190 likes | 288 Views
Rejection of out-of-vocabulary words using phoneme confidence likelihood. Author: takatoshi jitsuhiro Present: 黃怡寧. Outline. Introduction Rejection by phoneme confidence likelihood Minimum error discriminative training Experiments Conclusion. Introduction.
E N D
Rejection of out-of-vocabulary words using phoneme confidence likelihood Author: takatoshi jitsuhiro Present: 黃怡寧
Outline • Introduction • Rejection by phoneme confidence likelihood • Minimum error discriminative training • Experiments • Conclusion
Introduction • Goal: A task independent rejection method that offers high rejection performance.
Rejection by phoneme confidence likelihood • Recognition processing using phoneme confidence likelihood : log of phoneme confidence likelihood : accumulated log likelihood Where is feature vector from time t1 to t2 αis constant
Definition of phoneme confidence likelihood Where is the log likelihood of the i phoneme model of the candidate for feature of the input voice at time t, and N is the total number of phoneme models, and di=t2-t1 is the duration.
Phoneme confidence likelihood Where a and b are constant. • is taken between 0 and 1. • It approaches 1 when the likelihood of the current phoneme model is relatively larger than of other phoneme models.
Use past records of confidence likelihood Where is the j-th past record of the i-th phoneme, and M is the number of records. • The record of confidence likelihood is not used when M=0.
Minimum error discriminative training (MCE) • : parameter set • : discriminant function • : misclassification function • K: number of competition candidates • η: constant
The class loss function • β and γ: constants • : a small positive real number • Vt: a positive definite matrix
Experimental resulrs • No phoneme confidence likelihood (baseline) • Phoneme confidence likelihood (no past record) • 1 past record • 2 past record
Conclusion • The rejection accuracy improvement of unknown words was achieved by introducing the phoneme confidence likelihood at each phoneme during the search process. • The rejection performance was further improved by MCE training algorithm because the phoneme confidence likelihood was • made more accurate by this algorithm.