Perception of Place of Articulation for Syllable-Initial Consonants … · 2004-07-29 · For my...

University of California

Los Angeles

Perception of Place of Articulation for

Syllable-Initial Consonants in Noise

A thesis submitted in partial satisfaction

of the requirements for the degree

Master of Science in Electrical Engineering

by

Willa Shiao-Wei Chen

2001

c© Copyright by


2001

The thesis of Willa Shiao-Wei Chen is approved.

Nhan Levan

Kung Yao

Abeer Alwan, Committee Chair

University of California, Los Angeles

2001

ii

For my incredible parents, who make everything possible for me.

For my sisters Shirley, Rita and Connie to comfort and support when I need.

and

For all my wonderful friends to make my school experience a pleasant one.

iii

Table of Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Speech Production . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.3 Representing Speech Signals . . . . . . . . . . . . . . . . . . . . . 3

1.4 Vowels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.5 Consonants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.6 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.6.1 Plosive Consonants . . . . . . . . . . . . . . . . . . . . . . 7

1.6.2 Fricative Consonants . . . . . . . . . . . . . . . . . . . . . 11

1.7 Modeling of Speech Confusion in Noise . . . . . . . . . . . . . . . 14

1.8 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2 Acoustic Measurements . . . . . . . . . . . . . . . . . . . . . . . . 17

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.2 Measurements for Stops . . . . . . . . . . . . . . . . . . . . . . . 18

2.2.1 Formant Measurements . . . . . . . . . . . . . . . . . . . . 18

2.2.2 Noise Measurements . . . . . . . . . . . . . . . . . . . . . 20

2.2.3 Locus Equations . . . . . . . . . . . . . . . . . . . . . . . 24

2.3 Measurements for Fricatives . . . . . . . . . . . . . . . . . . . . . 25

2.3.1 Formant and Time Duration Measurements . . . . . . . . 25

2.3.2 Noise Measurements . . . . . . . . . . . . . . . . . . . . . 26

iv

2.3.3 Locus Equations . . . . . . . . . . . . . . . . . . . . . . . 32

2.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.4.1 Results for the Stops . . . . . . . . . . . . . . . . . . . . . 32

2.4.2 Percent Correct Classification for Stops . . . . . . . . . . . 38

2.4.3 Results for the Fricatives . . . . . . . . . . . . . . . . . . . 50

2.4.4 Percent Correct Classification for Fricatives . . . . . . . . 54

2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

3 Perceptual Experiments . . . . . . . . . . . . . . . . . . . . . . . . 68

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

3.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

3.2.1 Subjects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

3.2.2 CV syllables . . . . . . . . . . . . . . . . . . . . . . . . . . 69

3.2.3 Masking Noise . . . . . . . . . . . . . . . . . . . . . . . . . 69

3.2.4 Experiment Setup . . . . . . . . . . . . . . . . . . . . . . . 69

3.2.5 Experimental Protocol . . . . . . . . . . . . . . . . . . . . 70

3.3 Threshold Definition . . . . . . . . . . . . . . . . . . . . . . . . . 71

3.3.1 Sigmoid . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

3.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . 72

3.4.1 Vowel Context Comparison . . . . . . . . . . . . . . . . . 74

3.4.2 Effect of the Feature ‘Manner’ . . . . . . . . . . . . . . . . 74

3.4.3 Voicing Comparison . . . . . . . . . . . . . . . . . . . . . 77

3.5 Summary and Conclusion . . . . . . . . . . . . . . . . . . . . . . 81

v

4 Correlation Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

4.2 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

4.2.1 Threshold Values . . . . . . . . . . . . . . . . . . . . . . . 84

4.2.2 Acoustic Measurements . . . . . . . . . . . . . . . . . . . . 86

4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

4.3.1 Talker Independent Analysis . . . . . . . . . . . . . . . . . 87

4.3.2 Talker Dependent Analysis . . . . . . . . . . . . . . . . . . 92

4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

5 Summary and Future Directions . . . . . . . . . . . . . . . . . . . 104

5.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

5.2 Future Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

A Formant Measurements . . . . . . . . . . . . . . . . . . . . . . . . . 108

B Duration Measurements . . . . . . . . . . . . . . . . . . . . . . . . 157

C Spectral Measurements . . . . . . . . . . . . . . . . . . . . . . . . . 182

D Confusion Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . 207

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220

vi

List of Figures

1.1 Schematic representation of the complete physiological mechanism

of speech production[28] . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Waveform of /ba/ . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.3 FFT spectrum of /ba/ at steady state . . . . . . . . . . . . . . . 5

1.4 Spectrogram of /ba/ . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.5 Spectrogram of the Vowels /a/, /i/ and /u/ . . . . . . . . . . . . 6

2.1 Successive frames of spectra at the vowel onset in /ba/ that show

abrupt changes in spectral characteristics . . . . . . . . . . . . . . 19

2.2 F1 v.s. Time showing the offset of the first formant in /ba/ . . . . 20

2.3 Waveform of /ba/ by Male Talker No. 1. . . . . . . . . . . . . . . 21

2.4 Spectrogram of /ba/ by Male Talker No. 1. . . . . . . . . . . . . 21

2.5 Waveform of /da/ by Male Talker No. 1. . . . . . . . . . . . . . . 21

2.6 Spectrogram of /da/ by Male Talker No. 1 . . . . . . . . . . . . . 21

2.7 Waveform of /pa/ by Male Talker No.1. . . . . . . . . . . . . . . 22

2.8 Spectrogram of /pa/ by Male Talker No.1. . . . . . . . . . . . . . 22

2.9 Waveform of /ta/ by Male Talker No.1. . . . . . . . . . . . . . . . 22

2.10 Spectrogram of /ta/ by Male Talker No.1. . . . . . . . . . . . . . 22

2.11 Short-time Spectra of /ba/ by Male Talker No.1.The window lengths

was 6 ms for the burst/aspiration and 16 ms for the vowel onset. . 27

2.12 Short-time Spectra of /da/ by Male Talker No. 1 . . . . . . . . . 27

2.13 Short-time Spectra of /pa/ by Male Talker No. 1 . . . . . . . . . 27

vii

2.14 Short-time Spectra of /ta/ by Male Talker No. 1 . . . . . . . . . . 27

2.15 Scatter plots for voiced and voiceless stops in all vowel contexts . 28

2.16 Waveform of /fa/ by Male Talker No. 1. . . . . . . . . . . . . . . 29

2.17 Spectrogram of /fa/ by Male Talker No. 1. . . . . . . . . . . . . . 29

2.18 Waveform of /sa/ by Male Talker No. 1. . . . . . . . . . . . . . . 29

2.19 Spectrogram of /sa/ by Male Talker No. 1. . . . . . . . . . . . . . 29

2.20 Waveform of /va/ by Male Talker No. 1. . . . . . . . . . . . . . . 30

2.21 Spectrogram of /va/ by Male Talker No. 1. . . . . . . . . . . . . . 30

2.22 Waveform of /za/ by Male Talker No. 1. . . . . . . . . . . . . . . 30

2.23 Spectrogram of /za/ by Male Talker No. 1. . . . . . . . . . . . . . 30

2.24 Short-time Spectra of /fa/ by Male Talker No. 1. The window

length was 16 ms for both vowel onset and noise segment. . . . . 31

2.25 Short-time Spectra of /sa/ by Male Talker No. 1. . . . . . . . . . 31

2.26 Short-time Spectra of /va/ by Male Talker No. 1. . . . . . . . . . 31

2.27 Short-time Spectra of /za/ by Male Talker No. 1. . . . . . . . . . 31

2.28 Scatter plots for voiced and voiceless fricatives in all vowel contexts 33

2.29 Ahi-A23 for Stops. The small rectangles, and numbers next to

them, represent average values. The minimum and maximum num-

bers are shown by the lines. . . . . . . . . . . . . . . . . . . . . . 42

2.30 Av-maxA23 for Stops . . . . . . . . . . . . . . . . . . . . . . . . . 43

2.31 Av-Ahi for Stops . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

2.32 Amid-Avmid for Stops . . . . . . . . . . . . . . . . . . . . . . . . 45

2.33 Av4-A45 for Stops . . . . . . . . . . . . . . . . . . . . . . . . . . 46

viii

2.34 Av4-maxA45 for Stops . . . . . . . . . . . . . . . . . . . . . . . . 47

2.35 Slope and Y-intercept Values for Stops . . . . . . . . . . . . . . . 48

2.36 Ahi-A23 for Fricatives. The small rectangles, and the number next

to them, are the average values. The minimum and maximum

numbers are shown by the lines. . . . . . . . . . . . . . . . . . . . 62

2.37 Av-Ahi for Fricatives . . . . . . . . . . . . . . . . . . . . . . . . . 63

2.38 Av-Anoise for Fricatives . . . . . . . . . . . . . . . . . . . . . . . 64

2.39 Av4-A45 for Fricatives . . . . . . . . . . . . . . . . . . . . . . . . 65

2.40 Av4-maxA45 for Fricatives . . . . . . . . . . . . . . . . . . . . . . 66

2.41 Slopes and Y-intercepts Comparisons for Fricatives . . . . . . . . 67

3.1 Percent correct identification data for /ba,da/ and the sigmoidal

fit curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

3.2 Percent correct identification for stops. The dotted line in each

plot represent the 79% correct point. . . . . . . . . . . . . . . . . 76

3.3 Percent correct identification for fricatives . . . . . . . . . . . . . 77

3.4 Summary of the perceptual thresholds . . . . . . . . . . . . . . . 78

3.5 Thresholds grouped by vowel . . . . . . . . . . . . . . . . . . . . 78

3.6 Voiced CVs grouped by manner . . . . . . . . . . . . . . . . . . . 79

3.7 Voiceless CVs grouped by manner . . . . . . . . . . . . . . . . . . 79

3.8 Stop CVs grouped by voicing . . . . . . . . . . . . . . . . . . . . 80

3.9 Fricative CVs grouped by voicing . . . . . . . . . . . . . . . . . . 80

ix

List of Tables

2.1 Ranges of Formant Frequencies (in Hz) . . . . . . . . . . . . . . . 23

2.2 Onset of F1 for stops (in Hz) averaged across the four speakers.

Numbers in parentheses are the variances. . . . . . . . . . . . . . 34

2.3 Onset of F2 for stops(in Hz) averaged for all speakers. Numbers

in parenthesis are the variances. . . . . . . . . . . . . . . . . . . . 35

2.4 Frequency extent of F2 for stops (in Hz) averaged for all speakers.

Numbers in parenthesis are the variances. . . . . . . . . . . . . . . 35

2.5 Spectral quantity difference between /b,d/and /p,t/ for /a,i,u/,

averaged values for voiced and voiceless stops and the averaged for

all stops. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

2.6 Average data (in dB) from measurements on bursts in Stevens et

al.’s study. [36] . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

2.7 Slope, Y-intercept and R2 summary for stops with /a/ . . . . . . 38

2.8 Slope, Y-intercept and R2 summary for stops with /i/ . . . . . . . 38

2.9 Slope, Y-intercept and R2 summary for stops with /u/ . . . . . . 39

2.10 Slope and Y-intercept Values for Plosives when Considering all

Vowel Contexts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

2.11 Percent Correct Classification using F1, F2, or F3 onset values

for stop CVs in each vowel context and across vowel contexts.

Numbers in parentheses are the threshold values in Hz. . . . . . 39

x

2.12 Percent Correct Classification using F1∆, F2∆, or F3∆ onset val-

ues for stop CVs for each vowel context and across vowel contexts.

A negative frequency extent value indicates that the formant is

decreasing in time (steady state formant frequency is smaller than

formant onset frequency.) Numbers in boldface represent the high-

est percent correct classification for place of articulation using a

particular feature. Numbers in parentheses are the threshold val-

ues in Hz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

2.13 Percent Correct Classification using noise measurements Ahi-A23,

Av-Ahi, Av-maxA23 or Burst duration for stop CVs in each vowel

context and across vowels. Numbers in boldface represent the high-

est percent correct classification for place of articulation using a

particular feature. Numbers in parentheses are threshold values in

dB for the first three columns and ms for the last one. . . . . . . 41

2.14 Percent Correct Classification using noise measurements Amid-

Avmid, Av4-A45 or Av4-maxA45 for stop CVs in each vowel con-

text and across vowel contexts. Numbers in parentheses are thresh-

old values in dB. . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

2.15 Noise Duration for Fricatives (in ms). Numbers in parenthesis are

the variances. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

2.16 Onset of F2 for Fricatives (in Hz). Numbers in parenthesis are the

variances. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

2.17 Frequency Extent of F2 for Fricatives (in Hz). Numbers in paren-

thesis are the variances. . . . . . . . . . . . . . . . . . . . . . . . 52

2.18 Onset of F3 for Fricatives (in Hz). Numbers in parenthesis are the

variances. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

xi

2.19 Frequency Extent of F3 for Fricatives (in Hz). Numbers in paren-

thesis are the variances. . . . . . . . . . . . . . . . . . . . . . . . 52

2.20 Spectral quantity difference (in dB) between /f,s/ and /v,z/ for

/a,i,u/, averaged values for voiced and voiceless fricatives and the

averaged for all fricatives. . . . . . . . . . . . . . . . . . . . . . . 53

2.21 Slope, Y-intercept and R2 values for fricatives with /a/ . . . . . . 54

2.22 Slope, Y-intercept and R2 values for fricatives with /i/ . . . . . . 54

2.23 Slope, Y-intercept and R2 values for fricatives with /u/ . . . . . . 55

2.24 Slope and Y-intercept Values for fricatives when considering all

vowel contexts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

2.25 Percent Correct Classification using F1, F2, or F3 onset values

for fricative CVs in each vowel context and across vowel contexts.

Numbers in boldface represent the highest percent correct classifi-

cation for place of articulation using a particular feature. Numbers

in parentheses are threshold values in Hz. . . . . . . . . . . . . . 56

2.26 Percent Correct Classification using F1∆, F2∆, or F3∆ onset val-

ues for fricative CVs in each vowel context and across vowel con-

texts. Numbers in parentheses are threshold values in Hz. . . . . . 57

2.27 Percent Correct Classification using noise measurements Ahi-A23,

Av-Ahi, or noise duration for fricative CVs in each vowel context

and across vowel contexts. Numbers in parentheses are threshold

values in dB for the first two columns and ms for the last one. . . 58

xii

2.28 Percent Correct Classification using noise measurements Av-Anoise,

Av4-A45, or Av4-maxA45 for fricative CVs for each vowel context

and across vowel contexts. Numbers in parentheses are threshold

values in dB. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

2.29 Summary for the highest percent correct classification for each

syllable. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

3.1 Thresholds from each stop syllable and the thresholds from aver-

aged correct responses (in dB SPL) . . . . . . . . . . . . . . . . . 73

3.2 Thresholds from each fricative syllable and the threshold from av-

eraged correct responses (in dB SPL) . . . . . . . . . . . . . . . . 73

3.3 Thresholds for /Ca/ syllables in dB SPL . . . . . . . . . . . . . . 75

3.4 Thresholds for /Ci/ syllables in dB SPL . . . . . . . . . . . . . . 75

3.5 Thresholds for /Cu/ syllables in dB SPL . . . . . . . . . . . . . . 75

4.1 Threshold SNR (in dB SPL) from each talker and overall values . 85

4.2 Correlation coefficients for all CVs across all talkers . . . . . . . . 86

4.3 Correlation coefficients for /a/, /i/ and /u/ across all talkers for

both fricatives and stops . . . . . . . . . . . . . . . . . . . . . . . 87

4.4 Correlation Coefficients for Fricatives in the /a/, /i/ and /u/ con-

texts across all talkers. . . . . . . . . . . . . . . . . . . . . . . . . 88

4.5 Correlation coefficients for stops of /a/, /i/ and /u/ contexts across

all talkers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

4.6 Correlation coefficients for fricatives (voiced or voiceless) across all

vowel contexts for all talkers. Numbers in italic are percent correct

acoustic classification. . . . . . . . . . . . . . . . . . . . . . . . . 89

xiii

4.7 Correlation coefficients for stops (voiced or voiceless) across all

vowel contexts for all talkers. Numbers in italic are percent cor-

rect acoustic classification. Number in bold is the highest percent

correct acoustic classification for the syllable. . . . . . . . . . . . . 91

4.8 Correlation coefficients for /fa,sa/, /fi,si/ and /fusu/ for all talkers.

Numbers in italic are percent correct acoustic classification. Num-

ber in bold is the highest percent correct acoustic classification for

the syllable. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

4.9 Correlation coefficients for /va,za/, /vi,zi/, and /vu,zu/ for all

talkers. Numbers in italic are percent correct acoustic classifi-

cation. Number in bold is the highest percent correct acoustic

classification for the syllable. . . . . . . . . . . . . . . . . . . . . . 93

4.10 Correlation coefficients for /pa,ta/, /pi,ti/, /pu,tu/ for all talkers.

Numbers in italic are percent correct acoustic classification. . . . . 94

4.11 Correlation coefficients for /ba,da/, /bi,di/, /bu,du/ for all talkers.

Numbers in italic are percent correct acoustic classification. Num-

ber in bold is the highest percent correct acoustic classification for

the syllable. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

4.12 Correlation coefficients for all CVs from each talker (ie. T1 repre-

sents Talker No.1). . . . . . . . . . . . . . . . . . . . . . . . . . . 96

4.13 Correlation for /a/,/i/, /u/ from Talker No. 1 . . . . . . . . . . . 97

4.14 Correlation for /a/,/i/, /u/ from Talker No.2 . . . . . . . . . . . . 97

4.15 Correlation for /a/,/i/, /u/ From talker No. 3 . . . . . . . . . . . 98

4.16 Correlation for /a/,/i/, /u/ from Talker No. 4 . . . . . . . . . . . 98

4.17 Correlation for fricatives from Talker No. 1 . . . . . . . . . . . . . 98

xiv




4.21 Correlation for stops from Talker No. 1 . . . . . . . . . . . . . . . 100




A.1 Onset and Offset frequencies for /ba/ (Hz). Measures are listed as

CV.TalkerNum.TokenNum . . . . . . . . . . . . . . . . . . . . . . 109

A.2 Onset and offset frequencies for /da/ (Hz). Measures are listed as


A.3 Onset and offset frequencies for /pa/ (Hz). Measures are listed as


A.4 Onset and offset frequencies for /ta/ (Hz). Measures are listed as


A.5 Onset and offset Frequencies for /fa/ (Hz) Measures are listed as


A.6 Onset and offset Frequencies for /sa/ (Hz) Measures are listed as


A.7 Onset and offset Frequencies for /va/ (Hz) Measures are listed as


A.8 Onset and offset Frequencies for /za/ (Hz) Measures are listed as


xv

A.9 Steady-state frequency and frequency translation for /ba/ (Hz).

Measures are listed as CV.TalkerNum.TokenNum . . . . . . . . . 117

A.10 Steady-state frequency and frequency translation for /da/ (Hz).


A.11 Steady-state frequency and frequency translation for /pa/ (Hz).


A.12 Steady-state frequency and frequency translation for /ta/ (Hz).


A.13 Steady-state frequency and frequency translation for /fa/ (Hz).


A.14 Steady-state frequency and frequency translation for /sa/ (Hz).


A.15 Steady-state frequency and frequency translation for /va/ (Hz).


A.16 Steady-state frequency and frequency translation for /za/ (Hz).


A.17 Onset and offset Frequencies for /bi/ (Hz) Measures are listed as


A.18 Onset and offset Frequencies for /di/ (Hz) Measures are listed as


A.19 Onset and offset Frequencies for /pi/ (Hz) Measures are listed as


A.20 Onset and offset Frequencies for /ti/ (Hz) Measures are listed as


xvi

A.21 Onset and offset Frequencies for /fi/ (Hz) Measures are listed as


A.22 Onset and offset Frequencies for /si/ (Hz) Measures are listed as


A.23 Onset and offset Frequencies for /vi/ (Hz) Measures are listed as


A.24 Onset and offset Frequencies for /zi/ (Hz) Measures are listed as


A.25 Steady-state frequency and frequency translation for /bi/ (Hz)


A.26 Steady-state frequency and frequency translation for /di/ (Hz)


A.27 Steady-state frequency and frequency translation for /pi/ (Hz)


A.28 Steady-state frequency and frequency translation for /ti/ (Hz)


A.29 Steady-state frequency and frequency translation for /fi/ (Hz) Mea-

sures are listed as CV.TalkerNum.TokenNum . . . . . . . . . . . . 137

A.30 Steady-state frequency and frequency translation for /si/ (Hz)


A.31 Steady-state frequency and frequency translation for /vi/ (Hz)


A.32 Steady-state frequency and frequency translation for /zi/ (Hz)


xvii

A.33 Onset and offset Frequencies for /bu/ (Hz) Measures are listed as


A.34 Onset and offset Frequencies for /du/ (Hz) Measures are listed as


A.35 Onset and offset Frequencies for /pu/ (Hz) Measures are listed as


A.36 Onset and offset Frequencies for /tu/ (Hz) Measures are listed as


A.37 Onset and offset Frequencies for /fu/ (Hz) Measures are listed as


A.38 Onset and offset Frequencies for /su/ (Hz) Measures are listed as


A.39 Onset and offset Frequencies for /vu/ (Hz) Measures are listed as


A.40 Onset and offset Frequencies for /zu/ (Hz) Measures are listed as


A.41 Steady-state frequency and frequency translation for /bu/ (Hz).


A.42 Steady-state frequency and frequency translation for /du/ (Hz).


A.43 Steady-state frequency and frequency translation for /pu/ (Hz).


A.44 Steady-state frequency and frequency translation for /tu/ (Hz).


xviii

A.45 Steady-state frequency and frequency translation for /fu/ (Hz).


A.46 Steady-state frequency and frequency translation for /su/ (Hz).


A.47 Steady-state frequency and frequency translation for /vu/ (Hz).


A.48 Steady-state frequency and frequency translation for /zu/ (Hz).


B.1 Voiced Bar Duration,Burst Duration and VOT for /ba/ (ms) . . . 158

B.2 Voiced Bar Duration,Burst Duration and VOT for /da/ (ms) . . . 159

B.3 Voiced Bar Duration,Burst Duration and VOT for /pa/ (ms) . . . 160

B.4 Voiced Bar Duration,Burst Duration and VOT for /ta/ (ms) . . . 161

B.5 Voiced Bar Duration,Burst Duration and VOT for /bi/ (ms) . . . 162

B.6 Voiced Bar Duration,Burst Duration and VOT for /di/ (ms) . . . 163

B.7 Voiced Bar Duration,Burst Duration and VOT for /pi/ (ms) . . . 164

B.8 Voiced Bar Duration,Burst Duration and VOT for /ti/ (ms) . . . 165

B.9 Voiced Bar Duration,Burst Duration and VOT for /bu/ (ms) . . . 166

B.10 Voiced Bar Duration,Burst Duration and VOT for /du/ (ms) . . . 167

B.11 Voiced Bar Duration,Burst Duration and VOT for /pu/ (ms) . . . 168

B.12 Voiced Bar Duration,Burst Duration and VOT for /tu/ (ms) . . . 169

B.13 Voiced bar duration,VOT, noise duration and aspiration for /fa/

(ms) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170

xix

B.14 Voiced bar duration,VOT, noise duration and aspiration for /sa/

(ms) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

B.15 Voiced bar duration,VOT, noise duration and aspiration for /va/

(ms) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172

B.16 Voiced bar duration,VOT, noise duration and aspiration for /za/

(ms) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

B.17 Voiced bar duration,VOT, noise duration and aspiration for /fi/

(ms) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174

B.18 Voiced bar duration,VOT, noise duration and aspiration for /si/

(ms) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

B.19 Voiced bar duration,VOT, noise duration and aspiration for /vi/

(ms) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176

B.20 Voiced bar duration,VOT, noise duration and aspiration for /zi/

(ms) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

B.21 Voiced bar duration,VOT, noise duration and aspiration for /fu/

(ms) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178

B.22 Voiced bar duration,VOT, noise duration and aspiration for /su/

(ms) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

B.23 Voiced bar duration,VOT, noise duration and aspiration for /vu/

(ms) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

B.24 Voiced bar duration,VOT, noise duration and aspiration for /zu/

(ms) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181

C.1 Spectral measures for /ba/. . . . . . . . . . . . . . . . . . . . . . 183

C.2 Spectral measures for /da/. . . . . . . . . . . . . . . . . . . . . . 184

xx

C.3 Spectral measures for /pa/. . . . . . . . . . . . . . . . . . . . . . 185

C.4 Spectral measures for /ta/. . . . . . . . . . . . . . . . . . . . . . . 186

C.5 Spectral measures for /bi/. . . . . . . . . . . . . . . . . . . . . . . 187

C.6 Spectral measures for /di/. . . . . . . . . . . . . . . . . . . . . . . 188

C.7 Spectral measures for /pi/. . . . . . . . . . . . . . . . . . . . . . . 189

C.8 Spectral measures for /ti/. . . . . . . . . . . . . . . . . . . . . . . 190

C.9 Spectral measures for /bu/. . . . . . . . . . . . . . . . . . . . . . 191

C.10 Spectral measures for /du/. . . . . . . . . . . . . . . . . . . . . . 192

C.11 Spectral measures for /pu/. . . . . . . . . . . . . . . . . . . . . . 193

C.12 Spectral measures for /tu/. . . . . . . . . . . . . . . . . . . . . . . 194

C.13 Spectral measures for /fa/. . . . . . . . . . . . . . . . . . . . . . . 195

C.14 Spectral measures for /sa/. . . . . . . . . . . . . . . . . . . . . . . 196

C.15 Spectral measures for /va/. . . . . . . . . . . . . . . . . . . . . . 197

C.16 Spectral measures for /za/. . . . . . . . . . . . . . . . . . . . . . . 198

C.17 Spectral measures for /fi/. . . . . . . . . . . . . . . . . . . . . . . 199

C.18 Spectral measures for /si/. . . . . . . . . . . . . . . . . . . . . . . 200

C.19 Spectral measures for /vi/. . . . . . . . . . . . . . . . . . . . . . . 201

C.20 Spectral measures for /zi/. . . . . . . . . . . . . . . . . . . . . . . 202

C.21 Spectral measures for /fu/. . . . . . . . . . . . . . . . . . . . . . . 203

C.22 Spectral measures for /su/. . . . . . . . . . . . . . . . . . . . . . . 204

C.23 Spectral measures for /vu/. . . . . . . . . . . . . . . . . . . . . . 205

C.24 Spectral measures for /zu/. . . . . . . . . . . . . . . . . . . . . . 206

xxi

D.1 Confusion Matrices for /ba,da/ . . . . . . . . . . . . . . . . . . . 208

D.2 Confusion Matrices for /bi,di/ . . . . . . . . . . . . . . . . . . . . 209

D.3 Confusion Matrices for /bu,du/ . . . . . . . . . . . . . . . . . . . 210

D.4 Confusion Matrices for /pa,ta/ . . . . . . . . . . . . . . . . . . . . 211

D.5 Confusion Matrices for /pi,ti/ . . . . . . . . . . . . . . . . . . . . 212

D.6 Confusion Matrices for /pu,tu/ . . . . . . . . . . . . . . . . . . . 213

D.7 Confusion Matrices for /fa,sa/ . . . . . . . . . . . . . . . . . . . . 214

D.8 Confusion Matrices for /fi,si/ . . . . . . . . . . . . . . . . . . . . 215

D.9 Confusion Matrices for /fu,su/ . . . . . . . . . . . . . . . . . . . . 216

D.10 Confusion Matrices for /va,za/ . . . . . . . . . . . . . . . . . . . . 217

D.11 Confusion Matrices for /vi,zi/ . . . . . . . . . . . . . . . . . . . . 218

D.12 Confusion Matrices for /vu,zu/ . . . . . . . . . . . . . . . . . . . 219

xxii

Acknowledgments

My sincere thanks to Prof. A. Alwan for providing me with the background and

the opportunity for this project. Thanks to James Hant and Marcia Chen for

helping me through out this project. In addition, I would like to thank all the

subjects who gave their time to help making the experiments possible. Finally,

I would also like to thank everyone at the SPAPL lab for their comments and

support in helping my two years of graduate school a joyful and exciting one.

This thesis was supported in part by NIH - NIDCD.

xxiii

Abstract of the Thesis

Perception of Place of Articulation for

Syllable-Initial Consonants in Noise

by


Master of Science in Electrical Engineering

University of California, Los Angeles, 2001

Professor Abeer Alwan, Chair

This thesis studies the perception of the place of articulation feature for conso-

nants when listening in noisy environments, and attempts to relate perceptual

thresholds to physical measurements of the speech signal. Acoustic correlates of

the labial and alveolar place of articulation for both plosive and fricative conso-

nants are investigated, and perceptual experiments were conducted to find out

how the place feature is perceived in noise. Consonant-Vowel pairs of the plosives

/p,t,b,d/ and fricatives /f,s,v,z/ with the vowels /a/,/i/, and /u/ are used in the

study. Acoustic analysis includes measurements of the formant frequencies (res-

onances of the vocal-tract transfer function) and noise measurements. Results

show that noise measurements are generally more reliable cues than formants

for place of articulation although the second formant frequency (F2) informa-

tion classifies place well for some syllables. Perceptual experiments show that

fricatives are generally more robust than stops, and that /Ca/ syllables are more

robust than /Cu/ syllables which are in turn more robust than /Ci/ syllables.

Correlation analysis was performed between acoustic measures and perceptual

thresholds. Results suggest that talker variability plays an important role in

xxiv

correlating acoustics and perception. Acoustic features such as F1 and F2 infor-

mation and the amplitude of the noise segment relative to the vowel onset in the

F4-F5 region show high correlation with perceptual thresholds. A major result

of this study is that no single acoustic feature can cue perception of place of

articulation in noise. The results are dependent upon the talker, vowel context,

voicing and manner of articulation.

xxv

CHAPTER 1

Introduction

1.1 Motivation

The goal of this study is to investigate cues for predicting the perceptual confu-

sion of the place of articulation feature in speech signals when listening in noisy

environments. The approach is to integrate knowledge of the acoustic correlates

for place of articulation and properties of human speech perception.

Most previous studies that examined the perception of place of articulation

had been done in quiet environments and the results from these studies may be

used as a guideline to predict confusion of place in noisy environments.

Modeling speech perception in noise has long been a challenging topic. Sci-

entists and engineers have been trying to find better ways for noise robust ASR

(Automatic Speech Recognition) systems and hearing aids. If we can find speech

cues that are perceptually salient in a noisy environment, these cues can be en-

hanced in speech processing algorithms to design better ASR systems, coders or

hearing aids.

1.2 Speech Production

The vocal tract is defined from the opening of the vocal cords, or the glottis, to

the lips. The cross-sectional area of the vocal tract is determined by the position

1

Figure 1.1: Schematic representation of the complete physiological mechanism of

speech production[28]

of the lips, tongue, jaw and velum.

While breathing, air enters and exits the lungs through the trachea, or wind-

pipe, and can cause the vocal cords within the larynx to vibrate as an oscillator.

Air flow then turns into quasi-periodic pulses that modulate in frequency and pass

through the throat cavity, or the pharynx, the mouth cavity, and sometimes, the

nasal cavity. Different sounds are then produced depending on the positions of

the various articulators.

A simplified representation of the complete physiological mechanism for pro-

ducing speech is shown in Figure 1.1 [28].When we speak, air from the lungs

passes through the vocal folds in the larynx. Voiced speech sounds are produced

2

when the vocal cords are tensed and the resulting vibration is periodic; the fre-

quency of this vibration determines the pitch of the voiced sound. When the

vocal cords are relaxed, in order to produce a sound, the air flow must either

pass through a constriction in the vocal tract, or it can build up pressure behind

a point of total closure within the vocal tract. Turbulence occurs if air passes

through a narrow constriction in the vocal tract. On the other hand, a brief

transient sound is produced when the pressure behind the closure is suddenly

released [28].

The excitation source is then filtered by the vocal tract transfer function.

The shape, cross-sectional area, and the length of the vocal tract determine the

amplitude and location of the resonant, or formant frequencies and the anti-

resonances of the vocal tract transfer function. These formant frequencies are

very important in determining the identity of speech sounds.

1.3 Representing Speech Signals

Speech signals are fast changing signals in time. However, when the signals are

examined over sufficiently short periods of time, typically between 5 and 20 ms,

speech signals are relatively stationary. To illustrate the various properties of the

speech signal, we show in Figure 1.2 the waveform of the syllable /ba/ as spoken

by a male talker. Speech waveforms can be classified through the state of the

speech production apparatus and we typically notice the following:

• Silence, when no speech is produced.

• The time-waveform of “unvoiced” segment is aperiodic and typically low in

amplitude.

3

Figure 1.2: Waveform of /ba/

• The time-waveform of “voiced” segment is quasi-periodic reflecting the vo-

cal cords vibration.

Another way of representing speech is by the Fourier Transform spectrum.

Figure 1.3 is the spectrum of the steady-state part of the /ba/ waveform. The

peaks of the spectrum envelope represent the formant frequencies of the sound.

The spacing between the harmonics is the pitch frequency.

One other popular representation of the speech signal is the spectrogram. A

spectrogram is a three dimensional representation which plots the signal in fre-

quency bands over time and the intensity of the signal is represented by the

darkness of the spectrogram. Figure 1.4 shows the spectrogram of the /ba/

syllable.

4

Figure 1.3: FFT spectrum of /ba/ at steady state

Figure 1.4: Spectrogram of /ba/

5

/a/

/i/

/u/

Time

Frequency

Figure 1.5: Spectrogram of the Vowels /a/, /i/ and /u/

1.4 Vowels

Vowels are produced by exciting the vocal tract with quasi-periodic pulses of air

caused by the vibration of the vocal cords. The way the cross-section of the

vocal tract varies determines the resonance, or formant, frequencies of the sound

produced. Vowel sounds are usually long in duration and spectrally defined, thus

play a significant role in speech recognition both by human and machine.

One way of classifying vowels is by their tongue position. For example, /i/ is a

front vowel and /u/ is a back vowel. Front vowels have the second (F2) and third

(F3) formant frequencies close together while back vowels usually have F1 and

F2 close together. These characteristics can be observed in their spectrograms

(Figure 1.5).

6

1.5 Consonants

Consonants can be broadly characterized by their 1) place of articulation, 2) man-

ner of articulation, and 3) voicing. Voicing is determined by a periodic excitation

source while the excitation source for unvoiced sounds is noise. The manner of

articulation groups sounds into, for example, plosives, which are characterized

by the build up of pressure behind a closure in the vocal tract and the sudden

release of pressure, and fricatives, which are characterized by turbulence in the

region of maximum constriction in the vocal tract. The location of the maxi-

mum constriction is called “ place of articulation ”. In this study, we examine

Consonant-Vowel (CV) pairs that only differ by their place of articulation. Such

pairs include labial stops (/b,p/), or labio-dentals fricatives (/f,v/), both have

the constriction at the lips, and alveolars (/d,t,s,z/), which have the constriction

at the roof of the mouth.

1.6 Background

1.6.1 Plosive Consonants

Over the last several decades, various studies have been performed to find invari-

ant acoustic cues for the production and perception of stop consonants. One of

the earliest works that characterized plosive formant transition was the “hub”

by Potter et al. in 1947 [27]. The “hub” was defined as the location of F2 (in

Hz). Potter et al. found that labial stops have the hub at the bottom of their

spectrogram while alveolars had their hub near the middle. In 1973, Fant ana-

lyzed spectrograms of six Swedish plosive consonants in nine vowel contexts and

concluded that F2 and F3 formant transition pattern did not sufficiently reflect

place of articulation and argued that the “loci” concept only applies to synthetic

7

speech sounds [6].

In 1982, Kewley-Port [19] studied acoustic correlates of place of articulation

in naturally produced voiced stops. For one male talker, she measured the first

three formant transitions (including the onset, steady-state, transition duration,

and time of onset relative to the burst) for stop CVs of /b/, /d/, /g/ in eight

vowel contexts ranging from front to back vowels. These CVs were embedded in

one sentence. She found that F2 and F3 transition onset values are not sufficient

to cue place of articulation across all vowel contexts.

Other studies for cuing place of articulation involved the characteristics of the

burst. In 1977, Zue [40] analyzed natural plosive bursts and found that an alveolar

burst spectrum has a broad shaped spectral peak in the high frequency region,

a velar burst has a compact peak at the mid-frequency range when followed

by a front vowel (such as /i/) and at a lower frequency range when followed

by a back vowel (such as /a/). As for the labials, Zue could not find general

burst characteristics. Stevens and Blumstein [3],[35] analyzed burst spectra in

an attempt to find an invariant acoustic pattern for place of articulation. They

defined labial bursts to be “diffused falling” (spectral energy is wide spread with a

concentration at the low to mid-frequency range), alveolar bursts to be “diffused

rising” (spectral energy is wide spread with a concentration at the high frequency

range), and velars to have a mid-frequency compact spectral peak. These spectral

templates exhibited approximately 85 % correct categorization of initial stop

place. A modification of the Stevens and Blumstein study was done by Kewley-

Port in 1983 [20]; she modified the fixed time window(26 ms) of Stevens and

Blumstein with a running spectral display, advancing by 5 ms, to include the

initial 40 ms of the vowel onset. Kewley-Port was able to categorize place of

articulation for stops at 88%.

8

More recently, researchers have suggested that the spectral amplitude of the

consonant portion of a consonant-vowel syllable relative to the onset of the vowel

cue place of articulation. Past studies of /p/-/t/ contrast showed that greater

burst amplitude in the high frequencies (F4 region) and higher presentation levels

result in more alveolar responses , [8], [26], [34]. One of the most recent studies of

burst cues was done by Stevens et al. in 1999 [36]. In that study, three spectral

quantities were measured: (1) the peak spectrum amplitude of the burst in the

frequency range above 3500 Hz (and 3000 Hz for males) relative to the average of

spectral peaks in the F2 and F3 range (in dB), denoted by Ahi-A23, measuring

the spectral tilt of the burst; (2) the spectrum amplitude of the F1 prominence

in the vowel onset relative to Ahi, denoted by Av-Ahi, measuring the burst am-

plitude relative to the vowel amplitude; and (3) the difference between Av and

the maximum spectrum amplitude in the F2-F3 range, denoted by Av-maxA23,

measuring the mid-frequency spectral prominence. These measurements were

done on a number of syllable-initial stop consonants, 15 tokens each, drawn from

100 read sentences spoken by 2 male and 2 female talkers. The results showed

that Ahi-A23 was smaller (meaning the burst spectra has smaller amplitude in

higher frequency (Ahi) than F2-F3 prominence (A23)) for labials but relatively

flat (similar Ahi and A23) for the alveolars. Av-Ahi showed that the labial burst

at high frequencies was weaker than the alveolar burst. Av-maxA34 best classified

velars because velars have a more prominent mid-frequency peak. Classification

accuracy showed a dependency on talkers and vowel contexts.

Another way for classifying place of articulation was done by the locus equa-

tion approach. The concept of “locus equations” was first proposed by Lindblom

in 1963 [23]. Locus equations are linear regression fits to data points formed by

the onset of F2 (at the y-axis) and the corresponding mid-vowel (nuclei) frequen-

cies at the x-axis. When performing this analysis on naturally spoken utterances

9

from one speaker with consonants /b,d,g/ at initial and final positions, Lindblom

found that the slope and y-intercept of these regression lines varied as the place

of articulation changes (slopes were 0.69, 0.28 and 0.95 for /b/, /d/ and /g/,

respectively, with y-intercept values at 410, 1225 and 360Hz). When the same

analysis was performed using data points formed by the third formant informa-

tion, no clustering was found. However, when Nearey and Shammass performed

analysis on both F2 and F3 using the onset and the steady-state vowel frequency

of /CVd/ syllables where C=/b,d,g/ from 5 female and 5 male speakers, they con-

cluded that slopes and intercepts for the three consonants are distinct and were

good representations for place of articulation [25]. In 1991 Sussman et al. tried

to characterize invariant cues for place of articulation based on locus equation

analysis [37]. They recorded plosive CVCs from 20 talkers and plotted scatter

plots of F2 onset v.s. F2 steady-state frequencies. When linearly fitting the data,

they could categorize 100% of the stops into their specific place categories.

Studies were also conducted to find perceptual cues of place of articulation.

In the 1950s, Liberman et. al performed simple synthesis and perception ex-

periments and found the role of direction and extent of F2 cue place for stop

consonants [22]. Similar synthetic speech perceptual experiments were done by

Cooper et. al [4] and Delattre et al. [5]. In 1955, Delattre et al. synthesized stop

consonant-vowel syllables by varying the onset frequency of F2 and performed

perceptual experiments on these synthesized syllables [5]. They found that the

perception of /b/ was the best when F2 onset frequency was at 720Hz and /d/

at 1800 Hz. The basic ideas behind these studies is that the perception of place

of articulation for plosive consonants across vowel contexts can be characterized

by a set of “loci”positions for F2 and F3 frequencies at the onset of the following

vowel. When similar studies were done in 1961 using natural CV syllables rather

than synthetic ones, the results were not as successful [21].

10

Hedrick et al. in 1995 and 1996 [13], [14] performed perceptual experiments

on normal and hearing impaired subjects using voiceless plosive CV syllables and

varied the burst amplitude in the F4-F5 frequency region relative to the vowel

onset amplitude and F2 and F3 onset frequency. The results for normal hearing

subjects showed that increasing the relative presentation level of the burst yields

more alveolar responses. The increase in alveolar responses also co-vary with the

F2 and F3 onset frequency.

In 1996, Smits et al. [30] performed perceptual experiments on Dutch stops

/b,d,p,t,k/ to investigate the role of the burst in cuing place of articulation.

They performed burst-splicing methods and conducted perceptual experiments

for burst only and burstless CV syllables. They concluded that with burst only

stimuli, the percent correct identification was about 60 %, while it was 90 % for

burstless stimuli. Even though the results were highly dependent upon the stop

consonants, (whether they are labials, alveolars or velars), and vowel contexts,

(front, mid or back vowels), the overall results suggested that humans can identify

place of articulation without the burst.

1.6.2 Fricative Consonants

Fricatives exhibit frequency characteristics that differed somewhat from plosives.

In 1956 and 1981, Hughes and Halle [16] and Soli [31] showed that the formant

location and transition of fricatives depend on talkers and vowel contexts. An-

other important cue for fricatives is the duration of frication noise. In 1979, You

[39] showed that the duration of frication noise varies with place of articulation;

the frication noise of alveolo-palatal fricatives averaged 176 ms, alveolars 155 ms,

labials 103 ms and dentals 99 ms.

Another approach for classifying fricatives are based on the amplitude of noise.

11

Fricative studies have shown differences in the amplitude of noise as a function of

place. Shadle et al. [29] tried to model fricatives by measuring spectral moments,

dynamic amplitude and spectral slope. Their database consisted of fricatives at

different effort levels and in different vowel contexts. They found that moments

varied significantly by frequency range and that the difference between the min-

imum amplitude value between 0-2 kHz and the maximum amplitude between

0.5-17 kHz is higher for /s/ than /f/.

There were also several perceptual studies on place cues for fricatives. In

1958 and 1961, perceptual experiments performed by Harris[11] and Heinz and

Stevens [15], respectively, using natural and synthetic speech showed that spectral

properties of frication noise are critical perceptual attributes for place of articu-

lation. They also investigated the amplitude of frication in synthetic stimuli with

different initial frequencies for the frication portion of the syllable and different

starting frequencies of the second formant of the vowel [15]. They then varied the

amplitude of the fricative noise relative to the vowel. The results showed that

stimuli with resonance frequencies of 6500-8000 Hz usually produced /f/ and /θ/

responses, but these responses only began to emerge when the fricative noise was

-15 and -25 dB relative to the vowel. In 1981, Guerlekian [9] explored the overall

amplitude of the noise relative to the vowel in the perception of /fa/ and /sa/.

Using several synthesized stimuli with conflicting cues, he found that low ampli-

tude of noise relative to the vowel was perceived by both Spanish and English

listeners as /fa/, and high amplitude noise relative to the vowel was perceived as

/sa/.

In 1985, Stevens studied the relative amplitude effect at different frequency

regions [33]. He varied the amplitude of the fricative relative to the vowel at

specific frequency regions. The results showed that responses shifted from /s/ to

12

/sh/ when the relative amplitude in the F3 region increases; stimuli with increas-

ing relative amplitude in the F5-F6 frequency region resulted in shifted responses

from /θ/ to /s/. In the same year, Jongman found that correct identification for

place of articulation does not change significantly even when the frication noise

of the naturally produced syllables were presented alone, without the transitions

and vowel information [17]. He also found that the duration of frication noise

seemed to play a minimal role in the perception of place of articulation in voiceless

fricatives. In 1988, Jongman conducted a perceptual experiment using natural

speech CV syllables (consisted of seven voiced and voiceless fricatives followed

by vowels /a,i,u/). The syllables were computer-edited to include 20-70 ms, in-

crement by 10 ms, of their natural frication duration. Results showed that the

listeners did not require the entire fricative-vowel syllable in order to correctly

perceive a fricative. However, the study showed that listeners required shorter

noise duration for alveolars (approximately 30ms) than for labials (approximately

50ms) [18].

A number of studies have explored the perceptual role of the amplitude prop-

erties of the fricative noise as a cue to place of articulation in fricatives. Behrens

and Bumstein [2] found that when the frication spectrum and formant transition

cues were appropriate for a specific fricative, perception of place of articulation

of fricatives was generally not influenced by overall frication amplitude. This re-

sult contradicts with other theories based on the amplitude feature for fricatives.

However, Behrens and Blumstein suggested that the relevant property of ampli-

tude may not be the “overall” amplitude of the frication, but rather a change in

amplitude of the fricative noise relative to the vowel in a specific region.

In 1993, Hedrick et al. [12] performed a more extensive study on the amplitude

of frication relative to vowel onset amplitude in the F3 and F5 formant frequency

13

region. Four sets of studies were done: (1) varying the duration of the frication

from 30 ms to 140 ms, (2) pairing frication noise with different vowels, (3) placing

formant transition in conflict with relative amplitude, and (4) varying formant

transitions and frication amplitudes of spectral regions but keeping the ratio

between frication and vowel amplitude constant. The stimuli were presented

either in isolation or by inserting a short gap between the noise and the vowel.

The results showed that the amplitude of the frication noise relative to the vowel

in the F3-F5 region affects perception of place across different vowel contexts

and frication duration. The relative amplitude was a robust cue for place even

when formant transitions and spectral levels other than those in the relative

amplitude region varied. Evidence from temporal gap studies showed that both

a short and long term memory process mediate the relative amplitude comparison,

especially in the short term case. Moreover, although relative amplitude is only

a component of spectral prominence, the spectral frication peak is primary in

perception while the comparison between the frication spectral peak and a vowel

peak in the same frequency region is of secondary perceptual importance [12].

Overall, labial fricatives seemed to have weaker intensity noise relative to the

vowel while alveolar fricatives have stronger noise relative to the vowel.

1.7 Modeling of Speech Confusion in Noise

One of the earliest studies on perceptual confusions in noise were done by Miller

and Nicely in 1954 [24]. In their study, sixteen English consonants followed by

the vowel /a/ were recorded in a list of 200 non-sense syllables. They selected

consonants across five articulatory features (voicing, nasality, affrication, duration

and place of articulation), with initial or final consonant positions of /a/ context.

Frequency distortion was imposed and different levels of random noise masks were

14

added to the recorded syllables. Subjects were asked to identify the consonant

in presence of the frequency distortion or noise masker. Confusion matrices were

generated to show the responses. One major result drawn from their study was

that voicing was much less affected by a noise masking (discrimination of voicing

is still good at SNR as poor as -12 dB), where as place of articulation is affected

more by masking noise (confusion becomes significant at around 6 dB). More

specifically, they found the plosives are much less robust than fricatives in a

noisy environment.

More recently, scientists used an information/theoretic approach to try to

model the confusion of speech in noise [32], [38]. These studies attempt to find

out which acoustic and phonetic cues account for perceptual tests in noise by

analyzing confusion data statistically. Results from these studies were highly

dependent upon the acoustic and phonetic dimensions chosen.

In 1987, Farar et al. [7] adopted an approach to quantify perceptual con-

fusions in noise by incorporating speech into psycho-acoustic masking models.

Using stationary broad-band noises with spectral shapes resembling certain plo-

sives, they measured the discrimination thresholds for different stop burst pairs

as a function of the burst’s duration. The results showed that discrimination

thresholds decreased nearly 20 dB as the bursts’ duration increased from 10 to

300 ms. However, they were unable to model the data to predict these durational

effects.

In 1992, Alwan [1] conducted discrimination experiments with synthetic /ba,da/

stimuli while masking their F2 trajectories with a bandpass noise masker. The

discrimination results suggested that high frequency cues (such as relative ampli-

tude differences in F3 and F4 frequency regions) can be used as place cues since

subjects were able to identify the consonants when F2 was completely masked.

15

Hant in 2000 [10] performed extensive perception experiments to investigate mod-

els for predicting human perception of speech in noise. He developed a general

time/frequency detection model to fit the noise-masked thresholds of bandpass

noises which varied in noise duration, bandwidth, and center-frequency. The

resulting models were used to predict the masking of glides, single formant tran-

sitions, synthetic plosive bursts, 4-formant transitions similar to those found in

plosive consonants and the discrimination and identification of synthetic and nat-

ural plosive CVs in various noise environments.

1.8 Thesis Outline

In this thesis, the approach of Alwan [1] is expanded by using both fricative and

plosive CVs to investigate the perception of the place feature in white noise. The

perceptual experiments are 2AFC (2 alternate-forced-choice) format. In order to

investigate cues for the perception of place of articulation in noise, it is necessary

to understand perceptual cues of place in quiet.

In Chapter 2, acoustic measurements (formant frequency and duration mea-

surements, and noise measurements) of plosive and fricative CVs in /a,i,u/ are

reported. The measurements are compared between alveolars and labials. The

procedures and results for the perceptual experiments are reported in Chapter

3. In Chapter 4, a correlation analysis is performed to seek the relation between

the perceptual experiment thresholds and the acoustic cue distinctions between

labials and alveolars. Finally, the results are summarized and possible future

studies are mentioned in Chapter 5.

16

CHAPTER 2

Acoustic Measurements

2.1 Introduction

In this chapter, the acoustic correlates of the labial and alveolar place of articula-

tion for both plosive and fricative consonants are investigated.The stimuli used for

this study consisted of naturally-spoken consonant-vowel utterances(CVs). Plo-

sives, /p,t,b,d/ and fricatives /f,s,v,z/ in syllable-initial position with the vowels

/a/,/i/, and /u/ spoken by 2 male and 2 female talkers, four repetitions each, of

American English were studied. The speech tokens were recorded digitally with

16 bits and sampling frequency of 16 kHz using a head mounted microphone.

Two types of acoustic measurements were made: formant frequency measure-

ments, and noise measurements. The first set of measurements consisted of the

duration, frequency and amplitude of the formant transitions in the vowel. The

second set included measures which attempt to quantify spectral characteristics

of the noisy segments (frication, burst and aspiration).

17

2.2 Measurements for Stops

2.2.1 Formant Measurements

A stop consonant is generated by a complete closure in the vocal tract via a

constriction at the place of articulation followed by the release of pressure built

up behind the constriction. The closure of the vocal tract is characterized acous-

tically by a silence and a voiced bar (in some voiced stops), while the release of

pressure is characterized by a short noise burst. The interval between the release

of the burst and the beginning of voicing in the vowel is referred to as the voice

onset time (VOT). Silence and/or aspiration can be found during this period.

Following the VOT are the onset and steady state formant frequencies of the

vowel. Measurements included:

• F1, F2, F3 transition duration and frequency Measurements of the

formant transitions were obtained visually by inspecting the time waveform,

wide-band spectrogram, short-time DFT (Discrete Fourier Transform) and

LPC (Linear Predictor Coefficient) spectra using Matlab. Spectral analysis

was done by using 20 ms (for male speakers) or 15 ms (for female speakers)

Hamming windows, overlapped by 2.5 ms. LPC spectra were calculated

from 8-12 LPC coefficients (depending on the talker). The onset of the

vowel was defined as the center point of the frame which shows an abrupt

change in the waveform, spectrogram, and spectral features. The total

energy at the onset of the vowel also shows a sudden increase. Figure 2.1

shows the abrupt change in F1, from left to right, to indicate the onset of

voicing. The onset of F1 usually occurs at the same frame where the onset

of voicing occurs, but not necessarily for F2 and F3. The end of formant

transitions, chosen automatically, was defined as the frame during which

18

Figure 2.1: Successive frames of spectra at the vowel onset in /ba/ that show

abrupt changes in spectral characteristics

the rate of change of the formant frequency fell to less than 5 Hz per 2.5

ms, see Figure 2.2, and the average rate of change for the next 5 frames

was also less than 5 Hz per 2.5 ms.

Since the determination of the transition offset is prone to error, a third

point, called the steady-state, was also measured at 95 ms after vowel onset.

At each of these three points (vowel onset, offset and steady-state), the time,

frequency and amplitude of F1, F2, and F3 were recorded.

• Burst and VOT duration: The measurements of the burst and VOT

duration were obtained by visually inspecting the time waveform and the

wide-band spectrogram of the signal using CoolEdit Pro (Syntrillium Soft-

ware Corporation). Waveforms and spectrograms of /ba, da, pa, ta/ by

Talker No. 1 are shown in Figures 2.3 - 2.10

19

Figure 2.2: F1 v.s. Time showing the offset of the first formant in /ba/

2.2.2 Noise Measurements

To obtain noise measurements, we need to first splice the CV into burst and

vowel segments. We performed a “burst-splicing” method manually with the aid

of the time waveform and the spectrogram. As for the vowel segment, we picked

the first impulse of the onset of the vowel after the burst segment [30], [36].

The spectrum of the combined transient and burst was estimated using the

Welch’s averaged periodogram method [36]. The signal is divided into overlapping

sections of specified window length. To estimate the spectrum of the stop bursts,

we need to first determine the duration of the burst release to voicing onset. If

the duration of the burst is shorter than 9 ms, then we average the burst by

using a 3 ms window with 1.5 ms overlap, otherwise, a 6 ms window was used

with a 3 ms overlap. The spectrum was obtained using a 256 point Fast Fourier

Transform method. Short time spectra of the burst and vowel onset for the CVs

/ba, da, pa, ta/ by Talker No.1 are shown in Figures 2.11- 2.14

20

1000 2000 3000 4000 5000 6000 7000 8000 9000−3

−2

−1

0

1

2

3

x 104

Figure 2.3: Waveform of /ba/ by Male

Talker No. 1.

Time

Fre

quen

cy

0 500 1000 1500 2000 2500 3000 3500 4000 45000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Figure 2.4: Spectrogram of /ba/ by Male

Talker No. 1.

1000 2000 3000 4000 5000 6000 7000 8000 9000−2

−1.5

−1

−0.5

0

0.5

1

1.5

2x 10

4

Figure 2.5: Waveform of /da/ by Male

Talker No. 1.

Time

Fre

quen

cy

0 500 1000 1500 2000 2500 3000 3500 4000 4500 50000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Figure 2.6: Spectrogram of /da/ by Male

Talker No. 1 .

The formant frequency regions were determined from vowel spectra with care-

ful comparison with the spectrogram of the CV. The frequency regions are sum-

marized in Table 2.1

• Ahi-A23: Ahi represents the peak spectrum amplitude (in dB) of the

burst in the frequency range above 3500 Hz for females and 3000 Hz for

males. A23 is the average of the spectral amplitude in the second and

third formant frequency range. The quantity “Ahi-A23” is the difference

21

1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 11000

−1.5

−1

−0.5

0

0.5

1

1.5

x 104

Figure 2.7: Waveform of /pa/ by Male

Talker No.1.

Time

Fre

quen

cy

0 500 1000 1500 2000 2500 3000 3500 4000 4500 50000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Figure 2.8: Spectrogram of /pa/ by Male

Talker No.1.

1000 2000 3000 4000 5000 6000 7000 8000 9000−2

−1.5

−1

−0.5

0

0.5

1

1.5

x 104

Figure 2.9: Waveform of /ta/ by Male

Talker No.1.

Time

Fre

quen

cy

0 500 1000 1500 2000 2500 3000 3500 4000 45000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Figure 2.10: Spectrogram of /ta/ by

Male Talker No.1.

22

/a/ /i/ /u/

F1-F2 200-1000 200-1000 200-1000

F2-F3 1000-3000 1500-3500 1000-3000

F4-F5 3000-5000 4000-6000 3000-5000

Table 2.1: Ranges of Formant Frequencies (in Hz)

between Ahi and A23 (in dB). This measurement characterizes the spectral

tilt of the burst [36].

• Av-Ahi: Av represents the spectrum amplitude peak of the vowel spec-

trum at the first formant frequency prominence. Av-Ahi is the high fre-

quency burst amplitude relative to F1 in the vowel[36].

• Av-maxA23: The quantity, maxA23, here is similar to A23. Instead of

taking the average of the spectral amplitude in the F2-F3 formant region,

we took the maximum of the burst spectral amplitude around the same

region. Av-maxA23 was calculated for stops to determine a mid-frequency

spectral prominence[36].

• Av4-A45 Similar to the definition for Av & A23, Av4 here represents

the spectrum peak amplitude of the vowel in the forth formant frequency

region. A45 is the average amplitude of the burst in the forth and fifth

formant frequency region as described in Table 2.1. Previous studies have

suggested that F4 and F5 play an important role in the place of articulation

distinction [13], [14]. This measurement, Av4-A45 characterizes the rela-

tive amplitude of the vowel versus the burst in the F4-F5 formant frequency

region.

• Av4-maxA45: This quantity is very similar to Av4-A45 except that we

23

calculate the maximum amplitude of the burst in the forth and fifth formant

frequency regions.

• Amid-Avmid: Amid is the average of the burst spectral amplitude at

mid frequency, between 3200 Hz and 4800 Hz. Similarly, Avmid is the

average of the vowel spectral amplitude between 3200 Hz and 4800 Hz.

This quantity characterized the difference between burst and vowel spectral

amplitude at the mid-frequency range.

Measurements Av-Ahi, Ahi-A23, and Av-maxA23 were inspired by Stevens

et al. [36] except there are two differences in calculating these noise measures.

First, the burst segment in the studies by Stevens et al. did not include aspiration

in voiceless stops. Second, Stevens et al. used same window length for both the

vowel onset and the burst segment.

2.2.3 Locus Equations

The locus equation approach was proposed by Sussman et al. in 1991 [37] in an

investigation of the invariant cues for place of articulation for voiced initial stop

consonants. Locus equations are derived from linear regression lines that fit the

data points of the onset and steady-state F2 frequencies. A scatter plot of the

F2 onset frequency (denoted by F2onset) at the y-axis and their corresponding

steady-state F2 frequency (denoted by F2ss) at the x-axis were plotted for all

speech tokens and CVs (Figure 2.15). The linear regression fit was then generated

from all the data points for each CV. The Y-intercept value (denoted by C) and

the slope (denoted by k) are calculated for analysis purposes. The locus equation

has the following form:

F2onset = k × F2ss + C

24

where the coefficient of determination is:

R2 = 1−

∑allj′s(Yj − Yj)

2

(∑

allj′s Y 2j )− (

∑allj′s Yi)2

n

where Yj is the actual data point series and Yj is the fitted data point series,

and n is the total number of points.

The coefficient of determination R2 was calculated to show the correlation

between the data points and the linear fit.

2.3 Measurements for Fricatives

2.3.1 Formant and Time Duration Measurements

Fricatives /f,s,v,z/ are produced by exciting the vocal tract with turbulence in the

region of the maximum constriction in the vocal tract and by a voicing source

for the voiced ones [28]. The place of articulation is the point of maximum

constriction. For labials /f,v/, the constriction is near the lips; for alveolars,

/s,z/, the constriction is near the middle of the oral tract. For voiceless fricatives,

/f,s/, a steady air flow, or noise, is the exciting source. For voiced fricatives,

/v,z/, which share the same place of constriction with /f,s/ respectively, have

two excitation sources. One excitation source is at the glottis, and the other is

turbulence around the point(s) of maximum constriction. Since the place of the

constriction determines where the energy is trapped, or the pole-zero patterns of

the vocal tract transfer functions, the spectrograms for labials and alveolars will

have energy concentrations at different frequency regions. The waveforms and

spectrograms of /fa, sa, va, za/ by Talker No.1 are shown in Figures 2.16 - 2.23.

The measurements for fricatives and stops are done in a similar manner. These

measurements are the following:

25

• F1, F2, F3 Transition Duration and Frequency These measurements

were similar to those described for the plosives.

• Noise, Aspiration and VOT duration Measurements of these durations

were done manually from the time waveform and the spectrogram using

CoolEdt Pro(Syntrillium Software Corporation).

2.3.2 Noise Measurements

Fricative spectra were obtained with a similar method as the stops. We first split

the speech signal into the fricative and vowel segments. Noise segment ranges

from the beginning of the noise to the onset of the vowel. This can be done

manually from the time waveform and the spectrogram. Short-time spectra were

obtained using a 16 ms window and 256 point Fast Fourier Transform. Figures

2.24- 2.27 are the short-time spectra of the noise (solid line) and the onset vowel

(dotted line) for the CVs /fa, sa, va, za/. In these figures, the first period of the

vowel was analyzed.

26

0 1000 2000 3000 4000 5000 6000 7000 800040

60

80

100

120

140

160

180

frequency

db

/ba/

burstvowel onset

Figure 2.11: Short-time Spectra of /ba/

by Male Talker No.1.The window lengths

was 6 ms for the burst/aspiration and 16

ms for the vowel onset.

0 1000 2000 3000 4000 5000 6000 7000 800040

60

80

100

120

140

160

180

frequency

db

/da/

burstvowel onset

Figure 2.12: Short-time Spectra of /da/

by Male Talker No. 1

0 1000 2000 3000 4000 5000 6000 7000 800040

60

80

100

120

140

160

180

frequency

db

/pa/

burstvowel onset

Figure 2.13: Short-time Spectra of /pa/


0 1000 2000 3000 4000 5000 6000 7000 800060

80

100

120

140

160

180

frequency

db

/ta/

burstvowel onset

Figure 2.14: Short-time Spectra of /ta/


27

Figure 2.15: Scatter plots for voiced and voiceless stops in all vowel contexts

28

1000 2000 3000 4000 5000 6000 7000 8000 9000

−2

−1.5

−1

−0.5

0

0.5

1

1.5

2

x 104

Figure 2.16: Waveform of /fa/ by Male

Talker No. 1.

Time

Fre

quen

cy

500 1000 1500 2000 2500 3000 3500 4000 45000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Figure 2.17: Spectrogram of /fa/ by

Male Talker No. 1.

1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 11000

−2.5

−2

−1.5

−1

−0.5

0

0.5

1

1.5

2

2.5

x 104 waveform for /sa/ by speaker bh

Figure 2.18: Waveform of /sa/ by Male

Talker No. 1.

Time

Fre

quen

cy

0 500 1000 1500 2000 2500 3000 3500 4000 4500 50000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Figure 2.19: Spectrogram of /sa/ by

Male Talker No. 1.

29

1000 2000 3000 4000 5000 6000 7000 8000 9000

−2

−1.5

−1

−0.5

0

0.5

1

1.5

2

x 104

Figure 2.20: Waveform of /va/ by Male

Talker No. 1.

Time

Fre

quen

cy

0 500 1000 1500 2000 2500 3000 3500 4000 45000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Figure 2.21: Spectrogram of /va/ by

Male Talker No. 1.

1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 11000

−1.5

−1

−0.5

0

0.5

1

1.5

2x 10

4

Figure 2.22: Waveform of /za/ by Male

Talker No. 1.

Time

Fre

quen

cy

0 500 1000 1500 2000 2500 3000 3500 4000 4500 50000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Figure 2.23: Spectrogram of /za/ by

Male Talker No. 1.

30

0 1000 2000 3000 4000 5000 6000 7000 800060

80

100

120

140

160

180

frequency

db

fa

noisevowel onset

Figure 2.24: Short-time Spectra of /fa/

by Male Talker No. 1. The window

length was 16 ms for both vowel onset

and noise segment.

0 1000 2000 3000 4000 5000 6000 7000 800040

60

80

100

120

140

160

180

frequency

db

sa

noisevowel onset

Figure 2.25: Short-time Spectra of /sa/

by Male Talker No. 1.

0 1000 2000 3000 4000 5000 6000 7000 800040

60

80

100

120

140

160

180

frequency

db

/va/

noisevowel onset

Figure 2.26: Short-time Spectra of /va/


0 1000 2000 3000 4000 5000 6000 7000 800040

60

80

100

120

140

160

180

200

frequency

db

/za/

noisevowel onset

Figure 2.27: Short-time Spectra of /za/


31

• Av-Anoise: Av is the same as the one measured for the stops. It is the

spectrum amplitude of the vowel at F1 prominence. Anoise here represents

the average amplitude of the entire noise spectrum. This measurement

quantifies the amplitude of the noise relative to the vowel [12].

• Av-Ahi: This measurement is the same as the Av-Ahi for the stops. It is

characterized by the relative amplitude between the higher frequency noise

and the vowel.

• Ahi-A23: Similar to the measurement for the stops, this quantity was

measured to show the spectral tilt of the noise.

• Av4-A45: This is similar to the measurement for stops except now we are

considering the frication noise spectrum instead of the burst spectrum.

• Av4-maxA45 Similar to the stop measurement.

2.3.3 Locus Equations

The locus equation quantities, slope and Y-intercept, are calculated in the same

manner as the stops. Figure 2.28 shows the scatter plots for voiced and voiceless

fricatives.

2.4 Results

2.4.1 Results for the Stops

The measurements in the above section were calculated for all speech samples

used in our perceptual experiments which will be described in a later chapter.

Four tokens from each of the four speakers for each consonant-vowel combination

32

Figure 2.28: Scatter plots for voiced and voiceless fricatives in all vowel contexts

33

/a/ /i/ /u/

/b/ 538.1(49.11) 306.6 (36.87) 345 (41.48)

/d/ 432.1(35.06) 293.5 (40.11) 287 (27.40)

/p/ 878.2(149.2) 236.1 (43.05) 267 (39.09)

/t/ 815.2 (165.0) 262.7 (27.79) 274 (41.4)

Table 2.2: Onset of F1 for stops (in Hz) averaged across the four speakers. Num-

bers in parentheses are the variances.

were measured separately and average values were then obtained.

1. Formant and Time Duration Changes

A complete list of the measurements are shown in Appendices A- B. In

general, the VOT does no show a pattern in the difference between labials

and alveolars. As for the formant information, the difference between labials

and alveolars is not very distinct. For F1, the only pattern is shown in the

onset frequency. Labials have higher onset F1 frequency than alveolars,

except for /piti, putu/. Table 2.2 summarizes the onset frequency of F1

for /a, i, u/, numbers in parentheses represent variances.

For F2, both the onset of F2 and the frequency extent of F2 show distinc-

tions between labials and alveolars. The onset of F2 is lower for labials and

higher for the alveolars, except for /Ci/ context where the onsets are ap-

proximately the same. The values are shown in Table 2.3. One even more

distinct measurement can be drawn from the frequency extent of F2. Fre-

quency extent here means the difference between the steady state F2 and

F2 onset value. The values are shown in Table 2.4. Note that when the

frequency extent is negative, the formant of the syllable is falling instead

34

/a/ /i/ /u/

/b/ 1180 (141.7) 2192 (261.0) 1419 (213.2)

/d/ 1801 (297.5) 2218 (323.1) 1961 (219.0)

/p/ 1165 (141.9) 2417 (265.5) 1453 (219.3)

/t/ 1381 (262.7) 2405 (279.4) 1903 (205.0)

Table 2.3: Onset of F2 for stops(in Hz) averaged for all speakers. Numbers in

parenthesis are the variances.

/a/ /i/ /u/

/b/ 21.39 (87.23) 192.4 (154.6) 11 (121.5)

/d/ -500 (197.5) 154.5 (106.7) -115 (130.0)

/p/ 27.44 (101) 8.251 (64.99) -120 (106.5)

/t/ -166 (217) -2.15 (103.2) -249 (96.2)

Table 2.4: Frequency extent of F2 for stops (in Hz) averaged for all speakers.

Numbers in parenthesis are the variances.

of rising. Transition durations of the formants do not show any distinction

between labials and alveolars.

F3 is not a very distinctive cue for place of articulation for stops.

2. Noise Measurements

Figures 2.29 - 2.34 and Table 2.5 summarize the noise measurements for

stop CVs with the vowels /a,i,u/. Appendix C lists the measurements for

each token and talker. Note that Table 2.5 shows the noise measure differ-

ences, (labials - alveolars). Alveolars have higher relative values than labi-

als for Ahi-A23 and Amid-Avmid. The difference is specially prominent

35

for the mid-frequency relative amplitude, Amid-Avmid. On average, the

alveolars have approximately 20 dB higher values for Amid-Avmid than

those of the labials. Conversely, Av-Ahi, Av-maxA23, Av4-A45 and

Av4-maxA45 are higher for labials than for alveolars. The most promi-

nent quantities are Av-Ahi, Av4-A45 and Av4-maxA45 where labials

are, on average, about 23 dB, 19 dB and 16 dB higher than alveolars re-

spectively. These observations are consistent across all vowels. For Av-

maxA23, labials have higher values than alveolars except for the /pi,ti/

case.

Table 2.6 summarized the average data for labials and alveolars from the

study of Stevens et al. [36]. Their results showed that alveolars have smaller

Av-Ahi and Av-maxA23 than labials but the opposite is true for Ahi-

A23. Measurements from our study showed that Av-Ahi,Ahi-A23 and

Av-maxA23 agree with Stevens et al. [36] on voiced and voiceless stops in

initial position. Note that the burst segment in the studies by Stevens et al.

did not include aspiration in voiceless stops. Av4-A45 and Av4-maxA45

results are also consistent with the perceptual experiments conducted by

Hedrick et al. [14] on synthetic stimuli.

3. Locus Equations For voiced and voiceless stops, alveolars consistently

have a smaller (or flatter) slope values and the Y-intercept for alveolars

are bigger than labials. These results are consistent with the findings by

Sussman et al. in 1991 [37].

To examine the locus equation analysis in different vowel contexts, the slope

and Y-intercept values are also calculated for each /Ca/, /Ci/ and /Cu/

syllable. The results for the Y-intercept, the slope and the R2 values are

summarized in Tables 2.7- 2.9. A bar graph of the slopes and Y-intercept

36

bada pata bidi piti budu putu bd pt Average

Ahi-A23 -1.4 -11.8 -6.3 -17.1 -6.9 -20.5 -4.8 -16.5 -10.7

Av-Ahi 22.8 19.5 27.0 16.0 20.5 29.5 23.4 21.7 22.5

Av-maxA23 21.2 5.7 22.6 -4.3 15.3 9.4 19.7 3.6 11.6

Amid-Avmid -23.9 -17.5 -29.4 -12.9 -18.1 -20.6 -23.8 -17.0 -20.4

Av4-A45 17.1 14.6 20.6 21.7 3.9 16.8 13.9 17.7 15.8

Av4-maxA45 20.8 17.7 22.9 22.5 8.7 23.0 17.5 21.0 19.3

Table 2.5: Spectral quantity difference between /b,d/and /p,t/ for /a,i,u/, aver-

aged values for voiced and voiceless stops and the averaged for all stops.

p b t d

Av-Ahi 41 45 25 31

Ahi-A23 -9 -10 2 -2

Av-maxA23 28 31 25 27

Table 2.6: Average data (in dB) from measurements on bursts in Stevens et al.’s

study. [36]

values are shown in Figure 2.35. The F2 slope measures cue place for

voiced plosives but not for voiceless ones (with the exception of /pa,ta/).

Y-intercept values are higher for the voiceless alveolars in the /Ci/ and

/Cu/ contexts, but these values are higher for the voiced labials in the

/Ca/ and /Ci/ contexts. However, when analyzing CVs across all vowel

contexts, alveolars consistently show smaller (or flatter) slope values and

larger Y-intercepts. Slope and Y-intercept values are summarized in Table

2.10

37

slope Y-int R2

ba 1.332 -420 0.6623

da 1.929 -790 0.7286

pa 1.1125 -161 0.4984

ta 2.2607 -1366.02 0.4617

Table 2.7: Slope, Y-intercept and R2 summary for stops with /a/

slope Y-int R2

bi 0.7439 418.36 0.7363

di 1.0378 -244.25 0.8921

pi 0.9941 6.0036 0.9401

ti 0.898 247.03 0.8748

Table 2.8: Slope, Y-intercept and R2 summary for stops with /i/

2.4.2 Percent Correct Classification for Stops

In order to better quantify the classification of these acoustic cues, a simple non-

parametric classification procedure was performed to obtain the percent correct

classification for each acoustic feature. This method first defines the threshold

to be at the lowest value among all measurements of each acoustic feature. The

threshold values are increased by small increments and the final percent classifi-

cation was defined as the largest percent correct separation between labials and

alveolars for the corresponding threshold value. The results are summarized in

Tables 2.11 - 2.14.

38

slope Y-int R2

bu 0.6778 499.76 0.8722

du 0.7421 594.75 0.7366

pu 0.6976 519.78 0.9409

tu 0.7136 722.54 0.929

Table 2.9: Slope, Y-intercept and R2 summary for stops with /u/

/b/ /d/ /p/ /t/

slope 0.8232 0.524 0.9476 0.8458

Y-intercept 220.62 1030.0 113.04 409.81

Table 2.10: Slope and Y-intercept Values for Plosives when Considering all Vowel

Contexts

F1 onset (threshold) F2 onset (threshold) F3 onset (threshold)

/ba,da/ 93.75% (476.95) 96.88% (1415.40) 75% (2744.00)

/pa,ta/ 65.63% (680.99) 68.75% (1293.50) 62.5% (2745.30)

/bi,di/ 68.75% (288.28) 68.75% (1925.00) 68.75% (3018.90)

/pi,ti/ 68.75% (264.06) 56.25% (2081.50) 62.5% (3174.20)

/bu,du/ 81.25% (318.36) 87.5% (1467.60) 87.5% (2308.60)

/pu,tu/ 59.38% (315.88) 87.5% (1523.40) 65.63% (2215.60)

/b,d/ 64.58% (326.30) 78.13% (1395.70) 67.71% (2395.40)

/p,t/ 56.25% (688.15) 67.71% (1487.20) 55.21% (2768.00)

Table 2.11: Percent Correct Classification using F1, F2, or F3 onset values for

stop CVs in each vowel context and across vowel contexts. Numbers in paren-

theses are the threshold values in Hz.

39

F1∆ (threshold) F2∆ (threshold) F3∆ (threshold)

/ba,da/ 78.13% (240.935) 100% (-313.9) 96.88% (-34.2663)

/pa,ta/ 75% (171.5143) 78.13% (-102.73) 62.5% (156.3873)

/bi,di/ 65.63% (9.9473) 59.38% (272.7637) 75% (167.48)

/pi,ti/ 65.63% (52.7283) 65.63% (-51.202) 71.88% (-1.8733)

/bu,du/ 78.13% (28.124) 75% (-137.583) 90.63% (-80.8567)

/pu,tu/ 65.63% (125.2607) 78.13% (-241.4067) 75% (-133.594)

/b,d/ 61.46% (3.5413) 73.96% (-117.575) 78.13% (-45.31)

/p,t/ 61.46% (119.535) 66.67% (-189.84) 56.25% (-118.901)

Table 2.12: Percent Correct Classification using F1∆, F2∆, or F3∆ onset values

for stop CVs for each vowel context and across vowel contexts. A negative fre-

quency extent value indicates that the formant is decreasing in time (steady state

formant frequency is smaller than formant onset frequency.) Numbers in boldface

represent the highest percent correct classification for place of articulation using

a particular feature. Numbers in parentheses are the threshold values in Hz.

40

Ahi-A23 Av-Ahi Av-maxA23 Burst Duration

(threshold) (threshold) (threshold) (threshold)

/ba,da/ 59.38% (-5.98) 90.63% (75.09) 81.25% (64.01) 81.25% (4.43)

/pa,ta/ 90.63% (-2.41) 84.38% (59.54) 65.63% (41.25) 96.88% (13.69)

/bi,di/ 75% (1.76) 93.75% (76.53) 78.13% (56.68) 84.38% (5.2)

/pi,ti/ 96.88% (5.06) 81.25% (45.75) 71.88% (39.78) 81.25% (4.77)

/bu,du/ 78.13% (1.03) 90.63% (78.95) 71.88% (63.47) 59.38% (5.09)

/pu,tu/ 93.75% (1.61) 100% (59.79) 75% (55.65) 68.75% (5.23)

/b,d/ 67.71% (1.0345) 89.58% (74.82) 73.96% (61.57) 71.88% (5.13)

/p,t/ 88.54% (-2.55) 86.46% (59.54) 60.42% (59.34) 68.75% (8.2)

Table 2.13: Percent Correct Classification using noise measurements Ahi-A23,

Av-Ahi, Av-maxA23 or Burst duration for stop CVs in each vowel context and

across vowels. Numbers in boldface represent the highest percent correct classifi-

cation for place of articulation using a particular feature. Numbers in parentheses

are threshold values in dB for the first three columns and ms for the last one.

41

Figure 2.29: Ahi-A23 for Stops. The small rectangles, and numbers next to them,

represent average values. The minimum and maximum numbers are shown by

the lines.

42

Figure 2.30: Av-maxA23 for Stops

43

Figure 2.31: Av-Ahi for Stops

44

Figure 2.32: Amid-Avmid for Stops

45

Figure 2.33: Av4-A45 for Stops

46

Figure 2.34: Av4-maxA45 for Stops

47

Figure 2.35: Slope and Y-intercept Values for Stops

48

Amid-Avmid Av4-A45 Av4-maxA45

(threshold) (threshold) (threshold)

/ba,da/ 78.13% (3.6133) 84.38% (21.7956) 84.38% (18.188)

/pa,ta/ 75% (13.1337) 71.88% (-1.2537) 75% (-7.9619)

/bi,di/ 84.38% (-10.7814) 68.75% (1.3268) 75% (10.2526)

/pi,ti/ 75% (13.7131) 78.13% (3.9696) 81.25% (-10.4291)

/bu,du/ 78.13% (-4.9194) 65.63% (39.5528) 68.75% (38.3827)

/pu,tu/ 87.5% (18.2035) 84.38% (8.9663) 90.63% (-0.7965)

/b,d/ 76.04% (-10.7814) 64.58% (15.0479) 71.88% (12.0115)

/p,t/ 77.08% (13.1337) 79.17% (5.8615) 79.17% (-8.0359)

Table 2.14: Percent Correct Classification using noise measurements

Amid-Avmid, Av4-A45 or Av4-maxA45 for stop CVs in each vowel context and

across vowel contexts. Numbers in parentheses are threshold values in dB.

Numbers in bold face represent the highest percent correct classification for

place of articulation for that pair of consonants. For example, in the /ba,da/

case, F2 frequency extent (F2∆) values separate well (100% correct classification)

between labials and alveolar place of articulation, but not in the /bi,di/ case

(59%).

Formant frequency measures (especially F1 and F2 onset values, F2∆ and

F3∆) cue place for several /Ca/ and /Cu/ syllables but not for /Ci/ syllables.

Relative noise spectral measurements seem to provide better discrimination cues

for place of articulation than formants. For example, the Av-Ahi measures re-

sulted in more than 81% correct place classification for both voiced and voiceless

consonants and for both vowel-dependent and vowel-independent measures. VOT

(not shown in the tables) does not cue place for plosives. Burst duration signals

49

place of articulation for plosives in /Ca/ and /Ci/ syllables (above 81% correct

classification) but not for /Cu/ syllables. Measurements that showed 100% clas-

sification for place include F2∆ for /ba,da/ and Av-Ahi for /pu,tu/. In addition,

burst duration and Ahi-A23 showed 97% correct classification for /pa,ta/ and

/pi,ti/ respectively.

When classification was performed across all vowel contexts (ie. using mea-

surements of CVs from all three vowel contexts), Av-Ahi separated out /b,d/

syllables at 90% and Ahi-A23 separated out /p,t/ at 89%.

2.4.3 Results for the Fricatives

The above measurements are also calculated for all 16 tokens for each consonant-

vowel sample for all fricatives in the three vowel contexts.

1. Formant and Duration Measurements Noise duration for fricatives

appears to be a cue for place. The duration of noise is about 40 ms longer

for alveolars than for labiodentals. The durations, in millisecond, are sum-

marized in Table 2.15. Numbers in parentheses represent corresponding

variances. Detailed measured values for formants and durations are listed

in Appendix A and Appendix B. F1 information does not cue place for

fricatives. The onset of F2 is approximately 300-400 Hz higher for alveolars

than for labiodentals for the /Ca/ and /Cu/ syllables. For the /Ci/ syl-

lables, alveolars had, on average, an onset F2 that was approximately 100

Hz lower than labiodentals. As for the F2 frequency extent, labiodentals

have less frequency extent than alveolars. The difference in F2 frequency

extent is most prominent for /Ca/ syllables and least for the /Ci/ syllables.

Summaries of the F2 onset frequency and the frequency extent are shown

in Tables 2.16 and 2.17. Once again, negative frequency extent values

50

/a/ /i/ /u/

/f/ 141.4 (44.5) 168.8 (24.15) 176.7 (29.11)

/s/ 175.4 (14.04) 202.0 (22.54) 217.7 (21.78)

/v/ 113.95 (23.20) 132.3 (21.02) 142.6 (29.25)

/z/ 135.5 (16.40) 170.7 (21.40) 173.8 (26.70)

Table 2.15: Noise Duration for Fricatives (in ms). Numbers in parenthesis are

the variances.

/a/ /i/ /u/

/f/ 1104 (120.8) 2216 (304.6) 1426 (203)

/s/ 1468 (196.3) 2096 (251.8) 1802 (196)

/v/ 1214 (169.1) 2084 (148.3) 1428 (221)

/z/ 1597 (209.9) 1968 (98.14) 1799 (230)

Table 2.16: Onset of F2 for Fricatives (in Hz). Numbers in parenthesis are the

variances.

mean that F2 is falling instead of rising into the vowel .

F3 information for fricatives show some difference between labiodentals and

alveolars. F3 onset frequency values are slightly higher for alveolars than for

labiodentals. As for the frequency extent, alveolars have shorter frequency

extent than labiodentals for /Ca/ and /Ci/ syllables and the opposite is true

for /Cu/ syllables. Tables 2.18 and 2.19 summarize the measurements.

Durational measurements of the formant frequency transitions do not cue

place of articulation for fricatives.

2. Noise Measurements The differences between labiodental and alveolar

51

/a/ /i/ /u/

/f/ 45.02 (89.68) 185(163.7) -33.4 (93.87)

/s/ -249 (191.3) 289 (93.52) -138 (142.8)

/v/ -22.1 (121.7) 185 (108.4) -33.4 (208.8)

/z/ -343 (117) 384 (88.78) -61.1 (87.21)

Table 2.17: Frequency Extent of F2 for Fricatives (in Hz). Numbers in parenthesis

are the variances.

/a/ /i/ /u/

/f/ 2363 (200.6) 2746 (273.3) 2425 (430)

/s/ 2678 (286.5) 2778 (291.4) 2679 (302)

/v/ 2444 (208.2) 2560 (114.2) 2344 (326)

/z/ 2728 (102.2) 2705 (72.09) 2663 (331)

Table 2.18: Onset of F3 for Fricatives (in Hz). Numbers in parenthesis are the

variances.

/a/ /i/ /u/

/f/ 269.7 (239) 273.6(111.8) 18.5 (236)

/s/ 5.176 (37.58) 193.9 (109.8) -230 (142.8)

/v/ 179 (207.5) 443.1 (102.2) 74.4 (208.8)

/z/ -29.7 (99.02) 213.5 (63.38) -145 (87.21)

Table 2.19: Frequency Extent of F3 for Fricatives (in Hz). Numbers in parenthesis

are the variances.

52

syllables are summarized in Table 2.20. With the exception of Ahi-A23,

labiodental syllables have higher noise measures for Av-Ahi, Av-Anoise,

Av4-A45 and Av4-maxA45. The most significant distinction occurs

at Av4-maxA45 and the difference is about 30 dB. Av4-A45 and Av-

Anoise also have differences of 20 dB and 17 dB, respectively. Labiodentals

have slightly higher values for the high-frequency amplitude, Av-Ahi, and

the difference is about 3.5 dB. The spectral tilt, Ahi-A23, is higher for alve-

olars than for labiodentals but the difference is not very significant(about 2

dB). Figures 2.36 - 2.40 summarize the noise measurements for fricatives.

fasa vaza fisi vizi fusu vuzu fs vz average

Ahi-A23 -1.1 -1.6 -0.02 -2.2 -2.8 -3.1 -1.3 -2.1 -1.8

Av-Ahi 5.4 4.3 0.1 2.9 4.8 4.5 3.5 3.9 3.7

Av-Anoise 18.9 17.0 12.9 15.1 18.4 18.1 16.7 16.8 16.8

Av4-A45 7.9 12.7 35.2 30.8 18.8 12.1 20.6 18.5 19.6

Av4-maxA45 20.5 20.4 43.7 37.3 33.0 25.1 32.4 27.6 30.0

Table 2.20: Spectral quantity difference (in dB) between /f,s/ and /v,z/ for

/a,i,u/, averaged values for voiced and voiceless fricatives and the averaged for

all fricatives.

3. Locus Equations

As expected, when locus equation analysis is performed separately for dif-

ferent vowel contexts, the results are less distinctive. The results for the

Y-intercept, the slope and the R2 values when separating vowel contexts

are summarized in Tables 2.21- 2.23. The bar graphs are also shown for

comparison in Figure 2.41.

53

Fricatives do not show a consistent pattern in the slope values when consid-

ering individual vowel contexts, and the Y-intercept values are higher for

the alveolars except for /fa,sa/ and /fi,si/ cases. Similar to the plosive re-

sults for locus equations, when considering all vowels contexts, the results

are consistent through out voicing and vowel contexts. Alveolars consis-

tently have smaller (or flatter) slope values and larger Y-intercept values.

Slope and Y-intercept values for fricatives are summarized in Table 2.24.

slope Y-int R2

fa 0.8924 78.55 0.4555

sa 1.6152 -500.34 0.46

sa 1.3604 -407.58 0.5184

za 1.3821 -135.56 0.6982

Table 2.21: Slope, Y-intercept and R2 values for fricatives with /a/

slope Y-int R2

fi 0.9205 5.8230 0.9180

si 0.9005 -52.028 0.8944

vi 0.8156 147.21 0.7496

zi 0.7711 153.61 0.9454

Table 2.22: Slope, Y-intercept and R2 values for fricatives with /i/

2.4.4 Percent Correct Classification for Fricatives

A similar method was applied to fricatives to obtain percent correct classification

for each feature. The results are summarized in Tables 2.26 - 2.28.

54

slope Y-int R2

fu 0.7345 402.97 0.9042

su 0.6124 782.85 0.8012

vu 0.6081 572.54 0.7597

zu 0.6733 628.90 0.7101

Table 2.23: Slope, Y-intercept and R2 values for fricatives with /u/

/v/ /z/ /f/ /s/

slope 0.7218 0.4315 0.8559 0.5890

Y-intercept 378.75 1008.5 171.92 754.34

Table 2.24: Slope and Y-intercept Values for fricatives when considering all vowel

contexts

Formant frequency measures, especially F1, F2 onset frequencies, F2∆ and

F3∆, classify place of articulation for /Ca/ syllables and to some extent /Cu/

syllables. However, formant information does not cue /Ci/ syllables. F1 onset

frequencies can classify /fa,sa/ and /va,za/ syllables at 100%. Relative noise

measurements provide better discrimination cues for place of articulation than

formant measurements. The most distinctive feature is Av-Anoise. Av-Anoise

resulted in more than 84% classification for CV syllables across all three vowel

contexts. It resulted in 100% classification for /fu,su/ syllables and 97% for

/fa,sa/, /vi,zi/ and /vu,zu/ syllables. Noise duration does signal some place in-

formation for fricatives in all vowel contexts (above 75 %). When classification

was performed on all CVs, (ie. using measurements for CVs of all vowel contexts),

Av-Anoise achieved 90% correct percent classification for both voiced and voice-

less fricative syllables; Av4-maxA45 also resulted in more than 81% classification

55

F1 onset (threshold) F2 onset (threshold) F3 onset (threshold)

/fa,sa/ 100% (485) 84.38% (1200.00) 81.25% (2352.30)

/va,za/ 100% (430) 84.38% (1301.20) 78.13% (2468.70)

/fi,si/ 59.38% (264.58) 68.75% (1836.70) 62.5% (2968.40)

/vi,zi/ 59.38% (310.15) 68.75% (1744.30) 65.63% (2353.50)

/fu,su/ 62.5% (331.25) 84.38% (1380.50) 75% (2274.20)

/vu,zu/ 68.75% (339.85) 81.25% (1484.40) 75% (2298.80)

/f,s/ 66.67% (86.32) 67.71% (1415.10) 69.79% (2384.10)

/v,z/ 67.71% (189.45) 71.88% (1474.00) 66.67% (2343.80)

Table 2.25: Percent Correct Classification using F1, F2, or F3 onset values for

fricative CVs in each vowel context and across vowel contexts. Numbers in bold-

face represent the highest percent correct classification for place of articulation

using a particular feature. Numbers in parentheses are threshold values in Hz.

across all vowel contexts.

2.5 Summary

In general, the noise measurements are reliable cues for place of articulation for

both stops and fricatives. For stops, labials have higher Av-Ahi(by 23 dB),

Av5-A45 (by 19 dB) and Av4-maxA45 (by 16 dB), than the alveolars, but

have smaller Amid-Avmid (by 20 dB) and Ahi-A23 (by about 4 dB for voiced

stops and 14 dB for voiceless stops). These results agree with the study by

Stevens et. al in 1999 [36] that Av-Ahi is larger for labials than for alveolars but

Ahi-A23 is smaller for labials than for alveolars. Noise measures Av-Ahi, Av-

Anoise, Av4-A45 and Av4-maxA45, for fricatives are higher for labiodentals

56

F1∆ (threshold) F2∆ (threshold) F3∆ (threshold)

/fa,sa/ 78.13% (205.47) 93.75% (-54.1147) 96.88% (54.4807)

/va,za/ 71.88% (255.4173) 93.75% (-245.6773) 84.38% (34.216)

/fi,si/ 59.38% (6.874) 75% (299.5827) 68.75% (120.2573)

/vi,zi/ 68.75% (24.532) 71.88% (247.344) 87.5% (286.2957)

/fu,su/ 62.5% (11.723) 71.88% (-240.474) 78.13% (-90.4677)

/vu,zu/ 59.38% (124.2167) 59.38% (-172.0847) 84.38% (-26.0727)

/f,s/ 60.42% (213.1757) 67.71% (-89.274) 72.92% (74.0603)

/v,z/ 57.29% (259.4813) 67.71% (-190.1) 69.79% (13.5367)

Table 2.26: Percent Correct Classification using F1∆, F2∆, or F3∆ onset values

for fricative CVs in each vowel context and across vowel contexts. Numbers in

parentheses are threshold values in Hz.

than for alveolars. The differences are, on average, 4 dB, 17 dB, 20 dB, and

30 dB, respectively. Percent correct classification showed that Av-Anoise (for

fricatives) and Av-Ahi (for plosives) result in more than 81% correct place clas-

sification for both voiced and voiceless consonants and for both vowel-dependent

and vowel-independent measures. Results showed that Av-Ahi and Av-Anoise

classify /pu,tu/ and /fu,su/ at 100% respectively.

Burst duration signals place of articulation for /Ca/ and /Ci/ stop syllables

(above 81% correct classification) but not for /Cu/ syllables; noise duration sig-

nals place for fricatives at above 75% in all vowel contexts. For fricatives, the

duration of the noise segments are approximately 40 ms longer for alveolars than

for labiodentals but do not show a pattern for the stops.

As for the formant frequency information, F2 is the most distinct measure

to cue place for both stops and fricatives. Both the onset F2 frequency and the

57

Ahi-A23 Av-Ahi Noise Duration


/fa,sa/ 65.63% (23.9981) 75% (62.6417) 75% (155.63)

/va,za/ 65.63% (23.4571) 65.63% (62.4738) 75% (112.35)

/fi,si/ 59.38% (21.4034) 59.38% (59.4979) 78.13% (210.9)

/vi,zi/ 62.5% (22.9824) 62.5% (53.8694) 90.63% (157.18)

/fu,su/ 71.88% (26.0952) 71.88% (57.5058) 75% (196.5667)

/vu,zu/ 65.63% (23.7072) 68.75% (60.8285) 75% (144.0667)

/f,s/ 59.38% (24.3373) 66.67% (60.4754) 67.71% (176.44)

/v,z/ 64.58% (23.4352) 63.54% (62.5444) 69.79% (136.44)

Table 2.27: Percent Correct Classification using noise measurements Ahi-A23,

Av-Ahi, or noise duration for fricative CVs in each vowel context and across

vowel contexts. Numbers in parentheses are threshold values in dB for the first

two columns and ms for the last one.

58

Av-Anoise Av4-A45 Av4-maxA45


/fa,sa/ 96.88% (71.5548) 75% (11.8661) 78.13% (-5.7108)

/va,za/ 90.63% (78.08) 71.88% (27.0353) 78.13% (27.4373)

/fi,si/ 84.38% (65.3148) 78.13% (-5.0862) 93.75% (-6.5434)

/vi,zi/ 96.88% (73.4701) 78.13% (-0.9298) 81.25% (-10.2633)

/fu,su/ 100% (70.1994) 84.38% (11.6187) 87.5% (-7.6032)

/vu,zu/ 96.88% (70.857) 68.75% (30.3214) 87.5% (11.4284)

/f,s/ 89.58% (70.172) 76.04% (12.4145) 85.42% (-6.5434)

/v,z/ 89.58% (73.5499) 70.83% (28.2204) 81.25% (17.1004)

Table 2.28: Percent Correct Classification using noise measurements Av-Anoise,

Av4-A45, or Av4-maxA45 for fricative CVs for each vowel context and across

vowel contexts. Numbers in parentheses are threshold values in dB.

59

F2 frequency transition extent are cues for place of articulation, but the results

vary with vowel contexts. For /Ca/ and /Cu/ syllables, F2 onset frequencies

for alveolars are about 200-400 Hz higher than labials. F2 frequency extents for

alveolar plosives are longer, and falling into the vowel, compared to the short

and rising F2 transition for labial plosives. These results agree with the studies

by Potter et al. [27] that labials have lower formant frequencies, or the “hub”,

than alveolars. For fricatives, the frequency extent of F2 both rise into the vowel

for labiodentals and alveolars and are shorter for alveolars than labiodentals.

F2 differences are usually most prominent for /Ca/ syllables and least for /Ci/

syllables. Percent correct classification show that F1 onset frequencies classify

/fa,sa/ and /va,za/ at 100% and F2∆ classifies /ba,da/ at 100%. Durational

measurements for formant information, on the other hand, are not reliable cues

for place of articulation for both stops and fricatives. Table 2.29 summarizes the

features and percentages that achieved highest percent correct classification for

each syllable.

The locus equation information, the slopes and the Y-intercept, are distinct cues

for place of articulation for both plosives and fricatives only when the analysis

is performed for all CVs in all vowel contexts. For both stops and fricatives,

alveolars have smaller (or flatter) slopes than labials and have higher Y-intercept

values. These results agree with the study done by Sussman et al. in 1991 [37]

that labials have bigger slopes and smaller Y-intercepts than alveolars for voiced

stops CVs. When performing the locus equation analysis on /Ca/,/Ci/ and /Cu/

syllables separately, the results are inconsistent and no distinctive patterns are

found.

60

/ba,da/ /bi,di/ /bu,du/ /b,d/

features F2∆ Av-Ahi Av-Ahi, F3∆ Av-Ahi

100% 93.75% 90.63% 89.58%

/pa,ta/ /pi,ti/ /pu,tu/ /p,t/

features Burst Duration Ahi-A23 Av-Ahi Ahi-A23

96.88% 96.88% 100% 88.54%

/va,za/ /vi,zi/ /vu,zu/ /v,z/

features F1 onset Av-Anoise Av-Anoise Av-Anoise

100% 96.88% 96.88% 89.58%

/fa,sa/ /fi,si/ /fu,su/ /f,s/

features F1 onset Av4-maxA45 Av-Anoise Av-Anoise

100% 93.75% 100% 89.58%

Table 2.29: Summary for the highest percent correct classification for each sylla-

ble.

61

Figure 2.36: Ahi-A23 for Fricatives. The small rectangles, and the number next

to them, are the average values. The minimum and maximum numbers are shown

by the lines.

62

Figure 2.37: Av-Ahi for Fricatives

63

Figure 2.38: Av-Anoise for Fricatives

64

Figure 2.39: Av4-A45 for Fricatives

65

Figure 2.40: Av4-maxA45 for Fricatives

66

Figure 2.41: Slopes and Y-intercepts Comparisons for Fricatives

67

CHAPTER 3

Perceptual Experiments

3.1 Introduction

The goal of these experiments is to discover which acoustic cues for place of

articulation are most perceptually robust in noise. The stimuli are Consonant

Vowel (CV) syllables which only differ by the place feature. Such pairs include the

voiced stops, /b, d/, voiceless stops, /p, t/, voiced fricatives,/v, z/ and voiceless

fricatives, /f, s/. Each syllable is presented in three vowel contexts, /a/,/i/ and

/u/, and in different levels of white noise. For each syllable, a threshold is found.

The threshold here is defined by the SNR (Signal to Noise Ratio) at which percent

correct identification is 79%.

3.2 Methods

3.2.1 Subjects

Four subjects, two males and two females, participated in the forced choice per-

ceptual experiments. Each subject was tested to have normal hearing (was able

to hear a +10 dB Sound Pressure Level (SPL) sound at frequencies from 250Hz -

8kHz). All subjects were native talkers of American English and their age ranged

between 18 - 36 years old. Each subject was paid $10 /hr. None of the subjects

68

was experienced with perceptual experiments but all were given an one-hour

training session before the experiments started.

3.2.2 CV syllables

The CVs are part of the UCLA Speech Processing and Auditory Perception

Laboratory speech database. The sound samples were recorded from four native

American English talkers, two males and two females. Each token was normalized

so that the maximum energy of the syllable segment is the same for all tokens.

The signals were played at 60 dB SPL for 160 ms. There are four tokens per

speaker per syllable. Each token was played twice, thus each CV was played 32

times.

3.2.3 Masking Noise

The masker used in the experiments was white noise. White noise has a flat

spectrum and is un-correlated in time. The signal level in the experiments was

fixed and the SPL of the white noise was adjusted to result in different SNRs.

3.2.4 Experiment Setup

Listening subjects were brought into a sound-proof room located at the UCLA

Speech Processing and Auditory Perception Laboratory. The stimuli were pre-

sented via Telephonics TDH49P headphones. The tokens were generated by

computer software to digital bits and passed through an Ariel ProPort 656 board

Digital-to-Analog converter. The analog signals were then amplified by a Sony

59ES DAT recorder. The analog signals were played through the headphones

which were connected to the amplifier. The entire system was calibrated within

69

0.5 dB using the Larson Davis 800B Sound Level Meter before each experiment.

3.2.5 Experimental Protocol

Each experiment lasted about one hour. Subjects were given a short break if

the experiment was longer than one hour. The experiment was a 2AFC (Two

Alternate Forced Choice) format. Each experiment consisted of one pair of CVs

that only differed by place of articulation, ie. /ba,da/, /fi,si/ or /vu,zu/, and

was divided into seven smaller sections. Each section of an experiment involved

stimuli at a different SNR. Noise levels were set at -15, -10, -5, 0, 5, 10 dB SNR

and one section was without noise (e.g., CVs in quiet). Subjects were played one

CV at a time and prompted to choose between one of the two consonants. The

same experiment was performed for all CV combinations. An error message was

displayed and no sound was played if the total SPL was more than 92 dB SPL.

A confusion matrix was generated for each CV and each SNR to summarize the

subjects’ responses. From the confusion matrix, we were able to determine the

number of correct responses for each CV. Using these data, we can determine the

identification thresholds of each CV pair.

The confusion matrix looks like the following:

• The confusion matrix for: A w -5.0

B D

29 3

2 30

where A stands for vowel context /a/, w stands for white noise, and -5.0 is

the SNR in dB for this particular section of one experiment. Appendix D

lists the confusion matrices from each subject for each syllable.

70

3.3 Threshold Definition

The threshold was defined by the SNR level (in dB) at which the number of

correct responses was at 79 %. Since this was a two forced-choice experiment,

bias responses were likely to occur. Bias represents the tendency for the subject

to choose one CV over the other. For example, one subject may tend to hear /ba/

more than /da/, thus having a higher number of correct responses for /b/. The

results from Tables 3.1 and 3.2 showed that bias occurred in some experiments.

For example, /pa/ had a threshold of -0.0042 dB SPL but /ta/’s threshold was

6.85 dB SPL; results from /pa,ta/ showed that the subjects were biased toward

/pa/ whenever a confusion occurred. In order to solve this problem, the threshold

for one set of CVs is determined from the average percent correct of the two CVs

in each set. For example, the percent correct for the set /ba,da/ will be the

average percent correct from /ba/ and /da/. In this experiment, a sigmoidal fit

is performed on the percent correct data to obtain a more accurate threshold

value.

3.3.1 Sigmoid

A simple parameter fit is necessary to obtain a more accurate threshold SNR.

One such fitting function is the sigmoid. The equation for the sigmoidal fits is

shown below:

y = bottom + (top− bottom)

1

2− 1

2

1− e(

x−ba )

1 + e(x−b

a )

where top and bottom are the maximum and minimum values for the percent

correct. In our case, we set these values to 30 and 100, respectively. a and b

are parameters to adjust the slope and position of the transition of the sigmoid

function between the top and bottom flat areas. These values are determined

71

−20 −15 −10 −5 0 5 10 1540

50

60

70

80

90

100

110

SNR (dB)

Per

cent

Cor

rect

(%

)

Figure 3.1: Percent correct identification data for /ba,da/ and the sigmoidal fit

curve

through an automatic iterative process. The 79% threshold is calculated from

the sigmoidal fitting function.

From the data in the confusion matrices, we are able to plot the percent

correct with respect to SNR along with the model fit. An example of such fitting

for /ba,da/ is plotted in Fig 3.1. The solid line is the sigmoidal fit to the actual

percent identification (in circles). The error bars represent the minimum and

maximum numbers among the four listening subjects.

3.4 Experimental Results

As expected, the resulting data showed a decrease in percent correct with decreas-

ing SNR. Percent correct ranged from 100% for clean tokens to about 40% at the

72

/b/ /d/ average /p/ /t/ average

/a/ -7.301 -7.321 -7.3050 -0.0042 6.850 6.7458

/i/ 3.489 3.9324 4.1592 1.416 -2.218 0.1254

/u/ -1.189 -0.512 -1.6404 0.088 -0.247 -0.0096

Table 3.1: Thresholds from each stop syllable and the thresholds from averaged

correct responses (in dB SPL)

/f/ /s/ average /v/ /z/ average

/a/ -7.700 -3.763 -5.1342 -7.040 -1.792 -4.5402

/i/ -4.190 -4.330 -3.7842 -1.457 -1.273 -1.1976

/u/ -7.219 -4.245 -5.0100 -1.943 -3.838 -3.4170

Table 3.2: Thresholds from each fricative syllable and the threshold from averaged

correct responses (in dB SPL)

lowest SNR (-15dB). Results varied from subject to subject and were largely de-

pendent upon the CV pair, including the consonant and vowel contexts. Thresh-

olds were determined by fitting the experimental data with a sigmoid function

and finding the SNR where the sigmoidal fitting curve crossed 79%. Using this

method, we can find the threshold values for the 12 sets of CVs for each of the

four listening subjects. The four subjects are referred to as Subject I, II, III and

IV.

Tables 3.1- 3.2 summarized the threshold SNR (in dB) for each syllable

and the threshold value from the average correct response of the corresponding

CV pair (sigmoidal fit was performed on the ratio between the sum of correct

responses to both syllables in each CV pair to the total number of tokens played

for both syllables). The thresholds in these tables were taken from the average

correct responses from all four listeners.

73

As mentioned in Section 3.3, biasing is a factor for any two forced-choice

experiment, thus the average correct responses from each CV pair are used and

analyzed for the purpose of this study to examine the confusion of place of artic-

ulation in noise.

Figures 3.2 and 3.3 summarized the results for the stops and fricatives,

respectively. Data were averaged across the four listening subjects with minimum

and maximum values shown by error bars. The solid lines were the sigmoidal fit

for the average data (in circles) and the dotted lines represented threshold of 79%

correct. Threshold values were summarized in Tables 3.3, 3.4 and 3.5. The

perceptual experiment thresholds were summarized in Figure 3.4.

3.4.1 Vowel Context Comparison

Threshold values across vowel context are summarized in Figure 3.5. In general,

the context /i/ was the hardest to distinguish than /a/ or /u/, except for the /p,

t/ case. It also seemed that the vowel context /a/ was the most robust context

(lowest threshold) across all consonant pairs except for /p,t/. We will attempt

to explore the reasons behind these results in a later chapter.

3.4.2 Effect of the Feature ‘Manner’

A summary of the thresholds grouped by manner is shown in Figures 3.6 and 3.7

for voiced and voiceless CVs, respectively. As described in the previous chapter,

stops are characterized by their bursts and aspiration noise while fricatives are

characterized by their frication noise.

Figure 3.6 is a comparison of manner for voiced CVs, and Figure 3.7 is

for voiceless CVs. From the results, we can see that the perception of place

74

Average Subject I Subject II Subject III Subject IV

/ba,da/ -7.3050 -8.9088 -6.4302 -7.7748 -4.9074

/pa,ta/ 6.7458 3.7974 5.8926 3.9702 8.1012

/fa,sa/ -5.1342 -4.7292 -5.7606 -5.5716 -4.3404

/va,za/ -4.5402 -4.0164 -6.7326 -6.2790 -2.8230

Table 3.3: Thresholds for /Ca/ syllables in dB SPL


/bi,di/ 4.1592 2.8146 4.8342 3.4788 4.0404

/pi,ti/ 0.1254 2.8308 0.9948 -2.9040 -1.3272

/fi,si/ -3.7842 -3.0930 -4.3998 -5.8416 -2.4882

/vi,zi/ -1.1976 -1.6728 -6.6300 -1.9698 -1.0410

Table 3.4: Thresholds for /Ci/ syllables in dB SPL


/bu,du/ -1.6404 -3.6654 -2.1696 -0.0420 -2.3370

/pu,tu/ -0.0096 0.0444 -2.5584 1.2540 1.1460

/fu,su/ -5.0100 -6.4626 -7.7724 -5.0802 -4.0596

/vu,zu/ -3.4170 -4.5672 -2.2020 -3.5034 -1.8078

Table 3.5: Thresholds for /Cu/ syllables in dB SPL

75

−20 0 2020

40

60

80

100

SNR

% c

orre

ct

/ba,da/

−20 0 2020

40

60

80

100

SNR

% c

orre

ct

/bi,di/

−20 0 2020

40

60

80

100

SNR

% c

orre

ct

/bu,du/

−20 0 2020

40

60

80

100

SNR

% c

orre

ct

/pa,ta/

−20 0 2020

40

60

80

100

SNR

% c

orre

ct

/pi,ti/

−20 0 2020

40

60

80

100

SNR%

cor

rect

/pu,tu/

Figure 3.2: Percent correct identification for stops. The dotted line in each plot

represent the 79% correct point.

for stops and fricatives depends on whether the consonant is voiced or voiceless.

For voiced CVs, we observe that stops are harder to distinguish for /i/ and

/u/ but the opposite is true for /a/. The voiceless pairs, however, have a clear

trend. Fricatives consistently have a lower threshold than the stops. Differences

in threshold are approximately 11.8 dB, 3.8 dB, and 5 dB for /a/, /i/ and /u/,

respectively.

76

−20 0 2020

40

60

80

100

SNR

% c

orre

ct

/fa,sa/

−20 0 2020

40

60

80

100

SNR

% c

orre

ct

/fi,si/

−20 0 2020

40

60

80

100

SNR

% c

orre

ct

/fu,su/

−20 0 2020

40

60

80

100

SNR

% c

orre

ct

/va,za/

−20 0 2020

40

60

80

100

SNR

% c

orre

ct

/vi,zi/

−20 0 2020

40

60

80

100

SNR%

cor

rect

/vu,zu/

Figure 3.3: Percent correct identification for fricatives

3.4.3 Voicing Comparison

The major distinction between a voiced syllable and a voiceless syllable is the

excitation source. Summary of the thresholds grouped by voicing is shown in

Figures 3.8 and 3.9.

For stop CVs, voiced pairs have lower thresholds than the voiceless except for

the /i/ case; this difference was greatest for the /a/ context. A 14 dB difference

is observed between /ba,da/ and /pa,ta/. However, /bi,di/ seems to be harder

to distinguish than /pi,ti/ since the threshold for /bi,di/ is higher. As for the

77

Figure 3.4: Summary of the perceptual thresholds

Figure 3.5: Thresholds grouped by vowel

78

Figure 3.6: Voiced CVs grouped by manner

Figure 3.7: Voiceless CVs grouped by manner

79

Figure 3.8: Stop CVs grouped by voicing

Figure 3.9: Fricative CVs grouped by voicing

80

fricatives, there is a clear trend across all three vowel contexts: voiceless fricatives,

/f,s/ perform better than their voiced counterparts /v,z/ by about 1.5 dB on

average.

3.5 Summary and Conclusion

From the data described in this chapter, we can conclude that the perception of

place of articulation in noise is dependent upon the vowel context, voicing and

manner. In general, fricatives are more robust than stops. For voiced consonants,

/Ci/ and /Cu/ fricatives are more robust than stops but the opposite is true for

/Ca/ syllables. As for voiceless consonants, fricatives are more robust than stops

in all vowel contexts. When comparing the voicing feature in CVs, voiceless

CVs are typically more robust than voiced CVs, especially for the fricatives.

Voiceless fricatives performed better than their voiced counterparts in all three

vowel contexts by about 1.5 dB on average. Voicing in stops does not show a

distinct pattern. Voiced stops /ba,da/ are much more robust than voiceless stops

/pa,ta/ (14 dB difference) but /bi,di/ is less robust than /pi,ti/. This agrees

with Hant’s [10] study that /bi,di/ is the least robust while /ba,ga/ is the most

robust. /Cu/ syllables for voiced stops is slightly more robust than voiceless /Cu/

syllables (by 1dB). As for the vowel context, /a/ is more robust than /u/, which

is more robust than /i/ except for the /p,t/ case.

Overall, we observed that the syllable pair /ba,da/ achieved the best percep-

tion in noise. The reason behind this may be that the distinction between the

formant information of /ba/ and /da/ are most prominent. F2 ∆ showed that

/ba/ has a rising F2 while /da/ has a falling one. In addition, the onset frequen-

cies for /da/ are higher than /ba/ for F1, F2 and F3. Another observation was

that /Ci/ syllables have the lowest perceptual threshold. One possible reason

81

may be that the frequency extent for /Ci/ syllables are the shortest. We will try

to explore the reasons that led to these results in more detail in the next chapter.

Miller and Nicely [24] performed similar perception experiments in noise in

1955. They used 16 consonants across five articulatory features (voicing, nasality,

affrication, duration and place of articulation), with initial or final consonant po-

sitions in the /a/ context. In their study, each of the 16 phonemes has a chance

to be confused with every other phoneme. One major generalization made from

their study was that voicing is much less affected by a random masking noise

(voicing was still discriminable at SNR as poor as -12 dB) whereas place of ar-

ticulation is affected more by masking noise (confusion becomes significant at

about 6dB), more specifically, plosive consonants are less robust than fricatives.

Our study showed that fricatives, in general, are more robust than plosives and

this agrees with the generalization from the study of Miller and Nicely. How-

ever, in contrast to Miller and Nicely, our perceptual study showed that voiceless

fricatives, in particular, are slightly more robust than the voiced ones.

82

CHAPTER 4

Correlation Analysis

4.1 Introduction

In Chapter 2, acoustic cues for place of articulation in CV syllables were investi-

gated. In Chapter 3, experiments were performed to investigate the perception of

place of articulation in noise for these CVs. In this chapter, a correlation analysis

is performed in an attempt to explain the observed perceptual confusions.

4.2 Methods

In this study, a simple correlation coefficient was calculated to show the relation-

ship between two properties, which are acoustic measurements and perceptual

thresholds. The equation for the correlation coefficient (ρx,y) between properties

X and Y , assuming both properties have a Gaussian distribution, is:

ρx,y =Cov(X, Y )

σxσy

where:

−1 ≤ ρx,y ≤ 1

83

and:

Cov(X,Y ) =1

n

n∑

j=1

(xj − µx)(yj − µy)

where σx and σy are the variances of properties X and Y , respectively, and µx

and µy are the means of the properties X and Y , respectively.

As mentioned earlier, the perceptual experiments focuses on the identification

of CV syllable pairs that only differed by their place of articulation. The differ-

ence between acoustic measures for CV pairs was considered for the correlation

analysis. It is reasonable to assume that it is harder to confuse between two

syllables when the difference in an acoustic cue(s) is bigger. In other words, the

assumption for this analysis is that the bigger the differences between each CV

pair, the lower the threshold SNR should be, or, the two properties are nega-

tively correlated. Thus, we are only interested in correlation coefficients that are

negative (−1 ≤ ρx,y ≤ 0).

4.2.1 Threshold Values

Threshold values were obtained from perceptual experiments as described in

Chapter 3. In Section 3.4, thresholds for each listening subject and average

thresholds were listed. For correlation analysis purposes, the threshold corre-

sponding to each ”talker” was also obtained. To obtain the threshold of each

talker, the number of correct responses were summed from each of the four lis-

tening subjects then distributed according to the talker, and confusion matrices

per talker were generated. Thresholds were obtained by performing a sigmoid fit

for each set of data and are summarized in Table 4.1. Notice that the values in

the first column of the table are the same as those shown in Tables 3.3 - 3.5.

84

Overall Talker 1 Talker 2 Talker 3 Talker 4

/ba,da/ -7.3050 -8.0016 -5.9018 -7.8126 -8.1258

/bi,di/ 4.1592 6.0546 3.7164 2.1288 4.1592

/bu,du/ -1.6404 -2.1480 -6.7920 -2.5584 -1.6404

/pa,ta/ 6.7458 11.3898 -1.7052 5.9142 6.7458

/pi,ti/ 0.1254 0.6168 0.2118 -5.8416 0.1254

/pu,tu/ -0.0096 -0.4794 -0.204 -2.2344 -0.0096

/fa,sa/ -5.1342 -4.6806 -7.3104 -4.1082 -5.1342

/fi,si/ -3.7842 -4.9506 -2.1804 -3.1578 -3.7842

/fu,su/ -5.0100 -7.5210 -3.6708 -1.9806 -5.0100

/va,za/ -4.5402 -3.4062 -8.8062 -3.7734 -4.5402

/vi,zi/ -1.1976 -0.9438 -0.4200 0.1254 -1.1976

/vu,zu/ -3.4117 -3.5088 -4.9668 -5.0802 -3.4170

Table 4.1: Threshold SNR (in dB SPL) from each talker and overall values

85

Correlation using All points

F2 onset -0.440

VOT -0.410

F2 ∆ -0.382

Table 4.2: Correlation coefficients for all CVs across all talkers

4.2.2 Acoustic Measurements

The acoustic measurements from Chapter 2 are used in this analysis. In this

study, we considered the difference in acoustic measurements between consonants

in each CV pair studied.

As noted earlier, four repetitions were recorded from each of the four talkers.

For each CV, acoustic measurements from the four repetitions were averaged to

obtain the average acoustic characteristics of each talker. The average of the

sixteen tokens (across tokens and talkers) was also calculated.

4.3 Results

The first set of correlations was done by using all data points, including thresholds

and acoustic measures from all talkers and all CVs. The correlation involved 48

points (12 CVs times 4 talkers). Results (shown in Table 4.2) did not show

any high correlation results. The highest correlation resulted in a correlation

coefficient of −0.440 for the F2 onset distinction. In order to better investigate

the correlations between perceptual results and acoustic cues, correlation analysis

was performed according to talker, vowel context, and consonant.

86

/a/ /i/ /u/

F2 onset -0.630 VOT -0.63 burst or noise -0.614

F2 ∆ -0.55 Av4-maxA45 -0.53 VOT -0.59

F2end -0.51 Av-A45 -0.420 Av4-maxA45 -0.374

Table 4.3: Correlation coefficients for /a/, /i/ and /u/ across all talkers for both

fricatives and stops

4.3.1 Talker Independent Analysis

4.3.1.1 Vowel Context Effects

In this section, we investigate the correlation between perception thresholds and

the acoustic measurements according to vowel context. The analysis involved 16

data points for each vowel (4 CVs times 4 talkers). The correlation coefficients

are summarized in Table 4.3. The highest correlation when separating vowels

across all talkers was about -0.63 (onset of F2) for /a/, -0.63 (VOT duration) for

/i/ and -0.61 (burst or noise duration) for /u/. Recall that VOT for stops is the

duration from the release of the burst to the onset of the following vowel and that

Fonset, Fend, Flength and F∆ represent the onset frequency, offset frequency,

transition duration and frequency extent of F1, F2 and F3. These values suggest

that the data are not very highly correlated. Overall, F2 information is the most

correlated with the perception threshold in the /a/ context. For both /i/ and

/u/ contexts, the voice onset time, relative amplitude of the burst to the vowel

at the F4-F5 region and burst or noise duration showed some correlation with

the thresholds. Further analysis can be made by separating fricatives from the

stops where 8 data points from each set will be used for the analysis.

• Fricatives

87

Fric /a/ Fric /i/ Fric /u/

Av4-maxA45 -0.38 Av4-maxA45 -0.51 noise duration -0.65

Av4-A45 -0.37 F1 ∆ -0.37 F3 ∆ -0.59

Table 4.4: Correlation Coefficients for Fricatives in the /a/, /i/ and /u/ contexts

across all talkers.

The acoustic features for the highest three correlation coefficients are listed

in Table 4.4. Once again, the data are not highly correlated. The highest

correlation occurs at Av4-maxA45 for /a/ (at -0.38), Av4-maxA45 for /i/

(at -0.51), and noise duration for /u/ (-0.65).

• Stops

The results for the stops are better than the fricatives. The highest corre-

lation coefficients for /a/, /i/, and /u/ are -0.80, -0.90, and -0.80 for the

onset of F2, VOT and onset of F1, respectively. Table 4.5 showed the

highest three correlations for each vowel. For /Ca/ syllables, F2 onset and

F2∆ were the highest correlated features, and these two showed more than

82% correct classification from Chapter 2. For /Ci/ syllables, Ahi-A23 was

acoustically distinctive and showed a correlation coefficient of -0.69. For

/Cu/ syllables, F1 onset, F1∆, and Av-maxA23 showed better than -0.7

correlation and are all somewhat acoustically distinctively. Their percent

correct acoustic classification are above 70%.

4.3.1.2 Consonant Effects

To investigate consonant effects, we performed similar correlation analysis with

respect to the characteristics of the consonants. The first analysis was separating

88

Stops /a/ Stops /i/ Stops /u/

F2 onset -0.80 VOT -0.90 F1 onset -0.80

F2 ∆ -0.76 Ahi-A23 -0.72 F1 ∆ -0.73

F3 ∆ -0.66 F1 onset -0.58 Av-maxA23 -0.71

Table 4.5: Correlation coefficients for stops of /a/, /i/ and /u/ contexts across

all talkers.

/f,s/ /v,z/ All Fricatives

F3 onset -0.60 F1 onset -0.53 F2 onset -0.37

69.8% 67.7% 69.8%

F3 ∆ -0.57 F2 delta -0.48 Av-Anoise -0.340

72.9% 67.7% 89.58%

Av-Anoise -0.53 F2 onset -0.46 F1 onset -0.34

89.58% 71.88% 67.1%

Table 4.6: Correlation coefficients for fricatives (voiced or voiceless) across all

vowel contexts for all talkers. Numbers in italic are percent correct acoustic

classification.

fricatives from the stops; 24 data points (6 CVs times 4 talkers) were used in this

correlation analysis. The highest correlation values obtained for this analysis

were -0.34 for fricatives (F2 onset) and -0.54 for stops (F2 onset). The top

three features were listed in the last two columns of Tables 4.6 and 4.7. Further

analysis separated out the voicing component of the consonants, which will reduce

the number of data points to 12 (3 CVs times 4 talkers).

• Voiced and Voiceless Fricatives

89

For voiced and voiceless fricatives, the correlation results were still not clear.

From Table 4.6, we observe that the highest correlations were -0.60 (F3

onset) for voiceless fricatives and -0.53 (F1 onset) for voiced fricatives. For

voiced fricatives, F1 onset and F2∆ and F2 onset only classified acoustic

cues at 68%, 68% and 72%, respectively. The most acoustically distinctive

features, such as Av-Anoise and Av4-maxA45 (at 90% and 82% classifi-

cation) did not show any high correlation results. For voiceless fricatives,

F3 onset, F3∆, and Av-Anoise showed 70%, 73% and 90% correct acoustic

classification in Chapter 2 and the high correlation agreed with our results.

• Voiced and Voiceless Stops

Stops resulted in slightly higher correlations than the fricatives, especially

for the voiceless syllables. Table 4.7 listed the highest three correlation co-

efficients and corresponding acoustic features for voiced and voiceless stops

across all vowel contexts. The best correlations were F1 onset, for voiced

stops (at -0.88) and Ahi-A23 for voiceless stops (at -0.84). For voiceless

stops, Ahi-A23 and Av-Ahi resulted in very high percent correct classifica-

tion for /p,t/ syllables (89% and 86%) in Chapter 2. The correlation results

agreed with the acoustics in Chapter 2. VOT, once again, was not a dis-

tinctive acoustic feature for place and it only showed 55% correct acoustic

classification. Voiced stops, on the other hand, showed worse correlation

between acoustics and perception. F1 information only exhibited 61-65%

correct classification for /b,d/ syllables, but the top three features for cor-

relation all relate to F1. Acoustically distinctive features, such as Av-Ahi

(at 90% classification) did not appear to be correlated for /b,d/.

90

/p,t/ /b,d/ Stops

Ahi-A23 -0.84 F1 onset -0.88 F2 onset -0.57

88.54% 64.58% 72.9%

Av-Ahi -0.46 F1 ∆ -0.81 F1 onset -0.54

86.46% 61.46 60.4%

VOT -0.41 F2 ∆ -0.77 F2 ∆ -0.53

55.21% 73.96% 70.3%

Table 4.7: Correlation coefficients for stops (voiced or voiceless) across all vowel

contexts for all talkers. Numbers in italic are percent correct acoustic classifica-

tion. Number in bold is the highest percent correct acoustic classification for the

syllable.

4.3.1.3 Correlation for Each Syllable

Tables 4.8 - 4.11 summarized the correlation coefficients for each syllable. Each

set of analysis only consists of 4 data points (from 4 talkers). Analysis results were

much better once we separate vowel context, voicing, and manner effects. For

fricatives, the highest coefficient values were -0.88 (F2 offset) for /fa,sa/, -0.954

(Ahi-A23) for /fi,si/, -0.94 (F1∆) for /fu,su/, -0.77 (noise duration) for /va,za/,

-0.99 (F2∆) for /vi,zi/ and -0.81 (noise duration) for /vu,zu/. Acoustically dis-

tinct features, such as Av-Anoise showed 96.88% correct classification for /fa,sa/

and /vi,zi/ but only showed correlation values of -0.58 and -0.70, respectively. For

stops, correlations were even better; the highest correlation coefficients and cor-

responding features for /pa,ta/, /pi,ti/, /pu,tu/, /ba,da/, /bi,di/, and /bu,du/

were F1 offset (-0.96), VOT (-0.98), F1∆ (-0.87), F3 offset (-0.996), F1 onset

(-0.95) and F1∆ (-0.86), respectively. The top five features that exhibit high

correlations agreed with acoustic results in Chapter 2. For fricatives, all noise

91

fasa fisi fusu

F2end -0.881 Ahi-A23 -0.954 F1 ∆ -0.936

62.5% 59.38% 62.5%

Av-Ahi -0.845 Av-Ahi -0.905 F3 onset -0.906

75% 78.13% 75%

Av4-A45 -0.833 Av4-maxA45 -0.888 F2 end -0.816

75% 93.75% 87.5%

Av4-maxA45 -0.747 F3 ∆ -0.810 F3∆ -0.723

78.13% 68.75% 78.13%

Av-Anoise -0.575 F3end -0.564 F2 onset -0.723

96.88% 68.75% 84.38%

Table 4.8: Correlation coefficients for /fa,sa/, /fi,si/ and /fusu/ for all talkers.

Numbers in italic are percent correct acoustic classification. Number in bold is

the highest percent correct acoustic classification for the syllable.

measurements and F2 information showed reasonable distinction in acoustics, and

these features appeared to be highly correlated. For stops, F1 and F3 informa-

tion appeared to be better correlated than F2 information, but acoustic results

in Chapter 2 showed otherwise. However, noise information, such as Av-Ahi, and

Ahi-A23 agreed with acoustic results. In general, noise measures correlate better

with voiceless syllables than voiced syllables, especially for fricatives.

4.3.2 Talker Dependent Analysis

For each talker (listed in the tables as T1, T2, T3 and T4), the correlation for

all CVs was calculated and listed in Table 4.12. The analysis used 12 points (12

CV pairs). Correlation results were better than Table 4.2, which considered all

92

vaza vizi vuzu

noise duration -0.769 F2 ∆ -0.990 noise duration -0.809

75% 71.88% 75%

F2end -0.654 F2 onset -0.987 Ahi-A23 -0.659

87.5% 68.75% 65.63%

F1 onset -0.336 F2end -0.879 F1∆ -0.583

100% 62.5% 59.38%

F2∆ -0.332 Av-Anoise -0.702 F3∆ -0.501

84.38% 96.88% 84.38%

Av4-A45 -0.321 F1∆ -0.658 F3 onset -0.472

71.88% 68.75% 75%

Table 4.9: Correlation coefficients for /va,za/, /vi,zi/, and /vu,zu/ for all talkers.



93

pata piti putu

F1end -0.960 VOT -0.980 F1 ∆ -0.874

68.75% 68.63% 65.63%

VOT -0.950 Ahi-A23 -0.967 VOT -0.782

68.75% 96.88% 68.75%

Av-Ahi -0.819 burst duration -0.937 F1 onset -0.770

84.38% 81.25% 59.38%

F2end -0.758 F3 ∆ -0.757 Av-maxA23 -0.713

71.88% 75%

Ahi-A23 -0.755 F2end -0.621 Av-Ahi -0.675

90.63% 100%

Table 4.10: Correlation coefficients for /pa,ta/, /pi,ti/, /pu,tu/ for all talkers.

Numbers in italic are percent correct acoustic classification.

94

bada bidi budu

F3end -0.996 F1 onset -0.947 F1 ∆ -0.856

65.63% 68.75% 78.13%

F1length -0.677 F1 ∆ -0.815 F1 onset -0.843

65.63% 81.25%

F2 ∆ -0.506 VOT -0.800 Av4-maxA45 -0.648

100% 71.88% 68.75%

F2 onset -0.431 Av-maxA23 -0.566 Av-maxA23 -0.612

96.88% 78.13% 71.88%

F3 onset -0.424 F3 onset -0.563 Amid-Avmid -0.598

75% 68.75% 78.13%

Table 4.11: Correlation coefficients for /ba,da/, /bi,di/, /bu,du/ for all talkers.



95

T1 T2 T3 T4

F2 ∆ -0.595 F1end -0.826 VOT -0.553 Av4-maxA45 -0.447

F2onset -0.572 F1length -0.744 F3 ∆ -0.496 F2 ∆ -0.437

VOT -0.511 F2onset -0.741 F2length -0.484 F2onset -0.352

Table 4.12: Correlation coefficients for all CVs from each talker (ie. T1 represents

Talker No.1).

CVs among all talkers. The better correlation results showed that talker effect

played an important role in perception.

4.3.2.1 Vowel Context Effect

Tables 4.13- 4.16 summarized the correlation results for /a/,/i/ and /u/ using

acoustic measures from each talker v.s. corresponding perception thresholds from

Table 4.1. For /Ca/ syllables, F1, F2 and F3 information showed very high cor-

relation values, especially F2 onset, which showed -0.99, and -0.91 correlation

from Talker No.1 and Talker No.4. For /Ci/ syllables, Av4-maxA45, F2 offset

and VOT all exhibited good correlation. As for /Cu/ syllables, F3∆ from Talker

No. 3 showed perfect correlation and F2 transition duration from Talker No.

2 showed -0.99 correlation as well. Overall, features that exhibit high correla-

tions were similar to the features that appeared in Table 4.3 but with higher

correlation values. The perceptual experiments showed that /Ci/ syllables are

usually the hardest to identify for all consonants, and the correlation analysis

showed that thresholds for /Ci/ syllables correlate with VOT, which is a weak

feature in noise. It is worth mentioning that a spectral measurement in the F4-F5

region (Av4-maxA45), which was acoustically distinctive (at 89% correct classifi-

cation), showed reasonable correlation for all talkers. This agrees with the study

96

T1-/a/ T1-/i/ T1-/u/

F2 ∆ -0.987 VOT -0.916 VOT -0.967

F2onset -0.964 F1onset -0.882 burst or noise -0.950

F1 ∆ -0.879 F1 ∆ -0.852 Av4-maxA45 -0.853

F1onset -0.863 burst or noise -0.825 Av4-A45 -0.849

F1length -0.855 F2length -0.802 F1length -0.660

Table 4.13: Correlation for /a/,/i/, /u/ from Talker No. 1

T2-/a/ T2-/i/ T2-/u/

F3end -0.885 F2onset -0.816 F2length -0.985

F1length -0.865 Av4-maxA45 -0.806 F1end -0.879

F3onset -0.802 F2 ∆ -0.793 F3onset -0.836

F2end -0.652 burst or noise -0.776 F3end -0.592

F1end -0.645 F2end -0.763 F1length -0.553

Table 4.14: Correlation for /a/,/i/, /u/ from Talker No.2

of Hedrick et al. [12], [14] in which they proposed that F4, F5 frication amplitude

relative to the vowel amplitude in the same region is a prominent cue for place

of articulation.

4.3.2.2 Consonant Effects

Results of the correlation analysis when separating out both the talker and con-

sonant effects were good. One reason could be the small number of data points

taken for each set of analysis and the other may be that the variabilities among

tokens were reduced.

97

T3-/a/ T3-/i/ T3-/u/

F3onset -0.908 F2end -0.939 F3 ∆ -1.000

F3 ∆ -0.873 F1end -0.751 F3onset -0.993

VOT -0.754 F1onset -0.705 burst or noise -0.958

F2length -0.748 Ahi-A23 -0.556 VOT -0.949

F3end -0.536 F2 ∆ -0.429 F3end -0.770

Table 4.15: Correlation for /a/,/i/, /u/ From talker No. 3

T4-/a/ T4-/i/ T4-/u/

F2 onset -0.902 Av4-maxA45 -0.958 burst or noise -0.887

Av4-maxA45 -0.885 F3 ∆ -0.905 F1 ∆ -0.769

F3onset -0.859 F3end -0.763 VOT -0.604

F2end -0.858 VOT -0.738 Av4-maxA45 -0.360

F3 ∆ -0.763 burst or noise -0.664 F3onset -0.125

Table 4.16: Correlation for /a/,/i/, /u/ from Talker No. 4

fs-T1 vz-T1 Fric-T1

F3 ∆ -0.999 F2onset -0.902 F3 ∆ -0.565

Ahi-A23 -0.690 F2end -0.887 F2end -0.457

F1end -0.628 F3onset -0.757 F3onset -0.396

F2 ∆ -0.596 F2onset -0.592 noise -0.368

F3 ∆ -0.544 F1end -0.571 F2onset -0.323

Table 4.17: Correlation for fricatives from Talker No. 1

98

fs-T2 vz-T2 Fric-T2

F1end -0.963 F2onset -0.993 F1length -0.867

F1length -0.933 Av-Anoise -0.991 F1end -0.854

F2end -0.967 F1 ∆ -0.983 F2end -0.839

F1end -0.961 F1length -0.956 F2onset -0.832

F1length -0.885 F1end -0.708 F1onset -0.804


fs-T3 vz-T3 Fric-T3

F3length -1.000 F3onset -0.963 F2onset -0.738

F2length -0.997 F2end -0.945 F1length -0.703

F3 ∆ -0.954 F1end -0.736 F1end -0.643

F2onset -0.905 Av4-A45 -0.687 F1onset -0.605

F1end -0.872 F3end -0.778 F3end -0.594


fs-T4 vz-T4 Fric-T4

Av-Ahi -1.000 Av4-maxA45 -0.965 F3 ∆ -0.634

noise -0.993 F2 end -0.948 Ahi-A23 -0.595

F1end -0.990 Av-A45 -0.636 Av4-maxA45 -0.355

Av4-maxA45 -0.563 F1 ∆ -0.314 Av4-A45 -0.290

F2 ∆ -0.410 noise -0.281 F3onset -0.279


99

pt-T1 bd-T1 Stop-T1

Ahi-A23 -0.984 F1onset -0.983 F2onset -0.756

Av-Ahi -0.923 F2length -0.980 F1end -0.702

F2 ∆ -0.782 F1 ∆ -0.978 F2 ∆ -0.688

F2onset -0.895 F1onset -0.899 F1length -0.682

F1length -0.814 F1 ∆ -0.847 F1onset -0.568

Table 4.21: Correlation for stops from Talker No. 1

pt-T2 bd-T2 Stop-T2

VOT -0.995 F2onset -0.987 F1end -0.892

F2 ∆ -0.968 F3 ∆ -0.976 F2onset -0.869

burst -0.959 F3 onset -0.953 F1 ∆ -0.801

F1end -0.917 F1∆ -0.931 F1length -0.787

F1onset -0.879 F2 ∆ -0.927 F1onset -0.769


pt-T3 bd-T3 Stop-T3

Av-maxA23 -0.995 Av4-A45 -1.000 Av4-A45 -0.805

VOT -0.969 F1onset -0.991 VOT -0.687

Ahi-A23 -0.928 F2end -0.984 F3 ∆ -0.609

F2onset -0.991 aspiration -0.870 F2length -0.569

F1length -0.980 F2onset -0.866 Av4-maxA45 -0.507


100

pt-T4 bd-T4 Stop-T4

Ahi-A23 -0.998 F2 ∆ -0.995 F2 ∆ -0.958

F3 ∆ -0.972 F1onset -0.981 F1onset -0.937

F3onset -0.971 F1length -0.956 F2onset -0.679

F1length -0.956 F1end -0.880 Amid-Avmid -0.662

F1end -0.880 F1 ∆ -0.655 F1end -0.611


• Fricatives

Correlations of all fricatives showed that formant information correlated

better than noise measures, but the correlation results were not very good.

To further investigate consonant effects, voicing was separated out for anal-

ysis. The results showed great improvement. These analysis used 3 points

(3CVs from each vowel context). For /f,s/, Av-Ahi, F2, F3 duration, F3∆

F1 offset and noise duration all resulted in better than -0.99% correlation

values. These results showed almost perfect linear correlation between per-

ception and acoustics. As for /v,z/ syllables, F2 onset, Av-Anoise and F1∆

showed better than -0.98 correlation. For both voiceless and voiced frica-

tives, the features listed in Tables 4.17-4.20 agreed with acoustic results

from Chapter 2. Moreover, recall from Chapter 2, fricatives have very dis-

tinctive measures for F1, F2 and noise measures. The correlation results

agree with the acoustic measures.

• Stops

Very good results are also observed for the stop syllables. Tables 4.21-4.24

listed the five highest correlated features for voiced and voiceless stops. The

101

results when correlating all stops showed that F2∆, F1 onset and offset

were few of the most highly correlated features (at -0.96, -0.94, and -0.89,

respectively). Acoustically, these features correlate stops to some extent,

but not as highly classified as noise measures such as Ahi-A23 or Av-Ahi.

Further analysis were performed to voiced and voiceless stops separately.

For /p,t/ syllables, Ahi-A23, Av-maxA23, VOT showed better than -0.98

for correlation values. Recall from Chapter 2, Ahi-A23 and Av-Ahi had two

of the highest percent correct classification for /p,t/ syllables. This result

not only agreed with the results in Table 4.7, but also agreed with the

acoustic results. For /b,d/ syllables, Av4-A45 from Talker No. 3 achieved

perfect linear correlation. However, Av4-A45 only classify /b,d/ acousti-

cally at 64%. Other features, such as F2∆, F2 onset, F1∆ and F3∆ , all

exhibited better than -0.90 correlation for /b,d/ and these features showed

reasonably good classification for /b,d/ for various vowel contexts in Chap-

ter 2.

4.4 Summary

Comparing results from Section 4.3.1 and 4.3.2, it is reasonable to conclude that

talker variation played an important role in correlating acoustics and perceptions.

Correlation coefficients that separated out talkers are much better than those

when considering CVs from all talkers. In general, perceptual threshold for stops

are better correlated with distinctive acoustic features than the fricatives. /Ca/

syllables are better than /Cu/ syllables, and both are better than /Ci/ syllables

in terms of correlations between perception and underlying acoustic features.

From Chapter 2, we observed that the distinctive acoustic features are most

prominent for /Ca/ syllables and least for /Ci/ syllables. Another generalization

102

can be drawn is that noise measurements better cue voiceless syllables than voiced

syllables, and that formant information are most robust in cuing voiced syllables,

particularly the stops. In Chapter 3, we noted that the threshold values for /Ci/

syllables were typically worse than /Cu/ syllables, which were worse than /Ca/

syllables.

Some of the features that are prominent both in the acoustic and correlation

analyses are the F1, F2 formant information and the amplitude of the noise

segment relative to the vowel in the F4-F5 frequency region (Av4-A45 and Av4-

maxA45). One major conclusion we can draw from the correlation analysis is

that no definite acoustic feature can cue perception of place of articulation in

noise. The results are dependent upon the vowel context, voicing and manner.

103

CHAPTER 5

Summary and Future Directions

5.1 Summary

In this thesis, the relationship between human perception and acoustic character-

istics of the place of articulation feature was studied. In Chapter 1, an overview

of speech production and background for this study were presented. In Chapter

2, results of acoustic analysis to classify stop and fricative place of articulation

were discussed. In Chapter 3, a description of perceptual experiments in noise

and their results were presented. Finally, in Chapter 4, correlation analysis be-

tween speech acoustics and perceptual experiment results were performed and

the results were provided. Some of the major results from this research can be

summarized as follows:

1. Chapter 2: Acoustic classification for place of articulation

In this chapter, various acoustic measurements, including formant and time

duration measurements, noise measurements and locus equation analysis,

were examined and analyzed on a CV database. The goal was to find

invariant cues to classify place of articulation. From our measurements,

we conclude that noise measures are more reliable cues for place of artic-

ulation for both stops and fricatives. Noise measures such as Av4-A45,

Av4-maxA45 and Av-Ahi classify well both stops and fricatives. In addi-

104

tion, Amid-Avmid and Ahi-A23 are good measures to classify stops and

Av-Anoise is good to classify fricatives.

As for formant frequency information, F2 is the most distinct measure to

cue place for both stops and fricatives. F1 show some distinction for several

syllables but F3 did not. Both the onset and the frequency extent of F2 are

cues for place of articulation. The distinction are typically most prominent

for /Ca/ syllables and least for /Ci/ syllables. Durational measures for

formants, such as formant transition duration are not reliable cues for place

of articulation for both stops and fricatives.

The duration of burst in stops signals place of articulation across vowel

contexts, especially /Ca/ and /Ci/ syllables (above 81% correct percent

classification), so does the duration of noise segments for fricatives (above

75%).

Locus equation information, the slopes and the Y-intercept, are distinct

cues for place of articulation for both plosives and fricatives only when

the analysis is performed for all CVs in all vowel contexts. However, no

significant consistency was found when the analysis was performed on /Ca/,

/Ci/ and /Cu/ syllables separately.

2. Chapter 3: Perceptual experiment in noise

In this chapter, perceptual experiments were conducted in an attempt to

seek robust acoustic cues that classify place of articulation in a noisy envi-

ronment. From the experimental data, we can conclude that the perception

of place of articulation in noise is dependent upon the vowel context, voicing

and manner of articulation. In general fricatives are more robust than stops

except voiced /Ca/ syllables where the opposite is true. Voiceless fricatives

are more robust than their voiced counterparts in all vowel contexts. Voiced

105

stops /ba,da/ are more robust than /pa,ta/, but /bi,di/, on the other hand,

is less robust than /pi,ti/. The voiced pair /bu,tu/ is slightly more robust

than /pu,tu/, by approximately 1 dB. As for vowel context, /a/ is more

robust than /u/, which is more robust than /i/ except for the /p,t,/ case.

3. Correlation between acoustic classification and perceptual exper-

iment results

In this chapter, correlation analysis was performed in order to seek acoustic

features that affect perception of place of articulation in noise. Correlation

analysis was performed between acoustic measures and perceptual thresh-

olds. One generalization from the analysis was that talker variation played

an important role in correlating acoustics and perception. Correlation co-

efficients that separated out talkers are much better than those when con-

sidering CVs from all talkers. In general, stops are better correlated with

their acoustic features than the fricatives, and that /Ca/ syllables are bet-

ter than /Cu/ syllables, and both are better than /Ci/ syllables in terms

of correlations between perception and underlying features. These trends

are similar to both the acoustic feature results and the perception results.

Some of the features that are prominent in both the acoustic analysis and

the correlation analysis are the F1, F2 frequency information and the am-

plitude of the noise segment relative to the vowel in the F4-F5 frequency

region (Av4-A45 and Av4-maxA45). One major conclusion we can draw

from this chapter is that no single acoustic feature can cue perception of

place of articulation in noise. The results are dependent upon the talker,

vowel context, voicing and manner.

106

5.2 Future Studies

For future work, several different directions can be taken. From the acoustic

stand point, one can further analyze the acoustic measures taking into account

the talker’s gender. It is known that speech characteristics, such as formant

frequencies, are different between males and females because of physiological

differences.

The perceptual studies can be expanded by masking the stimuli with different

types of noise maskers such as perceptually flat noise, speech shape noise, corre-

lated noise, car noise, party noise or auditorium noise ... etc. In addition, the

same experiments can be done on hearing impaired listeners. These studies can

help understand which speech cues hearing-impaired people rely on in a noisy

environment.

In our study, we attempt to classify place of articulation for stops and frica-

tives by performing a simple correlation between acoustic measurements and

perceptual thresholds. Studies on validating these results can be done by using

synthetic speech tokens. The tokens can have certain variable acoustic features

that are correlated with the perceptual thresholds. For example, we can gener-

ate several synthetic /ba,da/ tokens with different F1 onset frequencies since we

found that F1 onset frequency correlates with perceptual threshold of /ba,da/

at 0.859. We can validate our correlation results by performing perceptual ex-

periments on these synthesized /ba,da/ and investigate the thresholds as the F1

onset frequency changes. By performing perceptual studies on these synthetic

speech tokens, we can further validate (or not) our correlation results.

107

APPENDIX A

Formant Measurements

Formant measurements for all CVs are listed in this appendix. Measurements

include the onset, offset, steady-state of F1, F2 and F3 as well as each of the

three formant transition extents. Tokens are listed as CV.TalkerNum.TokenNum.

For example, Talker No.1 speaking token No.1 of /ba/ will be listed as ba.1.1.

Talkers No. 1 and 2 are male. Talkers No. 3 and 4 are female.

108

F1onset F2onset F3onset F1offset F2offset F3offset

ba.1.1 546.88 988.28 2308.59 683.59 988.28 2402.34

ba.1.2 578.13 1015.62 2445.31 664.06 1023.44 2550.78

ba.1.3 550.78 1011.72 2363.28 660.16 1035.16 2328.13

ba.1.4 539.06 1031.25 2425.78 656.25 1031.25 2453.13

ba.2.1 500 1101.56 2246.09 652.34 1097.66 2246.09

ba.2.2 593.75 1140.63 2300.78 718.75 1140.63 2312.50

ba.2.3 554.69 1179.69 2363.28 707.03 1207.03 2355.47

ba.2.4 539.06 1058.59 2261.72 718.75 1125.00 2261.72

ba.3.1 632.81 1382.81 2664.06 796.87 1390.63 2898.44

ba.3.2 519.53 1390.63 2707.03 875.00 1375.00 2765.63

ba.3.3 546.88 1277.34 2457.03 773.44 1277.34 2574.22

ba.3.4 570.31 1316.41 2496.09 847.66 1316.41 2703.13

ba.4.1 414.06 1101.56 2722.66 785.16 1292.97 2656.25

ba.4.2 511.72 1304.69 2625.00 804.69 1386.72 2722.66

ba.4.3 527.34 1343.75 2679.69 796.87 1367.19 2695.31

ba.4.4 484.38 1230.47 2648.44 777.34 1324.22 2667.97

Table A.1: Onset and Offset frequencies for /ba/ (Hz). Measures are listed as

CV.TalkerNum.TokenNum

109


da.1.1 414.06 1464.84 2535.16 691.41 1148.44 2453.13

da.1.2 421.88 1476.56 2511.72 664.06 1160.16 2539.06

da.1.3 402.34 1480.47 2578.12 722.66 1480.47 2609.38

da.1.4 425.78 1414.06 2535.16 710.94 1335.94 2527.34

da.2.1 472.66 1628.91 2433.59 589.84 1289.06 2355.47

da.2.2 433.59 1628.91 2488.28 675.78 1289.06 2261.72

da.2.3 421.88 1574.22 2429.69 644.53 1296.88 2175.78

da.2.4 386.72 1660.16 2476.56 660.16 1304.69 2269.53

da.3.1 445.31 2269.53 3078.13 812.50 1367.19 2859.38

da.3.2 437.5 1890.63 3070.31 835.94 1277.34 2750.00

da.3.3 464.84 2187.5 3179.69 824.22 1343.75 2847.66

da.3.4 453.13 2179.69 3054.69 746.09 1468.75 2828.13

da.4.1 398.44 1996.09 3066.41 792.97 1847.66 2882.81

da.4.2 523.44 1832.03 3039.06 789.06 1753.91 2699.22

da.4.3 386.72 2183.59 3101.56 562.50 1796.88 2777.34

da.4.4 425.78 1945.31 2769.53 542.97 1816.41 2769.53

Table A.2: Onset and offset frequencies for /da/ (Hz). Measures are listed as


110


pa.1.1 761.72 1058.59 2597.66 730.47 1058.59 2437.50

pa.1.2 781.25 1062.5 2402.34 761.72 1082.03 2296.88

pa.1.3 718.75 1035.16 2468.75 710.94 1035.16 2437.50

pa.1.4 765.63 1089.84 2457.03 703.13 1097.66 2386.72

pa.2.1 824.22 1128.91 2257.81 730.47 1140.63 2289.06

pa.2.2 691.41 1101.56 2312.50 734.38 1132.81 2312.50

pa.2.3 707.03 1121.09 2164.06 707.03 1140.63 2222.66

pa.2.4 707.03 1097.66 2253.91 707.03 1097.66 2253.91

pa.3.1 1019.53 1132.81 2988.28 945.31 1289.06 3085.94

pa.3.2 1042.97 1082.03 2976.56 871.09 1246.09 2960.94

pa.3.3 1136.72 1136.72 2921.88 843.75 1167.97 2945.31

pa.3.4 1039.06 1039.06 2871.09 875.00 1210.94 3031.25

pa.4.1 1007.81 1265.63 2656.25 902.34 1316.40 2621.09

pa.4.2 917.97 1468.75 2570.31 957.03 1289.06 2613.28

pa.4.3 925.78 1441.41 2703.13 917.97 1328.13 2687.50

Table A.3: Onset and offset frequencies for /pa/ (Hz). Measures are listed as


111


ta.1.1 632.81 1078.13 2402.34 699.22 1144.53 2355.47

ta.1.2 796.87 1050.78 2269.53 777.34 1125.00 2382.81

ta.1.3 769.53 1054.69 2496.09 738.28 1128.91 2214.84

ta.1.4 621.09 1128.91 2136.72 691.41 1187.50 2316.41

ta.2.1 687.5 1308.59 2363.28 710.94 1328.13 2328.13

ta.2.2 589.84 1359.37 2238.28 695.31 1378.91 2300.78

ta.2.3 667.97 1300.78 2277.34 679.69 1250.00 2203.13

ta.2.4 644.53 1347.66 2132.81 683.59 1308.59 2257.81

ta.3.1 906.25 1703.13 2847.66 898.44 1320.31 3097.66

ta.3.2 835.94 1539.06 2800.78 847.66 1332.03 3066.41

ta.3.3 960.94 2003.91 2746.09 824.22 1359.37 2789.06

ta.3.4 1121.09 1203.13 2898.44 898.44 1367.19 2953.13

ta.4.1 945.31 1593.75 2687.50 988.28 1289.06 2703.13

ta.4.2 921.88 1492.19 2597.66 929.69 1386.72 2734.37

ta.4.3 1062.5 1339.84 2515.63 976.56 1367.19 2679.69

ta.4.4 878.91 1589.84 2589.84 921.88 1363.28 2578.12

Table A.4: Onset and offset frequencies for /ta/ (Hz). Measures are listed as


112


fa.1.1 636.72 914.06 2347.66 636.72 1046.88 2414.06

fa.1.2 558.59 1058.59 2238.28 558.59 1089.84 2304.69

fa.1.3 800.78 921.88 2343.75 609.38 976.56 2410.16

fa.1.4 515.63 957.03 2304.69 562.5 1019.53 2304.69

fa.2.1 496.09 1054.69 2128.91 593.75 1104.37 2144.53

fa.2.2 527.34 1070.31 2125 722.66 1152.45 2312.5

fa.2.3 564.41 1003.91 2242.19 578.13 1062.5 2234.18

fa.2.4 531.36 1074.22 2257.81 609.38 1074.22 2257.81

fa.3.1 589.84 1214.84 2683.59 875 1261.72 2683.59

fa.3.2 570.31 1246.09 2449.22 804.69 1253.91 2531.36

fa.3.3 582.03 1300.78 2640.63 812.5 1308.59 2613.28

fa.3.4 609.38 1199.22 2687.5 792.97 1167.97 3031.25

fa.4.1 500 1064.41 2078.13 753.91 1179.69 2664.06

fa.4.2 914.06 1128.91 2468.75 863.28 1257.81 2500

fa.4.3 894.53 1199.22 2574.22 984.18 1199.22 2757.81

fa.4.4 796.87 1246.09 2230.47 871.09 1261.72 2617.19

Table A.5: Onset and offset Frequencies for /fa/ (Hz) Measures are listed as


113


sa.1.1 441.21 1218.75 2468.75 693.31 1089.84 2476.56

sa.1.2 476.56 1250 2515.63 667.97 1152.45 2515.63

sa.1.3 441.21 1250 2480.47 687.5 1152.45 2480.47

sa.1.4 410.16 1199.22 2515.63 687.5 1226.56 2515.63

sa.2.1 443.31 1300.78 2441.21 660.16 1289.06 2354.37

sa.2.2 457.03 1292.97 2390.63 574.22 1265.63 2335.94

sa.2.3 390.63 1339.84 2429.69 414.06 1359.37 2351.56

sa.2.4 382.81 1378.91 2449.22 604.37 1378.91 2367.19

sa.3.1 496.09 1683.59 2992.19 832.03 1281.36 3054.69

sa.3.2 378.91 1867.19 2980.47 792.97 1335.94 2921.88

sa.3.3 402.45 1753.91 2980.47 832.03 1281.25 2814.41

sa.3.4 443.31 1835.94 2992.19 812.5 1300.78 2957.03

sa.4.1 339.84 1761.72 2789.06 792.97 1480.47 2832.03

sa.4.2 808.59 1398.44 2808.59 878.91 1437.5 2808.59

sa.4.3 421.88 1671.88 2863.28 949.22 1402.45 2781.36

sa.4.4 314.41 1277.34 2753.91 814.41 1367.19 2730.47

Table A.6: Onset and offset Frequencies for /sa/ (Hz) Measures are listed as


114


va.1.1 519.53 1109.38 2378.91 519.53 1054.69 2410.16

va.1.2 492.19 1054.69 2320.31 660.16 1054.69 2390.63

va.1.3 496.09 1078.13 2354.37 691.21 1078.13 2414.06

va.1.4 531.25 1085.94 2410.16 652.45 1062.5 2410.16

va.2.1 453.13 1082.03 2328.13 558.59 1082.03 2277.34

va.2.2 507.81 1046.88 2300.78 703.13 1046.88 2281.36

va.2.3 472.66 1109.38 2265.63 679.69 1132.81 2281.25

va.2.4 511.72 1093.75 2187.5 753.91 1121.09 2234.18

va.3.1 437.5 1414.06 2746.09 812.5 1304.69 2746.09

va.3.2 488.28 1578.13 2164.06 824.22 1371.09 2726.56

va.3.3 464.84 1363.28 2429.69 789.06 1281.36 2867.19

va.3.4 492.19 1386.72 2460.94 878.91 1328.13 2753.91

va.4.1 300.78 1300.78 2523.24 812.5 1402.14 2613.28

va.4.2 320.31 1109.38 2730.47 777.34 1257.81 2746.09

va.4.3 320.31 1193.31 2613.28 814.41 1261.72 2640.63

va.4.4 343.75 1421.88 2894.53 851.56 1339.84 2699.22

Table A.7: Onset and offset Frequencies for /va/ (Hz) Measures are listed as


115


za.1.1 359.38 1347.66 2492.19 699.22 1351.56 2492.19

za.1.2 394.53 1375 2531.25 718.75 1367.19 2527.34

za.1.3 398.44 1371.09 2511.72 693.31 1104.37 2554.69

za.1.4 359.38 1402.14 2539.06 679.69 1347.66 2480.47

za.2.1 382.81 1480.47 2476.56 652.45 1414.06 2449.22

za.2.2 386.72 1500 2480.47 625 1386.72 2453.13

za.2.3 375 1511.72 2281.25 664.06 1531.25 2363.28

za.2.4 402.45 1484.18 2421.88 578.13 1449.22 2421.88

za.3.1 382.81 1882.81 2953.13 789.06 1804.69 2953.13

za.3.2 398.44 1875 2976.56 789.06 1828.13 2953.13

za.3.3 354.37 1832.03 3031.36 804.69 1253.91 2843.75

za.3.4 351.56 1894.53 3023.24 843.75 1851.56 3054.69

za.4.1 261.72 1765.63 3179.69 812.5 1570.31 2812.5

za.4.2 300.78 1632.81 3064.41 824.22 1359.37 2941.21

za.4.3 390.63 1628.91 2832.03 949.22 1449.22 2933.59

za.4.4 398.44 1562.5 2843.75 835.94 1433.59 2820.31

Table A.8: Onset and offset Frequencies for /za/ (Hz) Measures are listed as


116

ssF1 ssF2 ssF3 F1 ∆ F2 ∆ F3 ∆

ba.1.1 737.50 1056.25 2471.09 190.62 67.97 162.50

ba.1.2 735.94 1069.53 2437.50 157.81 53.91 -7.81

ba.1.3 738.28 1092.97 2439.06 187.50 81.25 75.78

ba.1.4 731.25 1084.38 2498.44 192.19 53.13 72.66

ba.2.1 717.97 1192.97 2349.22 217.97 91.41 103.13

ba.2.2 755.47 1229.69 2353.91 161.72 89.06 53.13

ba.2.3 737.50 1182.03 2331.25 182.81 2.34 -32.03

ba.2.4 755.47 1169.53 2358.59 216.41 110.94 96.87

ba.3.1 871.88 1297.66 3043.75 239.07 -85.15 379.69

ba.3.2 880.47 1228.91 2946.88 360.94 -161.72 239.85

ba.3.3 813.28 1214.06 2757.03 266.40 -63.28 300.00

ba.3.4 871.09 1244.53 2965.63 300.78 -71.88 469.54

ba.4.1 923.44 1260.16 2822.66 509.38 158.60 100.00

ba.4.2 958.59 1340.63 2828.13 446.87 35.94 203.13

ba.4.3 923.44 1275.78 2816.41 396.10 -67.97 136.72

ba.4.4 936.72 1278.13 2805.47 452.34 47.66 157.03

Table A.9: Steady-state frequency and frequency translation for /ba/ (Hz). Mea-

sures are listed as CV.TalkerNum.TokenNum

117


da.1.1 738.28 1142.19 2446.88 324.22 -322.65 -88.28

da.1.2 713.28 1148.44 2400.00 291.40 -328.12 -111.72

da.1.3 756.25 1109.38 2532.03 353.91 -371.09 -46.09

da.1.4 738.28 1096.88 2449.22 312.50 -317.18 -85.94

da.2.1 721.87 1269.53 2358.59 249.21 -359.38 -75.00

da.2.2 712.50 1270.31 2296.88 278.91 -358.60 -191.40

da.2.3 701.56 1255.47 2234.38 279.68 -318.75 -195.31

da.2.4 724.22 1302.34 2281.25 337.50 -357.82 -195.31

da.3.1 857.03 1415.63 2904.69 411.72 -853.90 -173.44

da.3.2 867.97 1289.06 2900.00 430.47 -601.57 -170.31

da.3.3 850.78 1489.06 2924.22 385.94 -698.44 -255.47

da.3.4 853.13 1461.72 2850.00 400.00 -717.97 -204.69

da.4.1 867.97 1280.47 2923.44 469.53 -715.62 -142.97

da.4.2 941.41 1489.84 2828.91 417.97 -342.19 -210.15

da.4.3 896.88 1391.41 2682.03 510.16 -792.18 -419.53

da.4.4 939.84 1401.56 2775.00 514.06 -543.75 5.47

Table A.10: Steady-state frequency and frequency translation for /da/ (Hz).

Measures are listed as CV.TalkerNum.TokenNum

118


pa.1.1 696.09 1071.88 2442.19 -65.63 13.29 -155.47

pa.1.2 726.56 1058.59 2215.63 -54.69 -3.91 -186.71

pa.1.3 710.94 1074.22 2417.97 -7.81 39.06 -50.78

pa.1.4 726.56 1098.44 2307.03 -39.07 8.60 -150.00

pa.2.1 744.53 1174.22 2292.19 -79.69 45.31 34.38

pa.2.2 750.00 1126.56 2467.19 58.59 25.00 154.69

pa.2.3 736.72 1150.78 2256.25 29.69 29.69 92.19

pa.2.4 720.31 1139.84 2328.91 13.28 42.18 75.00

pa.3.1 888.28 1265.63 3119.53 -131.25 132.82 131.25

pa.3.2 875.78 1255.47 3031.25 -167.19 173.44 54.69

pa.3.3 839.84 1214.84 3071.09 -296.88 78.12 149.21

pa.3.4 867.19 1264.84 3120.31 -171.87 225.78 249.22

pa.4.1 931.25 1283.59 2654.69 -76.56 17.96 -1.56

pa.4.2 971.09 1304.68 2664.06 53.12 -164.07 93.75

pa.4.3 991.41 1310.16 2715.62 65.63 -131.25 12.49

pa.4.4 982.81 1278.13 2683.59 -21.10 -92.96 19.53

Table A.11: Steady-state frequency and frequency translation for /pa/ (Hz).


119


ta.1.1 734.38 1126.56 2316.41 101.57 48.43 -85.93

ta.1.2 750.00 1112.50 2450.00 -46.87 61.72 180.47

ta.1.3 714.84 1058.59 2242.97 -54.69 3.90 -253.12

ta.1.4 714.06 1111.72 2352.34 92.97 -17.19 215.62

ta.2.1 715.63 1175.78 2591.41 28.13 -132.81 228.13

ta.2.2 743.75 1242.19 2368.75 153.91 -117.18 130.47

ta.2.3 708.59 1196.09 2347.66 40.62 -104.69 70.32

ta.2.4 742.19 1207.03 2372.66 97.66 -140.63 239.85

ta.3.1 896.88 1254.69 3104.69 -9.37 -448.44 257.03

ta.3.2 875.00 1228.91 3100.78 39.06 -310.15 300.00

ta.3.3 839.06 1291.41 2962.50 -121.88 -712.50 216.41

ta.3.4 850.78 1303.91 3056.25 -270.31 100.78 157.81

ta.4.1 991.41 1242.19 2700.78 46.10 -351.56 13.28

ta.4.2 987.50 1292.97 2758.59 65.62 -199.22 160.93

ta.4.3 972.66 1321.09 2731.25 -89.84 -18.75 215.62

ta.4.4 956.25 1276.56 2505.47 77.34 -313.28 -84.37

Table A.12: Steady-state frequency and frequency translation for /ta/ (Hz). Mea-


120


fa.1.1 709.38 1050.78 2467.97 72.66 136.72 120.31

fa.1.2 739.84 1093.75 2324.22 181.25 35.16 85.94

fa.1.3 717.19 1042.19 2421.88 -83.59 120.31 78.13

fa.1.4 726.56 1075.78 2367.19 210.93 118.75 62.5

fa.2.1 675.78 1170.31 2278.13 179.69 115.62 149.22

fa.2.2 769.53 1179.69 2343.75 242.19 109.38 218.75

fa.2.3 698.44 1131.36 2343.75 134.03 127.45 101.56

fa.2.4 731.36 1157.81 2493.31 200 83.59 235.5

fa.3.1 899.22 1223.24 3035.94 309.38 8.4 352.35

fa.3.2 833.59 1258.59 2871.88 263.28 12.5 422.66

fa.3.3 825 1279.69 3002.14 242.97 -21.09 361.51

fa.3.4 882.81 1164.84 3031.25 273.43 -34.38 343.75

fa.4.1 1042.97 1042.97 2943.75 542.97 -21.44 865.62

fa.4.2 982.81 1223.24 2665.63 68.75 94.33 196.88

fa.4.3 993.75 993.75 2786.72 99.22 -205.47 212.5

fa.4.4 997.66 1284.18 2735.94 200.79 38.09 505.47

Table A.13: Steady-state frequency and frequency translation for /fa/ (Hz). Mea-


121


sa.1.1 714.84 1069.53 2446.09 273.63 -149.22 -22.66

sa.1.2 707.81 1033.59 2532.03 231.25 -216.41 16.4

sa.1.3 718.75 1081.36 2517.19 277.54 -168.64 36.72

sa.1.4 702.45 1049.22 2517.97 292.29 -150 2.34

sa.2.1 714.84 1189.06 2409.38 271.53 -111.72 -31.83

sa.2.2 678.13 1214.84 2353.13 221.1 -78.13 -37.5

sa.2.3 675.78 1236.72 2464.41 285.15 -103.12 34.72

sa.2.4 654.37 1243.31 2444.53 271.56 -135.6 -4.69

sa.3.1 850.78 1275.78 3031.36 354.69 -407.81 39.17

sa.3.2 871.88 1288.28 3007.03 492.97 -578.91 26.56

sa.3.3 839.84 1250.78 3054.69 437.39 -503.13 74.22

sa.3.4 883.59 1274.22 3017.19 440.28 -561.72 25

sa.4.1 965.63 1341.21 2821.88 625.79 -420.51 32.82

sa.4.2 1027.34 1321.09 2784.18 218.75 -77.35 -24.41

sa.4.3 957.03 1321.09 2792.97 535.15 -350.79 -70.31

sa.4.4 954.37 1297.66 2734.18 639.96 20.32 -19.73

Table A.14: Steady-state frequency and frequency translation for /sa/ (Hz). Mea-


122


va.1.1 725.78 1089.06 2498.44 206.25 -20.32 119.53

va.1.2 718.75 1077.34 2467.19 226.56 22.65 146.88

va.1.3 737.5 1031.36 2389.84 241.41 -46.77 35.47

va.1.4 718.75 1101.56 2494.53 187.5 15.62 84.37

va.2.1 776.56 1152.14 2425.78 323.43 70.11 97.65

va.2.2 751.56 1164.41 2347.66 243.75 117.53 46.88

va.2.3 703.91 1190.63 2392.19 231.25 81.25 126.56

va.2.4 753.13 1143.31 2331.25 241.41 49.56 143.75

va.3.1 880.47 1264.41 3012.5 442.97 -149.65 266.41

va.3.2 852.45 1308.59 2952.45 364.17 -269.54 788.39

va.3.3 817.19 1199.22 2867.97 352.35 -164.06 438.28

va.3.4 879.69 1286.72 3023.24 387.5 -100 562.3

va.4.1 997.66 1348.44 2657.81 696.88 47.66 134.57

va.4.2 928.91 1250.78 2714.41 608.6 141.4 -16.06

va.4.3 942.19 1242.19 2704.69 621.88 48.88 91.41

va.4.4 973.24 1215.63 2686.72 629.49 -206.25 -207.81

Table A.15: Steady-state frequency and frequency translation for /va/ (Hz).


123


za.1.1 730.47 1037.5 2573.24 371.09 -310.16 81.05

za.1.2 734.18 1080.47 2534.18 339.65 -294.53 2.93

za.1.3 727.34 1108.59 2491.21 328.9 -262.5 -20.51

za.1.4 707.81 1094.53 2543.31 348.43 -307.61 4.25

za.2.1 704.37 1233.59 2386.72 321.56 -246.88 -89.84

za.2.2 688.28 1231.36 2397.66 301.56 -268.64 -82.81

za.2.3 703.13 1265.63 2414.41 328.13 -246.09 133.16

za.2.4 689.84 1235.94 2440.63 287.39 -248.24 18.75

za.3.1 818.75 1352.45 2887.5 435.94 -530.36 -65.63

za.3.2 832.81 1394.53 2959.38 434.37 -480.47 -17.18

za.3.3 834.18 1247.66 2944.53 479.81 -584.37 -86.83

za.3.4 854.69 1414.06 2909.37 503.13 -480.47 -113.87

za.4.1 932.81 1333.59 2889.06 671.09 -432.04 -290.63

za.4.2 846.09 1300.78 3010.94 545.31 -332.03 -53.47

za.4.3 962.5 1374.22 2912.5 571.87 -254.69 80.47

za.4.4 975 1341.21 2870.31 576.56 -221.29 26.56

Table A.16: Steady-state frequency and frequency translation for /za/ (Hz). Mea-


124

Token F1onset F2onset F3onset F1offset F2offset F3offset

bi.1.1 296.88 1941.41 2671.88 296.88 1960.94 2886.72

bi.2.1 296.88 1960.94 2550.78 296.88 1960.94 2722.66

bi.3.1 332.03 2113.28 2777.34 332.03 2632.81 3183.59

bi.4.1 328.13 2464.84 2984.38 378.91 2558.59 3179.69

bi.1.2 214.84 2164.06 2695.31 296.88 2019.53 2945.31

bi.2.2 328.13 1914.06 2332.03 328.13 1984.38 2312.5

bi.3.2 355.47 2625 3089.84 355.47 2691.41 3179.69

bi.4.2 253.91 2433.59 2906.25 312.5 2511.72 3144.53

bi.1.3 296.88 1941.41 2656.25 296.88 1941.41 2871.09

bi.2.3 316.41 2003.91 2570.31 316.41 2039.06 2808.59

bi.3.3 347.66 2578.12 2953.13 355.47 2585.94 3261.72

bi.4.3 273.44 2230.47 2996.09 312.5 2531.25 3269.53

bi.1.4 304.69 1941.41 2761.72 304.69 1941.41 2964.84

bi.2.4 320.31 1945.31 2429.69 320.31 2121.09 2640.63

bi.3.4 347.66 2539.06 2980.47 359.38 2539.06 3210.94

bi.4.4 292.97 2281.25 2996.09 355.47 2503.91 3179.69

Table A.17: Onset and offset Frequencies for /bi/ (Hz) Measures are listed as


125


di.1.1 316.41 1867.19 2664.06 316.41 1972.66 2902.34

di.2.1 296.88 1917.97 2597.66 296.88 2023.44 2796.88

di.3.1 398.44 2425.78 3117.19 339.84 2523.44 3195.31

di.4.1 253.91 2500 3085.94 324.22 2578.12 3167.97

di.1.2 285.16 1910.16 2644.53 292.97 1910.16 2835.94

di.2.2 285.16 1976.56 2636.72 285.16 2007.81 2671.88

di.3.2 320.31 2503.91 3109.38 343.75 2539.06 3148.44

di.4.2 277.34 2511.72 2976.56 351.56 2605.47 3218.75

di.1.3 261.72 1917.97 2734.37 277.34 1953.13 2828.13

di.2.3 320.31 1890.63 2484.38 304.69 2011.72 2625

di.3.3 324.22 2511.72 3156.25 335.94 2699.22 3156.25

di.4.3 250 2476.56 3050.78 343.75 2558.59 3132.81

di.1.4 257.81 1898.44 2722.66 304.69 1914.06 2812.5

di.2.4 324.22 1910.16 2562.5 324.22 2058.59 2683.59

di.3.4 285.16 2527.34 3101.56 332.03 2519.53 3101.56

di.4.4 238.28 2734.37 3078.13 332.03 2511.72 3187.5

Table A.18: Onset and offset Frequencies for /di/ (Hz) Measures are listed as


126


pi.1.1 246.09 2082.03 2855.47 304.69 2074.22 2890.62

pi.2.1 230.47 2246.09 2730.47 320.31 2187.5 2718.75

pi.3.1 289.06 2625 3261.72 289.06 2691.41 3203.12

pi.4.1 246.09 2648.44 3144.53 316.41 2589.84 3250

pi.1.2 210.94 2109.38 2828.13 292.97 2078.13 2910.16

pi.2.2 234.37 2191.41 2757.81 324.22 2222.66 2675.78

pi.3.2 281.25 2722.66 3253.91 328.13 2707.03 3140.63

pi.4.2 242.19 2699.22 3218.75 242.19 2601.56 3234.37

pi.1.3 261.72 2140.63 2859.38 312.5 2058.59 2859.38

pi.2.3 101.56 2316.41 2792.97 300.78 2257.81 2832.03

pi.3.3 250 2644.53 3160.16 324.22 2636.72 3132.81

pi.4.3 261.72 2675.78 3230.47 261.72 2648.44 3226.56

pi.1.4 246.09 2078.13 2882.81 316.41 1992.19 2843.75

pi.2.4 191.41 2175.78 2738.28 285.16 2156.25 2738.28

pi.3.4 230.47 2718.75 3191.41 292.97 2675.78 3105.47

pi.4.4 253.91 2593.75 3285.16 269.53 2503.91 3183.59

Table A.19: Onset and offset Frequencies for /pi/ (Hz) Measures are listed as


127


ti.1.1 230.47 2183.59 2871.09 292.97 2078.13 2898.44

ti.2.1 281.25 2152.34 2855.47 328.13 2164.06 2765.63

ti.3.1 300.78 2710.94 3171.87 367.19 2648.44 3175.78

ti.4.1 246.09 2636.72 3121.09 359.38 2632.81 3105.47

ti.1.2 246.09 2035.16 2812.5 312.5 2007.81 2820.31

ti.2.2 269.53 2097.66 2777.34 304.69 2097.66 2734.37

ti.3.2 285.16 2664.06 3152.34 363.28 2585.94 3128.91

ti.4.2 250 2730.47 3214.84 332.03 2617.19 3218.75

ti.1.3 265.63 2242.19 2792.97 300.78 2019.53 2820.31

ti.2.3 277.34 2070.31 2730.47 335.94 2105.47 2722.66

ti.3.3 304.69 2558.59 3121.09 375 2550.78 3113.28

ti.4.3 253.91 2703.13 3160.16 359.38 2585.94 3191.41

ti.1.4 238.28 2074.22 2781.25 320.31 1984.38 2753.91

ti.2.4 207.03 2289.06 2746.09 308.59 2132.81 2742.19

ti.3.4 300.78 2671.88 3195.31 367.19 2656.25 3210.94

ti.4.4 246.09 2652.34 3167.97 367.19 2597.66 3164.06

Table A.20: Onset and offset Frequencies for /ti/ (Hz) Measures are listed as


128


fi.1.1 632.81 1925.78 2527.34 328.13 1914.06 2941.41

fi.2.1 250 1972.66 2546.87 343.75 2183.59 2812.5

fi.3.1 292.97 2792.97 3222.66 332.03 2730.47 3226.56

fi.4.1 289.06 2335.94 2875 371.09 2535.16 3058.59

fi.1.2 281.25 1871.09 2425.78 332.03 1835.94 2910.16

fi.2.2 257.81 2003.91 2609.38 312.5 2082.03 2312.5

fi.3.2 378.91 2523.44 2941.41 386.72 2683.59 3238.28

fi.4.2 277.34 2332.03 3031.25 343.75 2519.53 3109.38

fi.1.3 250 1886.72 2464.84 324.22 1929.69 2957.03

fi.2.3 324.22 1914.06 2375 332.03 1941.41 2671.88

fi.3.3 320.31 2453.13 2878.91 363.28 2605.47 3187.5

fi.4.3 269.53 2414.06 2980.47 371.09 2578.12 3207.03

fi.1.4 261.72 1925.78 2531.25 324.22 1949.22 2910.16

fi.2.4 285.16 2066.41 2558.59 328.13 2050.78 2652.34

fi.3.4 308.59 2476.56 2921.88 339.84 2636.72 3167.97

fi.4.4 257.81 2562.5 3042.97 332.03 2562.5 3164.06

Table A.21: Onset and offset Frequencies for /fi/ (Hz) Measures are listed as


129


si.1.1 277.34 1750 2496.09 332.03 1937.5 2765.63

si.2.1 250 1789.06 2554.69 328.13 2082.03 2593.75

si.3.1 359.38 2359.38 3031.25 386.72 2800.78 3101.56

si.4.1 265.63 2312.5 2972.66 332.03 2550.78 2964.84

si.1.2 238.28 1859.37 2496.09 304.69 1871.09 2496.09

si.2.2 304.69 1761.72 2484.38 332.03 1992.19 2484.38

si.3.2 320.31 2398.44 3183.59 359.38 2710.94 3101.56

si.4.2 277.34 2390.63 3078.13 363.28 2542.97 3019.53

si.1.3 269.53 1816.41 2472.66 320.31 1957.03 2746.09

si.2.3 351.56 1730.47 2457.03 359.38 2050.78 2687.5

si.3.3 300.78 2621.09 3125 351.56 2777.34 3093.75

si.4.3 273.44 2347.66 3042.97 378.91 2585.94 3058.59

si.1.4 277.34 1878.91 2500 292.97 1957.03 2816.41

si.2.4 332.03 1828.13 2488.28 355.47 1816.41 2746.09

si.3.4 359.38 2371.09 3089.84 359.38 2636.72 3003.91

si.4.4 269.53 2320.31 2976.56 367.19 2605.47 3062.5

Table A.22: Onset and offset Frequencies for /si/ (Hz) Measures are listed as


130


vi.1.1 250 1847.66 2371.09 312.5 1875 2792.97

vi.2.1 339.84 1773.44 2191.41 339.84 2003.91 2667.97

vi.3.1 289.06 2191.41 2636.72 312.5 2746.09 3277.34

vi.4.1 214.84 2390.63 2843.75 343.75 2667.97 3003.91

vi.1.2 183.59 1765.63 2351.56 304.69 1902.34 2824.22

vi.2.2 339.84 1777.34 2226.56 339.84 1960.94 2667.97

vi.3.2 304.69 2648.44 2917.97 324.22 2671.88 3226.56

vi.4.2 304.69 2421.88 2906.25 351.56 2558.59 3093.75

vi.1.3 242.19 1859.37 2464.84 312.5 1933.59 2875

vi.2.3 335.94 1859.37 2222.66 335.94 1949.22 2722.66

vi.3.3 296.88 2226.56 2730.47 296.88 2753.91 3222.66

vi.4.3 285.16 2253.91 2730.47 351.56 2605.47 3164.06

vi.1.4 308.59 1863.28 2601.56 308.59 1914.06 2871.09

vi.2.4 332.03 1812.5 2187.5 332.03 2035.16 2648.44

vi.3.4 351.56 2093.75 2660.16 351.56 2507.81 3199.22

vi.4.4 210.94 2566.41 2917.97 351.56 2578.12 3171.87

Table A.23: Onset and offset Frequencies for /vi/ (Hz) Measures are listed as


131


zi.1.1 324.22 1722.66 2394.53 308.59 1914.06 2488.28

zi.2.1 328.13 1710.94 2394.53 328.13 2074.22 2589.84

zi.3.1 320.31 2367.19 3000 320.31 2734.37 3007.81

zi.4.1 238.28 2191.41 2992.19 359.38 2394.53 3066.41

zi.1.2 277.34 1722.66 2445.31 308.59 1941.41 2820.31

zi.2.2 394.53 1800.78 2628.91 347.66 1878.91 2558.59

zi.3.2 308.59 2246.09 3183.59 308.59 2761.72 3179.69

zi.4.2 214.84 2082.03 2789.06 347.66 2519.53 2949.22

zi.1.3 289.06 1750 2363.28 296.88 1914.06 2363.28

zi.2.3 332.03 1718.75 2453.13 339.84 2058.59 2570.31

zi.3.3 328.13 2222.66 3058.59 343.75 2234.38 2980.47

zi.4.3 234.37 2167.97 2945.31 343.75 2488.28 3066.41

zi.1.4 273.44 1679.69 2371.09 312.5 1898.44 2371.09

zi.2.4 328.13 1742.19 2500 351.56 2042.97 2546.87

zi.3.4 320.31 2195.31 2992.19 320.31 2734.37 3085.94

zi.4.4 199.22 2160.16 2761.72 347.66 2437.5 2863.28

Table A.24: Onset and offset Frequencies for /zi/ (Hz) Measures are listed as


132

Token ssF1 ssF2 ssF3 F1 ∆ F2 ∆ F3 ∆

bi.1.1 272.66 2045.31 2971.88 -24.22 103.9 300

bi.2.1 311.72 2111.72 2773.44 14.84 150.78 222.66

bi.3.1 334.38 2732.03 3244.53 2.35 618.75 467.19

bi.4.1 359.38 2664.06 3167.97 31.25 199.22 183.59

bi.1.2 278.13 2083.59 2991.41 63.29 -80.47 296.1

bi.2.2 334.38 2101.56 2674.22 6.25 187.5 342.19

bi.3.2 343.75 2775 3028.13 -11.72 150 -61.71

bi.4.2 357.03 2548.44 3199.22 103.12 114.85 292.97

bi.1.3 272.66 2017.19 2943.75 -24.22 75.78 287.5

bi.2.3 312.5 2175.78 2862.5 -3.91 171.87 292.19

bi.3.3 342.19 2699.22 3174.22 -5.47 121.1 221.09

bi.4.3 335.16 2613.28 3208.59 61.72 382.81 212.5

bi.1.4 275 2058.59 2979.69 -29.69 117.18 217.97

bi.2.4 318.75 2214.06 2698.44 -1.56 268.75 268.75

bi.3.4 325 2710.94 3103.13 -22.66 171.88 122.66

bi.4.4 355.47 2606.25 3204.69 62.5 325 208.6

Table A.25: Steady-state frequency and frequency translation for /bi/ (Hz) Mea-


133


di.1.1 292.97 2060.16 2910.94 -23.44 192.97 246.88

di.2.1 307.81 2139.84 2855.47 10.93 221.87 257.81

di.3.1 323.44 2690.63 3197.66 -75 264.85 80.47

di.4.1 335.94 2614.84 3217.97 82.03 114.84 132.03

di.1.2 281.25 2068.75 2899.22 -3.91 158.59 254.69

di.2.2 304.69 2146.88 2639.06 19.53 170.32 2.34

di.3.2 348.44 2761.72 3248.44 28.13 257.81 139.06

di.4.2 347.66 2585.16 3179.69 70.32 73.44 203.13

di.1.3 281.25 2053.91 2886.72 19.53 135.94 152.35

di.2.3 322.66 2127.34 2632.81 2.35 236.71 148.43

di.3.3 324.22 2753.91 3214.06 0 242.19 57.81

di.4.3 348.44 2564.84 3190.63 98.44 88.28 139.85

di.1.4 285.16 2039.84 2952.34 27.35 141.4 229.68

di.2.4 338.28 2114.84 2744.53 14.06 204.68 182.03

di.3.4 337.5 2675 3221.09 52.34 147.66 119.53

di.4.4 345.31 2554.69 3244.53 107.03 -179.68 166.4

Table A.26: Steady-state frequency and frequency translation for /di/ (Hz) Mea-


134


pi.1.1 283.59 2107.03 2971.09 37.5 25 115.62

pi.2.1 312.5 2303.91 2768.75 82.03 57.82 38.28

pi.3.1 343.75 2716.41 3167.19 54.69 91.41 -94.53

pi.4.1 351.56 2618.75 3225 105.47 -29.69 80.47

pi.1.2 285.94 2093.75 2982.81 75 -15.63 154.68

pi.2.2 316.41 2329.69 2696.09 82.04 138.28 -61.72

pi.3.2 329.69 2745.31 3110.94 48.44 22.65 -142.97

pi.4.2 340.62 2602.34 3187.5 98.43 -96.88 -31.25

pi.1.3 281.25 2071.09 2934.38 19.53 -69.54 75

pi.2.3 328.13 2309.38 2828.91 226.57 -7.03 35.94

pi.3.3 326.56 2705.47 3091.41 76.56 60.94 -68.75

pi.4.3 339.84 2625.78 3224.22 78.12 -50 -6.25

pi.1.4 285.94 2047.66 2955.47 39.85 -30.47 72.66

pi.2.4 299.22 2267.19 2707.03 107.81 91.41 -31.25

pi.3.4 289.84 2689.06 3107.81 59.37 -29.69 -83.6

pi.4.4 314.06 2567.19 3210.94 60.15 -26.56 -74.22

Table A.27: Steady-state frequency and frequency translation for /pi/ (Hz) Mea-


135


ti.1.1 275 2107.03 2980.47 44.53 -76.56 109.38

ti.2.1 317.97 2248.44 2857.03 36.72 96.1 1.56

ti.3.1 344.53 2761.72 3183.59 43.75 50.78 11.72

ti.4.1 364.06 2575 3198.44 117.97 -61.72 77.35

ti.1.2 284.38 2070.31 2971.88 38.29 35.15 159.38

ti.2.2 304.69 2176.56 2707.03 35.16 78.9 -70.31

ti.3.2 360.94 2734.38 3210.16 75.78 70.32 57.82

ti.4.2 355.47 2591.41 3246.09 105.47 -139.06 31.25

ti.1.3 283.59 2046.09 2916.41 17.96 -196.1 123.44

ti.2.3 340.62 2207.03 2702.34 63.28 136.72 -28.13

ti.3.3 342.97 2689.06 3169.53 38.28 130.47 48.44

ti.4.3 366.41 2635.94 3199.22 112.5 -67.19 39.06

ti.1.4 289.84 2005.47 2886.72 51.56 -68.75 105.47

ti.2.4 310.94 2192.97 2867.19 103.91 -96.09 121.1

ti.3.4 351.56 2783.59 3222.66 50.78 111.71 27.35

ti.4.4 365.62 2613.28 3174.22 119.53 -39.06 6.25

Table A.28: Steady-state frequency and frequency translation for /ti/ (Hz) Mea-


136


fi.1.1 289.06 2019.53 2940.63 -343.75 93.75 413.29

fi.2.1 328.13 2245.31 2841.41 78.13 272.65 294.54

fi.3.1 314.84 2797.66 3200 21.87 4.69 -22.66

fi.4.1 366.41 2543.75 3114.84 77.35 207.81 239.84

fi.1.2 281.25 1960.16 2939.06 0 89.07 513.28

fi.2.2 302.34 2226.56 2808.59 44.53 222.65 199.21

fi.3.2 372.66 2798.44 3162.5 -6.25 275 221.09

fi.4.2 355.47 2617.97 3197.66 78.13 285.94 166.41

fi.1.3 280.47 2032.03 2965.63 30.47 145.31 500.79

fi.2.3 331.25 2153.13 2715.63 7.03 239.07 340.63

fi.3.3 339.84 2686.72 3147.66 19.53 233.59 268.75

fi.4.3 359.38 2689.84 3222.66 89.85 275.78 242.19

fi.1.4 282.81 2064.06 2940.63 21.09 138.28 409.38

fi.2.4 341.41 2204.69 2679.69 56.25 138.28 121.1

fi.3.4 338.28 2750.78 3232.81 29.69 274.22 310.93

fi.4.4 345.31 2625.78 3202.34 87.5 63.28 159.37

Table A.29: Steady-state frequency and frequency translation for /fi/ (Hz) Mea-


137


si.1.1 296.88 2006.25 2789.06 19.54 256.25 292.97

si.2.1 320.31 2144.53 2646.88 70.31 355.47 92.19

si.3.1 349.22 2805.47 3214.06 -10.16 446.09 182.81

si.4.1 343.75 2557.81 3033.59 78.12 245.31 60.93

si.1.2 306.25 1956.25 2877.34 67.97 96.88 381.25

si.2.2 309.38 2125 2687.5 4.69 363.28 203.12

si.3.2 331.25 2776.56 3196.88 10.94 378.12 13.29

si.4.2 367.19 2625 3121.88 89.85 234.37 43.75

si.1.3 299.22 2038.28 2782.03 29.69 221.87 309.37

si.2.3 339.84 2200.78 2762.5 -11.72 470.31 305.47

si.3.3 333.59 2787.5 3225 32.81 166.41 100

si.4.3 363.28 2682.03 3160.94 89.84 334.37 117.97

si.1.4 275 2000 2935.94 -2.34 121.09 435.94

si.2.4 337.5 2154.69 2767.97 5.47 326.56 279.69

si.3.4 327.34 2703.13 3180.47 -32.04 332.04 90.63

si.4.4 364.06 2600 3169.53 94.53 279.69 192.97

Table A.30: Steady-state frequency and frequency translation for /si/ (Hz) Mea-


138


vi.1.1 308.59 1984.38 2829.69 58.59 136.72 458.6

vi.2.1 332.03 2137.5 2709.38 -7.81 364.06 517.97

vi.3.1 346.09 2765.63 3258.59 57.03 574.22 621.87

vi.4.1 349.22 2675 3191.41 134.38 284.37 347.66

vi.1.2 289.06 2025.78 2841.41 105.47 260.15 489.85

vi.2.2 332.03 2115.63 2740.63 -7.81 338.29 514.07

vi.3.2 329.69 2739.84 3207.03 25 91.4 289.06

vi.4.2 355.47 2658.59 3214.84 50.78 236.71 308.59

vi.1.3 291.41 2052.34 2889.84 49.22 192.97 425

vi.2.3 318.75 2069.53 2729.69 -17.19 210.16 507.03

vi.3.3 346.88 2825 3191.41 50 598.44 460.94

vi.4.3 359.38 2613.28 3174.22 74.22 359.37 443.75

vi.1.4 292.97 2047.66 2914.06 -15.62 184.38 312.5

vi.2.4 342.97 2126.56 2717.97 10.94 314.06 530.47

vi.3.4 313.28 2589.84 3267.19 -38.28 496.09 607.03

vi.4.4 353.13 2579.69 3173.44 142.19 13.28 255.47

Table A.31: Steady-state frequency and frequency translation for /vi/ (Hz) Mea-


139


zi.1.1 289.06 1999.22 2799.22 -35.16 276.56 404.69

zi.2.1 342.19 2117.19 2657.03 14.06 406.25 262.5

zi.3.1 339.84 2753.91 3217.19 19.53 386.72 217.19

zi.4.1 354.69 2603.13 3120.31 116.41 411.72 128.12

zi.1.2 296.88 1971.88 2708.59 19.54 249.22 263.28

zi.2.2 331.25 2071.88 2658.59 -63.28 271.1 29.68

zi.3.2 328.91 2788.28 3225.78 20.32 542.19 42.19

zi.4.2 360.16 2540.63 3008.59 145.32 458.6 219.53

zi.1.3 292.19 2015.62 2732.81 3.13 265.62 369.53

zi.2.3 340.62 2146.88 2617.19 8.59 428.13 164.06

zi.3.3 336.72 2660.16 3173.44 8.59 437.5 114.85

zi.4.3 335.16 2593.75 3182.81 100.79 425.78 237.5

zi.1.4 296.88 2025.78 2717.19 23.44 346.09 346.1

zi.2.4 347.66 2103.13 2612.5 19.53 360.94 112.5

zi.3.4 308.59 2756.25 3267.97 -11.72 560.94 275.78

zi.4.4 355.47 2489.06 2989.84 156.25 328.9 228.12

Table A.32: Steady-state frequency and frequency translation for /zi/ (Hz) Mea-


140


bu.1.1 304.69 1343.75 2222.66 324.22 1183.59 1949.22

bu.1.2 308.59 1218.75 2003.91 339.84 1218.75 1988.28

bu.1.3 355.47 1191.41 1984.38 371.09 1191.41 1984.38

bu.1.4 367.19 1281.25 1941.41 367.19 1195.31 1929.69

bu.2.1 378.91 1433.59 2273.44 378.91 1343.75 2195.31

bu.2.2 335.94 1339.84 2207.03 335.94 1394.53 2312.5

bu.2.3 367.19 1355.47 2066.41 367.19 1378.91 2066.41

bu.2.4 367.19 1390.63 2179.69 367.19 1437.5 2226.56

bu.3.1 402.34 1773.44 2804.69 402.34 1804.69 2808.59

bu.3.2 378.91 1652.34 2781.25 378.91 1828.13 2796.88

bu.3.3 417.97 1808.59 2917.97 441.41 1945.31 2890.62

bu.3.4 324.22 1777.34 2953.13 371.09 1792.97 2820.31

bu.4.1 277.34 1328.13 2300.78 402.34 1527.34 2601.56

bu.4.2 277.34 1199.22 2160.16 394.53 1257.81 2695.31

bu.4.3 335.94 1378.91 2136.72 421.88 1289.06 2449.22

bu.4.4 320.31 1230.47 2082.03 421.88 1289.06 2585.94

Table A.33: Onset and offset Frequencies for /bu/ (Hz) Measures are listed as


141


du.1.1 316.41 1722.66 2347.66 304.69 1722.66 2164.06

du.1.2 246.09 1707.03 2476.56 296.88 1707.03 2042.97

du.1.3 269.53 1644.53 2410.16 308.59 1726.56 2066.41

du.1.4 265.63 1808.59 2578.12 296.88 1714.84 2312.5

du.2.1 324.22 1781.25 2320.31 324.22 1781.25 2238.28

du.2.2 273.44 1855.47 2535.16 312.5 1792.97 2234.38

du.2.3 265.63 1886.72 2539.06 343.75 1789.06 2289.06

du.2.4 292.97 1835.94 2441.41 300.78 1835.94 2332.03

du.3.1 289.06 2375 2996.09 343.75 2242.19 2992.19

du.3.2 312.5 2078.13 3000 347.66 2144.53 2957.03

du.3.3 328.13 2281.25 3003.91 351.56 2230.47 2976.56

du.3.4 296.88 2183.59 3011.72 320.31 2238.28 2972.66

du.4.1 308.59 2175.78 3042.97 355.47 2125 2574.22

du.4.2 304.69 2039.06 2875 367.19 2121.09 2777.34

du.4.3 242.19 2089.84 2808.59 371.09 2058.59 2636.72

du.4.4 261.72 1910.16 2781.25 367.19 2074.22 2687.5

Table A.34: Onset and offset Frequencies for /du/ (Hz) Measures are listed as


142


pu.1.1 312.5 1230.47 1992.19 339.84 1156.25 2000

pu.1.2 324.22 1257.81 2023.44 351.56 1218.75 2042.97

pu.1.3 265.63 1140.63 2011.72 332.03 1121.09 1964.84

pu.1.4 292.97 1218.75 2085.94 351.56 1179.69 2019.53

pu.2.1 246.09 1312.5 2281.25 386.72 1222.66 2281.25

pu.2.2 230.47 1394.53 2230.47 367.19 1351.56 2230.47

pu.2.3 289.06 1417.97 2191.41 386.72 1417.97 2246.09

pu.2.4 277.34 1386.72 2164.06 375 1312.5 2210.94

pu.3.1 160.16 1824.22 3089.84 304.69 1917.97 2839.84

pu.3.2 292.97 1781.25 2980.47 363.28 1859.37 2968.75

pu.3.3 253.91 1796.88 2863.28 386.72 1898.44 2769.53

pu.3.4 304.69 1726.56 3007.81 386.72 1671.88 2734.37

pu.4.1 250 1453.12 2726.56 390.63 1406.25 2585.94

pu.4.2 246.09 1359.37 2429.69 402.34 1269.53 2550.78

pu.4.3 273.44 1519.53 2527.34 421.88 1343.75 2488.28

pu.4.4 257.81 1425.78 2945.31 398.44 1328.13 2546.87

Table A.35: Onset and offset Frequencies for /pu/ (Hz) Measures are listed as


143


tu.1.1 320.31 1722.66 2230.47 320.31 1683.59 2042.97

tu.1.2 269.53 1710.94 2250 332.03 1718.75 2140.63

tu.1.3 238.28 1738.28 2285.16 328.13 1687.5 2074.22

tu.1.4 289.06 1683.59 2273.44 328.13 1664.06 2207.03

tu.2.1 261.72 1753.91 2394.53 363.28 1734.38 2320.31

tu.2.2 238.28 1781.25 2324.22 347.66 1753.91 2210.94

tu.2.3 324.22 1789.06 2152.34 355.47 1718.75 2230.47

tu.2.4 292.97 1710.94 2347.66 343.75 1640.63 2304.69

tu.3.1 250 2289.06 3109.38 375 2101.56 2902.34

tu.3.2 332.03 2183.59 2929.69 371.09 2148.44 2824.22

tu.3.3 339.84 2183.59 2945.31 367.19 2109.38 2722.66

tu.3.4 300.78 2179.69 3015.63 394.53 2097.66 2796.88

tu.4.1 257.81 1921.88 2675.78 386.72 1921.88 2574.22

tu.4.2 242.19 1980.47 2859.38 355.47 1988.28 2656.25

tu.4.3 238.28 1847.66 2949.22 375 1667.97 2601.56

tu.4.4 195.31 1976.56 2625 386.72 1812.5 2597.66

Table A.36: Onset and offset Frequencies for /tu/ (Hz) Measures are listed as


144


fu.1.1 320.31 1292.97 1972.66 359.38 1292.97 1957.03

fu.1.2 308.59 1265.63 1964.84 367.19 1269.53 1964.84

fu.1.3 289.06 1316.41 2261.72 347.66 1167.97 1937.5

fu.1.4 351.56 1199.22 2082.03 351.56 1187.5 1945.31

fu.2.1 296.88 1312.5 2210.94 378.91 1304.69 2210.94

fu.2.2 343.75 1300.78 2148.44 367.19 1300.78 2312.5

fu.2.3 312.5 1199.22 2277.34 375 1265.63 2234.38

fu.2.4 363.28 1371.09 2164.06 382.81 1417.97 2171.88

fu.3.1 355.47 1796.88 3003.91 390.63 1628.91 2761.72

fu.3.2 335.94 1812.5 3125 378.91 1792.97 2867.19

fu.3.3 359.38 1640.63 2906.25 394.53 1648.44 2746.09

fu.3.4 382.81 1675.78 3101.56 398.44 1757.81 2777.34

fu.4.1 269.53 1355.47 2937.5 417.97 1250 2636.72

fu.4.2 261.72 1367.19 2144.53 406.25 1320.31 2582.03

fu.4.3 273.44 1343.75 2042.97 402.34 1265.63 2519.53

fu.4.4 273.44 1558.59 2453.13 437.5 1488.28 2566.41

Table A.37: Onset and offset Frequencies for /fu/ (Hz) Measures are listed as


145


su.1.1 253.91 1539.06 2390.63 304.69 1679.69 2089.84

su.1.2 285.16 1640.63 2394.53 335.94 1609.37 2187.5

su.1.3 273.44 1621.09 2292.97 332.03 1613.28 2148.44

su.1.4 257.81 1683.59 2476.56 332.03 1621.09 2203.13

su.2.1 304.69 1726.56 2691.41 351.56 1566.41 2144.53

su.2.2 324.22 1585.94 2429.69 324.22 1554.69 2292.97

su.2.3 328.13 1574.22 2500 328.13 1644.53 2421.88

su.2.4 308.59 1593.75 2167.97 339.84 1648.44 2257.81

su.3.1 343.75 1984.38 3000 394.53 1984.38 2957.03

su.3.2 359.38 2003.91 3097.66 367.19 2031.25 2914.06

su.3.3 339.84 2105.47 3070.31 406.25 2027.34 2910.16

su.3.4 328.13 1992.19 3031.25 351.56 2003.91 2917.97

su.4.1 285.16 1984.38 2925.78 398.44 1929.69 2601.56

su.4.2 261.72 1890.63 2832.03 402.34 1804.69 2667.97

su.4.3 300.78 1964.84 2765.63 406.25 1914.06 2546.87

su.4.4 281.25 1933.59 2796.88 394.53 1859.37 2613.28

Table A.38: Onset and offset Frequencies for /su/ (Hz) Measures are listed as


146


vu.1.1 300.78 1125 1960.94 347.66 1132.81 1972.66

vu.1.2 390.63 1148.44 1894.53 332.03 1117.19 1910.16

vu.1.3 273.44 1195.31 1937.5 273.44 1195.31 1960.94

vu.1.4 347.66 1132.81 2062.5 347.66 1132.81 1976.56

vu.2.1 343.75 1269.53 2199.22 375 1339.84 2179.69

vu.2.2 339.84 1386.72 2312.5 355.47 1324.22 2148.44

vu.2.3 304.69 1433.59 2324.22 398.44 1414.06 2160.16

vu.2.4 371.09 1464.84 2136.72 371.09 1359.37 2203.13

vu.3.1 398.44 1804.69 2609.38 414.06 1906.25 2878.91

vu.3.2 398.44 1683.59 2648.44 406.25 1773.44 2921.88

vu.3.3 289.06 1621.09 2457.03 371.09 1710.94 2664.06

vu.3.4 390.63 1707.03 2597.66 378.91 1691.41 2597.66

vu.4.1 281.25 1320.31 2238.28 347.66 1320.31 2523.44

vu.4.2 277.34 1371.09 2902.34 359.38 1273.44 2683.59

vu.4.3 265.63 1640.63 2953.13 351.56 1285.16 2578.12

vu.4.4 277.34 1539.06 2269.53 375 1410.16 2460.94

Table A.39: Onset and offset Frequencies for /vu/ (Hz) Measures are listed as


147


zu.1.1 273.44 1585.94 2328.13 300.78 1578.13 2300.78

zu.1.2 292.97 1621.09 2355.47 304.69 1597.66 2312.5

zu.1.3 289.06 1621.09 2378.91 320.31 1621.09 2363.28

zu.1.4 285.16 1550.78 2308.59 312.5 1558.59 2281.25

zu.2.1 339.84 1636.72 2500 339.84 1671.88 2429.69

zu.2.2 339.84 1609.37 2316.41 347.66 1644.53 2332.03

zu.2.3 343.75 1601.56 2460.94 351.56 1656.25 2335.94

zu.2.4 320.31 1570.31 2367.19 320.31 1570.31 2390.63

zu.3.1 316.41 2015.62 3054.69 316.41 2007.81 3031.25

zu.3.2 289.06 2039.06 3085.94 324.22 2148.44 3066.41

zu.3.3 292.97 2132.81 3242.19 343.75 2105.47 2949.22

zu.3.4 339.84 2039.06 3085.94 339.84 2148.44 2976.56

zu.4.1 222.66 1906.25 2683.59 367.19 2015.62 2828.13

zu.4.2 230.47 1722.66 2742.19 285.16 1898.44 2617.19

zu.4.3 308.59 2203.13 2714.84 308.59 2003.91 2664.06

zu.4.4 238.28 1929.69 2980.47 359.38 2050.78 2738.28

Table A.40: Onset and offset Frequencies for /zu/ (Hz) Measures are listed as


148


bu.1.1 292.97 1133.59 2023.44 -11.72 -210.16 -199.22

bu.1.2 303.91 1182.03 1983.59 -4.68 -36.72 -20.32

bu.1.3 308.59 1123.44 2053.91 -46.88 -67.97 69.53

bu.1.4 320.31 1153.91 1916.41 -46.88 -127.34 -25

bu.2.1 371.09 1365.63 2286.72 -7.82 -67.96 13.28

bu.2.2 364.84 1411.72 2235.94 28.9 71.88 28.91

bu.2.3 356.25 1314.84 2127.34 -10.94 -40.63 60.93

bu.2.4 345.31 1302.34 2214.84 -21.88 -88.29 35.15

bu.3.1 421.88 1883.59 2817.19 19.54 110.15 12.5

bu.3.2 401.56 1882.81 2867.97 22.65 230.47 86.72

bu.3.3 400.78 1952.34 2923.44 -17.19 143.75 5.47

bu.3.4 374.22 1861.72 2923.44 50 84.38 -29.69

bu.4.1 417.97 1475 2610.16 140.63 146.87 309.38

bu.4.2 396.09 1285.94 2754.69 118.75 86.72 594.53

bu.4.3 412.5 1265.63 2641.41 76.56 -113.28 504.69

bu.4.4 407.81 1284.38 2645.31 87.5 53.91 563.28

Table A.41: Steady-state frequency and frequency translation for /bu/ (Hz).


149


du.1.1 303.91 1427.34 2030.47 -12.5 -295.32 -317.19

du.1.2 296.88 1470.31 2004.69 50.79 -236.72 -471.87

du.1.3 309.38 1797.66 2038.28 39.85 153.13 -371.88

du.1.4 319.53 1720.31 2157.03 53.9 -88.28 -421.09

du.2.1 354.69 1560.16 2153.13 30.47 -221.09 -167.18

du.2.2 346.88 1692.97 2255.47 73.44 -162.5 -279.69

du.2.3 363.28 1745.31 2276.56 97.65 -141.41 -262.5

du.2.4 345.31 1746.88 2264.06 52.34 -89.06 -177.35

du.3.1 367.97 2184.38 2968.75 78.91 -190.62 -27.34

du.3.2 360.94 2184.38 2918.75 48.44 106.25 -81.25

du.3.3 362.5 2237.5 2987.5 34.37 -43.75 -16.41

du.3.4 339.06 2174.22 2914.06 42.18 -9.37 -97.66

du.4.1 398.44 1939.06 2627.34 89.85 -236.72 -415.63

du.4.2 386.72 1872.66 2720.31 82.03 -166.4 -154.69

du.4.3 378.91 1823.44 2635.16 136.72 -266.4 -173.43

du.4.4 375 1880.47 2656.25 113.28 -29.69 -125

Table A.42: Steady-state frequency and frequency translation for /du/ (Hz).


150


pu.1.1 312.5 1052.34 2006.25 0 -178.13 14.06

pu.1.2 296.88 1078.13 2061.72 -27.34 -179.68 38.28

pu.1.3 299.22 1001.56 1987.5 33.59 -139.07 -24.22

pu.1.4 296.88 1053.13 2026.56 3.91 -165.62 -59.38

pu.2.1 373.44 1081.25 2348.44 127.35 -231.25 67.19

pu.2.2 368.75 1243.75 2310.94 138.28 -150.78 80.47

pu.2.3 361.72 1327.34 2302.34 72.66 -90.63 110.93

pu.2.4 382.03 1224.22 2269.53 104.69 -162.5 105.47

pu.3.1 425.78 1892.97 2898.44 265.62 68.75 -191.4

pu.3.2 397.66 1878.91 2973.44 104.69 97.66 -7.03

pu.3.3 382.81 1861.72 2714.84 128.9 64.84 -148.44

pu.3.4 378.91 1589.06 2718.75 74.22 -137.5 -289.06

pu.4.1 406.25 1364.06 2600.78 156.25 -89.06 -125.78

pu.4.2 407.81 1272.66 2652.34 161.72 -86.71 222.65

pu.4.3 406.25 1271.09 2412.5 132.81 -248.44 -114.84

pu.4.4 396.87 1210.16 2574.22 139.06 -215.62 -371.09

Table A.43: Steady-state frequency and frequency translation for /pu/ (Hz).


151


tu.1.1 285.16 1314.06 2055.47 -35.15 -408.6 -175

tu.1.2 296.09 1366.41 2075.78 26.56 -344.53 -174.22

tu.1.3 289.06 1437.5 2000 50.78 -300.78 -285.16

tu.1.4 289.06 1438.28 1912.5 0 -245.31 -360.94

tu.2.1 357.03 1574.22 2260.94 95.31 -179.69 -133.59

tu.2.2 346.09 1517.97 2204.69 107.81 -263.28 -119.53

tu.2.3 348.44 1562.5 2175.78 24.22 -226.56 23.44

tu.2.4 351.56 1424.22 2183.59 58.59 -286.72 -164.07

tu.3.1 372.66 2068.75 2864.06 122.66 -220.31 -245.32

tu.3.2 361.72 2045.31 2781.25 29.69 -138.28 -148.44

tu.3.3 343.75 2137.5 2763.28 3.91 -46.09 -182.03

tu.3.4 375 2048.44 2804.69 74.22 -131.25 -210.94

tu.4.1 382.81 1660.94 2682.03 125 -260.94 6.25

tu.4.2 386.72 1721.88 2647.66 144.53 -258.59 -211.72

tu.4.3 379.69 1436.72 2639.84 141.41 -410.94 -309.38

tu.4.4 386.72 1719.53 2483.59 191.41 -257.03 -141.41

Table A.44: Steady-state frequency and frequency translation for /tu/ (Hz). Mea-


152


fu.1.1 330.47 1218.75 1984.38 10.16 -74.22 11.72

fu.1.2 325 1242.97 1981.25 16.41 -22.66 16.41

fu.1.3 298.44 1141.41 2005.47 9.38 -175 -256.25

fu.1.4 304.69 1076.56 1978.13 -46.87 -122.66 -103.9

fu.2.1 377.34 1309.38 2343.75 80.46 -3.12 132.81

fu.2.2 375.78 1278.13 2260.16 32.03 -22.65 111.72

fu.2.3 382.03 1260.94 2312.5 69.53 61.72 35.16

fu.2.4 389.84 1327.34 2296.88 26.56 -43.75 132.82

fu.3.1 406.25 1808.59 2878.13 50.78 11.71 -125.78

fu.3.2 372.66 1840.63 2866.41 36.72 28.13 -258.59

fu.3.3 387.5 1825.78 2847.66 28.12 185.15 -58.59

fu.3.4 389.06 1748.44 2909.38 6.25 72.66 -192.18

fu.4.1 414.06 1266.41 2607.81 144.53 -89.06 -329.69

fu.4.2 402.34 1209.38 2547.66 140.62 -157.81 403.13

fu.4.3 410.16 1220.31 2528.91 136.72 -123.44 485.94

fu.4.4 421.88 1499.22 2744.53 148.44 -59.37 291.4

Table A.45: Steady-state frequency and frequency translation for /fu/ (Hz). Mea-


153


su.1.1 290.63 1274.22 2050 36.72 -264.84 -340.63

su.1.2 306.25 1293.75 2043.75 21.09 -346.88 -350.78

su.1.3 300.78 1349.22 2011.72 27.34 -271.87 -281.25

su.1.4 320.31 1433.59 2041.41 62.5 -250 -435.15

su.2.1 353.13 1456.25 2160.16 48.44 -270.31 -531.25

su.2.2 365.63 1488.28 2189.06 41.41 -97.66 -240.63

su.2.3 357.03 1521.88 2300.78 28.9 -52.34 -199.22

su.2.4 350 1542.97 2219.53 41.41 -50.78 51.56

su.3.1 382.81 2063.28 2900 39.06 78.9 -100

su.3.2 373.44 2026.56 2935.94 14.06 22.65 -161.72

su.3.3 394.53 2087.5 2867.19 54.69 -17.97 -203.12

su.3.4 369.53 2067.97 2885.94 41.4 75.78 -145.31

su.4.1 410.16 1713.28 2604.69 125 -271.1 -321.09

su.4.2 386.72 1755.47 2712.5 125 -135.16 -119.53

su.4.3 403.91 1682.03 2621.09 103.13 -282.81 -144.54

su.4.4 390.63 1857.81 2642.97 109.38 -75.78 -153.91

Table A.46: Steady-state frequency and frequency translation for /su/ (Hz). Mea-


154


vu.1.1 312.5 1059.38 1985.94 11.72 -65.62 25

vu.1.2 315.63 1097.66 1924.22 -75 -50.78 29.69

vu.1.3 331.25 1123.44 1960.16 57.81 -71.87 22.66

vu.1.4 299.22 1107.03 2046.09 -48.44 -25.78 -16.41

vu.2.1 366.41 1310.94 2263.28 22.66 41.41 64.06

vu.2.2 372.66 1328.91 2230.47 32.82 -57.81 -82.03

vu.2.3 385.16 1412.5 2229.69 80.47 -21.09 -94.53

vu.2.4 398.44 1389.06 2254.69 27.35 -75.78 117.97

vu.3.1 446.09 1962.5 2875.78 47.65 157.81 266.4

vu.3.2 446.88 1942.97 2922.66 48.44 259.38 274.22

vu.3.3 393.75 1796.09 2888.28 104.69 175 431.25

vu.3.4 444.53 1908.59 2720.31 53.9 201.56 122.65

vu.4.1 390.63 1188.28 2594.53 109.38 -132.03 356.25

vu.4.2 382.81 1253.13 2714.84 105.47 -117.96 -187.5

vu.4.3 388.28 1218.75 2577.34 122.65 -421.88 -375.79

vu.4.4 397.66 1402.34 2506.25 120.32 -136.72 236.72

Table A.47: Steady-state frequency and frequency translation for /vu/ (Hz).


155


zu.1.1 305.47 1387.5 2118.75 32.03 -198.44 -209.38

zu.1.2 313.28 1426.56 2125.78 20.31 -194.53 -229.69

zu.1.3 325 1560.16 2150 35.94 -60.93 -228.91

zu.1.4 306.25 1300.78 2046.88 21.09 -250 -261.71

zu.2.1 364.84 1601.56 2300 25 -35.16 -200

zu.2.2 364.06 1534.38 2259.38 24.22 -74.99 -57.03

zu.2.3 355.47 1522.66 2296.88 11.72 -78.9 -164.06

zu.2.4 347.66 1595.31 2390.63 27.35 25 23.44

zu.3.1 380.47 2094.53 2966.41 64.06 78.91 -88.28

zu.3.2 360.94 2171.09 2972.66 71.88 132.03 -113.28

zu.3.3 380.47 2092.97 2955.47 87.5 -39.84 -286.72

zu.3.4 412.5 2167.19 2985.94 72.66 128.13 -100

zu.4.1 386.72 1821.09 2650 164.06 -85.16 -33.59

zu.4.2 392.19 1928.13 2609.38 161.72 205.47 -132.81

zu.4.3 382.81 1806.25 2630.47 74.22 -396.88 -84.37

zu.4.4 382.81 1797.66 2828.91 144.53 -132.03 -151.56

Table A.48: Steady-state frequency and frequency translation for /zu/ (Hz).


156

APPENDIX B

Duration Measurements

Durational measurements for all CV’s are listed in this appendix. For stops,

measurements include the duration of voiced bar and the burst segment as well

as the VOT (voice onset time). For fricatives, measurements include the duration

of voiced bar, VOT (voice onset time), noise duration and the aspiration. Note

that noise duration and aspiration duration for fricatives will sum up to the

duration of VOT. Measurements are in ms.

157

Voice Bar Burst Duration VOT

ba.1.1 0 3.3 7.3

ba.1.2 0 3.9 6.9

ba.1.3 0 3.6 6.7

ba.1.4 0 2.8 7.4

ba.2.1 0 3.4 7.7

ba.2.2 0 3.6 10.8

ba.2.3 0 3.6 6.2

ba.2.4 0 4.2 10.5

ba.3.1 102.9 0 102.9

ba.3.2 100.2 0 100.2

ba.3.3 87.2 0 87.2

ba.3.4 0 3.6 3.8

ba.4.1 104.6 3.1 11

ba.4.2 91 4.7 5.3

ba.4.3 91 5.6 4.7

ba.4.4 105.2 8.3 3.6

Table B.1: Voiced Bar Duration,Burst Duration and VOT for /ba/ (ms)

158


da.1.1 0 6.6 7.4

da.1.2 0 5.3 8.3

da.1.3 0 6.9 7.4

da.1.4 0 3.6 9.4

da.2.1 0 7.1 18.9

da.2.2 0 5.1 7.2

da.2.3 0 4.6 8.5

da.2.4 0 5.2 8.2

da.3.1 87.8 5 6.1

da.3.2 73.3 6.6 4.8

da.3.3 86 4.6 3.1

da.3.4 91.9 6.4 2.4

da.4.1 73.7 3 7.2

da.4.2 95.4 5.9 12.5

da.4.3 700.9 5.3 7.3

da.4.4 74.7 3.1 3.9

Table B.2: Voiced Bar Duration,Burst Duration and VOT for /da/ (ms)

159


pa.1.1 0 9.3 50.1

pa.1.2 0 9.9 49.9

pa.1.3 0 7.2 51.3

pa.1.4 0 7.1 75.8

pa.2.1 0 6.3 108.2

pa.2.2 0 19.4 67.6

pa.2.3 0 5.4 81

pa.2.4 0 5 83.5

pa.3.1 0 4.8 57.8

pa.3.2 0 7.3 32.4

pa.3.3 0 10.8 28.5

pa.3.4 0 7.3 36.4

pa.4.1 0 2.7 91.7

pa.4.2 0 8.3 67.7

pa.4.3 0 5.6 81.9

pa.4.4 0 12.9 73.1

Table B.3: Voiced Bar Duration,Burst Duration and VOT for /pa/ (ms)

160


ta.1.1 0 20.4 63.5

ta.1.2 0 19.9 47.1

ta.1.3 0 17.3 59.9

ta.1.4 0 17.8 37.3

ta.2.1 0 29.4 48.7

ta.2.2 0 49.8 47.9

ta.2.3 0 32.7 60.7

ta.2.4 0 39.4 43.1

ta.3.1 0 35.9 29.8

ta.3.2 0 38.2 17.7

ta.3.3 0 44.3 19.8

ta.3.4 0 39 47

ta.4.1 0 20.3 74.4

ta.4.2 0 23.7 71.8

ta.4.3 0 22.6 68.9

ta.4.4 0 32.8 52.9

Table B.4: Voiced Bar Duration,Burst Duration and VOT for /ta/ (ms)

161

Voice Bar Burst Duration Duration VOT

bi.1.1 0 2.4 5.3

bi.1.2 0 3.2 12.4

bi.1.3 60.6 2.7 2.9

bi.1.4 0 11.3 7.1

bi.2.1 0 3.1 4.3

bi.2.2 11.1 4.4 7.8

bi.2.3 86.9 3.1 2.9

bi.2.4 123.4 7.4 3.8

bi.3.1 0 2.7 6.4

bi.3.2 0 3.9 4.9

bi.3.3 85.6 5.1 3.6

bi.3.4 64.6 4.1 10.9

bi.4.1 0 4.6 5.7

bi.4.2 0 3.3 5.1

bi.4.3 94.3 2 2.1

bi.4.4 90.9 6.8 3.2

Table B.5: Voiced Bar Duration,Burst Duration and VOT for /bi/ (ms)

162


di.1.1 0 9.9 7.7

di.1.2 0 6.4 2.9

di.1.3 62.7 3.6 3.7

di.1.4 102.1 8.2 11.3

di.2.1 0 14 4.4

di.2.2 0 8.3 3.8

di.2.3 100.3 8.2 9.3

di.2.4 172.6 5.4 8.8

di.3.1 0 10.8 8.1

di.3.2 0 6.9 5

di.3.3 94 10.6 10.8

di.3.4 162.6 5.8 10.3

di.4.1 0 9.6 10.5

di.4.2 0 4.6 6.6

di.4.3 86.4 9.8 13.1

di.4.4 107.8 7.1 15.1

Table B.6: Voiced Bar Duration,Burst Duration and VOT for /di/ (ms)

163


pi.1.1 0 2 58.9

pi.1.2 0 3.2 103.4

pi.1.3 0 3.1 71.6

pi.1.4 0 4.6 91.3

pi.2.1 0 3.2 68.1

pi.2.2 0 3.8 105

pi.2.3 0 4.6 85.8

pi.2.4 0 4.7 90.3

pi.3.1 0 4.1 51

pi.3.2 0 3.4 106.3

pi.3.3 0 3.4 90.6

pi.3.4 0 12.4 98.8

pi.4.1 0 4.1 50.8

pi.4.2 0 5.2 97.2

pi.4.3 0 3.6 119.6

pi.4.4 0 2.6 96.5

Table B.7: Voiced Bar Duration,Burst Duration and VOT for /pi/ (ms)

164


ti.1.1 0 6.4 55.2

ti.1.2 0 9.9 87.2

ti.1.3 0 8.4 74.9

ti.1.4 0 4.1 87.9

ti.2.1 0 6.4 78.3

ti.2.2 0 8.9 100.3

ti.2.3 0 6 63.2

ti.2.4 0 4.3 110.4

ti.3.1 0 10.3 75

ti.3.2 0 3.9 97.9

ti.3.3 0 10.9 68.3

ti.3.4 0 6 88.9

ti.4.1 0 5.1 60.1

ti.4.2 0 7.7 98.9

ti.4.3 0 9.4 80.8

ti.4.4 0 3.9 90.6

Table B.8: Voiced Bar Duration,Burst Duration and VOT for /ti/ (ms)

165

Voiced Bar Burst Duration Duration VOT

bu.1.1 0 4.3 8.9

bu.1.2 0 2.3 17.7

bu.1.3 91.3 2.9 9.3

bu.1.4 99.5 10.6 4

bu.2.1 0 4.6 7.3

bu.2.2 0 3.4 10

bu.2.3 95.2 4.7 3.6

bu.2.4 130.6 4.3 8.1

bu.3.1 0 3.9 8.7

bu.3.2 0 4.9 5.6

bu.3.3 87.2 2.3 3.4

bu.3.4 0 7.1 12

bu.4.1 0 5 9.4

bu.4.2 0 3.2 9.1

bu.4.3 47.9 6.5 9.2

bu.4.4 0 9.4 11.9

Table B.9: Voiced Bar Duration,Burst Duration and VOT for /bu/ (ms)

166


du.1.1 0 2.7 9.1

du.1.2 0 5.9 6.8

du.1.3 91.6 5.7 23.4

du.1.4 85.8 2.1 13.6

du.2.1 0 6.1 14.1

du.2.2 0 4.6 11.6

du.2.3 78.3 4.8 7.2

du.2.4 117.6 3.5 6.3

du.3.1 0 6.6 18.2

du.3.2 0 6.3 8

du.3.3 89.1 8.1 10.6

du.3.4 114.3 1.9 15.8

du.4.1 0 4.5 17.2

du.4.2 0 5.4 6.4

du.4.3 97.3 3.2 11.2

du.4.4 101.9 3.7 14.2

Table B.10: Voiced Bar Duration,Burst Duration and VOT for /du/ (ms)

167


pu.1.1 0 4.5 50.8

pu.1.2 0 4.3 109

pu.1.3 0 7.2 58.4

pu.1.4 0 4.2 77.9

pu.2.1 0 2.9 74.7

pu.2.2 0 4.5 65.3

pu.2.3 0 6.1 65.8

pu.2.4 0 5 96.2

pu.3.1 0 3.2 68

pu.3.2 0 3.7 74.4

pu.3.3 0 6.6 79.8

pu.3.4 0 4.6 94.6

pu.4.1 0 3.7 69.1

pu.4.2 0 4.4 99.1

pu.4.3 0 5.3 94.6

pu.4.4 0 3.3 93.1

Table B.11: Voiced Bar Duration,Burst Duration and VOT for /pu/ (ms)

168


tu.1.1 0 7.6 77.9

tu.1.2 0 2.4 97.1

tu.1.3 0 9.8 93

tu.1.4 0 2.5 84.5

tu.2.1 0 6.3 61.4

tu.2.2 0 4.1 102.8

tu.2.3 0 5.4 103.2

tu.2.4 0 3.3 99

tu.3.1 0 12.1 72.4

tu.3.2 0 3.4 111.8

tu.3.3 0 7.3 90.4

tu.3.4 0 6.2 83

tu.4.1 0 7.7 66.5

tu.4.2 0 1.8 98.8

tu.4.3 0 5.3 89.9

tu.4.4 0 10.6 86.9

Table B.12: Voiced Bar Duration,Burst Duration and VOT for /tu/ (ms)

169

Voiced Bar VOT Noise Duration Aspiration

fa.1.1 0 112.9 109.3 3.6

fa.1.2 0 169.1 169.1 0

fa.1.3 0 180.1 180.1 0

fa.1.4 0 182.1 175.3 6.8

fa.2.1 0 135.7 134 1.7

fa.2.2 0 180.4 180.4 0

fa.2.3 0 59.9 53 6.9

fa.2.4 0 204.2 191.5 12.7

fa.3.1 0 150.7 150.7 0

fa.3.2 0 122.7 122.7 0

fa.3.3 0 168.9 162.3 6.6

fa.3.4 0 145.1 126.7 18.4

fa.4.1 0 132.1 132.1 0

fa.4.2 0 175.8 168.7 7.1

fa.4.3 0 42.3 37.5 4.8

fa.4.4 0 178.1 168.8 9.3

Table B.13: Voiced bar duration,VOT, noise duration and aspiration for /fa/

(ms)

170


sa.1.1 0 157.2 157.2 0

sa.1.2 0 176.6 176.6 0

sa.1.3 0 176.8 176.8 0

sa.1.4 0 186.5 168.9 17.6

sa.2.1 0 169 169 0

sa.2.2 0 195.6 195.6 0

sa.2.3 0 162.3 162.3 0

sa.2.4 0 180.8 169.5 11.3

sa.3.1 0 157.8 157.8 0

sa.3.2 0 193.3 193.3 0

sa.3.3 0 193.8 193.8 0

sa.3.4 0 182.3 173.8 8.5

sa.4.1 0 162.6 154.3 8.3

sa.4.2 0 190.8 190.8 0

sa.4.3 0 175 175 0

sa.4.4 0 200.2 191.3 8.9

Table B.14: Voiced bar duration,VOT, noise duration and aspiration for /sa/

(ms)

171


va.1.1 111 111 0 111

va.1.2 99.7 99.7 0 99.7

va.1.3 123.6 123.6 0 123.6

va.1.4 96.4 96.4 0 96.4

va.2.1 102.1 102.1 0 102.1

va.2.2 152.6 152.6 0 152.6

va.2.3 132.7 132.7 0 132.7

va.2.4 24.4 96.2 0 96.2

va.3.1 97.3 97.3 0 97.3

va.3.2 129.2 129.2 0 129.2

va.3.3 95.6 95.6 0 95.6

va.3.4 146.5 146.5 0 146.5

va.4.1 121.6 121.6 0 121.6

va.4.2 109.6 109.6 0 109.6

va.4.3 65.8 65.8 0 65.8

va.4.4 143.3 143.3 0 143.3

Table B.15: Voiced bar duration,VOT, noise duration and aspiration for /va/

(ms)

172


za.1.1 131.1 131.1 131.1 0

za.1.2 157.9 157.9 157.9 0

za.1.3 123.9 123.9 123.9 0

za.1.4 114.4 114.4 114.4 0

za.2.1 128.4 128.4 128.4 0

za.2.2 158.9 158.9 158.9 0

za.2.3 136.4 136.4 136.4 0

za.2.4 144.4 144.4 144.4 0

za.3.1 132.8 132.8 132.8 0

za.3.2 143.9 143.9 143.9 0

za.3.3 93.8 93.8 93.8 0

za.3.4 143.8 143.8 143.8 0

za.4.1 137.3 137.3 137.3 0

za.4.2 144.5 144.5 144.5 0

za.4.3 126 126 126 0

za.4.4 149.9 149.9 149.9 0

Table B.16: Voiced bar duration,VOT, noise duration and aspiration for /za/

(ms)

173


fi.1.1 0 155.6 147.4 8.2

fi.1.2 0 160.4 146.8 13.6

fi.1.3 0 181.2 172.1 9.1

fi.1.4 0 164.6 150.4 14.2

fi.2.1 0 124.5 118.7 5.8

fi.2.2 0 205.3 193.4 11.9

fi.2.3 0 196.2 185.3 10.9

fi.2.4 0 202.2 189.9 12.3

fi.3.1 0 161.9 151.6 10.3

fi.3.2 0 206 195.4 10.6

fi.3.3 0 204.7 197.4 7.3

fi.3.4 0 171 158.6 12.4

fi.4.1 0 162 157 5

fi.4.2 0 191.5 175.9 15.6

fi.4.3 0 174.6 167.3 7.3

fi.4.4 0 209.7 193.4 16.3

Table B.17: Voiced bar duration,VOT, noise duration and aspiration for /fi/ (ms)

174


si.1.1 0 189.8 183.4 6.4

si.1.2 0 254.1 250 4.1

si.1.3 0 191.8 184.1 7.7

si.1.4 0 231.4 219.4 12

si.2.1 0 189.2 186.1 3.1

si.2.2 0 223.5 213.7 9.8

si.2.3 0 214.9 201.8 13.1

si.2.4 0 200.4 186.9 13.5

si.3.1 0 236.1 231.3 4.8

si.3.2 0 212.8 210.6 2.2

si.3.3 0 227.4 225.5 1.9

si.3.4 0 182.3 160.6 21.7

si.4.1 0 224.9 219.3 5.6

si.4.2 0 222.1 220.9 1.2

si.4.3 0 177.2 174.8 2.4

si.4.4 0 186.6 164.3 22.3

Table B.18: Voiced bar duration,VOT, noise duration and aspiration for /si/

(ms)

175


vi.1.1 113.7 113.4 113.4 0

vi.1.2 102.9 102.9 102.9 0

vi.1.3 133.1 133.1 133.1 0

vi.1.4 153.8 153.8 153.8 0

vi.2.1 151.8 151.8 151.8 0

vi.2.2 146.8 146.8 146.8 0

vi.2.3 128.1 128.1 128.1 0

vi.2.4 155.4 155.4 155.4 0

vi.3.1 118.8 118.8 118.8 0

vi.3.2 153.8 153.8 153.8 0

vi.3.3 83.5 83.5 83.5 0

vi.3.4 0 116.4 116.4 0

vi.4.1 128.9 128.9 128.9 0

vi.4.2 150.6 150.6 150.6 0

vi.4.3 134.1 134.1 134.1 0

vi.4.4 145.5 145.5 145.5 0

Table B.19: Voiced bar duration,VOT, noise duration and aspiration for /vi/

(ms)

176


zi.1.1 133.8 136 133.8 2.2

zi.1.2 205.1 205.1 205.1 0

zi.1.3 159.6 159.6 159.6 0

zi.1.4 180.6 180.6 180.6 0

zi.2.1 154.8 154.8 154.8 0

zi.2.2 186.5 186.5 186.5 0

zi.2.3 168 168 168 0

zi.2.4 191.8 191.8 191.8 0

zi.3.1 162.9 162.9 162.9 0

zi.3.2 170.2 170.2 170.2 0

zi.3.3 128.1 128.1 128.1 0

zi.3.4 206.3 206.3 206.3 0

zi.4.1 159.8 159.8 159.8 0

zi.4.2 171.4 171.4 171.4 0

zi.4.3 176.9 176.9 176.9 0

zi.4.4 173.1 173.1 173.1 0

Table B.20: Voiced bar duration,VOT, noise duration and aspiration for /zi/

(ms)

177


fu.1.1 0 132.3 120.4 11.9

fu.1.2 0 191.8 180.4 11.4

fu.1.3 0 200.2 190 10.2

fu.1.4 0 194.2 173.8 20.4

fu.2.1 0 134.8 115.5 19.3

fu.2.2 0 206.6 198.8 7.8

fu.2.3 0 211.7 211.7 0

fu.2.4 0 225.4 198.6 26.8

fu.3.1 0 194.5 186.8 7.7

fu.3.2 0 184.2 178.8 5.4

fu.3.3 0 224.3 208.3 16

fu.3.4 0 162.8 148.6 14.2

fu.4.1 0 182.4 166.4 16

fu.4.2 0 163.5 152.1 11.4

fu.4.3 0 195.4 185.2 10.2

fu.4.4 0 228.2 212.1 16.1

Table B.21: Voiced bar duration,VOT, noise duration and aspiration for /fu/

(ms)

178


su.1.1 0 252.8 252.8 0

su.1.2 0 227.5 227.5 0

su.1.3 0 243.4 237.6 5.8

su.1.4 0 236.5 233.3 3.2

su.2.1 0 245.7 245.7 0

su.2.2 0 178 178 0

su.2.3 0 202.8 191.1 11.7

su.2.4 0 222 212.7 9.3

su.3.1 0 232 232 0

su.3.2 0 229.4 229.4 0

su.3.3 0 192.9 190.9 2

su.3.4 0 201 187.7 13.3

su.4.1 0 229.8 229.8 0

su.4.2 0 244 237.9 6.1

su.4.3 0 210.4 210.4 0

su.4.4 0 200.1 186.9 13.2

Table B.22: Voiced bar duration,VOT, noise duration and aspiration for /su/

(ms)

179


vu.1.1 132 132 132 0

vu.1.2 149.1 149.4 149.4 0

vu.1.3 131.3 131.3 131.3 0

vu.1.4 196.8 196.8 196.8 0

vu.2.1 159.1 159.1 159.1 0

vu.2.2 155.4 155.4 155.4 0

vu.2.3 103.6 103.6 103.6 0

vu.2.4 123.1 123.1 123.1 0

vu.3.1 94.9 94.9 94.9 0

vu.3.2 142.4 142.4 142.4 0

vu.3.3 95.6 95.6 95.6 0

vu.3.4 163.5 163.5 163.5 0

vu.4.1 158.2 158.2 158.2 0

vu.4.2 172.4 172.4 172.4 0

vu.4.3 129 129 129 0

vu.4.4 174.8 174.8 174.8 0

Table B.23: Voiced bar duration,VOT, noise duration and aspiration for /vu/

(ms)

180


zu.1.1 148.4 148.4 148.4 0

zu.1.2 159.3 159.3 159.3 0

zu.1.3 150.6 150.6 150.6 0

zu.1.4 211.8 211.8 211.8 0

zu.2.1 167.4 167.4 167.4 0

zu.2.2 199.6 199.6 199.6 0

zu.2.3 187.8 187.8 187.8 0

zu.2.4 166.6 166.6 166.6 0

zu.3.1 153.4 153.4 153.4 0

zu.3.2 242.4 242.4 242.4 0

zu.3.3 186.6 186.6 186.6 0

zu.3.4 147.9 147.9 147.9 0

zu.4.1 154.6 154.6 154.6 0

zu.4.2 166.5 166.5 166.5 0

zu.4.3 153.4 153.4 153.4 0

zu.4.4 184.3 184.3 184.3 0

Table B.24: Voiced bar duration,VOT, noise duration and aspiration for /zu/

(ms)

181

APPENDIX C

Spectral Measurements

Spectral measurements for stops and friatives are summarized in the following

appendices. These measurements include Ahi-A23, Av-Ahi, Av-maxA23, Amid-

Avmid, Av4-A45 and Av4-maxA45 for stops and Ahi-A23, Av-Ahi, Av-Anoise,

Av4-A45 and Av4-maxA45 for fricatives. Measurements are in dB.

182

Ahi- Av- Av- Amid- Av4- Av4-

A23 Ahi maxA23 Avmid A45 maxA45

ba.1.1 -6.90 90.79 73.68 -8.94 22.11 12.38

ba.1.2 -2.96 74.92 65.78 12.91 34.86 26.51

ba.1.3 -6.43 92.79 57.45 -0.23 17.53 4.75

ba.1.4 2.06 96.75 81.55 3.17 27.95 19.63

ba.2.1 -7.52 91.31 74.05 -6.06 31.95 28.78

ba.2.2 1.59 92.58 89.58 -10.80 30.29 24.52

ba.2.3 8.63 79.14 78.59 0.66 29.64 21.72

ba.2.4 7.24 92.69 84.34 -1.27 24.49 18.94

ba.3.1 2.85 101.45 98.25 -46.13 52.07 49.80

ba.3.2 3.91 94.37 94.97 -10.35 36.07 34.20

ba.3.3 0.03 107.27 105.52 -37.46 47.62 45.64

ba.3.4 -6.60 78.47 58.90 -14.61 29.15 16.34

ba.4.1 2.16 76.18 66.65 -21.98 27.00 20.86

ba.4.2 -11.31 85.48 65.27 -19.79 34.96 28.88

ba.4.3 0.62 89.64 87.59 -24.36 32.82 28.64

ba.4.4 3.02 97.63 96.76 -33.33 46.00 43.23

average -0.60 90.09 79.93 -13.66 32.78 26.55

min -11.31 74.92 57.45 -46.13 17.53 4.75

max 8.63 107.27 105.52 12.91 52.07 49.80

Table C.1: Spectral measures for /ba/.

183



da.1.1 11.51 74.06 78.68 12.95 13.16 3.40

da.1.2 -7.96 70.60 52.79 21.50 17.69 2.05

da.1.3 7.84 69.49 70.14 23.26 13.84 -3.92

da.1.4 -1.07 81.04 72.42 7.41 26.53 17.71

da.2.1 0.96 61.08 52.81 23.98 7.84 1.26

da.2.2 1.72 62.61 52.50 32.41 4.34 -2.17

da.2.3 -9.51 74.77 58.28 24.24 1.82 -7.46

da.2.4 7.89 59.04 57.80 22.95 -1.35 -10.99

da.3.1 -4.26 67.01 59.20 -8.37 30.76 24.84

da.3.2 -2.14 58.31 50.56 -9.50 20.09 6.01

da.3.3 2.52 64.28 61.97 21.02 2.55 -4.70

da.3.4 2.55 70.14 63.19 -0.06 21.43 15.86

da.4.1 -4.82 61.20 32.27 -4.46 36.35 18.16

da.4.2 1.97 56.45 42.43 -0.27 17.30 11.01

da.4.3 1.76 63.09 52.27 8.46 4.74 -7.59

da.4.4 3.83 83.94 82.80 -12.19 33.61 28.43

average 0.80 67.32 58.76 10.21 15.67 5.74

min -9.51 56.45 32.27 -12.19 -1.35 -10.99

max 11.51 83.94 82.80 32.41 36.35 28.43

Table C.2: Spectral measures for /da/.

184



pa.1.1 0.98 79.93 69.014 -24.74 25.53 8.86

pa.1.2 -10.36 85.00 58.373 -2.41 17.17 11.23

pa.1.3 -5.56 89.71 72.501 -11.56 28.87 23.62

pa.1.4 -2.94 68.60 55.836 8.43 3.66 -3.31

pa.2.1 -9.13 71.81 54.215 13.58 2.52 -7.58

pa.2.2 -10.57 79.75 58.097 15.62 6.03 -2.57

pa.2.3 -6.25 78.39 61.073 -1.90 14.34 5.34

pa.2.4 -5.55 82.56 68.287 13.11 6.98 -5.06

pa.3.1 -5.20 53.61 41.556 -5.34 32.60 23.71

pa.3.2 0.72 60.74 48.985 -11.38 21.46 12.58

pa.3.3 -5.55 68.15 44.021 -11.67 30.13 21.36

pa.3.4 -8.17 71.24 43.616 -13.63 25.44 17.43

pa.4.1 -10.23 68.75 53.024 -6.00 17.89 10.20

pa.4.2 -7.60 61.25 44.496 3.02 7.18 -4.67

pa.4.3 -6.14 74.92 53.014 -15.14 30.59 23.12

pa.4.4 -2.76 60.44 46.118 6.42 16.32 4.20

average -5.90 72.18 54.514 -2.72 17.92 8.65

min -10.57 53.61 41.556 -24.74 2.52 -7.58

max 0.98 89.71 72.50 15.62 32.60 23.71

Table C.3: Spectral measures for /pa/.

185



ta.1.1 -2.20 57.66 41.09 36.62 -25.39 -33.87

ta.1.2 5.85 56.97 57.35 31.91 -12.24 -18.40

ta.1.3 0.11 68.50 59.50 13.54 4.02 -8.82

ta.1.4 -0.93 74.00 63.04 19.88 5.07 -1.94

ta.2.1 8.22 33.47 34.31 42.09 -25.54 -45.14

ta.2.2 30.24 33.13 58.02 29.44 -7.52 -18.14

ta.2.3 6.48 58.68 56.82 37.86 -17.47 -26.04

ta.2.4 7.82 64.18 66.42 30.31 -3.31 -10.88

ta.3.1 14.63 38.10 45.19 16.08 4.10 -4.67

ta.3.2 8.19 55.54 56.03 -12.92 20.76 1.35

ta.3.3 9.09 54.70 54.08 -11.90 20.12 2.68

ta.3.4 16.83 43.17 54.78 -12.52 35.17 4.25

ta.4.1 1.60 51.42 45.55 0.01 18.09 12.52

ta.4.2 -9.72 63.08 40.98 -1.43 24.96 17.10

ta.4.3 -0.17 48.87 23.16 1.06 15.99 -2.68

ta.4.4 -1.68 41.53 25.42 16.86 -3.19 -11.97

average 5.90 52.69 48.86 14.81 3.35 -9.04

min -9.72 33.13 23.16 -12.92 -25.54 -45.14

max 30.24 74.00 66.42 42.09 35.17 17.10

Table C.4: Spectral measures for /ta/.

186



bi.1.1 -0.84 83.80 72.94 -21.78 10.10 2.79

bi.1.2 3.54 89.66 82.56 -13.91 19.61 8.13

bi.1.3 -1.34 78.28 66.98 3.41 2.50 -5.09

bi.1.4 -8.29 86.49 60.75 -14.89 16.66 13.15

bi.2.1 -1.85 82.30 68.93 -18.19 14.03 6.02

bi.2.2 -3.47 85.46 63.16 -13.55 17.55 5.22

bi.2.3 -7.38 83.69 67.47 -15.25 14.47 4.31

bi.2.4 1.12 82.54 75.27 -0.45 0.09 -15.91

bi.3.1 5.09 99.58 100.48 -56.68 19.57 22.16

bi.3.2 -1.84 94.70 87.87 -53.27 20.57 30.55

bi.3.3 4.29 106.63 101.25 -62.09 15.47 28.49

bi.3.4 0.29 96.97 95.79 -56.24 23.88 24.66

bi.4.1 -8.04 78.25 60.15 -39.79 54.16 49.26

bi.4.2 -1.11 74.02 62.49 -15.81 41.42 27.50

bi.4.3 -0.20 95.40 94.16 -32.81 73.36 21.09

bi.4.4 2.20 85.05 85.45 -34.24 50.01 34.85

average -1.11 87.67 77.85 -27.85 24.59 16.07

min -8.29 74.02 60.15 -62.09 0.09 -15.91

max 5.09 106.63 101.25 3.41 73.36 49.26

Table C.5: Spectral measures for /bi/.

187



di.1.1 2.77 58.79 44.45 18.45 -13.06 -21.75

di.1.2 5.50 52.24 47.32 34.12 -24.55 -43.62

di.1.3 11.65 57.42 60.20 28.86 -29.55 -36.78

di.1.4 7.19 46.43 42.51 20.95 -12.51 -22.72

di.2.1 1.89 56.60 42.14 3.04 -8.40 -24.41

di.2.2 3.65 71.94 70.11 4.43 -0.88 -9.18

di.2.3 -2.57 75.49 64.29 -3.22 -0.05 -9.64

di.2.4 5.85 72.91 68.58 -7.93 4.47 -2.72

di.3.1 -3.73 77.93 67.74 -8.48 27.33 18.96

di.3.2 19.15 48.87 56.04 -3.39 8.20 4.14

di.3.3 13.51 69.96 75.98 -16.25 21.16 9.04

di.3.4 18.82 65.39 76.42 -16.33 21.57 8.10

di.4.1 3.55 58.43 50.12 -13.05 23.45 9.76

di.4.2 -4.92 51.94 34.40 -5.68 14.88 6.30

di.4.3 1.95 49.51 35.28 -8.50 17.75 7.75

di.4.4 -1.51 57.50 48.32 -2.05 14.71 -2.40

average 5.17 60.71 55.24 1.56 4.03 -6.82

min -4.92 46.43 34.40 -16.33 -29.55 -43.62

max 19.15 77.93 76.42 34.12 27.33 18.96

Table C.6: Spectral measures for /di/.

188



pi.1.1 1.22 79.21 72.02 -8.28 23.92 16.86

pi.1.2 -2.66 71.14 52.18 12.84 -2.84 -9.97

pi.1.3 -10.70 76.72 55.38 -2.09 11.30 1.01

pi.1.4 -4.81 72.32 51.00 15.91 -14.84 -24.81

pi.2.1 -6.81 63.08 44.47 4.28 6.20 3.05

pi.2.2 -2.85 66.94 50.54 -3.50 9.13 4.14

pi.2.3 -2.03 70.97 49.97 -7.99 20.80 14.60

pi.2.4 -4.07 69.15 47.71 2.92 11.28 8.52

pi.3.1 -7.49 61.13 39.38 -4.75 22.67 10.16

pi.3.2 -4.69 54.62 34.75 -14.37 32.02 19.28

pi.3.3 -10.92 55.47 29.15 -12.81 23.49 9.08

pi.3.4 -7.64 52.55 28.06 0.55 14.77 -1.15

pi.4.1 4.05 45.53 35.70 9.16 7.60 -3.16

pi.4.2 9.60 46.94 38.15 1.06 7.72 -2.94

pi.4.3 2.56 49.60 30.35 8.87 4.30 -6.28

pi.4.4 2.26 53.40 37.15 4.84 6.44 0.38

average -2.81 61.80 43.50 0.41 11.50 2.42

min -10.92 45.53 28.06 -14.37 -14.84 -24.81

max 9.60 79.21 72.02 15.91 32.02 19.28

Table C.7: Spectral measures for /pi/.

189



ti.1.1 5.10 57.24 50.10 18.71 -10.88 -18.02

ti.1.2 20.16 52.08 59.01 38.28 -38.11 -45.30

ti.1.3 9.17 61.18 53.53 27.09 -26.95 -39.29

ti.1.4 12.45 45.71 49.68 20.73 -25.20 -33.53

ti.2.1 17.85 39.56 54.27 21.95 -31.24 -37.05

ti.2.2 8.71 44.06 32.69 37.45 -30.93 -35.92

ti.2.3 20.04 54.08 52.95 23.88 -14.78 -21.12

ti.2.4 12.91 44.30 42.43 21.59 -24.94 -30.58

ti.3.1 16.30 43.28 48.76 -9.04 19.84 9.63

ti.3.2 20.46 33.58 40.58 -3.65 5.60 -10.89

ti.3.3 12.02 46.77 50.80 16.79 -19.66 -28.83

ti.3.4 19.46 37.26 44.31 -10.06 16.68 1.55

ti.4.1 9.64 44.00 44.92 0.56 5.49 -6.02

ti.4.2 10.55 41.77 42.37 8.99 2.99 -15.03

ti.4.3 11.08 44.43 47.00 4.60 3.02 -1.68

ti.4.4 23.31 43.93 51.63 -5.24 6.59 -8.55

average 14.33 45.83 47.81 13.29 -10.16 -20.04

min 5.10 33.58 32.69 -10.06 -38.11 -45.30

max 23.31 61.18 59.01 38.28 19.84 9.63

Table C.8: Spectral measures for /ti/.

190



bu.1.1 -9.07 85.76 67.33 -1.98 24.67 19.58

bu.1.2 -12.49 86.11 59.33 9.67 32.87 22.29

bu.1.3 -0.45 83.14 72.19 3.44 28.67 25.51

bu.1.4 0.37 84.31 71.99 3.07 18.42 12.71

bu.2.1 -2.02 67.54 53.18 8.00 7.50 -0.66

bu.2.2 2.66 86.99 77.02 -17.15 58.32 50.70

bu.2.3 -4.99 79.42 67.88 3.62 40.37 34.43

bu.2.4 0.35 82.35 72.52 -11.23 47.07 41.57

bu.3.1 0.38 94.27 90.31 -32.55 47.32 44.81

bu.3.2 0.03 103.66 98.86 -24.37 42.59 40.05

bu.3.3 -1.46 93.47 85.83 -27.28 62.64 59.40

bu.3.4 -4.91 89.54 80.12 -19.80 46.09 41.73

bu.4.1 -15.79 91.28 67.06 -6.11 4.92 -2.42

bu.4.2 1.94 100.05 98.86 -19.59 17.33 15.13

bu.4.3 -11.99 87.10 70.26 -8.90 20.94 14.84

bu.4.4 -2.84 93.26 80.00 -14.64 17.12 13.60

average -3.77 88.02 75.80 -9.74 32.30 27.08

min -15.79 67.54 53.18 -32.55 4.92 -2.42

max 2.66 103.66 98.86 9.67 62.64 59.40

Table C.9: Spectral measures for /bu/.

191



du.1.1 5.38 53.70 46.77 24.73 27.28 15.51

du.1.2 3.33 66.28 61.27 27.66 37.94 29.07

du.1.3 4.81 46.63 32.51 31.21 19.99 12.29

du.1.4 -5.35 57.15 42.33 18.79 18.07 5.64

du.2.1 -2.09 54.64 44.95 18.79 26.43 13.45

du.2.2 1.46 67.91 55.89 18.16 10.32 -0.94

du.2.3 4.46 63.10 55.85 18.71 10.02 1.50

du.2.4 -14.32 71.30 47.16 3.56 42.64 28.33

du.3.1 23.03 65.73 76.83 2.71 37.98 18.81

du.3.2 15.46 61.86 70.16 -4.42 37.80 27.10

du.3.3 8.70 67.97 60.24 -3.04 40.15 35.08

du.3.4 5.69 78.38 68.06 -2.27 38.41 29.18

du.4.1 9.46 72.05 77.72 3.72 24.36 10.26

du.4.2 -7.58 85.18 73.34 -3.85 30.63 23.82

du.4.3 1.10 77.08 69.25 8.18 20.58 7.02

du.4.4 -3.00 91.48 85.71 -28.25 32.16 37.80

average 3.16 67.53 60.50 8.40 28.42 18.37

min -14.32 46.63 32.51 -28.25 10.02 -0.94

max 23.03 91.48 85.71 31.21 42.64 37.80

Table C.10: Spectral measures for /du/.

192



pu.1.1 0.76 75.76 69.80 16.84 9.91 2.86

pu.1.2 -16.60 84.40 47.16 18.10 14.68 11.20

pu.1.3 -12.93 83.10 51.26 6.58 11.58 8.75

pu.1.4 -6.76 76.71 66.72 17.74 14.14 9.91

pu.2.1 -7.55 70.39 48.12 10.82 10.57 3.76

pu.2.2 -7.08 72.59 49.18 -2.12 26.12 22.17

pu.2.3 -2.82 80.54 59.39 4.42 15.66 7.91

pu.2.4 0.71 76.14 67.32 -4.74 29.51 21.78

pu.3.1 -9.97 79.11 60.66 6.01 10.76 8.55

pu.3.2 -7.69 82.95 60.87 3.79 21.40 15.92

pu.3.3 -11.63 79.67 62.78 2.12 17.34 9.73

pu.3.4 -3.48 82.27 69.66 -10.46 18.26 14.94

pu.4.1 -10.04 80.07 55.72 -0.22 14.56 8.95

pu.4.2 -12.01 79.14 52.26 4.44 5.22 -0.21

pu.4.3 -7.51 82.84 63.97 0.02 19.88 18.92

pu.4.4 -12.31 78.53 49.46 -1.23 9.44 4.95

average -7.93 79.01 58.39 4.51 15.56 10.63

min -16.60 70.39 47.16 -10.46 5.22 -0.21

max 0.76 84.40 69.80 18.10 29.51 22.17

Table C.11: Spectral measures for /pu/.

193



tu.1.1 -1.98 54.48 31.24 37.31 -15.28 -28.42

tu.1.2 0.55 53.66 36.90 31.70 5.17 -2.71

tu.1.3 12.83 56.31 55.48 22.17 19.19 11.76

tu.1.4 12.34 57.77 55.34 36.73 7.71 -1.83

tu.2.1 20.46 40.70 48.70 35.27 -2.66 -14.01

tu.2.2 15.21 43.85 47.13 25.10 5.68 -2.97

tu.2.3 13.15 46.67 49.50 35.22 -10.41 -22.61

tu.2.4 28.92 38.25 40.94 32.33 -15.26 -20.29

tu.3.1 3.68 55.63 47.37 25.98 1.41 -9.88

tu.3.2 16.63 43.31 45.80 8.64 15.22 3.30

tu.3.3 25.20 45.66 61.79 2.56 15.81 -8.26

tu.3.4 23.67 39.71 54.30 11.30 9.33 1.41

tu.4.1 8.50 48.99 50.10 27.80 -21.85 -35.24

tu.4.2 6.37 54.36 53.58 25.94 -10.10 -21.00

tu.4.3 3.81 53.68 46.70 29.67 -17.50 -29.09

tu.4.4 12.11 59.20 58.72 13.89 -6.72 -17.55

average 12.59 49.52 48.97 25.10 -1.26 -12.34

min -1.98 38.25 31.24 2.56 -21.85 -35.24

max 28.92 59.20 61.79 37.31 19.19 11.76

Table C.12: Spectral measures for /tu/.

194

Ahi-A23 Av-Ahi Av-Anoise Av4-A45 Av4-maxA45

fa.1.1 30.8196 68.6742 90.9686 19.3946 11.4087

fa.1.2 28.3517 72.1773 93.6805 12.7369 5.3131

fa.1.3 26.9374 64.9029 85.8914 8.4323 -1.2649

fa.1.4 23.6129 67.758 86.1659 17.5966 7.4782

fa.2.1 24.1788 63.6454 83.8454 39.3546 34.0368

fa.2.2 23.0601 67.8684 86.2271 30.4783 23.6479

fa.2.3 26.3911 63.5642 85.3252 14.4879 10.347

fa.2.4 26.3606 65.0188 86.9555 15.1882 11.2266

fa.3.1 23.7832 58.7613 80.0085 29.0404 23.8333

fa.3.2 28.7519 62.8867 86.5399 18.4086 12.8639

fa.3.3 24.4785 61.7037 82.3066 28.6689 24.3793

fa.3.4 22.7656 68.0784 85.2245 32.2195 21.4151

fa.4.1 22.3734 51.0917 72.1889 20.6874 13.8563

fa.4.2 34.5589 58.618 82.169 20.4792 14.0996

fa.4.3 25.5963 52.6561 72.1113 21.4057 13.2186

fa.4.4 29.3869 62.8452 83.4519 35.9425 28.2526

average 26.33793125 63.14064375 83.9412625 22.7826 15.88200625

min 22.3734 51.0917 72.1113 8.4323 -1.2649

max 34.5589 72.1773 93.6805 39.3546 34.0368

Table C.13: Spectral measures for /fa/.

195


sa.1.1 28.23 58.48 64.53 6.58 -23.52

sa.1.2 24.93 67.82 69.19 6.99 -18.16

sa.1.3 25.45 61.46 63.51 -4.87 -36.11

sa.1.4 30.18 63.31 71.10 11.28 -22.37

sa.2.1 25.48 61.99 67.39 9.56 -15.52

sa.2.2 25.97 52.86 58.31 0.44 -24.30

sa.2.3 27.39 57.11 65.37 -6.46 -29.31

sa.2.4 24.39 59.28 65.01 6.47 -17.69

sa.3.1 24.65 56.00 67.67 27.85 21.81

sa.3.2 27.08 59.28 73.74 27.45 22.51

sa.3.3 27.11 54.81 65.25 25.45 17.95

sa.3.4 24.40 63.73 70.30 30.69 21.10

sa.4.1 30.57 54.77 64.09 31.76 15.31

sa.4.2 30.57 54.77 64.09 31.76 15.31

sa.4.3 30.28 46.17 52.19 23.98 5.97

sa.4.4 31.71 51.46 58.69 8.85 -6.31

average 27.40 57.71 65.03 14.86 -4.58

min 24.39 46.17 52.19 -6.46 -36.11

max 31.71 67.82 73.74 31.76 22.51

Table C.14: Spectral measures for /sa/.

196


va.1.1 18.99 65.54 85.93 29.00 23.01

va.1.2 20.89 75.89 94.15 30.37 20.61

va.1.3 25.50 76.56 102.03 52.39 44.32

va.1.4 18.64 74.21 94.17 42.49 32.87

va.2.1 32.48 63.60 91.09 50.37 41.58

va.2.2 26.34 68.47 90.40 62.27 57.72

va.2.3 29.39 60.99 88.55 41.51 33.15

va.2.4 24.96 67.82 87.65 49.53 38.47

va.3.1 20.72 64.34 81.27 46.49 34.73

va.3.2 21.62 64.97 84.70 38.94 33.02

va.3.3 22.75 74.28 92.65 33.31 19.23

va.3.4 33.63 68.59 96.80 49.86 35.77

va.4.1 34.51 75.47 103.21 45.52 35.55

va.4.2 26.42 67.51 90.23 42.36 33.11

va.4.3 38.35 54.83 82.15 27.58 17.91

va.4.4 27.56 57.78 78.75 35.89 28.70

average 26.42 67.55 90.23 42.37 33.11

min 18.64 54.83 78.75 27.58 17.91

max 38.35 76.56 103.21 62.27 57.72

Table C.15: Spectral measures for /va/.

197


za.1.1 24.03 69.0 73.39 11.36 -18.59

za.1.2 24.78 72.4 79.51 12.67 -18.29

za.1.3 24.73 70.0 75.67 14.28 -7.01

za.1.4 26.90 59.8 63.53 0.09 -24.87

za.2.1 23.69 66.8 73.42 26.57 7.57

za.2.2 29.93 65.0 73.25 22.93 4.41

za.2.3 19.40 69.4 71.16 42.94 25.56

za.2.4 26.61 61.7 68.06 34.58 14.57

za.3.1 29.11 68.4 80.87 46.54 39.35

za.3.2 30.19 56.6 70.58 48.45 38.33

za.3.3 26.54 66.3 82.69 33.26 20.22

za.3.4 27.50 62.1 76.28 46.38 38.96

za.4.1 31.78 57.5 67.75 22.35 7.80

za.4.2 32.10 55.0 66.18 38.75 26.85

za.4.3 42.72 50.1 77.58 35.21 26.47

za.4.4 28.35 62.3 71.22 37.76 21.43

average 28.02 63.3 73.20 29.63 12.67

min 19.40 50.1 63.53 0.09 -24.87

max 42.72 72.4 82.69 48.45 39.35

Table C.16: Spectral measures for /za/.

198


fi.1.1 27.18 65.92 88.55 10.23 3.39

fi.1.2 21.33 62.03 83.61 11.63 2.18

fi.1.3 25.55 40.26 60.78 9.44 -2.25

fi.1.4 24.19 64.80 86.36 3.99 -3.05

fi.2.1 21.20 59.08 78.35 23.28 17.46

fi.2.2 22.97 61.52 83.62 2.83 -5.71

fi.2.3 20.73 61.11 80.70 1.59 -5.70

fi.2.4 20.06 54.71 73.03 7.28 0.16

fi.3.1 22.02 45.42 68.81 30.39 23.04

fi.3.2 22.36 60.59 82.22 68.42 59.18

fi.3.3 28.93 51.09 76.78 39.58 32.99

fi.3.4 25.46 64.65 86.97 48.95 43.20

fi.4.1 25.30 59.98 80.26 38.89 32.58

fi.4.2 20.47 50.26 70.28 34.38 26.65

fi.4.3 19.85 52.35 68.63 41.35 32.68

fi.4.4 18.74 48.68 66.80 43.35 30.07

average 22.90 56.40 77.24 25.97 17.93

min 18.74 40.26 60.78 1.59 -5.71

max 28.93 65.92 88.55 68.42 59.18

Table C.17: Spectral measures for /fi/.

199


si.1.1 21.45 59.32 65.25 -29.60 -46.18

si.1.2 17.68 57.96 58.90 -36.59 -51.76

si.1.3 21.04 62.88 64.94 -35.41 -56.80

si.1.4 24.21 69.00 72.75 -36.26 -53.31

si.2.1 24.49 59.80 68.71 -22.88 -39.21

si.2.2 24.99 54.15 63.10 -22.93 -33.91

si.2.3 21.76 54.10 61.81 -21.89 -31.61

si.2.4 16.84 58.99 58.77 -7.49 -22.81

si.3.1 25.20 64.59 80.52 13.68 6.00

si.3.2 23.99 48.96 63.16 -9.59 -25.00

si.3.3 22.16 46.86 60.07 1.29 -13.12

si.3.4 19.41 69.12 79.43 23.36 15.96

si.4.1 23.39 50.37 58.33 7.65 -12.45

si.4.2 26.90 48.19 58.54 8.10 -13.77

si.4.3 30.53 47.17 58.13 9.39 -25.31

si.4.4 22.59 49.03 56.86 11.32 -8.54

average 22.91 56.28 64.33 -9.24 -25.74

min 16.84 46.86 56.86 -36.59 -56.80

max 30.53 69.12 80.52 23.36 15.96

Table C.18: Spectral measures for /si/.

200


vi.1.1 22.88 52.35 76.00 11.10 3.69

vi.1.2 19.34 63.72 81.80 8.51 1.01

vi.1.3 22.05 61.49 86.82 35.62 26.92

vi.1.4 18.84 75.23 93.29 35.89 27.55

vi.2.1 25.11 55.00 78.07 10.82 -1.24

vi.2.2 10.58 66.48 77.99 7.66 0.19

vi.2.3 26.08 54.87 77.98 23.53 16.01

vi.2.4 21.47 58.64 76.52 8.83 3.11

vi.3.1 20.27 55.51 74.67 57.73 51.55

vi.3.2 23.66 65.71 87.73 37.51 28.88

vi.3.3 13.69 67.76 83.98 66.12 44.65

vi.3.4 19.10 63.77 82.97 64.05 56.87

vi.4.1 27.29 54.92 76.05 43.66 35.04

vi.4.2 27.67 59.97 84.47 53.85 47.53

vi.4.3 27.55 55.78 79.16 29.26 15.61

vi.4.4 29.54 57.02 80.65 42.64 33.11

average 22.19 60.51 81.13 33.55 24.41

min 10.58 52.35 74.67 7.66 -1.24

max 29.54 75.23 93.29 66.12 56.87

Table C.19: Spectral measures for /vi/.

201


zi.1.1 27.83 52.04 63.59 -2.08 -19.84

zi.1.2 24.73 61.77 68.84 -16.13 -35.13

zi.1.3 16.82 67.08 68.02 -21.33 -36.16

zi.1.4 19.14 60.02 59.44 -20.67 -43.83

zi.2.1 23.82 57.65 63.55 -12.09 -22.78

zi.2.2 20.58 65.06 67.33 -6.96 -16.98

zi.2.3 21.44 66.55 67.95 -15.41 -30.26

zi.2.4 25.14 59.04 65.51 -10.64 -22.11

zi.3.1 27.43 53.21 73.76 14.81 6.52

zi.3.2 17.62 60.40 72.45 17.86 5.47

zi.3.3 23.63 59.62 71.29 27.03 7.60

zi.3.4 24.28 55.01 71.59 22.17 15.78

zi.4.1 29.56 55.60 64.74 37.52 16.06

zi.4.2 30.16 48.53 62.94 16.78 1.53

zi.4.3 28.60 48.60 56.13 16.62 -10.34

zi.4.4 29.66 52.34 58.87 -3.35 -21.92

average 24.40 57.66 66.00 2.76 -12.90

min 16.82 48.53 56.13 -21.33 -43.83

max 30.16 67.08 73.76 37.52 16.06

Table C.20: Spectral measures for /zi/.

202


fu.1.1 30.91 59.03 82.93 19.08 10.41

fu.1.2 24.32 61.99 80.05 23.47 15.58

fu.1.3 28.62 66.44 89.23 26.52 21.56

fu.1.4 25.80 61.55 81.39 24.08 15.26

fu.2.1 28.10 57.71 77.86 28.19 19.47

fu.2.2 30.58 70.91 92.24 52.62 47.19

fu.2.3 22.69 54.74 74.92 6.15 -0.65

fu.2.4 24.59 62.74 82.09 42.37 33.44

fu.3.1 26.09 65.34 83.56 31.26 26.85

fu.3.2 21.41 55.37 71.03 27.80 22.25

fu.3.3 24.22 67.83 87.53 23.21 18.61

fu.3.4 25.09 63.38 86.30 25.13 17.10

fu.4.1 20.60 66.51 83.89 15.84 8.82

fu.4.2 24.71 62.67 81.86 13.08 6.44

fu.4.3 15.39 58.81 75.68 11.84 4.43

fu.4.4 28.84 67.35 89.58 33.79 28.45

average 25.12 62.65 82.51 25.28 18.45

min 15.39 54.74 71.03 6.15 -0.65

max 30.91 70.91 92.24 52.62 47.19

Table C.21: Spectral measures for /fu/.

203


su.1.1 27.12 62.74 65.26 -1.62 -36.23

su.1.2 24.04 62.43 64.04 6.63 -20.07

su.1.3 23.29 63.43 64.46 -7.86 -32.45

su.1.4 27.05 57.40 59.27 -4.03 -28.48

su.2.1 28.91 47.25 57.44 -11.52 -24.81

su.2.2 27.91 51.92 60.37 -6.27 -24.26

su.2.3 31.81 52.38 63.78 9.62 -11.74

su.2.4 27.95 53.59 57.86 -12.12 -44.13

su.3.1 28.23 64.11 69.48 35.42 11.44

su.3.2 31.16 57.41 67.34 23.89 10.44

su.3.3 23.11 67.06 69.20 31.50 14.41

su.3.4 31.53 55.56 69.06 35.75 18.64

su.4.1 26.65 59.82 64.57 1.10 -20.92

su.4.2 25.65 58.83 65.11 3.29 -14.12

su.4.3 32.29 56.00 64.10 3.48 -10.59

su.4.4 30.58 55.66 65.23 -2.77 -20.44

average 27.95 57.85 64.16 6.53 -14.58

min 23.11 47.25 57.44 -12.12 -44.13

max 32.29 67.06 69.48 35.75 18.64

Table C.22: Spectral measures for /su/.

204


vu.1.1 21.04 63.59 80.99 32.00 25.01

vu.1.2 26.57 70.24 90.51 36.29 31.41

vu.1.3 23.44 63.13 83.69 26.81 18.93

vu.1.4 22.79 69.83 87.91 33.78 27.76

vu.2.1 33.31 67.86 93.87 62.92 54.50

vu.2.2 18.72 59.73 76.06 32.26 21.83

vu.2.3 20.71 55.92 72.87 41.67 32.75

vu.2.4 23.28 67.76 85.56 44.29 36.71

vu.3.1 18.16 61.57 77.71 45.21 37.15

vu.3.2 14.86 56.85 72.87 35.84 21.93

vu.3.3 25.21 63.21 84.20 35.55 24.44

vu.3.4 24.80 60.95 82.17 45.17 36.90

vu.4.1 29.57 59.76 84.45 33.09 25.11

vu.4.2 30.93 71.68 96.48 20.31 13.54

vu.4.3 30.97 54.96 79.37 26.12 16.46

vu.4.4 31.73 64.76 88.56 16.19 10.55

average 24.76 63.24 83.58 35.47 27.18

min 14.86 54.96 72.87 16.19 10.55

max 33.31 71.68 96.48 62.92 54.50

Table C.23: Spectral measures for /vu/.

205


zu.1.1 29.92 58.21 67.10 4.93 -21.51

zu.1.2 28.89 66.75 75.31 31.30 0.59

zu.1.3 24.48 65.18 67.78 23.26 -5.78

zu.1.4 19.04 64.13 65.98 17.47 -6.44

zu.2.1 30.98 59.73 67.63 27.04 11.18

zu.2.2 23.54 63.46 64.86 8.62 -6.98

zu.2.3 30.92 60.39 69.04 29.26 16.99

zu.2.4 26.12 62.20 69.79 34.92 15.76

zu.3.1 21.56 53.44 58.05 35.63 8.45

zu.3.2 24.33 52.53 59.46 30.98 11.23

zu.3.3 28.21 53.70 61.38 36.80 15.42

zu.3.4 24.38 52.89 53.78 43.03 7.47

zu.4.1 32.52 59.64 67.05 26.56 5.84

zu.4.2 35.28 55.22 66.45 1.80 -12.76

zu.4.3 30.85 57.00 66.27 13.77 2.86

zu.4.4 34.58 55.50 67.34 8.94 -8.81

average 27.85 58.75 65.45 23.39 2.09

min 19.04 52.53 53.78 1.80 -21.51

max 35.28 66.75 75.31 43.03 16.99

Table C.24: Spectral measures for /zu/.

206

APPENDIX D

Confusion Matrices

Shown in this appendix are tables for the confusion matrices from the perceptual

experiments described in Chapter 3. There are 32 tokens played for each syllable.

White noise was used as the noise mask. Each table summarized the experimental

results for each CV pair (differed by their place of articulation) with 7 noise levels.

207

Subject 1 Subject 2 Subject 3 Subject 4

noiseless ba da ba da ba da ba da

ba 32 0 32 0 32 0 32 0

da 0 32 0 32 0 32 0 32

10 SNR

ba 28 4 32 0 32 0 32 0

da 0 32 0 32 0 32 0 32

5 SNR

ba 23 9 32 0 32 0 32 0

da 1 31 1 31 0 32 0 32

0 SNR

ba 21 11 31 1 31 1 27 5

da 11 21 2 30 0 32 2 30

-5 SNR

ba 19 13 28 4 30 2 23 9

da 9 23 10 22 2 30 5 27

-10 SNR

ba 22 10 27 5 22 10 18 14

da 16 16 13 19 12 20 7 25

-15 SNR

ba 12 20 20 12 22 10 19 13

da 18 14 20 12 18 14 15 17

Table D.1: Confusion Matrices for /ba,da/

208


noiseless bi di bi di bi di bi di

bi 32 0 32 0 32 0 32 0

di 0 32 0 32 0 32 1 31

10 SNR

bi 28 4 29 3 32 0 28 4

di 0 32 4 28 5 27 1 31

5 SNR

bi 23 9 27 5 29 3 23 9

di 1 31 6 26 13 19 2 30

0 SNR

bi 21 11 27 5 25 7 20 12

di 11 21 20 12 12 20 11 21

-5 SNR

bi 19 13 24 8 15 17 20 12

di 9 23 21 11 15 17 16 16

-10 SNR

bi 22 10 19 13 21 11 7 25

di 16 16 22 10 18 14 14 18

-15 SNR

bi 12 20 15 17 20 12 6 26

di 18 14 22 10 18 14 17 15

Table D.2: Confusion Matrices for /bi,di/

209


noiseless bu du bu du bu du bu du

bu 32 0 32 0 32 0 32 0

du 0 32 0 32 0 32 0 32

10 SNR

bu 24 8 32 0 32 0 32 0

du 9 23 1 31 0 32 0 32

5 SNR

bu 32 0 32 0 32 0 32 0

du 3 29 0 32 0 32 3 29

0 SNR

bu 31 1 30 2 26 6 31 1

du 7 25 7 25 6 26 6 26

-5 SNR

bu 26 6 26 6 11 21 24 8

du 13 19 14 18 10 22 13 19

-10 SNR

bu 22 10 19 13 23 9 21 11

du 15 17 16 16 11 21 17 15

-15 SNR

bu 21 11 21 11 20 12 25 7

du 14 18 18 14 14 18 25 7

Table D.3: Confusion Matrices for /bu,du/

210


noiseless pa ta pa ta pa ta pa ta

pa 32 0 31 1 32 0 32 0

ta 0 32 1 31 0 32 0 32

10 SNR

pa 31 1 26 6 31 1 31 1

ta 8 24 15 17 5 27 11 21

5 SNR

pa 28 4 28 4 29 3 20 12

ta 12 20 9 23 9 23 4 28

0 SNR

pa 30 2 30 2 20 12 22 10

ta 12 20 19 13 8 24 11 21

-5 SNR

pa 27 5 27 5 21 11 16 16

ta 15 17 15 17 17 15 9 23

-10 SNR

pa 27 5 22 10 15 17 15 17

ta 21 11 17 15 18 14 12 20

-15 SNR

pa 18 14 18 14 13 19 14 18

ta 18 14 17 15 19 13 17 15

Table D.4: Confusion Matrices for /pa,ta/

211


noiseless pi ti pi ti pi ti pi ti

pi 31 1 31 1 32 0 29 3

ti 0 32 1 31 0 32 4 28

10 SNR

pi 29 3 31 1 32 0 32 0

ti 0 32 1 31 0 32 0 32

5 SNR

pi 19 13 23 9 32 0 27 5

ti 0 32 2 30 1 31 0 32

0 SNR

pi 18 14 22 10 23 9 26 6

ti 4 28 4 28 5 27 4 28

-5 SNR

pi 12 20 27 5 20 12 24 8

ti 3 29 14 18 14 18 15 17

-10 SNR

pi 20 12 23 9 16 16 18 14

ti 17 15 24 8 12 20 19 13

-15 SNR

pi 15 17 22 10 19 13 20 12

ti 20 12 20 12 15 17 19 13

Table D.5: Confusion Matrices for /pi,ti/

212


noiseless pu tu pu tu pu tu pu tu

pu 32 0 31 1 32 0 31 1

tu 0 32 1 31 0 32 0 32

10 SNR

pu 32 0 32 0 32 0 32 0

tu 0 32 0 32 0 32 0 32

5 SNR

pu 32 0 31 1 32 0 32 0

tu 1 31 1 31 1 31 2 30

0 SNR

pu 22 10 29 3 23 9 24 8

tu 6 26 4 28 5 27 9 23

-5 SNR

pu 20 12 19 13 20 12 17 15

tu 12 20 8 24 14 18 27 5

-10 SNR

pu 15 17 14 18 16 16 21 11

tu 15 17 13 19 12 20 21 11

-15 SNR

pu 10 22 11 21 19 13 20 12

tu 15 17 17 15 15 17 17 15

Table D.6: Confusion Matrices for /pu,tu/

213


noiseless fa sa fa sa fa sa fa sa

fa 32 0 32 0 32 0 32 0

sa 0 32 0 32 0 32 0 32

10 SNR

fa 32 0 32 0 32 0 32 0

sa 1 31 1 31 0 32 0 32

5 SNR

fa 32 0 31 1 32 0 32 0

sa 0 32 1 31 0 32 0 32

0 SNR

fa 30 2 30 2 32 0 27 5

sa 0 32 0 32 3 29 2 30

-5 SNR

fa 31 1 32 0 26 6 25 7

sa 14 18 11 21 7 25 7 25

-10 SNR

fa 21 11 25 7 23 9 21 11

sa 20 12 19 13 17 15 18 14

-15 SNR

fa 21 11 22 10 17 15 15 17

sa 27 5 20 12 26 6 22 10

Table D.7: Confusion Matrices for /fa,sa/

214


noiseless fi si fi si fi si fi si

fi 31 1 32 0 32 0 32 0

si 0 32 0 32 2 30 0 32

10 SNR

fi 32 0 31 1 32 0 32 0

si 0 32 0 32 0 32 0 32

5 SNR

fi 32 0 31 1 32 0 32 0

si 1 31 1 31 3 29 1 31

0 SNR

fi 26 6 32 0 31 1 30 2

si 0 32 0 32 3 29 5 27

-5 SNR

fi 20 12 25 7 29 3 22 10

si 8 24 12 20 7 25 12 20

-10 SNR

fi 17 15 25 7 22 10 18 14

si 21 11 16 16 20 12 22 10

-15 SNR

fi 18 14 22 10 15 17 19 13

si 16 16 18 14 18 14 22 10

Table D.8: Confusion Matrices for /fi,si/

215


noiseless fu su fu su fu su fu su

fu 32 0 32 0 32 0 32 0

su 0 32 0 32 0 32 0 32

10 SNR

fu 32 0 32 0 32 0 32 0

su 0 32 0 32 0 32 0 32

5 SNR

fu 32 0 28 4 32 0 30 2

su 0 32 5 27 0 32 0 32

0 SNR

fu 32 0 32 0 32 0 31 1

su 3 29 2 30 1 31 3 29

-5 SNR

fu 32 0 24 8 25 7 25 7

su 9 23 5 27 5 27 11 21

-10 SNR

fu 29 3 22 10 14 18 26 6

su 22 10 14 18 13 19 18 14

-15 SNR

fu 5 27 20 12 29 3 18 14

su 8 24 20 12 18 14 19 13

Table D.9: Confusion Matrices for /fu,su/

216


noiseless va za va za va za va za

va 32 0 32 0 32 0 32 0

za 0 32 0 32 0 32 0 32

10 SNR

va 32 0 31 1 32 0 30 2

za 2 30 2 30 0 32 1 31

5 SNR

va 32 0 32 0 32 0 32 0

za 4 28 2 30 1 31 1 31

0 SNR

va 32 0 31 1 32 0 29 3

za 8 24 6 26 2 30 4 28

-5 SNR

va 29 3 31 1 26 6 27 5

za 12 20 11 21 3 29 15 17

-10 SNR

va 23 9 27 5 19 13 18 14

za 20 12 13 19 12 20 13 19

-15 SNR

va 18 14 25 7 18 14 20 12

za 14 18 13 19 15 17 19 13

Table D.10: Confusion Matrices for /va,za/

217


noiseless vi zi vi zi vi zi vi zi

vi 32 0 31 1 32 0 32 0

zi 0 32 1 31 2 30 0 32

10 SNR

vi 32 0 32 0 31 1 31 1

zi 0 32 1 31 1 31 0 32

5 SNR

vi 32 0 32 0 32 0 30 2

zi 2 30 6 26 1 31 1 31

0 SNR

vi 26 6 28 4 31 1 21 11

zi 5 27 7 25 8 24 0 32

-5 SNR

vi 28 4 22 10 23 9 14 18

zi 15 17 14 18 12 20 4 28

-10 SNR

vi 17 15 21 11 19 13 13 19

zi 15 17 20 12 17 15 9 23

-15 SNR

vi 14 18 20 12 18 14 20 12

zi 12 20 21 11 20 12 15 17

Table D.11: Confusion Matrices for /vi,zi/

218


noiseless vu zu vu zu vu zu vu zu

bu 32 0 32 0 32 0 32 0

pu 0 32 0 32 0 32 0 32

10 SNR

bu 32 0 32 0 32 0 32 0

pu 0 32 0 32 0 32 0 32

5 SNR

bu 32 0 28 4 32 0 29 3

pu 0 32 2 30 0 32 0 32

0 SNR

bu 29 3 26 6 28 4 26 6

pu 1 31 0 32 3 29 2 30

-5 SNR

bu 29 3 23 9 23 9 19 13

pu 12 20 7 25 7 25 11 21

-10 SNR

bu 18 14 15 17 16 16 19 13

pu 12 20 16 16 16 16 20 12

-15 SNR

bu 16 16 12 20 19 13 22 10

pu 16 16 20 12 15 17 14 18

Table D.12: Confusion Matrices for /vu,zu/

219

References

[1] A. Alwan. The role of F3 and F4 in identifying the place of articulationfor stop consonants. Proceedings for International Conference on SpokenLanguage Processing ( ICSLP ), 2:1063–1066, 1992.

[2] S. Behrens and S. E. Blumstein. On the role of the amplitude of the frica-tive noise in the perception of place of articulation in voiceless fricativeconsonants. Journal of the Acoustical Society of America, 84(3):861–867,September 1988.

[3] S. E. Blumstein and K. N. Stevens. Acoustic invariance in speech produc-tion: Evidence from measurements of the spectral characteristics of stopconsonants. Journal of the Acoustical Society of America, 66:1001–1017,1979.

[4] F. S. Cooper, P. C. Delattre, J. M. Borst, and L.J. Gerstman. Some experi-ments on the perception of synthetic speech sounds. Journal of the AcousticalSociety of America, 24:597–606, 1952.

[5] P. C. Delattre, A. M. Liberman, and F. S. Cooper. Acoustic loci and tran-sitional cues for consonants. Journal of the Acoustical Society of America,27:769–773, 1955.

[6] G. Fant. Speech Sounds and Features, chapter Stops in CV-syllables. MIT,Cambridge, MA, 1973.

[7] C. L. Farar, C. M. Reed, Y. Ito, N. I. Durlach, L. A. Delhorne, P. M. Zurek,and L. D. Braida. Spectral-shape discrimination. I. results from normal-hearing listeners for stationary broadband noises. Journal of the AcousticalSociety of America, 81:1085–1092, 1987.

[8] J. S. Gravel and R. N. Ohde. Perception of stop place of articulation: Ef-fects of stimulus amplitude. American Speech-Language-Hearing Associa-tion, 25:101, 1983.

[9] J. A. Guerlekian. Recognition of the Spanish fricatives /s/ and /f/. Journalof the Acoustical Society of America, 70:1624–1627, 1981.

[10] J. J. Hant. A Computational Model to Predict Human Perception of Speechin Noise. PhD thesis, Department of Electrical Engineering, UCLA, 2000.

[11] K. S. Harris. Cues for discrimination of American English fricatives in spo-ken syllables. Language and Speech, 1:1–17, 1958.

220

[12] M. Hedrick and R. N. Ohde. Effect of relative amplitude of frication onperception of place of articulation. Journal of Acoustical Society of America,94(4):2005–2026, October 1993.

[13] M. S Hedrick and W. Jesteadt. Effect of relative amplitude, presentationlevel and vowel duration on perception of voiceless stop consonants by normaland hearing impaired listeners. Journal of the Acoustical Society of America,100(5):3398–3407, November 1996.

[14] M. S. Hedrick, L. Schulte, and W. Jesteadt. Effect of relative and overallamplitude on perception of voiceless stop consonants by listeners with normaland impaired hearing. Journal of Acoustical Society of America, 98(3):1292–1303, September 1995.

[15] J. M. Heinz and K. N. Stevens. On the properties of voiceless fricativeconsonants. Journal of the Acoustical Society of America, 33:589–596, 1961.

[16] G. W. Hughes and M. halle. Spectral properties of fricative consonants.Journal of the Acoustical Society of America, 28:303–310, 1956.

[17] A. Jongman. Duration of frication noise as a perceptual cue to place andmanner of articulation in english fricatives. Journal of the Acoustical Societyof America Supplement, 77:S26, 1985.

[18] A. Jongman. Duration of frication noise required for identification of Englishfricatives. Journal of the Acoustical Society of America, 85(4):1718–1725,1988.

[19] D. Kewley-Port. Measurement of formant transitions in naturally producedstop consonant-vowel syllables. Journal of the Acoustical Society of America,72(2):379–389, August 1982.

[20] D. Kewley-Port. Time varying features as correlates of place of articulationin stop consonants. Journal of the Acoustical Society of America, 73:322–335, 1983.

[21] I. Lehiste and G. E. Peterson. Transitions, glides and diphthongs. Journalof the Acoustical Society of America, 33:268–277, 1961.

[22] A. M. Liberman, P. C. Delattre, F. S. Cooper, and L. J. Gerstman. Therole of consonant-vowel transitions in the perception of the stop and nasalconsonants. Psychological Monographs, 68(8):13, 1954.

[23] B. Lindblom. Spectrographic study of vowel reduction. Journal of theAcoustical Society of America, 35:1773–1781, 1963.

221

[24] G.A. Miller and P. E. Nicely. An analysis of perceptual confusions amongsome English consonants. The Journal of The Acoustical Society of America,27(2):338–352, 1955.

[25] T. M. Nearey and S. E. Shammass. Formant transitions as partly distinctiveinvariant properties in the identification of voiced stops. Canadian Acoustics,15(4):17–24, 1987.

[26] R. N. Ohde and K. N. Stevens. Effect of burst amplitude on the perceptionof stop consonant place of articulation. Journal of the Acoustical Society ofAmerica, 74:706–714, 1983.

[27] R. K. Potter, G.A. Kopp, and H. Green. Visible Speech. Van Nostrand,Princeton, N.J, 1947.

[28] L. Rabiner and B-H Juang. Fundamentals of Speech Recognition. PrenticeHall PTR, Englewood Cliffs, New Jersey 07632, 1993.

[29] C. H. Shadle and S. J. Mair. Quantifying spectral characteristics of frica-tives. Proceedings for International Conference on Spoken Language (IC-SLP), 3:1521–1524, 1996.

[30] R. Smits, L. ten Bosch, and R. Collier. Evaluation of various sets of acousticcues for the perception of prevocalic stop consonants I: perception experi-ment. Journal of Acoustical Society of America, 100(6):3852–3863, Decem-ber 1996.

[31] S. D. Soli. Second formants in fricatives: Acoustic consequences of fricative-vowel coarticulation. Journal of the Acoustical Society of America, 70:976–984, 1981.

[32] S. D. Soli and P. Arabie. Auditory versus phonetic accounts of observedconfusions between consonant phonemes. Journal of the Acoustical Societyof America, 66:46–59, 1979.

[33] K. N. Stevens. Phonetic Linguistics: Essays in Honor of Peter Ladefoged,chapter Evidence for the role of Acoustic boundaries in the perception ofspeech sounds. Academic, New York ,NY, 1985.

[34] K. N. Stevens. On the quantal nature of speech. Journal of Phonetics,17:3–45, 1989.

[35] K. N. Stevens and S. E. Blumstein. Invariant cues for place of articulationin stop consonants. Journal of the Acoustical Society of America, 64:1358–1368, 1978.

222

[36] K. N. Stevens, S. Y. Manuel, and M. Metthies. Revisiting place of articu-lation measures for stop consonants : Implications for models of consonantproduction. ICPhS Proceedings 1999, 2:1117 1120, 1999.

[37] H. M. Sussman, H. A. McCaffrey, and S. A. Matthews. An investigation oflocus equations as a source of relational invariance for stop place categoriza-tion. Journal of Acoustical Society of America, 90(3):1309–1325, 1991.

[38] M. D. Wang and R. C. Bilger. Consonant confusion in noise: a study ofperceptual features. Journal of the Acoustical Society of America, 54:1248–1265, 1973.

[39] H. Y. You. An acoustical and perceptual study of English fricatives. M.S.thesis, University of Edmonton, Edmonton, Canada, 1979.

[40] V. Zue. Acoustic characteristics of stop consonants: a controlled study. PhDthesis, MIT, 1976.

223

Perception of Place of Articulation for Syllable-Initial Consonants … · 2004-07-29 · For my...

Documents

Transcript of Perception of Place of Articulation for Syllable-Initial Consonants … · 2004-07-29 · For my...