WEIGHTED PRINCIPAL COMPONENT MLLR FOR SPEAKER ADAPTATION
Sam-Joo Doh, Richard M. Stern · 1999
We present two new speaker adaptation methods which apply principal component analysis to maximum likelihood linear regression (MLLR) framework. If we apply MLLR after transforming the baseline mean vectors by their eigenvectors, the variance of the estimates for the MLLR matrix are inversely proportional to their corresponding eigenvalues. We describe two techniques to reduce the variance of the estimation, Principal Component MLLR (PC-MLLR) and Weighted Principal Component MLLR (WPC-MLLR). In experiments using sentences 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. Estimate MLLR matrix Update Recognize test data Update Gaussian means Low recognition accuracy A small amount of data A large estimation error Better estimation Better recognition accuracy We would like to achieve Related Work . Block diagonal MLLR: Reduce...