Recognition of Voice Onset Time for Use in Detecting Pronunciation Variation

1. Recognition of Voice Onset Time for Use in Detecting Pronunciation Variation • Project Description • What is Voice Onset Time (VOT)? • Physical Realization • Linguistic Significance • Motivation for studying VOT • Methodology for automatically analyzing VOT contrasts • Evaluation Method • Results • Discussion

2. Project Description • Automatically distinguish whether a voiceless stop consonant is pronounced with a native or accented pronunciation based on voice onset time characteristics. • Use data from the Tball corpus: ESL children doing oral reading tasks. • Evaluate different methods of accomplishing this. • State duration measurements • Explicit modeling of aspiration • Model probablility discrimination

3. What is VOT? • Voice onset time is defined for stops • e.g. /p,b,t,d,k,g/ • It is the inverval between the release of closure of an articulator (the transient “burst”) and the start of voicing. • VOT has a continuum of values: • When the start of voicing precedes the release of closure for a stop, the VOT takes on a negative value. • When the release of closure and onset of voicing are coincident, VOT is zero. • When voicing comes after release of closure, VOT is positive.

4. Physical Realization of VOT • Stop consonants are produced with a closure of the vocal tract at a specific point, the place of articulation • During the closure, there is a build up of sub-laryngeal pressure. • When the closure is released there is a transient burst of air, frication due to turbulence at the place of articulation, aspiration noise from turbulence at the glottis • Voicing may occur before, during, or after the closure.

5. Linguistic Significance of VOT • VOT distinguishes consonants with the same place of articulation (/p/ vs. /b/, /t/ vs. /d/, etc.) • However, different languages use different VOT intervals in contrasts (e.g. “taco”, “pasta”). • English voiceless stops: VOT= +40-50 ms • Spanish voiceless stops: VOT= near zero • English voiced stops: VOT = near zero • Spanish voiced stops: negative VOT (voicing before closure

6. Linguistic Significance Cont'd • In English, voiceless stops are have a long VOT at the beginning of a word and before stressed vowels, so aspiration is a perceptual cue to word boundaries and stress • Since the frication and aspiration during the VOT is due to build up of pressure from the lungs, it may correspond with emphasis.

7. Motivation for Studying VOT • This study was motivated by a desire to determine if a phone was pronounced with a non-standard pronuniation • Other reasons to study VOT • It is an important contrastive feature • It gives information about stess • It gives information about word segmentation • It may give information about emphasis

8. Methodology • Baseline: use duration measurements from a forced alignment. • Insert an /h/ symbol in the transcriptions with standard pronunciation, train accordingly and decode the test files to see if the /h/ phone is recognized. • Cut out the phones of interest from the audio file, train separate models and a combined model, and evaluate the likelihood of the separate models w.r.t. the combined model.

9. Methodology (cont'd) • The data was transcribed by ear with special symbols for non-standard pronunciations. • b/c the data for non standard pronunciatons was sparse, the symbol for dental /t/ was included as short VOT. • Standard 3 state HMM models • 4 mixtures, T-state silence model • Different frame rates were tested • Bootstrap and flat start methods were tested

10. Evaluation Method • The evaluation metric used was the error rate for both classes evaluated separately. • This was necessary because the there were much fewer instances of the non-standard pronunciations. • When using thresholds, the point of equal error rate for both classes was used. • This was necessary b/c moving the threshold would tilt the error rate toward one class or the other.

11. Results • Baseline method error rates: • p: 55% t:23% k:29% • p: 19% t:20% k:48% using duration of 3rd HMM state • With aspiration model: • ShortVOT/ LongVOT • p: 5% / 36% • t: 11% / 38% • k: 57% / 17% • With probability comparision: • p: 36% / 4% • t: 0% / 5% • k: 0% / 6% • (trained on test data—over trained?)

12. Discussion • Studies have noted that for VOT k>t>p • This could explain why the baseline gets poor results for p • and why the aspiration model predicts the short VOT class best for /p,t/ but predicts the long VOT class best for /k/ • Roughly, each method increased in difficulty. • The results improved from the baseline, but the last approach (comparing probabilities) may have been over-trained. • Comparing probabilities may be easier to extend to other pronunciation modeling tasks.

13. Discussion • Increasing the frame rate didn't help much. • Don't use a 1ms frame rate Unless you want to test your patience. • If an Inintial consonant has a short VOT, this does not necessarily imply non-standard accent. • Words like “today” and “together” have stress on the 2nd syllable, so the VOT of the initial consonant is shorter for even for standard pronunciation.

14. Conclusion • When classifying stop consonants based on VOT characteristics, different approaches work better on different stops • Measuring duration of stop state works reasonably well for /t,k/ b/c longer VOT than /p/. • Detecting insertion of an aspiration model during decoding works well for /p,t/ but not k, which has too many false positives. • Comparing phone probabilities worked well except for unaspirated /p/

15. Future Work • Since VOT is a time/timing related phenomenon, it may help to explicitly model the state duration density in the HMMs. • Other optimization criteria might be be better suited than maximum likelihood extimation to train models for this purpose