Combining Gaussian Mixture Models and Segmental Feature Models for Speaker Recognition

In most speaker recognition systems speech utterances are not constrained in content or language. In a text-dependent speaker recognition system lexical content of speech and language are known in advance. The goal of this paper is to show that this information can be used by a segmental features (SF) approach to improve a standard Gaussian mixture model with MFCC features (GMM-MFCC). Speech features such as mean energy, delta energy, pitch, delta pitch, the formants F1 – F4 and their bandwidths B1 – B4 and the difference between F2 and F1 are calculated on segments and are associated to phonemes and phoneme groups for each speaker. The SF and GMM-MFCC approaches are combined by multiplying the outputs of two classifiers. All the experiments are performed on the two versions of TEVOID: TEVOID16 with 16 and the upgraded TEVOID50 with 50 speakers. On TEVOID16, SF achieves 84.23%, GMM-MFCC 91.75%, and the combined approach gives 95.12% recognition rate. On TEVOID50, the SF approach gives 68.69%, while both GMM-MFCC and the combined model achieve 95.84% recognition rate.. On both databases, the number of male/female confusions decreased for the combined model. These results are promising for using segmental features to improve the recognition rate of textdependent systems.