Weighted Principal Component Mllr for Speaker Adaptation

We present and describe two new speaker adaptation methods which apply principal component analysis to maximum likelihood linear regression (MLLR). If we apply MLLR after transforming the baseline mean vectors by their eigenvectors, the contributions of these eigenvalues to the variance of the estimates for the MLLR matrix are inversely proportional to their corresponding eigenvalues. In the first new technique, called Principal Component MLLR (PC-MLLR), we reduce estimator variance (at the expense of increased bias) by eliminating the contributions of principal components corresponding to smaller eigenvalues. The second technique, called Weighted Principal Component MLLR (WPC-MLLR) makes use of the contributions of all principal components, but weights them according to the inverse of their putative variance. In experiments using sentences from Spokes 0 and 3 from the 1994 DARPA Wall Street Journal evaluation, the use of WPC-MLLR provided a relative reduction in word error rates of 15.1% for non-native speakers and 6.0% for native speakers compared to conventional MLLR.