Speaker Verification using GMM-UBM with Embedded TDNN

This paper proposes a modified Gaussian Mixed Model-Universal Background Model (GMM-UBM) with an embedded Time Delay Neural Network (TDNN) It integrates the merits of GMM which is a generative model and TDNN as a Discriminative model. TDNN digests the time information of the feature sets, and transmits the information to GMM. Also through the transformation of the feature vectors it makes the hypothesis of variable independence that maximum likelihood needed more reasonable. We train GMM and TDNN as a whole by means of maximum likelihood. In the process of training, the parameters of GMM and TDNN are updated alternately. Experiments show that using the method with TNorm can reduce EER about 28% against baseline GMM-UBM