Fast Speaker Adaptation Based on Triple Diagonal Transform Matrices and Shared Block Matrices
Bo Xu · Dianzi xuebao · 2004
In the Maximum Likelihood Linear Regression (MLLR) framework,this paper proposes two fast speaker adaptation approaches,which are called Speaker Adaptation using Triple Diagonal matrices in the log-spectral domain (SATD) and Speaker Adaptation using Shared Block Diagonal matrices (SASBD) in the cepstral domain,respectively.Based on some prior knowledge,the proposed approaches utilize fewer parameters to describe the variation between speakers,and thus fewer adaptation data are needed to give robust estimation.Experimental results in both the whole-word-modeled isolated word recognition system and the isolated word recognition system using triphones as modeling units show that the proposed approaches can provide faster performance than the traditional MLLR approaches.