Efficient Training of GMM Based Speaker Recognition System

Automatic speaker recognition (ASR) is based on speech feature vectors, models, and classifiers. To improve the speaker recognition performance, we must affect at least one of these modules. In this paper we propose to use subband spectral centroids (SSCs) as a complementary features with the traditional MFCC features, and a new GMM training algorithm, with the ultimate goal to search the better mixture component number N for each speaker model, which is fixed in the most speaker recognition systems based on GMM without any priori information, and all speaker models have the same number of components. In experiments, we compared the performance of the proposed scheme with the conventional GMM to show its robustness.