Using GMM with Embedded TDNN to Speaker Identification

Abstract

This paper proposes a modified Gaussian Mixed Model (GMM) with an embedded Time Delay Neural Network (TDNN). It integrates the merits of GMM which is a generative model and TDNN which is a discriminative model. TDNN digests the timing information of the feature sequences, and through the transformation of the feature vectors it makes the hypothesis of variable independence that maximum likelihood needed more reasonable. GMM and TDNN are trained as a whole by means of maximum likelihood probability. In the process of training, the parameters of GMM and TDNN are updated alternately. Experiments show that the proposed model improves accuracy rate against baseline GMM at all SNR, maximum to 21%.