Combination of SVM and large margin GMM modeling for speaker identification

Most state-of-the-art speaker recognition systems are partially or completely based on Gaussian mixture models (GMM). GMM have been widely and successfully used in speaker recognition during the last decades. They are traditionally estimated from a world model using the generative criterion of Maximum A Posteriori. In an earlier work, we proposed an efficient algorithm for discriminative learning of GMM with diagonal covariances under a large margin criterion. In this paper, we evaluate the combination of the large margin GMM modeling approach with SVM in the setting of speaker identification. We carry out a full NIST speaker identification task using NIST-SRE'2006 data, in a Symmetrical Factor Analysis compensation scheme. The results show that the two modeling approaches are complementary and that their combination outperforms their single use.