ISCA Archive Eurospeech 1993
ISCA Archive Eurospeech 1993

Bayesian learning of the parameters of discrete and tied mixture HMMs for speech recognition

Qiang Huo, Chorkin Chaw, Chin-Hui Lee

In this paper a theoretical framework for Bayesian adaptive learning of discrete HMM parameters is presented. Formulations of MAP and segmental MAP estimation of DHMM parameters are developed. An empirical Bayes method to estimate the hyperparameters of prior density based on the moment estimate is proposed. We applied the proposed method to speaker adaptation problems using a 26-word English alphabet vocabulary. Speaker-adaptive training algorithm is shown to be effective in improving the performance of both speaker-dependent and speaker-independent speech recognition problems. The method proposed in this paper will also be applicable to other problems in HMM training for speech recognition such as sequential or batch training, context adaptation, parameter smoothing, and so on.

Keywords: Bayesian learning; hidden Markov model; speaker adaptation.