Structural MAP speaker adaptation using hierarchical priors

Abstract

Most adaptation methods for speech recognition using hidden Markov models fall into two categories; one is the Bayesian approach, where prior distributions for the model parameters are assumed, and the other is the transformation-based approach, where a pre-determined simple transformation form is employed to modify the model parameters. It is known that the former is better when the amount of data for adaptation is large, while the latter is better when the amount of data is small. In this paper, we propose a new approach, the structural maximum a posteriori (SMAP) approach, in which hierarchical priors are introduced to combine the two approaches above. Experimental results showed that SMAP achieved a better recognition accuracy than the two individual approaches for both small and large amounts of adaptation data.