Dept. for Speech, Music and Hearing

Quarterly Progress andStatus Report

Distinctive features andphonetic dimensions

Fant, G.

journal: STL-QPSRvolume: 10number: 2-3year: 1969pages: 001-018

http://www.speech.kth.se/qpsr

STL-QPSR 2-3/1969

I. SPEECH ANALYSIS

A. DISTINCTIVE FEATURES AND PHONETIC DIMENSIONS*

G. Fant

The purpose of this paper is to express some comments on the recent

developments of distinctive feature theory with specific reference to the

work of Chomsky and Halle (1968). On the whole I consider the i r feature

sys tem to be an improvement over that of Jakobson, Fant , and Halle (1952),

one of the main advantages being the introduction of a se t of tongue body

features in common for vowels and consonants but separate f rom the con- I

sonantal "place of articulation" features. The basic philosophy of treating

phonetics a s an integral par t of general linguistics demands that features

i n addition to the i r classificatory function shall have a definite phonetic

function reflecting independently controllable aspects of the speech event

o r independent elements of perceptual representation. However, there is

a danger that the impact of the theoretical f r ame with its apparent mer i t s

of operational efficiency will give some readers the impression that the se t

of features i s once for a l l established and that their phonetic bas is has been

thoroughly investigated. This i s not so. Many of the i r propositions a r e in-

teresting and stimulating starting points fo r fur ther r e sea rch whereas

others I find in need of a revision.

As pointed out by Chomsky and Halle there a r e s t i l l ser ious short-

comings in our general knowledge of the speech event. Their feature sys-

t e m i s a lmost entirely based on speech production categorizations. The

exclusion of acoustical and perceptual cor re la tes was a pract ical limitation

in the scope of the i r work but a l so appears to note the importance layed on

the production stage. It is f a r eas i e r to construct hypothetical feature

sys tems than to t e s t them on any level of the speech communication chain.

This i s really our present dilemma. Until we have reached a more solid

bas is in general phonetics any feature theory will remain "preliminary".

Here follows m y reaction to some of the basic i ssues in chapter seven

of Sound Pa t t e rn of English. My ea r l i e r comments on distinctive feature

theory may be found in the l i s t of references, Fant (1960a, b, 1966, 1967,

1968).

9 submitted f o r publication in the proceedings of the Second International Congress of Applied Linguistics, Cambridge, England, Sept. 8- 12, 1969.

STL-QPSR 2-3/1969

1. Will we ever have a language universal, finite, and unique se t of distinctive features?

The universality aspects a r e attractive but I a m somewhat pessimistic

about the outlooks. Fea tures a r e a s universal a s the sound producing con-

s t raints of the human speech producing mechanism and a finite number

should suffice for the classificatory function. However, I a m ra ther scep-

t ical concerning the uniqueness and thereby a definite number of features

since one and the same facts often can be described in alternative forms

and the c r i t e r i a f o r selecting an optimum sys tem a r e not very rigid. Even

if we had a l l the knowledge we needed the choice of features would be de-

pendent on the part icular weight given to phonetic and general linguistic

considerations and the preferences of the investigator would in the last in-

stance determine some of the selections. The problem is the following.

2. Are the demands on a feature sys tem different on the classificatory level and the phonetic level?

There a r e two ways of arriving a t features: (1) by selecting an inven-

tory of c l a s ses suitable f o r encoding of language s t ruc tures and then deter-

mine the i r phonetic cor re la tes o r (z), to s t a r t with an exhaustive analysis

of the modes and constraints of the speech producing mechanisms and per -

ception and determine their distinctive function i n language. Fea ture

theory has to develop along both lines and investigators differ only in the

relative importance layed on one o r the other. The main approach of

Jakobson et a1 (1 952) was t o s t a r t out with an ordering of phonemic oppo-

sitions and to identify minimal distinctions a s the same if motivated by

phonetic s imilar i t ies . The demand for a smal les t possible number of

features and the fargoing identification of features within the vowel and

consonant systems, e.g. that of identifying the relation between dentals

and labials with that of front and back vowels, resulted in an unavoidable

pay-off between encoding efficiency and phonetic reali ty and specifiability.

Chomsky and Halle (1968) avoided some of these difficulties by introducing

a g rea te r number of features.

One of the i r basic issues i s that a feature sys tem in addition to the

classificatory efficiency should conform with a natural phonetic systemati-

zation. How have they managed in this respec t? In many instances such

a s dealing with the c l a s ses of fricatives, stops, nasals, l a te ra ls , etc. , the solution is s t raight forward. On the other hand, I find the encoding of

STL-QPSR 2-3/1969 3. - .

the class of labial consonants a s [+ anterior) and [ - coronal] to constitute

a c lear case of departure from the unifying principles. One single phonetic

dimension, "labiality", which has a distinctive function has here lost i ts

identity on the phonological level. It appears to be a rather far-fetched

hypothesis that the actual neural encoding of labial consonants a t some

stage should include a selection of a maximal anterior point of articula-

tion in the vocal t rac t and a lack of tongue t ip evaluation in order for a

lower level to find out that this command has to be executed by the lips

and not the tongue.

The major c lass features "vocalic" and "consonantal" introduced al-

ready in the work of Jakobson et a1 and the features "sonorant" and syl-

labic display a complicated system of interdependencies as will be de-

scribed in la ter sections.

The starting point for the major c lass features appears to have been

the need to encode certain pre-established phonetic classes whereas the

voiced-voiceless feature i s a typical example of the opposite approach,

i. e. to s tar t out with a natural phonetic dimension and study i t s distinctive

role in language. A natural linguistic class , i. e. a l l [ r ] -phonemes, may

have rather complicated sets of phonetic correlates and a natural phonetic

dimension a s voicing may have to be studied together with several other

dimensions a s tensening, durations, and coarticulation when i t comes t o

the discussion of i t s distinctive role,

Before we can accomplish the happy marriage between phonology and

phonetics we have to work out the rules for predicting the speech event

given the output of the phonological component of grammar. To me this

i s the central, though much neglected, problem of phonetics and i t i s of

the same magnitude a s that of generative g rammar in general and will re-

quire a s imilar se t of transformational rules, The starting point i s the

feature matrix of a message a s successive phonological segments, i. e,

columns each with i ts specific bundle of features, i. e. rows, The parti-

cular choice of classificatory features a t this stage i s not very important

providing the conventions relating phonemes to alternative features sys-

tems a r e known.

The derivation of the rules of this "phonetic component" of language

aims a t describing the speech production, speech wave, o r perception cor-

relates of each feature given the "context" in a very general sense of

STL-QPSR 2-3/1969 4.

co-roccurking features within the phohological segment a s well a s those of

following and preceding segments. One se t of sequential constraints a r e

expreeraible a s coakticulatibn rules which may be both universal and lan-

guage ispecificr

In additidn to these more o r l e s s iner t ia dependent laws of connecting

vocal ges tures there may exist rules of neural reorganization of control

signals fo r modifying the physi ca l manifestation of a feature in accordance

with a principle of leas t effort articulation, o r the contrary, a compensa-

tion for maintaining o r sharpening of a phonetic distinction dependent on

what features occur o r follow in the t ime domain. In addition there enter I rules for modifications dependent on s t r e s s patterns, intonation, tempo,

speaker , sex, type, and dialect, attitude etc. Rules fo r speech segment

durations and sound shapes have to be expressed in t e r m s of l a r g e r phono-

logical segments, generally severa l syllables defining a natural rhyth-

mical unit in t e r m s of s t r e s s and intonation. Very l i t t le is known about

these rules. There is some evidence that the phase of maximal intensity

increase within a syllable is a reference point for ordering rules concern-

ing segment durations (B . Lindblom, personal communication).

This "phonetic component!' of the speech event receives very l i t t le

attention in the work of Chomsky and Halle who merely re fer to the phone-

tic cor re la tes of a feature a s a sca le with many steps instead of the binary

scaling on the classificatory level. A knowledge of linguistic structuring

is of grea t importance in practical communication engineering undertakings

such a s the administration of synthesis by rule o r automatic identifications.

However, without access to the rules of the "phonetic component" the

phonetic aspect of features becomes a s imaginary and empty a s the "Em-

peror ' s New Clothes" in the s tory of H. C. Andersen. Observing the

speech wave we a r e not faced with phonemes o r features but sound seg-

ments and more o r l e s s continuous sound shapes with a reciprocal many-

to-one relation between phonol.ogica1 and physical units. The same is t rue

of speech production studied in relation to the phonological transcript. In

both cases there is the need to define inventories of physical units, Fant

(1968), which a r e not identical to the distinctive features but a r e used to

define the i r phonetic correlates . It may be quite pract ical t o r e fe r to a

specific sequence of segments a s a stop followed by a fricative a t the phon-

etic level while we may want to re fer to the whole unit a s a n affricate on

the phonological level.

STL-QPSR 2-3/1969 5.

Those who want to increase their ? e r s ~ e c t i v e s on phonology in relation

to phonetics should read Ladefoged' s monograph "Linguistic phonetics l1

(1 967a). A pure phonetic sys tem was outlined by G. E, Pe terson (1 968).

3. What is the psychological reali ty of features?

As demonstrated in the previous section features must, a t l eas t under

prototype conditions, have physical cor re la tes a s observed by a n external

observer of the speech communication ac t and they should hopefully re -

flect categorical phenomena i n the encoding and decoding mechanism.

This is not the s a m e a s ascribing each feature to a specific brain alloca-

tion. We can be aware of a feature by introspection but otherwise it may

lack immediate neurophysiological correlates . The important thing is that

the actual processes a r e phenomena that have some abs t rac t relation to our

feature mat r ices ,

4. Is the binarv ~ r i n c i ~ l e i m ~ o r t a n t ?

No, not necessar i ly , but i t is convenient. Language regularit ies and

language developments may in some instances b e more easily described

by sca les of three o r more levels, cif. Ladefoged (1967a). It i s a l so

questionable whether formulations in t e r m s of feature mat r ices always

reveals m o r e fundamental rules than formulations in t e r m s of phonemes.

5. Are features independent and orthogonal?

This question can pertain both to the classificatory, "phonological

level", and to the phonetic level discussing the production speech wave

and perceptual correlates . Besides the apparent constraints on possible

sequences of phonological segments there exist universal constraints on

feature combinations within one and the same segment. As discussed by

Chomsky and Halle [t high] would contradict [t low]. Also, some features

o r combinations of features imply specific signs of other feature i n the

same bundle, a s exemplified by [t vocalic] implying [t sonorant]. A

c loser analysis of interdependencies within the major c l a s s features r e -

veals that the c l a s s of [+ sorlorants] by definition a lso incorporates all

[t syllabics] and a l l C - consonantal] segments. Such constraints will be

discussed in g rea te r detail in the section of major c l a s s features. The

phonological dependencies within this se t of features a r e paralleled by

phonetic s imilar i t ies . Thus the c l a s s of C - cons onantala incorporating

vowels and glides must have much in common with the c l a s s of [t vocalic]

incorporating vowels and liquids. In other words "vocalic" i s a lmost the

negative of the "consonantal" feature.

The phonetic interdependencies a r e apparent even when they a r e not

paralleled by classificatory constraints. The situation had been ideal in I the vowel sys tem if the perceptually relevant number of dimensions had

been the same a s the number of classificatory features. We would have

had a perfect orthogonal sys tem if limited to the [t low] o r [ - high] and '

the [ - back] dimensions corresponding to the +F1 and +F2 dimensions,

respectively. The feature "rounding" i s correlated with - ( F ~ + F ~ + P ~ ) and

thus only partially independent of other features. The same i s t r u e of the

feature "tense" which is related to the formant pattern (direction towards

an extreme target) and duration. Additional features and/or sca le values

a r e needed f o r the Swedish vowel sys tem as will be discussed la ter .

We accordingly have to r e so r t to the minimal c la im of Chomsky and

Halle that features should be a t least partially independent. At the same

t ime we have to be aware of considerable interdependencies. This applies

to the i r classificatory function a s well as to the i r phonetic cor re la tes .

6 . Are differences in feature contents of mat r ices a reliable measure of phonetic distance?

No, not always. On a n average basis i t might be permissible to ex-

p r e s s differences between languages o r dialects by summing binary units

i n the classificat ory domain and expect such differences to represent the i r

phonetic differences, Ladefoged (1969). However, one cannot expect the

phonetic difference between any two phonemes to be proportional to the

number of features by which they differ. The situation was especially

severe in the Jakobson, Fant, and Halle system, It was stated that the

[ Q ] and the [i] of the word "wing" do not have any features in common,

the [i] being [$. voc] [ - cons)[ - compact][ - grave], the [Q] being [ - voc]

[t cons] [+ nasal] [t compact]. On the phonetic level, on the other hand,

the difference between the [i] and the [ n l is minimal since the ent i re [i]

is nasalized and the transit ion f rom [i] to [n] merely involves a ges ture

of tongue c losure which in dialectal variants is omitted. Within the

C homs ky-Halle f ramework the situation i s indeed improved since the tongue

body features [ -back][-low][+high] a r e in common for the two segments.

Consonantal sound{ a r e produced with a radical constriction in the

midsagittal region of the vocal t rac t , This constriction l imits the flow of

a i r in the obstruents and in the closed phase of r-sounds whereas i t is

"shunted", i. e. by-passed in la terals and nasals. Because of the variety

of sounds to be included by the feature a formulation of the acoustical cor-

re lates becomes ra ther complex, the common denominator being a devia-

tion f rom the ideal "vocalic" pattern by a reduction of the second and/or

higher formants.

Vocalic sounds a r e produced with an ora l opening that shal l not exceed

that of the high vowels [i] and [u] and which by definition shall be g rea te r

than that of glides. In addition the vocal cords shal l be positioned to allow

for spontaneous voicing. This requirement rules out unvoiced vowels a s

being nonvocalic. Ora l opening here includes la te ra l opening and in case

of sonorant [ r] -sounds the more open intervals. The acoustic cor re la te

i s a higher F1 and higher overall intensity than in nonvocalic sounds.

Syllabic sounds form a syllabic peak in the sequence of sound events.

Obstruents a r e by definition exchided f r o m the possibility of forming syl-

labic peaks, whereas syllabic nasals and liquids between obstruents a r e

basically characterized by the same cr i ter ion a s that of vowels between

obstruents o r glides. A weighted s u m of second and f i r s t formant inten-

sity relative to that of adjacent phonetic segment would be the s implest

acoustic correlate .

Sonorant sounds. The relative degree of sonority can be based on

exactly the same c r i t e r i a a s for syllabicity except that the relative degree

of sonority i s related to alternative compositions of one and the same seg-

ment whereas syllabicity implies comparisons in the t ime domain. The

production cc r re l a t e of sonority i s the s u m of vocal t r ac t openings includ-

ing ora l , nasal, and la te ra l passages which i s l a rge r than that found in ob-

struents. Thus [-sonorant] = obatruent. An interesting c la im not yet

verified i s that nonsonorarit sounds would not allow "spontaneous voicing"

and that a compensation of glottal adjustment to counteract the impaired

flow would be necessary.

The interdependencies between basic c lass features a r e a s apparent

on the phonetic level a s on the classificatory level. The situation i s even

more complicated by the fact that the continuant -noncontinuant (stop)

feature i s the same a s the consonantal feature, except that the degree of

STL-QPSR 2-3/1969 9.

pr imary s t r i c tu re is total in stops and in the closed interval of affricates

but not total in the [+consonantal][+continuant] fricatives.

I fully agree with Chomsky and Kalle on the need for replacing the

"vocalic" feature by the "syllabic" feature. The syllabicity seems to be

more easily testable than the vocalicity which employs a disputable thres-

hold between liquids and glides which does not focus on the important dif-

ferences. Fur thermore , I suggest a fur ther reduction of the number of

features dealing with vocal t r ac t opening by replacing the features "con-

sonantal" and "continuant" by one single feature (medially) "closed" which

i s identical to the "consonantal" feature but for an extension to separate

stops and affricates f rom fricatives, Before applying this feature we shal l

study hov.1 some of the main phonetic c lasses a r e encoded.

TABLE I-A-1

vowels nasals

syllabic + -. + consonantal - ( + ) +

s ono rant (+) + ( + )

na s a1 + + l a te ra l ( - ) ( - ) continuant

inst. re lease

affri- f r ica- la te ra ls r-sounds glides+h stops cates tives

- t - t - (-) (-) (-1 ( + I + (+) + - (4 (+I (+)

+ (+> + (+I (+I - - -

Features that by definition a r e implied by other features of the s a m e

phonological segment a r e marked with parantheses. Blank spaces repre-

sent other instances of "unmarkedness", i. e, (a) not applicable because

of physiological constraints, (b) i r relevant for the classificatory function,

o r (c ) occurrence in r a r e cases only. In detailed feature-analysis it

would be valuable to have separate notations for these four different a s -

pects of unmarkedness and also fo r the fifth aspect, that related to sequen-

t ia l constraints a s implied by a l l higher levels of analysis. According to

Chomsky and Halle the [+nasal] feature when added to stops could stand

fo r prenasalization, i. e. instance (c) above, whereas +nasal, when added

to vowels o r liquids, is a contextual variant due to adjacent nasal conso-

nants and can thus be omitted f rom the mat r ix (case (b) above). 1

STL-QPSR 2-3/1969

It i s interesting to note that if the feature matr ix is to be used for de-

scription of actual phonetic s ta tes , i t would not be possible to distinguish

between proper nasal consonants and nasalized [ r ) -sounds. This is a

consequence of liquids being opposed to nasal consonants in t e r m s of

[-nasal1 feature alone instead of by a specific complex a s the [+vocalic]

[+consonantal] in the ea r l i e r conventions.

A s imi l a r case of defining a phonetic category by the negative of an

other not directly related category i s the encoding of [r]-sounds a s

[ -lateral] . It is questionable whether an inhibition of the l a t e ra l command

in the production of an [I] automatically resul ts in an [ rl-sound. Addi-

tional adjustment may be necessary. These examples a r e analogous to

the [ -coronal, +anter ior] encoding of labial consonants which I consider

more objectional, All these instances of classification i n t e r m s of com-

binations and selections f rom a finite se t a r e acceptable provided we give

up the demand that each feature shall represent an independent and speci-

fic production category.

A coding t r e e related to Table I-A-1 i s shown in Fig. I-A-1, The syl-

labic feature presides in the top but this is not crucial. The same number

of yes-no branching points would have been needed if we put the sonority

feature on top. Now, coding t r ees a r e deceptive in a way since a l l so r t s

of variations and hierarchies a r e possible because of inherent redundan-

cies, However, the manipulation of coding t r e e s has the pedagogical

mer i t of bringing out these redundancies.

Examples of coding t r e e s for the reduced se t of features I have pro-

posed a r e shown in Figs. I-A-2 and I-A-3. In one the syllabic feature is

placed on the top, in the other i t i s given the lowest place and sonorant

the top place. The economy in t e r m s of branching points i s the same in

a l l the three figures. Figs. I-A-2 and I-A-3 merely have the mer i t of a

sma l l e r number of features. It was actually during the construction of

such t r e e s that I observed the complementary distribution of [-continuant]

and [+consonantal]. I p refer the t r e e of Fig. I-A-2 which s t a r t s out with

the sonorant feature related to vocal t r ac t opening i r respect ive of where

i t occurs. Then, logically follows the feature of c losure in the vocal

t r ac t midsagittal plane, then the manner of re lease of this c losure which

applies to [ -sonorants2 only. The medially closed sonorants a r e then

separated into nasals , l a te ra ls , and r-sounds a s previously discussed

Fig. I-A-l . Coding tree with the basic Chom~ky-Halle features, "syllabic" replacing "vocalic".

SYLLABIC + - -

. - --

CONSONANTAL

b 0 0 0 vow L r nos nos 1 r stop affr fric

0 glides + h

0 0 0

SYLLABIC

SONOR ANT

NASAL

LATERAL

CONTINUANT

INST. RELEASE

-- - -- - .

- * 4 -

SONOR ANT +-- r - 1

Fig. I -A -2 . Coding tree with the features contonantal and continuant replaced by a single feature l'mid-closure'l, The feature "sonorant" is given the top level.

MID-CLOSU E 7 -I+ +I - 7-

(consonantal INST. RELEASE

NASAL -

LATERAL

b -

b b b C, vow glidesnosnos 1 1 r r stop affr f r i c

+h 0 , o 0

- - -

SYLLABIC +a-

SYLLABIC + * - I -

INST. RELEASE

vow nas t r glid,ees nas 1 r stop affr f r i c

SYLLABIC

Fig. I -A-3 . Alternative coding tree with the s a m e features as in Fig. I-A-2 arranged in a different order, the feature "syllabic" in the top. Note the relation to Fig. I -A-1.

STL-QPSR 2-3/1969 11.

and glides a r e opposed to v,>wels a s nonsyllabic. The main acoustic cor -

re late of voiced sonorants i s their higher Fl intensity, whereas the acous-

t ic cor re la tes of "closure" is a reduction of formants higher than F1. The

specification of the nasal and the l a t e ra l cor re la tes a r e not s o simple.

They will not be discussed here.

Some detailed comments

The c lass of h-sounds has always been a problem in feature analysis.

I accep t the classification of glides (semivowels) and h-sounds given by

Chomsky and Halle a s [t sonorant], [ -consonantal]<* and [ -syllabic but I

object to their contrasting of h-sounds to other glides a s [+low]. This

solution i s an apparent mistake since h-sounds display perfect coarticula-

tion with vowels whether [+low] o r [ -low]. The h-sounds, voiced o r un-

voiced, a r e produced with an active glottal readjustment.

The presence of the unvoiced h-sound in the c lass of sonorants weakens

the simple acoustic cor re la te of intensity if this c l a s s since velar f r ica-

tives display s imi lar acoustic patterns but with more noise in the region

above F2. The degree to which the intensity i s associated with the vocalic

formant patterns i s accordingly a necessary aspect to take into account.

This fact a l so cor re la tes with the affinity of sonorants to be found next to

the syllabic nucleus.

Directly related to the classification of h-sounds i s the treatment of

aspiration. The statement of Ghomsky and Halle that a feature of height-

ened subglottal p res su res is a necessary requirement fo r zspiratior, is not

tenable, s ee Fant, Acoustic Theory of Speech Product ion, pp. 277 -27 9.

Instead we need a new feature of "glottal relaxation" yet to be defined that

covers aspiration in general a s well a s the c l a s s of h-sounds.

On the whole, there is a need for fur ther studies of the phonatory

mechanism in various situations before we can single out the various phon-

etic components involved in the various manner of articulations of stop

sounds. The difference between English o r Swedish [P, t , k] and [b , d , g]

involves both aspiration, tenseness and voicing a s phonetic parameters .

In initial s t r e s sed position the aspiration, i. e. glottal relaxation is the ob-

vious cause of the delay of voicing in L p , t , and k]. A higher in t raora l

stop p res su re , when present , appears to reflect a l a r g e r glottal opening

-:+ o r [ -"midclosure) instead of -consonantal].

STL-QPSR 2-3/1969 12.

ra ther than a higher subglottal pressure. At the same t ime there appears

to be a prolongation of the state of articulatory narrowing in [ p, t, and k]

which accounts fo r a high frequency "fricative" noise superimposed on the

f i r s t par t of the aspiration.

There a r e a l so coarticulation differences. The range of F2-locus a t

the instant of re lease is g rea te r fo r the voiced than for the unvoiced stops,

especially s o with [b] compared with [p]. This can be seen i n the data

of Lehiste and Peterson (1961) and I have measured s imi l a r distributions

f o r Swedish (forthcoming article). At the instant of re lease of [b) before

a back vowel the tongue takes a position close to that of the following vow-

el whilst the instant of re lease of the [p] before the same vowel displays

a much higher locus, typical of neutral tongue articulation. After about

40 msec f rom the release of the [p] the formant pattern follows essentially

that observed immediately a f te r the re lease of the [b]. These temporal

relations should be studied closer .

It could be, a s stated by Chomsky and Halle, that the amount of vocal

wall tensening could affect the possibility to maintain a prevoicing (before

the release) but I consider the glottal adjustment to be p r imary and that i t

a l so i s the pr imary cause of the smal l difference found in the t ime lag of

voicing af te r re lease comparing the intervocalic [k, p, t f and [g, b, d]

and associated with this t ime lag a difference in the F1 contour ( ~ 1 cut back).

The me r e fact that there a r e cer tain "tens e-lax" elements associated

with the distinction between the English o r Swedish [k, p, t] versus [b, d , g]

in addition to the obvious glottal adjustments i s not a sufficient basis fo r sel-

ecting the feature "tense" ra ther than the feature "vcice?!'. According to

Chomsky and Halle the cr i ter ion for classifying [p , t, k] a s [+tense] ra ther

than [-voiced? would be that vocal vibrations a r e stopped because of a r t ic -

ulatory interaction ra ther than by glottal relaxation. With this cr i ter ion I

would lay a g rea te r importance in the voicing component than in the tense-

ness component. Fur the r studies a r e needed.

The feature "distributed" which on the articulatory level i s defined a s

a long versus short constriction in the direction of the a i r flow has not been

analyzed very closely a s to i ts acoustic cor re la tes , and these a r e f a r f rom

obvious. Differences in source location, s i ze of front cavity, and the de-

g r e e of coupling to the back cavities may be affected. A high frequency ex-

tension of the noise could be an acoustic cor re la te but I cannot really say

STL-QPSR 2-3/1969 13.

anything definite before I have studied actual samples of spectrograms and

cineradiograms. It appears to me that the main difference between labials

and labiodentals is that of a l e s s effective versus a more effective source

and I a m ra ther hesitant to equate i t with differences in tongue articulations.

In Swedish there a r e both dental and apical alveolar stops, the la t te r

being lexically induced by a previous /r/. The phonological component

would have to work with classifications that differentiate these articulations.

It i s indeed questionable whether the phonetic difference i s that of dis t r ib-

uted-nondistributed.

Swedish vowels

The feature "covered" pertaining to narrowed, tensed pharynx wall and

an elevated larynx is suggested to have some relevance for the difference

between the Swedish vowels L y ] and [a]. There is no evidence to support

this suggestion a s f a r a s I can see.

The Swedish vowel sys tem is of considerable interest inview of the

la rge number of sounds contained. I shall attempt here to construct a phon- I etic feature mat r ix of Swedish long vowel phonemes, [u:], [o:], [a:], C &:I, [e:], [i:], Ly:], [u:], [&:I ,and the p r e - r allophones, [ae:] and [e:] of LC:] and [b:], respectively. I shall f i r s t attempt to use the Chomsky-Halle

tongue-body features back, low, high, and the rounding feature. In addition,

I have defined two new features , which in the consistent art iculatory termin-

ology a r e named "palatal" and "labial". These function a s extreme degrees

of tongue-height and lip-rounding, respectively. It has been long recognized

that 211 Swedish long vowels of extreme low f i r s t formant frequency, [i:],

[Y:], [a:], and [u:] a r e pronounced a s diphthongs towards a homorganic

glide o r fricative. However, what is not s o obvious and often overlooked i s

that the vowel Ly:] is made with a palatal closing ges ture just a s in [i:] but

with added lip- rounding and that the front vowel [w:] is produced with a

labial ges ture towards closure just a s in the back vowel [u:], Fant (1968).

The historical origin of [u:] i s a tongue fronting of [u:) which was replaced

by an Lo:] in a vowel shift, In the Swedish spoken in Finland Cu:] and [uc]

a r e not differentiated and a r e realized with a single sound shape. The

tongue fronting of the "long" [a:] has now progressed to a n articulation

close to that of [i:], [y:], [ e:] , and [+:I, generally a l i t t le more open

than [Y:] and a l i t t le m o r e close than [ b : ] . As f a r a s I can judgd the

STL-QPSR 2-3/1969 14.

element of velarization has been completely lostre. The position of the

m a s s of the tongue in the palatal-velar direction i s not m o r e "velar" than

that of the other front vowels, and the apex i s often slightly raised thus

tending to shift the location of the tongue-palate constriction somewhat an-

t e r i o r of [i:]. However, in the c lass of "short", i. e. lax Swedish vowels, ** the tongue of [u] is lower than that of [b] but more velarized.

When sampling formant data on vowels the distinction between Swedish

[o:] and [u:] and between [b:] and [u:] may be obscured i f [u:] and [u:]

a r e sampled a t the i r onset and not a t their target values where F1 and F2

a r e lower. Similarly, the contrast between [y:] and [a:] is increased if

the sampling i s performed a t the l a t e r par t of the vowel where F2 of [u:]

has been progressively lowered and F3 of Cy:] has been progressively in-

creased. At the place of the vowel target the main constriction is a t the

l ips fo r [u:] and [u:] but a t the tongue-palate region for Cy:] and [i:].

The progressingly decreasing tongue-height in the s e r i e s [u:], [ o:], [ a :]

and in [i:], [ e : ] , [ E:] , [=:I and in [u:], [ b : ] , [ce:] is paralleled by an in-

creasing jaw opening, Lindblom (1967). It has been demonstrated by

Lindblom and Sundberg (1969a and b) that with a minimum jaw opening but

otherwise normal tongue movements the F1 range is considerably reduced.

The jaw opening thus adds not only to the tongue-palate distance but a l so

to the effective lip-opening, everything e lse being equal. The oix vowel

features classify the Swedish long vowels a s follows.

TABLE I-A-2

Swedish long vowels

[=:I and [a:] a r e p r e - r allophones of [E : ] and [ b : ] Binary sys tem

u: o: a: E: E: e: i: y: u: 6: a:

back + + + - - .. - - - - - low - - + + - - - - - - - high + - - .. - + t + + + -

palatal - - - - - + + . - - - round t t - - - - - + - ( + - t

labial t - - - - - - - t - -

* Lindblom and Sundberg (1969a) classified [u) a s "velar" but expressed doubts a s to the phonetic validity.

** The quality of shor t /u/ is generally t ranscr ibed a s [el.

In the consonant sys tem the feature should be used instead of

[+anter ior] [ &coronal] to define the c lass of labial consonants. Labialized

vowels a r e analogous t6 "retroflei", i. e. [+coronal] vowels. Long (tense)

SwediSh vowels a r e accordingly diphthongized if they possess the features

"palatal" o r "labial1'. These a r e the maximally "close" vowels, compare

Lindblom and Sundberg (1969 a).

An alternative matr ix may be se t up with "jaw closure" instead of the

"palatal" feature. The maximum degree of jaw closure i s found in [i:],

[Y:], [u:], and [u:] which would be labeled [+closed]. With this solution

one gains the distinction in actual tongue-plat15 opening comparing [u:]

and [+:I whilst the distinction between [u:] and Cy:] is reduced t c one of

labializatiox~ only. One then has to add the rule that labialization always

determines the diphthongal element when present in the close vowels.

Note the minimal distinction of [-back] separating [w:] f rom [u:] in either

system, A third and ra ther different alternative sys t em was suggested

by Lindblom and Sundberg (1 969 a).

The variety of solutions possible in a sys tem of interrelated physiolog- 1 ical dimensions scaled according to binary principles i s indeed a problem.

One source of variability is that the number of possible combinations gen-

erated f rom a given ensemble i s l a r g e r than the number of sounds to be I encoded. Therefore there may resul t an ambiguity in feature selection.

Two o r m o r e physiological parameters may contribute to one and the s a m e

acoustical and perceptual effect which may constitute a more natural

candidate for the role of feature, a t leas t in the sense of phonetic feature.

Let us s e e what happens if we t r y to simplifjr the inventory of articulatory

parameters by grouping together the features "low", "high", and "palatal"

to one single dimension assigning the value 0 fo r the most "open" degrees

[a:], [z:], and [ce:] and the value 3 for the maximally palatal [i:]. Sim-

i la r ly the feature labial i s added to that of rounding accounting for

TABLE I-A-3 u: o: a : z: : e: i: y: u: : a :

back 1 1 1 0 0 0 0 0 0 0 0

h i g h 2 1 0 0 1 2 3 3 2 2 0

r o u n d 2 1 0 0 0 0 0 1 2 1 1

A matr ix of this so r t i s eas i e r to comprehend than a multidimensional

binary system. There a r e apparently three major c l a s ses within the sys-

tem, the back vowels [u:], [o:], [a:] in which an increase in tongue

STL-QPSR 2-3/1969 16.

height goes with increasing lip rounding (partially jaw dependent). The

unrounded front vowels a r e differentiated by tongue (and jaw) height and

the rounded front vowels a r e a l so differentiated by height and by extra

rounding a s a special feature of La:), cif. Malmberg (1 956) and Fant (1966).

At this stage we might a s k for the acoustic and perceptual cor re la tes

of these articulatory categories. The phonetic color i s mainly dependent

of F I J F2' and F but it should be possible to find an optimal projection 3

of this three-dimensional space on a plane. Pilot experiments now in pro-

g r e s s a t the Dept. of Speech Communication, KTH (Fant, Carlson and

~ r a n s t r b m ) indicate that an F1 versus F ' plot would se rve this purpose. 2 F A is the frequency of the second formant in a two-formant approximation

to the vowel. In mid- and back vowels F' is identical to F2 and in high 2

front vowels close to Fj.

A tentative F versus plot of Swedish long vowels and some short 1 2 vowels of specific identity have been plotted on a me1 scale , Fig. I-A-4.

In this d iagram we find evidence of a fairly even spread in the perceptual

domain. The average distance between any of the sounds and i t s c losest

neighbor i s 180 mels. The articulatory cor re la te of increasing F is in- 1 creasing jaw opening and a shift of tongue place towards a pharyngeal

position. The articulatory cor re la te of the ordinate F'h is a shift of the

tongue away f rom the velum and towards the palate.

It can be seen that back vowels may be separated f rom front vowels

by a line of the slope t45 degrees and rounded vowels f rom unrounded

vowels with a line of -45 degrees slope. Therefore a rotation of coord-

inates a s in Fig. I-A-5 brings out the d i rec t cor re la tes to the main vowel

classes . Back vowels a r e characterized by a distance between the f i r s t

and the second formant lower than 400 mels. All unrounded front vowels

l ie c lose to a line of M ~ + M ; = 2200 me1 and the rounded front vowels have

an absc issa of Ml+M; l e s s than 2100 mel. The quantal s teps in the ordi-

nate comparing [i:, e:, E:, and z:] a r e of the o rde r of 250-300 mels where

a s the quantal s teps in the absc issa a r e of the o rde r of 200-250 mels.

Since we now have condensed the vowel space to a plane we have only two

orthogonal parameters .

I The absc issa (Ivl1+M2) i s twice the center of gravity of the spectrum,

giving equal weight to M1 and M ' and will be identified with the negative 2' of the old feature "flat". - Labialization, velarization, jaw closing, larynx

STL-QPSR 2-3/1969 17.

lowering will a l l lower the center of gravity whilst the ordinate, he re

re fer red to a s the spec t ra l feature "spread" i s a measure of disperson.

Note that it is related to but not identical to any of the old features such

a s [ -compactness], [+diffuseness 1, o r [-gravity]. The spec t ra l spread

is increased with moving the tongue f rom a pharyngeal to a palatal place

of articulation. Five levels a r e indicated by the points [a:], [ax], [ E:],

[e:], [i:]. Note that increasing jaw opening increases in the f i r s t hand

M1 and thus makes the spectrum l e s s flat and l e s s spread. Fig. I-A-5

would motivate a quantization of the long vowels in sca les of "flat" and

"spread" a s follows.

TABLE I-A-4 u: o: a: ce: E: e: i: y: at : Q: parameter

"spread" 0 0 0 1 2 3 4 3 2 2 1 M2-M1

"flat" 5 4 2 0 0 0 0 1 2 1 2 -(MI-2)

These sca les a r e absolute but can of course be reduced according to the

principle of complimetltary distributions. The progressing "flatness"

f rom f rom [a:] over Lo:] to [u:] i s the effect of rounding + velarization

whereas the flatness of [u:] i s pr imari ly a mat te r of sma l l l ip opening.

As previously d i scusse i no velarization appears to be involved in [u:]

but possible an "anteriorization". The possibility of compensatory forms

of articulations in the flatness domain a r e apparent. In the c l a s s of "short"

i. e. lax Swedish vowels, the /u/, phonetically [ o ] i s m o r e "velar" than

the short [ b ] , s e e Fig. I-A-6. These facts support a perceptual ra ther

than an ar t iculatory feature basis.

It has often been suggested that ar t iculatory descriptions of vowels

actually rely on underlying perceptual classifications, Ladefoged (1 967b).

Our data indicate that the Swedish vowels a r e not a rb i t ra r i ly spaced in-

dividuals in the space of physically producible sounds but show a c l e a r

organization in t e r m s of l inear sequences and a tendency of equili3tant

spacings in an orthogonal perceptual space. This ordering appears to be

a subset of a language universal sys tem of maximal contrast . This idea

was also expressed b y L indblo~n an'3 Sunclbcrg (1 963 2). Fur the r work

along these l ines is continuing, E a r l i e r work on me1 sca le mapping of

Swedish vowels was published by Fant (1959).

references on next page

SWEDISH VOWELS

Feature "spread"

to palata l place - Jaw closing

I I I I I I I I I I I I 1 I 1

FEATURE "FLAT" - \ rnels LABIALIZATION \ JAW OPENING VELARIZATION VECTOR JAW CLOSING (INCREASING Ft) LARYNX LOWERING

(FORMANTS MOVING DOWN)

F i g . I - A - 5 . Swedish vowels in a "spread" versus "flat" me1 scale plot bringing out s o m e orthogonal vowel ca tegor ie s (back and front vowels) and a trndency of equidistant me1 spacings .

Fig . I-A-6. X - r a y t rac ings of Swedish vowels. ( ~ r o m Fan t , G. : "The acoust ics of speech", i n P r o c . of the Th i rd International Congress on Acoust ics , Stuttgart 1 959, pp. 188-201, Fig . 9, A m s t e r d a m 1961.)