Emotional speaker verification with linear adaptation

Speaker verification suffers from significant performance degradation on emotional speech. We present an adaptation approach based on maximum likelihood linear regression (MLLR) and its feature-space variant, CMLLR. Our preliminary experiments demonstrate that this approach leads to considerable performance improvement, particularly with CMLLR (about 10% relative EER reduction in average). We also find that the performance gain can be significantly increased with a large set of training data for the transform estimation.