Segmental quasi-Bayesian learning of the mixture coefficients in SCHMM for speech recognition

A theoretical formulation of the segmental quasi-Bayes learning of mixture coefficients in semi-continuous hidden Markov models (SCHMM) is presented. Its viability is confirmed by a series of comparative experiments using different adaptive training algorithms in estimating the mixture coefficients of the SCHMM for speaker adaptation (SA) application. Despite the fact that a batch (or block) adaptation scheme is adopted in this study, the proposed segmental quasi-Bayes method is also very suitable for performing an incremental (or on-line) adaptation of the HMM parameters, in view of its sequential nature in updating both the hyper-parameters of the prior distribution and the mixture coefficients themselves