Integrating visualized and auditory features for speaker recognition in reverberant fields

Many applications of the technique of speaker recognition are limited in enclosed spaces such as meeting rooms, subway stations and so on, where the influence of reverberation on speech signals must be suppressed. In this paper, the auditory features such as loudness,Mel frequency cepstrumcoefficient are extracted at first, and then the reverberant speech signals are transformed into images, from which the visualized features such as gray-scale gradient co-occurrence matrix are extracted and then are processed by the principle component analysis (PCA) method. Finally, both modes of features are combined by GMM algorithm to classify speaker signals. The experimental results show that the combined method performs well and achieve a higher recognition rate than traditional single mode features in reverberant fields. Copyright