A New Framework for Speaker Adaptation using Uniform Styled Bilinear Model

Chunyi Guo, Weiqian Liang, Ming Fan, and Kejun Liu

Keywords

bilinear model, speaker adaptation, singular value decompostion(SVD), speech recognition

Abstract

Speaker adaptation is a key technology for the practical applications of speech recognition. The traditional bilinear model-based method decomposes a matrix composed of speaker-dependent HMMs into the style factor specific to each speaker and content factor across all speakers. In this paper, we present a new framework for speaker adaptation using a uniform styled bilinear model which decomposes original trained speaker-independent HMMs into one single matrix to express all the speaker's styles and a matrix to describe the speech content classes. Using adaptation data from a new speaker, the elements of the style matrix are trained to be adjusted in order to be adaptive to the new speaker, rather than generating a new style matrix. The effectiveness of the new method is demonstrated with experimental results on the connected Chinese digits speech recognition.

Important Links:

Go Back