New technique to use the GMM in speaker recognition system (SRS)

Gaussian mixture models (GMM) have been widely applied in speaker recognition system (SRS); it is the baseline speaker modeling approach. A GMM is composed of a joint probability distribution function (PDF) described by the weighted sum of several multivariate Gaussian PDFs, each multivariate Gaussian PDF is termed as a Mixture Component, The Mixture Component Number (N) is fixed at the classical method in the beginning of training phase in this case all speakers have a GMM model with the identical mixture number exp (16, 64, 128). To enhance effectiveness of speaker recognition system based on GMM we propose in this article a new technique used training GMM algorithm to calculate the best number mixture component for each speaker model. Results show that the new method can improve the performance compared with the basic GMM.