Discriminative adaptation for speaker verification

This paper describes a speaker verification system in which the talker and imposter models are adapted to achieve maximum discrimination, or equivalently minimum verification error. This goal is accomplishedby extending the minimum error classificationcriterion (MCE) and generalized probabilistic descent (GPD) algorithm to the task of adapting talker model parameters and the corresponding anti-talker model parameters to the test environments so as to minimize an empirical estimate of the verification error rate. We address in the current study adaptation of two types of parameters: the model parameters and the decision threshold. We have obtained substantial improvements in equal error rate by applying combined techniques involving a simplified MAP (maximum a posteriori) method and the GPD algorithm. The equal error rate for a database of 43 talkers with 5 adaptation utterances each was reduced from the previously reported best result of 5.41% [1] to 2.17%. We will discuss several alternative methods that have been investigated in this work to provide comparative insights for the use of discriminative methods in speaker verification tasks.