Dynamic selection of magnitude and phase based acoustic feature streams for speaker verification

The use of joint acoustic features at the feature level leads to vectors of large dimensions and computational complexity. In this paper we propose a method for dynamic acoustic feature stream selection based on mutual information criterion for speaker verification. The method is based on the intuition that different acoustic features are better suited for recognising different speakers. An optimal selection of a particular feature stream for a particular speaker is necessary, assuming that the feature streams have complementary information. We use acoustic features derived from the magnitude and the phase spectrum for this method as they have empirically demonstrated diversity in previous work. An information theoretic measure based on mutual information is proposed to hypothesise the appropriate feature stream for the appropriate speaker. Separability analysis based on the Bhattacharya distance is also presented to verify this hypothesis. The proposed technique requires a claimed identity for dynamic feature selection during the test phase. We therefore apply it to the task of speaker verification. Reasonable improvements in verification performance are noted from the DET curves. The proposed method also significantly reduces the computational complexity compared to the use of joint feature streams.