Extended maximum a posterior linear regression (EMAPLR) model adaptation for speech recognition

In this paper, a new approach for model adaptation, extended maximum a posterior linear regression (EMAPLR), is described and studied. EMAPLR is an extension of maximum a posterior linear regression (MAPLR) for transform based model adaptation. The proposed approach has a close form solution under the elliptic symmetric matrix variate priors, and it is effective in our speech recognition experiments. EMAPLR is based on a direct MAPLR solution of the transform imageW s without explicitly solving the transformation matrix W . This is fundamentally different from conventional MAPLR and MLLR. Moreover, the proposed EMAPLR approach is incorporated with the structured prior evolution which significantly improves the algorithm efficiency and robustness. The structure of prior evolution in MAPLR is studied and it is shown that under the structured prior evolution, the priors in MAPLR follows a recursive formulation. Experimental results on WSJ (Spoke 3) non-native speaker adaptation task indicates that significant gain over MLLR and MAPLR can be obtained with same amount of adaptation data.