Enhancing the robustness of Bayesian methods for text-independent automatic speaker verification

In this paper we present the main advances of the IRISA speech group from 2001 to 2004 in robust methods for Bayesian adaptation of speaker models and Bayesian decision. The probabilistic framework and the state-of-the-art Bayesian approach for automatic speaker verification are first recalled. We then describe two original contributions for robust Bayesian decision. The first one is a score normalization technique whose main advantage is that it does not need any external data as opposed to other score normalizations. The second technique is a constrained Bayesian adaptation scheme which operates a normalization of the speaker models in order to compensate for speakerdependent biases in the verification scores. Experiments using these two methods showed significant improvements over the baseline systems. Finally, theoretical developments of a hierarchical Bayesian adaptation scheme based on a dependency tree structure is presented, with preliminary experiment results.