Speaker identification using the VQ-based discriminative kernels

Abstract

In this paper, a class of VQ-based discriminative kernel is proposed for speaker identification. Vector quantization is a well known method in speaker recognition, but its performance is not superior. The distortion of an utterance is accumulated, but the distortion source distribution on the codebook is discarded. We map an utterance to a vector by adopting the distribution and the average distortions on every code vector. Then the SVMs are used for classification. A one-versus-rest fashion is used for the problem of multiple classifications. Results on YOHO in text-independent case show that the method can improve the performance greatly and is comparative with the VQ and the basic GMM's performances. © Springer-Verlag Berlin Heidelberg 2005.