Maximum a posteriori linear regression for speaker adaptation with the prior of mean

An efficient method for speaker adaptation (SA) is proposed in this paper. Let the relationship between the mean parameters of adapted model and the mean parameters of the speaker independent (SI) model be represented by sets of linear transformations like that of maximum likelihood linear regression (MLLR) approach, we try to estimate the transformations by maximum a posteriori (MAP) criterion. The prior mean distribution is considered in the estimation. The experiments on Mandarin speech recognition show the proposed approach is superior to the MLLR approach when only little speech is available for speaker adaptation.