstruents.

(7) [−sonorant] → [−voice] / __ #

This rule describes the infinite set of pairs, indicated in (8).

(8) { (rat, rat), (sap, sap), (rad, rat), (sab, sap), (sag, sat), (flugenrat,
flugenrat), (flugenrad, flugenrat), . . . }

OT descriptions of [ɨ]-epenthesis and word-final obstruent devoicing

describe the SAME extensions. Baković (2013, Chapter 4) shows how to
translate any rule of the form A→ B / C __ D into a core ranking where a
markedness constraint like ∗CAD outranks those faithfulness constraints
violated by A→B. As he explains, this ranking “is assumed to be embedded
within a constraint hierarchy” whose other constraints must also be ranked a
certain way.
In other words, the ranking in (9) is the core ranking necessary to
describe the pairs of strings in (6).

(9) ∗[+STRIDENT][+STRIDENT] >> DEP(ɨ)

Obviously, if there is a candidate which does not violate ∗[+STRIDENT]

[+STRIDENT] but violates some other faithfulness constraint F, then F must
outrank dep(ɨ). These and other constraints are part of the presumed
constraint hierarchy to which Baković refers.
Similarly, in order to describe the set of pairs in (8), the core ranking in
(10) must be embedded in the constraint hierarchy.

(10) ∗[+VOICE,-SONORANT]# >> ID(VOICE)

Both OT grammars and rule-based grammars can be used to describe the

same sets of pairs. In cases where they define the same extension, they are
like the polar and Cartesian systems which can describe the same circles
with different equations, which are interpreted differently according to the
system they inhabit.
Here we have focused on simple transformations — ones that in rule-
based theories could be described by a single rule. But phonologies in the
world’s languages are more complex than that. There are multiple,
interacting factors. Still, both OT grammars and rule-based grammars
ultimately generate pairs of strings like the ones in (6) and (8). They do this
in different ways, but they do it nonetheless. Furthermore, the way
phonology is taught, practiced and studied — both rule-based and constraint-
based theories — is exactly by examining fragments of grammars and
building up to larger and larger analyses.78 The approach here is no different.
Thus, the object of interest in both cases is these sets of pairs, which are the
transformations from underlying to surface forms. As discussed earlier in
(1), this is the basis for modern generative phonology. What is the nature of
these maps?

4 Expressivity and restrictiveness

Before continuing, I would like to emphasize a common typological goal of
every theory of phonology. A good theory must be both SUFFICIENTLY
EXPRESSIVE to accurately describe the actual phonologies in the world’s
languages and MAXIMALLY RESTRICTIVE. It is clear enough why expressive
adequacy matters. To be clear, by “sufficiently expressive”, I am referring to
theories that, for every natural language phonology P, provide a descriptively
adequate grammar for P. It may be less clear why maximal restrictiveness
matters, but it does and no less so.

4.1 Why restrictiveness matters

There are three reasons why restrictiveness matters. First, it helps address
the problem of how children quickly learn the phonology of their language
(so it helps us reach an explanatorily adequate theory of phonology; cf.
Chomsky 1965, Chapter 1).
Second, scientific hypotheses are stronger when they are more
restrictive. The hypothesis that outlaws the most logically possible
phonologies as humanly impossible can be said to be the strongest because it
is the most readily falsifiable (Popper 1959). For a restrictive theory, it is
possible to identify logically possible patterns, which would serve as a
counterexample to the theory, if in fact it were found in the phonology of
some language.
Third, it is easy to find a sufficiently expressive theory of phonology
which is not restrictive. The widely held Church-Turing thesis states that
anything that can be calculated or computed can be computed by Turing
machines (and equivalently Church’s lambda calculus). If phonologists
believe their theories and models of phonology are computable (no matter
how complex or intricate the computations) then there is already a
sufficiently expressive theory of phonology available. The problem with this
theory is that it is unrestrictive because it says everything that is computable
is possible. The mere existence of a phonology will never be sufficient
grounds for dismissing the Church-Turing Theory of Phonology. All of this
is a way of saying that a theory not only needs to explain WHAT THERE IS, BUT
ALSO WHAT THERE IS NOT.
The larger point is that expressiveness needs to be balanced against
restrictiveness. Failure to be sufficiently expressive does not automatically
disqualify a theory. It is enough for theories to be “nearly sufficiently
expressive” to be viable. I say this because sometimes theories, especially
when newly posited, are not sufficiently expressive (though they are very
restrictive). For example, the Copernican theory of the solar system was
originally not sufficiently expressive to predict the retrograde motion of the
planets, at least as compared to the best Ptolemaic models at the time. Still
the predictions were close. The Copernican theory was more restrictive
however (since fewer types of retrograde motions were possible with the sun
as the center of the system.) Closer to home, classic OT was not abandoned
simply because of its inherent inability to represent opaque maps (see
McCarthy 2008a; Baković 2007; Baković 2011 for discussion and examples
of opaque maps in phonology). Instead, subsequent research focused on
trying to modify the theory to make it more expressive.
Theories are rarely monolithic entities; they contain many parts working
together. Ultimately, restrictiveness matters for the whole theory, as the sum
of its parts, as opposed to the individual parts themselves. One aspect of the
theory may overgenerate in an unrestricted manner, provided some OTHER
COMPONENT of the theory excludes the problematic cases. Thus evaluating
some aspect of the restrictiveness of a theory is by no means a
straightforward affair: it means evaluation must occur with respect to other
components which can be said to plausibly exclude the problematic cases.
There are degrees of expressivity and degrees of restrictiveness. Once we
recognize the extensions of the constraints and transformations posited in
phonological theories are infinite sets and functions, then we will see that the
theory of computation naturally provides an encyclopedia of categories
which measures these degrees of expressivity and restrictiveness.
Furthermore, in this regard, the theory of computation is without peer.

4.2 The Chomsky Hierarchy

In this section, I will provide an overview of why the theory of computation
provides a valuable way to examine the expressivity and restrictiveness of
linguistic theories. Figure 5.1 shows the Chomsky Hierarchy which
classifies stringsets according to the kind of grammars that generates them.
Points in the space represent stringsets. The larger regions properly include
the smaller ones so for instance all regular stringsets are context-free but not
vice versa. As shown in the figure, linguistic generalizations (modeled as
stringsets) have been argued to belong to certain regions and not others
within the hierarchy.
(The issue of representation — whether we want to model linguistic
forms with string structures, tree structures, autosegmental structures, or
other kinds of graph structures is taken up in the discussion Section 8. There,
I will argue that even if string structures are left behind, computational
theory still provides an unmatched encyclopedia of categories for these other
structures, analogous to the ones I discuss here for strings.)
At the top of the hierarchy is the “computably enumerable” region which
includes everything. These are essentially the stringsets whose elements are
computable.79 This is the most expressive, but least restrictive class.

Figure 5.1: Natural language patterns in the Chomsky Hierarchy.

At the bottom of the hierarchy are the “finite stringsets”. These stringsets are
of finite cardinality. Unlike infinite sets, which require a generative grammar
to generate or recognize them, elements of finite sets can be listed. In
introduction to linguistics courses, we learn that linguistic generalizations
cannot be modeled with finite sets because there is no principled upper
bound on the length of possible words or sentences. The finite languages are
the most restrictive, but least expressive class. In between the computably
enumerable and finite classes are the regular, context-free and context-
sensitive regions.
An important aspect of the hierarchy is that several regions have
independently motivated, equivalent descriptions. Regular stringsets for
instance can be defined with monadic second order logical formulae, finite-
state acceptors, or regular expressions. Computer scientists Engelfriet and
Hoogeboom explain: “It is always a pleasant surprise when two formalisms,
introduced with different motivations, turn out to be equally powerful, as
this indicates that the underlying concept is a natural one. Additionally, this
means that notions and tools from one formalism can be made use of within
the other, leading to a better understanding of the formalisms under
consideration” (Engelfriet & Hoogeboom 2001: 216). At a high level of
abstraction, the different characterizations can be thought of as different
views on the same underlying object roughly in the same way different
equations in different coordinate systems can describe the same circle. The
more views we have, the better we can understand what it is we are looking
at.
Also, each region X describes a linguistic hypothesis: linguistic
generalizations must belong to X. Early work in generative grammar was
interested in establishing evidence for or against such hypotheses in order to
establish upper bounds on the nature of linguistic generalizations. The
weakest scientific hypothesis is that they are computably enumerable, which
is what I called the Church-Turing Theory of Phonology. As X moves down
the hierarchy, the hypotheses become stronger, so the claim that the weak
generative capacity of human syntax is a regular stringset is a strong
scientific hypothesis. However, it is generally considered to be false
(Chomsky 1956; Shieber 1985).

4.3 Phonology is Regular

The Chomsky Hierarchy is the best-known hierarchy in formal language
theory, but it is not the only one. In fact there are several other hierarchies,
some which only became well understood in recent decades, and others
which are still being formed. Also, the Chomsky Hierarchy in Figure 5.1
classifies stringsets, but hierarchies exist (and are being developed) for sets
of pairs of strings (relations/maps/ functions) as well. It is important to
distinguish hierarchies for one kind of set (e.g. stringsets) from another (e.g.
sets of pairs of strings).
An important region in a hierarchy for relations (sets of pairs of strings)
is also called Regular. It is called this because it shares much in common
with the regular class of stringsets. For instance, one way to define the
regular class of stringsets is with non-deterministic finite-state acceptors and
one way to define the regular class of string-to-string maps is with non-
deterministic finite-state transducers. Readers are referred to other texts for
more information about finite-state grammars (Sipser 1997; Beesley &
Kartunnen 2003; Roark & Sproat 2007; Jurafsky & Martin 2008; Hulden
2009).
The primary result in computational phonology to date is that the
transformations from underlying to surface forms — these phonological
maps — are in fact REGULAR. The argument goes something like this:
Optional, left-to-right, right-to-left, and simultaneous application of SPE-
style rules A→B/C―D (where A,B,C,D are regular stringsets) DESCRIBE
REGULAR RELATIONS, provided the rule cannot reapply to the locus of its
structural change (Johnson 1972; Koskenniemi 1983; Kaplan & Kay 1994).
Rule ordering is functional composition (finite-state transducer
composition). Regular relations are closed under composition (so the
composition of two regular relations is also a regular relation). Rule-based
grammars (finitely many ordered rewrite rules of the above type) can
describe virtually all attested phonological patterns. This does not mean
these grammars do so elegantly, that the rules correspond to psychologically
real constructs, or that they have any other desirable trait. It just means that
the input/ output map is describable with such a grammar.
The above argument constitutes significant evidence for the following
statement:

(11) (Regular Hypothesis) Phonological maps are regular relations.

If this is true, then it is true REGARDLESS of whether they are described with
SPE, OT, or other grammar formalisms! Here are some other ways of saying
the same thing:
– There are no non-regular phonological maps.
– A UNIVERSAL property of phonological maps is that they are regular.
Again, the fact that every rule-based grammar describes a regular relation, in
addition to the fact that there is no counterexample to the hypothesis that
phonological maps are regular, is strong evidence that the hypothesis in (11)
is correct.
One consequence of this result is that finite-state grammars become a
lingua franca for different phonological theories describing some aspect of
the phonology of a language. Hence in addition to the work mentioned
above which translates rule-based grammars into finite-state machines, there
exists much work which shows how to translate OT grammars into finite-
state machines (Frank & Satta 1998; Karttunen 1998; Gerdemann & van
Noord 2000; J¨ager 2002; Riggle 2004). Thus, for attested phonological
patterns — just as with circles — there at several ways we can describe
them. Those stringsets and maps can be described with rule-based grammars,
OT grammars, finite-state machines, and other tools (e.g. logical formulae).
Another consequence of (11) follows from a theorem by Scott & Rabin
(1959). This theorem establishes that the domain and image of regular
relations are regular sets of strings. This means the set of possible underlying
representations and the set of possible surface representations are also
regular. In other words, phonotactic knowledge and markedness constraints
describe regular stringsets. Or equivalently, every stringset defined by a
markedness constraint has the property of “being regular”.

4.4 The Subregular Hypothesis

“Being regular” is therefore plausibly a universal property of phonological
patterns (both stringsets and maps). Furthermore, it is restrictive: there are
many logically possible, NON-REGULAR patterns.
However, while “being regular” may be a necessary property, it is not
restrictive enough. There are many logically possible, regular patterns that
are still bizarre from a phonological perspective. We will encounter some of
these strange creatures shortly.
There are many interesting SUBREGULAR classes of stringsets, as shown in
Figure 5.2. Figure 5.2 shows a “close-up” view of the regular region shown
in Figure 5.1. This will constitute the “encyclopedia of categories” and it
will be explored in more detail in Section 5.2, though an overview will be
given here.
 The subregular hierarchies are bounded by the Regular region at the top
and Finite region at the bottom. A region higher up in the diagram which is
connected by a line to a region lower down in the diagram indicates the
lower region is a subset of the higher region. So every generalization in the
lower region is expressible in the higher one, but not vice versa.
There are two main branches in these hierarchies, the successor branch
(+1) and the precedence branch (<) (for now the Tier-Based Strictly Local
class can be ignored). The successor branch is also known as the Local
branch, and the precedence branch is also known as the Piecewise branch.
Along each branch, the regions are defined in logical terms. The Monadic
Second Order regions are the most expressive and least restrictive. This is
followed by (in order of decreasing expressivity/ increasing restrictiveness)
the First Order, Propositional regions and then the least expressive and most
restrictive regions, the Conjunction of Negative Literals. While the hierarchy
is presented here in logical terms these regions can also be defined in other
ways and have multiple characterizations just like the regions in the
Chomsky Hierarchy.

Figure 5.2: Subregular hierarchies of stringsets.

 The Subregular Hypothesis refers to the idea that phonological patterns
belong to small, well-defined regions of regular stringsets and maps. Thus
the term “Subregular Hypothesis” on its own does not say much because it
itself does not say WHICH subregular regions are at stake. To anticipate the
remainder of this chapter, we distinguish between a strong and weak
subregular hypothesis for constraints (a similar hypothesis will be put
forward for maps). The “strong” subregular hypothesis is that phonological
markedness constraints are Strictly Local and Strictly Piecewise (at the
bottom of the hierarchy). The “weak” subregular hypothesis is that they are
Tier-Based Strictly Local, which is a particular generalization of the Strictly
Local class (inspired by phonological tiers). Both of these hypotheses are
discussed explicitly in Sections 5.2.1 and 5.2.3.

5 Constraints
This section is devoted to markedness constraints. The primary purpose is to
describe the Subregular Hierarchies in Figure 5.2, which constitutes the
encyclopedia of categories, that I am arguing is important for understanding
the nature of markedness constraints in phonology. To help motivate the
discussion, and help make it more accessible, I will begin by discussing part
of an encyclopedia of types (the actual constraints found in natural
language).

5.1 The encyclopedia of types: Stringsets

In this section, I present some constraints known to be attested in the world’s
languages. These will be contrasted with constraints that are unattested. This
is not intended to be an exhaustive or comprehensive encyclopedia of types.
Only four types of markedness constraints are presented. This is intended to
be sufficient to motivate the encyclopedia of categories, presented
afterwards.

5.1.1 Four types of constraints

The first type of markedness constraint we encountered penalizes certain

contiguous sequences of sounds (substrings). The impossible English words
in Halle’s (1978) example all begin with illicit consonant clusters. We saw
that ∗NC˳ also penalizes substrings. This constitutes one kind of markedness
constraint.
There are other kinds of constraints employed by phonologists which
identify structures other than substrings as marked. For instance, if we were
to ask native speakers of Samala (Applegate 1972, 2007) about the words in
Table 5.2, they would reliably and uniformly distinguish them as shown in
the table. How do Samala speakers know which of these words belong to
different columns?80 Well, it appears that Samala speakers know that words
cannot contain both [+anterior] sounds like [s] and [−anterior] sounds like
[ʃ].81 Applegate (1972) modeled this knowledge as the result of a productive,
regressive sibilant harmony process. In OT, this knowledge would be a
consequence of a high ranking constraint of the form ∗[+strident,α anterior].
. . [+strident,−α anterior] (Hansson 2001; Rose & Walker 2004; Heinz
2010a). This constitutes a second type of attested markedness constraint in
the world’s languages.
Another logically possible type of constraint is illustrated with by
speakers of a language I will call “Language X” (its true identity will be
revealed momentarily), shown in Table 5.3. What constraint are the speakers
of this language utilizing to reliably distinguish the logically possible words
shown there? In this case, we observe speakers of Language X reject words
that begin and end in sibilants that disagree in the feature [anterior]. Unlike
the Samala example, sibilants interior to the word may disagree with edge-
bound sibilants as evidenced by possible words like [soʃokos]. But if there
are sibilants at word edges, they must agree in order for the word to be a
possible word of Language X. This type of constraint is distinct type of
constraint from the previous two types mentioned.
Table 5.2: Phonotactic knowledge in Samala.
possible Samala words impossible Samala words
ʃtojonowonowaʃ stojonowonowaʃ
stojonowonowas ʃtojonowonowas
pisotonosikiwat pisotonoʃikiwat
nasipisotonosikiwa naʃipisotonoʃikiwa

Table 5.3: Phonotactic knowledge in Language X.

possible words of Language X impossible words of Language X
ʃotkoʃ sotkoʃ
ʃoʃkoʃ ʃotkos
ʃosokoʃ ʃoʃkos
soʃokos soskoʃ
sokosos
pitkol
pisol
piʃol

Next we consider Language Y. Table 5.4 shows how speakers of this

language discriminate logically possible words. How do they do it? The two
columns in Table 5.4 are distinguished as follows. Possible words have an
EVEN NUMBER of sibilant sounds, but impossible words have an odd number
of sibilant sounds. So speakers of Language Y are sensitive to the even/odd
parity of the number of sibilant sounds. This constitutes a fourth type of
constraint, distinct from the ones mentioned earlier.
So far we have considered four logically possible kinds of constraints.
What is the actual typology, the encyclopedia of types?
The actual typology of course looks like this. Attested phonotactic
patterns include those which forbid substrings of words (such as ∗NC˳. They
also include ones where words don’t contain both sounds like ʃ and s (as in
Samala). However, the logically possible phonotactic patterns represented by
languages X and Y are unattested. There are no known phonotactic patterns
where the last sound in a word depends in some fashion on its first sound (as
in Language X). And there are no known phonotactic patterns where the
right generalization is that words must contain an even number of members
of a particular natural class (as in Language Y).
Table 5.4: Phonotactic knowledge in Language Y.
possible words of Language Y impossible words of Language Y
ʃotkoʃ ʃoʃkoʃ
sotkoʃ ʃoskoʃ
ʃotkos soʃkos
pitkol ʃoʃkos
soʃkostoʃ soskoʃ
soksos
piskol
piʃkol
5.1.2 Explaining the typology

We would like to have an explanation for this fact. We would like our theory
of markedness to explain why constraints like those found in English and
Samala are possible, but the ones found in Language X and Language Y are
not.
So what’s the explanation? In OT, constraints like ∗#mgl and
∗[+strident,α anterior] . . . [+strident,−α anterior] structures would be part of

CON. But constraints like ∗ODD-SIBILANTS or ∗#[+strident,α anterior]. . .

[+strident,−α anterior]# would not be. The explanation in OT largely comes
down to constraints that are present in CON and those that are absent from
CON. (Whether complex markedness constraints can be derived via
constraint interaction is a matter I take up later in Section 7.)
This is not controversial. The basic syllable typology is derived in OT by
including the constraints NOCODA,ONSET and excluding the constraints
NOONSET,CODA. If the constraints NOONSET,CODA were included in CON,
then it would not be possible to derive a typology where onsets may be
required (but are never forbidden) and codas may be forbidden (but are
never required).
In phonetically-based phonology (Hayes et al. 2004), the explanation
would be that there are perceptual and/ or articulatory reasons for constraints
like ∗#mgl and ∗[+strident,α anterior]. . . [+strident,−α anterior]. But there
would be no such reasons for constraints like ∗ODD-SIBILANTS or #
[+strident,α anterior]. . . [+strident,−α anterior]#. More generally, this
research program hypothesizes that constraints in CON are based on
phonetic principles, and that certain rankings of these constraints are also
fixed according to these principles. This is a significant improvement over
stipulating certain constraints as belonging to CON to the exclusion of
others.
However, we should carefully examine the proposed perceptual and/ or
articulatory reasons. Consider the case of the pattern in Language X.
Following Lai (2012, 2015), let us now refer to this pattern as First/ Last
Harmony. We know long-distance assimilation is well-attested (Hansson
2001; Rose & Walker 2004) and arguments have been made for its
perceptual basis (Gallagher 2010). We also know word edges in phonology
are privileged positions (Fougeron & Keating 1997; Beckman 1998; Endress
et al. 2009). So what theory of perception or articulation prevents there from
being harmony only in privileged positions?82 This is a case where the
phonetic principles seem to overpredict the attested typology.

Figure 5.3: Pattern templates for Sibilant Harmony (left) and First/Last
Harmony (right).

We may also wonder to what extent memory requirements could explain the
difference between the attested pattern in Samala and First/ Last Harmony.
In fact, however, it comes down to the pattern type, or template. This is
because both types can be described simply by marking which pairs of
sounds are permitted or forbidden in a given template as shown in Figure
5.3. The 2x2 cells for are identical — it is only the templates that differ.
As for the pattern in Language Y, it is plausible that perception or
articulation should be able to explain the absence of even/odd parity
constraints (or more generally constraints which count mod n) in phonology,
but I haven’t seen any explicit connection. Whatever the explanation may
be, it SHOULD connect to the computational properties discussed here. More
generally, if phonology is truly reducible ENTIRELY to phonetic principles
then there ought to be research showing how the computational laws being
posited in this chapter can be clearly derived from such phonetic principles.
This is not meant to deny any role to phonetic explanation in phonology.
Instead this discussion is intended to make clearer some of the limits of
those explanations and to persuade researchers in those areas that the
computational principles discussed here are worth connecting their work to.
At the very least, a complete theory of of phonology will refer to phonetic
factors IN ADDITION TO the computational principles discussed here. I return
to this issue in Section 7.3.
The computational explanation offered in this chapter is simply this. The
extensions of constraints on substrings (like ∗NC˳) and constraints on
subsequences (like [+strident, α anterior]. . . [+strident, −α anterior] in
Samala) are Strictly Local and Strictly Piecewise stringsets respectively.
With the exception of the finite languages, these are the most restrictive,
least expressive regions in the Subregular Hierarchies shown Figure Table 5.
2. On the other hand, First/Last Harmony and ∗ODD-SIBILANTS belong to the
Locally Testable and Regular regions, respectively. In other words, the
widely-attested constraints are the formally simple ones, where the measure
of complexity is determined according to these hierarchies.

5.2 The encyclopedia of categories

In the last section, extensions of phonological constraints and
transformations were introduced and it was argued that conceiving them as
stringsets (formal languages), and string-to-string maps was reasonable and
potentially insightful. In this section, we explain what the theory of
computation has to say about this cast of characters. To do so requires some
technical details. I will try to keep them light and only provide a sketch.
Readers interested in the all the details are referred to Rogers et al. (2013),
which also provides cognitive interpretations of the different regions.
We are going to explore LANGUAGE-THEORETIC and LOGICAL descriptions
of stringsets and string-to-string maps from a generative perspective. The
word “language” in “language-theoretic” refers to a formal language, what
we have been calling a stringset. We may as well call it “stringset-theoretic”.
The idea behind language-theoretic descriptions is that they are completely
independent of any grammatical description. In other words, these
descriptions are statements that are simply true of the stringset itself (and not
the grammar that generates it). They can therefore be thought of as essential
properties of the stringsets. Examples will be provided shortly.
Logical descriptions are not agrammatical. Logical formulae are
grammars in the sense that they generate stringsets as extensions. However,
they are useful here because the expressive power of different logics is well
understood. In the encyclopedia of categories — the Subregular Hierarchies
— presented in Figure 5.2, there were four types of logic. As mentioned
earlier, in order of strictly increasing expressive power, they are:
conjunctions of negative literals (CNL), propositional logic (P), first order
logic (FO), and monadic second order logic (MSO). The type of logic forms
one parameter that is used to define the regions in the Subregular Hierarchies
(Rogers & Pullum 2011).
There is one other parameter used to define the regions. This parameter
specifies the kind of structures used to model strings. The parameter
specifies the kind of relation used to handle the order of elements in the
string. The relations that have been studied are SUCCESSOR (+1) and
PRECEDENCE (<).
If strings are modeled so that the order of elements is handled by the
successor relation, then SUBSTRINGS will be sub-structures of strings. On the
other hand, if strings are modeled so that the order of elements is handled by
the precedence relation, then SUBSEQUENCES will be sub-structures of strings.
As an example, with the successor relation, bcc is a sub-structure of abccab,
but with the precedence relation, aab is a sub-structure of abccab.
We will now examine the consequences of the ordering relation in terms
of the logical power.

5.2.1 Conjunctions of Negative Literals

Stating constraints as a “negative literal” is logical talk for what

phonologists simply call marked structure. When conjunctions of such
constraints are considered, we define stringsets which do not contain any of
the marked structures. The order parameter (successor or precedence) tell us
whether to interpret literals (the string structures) as substrings or
subsequences.
Let us begin with successor, so the literals are interpreted as substrings.
The negative literal ¬aa is thus interpreted to mean the substring aa is a
marked structure. So any string containing this marked structure violates the
constraint and is not in the extension of the constraint.
Here is an example. I will use φ to stand for logical formulae, and L (φ)

to stand for the stringset extension of φ. Following Rogers et al.

(2013), I also will use ⋊ and ⋉ for the left and right word boundaries. The
the formula below can be read as “Strings which do not begin with a b, do
not contain aa as a substring, do not contain bb as a substring, and do not
end with an a, are well-formed”.

φ = (¬ ⋊ b) ∧ (¬aa) ∧ (¬bb) ∧ (¬a⋉)

The extension L (φ) is easy to write since conjunction is
interpreted as set intersection. A word about notation: Σ is a finite set of
symbols (the alphabet); Σ∗ means all logically possible strings one can write
with this alphabet; and S means the complement of stringset S with respect
to Σ∗. Thus, a term by term translation of φ above into its extension is shown
below.

L (φ) = ¯
¯¯¯
¯¯¯ ∩¯¯
¯¯¯¯¯¯
¯¯¯
¯¯¯ ∩¯¯¯
¯¯¯¯¯¯
¯¯¯¯ ∩¯
¯¯¯
¯¯¯
∗ ∗ ∗ ∗ ∗ ∗
bΣ Σ aaΣ Σ bbΣ Σ a

It is not difficult to see that this is the same as the infinite set {ab, abab,
ababab, . . .}.
So now we can provide one definition of the Strictly Local stringsets. A
Strictly k-Local (SLk) stringset is one which can be defined as the
conjunction of negative literals, where the literals are interpreted as
substrings, and whose longest forbidden literal (substring) is of length k. The
Strictly Local stringsets are those that are SLk for some k.
If the order relation is precedence, then the literals are interpreted as
subsequences. The negative literal ¬aa is thus interpreted to mean the
subsequence aa is a marked structure. So any string containing this marked
structure violates the constraint and is not in the extension of the constraint.
Here is an example. The formula below can be read as “Strings which do
not contain an a followed by an a nor a b followed by a c are well-formed”.
So here the literals aa and bc are interpreted as subsequences, and not as
substrings.

φ = (¬aa) ∧ (¬bc)

A term by term translation of φ above into its extension is shown below.

 L (φ) = ¯¯¯
¯¯¯¯¯
¯¯¯
¯∗¯
¯¯¯
¯¯¯ ∩¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯
∗ ∗ ∗ ∗
Σ aΣ aΣ Σ bΣcΣ

Strictly Piecewise stringsets are defined analogously to Strictly Local

stringsets. A Strictly k-Piecewise (SLk ) stringset is one which can be
defined as the conjunction of negative literals, where the literals are
interpreted as subsequences, and whose longest forbidden literal (sub-
sequence) is of length k. The Strictly Piecewise stringsets are those that are
SPk for some k.
That many attested markedness constraints define SL and SP stringsets is
not in dispute. Clearly, constraints like ∗NC˳ are SL and constraints like
∗[+strident, α anterior]. . . [+strident, −α anterior] are SP. The strong
subregular hypothesis states that all markedness constraints are either SL or
SP (Heinz 2010a).

(12) (Strong Subregular Hypothesis) Markedness constraints are SL or SP.

The one notable outstanding case Heinz (2010a) discusses is the set of
surface forms derived from long-distance dissimilation. These appear to be
Non-Counting but do not belong to any lower class (hence they are called
‘Properly Non-Counting’) (Heinz et al. 2011).83 They are discussed further
below.
Whether constraints like Onset are SL or not depends on the choice of
representation. If syllable boundaries are included in string representations,
which is a common practice, then constraints like Onset are SL since they
can be represented this way: (¬ .V). The importance of representations will
be further discussed in Section 8.
I would like to conclude the discussion of the “Strict” classes by
providing their language-theoretic characterizations. This characterization
for SL stringsets is provided in (13), which Rogers & Pullum (2011) name
Suffix Substitution Closure.

(13) (Suffix Substitution Closure) A stringset L is SL if there is a k such

that for all strings u1, v1, u2, v2, x with the length x equal to k − 1, it is
 the case that if u1xv1 and u2xv2 belong to L then u1xv2 belongs to L as
well.

Suffix Substitution Closure is ultimately a Markovian principle: the well-

formedness of the next symbol in the string depends only on the previous k −
1 symbols (given as x above). So if both v1 and v2 can follow x, what comes
before x (u1 and u2) does not matter. This Markovian notion is important in
generalizing SL stringsets to SL functions discussed in Section 6.
On the other hand, subsequence closure characterizes the SP class
(Rogers et al. 2010).

(14) (Subsequence Closure) A stringset L is SP if for all u belonging to L,

every subsequence of u also belongs to L.

What is remarkable about the language-theoretic characterizations above is

not only that stringsets which have these properties are exactly the ones that
can be defined as the conjunction of negative literals, but also that these
characterizations DO NOT MENTION ANY SORT OF GRAMMAR AT ALL. In this way,
they are more like definitions of circles which do not refer to either the
Cartesian or polar coordinate system. They are characterizations which
speak directly to the nature of the stringsets without getting bogged down in
any particular grammatical formalism. Language-theoretic characterizations
are about the SHAPE of the language to which any grammar must conform
itself.
Another remarkable fact about these characterizations is that they
immediately suggest INFERENCE procedures. If one is observing words from a
language L THAT IS A PRIORI KNOWN to be SLk and one observes u1xv2, u2xv2,
one can immediately deduce that u1xv2 also belongs to L. Similarly, if one is
observing words from a language L THAT IS A PRIORI KNOWN to be SP, and one
observes the word u, one can immediately determine all subsequences of u
also belong to L.
These facts lay at the basis of learning algorithms developed for the SLk
and SPk classes. Garcia et al. (1990) first proved that the SLk stringsets are
identifiable in the limit from positive data. I have argued elsewhere that
IDENTIFICATION IN THE LIMIT FROM POSITIVE DATA is a rigorous and insightful
learning paradigm (Heinz 2016), and I will not review the arguments here.
Heinz (2010a) shows how a similar algorithm provably identifies SPk
languages in the limit from positive data, and Heinz et al. (2010b)
generalizes these ideas to a family of learning algorithms, a result which was
generalized even further by Heinz et al. (2012).
If the strong subregular hypothesis is correct, these learning results
provide a deep explanation of it. Constraints on phonological well-
formedness are SL and SP because people learn phonology in the way
suggested by these algorithms. More specifically, people generalize in
accordance to inference procedures suggested by the closure properties in
(13) and (14). But it is the inference procedures themselves that are basic
which structure the SLk and SPk classes; the inference procedures are not
auxiliaries to the classes.

5.2.2 Propositional Logic

Next we move up one level to the next kind of logic: propositional. Unlike
the conjunction of negative literals, where all formulae had the form (¬l1) ∧
(¬l2) ∧ . . . ∧ (¬ln) for n literals (li), propositional logic allows any well-
formed propositional formulae to generate a stringset. Not only is any
combination/ordering of negation and conjunction now permitted, but
disjunction (∨) is also allowed. As a consequence, mainly familiar
propositional connectives are also allowed, such as implication (→) and the
biconditional (↔). Propositional logic is therefore more expressive (and less
restrictive) than the conjunction of negative literals.
For example, the following formula is a well-defined formula in
propositional logic.

φ = b ∨ (aa → ac)

If these literals are interpreted with respect to the successor model of strings,
then this formula translates to the following English: “Words are well-
formed if they contain the substring b or if it is the case that if they contain
the substring aa they also contain the substring ac.” Below I provide the
extension of φ under the successor interpretation of the literals.
 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗
L (φ) = Σ bΣ ∪ (Σ aaΣ acΣ ∪ Σ acΣ aaΣ )

If these literals are interpreted with respect to the precedence model of

strings, then this formulae translates to the following English: “Words are
well-formed if they contain the subsequence b or if it is the case that if they
contain the subsequence aa they also contain the subsequence ac.” Here is
the extension of φ under the precedence interpretation of the literals.
∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗
L (φ) = Σ bΣ ∪ (Σ aΣ aΣ cΣ ∪ Σ aΣ cΣ aΣ )

I submit that both of these logically possible constraints seem more odd from
a phonological perspective than the SL or SP constraints. At first glance, it
seems strange to have a markedness constraint which requires that if one
sub-structure is present another one must be present as well.
This is perhaps the most notable difference between the kinds of
constraints permitted using propositional logic. Such constraints can require
sub-structures to be present in well-formed words (Rogers & Pullum 2011).
The interpretation of the simple formulae φ = b is that well-formed words
must contain the sub-structure b. Such examples exist in the phonological
literature. For instance, it is true that the constraint ONSET has this flavor. As
we have mentioned with ONSET, however, the choice of representation
matters: this can be construed as SL provided syllable boundaries are
introduced as symbols in strings. Another constraint like this is what Hyman
(2009) calls Obligatoriness, the requirement that all well-formed words bear
an accent (or stress). Unlike ONSET, there is no straightforward
representational “fix” for this constraint. I return to this issue in Section 5.3.
 Now we can provide one definition of the Locally Testable stringsets. A
Locally k-Testable (LTk ) stringset is one which can be defined with a formula in
propositional logic, where the literals are interpreted as substrings, and whose
longest literal (substring) is of length k. The Locally Testable stringsets are those
that are LTk for some k.
Similarly, a definition of the Piecewise Testable stringsets can be given. A
Piecewise k-Testable (PTk ) stringset is one which can be defined with a formulae
in propositional logic, where the literals are interpreted as subsequences, and
whose longest literal (subsequence) is of length k. The Piecewise Testable
stringsets are those that are PTk for some k.
There are language-theoretic characterizations of these classes too. This
characterization is given in (15) for the Locally Testable class.

(15) (Substring Equivalence) A stringset L is LT if there is a k such that for all

strings u and v, if u and v have the same set of substrings of length k then
either both u and v belong to L or both u and v do not belong to L.

In other words, Substring Equivalence means that membership in a LT stringset L

only depends on the set of substrings of some length k. If two distinct strings have
the same substrings up to some length k then no LTk stringset is able to distinguish
them.
A similar characterization is given in (16) for the Piecewise Testable class.

(16) (Subsequence Equivalence) A stringset L is PT if there is a k such that for all

strings u and v, if u and v have the same set of subsequences of length k then
either both u and v belong to L or both u and v do not belong to L.

Subsequence Equivalence means that membership in a PT stringset L only

depends on the set of subsequences up to some length k. If two distinct strings
have the same subsequences of some length k then no PTk stringset is able to
distinguish them.
Like the characterizations for the “Strict” classes, these characterizations
naturally suggest inference procedures. If one is observing words from a language
L THAT IS A PRIORI KNOWN to be LTk and one observes u, one can immediately
deduce that all words with the exactly the same substrings up to length k also
belong to L. Similarly, if one is observing words from a language L THAT IS A
PRIORI KNOWN to be PTk, and one observes the word u, one can immediately
determine all words with the exactly the same subsequences up to length k also
belong to L.
 That the PTk and LTk stringsets are identifiable in the limit from positive data
was established by Garcı´a & Ruiz (2004). Heinz et al. (2011) and Heinz et al.
(2012) show there are learning algorithms for these classes which have much in
common with the ones for the Strict classes. An interesting difference, however,
between these algorithms and the ones for the Strict classes has to do with time-
complexity: there is a clear computational sense in which learning these more
expressive classes takes significantly longer than learning the Strict classes.

5.2.3 First Order Logic

The next rung up the logical hierarchy brings us to First Order (FO) Logic. The
main differences between first order logic and propositional logic is that literals
disappear and variables appear. It is not necessary in this chapter to provide the
technical details regarding FO models of strings. For this, readers are referred to
Rogers et al. (2013).
There are only three important items readers need to to understand. First, FO
logic is strictly more powerful logic than Propositional logic. Second, as usual,
whether the ordering relation is given as the successor relation or the precedence
relation will determine the kinds of stringsets expressible with FO formulae. FO
logic with the successor relation yields the class called the Locally Threshold
Testable (LTT) class, and FO logic with the precedence relation yields the class
called Non-Counting (NC). Third, successor is FO-definable from precedence, but
not vice versa so the Non-Counting class properly includes the LTT class.
I will go straight to the language-theoretic properties. If the ordering relation is
the successor, then the class of stringsets that is FO-definable is called the Locally
Threshold Testable (LTT) class, and it properly includes the LT class.
One important difference between the FO-definable classes and the
Propositional-definable classes is that the FO-definable classes are able to
distinguish the presence of otherwise identical sub-structures. In this way, FO-
definable classes can count the number of sub-structures up to some threshold. On
the other hand, the Propositional classes can only detect the presence or absence of
sub-structures. So for a given sub-structure, Propositional logic can distinguish
zero of them from one of them. FO logic, however, can detect up to some number
n of sub-structures. So a limited ability to count is present at the FO-level. There is
always some finite number n after which the number of substructures cannot be
distinguished. FO-definable classes are not sufficiently expressive to be able to
count indefinitely. Thus the difference between LTT and LT is that in the LTT
class, the number of substrings can be counted, but only up to some threshold t
(Thomas 1997).
(17) (Substring Threshold Equivalence) A stringset L is LTT if there is a k and a t
such that for all strings u and v, if u and v have the same number, up to some
threshold t, of substrings of length k then either both u and v belong to L or
both u and v do not belong to L.

By now the reader may expect that language-theoretic characterization of FO-

definable classes with the precedence relation is similar except that the number of
subsequences up to some threshold is distinguishable. It is, however, in fact much
simpler than that.

(18) (Non-Counting) A stringset L is NC if there is a k such that for all strings u,

x, v if uxkv belongs to L then so does uxk+1v.

The reason for this is that the Non-Counting class can do much more than count
subsequences. This is partly because the successor ordering relation is FO-
definable from precedence, but not vice versa.84 Consequently, every stringset in
the LTT region is also in the Non-Counting region, but not vice versa. NC is
strictly more expressive than LTT. Thus, Figure 5.2 shows that the the NC class
properly includes the LTT class. McNaughton & Papert (1971) comprehensively
establish several other important characterizations of the Non-Counting class.
Are there markedness constraints that count up to some threshold? An example
of such a constraint would be something like ∗3NC˳ where words with zero, one or
two NC˳ substrings are considered well-formed, but words with three or more are
ill-formed. Needless to say, such constraints do not seem like the kinds of
constraints found in natural language.
On the other hand, there are constraints in natural language that have been
argued to be properly Non-Counting. These are the stringsets that are definable
from long-distance dissimilation (Heinz et al. 2011). Heinz et al. (2011) also show
that such constraints belong to subclasses of the Non-Counting region they call
Tier-based Strictly Local (TSL). These stringsets are defined with the common
notion of phonological tier (Goldsmith 1976). Like the Strictly Local class, TSL
stringsets can be defined with formulae that are conjunctions of negative literals,
interpreted under the successor relation after non-tier elements are ignored. Thus
the kind of long-distance behavior is limited in some kind of way. TSL stringsets
are not as well understood as the other classes (there are not multiple
characterizations), but Heinz et al. (2011) argue that every markedness constraint
in natural language is describable with TSL constraints. Of course an important
issue is here what the tier is. Jardine & Heinz (2016) show that the tier can be
identified from positive data when the bound k on the size of the constraints are
known a priori (the tier is not known a priori).
 I will refer to the hypothesis that all markedness constraints are TSL as the
weak subregular hypothesis.

(19) (Weak Subregular Hypothesis) Markedness constraints are TSL.

Whether the evidence favors the strong or weak subregular hypothesis will be
addressed in Section 7.

5.2.4 Monadic Second Order Logic

The next rung up the logical hierarchy and the highest to which we attend is
Monadic Second Order (MSO) Logic. The difference between first order and
monadic second order logic is that variables over sets of elements in the domain
are allowed in addition to the variables which vary over individual elements
(which FO logic allows). There are several interesting consequences of adding
such variables, which I will now review.
First, the two branches in the subregular hierarchies merge at this point
because precedence is MSO-definable from successor.85 So the stringsets that are
MSO-definable with successor are exactly the stringsets that are MSO-definable
with precedence.
Second, this class of stringsets corresponds exactly to the class of stringsets
definable with finite-state acceptors, i.e. the regular class of stringsets (Bu¨chi
1960).
Third, this class is strictly more expressive than both the Non-Counting and
Locally Threshold Testable class (McNaughton & Papert 1971; Thomas 1997). It
can be shown that the stringset defined by the constraint ∗ODD-SIBILANTS (see Sec
tion 5.1) is not Non-Counting, but it is a regular stringset.86

5.3 Further evidence supporting the Subregular

Hypotheses
So far in this section, we mentioned the most common types of attested
markedness constraints. We did not provide an exhaustive encyclopedia of types
in Humboldt’s sense, but enough of one to motivate the encyclopedia of categories
that was presented. The discussion was designed to convince readers that the
markedness constraints found in natural language were present at the lowest levels
of the hierarchy and that as one moves up the hierarchy, the kinds of constraints
describable at these higher levels become less and less natural from a phonological
point of view.
 This helped motivate two hypotheses. The Strong Subregular Hypothesis (12)
says that markedness constraints are SP or SL. The Weak Subregular Hypothesis
(19) says that markedness constraints are TSL.
While these constraints were motivated by appealing to common types of
constraints, readers may wonder whether the hypotheses have been subjected to
more rigorous empirical investigation. I would like to now give further evidence
for the Strong and Weak Subregular hypotheses. First I will discuss studies of
stress patterns in terms of the Subregular Hierarchies.
Jim Rogers and his students examined the stress patterns in the stress typology
in Heinz (2007, 2009) with respect to the Strictly Local languages. There are 109
distinct patterns in this typology from over 400 languages. Edlefsen et al. (2008)
report that 72% (of the 109 patterns) are SLk with k ≤ 6 and 49% are SL3. The
28% are which are not SL6 are unbounded stress patterns and are shown to not be
SL for any k. Heinz (2014) studies the four simplest types of unbounded stress
patterns and shows that these are SP2 once Culminativity (every word contains
exactly one stress) is factored out. Culminativity has been argued to be a universal
property of stress languages (Halle & Vergnaud 1987; Hayes 1995) and therefore
this LT constraint may be thought to come for free. While recently Hyman (2009)
suggests a more nuanced view, it seems ill-advised at this point to view the result
regarding Culminativity in Heinz (2014) as a rejection of the Sub-regular
Hypotheses. Rogers et al. (2013) argue that the other unbounded stress patterns
similarly factor into the conjunction of SL and PT constraints or SP and LT
constraints. The other unbounded stress patterns continue to be the subject of
current research.
Two potential counterexamples come from work by Thomas Graf. Graf
(2010a) provides a formal analysis of the stress patterns of Creek and Cairene
Arabic, as they have been characterized in the literature. According to Graf’s
analysis, these stress patterns are not Non-Counting and are properly regular. If the
posited linguistic generalizations are correct, by Graf’s analysis, these cases would
constitute clear counterexamples to the Subregular Hypotheses. A critical aspect
of the linguistic generalizations that Graf’s result relies on is that there is no
secondary stress in these languages. If the secondary stress were perceptible,
however, the constraints needed to describe the pattern would become SL.
Whether secondary stress is perceptible or not to speakers of these languages is
not a settled issue, and so there is some question regarding the accuracy of the
linguistic generalizations.
Thus, with only a couple of potential counterexamples meriting further study,
the current understanding of the stress typology supports the Strong and Weak
Subregular Hypotheses.
 A second source of evidence in favor of these hypotheses comes from
psycholinguistic experimentation (Lai 2012, 2015). In a series of artificial
language learning experiments, Lai compared how well native English speaking
young adults could internalize the phonotactic pattern expressed by a SP constraint
like ∗[+strident, α anterior]. . . [+strident, −α anterior] and a LT constraint like
First/ Last Harmony (∗#[+strident, α anterior]. . . [+strident, −α anterior] #).
Subjects in these experiments participated in a training session followed by a test
session. In the training session, they are told they are going to hear the words of a
foreign language. In the test session, they are given two words and are asked
which one more likely belongs to the language they just heard.
Subjects belonged to one of three conditions, which determined the kind of
training received. In the “Sibilant Harmony” (SH) condition, they were exposed to
words which were well-formed according to the constraint ∗[+strident, α anterior].
. . [+strident, −α anterior]. In the “First/Last Harmony” (FL) condition, they were
exposed to words which were well-formed according to the First/ Last Harmony
constraint. Subjects in the control condition received no training (and during test
were asked which word they thought was a better word).87 All subjects were given
the same test items in the test session.
As reported in Lai (2015), the results of this experiment were unambiguous.
As expected, subjects in the control condition behaved according to chance. In the
test session, subjects in the SH condition behaved in a manner consistent with
internalizing the SP constraint because they consistently chose words in the test
session that did not violate the constraint. On the other hand, subjects in the FL
condition did not consistently choose words in the test session that did not violate
the constraint. In fact, they behaved just like subjects in the SH condition! This is
despite the fact that they were exposed to words like [soʃos] in training, which
violate the constraint ∗[+strident, α anterior]. . . [+strident, −α anterior].
In sum, the experiments of Lai (2012, 2015) show that subjects find it easier to
learn SP (or TSL) stringsets as opposed to LT ones. This evidence is consistent
with both the Strong and Weak Subregular Hypotheses, but as far as I am aware,
no other theory explains these results.

5.4 Constraints: A summary

A summary of the foregoing section can be made very simply. Phonologists have
identified many kinds of constraints on string representations. Stringsets can be
classified according to two core computational parameters: the type of ordering
relation (successor or precedence) and the type of logical power. Together, these
provide a Constraint Definition Language in the sense of de Lacy (2011). With
only a few potential exceptions meriting further empirical investigation, the
stringsets corresponding to phonological constraints overwhelmingly belong to the
SL, SP, and TSL regions in the encyclopedia of categories shown in Figure 5.2,
which are arguably the simplest.
Readers may wonder whether it is necessary to allow both ordering relations to
be sub-structures in words.88 If only the successor relation is permitted, then it is
not possible to describe constraints like *[+strident, α anterior]. . . [+strident, −α
anterior] at the CNL, P, and FO levels. MSO logic is needed. MSO logic with
successor, however, permits any regular stringset to be described. Similarly, if
only the precedence relation is permitted, it is not possible to describe constraints
like ∗NC˳ at the CNL, and P levels. FO logic is needed. In other words, the most
restrictive theory is the Strong Subregular Hypothesis (12): phonological
constraints are defined by banning substrings or subsequences.
Finally, we may wonder why this would be the case. If the Strong Subregular
Hypothesis is correct, then the extensions of synchronic constraints are SL or SP
stringsets (or conjunctions thereof). Why would this be? The idea expressed in
Heinz (2010a) is that human learners generalize in particular ways — and the
ways they generalize yield exactly these classes. Synchronic constraints are
aspects of grammar, and grammars are learned. They are systems that grow and
develop in response to environmental stimuli. The learning biases structure the
classes; it is not the case that the nature of the class is independent of the learners.
As Dresher (1999) and Heinz (2009) argue, this kind of explanation is not
available to learners within OT settings.

6 Transformations
Now we turn to transformations. From an OT perspective, this section is about
faithfulness constraints and the map derived from the interaction of all the OT
constraints. It is also about the typology of maps generated from a given CON.
From a rule-based perspective, this section is about the extensions of individual
phonological rules and their composition.
The computational theory of subregular relations is not as well developed as
the Subregular Hierarchies. For example, logical characterizations of string
relations have not yet been fully carried out. Previous work on subclasses of
subregular relations is primarily limited to two classes known as the LEFT
SUBSEQUENTIAL and RIGHT SUBSEQUENTIAL. Essentially, these are classes of
transformations with finite look-ahead; so they are “myopic” in the sense of
Wilson (2003).
 More will be said about these classes momentarily. I will keep the discussion
at a high level and readers can find the definitions of them and other technical
details in many different places, including Berstel (1979), Mohri (1997), Roche &
Schabes (1997), Lothaire (2005), Sakarovitch (2009). A linguistically motivated
treatment is given in Heinz & Lai (2013).
Much recent work at the University of Delaware has sought to develop a
hierarchy for string relations that is analogous to the one for stringsets shown in Fi
gure 5.2. The most notable advance in this regard has been work by Jane Chandlee
(Chandlee 2014; Chandlee et al. 2014; Jardine et al. 2014; Chandlee & Heinz
2018), which establishes relational counterparts to the Strictly Local stringsets,
discusses their significant coverage of empirical phenomena, and explains how
they can be learned.

6.1 The encyclopedia of types: Maps

Phonologists are familiar with many ways in which underlying forms can differ
from surface forms. Underlying segments may be deleted. Their order may be
permuted (metathesis). The features composing the segments may change.
Additionally, there may be elements in surface forms which were not present
underlyingly (epenthesis).
Additionally, the contexts that trigger these changes are of different types. The
contexts may be local to where the changes occur, by which I mean the distance
between the trigger and the target falls within some fixed bound. A typical
example is where the triggering context is adjacent to the change. For instance
regressive nasal place assimilation is typically written in rule format as [+nasal] →
[αplace] / ___ [−sonorant, α place]. Alternatively, the triggers can be found
arbitrarily far away, as found in examples of long-distance consonantal harmony
(Hansson 2001; Rose & Walker 2004) and disharmony (Suzuki 1998; Bennett
2013). Unlike the local cases, there appears to be no fixed bound on the distance
between the trigger and the target.
The long-distance cases are of special interest, so I will largely follow the
analysis of vowel harmony from Heinz & Lai (2013) to motivate the encyclopedia
of categories introduced in the next section. Vowel harmony is a well-studied
phenomenon in phonology (van der Hulst & van de Weijer 1995; Baković 2000;
Finley 2008; Nevins 2010; Walker 2011, and many others). Vowel harmony refers
to a systematic pattern of pronunciation in which certain features of vowels which
are different at the underlying level are the same at the surface level. Thus vowel
harmony has been called a process of assimilation. One reason it has attracted
interest is because the affected vowels are not strictly speaking adjacent since
consonants may intervene between them.
Schematically, an example of harmony can be distilled to the following
mapping: /+−/↦[++]. We will read this mapping as follows. The underlying form
and surface form each contain two vowels. The mapping shows the values of the
feature F for each vowel. Virtually all features defining vowels (roundness, height,
backness, advanced/retracted tongue root) have been shown to participate in
harmony in some language, and F can be understood to be any of these features.
At the underlying level, the values of the first and second vowel are /+/ and /−/,
respectively. At the surface level, however, both vowels bear the value [+] for
feature F. Since consonants are irrelevant, they are not shown, but it should be
understood that the mapping above includes consonants which may precede or
succeed any of the vowels.
Heinz & Lai (2013) analyze six types of logically possible vowel harmony
maps discussed in the phonological literature. These will be discussed, along with
two other types, all of which are summarized in Table 5.5.
Following terminology introduced in the OT literature, I will refer to faithful
vowels as those whose value of feature F stays constant in the underlying and
surface forms. Unfaithful vowels are ones whose value of the feature F is not
constant across the two levels.
One logically possible — but unattested — VH map maintains that the
unfaithful vowels are always fewer in number than the faithful ones. (In words
with an equal number of faithful and unfaithful vowels, which feature values
would change would be determined according to some default.) This map has been
called been Majority Rules (MR) harmony. (20) shows some examples of this MR.

(20) (Majority Rules Harmony) { (+ + −, + + +), (+ − +, + + +), (− + +, + + +),

(− − +, −− −), (− + −, −− −), (+ − −, −− −), . . . }

As has been discussed in the OT literature, this map is the optimal outcome of two
very simple constraints: a markedness constraint banning successive vowels with
different values of feature F (Agree(F)) outranking the faithfulness constraint
Ident(F). Baković (2000: 26) defines the term this way:
When Agree[F] is dominant, it winnows the candidate set down to basically two candidates,
one with all [αF] segments and the other with all [−αF] segments. If IO-Ident[F] gets the next
crack at the evaluation process, it will choose the one of these candidates that is least deviant
from the input, regardless of the stem/ affix or +/− distinctions. In other words, what ends up
mattering is the relative percentages of [αF] and [−αF] vowels in the input: the underlyingly
greater number of [−αF] vowels in [a map where /+ − −/ ↦ [− − −]] gangs up on the lesser
number of [αF] vowels, yielding the problematic effect that I call ‘majority rule.’ [emphasis in
original]
As recognized by Baković, MR is unattested and considered phonologically
bizarre. His solution adds certain locally conjoined constraints to CON, which he
argues has the effect of ridding Majority Rules maps from the typology, and which
he argues is independently needed for analyzing dominant/ recessive types of
vowel harmony. The point I wish to emphasize here however is that Majority
Rules is a logically possible map, which is quite easy to generate in classic OT
with a simple markedness constraint (depending on whether arbitrary many
consonants may intervene between vowels determines whether Agree(F) is SL or
TSL) and standard faithfulness constraints.
Another possible map is one where the first or last vowel determines the
features of the other vowels in the word. This has been called progressive
harmony (PH) and regressive harmony (RH), respectively. Examples of a PH map
are shown in (21).

(21) (Progressive Harmony) { (++−, +++), (+−+, +++), (−++, −−−), (−−+, −−−),
(− + −, −− −), (+ − −, + + +), . . . }

The inclusion of neutral vowels alters this map only slightly. Neutral are vowels
which resist harmonizing and either are skipped (in which case they are called
transparent) or force subsequent vowels to harmonize with them (in which case
they are called opaque). Following Heinz & Lai (2013), I will use the symbols [⊖]
and [⊟] to represent [−F] vowels that are transparent and opaque, respectively, and
the symbols [⊕] and [⊞] to represent [+F] vowels that are transparent and opaque,
respectively. With this expanded alphabet, mappings like /+ ⊖ −/↦[+ ⊖ +] and /+
⊟ +/↦[+ ⊟ −] would also belong to the PH map.
In contrast to the above, sometimes vowel harmony is bounded, in the sense
that only the subsequent vowel is affected. I will call this Local Assimilation (LA)
and (22) below illustrates this map where only the initial vowel is the trigger.89

(22) (Local Assimilation) { (+ −−, ++ −), (−++, −−+), (+ −+, +++), (−+ −, −−−),.
..}

Early analyses of vowel harmony analyzed the extension of many patterns like
bounded or unbounded PH or RH (van der Hulst & van de Weijer 1995), but this
type of analysis is present in recent work as well (Nevins 2010).
Another logically possible, but unattested type of vowel harmony process has
been called “Sour Grapes” (SG) (Padgett 1995; Wilson 2003). Informally, SG is
like progressive harmony except that later vowels only harmonize if no opaque
vowels occur later in the word. If an opaque vowel occurs somewhere after the
initial vowel, then non-neutral vowels between it and the initial vowel will not
harmonize. (23) illustrates a SG map.

(23) (Sour Grapes) { (+−−, +++), (+−⊟, +−⊟), (+−−−, ++++), (+−−⊟, +−−⊟), . .
.}

Like Majority Rules, Sour Grapes has been argued to be a phonologically bizarre
vowel harmony process. In particular, Wilson (2003) argues that harmony
processes never look ahead beyond immediately adjacent segments. Wilson refers
to the absence of look-ahead as a kind of myopia, and characterizes spreading
processes as “myopic”. The Sour Grapes pattern disobeys Wilson’s (2003)
principle that phonological laws are myopic. In a Sour Grapes pattern, each vowel
gets to “look ahead” arbitrarily far to the end of the word to see if there is an
opaque vowel downstream, and only harmonizes if it does not find one.
Classic OT has no difficulty generating SG maps. Under a typical analysis,
there is a markedness constraint against segments that are [+F] but share all other
features with ⊟ (such as ∗⊞). Consequently, underlying /⊟/ can never surface as
[⊞]. Finley (2008: 32) describes the rest of the OT analysis this way.
Sour grapes harmony patterns occur when a blocker prevents spreading to vowels intervening
between the source and the blocker. For the input [/+ − ⊟/] . . . the output [[+ − ⊟]] will be
optimal rather than the desired [+ + ⊟]. This type of pathology is produced when the harmony-
inducing constraint [Agree(F)] does not localize the violation of harmony. In both the sour
grapes candidate [+−⊟] and the spreading candidate [+ + ⊟], there is only one locus of
disagreement. [. . .] However, because the sour grapes candidate incurs no faithfulness
violations, it will emerge as optimal.

Like MR, SG has received attention in the literature because it the optimal
outcome of relatively simple constraints in OT (Wilson 2003; McCarthy 2004;
Finley 2008).
Interestingly, in the domain of tone, there do seem to be some patterns that
exhibit Sour Grapes-like behavior. See Hyman (2007) and Kula & Bickmore
(2015) for cases in Kuki Thaadow and Copperbelt Bemba, respectively, and
Jardine (2016) for extensive analysis and discussion.
Other analyses of vowel harmony argue that the right generalization is that
vowels in a word harmonize to a particular feature value, if it is present anywhere
in the word. This analysis has been called dominant/ recessive (DR) since the
feature F appears to have a dominant value (the one that vowels harmonize with)
and a recessive value (the one that vowels don’t harmonize with). In the example
DR map below, the [+] value is the dominant one; so any underlying
representation containing the harmonizing feature with the value [+] will surface
so that the harmonizing feature in all vowels will also be [+].

(24) (Dominant/Recessive) { (++−, +++), (+−+, +++), (−++, +++), (−−+, +++),
(− + −, + + +), (+ − −, + + +), (− − −, −− −), . . . }

A similar analysis of VH patterns (shown in (25)) is one where the root vowel
determines the features of the other vowels in the word. This kind of analysis has
been termed “Stem Control” (SC). Here the feature that spreads is determined by
its morphological status, and not its inherent vowel as in DR harmony. Typically
vowels agree with the closest stem vowel. Following Baković (2000), I use the

and to indicate root vowels that are +F and −F, respectively.

(25)

The last logically possible harmony pattern to be discussed I will call

Circumambient Unbounded harmony (CU), following Jardine (2016). This pattern
is also like dominant/recessive harmony in that only one value of the feature
triggers the harmony. However, CU harmony requires two /+F/ triggers which
must SURROUND the affected vowels, and which may be ARBITRARILY FAR from
them. The examples in (26) illustrate.

(26) (Circumambient Unbounded) { (+ − −, + − −), (+ + −, + + −), (+ − +, + +

+), (+ −− −, + −− −), (+ − −+, + + ++), (− + −− +, − + + + +), . . . }

The term “circumambient” refers to two surrounding triggers and the term
“unbounded” refers to the absence of a bound on the distance between the two
triggers. Yaka is the only language which appears to have CU vowel harmony
(Hyman 1998), though Jardine (2016) argues Sanskrit n-retroflexion is formally
similar, and (Graf 2010a) provides a logical analysis of it. (As discussed further
below in 6.3, unbounded high tone plateauing is a well-attested, a common tonal
pattern which is circumambient unbounded (Hyman 2011; Jardine 2016).)
Table 5.5 is a reproduction of Table 5.3 from Heinz & Lai (2013), with the
additions of local assimilation and circumambient unbounded harmony. It
summarizes the encyclopedia of types outlined above. Phonological theory has
posited maps like LA, PH, RH, DR, and SC, the consensus appears to be that MR
and SG are not only unattested but bizarre. In fact they have been called
pathological patterns in some works (Wilson 2003, 2004; Finley 2008). Lastly,
setting aside tonal phonology for now, maps like CU are only marginally attested.
As before, we ask the question: What principle or principles separate the
linguistically- motivated generalizations (PH, RH, DR, SC) from the pathological
ones (MR and SG) and the marginally attested ones (CU)?
Table 5.5: Example mappings of underlying forms (w) given by local assimilation (LA),
progressive harmony (PH), regressive harmony (RH), dominant/recessive harmony (DR),
sour grapes harmony (SG), majority rules harmony (MR), and circumambient unbounded
harmony (CU). Symbols [+] indicates a [+F] vowel and [−] indicates a [−F] vowel where F
is the feature harmonizing. Symbols [⊟] and [⊖] are [−F] vowels that are opaque and
transparent, respectively. (From Heinz & Lai 2013: 57.)

6.2 An encyclopedia of categories: String-to-string

maps
Hierarchies of string-to-string transductions are not as well studied nor understood
as classes of stringsets. Part of the issue is that string-to-string maps are inherently
more complex; they have more parts than a stringset because in fact they are two
stringsets — the domain and co-domain — and a map from elements of one to the
other. Consequently, properties that converge for stringsets can diverge for string-
to-string maps. One example is the class of regular stringsets. Regular stringsets
are exactly those describable with MSO logical formulae with successor,
deterministic finite-state acceptors and non-deterministic finite-state acceptors.
(Informally, a finite-state machine is deterministic only if there is at most one path
through the machine for each input; if there are some inputs with more than one
path, it is non-deterministic.) However when the corresponding classes of string-
to-string maps are considered, these three classes are distinct (Engelfriet &
Hoogeboom 2001).
The hierarchies of transductions that are known to exist, such as those
described by Roche & Schabes (1997, Chapter 1) and Engelfriet & Hoogeboom
(2001) focus on the more expressive regions. There is very little work on regions
that make the kinds of distinctions made in the Subregular Hierarchies discussed
in the previous section.
Therefore, current answers to the questions at the end of the preceding section
at present are unlikely to satisfy all readers since they are incomplete.
In this section, I will not explicate all the known regions, but only those that I
think are currently most relevant for phonology. These regions are shown in Figur
e 5.4. As before, lines connecting two regions indicate that the higher region
properly includes the lower region. I will begin at the bottom and go up.

Figure 5.4: Subregular hierarchies of regular relations.

6.2.1 Input Strictly Local functions

Input Strictly Local functions generalize the notion of Strictly Local stringset.
Recall the Strictly Local stringsets are Markovian in nature: the well-formedness
of a string can be determined by examining the substrings of length k.
Equivalently, this means that the well formedness of any position in the string can
be determined by checking the k − 1 previous symbols. This is illustrated in Figure
5.5, for the case where k = 2.
Input Strictly Local functions are similarly Markovian. The idea is that every
element in the input string corresponds to a STRING of symbols in the output string.
For any input symbol x its output string u will only depend on x and the previous k
− 1 elements of x in the input string. Figure 5.6 illustrates, for the case where k =
2.

Figure 5.5: Schematic illustration of the Markovian nature of Strictly k-Local

stringsets. Each element x of a string belonging to a strictly 2-local stringset
depends only on the previous element. In other words, the lightly shaded cell
only depends on the darkly shaded cell.

Local Assimilation (LA) is ISL with k = 3. Basically if the previous two

elements are /⋊+/ (or /⋊−/) then if the current input element is a non-neutral
vowel, the output string will be [+] (or [−]).
Chandlee (2014) shows that ISL functions can model a range of local
phonological processes, including substitution, insertion, deletion, and synchronic
metathesis. More generally, she shows that given a mapping describable with a
rule of the form A → B / C __D where the set of strings in CAD is finite and the
rule applies simultaneously then it is ISL for some k.
This result may seem counter-intuitive given the current discussion. A reader
may wonder whether, especially given the diagram in Figure 5.6, how ISL
functions can model any transformation triggered by any right context at all. As
mentioned, every element in the input string corresponds to a STRING of elements
in the output. These output strings can be any length, including length zero (the
so-called “empty” string). The option to output the empty string allows the
function to wait until it has enough information to decide what to output. But
importantly, the amount of input it needs to see to make this decision is bounded,
by the specified value of k. For example consider regressive nasal place
assimilation where underlying /inpa/→[impa]. Each row in Table 5.6 shows how
the output string is determined by each input element x and the input element
preceding x. Since the output string at each point is determined by a window
whose size is bounded by k, ISL maps are myopic in Wilson’s (2003) sense.
Chandlee also investigated the approximately 5500 phonological processes
(from over 500 languages) reported in the P-Base database (v1.95 Mielke 2008). It
was determined that over 95% of these patterns are ISL. Chandlee acknowledges
that P-Base ought not be taken as representative of the cross-linguistic distribution
of processes that target contiguous versus non-contiguous segments. However,
given that it is the most comprehensive collection of processes of which we are
aware, she deemed it necessary to survey.

Figure 5.6: Schematic illustration of the Markovian nature of Input Strictly k-

Local functions. For every Input Strictly 2-Local function, the output string u of
each input element x depends only on x and the input element previous to x. In
other words, the lightly shaded cell only depends on the darkly shaded cells.

Table 5.6: Illustrating why transformations with right contexts can still be ISL. The symbol
λ represents the empty string (the string of length zero).
element preceding input element output string
x x u
⋊ i i
i n λ
n p mp
p a a

Furthermore, Chandlee (2014) and Chandlee et al. (2014a) also show how ISL
functions can be efficiently learned from finitely many examples in the sense of
Gold (1967) and de la Higuera (1997). This stands in stark contrast to the class of
regular functions which cannot be so learned. Remarkably, Jardine et al. (2014)
generalize this result to obtain an even more efficient learning algorithm for this
class of functions.

6.2.2 Output Strictly Local functions

A notable example of a map that Input Strictly Local functions are unable to
model are ones like progressive harmony (PH) (21) above. Recall that a mapping
like /+ − −−/↦[+ + ++] belongs to this map, and more generally for all numbers k,
/+ −k −/↦[+ +k +] and /− −k −/↦[− −k −]. Such a map cannot be Input Strictly Local
for any k. This is because whether the last input element surfaces as [+] or [−]
depends on an input element which is more than k input elements away.
 Chandlee (2014) defines Left and Right Output Strictly Local functions
(LOSL and ROSL) to address such maps. These capitalize on the output-oriented
nature of many phonological processes (Kisseberth 1970; Prince & Smolensky
1993, 2004). They are Markovian like ISL functions, but this time the context is
found in the output string, not the input string. Specifically for Left (Right) OSL
functions, for any input element x, its output string u will only depend on x and the
previous (following) k − 1 elements of the output string. The idea is that a function
is Left or Right, depending on whether the left or right context in the output string
matters. Figures 5.7 and 5.8 illustrate Left and Right OSL functions, respectively,
for the case where k = 2.

Figure 5.7: Schematic illustration of the Markovian nature of Left Output

Strictly k-Local functions. For every Left Output Strictly 2-Local function, the
output string u of each input element x depends only on x and the output
element previous to u. As before, the lightly shaded cell only depends on the
darkly shaded cells.

Figure 5.8: Schematic illustration of the Markovian nature of Right Output

Strictly k-Local functions. For every Right Output Strictly 2-Local function, the
output string u of each input element x depends only on x and the output
element succeeding u. As before, the lightly shaded cell only depends on the
darkly shaded cells.

Informally, Left and Right OSL functions can be thought of as characterizing the
maps one can describe with rewrite rules that apply left-to-right or right-to-left
(Howard 1972) (cf. the treatment of rule-application by Kaplan & Kay 1994). This
appears to be approximately correct, though certain details are still being worked
out. However, we can say with certainty that the map PH is LOSL and the map
RH is ROSL. More generally, such functions capture spreading processes such as
progressive and regressive nasal spreading.
Left and Right OSL functions can both be computed by subsequential
transducers. For Right OSL functions, the input string must be processed right-to-
left by the transducer and the resulting output will then be reversed. See Heinz &
Lai (2013) for details.

6.2.3 Subsequential functions

In the abstract maps for vowel harmony discussed earlier, consonants were
ignored. If arbitrary many consonants are allowed to intervene between the vowels
then the PH and RH maps will not be LOSL nor ROSL, respectively. For the PH
case, this means for all numbers k, /+Ck −/↦[+Ck +] and /−Ck −/↦[−Ck −]. Such a
map cannot be Left nor Right Output Strictly Local for any k because whether the
last input element surfaces as [+] or [−] depends on an output element which is
more than k output elements away. In a sense, at input element x, the functions
cannot remember whether the preceding vowel in the output string was [+] or [−]
because too many [C]s intervene.
We therefore move up the hierarchy in Figure 5.4. I note that as of yet there are
no regions for string-to-string maps corresponding to the SP, LT, PT, TSL, or LTT
stringsets.90
Informally, for Left (Right) Subsequential functions, each logically possible
input string is classified as belonging to exactly one of finitely many regular
stringsets. For any input element x, the output string u will only depend on x and
the regular stringset to which its preceding input string belongs. Figure 5.9
illustrates left subsequential functions.
Even if arbitrary many consonants are allowed to intervene between the
vowels then the PH and RH are in fact left and right subsequential, respectively.
To see why, consider Table 5.7. Subsequential functions can “remember” up to
finitely many pieces of information about the left context; in Table 5.7, that the
first vowel was [+F]. Thus even if k or more Cs then occur in the input string, the
function simply outputs each C as it reads each C, without changing its memory
state.
 Figure 5.9: Schematic illustration of Left Subsequential functions. For every
Left Subsequential function, the output string u of each input element x
depends only on x and the stringset S to which the preceding input string
belongs. As before, the lightly shaded cell only depends on the darkly shaded
cells.

Table 5.7: Illustrating why PH is left subsequential even if arbitrarily many Cs intervene
between vowels.
set to which string preceding x belongs input element x output string u
⋊ + +
⋊ + C∗ C C
⋊ + C∗ C C
... ... ...
⋊ + C∗ C C
⋊ + C∗ – +

In the same way that ISL functions could “look ahead” by writing empty
output strings, subsequential functions can do so do as well. However, like the ISL
functions, there is a sense in which left subsequential functions can look into the
right context of the input element ONLY some finite distance. There is a bound k on
how far they can look ahead, which relates to the fact that it can only remember
finitely much information about the input string. For this reason it is not possible
to remember the EXACT preceding input string.
An example will help make this idea clear. The dominant/recessive (DR) map
is neither left nor right subsequential. This is because, for all numbers k, /−k + −k /
↦[+k + +k ] and /−k − −k /↦[−k − −k ]. Such a map cannot be left subsequential
because whether the first k input elements all surface as [+] or [−] depends
whether the next element is [+] or not. Therefore, even though these functions
might output the empty string for the first k input elements, if the [+] comes next,
such functions would have to output k [+] symbols (and one more). But this is
impossible because k can be ANY number and left subsequential functions can only
classify the preceding input string into one of FINITELY many categories. Table 5.8
illustrates this conundrum. For this reason, left subsequential functions are myopic
in the sense that they cannot look unboundedly far into the right context.
Right subsequential functions are similar except that input strings following
the input element are categorized into finitely many regular stringsets. Also, right
subsequential functions can only “look ahead” into the left context a finite
distance (and an argument similar to the one made above shows why). It may be
useful to think of right subsequential maps as the “reverse” of left subsequential
maps: if L is a left subsequential map then there is a right subsequential map Lr
such that (w, v) ∈ L iff (wr, vr ) ∈ Lr (where xr is the reverse of the string x so
(abc)r = cba). From a processing perspective, one could say that left subsequential
functions process strings left-to-right, and right subsequential functions process
strings right-to-left.
Table 5.8: Illustrating why the dominant/recessive DR map is not left subsequential. The
symbol λ represents the empty string. The problem is that the left subsequential function
cannot remember exactly how many − symbols occurred before the first + (it cannot
always correctly fill in the ‘. . . ’).
set to which string preceding x input element output string
belongs x u
λ — λ
— − λ
−− − λ
... ... ...
−− − . . . − λ
−− − . . . − + +++...++

At the University of Delaware in 2010, the question was asked whether the
transformations from underlying to surface forms are left or right subsequential. In
other words, we investigated what I will call the Subsequential Hypothesis.

(27) (Subsequential Hypothesis) Phonological transformations are left or right

subsequential.

With one interesting class of exceptions discussed below, this hypothesis appears
to be well supported. This matters for two reasons. First, it is a stronger more
restrictive hypothesis than the previously understood bound (phonology is regular,
see Section 4.3). Second, it has been known for quite some time that left
subsequential (and right subsequential) functions are learnable in a particular
sense (Oncina et al. 1993). The algorithm presented there has even been adapted
for use in phonology (Gildea & Jurafsky 1996). In other words, if phonological
transformations are subsequential, then the computational nature of phonological
transformations directly provides purchase on the learning problem.
So what is the evidence which favors (27)? I will use the term “subsequential”
to mean either left or right subsequential. Chandlee (2014) proves that ISL, LOSL,
and ROSL functions are subsequential; therefore, all the maps they cover are
subsequential. Synchronically attested metathesis is also subsequential (Chandlee
et al. 2012; Chandlee & Heinz 2012). Gainor et al. (2012) study the extensions of
the vowel harmony maps in Nevins (2010) and conclude they are subsequential.
Since Nevins assumes a certain degree of underspecification, Heinz & Lai (2013)
show that progressive and regressive vowel harmony with no underspecification
pace OT (maps PH and RH above) are subsequential. Payne (2017) shows that
long-distance consonant dissimilation maps described by Suzuki (1998) and
Bennett (2013) are subsequential, and Luo (2017) shows that long-distance
consonant assimilation maps described by Hansson (2008) and Rose & Walker
(2004) are subsequential.
In some sense, these results are not too surprising because ultimately these
results support Wilson’s intuition that phonology is myopic. Nonetheless, if
phonological myopia is best characterized as subsequentiality (or something
stronger like ISL), then there is much concrete to gain: a theory which not only
appears sufficiently expressive, but which is also more restrictive than previously
entertained, and which has desirable learnability properties.

6.2.4 Weakly Deterministic functions

Weakly Deterministic functions are defined by Heinz & Lai (2013) as those maps
that can be defined as the composition of a left subsequential and right
subsequential function without the introduction of new alphabetic symbols. Heinz
& Lai (2013) show that the dominant/ recessive (DR) and stem-control (SC) maps
are properly weakly deterministic. In fact the DR map is the composition of a map
like progressive harmony (DRP ), which only spreads the dominant feature
progressively, and a map like regressive harmony (DRR), which only spreads the
dominant feature regressively. Table 5.9 illustrates. Since DRp and DRR are left
and right subsequential, respectively, their composition is weakly deterministic.
Heinz & Lai (2013) conjecture that sour grapes (SG) is not weakly
deterministic. They explain that SG can be described as the composition of a left
subsequential function and a right subsequential function, but that crucially the
intermediate form requires the use of an additional alphabetic symbol which they
write as [?]. Table 5.10 (adapted from their paper) illustrate the role an additional
symbol plays in the decomposition. Essentially, [?] records the fact that this is a
minus, which has a [+] in its left context. So the right subsequential function R
will rewrite this as [−] or [+] depending on whether there is a [⊟] in the right
context of the [?].
Table 5.9: Map DRp converts every − after a + to + (like PH), and map DRR convert
every − before a + to + (like RH). As indicated, the composition of these two maps yields
the DR map.

Table 5.10: Illustrations of the role of the new symbol [? ] in the deterministic
decomposition into a left subsequential function L and a right subsequential function R.

While many theorists have argued in favor of dominant/recessive and stemcontrol

analyses of vowel harmony(Baković 2000; Kr¨amer 2003), this is not a settled
debate (Nevins 2010). What Heinz & Lai (2013) show is that DR and SC maps are
more computationally complex than PH and RH maps, and that SG maps are even
more complex than these. If the debate is settled in favor of DR and SC maps then
it means the subsequential hypothesis would be incorrect, and then the most
restrictive hypothesis would retreat to the weakly deterministic region.
Jardine (2016) argues that circumambient unbounded (CU) maps are also not
weakly deterministic. He shows that CU maps and SG maps are similar in that in
both cases, the information which determines whether a vowel is unfaithful is
located in two distinct places: one location is arbitrarily far before the vowel and
the other location is arbitrarily far after the vowel. While both CU and SG maps
are the same with respect to the first place (an earlier + is required), the CU map
also requires a + in the second place but the SG map requires the ABSENCE of a ⊟.
Jardine’s result is perhaps the most serious challenge to the Subsequential
Hypothesis (or a revised Weakly Deterministic hypothesis) because the best
characterization of Yaka vowel harmony seems to be that it is circumambient
unbounded (Hyman 1998). However, this is the only known example of this type,
and it is probably premature to reject the hypothesis on these grounds alone. I will
return to this issue below.
6.2.5 Non-deterministic Regular functions and Regular relations

Non-deterministic Regular functions can be defined in at least two ways (Elgot &
Mezei 1965). First, they can be defined as the composition of a left and right
subsequential function, provided the intermediate string is allowed to use
additional alphabetic symbols. As we have seen in the example of Sour Grapes
(SG) in Table 5.10, these additional symbols allow certain types of information to
become present in the string. Second, non-deterministic regular functions can be
defined as those string-to-string functions that can be described with a non-
deterministic finite-state transducer.
Both SG and CU maps thus properly belong to the non-deterministic regular
function region (Heinz & Lai 2013; Jardine 2016).
Non-deterministic finite-state transducers are a grammatical formalism that
can also describe transformations which have more than one output for each input.
In fact the class of transformations describable with non-deterministic finite-state
transducers are called regular relations. Beesley & Kartunnen (2003) and Hulden
(2009) develop toolkits for manipulating regular relations for describing the
phonology and morphology of languages.
The Majority Rules map cannot even be described with a non-deterministic
finite-state transducer. It is in fact non-regular (Riggle 2004; Heinz & Lai 2013).
According to the hierarchy presented here, it is the most complex kind of map
under discussion.
Given that many phonological rules are optional, one may wonder whether it
is appropriate to model individual transformations (as we have here) as functions
instead of relations. There are two responses to this.
The first response is to say that the optionality is handled at a higher level of
control than the individual transformation. This is essentially the position adopted
in rule-based and constraint-based phonology. In rule-based phonology, the idea
was that a rule was marked as optional. When deriving the output from an input
and a rule marked as optional is encountered, additional, usually random,
information is consulted (such as a coin flip) and the outcome determines whether
the rule is applied or skipped. Thus the extension of the rule itself is still
functional. Similarly, in stochastic OT (Boersma 1998; Boersma & Hayes 2001), a
given constraint ranking has a functional extension, but a higher-level control
process determines which particular constraint ranking will be utilized at any
particular time.
The second response is to say that subclasses of regular relations are likely to
follow the same lines developed here. Mohri (1997) establishes that as long as
there is a bound on the amount of optionality, many properties of subsequential
functions are preserved. More recently, Beros & de la Higuera (2014) also show
how to generalize subsequential functions in a way that permits a degree of
optionality. While subclasses of classes have not been studied the fact that they
preserve important aspects of the underlying finite-state transducers and that
classes like ISL have automata-theoretic characterizations based on subsequential
transducers (Chandlee 2014; Chandlee et al. 2014a) strongly suggests that
subclasses like ISL which permit a degree of optionality are only waiting to be
discovered.

6.3 Further evidence

There is some psycholinguistic evidence which supports the hypothesis that
phonological transformations are regular. In a series of artificial language learning
experiments, Finley (2008) compared how well native English speaking young
adults could learn a Majority Rules (MR) type map (which is non-regular) as
compared to a Progressive Harmony (PH) type map (which is regular). Subjects
were either assigned to the MR, PH, or control conditions. Each subject received a
training session and then performed in a test session. The results clearly
established that subjects learned the PH harmony pattern, but not the MR pattern.
Finley (2008) also conducted a series of experiments to investigate the
learnability of the Sour Grapes (SG) map. This is one way to test the
Subsequential Hypothesis. Here the results were inconclusive, probably due to an
interference of neutral vowels in the particular paradigm (though see Finley 2015).
However the subsequential hypothesis clearly predicts that SG should be more
difficult to learn that PH, and so future research in this vein should be conducted
to see whether the prediction is borne out.
An interesting source of evidence in favor of (a revised) Subsequential
Hypothesis comes from work by Jardine (2016). Jardine studies Unbounded Tone
Plateauing (UTP) (Kisseberth & Odden 2003; Hyman 2011) and concludes that
such transformations are also circumambient unbounded. In UTP, a string of
underlying low tones (or unmarked vowels) are realized as high only if there is a
high tone at both the left and right edges of this string. Jardine makes a persuasive
case for a typological asymmetry between tonal patterns and segmental patterns.
Several well-documented cases of UTP exist in the literature, despite the absence
of comprehensive typological surveys. On the other hand, despite the existence of
several surveys on long-distance harmony, the only CU maps known to be present
in segmental phonology come from Yaka vowel harmony and Sanskrit n-
retroflexion. Furthermore, Jardine shows that the evidence that the segmental
maps are truly unbounded is weaker than the evidence that the UTP cases are
unbounded.
 In other words, Jardine’s work shows that UTP — because of its widespread
and well documented existence — is a counterexample to the Subsequential
Hypothesis (27). He suggests that it be revised as follows.

(28) (Revised Subsequential Hypothesis) Segmental transformations in

phonology are left or right subsequential.

While Yaka vowel height harmony and Sanskrit n-retroflexion are arguably
counterexamples to this revised hypothesis, I think it is prudent not to reject the
hypothesis on these grounds. Unlike UTP, these cases are rare and the evidence
that they are truly CU maps, while compelling, is not as strong as it is for the UTP
cases. Future research may lead to a better understanding of these languages.
Thus, Jardine shows that tonal patterns are different from segmental patterns,
since they are arguably more complex. Paraphrasing Hyman (2011), tonal
phonology really can do more than segmental phonology!

6.4 Transformations: A summary

A summary of the foregoing section can be made very simply. Phonologists have
identified many ways in which underlying forms are transformed into surface
forms. The study of subregular string-to-string maps has not yet been as
articulated as the one of subregular stringsets. Nonetheless, the study of the
typology of the attested transformations in the light of the existing categories
yields similar conclusions. With one interesting class of exceptions which is
largely confined to tonal phonology (CU maps), segmental transformations appear
to overwhelmingly belong to the simplest maps in the encyclopedia of categories
shown in Figure 5.4.
Thus, even though an encyclopedia of categories for string-to-string maps as
fine-grained as the one for stringsets does not yet exist, the work to date has
nonetheless yielded important insights. The simplest maps are Markovian on the
input or the output (ISL, LOSL, and ROSL), and very many phonological
transformations belong to these classes. Transformations which are not Markovian
in this sense involve long-distance harmony. Such patterns however are
subsequential, which means they are still myopic in an important sense. This
stands in contrast to the SG map which is not subsequential and the MR map,
which is not even a regular relation. It is anticipated that other subregular regions
for maps analogous to the SP or TSL regions will be developed that better
characterize long-distance transformations.
The asymmetry between tonal CU maps and segmental CU maps noticed by
Jardine (2015) is perhaps the most difficult to interpret. How can it be that some
formal mechanism is available to one aspect of the grammar but not to another?
Perhaps it is an indicator of the modularity of grammar (Heinz & Idsardi 2011,
2013). Jardine’s work then can be understood as providing support for Hyman’s
thesis, that tonal phonology is in fact different from segmental phonology (Hyman
2011). As a separate module of the grammar, it has resources available to it that
are not available everywhere else.

7 Summary and implications for the phonological

component
In this section, I would like to review the main lessons for phonological theory to
be taken from the computational analyses reviewed so far in this chapter.

7.1 Phonological generalizations have strong

computational properties
From the computational perspective, most, if not all, phonological generalizations
obey very strong computational laws. This is what typological analysis of
phonological constraints and transformations that was reviewed in Sections 5 and
6 reveals. Phonological generalizations (both stringsets and maps) belong to small,
well-defined regions WITHIN the region of regular stringsets and maps, and these
regions are at the BOTTOM of the subregular hierarchies shown in Figures 5.2 and
5.4.

7.2 Problems with optimization

Current phonological theories do not account for these laws. Since Prince &
Smolensky (1993), optimization has been a central feature of phonological theory
including classical OT (Kager 1999), Stratal OT (Kiparsky 2000), Harmonic
Grammar (Smolensky & Legendre 2006; Potts et al. 2008), Maximum Entropy
(Goldwater & Johnson 2003; Hayes & Wilson 2008), and Harmonic Serialism
(McCarthy 2008b). One of the most compelling features of optimization is the
idea that complex patterns within and across languages arise from the interaction
of simple constraints. The celebrated examples of syllabification in Berber (Dell &
Elmedlaoui 1985), complex margin avoidance in Yokuts (Kisseberth 1970), and
the many solutions to the international conspiracy ∗NC˳ (Pater 2000) in terms of
optimization all attest to this. However, if “complex patterns arising from the
interaction of simple constraints” is optimization’s greatest strength, it is also its
greatest weakness.
As explained in Section 4.3, computational analysis has revealed that
phonological transformations are regular. But even with regular constraints and a
regular GEN, optimization can result in non-regular maps (Frank & Satta 1998).
Optimization is very powerful because very complex patterns can indeed arise
from simpler constraints. Majority Rules is a case in point (Riggle 2004; Heinz &
Lai 2013).
This particular overgeneration problem is not specific to classical OT. It is a
problem for Stratal OT, Harmonic Serialism and Maximum Entropy theories as
well. This is not a controversial point. It is in fact just one more example of how
“complex patterns arise from the interaction of simple constraints”. It may be
possible to add constraints to CON as in Baković(2000) to avoid generating MR
type maps. Different types of constraints, such as targeted constraints (Wilson
2001; Baković & Wilson 2000; Baković 2004), or ones which operate over turbid
structures (Finley 2008) may be invoked.
However, it is not enough to show that MR is avoided. One must show that all
non-regular maps are avoided. Frank & Satta (1998) write: “It remains an open
problem to characterize precisely the generative capacity [. . .] of [Optimality
Systems]’s with other assumptions about the formal power of GEN and the
constraints”. While it still remains an open problem today, the markedness
constraints involved in generating Majority Rule patterns are SL or TSL. In other
words, even the simplest constraints under optimization can generate non-regular
maps. Perhaps future research will show that there is an straightforward way to
prevent non-regular maps from occurring in the factorial typology, but to me the
prospects seem dim.
For even if it were possible to add new constraints (or constraint types) to
CON to avoid deriving non-regular maps, there is a problem. These constraints
would be in service of deriving a generalization that is already very simple to
state: phonological transformations are regular. It means that in order for
optimization to be the right theory of phonology that the constraints over which
optimization operates have to be designed to blunt the power of optimization!
 And these problems are just to avoid generating non-regular maps. The
revised Subsequential Hypothesis limits the kinds of humanly possible
segmental maps. The same kinds of problems mentioned above exist for
developing a theory of CON which guarantees that segmental maps are
always subsequential and avoid generating non-myopic maps like Sour
Grapes (Jardine 2016a).
From a computational perspective then, optimization appears to be too
powerful a tool. Critics of optimization have generally focused on the fact
that OT undergenerates the typology with respect to opacity (Idsardi 1998,
2000; McCarthy 2007; Buccola 2013), but the overgeneration problem is
equally pressing. In both cases, new constraint types or optimizing
architectures are introduced with the ultimate purpose to make the optimal
maps ones that conform to the computational generalizations stated in the
present chapter.
Thus, from a computational perspective, with respect to the dual goals
of developing both an adequately expressive theory and a maximally
restrictive theory of phonology, optimization misses the mark. It is neither
adequately expressive (because of opaque maps) nor sufficiently restrictive
(because it generates non-regular maps like Majority Rules).
Of course the question can never be “Is Optimization the Right
Theory?”. After all, what is the alternative?

7.3 Organizing phonological theory around

computational properties
One alternative suggested by the work reviewed in this chapter is that the
computational properties highlighted here — and not optimization — be
taken as the organizing principles of the phonological component of
grammar. Constraints on surface forms are, with few exceptions, banning
substrings and subsequences (Section 5). Phonological transformations
which are intuitively “local” are also among the simplest types of logically
possible maps (Section 6).
But is this theory adequately expressive? We have focused on maps
which correspond to individual rules, so can the theory being suggested
handle opacity? This is work in progress, but it appears so. Chandlee et al.
(to appear) show that ISL functions can represent opaque maps. In fact,
they show that every opaque map described in Bakovi ´c ( 2007) is ISL
(and as mentioned, Chandlee and her collaborators have established
algorithms for learning the ISL functions).
Are there overgeneration problems as is the case with optimization-
based theories? It is true that even the theory which adopts the strong
subregular hypothesis (12) overgenerates in some sense. For instance, it is
straightforward to write a grammar for a language which bans subsequences
like ∗sg. This means words in this language are ill-formed if [g] occurs
anywhere after [s]. So there is some overgeneration, though it is plausibly
due to phonetic factors (see below). Joe Pater helpfully points out (p.c.) that
when two theories overgenerate in different ways, it can be difficult to
determine which overgenerates “less” or which is preferable along the
dimension of overgeneration. I agree and this is another area where
computational analysis has something to offer. The theory which
overgenerates in a computationally simpler way ought to be preferred.
Under this assumption, theories of phonology which adopt subregular
properties overgene-rate in a computationally simpler way as compared to
theories organized around optimization. This is for the simple reason that
they cannot possibly generate nonregular patterns like Majority Rules.
Another potential criticism is that computational properties highlighted
here do not take into account phonetic substance. As mentioned, this means
many possible subregular constraints and maps are not likely phonological
ones because they are phonetically unnatural. I will put aside the important
question of whether phonetically unnatural constraints and maps are
phonologically possible and learnable, and just assume for the purposes of
discussion that they are not.
In my view, it is a feature and not a bug that formal and substantive
issues are separated. I am not so extreme as Hale & Reiss (2000) as to deny
substance a role altogether, but I do think science proceeds by factoring
complex systems. Ever since Chomsky & Halle (1968, Chapter 9), it has
been clear that formal and substantive constraints on phonological systems
are distinct. While details have not yet been worked out, substantive
constraints on phonological systems would simply be IN ADDITION to the
formal constraints being proposed here. This is really no different than the
program offered by Chomsky & Halle (1968) or, for that matter, by Hayes
et al. (2004), which adopts optimization as a formal system and ADDS
substantive constraints to the nature of CON.
 Finally, another argument that could be put in favor of optimization is
that it readily lends itself to a theory of learning (Tesar 1995; Tesar &
Smolensky 1998, 2000; Tesar 2014, and many others). This is a good
argument. However, I have tried to highlight the fact that the subregular
regions to which phonological patterns appear to belong also readily lend
themselves to a theory of learning (Oncina & Garcia 1991; Heinz 2007,
2010a,b; Chandlee 2014; Chandlee et al. 2014a; Jardine et al. 2014).
Furthermore, if people generalize from data in the way suggested by these
learning procedures, it explains the computational nature of the
phonological patterns. This is in contrast to the learners in optimization-
based theories (and variants thereof), where the nature of the phonological
patterns is completely explained by what belongs and what does not belong
to CON.

7.4 Next steps

If the computational properties highlighted in this chapter are taken
seriously, there are many subsequent issues to attend to. There are several
possibilities, but I will focus on the following four:
– better characterizing long-distance transformations;
– better understanding the role of abstractness;
– better understanding non-string representations of words; and
– testing the predictions of these hypotheses in imaginative ways.
Each of these is a current focus of research.
Chandlee (2014) provides a clear definition of locality and shows that
many maps which are intuitively local (and some that are not) meet this
definition. Her definition generalizes the notion of Strictly Local stringsets
to maps. Strictly Piecewise stringsets (Rogers et al. 2010) can describe
many long-distance constraints (Heinz 2007, 2010a) as can Tier-Based
Strictly Local stringsets (Heinz et al. 2011). What are the corresponding
generalizations of the SP and TSL stringsets for maps and what range of the
phonological transformations do they cover? Do they have learnability
properties like the SP and TSL regions?
It is well known that deterministic regular functions are less expressive
than non-deterministic functions (see Figure 5.4). Elgot & Mezei (1965)
show a deep connection between non-determinism and abstractness:
basically any nondeterministic function can be described as the composition
of two deterministic functions provided the intermediate form is allowed to
make use of symbols NOT IN THE INPUT ALPHABET. The fact that the
“intermediate” alphabet contains symbols not in the input alphabet
introduces a degree of abstractness (the extra symbols represent abstract
information). This appears to be exactly the dividing line between
dominant/recessive harmony patterns on one side and sour grapes harmony
on the other. Better understanding the role of abstractness in maps and
stringsets is important.
Another example of the interplay comes from a theorem by Medvedev
(1964), which says that every regular language is the homomorphic image
of strictly 2-local languages. In layman’s terms this means every regular
stringset, which can be used to model complex kinds of constraints, can be
derived from a strictly 2-local stringset, which belongs to the lowest level of
the Subregular Hierarchies. It suggests that what looks complex is actually
very simple. But the trick is that the SL2 language has a bigger alphabet and
the latent information hidden in the more complex regular language is made
explicit in the SL2 language. Thus, by making our alphabet larger and more
abstract we simplify constraints we may want to state. But the price is that
the alphabet no longer represents observables (so one consequence is
learning remains as difficult as before).
A related issue is non-string based representations. The main idea is that
there is an interplay between the generative capacity of a formalism, the
representation, and the power of logic. We have seen this already with
respect to how order is represented. Representing order with successor and
precedence has the following implication: if you want to represent
∗[+strident,α anterior]. . . [+strident,−α anterior] you will need MSO logic
with successor, but only CNL logic with precedence. Another example is
discussed in Section 8 below has to do with constraints on syllabic
structure.
Finally, if these computational properties are to become part of the
ontology of a theory of phonology, then much research is possible which
tests the psychological validity of this ontology. The artificial language
learning experiments by Finley (2008) and Lai (2012, 2015) are just one
example of the kind of research that can be carried out. So far this work
supports the hypotheses offered here. But currently there are more
hypotheses than there are experiments testing them.
8 Representational issues
Before concluding, I would like to address the issue of representation. The
above computational analysis appears to rest on the assumption that words
must be modeled as strings. We modeled phonotactic knowledge and
markedness constraints with infinite sets of STRINGS and we modeled the
transformation from underlying to surface forms with infinite sets of STRING
pairs. We argued that it is these objects whose nature we are interested in.
The nature of these objects directly informs the questions in (2). As with
circles (remember circles?), the nature of these objects is to some extent
independent of the grammars used to describe them.
Phonologists are well aware that other representations of words exist,
which are not based on strings and that phonological theories have
employed many kinds of data structures. Nonlinear representations of
words including autosegmental representations (Goldsmith 1976), the grid
(Liberman & Prince 1977), feature-geometric representations (Clements
1985), and gestural scores (Browman & Goldstein 1992). Instead of strings,
these theories employ graph-like data structures. Therefore, it is reasonable
to wonder how much of the foregoing analysis depends on the string-based
representations employed.
In this section, I would first like to explain that the extensions are not
limited to string representations. The concept of “extension” of a grammar
is much more general.
Then I would like to explain why the results in this chapter still matter,
even though they use string-based representations. There are two parts to
this. The first part reiterates the fact that string-based representations are in
fact widely adopted in phonology and explains why they are widely
adopted. Insofar as these reasons are compelling, the results in this chapter
matter.
The second part argues that even if the right representations are not
string-based, studies like the ones reviewed in this chapter are a necessary
step for understanding the computational nature of phonological patterns. I
will explain why subregular hierarchies like the ones presented in this
chapter for strings exist for these other representational schemes for words,
even if they have not yet all been discovered.
The main conclusion is that the interplay between the choice of
representation and the computational principles presented here are likely to
be fruitful areas of research in the coming decades.

8.1 Extensions without strings

In Section 3, I discussed the concept “extension of a grammar”, and
suggested that for constraints the extensions are infinite sets of strings and
for transformations they are infinite sets of pairs of strings. More generally,
the extension is an infinite set of objects, and the objects can be anything so
long as it is well-defined.
For instance, a constraint like NOCODA could be defined so that the
well-formed elements of its extension includes objects like the ones shown
below.

(29)

Non-linear representations such as these are common in phonological

theory, and have brought much insight. Similarly, when discussing
transformations from underlying to surface forms, the left and right
elements in each pair can also be represented with a non-linear, graph-like
representation similar to the ones shown in (29).
In order to think about the extensions of constraints and
transformations, it is necessary to think about the representations of words.
The constraint definition will determine which representations of the
“logically possible” representations are well-formed and which are not. For
instance, we may wonder whether the following representation is permitted
by the theory and if so whether it violates NOCODA.

(30)
At issue of course is the marked structure(s) NOCODA, which is supposed to
define when graph -like representations such as those in the preceding
examples are adopted. There are several possibilities. Just having a node in
the graph labeled with “C” for instance could be sufficient, in which case
(30) would violate it. Or it may be necessary for the node labeled “C” to
dominate a phonetic element (in which case (30) would not violate it). Or it
may be that the labels in the preceding examples are just ornaments for
phonologists for readability, and that what really matters is that there is
consonantal material after the nucleus dominated by the same syllable node.
A more difficult question raises itself with these representations with
respect to the constraint ONSET. At the very least this constraint REQUIRES an
“O” node to be present. It does not ban a sub-structure, suggesting this
constraint is at the Propositional level. Here again we see an interplay
between the choice of representation and the power of the constraint. To
state ONSET, this enriched representation requires a more complex constraint
type than the string representation with the latent syllable boundary, which
can define ONSET as the Strictly 2-Local constraint (∗.V).
These are all interesting possibilities that have been explored to various
degrees in the phonological literature. The point I wish to make is a fairly
obvious one: representations of words matter when defining constraints or
transformations. The extensions of the constraints and transformations will
be in terms of these representations. String representations were used
throughout this chapter, but some other representations could have been
used, and computational analysis could have preceded on those
representations instead.

8.2 Why string representations matter

Strings are one of the most basic data structures (Lothaire 1997, 2005).
They are sequences of events. They are typically defined inductively with
the primitive operation of concatenation and an alphabet: there is a unique
string of zero length (the base case) and then if w is a string and there is a
symbol a in the alphabet then the concatenation of w and a (written wa) is
also a string (the inductive case). It is natural to think of phonological forms
in terms of strings. The act of speech can be thought of as a sequence of
articulatory or acoustic events. Writing systems are string-based
representations of speech.
The most compelling reason to study string representations is that
phonologists use them. One reason may be practical: it is more convenient
to typeset strings than graph-like representations. But even from a
theoretical perspective, the fact that representations were prominently
debated at one point in the history of phonological theory (mainly during
the 1970s and 1980s) does not mean that such issues are necessarily live
today. As Hyman (2014) discusses, phonological theory has moved away
from issues of representation (see also Anderson 1985). String-based
representations are prominent nowadays, even in explaining long-distance
harmony (Rose & Walker 2004; Walker 2011; Nevins 2010), a domain in
which at one time autosegmental spreading analyses seemed particularly
well suited (Hayes 1986). Tonal phonology is less amenable to string-based
representations (for discussion, see also Marlo 2007), but even the
principles governing autosegmental representations have yet to address
every question posed by every language (Hyman 2014).
A second reason to work with strings is that the string is a fundamental
data structure that has been well-studied (Lothaire 1997, 2005). This helps
to make studying stringsets and string-to-string maps the easiest place to
start. If we want to understand how computational principles play out with
complicated data structures, we better first understand how they play out
with simpler structures like strings.
The particular subregular hierarchies being established for strings are
likely to have analogues for other data structures. For instance, Strictly 2-
Local sets of tree structures have been studied and they can be used to
describe context-free string sets (Rogers 1998). More generally, regular tree
languages are also well-studied (Comon et al. 2007). While hierarchies as
articulated as the ones for strings do not yet exist for these other data
structures, they are in fact a focus in theoretical computer science
(Rozenberg & Salomaa 1997).
 Closer to issues in phonological theory, Kornai (1995) shows how
autosegmental representations can be processed by finite-state automata for
recognition and generation. Similarly, Jardine & Heinz (2015) show that
autosegmental representations have important string-like properties and can
be thought of as the concatenation of finitely many autosegmental
primitives. Jardine (2016b) shows how Strictly Local autosegmental
representations can be defined (and how in certain circumstances the No
Crossing Constraint is a Strictly Local constraint).
In short, the methodological points being made in this chapter stand
regardless of whether or not the representations are strings are something
else. Computational principles provide a natural encyclopedia of categories
with which the typology of phonological generalizations can be illuminated.

9 Conclusion
This has been a long chapter. Thankfully, the conclusion can be brief.
Phonology is about how underlying lexical representations are transformed
into surface ones. An important question asks about the cross-linguistic
nature of these transformations. Grammars are typically conceived as
generating patterns; these patterns are extensions of the grammar in the
same way the extension of an algebraic equation is the set of points
satisfying that equation. Computational analysis studies these EXTENSIONS,
and such analysis of phonological generalizations is ongoing. Nonetheless,
the results so far reveal that despite the cross-linguistic diversity, there are
very strong, specific, universal computational properties shared by almost
all phonological patterns. The few potential counter-examples are of special
interest and deserve further study. Explaining these plausibly universal
computational properties of phonological patterns is hard for theories that
rely on optimization as a central organizing feature of the theory, but is
straightforward if the computational properties highlighted within this
chapter become the organizing principles themselves. These principles are
natural for many reasons, only some of which could be covered here. Also,
there is a clear sense in which these principles derive from principles of
inference and learning. While there is still much work to do, a theory of
phonology built around these computational principles promises to be
sufficiently expressive, maximally restrictive, and learnable.
Acknowledgement: I am in indebted to Jane Chandlee, Rémi Eyraud,
Bill Idsardi, Adam Jardine, Regine Lai, Jason Riggle, and Jim Rogers for
invaluable discussion. I am also grateful to Jane Chandlee, Alex Cristia,
Thomas Graf, Larry Hyman, and Lisa Pearl for helpful comments on an
earlier draft. I also thank the students in my computational phonology
course at the 2015 LSA Summer Institute, and the students in the Spring
2015 computational phonology seminar at the University of Delaware, in
particular Hossep Dolatian, Hyun Jin Hwangbo, Huan Luo, Amanda Payne,
Kristina Strother-Garcia, and Mai Ha Vu. Of course I assume full
responsibility for flaws present in this chapter.

References
Anderson, Stephen R. 1985. Phonology in the twentieth century. Chicago:
University of Chicago Press.
Applegate, Richard B. 1972. Ineseño Chumash grammar. Doctoral dissertation,
University of California, Berkeley.
Applegate, Richard B. 2007. Samala-English dictionary: A guide to the Samala
language of the Ineseño Chumash people. Santa Ynez, CA: Santa Ynez Band
of Chumash Indians.
Bach, Emmon. 1975. Long vowels and stress in Kwakiutl. Texas Linguistic Forum
2. 9–19.
Baković, Eric. 2000. Harmony, dominance and control. Doctoral dissertation,
Rutgers University, New Brunswick, NJ.
Baković, Eric. 2004. Unbounded stress and factorial typology. In John J. McCarthy
(ed.), Optimality Theory in phonology: A reader. Oxford: Blackwell. ROA-244,
Rutgers Optimality Archive, http://roa.rutgers.edu/
Baković, Eric. 2007. A revised typology of opaque generalizations. Phonology 24.
217–259.
Baković, Eric. 2011. Opacity deconstructed. In van Oostendorp et al. (eds.) The
Blackwell companion to phonology, 2011.
Baković, Eric. 2013. Blocking and complementarity in phonological theory.
Sheffield: Equinox.
Baković, Eric & Colin Wilson. 2000. Transparency, strict locality, and targeted
constraints. In Roger Billerey & Brook Danielle Lillehaugen (eds.),
Proceedings of the 19th West Coast Conference on Formal Linguistics, 43–
56. Somerville, MA: Cascadilla Press.
Beckman, Jill. 1998. Positional faithfulness. Doctoral dissertation, University of
Massachusetts, Amherst.
Beesley, Kenneth & Lauri Kartunnen. 2003. Finite state morphology. Stanford:
CSLI.
Bennett, William. 2013. Dissimilation, consonant harmony, and surface
correspondence. Doctoral dissertation, Rutgers University, New Brunswick,
NJ.
Benua, Laura. 1995. Identity effects in morphological truncation. In Jill Beckman,
Laura Walsh Dickey & Suzanne Urbanczyk (eds.), Papers in Optimality
Theory, 77–136. Amherst, MA: GLSA Publications.
Benua, Laura. 1997. Transderivational identity: Phonological relations between
words. Doctoral dissertation, University of Massachusetts, Amherst.
Beros, Achilles & Colin de la Higuera. 2014. A canonical semi-deterministic
transducer. In Alexander Clark, Makoto Kanazawa, & Ryo Yoshinaka (eds.),
Proceedings of the 12th International Conference on Grammatical Inference
(ICGI 2014), vol. 34, 33–148. JMLR: Workshop and Conference Proceedings.
Berstel, Jean. 1979. Transductions and context-free languages. Dordrecht:
Springer.
Blevins, Juliette. 2004. Evolutionary phonology. Cambridge: Cambridge University
Press.
Boersma, Paul. 1998. Functional phonology: Formalizing the interactions between
articulatory and perceptual drives. Doctoral dissertation, Universiteit van
Amsterdam. Published as LOT International Series 11. The Hague: Holland
Academic Graphics.
Boersma, Paul & Bruce Hayes. 2001. Empirical tests of the gradual learning
algorithm. Lingustic Inquiry 32. 45–86.
Browman, Catherine P. & Louis Goldstein. 1992. Articulatory phonology: An
overview. Phonetica 49. 155–180.
Buccola, Brian. 2013. On the expressivity of Optimality Theory versus ordered
rewrite rules. In Glyn Morrill & Mark Jan Nederhof (eds.), Proceedings of
Formal Grammar 2012 and 2013, vol. 8306 of Lecture notes in computer
science, 142–158. Berlin: Springer.
Büchi, J. Richard. 1960. Weak second-order arithmetic and finite automata.
Mathematical Logic Quarterly 6. 66–92.
Chandlee, Jane. 2014. Strictly local phonological processes. Doctoral dissertation,
University of Delaware.
Chandlee, Jane, Angeliki Athanasopoulou & Jeffrey Heinz. 2012. Evidence for
classifying metathesis patterns as subsequential. In Proceedings of the 29th
West Coast Conference on Formal Linguistics, 303–309. Somerville, MA:
Cascadilla Press.
Chandlee, Jane, Rémy Eyraud & Jeffrey Heinz. 2014a. Learning strictly local
subsequential functions. Transactions of the Association for Computational
Linguistics 2. 491–503.
Chandlee, Jane & Jeffrey Heinz. 2012. Bounded copying is subsequential:
Implications for metathesis and reduplication. In Proceedings of the 12th
 Meeting of the ACL Special Interest Group on Computational Morphology and
Phonology, 42–51. Montreal: Association for Computational Linguistics.
Chandlee, Jane & Jeffrey Heinz. 2018. Strict locality and phonological maps.
Linguistic Inquiry 49. 23–60.
Chandlee, Jane, Jeffrey Heinz & Adam Jardine. 2018. Input strictly local opaque
maps. Phonology, to appear.
Chandlee, Jane, Adam Jardine & Jeffrey Heinz. 2014. Learning repairs for marked
structures. Poster at the Annual Meeting of Phonology. MIT.
Chomsky, Noam. 1951. Morphophonemics of Modern Hebrew. Doctoral
dissertation, University of Pennsylvania, Philadelphia. Published New York:
Garland Press, 1979.
Chomsky, Noam. 1956. Three models for the description of language. IRE
Transactions on Information Theory 113124. IT-2.
Chomsky, Noam. 1965. Aspects of the theory of syntax. Cambridge, MA: MIT
Press.
Chomsky, Noam & Morris Halle. 1968. The sound pattern of English. New York:
Harper & Row.
Clements, George N. 1985. The geometry of phonological features. Phonology
Yearbook 2. 225–252.
Comon, Hubert, Max Dauchet, Rémi Gilleron, Florent Jacquemard, Denis Lugiez,
Christof Löding, Sophie Tison & Marc Tommasi. 2007. Tree automata
techniques and applications. http://www.grappa.univ-lille3.fr/tata (accessed 12
October 2007).
Dell, François & Mohamed Elmedlaoui. 1985. Syllabic consonants and
syllabification in Imdlawn Tashlhiyt Berber. Journal of African Languages and
Linguistics 7. 105–130.
Dresher, B. Elan. 1999. Charting the learning path: Cues to parameter setting.
Linguistic Inquiry 30. 27–67.
Dresher, B. Elan. 2011. The phoneme. In van Oostendorp et al. (eds.) 2011, vol. 1,
241–266.
Edlefsen, Matt, Dylan Leeman, Nathan Myers, Nathaniel Smith, Molly Visscher &
David Wellcome. 2008. Deciding strictly local (SL) languages. In Jon
Breitenbucher (ed.), Proceedings of the Midstates Conference for
Undergraduate Research in Computer Science and Mathematics, 66–73.
Elgot, C. C. & J. E. Mezei. 1965. On relations defined by generalized finite
automata. IBM Journal of Research and Development 9. 47–68.
Endress, Ansgar D., Marina Nespor & Jacques Mehler. 2009. Perceptual and
memory constraints on language acquisition. Trends in Cognitive Science 13.
348–353.
Engelfriet, Joost & Hendrik Jan Hoogeboom. 2001. MSO definable string
transductions and two-way finite-state transducers. ACM Transactions on
Computational Logic 2. 216–254.
Finley, Sara. 2008. The formal and cognitive restrictions on vowel harmony.
Doctoral dissertation, Johns Hopkins University, Baltimore.
Finley, Sara. 2015. Learning non-adjacent dependencies in phonology:
Transparent vowels in vowel harmony. Language 91. 48–72.
Finley, Sara & William Badecker. 2009. Artificial language learning and feature-
based generalization. Journal of Memory and Language 61. 423–437.
Fougeron, Cécile & Patricia A. Keating. 1997. Articulatory strengthening at edges
of prosodic domains. Journal of the Acoustical Society of America 101. 3728–
3740.
Frank, Robert & Giorgo Satta. 1998. Optimality Theory and the generative
complexity of constraint violability. Computational Linguistics 24. 307–315.
Gainor, Brian, Regine Lai & Jeffrey Heinz. 2012. Computational characterizations
of vowel harmony patterns and pathologies. In Proceedings of the 29th West
Coast Conference on Formal Linguistics, 63–71. Somerville, MA: Cascadilla
Press.
Gallagher, Gillian. 2010. Perceptual distinctness and long-distance laryngeal
restrictions. Phonology 27. 435–480.
García, Pedro & José Ruiz. 2004. Learning k-testable and k-piecewise testable
languages from positive data. Grammars 7. 125–140.
García, Pedro, Enrique Vidal & José Oncina. 1990. Learning locally testable
languages in the strict sense. In Proceedings of the Workshop on Algorithmic
Learning Theory, 325–338. Tokyo.
Gazdar, Gerald & Geoffrey K. Pullum. 1982. Natural languages and context-free
languages. Linguistics and Philosophy 4. 469–470.
Gerdemann, Dale & Gertjan van Noord. 2000. Approximation and exactness in
finite state optimality theory. In Proceedings of the 5th Meeting of the ACL
Special Interest Group in Computational Phonology, 34–45.
Gildea, Daniel & Daniel Jurafsky. 1996. Learning bias and phonological-rule
induction. Computational Linguistics 24. 497–530.
Gold, E. Mark. 1967. Language identification in the limit. Information and Control
10. 447–474.
Goldsmith, John A. 1976. Autosegmental phonology. Doctoral dissertation,
Massachusetts Institute of Technology, Cambridge, MA.
Goldwater, Sharon & Mark Johnson. 2003. Learning OT constraint rankings using
a maximum entropy model. In Jennifer Spenader, Anders Eriksson & Östen
Dahl (eds.), Proceedings of the Stockholm Workshop on Variation within
Optimality Theory, 111–120. Stockholm: Stockholm University.
Gorman, Kyle. 2013. Generative phonotactics. Doctoral dissertation, University of
Pennsylvania, Philadelphia.
Graf, Thomas. 2010a. Comparing incomparable frameworks: A model theoretic
approach to phonology. University of Pennsylvania Working Papers in
Linguistics 16, Article 10. http://repository.upenn.edu/pwpl/vol16/iss1/10.
Graf, Thomas. 2010b. Logics of phonological reasoning. Master’s thesis,
University of California, Los Angeles.
Graf, Thomas. 2013. Local and transderivational constraints in syntax and
semantics. Doctoral dissertation, University of California, Los Angeles.
Hale, Mark & Charles Reiss. 2000. Substance abuse and dysfunctionalism:
Current trends in phonology. Linguistic Inquiry 31. 157–169.
Halle, Morris. 1959. The sound pattern of Russian. The Hague: Mouton.
Halle, Morris. 1978. Knowledge unlearned and untaught: What speakers know
about the sounds of their language. In Morris Halle, Joan Bresnan & George
Miller (eds.), Linguistic theory and psychological reality, 294–303. Cambridge,
MA: MIT Press.
Halle, Morris & Jean-Roger Vergnaud. 1987. An essay on stress. Cambridge, MA:
MIT Press.
Hansson, Gunnar. 2001. Theoretical and typological issues in consonant harmony.
Doctoral dissertation, University of California, Berkeley.
Hansson, Gunnar. 2008. Diachronic explanations of sound patterns. Language
and Linguistics Compass 2. 859–893.
Hayes, Bruce. 1986. Assimilation as spreading in Toba Batak. Linguistic Inquiry
17. 467–499.
Hayes, Bruce. 1995. Metrical stress theory. Chicago: University of Chicago Press.
Hayes, Bruce, Robert Kirchner & Donca Steriade (eds.). 2004. Phonetically-based
phonology. Cambridge: Cambridge University Press.
Hayes, Bruce & Colin Wilson. 2008. A maximum entropy model of phonotactics
and phonotactic learning. Linguistic Inquiry 39. 379–440.
Heinz, Jeffrey. 2007. The inductive learning of phonotactic patterns. Doctoral
dissertation, University of California, Los Angeles.
Heinz, Jeffrey. 2009. On the role of locality in learning stress patterns. Phonology
26. 303–351.
Heinz, Jeffrey. 2010a. Learning long-distance phonotactics. Linguistic Inquiry 41.
623–661.
Heinz, Jeffrey. 2010b. String extension learning. In Proceedings of the 48th Annual
Meeting of the Association for Computational Linguistics, 897–906. Uppsala:
Association for Computational Linguistics.
Heinz, Jeffrey. 2014. Culminativity times harmony equals unbounded stress. In
Harry van der Hulst (ed.), Word stress: Theoretical and typological issues,
255–275. Cambridge: Cambridge University Press.
Heinz, Jeffrey. 2016. Computational theories of learning and developmental
psycholinguistics. In Jeffrey Lidz, William Synder & Joe Pater (eds.), The
Oxford handbook of developmental linguistics. Oxford: Oxford University
Press.
Heinz, Jeffrey & William Idsardi. 2011. Sentence and word complexity. Science
333. 295–297.
Heinz, Jeffrey & William Idsardi. 2013. What complexity differences reveal about
domains in language. Topics in Cognitive Science 5. 111–131.
Heinz, Jeffrey, Anna Kasprzik & Timo Kötzing. 2012. Learning with lattice-
structured hypothesis spaces. Theoretical Computer Science 457. 111–127.
Heinz, Jeffrey & Regine Lai. 2013. Vowel harmony and subsequentiality. In András
Kornai & Marco Kuhlmann (eds.), Proceedings of the 13th Meeting on the
Mathematics of Language (MoL 13), 52–63. Sofia, Bulgaria.
Heinz, Jeffrey, Chetan Rawal, & Herbert G. Tanner. 2011. Tier-based strictly local
constraints for phonology. In Proceedings of the 49th Annual Meeting of the
Association for Computational Linguistics, 58–64. Portland, OR: Association
for Computational Linguistics.
Higuera, Colin de la. 1997. Characteristic sets for polynomial grammatical
inference. Machine Learning 27. 125–138.
Howard, Irwin. 1972. A directional theory of rule application in phonology. Doctoral
dissertation, Massachusetts Institute of Technology, Cambridge, MA.
Hulden, Mans. 2009. Finite-state machine construction methods and algorithms for
phonology and morphology. Doctoral dissertation, University of Arizona,
Tucson.
Hulst, Harry van der & Jeroen van de Weijer. 1995. Vowel harmony. In John A.
Goldsmith (ed.), The handbook of phonological theory, 495–534. Oxford:
Blackwell.
Humboldt, Wilhelm von. 1999. On language. Edited by Michael Losonsky.
Translated by Peter Heath. Cambridge: Cambridge University Press.
Originally published 1836.
Hyman, Larry M. 1998. Positional prominence and the “prosodic trough” in Yaka.
Phonology 15. 41–75.
Hyman, Larry M. 2007. Kuki-Thaadow: An African tone system in Southeast Asia.
Berkeley Phonology Lab Annual Report. University of California, Berkeley:
Department of Linguistics.
Hyman, Larry M. 2011. Tone: Is it different? In John A. Goldsmith, Jason Riggle, &
Alan C. L. Yu (eds.), The handbook of phonological theory, 2nd edn. 197–238.
Oxford: Wiley-Blackwell.
Hyman, Larry M. 2014. How autosegmental is phonology? The Linguistic Review
31. 363–400.
Hyman, Larry M. 2009. How (not) to do phonological typology: The case of pitch-
accent. Language Sciences 31. 213–238 (Data and Theory: Papers in
Phonology in celebration of Charles W. Kisseberth).
Idsardi, William. 1998. Tiberian Hebrew spirantization and phonological
derivations. Linguistic Inquiry 29. 37–73.
Idsardi, William J. 2000. Clarifying opacity. The Linguistic Review 17. 337–350.
Jäger, Gerhard. 2002. Some notes on the formal properties of bidirectional
optimality theory. Journal of Logic, Language, and Information 11. 427–451.
Jardine, Adam. 2016a. Computationally, tone is different. Phonology 32. 247–283.
Jardine, Adam. 2016b. Locality and non-linear representations in tonal phonology.
Doctoral dissertation, University of Delaware.
Jardine, Adam, Jane Chandlee, Rémi Eyraud & Jeffrey Heinz. 2014. Very efficient
learning of structured classes of subsequential functions from positive data. In
Alexander Clark, Makoto Kanazawa & Ryo Yoshinaka (eds.), Proceedings of
the 12th International Conference on Grammatical Inference (ICGI 2014), vol.
34, 94–108. JMLR: Workshop and Conference Proceedings.
Jardine, Adam & Jeffrey Heinz. 2015a. A concatenation operation to derive
autosegmental graphs. In Proceedings of the 14th Meeting on the
Mathematics of Language (MoL 2015), 139–151. Chicago.
Jardine, Adam & Jeffrey Heinz. 2016. Learning tier-based strictly local languages.
Transactions of the Association for Computational Linguistics 4. 87-98.
Johnson, C. Douglas. 1972. Formal aspects of phonological description. The
Hague: Mouton.
Jurafsky, Daniel & James Martin. 2008. Speech and language processing: An
introduction to natural language processing, speech recognition, and
computational linguistics, 2nd edn. Upper Saddle River, NJ: Prentice-Hall.
Kager, René. 1999. Optimality Theory. Cambridge: Cambridge University Press.
Kaplan, Ronald & Martin Kay. 1994. Regular models of phonological rule systems.
Computational Linguistics 20. 331–378.
Karttunen, Lauri. 1998. The proper treatment of optimality in computational
phonology. In FSMNLP’98, 1–12. International Workshop on Finite-State
Methods in Natural Language Processing, Bilkent University, Ankara.
Kiparsky, Paul. 2000. Opacity and cyclicity. The Linguistic Review 17. 351–366.
Kisseberth, Charles W. 1970. On the functional unity of phonological rules.
Linguistic Inquiry 1. 291–306.
Kisseberth, Charles W. & David Odden. 2003. Tone. In Derek Nurse & Gérard
Philippson (eds.), The Bantu languages, 59–70. London: Routledge.
Kobele, Gregory. 2006. Generating copies: An investigation into structural identity
in language and grammar. Doctoral dissertation, University of California, Los
Angeles.
Kornai, András. 1995. Formal phonology. Outstanding Dissertations in Linguistics.
New York: Garland Publishing.
Koskenniemi, Kimmo. 1983. Two-level morphology. Publication no. 11, Department
of General Linguistics. Helsinki: University of Helsinki.
Krämer, Martin. 2003. Vowel harmony and Correspondence Theory. Berlin:
Mouton de Gruyter.
Kula, Nancy C. & Lee S. Bickmore. 2015. Phrasal phonology in Copperbelt
Bemba. Phonology 32. 147–176.
Lacy, Paul de. 2011. Markedness and faithfulness constraints. In van Oostendorp
et al. (eds.) 2011, Chapter 74.
Lai, Regine. 2012. Domain specificity in phonology. Doctoral dissertation,
University of Delaware.
Lai, Regine. 2015. Learnable vs. unlearnable harmony patterns. Linguistic Inquiry
46. 425–451.
Lautemann, Clemens, Pierre McKenzie, Thomas Schwentick & Heribert Vollmer.
2001. The descriptive complexity approach to {LOGCFL}. Journal of
Computer and System Sciences 62: 629–652.
Liberman, Mark & Alan Prince. 1977. On stress and linguistic rhythm. Linguistic
Inquiry 8. 249–336.
Lothaire, M. (ed.). 1997. Combinatorics on words. Cambridge: Cambridge
University Press.
Lothaire, M. (ed.) 2005. Algebraic combinatorics on words, 2nd edn. Cambridge:
Cambridge University Press.
Luo, Huan. 2017. Long-distance consonant harmony and subsequantiality. Glossa
2, 52.
Marlo, Michael. 2007. The verbal tonology of Lumarachi and Lunyala: Two dialects
of Luluyia. Doctoral dissertation, University of Michigan, Ann Arbor.
McCarthy, John J. 2003. OT constraints are categorical. Phonology 20. 75–138.
McCarthy, John J. 2004. Headed spans and autosegmental spreading.
Unpublished manuscript, University of Massachusetts, Amherst.
McCarthy, John J. 2007. Hidden generalizations: Phonological opacity in
Optimality Theory. London: Equinox.
McCarthy, John J. 2008a. Doing Optimality Theory. Oxford: Blackwell.
McCarthy, John J. 2008b. The gradual path to cluster simplification. Phonology 25.
271–319.
McNaughton, Robert & Seymour Papert. 1971. Counter-free automata.
Cambridge, MA: MIT Press.
Medvedev, Yu. T. 1964. On the class of events representable in a finite automaton.
In Edward F. Moore (ed.), Sequential machines: Selected Papers, 215–227.
Boston: Addison-Wesley. Originally published in Russian in Avtomaty, 1956,
385–401.
Mielke, Jeff. 2008. The emergence of distinctive features. Oxford: Oxford
University Press.
Mohri, Mehryar. 1997. Finite-state transducers in language and speech
processing. Computational Linguistics 23. 269–311.
Nevins, Andrew. 2010. Locality in vowel harmony. Cambridge, MA: MIT Press.
Odden, David. 1994. Adjacency parameters in phonology. Language 70. 289–330.
Ohala, John J. 1981. The listener as a source of sound change. In C. S. Masek, R.
A. Hendrik & M. F. Miller (eds.), Papers from the Parasession on Language
and Behavior, Chicago Linguistics Society, 178–203. Chicago: CLS.
Oncina, José & Pedro García. 1991. Inductive learning of subsequential functions.
Technical Report DSIC II-34, University Politécnia de Valencia.
Oncina, José, Pedro García & Enrique Vidal. 1993. Learning subsequential
transducers for pattern recognition tasks. IEEE Transactions on Pattern
Analysis and Machine Intelligence 15. 448–458.
Oostendorp, Marc van, Colin J. Ewen, Elizabeth Hume & Keren Rice (eds.). 2011.
The Blackwell companion to phonology. 5 volumes. Oxford: Blackwell.
Padgett, Jaye. 1995. Partial class behavior and nasal place assimilation. In
Keiichiro Suzuki & Dirk Elzinga (eds.), Proceedings of the 1995 Southwestern
Workshop on Optimality Theory. University of Arizona: Coyote Papers.
Pater, Joe. 2000. *NC. In Kiyomi Kusumoto (ed.), Proceedings of the 26th Annual
Meeting of the North East Linguistics Society, 227–239. Amherst, MA: GLSA.
Pater, Joe. 2001. Austronesian nasal substitution revisited: What’s wrong with *NC
(and what’s not). In Linda Lombardi (ed.), Segmental phonology in Optimality
Theory: Constraints and representations, 159–182. Cambridge: Cambridge
University Press.
Payne, Amanda. 2017. All dissimilation is computationally subsequential.
Language 93. e353–e371.
Popper, Karl. 1959. The logic of scientific discovery. New York: Basic Books.
Potts, Christopher, Joe Pater, Rajesh Bhatt & Michael Becker. 2008. Harmonic
grammar with linear programming: From linear systems to linguistic typology.
Rutgers Optimality Archive ROA-984.
Potts, Christopher & Geoffrey K. Pullum. 2002. Model theory and the content of
OT constraints. Phonology 19. 361–393.
Prince, Alan & Paul Smolensky. 1993. Optimality Theory: Constraint interaction in
generative grammar. Technical Report 2, Rutgers University Center for
Cognitive Science.
Prince, Alan & Paul Smolensky. 2004. Optimality Theory: Constraint interaction in
Generative Grammar. Oxford: Blackwell.
Riggle, Jason. 2004. Generation, recognition, and learning in finite state Optimality
Theory. Doctoral dissertation, University of California, Los Angeles.
Roark, Brian & Richard Sproat. 2007. Computational approaches to morphology
and syntax. Oxford: Oxford University Press.
Roche, Emmanuel & Yves Schabes. 1997. Finite-state language processing.
Cambridge, MA: MIT Press.
Rogers, James. 1998. A descriptive approach to language-theoretic complexity.
Stanford: CSLI Publications.
Rogers, James, Jeffrey Heinz, Gil Bailey, Matt Edlefsen, Molly Visscher, David
Wellcome & Sean Wibel. 2010. On languages piecewise testable in the strict
sense. In Christian Ebert, Gerhard Jäger, & Jens Michaelis (eds.), The
mathematics of language, vol. 6149 of Lecture notes in Artifical Intelligence,
255–265. Dordrecht: Springer.
Rogers, James, Jeffrey Heinz, Margaret Fero, Jeremy Hurst, Dakotah Lambert, &
Sean Wibel. 2013. Cognitive and sub-regular complexity. In Glyn Morrill &
Mark-Jan Nederhof (eds.), Formal grammar, vol. 8036 of Lecture notes in
computer science, 90–108. Dordrecht: Springer.
Rogers, James & Geoffrey K. Pullum. 2011. Aural pattern recognition experiments
and the subregular hierarchy. Journal of Logic, Language and Information 20.
 329–342.
Rose, Sharon & Rachel Walker. 2004. A typology of consonant agreement as
correspondence. Language 80. 475–531.
Rozenberg, Grzegorz & Arto Salomaa (eds.). 1997. Handbook of formal
languages: Beyond words, vol. 3. Dordrecht: Springer.
Sakarovitch, Jaques. 2009. Elements of automata theory. Cambridge: Cambridge
University Press. (Translated by Reuben Thomas from the French original,
Paris: Vuibert, 2003.)
Scott, Dana & Michael Rabin. 1959. Finite automata and their decision problems.
IBM Journal of Research and Development 5. 114–125.
Shieber, Stuart. 1985. Evidence against the context-freeness of natural language.
Linguistics and Philosophy 8. 333–343.
Sipser, Michael. 1997. Introduction to the theory of computation. Boston: PWS
Publishing.
Smolensky, Paul & Géraldine Legendre. 2006. The harmonic mind: From neural
computation to Optimality-Theoretic grammar. Cambridge, MA: MIT Press.
Suzuki, Keiichiro. 1998. A typological investigation of dissimilation. Doctoral
dissertation, University of Arizona, Tucson.
Tesar, Bruce. 1995. Computational Optimality Theory. Doctoral dissertation,
University of Colorado at Boulder.
Tesar, Bruce. 2014. Output-driven phonology. Cambridge: Cambridge University
Press.
Tesar, Bruce & Paul Smolensky. 1998. Learnability in Optimality Theory. Linguistic
Inquiry 29. 229–268.
Tesar, Bruce & Paul Smolensky. 2000. Learnability in Optimality Theory.
Cambridge, MA: MIT Press.
Thomas, Wolfgang. 1997. Languages, automata, and logic. In Rozenberg &
Salomaa (eds.) 1997, 389–455.
Turing, Alan. 1937. On computable numbers, with an application to the
Entscheidungsproblem. Proceedings of the London Mathematical Society 2.
230–265.
Walker, Rachel. 2011. Vowel patterns in language. Cambridge: Cambridge
University Press.
Wilson, Colin. 2001. Consonant cluster neutralization and targeted constraints.
Phonology 18. 147–197.
Wilson, Colin. 2003. Analyzing unbounded spreading with constraints: Marks,
targets, and derivations. Unpublished manuscript, UCLA.
Wilson, Colin. 2004. Experimental investigation of phonological naturalness. In
Proceedings of the 22nd West Coast Conference on Formal Linguistics, 534–
546. San Diego: University of California.
Anthony Brohan and Jeff Mielke
Frequent segmental alternations in P-
base 3

Abstract: P-base (Mielke 2008) is a large typological database of several

thousand phonological rules and distributions in 537 languages. In this
chapter we demonstrate how it can be applied to a wide range of questions
in phonological typology, making use of several new features of version 3.
This includes possible follow-up studies to many of the other chapters in
the volume. As a starting point, we show that the crosslinguistic frequency
of segments in segment inventories are largely similar to UPSID
(Maddieson 1984; Maddieson & Precoda 1990), a similarly-sized
genetically balanced database of segment inventories. Second, we show that
a considerable number of the phonological patterns in P-base fall into a
small number of categories defined in terms of the classes of segments
involved, the features changing, and the position of the trigger relative to
the target. For instance, regressive preconsonantal nasal place assimilation
accounts for 4.54% of the sound patterns, and because place changes are
quite rare, this constitutes more than half of all cases of place assimilation.
Other types of sound patterns are shown to be distributed in ways that are
consistent with phonetic accounts of phonological typology, e.g., consonant
epenthesis is dominated by glottals and glides, particularly in contexts that
have been argued to be the locus of epenthesis as a sound change.

1 Introduction
Much of what is currently known about phonological typology is based on
the UCLA Phonological Segment Inventory Database (UPSID), compiled
by Ian Maddieson and colleagues. The UPSID database was published with
317 languages in Maddieson’s 1984 book Patterns of sounds, later
expanded to 451 inventories (Maddieson & Precoda 1990), and now being
expanded as LAPSyD (Maddieson 2014). The availability of UPSID has
meant that studies of phonological universals have preferentially favored
segment inventories (Lindblom & Maddieson 1988; Hyman 2008) over
phonological alternations. Crosslinguistic databases of phonological
alternations have often involved custom databases of particular types of
phonological patterns such as lenition (Lavoie 2001; Kirchner 2001) and
metathesis (Hume 2004). P-base (Mielke 2008) is a general database of
phonological alternations in several hundred languages, but until recently it
was organized in terms of classes of sounds, making it difficult to use to
study phonological patterns more generally. We report a study of frequent
phonological patterns in a newly reorganized version of P-base.

2 P-base
P-base was compiled from descriptions available on library stacks at the
Ohio State University and Michigan State University. It was originally
collected to test models of natural classes (Mielke 2005, 2008), and was
structured according to classes involved in phonological patterns (6170
classes involved in alternations, and about 3000 others involved in
phonotactic restrictions. These alternations include general phonological
patterns as well as morphologically conditioned patterns.
We report on a reorganized P-base that is structured according to
phonological alternations and distributional restrictions. As an example of
this reorganization, (1) shows how Japanese high vowel devoicing (Vance
1987) was entered in the original P-base as two phonologically active
classes, and (2) shows how the same phonological pattern is entered in P-
base 3, as a single phonological pattern involving two classes of sounds.

In the P-base 1 version, a user could easily query the classes of sounds
involved in patterns such as this one, but the details of the alternation were
stored in text strings, leaving it up to the user to perform keyword searches
in order to identify similar patterns, and even to associate these two entries
(the target and the trigger) with one another.

(2) Japanese high vowel devoicing in P-base 3:

p, t, k, s, ∫ , h
jpn 6795 {i,ɯ} → {i ̥,ɯ̥}/{p,t,k,s,ʃ,h} __ { }
#

P-base 3 is a MySQL database with a web interface.91 It contains

phonological patterns from 630 language varieties, corresponding to 537
distinct ISO639.3 codes, and matched to AUTOTYP family trees (Bickel &
Nichols 1996). Sound pattern entries contain fields for input, output,
segmental context, prosodic domain, morphological effects, prosody, and
optionality.
Automated querying of sound pattern properties is enhanced with
feature analysis, which is implemented by associating each of the 1274
distinct IPA transcriptions with a vector of feature values in one of several
feature systems. P-base currently supports seven feature systems: a
descriptive feature system, the systems described in Preliminaries to the
analysis of speech (Jakobson, Fant & Halle 1952), The sound pattern of
English (Chomsky & Halle 1968), Problem book in phonology (Halle &
Clements 1983), and three versions of Unified Feature Theory (Clements &
Hume 1995). Mielke et al. (2011) compares the performance of six of these
systems in characterizing the phonologically active classes of sounds in P-
base. This example and the rest of this chapter use the familiar Halle &
Clements (1983) feature system.
IPA symbols are used to explicitly represent the segments that undergo,
condition, and result from phonological patterns. Types of sound patterns
such as assimilation are recognized by converting the segment-based
description to a feature-based description. A feature analysis of the pattern
in (2) is shown in (3). The feature analysis is computed automatically by P-
base, meaning that the user can symbolically search for patterns involving
various segments, features, and types of changes.
Associating each of the segments in (2) with its feature representation
automates a big part of the process of analyzing and classifying the
alternation. An algorithm determines common features between each cell of
the structural description, allowing the sound patterns to be described in
terms of natural classes. This natural class analysis is instantiated on a given
language’s inventory, so we are able to query for patterns which target a
particular natural class given a particular language. The feature analysis
also determines which features in an alternation are implicated in a change
by looking for differences between the common features of the input and
output.92 Comparing the set of changes detected with common features in
the environment can detect assimilations. So, although the change
(devoicing) is not explicit in (2), this information is easily extracted by
observing that output segments differ from the input segments only by
being devoiced. Assimilation is recognized similarly, by observing that the
feature value that describes the change is shared by the left or right context,
or both.
This automated feature analysis is available through P-base 3’s online
query tool, and here it contributes to classifying 4560 phonological
alternations in the database. In addition, there are approximately 1300
distributional patterns in P-base 3 that we are not considering here. P-base 3
entries for an additional 1000 or so phonological alternations from P-base 1
remain to be completed. These are mostly patterns that do not easily fit the
basic segmental pattern template (e.g., metathesis, coalescence, prosodic
patterns). The purpose of this chapter is to describe the most frequent
patterns, so these are not very consequential here.

3 Database comparison and sampling

When considering questions about what properties of phonological systems
are crosslinguistically frequent and why, it is useful to be able to determine
whether a feature is frequent because of essentially universal forces that
shape sound systems, or due to shared inheritance within a particularly
well-attested language family, or to areal factors. In UPSID, shared
inheritance has been dealt with by only including languages that are
sufficiently distant from languages that are already included.
The phonological patterns in P-base are the result of an exhaustive
search of descriptive grammars available in a large university library. As
such, it is not a genetically balanced sample, and its bias toward better-
described language families is discussed in Mielke (2008: 47–48). An
UPSID-like inclusion criterion would reduce the number of included
languages considerably. Our approach has been to include all available
language descriptions and then take genetic balance into account when
interpreting results. LAPSyD (Maddieson 2014) has followed a similar
approach in expanding upon UPSID, by including more closely related
languages in the database and allowing the user to employ appropriate
criteria for considering their relationships.
Since the 451-language version of UPSID (Maddieson & Precoda 1990)
includes about the same number of languages as P-base, and is genetically
balanced, we turn to segment inventories as a basis for comparison between
the samples of the two databases, before considering the role of genetic
sampling in a typological survey of phonological patterns. Figure 6.1
provides an overview of the distribution of consonants in UPSID and P-
base. These are density-equalizing maps (Gastner & Newman 2004) of a
chart of basic consonants (Mielke 2017), where the area of each cell reflects
the number of languages having the sound in their inventories. The two
charts tell similar stories about frequent and infrequent consonants. [m] is
the most frequent consonant in both samples, followed by [k] and [j].
Differences that are apparent include the higher frequency of approximants

such as [β]and[ ∂ ] in UPSID (which may really

T
T

be a difference in the methods for selecting representative IPA symbols),

and the higher frequency of [r] and voiced stops in P-base.
 Figure 6.1: Frequency counts of basic consonants in UPSID and P-
base inventories.

Since the genetic relationships between the languages in both samples

are known (or at least posited), it is possible to compensate for genetic
imbalances by applying a sampling algorithm such as G-sampling (Bickel
2008). The frequencies depicted in Figure 6.1 are not adjusted for genetic
relationships, meaning that the frequency of a segment in a sample is
literally the number of languages in the sample that have that segment. This
technique does not distinguish between a case where a segment has been
innovated many times in many different language families, and a case
where a segment has been innovated once in one family and all the instance
of that segment are found in that family. UPSID minifies this concern by
not sampling closely related languages, but it is still vulnerable to over-
counting segments found in large families.
In short, G-sampling works by descending a language family tree and
counting only the number of subgroups where the presence or absence of a
linguistic feature is significantly skewed in comparison to the higher level.
If all the members of a language family have a particular segment, it is
counted once for that family. If only one language in a large family has a
particular segment, it is counted once for that language, and the rest of the
family (if it is significantly skewed relative to the other language families,
would be counted once as not having the segment). See Bickel (2008: 7–8)
for a more explicit statement of the algorithm.
 Figure 6.2: Effects of G-sampling on segment frequencies.

Figure 6.2 shows the result of applying G-sampling to the UPSID and
P-base samples of inventories. The coordinates of each IPA symbol in the
figure represent the proportion of P-base and UPSID inventories containing
the segment. The gray arrow originating from each IPA symbol represents
the effect of G-sampling: an arrow pointing down and to the left means that
G-sampling reduced the proportion of languages having the segment
(indicating that frequent occurrence in particular language families led to its
higher raw numbers). The fact that the segments are close to the diagonal
indicates that they are generally similar in both databases. The vast majority
of distinct IPA transcriptions are attested in only a handful of languages in
either sample. Segments appearing substantially above the diagonal are
more frequent in P-base, and segments appearing substantially below the
diagonal are more frequent in UPSID. Where they diverge, UPSID can be
considered to be more reliable, because the primary difference is the
sampling techniques.
A major effect of G-sampling is to move everything closer to the
middle, along the diagonal. If G-sampling were having the effect of
reconciling differences between the UPSID and P-base samples, the arrows
would be pointing closer to the diagonal, not parallel to it. The two
databases are the most divergent on [r], and G-sampling indicates that this
is not due to sampling. On the other hand, [v] and [ɾ] both occur in about
25% of languages, but G-sampling affects them in opposite ways, that are
similar across the two databases, suggesting that the frequency of [v] is
overestimated due shared inheritance, while the frequency of [ɾ] is
underestimated due to language families that lack it.
The rest of this chapter is concerned with phonological alternations, for
which there is no other database to compare to. The fact that segment
frequencies in P-base are similar to UPSID, which was collected with
genetic balance in mind, lead us to be less concerned about genetic balance
in our interpretation of other patterns in the P-base language sample. While
it would be possible to apply G-sampling to our counts of phonological
alternations, we elect not to, on the basis of the inventory comparison, and
because G-sampling works best for phenomena whose presence or absence
is easily counted.
Counting the absence of a feature is where phonological patterns
become challenging to genetic sampling techniques. It is reasonable to
expect language descriptions to report binary features such as verb-object
vs. object-verb word order in a consistent way, and only a little less
reasonable to expect phonological descriptions to indicate in similar terms
whether a particular type of phonological segment is present or absent in
the inventory of a language. By comparison, descriptions of phonological
alternations and distributional restrictions vary widely in their
exhaustiveness, and it is much harder to interpret a pattern that is
unreported as truly absent from the language. A detailed phonological
description may include a range of postlexical rules that require some
phonetic sophistication to recognize, and the fact that such a pattern is not
included in a description of a related language is not conclusive.93 To be
confident enough in the absence of a pattern to conclude that it is innovative
in a related language would require going and looking for each pattern in
each language where it wasn’t reported. We have not done this.

4 Overview of segmental changes

Figure 6.3 provides an overview of all the changes involving basic
consonants and vowels. Input-output mappings between pairs of segments
are indicated by arrows. The size of the arrow represents how many time
each input-output mapping is observed. The arrows are superimposed on a
density-equalized IPA chart, where cell area indicates the G-sampled
frequency of occurrence in P-base inventories. Vowels and consonants have
been placed in the same chart, with front unrounded vowels aligned to
palatal consonants and back rounded vowels aligned to labiovelar
consonants. Cells and arrows in Figure 6.3 are colored by place of
articulation. Arrows reflect the properties of the input, so arrows that extend
over a region of a different color indicate changes of place of articulation.

Figure 6.3: All input-output mappings colored by place.

A striking observation from Figure 6.3 is that there are few kinds of
frequent place changes (represented by horizontal or diagonal arrows
passing over a different color). The most salient manner changes represent
voicing and devoicing of obstruents, spirantization of stops, nasalization of
voiced stops, /ɾ/, and /l/, gliding of high vowels, and height changes among
vowels. The most frequent place changes involve nasal changing into other
nasals, palatalization of the consonants /s k ɡ/, debuccalization, including
/k/ spirantizing to [h], and vowels changing backness and rounding,
especially by changing to [ə].
Another set of place changes requires comment: /w/ is involved in many
place changes with bilabial stops and voiced labial fricatives. This reflects
the arbitrary choice of putting labiovelar /w/ near the velars instead of the
labials. Additionally, we do not have much articulatory evidence for how
many instances of /w/ really involve a velar constriction, and how many
should really be transcribed as /β̞/ and placed in the bilabial column. UPSID
has transcribed more approximants as /β̞/ than P-base has. /k/ → [h] is
notable because most instances of spirantization do not involve place
changes, but /k/’s fricative counterpart /x/ is considerably less frequent than
/h/ in segment inventories. Reinterpretation of /k/ lenition as /k/ → [h] in
languages with /h/ but no /x/ is consistent with the principle of structural
analogy (Blevins 2004: 154):
In the course of language acquisition, the existence of a (non-ambiguous) phonological
contrast between A and B will result in more instances of sound change involving
shifts of ambiguous elements to A or B than if no contrast between A and B existed.

The input-output mappings of Figure 6.3 are summarized in Figures 6.4 and
6.5. In both figures, the background color indicates the place or
manner/voicing of the output segment, and color of the outer ring of the
balloon indicates the place or manner/voicing of the input segment. The
area of each colored outer ring represents the number of structure
preserving input-output mappings, and the area of each inner white circle
represents the number of structure changing input-output mappings. The
numbers used to generate the balloons are segment counts, not pattern
counts.
The manner mappings in Figure 6.4 are mostly off the diagonal,
meaning that there are few changes that do not affect manner or voicing.
These are dominated by the categories involved in place changes observed
above (nasals, palatalization of obstruents, and vowel changes). All of these
are about evenly split between structure changing and structure preserving.
The leading (off-diagonal) manner changes include obstruent voicing,
which is mostly structure changing, and obstruent devoicing, which is
mostly structure preserving. This asymmetry can potentially be explained
entirely in terms of the rarity of voiced obstruents without voiceless
counterparts.

Figure 6.4: Changes by manner of articulation and voicing. Input is on

the left and output is on the top. Area of colored outer ring represents
the number of structure preserving changes in the category. Area of
inner white circle represents the number of structure changing changes
in the category.

Vowel height changes are abundant, especially between high and mid
vowels, and are more likely to be structure preserving than non-height
changes. High vowels frequently turn into glides, and the converse is less
frequent. Stops change manner in many different ways, and a salient fact is
that stop changes that result in a fricative, flap, implosive, or ejective are
typically structure changing, but stop changes that result in a nasal, trill,
glide, or lateral approximant are more likely to be structure preserving.
 The place mappings in Figure 6.5 mostly fall into one of three types.
First, many are on the diagonal, meaning place is unchanged (as seen above
in Figure 6.4). Second, many involve pairings of bilabial, alveolar, palatal,
and velar, which are by far the most frequent nasal places of articulation (as
can be seen in Figure 6.3, and abovein Figure 6.1). This is because nasal
place assimilation accounts for a large portion of place changes. Third,
there are many vowel backness/rounding changes, which in this figure are
overlaid on palatal, labialpalatal, velar, and labial- velar. The infrequency of
changes involving labialpalatal/front rounded segments reflects the rarity of
these segments to begin with. Similarly, the fact that place changes resulting
in bilabial or alveolar outputs are mostly structure preserving reflects the
fact that almost all inventories contain /m/ and /n/, so nasal place
assimilation to these places of articulation is almost never structure
changing.

Figure 6.5: Changes by place of articulation. Input is on the left and

output is on the top. Area of colored outer ring represents the number of
structure preserving changes in the category. Area of inner white circle
represents the number of structure changing changes in the category.
 Changes from velar to glottal place are mostly structure preserving,
consistent with the idea that /k/ → [h] often appears as a structure
preserving alternative to /k/ → [x]. /k/ → [ʔ] is also usually structure
preserving, and so are most other place changes resulting in glottals. The
most frequent of the rest is alveolar- glottal, which is mostly /s/ → [h]94.
Blevins (2008b) draws a connection between structure preservation and
different outcomes of preconsonantal /s/ weakening in Romance languages,
mediated by the role of contrast in the perceptibility of [h]: s > h / __C is a
recurrent sound change, seen in many varieties of Spanish, which has /h/ as
a phoneme. French, which does not have /h/, instead underwent the sound
change s > Ø / __C (Hall 1949). This is consistent with the finding that h ~
Ø is highly confusable for speakers of languages that lack the contrast,
especially in preconsonantal position (Mielke 2003).
This analysis of input-output mappings has illustrated that some
changes are far more frequent than others. A logical next step to
understanding these facts is to examine the types of sound patterns these
segmental changes are involved in.

5 Top-down analysis
We will analyze the phonological alternations in P-base in two ways. The
first one involves labeling the patterns according to established
phonological criteria and finding the most frequently co-occurring labels.
The second one will involve clustering the patterns in order to induce
categories that may not be detected through expected labels.

5.1 Methods
The 4560 phonological alternations were partitioned according the criteria
listed in Tables 6.1–6.3 on the basis of the automated phonological feature
analysis provided by P-base, in order to count the most frequently-occurring
patterns.
All 4560 patterns under consideration meet the criteria for Rule (having
an input-output mapping), and do not meet the criteria for Distribution
(describing what can or cannot occur in a particular context) so we omit
these criteria.95
Table 6.1: Phonological pattern labels for contexts

Tables 6.1–6.3 show the labels used to classify phonological patterns, and
the number of matching patterns for each label. These labels were selected
in order to systematize what we as phonologists expect to be frequent, and
what we find to be frequent on the basis of our experience querying P-base.
Most phonological alternations in P-base fall into more than one of the
categories, and in the next section we will examine the co-occurrence of
labels. To select label co-occurrences to report, we conducted a chi-squared
test for the combinations of each process label with every other label, in
order to identify pairs of labels that co-occur more than chance.
An example is shown in (4), comparing the co-occurrence of the labels
Deletion and I=h.

For this test, χ2(1) = 123.31, exceeding 6.635, which is the critical value for
α = 0.01. This captures the fact that the 62 cases of /h/ deletion are more
than would be expected if the two labels co-occurred at random. Rather, /h/
is the input for 2.4% of all patterns, but it is the input for 8.0% of all
deletion patterns, and /h/ deletion accounts for 57% of all occurrences of /h/
as input. Trivial co-occurrences were omitted (e.g., I=Nasal implies I=C,
and I=Nasal is not expected to be independent from O=Nasal) and labels
which imply the same subset of labels were compared on that subset (e.g.,
Vowel Raising and I=a were compared on the basis of the 1383
V→patterns, not all 4560 patterns, because both labels imply that the input
is a vowel).
Table 6.2: Phonological pattern labels for inputs and outputs
Table 6.3: Phonological pattern labels for processes
5.2 Results by pattern type
Figures 6.6, 6.7, 6.10, and 6.12 illustrate some of the patterns in the
structure of the label counts in Tables 6.1–6.3. Each circle’s area represents
the number of patterns matching one or more labels. The rectangular nodes
are broad categories of patterns, and edges connect groups that are in a
subset-superset relationship. The purpose of displaying them this way is for
the relative size of the circles to provide a visual gestalt impression of what
the bulk of phonological patterns are.
From Tables 6.1–6.3, above, a few facts are apparent, each investigated
more closely below. Deletion is 2.4 times as frequent as epenthesis, and
glides and glottals make up a disproportionately large portion of the cases
of epenthesis, as compared to deletion. Assimilatory and non-assimilatory
changes each make up a little over a third of phonological patterns in the
sample. Among assimilatory changes, regressive assimilation is more
frequent than progressive assimilation. Regressive assimilation accounts for
more than twice as many patterns as progressive assimilation. These
asymmetries between formally symmetrical processes such deletion and
epenthesis and between progressive and regressive assimilation are an
important contribution of phonological typology to phonological theory: an
important goal of phonological theory is to account for why some things
happen more often than other things, whether it is because of markedness,
the role of sound change, or something else.

5.2.1 Epenthesis and deletion

Figure 6.6 illustrates some major patterns for epenthesis. About one third of
epenthesis patterns in P-base involve epenthetic vowels, and 64% of these
are one of the vowels [i u ə]. Among many factors thought to contribute to
epenthetic vowel quality, (e.g., Hume & Bromberg 2005), it has often been
observed that epenthetic vowels tend to be short and otherwise perceptually
non-salient. High vowels are often shorter than lower vowels (Catford
1977; Maddieson 1997), so epenthetic high vowels are consistent with the
general idea that phonological repairs make minimal changes (Steriade
2001). [i u] are likely the shortest underlying vowels in the inventories of
many languages in P-base, and while [ə] occurs in fewer inventories than [i
u], [ə] epenthesis accounts for the majority of vowel epenthesis patterns that
are structure-changing with respect to segments (i.e., [ə] epenthesis in
languages without /ə/).96
There are 210 cases of epenthesis of glides, glottals, and other
consonants. Of these, 43.7% are glide epenthesis, 33.3% are glottal
epenthesis, and all other types of consonants combined amount to 24.0%.
This is consistent with two recent accounts of the typology of consonant
epenthesis. Vaux (2002) showed that consonant epenthesis is not restricted
to a few default consonants, but that nearly every familiar consonant is
epenthetic in at least one language, and that many of the more obscure
epenthetic consonants (such as [ɹ] in some varieties of English) are due to
restructuring of deletion patterns. Blevins (2008a) argued that the record of
sound changes supports two basic sources of epenthetic consonants:
reinterpretation of vowel-vowel sequences as vowel-glide-vowel, and the
phonologization of naturally occurring irregular phonation at prosodic
boundaries (Pierrehumbert & Talkin 1992) as glottal consonants. Blevins
attributes epenthetic consonants other than glides and glottals to complex
and/ or unnatural sources such as subsequent glide fortition and the
restructuring of consonant deletion patterns. Blevins’ account predicts that
non-glide/non-glottal epenthesis will be sparse, because they require
telescoping or restructuring of existing patterns, and the particular
epenthetic consonant involved is dependent on fortition and deletion
patterns occurring in any particular language, which are likely not to be
nearly as specific as the sources of epenthetic glides and glottals. The
contexts in which glottals and glides are epenthesized in P-base is also
consistent with the historical account. 61% of glide epenthesis is
intervocalic, and most of the rest is either prevocalic or postvocalic. It
generally is not sensitive to word boundaries. On the other hand, only the
word-initial prevocalic context is significantly associated with glottal
epenthesis, accounting for 33% of [ʔ] epenthesis.

Figure 6.6: Summary of epenthesis patterns.

 Figure 6.7 illustrates some major patterns for deletion. 37% of deletion
patterns involve vowels, which is slightly higher than the proportion of
epenthesis patterns involving vowels. A major difference between vowel
epenthesis and vowel deletion is that epenthesis often involves a single
default vowel, while deletion may target particular vowels or all of the
vowels. Low vowels play a bigger role in deletion than in epenthesis.
Unsurprisingly, wholesale vowel deletion typically occurs next to vowels,
especially before other vowels, and word-finally.

Figure 6.7: Summary of deletion patterns.

The biggest difference between consonant deletion and consonant

epenthesis is that non-glide/nonglottal consonants get deleted in large
numbers. Much of this applies to large groups of consonants, particularly in
consonant clusters or between vowels, and word-finally. Stops,97 nasals,
and glottal consonants are each specifically targeted by 10% of deletion
patterns, and glides account for 12%. Nasals and stops are lumped together
with other consonants as “Other” because they are not significantly more
frequent as deletion targets than they are as targets in general, but the
distribution of nasal and stop deletion can be seen below in Figure 6.8.
 Glide deletion is triggered by an adjacent vowel 43% of the time and
triggered by an adjacent consonant 35% of the time. This is relevant for
considering multiple ways for deletion and epenthesis contexts to be
associated: they can be unrelated, such as when there is a specific phonetic
source for deletion or epenthesis with, they can be opposite, as in the case
of vowels (deleted next to vowels and epenthesized next to consonants), or
they can be the same, because deletion and epenthesis both involve the
same phonological or phonetic ambiguity (i.e., XY and Y are confusable, so
X is deleted before Y in some languages and inserted before Y in some
languages). Glide deletion appears to be a mixture of at least two of these.
Vowel-adjacent glide deletion could involve the same phonological
ambiguity that enables glide epenthesis, and consonant-adjacent glide
deletion could involve completely independent factors.
Like glides, glottals are deleted only about as frequently as they are
epenthesized (unlike all other consonants and vowels), which is still quite
frequent. However, while [ʔ] dominates glottal epenthesis, [h] dominates
glottal deletion. This is understandable in terms of the phonetic bases for [ʔ]
epenthesis and /h/ deletion. Glottal gestures at prosodic boundaries provide
fodder for phonological [ʔ] epenthesis, and probably bear less resemblance
to [h], but /h/ is a deletion target due to its perceptual weakness, both in
terms of its tendency to be reduced in production and its weak acoustic cues
even when produced successfully (e.g., Mielke 2003).
Figure 6.8 shows the distribution of epenthesis and deletion across
contexts. Consistent with the greater frequency of deletion overall, the norm
is for epenthesis to be less frequent then deletion in each specific case, with
notable exceptions mostly discussed above, such as glottal epenthesis at
prosodic boundaries and glide epenthesis next to vowels. The most
important environment for stop and nasal deletion is preconsonantal, which
is broadly consistent with consonant clusters as a locus for consonant
change, and with perceptual accounts of consonant cluster simplification
(Coˆte´ 2000; Steriade 2001). Vowel deletion is triggered both by
consonants and vowels, probably for very different reasons.
Deletion generally targets segments that are present in lexical
representations, and by definition it is structure preserving in the segmental
sense, because while it may produce novel sequences, it cannot produce
novel segments (because it produces no segments). Epenthesis, on the other
hand, may be either structure preserving or structure changing, and it is
useful to examine which kinds of epenthesis are most likely to result in
novel segments in the language where they occur. This is illustrated in Figur
e 6.9 98.

Figure 6.8: Epenthesis and deletion by segment and context.

Figure 6.9: Structure-changing epenthesis by segment and context.

Among the glottals, [ʔ] epenthesis is frequently structure changing, but

[h] is rarely structure changing. Glide epenthesis is structure changing at a
similar rate across context, and the two main contexts where it is always
structure preserving (V_ and C_V) may reflect a tendency to describe glide
epenthesis in these contexts as diphthongization and palatalization,
respectively. Most of the cases of structure changing vowel epenthesis
involve epenthetic [ə], which is consistent both with the idea that an
epenthetic vowel will have a neutral vowel quality if it is not being
interpreted as a lexical vowel, and that most languages already have /i/ and
/u/, so there are few opportunities for them to be introduced to a language
by epenthesis.

5.2.2 Assimilation

Figure 6.10 illustrates some major patterns for assimilation, which accounts
for 35.8% of the sound patterns in P-base, or a little more than half of the
segmental changes (phonological alternations that are not deletion or
epenthesis). 24.2% of assimilation patterns change vowels into other
vowels, and 24.6% of these are nasalization. The rest are vowel quality
changes that are predominantly conditioned by other vowels. 51.4% of
consonant assimilation patterns occur within consonant clusters, which is
reminiscent of consonant deletion.
 Consonant-consonant assimilation patterns are even more biased toward
regressive assimilation (2.9 times as many regressive than progressive CC
assimilations, vs. a ratio of 2.0 for assimilations in general). The elephant in
the room for regressive vs. progressive CC assimilation is regressive nasal
place assimilation, which accounts for 46.4% of all regressive CC
assimilation. While this is consistent with the perceptual account of CC
repair strategies, it can also be attributed in part to nasal-consonant clusters
simply being more frequent than consonantnasal clusters. Total assimilation
accounts for 14.6% of regressive CC assimilation, and a negligible part of
progressive CC assimilation. Nasal place assimilation and total assimilation
account for much of the directional asymmetry in CC assimilation. If nasal
place assimilation is excluded, the ratio of regressive to progressive
assimilation in consonant clusters drops from 2.9 to 1.7.
Voicing and devoicing together account for 32.9% of CC assimilation.
Devoicing is biased toward regressive (with a ratio of 2.7) while voicing is
not. In addition to consonant cluster assimilation, intervocalic voicing
accounts for 32.3% of bidirectional assimilation patterns, and it is the
largest recognizable subgroup within prevocalic and postvocalic
assimilatory consonant changes, although it is not significantly more
frequent in those contexts. Palatalization and lenition are ill-served by the
split into assimilatory and non-assimilatory patterns, because only some of
each is formally assimilation in the feature system we are using. We have
excluded palatalization from the assimilation analysis (because we are using
a feature system that does not readily capture it as assimilation) and
included it below in Figure 6.12 with other miscellaneous patterns. If
included here, palatalization would increase the number of prevocalic
regressive assimilation among consonants. Many intervocalic lenition
patterns are assimilatory, and these are included in both figures
Figure 6.11 shows assimilatory changes by feature and context. Unlike
the previous balloon plots, the [+] and [−] values are not superimposed, and
the visible area of the black outer ring represents the number of times the
[−] value of the feature spreads.

5.2.3 Other recurrent changes

Figure 6.12 illustrates some major types of sound patterns not addressed in
the preceding sections. Lenition (including voicing, spirantization, and
debuccalization) accounts for 628 patterns (13.77%), many of which are
also classified as assimilation. Word-final devoicing occurs 48 times
(1.05%). Final devoicing can be considered to be assimilatory in utterance-
final position, and word- final devoicing has been analyzed as a
generalization of utterance-final deletion (see, e.g., Vennemann 1974;
Blevins 2006; Myers 2010, 2012). A possible connection between
devoicing in clusters and at edges is that assimilatory final devoicing is
regressive, and in consonant clusters, regressive devoicing assimilation is
quite a bit more frequent than progressive devoicing assimilation, while
voicing is symmetrical.
 Figure 6.11: Vowel (left) and consonant (right) assimilation by feature
and context.

Palatalization (defined in terms of the change, not the trigger) has 145
occurrences (3.18%), and 82 of these are prevocalic. Many of these are
triggered specifically by front and/ or high vowels, as expected. While we
did not code for subsets of vowels in this analysis, this is certainly relevant
for palatalization. There are 116 cases of vowel gliding, where a vowel
becomes a vocalic glide (or 2.54% of all patterns). For comparison, there
are only 17 instances where glides become vowels. 73% of vowel gliding
instances occur prevocalically and 66% involve only high vowels.
In the top-down analysis, the label “Lenition” has been used to
characterize a set of pre-determined changes (Degemination, Spirantization,
Debuccalization, Voicing, Vowel Shortening). We can instead unpack this
notion of lenition and look at changes which occur intervocalically, which
involve changes of the feature [sonorant], [voice], and [continuant]. Below
is a view of the feature changes, highlighting common pathways of
intervocalic lenition (Figure 6.13).
Common paths which emerge from this picture are the attestation of
stop voicing, voiced and voiceless stop spirantization, and voiced stop
flapping. Other changes along this lattice (gliding of voiceless stops (e.g., /p
/ → [w] / V——V)) are also seen to be represented. This lattice illustrates
that most of these lenition patterns typically make a minimal feature
change.
 Figure 6.13: Changes involved in intervocalic lenition.

6 Bottom-up analysis
A bottom-up analysis of the patterns in P-base was conducted to determine
whether there were significant groupings of features patterning together that
our top-down analysis was missing. The bottom-up analysis generates
generalizations based on the data, and induces potential categories of
patterns from the observed set of phonological patterns.
6.1 Methods
We employed Multiple Correspondence Analysis, a technique of vector
space reduction for multinomial data. This technique summarizes a large set
of variables with a small number of factors (similar to Principal
Components Analysis). This essentially can be used to reduce the data set
into a smaller number of factors, which can be inspected to see how
features pattern together. Patterns which are associated with each other
show up with similar weightings in their factors, for instance features like
[±tense] and [±ATR] generally group together in their behavior; so they
have a similar weighting in the factor space.
Each factor can be interpreted as summarizing a dimension along which
the set of phonological rules vary. The first factor splits the set of patterns
into left-triggered and right-triggered rules. We can inspect the grouping of
features in this factor space to see what relations exist between features,
with the idea that features which are close to each other generally pattern
together. We achieve this by running a hierarchical cluster analysis on a
trimmed set of features (features with a low factor loading have little
predictive value in the MCA analysis – they don’t pattern with other
features in a predictable manner). Hierarchical clustering works to identify
features which are close together in the factor loading space; which
effectively yields a tree of features which tend to co-occur with each other.
Two analyses were undertaken, the first takes binary features referring
to P-base’s feature analysis of the input, output, change, environment and
detected assimilations (to the left and right).99 Here the number of features
is tremendous (46 possible values in 7 possible positions), and after filtering
features which are farther away than 1.2 in our factor loadings we have the
following cluster dendrograms.

6.2 Results
With this set of data, the primary axis of variation is between left-triggered
and right-triggered rules (Figures 6.14 and 6.15). Inspecting the diagram
yields predictable clumps of features which pattern together generally
around triggering context, and a number of processes tend to cluster (such
as lenition) together. Sensible labels were attached to groups of clusters
based on querying P-base and seeing how these features which describe
patterns tend to group together in terms of set of patterns described.

Figure 6.14: Progressive Feature Cluster Dendrogram.

Multiple correspondence analysis seeks to build factor weights based on the

predictive quality of features, so this approach tends to under-emphasize the
size of particular groupings which have been seen in the top-down labeling
to be rather significant groupings. A feature such as ER-[+lateral] would
tend to get a significant weighting along an axis if it had a narrow
predictive value (there are only a small number of processes (deletion, total
assimilation)), then it would get a higher factor weighting with other
processes which match the same predictive value (occur with deletion, total
assimilations).

Figure 6.15: Regressive Feature Cluster Dendrogram.

 A second analysis considered only the binary features referring to P-
base’s feature analysis of the features changed in each described rule. The
intention of this analysis is to consider what clustering around particular
processes we should expect. This approach has the benefit of being more
interpretable as there are fewer features (46 features) in the analysis, so
more can be visualized in the dendrogram (Figure 6.16).
This second set of analyses are intended to determine features which
group together in terms of their changes. Here the dependence of features
on each other is being predicted. We have a similar grouping of features
which pattern together ([+round]/[+labial]) in terms of the changes, but
there seem to be significant groupings representing changes associated with
sonorant/ glide hardening, gliding, debuccalization, and a complex of
changes associated with vowels.
The interpretability of this dependence between feature changes is a
little loose, and somewhat dependent on the predictability of one feature
change given another, or the representation of changes in the database. We
see that a commonly changed feature [+voice] is not represented in this set
because it occurs in so many different contexts and doesn’t often pattern
with any one particular feature.
This bottom-up analysis has provided associations between features
which either describe a rule’s environment, assimilations, or changes.
Common environments and common change processes end up grouping
together under the cluster dendrograms.
Common processes such as place assimilation are unable to be captured
under this dendrogram-type approach because they are represented as the
change of several place-related features. Other processes which are
triggered in a diverse set of triggering contexts (intervocalic/pre-sonorant)
voicing don’t show up as having good predictive value and don’t end up
significantly populating a certain factor.
 Figure 6.16: Changes.

7 Discussion
Phonological alternations are clearly a diverse set of phenomena, often
reflecting idiosyncratic, arbitrary facts about the particular language in
which they occur. A major goal of phonological typology is to determine
what is frequent and why, and a major recurring theme in this exploration of
P-base has been that certain very specific types of phonological alternations
are extremely frequent, and their frequency is not predictable on the basis of
the frequency of their parts. It is useful to interpret the very frequent
patterns in terms of potential phonetic and structural sources, i.e., in terms
of phonetic factors that could drive sound change and independent
linguistic or cognitive factors that could drive learners to learn sound
patterns in a particular way.

7.1 Highly frequent patterns

A small number of sound patterns with quite specific descriptions are very
frequent. The most striking is preconsonantal regressive nasal place
assimilation, whose 207 cases account for 4.54% of the 4560 sound pattern
entries and 56% of all place assimilation. Most of the rest of place
assimilation is also in the preconsonantal context. Place changes are very
rare otherwise. Place assimilation has played a central role in the
development of distinctive feature systems, especially in feature geometry
(e.g., Clements 1985; Sagey 1986[1990]; Clements & Hume 1995; Halle et
al. 2000). It is clear that phonological theory needs a system for
representing place assimilation, but it is interesting to know that it is called
upon primarily for a narrow range of frequent patterns and only
sporadically otherwise.
The distribution of place assimilation patterns recalls the claims made
recently by Vaux and Blevins, which are superficially opposite but
compatible: Vaux (2002) argued that almost any consonant can be
epenthetic, and Blevins (2008a) argued that glottals and glides are the only
epenthetic consonants generated in great numbers by sound change. Here
we have seen a wide range of assimilation patterns but only regressive nasal
place assimilation is observed in huge numbers. This and some of the other
patterns we have reported in this chapter showcase what Moreton (2008)
terms “channel bias”: phonological typology reflects a bias that has an
explanation in recurrent sound change, and human language competence
can apparently handle a wide range of sound patterns that rarely occur, for
historical reasons. It is our hope that future research will examine P-base
more closely for patterns that reveal Moreton’s analytic bias, i.e., sound
patterns that have a clear phonetic basis like glide epenthesis and nasal
place assimilation, but nevertheless are rare in the record of synchronic
phenomena, due to bias in the system of phonological representations or the
learning process.

7.2 Nasal place assimilation

Preconsonantal regressive nasal place assimilation stands out as our
prototypical highly specific highly frequent sound pattern whose frequency
is not accounted for by the frequency of its parts. The imbalance between
nasals and oral stops in the frequency of place assimilation has been known
for a long time, and a lot of work has been devoted to accounting for why
nasals undergo place assimilation more frequently than oral stops (e.g., Cho
1990). While a perceptual explanation is intuitive (because nasalization
interferes with formant transitions that provide cues to place, exacerbating
perceptual weaknesses associated with being preconsonantal), Winters
(2003) showed that perceptibility does not account for the asymmetry. We
offer a potential articulatory basis for the asymmetry. Gick et al. (2013) and
Chiu & Gick (2013) have recently shown that nasals differ from oral stops
in the forcefulness of the closure, e.g., lip compression is weaker in [m]
than in [b]. This is thought to be because the lingual or labial closure for an
oral stop needs to withstand an increase in intraoral pressure, but the closure
in a nasal consonant does not, and the constrictions are executed differently.
We speculate that a consequence of this less forceful closure at the place of
articulation for nasal consonants makes them more susceptible to being
reduced, providing an articulatory basis for NC clusters being heard as
homorganic at the place of articulation of the second consonant, without
any appeal to misperception.

7.3 Conclusions
This study of frequent sound patterns in P-base has been deliberately
incomplete. We have avoided focusing a lot of attention on types of sound
patterns that have been the focus of a lot of crosslinguistic work already
(but could be revisited), and certainly we have overlooked interesting types
of sound patterns that have not been well studied before. We encourage
readers who have been intrigued by any of these possibilities, or by
questions raised in other chapters in this volume, to follow up with their
own P-base queries at http://pbase.phon.chass.ncsu.edu.

References
Bickel, Balthasar. 2008. A refined sampling procedure for genealogical control.
Sprachtypologie und Universalienforschung 61. 221–233.
Bickel, Balthasar & Johanna Nichols. 1996. The AUTOTYP database. http://www.s
pw.uzh.ch/autotyp/, electronic database.
Blevins, Juliette. 2004. Evolutionary phonology. Cambridge: Cambridge University
Press.
Blevins, Juliette. 2006. A theoretical synopsis of Evolutionary Phonology.
Theoretical Linguistics 32. 117–165.
Blevins, Juliette. 2008a. Consonant epenthesis: Natural and unnatural histories. In
Jeff Good (ed.), Linguistic universals and language change, 79–107. Oxford:
Oxford University Press.
Blevins, Juliette. 2008b. Natural and unnatural sound patterns: A pocket field
guide. In Klaas Willems & Ludovic de Cuypere (eds.), Naturalness and
 iconicity in language, 121–148. Amsterdam: Benjamins.
Catford, J. C. 1977. Fundamental problems in phonetics, volume 1. Edinburgh:
Edinburgh University Press.
Chiu, Chenhao & Bryan Gick. 2013. Producing whole speech events: Anticipatory
lip compression in bilabial stops. Proceedings of Meetings on Acoustics 19.
060252. http://asa.scitation.org/doi/abs/10.1121/1.4800579
Cho, Young-Mee. 1990. Parameters of consonantal assimilation. Ph.D. thesis,
Stanford University. (Published München: Lincom, 1999.)
Chomsky, Noam & Morris Halle. 1968. The sound pattern of English. New York:
Harper and Row.
Clements, G. N. 1985. The geometry of phonological features. Phonology
Yearbook 2. 225–252.
Clements, G. N. & Elizabeth V. Hume. 1995. The internal organization of speech
sounds. In John Goldsmith (ed.), The handbook of phonological theory, 245–
306. Oxford: Blackwell.
Côté, Marie-Hélène. 2000. Consonant cluster phonotactics: A perceptual
approach. Ph.D. thesis, Massachusetts Institute of Technology. https://rucore.li
braries.rutgers.edu/rutgers-lib/38478/
Gastner, Michael T. & M. E. J. Newman. 2004. Diffusion-based method for
producing density-equalizing maps. Proceedings of the National Academy of
Sciences of the United States of America 101. 7499–7504.
Gick, Bryan, I. Ian Stavness & C. Chenhao Chiu. 2013. Coarticulation in a whole
event model of speech production. Proceedings of Meetings on Acoustics 19,
060207. http://asa.scitation.org/doi/abs/10.1121/1.4799482.
Hall, Robert A. 1949. The linguistic position of Franco-Provençal. Language 25. 1–
14.
Halle, Morris & George N. Clements. 1983. Problem book in phonology.
Cambridge, MA: MIT Press.
Halle, Morris, Bert Vaux & Andrew Wolfe. 2000. On feature spreading and the
representation of place of articulation. Linguistic Inquiry 31. 387–444.
Hume, Elizabeth V. 2004. The indeterminacy/attestation model of metathesis.
Language 80. 203–237.
Hume, Elizabeth & Ilana Bromberg. 2005. Predicting epenthesis: An information-
theoretic account. Paper presented at 7èmes journées internationales du
réseau français de phonologie, Aix-en-Provence. https://www.researchgate.ne
t/publication/228973329_Predicting_epenthesis_An_Information-theoretic_acc
ount
Hyman, Larry M. 2008. Universals in phonology. The Linguistic Review 25. 83–
137.
Jakobson, Roman, C. Gunnar M. Fant & Morris Halle. 1952. Preliminaries to
speech analysis: The distinctive features and their correlates. Massachusetts
Institute of Technology, Acoustics Laboratory, Technical Report No. 13; 2nd
printing with additions and corrections.
Kirchner, Robert M. 2001. An effort based approach to consonant lenition. New
York: Routledge.
Lavoie, Lisa M. 2001. Consonant strength: Phonological patterns and phonetic
manifestations. New York: Garland.
Lindblom, Björn & Ian Maddieson. 1988. Phonetic universals in consonant
systems. In Larry M. Hyman & Charles N. Li (eds.), Language, speech, and
mind: Studies in honour of Victoria Fromkin, 62–78. London: Routledge.
Maddieson, Ian. 1984. Patterns of sounds. Cambridge: Cambridge University
Press.
Maddieson, Ian. 1997. Phonetic universals. In William J. Hardcastle & John Laver
(eds.), The handbook of phonetic sciences, 619–639. Oxford: Blackwell.
Maddieson, Ian. 2014. LAPSyD: Lyon-Albuquerque Phonological Systems
Database, CNRS, Lyon, France. http://www.lapsyd.ddl.ish-lyon.cnrs.fr/lapsyd/
Maddieson, Ian & Kristin Precoda. 1990. Updating UPSID. UCLA Working Papers
in Phonetics 74. 104–111.
Mielke, Jeff. 2003. The interplay of speech perception and phonology:
Experimental evidence from Turkish. Phonetica 60. 208–229.
Mielke, Jeff. 2005. Ambivalence and ambiguity in laterals and nasals. Phonology
22. 169–203.
Mielke, Jeff. 2008. The emergence of distinctive features. Oxford: Oxford
University Press.
Mielke, Jeff. 2017. Visualizing phonetic segment frequencies with density-
equalizing maps. Journal of the International Phonetic Association, https://doi.
org/10.1017/S0025100317000123.
Mielke, Jeff, Lyra Magloughlin & Elizabeth Hume. 2011. Evaluating the
effectiveness of Unified Feature Theory and three other feature systems. In
John Goldsmith, Elizabeth Hume & Leo Wetzels (eds.), Tones and features: In
honor of G. Nick Clements, 223–263. Berlin: Mouton de Gruyter.
Moreton, Elliot. 2008. Analytic bias and phonological typology. Phonology 25. 83–
127.
Myers, Scott. 2010. Regressive voicing assimilation: Production and perception
studies. Journal of the International Phonetic Association 40. 163–179.
Myers, Scott. 2012. Final devoicing: Production and perception studies. In Toni
Borowsky, Shigeto Kawahara, Mariko Sugahara & Takahito Shinya (eds.),
Prosody matters: Essays in honor of Elisabeth Selkirk, 148–180. London:
Equinox Press.
Pierrehumbert, Janet & David Talkin. 1992. Lenition of /h/ and glottal stop. In G. J.
Doherty & D. R. Ladd (eds.), Papers in laboratory phonology, vol. 2: Gesture,
segment, prosody, 90–117. Cambridge: Cambridge University Press.
Sagey, Elizabeth C. 1986. The representation of features and relations in non-
linear phonology. Ph.D. dissertation, Massachusetts Institute of Technology.
(Published New York: Garland, 1990.)
Simpson, Adrian P. 1999. Fundamental problems in comparative phonetics and
phonology: Does UPSID help to solve them? In Proceedings of ICPhS XIV,
Berkeley, CA, 349–352. http://www.personal.uni-jena.de/~x1siad/papers/icphs
99_fund.pdf
Steriade, Donca. 2001. Directional asymmetries in place assimilation: A perceptual
account. In Elizabeth Hume & Keith Johnson (eds.), Perception in phonology,
219–250. New York: Academic Press.
Vance, Timothy J. 1987. An introduction to Japanese phonology. Albany, NY: State
University of New York Press.
Vaux, Bert. 2002. Consonant epenthesis and the problem of unnatural phonology,
Paper presented at Yale University Linguistics Colloquium.
Vennemann, Theo. 1974. Words and syllables in natural generative phonology. In
A. Bruck, R. Fox & M. L. Galy (eds.), CLS 10: Parasession on natural
phonology, 364–374. Chicago: Chicago Linguistic Society.
Winters, Stephen J. 2003. Empirical investigations into the perceptual and
articulatory origins of crosslinguistic asymmetries in place assimilation. Ph.D.
dissertation, Ohio State University, Columbus, OH. https://etd.ohiolink.edu/pg_
10?0::NO:10:P10_ACCESSION_NUM:osu1054756426
Aditi Lahiri
Predicting universal phonological
contrasts

Abstract: FUL (Featurally Underspecified Lexicon) assumes that a

handful of features will account for the phonological systems in the world’s
languages. Such an assumption would not be unusual. However, FUL
makes several other assumptions including the following: (i) consonants
and vowels share place features which are not represented on separate tiers;
(ii) features are monovalent; (iii) there are no feature dependencies; (iv)
CORONAL and PLOSIVE are always underspecified in representation but
present on the surface, which in turn presupposes that both these features
must occur in all languages; (v) phonological activity is not the only way to
determine feature contrast. These assumptions are based on synchronic,
diachronic, and experimental evidence. Detailed case studies examine
whether these hypotheses hold in instances where the opposite claims have
been made.

1 Introducing FUL
In a landmark work, Jakobson, Fant, & Halle (1952, henceforth JFH)
proposed a set of 21 distinctive features for describing phonological
systems. Well defined acoustic and articulatory correlates were identified
for their features, and the same features were employed to classify place of
articulation for vowels and consonants. An example would be the feature
acute: front vowels (such as [i y e ø æ]) and fronted consonants (e.g.,
alveolars and palatals) were classified as acute and characterised as having
high frequency energy. First proposed in 1999, the FUL system (short for
Featurally Underspecified Lexicon) endorsed these two fundamental
assumptions of JFH’s. The following considerations, some differing from
JHF, are especially highlighted in FUL: (i) phonological features form a
hierarchical system; (ii) all features are monovalent; (iii) the contrasts
established by this set of features should account for phonological
alternations across the languages of the world; (iv) a small set of features
are universally underspecified, and these features should therefore always
be part of the inventory; (v) there are no feature dependencies; (vi)
underlying phonological representations, as part of the mental lexicon,
govern production and comprehension, with underspecification, thus,
implying asymmetries in processing; (vii) feature specification and building
the feature tree during acquisition initially follow a universal pattern; (viii)
feature specification and underspecification should also play a part in
language change.
Notions from the JHF tradition such as “markedness”, “specificity”,
“redundancy”, and “activity” have in one way or another been widely used
by phonologists. No one has ever assumed that all features have the same
“weight”, and most phonologists do not specify non-contrastive features.
Chomsky & Halle (1968) engaged in detailed discussions about markedness
combined with redundancy to obtain the right phonological alternations. In
the early eighties, underspecification was hotly debated (cf. Archangeli
1988), particularly with respect to coronality (Paradis & Prunet 1991), and
the concept was indeed frowned upon (McCarthy 1988). Halle et al. (2000)
emphasise that full specification for contrastive features should be the
norm. Despite the unease, there is no doubt that asymmetries and
markedness differences exist across feature distributions and directions of
the output of phonological rules, and various methods have been employed
to handle them. Calabrese (1995) distinguished different types of feature
representations such as contrastive, marked, and full, which in turn were
interspersed in the ordering of rules. Mohanan (1993) favoured what he
called “fields of attraction” and “dominance”, which allowed him to express
degrees of markedness. Clements (2001) proposed a complex model
combining both specification and underspecification, which allowed non-
contrastive features to be specified if they were “active” in phonology. He
distinguished between “active” features (which may form natural classes)
and “prominent” features (which, for instance, play a role in spreading).
Against this historical backdrop, this chapter sets out the FUL view of
underspecification and asymmetry, and specifically addresses two
questions:
(i) How do FUL’s features and their hierarchical organisation account for
the phonological contrasts of the languages of the world?
(ii) To what extent are (UNDER)SPECIFICATION and (IN)ACTIVITY correlated?

Through examining several test cases bearing on these issue in especially

challenging ways, we seek to further strengthen the case for the FUL
approach. Particular emphasis will be on the feature CORONAL, the focus of a
lot of attention in past decades. In the course of a brief historical overview,
Section 2 compares several approaches to coronals. Section 3 highlights the
specifics of FUL in relation to other models, in particular the Contrastive
Hierarchy proposed by Dresher and colleagues and Clements’ system of
underspecification. Finally, Section 4 returns to monovalent features, in
particular to account for complex vowel alternations like those of Kàlɔ̀ŋ
analysed in Hyman (2003). The typological moral is that coronal contrasts
and alternations involving coronal triggering on the face of it show a great
deal of variation, but analysis – along lines dictated by a particular
theoretical model, FUL – reveals fundamental unity of phonological
grammar behind crosslinguistic diversity.

2 Towards FUL: [coronal] vs. [−back]

A decade after JFH, articulatorily oriented features were in the ascendancy
in The sound pattern of English (Chomsky & Halle 1968, SPE), being
supposedly better suited to describe phonological patterns of the world’s
languages. A major change was the establishment of separate place features
for vowels and consonants. Vowels, for example, were all [−anterior], while
consonants could be both [+anterior] and [−anterior]; vowels were
characterised by [±back] and were always [−coronal]. Thus, there was no
way to pair [coronal] consonants like dentals and palatoalveolars with
[−back] vowels. A subset of features considered by SPE (1968: 407,
adapted from Table 3) for various main places of articulation (and not
including secondary articulations such as palatalised labials), is given in
(1):100
The eighties led the way to grouping features into natural classes rather than
listing them arbitrarily (e.g., Clements 1985, 1989; Sagey 1986; Clements
& Hume 1995; McCarthy 1988). Although controversies raged over the
precise grouping, one assumption remained unchanged: vowels and
consonants did not share all place features. The dorsal node dominated
vowels which were largely distinguished by [±back]. This had the unwanted
consequence of segregating front vowels ([−back], [−coronal] [−anterior])
from dentals, alveolars, and palatoalveolars, which were grouped under
[coronal]. Additionally, the height features [±high], [±low] were dominated
by [dorsal]. The feature tree in (2) gives the general idea (see further Lahiri
& Reetz 2010):101
In a novel proposal, Clements (1989) argued that, similar to JFH, vowels
and consonants ought to be brought together if the notion of constriction of
the vocal tract with the parameters of degree and location was to be taken
seriously. However, although the place features were accordingly the same,
the place nodes for vowels and consonants would be on separate tiers.
A fundamental difference from earlier models is that [coronal] here entirely
replaced [±back]. In response to Clements’ unified theory, Halle et al.
(2000) proposed to dispense with dependencies, such that [back] [high]
[low] were no longer dependents of DORSAL.102 Thus, any fronting that
would spread [dorsal] would not necessarily spread [−back]. Nevertheless,
vowels and consonants remained distinct in terms of place features. The
PLACE node proposed by Halle et al. (2000) is as in (4):
Taking JFH’s view of combining all consonantal and vocalic features
seriously, Lahiri & Evers (1991) and Lahiri & Reetz (2002) (cf. also Lahiri
2000; Ghini 2001a) pointed out that there was no necessity to duplicate the
V-place node for vowels and secondary articulations, and that the aperture
node was not only relevant for vowels but also for consonants. Thus, the
constriction relevant on the horizontal dimension along the vocal tract was
determined by the ARTICULATORS, and on the vertical dimension was
characterised by the height of the tongue. Consequently, as seen in (5), the
PLACE node dominated separate nodes ARTICULATOR and TONGUE
HEIGHT, as well as the TONGUE ROOT features, with the PLACE
features thus identical for vowels and consonants. Although, to honour
tradition, we have kept the basic articulatory names, each feature has well
defined acoustic cues as well. The features and feature organisation we will
defend are based on universal principles of phonological alternations as
well as production and perceptual mechanisms. The features are the same
for production and perception (cf. Lahiri & Reetz 2010; Lahiri 2010; Plank
& Lahiri 2015). Furthermore, the FUL processing model has clear-cut
processes of matching from the signal to the representation and the other
way around.
A subset of the phonemes grouped under each feature (vowels and
consonants) are given in (6).

(6) Features and segments

[LABIAL] labial consonants, rounded vowels
[CORONAL] front vowels, dental, palatal, palatoalveolar, retroflex
consonants
[DORSAL] back vowels, velar, uvular consonants
[RADICAL] pharyngealised vowels, glottal, pharyngeal consonants
[HIGH] high vowels, palatalised consonants, retroflex, velar,
 palatal, pharyngeal consonants
[LOW] low vowels, dental, uvular consonants
[ATR] palatoalveolar consonants, tense vowels
[RTR] retroflex consonants, lax vowels

Two pairs of opposing features – CONSONANTAL/VOCALIC and

SONORANT/OBSTRUENT – are the major class features available in all
languages. All phonemes must be either CONSONANTAL or VOCALIC and
SONORANT or OBSTRUENT. The members of each pair are conflicting – i.e.,
consonantal implies not vocalic and vice versa. The nodes LARYNGEAL,
CONSTRICTION, PLACE, TONGUE HEIGHT, TONGUE ROOT,
RADICAL are always present, although they may be empty if there are no
contrastive phonemes in the language concerned, as discussed below. All
features are monovalent; therefore, they are either present or absent, and
unlike in Halle et al. (2000) and Clements (2001), there is no mixture of
some binary (e.g., [±back]) and some monovalent features (e.g., [dorsal]).
Further, in contrast to earlier approaches – for instance, in Clements (2001:
114, 47) [±posterior] is dominated by [coronal] – no features dominate
other features in FUL.
A number of assumptions fall out from the feature tree and we take
them in turn.
First, the features under each node are mutually exclusive. An exception
is the LARYNGEAL node, where [VOICE] and [SPREAD GLOTTIS] can co-
occur, as attested in only a few languages (among them Bengali and
Hindi).103 Thus, [NASAL/ LATERAL/RHOTIC/STRIDENT], [CONTINUANT/PLOSIVE],
[LABIAL/CORONAL/DORSAL/ GLOTTAL], [HIGH/LOW], [ATR/RTR] are mutually
exclusive for consonants. For vowels, [LABIAL] may combine with
[CORONAL] and [DORSAL].
Second, the only dependencies we assume are universal implications
such as [NASAL] ⇒ [SONORANT] or [STRIDENT] ⇒ [OBSTRUENT].
Third, underspecification is fundamental to FUL. However, there are
only two features which are universally underspecified: [CORONAL] and
[PLOSIVE]. The reasons are based on contrast sensitivity and typological
prominence.
Fourth, only those features that are necessary to maintain contrasts
between the phonemes of a language are used. This is similar to Dresher’s
(2009) contrastivity, with the exception that there is no “activity”
requirement in FUL. Feature specifications are independent of whether or
not the features play an “active” role in phonological processes – as to be
discussed in the context of a case study in the next section.
Fifth, feature trees are built in language acquisition based on the
universal principle of PLACE-first (Ghini 2001b), where ARTICULATOR
contrasts precede TONGUE HEIGHT contrasts. For the ARTICULATOR
contrasts, [CORONAL] is the universal default: all languages must have this
feature. Since [CORONAL] is underspecified, the assumption is that when
during acquisition, a non-coronal phoneme is enountered, it becomes
specified. For instance, Levelt (1995) and Fikkert & Levelt (2008) observed
that a contrastive feature like [LABIAL] becomes specified first. For
TONGUE HEIGHT, we believe that the feature [low] will be assigned first.
These assumptions are fairly restrictive and we are aware that they are
in conflict with many assumptions made in the literature. Three issues are
especially critical and will be addressed presently:
(i) How can languages be accounted for where [coronal] is supposed to be
active?
(ii) If no dependent features such as [±anterior] and [±distributed] are
permitted, how can various sets of sounds be accounted for which
were classified by these features?
(iii) Does [coronal] always exist?

3 Underspecification of [coronal] and “activity”

3.1 Coronal activities

It has variously been proposed that it is essential that [CORONAL] is “active”
and therefore needs to be specified. We will discuss two relevant case
studies in some depth: palatalisation in Inuit dialects as analysed in
Compton & Dresher (2001, henceforth C&D), and Tahltan vowel harmony
as analysed in Clements (2001). In both instances, the presence of
[CORONAL] is indispensible.
3.2 Development of Proto-Eskimo vowels and
palatalisation
Proto-Eskimo had four vowels */i u a ə/ and in most Inuit dialects /ə/
merged with reflexes of */i/. Traditional descriptions distinguish between
“strong i” from original */i/, which triggers palatalisation, and “weak i”
from original */ə/, which fails to trigger palatalisation. In C&D’s story,
features are ordered according to a Contrastive Hierarchy as established by
the Sucessive Division Algorithm. Only features which are “active” are on
the top of the hierarchy. Based on various processes, it can be shown that
[low], [labial], and [coronal] are active in Proto-Eskimo and are treated as
the marked values and the opposites are unmarked. The hierarchy for Proto-
Eskimo is given in (7).

The hierarchy begins with [low], this being the first division by Jakobson &
Halle (1956) on the grounds of highest sonority. The next cut follows the
common pattern of place-next. The most important aspect is that /i/ is
[coronal] and this is the feature that is required to trigger palatalisation. All
four-vowel systems of this family have the strong i as coronal. The three-
vowel systems / i u a/, however, do not have /ə/, neither is palatalisation
being triggered. Thus, these vowel systems (again beginning with [low]) are
organised as follows:

This is an elegant analysis which produces the pattern of alternations set out
in (9). Weak i alternations show no effect of the feature CORONAL, while
strong i alternations do, since /i/ is specified for CORONAL. Strong i leads to
palatalisation with surface [tʃ] and [ʎ].
Although FUL agrees that the vowel /i/ is [CORONAL], it faces several
problems with the assumption that CORONAL alone triggers palatalisation.
The assumption in FUL is that palatalisation would normally occur with the
additional feature [HIGH]; certainly the vowel /i/ is involved, but not the
main place feature. Second, [CORONAL] would be underspecified, and thus
will not play an active role. How could it work under these assumptions and
would such an analysis be in any way preferable? We provide an alternative
below.
First, the Inuit palatalisations affect all places of articulation; a summary
from C&D’s data is in (10).

(10) Surface outputs in Inuit due to palatalisation

n→ɲ
l→ʎ
k→ʧ
t→s

Note that the obstruents become strident, which would be the phonetic
enhancement of the palatalisation process. It is actually not evident from
C&D’s analysis why [CORONAL] is the active feature relevant for
palatalisation, since the inputs / t l s/ are all [CORONAL] to begin with. The
only change in place of articulation is /k/ → /s/. All relevant features in
FUL are tabulated in (11); the consonants [λ ɲ s tʃ] are listed for
convenience, but they are in parentheses since they are derivatives of /l n t
k/ in the context of /i/.

(11) FUL features for vowels (4-vowel system)

(12) Features for relevant coronal consonants

Although the tables show the distribution of features, a proper tree diagram
is necessary to show the precise nature of the underlying representations
and how the palatalisation process should work. On our analysis, the
underlying representations of four contrasting vowels and consonants, / i u
k l/, are as follows:

We distinguish between UNDERSPECIFIED features, which are marked as [—]

with the relevant node (e.g., ART [—] for /i/), and NOT SPECIFIED such as the
empty cell for ARTICULATOR for /ə/. We will see presently that the
underspecified features will eventually have specific features filled in by
production rules, while those that are not specified or redundant will be
more variable. Thus, /ə/ will have a PLACE node but no ARTICULATOR
node properly assigned; it may, therefore, obtain different features in
production.
Palatalisation is triggered by /i/, which is ARTICULATOR-free but
specified for TONGUE HEIGHT [HIGH]. Two processes are involved in the
way /i/ “fronts” / k n l/, a not uncommon way to describe palatalisation. The
first entails ensuring that the sequence of ARTICULATOR features do not
mismatch, leading to [DORSAL] being deleted in the context of
ARTICULATOR-free /i/. The second, where the ARTICULATOR features
of the consonants /n l/ are unspecified, involves the spreading of [high] to
non-[HIGH] coronal consonants making them palatals and thereby [HIGH].
Both are illustrated in (14).
There is a third palatalisation which is an assibilation process, where /t/
does not change the place of articulation, but becomes a strident [s]. To
account for this, [STRIDENT] is incorporated as a fill-in surface rule as an
effect of the spreading of [high], as in (15). Languages differ in the way /t/
becomes a sibilant; in English, for instance, [high] leads to /t/ becoming a
/tʃ/, as in don’t you → don[tʃ]you.
To obtain the correct surface forms we need further steps, viz., rules which
fill in features for production (cf. Lahiri 2010; Plank & Lahiri 2015). An
ARTICULATOR-free consonant will be provided with [CORONAL] as in
(16). This will also include the addition of [STRIDENT] for /t/ in the context
of /i/.

(16) Surface production rules for vowels

ART [—] → [CORONAL]
ART [DORSAL] & TH [HIGH] → ADD [LABIAL]

There are a few additional points to be made. First, we are able to account
for palatalisation even if [CORONAL] is unspecified. However, why would
this analysis be preferred over that of C&D, who assume that the [coronal]
specification of /i/ can account for all the palatalisation processes? They
elegantly connect the presence and absence of palatalisation and the
specification of [CORONAL] for /i/. We do not deny that /i/ is [coronal] nor
that it plays a significant role. However, C&D do not discuss the various
ways in which [CORONAL] should affect the other consonants and in fact
they do not show how palatalisation is actually realised. For instance, why
is it that /l/ becomes /λ/ when [CORONAL] from /i/ spreads? Is /l/ not
[CORONAL]? What about other coronal consonants such as /n/? Why does the
addition of [coronal] from /i/ alone lead to palatalisation? Is it the vocalic
element that is crucial and [CORONAL] from consonants has no effect?104 For
/k/-palatalisation, it is obvious that the place feature of the consonant
changes. In our analysis, this is treated as an assimilation process whereby
the ARTICULATOR features merge; this can be achieved by spreading or
deletion. However, in our view palatalisation of the other consonants which
are inherently all [coronal] is different. Thus, we crucially distinguish
between palatalisations which affects back consonants and those which
share the place feature with /i/. It is not clear how this is accounted for in
C&D.
Second, the main aim of C&D’s analysis is to confirm that the four-
vowel and three-vowel systems have different feature distributions.
Accordingly, for them the difference lies in the four-vowel systems
requiring [CORONAL] to be specified for /i/, which triggers palatalisation,
while it is unspecified in the three-vowel system (see above (7), (8)). Can
our analysis account for this contrast, given that [CORONAL] is always
underspecified and will always be filled in in the surface representation
because it has an empty ARTICULATOR node? The answer is yes. Recall
that in FUL it is not coronality per se which triggers palatalisation: it is
[HIGH] that plays a crucial role. We compare the four- and three-vowel
systems in FUL:

(17) FUL features for vowels in Inuit dialects (3- and 4-vowel systems)
(a) 3-vowel system

(b) 4-vowel system

Unlike C&D, it is not the presence or absence of coronality which is of

central concern, but the TONGUE HEIGHT contrasts. In the three-vowel
system, [HIGH] is not required, since the ARTICULATOR features are
enough to distinguish /i/ and /u/. The first height feature is always [LOW]. If
that is sufficient, there is no further need for [high]. In the four-vowel
system, the presence of a fourth vowel /ə/ requires [HIGH] to be specified.
Since [high] is essential for /i/ to trigger either fronting or assibilation, the
three-vowel systems do not cater to palatalisation. Consequently, the lack of
specification of [CORONAL] universally does not prevent us from accounting
for the presence and absence of palatalisations.

3.3 Tahltan coronal harmony

Palatalisation in Inuit dialects directly leads us into a discussion of vowel
harmony where coronal consonants differ in terms of their transparency,
thereby either blocking or permitting harmony. Tahltan coronal harmony
was analysed in Shaw (1991: 144–152), reconsidered in Clements (2001),
and further discussed in Lahiri & Reetz (2010, henceforth L&R). In Shaw’s
account, Tahltan, an Athapaskan language, has a series of five coronal
obstruents, and coronal harmony is only applicable across three sets —
apical, laminal, and palatoalveolar consonants. The process involves
fricatives of these places of articulation, which assimilate to all coronal
place features, and stridency of any following coronal obstruent of one of
these three sets. The simple and lateral series are transparent to this
harmony process. Tahltan has four series of affricates and fricatives. The
only true stops belong to the simple series. Even in the lateral series, the
consonants are not sonorants, but obstruent affricates and fricatives.
We compare Clements’ analysis to L&R. The coronal obstruents and the
relevant features proposed by Clements are given in (18) along with the
corresponding features from FUL. Only the crucial features are listed.
In Clements (2001), the features [strident], [apical], and [posterior] are
dominated by the [coronal] node, and only the marked feature values are
specified, namely [+strident], [−apical], and [+posterior]. Thus, of the five
coronal sets of a consonants, two are not specified for [coronal], but the
others are. Stridency is a property only of coronal consonants; i.e., strident
is dependent on coronal. The simple and the lateral series are unmarked for
coronal, which is neither lexically specified nor active, and hence absent for
these consonants. Thus, only the specified coronal consonants are engaged
in harmony, but the others are transparent for harmony. In FUL, however,
all coronal consonants, including the simple and the lateral consonants, are
unspecified for the ARTICULATOR node and [STRIDENT] is independent of
the ARTICULATOR node. An important assumption in FUL is that all
fricatives and obstruent affricates in Tahltan are [STRIDENT]. Before we
discuss the actual harmony process, we will look at some examples.
The differences between the first set, with /s/ changing to /θ/ /ʃ/, and the
second set, with /θ/ changing to /s/ and /ʃ/, depend on the target: an apical
(or perhaps dental) strident /s/ changes to the palatoalveolar or interdental in
the respective contexts, while interdental /θ/ changes to palatoalveolar /ʃ/ or
dental /s/.
In Shaw’s rule for harmony, the rightmost specified coronal node
spreads leftwards with the concomitant delinking of the previous coronal
specification of the target (cf. Lahiri & Reetz 2010 for a detailed analysis).
The target is an immediately adjacent specified coronal node. Since both
trigger and target need to be specified coronal nodes, the lateral and simple
series are unaffected by the spreading and cannot block harmony. Clements
accomplishes consonantal harmony with a single AGREE constraint which
ensures that all coronal nodes containing the marked feature values must be
identical within the word. This means that the coronal laterals and plosives
remain untouched while the others are involved in harmony. Both Shaw and
Clements need to separate the lateral and plosives from the other
consonants in terms of coronal specification. They achieve the harmony
process by ensuring that the simple [d] and the lateral [dl] series are free of
the coronal node, while the other series require features which are
dominated by the coronal node.
How does it work in FUL? Under our analysis, harmony is restricted to
obstruents specified for STRIDENT; as mentioned before, [CORONAL] will
surface, but is unspecified in the underlying representation. Moreover, the
plosives and the laterals do not have the feature [STRIDENT].

(20) Tahltan harmony in FUL: for a sequence of [STRIDENT] obstruents,

TONGUE HEIGHT features spread regressively within a word when
no other [STRIDENT] obstruent intervenes
In (i) and (ii), [HIGH] dominates, turning both /θ and /s/ to /ʃ/, while in (iii)
[LOW] spreads and /θ/ becomes /s/. For (iv), since assimilation is
progressive, the TONGUE HEIGHT feature [low] of /ts/ is deleted to agree
with the unspecified TONGUE HEIGHT feature of /θ/. The TONGUE
HEIGHT features spread across the laterals (e.g., /tɬ’/) and plosives (e.g.,
/d/) which are both independent of the TONGUE HEIGHT node and does
not block any feature spreading. No other feature is altered.
Thus, for Tahltan, in both Shaw and Clements’ analyses, some of the
coronal consonants could not be specified for coronality to obtain correct
results for harmony. For both analyses, stridency was an additional
complication since the distinguishing features for the various places of
articulation with coronal consonants as well as the feature [STRIDENT] were
dependent on [CORONAL]. Consequently, to prevent some of the coronal
consonants from undergoing harmony or to be transparent to harmony, they
were prevented from having the coronal feature or in Clements’ terms,
“coronal was not active”. FUL does not have that choice. The feature
coronal is always there and is always underspecified. Furthermore, since
there are no features dependent on CORONAL, to obtain the other contrasts,
different features are employed.

4 Lack of dependent features and coronal

contrasts

4.1 Ways of accounting for coronal contrasts

As mentioned above, FUL disallows dependent features. Although this has
positive aspects, we still need to be able to distinguish many places of
articulation which are all [CORONAL]. Since SPE, the features [±anterior]
and [±distributed] have been used to distinguish between the various
coronal affricates and fricatives. The traditional SPE feature combinations
(cf. (1)) for coronal consonants made by the tip and blade of the tongue are
as follows:
In FUL, both palatal and retroflex consonants are [CORONAL] and [high],
and therefore are not distinguishable by these two features. These types of
consonantal contrasts could be potential problems: palatal versus retroflex
stops /c ʈ / and nasals /ɲ ɳ /, and palatoalveolar versus retroflex sibilants /ʃʂ/.
Lahiri & Reetz (2010) maintain that it is extremely rare (if attested at
all) that phonemic contrast occurs between dentals and alveolars on the one
hand and retroflex and palatals on the other. For example, Malayalam has
been claimed to have both alveolar and dental stops. However, the alveolar
stop is always derived from a rhotic and the minimal pairs that can be
obtained are via gemination. The rhotic, when geminated, becomes an
alveolar (Lahiri et al. 1984). Palatals are definitely coronal (Lahiri &
Blumstein 1984; Keating & Lahiri 1993), but we propose that a palatal
versus retroflex underlying contrast in stops is only possible if the palatal
stop is affricated or is an “alveopalatal” consonant, both of which would be
[STRIDENT] (cf. Hall 1997), or if one is derived from the other. In
Malayalam, which has both retroflex and palatal stops, only retroflex stops
occur in the underlying inventory. The palatal stops are derived in specific
morphological environments from intervocalic velars when preceded by
front vowels. Mohanan & Mohanan (1984: 589) also suggest that, given the
complex conditioning of the palatalisation rule, “[p]erhaps the right solution
is to say that Palatalization is blocked when the segment has some ad hoc
diacritical feature [−P]”. Mohanan & Mohanan also make a distinction
between underlying and lexical alphabet, the latter being derived by rules in
the lexicon. Their claim is that the lexical alphabet “has significant
consequences for human perception of speech sounds” (1984: 598).
Possibly Mohanan & Mohanan’s lexical contrasts and our underlying
contrasts are the same. A further possibility is that, like the dental/ alveolar
contrast, the feature [HIGH] is specified in one case and not in the other.
If a language has both palatal and retroflex nasals, FUL predicts that
they are not truly contrastive. Either the palatal nasal /ɲ/ is an assimilated
variant of an alveolar or dental /n/ in the context of a palatal or
palatoalveolar stop, or the retroflex nasal is derived, or it consists of a nasal-
plus-glide sequence. Again, Malayalam is a good example since it has
seven phonetic nasals derived from three underlying ones which are labial,
dental, and velar /m n ŋ/ (Mohanan & Mohanan 1984: 583–586, 596–598).
Mohanan & Mohanan show that palatal and retroflex nasals are derived by
a homorganic nasal assimilation rule in the context of following palatal and
retroflex stops, and the palatal stops are in turn derived from velars. Thus,
the feature contrast for a language like Pitta-Pitta, which has been claimed
to have a series including dentals, alveolar, palatal, and retroflex stops and
nasals (from Hall 1997) would be as follows.105

So far we have not considered secondary articulations, although we have

discussed palatalisation and the spreading of the feature [HIGH] of coronal
vowels. In the next section we compare the fronting of velars with the
palatalisation of other coronal consonants.

4.2 Palatalised consonants and palatalisation

In Bhat’s seminal work (1978: 60–61), three types of palatalisation have
been shown to recur across languages: (i) fronting of velars (24a); (ii)
change of place within coronal consonants, with alveolar/ dental becoming
palatoalveolar/ palatal (24b); and (iii) addition of secondary articulation to
any place of articulation (24c); the context is invariably high front vowel /i/
and glide /j/ and sometimes the front vowel /e/.

(24) Results of palatalisation

 (a) k, x → tʃ, ç
(b) t, s → tʃ, ʃ
(c) p, t → pj, kj

Examples of (23a) include Slavic languages where [k g x] became [tʃ, dʒ, ʃ]

(SPE, 421–422). Polish is known to have palatalisations as in (23b) where
coronal consonants [t d s z r n l] become pre-palatal consonants before front
vowels and glides (Rubach 1984: 60). English alveolars such as [t d]
become [tʃ dʒ] in the context of [j]. Finally “secondary palatal” articulation
occurs involving “raising the central part of the tongue while keeping the
main articulator intact” (Bhat 1978: 67). Secondary palatalisation of this
sort occurs in Dutch diminutive formation, to be discussed in more detail
below. Two issues need to be addressed: palatalised consonants, and
particularly palatalised velars such as [kj] vs. [c] or [ç], and palatalisation as
a process.
There have been many discussions of palatalised consonants, succinctly
summarised and discussed in Hall (1997). In our view, as argued above (cf.
also Lahiri & Evers 1991; Lahiri & Reetz 2010), palatalisation is triggered
by [CORONAL] and [HIGH]; to set the scene for its defense, let us look at a few
notable alternatives.
To repeat, the crucial features traditionally implicated were [±anterior]
and [±back]. The pertinent consonants had the following features:

Thus, a change from [k] to [tʃ] in the context of [i] or [j] would involve a
change in the primary change of articulation of dorsal to coronal in the
context of [−back], which was dominated by dorsal. This problem was
addressed in detail by Clements and taken up by Hume, leading to the
feature set we discussed above.
In the analysis of Clements (1989), the structure of palatalisation would
involve the following features:

The first step involves a palatalised [kj], which has a C-place dorsal as well
as a V-place coronal. This in turn undergoes tier promotion, complex
segment formation, and concomitant affrication to become [tʃ]. This was a
remarkable proposal, suggesting for the first time that [j] led to
palatalisation because of its coronal status. Our view is similar, except that
we do not have the independent tiers. However, before we delve into FUL’s
proposal, we briefly discuss Hall’s take on this.
According to Hall (1997), palatals (“alveopalatals” in his terminology)
differ from “true” palatals such as German [ç]. His features for these
consonants would be as follows.

This differs from the Clements & Hume’s feature set:

Hall argues that, in his terminology, the features deliberately do not contrast
[ç] and [xj], because these two sounds never co-occur. The crucial point
here is that under Hall’s analysis, unlike Clements & Hume’s, palatalisation
is governed by a feature [+P], which is essentially [−back].
However, as seen above, [−back] would normally be dominated by
[dorsal]. Under Hall’s analysis, the palatalisation feature [+P] must,
therefore, come under both [coronal] and [dorsal]. This is because under his
analysis palatal [ç] is [dorsal] but unlike Sagey-Halle’s analysis, front
vowels and glides are coronal and not dorsal (1997: 79–83). In (29) we
reproduce Hall’s analysis of a velar and alveolar palatalisation, with velar
[x] to [ç] being a regular phenomenon in German.
As Hall states, [+P] as [−back] presents an apparent formal problem,
because it requires this feature to be located under both [+coronal] and
[+dorsal] (1997: 83–84).

(30) Definition of the palatalisation feature [+P] (Hall 1997: 83)

“the segments that are marked [+P] include (a) front vowels like [i e
æ], (b) palatoalveolars [ʃ ʒ], (c) alveolopalatals [ɕ ʑ], (d) palatals [ç
ʝ], and (e) palatalized segments (e.g. pj bj tj dj kj gj). The property
shared by all of these segment types in (a)-(e) is a fronted tongue
body (see Sagey 1986: 278)”

Consequently, since front vowels are coronal (like in FUL), Hall’s analysis
dispenses with the awkwardness of having a dorsal [k] becoming a coronal
palatoalveolar in the context of dorsal [i] or [j] via [−back], which too is
dominated by dorsal. However, since the palatals are still dorsal (unlike
FUL), the [+P] feature has to be dominated by both coronal and dorsal.
Instead of this rather complex analysis, we follow Clements’
assumptions that all palatals and palatoalveolars and front vowels and
glides are coronal, and thus palatalisation which causes a fronting of velar
consonants is an assimilation to a coronal place of articulation and by a
coronal. However, as noted above, palatalisation involves also the
“backing” of dentals/alveolars [t d] to [tʃ dʃ] or [ʃ ʒ] as well as adding
secondary articulations. We turn to this below.
We have argued elsewhere (i) that neither fronted velars and palatals,
nor palatalised velar and regular velar in the context of [i] may contrast in
any single language (Keating & Lahiri 1993), and (ii) that alveopalatal and
palatal stops do not co-occur in the same language (Lahiri & Blumstein
1984). Thus, features for these various coronal consonants as compared to a
velar would be as in (31).

Note that FUL’s features for alveopalatal and palatal sounds are the same: if
there is a contrast it has to be via [STRIDENT]. In FUL, the palatalisation of
velars, as for example [k] → [ç] in German or [k] → [tʃ] in Slavic, would
always have to be as follows:
Earlier, in Lahiri & Evers (1991, henceforth L&E), where we permitted
dependent features, palatal and palatoalveolar consonants were
distinguished by [−anterior]. Thus, the various coronal consonants were
distinguished as follows:

Despite allowing both binary features and dependency, in L&E

palatalisation of velars was a change of dorsal to coronal and crucially
[+high] and not [–anterior]. However, [±anterior] played a crucial role in
converting dental-alveolar consonants into palatoalveolars:
We believe this was not the correct approach. In the FUL model, Lahiri &
Reetz (2010) argued that the contrasts enabled by [±anterior] and
[±distributed] were adequately dealt with by TONGUE HEIGHT [HIGH] &
[LOW] features along with [ATR] and [RTR]. Within FUL, a move from dental
to palatoalveolar or palatal would not be a change in main articulators. It
would be more of a change to stridency lead by a combination of the feature
[HIGH] and [CORONAL], both enhancing high frequency energy. Thus, the
usual dental/alveolar change can lead to [ʃ] or [tʃ], i.e., a change to fricative
or an affricate, both of which must be [STRIDENT]. Stridency comes as a
concomitant change because, as argued earlier in Lahiri & Blumstein
(1984), the palatoalveolar place of articulation cannot be articulated without
fricativisation. Clements (1986, 2001) came to the same conclusion that
there is concomitant affrication. Thus, dental palatalisation would be
formalised as follows:
What about secondary articulations involving palatalisation? We will argue
that these do not involve any ARTICULATOR feature, but only the feature
[HIGH].
For L&E, secondary palatalisations were not caused by a change of any
ARTICULATOR feature but by height features. In FUL, the palatalised versions
of all places of articulation likewise have a non-redundant [HIGH].

The contrast for other palatalised consonants as in the Finno-Ugric

language Ter Lapp would also be marked by [high]:

An analysis of the Dutch diminutive (Trommelen 1984; Gussenhoven &

Jacobs 2011, partially also discussed in L&E) sheds more light on
palatalisation where alveolars become palatoalveolars. Labials and velars
have a secondary articulation with devoiced [j].
The underlying form of the diminutive is assumed to have a coronal stop [t]
for two reasons. First, when the stem is disyllabic or contains a long vowel
and ends in a non-coronal sonorant such as [m] or [ŋ], the underlying [t]
assimilates in place, resulting in forms such as (38b iii, iv). Second, when
the stem contains a short vowel followed by a sonorant (38f), an [ə] is
inserted and the [t] becomes a palatalised stop [c]. Palatalisation is more
obvious in (38c) and (38d). The [t] of the diminutive is deleted after a stem
ending in an obstruent (38a). In (38c), after [t] is deleted, the stem-final
coronal obstruent ([s] or [t]) becomes [ʃ] or [c] respectively. The examples
in (xii) and (xiii) are particularly interesting because when there is a
sequence of two [t]s, one deletes and the remaining palatalises.
Gussenhoven & Jacobs (2011, 2017) state that the underlying form is
/tjə/ and the palatalised stop [c] is defined as [−anterior, −distributed] stop.1
07 For our purposes, since both /t/ and /j/ are placeless, we could simply
assume that there is only a CV morpheme. L&V analysed the diminutive
suffix in a similar fashion with an empty obstruent root followed by a
floating [−anterior] [+high] segment unspecified for any other feature and a
schwa.

However, given that all palatalised segments in FUL are represented by the
ARTICULATOR node with a [HIGH] under TONGUE HEIGHT, we could
represent the diminutive morpheme as in (40), with the features of the
relevant consonants in (41).

(40) Dutch diminutive morpheme in FUL

Palatalisation for the diminutive involves [s] becoming [ʃ] and [t] becoming
[c]. For FUL, both involve adding [high]. To illustrate, we show the features
of the relevant consonants in Dutch.
In (42) we state the rules which are required to obtain the diminutive forms.
The point we would like to make here is that the high front glide [j] which
is part of the diminutive morpheme has the feature [HIGH] which in turn
requires the obstruents [t s] to become [c ʃ] respectively. Since all
consonants in question are [CORONAL], there is nothing else that is required.
The only other relevant process is place assimilation where the place-
underspecified [t] acquires the place of the final consonant in words like
[raːmpjə] from /raːm – tjə/. Sample derivations are added in (43).

(42) Diminutive formation in FUL

(i) [ə] insertion: V [SONORANT] — DIMINUTIVE suffix
(ii) [t] deletion. [OBSTRUENT]DIMINUTIVE > Ø/ [OBSTRUENT] —
(iii) PLACE ASSIMILATION: spread specified features to the
ARTICULATOR node
(iv) PALATALISATION: spread feature [high], delete feature [cons]

(43) Sample derivations (the rule numbers refer to those in (42))

As for secondary palatalisation in forms such as [rɔkjə] ‘skirt-dim’, the
crucial point we have made is that there is no real necessity to mark
palatalised segments with special sets of features or as complex segments.
Secondary palatalisation need not involve a change in place; thus,
palatalised [k], which is [kj], would be represented as as ART[DORSAL],
TH[HIGH], without change of ARTICULATOR. But dental consonants can
change place of articulation. Palatalised dentals often undergo a change of
place of articulation such as the Dutch [t] becoming [c]. Under our
conception this can happen only if the context includes a high front vowel
or a glide, not any front vowel. The general process of assimilation invoves
the spreading of [high]. Palatalised coronal stops can often also become
strident affricates or fricatives because a combination of [CORONAL] and
[HIGH] would provide a greater energy in the higher frequencies, a
characteristic of strident segments. The addition of [strident] comes as a
“free ride” since the unmarked articulation of all obstruents in the
palatoalveolar region is with stridency (Lahiri & Blumstein 1984: 142).108
Second, it has been claimed that non-high front vowels can trigger
secondary palatalisation for example in Nupe and Kinyarwanda (Sagey
1986: 209–218, 227–240). In these languages, secondary palatalisation is
triggered not only by [i], but also by [e]. As in other languages, these
consonants are phonetically produced by an offglide. We would argue (as in
L&E, p. 95) that in these cases this glide could also be present
phonologically as an onglide of the vowel such that [e] would be
underlyingly [je]. If this is the case, it would predict that in the course of
time the [j] triggering the palatalisation would be absorbed by the vowel
and the consonant survives with a single major ARTICULATOR plus a
TONGUE HEIGHT feature [HIGH]. This can be reflected in the orthography
as in Russian, where the palatalisation mark of the consonant rests on the
following vowel. Note that the fronting of velars is a matter of place change
whereby a [DORSAL] [k] becomes a [CORONAL] consonant. Here, the context
need not always be a high front vowel, but it could be any front vowel: it is
[CORONAL] status that matters. However, as always, [i] and [j] are favoured.

5 Morphophonological alternation and

language-specific underspecification
FUL accepts only monovalent and privative features along with
underspecification of [CORONAL] and [PLOSIVE]. Consequently, minus
features are not allowed. Since the mid to late eighties, the main articulator
features LABIAL, CORONAL, dorsal have been assumed to be monovalent. The
question we ask is whether it is possible to account for complex analyses
which may involve language-specific underspecification. This section is
based on Hyman’s analysis of vowel harmony in Kàlɔ̀ŋ, and what we would
like to show is that despite the sparse nature of the FUL system, the
complex set of alternations involving both specified and underspecified
place features for vowels can be accounted for.
In Hyman’s analysis of Kàlɔ̀ŋ, a Mbam language (Niger-Congo) of
Cameroon, the stem-affix alternations are governed by a limited set of
features, and there are three types of harmony: ATR, FRONT, and ROUND
harmony. First, we examine the vowel /a/ and its realisations in different
contexts. These realisations are summarised in (44).

The affix /a/ remains unchanged when the root also contains /a/ (iv). When
the root has a high vowel, /a/ becomes /e/ (i), while it takes on the features
of the root if they contain mid vowels /e ɛ o ɔ/.
 The underlying features of the relevant vowels in Hyman’s analysis are
given in (45).

Like Dresher and Clements, Hyman invokes the notion of “activity” and
argues that only the “active” four features that are necessary to account for
the data should be relevant. Using a system like that of Clements, ATR and
OPEN fall under a single APERTURE node. The vowel [ə] does not surface,
but is assumed to be the intermediate fronted vowel of /a/ when the root has
a high vowel / i u/.
Examples for the first three harmony cases are given in (46).
The features ATR, FRONT, and round participate actively in the harmony
process and the last two are “parasitic” on FRONT (Hyman 2003: 90). Our
interest here is in the vowel /a/, which changes to [e] not only in the context
of /i/, but also in the context of /u/ where, in a parallel scenario, it ought to
change to [o]. Hyman argues that “the fronting of /a/ under ATR harmony is
a secondary development, the primary one being to lower its F1”. That is,
/a/ “first converts to a [+ATR] central vowel, here symbolised as schwa”,
which in turn becomes /e/ (44(i)). Why should this be so in a perfectly
regulated harmony system? Why does the spreading of ATR ignore the place
features for the high vowels? We turn to FUL for an answer. (47), including
a tree diagram representation, gives the features that FUL would assign on
universal principles; note that CORONAL remains underspecified.
A clear distinction needs to be made between the underspecified CORONAL
and the lack of an ARTICULATOR node. As always, CORONAL is not
specified in the representation, but if the vowel has an ARTICULATOR
node, it will get the feature on the surface by a fill-in rule. Thus, /a/ will not
get a CORONAL specification, but / i e ɛ/ will. A futher lack of feature
specification involves TONGUE HEIGHT (TH) as well as TONGUE
ROOT (TR) features: / i u/ are not specified for height, but they do have the
TH node and /ɛ ɔ a/ are not specified for TR. The feature filling rules,
which determine the surface features, are as follows:

(48) Surface feature filling rules

(i) LABIAL ⇒ dorsal; any surface LABIAL vowel will get a dorsal
feature
(ii) Unfilled ARTICULATOR nodes will be assigned CORONAL; i.e., /
i e ɛ/ will be assigned CORONAL
 (iii) Unfilled TONGUE ROOT nodes will be assigned rtr; i.e., /ɛ, ɔ,
a/ will be assigned RTR

Thus, /a/ has the feature low without precise place features suggesting that
phonetically it can be in between.
Under this representational hypothesis, it is clear why, when ATR spreads
from /i u/, the suffix /a/ will automatically become /e/: /a/ and /e/ share the
feature [LOW] and nothing else. Thus, spreading [ATR] from /i, u/ to /a/ turns
it into /e/. The difference between /a/ and /e/ is that /a/ does not have an
ARTICULATOR node. This gets filled in on the surface where it will then
emerge as /e/ since atr has spread. Consequently, unlike in Hyman’s
analysis, /a/ > [e] does not require an intermediate analysis which produces
[ə] (49i). The harmony processes for high and mid vowels look different
because of the mismatch between the TH features. These are spelt out with
relevant examples below.

Hyman mentions another interesting harmony process that includes two

different underlying vowels which sometimes surface as [i u] but not
always:
Underlying /I U/ are realised as /ɛ ɔ/ in closed syllables and as / i u/ in open
syllables. With the suffix /-a/, underlying /I U/ are not realised as *kù-lík-è
and *kù-lùk-è, since underlying /a/ does not undergo atr harmony, but
underlying /ɛ ɔ/ remain in open syllables and /a/ assimilates to them.
Consequently, Hyman analyses the vowels /I U/ as having only the features
front and round; neither atr nor open are specified, which accounts for
several distributions:

(51) Hyman’s analysis of /I, U/ compared with / i u/

Lack of ATR and OPEN, then, ensue in the lack of harmony alternations.
 Is it possible to account for this complex situation in FUL, where not
only CORONAL is underspecified, but the vowels /I U/ must lack height as
well as ATR features? Our proposal is outlined in (53).
Thus, in closed syllables, the addition of [low] for unspecified TH of /I U/
would give /ɛ ɔ/, while in open syllables they would receive the feature
[atr].

We have seen, then, that the underspecification of [CORONAL] does not

prevent FUL from accounting for such complex patterns as vowel harmony
in Kàlɔ̀ŋ in a principled way. What is additionally required here is the lack
of specification of ATR and OPEN (i.e., height) features, not a problem in the
FUL framework: the raising and fronting of /a/ to [e] can then be achieved
without an intermediate step.

6 Moving on
To sum up, FUL provides a set of monovalent features, along with
underspecification of [CORONAL] and [PLOSIVE], which are intended to be
universal. Thus, binary features like [±high] or [±voice] are not acceptable
and the automatic consequence is that negative features cannot form natural
classes. However, it is possible to refer to a node which does not contain a
fully specified feature. Thus, ARTICULATOR remains empty for CORONAL,
which gets filled in on the surface. Rules like English aspiration of
voiceless consonants (under the assumption that underlying stops are
unaspirated) could be realised as below:

The rule of aspiration says that when the LARYNGEAL node is “empty”
and does not contain either SPREAD GLOTTIS or VOICE, the feature SPREAD
GLOTTIS would be added. When, on the other hand, the feature VOICE is part
of the LARYNGEAL node, then SPREAD GLOTTIS would not be added. Thus,
in a word like /pɪn/, the initial consonant has no laryngeal feature and
acquires SPREAD GLOTTIS in syllable-initial position, but since the
LARYNGEAL node for /b/ (in words like / bɪn/) are already specified with
the feature voice, no other feature can be added. Consequently, /b/ remains
without aspiration.
In fleshing out a model like FUL, a host of further questions need to be
tackled. We will only broach three here: they are ones where significant
progress has been or is being made. First, since unlike contrastive theories
that assume activation we assume that universal features are acquired first
and always establish a contrast, then how do the other features become part
of the system? Second, if CORONAL and PLOSIVE are always underspecified,
then they must always be available in natural languages; but are they?
Finally, we have claimed that underspecification has consequences for
processing: but to what extent do we have evidence supporting this?
With respect to acquisition, if CORONAL must always be present, then
the first cut is CORONAL vs. something else. Following Ghini (2001a), we
maintain that PLACE-first is a universal principle. The acquisition literature
suggests that LABIAL is produced first (cf. Jakobson 1941; Levelt 1995;
Fikkert & Levelt 2008). Fikkert & Levelt find that words are
undifferentiated with respect to features and the word node itself has
LABIAL, with vowels and consonants sharing the same feature. Our
assumption is that CORONAL is underspecified but present, and in fact the
LABIAL VS. CORONAL contrast is the first one to be manifest on the surface.
We also assume that all languages have PLOSIVES – not necessarily all places
of articulation, but at least one. This tallies with Hyman (2008) who argues
that two of the valid universals about phonological inventories are that all
have oral stops and all have coronals. But CORONAL phonemes need not be
PLOSIVES; they could be CONTINUANT for instance. Thus, in acquisition, we
would first find a contrast of underspecified CORONAL vs. some other
ARTICULATOR (in all probability LABIAL) and PLOSIVE vs. probably
CONTINUANT. Recall that FUL assumes that vowels and consonants share
PLACE. Thus, for vowels as well, the first cut is probably CORONAL VS.
LABIAL. It could be the case that the LABIAL vowels are also DORSAL.
We have also suggested that in terms of TONGUE HEIGHT, [low] is
acquired first. But we do not believe that this needs to be underspecified
universally, because a language might only have one vowel, with no
necessity to specify any height contrast. Thus, other features are built very
much on the basis of contrast. The question is whether contrasts depend
entirely on “activity” or on distribution. The answer is probably both.
Initially, infants are not going to be exposed to lots of alternations which
would conclusively estabish activity. However, distribution is something
they inevitably enounter right away.
Challenging the assumption of the universality of coronals, Blevins
(2009) has suggested that Northwest Mekeo lacks CORONAL obstruents,
though it may acquire them via language contact. All Mekeo dialects,
however, have coronal sonorants; /l/ occurs in other Mekeo dialects and
Northwest Mekeo itself has a palatal glide /y/ (Blevins’ notation) which
alternates with /ɛ/. Blevins argues that /l/ can be seen as primarily lateral
with redundant coronal specification. That is not an assumption made by
FUL, where PLACE is primary. Consequently, it is not the case that this
universal “bites the dust”: CORONAL is very much present even in Northwest
Mekeo, albeit perhaps not in obstruents. In Blevins’ own terms, CORONAL
appears on the surface via assimilation, and with /i/.

This fits in with FUL’s assumptions perfectly. In FUL /i/ is CORONAL

(underspecified, with only the ARTICULATOR node), and in its context, as
we have seen earlier for palatalisation, DORSAL consonants will lose their
feature. Our analysis is in (57).

Whether the result is an affricate or is pronounced with a palatalised

affricate is a matter of phonetic implementation. The crucial point is that /g/
loses its DORSAL feature in the CONTEXT of a CORONAL underspecified high
vowel.
Finally, in the FUL approach, underspecification in representations is
intended to have consequences for processing. What is the evidence? We
have shown in several experimental studies that CORONAL underspecifiation
predicts asymmetries. For example, in an MMN (mismatch negativity)
paradigm in an EEG experiment in German, when listeners were saturated
with the nonsense syllable [egi] (played several times) and were then
provided with a deviant stimulus [edi], the surface coronal feature from /d/
was found to mismatch with the DORSAL representation of /g/, triggering a
high negative peak. However, the negative peak was significantly lower
when the presentation of stimuli was reversed: when [edi] was the standard
(surface CORONAL, mapping onto an underspecified representation) and was
followed by deviant [egi], then the surface DORSAL was tolerated by the
underspecified representation. The same asymmetric pattern is found with
underspecified [PLOSIVE] and specified [NASAL]. This is illustrated in (58)
and (59).
Aymmetries have also been observed in several other experimental designs
such as lexical decision tasks with semantic priming (Roberts et al. 2013;
Lahiri & Reetz 2010; Eulitz & Lahiri 2004).111
Typologically, FUL’s general goal is to define and regulate a set of
features which can cover all possible contrasts and alternations in the
languages of the world; the ability to account for acquisition and processing
are important added bonuses. Our focus here was on the CORONAL node
where the largest set of contrasts needs to be accommodated; but, naturally,
other contrasts, such as pharyngeal ones coming under the RADICAL node,
would equally be taken care of along similar lines. Insofar as contrasts and
alternations, however crosslinguistically diverse, fall into just those patterns
that are dictated by a particular theoretical model, FUL, and not into any
others conceivable, fundamental unity is revealed behind diversity.

Acknowledgement: This work was partially supported by the ERC

Advanced Research Grant MORPHON 695481, PI Aditi Lahiri.

References
Archangeli, Diana. 1988. Aspects of underspecification theory. Phonology 5. 183–
207.
Bhat, D. N. S. 1978. A general study of palatalization. In Joseph H. Greenberg,
Charles A. Ferguson, & Edith Moravcsik (eds.), Universals of language, vol. 2:
 Phonology, 47–92. Stanford: Stanford University Press.
Blumstein, Sheila & Kenneth Stevens. 1980. Perceptual invariance and onset
spectra for stop consonants in different vowel environments. Journal of the
Acoustical Society of America 67. 648–662.
Blevins, Juliette. 2009. Another universal bites the dust: Northwest Mekeo lacks
coronal phonemes. Oceanic Linguistics 48. 264–273.
Calabrese, Andrea. 1995. A constraint-based theory of phonological markedness
and simplification procedures. Linguistic Inquiry 26. 373–463.
Chomsky, Noam & Morris Halle. 1968. The sound pattern of English. New York:
Harper & Row.
Clements, George N. 1985. The geometry of phonological features. Phonology
Yearbook 2. 225–252.
Clements, G. Nick. 1989. A unified set of features for consonants and vowels.
Unpublished, Cornell University, Ithaca, NY.
Clements, G. Nick. 2001. Representational economy in constraint-based
phonology. In Hall (ed.) 2001, 71–146.
Clements, George N. & Elizabeth V. Hume. 1995. The internal organization of
speech sounds. In John A. Goldsmith (ed.), The handbook of phonological
theory, 245–306. Oxford: Blackwell.
Compton, Richard & B. Elan Dresher. 2011. Palatalization and “strong i” across
Inuit dialects. Canadian Journal of Linguistics 56. 203–228.
Cornell, Sonia A., Carsten Eulitz, & Aditi Lahiri. 2013. Inequality across
consonantal contrasts in speech perception: Evidence from mismatch
negativity. Journal of Experimental Psychology: Human Perception and
Performance 39. 757–772.
Dresher, B. Elan. 2009. The contrastive hierarchy in phonology. Cambridge:
Cambridge University Press.
Eulitz, Carsten & Aditi Lahiri. 2004. Neurobiological evidence for abstract
phonological representations in the mental lexicon during speech recognition.
Journal of Cognitive Neuroscience 16. 577–583.
Fikkert, Paula & Clara Levelt. 2008. How does Place fall into place? The lexicon
and emergent constraints. In Peter Avery, B. Elan Dresher, & Keren Rice
(eds.), Contrast in phonology: Theory, perception, acquisition, 231–270.
Berlin: Mouton de Gruyter.
Ghini, Mirco. 2001a. Asymmetries in the phonology of Miogliola. Berlin: Mouton de
Gruyter.
Ghini, Mirco. 2001b. Place of articulation first. In Hall (ed.) 2001, 71–146.
Gussenhoven, Carlos & Haike Jacobs. 2011. Understanding phonology. 3rd edn
2017. London: Hodder Education.
Hall, T. Alan. 1997. The phonology of coronals. Amsterdam: Benjamins.
Hall, T. Alan (ed.). 2001. Distinctive feature theory. Berlin: Mouton de Gruyter.
Halle, Morris, Bert Vaux, & Andrew Wolfe. 2000. On feature spreading and the
representation of place of articulation. Linguistic Inquiry 31. 387–444.
Hyman, Larry M. 2003. “Abstract” vowel harmony in Kàlɔ̀ŋ: A system-driven
account. In Patrick Sauzet & Anne Zribi-Hertz (eds.) Typologie des langues
d’Afrique et universaux de la grammaire, vol. 1, 85–112. Paris: L’Harmattan.
Hyman, Larry M. 2008. Universals in phonology. The Linguistic Review 25. 83–
137.
Jakobson, Roman. 1941. Kindersprache, Aphasie und allgemeine Lautgesetze.
Uppsala: Almqvist & Wiksell.
Jakobson, Roman, Gunnar Fant, & Morris Halle. 1952. Preliminaries to speech
analysis: The distinctive features and their correlates (Technical Report No.
13). Cambridge, MA: MIT, Acoustics Laboratory.
Jakobson, Roman & Morris Halle. 1956. Fundamentals of language. The Hague:
Mouton.
Kaplan, Lawrence D. 1981. Phonological issus in North Alaskan Inupiaq.
Fairbanks: Alaska Native Language Center.
Keating, Patricia & Aditi Lahiri. 1993. Fronted velars, palatalized velars, and
palatals. Phonetica 50. 73–101.
Kotzor, Sandra, Allison Wetterlin, & Aditi Lahiri. 2017. Symmetry or asymmetry:
Evidence for underspecification in the mental lexicon. In Aditi Lahiri & Sandra
Kotzor (eds.), The speech processing lexicon, 85–106. Berlin: De Gruyter
Mouton.
Lahiri, Aditi. 2000. Hierarchical restructuring in the creation of verbal morphology in
Bengali and Germanic: Evidence from phonology. In Aditi Lahiri (ed.), Analogy
and markedness: Principles of change in phonology and morphology, 71–123.
Berlin: Mouton de Gruyter, paperback 2003.
Lahiri, Aditi. 2012. Asymmetric phonological representations of words in the mental
lexicon. In Abigail C. Cohn, Cécile Fougeron, & Marie Huffman (eds.), The
Oxford handbook of laboratory phonology, 146–161. Oxford: Oxford University
Press.
Lahiri, Aditi & Sheila E. Blumstein. 1984. A re-evaluation of the feature ‘coronal’.
Journal of Phonetics 12. 133–145.
Lahiri, Aditi & Vincent Evers. 1991. Palatalization and coronality. In Paradis &
Prunet (eds.) 1991, 79–100.
Lahiri, Aditi, Letitia Gewirth, & Sheila E. Blumstein. 1984. A reconsideration of
acoustic invariance for place of articulation in diffuse stop consonants:
Evidence from a cross-language study. Journal of the Acoustical Society of
America 76. 391–404.
Lahiri, Aditi & Henning Reetz. 2002. Underspecified recognition. In Carlos
Gussenhoven & Natasha Warner (eds.), Labphon 7, 637–676. Berlin: Mouton
de Gruyter.
Lahiri, Aditi & Henning Reetz. 2010. Distinctive features: Phonological
underspecification in representation and processing. Journal of Phonetics 28.
44–59.
Levelt, Clara. 1995. Segmental structure of early words: Articulatory frames or
phonological constraints. In Eve V. Clark (ed.), The Proceedings of the
Twenty-seventh Annual Child Language Research Forum, 19–27. Stanford:
CSLI.
McCarthy, John. 1988. Feature geometry and dependency: A review. Phonetica
43. 84–108.
Mohanan, K. P. 1993. Fields of attraction in phonology. In John Goldsmith (ed.),
The last phonological rule: Reflections on constraints and derivations, 61–116.
Chicago: University of Chicago Press.
Mohanan, K. P. & Tara Mohanan. 1984. Lexical phonology of the consonant
system in Malayalam. Linguistic Inquiry 15. 575–602.
Paradis, Carole & Jean-François Prunet (eds.). 1991. The special status of
coronals (Phonetics and Phonology 2). San Diego: Academic Press.
Paulian, Christiane. 1986. Les voyelles en nù-kàlùŋɛ̀: Sept phonèmes, mais ...
Cahiers du Lacito 1986/1. 51–65.
Pierrehumbert, Janet B. 2016. Phonological representation: Beyond abstract
versus episodic. Annual Review of Linguistics 2. 33–52.
Plank, Frans & Aditi Lahiri. 2015. Macroscopic and microscopic typology: Basic
Valence Orientation, more pertinacious than meets the naked eye. Linguistic
Typology 19. 1–54.
Roberts, Adam, Allison Wetterlin, & Aditi Lahiri. 2013. Aligning mispronounced
words to meaning. The Mental Lexicon 8. 140–163.
Rubach, Jerzy. 1984. Cyclic and lexical phonology: The structure of Polish.
Dordrecht: Foris.
Sagey, Elizabeth C. 1986. The representation of features and relations in non-
linear phonology. Doctoral dissertation, MIT, Cambridge, MA.
Shaw, Patricia. 1991. Consonant harmony systems: The special status of coronal
harmony. In Paradis & Prunet (eds.) 1991, 125–157.
Stevens, Kenneth & Sheila E. Blumstein. 1978. Invariant cues for place of
articulation in stop consonants. Journal of the Acoustical Society of America
64. 1358–1368.
Trommelen, Mieke. 1984. The syllable in Dutch. Dordrecht: Foris.
B. Elan Dresher, Christopher Harvey, and Will Oxford
Contrastive feature hierarchies as a
new lens on typology

Abstract: We propose a way of looking at phonological typology that is

based on a fundamental distinction between a phonetic and phonological
analysis of the sound systems of languages. We build on approaches to
phonology pioneered by Sapir and the Prague School (Jakobson and
Trubetzkoy), instantiated within a generative grammar. We view phonemes
as being composed of contrastive features that are themselves organized
into language-particular hierarchies. We propose that these contrastive
feature hierarchies shed light on synchronic and diachronic phonological
patterns, and therefore offer a new lens on phonological typology. Thus, on
this view the subject matter for typological investigation is not a phonetic
sound (e.g., [i]) or a phoneme (/i/), or even a phonemic inventory (/i, a, u/),
but an inventory generated by a feature hierarchy: for example, / i, a, u/
generated by the hierarchy [low] > [round]. This yields a different set of
representations from the same terminal symbols generated by the hierarchy
[round] > [low].
We will illustrate this approach to phonological representations with a
synchronic analysis of Classical Manchu, and then show how it accounts
for the results of typological surveys of rounding harmony in Manchu-
Tungusic, Eastern Mongolian, and Turkic, and for the distribution of
palatalization in Yupik-Inuit dialects. We will then propose that contrast
shift should be recognized as a type of phonological change, and show how
it applies to diachronic developments of the Algonquian and Ob-Ugric
vowel systems. We find that feature hierarchies can be relatively stable, but
contrast shifts do occur, for various reasons, and these can result in dramatic
differences in patterning. Harvey’s analysis of Ob-Ugric also shows that
elements of feature hierarchies can spread and be borrowed, like other
aspects of linguistic structure. As Sapir (1925) proposed, languages whose
phonemes line up in similar ways (i.e., have similar contrastive feature
hierarchies) show similar phonological patterning, though they may differ
considerably in their phonetic realizations. We conclude that contrastive
feature hierarchies provide an interesting level of representation for
typological research.

1 Introduction
This article addresses a question raised in the proposal for the Workshop on
Phonological Typology (Oxford University, August 2013): Phonological
typology vs. phonetic typology – same or different? We will propose a way
of looking at phonological typology that is clearly DIFFERENT from phonetic
typology. In particular, we will propose that CONTRASTIVE FEATURE
HIERARCHIES offer a new lens on typology, while also shedding light on
synchronic and diachronic phonological patterns.
We will begin in Section 2 with some general remarks on typology,
phonological contrast, and contrastive feature hierarchies. Section 3
illustrates the relation between contrast and phonological activity, as
exemplified by the Classical Manchu vowel system. We then show how
contrastive hierarchies can lend insight to synchronic, diachronic, and areal
typology, with examples drawn from a typological survey of rounding
harmony and the relative ordering of features [round] and [front] (Section
4), the diachrony of Algonquian vowel systems (Section 5), and areal
typology of Ob-Ugric vowel systems (Section 6), respectively. Section 7 is
a brief conclusion.

2 Typology, phonological contrast, and

contrastive feature hierarchies

2.1 Phonological typology and contrast

Following Hyman (2007), the kind of typology we will be concerned with
is “an underlying one, based on phonological analysis, not on surface
inventories”. Hyman cites Vajda’s (2001) view of phonological typology:
“it is possible to classify languages according to the PHONEMES they contain
[. . .] Typology is the study of STRUCTURAL FEATURES across languages.
Phonological typology involves comparing languages according to the
number or type of sounds they contain” [emphasis added]. We will build on
this view by advancing a specific notion of the terms “phonemes”,
“structural features”, and “number or type of sounds”.
In the same article, Hyman (2007) cites Sapir’s (1925: 43) “intrinsically
typological” idea that “two languages, A and B, may have identical sounds
but utterly distinct phonetic [read: phonological] patterns”. Sapir also
constructs two languages C and D that illustrate the converse situation:
phonetically their sounds are different, but their “pattern alignments” are
isomorphic. Sapir (1925) arranges the phonemes as in (1).

The phonemes /v/ and /ʒ/ appear to be out of place in the chart of language
D, but Sapir justifies their positions by their phonological behaviour, in that
their places in the pattern are parallel to those of language C’s /w/ and /j/,
respectively. Sapir (1925: 47–48) allows that the “natural phonetic
arrangement” of sounds is a useful guide to how they pattern, but he goes
on: “And yet it is most important to emphasize the fact, strange but
indubitable, that a pattern alignment does not need to correspond exactly to
the more obvious phonetic one.”
The isomorphic alignments in C and D can be understood as indicating
that corresponding phonemes have the same CONTRASTIVE values. The chart
in (2) represents one possible way of suggesting what the contrastive
specifications might be for the consonants in (1). In each cell, the first
sound is from C, the second from D. The differences between them do not
involve contrastive specifications.

It was observed that the language D phonemes /v/ and /ʒ/ appear to be in the
“wrong place”, which in (2) translates into their having incorrect
specifications. In generative grammar, this mismatch can be resolved by
assigning them different underlying specifications, matching those of their
counterparts. These types of examples have been much discussed in
connection with how abstract Sapir’s theory of phonology was (cf.
McCawley 1967). Less attention has been paid to the other examples, which
do not appeal to abstractness, but which show the importance of
establishing the contrastive properties of segments. For example, the
obstruents in the third row in (2) are contrastively voiced and redundantly
stops or spirants. No abstractness is at issue here, but we have to distinguish
between contrastive and non-contrastive properties.
It follows that for Sapir the pattern alignment of a phoneme amounts to
its contrastive status, which is not determined by its phonetics, but is a
function of its phonetic and phonological behaviour. Thus, a synchronic
analysis of the phonology should, among other things, give an account of
the contrastive features of each phoneme.
Turning to diachrony, Prague School phonologists have argued that the
contrastive properties of phonemes also play an important role in
phonological change. The insight that phonological change may involve a
reorganization of the phonemes of a language goes back to an article by
Roman Jakobson first published in 1931 (Jakobson 1972 [1931]): “Once a
phonological change has taken place, the following questions must be
asked: What exactly has been modified within the phonological system? [. .
.] has the structure of individual oppositions [contrasts] been transformed?
Or in other words, has the place of a specific opposition been changed [. .
.]?”
It should be noted that phonological theories that put the emphasis on
contrast have not been unproblematic. In pre-generative structuralist
theories, synchronic grammars were composed of contrasting elements
locked into systems of oppositions. If one takes too literally Saussure’s
(1972 [1916]: 166) dictum that “dans la langue il n’y a que des différences
[. . .] sans termes positifs” then grammars become incommensurable, and
one has no way to relate successive stages of a language, or even closely
related dialects (Moulton 1960). Generative grammar (Chomsky & Halle
1968) solves this problem by construing phonology as a system of rules that
mediate between underlying (lexical) and surface (phonetic) forms. Now,
grammar change takes the form of the addition, loss, reordering, or
restructuring of rules.
Kiparsky (1965) demonstrated that a series of sound changes in
Armenian dialects, shown in (3), can be understood in terms of the
spreading of three rules, described informally in (4). Kiparsky (1965) points
out that these sound changes spread from one dialect to another, regardless
of how many contrasts they contained. If we were to classify the dialects in
terms of oppositions, we would arrive at meaningless groupings for
explaining any synchronic or diachronic facts. He writes:
 An incidental feature of the present example is that it highlights the pointlessness of a
structural dialectology that [. . .] distinguishes dialects according to points of structural
difference rather than according to the innovations through which they diverged [. . .] If
in the present example we were to divide the dialects into those with two stop series
and those with three, we would be linking together dialects that have nothing to do with
each other and separating dialects that are closely related. (Kiparsky 1965: 17)

(4) Armenian sound changes (Kiparsky 1965)

a. Aspiration: /d/ aspirates to [dh] (or [th]) in the Central and
Northwestern dialects.
b. Voicing: /t/ voices to [d] in the Western, West Central,
Northwestern, and Southern dialects.
c. Devoicing: /d/ devoices to [t] in the Eastern and Southern dialects.

The above considerations show the inadequacy of a phonology that deals

only in structural points of contrast (“differences”), without also including
substantive properties (“positive terms”), including features and a system of
rules or constraints. However, we believe that generative grammar went
overboard in jettisoning the structuralist notion of language-particular
contrast. We will argue that contrast plays a crucial role in synchronic and
diachronic phonology, and hence in phonological typology.

2.2 A theory of phonological contrast

To implement contrast in an explicit theory, we assume first that contrastive
features are assigned hierarchically, using a method that was called
“branching trees” in the literature of the 1950s and 60s (Jakobson, Fant, &
Halle 1952; Jakobson & Halle 1956), stated in (5). We call it the Successive
Division Algorithm (Dresher 1998, 2003, 2009), given informally in (6):

(5) The contrastive feature hierarchy (based on Jakobson, Fant, & Halle
1952, among others):
Contrastive features are assigned by language-particular feature
hierarchies.

(6) The Successive Division Algorithm:

Assign contrastive features by successively dividing the inventory
until every phoneme has been distinguished

As a first approximation we assume further that phonology computes only

contrastive features, in keeping with the Contrastivist Hypothesis in (7).

(7) The Contrastivist Hypothesis (Hall 2007):

The phonological component of a language L operates only on those
features which are necessary to distinguish the phonemes of L from
one another.

That is, only contrastive features can be PHONOLOGICALLY ACTIVE, where

feature activity is defined as in (8):

(8) Phonological activity (adapted from Clements (2001: 77):

A feature can be said to be active if it plays a role in the phonological
computation; that is, if it is required for the expression of phonological
regularities in a language, including both static phonotactic patterns
and patterns of alternation.

If the Contrastivist Hypothesis is correct, (9) follows as its corollary.

(9) Corollary to the Contrastivist Hypothesis:

If a feature is phonologically active, then it must be contrastive.
This corollary suggests a working heuristic: Assume that active features are
contrastive, and find, if possible, a feature ordering that fits the observed
patterns of activity. We believe that this heuristic represents the practice of
many descriptive phonologists, minus the requirement that all active
features are necessarily contrastive. That is, phonologists typically limit
their analyses to those features that are relevant to the workings of the
language, and these active features also serve as the contrastive features, as
far as possible.
A further assumption is that features are binary, and that every feature
has a MARKED and UNMARKED value. We assume, as in (10), that markedness
is language particular (Rice 2003, 2007) and accounts for asymmetries
between the two values of a feature, where these exist.112 Where the
asymmetry is substantial, a feature may appear to act in a privative manner,
so that the unmarked value may appear to be absent. In other cases, both
values of a feature may be referred to by the phonology (Mackenzie 2011,
2013). We will designate the marked value of a feature F as [F], and the
unmarked value as (non-F).113

(10) Feature markedness:

Each feature F has a marked value, [F], and an unmarked value,
(non-F). Where these values function asymmetrically, the marked
value is the more active one.

Finally, this theory of contrast does not need to make any assumptions
as to where features come from: the Successive Division Algorithm works
equally well if features are universal, as supposed by Chomsky & Halle
(1968), or emergent, as suggested by Mielke (2008) and Samuel (2011).
Dresher (2014) observes that the contrastive hierarchy itself ensures that
phonological representations across languages will look rather similar even
in the absence of a universal set of features.
To illustrate the workings of the feature hierarchy and the Contrastivist
Hypothesis, consider a hypothetical vowel inventory / i, u, a/. The
Successive Division Algorithm requires that an inventory of three
phonemes must be characterized by exactly two features, though both the
choice of features and their ordering may vary. In (11), we illustrate two
possible contrastive hierarchies that use the features [back] and [low]; in
(12), we give two more hierarchies using the features [front] and [round].
Other combinations of features are also possible, but these examples should
suffice to illustrate the concept.

The feature hierarchy constrains phonological activity in a number of ways.

First, it follows from (7) that both /a/ and /u/ can potentially trigger backing
in (11a), because both are contrastively [back]; in (11b), only /u/ is
contrastively [back], so that is the only potential phoneme that could cause
backing. In (12), the feature [back] is not contrastive in the vowel system at
all, and we do not expect any vowel to cause backing.
Second, the hierarchy constrains neutralization and merger: we make
the hypothesis in (13). In (11a) and (12a), we expect that /u/ could merge
with /a/, whereas in (11b) it would more likely merge with /i/. Note that this
restriction does not apply to ordinary synchronic processes. For example, in
both languages in (12) /i/ is contrastively [front] and /u/ is contrastively
[round]; therefore, both languages may have harmony processes whereby
/a/ is fronted in the environment of /i/ and rounded in the environment of
/u/, whether or not /a/ is a contrastive sister of /i/ or /u/. Though /a/ can
alternate synchronically with both /i/ and /u/, depending on position, it can
only merge diachronically with one of these vowels.114

(13) Hypothesis concerning diachronic mergers:

Mergers affect phonemes that are CONTRASTIVE SISTERS.

The typological generalizations we will be discussing can thus not be found

by looking at inventories alone (say, / i, a, u/), or at individual phonemes
(say, /a/), or phones ([a]), without also considering the relevant contrastive
feature hierarchy. Notice also that a consequence of this hierarchical
method for assigning contrastive features is that a contrastive specification
need not be unpredictable. For example, in (11a) /a/ is the only [low] vowel,
so its [back] feature is predictable; but it is still contrastive, for it
distinguishes between /a, u/ and /i/.

3 Example of contrast and activity: The

Classical Manchu vowel system
In this section we will illustrate the connection between contrast and
phonological activity, taking as an extended example the Classical Manchu
vowel system, following the analysis of Zhang (1996) and Dresher &
Zhang (2005). Classical Manchu has six vowel phonemes, as shown in (14).
3.1 Contrastive feature hierarchy for Classical
Manchu
Based on the phonological patterning of the Classical Manchu vowels,
Zhang (1996) proposes the feature hierarchy in (15), which yields the
marked feature representations in (16).115

The three most notable kinds of phonological activity involving vowels are
ATR harmony, rounding (labial) harmony, and palatalization. We will
briefly discuss them in turn, and show how the patterns of activity motivate
the hierarchy in (15).

3.2 ATR harmony

The vowels /ə/ and /u/ trigger ATR harmony within a word: /ə/ alternates
with /a/ (17a) and /u/ alternates with /ʊ/ (17b).

An apparent exception is caused by the fact that /ʊ/ changes to [u]

everywhere except after dorsal (velar ~ uvular) consonants; however, the
underlying contrast between /ʊ/ and /u/ emerges in the way they participate
in ATR harmony (17): underlying /u/ co-occurs with ATR vowels (18a),
underlying /ʊ/ co-occurs with non-ATR vowels (18b).

The vowel /i/ is neutral and co-occurs in stems with both ATR (19a) and
non-ATR vowels (19b). Similarly, suffix /i/ freely occurs with both types of
vowels (19c).
Perhaps unexpectedly, when /i/ is in a position to trigger harmony, it occurs
only with non-ATR vowels (20).

The evidence from activity, therefore, is that /ə/ and /u/ have an active
feature in common, that we are calling [ATR], that is not shared by the
other vowels; by hypothesis, this feature must be contrastive. The same is
evidently not the case with /i/, though /i/ is phonetically ATR. In the
representations proposed in (15) and (16), /ə/ and /u/, but not /i/, are
contrastively [ATR].

3.3 Round (labial) harmony

Two successive /ɔ/ vowels cause a suffix /a/ to become /ɔ/ (21a); a single /
ɔ/, short or long, does not trigger rounding (21b).116 Note that /u/ and /ʊ/ do
not trigger round harmony (22).
The evidence from activity here, then, is that /ɔ/ must have an active,
therefore contrastive, feature that causes rounding, which we are calling
[round]; the same is not the case with /u/ and /ʊ/, though they are also
phonetically rounded. The feature ordering in (15) has the result that /ɔ/ is
contrastively [round], but /u/ and /ʊ/ are not.

3.4 Palatalization
The vowel /i/ uniquely causes palatalization of a preceding consonant,
which suggests that it alone has a contrastive triggering feature we call
[front]. There is no evidence that it has any other active features.

3.5 Height contrast

The alternations /ə/ ~ /a/ ~ /ɔ/ and /u/ ~ /ʊ/ are limited to a height class, and
we still need to distinguish /ə/ from /u/ and /a/ from /ʊ/. It is simplest to
assume one height contrast, which we call [low] (as there are only two
height classes, [high] would also be possible here). As shown in (15), no
more features are required in ordered to make each vowel distinct from
every other, and there is no evidence that any other feature is active in this
vowel system.

4 Synchrony: Typology with contrastive feature

hierarchies
Contrastive feature hierarchies allow us to update Sapir’s approach to
phonological systems and view phonological typology in a new way. Rather
than considering only the number of segments in an inventory, or their
geometrical arrangement, we can look at inventories in terms of their
active/contrastive features and how they are ordered. As with Sapir’s
languages A–D, this approach reveals unexpected similarities between
inventories that do not superficially look very similar; conversely,
inventories that look quite similar may turn out to have different patterns of
phonological activity because they have different contrastive hierarchies.
To illustrate, we will consider a number of vowel systems which have
contrasts between front and back round vowels. Two features that could
potentially play a role in such inventories are [front] and [round]. We have
seen that both of these features are contrastive and active in Classical
Manchu vowels, but in asymmetrical ways: whereas /i/ is contrastively
[front], [round] is restricted to /ɔ/, and is not contrastive in /u/ and /ʊ/. This
is a result of ordering [front] > [round]; if the ordering were [round] >
[front], then /u/ and /ʊ/ would be contrastively [round], and /i/ would not be
assigned [front]. If the orderings of these features is allowed to vary cross-
linguistically, we expect to find vowel systems that manifest each ordering.
What the specific consequences of these orderings are in any given
language depends on the number of segments in the inventory, and the
ordering of other contrastive features.

4.1 Vowel systems with [front] > [round]

Contrastive feature hierarchies shed new light on the results of typological
surveys of rounding (labial) harmony in Manchu-Tungusic, Mongolian, and
Turkic (Korn 1969; Kaun 1995). We have seen that round harmony in
Classical Manchu is limited to the [low] vowels. On our account, only the
low vowel /ɔ/ is contrastively [round] in this inventory, because of the
ordering of [front] > [round]. The same holds for most Manchu-Tungusic
languages, which have similar vowel inventories. A Tungusic example is
Oroqen (Zhang 1996), shown in (23): again, only low vowels are triggers
(in the solid box) and targets (in the dashed box) of harmony. Oroqen has
both ATR and non-ATR low vowels. We assume it has the same feature
hierarchy as Classical Manchu (plus a length contrast that we omit from the
tree), as shown in (24).

The schematic representation of the Oroqen vowel system in (23) might

suggest a different explanation for the fact that [round] is active only in the
[low] part of the inventory. One might suppose that one could simply read
off this display that rounding is not contrastive in the high vowels, because
there is no /ɨ/ in the inventory. That is, one might think that the contrasts
between / o, oo, ɔ, ɔɔ/ and /ə, əə, a, aa/ are more minimal than the ones
between /u, uu, ʊ, ʊʊ/ and /i, ii/. The notion of “minimal contrast” or
“minimal difference” (Padgett 2003; Calabrese 2005; Campos-Astorkiza
2009; Nevins 2010) or “crowding” (Kaun 1995) has been proposed as the
principle governing contrast. This approach is correct in one direction: if
there is only one phonetic property that distinguishes between two
phonemes, then that property must be contrastive. However, the converse
does not hold: a feature may still be contrastive in a phoneme even if it is
not the only phonetic property that distinguishes that phoneme from any
other. Minimal contrast has been shown to be incorrect on both conceptual
and empirical grounds (Archangeli 1988; Dresher 2009, 2015, 2016); the
latter will become apparent when we look at Yowlumne Yokuts in the next
section. One of the merits of the hierarchical approach to contrast is that it
can operate smoothly even when minimal phonetic differences between
phonemes are lacking, as they often are.
Eastern Mongolian languages have round harmony, that, as in the
majority of Manchu-Tungusic languages, is limited to low vowels. A
typical example is Khalkha Mongolian (Svantesson 1985; Qinggertai 1982),
with the vowel inventory in (25). We assume that they have similar feature
hierarchies as most of the Manchu-Tungus languages (26). In these
languages, harmony triggers are non-high because only non-high vowels
are contrastive for [round], a limitation that follows from the fact that
[front] (as well as a height feature) is higher in the hierarchy than [round].
4.2 Vowel systems with [round] > [front]
It is interesting to compare the above languages with Yowlumne Yokuts
(Southwestern USA; Newman 1944), which has an underlying vowel
inventory whose basic configuration looks similar (minus the ATR
contrasts); but it is a completely different type of language. In Yokuts, BOTH
/u/ and /ɔ/ trigger height-bounded round harmony: /u/ rounds only /i/, and /
ɔ/ rounds only /a/ (27). Why can /u/ trigger harmony here, but not in
Manchu-Tungusic and Eastern Mongolian?

A simple solution is available in terms of the contrastive hierarchy: in

Yowlumne, [round] is ranked over [front]. Hence, both /u/ and /ɔ/ are
[round], and [front] is not a contrastive feature in this language, as shown in
(28) (we omit the length contrast in the tree).
In support of this analysis, we note that /i/ in Yowlumne is phonologically
inert, and serves also as the epenthetic vowel. This is in sharp contrast to
the [front] /i/ in Manchu-Tungusic and many Mongolian languages.
Another language family in which [round] is typically ordered ahead of
[front] are the Yupik and Inuit languages that descend from Proto-Eskimo,
which is reconstructed to have vowels */i/, */a/, */u/, and a fourth vowel
assumed to be */ə/ (Fortescue et al. 1994). In most dialects this vowel has
merged with /i/. In some of these dialects merger is total, resulting in a three
vowel system; other dialects retain a trace of the distinction between */i/
and */ə/.
Original */i/ could cause palatalization of consonants, and some Inuit
dialects show palatalization (or traces of former palatalization) (Dorais
2003: 33). In parts of Baffin Island, for example, the word ‘foot’ is
pronounced [isiɣak], where i has caused a following original t to change to
s (29a). This assibilation is the most common manifestation of
palatalization in Inuit. In such dialects, it is traditional to distinguish
between “strong i”, which descends from */i/ and causes palatalization
(29a), and “weak i”, which descends from */ə/ and does not (29b). In some
dialects the two types of i exhibit other kinds of distinct behaviour as well.

Compton & Dresher (2011) observe the generalization in (30) about dialects
in which /i/ causes or once caused palatalization:
(30) Generalization about Inuit palatalization (Compton & Dresher
2011):
Inuit /i/ can cause palatalization (assibilation) of a consonant only in
dialects where there is evidence for a (former) contrast with a fourth
vowel; where there is no contrast between strong and weak i, /i/ does
not trigger palatalization.

This generalization follows if we assume that the feature hierarchy for Inuit
and Yupik is [low] > [round] > [front] as in (31). When the fourth vowel is
in the underlying inventory, /i/ has a contrastive [front] feature that enables
it to cause palatalization (31a). But in the absence of a fourth vowel, [front]
is not a contrastive feature (31b).
4.3 Vowel systems where ordering of [round]
and [front] is not crucial
Turkic languages have symmetrical inventories. They are typically analyzed
with three features: one height feature and two place features, as in (32)
(see Kabak 2011 for Turkish). Here, every feature specification is
contrastive in any order; the vowels completely fill the eight-cell vowel
space defined by three binary features. A possible ordering of the features
of Turkish is given in (33); however, the same contrastive specifications
would result from any ordering of these three features.117 We predict,
therefore, that all round vowels could potentially be triggers of round
harmony in such languages. This prediction is correct, though harmony
observes limitations that are not due to contrast, but to other factors.

In Turkish, harmony triggers can be high or low, but targets are typically
limited to high vowels (34). In Kachin Khakass (Korn 1969), both triggers
and targets of round harmony must be high (35), the opposite of the
Manchu-Tungus-Eastern Mongolian pattern. Because all vowels have
contrastive [round] and [front] features, however they are ordered, these
restrictions cannot be due to considerations of contrast, but to other factors.
This way of classifying phonological systems allows us to account for two
Manchu languages that are notable exceptions to the prevailing Manchu-
Tungusic pattern of round harmony. Spoken Manchu and Xibe are modern
Manchu languages in which [ATR] has been lost and /ə/ has become a (non-
low) vowel (Zhang 1996; Dresher & Zhang 2005). The vowel system of
Xibe, for example, is given in (36). The reclassification of /ə/ as a (non-low)
vowel necessitates a new contrastive feature to distinguish it from /u/. The
most natural modification is to extend the feature [round], already in the
system, to /u/.

Evidence that /u/ is in fact contrastively [round] in Xibe can be found in the
creation of new phonemes /y/ and /oe/. The latter derives from sequences of
/ɔ/ and /i/, where the [front] feature derives from /i/ and the [round] feature
from /ɔ/. Similarly, the new phoneme /y/ derives from sequences of /u/ and
/i/, showing that /u/ had acquired a [round] feature. More evidence that /u/
is contrastively [round] in Xibe comes from a new form of round harmony
that arose in Xibe, whereby /ə/ alternates with /u/ in suffixes: /u/ occurs if
the stem-final vowel is round, /ə/ occurs otherwise.
The participation of /u/ in triggering round harmony, rare in the
Manchu-Tungusic family, is accounted for by the extension of the
contrastive [round] specification to /u/. The phonological patterning of the
vowels in Xibe points to a contrastive hierarchy and branching tree as in
(37). This tree very closely resembles the Turkish feature hierarchy in (33).

4.4 Summary
To sum up, we can classify languages into types based on the contrastive
scopes of the vowel features [front] and [round] as in (38). Whether a
feature is contrastive on a given vowel depends on the feature hierarchy and
the size and structure of the phonological inventory.

(38) Typology of contrastive [front] and [round]

a. If [front] > [round], /i/ can cause palatalization, but /u/ may or may
not trigger round harmony.
b. If [round] > [front], /u/ may trigger round harmony, but /i/ may or
may not cause palatalization.
c. In languages where [round] and [front] are contrastive for all
vowels, both these features may be active in all vowels.
5 The diachrony of Algonquian vowel systems:
Contrast shift as a type of change
Understanding the role of contrastive hierarchies in phonological patterning
allows us to implement the program for diachronic phonology set out by
Jakobson (1972 [1931]), which we alluded to in Section 2. That is, when a
phonological change occurs in a language we need to look at what effect the
change has had on the system of contrasts.
For example, we have seen a number of differences between the vowel
system of Classical Manchu (Section 3.1) and that of Xibe (Section 4.3); on
the assumption that Xibe descends from a language whose vowel system is
essentially the same as that of Classical Manchu, we can assume that the
Xibe vowel system derives from the Classical Manchu one by a series of
phonological changes (Zhang 1996; Dresher & Zhang 2005). Some of the
changes are overt at a phonetic level, such as the loss of /ʊ/ and the raising
of /ə/. These phonetic changes are accompanied by a change in
phonological features, namely, the loss of [ATR] as a contrastive feature
and change of /ə/ from a [low] vowel to a (non-low) vowel. Less overt, but
just as consequential for the phonological patterning of Xibe, is the change
in contrastive status of /u/ from lacking a specification for [round] in
Classical Manchu to being [round] in Xibe.
We will designate as a contrast shift any change in the contrastive
feature hierarchy or in the contrastive status of a phoneme. A contrast shift
can involve a reordering of features, or a change of features. A contrast shift
may come about as a result of an overt phonetic change, such as the loss of
a phoneme or a change in its phonetic realization. Of particular interest are
“silent” changes like the one involving Xibe /u/, whereby a segment that
does not appear to change phonetically from one synchronic stage to the
next nevertheless takes on different contrastive features, with consequences
for its synchronic patterning.
We propose that contrast shift is an important type of diachronic
phonological change that can have far-reaching effects on the phonology of
a language.118 As should by now be evident, contrast shift can only be
understood with reference to a particular feature hierarchy.
5.1 The vowel system of Proto-Algonquian
In a survey of the historical development of Algonquian vowel systems,
Oxford (2012a, 2015) identifies persistent patterns in vowel changes. In an
attempt to make sense of these patterns, Oxford posits the feature hierarchy
in (39) for Proto-Algonquian (the length contrast is omitted for ease of
exposition).

The hierarchy in (39) is motivated by feature activity that can be recovered

as having been present in Proto-Algonquian. Thus, */o/ triggers rounding,
an indication that it has an active, hence contrastive, [round] feature.
Similarly, */i/ triggers palatalization, indicating a contrastive feature we call
[front]. Patterns of partial neutralization relate */ɛ/ and */i/, suggesting that
they are contrastive sisters by (13). Finally, */a/ does not trigger any
processes, consistent with its being assigned no positive (marked)
contrastive features. This evidence is summarized in (40).119

(40) Proto-Algonquian feature activity

a. */o/ is [round]: triggers rounding
 b. */i/ is [front]: triggers palatalization
c. */ i, ɛ/ are sisters: partial neutralization
d. */a/ has no marked contrastive features: is never a trigger

5.2 The Central Algonquian languages and

Blackfoot
The Proto-Algonquian vowel feature hierarchy continues unchanged in the
Central Algonquian languages and in Blackfoot. It accounts for two
recurring patterns: (a) palatalization always includes */i/ as a trigger; and
(b) */ɛ/ regularly merges with */i/. Examples of these processes are listed in
(41). The patterns in (41a) support the view that palatalization is triggered
by a contrastive [front] feature, and favours vowels that are (non-low); the
mergers in (41b) are consistent with the idea (13) that mergers tend to
involve terminal nodes in the feature tree.

(41) Central Algonquian and Blackfoot feature activity

a. Palatalization always includes */i/ as a trigger
i. Proto-Algonquian */ t, θ/-palatalization is triggered by */ i, iː/;
ii. Innu */k/-palatalization is triggered by */ i, iː, ɛː/;
iii. Betsiamites Innu /t/-palatalization is triggered by /iː/;
iv. Blackfoot */k/-assibilation is triggered by PA */i, iː/;
v. Blackfoot /t/-assibilation is triggered by Blackfoot / i, iː/.
b. */ɛ/ regularly merges with */i/
i. Partial or complete mergers of short */ɛ/ > /i/ occur in Fox,
Shawnee, Miami-Illinois, Cree-Innu, Ojibwe, and Blackfoot;
ii. long */ɛː/ > /iː/ in Woods Cree, Northern Plains Cree, and
Blackfoot.

5.3 The Eastern and Western Algonquian

languages
On the eastern and western edges of the Algonquian area, developments
diverge from the predictions of the Proto-Algonquian hierarchy: in
particular, the high vowels, derived from Proto-Algonquian */o/ and */i/,
begin to pattern together. In the east, Proto-Eastern Algonquian lost the
length contrast only in the high vowels (i.e., the reflexes of */o/, */i/), and in
the west, Proto-Arapaho-Atsina and Pre-Cheyenne merged */o, o:/ with */i,
i:/.
Under the hierarchy inherited from Proto-Algonquian, however, [high]
is not a contrastive feature, and the old height feature, [low], is ordered at
the bottom of the vowel feature hierarchy. The result is that the high vowels
derived from */o/ and */i/ are not a natural class. If the hierarchy constrains
patterning, then a new height contrast with the feature [high] must have
come to outrank the place contrasts. That is, the Proto-Algonquian feature
[low] is reinterpreted as [high] and moves to the top of the hierarchy,
creating the new hierarchy and contrastive feature tree in (42).

Subsequent developments in the eastern and western daughter languages

follow the predictions of the new hierarchy. The patterns consistently differ
from those of Central Algonquian: (a) palatalization in these languages is
triggered by */ɛ/ but excludes */i/; and (b) */ɛ/ merges with or shifts to */a/
(not */i/). Instances of these processes are listed in (43).

(43) Eastern and Western Algonquian feature activity

a. Palatalization is triggered by */ɛ/ but excludes */i/
i. Massachusett */k/-palatalization is triggered by Proto-Eastern
Algonquian */ɛː/ but not */iː/;
ii. Cheyenne “yodation”, where */k/ > /kj/, is triggered by */ɛ(ː)/
only.
 b. */ɛ/ merges with or shifts to */a/
i. Partial or complete mergers of PA short */ɛ/ and */a/ occur in
Abenaki, Mahican, Mi’kmaq, and Maliseet-Passamaquoddy;
ii. Proto-Eastern Algonquian long */ɛː/ shifts to /aː/ in
Massachusett and merges with */a/ in Western Abenaki;
iii. long and short */ɛ(ː)/ shift to /a(ː)/ in Cheyenne;
iv. vowel harmony involves */ɛ(ː)/ and */a(ː)/ in Arapaho.

Again, these patterns support the view that palatalization is triggered by a

contrastive [front] feature: only /ɛ/ is contrastively [front] in these
languages.120 The mergers in (43b) follow from the sisterhood of */ɛ/ and
*/a/ under the new hierarchy. A single contrast shift thus accounts for the
patterning of a large number of phonological changes across the
Algonquian family.

6 Areal isoglosses: Borrowing contrast shifts in

the Ob-Ugric Mansi and Khanty languages
The Algonquian languages have relatively simple vowel systems, and the
types of phonological activity we observed follow from the contrastive trees
in a rather straightforward manner. To see how alternations work in the
context of more complex and asymmetric feature trees, we need to look at
languages with larger vowel systems. Harvey (2012) shows that the
principles of contrast shift can be used to describe the sound changes which
have occurred over time in the vowel systems of the Ob-Ugric languages,
from the reconstructed Proto-Ob-Ugric up until modern times, starting
approximately 3,400 years ago when Hungarian split from Ob-Ugric.
Moreover, he shows that contrastive shifts in the Ob-Ugric Mansi and
Khanty languages show clear isoglosses and are borrowed between
languages.
The Ob-Ugric languages are found in central Russia, to the east of the
Ural mountains along the Ob river system. The two branches of Ob-Ugric
are the Mansi languages, in the southwest, and the Khanty languages, to the
east and north. The Ob-Ugric languages inherited a complex vowel system:
Proto-Ob-Ugric has been reconstructed to have nineteen vowel phonemes
(Harvey 2012, based on Sammallahti 1988). Also characteristic of Ob-
Ugric was a pervasive front-back vowel harmony that affected all vowels;
we assume that the relevant feature is [front].

6.1 Proto-Mansi
We will focus here on Mansi. Starting from the Proto-Mansi first-syllable
vowel system reconstructed by Steinitz (1955), and taking into account the
phonological patterning attributed to that period, Harvey (2012) posits the
Proto-Mansi contrastive hierarchy in (44).
A major type of phonological activity that provides evidence for this
hierarchy is front vowel harmony (45a), which we suppose to be governed
by the feature [front]. The Ob-Ugric languages have no neutral vowels,
therefore all vowels must have a contrastive value for this feature. Proto-
Mansi also had a system of productive ablaut-like root-vowel alternations
(Honti 1988a:149, 1988b:174), where a certain set of suffixes causes roots
with long vowels to shorten, as in the Western Mansi examples in (45b).

(45) Feature activity reconstructed for Proto-Mansi

a. Front vowel harmony
Suffix vowels harmonize with root vowels in the feature
[front]; thus, the vowel in the 1st person future suffix in
Southern Mansi is front or back depending on the root vowel:
e.g., jām-ăm ‘I will go’ ~ wēr-ĕm ‘I will make’.
b. Root-vowel alternations
A [long] vowel in a monosyllabic root becomes (non-long)
when it occurs with specific lexically-defined inflectional or
derivational suffixes, as well as appearing in certain compound
environments: e.g., Western Mansi tyoels ‘I will sit down’ ~
taɫt- ‘put into (a boat)’, where WM /yoe/ ~ /a/ derives from
Proto-Mansi */ǣ/ ~ */æ̆/; wɤ̄ɣm ‘I see’ > wăj ‘he/she sees’.

Both front vowel harmony and root-vowel alternations have been

reconstructed for the proto-languages (Honti 1988a:149, 1988b:174). Any
contrastive hierarchy for Proto-Mansi must account for both of these
processes. Moreover, changes to the hierarchies leading from Proto-Ob-
Ugric to the modern languages must remain consistent with them in
languages where they remain productive. That is, the features active in
harmony and vowel alternation must remain contrastive.
The examples in (45) also illustrate the importance of markedness in the
operation of these alternations. In front vowel harmony (45a), the suffix -
ăm changes to -ĕm. Simply changing ă to be [front] yields the features
(non-long), [front], (non-high), (non-round), a combination that does not
exist in (44). In the branch of the tree under (non-long), [front], and (non-
high), there is no contrastive (non-round); rather, we must choose between
non-low */ĕ/ and [low] */æ̆/. The correct outcome is obtained by choosing
the unmarked branch, */ĕ/. In (45b), the alternation */ǣ/ ~ */æ̆/ is
straightforward (as is a similar alternation */ī/ ~ */ĭ/), but */ɤ̄/ ~ */ă/ again
shows the effects of choosing the unmarked value of an unspecified feature;
in this case, we choose (non-round) */ă/ rather than [round] */ŏ/.
One might question the inclusion of [long] as a feature in the hierarchy
in (44). A currently widespread view represents the difference between long
and short vowels in structural terms, rather than as a feature: a short vowel
associates to a single timing unit, and a long vowel associates to two such
units (see Odden 2011 for discussion). However, the long/short contrast
interacts with other features, and therefore has to be represented somewhere
in the contrastive tree. 121 Moreover, Oxford (2012b) shows how a length
contrast can be easily reinterpreted as a tense ~ lax contrast, which is often
represented in featural terms, depending on where in the feature hierarchy
the relevant contrast is located. Therefore, including [long] in a contrastive
feature hierarchy does not preclude representing length differences in
structural terms.

6.2 Western Mansi

Early Western Mansi (~600 ybp) has been reconstructed to have thirteen
vowels (Steinitz 1955:154; Sammallahti 1988:504; Honti 1998: 330).
Harvey (2012) details the changes by which the Proto-Mansi vowel system
in (44) evolved into that of Early Western Mansi. In addition to some
mergers, at the Proto-Eastern-Western Mansi stage a new feature [contour]
becomes contrastive for two vowels. Then in Early Western Mansi, [round]
is promoted one step above [high]. These changes cause [low] to be non-
contrastive, and yield the hierarchy in (46), where [contour] is realized as a
diphthong (as with Proto-Mansi, the hierarchies are for vowels in initial
syllables, which exhibit the full range of contrasts).
At some point a contrastive shift occurred whereby [front] dropped to the
bottom of the hierarchy (47). As a result, five vowels, over one third of the
inventory, no longer contrast for [front]. This change had important
consequences for phonological activity in Western Mansi. The first
consequence is that front harmony is lost. Ob-Ugric front/ back harmony
does not have neutral vowels: either all vowels participate in harmony, or
none do. The loss of the front/ back contrast in a significant portion of
vowels thus caused the demise of harmony.
Second, the shift rearranges the pairs of vowels that participate in root-
vowel alternations. Previously, */ī/ could alternate with */ĭ/ by changing
[long] to (non-long). Now, however, there are two short counterparts of */ī/
with features (non-round) and [high]; of these, */ ̆ /, not */ĭ/, is unmarked.
One would predict that root-vowel alternation would be lost as a productive
process when a pair like Southern Mansi ʎīχ ‘wedge’~ ʎĭχt ‘wedges’ is no
longer derivable by the phonology. Loss of productivity indeed occurs in
the modern language.
A third consequence has to do with sub-phonemic drift. This can be
illustrated by the diachronic changes that vowels with contrastive [contour]
underwent. The (non-contour) vowel */ɒ̄/ in (46b) developed from an
earlier */ā/, and the [contour] vowel */ɔɒ/ developed from earlier */ǣ/. Both
vowels underwent rounding, an ENHANCEMENT of their contrastive features
(Stevens, Keyser, & Kawasaki 1986; Hall 2011). The addition of phonetic
rounding may have contributed to the promotion of [round] as a contrastive
feature, with the consequence that [low] is no longer contrastive. Lacking
[low] the vowels are free to raise to */ō/ and */oɔ/, respectively. When
[front] is demoted as in (47), it no longer constrains the vowels’ contrastive
space. The result is that */oɔ/ is able to front to /øoe/ (e.g., Modern Western
Mansi /øoemp/ ‘dog’), while */ō/ is also fronted, though not as far (/ō̟tər/
‘prince’).

6.3 Northern Mansi

Northern Mansi has reduced phonological complexity more than any other
Mansi dialect group, to the extent that front harmony and productive root-
vowel alternations have been completely lost. As in the Western dialect, all
vowels in Early Northern Mansi were contrastive for [front], and as in that
dialect, [front] was demoted over time. Starting from the Proto-Mansi
vowel system in (44), Harvey (2012) posits that [long] dropped to the
bottom of the hierarchy, and [round] was demoted below [low], yielding the
tree in (48).
At a stage just prior to the modern Northern Mansi dialects, [front] was
demoted as in (49).

When the front feature is dropped to the lowest rank, about half of the
vowels lose their contrastive [front] feature. In the next stage the three
remaining [front] vowels are merged to their back counterparts: */ǣ/ > */ɤ̄/,
*/ĭ/ > */ ̆ /, and */ȳ/ > */ū/. Once complete, these mergers leave no vowels
with a contrastive [front] feature at all. As expected, front harmony is no
longer viable, and has disappeared from Northern Mansi.
We also expect that root-vowel alternation would become untenable.
For instance, in Proto-Mansi, /ū/ alternated with /ŭ/. After */ȳ/ has merged
with */ū/, there is no way for a speaker of the modern language to tell
which /ū/ should alternate and which should not. As predicted, vowel
alternation has almost completely vanished in Northern Mansi.
Although the evolution of the vowel systems of Western and Northern
Mansi differ in their details, in both the feature [front] was demoted, and in
both front harmony and root-vowel alternations were adversely affected.
Interestingly, the dropping of [front] has also produced two very different
results. In Western Mansi, front dropping has caused some back vowels to
become more front; in Northern Mansi, the loss of the same contrast has
caused some front vowels to merge with their back counterparts.

6.4 Contrastive isoglosses

The dropping of [front] occurred in three of the four Mansi languages, and
all three have lost front harmony. However, [front] dropping did not occur
in the early history of Mansi. It can be shown that the shift occurred later in
the daughter languages, as illustrated in (50), where X indicates when
[front] dropping occurred.
If [front] dropping is not a genetic inheritance common to the non-Southern
Mansi languages, could it have been spread by areal diffusion? That is, can
contrast shift show areal patterning, like other elements of linguistic
systems? To investigate this question, Harvey (2012) plotted a number of
contrast shifts on a map, and the results are shown in Figure 8.1. It is clear
that the contrast shifts have occurred in a way that is not at all random.
Figure 8.1 shows the Ob-Ugric language area, in central Russia to the
east of the Ural mountains along the Ob river system. A key to the dialect
groupings and language name abbreviations is given below the map. Mansi
languages (M) are in the southwest, and the Khanty languages (K) are east
and north. The dashed line labelled ft dropped shows all the languages
which had the [front] dropping contrast shift.
 Figure 8.1: Ob-Ugric isoglosses of feature contrasts and contrast shifts
(Harvey 2012)

Key to dialect groups and abbreviations

KN: Northern Khanty MN: Northern Mansi
K Obd: Obdorsk
K N-N: Kazym MW: Western Mansi
KE: Eastern Khanty
K Surg: Surgut ME: Eastern (Konda) Mansi
VVj: Vach-Vasjugan
KS: Southern (Irtysh) Khanty MS: Southern (Tavda) Mansi

It appears that the innovative dialect from which [front] dropping

radiated is Northern Mansi. Northern, Western, and Eastern Mansi all
participate in the shift. Interestingly, two of the Khanty languages, Kazym
and Obdorsk Khanty, also had a phase where [front] dropped. Those
languages that are geographically and culturally farther away from the
likely innovation centre have not borrowed the shift. The arrows indicate
the Ob river and its tributaries, which are the main routes for cultural
contact and communication.
We conclude that there a pattern to these contrastive changes: they
follow routes of cultural contact. Contrast shifts show clear isoglosses and
can be borrowed between languages. The contrastive analysis of the Ob-
Ugric languages presented here is also consistent with earlier dialect studies
(Honti 1998; Steinitz 1955), and matches earlier observations about which
dialects are conservative or innovative.

7 Conclusions
The approach to phonological typology we have sketched here is based on a
fundamental distinction between a phonetic and phonological analysis of
the sound systems of languages. This view builds on approaches to
phonology pioneered by Sapir and the Prague School (Jakobson and
Trubetzkoy), instantiated within a generative grammar. More specifically, it
views phonemes as being composed of contrastive features that are
themselves organized into language-particular hierarchies. Because of the
hypothesized connection between contrast and activity, we expect
languages with similar hierarchies and inventories to exhibit similar
patterns.
In some of the language families we have surveyed here, feature
hierarchies appear to be relatively stable, as exemplified by Manchu-
Tungusic, Eastern Mongolian, Yupik-Inuit, and branches of Algonquin.
Contrast shifts can occur, however, for various reasons, and these can result
in dramatic differences in patterning, as shown by the modern Manchu
languages, Eastern and Western Algonquin as compared with Central, and
extensive changes in Ob-Ugric vowel systems viewed over a relatively long
period of time. Finally, Ob-Ugric shows that elements of feature hierarchies
can spread and be borrowed, like other aspects of linguistic structure.
We have seen that, like Sapir’s languages C and D, languages with
similar contrastive structures may show varying phonetic realizations. For
example, the breakdown of the front-back contrast had different phonetic
results in Western and Northern Mansi: in the former it resulted in some
back vowels fronting, and in the latter a series of vowels that used to be
front retracted and merged with back vowels. What the two dialects have in
common is the dropping and subsequent loss of [front] as a contrastive
feature; thus, it no longer constrained the phonetic ranges of the vowels. In
Algonquian, the various palatalizations and mergers show phonetic
differences, and the phonetic descriptions of the vowels vary from dialect to
dialect. But dialects sharing the same contrastive hierarchy show similar
patterns at that level.
We hope to have demonstrated that contrastive feature hierarchies
provide an interesting and fruitful level of representation for typological
research in phonology.

Acknowledgements: This article is a revised and expanded version of

Dresher, Harvey, & Oxford (2014). For comments we are grateful to Larry
Hyman, and to audiences at the CRC Summer Phonetics/phonology
Workshop, University of Toronto, July 2012; NELS 43, CUNY Graduate
Center, October 2012; the GLOW 36 Workshop on Diachronic Workings in
Phonological Patterns, Lund University, April 2013; the annual meeting of
the CLA, Victoria, B.C., June 2013; and the Workshop on Phonological
Typology, Somerville College, University of Oxford, August 2013. We
would also like to thank members of the project on Markedness and the
Contrastive Hierarchy in Phonology at the University of Toronto (Dresher
& Rice 2007): http://homes.chass.utoronto.ca/~contrast/. This research was
supported in part by grants 410-2003-0913 and 410-08-2645 from the
Social Sciences and Humanities Research Council of Canada.

References
Archangeli, Diana. 1988. Aspects of underspecification theory. Phonology 5. 183–
207.
Barrie, Mike. 2003. Contrast in Cantonese vowels. Toronto Working Papers in
Linguistics 20. 1–19.
Bhat, D. N. S. 1978. A general study of palatalization. In Joseph H. Greenberg,
Charles A. Ferguson, & Edith A. Moravcsik (eds.), Universals of human
language, vol. 2: Phonology, 47–92. Stanford: Stanford University Press.
Calabrese, Andrea. 2005. Markedness and economy in a derivational model of
phonology. Berlin: Mouton de Gruyter.
Campos-Astorkiza, Judit Rebeka. 2009. The role and representation of minimal
contrast and the phonology – phonetics interaction. München: Lincom Europa.
Chomsky, Noam & Morris Halle. 1968. The sound pattern of English. New York:
Harper & Row.
Compton, Richard & B. Elan Dresher. 2011. Palatalization and “strong i” across
Inuit dialects. Canadian Journal of Linguistics/Revue canadienne de
linguistique 56. 203–228.
Dorais, Louis-Jacques. 2003. Inuit uqausiqatigiit: Inuit languages and dialects
(second, revised edition). Iqaluit: Nunavut Arctic College.
Dresher, B. Elan. 1998. On contrast and redundancy. Paper presented at the
annual meeting of the Canadian Linguistic Association, May, Ottawa. Ms.,
University of Toronto.
Dresher, B. Elan. 2003. Contrast and asymmetries in inventories. In Anna-Maria di
Sciullo (ed.), Asymmetry in grammar, vol. 2: Morphology, phonology,
acquisition, 239–257. Amsterdam: John Benjamins.
Dresher, B. Elan. 2009. The contrastive hierarchy in phonology. Cambridge:
Cambridge University Press.
Dresher, B. Elan. 2014. The arch not the stones: Universal feature theory without
universal features. Nordlyd 41(2). 165–181, special issue on Features, ed. by
Martin Krämer, Sandra Ronai, & Peter Svenonius. University of Tromsø – The
Arctic University of Norway.
Dresher, B. Elan. 2015. The motivation for contrastive feature hierarchies in
phonology. Linguistic Variation 15. 1–40.
Dresher, B. Elan. 2016. Contrast in phonology 1867–1967: History and
development. Annual Review of Linguistics 2. 53–73.
Dresher, B. Elan. 2017. Contrastive feature hierarchies in Old English diachronic
phonology. Transactions of the Philological Society, doi:10.1111/1467-
968X.12105.
Dresher, B. Elan, Christopher Harvey, & Will Oxford. 2014. Contrast shift as a type
of diachronic change. In Hsin-Lun Huang, Ethan Poole, & Amanda Rysling
(eds.), NELS 43: Proceedings of the Forty-Third Annual Meeting of the North
East Linguistic Society, The City University of New York, vol. 1, 103–116.
Amherst, MA: GLSA.
Dresher, B. Elan & Andrew Nevins. 2017. Conditions on iterative rounding
harmony in Oroqen. Transactions of the Philological Society 115. 365–394.
Dresher, B. Elan & Keren Rice. 2007. Markedness and the contrastive hierarchy in
phonology. http://homes.chass.utoronto.ca/~contrast/.
Dresher, B. Elan & Xi Zhang. 2005. Contrast and phonological activity in Manchu
vowel systems. Canadian Journal of Linguistics/Revue canadienne de
linguistique 50. 45–82.
Fortescue, Michael, Steven A. Jacobson, & Lawrence D. Kaplan. 1994.
Comparative Eskimo dictionary with Aleut cognates. Fairbanks: Alaska Native
Language Center.
Gardner, Matt Hunt. 2012. Beyond the phonological void: Contrast and the
Canadian Shift. Ms., Department of Linguistics, University of Toronto.
Hall, Daniel Currie. 2007. The role and representation of contrast in phonological
theory. Doctoral dissertation, University of Toronto.
Hall, Daniel Currie. 2011. Phonological contrast and its phonetic enhancement:
Dispersedness without dispersion. Phonology 28. 1–54.
Harvey, Christopher. 2012. Contrastive shift in Ob-Ugric vowel systems. Ms.,
University of Toronto.
Honti, László. 1988a. Die ob-ugrischen Sprachen I: Die wogulische Sprache. In
Sinor (ed.) 1988, 147–171.
Honti, László. 1988b. Die ob-ugrischen Sprachen II: Die ostjakische Sprache. In
Sinor (ed.) 1988, 172–196.
Honti, László. 1998. Ob Ugrian. In Daniel Abondolo (ed.), The Uralic languages,
327–357. London: Routledge.
Hyman, Larry M. 2007. Where’s phonology in typology? Linguistic Typology 11.
265–271.
Jakobson, Roman. 1972 [1931]. Principles of historical phonology. In Allan R.
Keiler (ed.), A reader in historical and comparative linguistics, 121–138. New
York: Holt, Rinehart and Winston. Translation of Prinzipien der historischen
Phonologie. Travaux du cercle linguistique de Prague 4. 247–267.
Copenhagen, 1931.
Jakobson, Roman, C. Gunnar M. Fant, & Morris Halle. 1952. Preliminaries to
speech analysis. MIT Acoustics Laboratory, Technical Report, No. 13.
Reissued by MIT Press, Cambridge, Mass., 11th printing, 1976.
Jakobson, Roman & Morris Halle. 1956. Fundamentals of language. The Hague:
Mouton.
Kabak, Bariş. 2011. Turkish vowel harmony. In van Oostendorp et al. (eds.) 2011.
Kaun, Abigail Rhoades. 1995. The typology of rounding harmony: An Optimality
Theoretic approach. Doctoral dissertation, University of California, Los
Angeles.
Kiparsky, Paul. 1965. Phonological change. Doctoral dissertation, MIT.
Ko, Seongyeon. 2010. A contrastivist view on the evolution of the Korean vowel
system. In Hiroki Maezawa & Azusa Yokogoshi (eds.), MITWPL 61:
Proceedings of the Sixth Workshop on Altaic Formal Linguistics, 181–196.
Ko, Seongyeon. 2011. Vowel contrast and vowel harmony shift in the Mongolic
languages. Language Research 47. 23–43.
Ko, Seongyeon. 2012. Tongue root harmony and vowel contrast in Northeast
Asian languages. Doctoral dissertation, Cornell University.
Kochetov, Alexei. 2011. Palatalization. In van Oostendorp et al. (eds.) 2011.
Korn, David. 1969. Types of labial vowel harmony in the Turkic languages.
Anthropological Linguistics 11. 98–106.
Li, Bing. 1996. Tungusic vowel harmony. The Hague: Holland Academic Graphics.
Li, Shulan & Qian Zhong. 1986. Xiboyu jianzhi [A brief introduction to the Xibe
language]. Beijing: Minzu Chubanshe.
Mackenzie, Sara. 2011. Contrast and the evaluation of similarity: Evidence from
consonant harmony. Lingua 121. 1401–1423.
Mackenzie, Sara. 2013. Laryngeal co-occurrence restrictions in Aymara:
Contrastive representations and constraint interaction. Phonology 30. 297–
345.
McCawley, James D. 1967. Edward Sapir’s “phonologic representation”.
International Journal of American Linguistics 33. 106–111.
Mielke, Jeff. 2008. The emergence of distinctive features. Oxford: Oxford
University Press.
Moulton, William G. 1960. The short vowel systems of Northern Switzerland: A
study in structural dialectology. Word 16. 155–182.
Nevins, Andrew. 2010. Locality in vowel harmony. Cambridge, MA: MIT Press.
Newman, Stanley. 1944. Yokuts language of California (VFPA 2). New York: The
Viking Fund Publications in Anthropology.
Odden, David. 2011. The representation of vowel length. In van Oostendorp et al.
(eds.) 2011.
Oostendorp, Marc van, Colin J. Ewen, Elizabeth Hume, & Keren Rice (eds.). 2011.
The Blackwell companion to phonology. Oxford: Blackwell.
Oxford, Will. 2012a. “Contrast shift” in the Algonquian languages. In A. McKillen &
J. Loughren (eds.), Proceedings from the Montreal-Ottawa-Toronto (MOT)
Phonology Workshop 2011: Phonology in the 21st Century: In Honour of
Glyne Piggott. McGill Working Papers in Linguistics 22 (1).
Oxford, Will. 2012b. On the contrastive status of vowel length. Presented at the
MOT Phonology Workshop, University of Toronto, March 2012. http://home.cc.
umanitoba.ca/~oxfordwr/papers/Oxford_2011_MOT.pdf.
Oxford, Will. 2015. Patterns of contrast in phonological change: Evidence from
Algonquian vowel systems. Language 91. 308–357.
Padgett, Jaye. 2003. Contrast and post-velar fronting in Russian. Natural
Language & Linguistic Theory 21. 39–87.
Purnell, Thomas & Eric Raimy. 2013. Contrastive features in phonetic
implementation: The English vowel system. Presented at the CUNY
Phonology Forum Conference On The Feature, January 2013.
Purnell, Thomas & Eric Raimy. 2015. Distinctive features, levels of representation,
and historical phonology. In Patrick Honeybone & Joseph Salmons (eds.), The
handbook of historical phonology, 522–544. Oxford: Oxford University Press.
Qinggertai (Chingeltei). 1982. Guanyu yuanyin hexielü [On the vowel harmony
rule]. Zhongguo Yuyanxue Bao 1. 200–220.
Rice, Keren. 2003. Featural markedness in phonology: Variation. In Lisa Cheng &
Rint Sybesma (eds.), The second Glot International state-of-the-article book:
The latest in linguistics, 387–427. Berlin: Mouton de Gruyter.
Rice, Keren. 2007. Markedness in phonology. In Paul de Lacy (ed.), The
Cambridge handbook of phonology, 79–97. Cambridge: Cambridge University
Press.
Roeder, Rebecca & Matt Hunt Gardner. 2013. The phonology of the Canadian
Shift revisited: Thunder Bay and Cape Breton. University of Pennsylvania
Working Papers in Linguistics: Selected Papers from NWAV 41, 19 (2). 161–
170.
Rohany Rahbar, Elham. 2008. A historical study of the Persian vowel system.
Kansas Working Papers in Linguistics 30. 233–245.
Sammallahti, Pekka. 1988. Historical phonology of the Uralic languages. In Sinor
(ed.) 1988, 478–554.
Samuels, Bridget D. 2011. Phonological architecture: A biolinguistic perspective.
Oxford: Oxford University Press.
Sapir, Edward. 1925. Sound patterns in language. Language 1. 37–51. Reprinted
in Martin Joos (ed.), Readings in linguistics I, 19–25. Chicago, IL: University of
Chicago Press, 1957.
Saussure, Ferdinand de. 1972 [1916]. Cours de linguistique générale. Publié par
Charles Bally et Albert Sechehaye; avec la collaboration de Albert Riedlinger.
Éd. critique préparée par Tullio de Mauro. Paris: Payot.
Sinor, Denis (ed.). 1988. Handbuch der Orientalistik: Handbook of Uralic studies,
vol. 1: The Uralic languages. Leiden: E. J. Brill.
Steinitz, Wolfgang. 1955. Geschichte des wogulischen Vokalismus. Berlin:
Akademie-Verlag.
Stevens, Kenneth N., Samuel Jay Keyser, & Haruko Kawasaki. 1986. Toward a
phonetic and phonological theory of redundant features. In Joseph S. Perkell
& Dennis H. Klatt (eds.), Symposium on invariance and variability of speech
processes, 432–469. Hillsdale, NJ: Lawrence Erlbaum.
Svantesson, Jan-Olaf. 1985. Vowel harmony shift in Mongolian. Lingua 67. 283–
327.
Vajda, Edward. 2001. Test materials dated August 17, 2001. Posted at http://pando
ra.cii.wwu.edu/vajda/ling201/test2materials/Phonology3.htm.
Walker, Rachel. 2001. Round licensing and bisyllabic triggers in Altaic. Natural
Language & Linguistic Theory 19. 827–878.
Walker, Rachel. 2014. Nonlocal trigger-target relations. Linguistic Inquiry 45. 501–
523.
Zhang, Xi. 1996. Vowel systems of the Manchu-Tungus languages of China.
Doctoral dissertation, University of Toronto.
Zhang, Xi & B. Elan Dresher. 1996. Labial harmony in written Manchu. Saksaha: A
Review of Manchu Studies 1. 13–24.
Ellen Broselow
Laryngeal contrasts in second
language phonology

Abstract: This chapter investigates the acquisition of obstruent laryngeal

contrasts in foreign language acquisition. The goal is to determine the
alignment between the putative universal markedness relationships
established through cross-linguistic investigation and the patterns found in
second language phonology, particularly those patterns that appear to be
independent of both the native and the foreign language systems.
Cross-linguistic research has revealed that obstruent laryngeal contrasts
are more common in nonfinal than in final positions; that when contrast is
limited in final position, voiceless obstruents are the preferred segment
type; and that the preferred repair for underlying voiced obstruents in final
position is devoicing of the obstruent rather than any of the logically
possible alternatives such as post-obstruent vowel insertion. A survey of
studies on the acquisition of foreign language laryngeal contrasts supports a
hierarchy of difficulty in the acquisition of foreign structures that is
consistent with the principles established by typological investigation:
learners from a wide range of native language backgrounds show earlier
success in mastering final voiceless than final voiced obstruents, even when
the native language has neither and the foreign language has both, while the
opposite order of acquisition is not attested. Furthermore, devoicing is
frequently found in second language phonology even in the absence of a
phonological devoicing process in either the native or the foreign language.
However, second language learners do exhibit some tendencies that are
not predicted by established typological generalizations, such as effects of
word size and of manner and/or place of articulation on the likelihood of
obstruent devoicing. We consider possible explanations for these tendencies
as well as the question of whether these tendencies ever become
phonologized as categorical processes in established native language
grammars.
1 Introduction
Among the strongest candidates for typological universals in phonology are
three generalizations that pertain to laryngeal contrasts. The first
generalization refers to favored contrast position: a language that exhibits
laryngeal contrasts in final position (whether syllable-final, word-final, or
phrase-final) will exhibit laryngeal contrast in initial positions as well. The
second generalization concerns favored segment type: where a language
limits or suppresses laryngeal contrasts in final position, the obstruents most
likely to surface in this position are voiceless. And the third generalization
addresses favored repair: languages that ban voiced obstruents in final
position typically enforce this ban by devoicing underlyingly voiced
obstruents, despite the in-principle availability of alternative strategies such
as consonant deletion, vowel insertion, and nasalization (see Blevins 2004
and Steriade 2001/2008, among many others).
If these generalizations are true universals, we expect them to meet the
criteria defined by Blevins (2010), following Kiparsky (2008):

(1) Prerequisites for true phonological universals

a. Phonological universals should have no exceptions.
b. Phonological universals should constrain change.
c. Phonological universals should emerge spontaneously within
grammars (e.g., as in the final devoicing often associated with
children’s L1 English productions).
d. Learners will not construct grammars that violate universals.
e. Universals are part of every grammar.

Potentially fruitful test cases for criteria (1c) and (1d) involve new linguistic
systems, among them the patterns of speakers acquiring a novel language.
Typological markedness has frequently been invoked to explain the
emergence of patterns in second language (L2) phonology that appear to
have no basis in either the native or the foreign language grammars (e.g.,
Eckman 1977, 1984). A surprising number of L2 studies have reported that
for speakers of native languages that lack final laryngeal contrasts, or that
lack any final obstruents, the mastery of final voiceless obstruents precedes
the mastery of final voiced obstruents. This finding has served as a veritable
poster child for arguments that second language learning is guided by
universal principles, even in the absence of direct supporting evidence in
the input to the learner.
The goal of this chapter is to survey the literature on the second
language acquisition of laryngeal contrasts, in order to determine, first, the
extent to which L2 patterns align with typological generalizations, and
second, whether the second language data can shed light on the nature and
source of these typological generalizations. To begin, we distinguish two
opposing views (along a broad spectrum) concerning the nature of
typological asymmetries. On one view, typology reflects what Moreton
(2008) calls “channel bias”: factors based in articulation and perception
make certain structures less likely to survive in the transmission of language
across generations (Blevins 2004 and Ohala 1981, among many others), and
listeners’ imperfect perception of more fragile contrasts ultimately results in
phonologization of a system lacking these contrasts (Hyman 1976). The
numerous aerodynamic and acoustic factors that make voicing difficult to
maintain and to perceive in final positions (reviewed in Blevins 2004, 2006
and Myers 2012) make the typological generalizations concerning laryngeal
contrasts very strong candidates for this sort of explanation. However, some
L2 evidence has been argued to support the view that at least some
typological generalizations reflect what Moreton calls “analytic bias”,
defined as “cognitive biases which facilitate the learning of some
phonological patterns and inhibit that of others” (Moreton 2008: 84). On
this view, language learners simply will not entertain the hypothesis that the
system they are learning fails to conform to the relevant typological
generalization. In surveying the second language literature, we will consider
the fit of the second language data, particularly the finding that L2 learners
frequently master some L2 structures earlier than other equally novel
structures, with typological generalizations. We will consider explanations
of the L2 patterns ranging from the articulatory and perceptual difficulty of
particular structures (channel bias effects) to learning biases potentially
rooted in universal grammatical constraints (analytic bias effects).
Before proceeding, a caveat is in order regarding the scope of this
survey. First, in considering the acquisition of final obstruents, we will
consider only single obstruents in final position, since the introduction of
consonant clusters introduces additional factors that cloud the debate.
Second, the term “second language acquisition” casts a wide net, including
learners ranging from children to adults, with varying levels of proficiency
and exposure, and situations ranging from naturalistic learning to formal
instruction. Furthermore, the studies in the second language literature below
include a wide range of methodologies which makes comparison across
studies difficult. We will see, however, that certain patterns emerge across a
wide range of subject populations and methodologies.
Section 2 reviews the typological claims concerning the favored
positions for laryngeal contrasts in native language systems, as well as
favored segment types in different positions. In Section 3 we will see that
studies of speakers from a wide range of native languages show more
success in mastering the typologically more natural structures, and we will
consider possible explanations of individual cases. Section 4 focuses on the
question of whether the preferred repair strategy for those learners who fail
to successfully produce final voiced obstruents is devoicing of the
obstruent, as predicted by Steriade’s (2001/2008) proposal. Here we will
consider the interaction of devoicing with speaker-dependent factors such
as proficiency as well as linguistic factors such as word size and the manner
and place of articulation of the target final obstruents. We conclude by
discussing the implications of the second language data for theories of
typology.

2 Preferred position of contrast and preferred

segment type
The typological literature on laryngeal contrasts presents convincing
evidence for two generalizations: (i) laryngeal contrasts in nonfinal
positions are more common than such contrasts in final positions; and (ii)
when contrast is absent or actively suppressed in final position, voiceless
obstruents are the likely survivors. Of the 51 languages surveyed by
Keating et al. (1983), 18 displayed “at least some neutralization of voicing-
related contrasts among stops” in final position (Westbury & Keating 1986:
160).
However, the facts demand more fine-grained distinctions than simply
final vs. nonfinal positions. Contrasting word-final and word-internal
syllable-final positions, both Wetzels & Mascaró (2001) and Myers (2012)
argue that devoicing in syllable-final position implies devoicing in word-
final position, but not vice versa. Thus, Myers (2012) presents examples of
languages in which devoicing affects only word-final obstruents (e.g.,
Russian, Walloon, and Uyghur) as well as languages in which syllable-final
obstruents, both within and at the end of words, are voiceless (e.g.,
Takelma, Breton, and Malay). However, languages with devoicing only in
word-internal syllable-final position appear to be unattested. A further
distinction between word-final and utterance-final positions is made by
Westbury & Keating (1986), who argue that “some effects commonly
reported as ‘word’ effects are in fact constrained by pause – i.e. they are
utterance effects” (Westbury & Keating 1986: 161). Blevins (2006) and
Myers (2012), reviewing the factors that favor devoicing in prepausal
position, provide convincing arguments for a diachronic scenario in which
both word-final and subsequent syllable-final devoicing develop from the
generalization of utterance-final devoicing. Consistent with this proposal,
Myers & Padgett (2015) provide evidence that participants exposed in
artificial language learning experiments to utterance-final obstruent
devoicing extended devoicing to utterance-medial word-final coda
obstruents (though the converse did not hold: participants exposed to
utterance-final voicing did not extend devoicing to utterance-internal
syllable codas).
In addition to position in the syllable, word, and utterance, preceding
and following segmental context are crucially important factors for
laryngeal contrast. Thus, even heterosyllabic obstruents within a cluster
frequently assimilate in voicing (see, e.g., Lombardi 1995 for a set of
grammatical constraints meant to reflect the typological possibilities).
Steriade (1999) argues that the relevant factor determining the possibility of
voicing contrasts is not prosodic structure per se but rather the extent to
which specific contexts allow the realization of acoustic cues that signal the
contrast (for example, the possibility of release). Steriade proposes the
hierarchy below (adapted here from Gordon’s 2007 summary), in which the
possibility of contrast in one position on the hierarchy implies the
possibility of contrast in all positions to the left (examples of languages
along the hierarchy appear below each cutoff point):
 The question of preferred member of contrasting laryngeal sets also
demands finer-grained distinctions than those embodied in the common
assumption that voiceless obstruents are less marked than voiced
obstruents; the choice of preferred laryngeal specification is related both to
the position in which the obstruent appears and to the nature of the
laryngeal contrast in individual languages. The term “voicing contrast” as
used in both the typological literature and the second language phonology
literature often conflates two distinct types of contrast: voicing contrasts,
which oppose voiceless and voiced consonants (short lag VOT vs.
prevoiced), and aspiration contrasts, which oppose aspirated and
unaspirated consonants (long lag vs. short lag VOT). These differences
have typological consequences in terms of preferred segment type. In final
position, the favored member of a voiceless-voiced contrast is typically the
voiceless member, while in languages relying on a contrast between
voiceless aspirated and unaspirated pairs, aspirated stops may emerge as the
output of neutralization (Westbury & Keating 1986; Vaux & Samuels
2005). Cross-linguistic study reveals interactions between contrast type and
contrast position; Keating et al. (1983) argue that languages with a true
initial voicing contrast (such as Arabic, Dutch, and Japanese) generally
show the same contrast possibilities in initial and medial positions, while
languages with an initial aspiration contrast (such as Gaelic, Mandarin, and
Swedish) frequently show deaspiration and/or voicing of intervocalic
medial stops.
The phonological specifications that best characterize the different types
of laryngeal contrast are a matter of some debate, rooted in larger debates
concerning binary vs. unary features, the role and extent of of
underspecification, and the notion of universal markedness hierarchies
implying a single least marked member of a series. Voiceless unaspirated
stops are characterized as [-voice] by Wetzels & Mascaró (2001), who
argue for binary specification of voicing, but as lacking any specified
laryngeal features by Lombardi (1995), who argues for privative [voice]
and [aspiration], with voiceless unaspirated stops representing the
unspecified (and least marked) value. However, Vaux, & Samuels (2005)
argue that the universal output of laryngeal neutralization in stops is a
segment that is underspecified for laryngeal features, and that languages
vary in the surface realization of the underspecified stop, which may be
either aspirated or unaspirated. Arguments for full specification are
provided by Beckman et al. (2011), who claim that the two-way laryngeal
contrast in Swedish is best characterized as a contrast between fully
specified [spread glottis] and [voiced]). Iverson & Salmons (2011) present a
typology of final laryngeal neutralization that provides ten options defined
in terms of insertion or deletion of the features [voice], [spread], and
[constricted] – of which at least seven are, they argue, attested. The
possibility that languages might differ not only in the phonetic realization of
a contrast but also in the phonological specification of what might appear to
be similar contrasts across languages highlights the difficulty of deciding at
what level a typological generalization must hold in order to be considered
universal (a point made in detail in Hyman’s 2008 discussion of the
distinction between “descriptive universals” and “analytical universals”).
An additional complication comes from disagreement in the correct
analyses of individual languages – see, for example, the disagreement
between Blevins (2004, 2006, 2010) and Yu (2004), who argue for the
existence of productive final voicing, and Kiparky (2008), who argues that
closer inspection of the putative voicing languages does not support the
existence of final obstruent voicing as a true phonological process.
Despite disagreements of analysis, however, some cross-linguistic
generalizations about the preferred position of voicing contrasts and the
preferred output of contrast suppression have emerged as relatively
uncontroversial: final position is the most likely position to exhibit
impoverished contrast, and the presence of final voiced obstruents implies
the presence of voiceless obstruents in final position. These generalizations
therefore serve as a starting point for the investigation of the acquisition of
voicing contrasts in second language learning. As we will see below,
numerous studies have documented that L2 learners tend to master the
typologically less marked laryngeal structures more readily than the more
marked structures, even when both are equally novel for the learner.
3 Hierarchy of difficulty in L2 voicing contrasts
Not surprisingly, the second language literature contains no studies of L2
production that have manipulated and controlled every positional variable
that may affect the realization of obstruent laryngeal contrasts (position in
the word and the utterance; context in terms of word and sentence stress and
intonation; and preceding and following segmental contexts). Nonetheless,
studies using a variety of methods and a variety of contexts have converged
on the finding of a hierarchy of difficulty (Broselow & Kang 2013): L2
learners are more successful in producing novel voicing contrasts in
typologically less marked (i.e., nonfinal) positions, and are more successful
in producing final voiceless than voiced obstruents. We consider the results
on the acquisition of L2 final laryngeal contrasts in terms of the learners’
native language backgrounds: languages with no final obstruents; languages
with only voiceless final obstruents; languages with a final laryngeal
contrast; and languages that lack a voicing contrast even in nonfinal
positions.

3.1 L1 has no final obstruents

Perhaps the most well documented cases of laryngeal contrast acquisition
involve speakers of a native language with no final obstruents whose target
language contains both voiceless and voiced final obstruents. For these
speakers, both classes of final obstruent are equally novel, so asymmetry in
the mastery of one class over the other cannot, at least at first glance, be
ascribed to either the native or the target language. Yet numerous studies
have documented greater accuracy in the production of English final
voiceless than voiced obstruents by native speakers of Mandarin (Eckman
1981; Flege & Davidian 1984; Weinberger 1987; Wang 1995; Yavas 2009),
Tswana (Wissing & Zonneveld 1996), the Tibeto-Burman languages
Angami and Ao (Wiltshire 2006), and Japanese (Eckman 1981; Edge 1991;
Yavas 2009) – all languages that lack word-final (and, in many cases,
syllable-final) obstruents. Despite large differences across studies in the
rates of target-like productions, final voiceless obstruents still show
substantially higher accuracy than final voiced obstruents. In Wang’s (1995)
word production task, for example, 19% of final voiceless stops produced
by Mandarin speakers in English pseudowords were identified as target-
like, as opposed to only 2% of final voiced stops. For Wissing &
Zonneveld’s (1996) Tswana speakers, percentages of target-like final
obstruents were much higher overall, but the difference between voiceless
vs. voiced final obstruents was still large (74% vs. 52%). The earlier
mastery of final voiceless stops is of course in line with typological
generalizations, and the literature reveals no reports of learners who are
more successful in producing final voiced than voiceless obstruents.
All of the studies mentioned above involve English as the target
language. Unfortunately, a study of Mandarin speakers learning Swedish
(Abrahamsson 2003) focused only on changes in rates of epenthesis and
deletion over time; productions with voicing change were coded as correct,
and no report was made of individual error rates for final voiceless and
voiced obstruents. The paucity of studies addressing the acquisition of final-
contrast languages other than English is a lamentable gap in the second
language acquisition literature, and without such studies it is difficult to
distinguish English-specific effects from more general effects.

3.2 L1 has only voiceless final obstruents

Earlier mastery of final voiceless than voiced obstruents is expected among
speakers whose native language limits final obstruents to voiceless, and this
pattern is well attested, both for speakers of languages with productive
voicing alternations (German, Smith et al. 2009; Dutch, Simon 2009, 2010;
Polish, Flege & Davidian 1984; Catalan, Cebrian 2000) and for speakers of
languages that lack such alternations but nonetheless restrict final
obstruents to voiceless (Cantonese, Edge 1991, Peng & Ann 2004;
Taiwanese Mandarin, Wang 1995; Japanese, Edge 2004; and Thai, Hancin-
Bhatt 2000).
Greater accuracy in the production of final voiceless obstruents has
been found not only for learners of English, but also for Cantonese-
speaking learners of French (Cichocki et al. 1993) and German-speaking
learners of Swedish (Hammarberg 1990). While we cannot eliminate the
native language as the (possibly sole) source of this asymmetry, it is
important to note that final voicing is realized very differently in French and
Swedish than in English. In French and Swedish, closure voicing in final
stops is a major cue to the final voicing contrast (Blevins 2004; Helgason &
Ringen 2008), whereas in English, final obstruents are often at least
partially devoiced, and voicing during closure does not appear to be a
necessary cue for English speakers to identify a final stop as voiced
(Hillenbrand et al. 1984). Thus, these studies suggest that the asymmetric
mastery of final voiceless vs. voiced obstruents in second language
acquisition cannot be ascribed solely to facts about the way this contrast is
realized in English.

3.3 L1 has a final voicing contrast

The assumption that L2 learners take their native language system as their
starting point predicts that speakers of languages with only final voiceless
obstruents will show greater difficulty in producing L2 voiced than
voiceless final obstruents. Similarly, where both the L1 and the L2 employ
laryngeal contrasts in final position, we might reasonably expect learners to
be equally proficient in producing L2 voiceless and voiced final obstruents.
In fact, however, devoicing of English final obstruents has been noted
among native speakers of two languages that are reported to have a two-
way laryngeal contrast in final position, Farsi (Eckman 1984) and
Hungarian (Altenberg & Vago 1983). These findings are unexpected if the
native language grammar serves as the starting point for second language
acquisition, and are perhaps equally surprising if we approach the problem
as one of cross-language differences in phonetic implementation; since
Hungarian final stops are reported to have significant voicing during
closure (Gosy & Ringen 2009), transfer of the L1 articulatory routines
would seem to suggest that Hungarian speakers’ final stops in English
should sound, if anything, even more fully voiced than native speakers’
final English stops. We will return to possible explanations of this
phenomenon in Section 3.5.

3.4 L1 has no voice contrast

An ideal testing ground for a hierarchy of difficulty involving position of
laryngeal contrast is provided by learners whose native languages lack
laryngeal contrast not only in final position but in any position. For
speakers of such languages, L2 contrasts in final and nonfinal positions are
equally novel. While there do not seem to be any studies of this ideal
combination (L1 lacking laryngeal contrasts in all positions and L2
allowing both final and nonfinal laryngeal contrasts), Hansen (2004) studies
English learners of Vietnamese, a language in which she reports that
laryngeal contrast, even in initial position, is possible only for coronal stops
and for labiodental and velar fricatives. Hansen carried out a longitudinal
study of two Vietnamese speakers learning English, including three lengthy
interviews and word list reading tasks spaced over one year. While the
focus of her study was on the speakers’ production of English codas, she
does informally report that accuracy was consistently higher for onsets than
for codas, supporting a hierarchy of difficulty effect for position. Hansen’s
data also provide additional support for a hierarchy of difficulty for
voiceless vs. voiced final obstruents. Although in Vietnamese neither
voiced nor voiceless fricatives occur in final position, where the only
possible consonants are voiceless stops, nasals, and glides, she reports that
“voiceless consonants emerged before their voiceless counterparts [. . .] a
finding consistent for every voiceless-voiced pair” (Hansen 2004: 113).
This asymmetry held for both participants, in both the interview data and
the word list task.
A language that lacks a voicing contrast in any position is Korean.
Although Korean does employ a three-way laryngeal contrast among tense,
aspirated, and plain voiceless stops in nonfinal positions, voicing per se is
never contrastive; unaspirated stops are predictably voiced in intersonorant
positions, voiceless elsewhere. Major & Faudree (1996) investigated
Korean speakers’ productions of English obstruents in initial, intervocalic
medial, and final positions. In a word list reading task, Major & Faudree
found no positional effect for voiceless obstruents, 98% of which were
judged as target-like in all three positions. For voiced obstruents, however,
productions in final position were significantly less target-like (38%) than
those in initial or medial positions (98% or above for both). Since voiced
obstruents do not occur in either word-initial or word-final position in
Korean, this result is not readily attributable to the native language, but is
consistent with the typological generalization that voicing contrasts are
more marked in final position than in presonorant positions.
 Major & Faudree’s (1996) study was designed to investigate the claim
of Eckman (1977) that the presence of voicing contrast in final position
implies a contrast in medial position, which in turn implies initial contrast.
This claim conflicts with Steriade’s (1999) hierarchy, which identifies
intersonorant position as the most favored position for voicing contrasts
(with the implication that intersonorant contrast implies initial contrast).
The fact that the Korean speakers’ productions of voiceless stops were
judged equally target-like in medial and initial positions, despite the native
language process of intersonorant voicing, might be taken as support for
Steriade’s claim that intersonorant position is the most favored for voicing
contrasts, since this is the position where native language effects could
make voiceless obstruents difficult to produce. However, an alternative
explanation is that the Korean speakers identified the English stops with
their native language aspirated stops, which do not undergo intersonorant
voicing. This is consistent with the adaptation of voiceless stops in English
words borrowed into Korean, which are typically adapted as aspirated (e.g.,
[pho:kha] ‘poker’), even when the original stop is unaspirated in English
(Oh 1996; Kenstowicz 2005).

3.5 Possible explanations of the difficulty

hierarchies
The studies reviewed above provide evidence for hierarchies of difficulty
consistent with typological generalizations: L2 learners show evidence of
greater accuracy in producing voicing contrasts in nonfinal than in final
positions, and greater accuracy in producing final voiceless than voiced
obstruents. These asymmetries are particularly interesting when they appear
to be emergent – that is, when neither the native nor the target language
provides evidence for the asymmetry. There are essentially three categories
of explanations for learners’ relative lack of success in producing final
voiced obstruents, which can be summarized as follows:

(3) Possible explanations of failure to produce final voiced obstruents in

L2
a. Perception-based: Non-native speakers perceive voiced final
obstruents as voiceless.
 L2 speaker’s output target = voiceless;
L2 productions perceived by native speakers as voiceless.
b. Articulation-based: Non-native speakers produce non-target-like
voiced obstruents.
L2 speaker’s output target = voiced;
L2 productions perceived by native speakers as voiceless.
c. Grammatically-based: Non-native speakers’ (interlanguage)
grammars ban final voiced obstruents.
L2 output target = voiceless;
L2 productions perceived by native speakers as voiceless.

These accounts are not, of course, necessarily mutually exclusive. Blevins

(2006) reviews multiple factors that make final voiced obstruents difficult
both to produce and to perceive: laryngeal spreading or closing at phrase
boundaries, which interferes with the maintenance of voicing; phrase-final
lengthening, which may obscure durational cues to voice contrasts; and the
absence of audible release in final position. Learners’ difficulties in
producing target-like final voiced obstruents have been demonstrated in
various studies; e.g., Flege, McCutcheon, & Smith (1987) found that
Mandarin speakers’ voicing during the closure of English final voiced stops
was significantly shorter than that of native speakers, and Wissing &
Zonneveld (1996) found that Tswana speakers differed significantly from
native speakers both in voicing into closure and in the lengthening of
vowels before voiced obstruents, a major cue to voicing in final position in
English (Raphael 1972).
Difficulties in perception are also in evidence in the L2 literature,
though in some cases production has been shown to lag perception. For
example, Wissing & Zonneveld’s Tswana speakers correctly identified 70%
of English final voiced stops as voiced in a forced choice task, but produced
only 52% of English final voiced stops as target-like. For speakers of
Dutch, where final obstruents are devoiced, Broersma (2005) found
categorization accuracy for both voiced and voiceless English final
obstruents comparable to that of native speakers of English, though Dutch
speakers are not necessarily entirely successful in producing English-like
final voicing contrasts (Simon 2009, 2010). A considerable body of
literature on cross-language perception indicates that even where speakers
perform well on tests of their ability to perceive contrasts, they may be
using different cues than are used by native speakers; for example, Flege &
Wang (1989) found that removing burst cues from English final stop stimuli
resulted in a significant worsening of discrimination by Cantonese speakers
but not by native speakers of English. Thus, perception and production are
crucially intertwined: listeners need to both recognize and produce the cues
that the target language relies on to signal a contrast.
Approaches that locate L2 patterns in the developing L2 grammar have
long appealed to markedness, defined as typologically-based preferences
for less marked over more marked structures (e.g., Eckman 1977). In
versions of Optimality Theory that directly connect typology and
acquisition by encoding typologically-motivated constraints as part of every
grammar, such hierarchies of difficulty can be analyzed as the effect of
markedness constraints, the effects of which become visible only when the
target language provides learners with novel structures that violate these
constraints (Broselow, Chen, & Wang 1998; Hancin-Bhatt 2000; Eckman
2004; Peng & Ann 2004; Wiltshire 2006; Cardoso 2007, among others).
Thus, the existence of languages like Mandarin and Tswana motivates a
universal (but violable) constraint banning all final obstruents
(*FINALOBSTRUENT). The existence of languages like German and Catalan,
which restrict final obstruents to voiceless, motivates a universal (though
violable) constraint *FINALVOICEDOBSTRUENT, which will be ranked low in
languages like English but high in languages like German (see, e.g.,
Lombardi 1995) – and, by default, in Mandarin, on the assumption that
markedness constraints rank as high as is consistent with the data of the
target language (e.g., Smolensky 1996). On this view, Mandarin speakers’
more successful production of final voiceless than voiced obstruents can be
seen as reflecting an intermediate stage between the native and target
language grammars, in which the constraint banning final obstruents of any
type has been demoted, but the constraint banning final voiced obstruents
remains highly ranked.

(4) Proposed Grammars (Broselow et al. 1998)

a. Mandarin Grammar:
*FINALVOICEDOBSTRUENT, *FINALOBSTRUENT, >> Faithfulness
b. English Grammar:
Faithfulness >> *FINALVOICEDOBSTRUENT, *FINALOBSTRUENT
c. Interlanguage Grammar (Mandarin-speaking learners of English):
 *FINALVOICEDOBSTRUENT >> Faithfulness >> *FINALOBSTRUENT

Broselow (2004) argues that the intermediate ranking falls out of the
Gradual Learning Algorithm approach (Boersma & Hayes 2001), in which
the rate of constraint demotion is an effect of the frequency of input tokens
that violate the constraint. Since the constraint banning all final obstruents
will of necessity be violated more frequently than the constraint banning
only voiced final obstruents, the general constraint is demoted more rapidly
than the more specific constraint. Another application of the Gradual
Learning Algorithm approach, in which constraint rankings are stochastic
and variable across different speech events, is to predict variation; Cardoso
(2007) uses the GLA to model the variable productions he finds in his study
of Brazilian Portuguese learners of English.
While the constraint-based approach appears to be compatible with the
earlier mastery of L2 voiceless than of voiced final obstruents for speakers
of languages like Mandarin (with no final obstruents) or like German (with
only voiceless final obstruents), this approach faces a challenge from the
cases of asymmetry mentioned above involving speakers of languages with
final voicing contrasts. Speakers of Hungarian (Altenberg & Vago 1983)
and Farsi (Eckman 1984), which allow both voiced and voiceless final
obstruents, should approach the L2 with a native language grammar that has
the same constraint ranking as English. To account for these learners’
greater success in producing English voiceless than voiced final obstruents,
a proponent of the grammar-based approach might argue that while the
grammars of Hungarian and Farsi permit final voiced obstruents, the
phonetic realization of voiced targets is sufficiently different from the
realization of voicing in English that the learners’ attempts to produce
voiced stops are not recognized as such by native English speakers.
However, this explanation faces the difficulty that in Hungarian, final
voiced stops are actually more fully voiced than are their English
counterparts (Gosy & Ringen 2009), a fact that might lead us to expect that
transfer of native language articulatory routines should make final voicing
easy to hear. An alternative explanation could appeal to a difference in the
phonological feature specifications that define the laryngeal contrasts in
English vs. in the other languages. If, as proposed in Iverson & Salmons
(1995), Jessen & Ringen (2002), Vaux & Samuels (2005), and many others,
the relevant feature distinguishing stops is [spread glottis] for aspirating
languages like English but [voice] for voicing languages like Hungarian,
then the grammars of Hungarian and Farsi may not in fact be identical to
that of English. The validity of these approaches can only be evaluated in
the context of detailed study of the acoustics of the relevant languages,
explicit analyses of the grammars of the two languages, and explicit
theories of phonological specification.
An additional explanation for the developmental asymmetry that should
be considered is the possibility of asymmetries in the data available to the
learner. If final voiceless obstruents are significantly more frequent in the
target language than are final voiced obstruents (that is, if input to the
learners contains significantly more tokens of final voiceless than voiced
obstruents), this could explain why learners might acquire the former before
the latter, with no recourse to markedness considerations. As Broselow &
Xu (2004) point out, the order in which new English structures are mastered
by Mandarin-speaking learners does not correlate in any obviously way
with the frequency of different English coda types as outlined by Kessler &
Treiman (1997), though systematic studies of frequency in learner input are
lacking.
In summary, the second language data provides convincing evidence for
hierarchies of difficulty: for learners from a variety of native language
backgrounds, L2 final voiced obstruents seem to be harder to successfully
produce than either L2 nonfinal voiced or final voiceless obstruents. We
now examine the nature of L2 learners’ unsuccessful productions of final
voiced obstruents.

4 L2 repair of final voiced obstruents: Too many

solutions?
Steriade (2001/2008) identifies the “too many solutions” problem in
Optimality Theory: the theory predicts that languages will vary in their
strategies for realizing underlying structures that are banned on the surface,
but all or most languages seem to converge on a single repair for particular
prohibited structures. This problem arises from the assumptions that the
choice of repair in a language is a function of the ranking of faithfulness
constraints defining preferred input-output correspondence relationships,
along with the assumption that constraints are freely rankable across
languages. We do indeed find cross-linguistic differences in the repair of
many structures: for example, structures containing vowel hiatus are
variously realized by deletion of a vowel, coalescence of the two vowels
into a single vowel or a diphthong, gliding of a vowel, or insertion of a
consonant between the vowels. Yet, although speakers have in principle a
number of options for repairing structures containing final voiced
obstruents (deletion of the obstruent; epenthesis of a vowel following the
obstruent; sonorization of the obstruent), Steriade argues that the only
productive repair found across languages is final devoicing.
Steriade proposes to solve this problem by assuming that faithfulness
constraints are not necessarily freely rankable; she argues that the ranking
of some constraints is set by the P-Map, which defines the perceptual
distance between different structures (Steriade 2001/2008). Rankings that
produce the maximal perceptual similarity between input and output
structures are favored by the P-Map. Her claim is that devoicing is the
chosen repair for final voiced obstruents because devoicing produces an
output that is perceptually more similar to the input – for example, input
/Vb/ is more similar to output [Vp] than it is to [VbV], [V], or [Vm]. The
universal ranking of faithfulness constraints penalizes other changes more
harshly than a change in voicing:

(5) Ranking of faithfulness constraints favoring devoicing MAX(C),

DEP(V) >> IDENT(VOICE)

Since Steriade’s claim is based in typology, we would expect the preference

for devoicing over other repairs to be instantiated in second languages
phonology as well as in first languages. However, before considering
whether the preference for final devoicing holds for second language
phonological data, we note that the typological facts are not entirely
straightforward, since languages using strategies other than final devoicing
are attested. One example is adduced by Kiparsky (2008), who points out,
citing Cahill (1999), that vowel epenthesis is used to prevent the creation of
a voiced obstruent in coda position in Konni (northern Ghana). This
language has a productive voicing assimilation process whereby an
obstruent assimilates in voicing to a following obstruent. However, where
voicing assimilation would give rise to a voiced coda obstruent,
assimilation is blocked by the insertion of a vowel between the obstruents:
These facts do not directly contradict Steriade’s claim that final devoicing is
the universally preferred strategy for transforming final voiced obstruents to
some other structure, since in Konni, vowel insertion functions not to
remove an underlying voiced obstruent, but rather to block the creation of a
new voiced obstruent in coda position. However, if perceptual similarity is
the major motivation for choice of repair, Steriade must argue that in this
case, the output biisibu is more similar to input /biisbu/ than would be the
output *biizbu.
A stronger challenge to the universality of final devoicing is posed by
the facts of Noon, a Cangin language of Senegal. Merrill (2015) provides
evidence that Noon systematically nasalizes voiced stops that are brought
into coda position by morpheme concatenation:

Merrill argues that the nasalization process arose from two earlier
processes: at an earlier stage, all voiced stops were prenasalized;
subsequently, prenasalized stops became plain stops in onset position but
became nasals in coda. Although the nasalization pattern arose through
separate sound changes, it seems to have become established in the
synchronic grammar by learners who have not been exposed to the separate
stages that gave rise to this pattern. This disqualifies the preference for final
devoicing for the status of true phonological universal, according to
Kiparsky’s criteria (reviewed in Section 1), which include the claim that
learners will never construct grammars that violate a true universal.
Nonetheless, it is clear that the overwhelming majority of languages do
choose final devoicing as the preferred option. A weaker version of
Steriade’s claim would be to ascribe the preference for final devoicing to a
default, initial-state ranking which holds in the absence of evidence to the
contrary, but which could be adjusted when learners are exposed to
evidence contradicting this ranking. On this view, the responsibility to
explain the rarity of repairs other than final devoicing would rest with
channel bias effects, rather than on the formal grammar.
However, attempts to investigate channel bias effects in repair of final
voiced obstruents are not entirely consistent with the perceptual similarity
hypothesis. Kawahara & Garvey (2010), in an online experiment, elicited
direct judgments of perceptual similarity by asking participants to compare
forms with final voiced obstruents (e.g., ab) with possible correponding
forms (e.g., am, a, aba, ap) and to rate the similarity of each pair. In trials
that involved orthographic presentation of forms, the devoicing option was
chosen as most similar to the finalobstruent form, consistent with Steriade’s
claim. But when forms were presented auditorily, the form with final
epenthetic schwa was judged most similar to the final-obstruent form.
Kawahara & Garvey note that the final obstruents in the auditory stimuli
were released, and although the release was spliced off, sufficient
information may have remained to bias listeners toward the vowel insertion
form. These facts suggest that determining the closest perceptual match
may rely on a complex combination of subtle phonetic details.
With these facts in mind, we now turn to the question of whether
learners’ non-target-like productions provide evidence for devoicing as the
preferred (if not necessarily universal) repair. We consider the relative
proportions of different repairs (consonant deletion, vowel insertion, and
final devoicing) in various studies, and the effect on choice of repair of
several factors: learner proficiency, task, and grammatical context; the
existence of an active devoicing process in the native language; word size
and stress; and manner and place of articulation.

4.1 Learner proficiency, task effects, and

grammatical context
While the second language literature contains numerous reports of
devoicing of final obstruents, it also contains many examples of vowel
insertion and consonant deletion. This is not in itself a counterexample to
Steriade’s claim that final devoicing is the preferred option, since devoicing
is only a possibility if the learner can produce some sort of obstruent in
final position. A more serious problem for Steriade would be cases in which
final voiced obstruents were systematically nasalized by learners, while
voiceless obstruents were produced faithfully. I am not aware of such
patterns in the second language literature.
The interesting question from the standpoint of Steriade’s proposal is
whether it is the case that once speakers begin to acquire final obstruents,
devoicing becomes the norm. Devoicing is indeed a common phenomenon,
though we find a good deal of variation in devoicing rates across different
studies, even for speakers of the same native language; for example, Wang’s
(1995) Mandarin speaker participants, who had been in the US less than one
year, had a devoicing rate of 9%, while Flege & Davidian’s (1984)
Mandarin participants, with five or more years in the US, show a devoicing
rate of 29.5%. The lower devoicing found among the less advanced learners
correlated with higher rates of consonant deletion and vowel epenthesis,
consistent with the course of development suggested by Abrahamsson
(2003) for speakers of native languages without final obstruents : consonant
deletion > vowel epenthesis > feature change (devoicing) > target value.
Choice of repair is also clearly affected by the experimental task. For
example, Edge (1991) found that in a word list reading task, Japanese
speakers produced 30% of the word-final voiced obstruents with an
epenthetic vowel, but that the rate of epenthesis dropped to less than 5% in
tasks involving connected speech. It is not surprising that reading tasks, in
particular, might favor epenthesis.
Another factor that may affect both accuracy and choice of repair is the
grammatical status of the final consonant. Hansen (2004) found much
higher accuracy for past tense /d/ (43%) than for stem-final /d/ (11%),
although for some reason, similar effects did not obtain for plural vs. stem-
final /z/.

4.2 Native language devoicing

Speakers whose native language has an active devoicing process can be
expected to transfer this process to a second language, and indeed, many
do; Cebrian (2000) reports 97.8% devoicing in prepausal forms by Catalan
speakers; Flege & Davidian (1984) report a devoicing rate of 48.3% for
Polish speakers; and Hammarberg (1990) reports that almost all the errors
of the Swedish-learning German speakers involved devoicing. However,
speakers of a language with alternations that support a productive devoicing
process do not necessarily devoice L2 forms more often than speakers
whose language lacks such a process – or even than learners with no native
language final obstruents. Comparison across studies is problematic, given
differences in methodology, proficiency of subjects, etc., but Flege &
Davidian’s study involved native speakers of three typologically distinct
languages: Mandarin (no final obstruents), Polish (active final devoicing),
and Mexican Spanish (final obstruents limited in occurrence; voiced stops
spirantize following continuants). The devoicing rates of the three groups
(29.5%, 48.3%, and 43%) were not significantly different, though there was
a high degree of within-group variation. The three native language groups
in Flege & Davidian’s study were chosen to investigate the extent to which
native language processes affected the production of the second language;
thus, the comparison of Spanish speakers, with an active spirantization
process, and Polish speakers, with an active devoicing process, is
instructive. While the likelihood of L2 devoicing was not significantly
higher for Polish speakers than for the other two groups, the rate of final
stop spirantization was much higher for the Spanish speakers (19.3%) than
for either Polish (1.2%) or Chinese (0.8%) speakers. Thus, while the
likelihood of L2 spirantization correlated with the existence of a native
language spirantization process, L2 devoicing appeared even when
unsupported by the native language.

4.3 Word size and stress

For speakers of a language which (like Mandarin) disallows obstruent
codas, deletion of a final obstruent and insertion of a vowel are equally
valid strategies for creating possible native language syllable types. Several
studies have presented evidence that the English of Mandarin-speaking
learners shows evidence of a correlation between repair strategy and output
word size, with a preference for disyllabic words as the determining factor
(Heyer 1986; Weinberger 1987; Wang 1995; Steele 2002). For example,
Wang (1995) found that the size of the source word had a significant effect
on the learners’ choice of repair of pseudoword forms, with a preference for
vowel insertion in monosyllables but deletion in disyllables:

Wang also investigated the effect of word stress on the choice of deletion
vs. insertion. Her disyllabic forms were equally divided between those with
initial stress and those with final stress. Epenthesis was more likely in the
final-stress disyllables than in the initial-stress disyllables, suggesting that
in the absence of word size effects, stress did have an effect. However, in a
comparison of monosyllables with final-stress disyllables, the overall rate
of epenthesis was still significantly higher for monosyllables than for final-
stress disyllables.
Additional evidence of word size effects comes from Cardoso’s (2007)
study of six speakers of Brazilian Portuguese, a language in which the only
possible coda obstruent is /s/. The speakers in this study either produced
coda stops correctly, or inserted a vowel following the coda stop (which he
argues is a productive native language strategy for syllabifying stops,
though he notes that devoicing has been reported in other studies of
Brazilian Portuguese-English interlanguage). Cardoso’s study included
learners at three levels, and while the lowest level speakers produced almost
no codas successfully (i.e., inserted a following vowel), the intermediate
and advanced learners were far more likely to produce coda stops in
polysyllabic words (37% and 59%, respectively), than in monosyllables
(16% correct production for intermediate and 31% correct production for
advanced learners). As Cardoso notes, Brazilian Portuguese contains a
number of highly frequent monosyllables, as does English, though in
English, monosyllabic content words must arguably be bimoraic. He argues
that “the language learner opts for minimal word disyllabicity, a structure
that is enforced neither in BP nor in English, over bimoraicity, which
represents the target-like structure” (Cardoso 2007: 227).
Thus, while devoicing is extremely common in second language
phonology, it is not necessarily the favored strategy, even for learners who
have the ability to produce obstruents in final position. These facts are
consistent with the view that the choice of final devoicing over other repairs
represents at most a strong preference rather than an absolute universal, and
one that may interact with other universal preferences. In fact, the word size
effects are reconcilable with Steriade’s claim that the universal preference
for final devoicing represents a default ranking of faithfulness constraints,
given the architecture of Optimality Theory grammars. So long as the
faithfulness constraints are outranked by markedness constraints demanding
a disyllabic word minimum, vowel insertion will be chosen over deletion or
devoicing for final obstruents in monosyllables, even when the ranking of
faithfulness constraints defines devoicing as the generally preferred option.
This is illustrated in the tableau below (where D indicates any voiced
obstruent):

In this ranking (essentially that proposed in Broselow et al. 1998 to describe

the Mandarin learners’ patterns) the constraint banning vowel insertion,
Dep(V), outranks IDENT(VOICE), the constraint forbidding devoicing,
consistent with Steriade’s proposed ranking. However, the higher-ranked
MINWORD (words must be at least disyllabic) will rule out the devoicing
option for monosyllables.
 Thus, a grammar that adheres to Steriade’s proposed ranking of
faithfulness constraints need not necessarily entail that devoicing will be the
only choice in every context. This brings us back to the typological
question: if this is a possible grammar, then we should expect to find native
languages in which the L2 pattern of insertion in monosyllables and
devoicing in polysyllables has become grammaticalized. One language in
which devoicing is related to word size is Turkish, where Becker et al.
(2011) demonstrate that monosyllables are more likely than longer words to
preserve voiced obstruents in final position (e.g., [ad] ‘name’). However,
the resistance to devoicing in these forms does not change word size;
Becker et al. attribute the preservation of voice to a cross-linguistic
tendency toward greater faithfulness to word-initial syllables. At this point
the tendency toward epenthesis in monosyllables and devoicing in
polysyllables seen in the English of native speakers of Mandarin and
Brazilian Portuguese does not appear to have been grammaticalized in any
language – though it is possible that such a pattern, if it did arise, might be
unstable, since the next generation of learners might be led to reanalyze the
originally monosyllabic forms as underlying disyllabic.

4.4 Manner effects

Final devoicing typically affects both stops and fricatives, despite
differences in the realization of voicing in these two classes. Thus, it is
interesting to see whether voiced stops and fricatives pattern similarly in
second language phonology.
An asymmetry between stops and fricatives emerges in Simon’s (2009,
2010) investigation of Dutch speakers’ productions: though all obstruents
are devoiced in final position in Dutch, the Dutch speakers in her study
produced English final voiced stops significantly more successfully than
final voiced fricatives (the respective rates of devoicing were 76% for
fricatives vs. 49% for stops). As Simon points out, this pattern is consistent
with the cross-linguistically greater rarity of voiced fricatives than voiced
stops, which can be explained by the fact that the glottal opening required
for frication is antithetical to the cross-glottal pressure differential required
to sustain voicing (Ohala 1983). Voiced stops and fricatives differ in terms
of perceptibility as well; Myers (2012) found a tendency for English
speakers to identify utterance-final voiced fricatives as voiceless, while a
similar tendency was not found for utterance-final voiced stops.
A pattern that appears to be the reverse of the Dutch pattern is attested
in Hansen’s (2004) study of two Vietnamese speakers’ productions. Hansen
includes all voicing errors in the category of feature change errors, but her
discussion makes it clear that the normal feature change for stops was
devoicing. While 49% of final voiced stops underwent feature change, only
4% of final fricatives did (the opposite of the Dutch pattern, in which
fricatives were more likely to be devoiced):

A closer look at the choice of repair for stops and fricatives is intriguing.
The rate of target-like productions was the same for voiced stops and voiced
fricatives; the major difference lies in the higher rates of deletion and
epenthesis for voiced fricatives (72% combined) vs. voiced stops (27%
combined). This may be a native language effect: Vietnamese has no final
fricatives, but does have final stops (albeit only voiceless ones). Thus, the
higher rate of devoicing for stops than fricatives may simply reflect the fact
that devoicing is not an option for final fricatives, since these speakers
cannot yet successfully produce fricatives in final position. However, it is
puzzling that for voiceless stops and fricatives, the rates of target-like
productions were comparable (leaving out the surprisingly high rates for
/p/). Thus, while the Vietnamese data provide clear support for a difficulty
hierarchy involving voiceless vs. voiced final fricatives, their significance
with respect to the relationship between manner and the likelihood of
devoicing is less clear.
If we take the Dutch pattern – greater likelihood of devoicing of final
fricatives than final stops – as representative of phonetically-grounded
factors disfavoring voiced fricatives, it seems likely that we should find
languages in which the pattern of devoicing final fricatives but not final
stops has become phonologized. Myers (2012) addresses this question in
the context of his proposal that word-final and syllable-final devoicing
processes arise historically from the generalization of utterance-final
devoicing: “One might expect from this that utterance-final fricative
devoicing should be the most common version of the pattern of final
devoicing [. . .] But it certainly does not seem as if such cases are more
common than [. . .] devoicing of all obstruents including stops” (Myers
2012: 173). Myers cites only one language, Gothic, where final devoicing is
limited to fricatives. Thus, while the Dutch speakers’ L2 patterns are
congruent with aerodynamic and perceptual factors that appear to make a
final voicing contrast in stops more natural than one in fricatives, it does not
seem to be the case that this asymmetry has become widely
grammaticalized. If such an asymmetry emerges frequently in second
language phonology but never as a pattern in a first language, it might
provide an argument for an analytic bias against a grammar that allows
devoicing for one set of obstruents but not another, although at this point
the evidence from second language production is too limited to support this
claim. A related question is whether the same or different feature
specifications govern laryngeal contrasts for stops and fricatives, and
whether stop devoicing and fricative devoicing should be treated as
different processes in grammars. (See Vaux 1998 for the proposal that the
unmarked opposition for voiceless and voiced fricatives is [+spread glottis]
vs. [-spread glottis], and van Oostendorp 2007 for the proposal that for at
least some Dutch dialects, the fricative contrast is better explained in terms
of length rather than laryngeal features.)

4.5 Place of articulation

Another factor related to voicing is place of articulation: because voicing
requires that supraglottal pressure be lower than air pressure below the
glottis, a smaller oral cavity makes voicing more difficult to sustain. Thus,
velar stops tend to be less fully voiced than stops made farther front (Ohala
1983; Maddieson 1984). Despite the connection between constriction
location and difficulty of sustaining voicing, there seem to be no languages
that, say, devoice final velar stops but maintain voicing contrasts for
coronals and labials. Nor are there clear examples of such patterns in
second language phonology, although there is evidence for differences in
degree of voicing across place of articulation. Yavas (2009) studied the
production of final English bilabial, coronal, and velar stops by native
speakers of Mandarin, Japanese, and Portuguese, and found that although
the amount of closure voicing did not differ significantly by place of
articulation, there was an interaction between the place of the stop and the
height of the preceding vowel: velars were significantly less voiced, but
only when they followed high vowels:

Native speakers also showed less voicing in the high vowel-velar case than
in other cases, but their percentage of voicing was, for each vowel-
consonant combination, significantly higher than that of native speakers
(e.g., 65.5% voicing during a velar closure following a high vowel).
While it makes sense that the smaller oral cavity associated with velars
and the narrower constriction of high vowels should have an additive effect
on voicing, this does not appear to be a phonologized pattern in languages;
Moreton (2008) argues that few languages show systematic interactions of
vowel height and voicing, and those interactions that are attested take the
form of the raising of vowel height before voiced consonants and the
lowering of vowel height before voiceless consonants (though see Yu 2011
for a different interpretation of Moreton’s data). Thus, place and vowel
height effects, though they appear in the phonetic detail of both native and
non-native speakers, appear not to have been grammaticalized in either first
language or interlanguage phonology. Again, this is an area where research
is relatively sparse.

5 Conclusion
We set out to determine first, whether the facts of second language
phonology are compatible with typological generalizations, and second,
whether the second language facts can shed light on the source of
typological generalizations.
We found numerous cases supporting a difficulty hierarchy for final
voiceless vs. voiced stops in second language phonology, and this difficulty
hierarchy aligns with the typological generalizations on preferred segment
type. Across a range of native languages, including those with no final
obstruents, those with only voiceless final obstruents, and those with a final
laryngeal contrast, speakers successfully produced L2 final voiceless
obstruents before final voiced obstruents. In no case was there evidence of
speakers acquiring the more marked structure (final voiced obstruents)
before the less marked structure (final voiceless obstruents). Whether these
facts reflect articulatory and perceptual factors or the effects of formal
grammatical constraints is difficult to resolve – since the structural
constraints of Optimality Theory are generally grounded in articulatory and
perceptual considerations, there is considerable overlap between the
approaches. However, we note that locating the difficulty of final voiced
obstruents in articulatory and perceptual difficulty alone predicts that the
likelihood of a difficulty hierarchy emerging in second language acquisition
should be a function of the phonetic robustness of the contrast in the target
language; for example, languages in which final stops are uniformly
released should provide the learner with more cues to the voicing contrast
than languages without such release. Systematic study of the productions of
both native and second language speakers across a range of languages is
necessary to address this question.
We also found that final devoicing was quite common in second
language phonology, although it was by no means the only strategy used.
Since speakers must be able to produce obstruents in final position before
they can devoice them, the fact that vowel insertion and consonant deletion
were also common repairs of L2 forms does not in itself invalidate
Steriade’s (2001/2008) claim that final devoicing is the only solution to the
final obstruent problem. We did, however, find evidence that some speakers
exhibit a systematic relationship between choice of repair and preferred
word size. It is intriguing that this pattern was found for speakers of two
different languages, Mandarin and Brazilian Portuguese, but is not clearly
attested in any native language system.
 Effects of aerodynamic factors that contribute to the difficulty of
sustaining voicing appeared in some studies, at the level of relatively fine
phonetic detail: Dutch speakers’ final fricatives were less voiced than stops
(Simon 2010), and Mandarin, Japanese, and Portuguese speakers’ velars
were less voiced than alveolars and bilabials, though only after high vowels
(Yavas 2009). On the channel bias account, we might expect these
differences to give rise to systems in which the phonetic asymmetries
become phonologized. Yet such systems seem either rare or unattested;
Myers (2012) cites only one language, Gothic, in which fricatives, but not
stops, are regularly devoiced.
A reasonable place to look for systems that have phonologized the
effects found in second language phonology is in regionalized varieties of
English, where what generally began as a second language has now become
standardized. A striking number of regional Englishes show evidence of at
least some final devoicing. In a survey of English varieties of Africa, South
Asia, and Southeast Asia, Mesthrie (2004) reports final devoicing in St.
Helens English, Cape Flats English, Black South African English, Nigerian
English, Ghanaian English, Cameroon English, Cameroon Pidgin,
Singapore English, and Malaysian English. Final devoicing is also reported
in Fiji English (Tent & Mugler 2004), Tok Pisin (Smith 2004), and Liberian
Settler English (Singler 2004). The prevalence of final devoicing suggests
that speakers did indeed converge on this repair as their systems stabilized.
It is notable that none of these descriptions identify epenthesis or deletion
as regular productive processes targeting single final voiced obstruents, and
no systems are identified as showing different treatment of final stops and
fricatives or systematic effects of place of articulation that are independent
of the substrate language. Thus, at least some of the well-founded phonetic
effects that emerge in second language phonology fail to acquire the status
of regular phonological processes.

Acknowledgements: I am grateful to audiences at Oxford, CUNY, and

Stony Brook and particularly to Larry Hyman for comments that have
improved this paper.

References
Abrahamsson, Niclas. 2003. Development and recoverability of L2 codas: A
longitudinal study of Chinese-Swedish interphonology. Studies in Second
Language Acquisition 25. 313–349.
Altenberg, Evelyn & Robert Vago. 1983. Theoretical implications of an error
analysis of second language phonology production. Language Learning 33.
427–448.
Becker, Michael, Nihan Ketrez, & Andrew Nevins. 2011. The surfeit of the stimulus:
Analytic biases filter lexical statistics in Turkish laryngeal alternations.
Language 87. 84–125.
Beckman, Jill, Pétur Helgason, Bob McMurray, & Catherine Ringen. 2011. Rate
effects on Swedish VOT: Evidence for phonological overspecification. Journal
of Phonetics 39. 39–49.
Blevins, Juliette. 2004. Evolutionary phonology: The emergence of sound patterns.
Cambridge: Cambridge University Press.
Blevins, Juliette. 2006. A theoretical synopsis of Evolutionary Phonology.
Theoretical Linguistics 32. 117–166.
Blevins, Juliette. 2010. Phonetically-based sound patterns: Typological tendencies
or phonological universals? In Cécile Fougeron, Barbara Kühnert, Mariapaola
D’Imperio, & Nathalie Vallée (eds.), Papers in Laboratory Phonology 10:
Variation, phonetic detail and phonological modeling, 201–224. Berlin: Mouton
de Gruyter.
Broersma, Miriam. 2005. Perception of familiar contrasts in unfamiliar positions.
Journal of the Acoustical Society of America 117. 3890–3901.
Broselow, Ellen. 2004. Unmarked structures and emergent rankings in second
language phonology. International Journal of Bilingualism 8. 51–65.
Broselow, Ellen, Su-I Chen, & Chilin Wang. 1998. The emergence of the unmarked
in second language phonology. Studies in Second Language Acquisition 20.
261–280.
Broselow, Ellen & Yoonjung Kang. 2013. Second language phonology and speech.
In Julia Herschensohn & Martha Young-Scholten (eds.), The Cambridge
handbook of second language acquisition, 529–554. Cambridge: Cambridge
University Press.
Broselow, Ellen & Zheng Xu. 2004. Differential difficulty in the acquisition of
second language phonology. International Journal of English Studies 4
(special issue: Advances in Optimality Theory). 13–163.
Cahill, Michael. 1999. Aspects of morphology and phonology of Konni. Ph.D.
dissertation, Ohio State University.
Cardoso, Walcir. 2007. The variable development of English word-final stops by
Brazilian Portuguese speakers: A stochastic Optimality Theory account.
Language Variation and Change 19. 219–248.
Cebrian, Juli. 2000. Transferability and productivity of L1 rules in Catalan-English
interlanguage. Studies in Second Language Acquisition 22. 1–26.
Cichocki, Wladyslaw, Anthony B. House, A. Murray Kinloch, & Anthony C. Lister.
1993. Cantonese speakers and the acquisition of French consonants.
Language Learning 43. 43–68.
Eckman, Fred. 1977. Markedness and the contrastive analysis hypothesis.
Language Learning 27. 315–330.
Eckman, Fred. 1981. On the naturalness of interlanguage phonological rules.
Language Learning 31. 195–216.
Eckman, Fred. 1984. Universals, typology, and interlanguage. In William E.
Rutherford (ed.), Language universals and second language acquisition, 79–
105. Amsterdam: John Benjamins.
Eckman, Fred. 2004. From phonemic differences to constraint rankings: Research
on second language phonology. Studies in Second Language Acquisition 26.
513–549.
Edge, Beverly. 1991. The production of word-final voiced obstruents in English by
L1 speakers of Japanese and Cantonese. Studies in Second Language
Acquisition 13. 377–393.
Flege, James Emil & Richard D. Davidian. 1984. Transfer and developmental
processes in adult foreign language speech production. Applied
Psycholinguistics 5. 323–347.
Flege, James Emil, Martin J. McCutcheon, & Steven C. Smith. 1987. The
development of skill in producing word-final English stops. Journal of the
Acoustical Society of America 82. 433–447.
Flege, James Emil & Chipin Wang. 1989. Native-language phonotactic constraints
affect how well Chinese subjects perceive the word-final English /t/-/d/
contrast. Journal of Phonetics 17. 299–315.
Gordon, Matthew. 2007. Typology in Optimality Theory. Language and Linguistics
Compass 1. 750–769.
Gosy, Maria & Catherine Ringen. 2009. Everything you always wanted to know
about VOT in Hungarian. Talk presented at the International Conference on
the Structure of Hungarian 9. Debrecen, Hungary.
Greenberg, Joseph, Charles Ferguson, & Edith Moravscik (eds.). 1978. Universals
of human language. Volume 2: Phonology. Stanford, CA: Stanford University
Press.
Hammarberg, Björn. 1990. Conditions on transfer in phonology. In Allan R. James
& Jonathan Leather (eds.), New sounds 90: Proceedings of the 1990
Symposium on the Acquisition of Second-Language Speech, 198–215.
Dordrecht: Foris.
Hancin-Bhatt, Barbara. 2000. Optimality in second language phonology: Codas in
Thai ESL. Second Language Research 16. 201–232.
Hansen, Jette G. 2004. Developmental sequences in the acquisition of English L2
syllable codas. Studies in Second Language Acquisition 26. 85–124.
Helgason, Pétur & Catherine Ringen. 2008. Voicing and aspiration in Swedish
stops. Journal of Phonetics 36. 607–628.
Heyer, Sarah. 1986. English final consonants and the Chinese learner.
Unpublished master’s thesis, Southern Illinois University Edwardsville.
Hillenbrand, James, Dennis R. Ingrisano, Bruce L. Smith, & James E. Flege. 1984.
Perception of the voiced-voiceless contrast in syllable-final stops. Journal of
the Acoustical Society of America 76. 18–26.
Hyman, Larry. 1976. Phonologization. In Alphonse Juillard (ed.), Linguistic studies
offered to Joseph Greenberg, volume 2, 407–418. Saratoga, CA: Anna Libri.
Hyman, Larry. 2008. Universals in phonology. The Linguistic Review 25. 83–137.
Iverson, Gregory & Joseph Salmons. 1995. Aspiration and laryngeal
representation in Germanic. Phonology 12. 369–396.
Iverson, Gregory & Joseph Salmons. 2011. Final devoicing and final laryngeal
neutralization. In Marc van Oostendorp, Colin Ewen, Elizabeth V. Hume, &
Keren Rice (eds.), The Blackwell companion to phonology, volume 3, 1622–
1643. Oxford: Blackwell Publishing.
Jessen, Michael & Catherine Ringen. 2002. Laryngeal features in German.
Phonology 19. 189–218.
Kawahara, Shigeto & Kelly Garvey. 2010. Testing the P-map hypothesis: Coda
devoicing. Rutgers Optimality Archive.
Keating, Patricia, Wendy Linker, & Marie Huffman. 1983. Patterns in allophone
distribution for voiced and voiceless stops. Journal of Phonetics 11. 277–290.
Kenstowicz, Michael. 2005. The phonetics and phonology of loanword adaptation.
In S.-J. Rhee (ed.), Proceedings of ECKL 1: Proceedings of First European
Conference on Korean Linguistics, 316–340. Seoul: Hankook Publishing.
Kessler, Brett & Rebecca Treiman. 1997. Syllable structure and the distribution of
segments in English syllables. Journal of Memory and Language 37. 295–311.
Kiparsky, Paul. 2006. The amphichronic program vs. evolutionary phonology.
Theoretical Linguistics 32. 217–236.
Kiparsky, Paul. 2008. Universals constrain change; change results in typological
generalizations. In Jeff Good (ed.), Language universals and language
change, 23–53. Oxford: Oxford University Press.
Lombardi, Linda. 1995. Laryngeal neutralization and syllable well-formedness.
Natural Language and Linguistic Theory 13. 39–74.
Maddieson, Ian. 1984. Patterns of sounds. Cambridge: Cambridge University
Press.
Major, Roy & Michael Faudree. 1996. Markedness universals and the acquisition
of voicing contrasts by Korean speakers of English. Studies in Second
Language Acquisition 18. 69–90.
Merrill, John. 2015a. Nasalization as a repair for voiced obstruent codas in Noon.
Talk presented at the Annual Meeting of the LSA, January, 2015.
Merrill, John. 2015b. Nasalization as a repair for voiced obstruent codas in Noon.
LSA Annual Meeting Extended Abstracts. http://journals.linguisticsociety.org/p
roceedings/index.php/ExtendedAbs/article/view/3014
Mesthrie, Rajend. 2004. Synopsis: The phonology of English in Africa and South
and Southeast Asia. In Schneider et al. (eds.) 2004, 1099–1110.
Moreton, Elliott. 2008. Analytic bias and phonological typology. Phonology 25. 83–
127.
Myers, Scott. 2012. Final devoicing: Production and perception studies. In Toni
Borowsky, Shigeto Kawahara, Takahito Shinya, & Mariko Sugahara (eds.),
Prosody matters: Essays in honor of Elisabeth Selkirk, 148–180. London:
Equinox.
Myers, Scott & Jaye Padgett. 2015. Domain generalisation in artificial language
learning. Phonology 31. 399–434.
Oh, Mira. 1996. Linguistic input to loanword phonology. Studies in Phonetics,
Phonology, and Morphology 2. 117–126.
Ohala, John. 1981. The listener as a source of sound change. In Carrie S. Masek,
Roberta A. Hendrick, & Mary Frances Miller (eds.), CLS: Papers from the
parasession on language and behavior, 178–203. Chicago: Chicago Linguistic
Society.
Ohala, John. 1983. The origin of sound patterns in vocal tract constraints. In Peter
MacNeilage (ed.), The production of speech, 189–216. New York: Springer.
Oostendorp, Marc van. 2007. Exceptions to final devoicing. In Jeroen van de
Weijer & Erik Jan van de Torre (eds.), Voicing in Dutch: (De)voicing –
phonology, phonetics, and psycholinguistics, 81–98. Amsterdam: John
Benjamins,
Peng, Long & Jean Ann. 2004. Obstruent voicing and devoicing in the English of
Cantonese speakers from Hong Kong. World Englishes 23. 535–564.
Raphael, Lawrence. 1972. Preceding vowel duration as a cue to the perception of
the voicing characteristic of word-final consonants in American English.
Journal of the Acoustical Society of America 51. 1296–1303.
Schneider, Edgar, Kate Burridge, Bernd Kortmann, Rajend Mesthrie, & Clive Upton
(eds.). 2004. A handbook of varieties of English. Berlin: Mouton de Gruyter.
Simon, Ellen. 2009. Acquiring a new second language contrast: An analysis of the
English laryngeal system of native speakers of Dutch. Second Language
Research 25. 377–408.
Simon, Ellen. 2010. Phonological transfer of voicing and devoicing rules: Evidence
from NL Dutch and L2 English conversational speech. Language Sciences 32.
63–86.
Singler, John. 2004. Liberian Settler English: Phonology. In Schneider et al. (eds.)
2004, 874–884.
Smith, Geoff. 2004. Tok Pisin in Papua New Guinea: Phonology. In Schneider et
al. (eds.) 2004, 710–728.
Smolensky, Paul. 1996. On the comprehension/production dilemma in child
language. Linguistic Inquiry 27. 720–731.
Steele, Jeffrey. 2002. L2 learners’ modification of target language syllable
structure: Prosodic licensing effects in interlanguage phonology. In Allan R.
 James & Jonathan Leather (eds.), New sounds 2000: Proceedings of the 4th
International Symposium on the Acquisition of Second Language Speech,
315–324. Klagenfurt: University of Klagenfurt.
Steriade, Donca. 1999. Phonetics in phonology: The case of laryngeal
neutralization. UCLA Working Papers in Phonology 3. 25–146.
Steriade, Donca. 2001/2008. The phonology of perceptibility effects: The P-Map
and its consequences for constraint organization. In Kristin Hanson & Sharon
Inkelas (eds.), The nature of the word, 151–179. Cambridge: MIT Press.
Tent, Jan & France Mugler. 2004. Fiji English: Phonology. In Schneider et al. (eds.)
2004, 750–779. Vaux, Bert. 1998. The laryngeal specification of fricatives.
Linguistic Inquiry 29. 497–511.
Vaux, Bert & Bridget Samuels. 2005. Laryngeal markedness and aspiration.
Phonology 22. 395–346.
Wang, Chilin. 1995. The acquisition of English word-final obstruents by Chinese
speakers. Unpublished doctoral dissertation, Stony Brook University.
Weinberger, Steven. 1987. The influence of linguistic context on syllable
simplification. In Georgette Ioup & Steven H. Weinberger (eds.), Interlanguage
phonology: The acquisition of a second language sound system, 401–417.
Rowley, MA: Newbury House.
Westbury, John & Patricia Keating. 1986. On the naturalness of stop consonant
voicing. Journal of Linguistics 22. 145–166.
Wetzels, Leo & Joan Mascaró. 2001. The typology of voicing and devoicing.
Language 77. 207–244.
Wiltshire, Caroline. 2006. Word-final consonant and cluster acquisition in Indian
Englishes. In David Bamman, Tatiana Magnitskaia, & Colleen Zaller (eds.),
Online proceedings supplement, Boston University Conference on Language
Development 30.
Wissing, Daan & Wim Zonneveld. 1996. Final devoicing as a robust phenomenon
in second language acquisition: Tswana, English and Afrikaans. South African
Journal of Linguistics 14. 3–23.
Yavas, Mehmet. 2009. Factors influencing the VOT of English long lag stops in
interlanguage phonology. In M. A. Watkins, A. S. Reuber, & B. O. Baptista
(eds.), Recent research in second language phonetics/phonology: Perception
and production, 244–255. Newcastleupon-Tyne: Cambridge Scholars
Publishing.
Yu, Alan C. 2004. Explaining final obstruent voicing in Lezgian: Phonetics and
history. Language 80. 73–97.
Yu, Alan C. 2011. On measuring phonetic precursor robustness: A response to
Moreton. Phonology 28. 491–518.
Tomas Riad
The phonological typology of North
Germanic accent

Abstract: The dialects of Swedish and Norwegian frequently exhibit a

tonal contrast within the intonational prominence that is superimposed on
primary stressed syllables. Realizational and distributional properties are
largely shared such that a rather tightly constrained typology is in evidence.
The aim of this article is to show that the linguistic properties that form the
basis of the North Germanic tonal accent typology are structural in nature,
relating to phonological representation in terms of value of the lexical tone
as high/ H or low/ L, tonal association patterns in compounds, and
spreading behaviour. The main type of structure to compare between
dialects is the long compound (several TBU’s, regularly assigned accent 2),
given that the accent distinction is privative, and that accent 2 is the marked
member of the distinction. Several previous typologies have been based on
phonetic categories (e.g., number of tonal peaks, presence/absence of a
separate focus gesture) and often also include functional notions like focus.
I argue that such typologies fail to capture several generalizations, once we
look at all major dialect types, and that they also tend to overgenerate in
their predictions of possible dialect types. The secondary aim of the article
is to provide an updated and coherent account of the major dialect types. To
bring out the typological variables, I make pairwise comparisons of
minimally different dialects on particular variables, and provide illustrating
examples from natural speech.
The typology of closely related varieties gives us an idea of the frame
for variation, which in turn allows for the formulation of what constitutes a
likely or less likely diachronic change. The North Germanic tonal typology
is very coherent with regard to structure, geographic distribution of
features, and also shared history. A relevant analysis will reveal the
prosodic relationship between dialects, and allow us to formulate
hypotheses regarding the relative structural distance between tonal varieties.
From there we can adduce arguments for the reconstruction of diachronic
developments.

1 Introduction
Many of the dialects of Swedish and Norwegian exhibit a tonal contrast
within the intonational prominence that is superimposed on a syllable
carrying primary stress. Many properties of these so-called accents are
shared between dialects, but the tonal variation makes the dialects sound
quite different from one another, and this constitutes the main source for the
man-in-the-street’s recognition of the major dialect areas. This article is
concerned with laying bare the linguistic properties that form the basis of
the typology. My main point is that they are all structural in a quite concrete
sense, relating to phonological representation in terms of value of the
lexical tone as high/H or low/ L, tonal association patterns in compounds,
and spreading behaviour. Several previous typologies have been based on
more phonetic and/ or functional categories (e.g., number of tonal peaks,
presence/ absence of a separate focus gesture) which may describe parts of
the typology well, but which are ultimately too superficial. They tend either
to be insufficient when the typology is extended to all major dialect types or
to overgenerate in their predictions of possible dialect types. The
importance of identifying the most relevant structural categories for
typology in the tonal domain is emphasized below. Following from this, I
also hope to provide an updated and coherent account of the major dialect
types.
Varieties of Germanic that exhibit a lexical tonal contrast occur not only
in North Germanic (NGmc), but also in the Central Franconian varieties
spoken in and around the Rhine delta (West Germanic). There too, the tonal
distinction is superimposed on stressed syllables. If we look more widely,
we find this type of system also in, for instance, Bosnian/Croatian/Serbian,
varieties of Basque, Latvian and Lithuanian. The North Germanic system is
differently constituted from the Franconian and Baltic ones, e.g., in
requiring two syllables for the expression of one of the tonal categories
(accent 2). There is also an organic relationship between the NGmc tonal
varieties and the Danish stød system (Gårding 1977; Ringgaard 1983; Riad
2000a, 2000b, 2009b), but the typology given below will not include
Danish in a principled way, as there are outstanding issues regarding the
representation and status of Danish stød.122
The identification of parameters for microvariation is of course always
of interest, descriptively as well as comparatively. In the case of the
typology of very closely related varieties, the interest is enhanced by the
fact that one has a chance of getting a handle on the general frame for
variation. Depending on the unity of the system as a whole, the access of
relatively rich linguistic information might allow for the formulation of
what constitutes a likely or less likely change. The North Germanic
typology is very coherent with regard to tonal structure, geographic
distribution of features, and also shared history. A good analysis should
reveal the prosodic relationship between dialects, and allow us to formulate
hypotheses regarding the relative structural distance between tonal varieties,
in its turn a prerequisite for the reconstruction of diachronic developments
within the tonal system.
It is our task, then, to identify the properties that best describe the
variation and thereby the individual varieties. My claim is that phonology
provides the most relevant level at which to formulate these things. In
particular, we must view phonetic and functional categories with
scepticism. While phonetics will provide a lot of relevant information, it is
the typology at the phonological/ grammatical level that best explains
relationships between varieties.123

2 Previous typologies
There are a number of earlier typological treatments of NGmc accent. The
ground-breaking study of Meyer (1937, 1954) provides accent contrasts in
chiefly disyllabic simplex forms for some one hundred informants from
various locations around Scandinavia, with most recordings made in the
Central Swedish and Dala regions, a fair amount from the Göta region
(West Swedish, WSw), some recordings from North Swedish (NSw) and
scattered items from Finland, Estonia, and Denmark. Meyer’s materials
form an initial basis for the description of dialects in terms of the number of
tonal peaks, i.e., as one- or two-peaked realizations of accent 2 (cf. (1)
below). This material was also used as basis for a hypothesis known as the
“Scandinavian accent orbit” by Öhman (1967), where varieties were lined
up according to the phonetic timing of peaks. In the phonetic tradition of
Norway, there are the studies of Fintoft (1970), Mjaavatn (1978), and
Fintoft, Mjaavatn, Møllergård, & Ulseth (1978), where four types of tonal
contour are identified (the same set for both accent 1 and accent 2). These
are then paired to get the various dialect areas. In Norway, tradition refers to
dialect types via the first tone of the accent 1 contour, as “high-tone” and
“low-tone” dialects.
In the Swedish tradition it is customary to talk about single peak
dialects and double peak dialects (or one- and two-peak dialects). This
refers exclusively to the accent 2 contour, as accent 1 invariably has only
one peak (Gårding & Lindblad 1973; Bruce & Gårding 1978). This
terminology remains in later work by Bruce (2005, 2007), where the
typology is more refined and more clearly extended to the broader
intonation. Below is the basic typology currently assumed for disyllabic
simplex words in citation form (given in Bye 2004, based on Gårding &
Lindblad 1973), illustrated with Meyer’s tonal contours.124
The general problem with this type of phonetic typology is overgeneration.
The reference points mentioned admit several types that are not attested
(e.g., a variant of 2A but with accent 1 having early timing of the peak, or a
type combining accent 1 of 1A with accent 2 of 1B), without articulating
expectations or reasons for why they should be excluded (or not). Some
answers to this type of problem will come out of a segmentation of the tonal
contours into constituent tones (Bruce 1977, 2004).
 Proper phonological typologies, where reference is made to
phonological categories, e.g., by breaking down the global contour into a
string of constituent tones, are given in work by Lorentz (1995), Riad
(1998b, 2006), Bye (2004), and Bruce (2005, 2007). In Lorentz (1995), for
instance, the contour is divided into lexical tone, prominence tone, and
boundary tone. It must be noted, however, that beyond the lexical tone, the
functional aspects are not reliably tied to individual tones (Riad 2006). For
one thing, there appears to be a bias to use H tones for the prominence
function, whether it is a separate tonal gesture or a boundary tone (cf. Sectio
n 7).
The phonological typologies vary in geographic coverage (as well as in
analysis), but have in later years come to heed the broader area of North
Germanic tonal dialects, including both Norway and Sweden, and
sometimes also the few remaining tonal dialects of Finland (Lorentz 1995;
Riad 1998b; Bye 2004). In addition, these systems can then be put in
relation with Danish stød, which is clearly related historically (as is evident
from lexical distributional patterns), but which also clearly stands out
within the linguistic area.
Part of the background assumptions made by Gårding & Lindblad
(1973), Bruce (2005) and others is that two peaks would never occur in
accent 1, at least not to the exclusion of two peaks in accent 2. Another
(related) fact is that accent 2 always has the richer tonal structure, e.g., by
requiring one more tonal feature than accent 1. Beside these things, which
have implications for the representation of accentual contours, there are
distributional facts of which scholars have different interpretations. For
instance, tradition has often considered accent 2 as the typical accent of
disyllabic forms with initial stress, a fact supported by sheer type and token
frequencies. In phonological analyses this has sometimes been interpreted
as grounds for assuming accent 2 as the default accent of disyllables (Kock
1878; Danell 1937: 51; Malmberg 1970: 157; Öhman 1966; Teleman 1969:
187; Nyström 1997; Lahiri et al. 2005; Wetterlin 2010). On the other hand,
accent 2 correlates robustly with a large class of suffixes, inviting a quite
different analysis where accent 2 results from lexically represented
information in those suffixes (Riad 2009a, 2012, 2014, 2015). This type of
difference will not be settled in this article.125 Instead, we shall make a
number of basic assumptions explicit and then move on to the typological
comparison.
3 The tonal accent system and the crucial forms
The terms “accent 1” and “accent 2” are usually used with reference to the
entire tonal contour of words in citation form, hence in a focused context.
This means that lexical and intonational tonal material is not distinguished
until further segmentation is made. (2) provides an overview of the
segmentation in Central Swedish, for the two prominence levels in which
the accent distinction is realized. We will refer to the higher prominence
level as the “big” accent, and the lower prominence level as the “small”
accent (Myrberg & Riad 2015). This keeps the terminology free from
functional implications, and directs attention to the phonological shape,
without tying it to a particular dialect.126 There is thus a categorical
prominence level distinction between big accent and small accent, and the
lexical distinction is realized in both of them. For the purposes of the
typology, it will suffice to look at the big accent, which (largely) includes
the tonal material of the small accent.127 Bolded tones are lexical, all other
tones are postlexical. The accent distinction is privative.

The initial H* tone of the accent 2 contour is what distinguishes the big
accent contours for accent 1 and accent 2, respectively. The rise in the big
accent (LH) is common to both accent 1 and 2, and is purely intonational
(hence postlexical). We will refer to it as the “prominence tone”. The tonal
sequence is the same in the two cases of accent 2, but the initial H tone has
different sources. In simplex accent 2, the initial H tone (bold) is lexical (in
a root or a suffix), whereas in compound accent 2, this tone is postlexical.
This postlexical accent 2 tone is assigned by a rule which is sensitive to the
number of stresses, and which overrules lexical specifications.128 Any tones
of accent 1 are assigned postlexically. We will use the term “lexical tone” or
“(post)lexical tone” in reference to the first tone in the accent 2 contour
(bolded). This tone is what instantiates the marked member of the accentual
opposition, and accent 1 consequently consists of just intonation tones. The
lexical tone is invariably associated to the primary stressed syllable. The
next tone is the “prominence tone”. In accent 1, it is associated to the
primary stressed syllable, while in accent 2, it is displaced to the right by
the lexical tone which occupies the stressed syllable. A lexical tone thus
always has precedence to the stressed syllable (= TBU), which will only
host a single, associated tone. In citation forms the prominence tone is
followed by the “boundary tone”, usually L%. The boundary tone is not
associated to a TBU, but is aligned with the end of the phrase. The three
terms are used in the overview in (9).
Let us now have a look at the privative contrast in simplex forms. In the
following panels the big accent (on 1ˈAllan and 2ˈAnna, respectively) is
followed by a small accent (on i1ˈgår ‘yesterday’). The presence of the
small accent here creates a stable endpoint for the big accent, allowing us to
compare the realization of the shared part of the big accent contour in the
two accents. The (small accent) HL* is coordinated with the beginning of
the stressed vowel of i1ˈgår in each panel.
The distinction between the two accents thus lies in the initial part of the
contour, where accent 2 contains an extra tone. The rest of the big tonal
contour of accent 2 is identical to all of the accent 1 contour and constituted
by intonation tones only. We can illustrate this fact by matching the two
contours as in (5), where the lexical tone of accent 2 is to the left of the first
vertical line. In the right-hand panels, the contour of accent 1 is compressed
and the identity of the intonational part of the contour is evident. The tonal
sequences for accent 2 and accent 1 are given above and below,
respectively.

The shared part of the contour is delayed in accent 2, due to the presence of
the lexical tone. For the big accent 1 there is, then, more space for the
prominence tone (L*H), and the low target of the following small accent
(HL*) is reached earlier in igår than following a big accent 2.
Simplex forms typically only contain a single association point since
minimal prosodic words in Swedish can contain only a single stressed
syllable, unlike, e.g., English and German (Riad 2014). Whenever there are
more stresses, more minimal prosodic words are created, and the structure
as a whole receives what we shall call “compound accent”.129 Compound
accent is melodically the same as accent 2, but the first tone is postlexical
rather than lexical, cf. (2). Accent is here sensitive to the number of stresses
in a form, and this holds of Central Swedish and several other dialects
(Dala, Narvik, Göta).130 The fact that more TBU’s become available in
compounds also makes it possible for the prominence tone to associate, to
the last stress of the compound (or other similar forms containing two
stresses, formal compounds or derivations with a stressed suffix, cf. (24)).
This is illustrated in (6).

This contour shows how the postlexical H* associates to the first stress and
the prominence L*H associates to the second and last stress. The trailing H
of the prominence tone floats and does not exhibit stable timing (Bruce
1987). Indeed, it may in some dialects drift to the right of the focused word
(Bruce 2003; Myrberg 2010).132 The example in (6) is cut out from the
middle of a phrase, so there is no final boundary L%, a fact that might
motivate the relatively late realization of the floating H. The fact that it is
the last stress that is the target for the secondary association is clear from
forms that contain several stressed syllables. This is illustrated in (7).
Longish compounds in focus position (i.e., exhibiting the big accent) that
carry accent 2 is the single most relevant type of data to use in dialect
comparison, since they exhibit more properties of the tonal grammar than
any other form. Unlike simplex forms, otherwise the typical data used in
typologies of North Germanic accent, long compounds show whether or not
there is a secondary association to a later TBU, as well as if there is
spreading or interpolation between association points. Both of these things
prove to be important parameters of the tone accent typology.133 We
comment on compounds with accent 1 separately, in Section 5.3.
The lexical accent distinction in terms of minimal pairs is of no
particular interest to the typology as such. The distinction carries a very
marginal functional load, and alleged minimal pairs are often not “clean”,
i.e., they are often constituted by inflected forms where the uninflected
forms do not form a minimal pair (8a). Also, it is incidental changes like
vowel reductions and consonant assimilations that determine whether there
are relatively many minimal pairs (Norwegian, around 3000, Leira 1998) or
not (Swedish, about 350, Elert 1972), cf. (8b). Furthermore minimality is
seldom instantiated by forms of the same grammatical category, (8c). The
many flaws of minimal pairs, however, do not mean that there is no
unpredictable lexical distinction. Monomorphemic word pairs like those in
(8d) show that lexical tones are real. These forms are also semantically
close to each other.

It is the near-minimal pairs in (8d) that establish the lexical contrast, rather
than the alleged minimal pairs in (8a) and (8c).

4 One typology
The tonal dialects form a coherent typology by virtue of sharing some basic
properties. For one thing the realization of accent 2 requires two syllables
(disregarding the few apocopating dialects; Lorentz 2008), whereas accent
1 requires only a single syllable. This is indeed the first argument for the
privative nature of the contrast. Since accent 2 requires more space, there
should be more tonal material in that contour.134
The fact that accent 2 requires so much space has an interpretation in
terms of tone bearing unit, which is the stressed syllable in Swedish and
Norwegian. Tones associate to primary stressed syllables in all dialects, and
in some dialects they associate to secondary stresses, too. This
understanding of the TBU coupled with the common assumption of one-
tone-per-TBU rather predicts the synchronic requirement of two or more
syllables for the occurrence of accent 2. If the TBU were the mora, as for
example in most eastern Central Franconian varieties (Peters 2007: 171),
there would be nothing in the way of a contrast in monosyllables.
 Another thing that makes the typology coherent is the tonal alternation,
which means that there are no challenges to the Obligatory Contour
Principle (OCP). While it has been proposed that there are OCP-induced
tonal epentheses (Lorentz 1995), the simplest analysis is achieved by
simply segmenting the tonal contour into three basic parts: lexical tone (if
any), prominence tone, and boundary tone, where each tone is (or begins
with) the opposite value of the preceding tone. The lexical tone is invariably
a single tone, and that seems to be the case also with the boundary tone,
though the issue has not been systematically studied. The prominence tone
may be single or complex. The apparent generalization here (which we
return to in Section 7) is that a prominence tone must contain an H tone,
unless the boundary tone is H% and employed for the expression of focus
(East Norwegian). At any rate, there is no need to postulate tonal
epenthesis. The more conservative view that there are no epenthetic tones
also has a restrictive effect on the typology as such, as it reduces the
variational space, and also has something to say for the structural similarity
between dialects irrespective of tonal values.
Further support for the coherence of the typology comes from the
lexical distribution of the accents, which is very stable across dialects.
Accent 2 shows up with the same set of unstressed, posttonic suffixes (Sw -
ar, -or, -are, -ing, -nad, -lig, -ig, -a, among others, and corresponding
Norwegian suffixes), a fact that points at both a morphological anchoring of
lexical tone and a shared origin.
Another shared property is the fact that the postlexical generalization of
accent in compounds is always in the direction of accent 2. No dialect that
has a tonal contrast exhibits prosodically motivated assignment of accent 1
in compounds. Most dialects, however, exhibit some prosodic assignment
of accent 2, usually at least in the core set of stem compounds containing a
clash (South Swedish; Strandberg 2014) and in recently formed formal
compounds arising from initial stress insertion (protes1ˈtere > 2ˈprotesˌtere
‘to protest’; e.g., East Norwegian; Kristoffersen 2000: 165). Often, accent 2
is broadly generalized to any form containing two stresses (e.g., Central
Swedish, Dala, Göta, North Norwegian).135
Finally, the geographic contiguity of the tonal systems obviously points
at a common historical core. Although some developments have taken place
– yielding the variation we study as a typology here – there are no
indications of a tonal variety that is radically different from that of any
other dialect. With Danish stød, however, there is reason to believe that a
radical change has taken place, as some basic conditions are different there:
beside the different phonetic exponent, there is the sonority requirement in
stressed syllables (“stød basis”), the possibility of more than one stød in a
single compound, and the largely (but not completely) inverse distribution
of stød compared with accent 2.136 The historical affinity is not in question,
but neither is the typological distance.

5 The variables and the dialects

A number of variables instantiate the typological variation and in this
section we go through them. Recall that we shall look only at compounds
carrying big accent 2, since most of the basic properties of the tonal system
are contained in, and therefore derivable from, that type of form.
First, the value of the lexical tone: H or L. The first tone in accent 2 is
either lexical or postlexical (compounds). It is invariably the same tone
value in simplex forms as in compounds, and it also behaves consistently,
associating to the primary stressed syllable.
Second, the number of associations for accent 2 in compounds: two or
one. Dialects differ with respect to whether or not they admit a second
association beside that of the initial postlexical tone. If a secondary
association occurs (in a system with generalized accent 2 in compounds), it
is at the last stressed syllable of the compound. Dialects which only admit a
single association in compounds assign tonal accent the same way in
simplex as in compounds, i.e., they exhibit no “compound rule”.
Third, the shape of transitions: spreading or interpolation. In dialects
that admit two associations in compounds, the transition from the
(post)lexical tone to the prominence tone may have a characteristic profile
that suggests backwards spreading of the prominence tone, or else that there
is phonetic interpolation between the two. To some extent, there is also
some variation in the transition from the prominence tone to the boundary
tone.
There is a fourth potential variable that concerns the big accent (in focal
contexts), specifically whether the big accent is expressed by enhancement
of the small accent (southern varieties, Dala), or by the addition of a
separate focus gesture (Central Swedish, Göta) to the small accent (Bruce
2004). We will return to this potential variable later (Section 7), and suggest
that while it is phonetically reasonable, it does not constitute a relevant
phonological parameter. Rather it is dependent on, and derivable from,
more general properties of the melodic make-up of the tonal contour.
Let us first take a look at a map of Sweden and Norway, indicating the
general distribution of some of these variables. In the following subsections
we go through the variables one at a time and illustrate how they instantiate
typological variation by pairwise comparing dialects which differ minimally
on one or the other variable (Figure 10.1).

5.1 Value of the lexical tone: H or L

In (9) we have divided the melody of compounds in each dialect into three
segments, called “(post)lexical tone”, “prominence tone”, and “boundary
tone”.137
 Figure 10.1: Map of Scandinavia, with some important locations
indicated. Darker areas have an L*, and lighter areas an H*, as
(post)lexical tone. The isogloss (dotted line) indicates the boundary
between dialects that have two (east, north) or one (west, south)
association points in compounds.

As can be seen, the lexical tone varies between H* and L*. All tones in the
first column are marked with a star to indicate that the (post)lexical tone is
invariably associated. We shall now make the comparison between a CSw
variety (Stockholm) and a Dala variety (Norberg). These varieties differ
primarily with respect to the value of the lexical tone. Grammatically, they
are otherwise the same, i.e., with regard to tonal associations and spreading
pattern. We use the word sommarledigheten ‘the summer holidays’ as our
sample word. This word contains three stresses and will have accent 2 in all
dialects, either by virtue of a prosodic compound rule, or by virtue of the
first morpheme sommar ‘summer’ being lexically accent 2. In Figure 10.2,
autosegmental representations are provided for both dialects compared,
above and below the sample word, and also stylized contours.

Figure 10.2: Schematic comparison of CSw and Dala. These dialects

differ primarily on tonal value of the (post)lexical tone.

We look first at two examples from CSw. The previous panels that we have
looked at were all taken from this variety. The compounds in (10) and (11)
both contain three minimal prosodic words each, hence three stresses. The
(post)lexical tone associates to the first and the prominence tone associates
to the last.
These CSw contours should now be compared with Dala. The two contours
are from speakers from Norberg and Dala-Järna.
In the analysis of both of these dialects the prominence tone spreads back
from the last stressed syllable to the (post)lexical tone. In the CSw variety
this means a tonal floor from the last stress back to right after the
(post)lexical H*. In Dala varieties this means a high plateau from the
prominence H* in the last stress back to the (post)lexical L* on the primary
stress.
The value of the (post)lexical tone is readily identified from panels like
these, and we maintain that this tone is also a phonologically relevant
category, therefore useful for typological concerns. As we saw in (5), the
lexical tone is quite easy to isolate, by simple comparison of accent 1 and
accent 2 forms, where the accent 2 forms will contain an extra tone before
the intonational tones that are common to both accents.138
Let us look at another example of a minimal contrast on tonal value.
This time we compare East Norwegian (ENw) with South Swedish (SSw),
shown in Figure 10.3. These dialects have parallel association patterns, but
each tone contrasts. H and L tones do not fulfill functional purposes in
exactly the same way, there being a bias for H tones to serve as markers of
prominence. In Norwegian it has long been maintained that the boundary
tone also carries the function of focus (Fretheim & Nilsen 1989;
Kristoffersen 2000: 278).139 This does not affect the tonal grammar,
however, which is parallel. Panels exemplifying these dialects are given in
(14) and (15), below for ENw, and in (20) for SSw.

Figure 10.3: Schematic comparison of ENw and Skåne. These dialects

differ primarily regarding the tonal value of the (post)lexical tone.

5.2 Number of associations in compounds: Two

or one
The second column of the table in (9) indicates whether or not a dialect
admits a secondary association in compounds. The second tone (usually
carrying the prominence function) is star-marked if it associates to the last
stressed syllable of a compound. Geographically this property defines an
isogloss cutting through the Scandinavian peninsula, as marked in Figure 1
0.1. The independence of this variable is demonstrated by the fact that there
are dialects with H and L lexical tone on either side of the isogloss. The
dialects to the north and east of this isogloss are sometimes referred to as
“connective” dialects since compounds exhibit a particular prosodic pattern
which connects the beginning and end of the compounds. It is invariably the
case in these dialects that the prominence tone goes to the last stress. On the
other side of the isogloss compounds are partly assigned accent according
to the lexical specification (or its absence) in the first compound member.
The pattern is sometimes called “non-connective”. These dialects also often
exhibit a prosodic rule assigning accent 2, in the narrow context of word
initial stress clash, i.e., where the first morpheme is monosyllabic and
directly followed by the stress of the second element. There are also
morphological requirements that further constrain the application of this
prosodic rule (cf. Strandberg 2014 for South Swedish). It looks like stress
clash is the original context for postlexical accent 2, and it has been argued
that this is also the context in which accent 2 arises (Riad 1998a). The
connective pattern to the east and north of the isogloss should thereby be
understood as an innovation, spreading from Central Swedish dialects, and
perhaps independently in the very north (Strandberg 2014: 174).

Figure 10.4: Schematic comparison of ENw (Oslo) and Göta/WSw

(Göteborg). These dialects differ primarily on the secondary association
in compounds.

The dialects that we will first compare regarding a secondary

association are East Norwegian (Oslo) and West Swedish/Göta (Göteborg),
which have the same tonal make-up but which differ crucially on the issue
of a secondary association, as shown in Figure 10.4.
A secondary association occurs in WSw, which keeps the contour down
until the last stress. Thereafter the contour rises to the final H%, which is
aligned with the end of the word. In ENw, by contrast, there is no secondary
association and the tonal contour begins to rise directly after the H*L at the
primary stress. We begin with ENw, where the focus function is in fact
placed on the boundary H%, often referred to as the focal H (Fretheim &
Nilsen 1989; Kristoffersen 2000: 278).
The lowest point in the ENw contours is right after the initial H*. From
there there is a steady rise until the end. The WSw contours, by contrast,
have the lowest point at the last stress of the compound.
We shall have a look at another example of the same type of alternation
between dialects. This time we look at two dialects of the two-peaked type,
namely Central Swedish and Stavanger in southwest Norway, compared in
Figure 10.5.
The Stavanger variety has been investigated by Hognestad (2012) and
the panels below are based on his recordings. The intonation system of this
dialect has clearly begun to change in the last couple of generations.140 The
panels in (18) and (19) are taken from a (young) speaker who exhibits the
system of the older generation in careful, elicited speech.141 This is where
the contrast vis-à-vis Central Swedish is the clearest. The tones after the
initial H* have been separately coordinated with the following low and high
points.142
 Figure 10.5: Schematic comparison of SWNw (Stavanger) and CSw
(Stockholm). These dialects differ primarily on the secondary
association in compounds.

In this example we can clearly see how the second H tone occurs relatively
close to the left edge, certainly not at the last stress. It might look as if it
were associated to the second stress, but that is not in fact the case. This is
clear from the next example where the peak is in an unstressed syllable.
These contours should be compared with the ones given for Central
Swedish above, such as (7), (10), and (11), where a secondary association at
the rightmost stress is clearly in evidence in a dialect with the same tonal
sequence.
Our third comparison regarding tonal association involves Dala and
South Swedish (Skåne). These varieties have the same tonal make-up but
differ regarding secondary association, as represented in Figure 10.6.

Figure 10.6: Schematic comparison of SSw (Skåne) and Dala

(Norberg). These dialects differ primarily regarding the secondary
association in compounds.

We have already seen examples of Dala (items (12) and (13), above). From
Skåne we have the following utterance, which contains two compounds
with the accent 2 contour.

The tonal accent 2 contour is L*HL, where the (post)lexical tone is L*. We
have registered the peak and the immediate drop as an HL prominence tone.
The status of the L tonal segment is somewhat unclear, e.g., whether it is
part of the prominence tone, or if it represents some kind of default. Further
research on SSw varieties is needed to clarify the issue.

5.3 Consequences of the association patterns

The dialects that allow only a single association in compounds assign tonal
structure in the same way in compounds as in simplex forms. While there
may be a limited prosodic rule that applies in compounds (and compound-
like structures), there will not be a representational difference in terms of
tonal association in these forms. This structural fact allows us to (correctly)
predict that the single-association (non-connective) dialects should admit
both accents in compounds, as a matter of principle. Below are panels that
contain accent 1 compounds from the dialects we have looked at that have a
single association point: ENw, SWNw, and SSw.
For Skåne we take a cut-out from the previous example (20). In this dialect
accent 1 is H*L, which amounts to the prominence tone.
The dialects that require two association points in compounds always have
a realization of accent 2 in this type of form. Thus, no dialect has
generalized accent 1 in compounds (while still maintaining a distinction in
simplex forms). This fact should ideally follow from representation, and we
would propose that it is the combination of which tonal segments there are
and the identification of the relevant TBU that actually predicts the pattern.
If the dialect allows association to more than one TBU, then there must also
be more than one tone that could associate. In the privative analysis of the
distinction, only accent 2 has more than one tonal segment (the (post)lexical
tone and the prominence tone), and so we can predict that there could be no
such thing as a connective realization of accent 1.143
Furthermore, the connective accent 2 pattern is not limited to the
morphological category of compounds. Rather, the generalization concerns
forms that contain more than one stress. This obviously complicates the
notions of “simplex” and “compound” such as we have used them up to
now. The “compound rule” has a purely prosodic basis, rather than a
morphological one. The tonal representation results from the presence of
two stresses (not from the presence of a morphological object as such).
Thus, when we say that there is a “compound rule”, it is important to note
that this is in fact a statement about phonological representation, triggered
by prosodic information. The same thing goes for the notion of “simplex”,
again normally a morphological notion. The typical prosodic shape of
simplex forms will be that of a minimal word, with one stress and a single
syllabification domain (Riad 2014). But there are also morphological
simplex forms that are prosodically like compounds. The three types of
forms that are targeted by the compound rule, which assigns postlexical
accent 2, are listed in (24).

(24) Targets of the “compound rule”, CSw

Proper 2ˈsommar ˌdag ‘summer day’, ba2ˈnanˌskal ‘banana
compounds: peel’, 2ˈunderˌsökningsˌdomare ‘examining
magistrate’
Formal 2ˈävenˌtyr ‘adventure’, 2ˈparaˌdis ‘paradise’, 2ˈarˌbete
compounds: ‘work’, 2ˈgeniˌtiv ‘genitive’
(Some) 2ˈsjukˌdom ‘illness’, 2ˈkraftˌfull ‘forceful’, 2ˈunder
derivations: ˌbar ‘wonderful’

Derivations form a sort of middle category between compounds and typical

simplex forms. Diachronically, this is not unreasonable since derivations
typically stem from compounds in Germanic languages. Synchronically, the
situation is that many derivational suffixes are unstressed. Nearly all of
them induce accent 2 lexically. Several derivational suffixes are (still)
stressed, however, like those exemplified in (24). These suffixes induce
accent 2 via the postlexical, “compound rule”. The data in (24) point very
clearly in the direction of structural properties determining both the fact that
only accent 2 could occur with double associations, and that it is the
number of TBU’s that determine the occurrence of a connective accent
pattern, not the morphological category of the form.

5.4 Shape of transitions 1: Spreading or

interpolation
We turn now to the third parameter of variation, which is whether dialects
have a phonological spreading rule from the prominence tone back to the
(post)lexical tone, or whether phonetic interpolation takes place. Dialects of
northern Sweden are traditionally considered to be prosodically closely
related to those of CSw, since they have the same tonal make-up with a
two-peaked realization of accent 2. In simplex forms and short compounds,
where the stresses are adjacent, the similarity is obvious (Bruce 2007: 135).
There are some contrasts, however, one of which is revealed when we look
at longer compounds, as in Figure 10.7, where the transition from the
postlexical tone to the prominence tone can be studied (Riad 2006; Bruce
2007).
In many varieties of North Swedish the fall from the postlexical H* to
the following L is much flatter than the corresponding fall found in CSw,
where it is steep (cf. (7) above). The effect of this dynamic is that the initial
H* in NSw tends to sound less prominent than the prominence L*H at the
end of the compound. Indeed, the phenomenon is usually described as stress
on the last member of compound (Bruce 1982). We would maintain that it is
really just the perception of the tonal contour that gives this impression
(Riad 2006). The phonological interpretation of the contrast that we propose
is that CSw exhibits backward spreading of the low tone of the L*H
prominence tone, creating a floor between the two anchor points, while
NSw lacks the spreading rule and instead exhibits phonetic interpolation, as
a gradual fall. This analysis is in line with the segmentation of the contour
into two elements, H* and L*H. The pattern in Figure 10.7 is in evidence
only in long compounds, with minimally one syllable between the
association points. Two examples are given below. They should be
contrasted with the CSw items (6), (7), (10) and (11) above.

Figure 10.7: Schematic comparison of CSw (Stockholm) and NSw

(Luleå). These dialects have the same tonal composition and
association patterns, but differ in the transition between the postlexical
tone and the prominence tone.
The same contrast is also found with dialects that have a (post)lexical L* in
accent 2. The contrast is then between Dala, which exhibits a high plateau
by spreading, and East Färnebo-north, which has a rising contour, shown in
Figure 10.8.
 Figure 10.8: Schematic comparison of Dala (Norberg) and East
Färnebo-north. These dialects have the same tonal composition and
association patterns, but differ regarding the transition between the
postlexical tone and the prominence tone.

East Färnebo is described by Kallstenius (1902). The area is on the border

between Dala and Göta varieties (cf. Figure 10.1), where the northern part
is more like Dala and the southern is more like Göta, i.e., WSw. We have
two examples of compounds from the SweDia 2000 materials (Bruce,
Engstrand, & Eriksson 1998), from Gåsborn in the Färnebo area.
Unfortunately, they both have adjacent stresses, and are therefore not ideal
for showing the rising slope.
These panels should be compared with the Dala items that we looked at in
(12) and (13), above. Closer study of the East Färnebo type is required to
establish the structure more firmly.

6 Transitions and dialect change

We shall now look at a couple of other contrasts that provide key
information on diachronic relationships between dialects. We will not make
a full diachronic argument here, but rather point to variations in tonal
alignment that make sense in terms of transitions between tonal varieties.

6.1 Floating H in Central and West Swedish

The first contrast is not obviously structural and concerns the timing of an
unassociated tonal segment in otherwise very similar dialects. This type of
phonetic variation is likely to be important for changes in the phonological
structure. We look at the contrast between CSw and two varieties of WSw
which we will call Göta 1 and Göta 2. It is well-known that the
unassociated H of the prominence tone L*H in CSw may occur at a variable
distance from the associated L* (Bruce 1987). Things like the size of the
word carrying accent and the sentential context surely have an influence,
but the matter has not been fully investigated. However, it is clear that the
timing varies between dialects, too. A comparison of items (6) and (7)
illustrate the variation within CSw, where talespersonerna ‘the
spokespersons’ in (6) is phrase medial (late H), and
uppmärksamhetssplittring ‘attention split’ in (7) is phrase final (earlier H).
In the Göta dialects the timing of the trailing H is routinely later, in all
contexts.144 This is seen in items (16) and (17), which are exponents of
Göta 2. In these forms, the boundary tone appears to be H%, just like in
ENw (items (14) and (15)). But there are also pronunciations in the Göta
region where the final boundary tone is L% just as in closely related CSw.
This form we shall call Göta 1 in contrast with Göta 2 illustrated in (16) and
(17) above. The comparison we are considering is given in Figure 10.9.
Here is an example of an emphatic pronunciation in what we call Göta
1, however without making claims for this being a separate dialect type,
rather than an alterative pronunciation in one and the same variety.

Comparing with CSw item (7) and Göta 2 item (17) above, we see clearly
that the floating H of L*H is timed later in Göta 1 compared with CSw, but
earlier than in Göta 2, where the H looks like it is also the boundary tone.
The invited conclusion to draw from this is that these variant realizations
represent possible developmental stages. Moving one dialect to the west of
WSw, i.e. into ENw, the typological difference is simply the absence of a
secondary association, as described in Section 5.3.

Figure 10.9: Schematic comparison of CSw (Stockholm) and Göta 1

(Göteborg). These dialects have the same tonal composition and
association patterns, but differ regarding the transition between the
associated and floating tonal segments of the prominence tone.

6.2 Boundary L% and displacement: Stockholm

and Eskilstuna
We will touch briefly on another difference regarding tonal transition
occurring within the CSw area where it borders on the Dala area. The issue
here is the fall from the prominence L*H tone to the boundary L%. In the
Stockholm area, this fall is relatively smooth and adjusted to the space
available. If space is limited, it is not always the case that an L target is
reached before the end of the phrase. In the Eskilstuna area, to the west of
Stockholm, by contrast, the L target of the boundary tone is reached with
much greater regularity. Indeed, the boundary tone is often pulled into the
last or even penultimate syllable. This gives rise to a phenomenon called the
Eskilstuna curl (Bleckert 1997; Riad 2000a, 2000b, 2009b), which involves
a sharp tonal drop followed by creaky voice or a full glottal stop
(phonetically much like Danish stød). This prosodic event is accompanied
by a characteristic diphthongization. The contrast we shall look at is
depicted in Figure 10.10.
An example of a contour from Eskilstuna-east is given in (30), where
the tonal contour is HLHL just as in Stockholm, but where the alignment of
tonal segments at the end of the contour is partly different. The panels in
(30) and (31) are taken from Nordberg’s recordings in Eskilstuna in the
1960s (Nordberg 1970, 1972, 1985).
 Figure 10.10: Schematic comparison of CSw (Stockholm) and
Eskilstuna-east (CSw). These dialects have the same tonal
composition, but differ regarding the association of the prominence tone
and the realization pattern for the boundary L%.

The (post)lexical tone is in the primary stressed syllable, but the following
L occurs before the last stressed syllable, in contrast with Stockholm where
it is normally associated in the stressed syllable. Instead, the H tone of LH
is associated in Eskilstuna. The reason for this behaviour is to be found at
the very end of the contour, where the L% boundary tone is pulled firmly
into the stressed syllable, leaving little room for all of the prominence tone
L*H. In this situation, the prominence contour is pushed back, the H
segment associating (H*) while the preceding L becomes a leading tone to
H*. The sharp fall from H* to L% in the last stressed syllable causes
phonetic stød in this case, as can in fact be seen in the contour in (30).
Many other recordings exhibit creaky voice in the corresponding place. At
the segmental level there is some centralizing diphthongization (Bleckert
1997).
 The association of an H* at the last stress of compounds is in fact a step
toward the system found in Dala (cf. Section 5.1). And a further step in that
direction is found in the northwestern side of the city of Eskilstuna.145 In
this variety, which we will call Eskilstuna-west, compounds no longer
exhibit an H* (post) lexical tone on the first stressed syllable. Instead, the
(post)lexical tone is now L*. This contrast is schematically illustrated in Fig
ure 10.11, still in comparison with CSw, though a comparison with Dala
would be equally warranted.

Figure 10.11: Schematic comparison of CSw (Stockholm) and

Eskilstuna-west (CSw/Dala). Eskilstuna-west has an L* (post)lexical
tone and is hence closer to Dala than to CSw.

The proposed analysis here is that the leading L of LH* spreads

backwards to the initial stress, where it gets reinterpreted as the (new)
lexical tone. Viewing this as a phonological change, the context for
reinterpretation is likely to be simplex forms where the space is limited,
especially in accent 1. In accent 1 monosyllables (or final stressed syllables)
L*HL% would change into H*L% as an effect of the boundary tone target
within the syllable. The parallel change in accent 2 simplex would entail a
change from H*LHL% to L*HL%. The tonal structures of accents 1 and 2
are never more than one tone off from each other, in any dialect, which
means that phonological changes in timing never happen in one condition
only.
Here is now an example of a compound from Eskilstuna-west,
illustrating the situation depicted in Figure 10.11. Notice that the lexical
tone is L* and that there is a rising tone within the final stressed syllable,
indicating the spreading behaviour of the leading L tone of the LH*
prominence tone.146
The boundary L% tone is realized in creaky voice following the sharp fall
from LH* to L% (therefore not visible in (31)). The next dialect over will
be of the Dala type that we saw in (12) and (13), i.e., with an immediate rise
after the initial (post) lexical L*. The Dala variety thus shows that the
prominence tone is now reduced to just the H*, which spreads backwards
until the (post)lexical L*.
Note that the grammar of these systems does not really change: the
associations remain the same. It is only the tonal values that change,
apparently as a function of a leftward shift of tones, in turn due to the curl
behaviour of the boundary L%.
There are other changes related to Eskilstuna-curl in the area, which
provide further clues also for the reconstruction of how stød might have
become phonologized and the markedness patterns reversed between a tonal
system of the Central Swedish and a stød system like that found in Danish
Sjælland (Riad 2000a, 2000b, 2009b).

6.3 Observations
The spreading and interpolation behaviours provide evidence of alignment
tendencies to the left and to the right (Riad 1998b). Any dialect that has two
association points in compounds, and which hence has accent 2 as a rule in
compounds and other forms with two stresses, show association to the first
and the last stressed morpheme. This should be interpreted as a solution to
the double desiderata on the part of tones. Leftward alignment is primary
and all dialects have the main stress as an obligatory association point,
whereas rightward alignment comes to the surface only when a secondary
association point is sought out. Given a secondary association, dialects may
then differ in how the transition between the two association points is
realized. Backwards spreading is a sign of the grammaticalized “desire” of
the prominence tone to be at the left edge, simultaneously with being
associated at the right edge. This leads to a tonal floor (Central Swedish,
Göta) or a plateau (Dala, Narvik). Interpolation is the case when no
backwards spreading takes place, a fact that could be interpreted as the
action of a constraint against (the markedness of) spreading (Riad 1998b).
In dialects that allow only one association point, it is invariably the first
TBU which receives a tonal mark. In this way, we can account for the fact
that there is no systematic association to any medial stresses in long
compounds.

7 The shape of the big accent

The shape of the big accent has been subjected to typologization in work by
Bruce (e.g., 2005), concerning the dialects of Sweden. The general
observation is that the big accent is taken to be an enhanced version of the
small accent contour in the dialects of Dala and South Swedish, whereas in
Central Swedish and the Göta dialects (WSw) the big accent involves a
separate tonal gesture, which is added linearly to the small accent contour.
A diagram depicting this for accent 2 is given in Figure 10.12, cited from
Bruce (2005: 421).
As can be seen, the SSw and Dala varieties have the same shape for
both the big and the small accent, where the big accent peaks are simply
higher than the small accent peaks in the same position. The CSw and Göta
varieties, by contrast, exhibit differences in the overall shape of the
contours for small and big accents. This is very clear in CSw where the
separate tonal gesture of the big accent is in evidence in both positions,
respectively, forming a plateau in 1st focus and a narrower peak in 2nd
focus. In Göta the separate peak of 1st focus merges with the following
lexical peak, but in 2nd focus there is a clearly separate tonal gesture.
 This might look as if there were a simple typological difference between
varieties that have a separate tonal gesture and those that do not, in their
respective big accents, compared to their small accents. But this contrast
can in fact be derived from a single generalization that concerns all dialect
types (and both accents). Notice first that the dialects that have a separate
tonal gesture in their big accent 2 also have a lexical H* tone, while those
that exhibit enhancement only have a lexical L* tone. This property
generalizes to other dialects that are not included in the diagram.
The broader generalization that we can formulate based on this
observation is that the big accent requires an H tone, and that a (post)lexical
H* cannot serve this task. Just in case a dialect has the tone value H for the
(post)lexical tone, another H must be available to get a proper prominence
tone in place. This is done either by adding a secondary peak as part of the
prominence tone (L*H, CSw, Göta 1), or by recruiting an H% boundary
tone for the purposes of prominence (ENw, Göta 2). If the (post)lexical tone
is L*, all is good, as the prominence tone will then predictably be H (by
OCP). It would thus appear to be a general property of all North Germanic
dialects that an H tone must be available somewhere after the (post) lexical
tone for the instantiation of a big accent. This is hardly a surprising fact, as
peaks typically are associated with prominence, but it also allows us to
derive the typological difference in phonological shape of the big accent
from one single source. To see this consider Figure 10.13, where the
dialects that we have looked at are depicted. Some varieties which we have
not looked at specifically but which share properties with other dialects
have been listed together in types.
 Figure 10.12: Accent realization for four prosodic dialect types. The
pitch contours schematically represent a phrase consisting of two
simplex accent 2 words, with narrow focus on either the first (1st focus,
big accent) or second (2nd focus, big accent) word. Arrows indicate the
CV-boundary of stressed syllables. An example of a possible text here
is långa nunnor, vs. långa nunnor ‘tall nuns’.

The five types in the upper half of Figure 10.13 have a (post)lexical H*
tone and the three types in the lower half have a (post)lexical L*. In all
varieties, the remaining tones of the big accent include an H tone. A good
way of illustrating the generality of this H tone is to look at accent 1. Bruce
(2005) does not discuss the presence of a separate focus gesture in accent 1,
but there is an indirect reference to it in Bruce (2003). Clearly though, the
null hypothesis must be that the focus gesture generalizes across big accent
in both accent 1 and accent 2. In Figure 10.14 the (post)lexical part of the
big accent 2 contour has been shaded over, leaving the big accent 1 contour
unshaded. As can be readily seen, each dialect has a remaining H tone that
can be engaged for prominence purposes.
Figure 10.13: Big accent 2 in long compounds in schematic
representation.
 Figure 10.14: Removing the (post)lexical tone, leaving the contour of
big accent 1. All instances of accent 1 contain an H tone.

The inclusion of accent 1 in the comparison explains why the

(post)lexical H*, part only of accent 2, is not available for the expression of
prominence. The requirement of an H tone for prominence purposes is a
condition on intonation generally. If all dialects have the same type of
privative accent distinction, as per our analysis above, then it follows that
(post)lexical H* should not be available for intonational prominence
purposes.147
East Norwegian and Göta 2 employ the boundary H% tone as part of
their big accent. This fact also supports the superordinate status of the H
tone required for prominence purposes within the big accent. The notion of
a “separate tonal gesture” should thus not be interpreted as the necessary
ADDITION of a gesture. What is already there can also be employed for this
purpose.
The asymmetry between dialects concerning the presence or absence of
a separate tonal gesture is visible also in another fact. The tonal contours for
the big accent 2 that we have seen contain either three or four tones. Given
the variable tonal values, this would predict four possible contours under
OCP.

However, we only find four alternating tones if the first tone, the
(post)lexical tone, is H. The absence of LHLH now follows from the
generalization that an H tone must be regularly available for prominence
purposes in the big accent. In fact it narrows this generalization to “exactly
one H tone”.
Our conclusion is that “separate tonal gesture” vs. “enhancement” are
surface observations that depend on other things, rather than proper
typological features. A prediction for diachrony would be that we should
not expect to see tonal behaviours that appear to preserve the separate tonal
gesture as such (e.g., in the transition from CSw to Dala, via Eskilstuna).
But we should expect to see preservation of an H tone in the big accent
contour. As shown by Myrberg (2013, 2016), enhancement is related to
function and information structural status in Central Swedish. This is likely
to generalize to many, perhaps all, dialects, and would thereby not be a
structural typological property either.

8 North Germanic and Central Franconian:

Comparative notes
The typology of the North Germanic tonal accents is strikingly strict,
employing just a few variables, which can be identified and demarcated by
studying structural properties within a dialect, and by looking at comparable
structures across dialects. The strictness found could possibly be due to
some element of simplicity in the model of analysis, or to our current lack
of detailed knowledge of intonation in several of the dialects. At the same
time, however, there is a robustness both to intuitions regarding which
dialects are similar and which are different, and to the tonal behaviours that
we have looked at, which are all readily manifest on the surface, such as
tonal values and the presence of a secondary association. At this junction,
then, it seems fair to assume that there are fundamental, shared properties
that make the tonal typology relatively coherent.
Another Germanic tonal system which is by now rather well-studied is
Central Franconian (Schmidt 1986; de Vaan 1999; Gussenhoven & van der
Vliet 1999; Gussenhoven 2000, 2007, 2012; Peters 2006, 2007; Köhnlein
2011, 2016). The Central Franconian accentual contrast is often mentioned
as typologically similar to the North Germanic one. Here, we shall make
some comparative notes that profile some differences between them.
One fundamental property of the NGmc accents is the dependence on
morphology, which keeps the lexical distribution rather constant between
dialects. The stable classes of words that occur with accent 2 in all dialects
contain posttonic suffixes (2ˈläk-are ‘doctor’, 2ˈfäng-else ‘prison’, etc.).
Such words also have higher token frequencies, compared with
monomorphemes with accent 2, which tend to vary in accent between
dialects, e.g., 1/2ˈsenap ‘mustard’, 1/2ˈAnders (name) (Bruce 1998: 50). Our
analysis of this situation is that it is the posttonic suffixes that carry the
tonal information (hence -are2, -else2, etc.), which is assigned and realized
in the preceding stressed syllable (Riad 2014). This connects the patterning
of accent 2 to morphology and also accounts for the often observed fact that
accent 2 only occurs in polysyllabic words.149 It also harmonizes with the
greater tonal richness of accent 2 compared with accent 1, see (5).
In Central Franconian, by contrast, the lexical tone contrast is fully
contained within the stressed syllable, such that the contrast can be
expressed in a pair of monosyllabic morphemes, e.g., 1man ‘basket’ vs.
2man ‘man’ (dialect of Mayen; Schmidt 1986). The situation with accent
being assigned from a suffix to the root therefore does not occur in
Franconian. Still, morphological factors do influence the contours. The
marked accent tends to occur in morphologically more complex forms than
the unmarked accent, often as the diachronic result of apocope, e.g.,
unmarked 2stain ‘stone, sg’ vs. marked 1stain ‘stone, pl’ (dialect of
Arzbach; Köhnlein 2011).
Importantly, the lexical distribution of the tonal contrast in Franconian
varies between dialects based on segmental factors, concerning both vowel
type in the stressed syllable and local environment (e.g., de Vaan 1999:
26ff.; Köhnlein 2011: 219ff.). This segmental dependence represents a
major constitutional difference between NGmc tone (discounting Danish
stød), where there is no segmental dependence whatsoever, and Central
Franconian. This difference is clearly of importance for the understanding
of typology as well as tonogenesis.
A second fundamental property relates to the TBU. In NGmc, the TBU
is the stressed syllable, a typical situation in languages with accents of the
North Germanic kind (Gussenhoven 2007), that is, languages where a tonal
contrast is only ever found in stressed syllables. Also, the normal situation
(in any tonal system) is that only one tone is associated to the TBU, a fact
that is strictly obeyed in most tonal dialects of NGmc. This puts a limit on
the possible tonal variation within stressed syllables. If one assumes a less
typical TBU, such as the syllable (stressed or unstressed) or the mora, then
that will admit a more fine-grained typology, but it is unlikely to refute the
generalizations found for NGmc. Rather, such an assumption is likely to
overgenerate and predict the existence of dialects that may never be found.
We have thus taken a conservative view from which it may be possible to
diversify (given clear evidence), rather than starting from an overly rich
structure, from which it may be hard to retreat.
In Central Franconian, the TBU is either the mora in stressed syllables,
as roughly in the tonal dialects in the Netherlands, or the stressed syllable,
as roughly in the tonal dialects in Belgium (Gussenhoven 2007; Peters
2007). This distinction of TBU defines a rather clear isogloss. With a TBU
the size of a mora, an association system is going to have more
phonological options within the stressed syllable, in principle, for tones
stemming from the lexical contrast or from intonation (e.g., as in the dialect
of Helden; Gussenhoven & van den Beuken 2012).
A third property, related to the above, is the absence of a sonority
restriction on syllables that carry tone in the NGmc system. Accent 2 can be
assigned in a word like 2ˈpappa ‘dad’ just as well as in 2ˈmamma ‘mom’,
and accent 1 works as well in 1ˈtaxi as in 1ˈmammon. In NGmc, the absence
of a sonority restriction seems to be part and parcel of the space
requirement for accent 2, which in turn seems to be one reason why the
dialect variation is so regimented.
The Central Franconian dialects that have the stressed syllable as TBU
are similar to NGmc regarding the lack of a sonority restriction. However,
unlike NGmc, these dialects admit a contrast within a single syllable,
regardless of the sonority of the segments in that syllable. An example of
this is the dialect of Hasselt (Peters 2006). In such dialects, the tonal
contrast can thus be realized in more types of words than in dialects with a
sonority restriction in the stressed syllable. In the latter type of dialect a
consonant must be sonorant in order for a tonal contrast to be realized. The
mora-based varieties of Central Franconian are of this type (e.g., the dialect
of Venlo; Gussenhoven & van der Vliet 1999), and so is Lithuanian, and
also the Danish stød system, which has developed from an earlier tonal
system. Such systems are analysed as having the (sonorant) mora as TBU,
though the requirement of a stressed syllable remains.
Furthermore, in systems of the Lithuanian and Central Franconian type
there is a stronger inclination for there to be durational and/ or segmental
effects that covary with tonal constellations. A tonal contrast may be
(partly) dependent on a segmental property (Franconian) and vice versa
(Lithuanian). In some such dialects, the primacy of tone over and above
vowel quality/ quantity is harder to ascertain and is sometimes questioned.
In at least some varieties of East Aukštaitian Lithuanian, the tonal
distinction has been replaced by one relating to vowel duration and quality
(Hualde & Riad 2014).150
A fourth fundamental property that likely has a constraining effect on
the variation within the NGmc tone system is the stability of realization of
tones across syntactic and pragmatic contexts. This is to say that an instance
of, say, big accent 2 will have basically the same shape in declaratives,
interrogatives, imperatives, etc. For example, there is no dedicated
interrogative tonal contour in Swedish, although an optional rise may occur
(House 2002, 2004). Although this is an understudied area for many NGmc
dialects, the typical situation is that accents are recognized as similar in
various intonational contexts.
This is not the situation in Central Franconian where the realization of
tonal contours often varies a lot between different intonational contexts,
within the same dialect (Gussenhoven 2000, 2012; Köhnlein 2011). To
some extent, this may be due to the intonational tones themselves being
more varied in Franconian compared with NGmc. But the interaction of
tone with intonation in Franconian must also be handled in a more flexible
tonal grammar, given that even the ordering between a lexical tone and an
intonational tone can vary between contexts (dialect of Roermond;
Gussenhoven 2000). Such phenomena have not been reported for Swedish
and Norwegian.
9 Conclusion
The presentation of data and the discussion in this chapter has aimed at
laying bare a number of properties that are important in any account of the
typology of North Germanic accent. The contention is that the relevant
level is phonological, pertaining to representation and grammatical
behaviour of phonological units, in this case tones. The basic shapes of big
accents within the typology are definable in just these terms: value of
(post)lexical tone, secondary association (or not), and tonal spreading (or
not). Functional properties like information structural concerns will
influence the realization of the big accent (e.g., scaling, local alignment),
but not define it. Phonetic variation will be important for identifying minor
variation within a type (e.g., timing of floating H, sharpness of falls), and
this will play a role in accounting for dialect transitions, i.e., diachronic
change. But we cannot rely on phonetic variation for the systematic
typology, as it does not properly constrain it.

Acknowledgements: Feedback from Sara Myrberg and Jan Hognestad

during the preparation of this paper is gratefully acknowledged.

Sources of the data here analysed are identified by these abbreviations: BG

= Bert Göteborg; JA = Jon Almaas; JH = Jonas Hallberg; LG = Lotta
Gröning; LGA = Lars-Gunnar Andersson; NG = Niklas Göteborg; NN =
unknown speaker; SG = Sverker Göransson; OLVI and FAØV = Stavanger
informants from Hognestad (2012); nrk = Norsk Rundkringkasting
(Norwegian Radio and TV); om = older male (SweDia); sr = Sveriges
Radio (Swedish Radio and TV).

References
Arvaniti, Amalia. 2002. The intonation of yes-no questions in Greek. In M. Makri-
Tsilipakou (ed.), Selected papers on theoretical and applied linguistics, 71–83.
Thessaloniki, Department of Theoretical and Applied Linguistics, School of
English, Aristotle University.
Basbøll, Hans. 2005. The phonology of Danish. Oxford: Oxford University Press.
Bleckert, Lars. 1987. Centralsvensk diftongering som satsfonetiskt problem
(Skrifter utgivna av institutionen för nordiska språk vid Uppsala universitet 21).
 Uppsala.
Bruce, Gösta. 1977. Swedish word accents in sentence perspective (Travaux de
l’institut de linguistique de Lund 12). Lund: CWK Gleerup.
Bruce, Gösta. 1982. Reglerna för slutledsbetoning i sammansatta ord i
nordsvenskan. In Claes-Christian Elert & Sigurd Fries (eds.), Nordsvenska,
123–148. Umeå University.
Bruce, Gösta. 1987. How floating is focal accent? In Kirsten Gregersen & Hans
Basbøll (eds.), Nordic prosody 4, 41–49. Odense: Odense University Press.
Bruce, Gösta. 1998. Allmän och svensk prosodi (Praktisk Lingvistik 16). Lund
University.
Bruce, Gösta. 2003. Late pitch peaks in West Swedish. Proceedings of ICPhS 15,
Barcelona, 245–248.
Bruce, Gösta. 2004. An intonational typology of Swedish. Speech Prosody 2004,
Nara, Japan, 175–178.
Bruce, Gösta. 2005. Intonational prominence in Swedish revisited. In Sun-Ah Jun
(ed.), Prosodic typology: The phonology of intonation and phrasing, 410–429.
Oxford: Oxford University Press.
Bruce, Gösta. 2007. Components of a prosodic typology of Swedish intonation. In
Riad & Gussenhoven (eds.) 2007, 113–146.
Bruce, Gösta. 2010. Vår fonetiska geografi: Om svenskans accenter, melodi och
uttal. Lund: Studentlitteratur.
Bruce, Gösta & Eva Gårding. 1978. A prosodic typology for Swedish dialects. In
Gårding et al. (eds.), 219–228.
Bruce, Gösta, Olle Engstrand, & Anders Eriksson. 1998. De svenska dialekternas
fonetik och fonologi år 2000 (SweDia 2000) – en projektbeskrivning.
Folkmålsstudier 39. 33–54.
Bye, Patrik. 2004. Evolutionary typology and Scandinavian pitch accent.
Manuscript, University of Tromsø.
Dalton, Martha & Ailbhe Ní Chasaide. 2007. Melodic alignment and micro-dialect
variation in Connemara Irish. In Riad & Gussenhoven (eds.), 293–316.
Danell, Gideon. 1937. Svensk ljudlära. 4th edition. Stockholm: Svenska
bokförlaget, Norstedt & söner.
Elert, Claes-Christian. 1972. Tonality in Swedish: Rules and a list of minimal pairs.
In Evelyn S. Firchow, Kaaren Grimstad, Nils Hasselmo, & Wayne O’Neil
(eds.), Studies for Einar Haugen, 151–173. The Hague & Paris: Mouton.
Engstrand, Olle. 1995. Phonetic interpretation of the word accent contrast in
Swedish. Phonetica 52. 171–179.
Engstrand, Olle. 1997. Phonetic interpretation of the word accent contrast in
Swedish: Evidence from spontaneous speech. Phonetica 54. 61–75.
Fant, Gunnar & Anita Kruckenberg. 2008. Multi-level analysis and synthesis of
prosody with applications to Swedish. Manuscript, Kungliga Tekniska
Högskolan.
Fintoft, Knut. 1970. Acoustical analysis and perception of tonemes in some
Norwegian dialects. Oslo: Universitetsforlaget.
Fintoft, Knut., P. E. Mjaavatn, E. Møllergård, & B. Ulseth. 1978. Toneme patterns in
Norwegian dialects. In Gårding et al. (eds.), 197–206.
Fretheim, Thorstein & Randi Alice Nilsen. 1989. Terminal rise and rise-fall tunes in
East Norwegian intonation. Nordic Journal of Linguistics 12. 155–182.
Gårding, Eva. 1977. The Scandinavian word accents (Travaux de l’institut de
linguistique de Lund 11). Lund: CWK Gleerup.
Gårding, Eva & Per Lindblad. 1973. Constancy and variation in Swedish word
accent patterns. Working Papers 7. 36–110. Dept. of Linguistics, Lund
University.
Gårding, Eva, Gösta Bruce, & Robert Bannert (eds.). 1978. Nordic prosody:
Papers from a symposium (Travaux de l‘institut de linguistique de Lund 13).
Lund University.
Grice, Martine, D. Robert Ladd, & Amalia Arvaniti. 2000. On the place of phrase
accents in intonational phonology. Phonology 17. 143–185.
Gussenhoven, Carlos. 2000. The lexical tone contrast of Roermond Dutch in
Optimality Theory. In Merle Horne (ed.), Prosody: Theory and experiment,
129–167. Dordrecht: Kluwer.
Gussenhoven, Carlos. 2007. Intonation. In Paul de Lacy (ed.), The Cambridge
handbook of phonology, 253–280. Cambridge: Cambridge University Press.
Gussenhoven, Carlos. 2012. Asymmetries in the intonation system of Maastricht
Limburgish. Phonology 29. 39–79.
Gussenhoven, Carlos & Peter van der Vliet. 1999. The phonology of tone and
intonation in the Dutch dialect of Venlo. Journal of Linguistics 35. 99–135.
Gussenhoven, Carlos & Frank van den Beuken. 2012. Contrasting the high rise
and the low rise intonations in a dialect with the Central Franconian tone. The
Linguistic Review 29. 75–107.
Hognestad, Jan K. 2012. Tonelagsvariasjon i norsk: Synkrone og diakrone
aspekter, med særlig fokus på vestnorsk. PhD dissertation, University of
Agder.
House, David. 2002. Intonational and visual cues in the perception of interrogative
mode in Swedish. Proceedings of ICSLP 2002, 1957–1960. Denver,
Colorado.
House, David. 2004. Final rises and Swedish question intonation. Proceedings of
Fonetik 2004, 56–59. Stockholm University.
Hualde, José Ignacio & Tomas Riad. 2014. Word accent and intonation in Baltic. In
N. Campbell, D. Gibbon, & D. Hirst (eds.), Speech Prosody 7, 669–671.
Dublin.
Kallstenius, Gottfrid. 1902. Värmländska Bärgslagsmålets ljudlära. Stockholm:
Norstedt & Söner.
Kock, Axel. 1878. Språkhistoriska undersökningar om svensk akcent. Lund:
Gleerup.
Köhnlein, Björn. 2011. Rule reversal revisited: Synchrony and diachrony of tone
and prosodic structure in the Franconian dialect of Arzbach. PhD dissertation,
University of Leiden.
Köhnlein, Björn. 2016. Contrastive foot structure in Franconian tone-accent
dialects. Phonology 33. 87–123.
Kristoffersen, Gjert. 2000. The phonology of Norwegian. Oxford: Oxford University
Press.
Kristoffersen, Gjert. 2007. Dialect variation in East Norwegian tone. In Riad &
Gussenhoven (eds.), 91–111.
Lahiri, Aditi, Allison Wetterlin, & Elisabet Jönsson-Steiner. 2005. Lexical
specification of tone in North Germanic. Nordic Journal of Linguistics 28. 61–
96.
Leira, Vigleik. 1998. Tonempar i bokmål. Norskrift 95. 49–86.
Lorentz, Ove. 1995. Tonal prominence and alignment. Phonology at Santa Cruz 4.
39–56.
Lorentz, Ove. 2008. Tonelagsbasis i norsk. Maal og Minne 1. 50–68.
Malmberg, Bertil. 1970. Lärobok i fonetik. Lund: Gleerup.
Meyer, Ernst A. 1937. Die Intonation im Schwedischen I: Die Sveamundarten.
Helsingfors: Fritzes bokförlags AB. Mercators tryckeri.
Meyer, Ernst A. 1954. Die Intonation im Schwedischen II: Die norrländischen
Mundarten (Stockholm Studies in Scandinavian Philology 11). Uppsala:
Almqvist & Wiksell.
Mjaavatn, Per Egil. 1978. Isoglosses of toneme categories compared with
isoglosses of traditional dialect geography. In Gårding et al. (eds.), 207–216.
Myrberg, Sara. 2010. The intonational phonology of Stockholm Swedish
(Stockholm Studies in Scandinavian Philology 53). Stockholm University.
Myrberg, Sara. 2013. Focus type effects on focal accents and boundary tones.
Proceedings of Fonetik 2013, 53–56. Linköping University.
Myrberg, Sara. 2016. Second occurrence focus in Stockholm Swedish.
Manuscript, Stockholm University.
Myrberg, Sara & Tomas Riad. 2015. The prosodic hierarchy of Swedish. Nordic
Journal of Linguistics 38. 115–147.
Myrberg, Sara & Tomas Riad. 2016. On the expression of focus in the metrical grid
and in the prosodic hierarchy. In Caroline Féry & Shinichiro Ishihara (eds.),
Oxford handbook of information structure, 441–462. Oxford: Oxford University
Press.
Naydenov, Vladimir. 2011. Issues in the phonology of the tonal accents in Swedish
and their Norwegian and Danish counterparts. PhD dissertation, University of
Sofia.
Nordberg, Bengt. 1970. Språket som socialt kännetecken: Rapport om ett
språksociologiskt försök. Uppsala, FUMS report 7.
Nordberg, Bengt. 1972. Morfologiska variationsmönster i ett centralsvenskt
stadsspråk. Uppsala, FUMS report 23.
Nordberg, Bengt. 1985. Det mångskiftande språket: Om variation i nusvenskan.
Malmö: Liber Förlag.
Nyström, Staffan. 1997. Grav accent i östra Svealands folkmål. In Maj
Reinhammar (ed.), Nordiska dialektstudier: Föredrag vid femte nordiska
dialektkonferensen, 215–222 (Skrifter utgivna av Språk- och
folkminnesinstitutet genom dialektenheten i Uppsala. Ser. A:27). Uppsala:
Språk- och folkminnesinstitutet.
Öhman, Sven. 1966. Generativa regler för det svenska verbets fonologi och
prosodi. In Sture Allén (ed.), Svenskans beskrivning 3, Göteborg.
Öhman, Sven. 1967. Word and sentence intonation: A quantitative model. Speech
Transmission Laboratory Quarterly Progress and Status Report (STL-QPSR)
2–3. 20–54. Dept. of Speech Transmission, Royal Institute of Technology,
Stockholm.
Peters, Jörg. 2006. The dialect of Hasselt. Journal of the International Phonetic
Association 36. 117–124.
Peters, Jörg. 2007. A bitonal lexical pitch accent in the Limburgian dialect of
Borgloon. In Riad & Gussenhoven (eds.), 167–198.
Riad, Tomas. 1998a. The origin of Scandinavian tone accents. Diachronica 15.
63–98.
Riad, Tomas. 1998b. Towards a Scandinavian accent typology. In Wolfgang
Kehrein & Richard Wiese (eds.), Phonology and morphology of the Germanic
languages, 77–109. Tübingen: Niemeyer.
Riad, Tomas. 2000a. The origin of Danish stød. In Aditi Lahiri (ed.), Analogy,
levelling and markedness: Principles of change in phonology and morphology,
261–300. Berlin: Mouton de Gruyter.
Riad, Tomas. 2000b. Stöten som aldrig blev av – generaliserad accent 2 i Östra
Mälardalen. Folkmålsstudier 39. 319–344. Helsingfors.
Riad, Tomas. 2006. Scandinavian accent typology, Sprachtypologie und
Universalienforschung 59. 36–55.
Riad, Tomas. 2009a. The morphological status of accent 2 in North Germanic
simplex forms. In Martti Vainio, Reijo Aulanko, & Olli Aaltonen (eds.), Nordic
prosody: Proceedings of the 10th Conference, Helsinki 2008, 205–216.
Frankfurt am Main: Peter Lang.
Riad, Tomas. 2009b. Eskilstuna as the tonal key to Danish. Proceedings Fonetik
2009, 12–17. Stockholm University.
Riad, Tomas. 2012. Culminativity, stress and tone accent in Central Swedish.
Lingua 122. 1352–1379.
Riad, Tomas. 2014. The phonology of Swedish. Oxford: Oxford University Press.
Riad, Tomas. 2015. Prosodin i svenskans morfologi. Stockholm: Morfem förlag.
Riad, Tomas & Carlos Gussenhoven (eds.). 2007. Tones and tunes I: Studies in
word and sentence prosody. Berlin: Mouton de Gruyter.
Riad, Tomas & My Segerup. 2008. Phonological association of tone: Phonetic
implications in West Swedish and East Norwegian. Proceedings Fonetik 2008,
 93–96. Göteborg.
Ringgaard, Kristian. 1983. Review of Liberman (1982). Phonetica 40. 342–344.
Schmidt, Jürgen Erich. 1986. Die mittelfränkischen Tonakzente (Rheinische
Akzentuierung). (Mainzer Studien zur Sprach- und Volksforschung 8).
Stuttgart: Steiner.
Segerup, My. 2004. Gothenburg Swedish word accents: A fine distinction.
Proceedings Fonetik 2004, 28–31. Stockholm University.
Strandberg, Mathias. 2014. De sammansatta ordens accentuering i Skånemålen.
PhD dissertation, Uppsala University.
Teleman, Ulf. 1969. Böjningssuffixens form i svenskan. Arkiv för nordisk filologi 84.
163–208.
de Vaan, Michiel. 1999. Towards an explanation of the Franconian tone accents.
Amsterdamer Beiträge zur älteren Germanistik 51. 3–44.
Wetterlin, Allison. 2010. Tonal accents in Norwegian: Phonology, morphology and
lexical specification. Berlin: Mouton de Gruyter.
Carlos Gussenhoven
Prosodic typology meets
phonological representations

Abstract: Universally, phonological grammars have three structural

features. First, there are segments (vowels, consonants, tones); second,
there are constituents that contain them (the prosodic hierarchy); and third,
there are two ways in which segments are anchored in constituents
(phonological alignment and association). The specific segments and
constituents a language has are not necessarily shared with other languages.
There are three points that arise from this conceptualization. First, while
word stress is a constituent (a foot), sentential prominence is derivative and
not to be represented in terms of additional mechanisms like metrical grids
or trees. Second, association appears to have been overused at the expense
of the simpler concept of phonological alignment, to the detriment of
descriptions of intonation. Third, the much-discussed word prosodic
concepts of tone, stress, and accent turn out to belong to each of the three
very different structural features of phonological grammars: tone is a
segment, stress a prosodic constituent, and accent an instruction for
association. As such, they will not easily fit into a single typological
taxonomy.

1 Introduction
Like Christmas presents, phonological grammars are best typologized under
three structural headings. First, there is the contents, the actual present, or
equivalently, the phonological features and segments that make up the
segmental strings. Next, there are the containers, the box, the wrapping
paper, and the ribbon, equivalent to a suite of constituents in the prosodic
hierarchy (Selkirk 1981; Nespor & Vogel 1986; Selkirk & Lee 2015).
Finally, there are ways of anchoring the present inside the box, airbags or
polystyrene beads perhaps, equivalent to phonological alignment and
segmental association (Goldsmith 1976; McCarthy & Prince 1993). The
purpose of this article is, first, to point out that tone, stress, and accent are
not easily typologized within a single taxonomy, since they are
instantiations of each of these three very disparate aspects of grammars,
respectively. The second goal is to draw attention to two aspects in prosodic
descriptions that do not fit this model. One is the burgeoning of
representations of prominence above the word level. The second is the
imbalance in the employment of phonological alignment and phonological
association, where phonological alignment has been underused and
phonological association overused, to the detriment of our understanding of
phonological representations of intonation.

2 Segments
In this section, I defend two positions. First, all languages have
phonological segments (Section 2.1), and second, there is no systematic
phonetic representation in the sense of Chomsky & Halle (1968) (Section 2.
2). A claim that there are no phonological segments is refuted in Section
2.3. In Section 2.4, I point out that the “suprasegmental” view of
phonological structure has blurred the segmental status of tones, with a
detrimental effect on typological discussions of word prosody.

2.1 Vowels, consonants, and tones

Vowels, consonants, and tones are the three featurally specified
phonological constituents providing phonological content. Linearly
arranged segments and specific features in them may form parallel
autosegmental tiers (Goldsmith 1976; Pierrehumbert 1980; McCarthy
1985). All languages have segments, but a language’s specific segments
ultimately result from varyingly probable responses to ergonomic
conditions in speech production and perception (Flemming 1995; Boersma
1996; Clements & Ridouane 2011). That is, if [a] occurs in all languages, it
is because conditions on its inclusion are highly favorable, but a language
without low vowels may well be encountered any time from now. Tones are
often excluded from the meaning of the term “segmental”, which in its
narrower sense only refers to vowels and consonants. Hyman (2011) brings
out the segmental status of tones particularly clearly, showing that they are
segments par excellence in displaying more versatile behavior than vowels
and consonants. In addition to representing morphemes by themselves,
being autosegmental, causing and undergoing assimilation, only tones are
routinely associated with numerous adjacent association units, interact with
other tones over large distances, are displaced over large distances, and are
frequently retained in the representation without association (floating
tones).

2.2 Segments and phonetic implementation

The phonology-phonetics interface which is assumed here is that of
Pierrehumbert (1980), in which the surface configuration of discrete
phonological constituents is translated into an analogue representation,
exemplified in her dissertation by a target-interpolation model for the
translation of tones into English f0 contours. Regardless of the particular
conception of phonetic implementation (target interpolation, as in
Pierrehumbert 1980, a gestural score, as in Browman & Goldstein 1986,
exemplars, as in Pierrehumbert 2002), this process is unlikely to leave
behind a one-to-one correspondence between phonological segments and
chunks in the signal. For instance, a nasalized vocoid represents no segment
if it is a context-dependent nasalization of a long oral vowel, as in English
[kʰɑɑ̃m] /kɑːm/ ‘calm’. It corresponds to one segment if it results from a
contrastively nasalized vowel, as in French /kã/, occurring twice in cancan,
while it results from two segments in Frisian [ ĩ̍fɔlə] ‘to substitute for’,
where /in/ is realized as [ĩ] before fricatives (compare [ı̍ ndɑmə] ‘to
contain’).
Arguing against the existence of a systematic phonetic representation,
Pierrehumbert (1990) brings up the case of the “segment” between /p/ and
/l/ in police in a token of the word in which a brief interval occurs between
the release of [p] and the contact for [l]. On the basis of the phonetics, no
principled choice is possible between schwa or the release of the plosive.
(See also Ladd 2014: Ch. 2, who extensively discusses the illusory nature of
a systematic phonetic representation.) As a result, speech signals can be
transcribed in many ways without any one of them having a superior claim
to correctness, as in the case of [bʱa], [bʱa̤], [b̤a̤], [ba̤] [bà̤], etc. for an
utterance [ba] pronounced with breathy voice. This situation differs sharply
from phonological transcriptions, only one of which is correct for any given
pronunciation of a linguistic expression and for which, moreover, no signal
is required for a transcriber who knows the language, as evidenced by
pronunciation dictionaries (Gussenhoven 2007; see also Abercrombie’s
1964: 17; 1967: 128 discussion of “impressionistic” and “systematic”
transcriptions).

2.3 Refuting a claim of segmentless phonology

One view that rejects segments altogether is represented by Port & Leary
(2005), who claim that the assumption of a segmental structure has been ‘a
fundamental mistake of the generative paradigm’. Their point is that
‘phonetic segments are not formal symbol tokens’, whereby we should bear
in mind that their objection is directed at the (presumed) existence of
phonological segments, since there is agreement on the non-existence of
phonetic segments (Section 2.2). It is important to see that the non-
existence of phonetic segments provides no basis for deciding the status of
phonological segments. As Ladd (2014: 22) points out, cursive writing
shows many contextual effects on the shapes of letters and does not present
obvious segmentation points between letters, much in the way that a
phonetic signals blurs the segments that they express, but that does not
invalidate the reality of the set of discrete letters that are realized (cf.
Hyman 2015b). Port & Leary present three main arguments. First, there is
no evidence of a segment-based processing of speech. Second, speakers are
unable to list the phonemes of their language with the same ease as they can
list the letters used to represent the language. Third, people are inconsistent
in writing transcriptions. A response to the first argument must ignore the
overwhelming evidence of segment manipulations in deletion, insertion,
spreading processes in phonological grammars, since these might not fall
under empirical evidence for Port & Leary (2005) and probably fall outside
their definition of processing. However, it can be weakened by evidence
from segment-based language change. The Dutch diphthongs /ɛi oey ʌu/, as
in bijt ‘bite 1sg’, buit ‘loot’, bout ‘bolt’, have recently lowered to [æi ɶy au]
in Netherlandic Dutch (Jacobi 2009: 66; van der Harst 2011: 298).
However, this lowering does not apply to monophthongal /ɛ ʏ ɔ/ in bed
‘bed’, put ‘water well’, and bot ‘bone’, whose phonetic qualities before the
change were close to those of the starting points of the diphthongs,
particularly in the case of /ɛ/, which has undergone raising in recent years.
This can be explained if the diphthongs are phonological segments that are
distinct from segments that are phonetically similar to their beginning and
end points. The argument can be widened to phonological change over
longer time spans, including the Neogrammarian reconstruction work on
Indo-European. McQueen et al. (2006) provide experimental evidence that
makes the same point. They showed that listeners’ adaptation to a spectral
change in either [s] or [f] affects the perceptual boundary between these
segments throughout the lexicon, not just in the test items, which finding
can only be explained if words are made up of segments. In addition,
speech error research strongly suggests that representations are segmental
as well as syllabic (Fromkin 1973; Nooteboom & Quené 2017).
The second argument touches on the tacit nature of human cognitive
functioning and the limited domains in it that become available for
awareness. Language is notoriously tacit, but some differentiation does
exist, in part depending on the structure of the language. For many
languages, syllables are relatively close to awareness, but segments,
intonation, and tone much less so. The fact that people find it easy to count
visible objects, like letters, punctuation marks, and Arabic digits, would
appear to be unrelated to the relative difficulty of achieving full awareness
of the segmental structure of one’s native language.
Third, the fact that people are bad at transcribing words segmentally and
show interference from the morphology and the orthography (Strange 1995)
testifies – again – to an incomplete process of segmental awareness as well
as interference of other aspects of the linguistic structure and of the
orthography. Total segmental awareness may perhaps not have been
achieved whenever the analysis of languages has remained controversial, as
in the case of the German affricate /ts/, which is more generally assumed to
be a segment in initial and final position (e.g., Zeit ‘time’, Herz ‘heart’) than
when it straddles a syllable boundary, as in Katze ‘cat’. Sequences like /ps/
have also been argued to be single segments (Wiese 1996: 13, 265). Still,
many people do acquire the ability to produce pronunciation dictionaries,
and with it, the ability to count the segments of their language.
Uncertainties may remain for less accessible aspects, like complexity of
segments, syllabic affiliation, and over decisions about whether or not to
include predictable assimilations. However, these aspects do not undermine
the otherwise strong evidence that segmental representations are relatively
easily introduced into a speaker’s awareness.

2.4 The distorting lens of the suprasegmental

view
The traditional classification into three types of segments (vowels,
consonants, and tones) competes with an alternative conceptualization of
phonological structure, one based on the even older division into segmental
and suprasegmental phenomena, whereby tones are divorced from a more
narrowly defined “segmental” class of vowels and consonants (e.g., Lehiste
1970). While endorsing the idea that suprasegmental features may
encompass whole strings of spectrally defined units (vowels and
consonants), Lehiste motivated the suprasegmental status of f0, intensity,
and duration by observing that only spectrally defined units can have values
for each of those three parameters, unlike any of those three themselves.
That is, intensity cannot be said to have an f0, say. Table 11.1 shows the
classic four phonetic parameters in the first column in each partition and the
phonological phenomena that are cued by them in the second. The multiple
cues to stress are an indication of a more general spread of cues involved in
the realization of phonological constituents.
Table 11.1: Four suprasegmental phonetic parameters and the phonological
phenomena most closely cued by them, dividing the segmental grouping of
vowels, consonants, and tones and classifying tones with prosodic constituents
(i.e., feet, the representation of stress).
 The view in Table 11.1 goes against the three-segment view defended
here in grouping tones together with a prosodic constituent, the foot, into a
class of entities that are not vowels or consonants. This has had a
detrimental influence on the typology of word prosody. Together with a
poorly defined notion “accent”, their grouping has blurred the tripartite
division into three aspects of phonological grammars identified in the
introduction. In addition, the metaphor of prosody as an overlay on vowels
and consonants (“suprasegmental”) has reified inappropriate notions of
stress as being a relational concept that needs another syllable for it to exist
(Trager & Smith 1951: 36; Liberman & Prince 1976), and perhaps also of
intonation as sets of utterance-wide melodies.

3 Prosodic constituents
In this section, I argue that in all languages segments are contained in a
hierarchically arranged set of otherwise empty constituents, with higher
ones encompassing lower ones (Selkirk 1981; Nespor & Vogel 1986;
Selkirk & Lee 2015). Prosodic constituents constrain the distribution of
segments. For instance, an English sequence like /pk/ can only occur across
a syllable boundary, as in napkin, because inside the syllable, there is no
legitimate way in which they can be located in that order. Similarly, they
may forbid specific mixes of segments. Pharyngealized and contrastively
plain consonants do not occur in the same syllable in Zwara Berber
(Gussenhoven 2017), and nasal and non-nasal segments may not occur in
the same phonological word in some languages (Walker 2011). They may
also restrict their number, like the single voiced obstruent in Japanese words
(Itô & Mester 1986) and the single glottalized consonant in the Indo-
European syllable (Hopper 1973; Gamkrelidze & 1973).
The specific suite of constituents is language-specific. First, constituents
with identical ranks may take different forms in different languages, and
second, languages do not have identical suites of prosodic constituents. The
first point in particular is true for syllables and feet, as discussed in Sections
3.1 and 3.2. The second point would appear to be true of feet, as discussed
in Section 3.3, phonological words (Schiering et al. 2010), and accentual
phrases. The situation becomes more varied at the higher end of the
hierarchy. For instance, only six out of the 14 chapters in Jun (2014) with
descriptions of (groups of) languages postulate an accentual phrase, a
constituent somewhat larger than the word (Basque, Bengali, Dalabon,
Georgian, Japanese, and Mongolian). Section 3.4 emphasizes that in the
interest of conceptual clarity, stress is to be equated with word stress, i.e.,
foot structure. Headedness of higher prosodic constituents than the foot and
the Pword is less obvious and the ways in which higher-level prominence
has been recruited to explain phonological phenomena have not been
successful in improving our understanding of prosodic phenomena.
Table 11.2: Indonesian and Japanese syllable structures with disyllabic examples
/salak/ ‘Salacca zalacca’ and /nikoN/ ‘brand name’.

3.1 Syllables
Mainstream analyses of syllable structure have assumed a binary division
into subconstituents.151 One such analysis is the majority onset-rhyme type,
as exemplified for Indonesian, and another the Japanese type, where the
first cut is between the onset plus the vowel in the first mora vs. the
segment in the second mora (Kubozono 1999). Table 11.2 illustrates these,
together with disyllabic examples.
The difference in Table 11.2 suggests that syllables emerge from data.
At the same time, if associations are universal, so will be the phonological
constituent that provides the elements with which segments are associated,
and the syllable will therefore have to be universal too. As was clear in Sect
ion 2.3, linguists have two ways of deciding if a constituent exists. In
theoretical linguistics, their existence depends on their role in the grammar.
If no distributional generalization of any kind refers to a specific
constituent, it has no reason for figuring in the grammar of that language.
For instance, Labrune (2012) proposed that Japanese equates syllables with
moras, thus denying the existence of heavy syllables. As it happens, in this
case such evidence is found in the fact that the syllable is the domain of
word accent, while there is at least one syllable-based generalization about
accent placement in longer loanwords (Kubozono 2011; Kawahara 2016).
In addition, unaccented word-initial syllables have high pitch if they are
long, but low pitch if they are short in Tokyo Japanese, a difference that is
hard to capture without the syllable (e.g., Pierrehumbert & Beckman 1988).
The other way is to provide behavioral evidence. For most linguists,
Kawahara’s (2016) review of durational evidence for the rhyme in the
Japanese syllable would be enough to refute Labrune’s (2012) claim.
Conversely, Hyman (2015a) shows that reference to the syllable can be
avoided altogether in Gokana, but that is the extent of the claim. A
demonstration that the language has no syllables would have to await the
negative results of phonetic and behavioral research.

3.2 Feet
The obligatory status of stress as well as its syllable-based nature (Hyman
2006) follow from the fact that feet are headed and directly dominate
syllables. That is, if prosodic constituents are obligatory and words are
parsed into feet, stess is obligatory if the language has feet. Stress thus has
no phonological substance. There will typically be a hyperarticulation of the
stressed syllable, often leading to greater duration, more even spectral
balance and relatively little undershooting of targets. Also, many languages
have developed different segmental profiles for stressed and unstressed
syllables, like English. However, there is in principle no problem with
languages that have feet which do not obviously lead to measurable
properties.
Compared to the shapes of syllables, those for the foot are more varied
still. In addition to the distinction between trochees and iambs, trochees
come in many shapes, depending on the language. Thus, there is the
syllabic trochee ([σ σ]), the even moraic trochee ([μ.μ] or [μμ]) (Hayes
1995), and the uneven moraic trochee ([μμ. μ], [μ. μ] or [μμ]), as widely
used in the analysis of Germanic languages (Zonneveld et al. 1999). And
there is the Germanic foot, which can have a disyllabic strong branch
(indicated by parentheses) and ranges across the structures [μμ.μ], [μ.μ],
[(μ.μμ) μ], [(μ.μ) μ] and [μμ], as in /wor.du/ ‘words’, /lo.fu/ ‘praises’,
/ky.niŋ. ga/ ‘king’, /we.ru.du/ ‘troops’ and /sel/ ‘hall’ (Dresher & Lahiri
1991). Like syllable structure, foot structure evidently emerges on the basis
of language input.

3.3 Languages without feet

Not all languages have stress. Perception experiments in which Indonesian
listeners had to choose between pronunciations with f0 obtrusions on
different syllables showed that two of the three regionally defined groups
had no clear preference, which was interpreted to mean that their language
had no word stress, but marked intonation phrases with the help of a
boundary accent (Goedemans & van Zanten 1993). Maskikit &
Gussenhoven (2016) established the Pearson correlations between the time
stamps of the f0 peak and six syllable-related landmarks in an Ambonese
Malay corpus of one trisyllabic and seven disyllabic words with varied
segmental structures recorded in a variety of prosodic conditions. The aim
was to see if the segmental landmark that best predicts the location of the f0
peak was located in the putatively stressed syllable. Most probably, this
would have to be the beginning of the rhyme or the end of that syllable. We
found a steady increase of the correlation as the landmark was later in the
word, the best value being the beginning of the rime of the last syllable.
Since descriptions have claimed that the PENULTIMATE syllables of these
words have word stress (van Minde 1997), the putatively stressed syllable
appeared to play no role in the alignment of the f0 peak. Other phonetic
measures equally failed to single out the penultimate syllable as having
stress. Tellingly, segmentally equivalent Dutch data with penultimate stress
did show the highest correlation with the beginning of the penultimate rime,
which coefficient differed significantly from that between the f0 peak and
the beginning of the penultimate syllable. Clearly, the attribution of stress in
Ambonese Malay, and by a justifiable extension in Indonesian, has been
due to the perception of stress caused by the phrase-final intonation contour
by speakers of languages with stress.
Ambonese Malay thus has no stress. Additionally, there is independent
evidence that it has no feet, though it does have syllables and Pwords.
Maskikit & Gussenhoven (2016) investigated the durational coherence of
the beginnings and endings of successive syllable rhymes in the same words
as used for predicting the location of the f0 peak, on the assumption that
absence of metrical structure might reveal itself in low correlations among
these landmarks. Table 11.3 gives the results of a Principal Component
Analysis showing greater coherence among the landmarks in Dutch than in
Ambonese Malay, as expressed in the percentage explained variance by the
two orthogonal variables (aka “components”) which were extracted from
the six segmental timing variables. Both are higher in the Dutch data,
meaning that overall the temporal coherence of the Dutch words is higher
than of the Ambonese Malay words. And to return to the first point, the
prediction of the location the f0 peak, it is best predicted in Ambonese
Malay by using THE TWO COMPONENTS, indicating the diffuseness of the
variation in the peak location relative to the syllabic landmarks. For Dutch,
a SINGLE LANDMARK, the beginning of the penultimate rhyme, predicts it
best, other landmarks being less relevant.
Table 11.3: Explained and cumulative explained variance of the six landmarks by
the first two components obtained from a Principal Component Analysis (columns
2 and 3) and the explained variance in the time stamp of the f0-peak by the time
stamp of the best a single landmark (column 4) and by the combination of the two
components (column 5) in a corpus of Ambonese Malay and Dutch words.

Our investigation indicated that Ambonese Malay does have Pwords.

There is an evident contrast between a reduplicated monosyllabic word and
a disyllabic word. This is due to the placement of the f0 peak within the
domain of the Pword, whereby the reduplicant is a separate Pword. A
declarative pronunciation of (1a), a non-reduplicated form (van Minde
1997: 29), has the peak around the boundary between the two syllables, but
the reduplicated form (1b) will have it centrally in the second syllable.
Significantly, this phrasal minimal pair will wrongly impress speakers of
languages like Dutch and English as showing contrastive stress. In
combination with the data showing the absence of stress and feet in the
language, these examples show that Pwords need not be headed, and can
exist without containing feet.

Distributional properties also indicate that the language has syllables, as

evidenced by its rejection of complex onsets, non-nasal word-internal codas
and word-internal insertion of glottal stops to separate identical vowels. The
case of Ambonese Malay does not exclude languages that demonstrably
have no stress, but still have feet. Their existence could reveal itself in the
way the foot constrains the distribution of segments.

3.4 Questionable extensions of word stress to

the phrasal domain
Headed feet and headed Pwords predict that words maximally have primary
and secondary stress, as in English ˈsalaˌmander and ˌsaluˈtation. The
discrete view of stress espoused here contrasts with a conception of stress
as graded levels of prominence extending beyond the level of the word. For
English, these have been used to account for pitch accent distributions in
the sentence (Chomsky & Halle 1968 and Liberman & Prince 1976, inter
alia). Of the three mechanisms that have been used to represent continuous
prominence, one has been widely rejected, the n-ary feature [stress] of
Chomsky & Halle (1968). The other two do not fit into my Christmas-
present model, the metrical tree and the metrical grid (Liberman & Prince
1976). Elsewhere I have argued that these metrical representations have
failed to provide successful generalizations to account for the postlexical
“Rhythm Rule” or “Stress Shift” (Gussenhoven 2010, 2015). By contrast, a
representation with prosodic constituents, notably the phonological phrase,
in combination with “accent” in the sense of syllables that are to be
provided with an intonational pitch accent (see 4.4) does provide full
coverage. Since other generalizations governing the distribution of accents
are morphosyntactic, the only remaining roles for a metrical representation
are the representation of utterance rhythm, of prominence and the provision
of matching templates in textsetting to respect poetic meter (cf. Giegerich
1983).
None of these tasks requires the presence of a metrical grid or tree.
First, even though they are quite capable of speaking is tune with some
poetic meter, speakers do not normally speak rhythmically at all (Nolan &
Jeon 2014). And while we can redundantly derive degrees of prominence
from a surface representation like (2) (from Gussenhoven 2015), it is not
clear how adding that information facilitates the phonetic implementation of
either speech rhythm or prominence.

Finally, while we can represent (2) in terms of a metrical score and point
out structural similarities between music and language (Lerdahl &
Jackendoff 1983), it is not the case that all sentences fit at least a single
meter; neither do we need a grid or tree to establish what utterances do fit a
given meter, assuming we know the conventions. In (2), bold print is used
for stressed syllables and small capitals for accented syllables; Pwords are
in parentheses, phonological phrases in square brackets and intonational
phrases in accolades. The point is that (2) gives all the information needed
to generate a flawless, canonical utterance.

4 Alignment and association

In Sections 4.1 and 4.2, I briefly discuss the two anchoring devices that
represent the third structural aspect of phonological grammars. In Section 4.
3, I discuss the ways in which association and alignment are contrastive and
how higher-ranking constraints may prevent expected associations or
alignments, citing data from Japanese and varieties of Dutch. Next, in the
spirit of Section 3.4, Section 4.4 critically discusses the semantic extensions
that ‘association’ has undergone in the intonational literature. Finally, in
Section 4.5 I discuss and defend the notion of accent, the device which is
used to enforce the association of a tone.

4.1 Alignment
All linguistic constituents are aligned with one or more other constituents
(McCarthy & Prince 1993). Alignments minimally amount to the
coincidence of an edge of one (prosodic or segmental) constituent with an
edge of another. One implication is that no constituent can be demanded to
make an appearance in random locations in linguistic expressions. A second
implication is that infixation is oriented towards the beginning or the end of
a constituent, as in the well-known case of Tagalog verbal um. The suffix’s
left edge aligns with the left edge of the derived verb, but because demands
on syllable structure prevent it from appearing word-initially if the base
begins with a consonant, it only ever aligns left on the surface with vowel-
initial hosts. Thus, /um/+/alis/ gives [umalis] ‘to leave’, but / um/+/sulat/
gives [sumulat] ‘to write’. In both cases, the prefix is leftmost in the word,
while respecting the minimization of onsetless syllables (Schachter &
Otanes 1972; Prince & Smolensky 1993).
Alignment of segments is particularly relevant for tones. If we disregard
the autosegmental behavior of vowels and consonants with respect to each
other in the phonology of Semitic languages (McCarthy 1985), vowels and
consonants typically appear in single strings in morphemes, even in quite
lengthy ones. Like plain-clothed policemen at a fancy ball or clowns in a
crowded village square, tones tend to be sparser on their tier compared to
vowels and consonants on theirs, making their alignments more
conspicuous. Examples of the less common situation of more tones than
vowels are given in (3a) and (3b). The lax vowel in the vocative chant in
(3a) is given as long, following the observation in Hayes & Lahiri (1991a)
that they neutralize with long vowels under this intonation contour.
Example (3b) has a lax vowel spanning four tones in an intonation contour.
In (3c), finally, the more typical situation is given of tones spanning
stretches of vowels and consonants. The L-tone spans the stretch between
the preceding H* and following H%, approximately coinciding with /əv
ðəʊz pʊ/.

These segmental timings have been described as lengthening (of vowels) in

(3a, b) and “spreading” or “control” by the L-tone of the post-accentual
stretch in (3c) (Beckman & Pierrehumbert 1986). Both the beginning and
the end of the stretch show up as targets, spectral change in (3a, b) and pitch
elbows in (3c) in the f0 contour.

4.2 Association
Association is a temporal integration between segments (or features) with
either rhymes or moras, the tone bearing units (TBUs) (Howie 1974;
Hyman 1985). The motivation for TBUs is (i) the relatively constant
phonetic timing of a tone’s target in relation to a location in the TBU and
(ii) the implication that strings of tones will be regularly distributed over
strings of TBUs. For the first aspect, Pierrehumbert (1980: 44) observed
that the target of the accented tone (T*) of English was timed in a fairly
constant fashion relative to an accented syllable, but that the following
“phrase accent”, by which she meant the tone after T*, showed “a fair
amount of variation” in a location near the end of the nuclear word. This
point is shown graphically in (4), with transcripion following e.g.
Gussenhoven (2016).
The timing of L is related to its distance from H* and is not governed by the
syllable structure of the post-accentual stretch, as shown for English by
Barnes et al. (2010). To a limited extent, the timing and scaling of
associated tones are affected by contextual factors, like tone crowding and
the nature of adjacent tones. Thus, in (4a), the target of H* occurs earlier
than in (4b, c), because it needs to make room for the targets of the
following L-tones. In (4b, c), the distance of H* to L is greater than that in
(4a), because despite accommodating the L-tone by being earlier, the target
of H* is close enough to the final boundary for LL% to be squeezed up to it.
Observe that the graphs are distorted because they are projected onto the
orthographic examples; an impression of their utterance durations is
provided by the horizontal bars.
The second motivation for TBUs is the way the strings of tones are
distributed over them, which figured emphatically in work by Leben (1970,
1973) and Goldsmith (1976). This is shown, for instance, by the different
distributions of LH in Mende words like /lèlèmá/ ‘praying mantis’ and
/ndàvúlá/ ‘sling’, where L is associated with the first two syllables in the
first word, but only with the first in the second word, other syllables having
H. Unlike the situation illustrated in (3), the synchronization of the target of
L with its syllables is of the same order of precision as that between H and
its syllables (see also Beckman & Pierrehumbert 1986: 281). That is, the
Mende contrast cannot be reproduced in (4).

4.3 Alignment and association are contrastive

and violable
Alignment is contrastive in two senses. First, a tone can be aligned with
edges of different prosodic constituents. An H-tone, for example, may be
aligned with the right edge of an intonation phrase or with the right edge of
a syllable. When the right edges of these constituents do not themselves
coincide, the alignment contrast is evident from the fact that the two tones
are in different locations, as shown in (5a, b). However, the same contrast
occurs when the two prosodic constituent edges do coincide, as in (6a, b),
where the edge immediately above the H-tone is the edge with which it
aligns. As a result, the H-tones will be treated differently by the phonetic
implementation rules.

A case in point concerns H* and Hα in Tokyo Japanese. Its accentual phrase

(α) begins with LαHα, while an interrogative utterance ends in H%, as seen
in (7a).

The two tones in the Tokyo Japanese pitch accent H*L associate with the
first mora in the accented syllable and the next mora, if there is one,
respectively (Pierrehumbert & Beckman 1988). In (7b) this is possible only
for H*, there being no second mora available for L. Each of the members of
this minimal pair now consists of a sequence of two H-tones, Hα in (7a) and
H* in (7b), where the second is the floating interrogative H%. Since the
boxed tones in (7b) are deleted for lack of a TBU, the difference between
(7a) and (7b) is that Hα (7a) left-aligns with the left edge of the accentual
phrase, while H* in (7b) left-aligns with the left edge of the rhyme of the
accented syllable, as expressed by the diacritics α and *, respectively. For
those speakers who maintain this contrast,152 the pitch movement in (7a) is
lower than that in (7b). Thus, while these forms have identical phonological
tones, LHH, in which the first H is associated with the one available mora,
the alignments of that first H are not identical, leading to a phonetic
distinction (Gussenhoven 2004: 187). A similar case was earlier reported by
Hayes & Lahiri (1991b), who noted that the IP-final string LHL in Bengali
had a systematically higher and later f0 peak when H is aligned with the IP,
in which case it signals a question (L* HɩLφ), than when it aligns with the
phonological phrase, in which case it signals a narrow-focus declarative (L*
HφLφ).
A second sense in which alignment is contrastive arises from OT, which
holds that alignment constraints are ranked amongst each other. When two
tones with different alignments compete for the same position, the order
they appear in will depend on their ranking. In (8a), the right-edge
alignment of H% to the right edge of the intonation phrase is ranked above
a constraint that aligns the right edge of a lexical H to the right edge of the
syllable, a situation is found in Venlo Limburgish interrogatives
(Gussenhoven & van de Vliet 1999).

By contrast, the ranking is reversed in the dialect of Roermond Limburgish,

as in (8b) (Gussenhoven 2000). In (8b), L% is aligned as close to the edge
of the intonation phrase as it can be, while granting a rightmost position to
the lexical H, just as H is as close to the edge of its syllable as it can be,
while granting a rightmost position for H% in (8a). (The parenthesis stands
for the right edge of the syllable and the curly bracket for that of the IP.)
Importantly, the spatial orientation of } to the right of ) is a metaphor for
HIGHER CONSTITUENCY, not for distance or time. Hume (1998) gave the same
analysis for Leti metathesis, where underlying VC# is CV# phrase-finally,
except that she could use the more encompassing LINEARITY constraint
instead of a constraint aligning C with the right word edge.
ASSOCIATION contrasts with non-association. To return to Tokyo
Japanese, observe that both (7b) and (9b) are accented by means of the H*L
pitch accent.
In word-final monomoraic accented syllables like (7b), the trailing L
(boxed) cannot associate, since there is no mora available for it and the
language disallows the association of more than one tone to a TBU. As a
result, the floating lexical tone, identified through its left-edge alignment
with the right edge of H* (cf. CONCATENATE in Riad 1998), is deleted.
Because in (9b) a second mora is available, [n], L can associate here. As a
result, the interrogative contours for the two monosyllables are notably
different, a mid-to-high rise in (7b) and a fallrise in (9d). The interrogative
contours for unaccented (9a) and accented (9b) are highly salient as a result,
quite unlike the more vulnerable contrast between (7a) and (7b). Non-
crucially, floating H α is deleted in both (7b) and (9b) (Gussenhoven 2004:
188).
Both alignment and association are violable. Alignment is violated in
Giryama, where the last lexical H-tone aligns right with the right edge of
the IP, regardless of the location of its sponsoring morpheme, which can be
a considerable distance from the phrase end (Volk 2011). However, instead
of associating with the last syllable, it associates with the penult. The
constraint that outranks it will either be one that locates a boundary tone on
that last syllable, rendering it unavailable for H, or one that attracts H to a
stressed penult. A similar frustration of right-edge alignment occurs in a
number of European languages (4.3).
Similarly, an expected association of an intonational tone may be
frustrated by the presence of a lexical tone in the accented syllable, as in
Hasselt Limburgish (Peters 2008).
In (10a), H* associates with the accented syllable /kiəs/ in the word for
‘cheeses’, while in (10b), a lexical L is associated with that syllable in the
word for ‘socks’. As a result, H* in (10b) needs to forgo its absolute left-
edge alignment in the accented syllable as well as its association, in favor of
the lexical tone. While this analysis is identical to that in Peters (2008), I
have not adopted his convention of attaching the star to the tone that
associates, intonational H* in (10a) and lexical L (Peters’ L*) in (10b).
While ‘*’ is an instruction to associate the segment bearing it to an accented
TBU, it is not equivalent to association, which is symbolized by an
association line. In (10b), the intonational pitch accent LH* aligns left with
the lexical tone, where it fails to associate its H*, due to the pre-emptive
association of L in the lexicon with the TBU that is designated as accented
postlexically.

4.4 Experimental support

Phonetic research has established the existence of variation between
floating and associated post-T* tones among varieties of West Germanic in
the Netherlands. In Roermond Limburgish, right-hand boundary tones
associate with stressed syllables (Gussenhoven 2000), while tones in
equivalent positions in Dutch remain floating (van de Ven & Gussenhoven
2013). The first situation was shown for the post-focal stressed syllables in /
ɛin(dələk)/ eindelik ‘at last’, /klɔɔʁ/ klaor ‘ready’, and /ɣɛt/ ‘something’, in
which the beginning of the rhyme is increasingly close to the right
boundary of the IP. Its association is reflected in the beginning of the f0 fall
in the interrogative melody L* HL%, as shown in Figure 11.1. The high
plateau illustrates a case of two-edge alignment, which creates two targets
with approximately the same f0 between which the pitch is interpolated. In
the Roermond case, the left-edge alignment does not, but the right-edge
alignment does result in an association. For more discussion of associating
post-T* tones see Grice et al. (2000).
By contrast, the timing of both the first and second targets of the medial
low level stretch in a Dutch H*L H% contour has been investigated by
Peters et al. (2012) and by van de Ven & Gussenhoven (2013), respectively.
The second target is shown in Figure 11.2 by the downward arrow. The data
is based on two unaccented words in final position in IP, which varied from
monosyllabic to disyllabic, with initial stress. The horizontal bars give
durations of onsets and rhymes and are aligned at the beginning of the first
word. From top to bottom, the words are two disyllables, a monosyllable
plus a disyllable, a disyllable plus a monosyllable, and two monosyllables.
The only effect on the timing of the low target is the end of the intonation
phrase. Such insensitivity to the phonological structure indicates a lack of
association. The alignment of this L is identical to that of the associated H
of the Roermond dialect, shown in Figure 11.1.

Figure 11.1: Synchronization of the beginning of the f0 fall in Roermond

Limburgish with the beginning of the rhyme of the last stressed syllable
in the intonation phrase.
▲ = antepenultimate stressed /ɛin(dələk)/, ● = final stressed /klɔɔʁ/; ∎ =
final stressed /ɣɛt/. N = 24.
 Figure 11.2: Synchronization of the pre-final f0 rise in Dutch with the
end of the intonation phrase.
= Word1; = unstressed syllable; = stressed rhyme; = onset
consonant. N= 432. Design after Lickley et al. (2005).

Two-edge alignment offers a principled solution to a phenomenon that

has been described as “spreading” of Pierrehumbert’s (1980) phrase accent
in English, reanalyzed as the final boundary tone of the intermediate phrase
(Beckman & Pierrehumbert 1986). This is the low level stretch due to their
L-, as occurring between H* and H% in cases in which the accented
(nuclear) syllable is located at some distance from the end of the intonation
phrase, as in (3c). By allowing a segment to align with two edges, we can
describe the way in which the realization of a single tone fills up the space
between two flanking targets as an interpolation between two targets. While
effectively this creates the effect of tone copying, including such a rule
would frequently run counter to the tonal grammar of the language, which
may apply the OCP without exception. Characterizing two edge-alignment
as spreading directly contradicts the findings showing that L- floats.
Spreading is equivalent to one of the four possible configurations predicted
by two-edge alignment, the one where an association exists at both ends
(and by implication all TBUs in between). Spreading thus retains its usual
meaning of multiple associations of the same tone.

4.5 Questionable extensions of association

There have been proposals to extend the concept of association to TBUs
other than rhymes or moras. Notably, Pierrehumbert & Beckman (1990:
158) described Japanese boundary tones as associating with the
phonological constituent with whose edge they align with, as shown in Figu
re 11.3, where the first L and the last H associate with the utterance node,
while that same L additionally associates with the first mora. The edge at
which the boundary tone appears was derived by a set of percolation
conventions that were meant to deal with cases of constituents with tones at
one edge.
This extension of association to higher constituents was widely adopted,
because it came with the related notion of SECONDARY ASSOCIATION. Its
motivation was the need to describe the contrastive association of a
boundary tone in Japanese to the first TBU of the accentual phrase, like the
first L in (11). In (11a), Hα associates with the first mora, a feature of
bimoraic long rhymes, as noted in Section 3.2, whereas initial Lα has a free
mora it can associate with in (11a). Pierrehumbert & Beckman (1988: 26ff.)
document the systematic and substantial phonetic effect of this difference
and called the phonetic interpretation of the associated Lα its strong
allophone and that of the floating one its weak allophone. Parallel situations
readily cropped up after the publication of their book, as in Grice (1995)
and Gussenhoven & van der Vliet (1999), inter alia. In terms of the
assumptions in this contribution, both (11a) and (11b) show the left edge
alignment of floating Hα with the accentual phrase, but only in (11b) will it
associate.
 Figure 11.3: Associations of tones to moras and high prosodic
constituents for Ane-no akai seetaawa doko desuka? ‘Where is big
sister’s red sweater?’ (from Pierrehumbert & Beckman 1988).

(11) a. {yoozinbo b. {omawarisan

| |||
Lα Hα . . . Lα H*L . . .
farming policeman

For cases like (11), association to a higher node looks like a complicated
way of describing alignment. Describing alignment as association in the
specific way in which Pierrehumbert & Beckman (1988: 154) interpret this
also leads to the false prediction that distinctions like (8a, b) cannot exist.
Observe that the nodes to which tones associate do not form a single tier in
the way strings of moras or rhymes do. As a result, depending on whether F
igure 11.3 is a two-dimensional object, as it appears on the page, or a three-
dimensional object, in which node ɩ does not lie in the same plane as node
α, for instance, the interpretation of the No CROSSING CONSTRAINT
(Goldsmith (1976) varies. Pierrehumbert & Beckman (1988: 154) in effect
take the two-dimensional interpretation when proposing that boundary
tones are sequenced so as to mirror their rank, with the tone at the highest
edge occurring outermost in the tone string.
 The resulting diffuseness of the meaning of association led to additional
interpretations. For instance, Prieto et al. (2005) proposed that tones might
associate with edges of these constituents by way of secondary association,
doubling down on the ambiguity between association and alignment. Ladd
(2004, 2006) instead argued that phonetic implementation rules should be
available for fine-tuning the synchronization of tone targets with the
syllabic, moraic, and CV-segmental structure. The concept of tone-to-node
association further gave rise to the notational convention of showing the
boundary tones as linked to the boundary bracket by means of an
association line (Hayes & Lahiri 1991b). I blithely took over that practice in
my (2004) book, but only up to Chapter 9. When formulating the
optimality-theoretic descriptions of Northern Bizkaian Basque and
Japanese, the meaning of constraints demanding association became
intractable, for which reason I reverted to the original notation of
Pierrehumbert (1980), announcing the change in Section 8.4 (p. 155).
Extensions of tone-to-TBU association at the tonal end were considered
and rejected in Arvaniti et al. (2000), who showed that the prenuclear rise in
Athenian Greek has its beginning synchronized with the end of the pre-
accentual vowel and its ending at the beginning of the post-tonic vowel,
thus embracing the consonants around the vowel of the accented syllable.
Because neither tone therefore had a privileged temporal relation with the
accented syllable, they briefly considered and rejected the option of
labelling the LH-pitch accent as accented, rather than either L or H, in the
spirit of a proposal of a dominating tonal node for the H*L pitch accent of
Japanese by Pierrehumbert & Beckman (1988). Arvaniti et al. (2000)
argued that this move would undermine the tenet of autosegmental
phonology which derives f0 movements from interpolations of level targets.
Within the terms of this contribution, the only option is to assign a star to
one of the tones in the pitch accent. For Japanese, this is clearly H, but that
decision is harder in the Greek case.153
The difference between alignment and association is in principle
applicable to other segments, too. An association difference well may
account for the phonetic behavior of “impure s” in Italian. As shown by
Hermes et al. (2000), initial /s/ in onset clusters, as in stella ‘star’, is neither
a separate syllable nor part of the onset, being durationally independent in a
way that /b/ is brilla ‘is shining’ is not. Since association is potentially
contrastive for all types of segments, the representation of “impure s” may
lack an association to the onset, making /s/ in stella a word-initial floating
consonant. In spirit, this is what Gierut (1999) and Barlow (2001) intended
to achieve when characterizing impure /s/ as extrasyllabic. To be sure,
despite a similarity in notation, association lines and lines indicating tree
structures of prosodic constituents are distinct concepts. Absence of an
association line neither implies non-parsing nor a parsing outside the
syllable, as suggested by the term “extrasyllabic”. That is, impure /s/ is just
as much part of its syllable as a floating boundary tone is part of the
constituent it aligns with. While consonants associate with onsets (or
syllables, if no onset node exists, e.g., Hayes 1995) and moras, syllables do
not ASSOCIATE with feet, but are dominated by them, or contained within
them, in the metaphor of Section 1. A conceptual blurring of association
lines and lines in tree structures may wreak havoc with notions like
“spreading” and “improper bracketing”. Segments can spread, i.e., associate
with more than one of its segment-bearing units, while prosodic
constituents ideally behave according to the Strict Layer Hypothesis. The
TBUs for tones are moras or rhymes, VBUs for vowels are moras, while
CBUs are onsets and moras. Just as vowels and tones may show multiple
associations, so may consonants, which as a result may well be
ambisyllabic. To quote Kessler (1998):
A major objection is that [ambisyllabicity] violates proper bracketing, or specifically,
the prosodic hierarchy, which teaches that elements at one prosodic level are properly
included in a single parent construct (Selkirk 1982: 355). To this one may well ask why
segments should be considered part of the prosodic hierarchy; or why geminates, which
autosegmental phonologists agree are typically single melodic constituents shared by
two syllables, are not an equally big problem.

4.6 Accent
Accent is neither phonological content, like a segment, nor is it a prosodic
constituent, like a foot. It amounts to a dual marking of a TBU and a tone
(or other segment, but there have been no proposals of accented consonants
or vowels), an instruction that the tone is to be associated with the TBU
(Goldsmith 1976: 47, 87). When a melody appears in different locations in
different words or expressions, it stands to reason to separate off the melody
and mark its location as an address label on the TBU concerned in each
word, i.e., to mark accent. Japanese and English are obvious examples.
Gomez-Imbert & Kenstowicz (2000) present the instructive case of
Barasana, which has the tone structures in (12).

These examples do not obviously suggest that the right analysis is one that
assumes two melodies, H and HL, plus a lexical accent mark on the first or
second mora, Barasana’s TBU, rather than the four melodies H, HL, LH,
and LHL of (12). One argument for the accentual analysis is based on the
way pronouns impose their tone pattern on the following noun, whose
original tones are deleted.

In (13a), the HL of [ín à] is copied onto [jáí], which loses its H, while in
(13b) the doubly linked H of [mání] ‘my’ is copied onto [mínì] ‘pet’, which
loses its HL melody. Unexpectedly, however, the LH melody of copies
as H onto [mínì], suggesting that the initial L does not count for the tone
copy rule. This suspicion is confirmed in (13d), where the initial L of
[wì˗hí˗bò] is preserved when H of is copied from of [mání] to create
[wì˗hí˗bó] as opposed to ill-formed *[wì˗hí˗bó]. Apparently, an L on the
first mora is exempt from being copied as well as from being overwritten.
Example (13e) confirms this, because [ínà] only manages to copy its HL on
the second and third moras of [bàbárá̃ ], which loses its H on those moras.
Finally, the H of [jì˗˗í] is seen on the final two TBUs of [wì˗hí˗bó].
The words in (12) now look like (14), where * over a vowel indicates
the accented mora. The empty first moras in (14c, d) receive a default L.
This analysis is supported by other processes, like the compound rule,
which deletes the accent on the second constituent, causing the tones of the
first constituent to spread through it, this time of course ignoring the accent
location in the second constituent.

Importantly, accent belongs to the third aspect of our barebones concept

of phonological grammars. That is, it concerns the anchoring of
phonological content inside prosodic constituents. There are thus no
necessary implications of the presence of a star for prominence or stress;
neither are there necessary implications of distributional limitations per
morphological word, phonological word, or accentual phrase. These
features are frequently attested, as in English, where only stressed syllables
can be accented, and which has led to the assumption that T* is somehow
metrically strong in and of itself (e.g., Prieto et al. 2005), or Japanese and
Northern Bizkaian Basque, where accent is limited to one per domain.154
Such correlations of course require explanations, which might be sought in
the segmentally privileged status of foot heads (e.g., English) and the fact
that melodies are frequently longer than the extent of one TBU.155

5 Discussion and conclusion

After attending a class during a course on intonation I taught in 2009,
Donca Steriade asked how it was that phonological representations of
intonation need to provide tones with diacritics, like H*, H-, and H%, so as
to alert the phonetic implementation rules to any special aspects of their
scaling or timing. Isn’t one H the same phonological entity as any other, the
way lexical H-tones in Mende are of a kind? A more coherent answer than I
gave at the time is that the diacritic notation is a stand-in for edge
alignment, just as the association line is a representation of association. The
point has implicitly been part of the work by Pierrehumbert (1980) and
Pierrehumbert & Beckman (1988), even though their interpretation of the
two phonological anchoring mechanisms changed. This is why there are no
diacritics in the tone string in Figure 11.3 and why Pierrehumbert (1980)
only provided association lines for the starred tone in English intonation, as
in (4).
The conservative position defended in this article has considered the
interpretations of a number of phonological concepts in detail. I have
argued that, together, those interpretations make for a coherent
conceptualization of a universal structure of phonological grammars. It
consists of three structural aspects. First, there is a vacancy for phonological
substance in the form of featurally specified tones, vowels, and consonants,
whereby languages differ in the features and segments they have. Second,
there is a prosodic hierarchy which provides the vacancy, the prosodic
constituents that contain the segments, whereby languages differ in the
specific prosodic constituents they have. And third, there are two ways in
which segments are anchored in the prosodic constituents, (i) association to
syllabic constituents, a relation which is unique to segments or features and
moras or rhymes, and (ii) alignment, a form of anchoring that applies to all
linguistic constituents. Finally, as an important byproduct of this exercise,
each of three word-prosodic concepts that often appear together in
typological discussions turned out to belong to these three very different
aspects of the grammar. TONES are segments, STRESS is a prosodic
constituent, while ACCENT is a diacritic indicating a segmental association in
grammatical processes. It should be no surprise that they will not easily fit
into a single taxonomy.

Acknowledgements: Parts of this article were presented at the 25th

Manchester Phonology Meeting (25–27 May 2017), at the UCL Phonology
Seminar (12 June 2017), at the conference on Approaches to Phonology and
Phonetics (APAP) in Lublin (23–25 June 2017), and at the 2nd
quinquennial Lahiri Fest at Eardisland Manor, Herefordshire (14 July
2017). I thank Jean Lekeneny for advice on data, Anna Balas for the
reference to Port & Leary’s article, and Larry Hyman and Aditi Lahiri for
commenting on an earlier version.

References
Abercrombie, David. 1964. English phonetic texts. London: Faber and Faber.
Abercrombie, David. 1967. Elements of general phonetics. Edinburgh: Edinburgh
University Press.
Arvaniti, Amalia, D. Robert Ladd, & Ineke Mennen. 2000. What is a starred tone?
Evidence from Greek. In Broe & Pierrehumbert (eds.) 2000, 119–131.
Barlow, Jessica A. 2001. A preliminary typology of initial clusters in acquisition.
Clinical Linguistics and Phonetics 15. 9–13.
Barnes, Jonathan, Nanette Veilleux, Alejna Brugos, & Stefanie Shattuck-Hufnagel.
2010. Turning points, tonal targets, and the English L-phrase accent.
Language and Cognitive Processes 25. 982–1023.
Beckman, Mary E. & Janet B. Pierrehumbert. 1986. Intonational structure in
Japanese and English. Phonology 3. 255–309.
Boersma, Paul. 1998. Functional phonology: Formalizing the interactions between
articulatory and perceptual drives. The Hague: Holland Academic Graphics.
Broe, Michael B. & Janet B. Pierrehumbert (eds.). 2000. Papers in Laboratory
Phonology, vol. 5: Acquisition and the lexicon. Cambridge: Cambridge
University Press.
Browman, Catherine P. & Louis M. Goldstein. 1986. Towards an articulatory
phonology. Phonology Yearbook 3. 219–252.
Chomsky, Noam & Morris Halle. 1968. The sound pattern of English. New York:
Harper and Row.
Clements, N. George & Samuel Jay Keyser. 1981. A three-tiered theory of the
syllable. Occasional Paper No. 19. The Center for Cognitive Science, MIT,
Cambridge, MA.
Clements, G. Nick & Rachid Ridouane (eds.). 2016. Where do phonological
features come from? Cognitive, physical and developmental bases of
distinctive speech categories. Amsterdam: Benjamins.
Dresher, Elan B. & Aditi Lahiri. 1991. The Germanic foot: Metrical coherence in
Old English. Linguistic Inquiry 22. 251–286.
Flemming, Edward. 2002. Auditory representations in phonology. Abington and
New York: Routledge.
Fromkin, Victoria A. (ed.). 1973. Speech errors as linguistic evidence. The Hague:
Mouton.
Gamkrelidze, Thomas & Vyacheslav Ivanov. 1973. Sprachtypologie und die
Rekonstruktion der gemeinindogermanischen Verschlüsse. Phonetica 27.
150–156.
Giegerich, Heinz J. 1983. On English sentence stress and the nature of metrical
structure. Journal of Linguistics 19. 1–28.
Gierut, Judith A. 1999. Syllable onsets: Clusters and adjuncts in acquisition.
Journal of Speech, Language, and Hearing Research 42. 708–726.
Goedemans, Rob & Ellen A. van Zanten. 2007. Stress and accent in Indonesian.
In Vincent J. van Heuven & E. A. van Zanten (eds.), Prosody in Indonesian
languages (LOT Occasional Series 9), 35–62. Utrecht: Netherlands School of
Linguistics.
Goldsmith, John A. 1976. Autosegmental phonology. Doctoral dissertation, MIT,
Cambridge, MA. Bloomington, IN: Indiana University Linguistics Club.
Gomez-Imbert, Elsa & Michael Kenstowicz. 2000. Barasana tone and accent.
International Journal of American Linguistics 66. 419–463.
Grice, Martine. 1995. The intonation of interrogation in Palermo Italian:
Implications for intonation theory. Tübingen: Niemeyer.
Grice, Martine, D. Robert Ladd, & Amalia Arvaniti. 2000. On the place of phrase
accents in intonational phonology. Phonology 17. 143–185.
Gussenhoven, Carlos. 2000. The boundary tones are coming: On the
nonperipheral realization of boundary tones. In Broe & Pierrehumbert (eds.)
2000, 132–151.
Gussenhoven, Carlos. 2004. The phonology of tone and intonation. Cambridge:
Cambridge University Press.
Gussenhoven, Carlos. 2007. Wat is de beste transcriptie voor het Nederlands?
Nederlandse Taalkunde 12. 331–350.
Gussenhoven, Carlos. 2011. Sentential prominence in English. In van Oostendorp
et al. (eds.) 2011, 2778–2806.
Gussenhoven, Carlos. 2015. Does phonological prominence exist? Lingue e
Linguaggio 14. 7–24.
Gussenhoven, Carlos. 2016. The analysis of intonation: The case of MAE-ToBI.
Laboratory Phonology 7. 1–35.
Gussenhoven, Carlos. 2017. Zwara (Zuwārah) Berber. Journal of the Association
for Laboratory Phonology 7. 1–17. DOI: http://doi.org/10.5334/labphon.30
Harst, Sander van der. 2011. The vowel space paradox: A sociophonetic study on
Dutch. Utrecht: LOT.
Hayes, Bruce. 1995. Metrical stress theory: Principles and case studies. Chicago:
Chicago University Press.
Hayes Bruce & Aditi Lahiri. 1991a. Durationally specified intonation in English and
Bengali. In Johan Sundberg, Lennart Nord, & Rolf Carlson (eds.), Music,
language, speech and brain (Wenner-Gren Center International Symposium
Series), 78–91. London: Palgrave.
Hayes, Bruce & Aditi Lahiri. 1991b. Bengali intonational phonology. Natural
Language and Linguistic Theory 9. 47–96.
Hermes, Anne, Doris Mücke, & Martine Grice. 2013. Gestural coordination of
Italian word-initial clusters: The case of Italian “impure s”. Phonology 30. 1–
25.
Hopper, Paul J. 1973. Glottalized and murmured occlusives in Indo-European.
Glossa 7. 141–166.
Howie, J. M. 1974. On the domain of tone in Mandarin: Some acoustical evidence.
Phonetica 30. 129–148.
Hulst, Harry van der. 2012. Deconstructing stress. Lingua 122. 1494–1521.
Hume, Elizabeth. 1998. Metathesis in phonological theory: The case of Leti.
Lingua 104. 147–186.
Hyman, Larry M. 1985. A theory of phonological weight. Dordrecht: Foris.
Hyman, Larry M. 2006. Word-prosodic typology. Phonology 23. 225–257.
Hyman, Larry M. 2011. Tone: Is it different? In John Goldsmith, Jason Riggle, &
Alan C. L. Yu (eds.), The handbook of phonological theory, 197–239. 2nd edn,
Oxford: Wiley-Blackwell.
Hyman, Larry M. 2015a. Does Gokana really have syllables? A postscript to
Hyman 2011. Phonology 32. 303–306.
Hyman, Larry M. 2015b. Why underlying representations? UC Berkeley Phonology
Lab Annual Report (2015): 210–226. Published in Journal of Linguistics 54.
Itô, Junko & Armin Mester. 1986. The phonology of voicing in Japanese:
Theoretical consequences for morphological accessibility. Linguistic Inquiry
17. 49–73.
Jacobi, Irene. 2009. On variation and change in diphthongs and long vowels of
spoken Dutch. Doctoral dissertation, University of Amsterdam.
Jun, Sun-Ah. 2014. Prosodic typology II. Oxford: Oxford University Press.
Kawahara, Shigeto. 2016. Japanese has syllables: A reply to Labrune. Phonology
33. 169–194.
Kessler, Brett. 1998. Ambisyllabicity in the language of the Rigveda.
http://spell.psychology.wustl.edu/ambisyll-sanskrit/Last change 27-08-2004
Kiparsky, Paul. 1982. Lexical morphology and phonology. In In-Seok Yang (ed.),
Linguistics in the morning calm: Selected papers from SICOL, 3–91. Seoul:
Hanshin.
Kubozono, Haruo. 1999. Mora and syllable. In Natsuko Tsujimura (ed.), The
handbook of Japanese linguistics, 31–61. Malden, MA: Blackwell.
Kubozono, Haruo. 2011. Japanese pitch accent. In van Oostendorp et al. (eds.)
2011, 2879–2907.
Labrune, Laurence. 2012. Questioning the universality of the syllable: Evidence
from Japanese. Phonology 29. 113–152.
Ladd, D. Robert. 2004. Segmental anchoring of pitch movements: Autosegmental
phonology or speech production? In Hugo Quené & Vincent J. van Heuven
(eds.), On speech and language: Studies for Sieb G. Nooteboom, 123–132.
Utrecht: LOT.
Ladd, D. Robert. 2006. Segmental anchoring of pitch movements: Autosegmental
association or gestural coordination? Italian Journal of Linguistics 18. 19–38.
Ladd, D. Robert. 2014. Simultaneous structure in phonology. Oxford: Oxford
University Press.
Leben, William R. 1970. The representation of tone. In Victoria A. Fromkin (ed.),
Tone: A linguistic survey, 177–219. New York: Academic Press.
Leben, William R. 1973. Suprasegmental phonology. Doctoral dissertation, MIT,
Cambridge, MA.
Lehiste, Ilse. 1970. Suprasegmentals. Cambridge MA: MIT Press.
Lerdahl, Fred & Ray Jackendoff. 1983. A generative theory of tonal music.
Cambridge, MA: MIT Press.
Liberman, Mark & Alan Prince. 1977. On stress and linguistic rhythm. Linguistic
Inquiry 8. 249–336.
Lickley, Robin, Astrid Schepman, & D. Robert Ladd. 2005. Alignment of “phrase
accent” lows in Dutch falling rising questions: Theoretical and methodological
implications. Language and Speech 48. 157–183.
Maskikit-Essed, Raechel & Carlos Gussenhoven. 2016. No stress, no pitch accent,
no prosodic focus: The case of Ambonese Malay. Phonology 33. 353–389.
McCarthy, John J. 1985. Formal problems in Semitic phonology and morphology.
New York: Garland.
McCarthy, John J. & Alan Prince. 1993. Generalized alignment. In Geert Booij &
Jaap van Marle (eds.), Yearbook of morphology, 79–154. Berlin: Springer.
McQueen, James M., Anne Cutler, & Dennis Norris. 2006. Phonological
abstraction in the mental lexicon. Cognitive Science 30. 1113–1126.
Minde, Don van. 1997. Malayu Ambong: Phonology, morphology, syntax. Doctoral
dissertation, University of Leiden.
Nespor, Marina & Irene Vogel. 1986. Prosodic phonology. Dordrecht: Foris.
Nolan, Francis & Hae-Song Jeon. 2014. Speech rhythm: A metaphor?
Philosophical Transactions of the Royal Society B: Biological Sciences 369
(1658): 20130396. DOI: 10.1098/ rstb.2013.0396.
Nooteboom, Sieb & Hugo Quené. 2007. The SLIP technique as a window on the
mental preparation of speech: Some methodological considerations. In Maria-
Josep Solé, Pamela S. Beddor, & Manjari Ohala (eds.), Experimental
approaches to phonology, 339–350. Oxford: Oxford University Press.
Oostendorp, Marc van, Colin J. Ewen, Elizabeth Hume, & Keren Rice (eds.). The
Blackwell companion to phonology. Oxford: Wiley-Blackwell.
Peters, Jörg. 2008. Tone and intonation in the dialect of Hasselt. Linguistics 46.
983–1018.
Peters, Jörg, Judith Hanssen, & Carlos Gussenhoven. 2015. The timing of nuclear
falls: Evidence from Dutch, West Frisian, Dutch Low Saxon, German Low
Saxon, and High German. Laboratory Phonology 6. 1–52.
Pierrehumbert, Janet B. 1980. The phonology and phonetics of English intonation.
Doctoral dissertation, MIT, Cambridge, MA.
Pierrehumbert, Janet. 1990. Phonological and phonetic representation. Journal of
Phonetics 18. 375–394.
Pierrehumbert, Janet B. 2002. Word-specific phonetics. In Carlos Gussenhoven &
Natasha Warner (eds.), Laboratory phonology vol. 7, 1001–1039. Berlin:
Mouton de Gruyter.
Pierrehumbert, Janet B. & Mary E. Beckman. 1988. Japanese tone structure.
Cambridge, MA: MIT Press.
Post, Mark W. 2009. The phonology and grammar of Galo “words”: A case study in
benign disunity. Studies in Language 33. 934–974.
Port, Robert & Adam Leary. 2005. Against formal phonology. Language 81. 927–
964.
Prieto, Pilar, Mariapaola D’Imperio, & Barbara Gili-Fivela. 2005. Pitch accent
alignment in Romance: Primary and secondary associations with metrical
structure. Language and Speech 48. 359–396.
Prince, Alan & Paul Smolensky. 1993. Optimality Theory: Constraint interaction in
Generative Grammar. Rutgers University Center for Cognitive Science
Technical Report 2.
Riad, Tomas. 1998. Towards a Scandinavian accent typology. In Wolfgang Kehrein
& Richard Wiese (eds.), Phonology and morphology of the Germanic
languages, 77–109. Tübingen: Niemeyer.
Schachter, Paul & Fe T. Otanes. 1972. Tagalog reference grammar. Berkeley:
University of California Press.
Schiering, René, Balthasar Bickel, & Kristine A. Hildebrandt. 2010. The
phonological word is not universal, but emergent. Journal of Linguistics 46.
657–709.
Selkirk, Elisabeth. 1981. On prosodic structure and its relation to syntactic
structure. In Thorstein Fretheim (ed.), Nordic prosody 2, 111–114. Trondheim:
Tapir.
Selkirk, Elisabeth. 1982. Syllables. In Harry van der Hulst & Norval Smith (eds.),
The structure of phonological representations. Part 2, 337–383. Dordrecht:
Foris.
Selkirk, Elisabeth & Seunghun J. Lee. 2015. Constituency in sentence phonology:
An introduction. Phonology 31. 1–18.
Strange, Winifred. 1995. Cross-language studies of speech perception: A historical
review. In Winifred Strange (ed.), Speech perception and linguistic
experience: Issues in cross-language speech research, 3–45. Timonium, MD:
York.
Trager, George L. & Henry Lee Smith. 1951. An outline of English structure.
Washington: American Council of Learned Societies.
Vance, Timothy. J. 1995. Final accent vs. no accent: Utterance-final neutralization
in Tokyo Japanese. Journal of Phonetics 23. 487–499.
Ven, Marco van de & Carlos Gussenhoven. 2011. The timing of the final rise in
falling-rising intonation contours in Dutch. Journal of Phonetics 39. 225–236.
Volk, Erez. 2011. Mijikenda phonology. Doctoral dissertation, Tel Aviv University.
Walker, Rachel. 2011. Nasal harmony. In van Oostendorp et al. (eds.) 2011, 1838–
1865.
Warner, Natasha. 1997. Japanese final-accented and unaccented phrases.
Journal of Phonetics 25. 43–60.
Wiese, Richard. 2000. The phonology of German. Oxford: Oxford University Press.
Zonneveld, Wim, Mieke Trommelen, Michael Jessen, Curtis Rice, Gösta Bruce, &
Kristján Árnason. 1999. Word stress in West-Germanic and North-Germanic
languages. In Harry van der Hulst (ed.), Word prosodic systems in the
languages of Europe, 477–603. Berlin: Mouton de Gruyter.
Subject Index
abstractness 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17
accent 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13
see also pitch accent; stress
acquisition 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
see also artificial language learning; learnability; second language acquisi
tion
active (of a contrast/feature) 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
17
alignment 1, 2, 3, 4, 5, 6, 7
analogy 1, 2, 3
apocope 1, 2, 3
artificial language learning 1, 2, 3, 4
aspiration 1, 2, 3, 4, 5, 6, 7, 8, 9
assimilation 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35
association 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17
autosegment(al) 1, 2, 3, 4, 5, 6, 7, 8, 9, 10

borrowing 1, 2, 3, 4, 5, 6, 7
see also loanword

clicks 1, 2, 3
coarticulation 1, 2, 3, 4, 5, 6
colour (vowels) 1, 2, 3, 4, 5, 6, 7, 8, 9
complexity 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11n, 12, 13, 14, 15, 16
consonant cluster 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12
consonant (dis)harmony 1, 2, 3, 4, 5
constraint-based phonology see Optimality Theory
contact 1, 2, 3, 4
contrast privative, equipollent, or gradual 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
contrast shift 1, 2
contrastive hierarchy 1, 2, 3
coronal 1, 2, 3, 4, 5, 6

databases 1, 2, 3, 4, 5, 6
 see also UPSID/LAPSyd database; P-base
deletion, of
– consonant 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
– feature 1, 2, 3, 4
– tone 1, 2, 3
– vowel 1, 2, 3, 4, 5, 6, 7, 8, 9
diachrony 1, 2, 3, 4, 5, 6, 7, 8, 9, 10n, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 2
1, 22, 23, 24, 25, 26, 27
diffusion (areal) see borrowing
diphthongization 1, 2, 3, 4, 5, 6, 7
dispersion 1, 2, 3, 4, 5
dissimilation 1, 2, 3
economy 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
enhancement 1, 2, 3, 4, 5, 6, 7, 8, 9
epenthesis 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15

features distinctive or contrastive 1, 2

features monovalent/unary, bivalent/binary, or scalar 1, 2, 3, 4, 5, 6, 7, 8, 9, 1
0, 11, 12, 13, 14
feature geometry 1, 2, 3, 4, 5, 6, 7
final devoicing 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
see also voicing contrast
floating tone 1, 2, 3, 4, 5, 6, 7
foot 1, 2, 3, 4, 5, 6, 7, 8
frequency (of occurrence, distribution) 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 1
4, 15, 16

geminate 1, 2, 3
gemination/degemination 1, 2, 3, 4, 5
geography and phonology 1
glottalization 1, 2
hiatus 1, 2, 3, 4, 5

intonation 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16

labialization 1, 2, 3
laryngeal contrast see voicing contrast
learnability 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
see also acquisition
lengthening see quantity contrast
lenition 1, 2, 3, 4, 5, 6
lexical/postlexical phonology 1, 2, 3, 4, 5, 6, 7n, 8, 9, 10, 11, 12, 13, 14, 15, 1
6
loanword 1, 2, 3, 4, 5, 6
see also borrowing

markedness 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33
merger 1n, 2, 3, 4, 5, 6, 7n, 8, 9, 10, 11
metathesis 1, 2, 3, 4, 5, 6, 7
meter (poetic) 1, 2, 3
mora 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
mora-timing see rhythm
morphophonology 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11

nasalization 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13

natural class 1, 2, 3, 4, 5
naturalness 1, 2, 3, 4
neutralization 1, 2, 3, 4, 5, 6, 7, 8, 9, 10n

Optimality Theory (OT) 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28

P-base 1, 2
palatalization 1, 2, 3n, 4, 5, 6n, 7n, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17n, 18, 1
9, 20, 21, 22, 23
pharyngealization 1, 2, 3
phonemic level of representation 1, 2, 3, 4, 5
phonetics in relation to phonology 1, 2, 3, 4, 5, 6, 7n, 8, 9, 10, 11, 12, 13, 14,
15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 3
5, 36, 37, 38, 39, 40, 41n, 42, 43, 44, 45, 46, 47, 48
phonetic implementation see phonetics in relation to phonology
phonological alternations 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12n, 13, 14, 15, 16, 1
7
phonological change see diachrony
phonologization 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
phonotactics 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18
pitch accent 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
play languages 1, 2, 3, 4
Prague School 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12
prosodic hierarchy 1, 2, 3, 4, 5
prosodic morphology 1, 2, 3, 4
prosody 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20

quantity contrast 1, 2, 3, 4, 5n, 6, 7, 8, 9, 10, 11, 12, 13

reconstruction 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19
redundancy see economy
reduplication 1, 2, 3, 4
rhythm 1, 2, 3, 4, 5, 6, 7, 8

sampling 1, 2, 3, 4, 5, 6n, 7, 8, 9, 10, 11

second language acquisition 1
segment inventories 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16
shortening see quantity contrast
song and chant 1, 2, 3
Sound Pattern of English (SPE) 1, 2, 3, 4, 5n, 6, 7, 8, 9, 10n, 11, 12, 13, 14, 1
5
speech perception 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 1
9, 20, 21, 22, 23, 24, 25
speech processing 1, 2, 3, 4, 5, 6, 7
speech production 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 1
9, 20n, 21, 22
spirantization 1, 2, 3, 4
spreading (features) 1, 2, 3n, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29
stability (diachronic) 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12
stød 1, 2, 3, 4, 5, 6
stress 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21,
22, 23, 24, 25
see also accent; rhythm
stress-timing see rhythm
suprasegmental 1, 2, 3
syllable 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21n,
22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35
syllable-timing see rhythm
syllable weight 1, 2, 3, 4, 5, 6
see also mora

tone 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13n, 14, 15, 16, 17, 18, 19, 20, 21
see also floating tone
typology defined 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14

underspecification 1, 2, 3, 4, 5n, 6, 7, 8, 9, 10, 11, 12, 13, 14

UPSID/LAPSyD database 1n, 2, 3, 4, 5n, 6, 7, 8, 9, 10, 11

voicing contrast 1, 2, 3, 4, 5, 6, 7, 8, 9, 10n, 11, 12, 13, 14, 15, 16, 17, 18, 19,
20, 21
see also final devoicing
vowel harmony 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 2
0, 21
vowel reduction 1, 2, 3, 4, 5, 6, 7
vowel systems 1, 2, 3, 4, 5, 6, 7, 8, 279–297
Language Index
Abenaki 1
Alawa 1
Alemannic 1n
Aleut 1
Algonquian 1, 2, 3, 4
Angami 1
Ao 1
Arabic 1, 2, 3, 4, 5, 6, 7, 8, 9
Arapaho 1, 2
Armenian 1, 2, 3
Arrernte 1, 2, 3n, 4, 5

Bambara 1
Basaá 1
Basque 1, 2, 3, 4, 5, 6, 7
Bella Coola 1
Bengali 1, 2, 3, 4
Berber 1, 2, 3, 4
Bininj Gun-Wok 1
Blackfoot 1
Breton 1

Campa 1
Cantonese 1n, 2, 3
Catalan 1, 2, 3, 4, 5, 6, 7, 8
Cheyenne 1, 2
Chichewa 1n
Chickasaw 1
Chinese (see Mandarin, Cantonese)
Chukchi 1
Chumash 1
Cree 1
Cree, Northern Plains 1
Czech 1
Danish 1n, 2, 3, 4n, 5, 6n, 7, 8, 9, 10
Desano 1
Doutai 1
Dutch 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21n

Ekagi 1
English 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14n, 15, 16, 17, 18, 19, 20n, 21
n, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32n, 33, 34, 35, 36, 37, 38, 39, 40, 4
1, 42, 43, 44, 45, 46, 47n, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 6
1, 62, 63, 64, 65, 66, 67, 68, 69
Eskimo (Proto) 1, 2, 3, 4n, 5
Eskimo 1, 2, 3, 4, 5n, 6
Estonian 1

Faroese 1
Farsi 1, 2, 3
Finnish 1, 2, 3, 4
Fox 1
Franconian (Central) 1, 2, 3n, 4n, 5
French 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11

German 1, 2, 3n, 4, 5n, 6, 7, 8, 9, 10, 11, 12, 13n, 14, 15, 16, 17, 18, 19, 20,
21, 22
Germanic 1, 2, 3, 4, 5, 6
Gokana 1, 2, 3, 4, 5
Gothic 1, 2
Greek 1, 2, 3, 4, 5, 6, 7n, 8n, 9, 10
Greenlandic 1

Haruai 1
Hindi 1
Hungarian 1, 2n, 3, 4, 5, 6, 7, 8n

Icelandic 1
Igbo 1
Ik 1
Ikwere 1
Ilocano 1
Indonesian 1, 2, 3
Inuit 1, 2, 3, 4, 5, 6, 7, 8n, 9
Inupiaq (Barrow) 1
Irish 1n, 2n, 3n
Italian 1n, 2n, 3, 4, 5, 6, 7
Japanese 1, 2, 3, 4, 5n, 6, 7n, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
21, 22, 23, 24, 25, 26, 27, 28, 29n
Jaqaru 1
Jimi 1

Kabardian 1, 2, 3, 4, 5n, 6
Kachin Khakass 1
Kalam 1, 2
Kàlɔ̀ŋ 1
Karen (Sgaw) 1
Khanty 1, 2, 3
Kinyarwanda 1
Kom 1
Konni 1
Korean 1, 2, 3, 4, 5, 6, 7n, 8, 9
Koromfe 1
Kuki-Thaadow 1
Kune 1
Kunwinjku 1
Kwakiutl 1

Latvian 1
Lithuanian 1n, 2, 3, 4n, 5

Mahican 1
Malagasy 1
Malay 1, 2, 3
Malayalam 1, 248m
Maliseet-Passamaquoddy 1
Manchu (Classical) 1
Mandarin (Chinese) 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
Mansi 1
Maori 1
Marshallese 1, 2
Massachusett 1
Maung 1
Mayá 1
Mayali 1
Mazatec 1
Mekeo 1
Menomini 1, 2
Mi’kmaq 1
Miami-Illinois 1
Mixtec 1, 2
Mlabri 1
Moloko 1, 2, 3, 4
Mongolian 1, 2, 3, 4, 5, 6, 7, 8

Nahuatl 1
Nenets (Tundra) 1n
Nez Perce 1
Nimboran 1
Noon 1
Norwegian 1, 2, 3, 4, 5n, 6, 7, 8, 9n, 10, 11
Nunggubuyu 1
Nupe 1
Nzadi 1

Ob-Ugric 1, 2, 3n, 4, 5, 6, 7, 8, 9, 10
Ojibwe 1
Oroqen 1n, 2, 3

Pitta-Pitta 1, 2
Pohnpeian 1
Polish 1, 2, 3, 4, 5, 6, 7, 8, 9
Portuguese 1, 2, 3, 4, 5, 6, 7, 8

Qawasar 1, 2
Qiang 1

Romance 1, 2, 3, 4, 5
Russian 1, 2n, 3, 4, 5, 6, 7

Salishan 1
Samala 1, 2, 3, 4n, 5
Sanskrit 1, 2n, 3, 4, 5, 6
Semitic 1, 2
Serbo-Croatian 1, 2, 3
Shawnee 1
Sheko 1
Skou 1
Slave 1
Slavic 1, 2, 3n
Somali 1
Spanish 1, 2, 3, 4n, 5, 6, 7, 8, 9, 10, 11
Sundanese 1
Swedish 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11n, 12, 13, 14, 15n, 16, 17, 18, 19, 20, 2
1, 22, 23

Tacana 1
Tahltan 1
Takelma 1, 2
Tashlhiyt (Berber) 1n
Teleéfoól 1
Thaayore 1
Thai 1, 2, 3
Totontepec (Mixe) 1
Towa 1
Tswana 1, 2, 3
Tungusic, Tungus 1, 2, 3, 4, 5, 6, 7
Turkic 1, 2, 3
Turkish 1, 2, 3, 4, 5, 6, 7, 8, 9

Usarufa 1
Uyghur 1, 2

Vietnamese 1, 2, 3, 4

Walloon 1
Welsh 1, 2
Wichita 1n

Xibe 1, 2, 3

Yaka 1, 2, 3
Yessan-Mayo 1
Yokuts (Yowlumne) 1, 2, 3, 4
Yoruba 1, 2
Yupik 1, 2, 3, 4
Author Index
Abercrombie, David 1, 2, 3, 4
Abitov, M. L. 1, 2
Abrahamsson, Niclas 1, 2, 3
Abry, Christian 1
Aikhenvald, Alexandra Y. 1, 2
Aissen, Judith 1, 2
Alber, Birgit 1, 2
Altenberg, Evelyn 1, 2, 3
Andersen, Henning 1, 2
Anderson, Stephen R. 1n, 2, 3, 4, 5, 6, 7, 8
Ann, Jean 1, 2
Aoun, Joseph 1
Applebaum, Ayla 1, 2n, 3, 4, 5
Applegate, Richard B. 1, 2, 3n, 4
Archangeli, Diana 1, 2, 3, 4
Árnason, Kristján 1, 2
Aronoff, Mark 1
Arvaniti, Amalia 1, 2, 3n, 4n, 5, 6, 7, 8, 9, 10
Atkinson. Quentin 1, 2, 3
Auer, Peter 1, 2
Austin, Peter 1, 2, 3

Bach, Emmon 1
Badecker, William 1n, 2
Bagov, P. M. 1, 2
Bailey, Gil 1
Baker, Mark 1
Baković, Eric 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
Balas, Anna 1
Balkarov, B. X. 1
Bannert, Robert 1
Barlow, Jessica A. 1, 2
Barnes, Jonathan 1, 2
Barreteau, Daniel 1, 2, 3
Barrie, Mike 1, 2
Basbøll, Hans 1, 2, 3
Beauzée, Nicolas 1, 2n
Becker, Michael 1, 2, 3
Beckman, Jill 1, 2, 3, 4
Beckman, Mary E. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13
Bedell, George 1, 2
Beesley, Kenneth 1, 2, 3
Bender, Byron W. 1, 2, 3, 4, 5, 6
Bennett, William 1, 2, 3
Benua, Laura 1n, 2
Bergsland, Knut 1
Bermúdez-Otero, Ricardo 1n, 2, 3, 4, 5, 6
Beros, Achilles 1, 2
Berstel, Jean 1, 2
Bhat, D. N. S. 1, 2, 3n, 4
Bhatt, Rajesh 1
Bickel, Balthasar 1, 2, 3, 4, 5, 6, 7, 8
Bickmore, Lee S. 1, 2
Blake, Barry J. 1
Bleckert, Lars 1, 2, 3
Blevins, Juliette 1n, 2, 3n, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 1
9, 20, 21
Bloch, Bernard 1, 2, 3
Bloomfield, Leonard 1, 2n, 3, 4, 5, 6
Blumenfeld, Lev 1, 2
Blumstein, Sheila 1n, 2, 3n, 4, 5, 6, 7, 8, 9
Boas, Franz 1
Boë, Louis-Jean 1
Boersma, Paul 1, 2, 3, 4, 5
Booij, Geert 1, 2
Borowsky, Toni 1, 2, 3n, 4, 5
Borsley, Robert 1
Bossong. Georg 1, 2
Breen, Gavan 1, 2n, 3, 4
Bresnan, Joan 1, 2, 3, 4, 5
Broe, Michael B. 1, 2
Broersma, Miriam 1, 2
Brohan, Anthony 1, 2, 3, 4
Bromberg, Ilana 1, 2
Broselow, Ellen 1, 2, 3, 4
Browman, Catherine P. 1, 2, 3, 4
Bruce, Gösta 1, 2, 3, 4n, 5, 6, 7, 8, 9, 10n, 11, 12, 13, 14, 15, 16, 17, 18
Brugos, Alejna 1
Buccola, Brian 1, 2
Buckley, Eugene 1, 2
Bulmer, Ralph 1, 2, 3, 4, 5, 6
Burnett, James (Lord Monboddo) 1, 2
Burzio, Luigi 1, 2
Butcher, Andrew 1
Bybee, Joan L. 1, 2, 3, 4
Bye, Patrik 1, 2, 3
Bynon, Theodora 1n, 2

Cahill, Michael 1, 2
Calabrese, Andrea 1, 2, 3, 4, 5
Campanella, Tommaso 1, 2, 3, 4
Campos-Astorkiza 1, 2n, 3
Cardoso, Walcir 1, 2, 3, 4
Carstairs, Andrew 1
Casali, Roderic F. 1, 2
Catford, J. C. 1, 2, 3, 4, 5, 6, 7
Cebrian, Juli 1, 2, 3
Chakraborti, Paromita 1, 2
Chandlee, Jane 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
Chen, Su-I 1, 2
Chiu, C. Chenhao 1, 2
Cho, Young-Mee 1, 2
Choi, John D. 1, 2n, 3
Chomsky, Noam 1, 2, 3, 4, 5, 6, 7, 8, 9, 10n, 11, 12, 13, 14, 15, 16, 17, 18,
19, 20, 21, 22, 23, 24, 25, 26n, 27, 28, 29, 30, 31, 32, 33, 34, 35
Cichocki, Wladyslaw 1, 2
Cinque, Guglielmo 1
Clairis, Christos 1, 2
Clark, Mary 1, 2
Clements, G. N. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19,
20, 21, 22n, 23
Cohn, Abigail 1, 2, 3
Colarusso, John 1n, 2, 3n, 4, 5, 6n, 7
Comon, Hubert 1, 2
Compton, Richard 1, 2, 3n, 4, 5n, 6
Comrie, Bernard 1, 2, 3, 4, 5, 6, 7, 8
Cope, Dana 1
Cornell, Sonia A. 1, 2, 270C
Côté, Marie-Hélène 1
Coupé, Christophe 1, 2, 3
Crane, Thera M. 1
Crawford, Penny 1, 2
Cristofaro, Sonia 1
Croft, William 1, 2, 3, 4, 5
Currie-Hall, Kathleen 1n, 2
Cutler, Anne 1

d’Andrade, Ernesto 1n
D’Imperio, Mariapaola 1, 2
Dauchet, Max 1
Dauer, Rebecca 1, 2, 3, 4
Davidian, Richard D. 1, 2, 3, 4, 5
DeCamp, David 1n, 2
Delisi, Jessica L. 1, 2
Dell, François 1, 2, 3, 4
Dellwo, Volker 1, 2, 3, 4
Dinnsen, Daniel A. 1, 2
Dixon, R. M. W. 1, 2
Dolatian, Hossep 1
Donegan, Patricia 1, 2
Donohue, Mark 1, 2
Dorais, Louis-Jacques 1, 2
Downing, Laura J. 1, 2
Dresher, B. Elan 1, 2, 3, 4, 5, 6n, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19
Dressler, Wolfgang U. 1, 2
Dryer, Matlthew S. 1, 2, 3, 4
Duanmu, San 1
Durand, Jacques 1

É. Kiss, Katalin 1
Easterday, Shelece 1, 2
Ebeling, C. L. 1, 2, 3
Eckman, Fred 1, 2, 3, 4, 5, 6, 7
Edge, Beverly 1, 2, 3, 4
Edlefsen, Matt 1, 2, 3
Elert, Claes-Christian 1, 2, 3
Elgot, C. C. 1, 2, 3
Elmedlaoui, Mohamed 1, 2
Ember, Carol R. 1, 2, 3
Ember, Marvin 1, 2, 3
Endress, Ansgar D. 1, 2
Engelfriet, Joost 1, 2, 3
Engstrand, Olle 1n, 2, 3
Eulitz, Carsten 1, 2
Evans, Nicholas 1, 2, 3, 4, 5, 6, 7n, 8, 9, 10, 11
Evers, Vincent 1, 2, 3, 4n, 5
Ewen, Colin J. 1, 2, 3, 4, 5, 6, 7
Eyraud, Rémy 1, 2, 3

Fant, C. Gunnar M. 1, 2n, 3, 4, 5, 6, 7, 8, 9, 10n, 11

Fast, P. W. 1, 2
Faudree, Michael 1, 2
Feuillet, Jack 1, 2
Fikkert, Paula 1, 2, 3
Finley, Sara 1n, 2, 3, 4, 5, 6, 7, 8
Fintoft, Knut 1, 2
Firth, John Rubert 1n
Flämig, Walter 1
Flavier, Sébastien 1
Flege, James Emil 1, 2, 3, 4, 5, 6, 7
Flemming, Edward 1, 2, 3, 4, 5
Fortescue, Michael 1, 2
Foss, Donald J. 1, 2
Fougeron, Cécile 1n, 2, 3, 4, 5, 6
Fought, Carmen R. 1, 2, 3, 4, 5
Fought, John R. 1, 2, 3, 4, 5
Frank, Robert 1, 2, 3
Franzen, Vivan 1
Fretheim, Thorstein 1, 2, 3, 4
Fromkin, Victoria A. 1, 2, 3

Gabelentz, Georg von der 1, 2, 3, 4, 5, 6, 7, 8, 9

Gainor, Brian 1, 2
Gallagher, Gillian 1, 2
Gamkrelidze, Thomas 1, 2
García, Pedro 1, 2, 3, 4
Gårding, Eva 1, 2, 3, 4, 5, 6
Gardner, Matt Hunt 1n, 2, 3
Garvey, Kelly 1, 2
Gastner, Michael 1, 2
Gazdar, Gerald 1, 2
Genetti, Carol 1
Gerdemann, Dale 1, 2
Gewirth, Letitia 1n, 2
Ghini, Mirco 1, 2, 3
Gick, Bryan 1, 2
Giegerich, Heinz J. 1, 2
Gierut, Judith A. 1, 2
Gildea, Daniel 1, 2
Gili-Fivela, Barbara 1
Gilleron, Rémi 1
Girard, Gabriel 1
Goedemans, Rob 1, 2, 3
Gold, E. Mark 1, 2
Goldsmith, John A. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18
Goldstein, Louis 1, 2, 3, 4, 5
Goldwater, Sharon 1, 2
Gomez-Imbert, Elsa 1, 2
Good, Erin M. 1
Gordon, Matthew K. 1, 2, 3n, 4, 5, 6, 7, 8n, 9, 10, 11, 12, 13
Gorman, Kyle 1n, 2
Gosy, Maria 1, 2, 3
Gouskova, Maria 1, 2
Grabe, Esther 1, 2, 3, 4
Graf, Thomas 1, 2n, 3, 4, 5, 6
Gravina, Richard 1, 2, 3, 4
Green, Jenny 1n, 2
Greenberg, Joseph H. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13
Grice, Martine 1n, 2n, 3, 4, 5, 6
Grimm, Jakob 1
Guimarães, Maximiliano 1, 2
Gussenhoven, Carlos 1, 2, 3, 4, 5, 6, 7, 8, 9n, 10, 11, 12, 13, 14n, 15, 16, 17,
18, 19, 20, 21, 22, 23

Haarmann, Harald 1
Hagège, Claude 1, 2, 3n
Hale, Mark 1, 2, 3, 4
Hall, Daniel Currie 1n, 2, 3
Hall, Robert A. 1, 2
Hall, Tracy Alan 1, 2, 3, 4, 5n, 6, 7, 8, 9, 10, 11
Halle, Morris 1, 2, 3, 4, 5, 6n, 7, 8, 9, 10n, 11, 12n, 13, 14, 15, 16, 17, 18, 19,
20, 21, 22, 23, 24, 25, 26, 27, 28n, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39,
40, 41, 42, 43, 44, 45, 46
Hammarberg, Björn 1, 2, 3
Hammond, Michael 1, 2
Hancin-Bhatt, Barbara 1, 2, 3
Hannahs, S. J. 1
Hansen, Jette G. 1, 2, 3, 4
Hansson, Gunnar 1, 2, 3, 4, 5
Harnad, Stevan 1, 2, 3
Harrington, Jonathan 1, 2
Harris, James 1, 2
Harris, John 1, 2, 3
Harris, Zellig 1
Harst, Sander van der 1, 2
Harvey, Christopher 1, 2, 3
Haspelmath, Martin 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
Haudricourt, André-Georges 1
Havers, Wilhelm 1
Hayes, Bruce 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20
Healey, P. 1
Heidolph, Karl Erich 1, 2
Heine, Bernd 1, 2, 3
Heinz, Jeffrey 1, 2, 3, 4, 5, 6
Helgason, Pétur 1, 2, 3
Hermes, Anne 1, 2
Heyer, Sarah 1, 2
Higuera, Colin de la 1, 2, 3, 4
Hildebrandt, Kristine A. 1
Hillenbrand, James 1, 2
Hjelmslev, Louis 1
Hockett, Charles F. 1, 2, 3, 4, 5, 6
Hognestad, Jan K. 1, 2, 3
Honey, P. J. 1
Honeybone, Patrick 1, 2, 3, 4, 5, 6
Honti, László 1, 2, 3, 4, 5
Hoogeboom, Hendrik Jan 1, 2, 3
Hopper, Paul J. 1, 2
House, Anthony B. 1
House, David 1, 2
Householder, Fred 1
Howard, Irwin 1, 2
Howie, J. M. 1, 2
Hsieh, Feng-Fan 1, 2
Hualde, José Ignacio 1, 2, 3, 4, 5, 6, 7
Huang, James 1, 2
Huber, Brad R. 1, 2
Huffman, Marie 1, 2
Hulden, Mans 1, 2, 3
Humboldt, Wilhelm von 1, 2, 3
Hume, Elizabeth 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19,
20
Hwangbo, Hyun Jin 1
Hyde, Brett 1, 2
Hyman, Larry M. 1, 2, 3, 4, 5, 6, 7, 8n, 9, 10, 11n, 12n, 13, 14, 15, 16, 17, 18,
19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 3
9, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55

Idsardi, William 1, 2, 3, 4, 5, 6, 7
Ineichen, Gustav 1
Ingrisano, Dennis R. 1, 2
Isidore of Seville 1
Itô, Junko 1, 2, 3, 4, 5
Ivanov, Vyacheslav 1
Iverson, Gregory 1, 2, 3

Jackendoff, Ray 1, 2
Jacobi, Irene 1, 2
Jacobs, Haike 1, 2, 3, 4, 5
Jacobson, Steven A. 1
Jäger, Gerhard 1, 2
Jakobson, Roman O. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
18, 19, 20, 21
Jakovlev, N. F. 1, 2, 3
Janda, Richard D. 1, 2, 3
Janker, Peter M. 1, 2
Jardine, Adam 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16
Jenkins, James J. 1, 2
Jeon, Hae-Song 1, 2
Jespersen, Otto 1, 2, 3, 4
Jessen, Michael 1, 2, 3
John, Tina 1, 2
Johnson, C. Douglas 1, 2, 3
Johnson, Keith 1
Johnson, Mark 1, 2
Jönsson-Steiner, Elisabet 1, 2
Jun, Sun-Ah 1, 2, 3, 4
Jung, Dagmar 1, 2
Jurafsky, Daniel 1, 2, 3, 4
Jurgec, Peter 1, 2

Kabak, Barıș 1, 2
Kager, René 1, 2, 3, 4
Kahl, Thede 1
Kahn, Daniel 1, 2, 3
Kaisse, Ellen M. 1, 2, 3
Kallstenius, Gottfrid 1, 2
Kang, Yoonjung 1, 2
Kaplan, Abby 1, 2
Kaplan, Lawrence D. 1, 2, 3
Kaplan, Ronald 1, 2, 3
Karlsson, Anastasia M. 1
Kartunnen, Lauri 1, 2, 3
Karvonen, Dan 1, 2
Kaschube, Dorothea 1, 2
Kasprzik, Anna 1
Kaufmann, Stefan 1
Kaun, Abigail Rhoades 1, 2, 3
Kawahara, Shigeto 1, 2, 3, 4, 5, 6, 7
Kawasaki-Fukumori, Haruko 1, 2, 3, 4
Kay, Martin 1, 2, 3
Keating, Patricia 1, 2, 3, 4, 5, 6, 7, 8, 9
Kenstowicz, Michael 1, 2, 3, 4, 5, 6, 7, 8, 9
Kessler, Brett 1n, 2, 3, 4, 5, 6
Ketrez, Nihan 1
Keyser, Samuel Jay 1, 2, 3, 4, 5, 6, 7, 8
Kibrik, A. E. 1
Kim, Susan 1, 2, 3
Kinloch, Murray 1
Kiparsky, Paul 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17
Kirchner, Robert M. 1, 2, 3
Kisseberth, Charles W. 1, 2, 3, 4, 5, 6, 7, 8
Kleber, Felicitas 1, 2
Ko, Seongyeon 1n, 2
Kobele, Gregory 1, 2, 192ko
Kochetov, Alexei 1n, 2n, 3
Kock, Axel 1, 2
Kodzasov, S. V. 1
Kohler, Klaus J. 1, 2
Köhnlein, Björn 1, 2, 3, 4
Kolly, Marie-José 1
König, Ekkehard 1
Korhonen, Mikko 1, 2, 3
Korn, David 1, 2, 3
Kornai, András 1, 2, 3
Koskenniemi, Kimmo 1, 2
Kötzing, Timo 1
Krämer, Martin 1, 2
Kristoffersen, Gjert 1, 2, 3, 4, 5
Kruckenberg, Anita 1, 2
Kuaševa, T. X. 1
Kubozono, Haruo 1, 2, 3, 4, 5, 6, 7, 8
Kuipers, Aert H. 1, 2, 3
Kula, Nancy C. 1, 2
Kumaxov, M. A. 1, 2, 3

Labov, William 1, 2
Lacy, Paul de 1, 2, 3, 4, 5, 6
Ladd, D. Robert 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
Ladefoged, Peter 1, 2
Lahiri, Aditi 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16
Lai, Regine 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18
Lautemann, Clements 1, 2
Lavoie, Lisa M. 1, 2
Laycock, D. C. 1, 2
Lazard, Gilbert 1
Leary, Adam 1, 2, 3
Leben, William R. 1, 2
Lee, Seunghun 1, 2, 3, 4
Leeman, Dylan 1
Leemann, Adrian 1
Lehiste, Ilse 1, 2
Leira, Vigleik 1, 2
Lekeneny, Jean 1
Lerdahl, Fred 1, 2
Levelt, Clara 1, 2, 3, 4
Levinson, Stephen C. 1, 2, 3, 4, 5
Li, Bing 1n, 2
Li, Shulan 1, 2
Liberman, Anatoly 1, 2, 3
Liberman, Mark 1, 2, 3, 4, 5
Lickley, Robin 1, 2
Lieber, Rochelle 1
Lightfoot, David 1
Lindblad, Per 1, 2, 3, 4
Lindblom, Björn 1, 2, 3, 4
Linhartova, Vendula 1
Linker, Wendy 1
Lister, Anthony C. 1
Lloyd James, Arthur 1, 2
Lombardi, Linda 1, 2, 3, 4, 5
Lorentz, Ove 1, 2, 3, 4
Lothaire, M. 1, 2, 3, 4
Louriz, Nabila 1, 2
Low, E. L. 1, 2, 3, 4
Luo, Huan 1, 2, 3

McCarthy, John J. 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 1
9
McCawley, James D. 1, 2, 3, 4
McCutcheon, Martin J. 1, 2
McKenzie, Pierre 1
MacMahon, April M. S. 1, 2, 3
McNaughton, Robert 1, 2, 3, 4
McQueen, James 1, 2
Mackenzie, Sara 1, 2
Maddieson, Ian 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19
Magloughlin, Lyra 1
Major, Roy 1, 2, 3
Mallinson, Graham 1
Manaster-Ramer, Alexis 1, 2
Marlo, Michael 1, 2
Marsico, Egidio 1
Martin, James 1, 2
Martin, Samuel 1, 2
Martinet, André 1, 2, 3, 4, 5
Martínez-Gil, Fernando 1, 2
Mascaró, Joan 1, 2, 3
Maskikit-Essed, Raechel 1, 2
Matasović, Ranko 1, 2
Mateus, Maria Helena 1
Matthews, Peter 1
Mchombo, Sam 1
Medvedev, Yu. 1, 2
Mehler, Jacques 1, 2
Melville, Herman 1
Mennen, Ineke 1
Merrill, John 1, 2, 3
Mesgnien-Meninski, François (de) 1, 2, 3
Mester, Armin 1, 2, 3, 4, 5
Mesthrie, Rajend 1, 2, 3
Metzeltin, Michael 1
Meyer, Ernst A. 1, 2
Mezei, J. E. 1, 2, 3
Mielke, Jeff 1, 2, 3, 4, 5, 6, 7, 8
Minde, Don van 1, 2, 3
Mithun, Marianne 1
Mjaavatn, Per Egil 1, 2, 3
Mohanan, K. P. 1, 2, 3, 4, 5
Mohanan, Tara 1, 2
Mohri, Mehryar 1, 2, 3
Møllergård, E. 1, 2
Moravcsik, Edith A. 1, 2, 3, 4, 5, 6, 7
Moreton, Elliot 1, 2, 3, 4, 5, 6, 7
Morpurgo Davies, Anna 1, 2
Motsch, Wolfgang 1
Mou, Xiaomin 1, 2
Moulton, William G. 1, 2
Moure, Teresa 1
Mücke, Doris 1
Mugler, France 1, 2
Munroe, R. H. 1
Munroe, Robert L. 1, 2, 3
Myers, Nathan 1
Myers, Scott 1, 2, 3, 4, 5, 6, 7, 8
Myrberg, Sara 1, 2n, 3, 4, 5, 6

Nagarajan, Hemalatha 1, 2
Naydenov, Vladimir 1, 2
Nedjalkov, Vladimir P. 1
Nespor, Marina 1, 2, 3, 4, 5
Nevins, Andrew 1, 2, 3, 4, 5, 6, 7n, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17
Newman, M. E. J. 1, 2
Newman, Stanley 1, 2, 3
Newmeyer, Frederick J. 1, 2
Ní Chiosáin, Máire 1, 2
Nichols, Johanna 1, 2, 3, 4, 5, 6
Nilsen, Randi Alice 1, 2, 3
Nolan, Francis 1, 2
Noord, Gertjan van 1, 2
Nooteboom, Sieb 1, 2, 3
Nordberg, Bengt 1, 2n, 3
Norris, Daniel 1
Nyström, Staffan 1, 2

Ó Siadhail, Mícheál 1n, 2

O’Dell, Michael 1, 2
Odden, David 1, 2, 3, 4n, 5, 6, 7, 8, 9
Oestereicher, Wulf 1
Oh, Mira 1, 2
Ohala, John J. 1, 2, 3, 4, 5, 6, 7, 8, 9
Öhman, Sven 1, 2, 3
Olovjannikova, I. P. 1
Oncina, José 1, 2, 3, 4
Oostendorp, Marc van 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
Osu, Sylvester 1, 2, 3, 4
Otanes, Fe T. 1, 2
Oxford, Will 1, 2, 3, 4

Padgett, Jaye 1, 2, 3, 4, 5, 6, 7, 8, 9
Pāṇini 1n
Papert, Seymour 1n, 2, 3, 4
Paradis, Carole 1, 2, 3, 4
Parker, Aliana 1, 2
Pater, Joe 1, 2, 3, 4, 5
Paulian, Christiane 1
Pawley, Andrew 1, 2, 3, 4, 5, 6
Payne, Amanda 1, 2, 3
Pellegrino, François 1
Peng, Long 1, 2, 3
Pensalfini, Rob 1, 2, 3, 4
Peters, Jörg 1, 2, 3, 4, 5, 6, 7, 8
Pierrehumbert, Janet 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15n, 16, 17, 1
8, 19
Pike, Kenneth L. 1n, 2, 3, 4
Piroth, Hans Georg 1, 2
Plank, Frans 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14
Popper, Karl 1, 2
Port, Robert F. 1, 2, 3, 4, 5, 6, 7
Post, Mark W. 1
Pott, August Friedrich 1n
Potts, Christopher 1n, 2, 3
Precoda, Kristin 1n, 2, 3, 4, 5, 6, 7, 8
Prieto, Pilar 1, 2, 3
Prince, Alan 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
21, 22, 23
Prunet, Jean-François 1, 2, 3
Pullum, Geoffrey K. 1, 2, 3, 4, 5, 6, 7, 8
Purnell, Thomas 1, 2n, 3

Qinggertai (Chingeltei) 1, 2
Quené, Hugo 1, 2, 3

Rabin, Michael 1, 2
Raible, Wolfgang 1
Raimy, Eric 1, 2, 3n, 4
Ramat, Paolo 1
Ramus, Franck 1, 2, 3, 4, 5, 6
Raphael, Lawrence 1, 2
Rask, Rasmus 1
Rawal, Chetan 1
Reetz, Henning 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
Reiss, Charles 1, 2
Rennison, John R. 1
Riad, Tomas 1, 2, 3, 4, 5, 6, 7
Rialland, Annie 1, 2
Rice, Curtis 1
Rice, Keren 1, 2, 3, 4, 5, 6, 7, 8, 9n, 10, 11, 12, 13, 14, 15
Ridouane, Rachid 1, 2, 3, 4, 5, 6
Riggle, Jason 1, 2, 3, 4, 5, 6, 7, 8, 9
Ringen, Catherine 1, 2, 3, 4, 5, 6
Ringgaard, Kristian 1, 2
Rischel, Jørgen 1
Roark, Brian 1, 2
Roberts, Adam 1, 2
Roberts, Ian 1
Robins, R. H. 1
Roca, Iggy 1, 2
Roche, James 1, 2, 3
Roeder, Rebecca 1, 2
Rogers, James 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12
Rogova, G. B. 1
Rohany Rahbar, Elham 1, 2
Rood, David S. 1n, 2
Rose, Sharon 1, 2, 3, 4, 5, 6
Rozenberg, Grzegorz 1, 2
Rubach, Jerzy 1, 2, 3, 4
Ruiz, José 1, 2
Ryan, Kevin 1, 2

Sagey, Elizabeth C. 1, 2, 3, 4, 5, 6, 7
Šagirov, A. K. 1, 2
Sakarovitch, Jacques 1, 2
Salmons, Joseph 1, 2, 3, 4, 5, 6
Salomaa, Arto 1, 2
Samedov, D. S. 1
Sammallahti, Pekka 1, 2, 3
Samuels, Bridget 1, 2n, 3, 4, 5, 6, 7
Sapir, Edward 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11
Satta, Giorgio 1, 2, 3, 4
Saussure, Ferdinand de 1, 2
Schabes, Yves 1, 2, 3
Schachter, Paul 1, 2
Scheibman, Joanne 1, 2
Schepman, Astrid 1
Schiering, René 1, 2
Schmidt, Jürgen Erich 1, 2
Schwartz, Geoffrey 1n, 2
Schwartz, Jean-Luc 1, 2
Schwentick, Thomas 1
Scobbie, James M. 1, 2, 3n, 4, 5
Scott, Dana 1, 2
Segerup, My 1, 2n, 3, 4
Seiler, Hansjakob 1, 2
Selkirk, Elisabeth 1, 2, 3, 4, 5, 6
Shattuck-Hufnagel, Stefanie 1
Shaw, Patricia 1, 2, 3, 4
Shibatani, Masayoshi 1, 2
Shieber, Stuart 1, 2, 3, 4
Shopen, Timothy 1, 2
Siewierska, Anna 1
Silander, Megan 1, 2
Simon, Ellen 1, 2, 3, 4, 5
Simpson, Adrian P. 1n, 2
Singler, John 1, 2
Sipser, Michael 1, 2
Siptár, Péter 1
Skalička, Vladimír 1
Slobin, Dan 1, 2
Smith, Adam 1, 2
Smith, Bruce L. 1, 2
Smith, Geoff 1, 2
Smith, Henry Lee 1, 2
Smith, Nathaniel 1
Smith, Neil 1
Smith, Norval 1, 2, 3
Smith, Steven C. 1
Smith, Tony 1
Smolensky, Paul 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17
Sommer, Bruce A. 1, 2
Song, Jae Sung 1, 2
Spencer, Andrew 1
Sproat, Richard 1, 2
Stampe, David 1, 2
Stavness, I. Ian 1
Steinitz, Wolfgang 1, 2, 3, 4
Steriade, Donca 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14
Stevens, Kenneth N. 1, 2, 3, 4, 5, 6, 7n, 8, 9, 10, 11
Strandberg, Mathias 1, 2, 3, 4
Strange, Winifred 1, 2
Strother-Garcia, Kristina 1
Stuart-Smith, Jane 1, 2, 3n, 4, 5
Suzuki, Keiichiro 1, 2, 3, 4
Svantesson, Jan-Olof 1, 2, 3, 4
Sweet, Henry 1, 2, 3
Tabain, Marija 1, 2
Talkin, David 1, 2
Tallerman, Maggie 1, 2
Talmy, Leonard 1, 2
Tangi, Oufae 1, 2
Tanner, Herbert G. 1
Teleman, Ulf 1, 2
Tent, Jan 1, 2
Tesar, Bruce 1, 2, 3, 4
Thomas, Wolfgang 1, 2, 3
Thráinsson, Höskuldur 1
Timm, Jason 1
Topinzi, Nina 1, 2n, 3
Törkenczy, Miklós 1
Trager, George 1, 2
Trask, R. L. 1, 2
Trommelen, Mieke 1, 2, 3
Trubetzkoy, Nikolaj S. 1, 2, 3, 4, 5, 6, 7n, 8, 9, 10, 11, 12, 13, 14, 15, 16
Tsagov, M. 1, 2
Tsendina, Anna 1
Tukumu, Simon Nsielanga 1
Turčaninov, G. 1, 2
Turing, Alan 1, 2, 3, 4
Turpin, Myfany 1, 2

Ulseth, B. 1, 2
Urbina, Jon Ortiz de 1

Vaan, Michiel de 1, 2
Vago, Robert 1, 2, 3
Vajda, Edward 1, 2, 3, 4
Vallée, Nathalie 1, 2
Vance, Timothy J. 1, 2, 3, 4
Vanhove, Martine 1
Vaux, Bert 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13
Veilleux, Nanette 1
Velupillai, Viveka 1
Ven, Marco van de 1, 2
Venditti, Jennifer J. 1, 2, 3
Vennemann, Theo 1, 2, 3, 4
Vergnaud, Jean-Roger 1, 2, 3, 4
Verner, Karl 1
Versteegh, Kees 1, 2
Vidal, Enrique 1, 2
Visscher, Molly 1, 2
Vliet, Pete van der 1, 2, 3, 4, 5
Vogel, Irene 1, 2, 3
Vollmer, Heribert 1
Vu, Mai Ha 1

Walker, Rachel 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
Wang, Chilin 1, 2, 3, 4, 5, 6, 7, 8, 9
Watson, Janet 1
Weijer, Jeroen van de 1, 2, 3, 4
Weinberger, Steven 1, 2, 3
Wellcome, David 1, 2
Wells, John 1, 2
Westbury, John 1, 2, 3
Wetterlin, Allison 1, 2, 3, 4, 5
Wetzels, Leo 1, 2, 3, 4, 5
Whaley, Lindsay J. 1, 2
Wheeler, Max 1, 2
Whorf, B. L. 1
Wibel, Sean 1
Wiese, Richard 1, 2, 3, 4, 5, 6, 7
Wilson, Colin 1n, 2, 3, 4, 5, 6, 7, 8, 9, 10
Wiltshire, Caroline 1, 2, 3
Winteler, Jost 1
Winters, Stephen J. 1, 2, 3
Wissing, Daan 1, 2, 3
Wolfe, Andrew 1, 2, 3, 4
Wood, Sidney 1, 2
Wurzel, Wolfgang 1, 2, 3

Xolodovič, Aleksandr A. 1
Xrakovskij, Viktor S. 1
Xu, Zheng 1, 2

Yavas, Mehmet 1, 2, 3, 4
Yu, Alan C. 1, 2, 3, 4, 5, 6, 7, 8, 9

Zagona, Karen 1
Zanten, Ellen van 1, 2, 3
Zhang, Xi 1, 2, 3, 4, 5, 6, 7, 8, 9n, 10, 11
Zonneveld, Wim 1, 2, 3, 4, 5
Zwart, Jan-Wouter 1
Zwicky, Arnold M. 1
Endnotes
1
In the UPSID database (Maddieson & Precoda 1992) I have however not found a
language which only has the two places of articulation, labial and alveolar. For
accessing UPSID I have used Henning Reetz’s online interface: http://web.phoneti
k.uni-frankfurt.de/upsid.
2
Since submitting this chapter, Gordon (2016) has appeared who defines
phonological typology as follows: “Phonological typology is concerned with the
study of the distribution and behavior of sounds found in human languages of the
world” (p. 1).
3
As a test, consult the World atlas of language structure (WALS, http://wals.info/)
and decide for yourself whether the variables labeled “Phonology” are phonological
or phonetic, as you would draw the line.
4
Syntax-and-inflection typologists will sometimes, upon reflection, express surprise
that they had not noticed before how little phonology there was to be met with in
their own circles. Perceptions similar to mine have been reported in Hyman
(2007), and also in this volume.
5
This whole subsection 3.1 draws on earlier historiographic writings of mine, in
particular Plank (1991, 1992, 1993, 1998, 2001), where the reporting and
referencing are more conscientious. My history is intended as an “inside narrative”
(to borrow a term of Herman Melville’s), and thereby differs from historians’
histories of typology, which standardly begin with ninetheenth century German
Romanticism, a period that in my view produced little of substance that was novel
or profound. Let me just emphasise that the typological speculations and insights
referred to here were not private musings or only shared in private correspondence
(subsequently buried in archives, if surviving at all), but saw contemporary
publication.
6
Beauzée was employed by the École Royale Militaire in Paris, where they still
maintain a centre linguistique (http://www.rma.ac.be/clng/fr/index.html).
Although, unlike his latter-day successors, Beauzée was a grammairien rather
than a grammatiste (language teacher), his job was not to train future linguists of
the kind he himself was one. Probably August Friedrich Pott (1802–87) – better
known as a historical-comparative Indo-Europeanist despite his numerous
Humboldtinspired contributions to typology – was the first to have been trained as
a general linguist (at least insofar as his doctoral dissertation at Göttingen was
about general linguistics, dealing with the semantics of prepositions across
languages), and whose academic responsibilities at the university of Halle an der
Saale then included the training of future general linguists (Plank 1995).
7
A “conjuncture” of vowel harmony and agglutinative morphology, where word
cohesion is otherwise rather loose, was suspected by Jan Baudouin de Courtenay in
the 1870’s.
8
Morpurgo Davies (1997) is a masterful portrait of this century, an “inside
narrative” in a class of its own. Morpurgo Davies (1975) highlights the relationship
between historical and typological comparison. For further details also see Plank
(1991, 1995).
9
And sounds mattered even before one got started: there was an inbuilt historical
limitation to the Comparative Method, insofar as cognates, if not lost, would at
some point (after 8,000 years or so) become impossible to recognise, as the sound
shapes of morphemes would inexorably change over time.
10
A similar point has been made on similar grounds by Hyman (2007) (and
elsewhere, including in this volume).
11
These research initiatives are instructively portrayed by group leaders in Shibatani
& Bynon (1995). Paris did see groundbreaking research on phonological typology,
namely work centred around André-Georges Haudricourt’s Phonologie
panchronique (1940 etc., with an interim summary in La phonologie
panchronique by Claude Hagège & Haudricourt, Paris: Presses Universitaires de
France, 1978); but it was only later that this attained formal project status.
12
Not without initial opposition: some reviewers of the programme proposal sought
to block it as pointless; for them, the only respectable comparison was historical.
13
Regrettably, over the centuries, negative results – demonstrations that variables do
not co-vary – have continuingly been deemed less worthy of reporting and
recording.
14
Only the programme committees will know how the rejection rates compared for
phonological and morphosyntactic abstracts. There was probably never a bias
against phonology at the stage of abstract selection; but there were very few
acknowledged phonologists on these ALT committees, and common sense suggests
that the perceived expertise of abstract selectors is a factor encouraging or
discouraging abstract submission.
15
Details, so far as they could be recovered other than from memory, at: https://ling
uistlist.org/issues/9/9-874.html; http://listserv.linguistlist.org/pipermail/alt/200
2-November/000039.html; http://typoling2016.sciencesconf.org/.
16
Figures supplied upon request. These figures would be similar for organisations
devoted to particular language families. In terms of specialised journals or also
specialised conferences, however, syntax does not seem far ahead of phonology
and morphology. Significantly, phonetics is the clear winner in this respect,
essentially forming a professional world of its own. Even passable all-rounders in
linguistics can yet be useless in phonetics (and vice versa).
17
In other grammars following this format, not written by phonologists, the
phonology sections are substantially shorter.
18
Jespersen, Sapir, and Bloomfield were all-rounders, but had substantial phonetic
or phonological work to their credit. Hyman has morphosyntax as a sideline. Evans
is a part-timer, but the best to be had among Australianists for present purposes.
19
On current evidence, the contributors to the present volume, with a single
exception (grammar-writing Hyman), are thus in the company of the likes of Jan
Baudouin de Courtenay, Mikołaj Kruszewski, Ferdinand de Saussure, Nikolaj S.
Trubetzkoy, Roman Jakobson, John Rupert Firth, Louis Hjelmslev, André
Martinet, Kenneth L. Pike, Morris Halle (who did write and co-author what could
have formed the phonology chapters of the grammars of two languages, Russian
and English).
20
Don’t be misled by dialect grammars: they tend to background what dialects
supposedly do not much differ in, namely syntax. For example, Die Kerenzer
Mundart des Kantons Glarus in ihren Grundzügen dargestellt by Jost Winteler
(Leipzig: Winter, 1876), renowned for its innovative phonology, devotes 147 pages
to this subject, 43 to inflection, and not one to the syntax of this Swiss variety of
Alemannic, a dialect of High German.
21
Heinz’ chapter in this volume is in the same spirit, but limits complexity
comparisons to phonological patterns.
22
If you define typology as only being about diversity, as is sometimes done in
phonological circles (and elsewhere), then matters are of course different. More on
this point presently, and also in Hyman in this volume.
23
The admirable inside history of phonology by Anderson (1985) doesn’t quite
highlight this theme of languages and typology in phonology vs. syntax. Nor does
the more recent collection of Honeybone & Bermúdez-Otero (2006), where
similarities between phonology and syntax are emphasised, rather than possible
differences in researching phonological and syntactic structures and architectures.
24
This overly spartan mode of description arguably prevented such typological
syntax from meaningfully engaging with diachronic syntax and from playing a
more significant role in experimental psycho- and neurolinguistics.
25
After Greenberg’s “dynamicised” typology and Haudricourt’s “panchronic
phonology”, the most determined single effort to explain co-variation as co-
evolution in phonology was Juliette Blevins’ Evolutionary phonology (Cambridge:
Cambridge University Press, 2007). Characteristically, the impact of this book was
felt more in phonology, theoretical as well as historical, than in typology, while
similar evolutionary work in morphosyntax typically had stronger reverberations
in typology than in syntax.
26
The by far longest phonology, at 624 pages, is that of Danish, and you wonder
whether to give credit to the language or the author (Hans Basbøll) for such
unusual opulence. Portuguese phonology only needs 170 pages (from Maria
Helena Mateus & Ernesto d’Andrade), almost as little as Chichewa syntax does
(166, Sam Mchombo); and it is moot to speculate whether such frugality reflects on
these languages or the describers of their phonology and syntax.
27
Admittedly it has been recognised that often what has been sampled were really
lower-level units such as “doculects”, namely those varieties of a language that
happen to have been documented (in a text or corpus, by a fieldworker, in a
particular grammar).
28
This useful work is often somewhat misleadingly called a “database” of “sound
inventories”. It contains phoneme inventories, which can be at considerable
remove from the primary data. Really they are theories, just as grammars are not
primary data but theories of languages.
29
UPSID’s convention of representing phonemes by their most frequent allophone,
rather than by an invariant feature bundle, could be seen as a partial
acknowledgment of this.
30
This is inconsistent with the contrastive hypothesis, according to which
phonological generalizations can only refer to contrastive features (Currie-Hall
2007, Dresher 2007).
31
Moloko allows as medial codas only the most sonorous consonants, the non-nasal
sonorants /r/, /l/, /w/ or /j/. Violations are eliminated by inserting /ǝ/.
32
I include the labial prosodies y and w within the phonemic representation, but they
are not phonemically attached to the last segment.
33
Perhaps through merger with an a-colored consonant inserted to satisfy ONSET.
34
See below for a defense of Jakobson’s generalization against claims that some
languages have only VC(C) syllables (Section 2.1) and that some languages have no
syllables (Section 2.2).
35
Lexical words should be distinguished from postlexical words formed by syntactic
processes such as cliticization, which are only subject to the postlexical phonology.
36
At the stem level, on the other hand, structure-preservation is a theorem of Stratal
OT, because the phonological inventory and stem structure of a language derives
from its stem-level constraint system. No special structure-preservation principle
is needed. Note further that with the equivalent update, rule-based Lexical
Phonology can provide essentially the same kind of rich word phonology as Stratal
OT.
37
REMOVING it by syllabification and desyllabification does; this will become
important below.
38
Such near-mergers had been reported in the earlier dialectological literature,
though their significance remained unappreciated. For example, DeCamp (1958,
1959) notes near-merger of four and for in what was then old-fashioned San
Francisco speech, since then replaced by complete merger.
39
These terms have also been used to refer to contextually restricted contrasts, such
as Spanish [r]:[ɾ] intervocalically (Hualde 2005), or Italian [ɛ]:[e] (only stressed
syllables), as well as to marginal, “fuzzy” contrasts (Scobbie & Stuart Smith 2008).
Currie-Hall (2013) sorts out these various uses.
40
There are of course vertical subsystems consisting of minimally complex reduced
central vowels, such as English /ɨ/ and /ǝ/ (Rosa’s roses). Irish has a subsystem of
three short vowels /ɯ/, /Ѳ/, /a/, plus six long vowels /i:/, /Ѳ:/, /u:/, /e:/, /a:/,
/o:/ (Ó Siadhail 1989: 35–37). /ɯ/, /Ѳ/ have back allophones [u] and [o]
respectively before broad (velarized) consonants and front allophones [i] and [e]
before slender (palatalized) consonants, e.g. /lʹɯm/ → [lʹum] ‘with me’, /lʹɯNʹ/ →
[lʹiNʹ] ‘with us’.
41
Even in vertical systems, when the non-low vowels are not colored by a consonant
or prosody, they are often front rather than central. In the two-vowel system of
Arrernte, the non-high vowel appears as [i] in initial position where there is no
consonant to influence it, and Hale therefore set it up as /i/ (quoted in Green 1994:
35). Wichita has a three-vowel system /i/, /e/, /a/, with three degrees of length;
phonetically also [o] and [u]. /i/ ranges between [i] and [e], /e/ between [ɛ] and
[æ], and /a/ between low back unrounded [ɑ] and (when short) [ʌ] as in but, with
rounding next to /w/, rarely [u] (Rood 1975). In the variety of Kabardian described
by Colarusso (1992, 2006), the vowels transcribed as [ǝ] and [ɨ] are actually front
vowels.
42
These are nominal exceptions to Stieber’s Law, which says that allophonic features
cannot spread by analogy (see Manaster-Ramer 1994). But if Stieber’s Law is taken
as a generalization about l-phonemes, it may well be exceptionless.
43
For example, artificial and beneficial rhyme, even though they differ underlyingly
({-s-} vs. {-t-}), and keep and coop alliterate, even though their initial consonants
differ phonetically in backness.
44
The same is true of the similar earlier claim about the syllable structure of the
Kunjen dialects (Sommer 1970a, 1970b). As Sommer makes clear in the latter
article, their output syllabification actually conforms to Jakobson’s CV
generalization.
45
Two other ways to simplify stress have been proposed, also with strange syllable
structure. Topintzi & Nevins (2017) make initial consonants in Arrernte moraic,
with stress falling on the second mora, so that [mǝ´ɳǝ] is /mµǝ´µ.ɳǝµ/ and [e.nǝ
´.kǝ] is /eµ.nǝ´µ.kǝµ/. Schwartz (2013) assigned vowels a “vocalic onset node” (≈
null onset), with other onset consonants being excluded by *COMPLEXONSET.
46
On one analysis onsetless syllables can be adjoined to an adjacent syllable to form
a “sesquisyllabic” complex (Kiparsky 2003).
47
Orthographic rr denotes an alveolar tap or trill, r a retroflex approximant [ɻ]
(transcribed as r. in Pensalfini’s and Breen’s work). rn, rt are retroflex [ɳ], [t]. The
orthography uses h to mark dental place of articulation in th, nh etc. I have
replaced them by the IPA symbols to prevent confusion with aspiration.
48
Compare stress-sensitive root allomorphy in Italian, e.g. vádo, andáte, andáre
(Kiparsky 1996).
49
The last two examples are from Breen; thanks to Toni Borowsky for passing them
on.
50
The sources don’t reveal the stress of the Rabbit Talk forms; my guess is that they
stay on the same syllable as in the original word, e.g. áŋkemamp.
51
Another case where theorizing has been led astray by a misconstrual of abstract
phonemic and morphophonemic representations as phonetic transcriptions are
Tundra Nenets word-final stops (Kiparsky 2006).
52
This is not to say that syllabicity has no phonetic correlates. For example, Fougeron
& Ridouane (2008) find that Tashlhiyt syllabic consonants are not longer than non-
syllabic consonants, but they are less coarticulated.
53
Tellingly, even Pa¯ṇini, whose rich descriptive apparatus includes phonological and
morphological features, ordered rules, constraints, blocking, Theta-roles, linking,
and inheritance hierarchies, among others, did not use syllables, even though
Sanskrit very clearly has them (Kessler 1994); they simply would not have made his
grammar shorter. This is a case of the ICEBERG PROBLEM, fatal for the project of
“describing each language on its own terms”: a single language, however rich and
precise the description, cannot reveal all aspects of UG.
54
A caveat: syllable structure can change in the course of a derivation. The more
careful formulation has to be that it must be consistent at any given level of
representation. For example, in English rhythm, spasm, plasm are monosyllabic at
the stem level. If the nasal were syllabic at this level, it would get stressed in words
like rhýthm-ic (cf. átom, atómic), and words like éctoplàsm, ángiospàsm,
cýtoplàsm, cátaclàsm, hólophràsm, would be stressed on the second part, on the
pattern of èctopárasite, àngiothlípsis, èndotóxin, cy`tocóccus, càtatónia,
hòmophóbia, hòmomórphism, rather than on the pattern of éctomòrph,
ángiospèrm, éndolymph, cý → cyst, péricàrp, hómophòbe, hómomòrph. Spanish
[je] is a diphthong in the lexical phonology and behaves as a heavy syllable for
purposes of stress, but postlexically [j-] is resyllabified from the nucleus into the
onset, as shown by its allophonic realization (Harris & Kaisse 1999).
55
Hyman suggests that this distribution could be due to the lack of inputs that would
yield *CVVVCV, *CVCVVV (an accident or a conspiracy?). For example, *CVCVVV
must be of the form Root + Derivational suffix + Inflectional suffix, and this can be
neither /CVC-V-VV/, because this sequence would undergo vowel shortening, nor
/CVC-VV-V/, because derivational suffixes must be minimal syllables of the form -
(C)V.
56
I am grateful to Junko Itô and Stefan Kaufmann for information on Japanese.
57
Similarly the attempt to replicate the phonology of the source language in
loanwords, as in Labrune’s example baiorin˚ /baioriN/ ‘violin’.
58
Of course positive arguments against a VP constituent were sometimes adduced,
one being that subjects and objects c-command each other for purposes of
Condition C of Binding Theory. However, these arguments turned out to be fragile,
and the formulation of Condition C on which it relies has itself been questioned.
59
Or even to no vowels, as Barreteau (1988: 429–437) does for Mofu-Gudur, also a
Chadic language. His analysis predicts the surface vowels just from the consonants,
prosodies, and tone.
60
This has been generally accepted at least since Jakobson, Fant, & Halle (1952),
irrespective of whether syllabicity is represented featurally as in Chomsky & Halle
(1968), or by position in a syllabic constituent as in Kahn (1976), Gussenhoven
(1986), and later work.
61
(24a-j) are from B&P, (24k, l) are from P&B.
62
The length mark is inadvertently (I think) omitted by the authors in (24h,i,j), where
I have inserted it.
63
Words beginning with a central vowel are excluded because the central vowel is only
added after otherwise unsyllabifiable consonants.
64
On the other hand, it is conceivable that final high vowels are actually pronounced
with an offglide, as in English, where a word like bee [biy] is transcribed as [bi:].
65
The transcriptions in (28) are inferred from the information in P&B (2011: 30).
66
B&P’s remark in this connection that “technical problems within OT grammars can
always be solved by invoking additional constraints” is unduly dismissive. Actually
the data are predicted by the theory, in the sense that if monoconsonantal words did
not undergo epenthesis, the analysis would require otherwise unwanted constraints,
as the reader can verify.
67
This is not inconsistent with their formulation that vowels are inserted after
consonants that require a release. It may be that syllable-final consonants must be
released. Another question is whether it is justified to attribute the release property
not only to plosives, but also to nasal stops and /l/, as B&P have to do; only the
continuants /s/, /y/, and /w/ are licit medial codas. Phoneticians normally use the
term “release” for the separation of articulators in plosives, and explain that
plosives prefer syllable onsets because that is where their release burst is most
easily preceptible. The prohibition of sonorant stops in word-internal codas in
Kalam cannot be explained the same way, because they are not released with a noisy
burst, and would be easily perceptible even in coda position, and are in fact
common as codas across languages.
68
The two-vowel analysis is originally due to Kuipers (1960). After presenting it, he
goes further by eliminating first /ɐ/ by doubling the consonant inventory with a set
of ɐ-colored consonants, and then, in an extreme tour de force, eliminates the
remaining vowel /ǝ/ by enriching the phonemic representation with an abstract
juncture marker “:”. In this analysis, not only were all phonemes consonants, but
every consonant was a morpheme, and every morpheme was a consonant. It was
conclusively refuted by Halle (1970), Kumaxov (1973), and Colarusso (1982), and
has found no followers since.
69
It is always tautosyllabic, if Colarusso (1992: 15) is right that all intervocalic
consonants are ambisyllabic, e.g. /dǝda/ [ˈdi˺.dæ].
70
E.g. /Ɂa/ [Ɂæ] ‘hand’, /psǝɁa/ [ˈpsεɁæ] ‘wetness’, [ˈpsε>Ɂ˺ ˌɁæ>] in Colarusso’s
narrower transcription. They are represented as central vowels [ǝ] and [ɐ] in
Gordon & Applebaum (2006), which would be more consistent with the
expectations of Dispersion Theory (Flemming 1995, 2016, Vaux & Samuels 2015).
But in the samples I have heard they are definitely front in agreement with
Colarusso’s description: https://www.youtube.com/watch?v=gtuU5_U-gL4, http
s://www.youtube.com/watch?v=4-BY1vYfM_Q, https://www.youtube.com/watch?
v=r_qQCUDaz-I.
71
Though in practice he writes the phonemes as /i,i,e,a/.
72
The data in (39) undermine Hale’s (2000) claim that the vowels of Marshallese are
not only phonologically but PHONETICALLY underspecified for backness and
rounding (he pointedly represents them at both levels with arbitrary dingbat
symbols), and that the twelve vowels in (36) and their diphthongal combinations
are introduced only in the acoustic/articulatory output. It would be hard to account
for the contrasts in (39) as resulting from coarticulation (at least under standard
assumptions about the phonology/phonetics interface). Deletion of glides would
also have to be an acoustic/ articulatory process, in counterbleeding order with
acoustic/ articulatory assimilation.
73
Choi (1992: 68) also concludes that the smooth transition between vowel qualities
must be due to phonetic coarticulation processes: the F2 trajectory for Marshallese
/tyeap/ ‘to return’ shows no steady-state position for the tongue during the
realization of the diphthong.
74
The exception is Pa¯ṇini’s Aṣṭa¯dhya¯yı¯, built from scratch strictly by using
minimum description length as the sole criterion for establishing both the
generalizations and the formalism in which these are expressed. This was done by
defining the technical terms and conventions of the system in the grammar itself, so
that minimum description length then requires that they are introduced if and only
if they reduce the overall length of the grammar — that is, if the minimum possible
cost of defining them is outweighed by the maximum possible grammatical
simplification they allow. Autochtonous philologies such as that of the Japanese
kokugakusha (Bedell 1968) and the Arabic tradition originating with Sibawayh
(Versteegh 1997) also describe their respective object language in its own terms, but
they were not comprehensive grammars in the modern sense. They were more
concerned with settling points of usage and philosophical issues than with
grammatical analysis per se.
75
The roadmap is not exhaustive. Notable earlier research which examines the nature
of phonological generalizations from a computational perspective but which will not
receive as much discussion as it should includes Potts & Pullum (2002) and Graf
(2010b), which also come from an intellectually similar perspective.
76
It is true that periodically some work is published in that direction, for example the
work on output-to-output correspondence (Benua 1995, 1997, and others).
77
A small minority of students suggest that vlas and sram might be English words but
they agree they are less sure about these than the others. For more on gradient
versus categorical distinctions in phonotactics, see Hayes & Wilson (2008) and
Gorman (2013). In this chapter, we assume a categorical distinction for expositional
purposes, but as discussed in Heinz (2010a) nothing really hinges on this.
78
In fact, both rule-based and OT grammars predict there to be complete
phonological grammars which only instantiate the process of inter-strident
epenthesis or word-final devoicing. The fact no known phonology only contains a
map which would correspond to a single traditional phonological rule has never
been taken as a problem for either rule-based theories or OT.
79
More formally, it is decidable whether or a not any particular string belongs to the
set. Interestingly, most logically possible sets of strings are not computably
enumerable (Turing 1937).
80
By the way, [∫toyonowonowa∫] means ‘it stood upright’ (Applegate 1972).
81
The relevant feature could also be [distributed].
82
Alan Yu also points out that there is a perceptually-motivated diachronic path to
arriving at this language (pace Ohala 1981 and Blevins 2004). A language like
Samala would be the precursor language to one with First/Last Harmony. Since
interior sibilants in the precursor language are not perceived as accurately as ones
at word edges, some of them may change over time to disagreeing sibilants. This
would result in a language whose words obey First/ Last Harmony.
83
The Non-Counting class also goes by the names Star-Free and Locally Testable with
Order (McNaughton & Papert 1971).
84
For those familiar with the FO formulae, here is the definition where x ⊲ y means y
is a successor of x, and x < y means x precedes y.
 def
x y = x < y  ∧ ¬ (∃z) [x < ∧ z < y]

85
For those familiar with MSO logic, here is a definition. Individual variables are
denoted with x, y, and X denotes a set variable. x ⊲ y means y is a successor of x,
and x < y means x precedes y.
def
closed (X) = (∀x, y) [(x ∈ X ∧ x y) → y ∈ X]

def
x < y = (∀X) [(x ∈ X ∧ closed (X) → y ∈ X]

86
To see why the set of strings L containing only an even number of sibilants is not
Non-Counting, the characterization in (18) can be used. For any k, observe that
os2ko belongs to L, but os2k+1o does not.
87
Finley & Badecker (2009) found no difference between the absence of training and
a control condition where the words in the training condition contained words
which were well-formed and ill-formed according to each targeted constraint type.
88
Readers may also wonder why the sub-structure that picks out the first/ last
template [#—. . .— #] is not available. Here the reason is simple: both the successor
and precedence relations allow every word to have a model and for distinct words to
have distinct models. This is not the case with the sub-structure indicated by the
template [#—. . . — #] is used to model words.
89
Bounded spreading may be more common with the feature nasal. For convenience,
the bound here is assumed to be the next relevant segment, but in fact the syllable
seems to be a natural domain (Odden 1994). See discussion in Nevins (2010,
Chapter 5).
90
Some work exists that characterizes string-to-string maps which correspond to the
NC stringsets (Lautemann et al. 2001). Also the author has work in progress
characterizing string-to-string maps for each of these regions.
91
http://pbase.phon.chass.ncsu.edu
92
This procedure works well for parallel patterns, but fails at detecting a common
feature change in chain shift alternations.
93
All the usual caveats about studies of phonological inventories apply to studies of
phonological alternations. A phoneme inventory and a phonological rule are both
very high-level descriptions, and many steps intervene between the numbers
reported in this paper, and the linguistic descriptions they are based on. The
criticisms of UPSID put forth by Simpson (1999: 349) apply equally to the segment
inventories in P-base, and the phonological alternations that this paper is mainly
concerned with are subject to similar issues. In general, these have been taken as
described from the grammars in which they appeared, except when contradicted by
available data. The aggregate data presented here is expected to highlight many
facts about phonological patterns that are interesting to typologists, and also to
reflect some conventions in how phonological patterns are typically described.
94
Other voiceless fricatives also turn to [h], but not as often as /s/ does, which is
consistent with /s/ being very frequent in inventories. Corresponding changes such
as /z/ → [ɦ] are not observed.
95
For the purpose of these definitions, place features = [anterior], [coronal], [back],
[high], or [labial];

but
not glottal, e.g., palatized = having the palatalization diacritic [j].
Palatalization has been excluded from assimilation in the reporting of these results.
There is overlap of 69 patterns between Progressive Assimilation and Regressive
Assimilation because some patterns operate in both directions but do not require
both at the same time (which would be Bidirectional Assimilation).
96
Since /i/ and /u/ occur in such a large number of inventories, there is very little
opportunity to observe structure changing epenthesis involving these vowels.
97
Here and elsewhere, “stops” is used to mean obstruent stops, not including [ʔ],
which typically has very different phonological behavior, and has often been treated
as a glide by phonologists (e.g., Chomsky & Halle 1968).
98
In these balloon plots, frequency is indicated by area, and the epenthesis and
deletion balloons are superimposed (e.g., vowel deletion is slightly more frequent
than vowel deletion in the C_C context). In Figure 6.9, the balloons are
superimposed, but the “all epenthesis” balloons are by definition as large or larger
than the subset of epenthesis that is structure changing.
99
I: Input, O: Output, C: Change, ER: Environment right, EL: Environment Left, AR:
Assimilated Right, AL: Assimilated Left
100
Notwithstanding the move towards articulatorily oriented features in phonology,
the acoustics of features continued to be investigated by Stevens, Blumstein and
colleagues (cf. Stevens & Blumstein 1978; Blumstein & Stevens 1980; Lahiri,
Gewirth, & Blumstein 1984), the goal being to locate invariant acoustic cues for
distinctive features rather than for segments, which had proved to be impossible (cf.
Lahiri et al. 1984 for cues to distinguish coronal and labial diffuse stops).
101
The tier structures in the feature trees (1)–(4) are not relevant for the present
discussion.
102
The feature tree given in Halle et al. (2000: 389) does not indicate +/− values.
However, from their discussion of Irish assimilation it is obvious that as before the
features high, low, distributed, round, anterior, back are binary.
103
It is possible that the features under the LARYNGEAL node should be independent
and not be subsumed under a single node.
104
We are assuming that these consonants should be [coronal] in C&D based on the
rest of their analysis.
105
In Hall’s terminology, rather confusingly, traditional “palatals” are called
“alveolopalatals”, and they differ in their coronality: “The term ‘palatals’ will used
here to refer to true palatals, such as German [ç ʝ] and not to sounds like
Hungarian [c ɟ], which are alveolopalatal” (1997: 70, §2.6). According to Hall,
alveolopalatals are coronal whereas true palatals are not; thus, “alveopalatals” [ c ɟ
ɲ ɕ ʑ] are [+coronal], “true palatals” [ç ʝ] are [−back, +dorsal]; also, he assumes
that a four-way contrast among a single series of [+coronal –cont] is maximal
(1997: 88, (4)). Since [±back] is not an option, in FUL all of these consonants are
[coronal]. Hall also states, and here we agree, that no language contrasts
alveolopalatals [ɕ] and palatoalveolars like [ʃ], and in fact the same holds true for
palatals and palatalised velars – which is why, in his model, they have the same
features. However, no language contrasts alveopalatal [ɕ] and palatal [c] either,
and moreover there cannot be stops in both positions: one of the consonants has to
be a continuant (cf. Lahiri & Blumstein 1984).
106 The change leads most often to a [high] consonant such as [ç tʃ ʃ]. Sometimes
/t/ also becomes /s/ in a similar context, but that is more of an assibilation
whereby the stop becomes a sibilant fricative, again in the context of a high vowel
or glide.
107
L&E assumed that the palatalisation of [t] lead to an affricate [tʃ]. This was an
incorrect assumption, as Carlos Gussenhoven points out, because it ought to be
more like [c], which is a stop. However, the second author of L&E, Vincent Evers,
finds that the diminutive of plaats ‘place’ ends up as [plaːtʃə] and is, thus, not very
different from the diminutive of plaat ‘plate’. What is important here is that for
FUL, both are [coronal], differing in affrication.
108
This comment has also been made by many phonologists including Sagey as well as
in SPE.
109
When the root has the “abstract” vowels /I U/, which in turn surface as [i~e] or
[u~ɔ], the suffix remains /a/. Our focus is not on the abstract vowels, which as
Hyman shows are entirely transparent and predictable, but on the first three
contexts.
110 Blevins provides these examples. If, however, /g/ > [dʒ] in Northwest Mekeo
in the context of /i/ it is not obvious to us where the example gina comes from.
111
Hybrid models which allow both abstract and episodic representations
(Pierrehumbert 2016) are hard to test. FUL does not deny that native listeners are
especially sensitive to familiar voices; surely one’s mother’s voice is easier to
identify in a noisy environment than the voice of a salesperson. Nor do we
disregard the fact that different dialects can cause hiccups in processing or that
hearing an unfamiliar dialect for many days at a time leads to familiarisation.
Nevertheless, we believe that individual lexical representations are abstract and do
not contain details of individual voices or dialects. Certainly representations can
change and become more flexible, but our claim is that basic contrasts and feature
representations along with concomitant processing implications are universal.
112
We do not exclude the possibility that there may be universal tendencies
concerning markedness; for example, we do not know of a language where [–nasal]
is marked. However, Rice (2003, 2007) shows that a number of presumed
universals of markedness are not empirically supported. Therefore, we adopt the
conservative position that all markedness relations are language specific. We are
prepared to modify this view where evidence exists in favour of a stronger position.
113
Markedness considerations thus dictate whether we name a feature [back] or
[front]: if a language has backing triggered by a back vowel but no fronting
triggered by a front vowel we call the harmony feature [back]; conversely, we
attribute fronting or palatalization to a feature [front]. In some cases the phonetic
ranges of vowels might influence the choice of label.
114
For example, in dialects descending from Proto-Eskimo that retain a four-vowel
system (either overtly or in underlying representations), the reflex of Proto-Eskimo
*/ə/ can assimilate to different vowels depending on context, but diachronically
this vowel has only merged with Proto- Eskimo */i/; see Compton & Dresher
(2011) and Section 4.2 below.
115
Zhang (1996) labels the features [labial] rather than [round], and [coronal] rather
than [front]. For our purposes these names are interchangeable and do not imply
any differences in the substance of these features.
116
Various proposals have been offered to account for why a single /ɔ/ does not cause
rounding harmony; a similar restriction occurs in Oroqen (Zhang & Dresher 1996;
Walker 2001, 2014). Based on the observation that a single irregular stem-internal
/ɔ/ does cause harmony in Baiyinna Oroqen (Li 1996; Walker 2014), Dresher &
Nevins (2017) propose that the restriction may actually be that a low suffix vowel
may obtain a [round] feature from a stem-internal /ɔ/, but not from an /ɔ/ that is
stem-initial.
117
This is not to say that there can be no other empirical evidence, for example from
synchronic alternations or diachronic mergers, that can choose between these
orderings.
118
Analyses that exploit the contrastive hierarchy in accounting for diachronic change
include: Zhang (1996) and Dresher & Zhang (2005) on Manchu; Barrie (2003) on
Cantonese; Rohany Rahbar (2008) on Persian; Dresher (2009: 215–225) on East
Slavic; Compton & Dresher (2011) on Inuit; Gardner (2012), Roeder & Gardner
(2013), and Purnell & Raimy (2013) on North American English vowel shifts;
Purnell & Raimy (2015) and Dresher (2017) on Old English; and large-scale studies
by Harvey (2012) on Ob-Ugric, Ko (2010, 2011, 2012) on Korean, Mongolic, and
Tungusic, and Oxford (2011, 2012a, 2015) on Algonquian.
119
See Oxford (2015) for the sources of these observations.
120
However, they are a counterexample to a proposed implicational universal to the
effect that “in a given language, low and mid front vowels apparently only trigger
palatalization if high front vowels trigger it too” (Kochetov 2011). Kochetov notes,
however, that “Bhat (1978) mentions some cases where mid front vowels palatalize
velars to the exclusion of high front vowels”. The typological rarity of such cases
may be due to the rarity of feature hierarchies like (42). More expected is the fact
that the palatalizations in question involve dorsal consonants, which, according to
Kochetov (2011), are “almost exclusively targeted by /i/ and other front vowels”,
unlike coronals which may be targeted by high vocoids.
121
See, for example, analyses of Lithuanian vowel length in Campos-Astorkiza (2009)
and Dresher (2009).
122
For instance, scholars do not agree on whether stød is a tonal configuration or a
separate phonological type of object.
123
This is not to say that there are no important distinctions of a finer kind. Meyer’s
(1937) average contours exhibit small timing differences between dialects,
employed by Öhman (1967) in his “Scandinavian accent orbit”, and Dalton & Ní
Chasaide (2007) have shown how systematic such differences can be between
varieties within what is considered the same dialect of Irish. The exact boundary, if
any, between phonology and phonetics in this regard is ultimately a matter of
model and interpretation.
124
Bruce (2007, 2010) adds an interesting and important discussion of North
Swedish, which we return to in Section 5.4.
125
For a general discussion of the analytical history of Scandinavian accent, see
Naydenov (2011).
126
Bruce (1977) called these “sentence accent” and “word accent”. Later they have
been referred to as “focus/focal accent” and “word accent”, “focal accent” and
“non-focal accent” (e.g., Bruce 2007), or “prominence level 2” and “prominence
level 1” (Myrberg 2010). For a fuller discussion of the reasons for adopting the
terms “big” and “small”, see Myrberg & Riad (2015).
127
The small accent is HL in both instances, but with different associations, yielding
the difference in timing (Bruce 1977). The leading H in accent 1 is sometimes in
evidence also in the big accent (hence HL*H), but the phonological status of this
tone is disputed. It is more stable in the small accent (Bruce 1977), and its presence
in the big accent is related to articulatory emphasis, and thereby the height of the
trailing H (Fant & Kruckenberg 2008). Engstrand (1995, 1997) has argued that the
leading H in the big accent is predictable from ambient intonation.
128
The tonal identity between lexical accent 2 and postlexical accent 2 is no
coincidence, but the result of a diachronic change, from postlexical to lexical (Riad
1998a).
129
Compound accent occurs in any form that contains two stresses, predominantly
but not exclusively compounds, cf. (24) below.
130
Several dialects admit either accent in compounds and thereby do not have a
“compound rule”. This is discussed in Section 5.3. In the comparisons we make we
shall use instances of accent 2 in compounds.
131
The orthographic form is divided into syllables according to retroflexion in the
phonetic form which is [ˈtʰɑːlɛspæˌʂuːnɛɳa]. Unknown speakers are coded as ‘NN’,
known ones with initials.
132
It has, however, not been reported that there would be a stable association point in
the following word, as has for example been reported for Standard Greek (Arvaniti
2002; Grice et al. 2000).
133
The secondary association here is motivated by the availability of TBU’s. Grice,
Ladd, & Arvaniti (2000) discuss cases of secondary association of phrase accents
(corresponding to the separate focus gesture in Bruce’s terminology) in Hungarian,
Greek, and Cypriot Greek. Intonation tones may then associate to a stressed
syllable in the following word (Greek) or another syllable down the line, depending
on prosodic context.
134
Unlike the case in Central Franconian and Lithuanian, where there is a basic
requirement of two sonorant moras for the tonal distinction to be realized, there is
no sonority requirement in Swedish and Norwegian that affects the tonal contour.
The situation for Danish stød resembles that of Central Franconian and Lithuanian
in this regard (Basbøll 2005: 272).
135
To make this point clear, the alternative to prosodically motivated accent
assignment is lexically motivated accent assignment, which looks at the properties
of the first compound member in free form.
136
Some Central Franconian dialects, too, exhibit the possibility of more than one
accent within compounds or compound-like forms (Peters 2006: 120). This points
at a difference regarding culminativity, between, on the one hand, NGmc tonal
dialects, where there is invariably one accent per (maximal prosodic) word, always
with association to the primary stressed syllable, and on the other hand, Danish
and Central Franconian dialects of the Hasselt type, where there can be more than
one accents in a word, hence in both primary and secondary stressed syllables.
137
The actual prominence FUNCTION is sometimes carried by the boundary tone,
when H%, cf. Section 7.
138
In the Norwegian tradition, where one distinguishes “high-tone” dialects from
“low-tone” dialects, the terminology refers to the value of the prominence tone
(i.e., the first tone of accent 1). This is a rather unfelicitous choice of category, since
it does not necessarily single out a phonological unit. In CSw the prominence tone
is L*H, i.e., a bitonal unit. Calling this a “low-tone” dialect would refer to just half
of the phonological category.
139
The boundary H% can also be followed by another boundary tone L%, which we
disregard here.
140
For instance, speakers have developed a sensitivity to a secondary association.
Hognestad proposes that this ongoing change might lead to generalized connective
accent 2 in compounds. At this stage, there is not (yet) a “compound accent”, and
all speakers still exhibit both accent 1 and accent 2 in compounds. The shape of
accents with sensitivity to the final secondary stress contains an early second peak
followed by a characteristic high plateau (rather than the low floor that occurs in
Central Swedish, which has the same tonal make-up). An example of such a high
plateau occurs in the accent 1 compound in (22).
141
In informal conversation this informant exhibits a pattern more like the other
informants his age.
142
The analyses of the three Stavanger forms do not necessarily agree with those
given by Hognestad (2012).
143
There have been proposals that there are dialects that have very nearly the same
tonal makeup in accent 1 and accent 2 (Kristoffersen 2007 for Oppdal, Segerup
2004 for Göteborg). These things would be expected to have a phonetic
explanation in terms of realization, under the analysis pursued here, and not lend
themselves to immediate phonological translation as identical melodies.
Otherwise, there should be nothing in the way of a generalized, connective accent 1
in compounds, which is unattested.
144
Bruce (2003: 246) states that the focus gesture, i.e., the second rise, is not
coordinated with the secondary stress in compounds in the Göta varieties.
However, Riad & Segerup (2008) found that the L*H was systematically associated
in the last (secondary) stress in compounds. It is the trailing H of L*H which is not
obviously timed with that syllable.
145
Thanks to Bengt Nordberg who pointed out (p.c.) that there appears to be a dialect
boundary cutting through the city of Eskilstuna.
146
It is not entirely clear that the association pattern for the prominence is LH* rather
than L*H. Closer investigation is required.
147
One could have imagined that the (post)lexical H* would have been employed for
prominence purposes in some dialect, in parallel with what we see for the
boundary H% in East Norwegian. However, that would have predicted the
possibility of a contrast between accent 2 as H*L and accent 1 as L*H, but this has
not been reported for any dialect. This adds weight to the generality of the
privative nature of the accent distinction.
148 As mentioned, it is not clear whether there is a strict separation between the
two consecutive L’s in South Swedish.
149
If the argumentation starts from the number of syllables as basic to accent 2, the
morphological patterning is left unexplained.
150
There is clearly a connection between the mora-based tonal assignment, the
sonority restriction and the segmental dependence, though the exact typological
implications, also with respect to tonal systems of the South East Asian types,
remain to be worked out.
151
An exception is Clements & Keyser (1981), who worked with multiply branching
syllables.
152
Many speakers neutralize the word accent contrast on word-final monomoraic
accented syllables in monosyllabic (cf. (5a) and (5b)) and disyllabic words, as in
/hana/ ‘nose’ vs. /haná/ ‘flower’ in utterance-final declarative and interrogative
contexts (cf. Vance 1995). In words of more than two syllables, the contrast is
reliably preserved in declarative intonation (L%), due to the presence of a
preceding syllable with Hα. In the unaccented case, the pitch falls from peninitial
Hα to mid pitch at the word end, but in the accented case, it rises towards H* and
then falls to mid in the last part of the final syllable (Warner 1997). In questions,
the situation in trisyllabic words is similar to that in the disyllabic cases, so that the
prediction is that trisyllabic minimal pairs are not distinct when spoken with
question intonation in the speech of speakers who neutralize disyllables and
monosyllables. I am not aware of experimental work on this question.
153
The prenuclear pitch accent might be H*L, with L being aligned rightmost, as in
my descriptions of English and Dutch, with somewhat delayed pronunciation of
H*. This is in tune with the fact that the L-target is undershot under tone
crowding. The motivation for LH as opposed to HL would appear to be the closer
proximity of the two tones, but since there is no constant timing relation between
them, that motivation is slim.
154
English has a vacuous restriction of one per foot, as in Cálifórnia (Pierrehumbert
1980), by virtue of its TBU, the stressed syllable, but Japanese for instance
routinely has more TBUs in the domain in which only one accent may occur.
155
The analytical approach to the nature of accent has no implications for its status in
the grammar. In Gussenhoven (2004: 42), I drew a distinction between analytical
concepts like “word melody” and “accent”, which are based on distributional
considerations, from phonological concepts like segments and prosodic
constituents, which typically have measurable properties, with no implication that
one type of concept is somehow superior to the other. On this, van der Hulst (2012:
1519) commented that ‘[CG] sees accent as an analytic device, suggesting that it is
the “invention” of the linguist, adopted in order to organize data, and that
therefore it is not as “real” as an H tone or “stress”, which we can “hear”.’ To
clarify, if a language were to have foot-based generalizations, while at the same
time failing to reveal foot heads in the phonetic record, the foot would be an
analytical notion, i.e., one based on distributional facts only, but it would still be a
prosodic constituent of that language.