Improving Viterbi Bayesian predictive classification via sequential bayesian learning in robust speech recognition

Abstract

In this paper, we extend our proposed Viterbi Bayesian predictive classification (VBPC) algorithm to a new class of prior probability density function (pdf), namely a family of natural conjugate prior pdf's of the complete-data density in continuous density hidden Markov model (CDHMM) and their mixtures. In this way, we can on-line adapt the prior pdf via a sequential Bayesian learning algorithm when some new data are available, so that the performance of VBPC can be continuously improved. Moreover, we also study a sequential Bayesian learning strategy for CDHMM based on a finite mixture approximation of its prior/posterior density which attempts to derive a more accurate prior pdf to describe the unknown mismatches. The experimental results on a speaker-independent recognition task of isolated Japanese digits confirm the viability and the usefulness of the proposed method.