Text-Independent Speaker Identification in Emotional Environments: A Classifier Fusion Approach

The study of text-independent speaker identification in emotional environments is presented in this paper. The study includes identifying the speaker using speech samples in five basic emotions viz, anger, happiness, sadness, disgust, and fear. The work presented compares the performance of four feature sets: Mel frequency cepstral coefficients (MFCC). Line spectral frequencies (LSF). Teager energy based mel cepstral coefficients (TMFCC) and Temporal energy of subband cepstral coefficients (TESBCC). Next, the performance of the speaker identification is studied with combination of two features MFCC-LSF and TESBCC-LSF. A novel classifier fusion method is proposed and its performance is compared with that of the individual classifiers. The database containing speech utterances recorded in the five basic emotions from thirty four speakers in one of the Indian languages (Marathi) is used for experimentation. Gaussian mixture model is used for classification. Fusion of classifiers enhances the speaker identification accuracy in both emotional and neutral environments.