Evaluation of modulation frequency features for speaker verification and identification

In this paper, we suggest the use of mutual information to explore the information provided by the modulation spectrum for speaker verification and identification purposes. The initial representation is first transformed to a lower-dimensional domain using Higher Order SVD and then, the mutual information between speaker identity and features in the transformed domain is computed. Projection of the relevant features back to the original dimensions reveals the modulation spectral components which discriminate speakers. Simulations carried out on YOHO database show that the relevance of modulation spectral features is speaker-dependent.