On the use of supra model information from multiple classifiers for robust speaker identification

In this paper, we propose a text-independent speaker identification (SI) scheme under uncertainty. In this scheme, extraction of supra model information about probability distributions in the feature space is proposed. Supra modeling is a model clustering technique which groups the speaker models into model sets where the speakers in these sets have similar properties. The scheme uses the Dempster-Shafer (D-S) theory of evidence to combine the model sets of two classifiers which are thought to provide complementary information about the speaker identity. A dependency analysis of classifiers to be combined is presented and it is shown to be effective in avoiding wrong decisions. Experimental results of the classifier combination system is given at the end of the paper.