Structural Bayesian Predictive Adaptation of Hidden Markov Models

Typical transformation-based model adaptation techniques (e.g. MLLR) in speech recognition systems rely on deriving point estimates of some fixed but unknown parameters (e.g. transformation matrices). These techniques face shortcomings in terms of accuracy and flexibility of modeling the mismatch, accuracy in estimating the parameters, and in efficiency of data usage. In this paper we present a unified framework to address these problems. Bayesian predictive (BP) techniques have been recently introduced which address the problems of accuracy and flexibility by explicitly taking into account uncertainties associated with the parameters to be estimated. This is done via the use of predictive densities in the decision rule. As any Bayesian technique, BP adaptation requires accurate specification of the prior densities. This paper describes how an accurate and efficient estimation of priors densities can be carried out for large vocabulary applications which typically involve a large number of prior densities. The proposed approach is evaluated on a non-native speaker adaptation task using the WSJ Spoke3 corpus.