A hybrid system based on GMM-SVM for speaker identification

Abstract

Gaussian mixture models (GMM) have become the standard method used for speaker recognition systems. A recent discovery is that combining GMM approach with another classifier is an effective method for speaker classification. We consider the GMM supervector in the context of support vector machines (SVM). We construct a support vector machine tested with two kernel functions employing the GMM supervectors. The main idea of the study is to combine the discriminative classifier SVM and the traditional GMM pattern classification with a new dimensional cepstral feature vector extracted from the speech to achieve better classification rate. This idea has been analytically formulated and tested on speakers from TIMIT database. First we describe the SVM-GMM system then we briefly discuss how the new low dimensional feature vector can feed to identification rate. We show comparative results obtained with GMM, SVM, GMM-SVM based system and existing works. Thereafter, we show that the new hybrid system can outperform the standard GMM-SVM based system and give remarkable increases in speaker identification rates.