International Symposium on Chinese Spoken Language Processing (ISCSLP 2002)

Taipei, Taiwan
August 23-24, 2002

Testing the Hypothesis of Multivariate Normality in Bayesian Approaches to Speaker Adaptation

Li-Wei Wang, Zuo-Ying Wang

Tsinghua University, Beijing, China

Bayesian approaches to speaker adaptation are popular in Automatic Speech Recognition (ASR) systems. In most kinds of Bayesian adaptation, there are parameters whose prior distributions are assumed to be multivariate normal. This paper presents a methodology, which can test the hypothesis of multivariate normality. When applied to Maximum A Posterior (MAP) adaptation, we found that the real prior distributions of the mean vectors are far from normal, which are always assumed in the MAP procedure. This result implies that better choice of the prior form may improve the adaptation result.

Full Paper

Bibliographic reference.  Wang, Li-Wei / Wang, Zuo-Ying (2002): "Testing the hypothesis of multivariate normality in bayesian approaches to speaker adaptation", In ISCSLP 2002, paper 18.