speaker variability in cue weighting for laryngeal...
TRANSCRIPT
gerFreq.pdf
§ Seoul Korean: Sound change in progress (Bang et al., under review):• F0 is replacing the role of VOT in producing aspirated/lax
stop contrasts (e.g. /pʰ/ vs /p/). • “Trade-off” between the use of VOT and f0
across words, vowel contexts, and individuals.§ Other languages: ∆VOT/∆F0 covariation observed
across individuals • English (Shultz et al., 2013): negative• English, Khmer, Thai, and Vietnamese
(Kirby, 2016): negativeo limited to the languages with long-lag VOT stop category• English (Clayards, 2017): positive
§ Limitations: • Inconsistent results possibly due to the small # of data• How are cues used across words and contexts?
Ø What is the relationship between synchronic covariation in VOT/F0, and diachronic sound change?
Q: What is VOT/F0 covariation in signaling contrasts across word frequencies, vowel contexts, and individuals?• Languages undergoing change (Korean) versus not
(German/English): “trade-off” as a precondition to change?
Bang, H.-Y., Sonderegger, M., Kang, Y., Clayards, M., & Yoon, T.-J. (2015). The effect of word frequency on the timecourse of tonogenesis in Seoul Korean. Proceedings of the 18th ICPHS, Glasgow, Scotland, UK. / Draxler, C. (1995). Introduction to the Verbmobil-PhonDat Database of Spoken German, PrologApplications Conference PAP 95, Paris. / Kang, Y. (2014). Voice Onset Time merger and development of tonal contrast in Seoul Korean stops:Acorpus study. J Phon, 45, 76–90. / Keshet, J., Sonderegger, M., & Knowles, T. (2014). AutoVOT: A tool for automatic measurement of voice onset time using discriminative structured prediction [Computer program]. Version 0.91. https://github.com/mlml/autovot. / Kirby, J. P. (2016). Cross-linguistic variability in cue weighting of consonant voicing. Oral presentation at the Workshop on Higher-order structure in speech variability, The 15th Conference on LabPhon, Cornell University, Ithaca, USA, Jul 13-17. / Shultz, A. A., Francis, A. L., & Llanos, F. (2012). Differential cue weighting in perception and production of consonant voicing. JASA, 132(2), EL95– E101.
§ The use of F0 and VOT are negatively correlated across speakers in all languages.§ Change seems to be progressing by strengthening the existing correlations.§ Correlations across words and contexts exist only in Seoul Korean.§ In Seoul Korean, change is more advanced in the conditions where total cue informativity
(VOT + F0) is smaller in other languages.§ VOT/f0 covariation across speakers may be an origin of sound change due to speakers’
use of efficiently integrated voicing cues.
Speaker Variability in Cue Weighting for Laryngeal Contrasts: the Relationship to Sound ChangeHye-Young Bang1, Morgan Sonderegger1, & Meghan Clayards1,2
[email protected] , [email protected], [email protected] of Linguistics, McGill University ,2 School of Communication Science and Disorders, McGill university
BACKGROUND QUESTIONS CUE COVARIATION ACROSS SPEAKERS
HEIGHT WORD FREQUENCY
4th Workshop on Sound ChangeApr 19 - 22, 2017, Edinburgh, Scotland
DATA
DISCUSSION
decrease
0
4
8
-50
0
50
100
Q1 Q2 Q3 Q4 Q5
Q1 Q2 Q3 Q4 Q5
f0 d
iffer
ence
(st)
vot d
iffer
ence
(ms)
(c) Koreanp = 0.07
F0 d
iffer
ence
(st)
VOT
diffe
renc
e (m
s)
0
4
8
-50
0
50
100
Q1 Q2 Q3 Q4 Q5
Q1 Q2 Q3 Q4 Q5
f0 d
iffer
ence
(st)
vot d
iffer
ence
(ms)
(a) Englishp = 0.12
0
4
8
-50
0
50
100
Q1 Q2 Q3 Q4 Q5
Q1 Q2 Q3 Q4 Q5
f0 d
iffer
ence
(st)
vot d
iffer
ence
(ms)
(b) German
0
2
4
6
8
-25
0
25
50
nonhighhigh
nonhighhigh
f0 d
iffer
ence
(st)
vot d
iffer
ence
(ms)
(c) Korean
F0 d
iffer
ence
(st)
VOT
diffe
renc
e (m
s)
-2
0
2
4
25
50
75
nonhighhigh
nonhighhigh
f0 d
iffer
ence
(st)
vot d
iffer
ence
(ms)
(b) German
0
2
4
6
8
20
40
60
80
nonhighhigh
nonhighhigh
f0 d
iffer
ence
(st)
vot d
iffer
ence
(ms)
(a) English
§ Korean: Trade-off between VOT & F0 contrasts• Total cue informativity constant across vowel contexts
§ German & English: No trade-off • Less informativity in [-high] contexts
p = 0.39
p = 0.01
p = 0.019
p = 0.9
p = 0.04
p = 0.002
p = 0.09 p = 0.8
p = 0.8
p < .001
−2.5
0.0
2.5
5.0
7.5
Coh
en's
D
cue f0 vot
Eng. raw: r = − 0.24, p = 0.006
−2.5
0.0
2.5
5.0
7.5
Coh
en's
D
cue f0 vot
Ger. raw: r = − 0.03, p = 0.7
−2.5
0.0
2.5
5.0
7.5
Coh
en's
D
cue f0 vot
Kor. raw: r = − 0.5, p < .001
0.0
2.5
5.0
0.0 2.5 5.0
Cohen's D for VOT
Cohe
n's
D fo
r f0
Kor. raw: r = − 0.5, p < .001
0.0
2.5
5.0
0.0 2.5 5.0
Cohen's D for VOT
Cohe
n's
D fo
r f0
Eng. raw: r = − 0.24, p = 0.006
0.0
2.5
5.0
0.0 2.5 5.0
Cohen's D for VOT
Cohe
n's
D fo
r f0
Ger. raw: r = − 0.03, p = 0.7
0
3
6
0 3 6VOT coefficient
f0 c
oeffi
cient
Eng. raw: r = − 0.38, p < .001
0
3
6
0 3 6VOT coefficient
f0 c
oeffi
cient
Ger. raw: r = − 0.26, p < .001
0
3
6
0 3 6VOT coefficient
f0 c
oeffi
cient
Kor. raw: r = − 0.44, p < .001
§ Negative correlation between the weights of VOT and F0 across speakers in all languages (Exception: D’ values in German)ü A stronger correlation in the language undergoing change
§ Korean: Greater speaker variability in F0 than VOT (also for almost all speakers, F0 weights are positive)
§ English and German: Greater speaker variability in VOT than F0 (also for all speakers, VOT weights are positive)
Language Korean English German
Corpus NIKL (NIKL, 2005) WPC G3 (Morgan et al., 2005) PhonDat (Draxler, 1995)
# of speakers 118 126 118# of words 60 76 79# of tokens 5559 4208 2660
§ Read speech corpora
§ F0 values converted into semitones§ VOT measured using AutoVOT (Keshet et al., 2014)
Language Korean English German
F0 contrast [-high] > [+high] [-high] < [+high] [-high] ? [+high]
VOT contrast [-high] < [+high] [-high] ? [+high] [-high] < [+high] Language Korean English German
F0 contrast High > Low (weak) High > Low (weak) High ? Low
VOT contrast High < Low (weak) High ? Low High < Low
§ Stops in low frequency words behave similarly to those in highvowel contexts.
§ Cohen’s D:
§ LDA weights:
speakers