Optimizing feature set for speaker verification

Abstract

This paper proposes a framework for the optimization of the feature set, in an HMM-based text-dependent speaker verification system, in which we distinguish the alignment task from the scoring task. The optimization is based on the search, among a set for potential features, for the feature subset that gives the minimal experimental Equal Error Rate. We have studied and compared various heuristics to find the optimal subset. We have also extended this optimization principle to the search for an optimal weighting of the different axes of the acoustic space. The optimal weighting was found by using a genetic algorithm.

The proposed framework was applied to study cepstral coefficients and their first and second derivatives, in order to find a speaker and text-independent optimal feature set. Experiments were conducted on a large scale telephone database. The results indicate that the selection of an appropriate feature set, or the appropriate weighting of the features, could significantly improve verification performance, especially when little training data is available. Practically, it was found that cepstral coefficients of high order and first derivatives of all cepstral coefficients are the most useful for speaker verification.