Speaker Adaptive Training of Appling MAP Estimation for Covariance

Recently there has been a growing interest in speaker adaptive training(SAT). However, errors can often arise when estimatingcovariance matrices in the original SAT framework due to the lack of observations in some Gauss components. This paper presents a novel approachwhich applies maximum a posteriori (MAP) covariance-estimating into original SAT. Experimental results in Switchboard corpus demonstrate thatthe proposed method can deliver significant reductions in word error rate (WER) and raise the robustness of SAT process.