4 Laryngeal features
Les changements phon´etiques sont les manifestations et les r´ealisations
de tendances, que la langue a contract´ee au cours de sa vie ant´erieure.
Ces changements sont d´esign´es par le nom de lois phon´etiques.
Grammont (1933: 166)
4.1 Phonological features and laryngeal features
If there is one point of agreement within phonetic and phonological
theory, it is that the segments which compose speech are not indivisible
primitive units of speech. Instead, the general view is that segments are
the simultaneous realization of smaller units, known as features. Features
play a signiﬁcant role in deﬁning sound change and sound patterns. At the
phonetic level, there are potentially gradient and potentially imperceptible
phonetic features; at the phonological level, there are distinctive features
which are typically privative (single-valued) or binary-valued, and which
deﬁne contrasts which are typically perceptible to all human newborns
(Werker and Pegg 1992). These distinctive features are the basis of all
attested phonological contrasts. Two primary arguments exist for distinctive
features. One argument is that they correspond quite closely to
innate perceptual categories demonstrated in newborns and young children
(9.1). Another argument is that they allow the statement of what
appear to be signiﬁcant generalizations across sound patterns.
In this chapter, characteristic patterns of laryngeal feature distribution
are investigated.1
Laryngeal features are those which characterize the state
1 Languages on which generalizations are based include: Lithuanian, German, Frisian,
Sanskrit, Russian, French, Catalan, English, Attic Greek, Polish, Maithili, Lamani (IndoEuropean);
Kiowa (Kiowa-Tanoan); Yokuts, Kashaya, Miwok, Klamath, Tsimshian,
Nez Perce; Afar (Cushitic); Dinka (Nilotic); Ngizim (Chadic); Korean; Hungarian
(Finno-Ugric); Kolami (Dravidian); Yate´e Zapotec (Otomanguean); Tamazight
Berber (Afro-Asiatic/Berber); Arabic, Syrian, Eastern, Moroccan, Iraqi Arabic (AfroAsiatic/Semitic);
Totontepec Mixe (Mixe-Zoque); Nhanda (Pama-Nyungan); Lac
Simon (Algonquian); Wantoat, Kalam (Papuan); Sre, Pacoh (Mon-Khmer); Chepang,
89
90 Sound patterns
of the larynx or vocal folds, and the acoustic and perceptual features associated
with these states. Though it is necessary to distinguish dozens of
phonetic categories which contrast gradient voice onset times, degrees of
laryngealization, and pulmonic versus laryngeal airstream mechanisms
(Gordon and Ladefoged 2001), for the purposes of describing contrasts
in phonological segment inventories, only three phonological features
appear necessary. These features are often labeled as [voiced], [spread
glottis], and [constricted glottis].2
While labeling and precise deﬁnitions
of these features may be disputed, there are two aspects of the feature
system which are non-negotiable. First, phonological features are distinct
from gradient phonetic properties, and typically reﬂect categories
which have multiple distinct phonetic instantiations or cues. Second, no
reference is made in phonological systems to non-contrastive features
(e.g. the release of stops, 2 ms of voice-onset time, laminal interdental
versus laminal dental, etc.) Each of these properties is integral to the
explanations provided for typical cases of regular sound change. Because
phonological features are distinct from gradient phonetic properties,
and there is typically no one-to-one relationship between a phonological
feature or category and a single phonetic characteristic, reinterpretations
of the phonetic signal in the form of slight shifts of whole categories within
the psycho-acoustic space are expected.
Within a single language, identical phonological feature speciﬁcations
can be associated with highly distinct phonetic features depending on
phonological context, or on other segment-internal feature speciﬁcations.
Consider the phonetic contrast between pre- and post-aspirated
segments. Though these two sounds clearly constitute different acoustic
and auditory categories, in Icelandic (Thr´ainsson 1978), regular alternations
condition pre- and post-aspirates, which are in complementary
distribution. In other languages, like Twampa (Thelwall 1983; 334)
post-aspirates have optional pre-aspirated variants in certain contexts.
The same sort of phonetic variation characterizes segments deﬁned as
[constricted glottis]. In Klamath (Barker 1964; Blevins 1993a), sonorants
with this feature are pre-glottalized, while obstruents are realized
as ejectives. For more examples of this kind, see Kingston (1990). The
Mikir, Sherpa (Tibeto-Burman); Lushootseed, Twana, Bella Coola, Montana Salish,
Shuswap (Salishan); Palauan, Marshallese (Austronesian); Nootka, Kwakiutl
(Wakashan); Haida (isolate); Mbe (Niger-Congo/Bantu), as well as others mentioned in
the text. This set includes 52 languages from 22 different language families.
2 Here I do not include laryngeal “tone” features. For an overview of the relationship
between laryngeal features and tone, see Yip (1995: 484–88). On the evolution of
tonal systems from F0 perturbations associated with consonantal laryngeal features, see
Hombert et al. (1979). And for phonological systems where tone features interact with
other laryngeal features, see Hyman (1973).
Laryngeal features 91
phonological identity of pre- and post-aspiration, and of pre-glottalization
and post-glottalization (or ejection), is strongly supported by the fact
that no languages are reported to contrast segments with respect to these
dimensions alone. Our starting point for the investigation which follows,
then, is that a limited number of phonological categories or segment types
are deﬁned by a limited set of phonological features.
4.2 The phonetic basis of sound patterns
Since the introduction of distinctive features into phonological theory,
the majority of phonotactic constraints have been stated in terms of natural
classes deﬁned by distinctive features. One fundamental observation
which emerges from the cross-linguistic study of the distribution of laryngeal
features is that segments with identical phonological feature representations
may have dramatically different patterns of distribution. For
example, there is a strong cross-linguistic preference for post-aspiration of
consonants to occur pre-vocalically, while pre-aspiration of consonants
is generally found post-vocalically. Accounting for this general type of
distributional pattern is problematic for most phonological approaches,
since pre-aspirates and post-aspirates are non-contrastive, and therefore
are assumed to have identical phonological representations. One might
think of this as an isolated problem, relating only to the characterization
of aspiration, but the same problem occurs again and again: recurrent
phonotactic patterns seem directly related to non-distinctive properties of
sound patterns, and yet, phonological characterization of these patterns
must be stated without reference to non-distinctive properties, missing
important generalizations.
How can this problem be overcome? Two logical alternatives present
themselves. One possibility is to abandon a pure phonological feature
system by introducing phonetic detail into phonological representations.
This is the position taken, for example, by Steriade (1993, 1999a),
Kirchner (2000), and Flemming (2001). In these accounts, sound patterns
are deﬁned with direct reference to phonetic features as well as
phonological ones. This alternative eliminates any principled distinction
between phonological and phonetic representations, or phonological
and phonetic features, and attempts to build phonetic explanations into
phonological representations and constraints themselves. A second alternative,
and the one defended throughout this book, is to maintain a pure
categorical phonology, free of phonetic detail. This view of phonology,
similar to that conceived of by the Prague school, represents all and only
the features, segments, and prosodic categories which are contrastive in
the world’s languages. Within Evolutionary Phonology, the phonotactic
92 Sound patterns
regularities exhibited by synchronic phonologies appear to be sensitive to
phonetic detail because they are arguably, in many cases, the transparent
result of phonetically motivated sound change. Since phonetically motivated
sound change is well evidenced in the historical record, and can be
simulated in the laboratory, relying on such factors to explain recurrent
relationships between feature distribution and non-distinctive properties
of speech appears to be the null hypothesis.
A primary argument for Evolutionary Phonology over alternative
approaches is that phonetically motivated sound change is already
accepted as one explanation for properties of synchronic sound systems.
Any model which incorporates phonetic detail into phonological systems
in order to explain synchronic regularities is essentially duplicating an
explanation which already exists in the diachronic dimension. Under this
account, there is no pre-determined set of phonological constraints in the
mind of the speaker which gives rise to recurrent sound patterns. Rather,
the phonotactic regularities examined in this and following chapters are
emergent universals in the sense of Deacon (1997: 115–16):
Grammatical universals exist, but . . . their existence does not imply that they are
preﬁgured in the brain like frozen evolutionary accidents. In fact, . . . the universal
rules or implicit axioms of grammar aren’t really stored or located anywhere,
and in an important sense, they are not determined at all. Instead . . . they have
emerged spontaneously and independently in each evolving language, in response
to universal biases in the selection processes affecting language transmission. They
are convergent features of language evolution in the same way that the dorsal ﬁns
of sharks, ichthyosaurs, and dolphins are independent convergent adaptations of
aquatic species. Like their biological counterparts, these structural commonalities
present in all languages have each arisen in response to the constraints imposed
by a common adaptive context. Some of the sources of universal selection on
the evolution of language structures include immature learning biases, human
mnemonic and perceptual biases, the constraints of human vocal articulation
and hearing . . . to name a few. Because of these incessant inﬂuences, languages
independently come to resemble one another, not in detail, but in terms of certain
general structural properties . . .
In the following section I summarize ways in which languages independently
come to resemble one another in terms of the distribution of laryngeal
features.
4.3 Recurrent patterns of laryngeal feature distribution
The distribution of distinctive laryngeal features on consonants is severely
limited in many languages. Obstruent voicing, aspiration, and ejection are
Laryngeal features 93
often limited to certain positions, while distinct distributional constraints
are found for sonorant voicelessness and sonorant laryngealization. As
documented by Lombardi (1991), Steriade (1999a), Fallon (2002), and
Blevins (2003a) among others, there is general similarity across languages
with respect to: (i) the neutralizing feature or features; (ii) the positions
in which neutralization occurs; (iii) the direction of neutralization; and
(iv) distinct neutralization patterns for laryngeal features of obstruents
and sonorants.
In (1)–(3) I summarize some of these recurrent patterns with respect
to their phonetic status as release features, realized at consonant release,
or closure features, realized at the onset of consonant closure.
(1) Release feature pattern: obstruent voicing, obstruent
post-aspiration, obstruent ejection
(i) PHONETIC FEATURES: post-aspiration, postglottalization,
ejective release, release cues for voicing
(burst strength, VOT, F0 values, F1 values)
(ii) PHONOLOGICAL FEATURES: [voiced], [spread glottis],
[constricted glottis]
(iii) COMMON POSITIONS OF CONTRAST: before sonorants
(iv) COMMON POSITIONS OF NEUTRALIZATION: before
obstruents, word-ﬁnally
(v) COMMON DIRECTION OF NEUTRALIZATION:
neutralization is either
a. to the voiceless unaspirated member of the series, or
b. in obstruent clusters, the result of regressive
assimilation
(2) Closure feature pattern: sonorant pre-glottalization
(i) PHONETIC FEATURE: pre-glottalization
(ii) PHONOLOGICAL FEATURES: [−syllabic, +sonorant,
constricted glottis]
(iii) COMMON POSITIONS OF CONTRAST: after vowels
(iv) COMMON POSITIONS OF NEUTRALIZATION: after
consonants, word-initially
(v) COMMON DIRECTION OF NEUTRALIZATION: to the
plain voiced member of the series
(3) Closure feature pattern: obstruent pre-aspiration
(i) PHONETIC FEATURE: pre-aspiration
(ii) PHONOLOGICAL FEATURES: [−sonorant, −continuant,
spread glottis]
(iii) COMMON POSITIONS OF CONTRAST: after vowels,
sometimes after sonorant consonants
94 Sound patterns
(iv) COMMON POSITIONS OF NEUTRALIZATION: after
obstruents
(v) COMMON DIRECTION OF NEUTRALIZATION:
neutralization is either
a. to the voiceless unaspirated member of the series, or
b. in obstruent clusters, the result of progressive
assimilation
No phonological system is known to contrast post-aspirated stops with
their pre-aspirated counterparts. Nevertheless, this phonetic difference
appears to determine the general sound patterns of aspirated stops both
within and across languages. Compare the common patterns involving
post-aspiration, a release feature, in (1), with those of pre-aspiration, a
closure feature, in (3). The patterns are near mirror images of each other.
Post-aspirates contrast most widely in pre-vocalic position, while preaspirates
are most likely to contrast in post-vocalic position. The crosslinguistic
regularities in this distribution are problematic for phonological
theory since all aspirated oral stops will have the same basic feature rep-
resentation.
The patterns in (1) and (2) are also near mirror images. Compare
ejective obstruents and their distribution with pre-glottalized sonorants.
Though both segment types are characterized by the same distinctive
feature, [constricted glottis], neutralization of ejection is common before
obstruents and word-ﬁnally, while neutralization of glottalized sonorants
is typical word-initially and after consonants. As far as I am aware, there
are no reported cases of ejectives neutralizing to plain stops word-initially
or post-consonantally, highlighting the distinctive patterns of contrast for
closure and release features documented by Steriade (1999a).3
Perhaps the most well-known and well-studied examples of laryngeal
neutralization involve obstruent devoicing. Given a language where
voiced and voiceless obstruents contrast in some environments, there are
typically other environments in which voicing is non-contrastive. In unrelated
languages across the world we ﬁnd that the voicing contrast is neutralized
for obstruents, that this neutralization tends to occur word-ﬁnally
and/or in pre-obstruent position, and that neutralization in word-ﬁnal
position is to the voiceless phone, while neutralization in pre-obstruent
position is typically either to the voiceless phone, or subject to regressive
voice assimilation. This pattern is included in (1).
The emergent sound patterns summarized in (1)–(3) do not appear to
be the result of chance occurrences. Word-ﬁnal neutralization of laryngeal
release features is common, while word-initial neutralization of the same
3 Context-free shifts from ejective to plain stops are reported in Fallon (2002, chapter 3).
Laryngeal features 95
release features is unattested. At the same time, word-initial neutralization
of sonorant pre-glottalization is common, while word-ﬁnal neutralization
of pre-glottalization is rare. The signiﬁcance of these distinct patterns of
contrast and neutralization cannot be overstated. Phonological features
show distinct patterns of distribution which appear to be dependent, at
least in part, on their phonetic realization. In 4.4, historical phonetic
explanations are proposed for these common and recurrent sound patterns.
However, before turning to these, it will be useful to brieﬂy review
how and why such patterns are best characterized in terms of parallel
evolution.
General accounts of similarity were reviewed in 2.4. Recurrent sound
patterns may be the result of shared inheritance, convergent evolution,
parallel evolution, or diffusion through language contact. For the majority
of languages included in the surveys cited above, diffusion can be ruled
out as the source of similarity. Many of these languages are spoken on
different continents or islands, with no evidence of contact between them.
For example, there is no evidence of linguistic contact between Korean
speakers, Dinka speakers of Saharan Africa, Wantoat speakers of New
Guinea or Klamath speakers of south-central Oregon, and yet all of these
languages have sound patterns of the type summarized in (1). Diffusion
of laryngeal sound patterns from neighboring languages can often be
ruled out as well. For example, although Yurok is spoken close to other
languages with glottalized sonorants, the same languages show no evidence
for the Yurok pattern of neutralization conforming to (2). As a
consequence, the sound patterns on which the generalizations in (1)–(3)
are based are ones which appear to have evolved spontaneously many
times in the natural course of language change.
Ruling out diffusion as an explanation for these common sound patterns
still leaves open the possibility that they are nothing more than
chance resemblances. As Blust (1990: 24) notes:
Ideally, in evaluating the role of chance as an explanation we should have access
to a statistical model which states how many times a given sound change might be
expected to occur in relation to the null hypothesis. Unfortunately, no such model
is available, and we are consequently forced to fall back upon an impressionistic
statement.
This is still the case over a decade later, and we are forced to rely on
linguistic intuitions as to what constitutes “more than chance frequency”
for a given pattern. However, if chance were entirely responsible for the
patterns of laryngeal feature distribution just summarized, then we would
expect similar frequencies for other “chance” events. If there is no principled
difference between word-ﬁnal devoicing and word-ﬁnal voicing, both
being the possible result of chance events, then how are we to explain the
96 Sound patterns
high frequency of the ﬁrst pattern in contrast to the low frequency of
the second? Likewise, if the association between release versus closure
features and positions of contrast is the result of chance, what could
possibly explain the common neutralization of post-aspiration in wordﬁnal
position, in contrast to the rarity of rules neutralizing pre-aspiration
in precisely the same environment? If common sound patterns are the
result of phonetically motivated sound change, then a starting point for a
statistical model is the observation that sound patterns which can result
directly from phonetically motivated sound change will be more common
than ones which cannot, and those which have multiple sources in sound
change will be more common than those with single sources. This starting
point will allow us to account for the very high frequency of word-ﬁnal
devoicing in contrast to the high frequency of word-ﬁnal de-aspiration,
and the apparent absence of regular word-ﬁnal voicing.
With diffusion and chance incapable of explaining the recurrent nature
and high frequency of the sound patterns in (1)–(3), we are left to conclude
that these patterns of laryngeal feature distribution are instances of
shared inheritance or parallel evolution. A great many of the genetic relationships
among the world’s languages have been established by use of
the comparative method. Widely agreed-upon genealogies exist for many
language families, including Algonquian, Austronesian, Indo-European,
Mayan, Niger-Congo, Semitic, and Sino-Tibetan, to name just a few.
Given well-understood genetic relationships among languages, it is often
possible to determine whether or not a particular sound pattern is the
result of direct inheritance, or not. While many similar patterns of laryngeal
feature distribution are the simple result of shared inheritance, a
great many are not. Historical records of Sanskrit allow us to reconstruct
ancient patterns of laryngeal feature distribution, and to see the
extent to which these have been directly inherited by modern Indic languages
like Gujarati and Punjabi. But for each case of direct inheritance,
there are near-identical patterns of laryngeal feature distribution in an
unrelated language or language family. Klamath, Korean, Dinka, and
Wantoat were mentioned earlier. Genetic afﬁliations of Klamath and
Korean are debated, but each language certainly qualiﬁes as independent
stock from any of the other languages with similar sound patterns. Dinka
is a Nilotic (Nilo-Saharan) language of the Sudan, and Wantoat, a language
of Morobe Province, Papua New Guinea, has been assigned to the
Trans-New Guinea phylum. Similarities between Proto-Indo-European,
Gujarati, and Punjabi might be the result of direct inheritance, but those
between Indic, Klamath, Korean, Dinka, and Wantoat are not.
The strongest evidence for parallel evolution, however, are cases where
it can be demonstrated that a particular sound pattern was not a feature
of the mother tongue, and that it is the result of an innovation. The
Laryngeal features 97
reconstruction of languages and language families has uncovered thousands
of sound changes, which, when studied closely, provide the empirical
base for parallel evolution as the source of common sound patterns.
To take just one example, consider the common distribution of contrastive
voicing summarized in (1). Lithuanian, an Indo-European language,
allows voicing contrasts in obstruents before vowels and sonorants,
but neutralizes these contrasts elsewhere. The same sound pattern has
evolved independently within the Slavic, Germanic, and Italic subgroups
of Indo-European. Within Italic, this pattern has evolved in Catalan, but
not in French. Within West Germanic, the pattern has evolved in German,
but not in English.4
In sum, ﬁndings in historical linguistics present
a wealth of evidence for parallel evolution in the world of sounds. Similar
instances of phonetically based sound change occur in language after language.
These sound changes are both the locus of phonetic explanation
and the source of synchronic regularities in Evolutionary Phonology. It
is to these sound changes that we now turn.
4.4 Explanations for patterns of laryngeal feature
distribution
The general hypothesis of Evolutionary Phonology, and the one argued
for throughout this book, is that parallel evolution is associated with phonetically
based sound changes which recur with more than chance frequency
due to inherent features of the human perceptual and articulatory
system. The formal model of sound change proposed in 2.2 predicts that
any sound change with sources in CHANGE, CHANCE, or CHOICE will
give rise to sound patterns that are more frequent than those which do
not have their sources in natural sound change. Common sound patterns
will typically reﬂect common instances of sound change. In this
section, the common patterns of laryngeal feature distribution in (1)–(3)
are associated with common instances of sound change. The overarching
generalization is that positions of contrast for a particular feature are
those in which neutralizing sound change is unattested, while positions
of neutralization are precisely those where phonetically motivated sound
change is common.
4.4.1 Release features
Let us focus on the most signiﬁcant generalizations which have emerged
from the cross-linguistic study of laryngeal feature distribution, many of
4 In fact, English dialects with word-ﬁnal devoicing are reported. One case in Appalachian
American English is discussed in 10.2.
98 Sound patterns
them originally due to Steriade (1999a). One important generalization
is that phonetic features timed with release have distributional patterns
which are distinct from those timed with the onset of constriction or closure.
A related generalization is that release features tend to be neutralized
before obstruents and word-ﬁnally, while closure features tend to be neutralized
after consonants and word-initially. In addition, there appears to
be a signiﬁcant tendency to devoice obstruents in word-ﬁnal position,
even in languages where voicing is primarily cued by segment-internal
voicing or duration, as opposed to differences in voice onset time.
Why are release features like post-aspiration and ejection neutralized
before obstruents and word-ﬁnally? Recall the discussion of variability
in articulation in 2.2. There it was observed that stop–release is
not distinctive in any attested spoken language. In some languages, like
English, the release of word-ﬁnal stops is optional. In other languages, like
Marshallese, there is a strong tendency for ﬁnal stops to be unreleased.
In still other languages, like German, there is a strong tendency for ﬁnal
stops to be released. While a given language, dialect, or speech style may
show a particular tendency for release or non-release of word-ﬁnal stops,
this phonetic feature is never contrastive. In fact, when careful speech is
elicited, one often ﬁnds that non-released stops are released. In general,
hyperarticulated speech may transform unreleased stops into released
stops. In 2.5 I suggest that release as a distinctive phonological feature
would be particularly nonaptive given the frequency of pre-vocalic contexts.
Whatever the explanation, however, the fact remains that release
is a variable feature of pronunciation for some sounds in nearly all welldescribed
languages.
Now, consider how variability in release might give rise to the observed
patterns of neutralization. There are many languages like Klamath, where
distinct laryngeal series of stops are distinguished primarily in terms of
release features. In Klamath, with plain voiceless, voiceless aspirated,
and voiceless ejectives, voice onset time and burst quality appear to be
the primary phonetic cues for the laryngeal contrast: plain stops differ
from aspirated stops in VOT values, while ejectives differ from the other
two series in possessing a glottal release which follows the oral release.
Ejectives also carry longer VOTs than plain stops, and ejectives will typically
differ from the aspirated stops in showing no formant structure
following the moment of oral release. What happens to the three-way
laryngeal contrast for oral stops when a word-ﬁnal or pre-obstruent stop
is unreleased? Without audible release, the primary phonetic cues for the
three-way contrast are no longer perceptible, resulting in high probabilities
of neutralizing sound change.
In this case, the phonetic source of sound change is two-fold involving
CHOICE and CHANGE. First, there is the fact that, for all languages, the
Laryngeal features 99
continuum of careful to casual speech allows for variation in the (audible)
release versus non-release of consonants involving oral closure. If the
frequency of unreleased variants is higher than that of released variants,
these forms can become new norms of pronunciation – a simple and
common case of CHOICE. For many segment types, like [m], release
versus non-release may not result in any general increased instances of
misperception. However, where laryngeal contrasts like those described
for Klamath are involved, the absence of release will result in reduction
of primary phonetic cues for laryngeal contrasts, with a greater likelihood
of the unreleased stops being perceived by language learners as
stops lacking aspiration or ejection. Note that the speaker may still be
carrying out the same laryngeal timing sequence for the production of
unreleased aspirates and ejectives; however, the absence of release will
make these articulatory movements less perceptually salient, or even
inaudible.
By expressing the relationship between non-release and neutralization
in terms of sound change, a restricted phonetic typology is deﬁned.5
First, there will be languages which maintain a surface contrast between,
e.g., plain obstruents and obstruents contrasting in laryngeal release features
in ﬁnal and pre-obstruent position. In these languages, by deﬁnition,
stops will be audibly released. Languages of this type include Lushootseed
(Urbanczyk 1995, 1996) and Bella Coola (Nater 1984), where ejectives
and plain obstruents are contrastive in all positions, and there is audible
release. Second, there will be languages in which neutralization of laryngeal
release features occurs precisely in positions where obstruents are
not released. A language of this type is the dialect of Bengali described
by Kenstowicz (1994: 193), where aspiration and breathy voiced release
are neutralized word-ﬁnally, and in pre-obstruent position, but obstruent
voicing is maintained. A third type of language will reﬂect a superﬁcial
development from the Bengali stage to a subsequent one: neutralization
patterns will reﬂect earlier non-release, but obstruents will be phonetically
released. Klamath, as described by Barker (1963, 1964) reﬂects
this language type. In this last case, the intrinsic variability of consonant
release plays a role in rendering a once transparent relationship between
phonological patterns and phonetic source translucent.6
5 This typology may have practical applications. For example, in efforts to reconstruct the
phonetics of Coahuilteco for the purposes of language revitalization, this typology has
been useful. Ejectives clearly contrast with plain stops, but should they be produced as
ejectives in pre-consonantal and ﬁnal position or not? Troike (1996: 651) has discovered at
least one example where an expected sequence of ejectives p’t’ is written as pt’, suggesting
that Coahuilteco obstruents are not released in pre-consonantal position.
6 This sort of translucency is problematic for Steriade’s (1999a) account in which parallel
phonetic and phonological phenomena are accounted for by the same constraints.
100 Sound patterns
But what of the many other cases where neutralization in VC1C2V is
to the laryngeal features of C2? What phonetic fact explains the common
pattern of regressive voice assimilation in obstruent clusters? As demonstrated
by Raphael (1981) and Slis (1986) for voicing, there is a general
perceptual preference for the cues present in stop bursts and C–V transitions
to take precedence over those in the V–C transition, all else being
equal. Viewed in this light, regressive voice assimilation involves a classic
case of CHANGE: a sequence like [apda] is misperceived as [abda],
resulting in regressive voice asssimilation. The same explanation can be
given for assimilation in longer obstruent clusters, like the pattern found
in Russian, where all obstruents take on the voicing feature of the last
obstruent in the cluster.7
Finally, notice that this released-based account, originally due to
Steriade (1999a), makes two predictions concerning voicing contrasts in
stops versus fricatives, assuming that fricative voicing is typically cued by
segment-internal noise (or duration) as opposed to release features. First,
there are predicted to be languages in which stops undergo laryngeal neutralization
due to non-release, but fricatives do not. This pattern is found
in Turkish, where the voicing contrast for stops and affricates is neutralized
word-ﬁnally and in pre-consonantal position, but the voicing contrast
for fricatives is maintained in the same contexts. Second, across-theboard
laryngeal neutralization of stops and fricatives under non-release
is unexpected, unless laryngeal features of fricatives are speciﬁcally associated
with release, an apparent cross-linguistic rarity.8
This second prediction
appears to be in conﬂict with the many languages which show
consistent word-ﬁnal devoicing of both stops and fricatives. In 4.4.4
I suggest an alternative pathway to ﬁnal devoicing which predicts ﬁnal
devoicing of both stops and fricatives.
4.4.2 Closure features
If the absence of release can explain the most general features of laryngeal
release feature distribution in (1), is there a mirror-image explanation for
7 Work with chinchillas and quails shows evidence of categorical perception for voicing
contrasts in oral stops (Kuhl and Miller 1978; Kluender, Diehl, and Killeen 1987). If
their categories are based on similar perceptual features, chinchillas and quails should
show the same tendencies in misperception, mistaking [apda] for [abda], and mistaking
[abta] for [apta]. As far as I know, this prediction has not been tested.
8 The segment types involved would be post-aspirated, ejective, or post-glottalized fricatives.
Burmese has post-aspirated fricatives; ejective fricatives are found in Hausa and
Tlingit; and post-glottalized fricatives are reported for Amharic and Yapese (Ladefoged
and Maddieson 1996: 178–79; Demolin 2000; Maddieson 1998). In Tlingit, where consonants
are typically audibly released in all positions, ejective fricatives occur in preconsonantal
and ﬁnal position (Story and Naish 1973; Maddieson et al. 2001).
Laryngeal features 101
the distributional patterns of the closure features of pre-glottalization and
pre-aspiration summarized in (2) and (3)? Before answering this question,
it is important to recognize that these two cases have signiﬁcant differences
which cannot be ignored. First, though there are many languages
where a contrast between plain and pre-aspirated obstruents is absent
in word-initial position, there are few languages in which phonological
alternations provide evidence of neutralization in precisely this position.
Since, within Evolutionary Phonology, regular phonological alternations
are often the fossilized reﬂexes of regular sound change, this gap is puzzling,
and suggests that the absence of pre-aspirates in word-initial position
may be partly independent of neutralizing sound change. On the
other hand, similar alternations in unrelated Yurok and Yokuts suggest
that the restriction of pre-glottalized sonorants to post-vocalic position
may, at least in some languages, be the direct result of neutralizing sound
change. Given these differences, this subsection will limit itself to potential
sound changes resulting in common distributional patterns of preglottalization
in (2). In 4.4.3 an additional factor is suggested in the
cross-linguistic distribution of pre-aspirates, one which is independent
of laryngeal neutralization.
If the primary cue for sonorant laryngealization in some languages is
creak on a preceding vowel, or a glottal stop preceding sonorant modal
voicing, then it is not surprising that in contexts where a preceding vowel
is absent, neutralization occurs. Unlike the account of release features,
there is no reference to variation in this account. In a language like Yurok
(Robins 1958: 9), with a contrast between /w/ and /w’/, where /w’/ is
phonetically realized as a glide with preceding creak or glottal closure,
utterance initial [ʔw] can be easily misheard by the listener as [w]. In this
context, primary phonetic cues of creakiness on a preceding vowel or the
post-vocalic silent interval of the glottal stop are absent. Over time, we
expect that word-initial glottalized sonorants in languages with consistent
sonorant pre-glottalization will tend to neutralize to plain sonorants,
based on the perception of these segments in phrase-initial and postconsonantal
positions.9
Timing of laryngeal articulations relative to oral constriction is often
reversed for obstruents and sonorants (Sapir 1938, Kingston 1985), with
laryngeal events typically timed to peak at release for obstruents, but at
onset of oral closure in sonorants. Along with these differences in timing,
9 It would be of great value to study the frequency of single-word utterances or single-word
phrases in child-directed speech. If more content words occur either phrase-initially or
phrase-ﬁnally in child-directed speech than in non-child-directed speech, a child may be
more likely to generalize from phrase-ﬁnal to word-ﬁnal context, or from phrase-initial to
word-initial context on the basis of raw phonetic token frequency effects.
102 Sound patterns
we ﬁnd signiﬁcant differences in distributional patterns. Neutralization of
obstruent release features is determined by the phonetic context immediately
following the obstruent, while neutralization of sonorant onset
features is determined by the phonetic context immediately preceding
the sonorant. In the case of obstruent release features, the overarching
generalizations can be explained by the presence versus absence of consonant
release. For pre-glottalized sonorants, seemingly signiﬁcant crosslinguistic
generalizations can be explained in terms of the presence versus
absence of a preceding phonetic context facilitating the perception of
creak or a silent glottal closure. In both cases, common sound changes
are suggested which lead to common sound patterns: in the ﬁrst case, the
absence of release leads to neutralizing sound change for non-released
stops; in the second case, regular sound change takes pre-glottalized sonorants
to plain sonorants in phrase- or word-initial position.
4.4.3 The origins of obstruent pre-aspiration
Though pre-glottalization of sonorants and pre-aspiration of obstruents
may both be classiﬁed as phonetic closure features, signiﬁcant differences
set them apart. Despite distributional constraints, there is little evidence
from phonological alternations or the historical record suggesting neutralization
of obstruent pre-aspiration in word-initial or post-consonantal
position. In fact, a contrast between post-aspirated and unaspirated stops
is found initially, in languages like Icelandic, Gaelic, Nukuoro, and West
Futuna, suggesting that initial position is not a position of neutralization
for the phonological contrast, but one where the contrast has a distinct
phonetic realization. In all of these languages, pre-aspirates are realized as
post-aspirates word-initially.
Another signiﬁcant difference between pre-aspirates and post-aspirates
involves duration. While languages with allophonic pre- and postglottalized
sonorants provide little evidence of consistent durational difference
between these two segment types, pre-aspiration differs signiﬁcantly
from post-aspiration in terms of duration. Thr´ainsson (1978) notes
for Icelandic that “pre-aspiration typically has a normal segment length
in Icelandic, whereas postaspiration is much shorter . . . This suggests
that pre-aspiration is not simply the inverse of postaspiration, as its name
and some phonetic descriptions might lead us to believe.”10
Both the distribution of pre-aspirates and their bisegmental durations
in some languages suggest potential historical origins in geminates or
10 The length associated with pre-aspiration in Icelandic is also found in Skye Gaelic, but
not in Lewis Gaelic where pre-aspirates are as short as other stops (Ladefoged and
Maddieson 1996: 72).
Laryngeal features 103
consonant clusters. Geminates and consonant clusters in many languages
are absent word-initially, and word-ﬁnally, but present intervocalically
(see chapters 5 and 7). Consonant clusters and geminates typically have
longer durations than single segments, in the same way that pre-aspirated
segments in languages like Icelandic are longer than their post-aspirated
counterparts. It is possible, then, that some of the synchronic distributional
constraints on pre-aspirates may reﬂect earlier distributional
constraints on geminates or consonant clusters. An investigation of the
history of obstruent pre-aspiration in a variety of genetically unrelated
languages yields some support for this hypothesis (Blevins and Garrett
1993): in all cases where evidence is available, and where language contact
is not involved, pre-aspirates seem to derive from earlier geminates
or clusters. For example, in Lule Sami, pre-aspirates reﬂect medial geminates
in other dialects, while in Cree, where pre-aspirates are found only
after vowels, they are the result of a ∗
CC > hC sound change (Bloomﬁeld
1946: 88–90).11
4.4.4 Final devoicing
Some laryngeal features are cued by phonetic features internal to a segment.
One case of this kind is the obstruent voicing contrast: in some
languages, the primary cue for voiced (versus voiceless) stops is closure
voicing and/or closure duration. While most languages combine segmentinternal
voicing and duration cues with differences in VOT, preceding
vowel length, F0 and F1 values on adjacent vowels, and burst duration
and amplitude, there are some languages like French, Catalan, and Thai
where primary phonetic cues for the voicing contrast are closure voicing
and closure duration. Since, in at least some languages, the cues for
voicing are internal to the segment, and not aligned with stop release,
presence versus absence of release should not result in neutralization of
the voicing contrast. Nevertheless, we ﬁnd that in languages like Catalan
and Thai there is no voicing contrast for obstruents in word-ﬁnal position.
Assuming that the phonetic realization of voicing in these languages
has been constant for some time, ﬁnal devoicing cannot be attributed to
the absence of release.
This ﬁnding should not be surprising since a purely release-based
approach to ﬁnal neutralization of voicing contrasts also makes the
wrong predictions with respect to stops and fricatives. Under a purely
11 In Ojibwe, clusters whose ﬁrst member continues Proto-Algonquian ∗h, ∗x, ∗ʔ, or ∗␪
are realized in some dialects as “fortis” or long stops pp, kk, cc, tt and in others as
pre-aspirated stops.
104 Sound patterns
release-based approach, word-ﬁnal fricative devoicing is unexpected,
since fricative voicing is cued by presence versus absence of noise during
the fricative noise portion of the segment. And yet, many languages,
including German and Russian, show devoicing of stops and fricatives
word-ﬁnally. If release cues are not involved, what other phonetic factors
might play a role in the evolution of word-ﬁnal obstruent devoicing?
In Smith’s (1997) detailed phonetic study of the devoicing of /z/ in
American English, devoicing is found to be most common in two different
contexts: ﬁrst, where the fricative is followed by another voiceless
consonant; and second, at the ends of words and phrases, and in
unstressed syllables. Smith (1997) identiﬁes two distinct causes of devoicing.
In the ﬁrst case there is glottal abduction, and transglottal airﬂow is
insufﬁcient for vocal-fold vibration. Glottal abduction occurs in anticipation
of a glottal opening in a following voiceless sound, and also prepausally.
A reduced glottal abduction gesture may also occur in contexts
of reduction. These ﬁndings are consistent with the common devoicing
of ﬁnal unstressed vowels in many languages (see 8.2.1), as well as
the offglides to voicelessness which, for example, are possible after all
Klamath word-ﬁnal segments (Blevins 1993a). A second cause of devoicing
identiﬁed by Smith is when the transglottal pressure differential is too
small to allow voicing. This study suggests then that there are quantiﬁable
articulatory and aerodynamic factors which may lead to obstruent devoicing
preceding other voiceless obstruents, and obstruent devoicing phraseﬁnally.
In the current model of sound change, these articulatory and aerodynamic
factors, in particular those which result in word- and phrase-ﬁnal
devoicing, shape the set of phonetic variants from which learners must
choose to model their own sound system. Smith’s (1997) study, then,
suggests a phonetic source of devoiced ﬁnal obstruents (and sonorants)
for sound change with sources in CHOICE which is independent of release
features.
An additional potential phonetic source for cross-linguistic word-ﬁnal
obstruent devoicing is also independent of closure and release cues and
non-assimilatory in nature.12
In many languages, a phrase-ﬁnal syllable
is longer in duration than a segmentally identical phrase-medial syllable
(Klatt 1975; Wightman et al. 1992; Fougeron and Keating 1997). Phraseﬁnal
or pre-pausal lengthening can often result in segments which are two
to three times longer than their non-lengthened counterparts. In some
languages, like Modern Hebrew (Berkovits 1993), ﬁnal lengthening has
the greatest effect on ﬁnal plosives and fricatives. In addition, as noted
ﬁrst by Catford (1977), there is a near universal association between
12 I thank John Ohala for his comments on early versions of this hypothesis.
Laryngeal features 105
consonant duration and voicing: in obstruents, voiced obstruents are
shorter than their voiceless counterparts. Together, these two observations
suggest that phrase-ﬁnal lengthening is one phonetic source of wordﬁnal
obstruent devoicing, and that perceptual and articulatory factors
may both be involved.
If phrase-ﬁnal lengthening results in lengthening of closure duration,
obstruent voicing can be inhibited.13
The standard explanation for the
association between length and voicelessness is aerodynamic: when the
vocal folds are in the position for modal voicing, soon after obstruent closure,
voicing ceases due to the supralaryngeal pressure resulting from the
oral closure if no other strategies (e.g. relaxation of oral soft tissue, tongue
root advancement, jaw lowering, etc.) are made use of to sustain vocalfold
vibration. In addition, longer stop duration can contribute to a percept
of voicelessness, since voiceless stops are typically longer than their
voiced counterparts, or have lower V-to-C durational ratios than their
voiced counterparts (e.g Kohler 1979). A working hypothesis then is that
due to common phrase-ﬁnal lengthening, phrase-ﬁnal consonants may be
lengthened. As a result of this lengthening, two things may happen: voicing
may be inhibited as a result of prolonged closure duration; or ﬁnal
lengthened voiced obstruents may come to have durations not unlike
voiceless obstruents in non-ﬁnal position, and be misperceived as voiceless.
Either factor could result in a listener categorizing a word-ﬁnal voiced
obstruent as voiceless.
The proposed analysis has neither of the weaknesses of the purely
release-based approach outlined above. In languages like French, Catalan,
and Thai where primary phonetic cues for the voicing contrast are
closure voicing and closure duration, ﬁnal devoicing can evolve through
phrase-ﬁnal lengthening. In addition, such devoicing should affect both
stops and fricatives, since there is no association between devoicing and
absence of release. Another welcome result of separate phonetic accounts
of general ﬁnal devoicing and neutralization of release features is that it
allows for loss of features like breathy voice under non-release, without
neutralization of ﬁnal voicing. This is precisely the pattern observed
in some dialects of Bengali (Kenstowicz 1994: 193), where the wordﬁnal
contrast between Th
, T, D , D is neutralized to T versus D, with
loss of release features, but no loss of the voicing contrast. Another prediction
of a model in which ﬁnal devoicing is an aerodynamic consequence
of ﬁnal lengthening is that phonetic devoicing will be most common
in velar consonants, less common in coronal consonants, and less
13 This is related to the ﬁndings of Klingenheben (1927), Jaeger (1978), and Ohala (1983b,
1994a) that geminates tend to be voiceless.
106 Sound patterns
common still in labial consonants (Ohala 1983a). Where ﬁnal devoicing
has been captured in its earliest stages, there is evidence for this pattern.
In Tonkawa, word-ﬁnal g is devoiced to [k], but b, d are not (Hoijer
1933: 4). In Frisian, where ﬁnal devoicing of b, d, g is of recent origin,
Tiersma (1985: 30) reports early studies of the language of Grou by
Eijkman (1907) where b and d, in contrast to g, are still voiced. Another
implication of the ﬁnal-lengthening account, also illustrated by Frisian,
regards vowel-length effects. If phrase-ﬁnal lengthening results in generally
longer syllables, and if a language already has a long versus short
vowel contrast in ﬁnal syllables preceding voiced stops, then phrase-ﬁnal
lengthening may, in some languages, have a greater effect on consonants
following long vowels than short vowels, due to upper limits on the length
of sonorant portions of the syllable rime. In Sipma’s (1913) grammar of
Frisian, there are signs that devoicing has started to take place. Though
his transcriptions are not entirely consistent, they suggest that devoicing
took place ﬁrst after long vowels, falling diphthongs, and liquids, and only
later after short vowels or rising diphthongs (Tiersma 1985: 30).
At the same time, the suggestion that phrase-ﬁnal lengthening may
play a role in ﬁnal devoicing is not meant to rule out other phonetically
based sound changes which may result in convergent sound patterns. If
ﬁnal lengthening and ﬁnal non-release both play a role in neutralization
of voicing contrasts, this neutralization is expected to be more common
than ﬁnal neutralization of aspiration or ejection, which occurs only under
non-release. This explanation is also not meant to rule out other potential
factors, such as those noted by Smith (1997). Vocal-cord abduction in
pre-pausal position is not uncommon. However, in languages where this
occurs and appears to be phonologized (e.g. Klamath), both obstruents
and sonorants are followed by audible voicelessness word-ﬁnally.14
4.4.5 Common cases of sound change
In (4) sound changes giving rise to some of the common patterns of
laryngeal feature distribution in (1)–(3) are classiﬁed and summarized.
14 Baudouin (1895/1972: 209–10) was perhaps the ﬁrst linguist to explicitly note that sound
changes like ﬁnal devoicing not only occur in mature linguistic systems, but that they
also independently arise in the course of language acquisition. Stampe (1969) refers
to similar sound patterns in child-language development as “phonological processes,”
and suggests that they are part of universal strategies used in the course of language
acquisition. However, see 9.1.4, where most aspects of sound patterns during early stages
of language acquisition are argued to be due to maturational constraints on production.
In child-language phonology, neutralization of voicing contrasts is found word-initially
as well as word-ﬁnally, supporting this general view.
Laryngeal features 107
(4) Some sound changes resulting in recurrent patterns of
laryngeal feature distribution
Sound change Type Phonetic basis
i. Th
> T ,
T’ > T , D > T
CHOICE + CHANGE i. variation provides released
and unreleased tokens of
stops, with unreleased tokens
most common in certain
positions (ﬁnally,
pre-consonantally); ii. If a stop
is unreleased, laryngeal
features typically associated
with release can be absent, or
misperceived as being absent.
ii. ’R > R/ X
(X not a vowel)
CHANGE Sonorant pre-glottalization is
cued by vowel shift from
modal voicing to creak or
glottal stop. When vowel is
absent, pre-glottalization is
commonly misperceived as
being absent.
iii. [−son] >
[−voiced] / //
CHOICE + CHANGE i. General phrase-ﬁnal
lengthening results in
lengthened ﬁnal consonants; if
these variants are taken as
basic, sound change can
result.
ii. Lengthening may inhibit
obstruent voicing directly, or
lengthened tokens may be may
interpreted as voiceless, since
voiceless consonants are
typically longer than voiced
ones.
iv. TiTi > h
T,
TiTi > Ch
CHOICE Variation in timing of oral and
laryngeal gestures can give rise
to voiceless geminates
produced as pre- or
post-aspirated stops.
Distribution of resulting
aspirates will be identical to
original distribution of
geminates.
108 Sound patterns
The sound changes proposed in (4i) involve a shift from released to unreleased
consonants, and associated neutralizations of release features. The
typical contexts for these changes are phrase-ﬁnal position, word-ﬁnal
position, and pre-obstruent position; less commonly release is lost before
other consonants. The changes in (4i) are unattested in pre-vocalic position
within the word, since in this context consonants are released. Sound
changes like those in (4i) are common because presence versus absence
of release can be highly variable across time, speakers, and utterances.
Final release may be common in hyperarticulated speech, but may be
absent elsewhere.
In (4ii), pre-glottalized sonorants are misheard as plain voiced sonorants
when a preceding vowel is absent. This sound change has been
phonologized in languages like Yurok, giving rise to alternations, but the
same sound change may also account for the limited distribution of glottalized
sonorants in languages like Yokuts and Shuswap.
The ﬁnal devoicing sound change in (4iii) is well documented crosslinguistically,
and appears to have multiple sources. By relating this sound
change to ﬁnal lengthening, it is entirely independent of stop release. This
is a welcome result since obstruent devoicing in many languages applies
to both stops and fricatives, and occurs in languages where voicing is cued
by segment-internal features like closure duration and closure voicing. In
addition, the possibility of ﬁnal-stop devoicing via non-release and ﬁnal
lengthening suggests that ﬁnal-stop devoicing will be more widespread
than other types of laryngeal neutralizations, since it has multiple sources.
This appears to be the case.
In (4iv), the distribution of pre-aspirates is accounted for, in part, by the
reanalysis of geminate obstruents as pre-aspirates. Under this account,
distributional constraints on pre-aspiration may be directly inherited, and
reﬂect earlier constraints on the distribution of geminates.
4.5 Exceptional patterns of laryngeal feature distribution
4.5.1 Final voicing
The phonetic-historical explanations proposed above for common patterns
of laryngeal feature distribution are also meant to explain the
absence or uncommon occurrence of other patterns. Certain patterns
of laryngeal feature distribution will be unattested, or extremely rare,
since common sound changes will not give rise to them, and other common
sound changes will eliminate them. Above, ﬁnal-obstruent devoicing
is suggested as the convergence of at least three independently attested
types of sound change: CHANGE based on the common misperception of
Laryngeal features 109
unreleased consonants as plain voiceless consonants; CHANGE⁄CHOICE
based on phrase-ﬁnal laryngeal spreading gestures; and CHANGE⁄
CHOICE based on ﬁnal lengthening of obstruents which can result in
voicelessness or percepts of voicelessness. Together, these three factors
are claimed to account for the high frequency of word-ﬁnal obstruent
devoicing cross-linguistically. But these same factors will render other
patterns, like ﬁnal voicing, rare or infrequent.
Natural sound changes giving rise to obstruent voicing are of two basic
types: assimilation to a neighboring voiced consonant; or gestural reduction
in lenition contexts, with spontaneous voicing or a percept of voicing,
as a potential consequence.15
Since word-ﬁnal VC# contexts do not provide
a natural context for voice assimilation to a neighboring consonant,
the only natural source of ﬁnal voicing for VC# is lenition. However,
cross-linguistic surveys of lenition show that voicing is common only in
intervocalic contexts (LaVoie 2001: 31–32). In short, there is no single
natural sound change which will give rise to sound patterns involving
word-ﬁnal voicing. At the same time, if such a pattern were to arise,
either through rule telescoping, analogy, or by chance, the same factors
which account for the widespread occurrence of ﬁnal devoicing would
play a role in speeding the decay of the ﬁnal-voicing pattern. Unsurprisingly,
then, there are very few cases of word-ﬁnal voicing reported in the
literature.
Three cases within Austronesian are reported by the same author, and
are therefore somewhat suspect. K¨ahler (1946/1949, 1960), as reported
by Robert Blust (personal communication, 2001), transcribes only ﬁnal
voiced stops in two dialects of Malay, though elsewhere (e.g. word-initially
and medially) he records a voicing contrast. However, the consistency of
K¨ahler’s transcriptions across these three languages, and his mention
that ﬁnal stops are unreleased, suggest that his voiced stop symbols have
been used to transcribe precisely the same neutralized segments found
in Standard Malay, or languages more distantly related to Simalur. All
of these languages have undergone neutralization of ﬁnal ∗
-b, ∗
-d, ∗
-g to
voiceless (unreleased) -p, -t, -k.
Breton (Ternes 1970) has essentially the same distribution of voicing
as Lithuanian, with one complication. As in Lithuanian, voicing is contrastive
before sonorants, but elsewhere contrasts are neutralized. The
difference between Breton and Lithuanian is that in Lithuanian wordﬁnal
obstruents are voiceless, while in Breton they are voiceless, except
15 In this second case, voicing is often associated with other phonetic features of lenited
segments, including shortening of closure duration, reduction of intraoral air pressure,
or a shift from stop to continuant. See LaVoie (2001) for many examples.
110 Sound patterns
when a vowel-initial word follows within the phrase, in which case they
are voiced. Since, in Breton, word-ﬁnal obstruents are always devoiced
in phrase-ﬁnal position, or when followed by a consonant, it would seem
inappropriate to treat this as an instance of general word-ﬁnal voicing.
Rather, voicing of the word-ﬁnal segment is an instance of intervocalic
lenition at the phrasal level.16
Lezgian, a Nakh-Daghestanian language, has a limited set of alternations
which suggest ﬁnal-obstruent voicing through telescoping and rule
inversion (Yu 2001). Lezgian has a four-way laryngeal contrast between
ejectives, voiceless aspirated stops, voiced stops, and voiceless unaspirated
stops. In certain monosyllabic nouns, a plain voiceless stop in pre-vocalic
position alternates with a voiced stop in word-ﬁnal position: pab/pap-a
‘wife,’ pad/pat-ar ‘side,’ mez/mets-i ‘tongue,’ etc. This alternation is not
found in polysyllabic unsufﬁxed nouns, or in other lexical categories.
Yu (2001) suggests that these synchronic alternations reﬂect a historical
sequence of intervocalic pre-tonic gemination, where geminates were
voiceless, followed by degemination, and rule inversion. Synchronic ﬁnal
voicing in Lezgian, then, appears to reﬂect historical medial devoicing.
What Breton, its hypothetical descendants, and Lezgian are meant to
illustrate is that word-ﬁnal voicing is not ruled out within Evolutionary
Phonology. It is not ruled out as a general sound pattern, and it is not ruled
out as the output of a limited set of phonological alternations. Rather, the
common sound changes which make word-ﬁnal devoicing common will
make sound patterns limiting ﬁnal obstruents to voiced ones uncommon.
Since exceptionless patterns of word-ﬁnal voicing will never be the direct
transparent result of a phonetically motivated sound change, they are
predicted to be less common than patterns of ﬁnal devoicing. Their rarity
follows not from any intrinsic property of synchronic grammars, but from
contingent facts about the world.
4.5.2 Initial devoicing
Another rare and unexpected laryngeal feature pattern is word-initial
obstruent devoicing. Word-ﬁnal obstruent devoicing is attributed both
to the absence of release and to phrase-ﬁnal lengthening. Since wordinitial
pre-vocalic position is a context where obstruents are consistently
released, devoicing with the same phonetic source is unexpected.
However, phrase-initial strengthening is well documented (Fougeron
and Keating 1997; Fougeron 1999). While the effects of phrase-ﬁnal
16 The French of Quimper has borrowed the word-ﬁnal devoicing rule from Breton, but
not the phrasal rule of intervocalic voicing.
Laryngeal features 111
lengthening on consonant length seem more extreme, it is possible that
the few reported cases of word-initial devoicing have their origins in
phrase-initial lengthening. These include Somali (Armstrong 1934), with
devoicing of all voiced stops, and Pennsylvania German, with initial
devoicing of /b/ to [p]. In these cases, initial lengthening may inhibit
obstruent voicing, or simply give rise to the percept of voicelessness.
In Chamic (Thurgood 1999: 72–73), reﬂexes of Proto-Chamic wordinitial
breathy voiced stops have undergone neutralization to voiceless
stops when the following syllable begins with a voiceless stop. Where
the following syllable begins with a sonorant, breathy voiced stops are
maintained initially. And where the following syllable begins with a voiced
obstruent, there is maintenance of breathy voicing in Northern Roglai,
loss of voicing in Chru and Phan Rang Cham, and variable maintenance
of breathy voice in Jarai, and Western Cham. In reﬂexes of Proto-Chamic
monosyllables, voicing is maintained word-initially. Within Chamic, the
disyllabic domain is divided into a pre-syllable and main syllable. The
pre-syllable is short and unstressed, while the main syllable is longer and
stressed. Initial devoicing is limited to pre-syllables, and is conditioned
by the laryngeal features of the onset of the main syllable. This then, is
not a case of general word-initial devoicing, but a case of obstruent voice
assimilation across a short unstressed vowel.
Another pattern which might be erroneously classiﬁed as word-initial
devoicing is typical of those Pama-Nyungan languages which have only
a single laryngeal series of obstruents, and prohibit word-ﬁnal obstruents.
In these languages, obstruents occur either word-initially or between
sonorants (intervocalically, or after a sonorant consonant and before a
vowel). Nearly all languages with these phonotactics have the same pattern
of obstruent voicing: oral stops are voiceless word-initially, but tend
to be voiced medially. This is the pattern found, for example, in Panyjima
(Dench 1991: 130). This distribution of stop allophones is a consequence
of simple intersonorant lenition.
4.6 Summary
In this chapter the most common patterns of laryngeal feature distribution
in the world’s languages have been attributed to common types of
sound change. Certain patterns of laryngeal feature distribution reﬂect
the fact that laryngeal features of obstruents are more perceptually salient
in certain contexts than in others. In contexts in which these features
are less salient, neutralizing sound changes may occur, and it is these
sound changes which are reﬂected in the most common distributional
patterns of laryngeal features. The same phonetic contexts of perceptual
112 Sound patterns
saliency deﬁne triggers in rules of laryngeal feature assimilation. Phraseﬁnal
lengthening as well as the association between phonation types and
prosodic boundaries may also play a role in the common process of wordﬁnal
devoicing.
At the same time, uncommon patterns of laryngeal feature distribution
can also be explained in terms of sound change. Final voicing is rare not
only because sound changes giving rise to ﬁnal devoicing are common,
but also because there is no single phonetically motivated sound change
whose immediate output will give rise to this pattern. Rule inversion
and rule telescoping may give rise to patterns of ﬁnal obstruent voicing,
suggesting that the rarity of such patterns does not follow from universal
markedness constraints.
Within Evolutionary Phonology, the relationship demonstrated by Steriade
(1999a) between positions of laryngeal feature neutralization and
phonetic cues for laryngeal features are accounted for by positing phonetically
based sound changes with sources in CHANGE, CHANCE, and
CHOICE. These sound changes give rise to synchronic sound patterns
with similar characteristics, but allow synchronic phonological systems
to remain free of reference to phonetic features, and to diverge dramatically
from phonetically natural sound patterns when rule inversion, rule
telescoping, analogy, or language contact are involved.