Inter-speaker weighted MAP adaptation for GMM-supervector speaker recognition

摘要

Gaussian Mixture Models (GMM) are ubiquitously used in state-of-the-art speaker recognition systems. The popular GMM-SVM paradigm uses Maximum A Posteriori (MAP) speaker-adapted GMM models by stacking the mean vectors into a supervector that is fed into a Support Vector Machine classifier. In this paper, we modify the standard relevance MAP algorithm to better fit the speaker recognition task. We propose to emphasize the adaptation of the Gaussian mixtures according to the inter-speaker variability exhibited on a training set, thus accounting for both the occupation count and the speaker discrimination ability during adaptation. We evaluate our proposal on a relevance MAP based GMM-SVM system using a large telephone speech corpus such as the one provided in the 2006 NIST Speaker Recognition Evaluation. We show that despite its simplicity this technique is effective.