a cross-linguistic investigation of locus equations as a ... · harvey m. sussman, kathryn a....

A cross-linguistic investigation of locus equations as a phonetic descriptor for place of articulation

Harvey M. Sussman, Kathryn A. Hoemeke, and Farhan S. Ahmed Department of Linguistics and Speech Communication, University of Texas at Austin, Austin, Texas 78712

(Received 8 August 1992; revised 22 February 1993; accepted 28 April 1993)

A previous study [H. $ussman, H. McCaffrey, and $. Matthews, J. Acoust. Soc. Am. 90, 1309-1325 (1991)] of American English CV coarticulation showed a remarkably linear relationship between onset frequencies of F2 transitions, plotted on the y axis, in relation to the F2 midvowel "target" frequencies, plotted on the x axis, for CVC tokens with initial [b d g] and ten medial vowel contexts. Slope and y-intercept values of regression functions fit to these scatterplots ("locus equations") were shown to serve as statistically powerful phonetic descriptors of place of articulation. The present study extends the locus equation metric to three additional languages--Thai, Cairene Arabic, and Urdu--having both two and four place contrasts for syllable-initial voiced stops. A total of 14 speakers (Thai = 6, Arabic= 3, Urdu = 5) produced 1740 CVC tokens that were acoustically analyzed using MacSpeech Lab II. Strong linear regression relationships were found for every stop category across all speakers. Slopes and y intercepts systematically varied as a function of place of articulation. Cross-language comparisons of stop place categories were performed but variability of slope and y intercept means tempered conclusions concerning the existence of CV "phonetic hot spots."

PACS numbers: 43.71.Hw, 43.70. Kv

INTRODUCTION

The most fundamental problem that has shaped both research and theory in speech perception for the past 50 years is the problem of contextual variability in relation to acoustic/phonetic invariance (Goldlinger et al., 1990; Per- kell and Klatt, 1986). The resucrection of locus equations has provided a promising algorithm to apply to the problem of coarticulation_induced variability of stop+vowel tokens that underlies the invariance dilemma (see Sussman et al., 1991). Locus equations are straight-line regression fits to coordinates formed by plotting onsets of F2 transitions in relation to their coarticulated F2 midvowel "tar-

get" frequencies. The frequency of F2 onset (for syllable- initial voiced stops [b d õ]) has been found to be a linear function of F2 as measured in the midvowel nucleus. The

particular linear function relating these two parameters is itself a function of the stop place of articulation. Both the lawfulness of the regression functions and the classificatory power of predictor variables, slope and y intercept, contributed to an interpretation of locus equations as a higher- order phonetic index for classifying and representing stop consonant place of articulation (Sussman et al., 1991 ).

While locus equations characterize acoustic/phonetic regularities of a whole distribution of tokens (the stop place equivalence class) rather than properties of single tokens, the well-behaved locus equation scatterplots are relevant to the invariance problem. By themselves, locus equations cannot provide a complete and sufficient set of invariant perceptual cues for identification of stop place of articulation, but rather "partial invariants" (Nearey and Shammass, 1987) that contribute to phonetic category distinctions. Their major significance at this time lies in their ability to illustrate, in a taxonomic sense, the relational

(rather than absolute) invariance existing within stop place categories, i.e., although the acoustic parameters themselves vary within a stop place category, the relationship between the F2 transition parameters (F2onse t and F2vowel) is invariant and well-defined. The problem of context-induced variability thus becomes more tractable when approached through a locus equation perspective.

The lawfulness of locus equations could be interpreted as an effect that is achieved in spite of context-induced coarticulation, or in contrast, achieved because of context- induced coarticulation. The latter option is in stark contrast to a motor theory position (Liberman and Mattingly, 1985), which has historically viewed coarticulation as pro- foundly obscuring the consequences of underlying invariance. This view has contributed to the impression that the speech signal, because of vowel-conditioned variation on consonants, borders on acoustic chaos.

Different coarticulation strategies underlie and con- strain the stop place acoustic/phonetic distinctions revealed by locus equation scatterplots. Coarticulatory effects, however, are not universally equivalent across world languages. An emerging array of cross-linguistic studies have pointed to language-specific effects for anticipatory coarticulation (e.g., Keating, 1985; Browman and Gold- stein, 1989, 1992; Smith, 1988, 1991; McAllister and Eng- strand, 1991 ). Since languages have need for sufficient dis- criminability or contrast among stop place categories, and since languages show language-specific patterns of CV coarticulation, it is of particular interest to ascertain how locus equations characterize stop place categories cross- linguistically.

A closely related issue pertains to how stop+vowel acoustic/phonetic space is filled across languages. Do languages utilize the entire potential extent of CV acousticZ

1256 J. Acoust. Soc. Am. 94 (3), Pt. 1, Sept. 1993 0001-4966/93/94(3)/1256/13/$6.00 ¸ 1993 Acoustical Society of America 1256

Downloaded 31 Aug 2013 to 128.119.168.112. Redistribution subject to ASA license or copyright; see http://asadl.org/terms

phonetic space as dictated by the bounds of human coarticulatory kinematics? Or, are there "phonetic hot spots" for stop + vowel sequencesmi.e., "clear preferences for certain specific regions in that space" (Krull and Lindblom, 1992, p. 1). The concept of "phonetic hot spots" can be motivated by either of two principlesre(a) quantal theory (Stevens, 1972,1989) or (b) adaptive dispersion theory (Liljencrants and Lindblom, 1972; Lindblom et al., 1983; Lindblom, 1986; Lindblom and Maddieson, 1988). The former holds that languages choose maximally stable acoustic regions of phonetic space to offset articulatory variability, and the latter holds that languages select phonemic inventories based on sufficient separation of auditory and hence perceptual distances. While locus equations cannot uniquely distinguish between these two explanations, the metric does allow for a quantitatively based cross- linguistic assessment of the validity of "phonetic hot spots" for stop+vowel gestures. Quantal theory would logically predict more narrowly focused preference zones for CVs cross-linguistically compared to adaptive dispersion theory. Adaptive dispersion theorists would expect more vari- ablity in cross-language dispersion patterns, but within- language maintenance of sufficient separability amongst stop place categories.

To summarize, the overall objectives of this study were to ( 1 ) investigate locus equations cross-linguistically to ascertain if the metric would yield distinct linear functions across stop place contrasts, (2) represent CV acoustic/ phonetic space in higher-order locus-equation-defined coordinates, slope Xy intercept, and (3) determine if there are preferred areas of this category-level CV acoustic space ("phonetic hot spots") for stop consonant place of articulation in languages differing in the number of voiced place contrasts.

Let us present a brief history of locus equations. Al- though locus equations have been known for almost three decades, their application to theoretical problems in speech research is rather recent. Lindblom (1963) first derived locus equations for a single Swedish speaker producing CVC tokens with initial [b d g] and eight medial vowel contexts. Plotting F2onset'S (at the first glottal pulse) along the ordinate and F2v6w½ • target frequencies along the ab- scissa, Lindblom reported that regression functions were, unexpectedly, quite linear and, moreover, slopes and y intercepts systematically varied as a function of stop place. By using the term "locus" Lindblom did not mean to imply the "virtual locus" notion of Delattre et al. (1955). The latter concept was fixed, invariant, and abstract, while Lindblom's use of the term "locus" was based on a real

value, the onset of F2 at the CV boundary, which systematically varied as a function of the following vowel.

Klatt (1979,1987) briefly experimented with locus equations in the context of deriving speech synthesis algo- rithms for velar [g] in varying vowel environments. Nearey and Shammass (1987), unaware of Lindblom's earlier work, independently replicated the strong linear relationship between g2onset'S plotted against g2steady state'S for [CVd] syllables using 11 medial vowels in ten Canadian English speakers. Regression lines fit to their data points

were distinct for stop place and described as "partly distinctive invariant properties" (p. 17). Krull ( 1988,1989) revived interest in locus equations in Swedish and reported, for syllable-initial stops, varying slope/y-intercept values as a function of stop place. Syllable-final stops failed to reveal the same linear relationships with preceding vowel context as shown by syllable-initial stops with following vowel contexts.

A principled link between locus equation slope and CV coarticulation was first pointed out in Krull (1988). Min- imal coarticulation was reflected by flatter slopes, such as those consistently found for dental/alveolar stops, as g2onset'S remain fairly stable across changes in vowel targets. Maximal coarticulation was reflected by steeper slopes, such as those found for labials and velar stops in back vowel contexts, as g2onset'S linearly changed with each vowel context.

Sussman et al. (1991), had ten male and ten female speakers produce [bvt], [dvt], and [gvt] tokens with ten medial vowels. They reported significant differences in both locus equation slopes and y intercepts as a function of stop place. Moreover, discriminant analyses showed nearly 76% correct classification rates overall when using F2onset/F2vowe 1 frequencies as predictor variables, but 100% correct classification into stop place categories when higher-order and derived slope and y-intercept values were used as predictor variables. In addition to its phonetic validity, Sussman (1989) suggested that locus equations could be conceptualized within a neurobiological frame- work, consistent with and analogous to, recent neuroetho- logical findings explicating how brain-based maps of auditory space are organized in animals specialized for hearingsthe bat and barn owl.

The aim of the present investigation was to extend the locus equation metric to several untested languages that differed in the number of constriction loci for syllable- initial voiced stops--Thai, Cairene Arabic, and Urdu--and to cross-linguistically compare stop categories as mapped in locus-equation-defined CV space.

I. METHOD

A. Subjects and stimulus materials

(1) Thai: Six native speakers of Thai enrolled in an English study program at the University of North Texas, Denton, TX, were recruited for this study. There were five female and one male speaker, with a mean age of 25.7 yr. Thai voiced stops were chosen for study because only bilabial and dental place contrasts exist, compared to Thai voiceless aspirated and unaspirated stops with an additional velar ([kh],[k]) contrast. It was of interest to determine how the presence of additional stop place contrasts across phonation types would possibly influence the real- ization of the two voiced stops, viz., would their distribu- tions behave as if a third place were being held, but not occupied? Bilabial [b] and dental [d] stops followed by one of nine Thai vowels ([i], [u], [ae:], [e:], [•:], [o:], [ae:], [v:], [a:]) were written in Thai script, embedded in a Thai phrase meaning "Say CV again," and produced by each

1257 J. Acoust. Soc. Am., Vol. 94, No. 3, Pt. 1, Sept. 1993 Sussman et aL: Cross-linguistic study of locus equations 1257


subject in randomized order five times each for a total of 90 tokens per subject. All vowels were produced with a mid- tone.

(2) Cairene Arabic: Three native male speakers of Cairene Arabic were recruited for this study. Two were graduate students enrolled at the University of Texas at Austin (mean age=24.3), and one was a professor of Ar- abic (age=60). Because of the diglossia (Ferguson, 1959) found in Arabic speech communities and because Classical/Modern Standard Arabic has only a three-vowel system, an Arabic-script prompt was inappropriate for eliciting all of the vowels found in the unwritten, spoken dialect. Hence, subjects were asked to read a matrix sen- tence "I did not say ...... ," followed by a test token of the form CVC(C) in Roman script. Initial stops were labial [b], dental [d], pharyngealized [df], and velar [g], followed by eight medial vowels and all ending in [t] or [tt]. The eight vowels were [i:], [e:], [ae:], [o:], [u:], [i], [a], and [•]. There were five randomized orderings yielding a total of 160 tokens per speaker.

(3) Urdu: Five native male speakers of Urdu were recruited for this study. All were undergraduate foreign students from Pakistan enrolled at the University of Texas at Austin. Mean age was 24.8 yr. Labial [b], dental [d], retroflex [d], and velar [g] initial stops were produced followed by nine medial vowel contexts ([i:], [a:], [e:], [o:], [u:], [o], [4, [•], [el) and ending with either [t], [d], [b], [p], [g], or [k], which maximized the number of real words in the inventory. All CVC(C)s were written in Arabic or- thography (Nastaliq) and read aloud. Five randomized orderings were used yielding a total of 180 tokens per speaker.

B. Measurements

All measurements and analysis procedures were iden- tical, insofar as possible, to those previously described in Sussman et al. (1991). All subjects were recorded in a soundproof room (IAC) using a high-quality microphone (Electro-voice RE 15) and a Marantz cassette tape re- corder (model PMD 201). The recorded signal was sampled at 10 kHz, digitized, and filtered using an Apple Ma- cintosh IIsi computer with MacAdios II hardware. The MacSpeech Lab II software package (version 1.8, G.W. Instruments, Inc.) was used for all display, editing, play- back, and measurement procedures. Formant measurements were obtained from three sources: direct cursor

readouts from wideband spectrogram displays, linear pre- dictive coding (LPC), and fast Fourier transforms (both narrow- and wideband FFTs). A default parameter of 13 coefficients was used for the autocorrelation function for

LPC analysis. When LPC and FFT spectra were obtained from midvowel target locations the time marker in the center of the analysis window coincided to within a milli- second to the time point sampled by the (mouse- controlled) cursor frequency from the spectrogram display. Vowel target frequencies were within 78.4 Hz as measured by the three techniques (direct spectrogram readouts, LPC, and FFTs).

TABLE I. Slope, y intercept, and R 2 for locus equations for six Thai speakers.

Slope y intercept R 2

[b] [d] [b] [d] [b] [d]

1 0.748 0.241 148 1646 0.97 0.86

2 0.753 0.248 161 1617 0.90 0.75

3 0.682 0.324 230 1349 0.94 0.86

4 0.619 0.319 316 1431 0.84 0.85

5 0.681 0.330 328 1212 0.92 0.89

6 0.726 0.305 186 1296 0.97 0.84

Mean 0.70 0.295 228 1425 0.92 0.84

C. Data sample points

Measurement loci for both F2onse t and F2vowe 1 were exactly the same as those previously followed in Sussman et al. ( 1991 ). Briefly, F2onse t was defined as the frequency of F2 taken at the first glottal pulse after the release burst, and F2vowe 1 was taken as the "target" frequency most closely corresponding to a midvowel nucleus location. A fixed time window for F2vowel measures (such as done by Nearey and Shammass, 1987) was not followed because of the variable vowel durations encountered both within and

across languages. Criteria for F2vowel placement were (a) if the F2 resonance was steady state or diagonally oriented the visual midpoint of the formant was taken (this point ranged from approximately 60 to 120 ms post-transition onset); (b) if the F2 resonance was U-shaped or the in- verse, the "minima/maxima" point was taken as the vowel target frequency. Recent evidence (Van Son and Pols, 1990) has shown consistent results in obtaining vowel "target" frequencies across several formant extraction techniques including a fixed midpoint, a linear formant averaging technique, and a maximal/minimal formant seeking algorithm. Although criteria for formant measurement loci were visually determined, no sample points from what might be considered the vowel off-glide were taken, and all measurement procedures were consistently followed across the three languages.

II. RESULTS

A. Thai

Locus equation slopes, y intercepts, and R 2 values cal- culated from raw scores for each Thai speaker for initial [b] and [d] tokens are shown in Table I. Labial slopes ranged from 0.619 to 0.753 and y intercepts ranged from 148 to 328 Hz. The group mean labial slope was 0.70 (s.d. =0.05) with a y intercept of 228 Hz (s.d. = 78). Ninety-seven percent of the variability of the dependent variable, F2onset, was accounted for by F2vowd. Mean [d] slope was 0.295 (s.d.=0.04) with a y intercept of 1425 Hz (s.d.= 175). Eighty-four percent of the variability in F2onse t was accounted for by F2vowd.

Figure 1 shows group mean locus equation regression functions for the two stop place categories obtained by averaging F2onse t and F2vow• 1 coordinates across all repetitions and subjects. Both scatterplots show highly linear

1258 d. Acoust. Soc. Am., Vol. 94, No. 3, Pt. 1, Sept. 1993 Sussman et al.: Cross-linguistic study of locus equations 1258


y = .717x + 207.363, R-squared: .96

3000 t , , , i , , , , ,

ø,0001 o 500 •ooo •5•o ' 2o•o ' 25•o '

ß / F2 Vowel Target (Hz)

y = .297x + 1409.96, R-squared: .968

0 500 1000 1500 2000 2500 3000 /d/ F2 Vowel Target (Hz)

FIG. 1. Mean locus equations across six Thai speakers for/b/and/d/.

patterns with relatively small standard error of estimates (SE), a measure reflecting the "goodness of fit" of data points around the regression line of only 113 and 70 Hz, respectively, for [b] and [d]. SEs obtained from labial and alveolar regression functions for English stops were 104 and 65 Hz, respectively (Sussman et al., 1991 ).

Figure 2 plots CV locus equation space for all six Thai speakers using the derived and higher-order parameters, slope and y intercept. It can be seen that the two phonemic categories occupy widely separated areas of this space-- relatively high slope and low y intercept for labials and low slope with high y intercepts for dentals.

A doubly multivariate repeated measures analysis with slope and y-intercept measured at each of the two stops [b] and [d] was conducted. Because of a sphericity problem, the omnibus F was adjusted with the Huynh-Feldt (1976) epsilon. This MANOVA revealed a significant effect for stop place [F (2,4) = 64.13, p < 0.001 ]. An effect for stop place was found in univariate analyses of slope [F(1,5) = 135.09, p < 0.00001] and y intercept [F(1,5) = 155.72, p < o.ooool].

-d

.d

.d

.d

.d

Thai (N=6)

©b©b

©b

slope

FIG. 2. Locus-equation-defined CV space for Thai voiced stop place contrasts.

B. Cairene Arabic

Slope, y intercept, and R 2 values for the four stop place categories in Cairene Arabic are shown in Table II for each speaker. The mean slopes (and standard deviations) for labial, dental, dental pharyngeal, and velar stop place were 0.77 (0.076), 0.25 (0.02), 0.21 (0.095), and 0.92 (0.07), respectively. Mean y intercepts (and standard deviations) were 206 Hz (130), 1307 Hz (43), 933 Hz (85), and 220 Hz (122), respectively, for labial, dental, dental pharyngeal, and velar stops. R 2 values were 0.95 for [b], 0.67 for [d], 0.59 for [d•], and 0.92 for [õ]. Across all place contrasts, 78.3% of the F2onse t variance was accounted for by the independent variable, F2vowe]. Interestingly, both coronal variants were characteristically lower in R 2 values compared to the exceptionally high values of R 2 for labial (0.95) and velar (0.92) stops.

Figure 3 shows group mean locus equation functions for the four stop place contrasts in Cairene Arabic. Once again, extreme linearity is seen for all stop place regression functions. SEs were 70.1, 57.5, 41.7, and 129.8 Hz for [b], [d], [d•], and [õ], respectively. The relative separability of the four stop place categories in locus-equation-defined CV

TABLE II. Slope, y intercept, and R 2 of locus equations for Cairene Arabic speakers.

Slope y intercept

[b] [d] [d •] [g] [b] [d]

a 2

[d •] [g] [b] [d] [d •]

1 0.808 0.267 0.153 0.923 162 1278

2 0.812 0.228 0.319 0.984 104 1286

3 0.679 0.240 0.155 0.845 352 1356

Mean 0.77 0.25 0.21 0.92 206 1307

954 220 0.96 0.83 0.43 0.91

839 111 0.94 0.59 0.88 0.94

1005 355 0.95 0.60 0.47 0.92

933 229 0.95 0.67 0.59 0.92



3000 t '

ß .

o 5•o •o•o

y = .762x + 214.928, R-squared: .978 i , i ,

/b/

ß 15'00 ' 20'00 ' 25'00 F2 Vowel (Hz)

3000

y = .246x + 1306.281, R-squared: .85

0 500 1000 1500 2000 F2 Vowel(Hz)

/d/

2500 3000

y = .199x + 944.508, R-squared: .865

/d ?/

0 500 1000 1500 2000 2500 3000 F2 Vowel (Hz)

y = .947x + 185.627, R-squared: .952

3000 t , , , , , , , , ,

15001 ••,••• 0 500 1000 1500

F2 Vowel (Hz) 20'00 ' 25•0 ' 3000

FIG. 3. Mean locus equations across three Cairene Arabic speakers for/b/,/d/,/d•/, and/g/.

space for each Cairene Arabic speaker is shown in Fig. 4. Slope and y-intercept coordinates do a very nice job in segregating the stop classes--[d] occupying a relatively low slope/high y-intercept zone, [d •] an even lower slope region combined with lower y intercept, [b] a moderately high slope zone with low y intercepts, and [g] a zone characterized by the highest slopes with low y intercepts. The dif-

120

Y .

i

e

r 8O c

e

z

)

4O

.d

.Dph

.Dph

.Dph

Cairen e Arabic

(N=3)

eb -•g

•.g

eb

eb '•g

slope

FIG. 4. Locus-equation-defined CV space for Cairene Arabic voiced stop place contrasts.

ference in y-intercept values for [d] vs [d •] illustrates the strong influence of the consonantal secondary articulatiøn--pharyngealizatiøn---øn g2onset'S.

A doubly multivariate repeated measures analysis with slope and y intercept measured for each of the four stop place contrasts was conducted. Once again, the omnibus F was adjusted with the Huynh-Feldt epsilon because of a sphericity problem. Despite the adjustment and the small number of subjects, the robustness of the effect was still much in evidence as the MANOVA revealed a significant effect for stop place [F ( 6,10) = 130.16, p < 0.00001 ]. Bon- ferroni tests, which adjust the alpha level of t-tests, were used to analyze differences between stops for slope and y-intercept values. The results of these comparisons can be seen in Table III. All comparisons were significant (at either p=0.01 or 0.05) except slope differences between the two coronals, [d] vs [d •] and y-intercept differences between [b] and [9].

TABLE III. Results of Bonferroni tests on differences between Cairene

Arabic stops for slope and y-intercept means. *= t is significant at 0.05 level; **= t is significant at 0.01 level.

Slope y intercept

[b] [d] [d •] [g] [b] [d] [d •] [g]

Mean 0.77 0.25 0.20 0.92 206 1307 933 229

[b] -- * * ** -- ** ** ns [d] -- -- ns ** -- -- * **

[g] ........

1260 J. Acoust. Soc. Am., Vol. 94, No. 3, Pt. 1, Sept. 1993 Sussman et al.: Cross-linguistic study of locus equations 1260


TABLE IV. Slope, y intercept, and R 2 for locus equations for Urdu speakers.

Slope

[b] [d] [d] [g] [b]

1 0.746 0.427 0.349 1.01 214

2 0.930 0.574 0.466 1.02 29

3 0.859 0.475 0.434 0.953 166

4 0.805 0.492 0.438 0.875 191

5 0.715 0.507 0.522 0.972 259

Mean 0.81 0.50 0.44 0.97 172

y intercept

948

722

899

939

777

857

a 2

[•] [g] [b] [d] [•] [g]

1220 127 0.949 0.859 0.848 0.929

1066 91 0.980 0.933 0.866 0.935

1042 255 0.970 0.882 0.886 0.915

1145 392 0.949 0.823 0.876 0.916

879 194 0.980 0.962 0.932 0.943

1070 212 0.97 0.89 0.88 0.93

C. Urdu

Table IV shows slopes, y intercepts, and R 2 values from locus equations obtained across the four stop place contrasts using raw data values from each Urdu speaker. The mean slopes (standard deviation) for [b], [d], [d], and [õ] stop place categories were 0.81 (0.087), 0.50 (0.053), 0.44 (0.063), and 0.97 (0.058), respectively. The y-intercept means (standard deviation) were 172 Hz (87), 857Hz (101), 1070 Hz (128), and 212 Hz (119), respectively. Here, R 2 values were consistently high averaging 92% across the four stop place contrasts and ranging from a high of 97% for labials to a low of 88% for retroflex [d]. Figure 5 presents group mean locus equations obtained across all Urdu speakers for each stop place of articulation. Once again highly linear regression functions characterized the relationship between F2onset'S and coarticulated F2vowe 1 targets. Standard error of estimates were once again extremely smallw52.3, 56.6, 37.3, and 139.1 Hz for labial, dental, retroflex, and velar stop places, respectively.

Figure 6 shows slope/y-intercept coordinates for the five Urdu speakers. All four stop place contrasts are clearly separate. Retroflex stops are differentiated from dentals by higher g2onset'S as reflected by higher y intercepts. The only overlap was for subject #5's retroflex stop falling amongst the dental cluster. Labial and velar stops are kept apart by slope differences, [õ] being steeper than [b].

A doubly multivariate repeated measures analysis with slope and y intercept measured for each of the four stop place contrasts was conducted. The sphericity problem was once again adjusted with the Huynh-Feldt epsilon. The MANOVA revealed a significant effect for stop [F(6,22) = 73.14, p < 0.00001]. Bonferroni tests were used to analyze differences between stops for slope and y-intercept values and the results of these comparisons can be seen in Table ¾. All slope comparisons were significant except for dental [d] versus retroflex [d] and all y-intercept comparisons were significantly different except for labial [b] versus

y = .829x + 149.907, R-squared: .988

3000 t ' , 2500 t

•2ooo] •15øø l t

0 500 ,000 ,5•0 20b F2 Vowel (Hz)

/b/

25;::)0 3000

y = .507x + 837.609, R-squared: .917

3000] 25001

•20001 • '1500

1000• /all

0 500 1000 1500 2000 25;::)0 F2 Vowel (Hz)

3000

3oooj . , . , . , .

•'1500• 0 500 1000 1500

F2 Vowel (Hz)

y -- .436x + 1070.868, R-squared: .966

/drilx I

,, .

20•0 ' 25•0 3000

y = .971x + 204.058, R-squared: .942

3oooj

/•/

501• ' . ß . s•o l o;•o ' •5bo ' 20bo

F2 Vowel (Hz) 3000

FIG. 5. Mean locus equations across five Urdu speakers for/b/,/d/,/d/, and/õ/.

1261 J. Acoust. Soc. Am., Vol. 94, No. 3, Pt. 1, Sept. 1993 Sussman et al.' Cross-linguistic study of locus equations 1261


1200

900

Y .

i

e

r 600 c

e

300

*D

*D

*D *D

-da d .d

*D

.d

.d

• o.•s o:s

slope

Urdu (N=5)

*g

©b •.g ©b

©b •g eb

•g

•g

©b

o.• ;

FIG. 6. Locus-equation-defined CV space for Urdu voiced stop place contrasts.

velar [g]. This was the same pattern seen in Cairene Arabic stop place comparisons for slope and y intercept.

2000

1500

lOOO

500

Arabic

+ /d/

•, /d/phyn

øo ' 5;0 ' 10'00 ' 15•00 ' 2;00 ' 25•00 F2 Vowel (Hz)

2000

1500

1000 -- /d/

C, /all rtflx 500

2000

F2 Vowel (Hz)

D. Coronal comparisons: Arabic and Urdu

Both sets of coronal stops, Arabic [d] vs [d •] and Urdu [d] vs [d], were the only stop contrasts that failed to reveal significant differences in locus equation slopes. To better. view these stop place contrasts representative locus equation functions for the two coronal stops for Arabic (taken from S# 1) and Urdu (taken from S# 1) were plotted using vowel means averaged across repetitions. These locus equation functions are shown in Fig. 7. It can be seen for Arabic that no overlap exists between the two categories as y-intercepts are distinct enough to keep dental [d] well above pharyngeal [d •] across all eight vowels. Urdu coronal functions are somewhat closer, but relatively little overlap is seen with only the vowel [i] following dental [d] over- lapping with the retroflex function. The remaining eight Urdu CV coordinates are apart with retroflex values consistently higher than dental values. Thus, while coarticulatory effects, as reflected by slope values, are relatively sim-

TABLE V. Results of Bonferroni tests on differences between Urdu stops for slope and y-intercept means. *--t is significant at 0.05 level; **= t is significant at 0.01 level.

Slope ),intercept

[b] [d] [_d] [g] [b] [d] [d_] [g]

Mean 0.81 0.50 0.44 0.97 172 857 1070 212

[b] -- ** ** * -- ** ** ns [d] -- -- ns ** -- -- * ** [d] -- -- -- ** -- -- -- ** [g] ........

FIG. 7. Locus equations for coronal stop place contrasts in a representative speaker of Cairene Arabic (top) and a representative speaker of Urdu (bottom).

ilar across these coronal contrasts, differing y intercepts due to differential F2onse t values are capable of phonetically indexing the stop place contrasts.

E. Cross-linguistic comparisons

One of the main purposes of this research was to use the quantitative indices provided by locus equation slope and y-intercept values to determine if "phonetic hot spots" exist for similar stop place of articulation categories across varied languages. Table VI presents slope and y-intercept means for three stop place categories, labial, dental/ alveolar, and velar across five languages [American English male means taken from Sussman et al., 1991, N= 10; Swedish [b] and [d] means taken from Krull, 1989, N= 5; Swedish [9] values taken from Lindblom, 1963, N= 1]. Figure 8 shows, for each language, the distribution of stop place categories within the 2-D space defined by locus equation slope and y intercept. Within each language contrastive stop place categories can be seen to occupy nonoverlapping areas of this attribute space. With only two voiced stop place categories, Thai speakers appear to maximally separate labial and alveolar stops. While Swedish and English show approximately the same overall separation of stop categories, velar and labial categories are re- versed. English [9] appears to be a phonetic anomaly with a lower slope and higher y intercept compared to other velar stops. Cairene Arabic and Urdu stop contrasts show



TABLE VI. Slope and y-intercept values for stop place categories across five languages.

Stop place Language Slope y intercept (Hz)

[b] Thai 0.70 228 [b] English 0.87 106 [b] Swedish 0.63 487 [b] Arabic 0.77 206 [b] Urdu 0.81 172

Mean 0.756 240

[d] Thai 0.30 1425 [d] English 0.43 1073 [d] Swedish 0.32 1096 [d] Arabic 0.25 1307 [d] Urdu 0.50 857

Mean 0.36 1152

[g] English 0.66 807 [g] Swedish 0.95 360 [g] Arabic 0.92 229 [g] Urdu 0.97 212 Mean 0.875 402

a similar pattern of occupying CV space, with within- language coronal classes differing from each other primarily in y-intercept values, not slope, while labials and velars differ in slope values, not y intercepts. Both slope and y

intercept serve to distinguish the coronals from noncoro- nals.

Figure 9 presents cross-language comparisons of the stop categories in locus-equation-defined coordinates. The range of coordinates for each stop place category can be seen to be quite extensive. Labials across Thai, Arabic, Urdu, and English cluster fairly tightly in the lower right quadrant of this space while Swedish labials are more cen- trally located with slightly lower slopes and considerably higher y intercepts. Dental?alveolar stops are generally more widely distributed compared to labials with slopes ranging from a low of 0.25 (Arabic) to a high of 0.50 (Urdu), and y intercepts ranging from a low of 857 Hz (Urdu) to a high of 1425 Hz (Thai). Swedish [d] and English [d] are similar in y intercepts but different in slope. Velars present an unusual pattern with English [9] as the sole outlier with closely spaced velar coordinates among Swedish, Arabic, and Urdu. Mean slopes only differed by an average of 0.03 and y intercepts differed by an average of 97 Hz for these three languages.

1. Interpretation of locus equation defined CV space

Since the plots of locus equation coordinates are second-order representations of the stop place categories, an interpretation of the meaning of differential loci in this space is called for. An attempt to provide such an interpretation is given in Fig. 10. As an example of how locus

1200'

800.

400,

Thai Stops

o o.•5 0'5

+ /hi

o.•s

1200-

800-

400-

Swedish

Stops

6 o •s

a /d/

+ /hi

alope

1200'

800.

400.

ENGLISH STOPS

a /d/

/h/

slope

o /d/ dental

1200'

Y .

I n /d/phaxyngrnd n

!

ß 800,

ß

CaJxene Arabic Stops

4oo.

ß /h/ +

• ø.•s o•s o.•s elope

1200.

800,

400

Urdu Stops

o o.•5

/d/retroflex

o /d/dental

,

0'5 slooe

FIG. 8. Locus-equation-defined CV space for voiced stop place contrasts in Thai, Swedish, American English, Cairene Arabic, and Urdu.

1263 J. Acoust. Soc. Am., Vol. 94, No. 3, Pt. 1, Sept. 1993 Sussrnan et al.: Cross-linguistic study of locus equations 1263


Labial

+ Ibl Swcd

[bl Thai + {b] Arabic

+ lb] U•lu

{b] Eng

o.hs o:s o.•s ;

[d] A•abic

{d] Thai

[d] Swed

o Id] Eng

o !dl U•lu

Dental A!veolar

o.•s o:s o.•s

1200-

Y .

i 800' ,i

4O

Velar

* [g] Eng

Igl Swed

{g] Arab .

[.g] Urdu

J o.•s o's o.•s ; $10po

FIG. 9. Locus-equation-defined cross-linguistic CV space for labial, dental/alveolar, and velar stop place contrasts.

equation data can be systematically related to underlying coarticulatory behavior a pair of stop place contrasts, En- glish [d] and Cairene Arabic [d], are depicted in the center plot. Arabic [d] is characterized by a lower slope (0.25) and higher y intercept (1306 Hz) than English [d] [0.42/ 1075 Hz: from male mean (N= 10), Sussman et al., 1991 )]. To put this contrast in pe,rspective, first recall that locus equation slope relates to the overall coarticulatory behavior being employed by a speaker across an entire phonetic category (Krull, 1988). The top four plots schemat- ically illustrate this point. The hypothetical bounds of such a relationship are shown for a stop place class with ( 1 ) an invariant onset locus and hence no coarticulation between

the initial stop and following vowel (upper left), and (2)

an infinitely variable consonantal onset locus and hence a maximal coarticulation strategy (upper right). A locus equation slope of 0 is indicative of an onset that is fixed across the vowel inventory and hence no anticipatory ad- justments to the following vowel are made. This spectrographic pattern would correspond to the Delattre et al. (1955) "virtual locus" notion for [d]. In contrast, a locus equation slope of 1.0 is obtained when each onset frequency corresponds to the F2vowe 1 steady state across the entire vowel inventory. The stop constriction is thus maximally affected by the following vowel. The spectrographic and relative locus equation correspondences shown in the lower quadrants illustrate how locus-equation-defined CV space relates to differing coarticulation strategies for pro-



Hz

IBverieBt L•m•= Ne C#rtic

Time

F 2 0

t

(Hi)

slope = 0

F2 revel (

Hz

i

lie Locu-I•x C#rticuletieB

Time

LOCUS EQUATION CV SPACE

F 2

0

ß

i

t

(•)

F2 Vevel (Hz)

INTER-

CEPT

/d/ Arabic ß

ß /d/ Engl I sh

Hz

Arebit /d/ ,,o leoo c#rtic

Time

(Hi) slope = .25

u-lnt = 1:506 Hz

F2 Yevel (Kz)

SLOPE

Hz

I.O

[•11iob Idl ,,0 more ceertic

Time

(•) y-int = 1075 Hz

F2 Yevel (Hz)

FIG. 10. Schematic interpretation of locus equation slope in spectrographic and coarticulatory parameters.

duction of Arabic versus English [d]. Arabic, with a lower slope, would be somewhat closer to the invariant locus end of the continuum than the steeper and hence more coarticulated English [d]. One would expect greater extents of F2 transitions for similar stop place constrictions to target vowels for Arabic [d] versus English [d].

2. Cross-linguistic statistical analyses

To quantify cross-linguistic comparisons a series of MANOVAs were run. Because all five languages were not crossed with [õ], two separate MANOVAs were conducted: the first compared only labial and dental/alveolar stops across five languages (English, Arabic, Urdu, Swed- ish, and Thai) and the second compared labial, dental/ alveolar, and velar stops across English, Arabic, and Urdu (Swedish [õ] was not included as data from only one subject was available). In all analyses sphericity problems were adjusted with the Huynh-Feldt epsilon and Bonfer- roni tests were used to assess all subsequent univariate pairwise comparisons.

A doubly multivariate repeated measures analysis with two sets of within-subjects variables (slope and y-intercept measured at each of two stops: [b] and [d]), and one between-subjects factor (language: English, Arabic, Urdu, Swedish, and Thai) was conducted. This MANOVA revealed a significant effect for language [F(8,66)= 13.11, p < 0.00001], for stop, [F(2,33) =273.79, p <0.00001], and for the interaction between language and stop [F(8,66) = 8.33, p < 0.00001]. Bonferroni tests were used to analyze

the effects of language and stop on slope and y-intercept values. Only [b] slope means between Swedish versus Urdu (p=0.05), Swedish versus English (p=0.01), and Thai versus English (p=0.01) were statistically significant, while five paired comparisons for slope reached statistical significance for [d]' Arabic versus English (p=0.01), Ar- abic versus Urdu (p=0.01), Thai versus English (p =0.01), Thai versus Urdu (p=0.01), and Swedish versus Urdu (p=0.05). The y intercept means for paired comparisons for [b] were statistically significant for English versus Thai (p =0.05), English versus Swedish (p--0.01), Urdu versus Swedish (p=0.05), Arabic versus Swedish (p =0.05), and Thai versus Swedish (p=0.05). The y-intercept means for paired comparisons for [d] were statistically significant for Urdu versus English (p=0.05), Urdu versus Arabic (p=0.01), Urdu versus Thai (p =0.01), and Swedish versus Thai (p=0.05).

The second cross-linguistic analysis was a doubly multivariate repeated measures analysis with two sets of within-subjects variables (slope and y intercept measured at each of three stops: [b], [d], and [õ]), and one between- subjects factor (language: English, Arabic, and Urdu). This MANOVA revealed a significant effect for language [F(4,48)=9.18, p<0.00001], for stop [F(4,98)=151.54, p < 0.00001], and for the interaction between language and stop [F(8,98) =9.95, p < 0.00001]. Subsequent univariate comparisons revealed no differences for labial slopes between English, Arabic, and Urdu; [d] slope differences were found for English versus Arabic (p=0.01) and Ara-



bic versus Urdu (p=0.01); [õ] slope differences were revealed only for English versus Urdu (p--0.01). There were no significant cross-linguistic differences in y-intercepts for [b]. Significant differences were seen for [d] y intercepts for Urdu versus English (p--0.05) and Urdu versus Arabic (p=0.01), while significant differences in [õ] y intercepts were found for Urdu versus English (p--0.01) and Arabic versus English (p=0.05).

3. Statistical assessment of "phonetic hot spots"

The univariate tests from the cross-linguistic MANO- VAs provided a quantitative means of initially assessing the validity of the concept of "phonetic hot spots" for stop place categories across languages. Lack of statistical differences along each linear distance space, as defined by the two dimensions, slope and y intercept, were used to determine similarity among regression coordinates. Thus, for a given stop place category, coordinates were regarded as being in the same locus-equation-defined CV space if language comparisons (e.g., lb] Thai vs lb] Arabic) were not significantly different in both slope and y intercept. The first MANOVA compared slope and y-intercept for [b] and [d] across Swedish, Thai, Cairene Arabic, English, and Urdu; the second MANOVA measured slope and y-intercept for [b d õ] across English, Cairene Arabic, and Urdu. For labials, two sets of three languages met the criterion of comprising a "hot spot" in stop+ vowel space-- Cairene Arabic•-•Thai•-•Urdu comprised one set and Cairene Arabic•-•Urdu•-•English the other (see Fig. 9). Labials were the only place contrast that had more than two languages meet the criterion for belonging to a "hot spot." Swedish labials remained statistically apart from all other labials. For alveolar/dental stops there was no clear clustering of languages within a statistically defined "phonetic hot spot." In fact, from the total of 20 possible two- way comparisons for slope+y intercept only three language pairs were found to be nonsignificant in both dimensions--Cairene Arabic•-•Thai, Cairene Arabic •-•Swedish, and Swedish•-•English. For the velar stop place category the only language pair having no significant difference in both slope and y intercept was Arabic•-•Urdu. English [õ] is statistically distant from all other velar stop categories.

III. DISCUSSION

This study attempted to provide initial answers to several basic but unexplored questions: (1) Would a cross- linguistic analysis of locus equations reveal the same linear relationship between F2onse t frequencies and coarticulated midvowel target frequencies as previously documented for English? (2) Would slope and y-intercept values reliably vary as a function of stop place of articulation? (3) Would similar stop place categories across languages occupy similar areas of CV locus equation space (i.e., do "phonetic hot spots" exist) ?

The locus equations obtained for Thai, Urdu, and Cairene Arabic were, without exception, highly linear and characterized by extremely tight clustering of data points

about the regression line. The overall mean R 2 value from all locus equation functions obtained across all speakers, stops, and languages analyzed was 0.86. This is a remarkably robust statistic indicative of the high inherent power of F2vowe 1 to predict, and thus account for, F2onse t .

Slopes and y intercepts systematically varied as a function of stop place, with the great majority of within language stop place comparisons reaching statistical significance in post hoc tests. Only coronal slopes, [d] vs [d •] in Cairene Arabic and [d] vs [d] slopes in Urdu, failed to show significant statistical differences, but these contrasts were clearly differentiated by locus equation y intercepts (as seen in Fig. 7). Thus, despite language specific tuning differences in coarticulation strategies, locus equations are robust enough to reliably serve as phonetic descriptors for contrastive stop place categories. In addition, the fundamental similarity of locus equations for stop categories across languages provides a simple graphical means for establishing family resemblances of allophonic or vowel- conditioned variants of a particular place of articulation.

A. Phonetic hot spots

One goal of this study was to ascertain if there were preferred regions of CV space for stop+ vowel sequences. One advantage of the locus equation metric, as compared to more traditional single token analyses, is that it allows quantitative-based comparisons to be made at the level of the phonological category across a language's vowel space. Locus-equation-defined coordinates map a second-order attribute space that reflects the overall coarticulatory behavior underlying the allophonic variants of a stop place class. Figure 9 showed, across five languages, relative slope/y-intercept coordinates for [b d õ]. The results of the subsequent univariate comparisons for slope and y-intercept means showed the existence of a moderately viable labial "hot spot" among Thai, Arabic, and Urdu and among the latter two with English. Labial stops were the only stop place category capable of being partially described as forming a "cluster" in the lower right quadrant of locus equation CV space. This is primarily attributable to the homogeneous constriction locus for labials compared to the degrees of freedom possible in forming coronal and dorsal tongue constriction loci. Swedish labials, however, exhibited less steep regression functions and hence higher y intercepts than the other four languages. A locus equation interpretation of this finding would be a relatively reduced coarticulation factor for Swedish labials. Further

analysis of Swedish speakers would be required to corrob- orate this claim.

No clear clustering of coordinates emerged from the dental/alveolar comparisons as the five languages were widely dispersed throughout the central area of locus equation C¾ space. Only three pairs of [d] stop categories met the selection criteria of having nonsignificant differences in both slope and y-intercept means as the majority of language comparisons were significantly different in at least one dimension. Clearly no "phonetic hot spot" can be said to exist for [d] in the five languages examined.

Only three languages were statistically compared for

1266 J. Acoust. Soc. Am., Vol. 94, No. 3, Pt. 1, Sept. 1993 Sussman et aL' Cross-linguistic study of locus equations 1266


velar stops with Cairene Arabic and Urdu meeting the criterion of lying Within the same locus equation zone. American English [õ] was an outlier with a relatively lower slope and higher y intercept than [õ] in the other languages. Nearey and Shammass (1987) reported velar slope and y intercept of 0.99 and 215 Hz, respectively for a group of Canadian English speakers. These values are close to the values obtained here for Cairene Arabic and Urdu. Suss-

man et al. ( 1991 ) showed two distinct locus equations for [õ] preceding front vowels and [õ] preceding back vowels. The latter had a mean slope of 0.87 and a y-intercept of 559 Hz (N=20 speakers). These slope/y-intercept values would place American English [õ]velar more in line with [õ]'s in the other languages shown in Fig. 9. Apparently, American English speakers have a stronger tendency to front velars before front vowels and this allophonic alter- ation greatly affects the overall (and Procrustean) regression line for [õ], ultimately yielding a lower slope and considerably higher y intercept compared to the other languages. Comparative language data using electro- palatography would be useful in further documenting this effect.

From the results of our preliminary analyses and based on only the few languages studied, one would be hard pressed to characterize labial, coronal, and dorsal stop place categories as narrowly focused clusters or "phonetic hot spots." Our initial judgement on this question is similar to that reached by Krull and Lindblom (1992) in their study attempting to map phonetic hot spots for vowels. In their study, Flvowe 1 and F2vowe 1 formant variability across 29 languages precluded a clear designation of formant areas as favored vowel "hot spots." Rather, it was concluded that "languages tend to use "i-like, .... u-like, .... e-like," etc., sounds, but the exact phonetic categories are subject to considerable language-specific tuning" (p. 4). Along the same lines, the data of this study point to the conclusion that languages tend to use "b-like, .... d-like," and "g-like" initial stops coarticulatorily tuned by the specific vowel inventory that follows. Phonetically favored "lukewarm" spots might be more appropriate for the coarticulated CV utterances seen across the languages analyzed in this study.

The relatively steep regression functions for both labials and velars indicate two stop place classes where the following vowel greatly influences articulation of the preceding stop closure. Flatter slopes for coronal stops in gen- eral indicate a region of relatively minimal coarticulatory effects and hence more acoustic stability. It is logical for the most acoustically stable place of articulation, coronals, to be universally favored among the world's languages. Furthermore, the acoustic stability of coronal stops allows for finer differentiation of contrastive place loci, as reflected in the examples provided by retroflex [d] in Urdu and the secondary articulation of pharyngealized [d •] in Arabic. Interestingly, the traditional view of a secondary articulation such as pharyngealization, as a superimposed vowel- like articulation, takes on another view when differences in the consonantal F2onse t frequencies (reflected by y intercepts) are seen as accounting for the major acoustic-related distinctions across the coronal stop contrast.

B. Acoustic phonetic maps

Neural systems coding auditory information impor- tant to a species appear to be laid out in terms of parallel and hierarchically arranged 2-D matrices. These hard- wired representations code independently processed but related sources of information [e.g., pulse-echo comparisons in the bat for biosonar navigation (Suga et al., 1983) and interaural time/intensity comparisons in the barn owl for sound localization (Wagner et al., (1987) ].

A parallel and hierarchical network of 2-D matrices representing "information-bearing" parameters from the polymorphous cue set that is thought to underlie stop + vowel tokens, would primarily include stop release burst +transition-based cues. Systematic regularities have al- ready been documented in studies of the properties of the spectral release burst (e.g., Blumstein and Stevens, 1979, 1980; Lahiri et al., 1984; Forrest et al., 1988). Regularities in F2 transitions have been harder to document. A contri-

bution of the locus equation data is in showing that systematic regularities in the signal may not emerge until the entire phonetic category, complete in its allophonic varia- tions, is represented and contrasted. If relationally invariant properties only emerge when higher levels of phonological abstraction are captured, then perhaps on-line processing models would do well to somehow incorporate the lawful regularities that we now know to exist in the coarticulatory behavior of stop+ vowel utterances.

ACKNOWLEDGMENTS

Portions of this research were conducted with the sup- port of Grant No. BNS-8919221 from the National Science Foundation to the first author. The Urdu data were based

on a Master of Arts thesis of Farhan S. Ahmed, submitted to the University of Texas at Austin, December 1992. The assistance of Dr. Peter Abboud and Ayman E1-Desouky in deriving the phonetic script for the Arabic citations is gratefully acknowledged. The assistance of Dr. Hoemeke in allowing us access to the Thai speakers is also greatly appreciated. Thanks to Bjorn Lindblom and John King- ston for their comments on earlier versions of the manu-

script and to Ann Repp for assisting in the statistical analyses.

Blumstein, S. E., and Stevens, K. N. (1979). "Acoustic invariance in speech production: Evidence from measurements of the spectral char- acteristics of stop consonants," J. Acoust. Soc. Am. 66, 1001-1017.

Blumstein, S. E., and Stevens, K. N. (1980). "Perceptual invariance and onset spectra for stop consonants in different vowel environments," J. Acoust. Soc. Am. 67, 648-662.

Browman, C. P., and Goldstein, L. (1989). "Articulatory gestures as phonological units," Phonology 6, 201-251.

Browman, C. P., and Goldstein, L. (1992). "Articulatory phonology: An overview," Phonetica, special issue on articulatory phonology.

Delattre, P. C., Liberman, A.M., and Cooper, F. S. (1955). "Acoustic loci and transitional cues for consonants," J. Acoust. Soc. Am. 27, 769-773.

Ferguson, C. (1959). "Diglossia," Word 15, 325-340. Forrest, K., Weismer, G., Milenkovic, P., and Dougall, R. N. (1988).

"Statistical analysis of word-initial voiceless obstruents: Preliminary data," J. Acoust. Soc. Am. 84, 115-123.



Goldlinger, S. D., Pisoni, D. B., and Luce, P. A. (1990), "Speech perception and spoken word recognition: Research and Theory," Research on Speech Perception, Progress Report No. 16, 4-63.

Huynh, H., and Feldt, L. S. (1976). "Estimation of the Box correction for degrees of freedom from sample data in randomized block and split-plot designs," J. Educ. Stat. 1, 69-82.

Keating, P. A. (1985). "CV phonology, experimental phonetics, and coarticulation," UCLA Working Papers in Phonetics 62, 1-13.

Klatt, D. H. (1979). "Synthesis by rule of consonant-vowel syllables," Speech Communication Group Working Papers 3, MIT, Cambridge, MA, 93-104.

Klatt, D. H. (1987). "Review of text-to-speech conversion for English," J. Acoust. Soc. Am. 82, 737-793.

Krull, D. (1988). "Acoustic properties as predictors of perceptual re- sponses: A study of Swedish voiced stops," Phonetic Experimental Re- search at the Institute of Linguistics, University of Stockholm, PER- ILUS VII, 66-70.

Krull, D. (1989). "Second formant locus patterns and consonant-vowel coarticulation in spontaneous speech," Phonetic Experimental Research at the Institute of Linguistics, University of Stockholm, PERILUS X, 87-108.

Krull, D., and Lindblom, B. (1992). "Comparing vowel formant data cross-linguistically," paper presented at FONETIK 92, Chalmers Insti- tute of Technology, Goltenburg, Sweden, May 20-22.

Lahiri, A., Gewirth, L., and Blumstein, S. E. (1984). "A reconsideration of acoustic invariance for place of articulation in diffuse stop consonants: Evidence from a cross-language study," J. Acoust. Soc. Am. 76, 391-404.

Liberman, A.M., and Mattingly, I. G. (1985). "The motor theory of speech perception revised," Cognition 21, 1-36.

Liljencrants, J., and Lindblom, B. (1972). "Numerical simulation of vowel quality systems: the role of perceptual contrasts," Language 48, 839-862.

Lindblom, B. (1963). "On vowel reduction," Rep. No. 29, The Royal Institute of Technology, Speech Transmission Laboratory, Stockholm, Sweden.

Lindblom, B. (1986). "Phonetic universals in vowel systems," in Exper- imental Phonology, edited by J. J. Ohala and J. J. Jaeger (Academic, Orlando, FL), pp. 13-44.

Lindblom, B., MacNeilage, P. F., and Studdert-Kennedy, M. (1983). "Self-organizing processes and the explanation of phonological univer-

sals," in Explanations of Linguistic Universals, edited by B. Butter- worth, B. Comrie, and O. Dahl (Mouton, The Hague), pp. 181-203.

Lindblom, B., and Maddieson, I. (1988). "Phonetic universals in consonant systems," in Language, Speech, and Mind, edited by L. M. Hyman and C. N. Li (Routledge, London), pp. 62-78.

McAllister, R., and Engstrand, O. (1991). "Some cross language aspects of coarticulation," Proceedings of the XIIth International Congress of Phonetic Sciences, Aug. 19-24, Aix-en-Provence, France, 5, 18-21.

Nearey, T. M., and Shammass, S. E. (1987). "Formant transitions as partly distinctive invariant properties in the identification of voiced stops," Can. Acoust. 15, 17-24.

Perkell, J., and Klatt, D. H. (1986). Invariance and Variability in Speech Processes (Lawrence Erlbaum, Hillsdale, NJ).

Smith, C. L. (1988). "A cross-linguistic contrast in consonant and vowel timing," J. Acoust. Soc. Am. 86, S84.

Smith, C. L. (1991). "The timing of vowel and consonant gestures in Italian and Japanese," Proceedings of the XIIth International Congress of Phonetic Sciences, Aug. 19-24, Aix-en-Provence, France, 4, 234-237.

Stevens, K. N. (1972). "The quantal nature of speech: Evidence from articulatory-acoustic data," in Human Communication: A Unified View, edited by E. E. David and P. B. Denes (Cambridge U.P., Cambridge), pp. 51-66.

Stevens, K. N. (1989). "On the quantal nature of speech," J. Phonet. 17, 3-45.

Suga, N., O'Neill, W. E., Kujirai, K., and Manabe, T. (1983). "Specific- ity of combination-sensitive neurons for processing of complex biosonar signals in auditory cortex of the mustached bat." J. Neurophys. 49, 1573-1626.

Sussman, H. M. (1989). "Neural coding of relational invariance in speech: Human language analogs to the barn owl," Psychol. Rev. 96, 631-642.

Sussman, H. M., McCaffrey, H. A., and Matthews, S. A. (1991). "An investigation of locus equations as a source of relational invariance for stop place categorization," J. Acoust. Soc. Am. 90, 1309-1325.

Van Son, R. J. J. H., and Pols, L. C. W. (1990). "Format frequencies of Dutch vowels in a text, read at normal and fast rate," J. Acoust. Soc. Am. 88, 1683-1693.

Wagner, H., Takahashi, T., and Konishi, M. (1987). "Representation of interaural time difference in the central nucleus of the barn owl's infe-

riorcolliculus," J. Neurosci. 7, 3105-3116.



a cross-linguistic investigation of locus equations as a ... · harvey m. sussman, kathryn a....

Documents