Sequential Bayesian learning of CDHMM based on finite mixture approximation of its prior/posterior density

Abstract

Proposes a sequential Bayesian learning strategy of a continuous-density hidden Markov model (CDHMM) based on a finite mixture approximation of its prior/posterior density. The initial prior density of the CDHMM is assumed to be a finite mixture of natural conjugate prior probability density functions (PDFs) of the complete-data density. With the new observation data, the true posterior PDF is approximated by the same type of finite-mixture PDFs which retain the required most significant terms in the true posterior density according to their contribution to the corresponding Bayesian predictive density by using an N-best beam search algorithm. Then, the updated mixture PDF is used in the VBPC (Viterbi Bayesian predictive classification) method to deal with unknown mismatches in robust speech recognition. Experimental results on a speaker-independent recognition task of isolated Japanese digits confirm the viability and the usefulness of the proposed method.