SLIDE 1 A CROSS-LANGUAGE VOWEL NORMALISATION PROCEDURE
Geoffrey Stewart Morrison & Terrance M. Nearey
*
University of Alberta
*now at Boston University
Social Sciences and Humanities Research Council of Canada Conseil de recherches en science humaines du Canada
Canada
Research supported by:
SLIDE 2
ln(F1) ln(F2)
Single language/dialect
SLIDE 3
ln(F1) ln(F2)
Single language/dialect
vocal-tract length differences
SLIDE 4 ln(F1) ln(F2)
Log-mean normalisation
Nearey (1978)
deviation from speaker mean ln(F1) = ln(F2)
SLIDE 5
ln(F1) ln(F2)
Log-mean normalisation
slide so speaker means have same reference value
SLIDE 6
ln(F1) ln(F2)
Log-mean normalisation
deviation from language/dialect reference value
SLIDE 7
Making a number of simplifying assumptions about language and dialect differences:
SLIDE 8
ln(F1) ln(F2)
Multiple languages/dialects
differences in inventory pattern number and distribution of phonemes (size & skew) affect speaker means Language B log mean Language A log mean
SLIDE 9
ln(F1) ln(F2) inter language correction
GL
Ideal bilingual
GL due to inventory differences, not vocal tract differences
SLIDE 10
ln(F1) ln(F2) inter language correction
GL
Ideal bilingual
Estimate GL from balanced samples of speakers from each language
SLIDE 11
Cross-Language Vowel Normalisation: perception of an instance of a vowel from in terms of vowel categories from languge B (Spanish) language A (English)
SLIDE 12
ln(F1) ln(F2) log-mean normalise all English speakers’ vowels, train model
SLIDE 13
ln(F1) ln(F2) normalise a single token of a Spanish vowel
SLIDE 14 Within-language normalised
within-language normalised token
vowel ln(F1) ln(F2)
SLIDE 15
ln(F1) ln(F2)
Cross-language normalised
add/subtract GL
GL
SLIDE 16 Evaluation data: Acoustic variables: F1, F2 F1, F2 duration
at 25% duration of vowel (difference from 25-75% duration of vowel)
Δ Δ English: / /, / /, / /, / /
/ /, / /, / /
SLIDE 17
Statistical model: discriminant analysis trained on English vowels used to classifiy instances of Spanish vowels a posteriori probabilities (APPs) 3 versions: non-normalised within-language normalised cross-language normalised
SLIDE 18
Monolingual English listeners: classified instances of Spanish vowels in terms of English vowel categories proportions Test value: correlation between model APPs and listener proportions
(pooled across listeners)
SLIDE 19 Results: model
- non-normalised
- within-language normalised
- cross-language normalised
correlation r = .848 r = .853 r = .869
SLIDE 20
Conclusion: The cross-language vowel normalisation procedure increased the correlation between the classification of Spanish vowels by a model trained on L1-English vowel productions and L1-English listeners’ perception of Spanish vowels.
SLIDE 21 250 300 350 400 450 500 600 700 800 900 1200 1400 1600 1800 2000 2400 2800 3200 F1 (Hz) F2 (Hz)
Eng Sp Eng Sp Eng Sp Eng
SLIDE 22 250 300 350 400 450 500 600 700 800 900 40 60 80 100 120 140 160 180 200 F1@25% (Hz) duration (ms)
Eng Sp Eng Sp Eng Sp Eng
SLIDE 23
5 10 15 20
5 10 15 Canonical Discriminant Function 1 Canonical Discriminant Function 2
SLIDE 24
5 10 15 20
5 10 Canonical Discrimiant Function 1 Canonical Discrimiant Function 2
SLIDE 25
Produced Perceived Eng // Eng // Eng // Eng // Sp // .997 .001 .001 Sp // 1.000 Sp // .014 .583 .286 .117
Model
Produced Perceived Eng // Eng // Eng // Eng // Sp // .951 .036 .009 .004 Sp // .005 .003 .982 .010 Sp // .004 .275 .473 .248
Listeners
