Speaker identification based on Classification Sub-space Gaussian Mixture Model

Abstract

This paper proposes a Classification Feature Sub-space Gaussian Mixture Model (CGMM), which can improve the training efficiency of conventional Gaussian Mixture Model (GMM) in speaker identification. By taking the advantage of the centralization tendency of similar features in phonetic signals, CGMM uses Vector Quantization (VQ) technique to cluster the similar features into a sub-space. In the procedure of training, it establishes a GMM for each sub-space instead of a GMM for all the feature vectors. Our experimental findings show that as the feature vectors were more concentrated in each sub-space, CGMM enhanced the training efficiency and recognition rate of speaker identification as compared with conventional GMM.