The phonetics and phonology of non-modal vowels: a cross-linguistic perspective

95

spreading  is  governed  by  the  same  restrictions  governing  the  realization  of

underlying non-modal vowels in Kedang and Jalapa Mazatec.  Realizing non-modal

voicing on the last half of a long vowel in Hupa still leaves a portion  of  modal

voicing from which place information may be easily recovered.  Laryngeal features

cannot spread onto a short vowel, since this would completely obscure the vowel.

Crucially, as in Kedang, there are no tonal contrasts present in Hupa which would

block laryngeal spreading in non-modal voicing; it is thus the desire to realize place

information saliently which is driving the data.  

The durational patterns of non-modal voicing can be modeled in a constraint-

based grammar using a few constraints.  First, there are two constraints against non-

modal vowels, one prohibiting non-modal short vowels (*N

ON

-

MODAL 

S

HORT 

V),

the other against non-modal long vowels (*N

ON

-

MODAL 

L

ONG 

V).  Let us assume

that the constraint against non-modal long vowels is violated once for each half of

the vowel which is non-modal, i.e. one violation for each timing position associated

with non-modal voicing.  Thus, a fully non-modal long vowel violates this constraint

twice, while a partially non-modal long vowel violates it once.  In Hupa, we must

also  assume  a  constraint  which  forces  non-modal  voicing  to  spread  from  a

preconsonantal consonant  onto  the  preceding  vowel.    The  relevant  constraint  is

motivated  by  the  requirement  that  laryngeal  features  of  a  consonant  not  be

completely overlapped by the consonant constriction (see Gordon 1998 for more

discussion).  Here I will simply formulate the constraint as *S

PREAD 

L

ARYNGEAL

F; this constraint requires that laryngeal features (creak or voicelessness depending

on the consonant) spread from preconsonantal obstruents onto an adjacent vowel.

By ranking *S

PREAD 

L

ARYNGEAL 

F below *N

ON

-

MODAL 

S

HORT 

V but above

*N

ON

-

MODAL 

L

ONG 

V, we get the Hupa facts.

We can also account for the Kedang and Jalapa Mazatec patterns if we assume

that non-modal vowels in these languages are underlyingly linked  to  one  timing

position reflecting their phonemic quantity as short vowels, but two timing positions

on the surface, reflecting their substantially longer surface duration.  By ranking the

constraint against insertion of timing positions not present underlyingly,  phrased

here as *D

EP

-X, following McCarthy and Prince’s (1995) Correspondence Theory,

below *N

ON

-

MODAL 

S

HORT 

V but above *N

ON

-

MODAL 

L

ONG 

V, we account for

the fact that, in Kedang and Jalapa Mazatec, non-modal vowels are phonetically quite

long, but non-modal for only portion of the vowel.

3. V

OICELESS VOWELS AND THE ROLE OF ARTICULATORY FACTORS

.  Thus far, I

have provided a perceptually-driven explanation for why  non-modal  vowels  have

such  a  limited  distribution  cross-linguistically  in  comparison  to  modal  voiced

vowels.  This account makes the prediction, borne out in the data presented thus far,

that, certain languages will disprefer non-modal voicing on short vowels. 

Interestingly, as it turns out, there are many languages which devoice short but

not long vowels, a  pattern  which  runs  opposite  to  the  predictions  made  by  the

perceptually driven explanation offered in the previous section.  The presence of

both  patterns  cross-linguistically,  devoicing  of  short  but  not  long  vowels,  and

devoicing of long but not short vowels, deserves explanation.  In sections 3.2-3.5, I

will address the asymmetries related to vowel length, as well as other asymmetries

gleaned  from  a  typology  of  approximately  50  languages  (see  the  Appendix).

Approximately half of the typology is drawn from Crother et al.’s database (1979),

while most of the remaining languages are mentioned or discussed in either Cho

(1993), Vine (1981), or Jun et al. (1997).  

96

3.1.  T

HE  STATUS  OF  NON

-

MODAL  VOICING

:   

PHONETIC  OR  PHONOLOGICAL

.

Before preceding with the typology, it is appropriate  to  address  the  question  of

whether non-modal voicing is a phonological or phonetic phenomenon or perhaps

both, depending on the language.  Clearly in languages with an underlying contrast

between non-modal voiced vowels and modal voiced vowels (e.g. Sedang, Gujarati,

Jalapa  Mazatec),  non-modal  vowels  are  a  synchronic  phonological  feature

1

.

However, as pointed out earlier, the number of languages with underlying or even

surface contrastive phonation type for vowels is quite small.  In  the  majority  of

languages in which they  occur,  non-modal  vowels  are  a  surface  non-contrastive

property, and thus less clearly belong to the phonology.  

The issue of the phonological vs. phonetic status of non-modal vowels has been

most thorough investigated for voiceless vowels, e.g. in relatively recent work by

Vine (1981), Cho (1993), Tsuchida (1994), Jun and Beckman (1993), Jun  et  al.

(1997).   In the majority of languages for  which  vowel  devoicing  has  been  the

subject of intensive acoustic analysis, vowel devoicing appears to be a gradient rather

than  a  categorical  phenomenon;  languages  falling  into  this  category  include

Japanese (Han 1961, Beckman 1982, Tsuchida 1994), Montreal French (Gendron

1966,  Cedergren  and  Simoneau  1985),  Greek  (Dauer  1980),  Turkish  (Jannedy

1995) and Korean (Jun and Beckman 1993, 1994, Jun et al.1997, 1998)

2

.  In these

languages, vowel devoicing operates on a continuum with token to token variation in

the presence or absence or degree of devoicing.  On one end of the continuum is a

voiced vowel, at the other extreme is vowel deletion; various degrees of devoicing fall

in between these two extremes.  The likelihood of devoicing is a function of various

phonetic  factors:    position  of  stress/accent,  distance  from  prosodic  boundaries,

vowel height, surrounding consonants, and  speech  rate.    The  gradient  nature  of

vowel devoicing is even suggested in many grammars which describe devoicing as

optional  but  not  required  in  a  given  environment  (e.g.  Tongan,  Acoma,  Tubu,

Boraana  Oromo,  Kawaiisu,  Big  Valley  Shoshoni,  Mokilese,  Cocama)  or  in

languages where the span of devoicing can vary in length (Acoma, Southern Paiute).

It  is  possible  that  instrumental  work  would  demonstrate  that,  in  a  great  many,

perhaps most, languages, devoicing is a gradient phenomenon.     

On the other hand, there are many languages in which vowel devoicing behaves

like a phonological phenomenon.  In some languages, voiceless vowels contrast on

the surface with voiced vowels.  For example, the word-internal contrast between

short and long vowels is realized as a contrast between voiceless and voiced vowels

in word-final position in Oromo and in Woleaian.  Similarly, in Hupa as discussed

in section 2, vowel devoicing is contrastive before many syllable-final consonants.

If we adopt the standard assumption that contrastive properties are  phonological,

vowel  devoicing  would  clearly  fall  under  the  purview  of  phonology  in  Oromo,

Woleaian and Hupa.  Furthermore, in other languages, vowel  devoicing  interacts

with  other  phenomena  which  are  typically  assumed  to  be  phonological.    For

example,  vowel  devoicing  influences  stress  assignment  in  Awadhi,  pitch  accent

placement in Tunica, and debuccalization and tone shift in Comanche.  Furthermore,

in  Tongan,  one  of  the  prerequisites  for  vowel  devoicing  is  that  vowels  be  in

morpheme-final  position;  such  morphological  conditioning  would  suggest  that

vowel devoicing is not merely a low level phonetic phenomenon.  In summary, vowel

devoicing thus appears to operate at a relatively deep level of the grammar in a fair

number of languages

3

.   

Perhaps not surprisingly, languages in which vowel devoicing clearly appears to

be phonological display vowel devoicing in the same  environments  (e.g.  domain

finally, adjacent to voiceless consonants) in which vowel devoicing is most likely to

câu 3

97

occur in languages where it has been demonstrated to be gradient.   Even if one

assumes a sharp distinction  between  vowel  devoicing  as  a  phonetic  process  vs.

devoicing as a phonological one, the striking similarity between the distributions of

phonetic and phonological vowel devoicing suggests that examination of phonetic

devoicing may also provide insight into phonological devoicing.  For this reason, the

typology  in  this  paper  includes  cases  of  vowel  devoicing  which  are  clearly

phonological as well as others which may not be  This paper will not address the

issue  of  where  to  draw  the  line  between  phonological  and  phonetic  processes.

Crucially, because substantially the same phonetic factors condition devoicing in all

languages, many aspects of the analysis of devoicing are more likely than not to be

quite similar for all languages with voiceless vowels.       

3.2. T

HE LENGTH ASYMMETRY

. Of the 32 languages in the survey with contrastive

vowel length in environments  targeted  by  devoicing,  devoice  short  but  not  long

vowels occurring in the same environment (in many languages, only high vowels

devoice; see section 3.4):  e.g. Awadhi, Big Smokey Valley Shoshoni, Bulu, Mbay,

Cheyenne,  Cocama,  Gadsup,  Galla,  Ik,  Inuit,  Oneida,  Goajiro,  Tarascan,  Zuni,

Japanese, Kawaiisu, Mokilese, Sámi, Sara, Shina, Bagirmi, Tongan, Tubu, Tunica,

Turkish, Woleaian).  Four languages (Boraana Oromo, Papago, Southern  Paiute,

Ket

4

) possess voiceless short vowels and also devoice a portion of long vowels in

certain  environments

5

,  one  (Acoma)  possesses  voiceless  short  vowels  and  long

vowels  which  are  completely  voiceless,  and  one  language  (Hupa)  has  partially

voiceless long vowels but lacks voiceless short vowels.  In  one,  Cheyenne,  long

vowels partially devoice in final position but do not devoice at all in word-medial

environments in which short vowels devoice.

3.3.

 

E

NVIRONMENT OF 

D

EVOICING

. In virtually all languages in the survey, vowel

devoicing is found at least  in  final  position;  in  many  languages,  devoicing  also

occurs in other environments as well.  Crucially, the occurrence of devoicing in non-

final  environments  almost  always  implies  devoicing  in  final  position.    The

languages I know of which are exceptional in this regard are Inuit (Crothers et al.

1979), Quechua (Crothers et al.), Turkish (Jannedy 1995), Azerbaijani (Crothers et

al.) and Montreal French (Gendron 1966, Cedergren and Simoneau 1985); vowels

in  these  languages  resist  devoicing  in  final  position,  but  allow  it  in  other

environments.    However,  in  three  of  these  languages  (Turkish,  Azerbaijani  and

Montreal  French),  final  vowels  are  stressed,  thereby  explaining  their  failure  to

devoice (see section 3.5).  In Inuit, phrase final position is typically associated with a

high tone which also often blocks devoicing cross-linguistically (see section 3.5).

Up to this point, I have been intentionally vague in defining “final position”.

The  reason  for  this  is  that  the  domain  of  devoicing  varies  from  language  to

language; however, these domains follow an implicational hierarchy.  Devoicing in

final position of a smaller domain (e.g. word) implies devoicing in larger domains

(e.g. phrase, utterance); the reverse of this statement is not necessarily true.  In 20

languages in the survey (Ik, Dafla, Cocama Galla, Bagirmi, Turkana, Sara,  Tubu,

Mbay,  Malagasy,  Campa,  Tarascan

6

,  Ticuna,  Ket,  Ainu

7

,  Island  Carib,  Zuni,

Washkuk, Goajiro, Woleaian

8

), voiceless vowels occur word-finally (and of course,

by implication, finally in larger domains as well).  In 14 languages, devoicing is

characteristic  only  of  final  position  of  larger  domains,  e.g.  phrase  or  utterance

(Alabama, Papago, Greek, Tarascan, Totonac, Chontal, Gadsup, Oneida, Apinaye

9

,

Mixtec, Nyangumata

10

, Boraana Oromo, Cheyenne, Kawaiisu

11

).  Note that, from

most descriptions, it is impossible to make distinctions among larger domains such

98

as the phonological phrase, intonational phrase or utterance.

12

  Interestingly, only

one language, Cocama, regularly devoices initial vowels in addition to final vowels;

in  both  environments,  devoicing  only  affects  vowels    adjacent  to  a  voiceless

consonant.

After final position, the next most common position in which vowels devoice is

adjacent to voiceless consonants.  Word-medial devoicing is found in 19 languages

in the survey (Mandarin, Brazilian Portugese, Malagasy, Mixtec, Quechua, Goajiro,

Azerbaijani,  Inuit,  Chontal,  Montreal  French,  Cheyenne,  Mokilese,  Big  Valley

Shoshoni, Japanese, Turkish, Korean, Tongan, Cocama, Papago).  It is interesting to

note that devoicing of final vowels in most languages (29 of 36), occurs not only

after voiceless but also after voiced consonants

13

.  In only 6 languages with final

devoicing (Japanese, Korean, Tongan, Turkana, Cocama, Mixtec) must  the  vowel

both be final  and next to a voiceless consonant for  devoicing  to  occur

14

.    In  8

languages with word-medial devoicing it is sufficient to have a voiceless consonant

on only one side of a vowel to trigger devoicing.  In  5  of  these  languages,  the

triggering  consonant  is  on  the  right  side  of  the  vowel  (Big  Valley  Shoshoni,

Comanche, Southern Paiute, Goajiro, Quechua), in 3, it is on the left side (Acoma,

Mandarin,  Chontal).  In  9  languages  with  word-medial  devoicing  (Cheyenne,

Mokilese, Japanese, Turkish, Korean, Montreal French, Tongan, Papago, Malagasy),

devoicing  is  described  as  affecting  vowels  (almost)  exclusively  between  two

voiceless consonants.  Further asymmetries between different voiceless consonants

will be discussed in section 4.

 

3.4. T

HE 

H

EIGHT 

A

SYMMETRY

.  Vowel devoicing is also sensitive to vowel height;

in many languages, high voiceless vowels but not mid and low  voiceless  vowels

occur  (Greek,  Korean,  Turkish,  Dafla,  Montreal  French,  Mokilese,  Brazilian

Portugese, Mandarin, Campa, Mixtec, Ainu, Azerbaijani, Gadsup, Inuit, Ticuna).

15

Similarly, in Tongan, the set of environments in which non-high vowels devoice is a

subset of the environments in which high vowels devoice.

16

  I know of no language

which devoices non-high vowels but not high vowels.

3.5. T

HE 

S

TRESS AND 

T

ONE

/I

NTONATION 

A

SYMMETRY

.  Two other asymmetries

in vowel devoicing relate to the closely related properties of tone and accent.  In all

languages in the survey for which data on accent location is reported and in which

the other necessary preconditions for devoicing are present, accented vowels resist

devoicing  (Montreal  French,  Turkish,  Tongan,  Comanche

17

,Cheyenne,  Brazilian

Portugese, Azerbaijani, Quechua).  In keeping with this pattern, in Papago, the set of

environments  in  which  stressed  vowels  devoice  is  a  subset  of  those  in  which

unstressed vowels devoice.  I know of no language with devoicing of stressed but

not unstressed vowels.  

A final asymmetry is that many tone and pitch accent languages fail to devoice

high-toned  vowels  (Japanese

18

,  Cheyenne,  Acoma

19

)    Furthermore,  in  some

languages with stress, intonational pitch accents (Greek, Boraana Oromo, Tunica)

and  high  boundary  tones  (Inuit)  can  inhibit  devoicing.    The  stress  and

tone/intonational  asymmetries  are  presumably  closely  related  since  accented

syllables often carry high pitch accents cross-linguistically, as in Japanese.  

4. A

N ARTICULATORY ACCOUNT  OF  VOWEL  DEVOICING

.    The  asymmetries  in

vowel devoicing discussed can, in large part, be explained in terms of a combination

of  articulatory  overlap  between  neighboring  glottal  gestures  and  aerodynamic

considerations.   First,  let  us  consider  the  patterns  which  are  compatible  with  a

gestural overlap account of devoicing. The reasoning given here basically follows

99

that of Dauer (1980), Jun and Beckman (1993, 1994), Jun et al. (1997, 1998).  The

types of vowels which devoice cross-linguistically are those which are likeliest to be

produced with voicing gestures which are durationally shortest.  Trivially, phonemic

short vowels are phonetically shorter than phonemic long vowels.  It is also well

known that high vowels are shorter than non-high vowels (Lehiste 1970) and that

unaccented vowels are shorter than accented ones.  

Because of their shorter duration, the glottal adduction gestures associated with

phonemic short vowels, unstressed vowels, and high vowels are more likely to be

overlapped by the glottal gestures of neighboring segments.  When the neighboring

gestures are abduction gestures, as in the case of voiceless consonants, they threaten

to overlap the adduction gestures for voicing of the vowel.  When sufficient overlap

occurs, vowel devoicing results.   The  overlap  account  makes  the  prediction  that

devoicing is more likely to occur when a vowel is surrounded on both sides  by

voiceless consonants.  This prediction is borne out by a number of instrumental

studies, e.g. Jun and Beckman 1994, Jun et al. 1997, 1998 on Korean, Han 1961 on

Japanese, Jannedy 1995 on Turkish, Dauer 1980 on Greek.  It also is supported by

the  fact  that  devoicing  in  many  languages  is  only  triggered  when  a  vowel  is

surrounded  by  voiceless  consonants.    The  overlap  account  also  predicts  that

devoicing is most likely to occur in the vicinity of voiceless consonants with the

greatest glottal abduction gestures and in the vicinity of voiceless consonants whose

glottal abduction peaks are timed to occur near the vowel.  In general, this prediction

is also borne out.  Voiceless consonants with the largest glottal openings, (fricatives-

-Löfqvist and Yoshioka 1980), and those whose peak glottal abductions fall close to

a vowel (aspirated stops--Kagaya 1974, Pétursson), tend to trigger devoicing most.

Thus,  in  languages  with  unaspirated  stops,  fricatives  are  most  likely  to  trigger

devoicing than stops; e.g. fricatives but not stops trigger devoicing in Comanche.  In

Mokilese, devoicing is most likely next to an /s/. In Goajiro, devoicing of vowels

occurs before voiceless fricatives and affricates, which presumably also often have

relatively  large  glottal  openings  (cf.  Kagaya).    In  Southern  Paiute  word-medial

devoicing is triggered by a following fricative or geminate stop; geminates have been

shown  to  have  greater  glottal  apertures  than  singletons  (Pétursson  1976).    In

Turkish, devoicing is more likely in the neighborhood of phonologically unaspirated

stops  than  fricatives;  however,  phonetically,  as  Jannedy  (1995)  points  out,  the

“unaspirated” stops of Turkish are characterized by substantial aspiration which

perhaps accounts for the preferential devoicing of vowels following stops.    This

hypothesized  link  between  aspiration  duration  and  likelihood  of  devoicing  is

compatible with the fact that /k/, the stop with the longest aspiration duration cross-

linguistically,  is  the  only  consonant  to  trigger  devoicing  in  Tunica.    However,

aspiration duration is not the entire story, as Jannedy points out, since a preceding

/

p/  is  more  likely  to  trigger  devoicing  than  /k/  in  Turkish  even  though  /p/’s

aspiration duration is shorter than /

k/’s.

Certain  languages show place asymmetries in the set of fricatives which trigger

devoicing; e.g. /h/ but not /s/ or /f/ triggers devoicing of /a/ in Tongan, /s/ is much

more likely than /f, 

T, x/ to trigger devoicing of high vowels in Greek, /h/ but not /s/

triggers devoicing of short vowels in Big Smokey Valley Shoshoni.  In Mandarin,

devoicing is most common after aspirated affricates and after  voiceless  fricatives

tải về 78.41 Kb.

Chia sẻ với bạn bè của bạn: