-- A corpus study using logistic regression Yao Yao @NWAV37
description
Transcript of -- A corpus study using logistic regression Yao Yao @NWAV37
-- A corpus study using logistic regression
Yao Yao @NWAV371
Vowel alternation in the pronunciation of THE in American English
BACKGROUND How do you say the word THE?
[dh ah], with a schwa [dh iy], with a high front tense vowel
What is the rule for vowel alternation? Canonical rule: [dh iy] / _ [+vowel]
[dh ah] / otherwise Other stories?
2
BACKGROUND Age (Keating et al, 1994)
TIMIT Corpus of read speech in English Age-dependent pronunciation
3
• Younger speakers have a higher probability of using other vowels than [iy] in “the” before vowel. • No speakers above 50 yrs use other vowels than [i] before vowels.
BACKGROUND Disfluency (Fox Tree & Clark, 1997)
More [dh iy] (81%) than [dh ah] (7%) before suspension of speech.
Ongoing sound change Age Gender? Social class? Dialect?
Online speech production Planning problem Speech rate? 4
DATA Buckeye corpus
40 speakers All residents at Columbus, Ohio Balanced in age and gender 1-hr interview Transcribed at word and phone level
Dataset All tokens of the from all speakers
5
PRELIMINARY COUNTS 8132 instances of the 172 different phonetic transcriptions 10 most common pronunciation cover
84.19% of the tokens Most common syllable structures
CV (N=7003); V (N=913); C (N=164) Most common vowels
[ah] (N=4426); [ih] (N=1808); [iy] (N=1130)
6
At least three vowel variants, instead of two!
PRELIMINARY ANALYSIS Vowel name and duration
[ə] [ɪ] [i] 7
PRELIMINARY ANALYSIS General vowel alternation pattern regarding
the following segment
8
STUDY DESIGN Use logistic regression to model the
alternation among the three vowels ([ah], [ih], [iy]).
Predictor variables include phonological factor: following segment speaker characteristics: age, gender contextual features: disfluency, speech rate
9
CODING VARIABLES Vowel variant (outcome variable)
ah: [ə] ih: [ɪ] Iy: [i]
Following segment C: Consonant V: Vowel U: Non-lingusitic
Age Y: Young (<40 yr) O: Old (>=40 yr) 10
CODING VARIABLES (CONT’D) Gender
F: Female M: Male
Following Disfluency D: Disfluent
Pause Filled pause (um, uh, you know). Repetition (the) Hesitation, cutoff, extended pronunciation
F: Fluent otherwise
11
CODING VARIABLES (CONT’D) Preceding Disfluency
D: Disfluent Similar to following disfluency
F: Fluent
Speed Average speed of the pause-bounded stretch (in
# of syll per second)
12
SIMPLEST MODEL [ah] vs. [iy] Exclude cases followed by non-linguistic
sounds. 5046 cases remain. Predictor variables
Block 1: following segment Block 2: age, gender, and their interaction with
following segment Block 3: speed, presence of disfluency, and their
interaction with other variables Method = Forward stepwise (conditional)
13
SIMPLEST MODEL (CONT’D) Results
Following segment is most significant. Percentage of right prediction: 80.3% 90.6%
Following disfluency is also significant. No other factor or interaction appears significant.
Temporary conclusion Old/young male/female speakers respect the
canonical phonological rule equally well.
14
ABOUT [IH] Some basic facts
Women produce [ih] more often than men (28.2% vs. 21.3%)
Young people produce [ih] more often than older people (23.3% vs. 26.1%)
The majority are followed by consonants (84.5%).
Are these also the factors that would favor [ih] over [ah] or [iy]?
15
A TAD MORE COMPLICATED: [IH] VS. [IY] Exclude cases followed by non-linguistic
sounds. 2675 cases remain. Same independent variables as the previous
model Results
Following segment is the most significant condition (right prediction: 62.8% 80.7%)
Following disfluency is also significant (80.7% 81.4%)
Other significant factors: gender, gender X following segment, speed X following segment
16
[IH] VS. [AH] Exclude cases followed by non-linguistic
sounds. 5747 cases remain. Same independent variables as the previous
model Results
Following segment is still significant, but the significance is reduced (right prediction: 70.8% 71.5%)
Other significant factors: gender X following segment, age, age X gender, following disfluency
17
TEMPORARY CONCLUSIONS Most important factor is following segment,
but the effect is weakest in the ah/ih model. The presence of following disfluency also
affects vowel alternation consistently, and the effect is strongest in iy/ih alternation.
18
EFFECT OF FOLLOWING DISFLUENCY IN IH/IY COMPARISON
19
Speaker characteristics (age, gender) and speech rate fail to enter the model for ah/iy distinction, but do show in the other two models considering the [ih] vowel. In particular, the interaction of gender and following segment shows in both models.
MOVING ON TO CASES FOLLOWED BY NON-LINGUISTIC SOUNDS [ah] vs. [iy] Same model, but with all cases (N=5556) Significant factors
Block 1: Following segment (79.7% 89.0%) Block 2: Age X following segment, age, age X
gender. Block 3: Following disfluency, speed and their
interaction. Speed X following segment. (89.0% 89.3%)
20
MOVING ON TO CASES FOLLOWED BY NON-LINGUISTIC SOUNDS [ih] vs. [iy] Same model, but with all cases (N=2938) Significant factors
Block 1: Following segment (61.5% 78.1%) Block 2: age, gender, age X following segment,
gender X following segment. (78.1% 79.1) Block 3: Following disfluency, speed and their
interaction. (79.1% 80.7%)
21
MOVING ON TO CASES FOLLOWED BY NON-LINGUISTIC SOUNDS [ah] vs. [ih] Same model, but with all cases (N=6234) Significant factors
Block 1: Following segment (71.0% 71.6%) Block 2: age, gender, age X gender. (71.6%
71.7%) Block 3: Following disfluency X speed.
22
TEMPORARY CONCLUSIONS When all cases are included (followed by
consonant, vowel, or non-linguistic sounds) Speaker characteristics enter the models, even
the one for ah/iy distinction. Following disfluency and speed continue to
contribute in all models. The ah/ih distinction is still the hardest to model.
23
EFFECT OF GENDER
24
EFFECT OF AGE
25
GENERAL DISCUSSION Ongoing sound change? - Yes…
The new pronunciation [dh ih] A variant form of [dh ah]? Speaker characteristics at play?
What about elongated [dh ah]? A variant form of [dh iy]? Vowel alternation duration alternation?
Disfluency and speech rate affecting the pronunciation? - Yes… Following (un)filled pauses and repetition Preceding disfluency has no effect 26
NEXT STEP Examine the phonetic makeup of the vowels
Moving from modeling vowel name distinction to modeling continuous variables, such as formants and durations
Include more speaker variables More specific age variable Social class?
Include more contextual measures More types of disfluency Contextual predictability?
27
THANKS! Questions and comments are more than
welcome…
28
REFERENCES Fox Tree, J.E., Clark, H.H. (1997) . Pronouncing "the" as
"thee" to signal problems in speaking . Cognition, 62, 151-167
Keating, P., MacEachern, M., Shryock, A., Dominguez, S. (1994) . A manual for phonetic transcription: Segmentation and labeling of words in spontaneous speech . Manual written for the Linguistic Data Consortium, UCLA Working Papers in Phonetics 88, 91-120
Pitt, M.A., Dilley, L., Johnson, K., Kiesling, S., Raymond, W., Hume, E. and Fosler-Lussier, E. (2007) Buckeye Corpus of Conversational Speech (2nd release) [www.buckeyecorpus.osu.edu] Columbus, OH: Department of Psychology, Ohio State University (Distributor).
29