Study of mutual information for speaker recognition features

Feature extraction is an important stage in speaker recognition systems since the overall performance depends on the type of the extracted features. In the framework of speaker recognition, the extracted features are mainly based on transformations of the speech spectrum. In spite of the great variety of features extracted from the speech, the common empirical approach to select features is based on a complete performance evaluation of the system. In this paper, we propose an information theory approach to evaluate the information about the speaker identity contained on the speech features. The results show that this approach can help on a more efficient feature selection. We also present an alternative AM-FM based magnitude representation of the speech that attains better performance than the MFCCs. Moreover, we show that phase information features can perform well in speaker verification systems.