Section: New Results

Discriminative learning for automatic speaker recognition

Participants : Khalid Daoudi, Reda Jourani, Régine André-Obrecht.

Most of the speaker recognition systems rely on generative learning of Gaussian Mixture Models (GMM). During the last decade, discriminative approaches have been an interesting and valuable alternative to address directly the classification problem. For instance, Support Vector Machines (SVM) combined with GMM supervectors are among state-of-the-art approaches in speaker recognition. Recently a new discriminative approach for multiway classification has been proposed, the Large Margin Gaussian mixture models (LM-GMM). These latter methods have the same advantage as SVM in term of the convexity of the optimization problem to solve. However they differ from SVM because they draw nonlinear class boundaries directly in the input space, and thus no kernel trick is required. We continued our work on investigating simplified versions of LM-GMM for speaker recognition that can handle large scale databases. We developed a new and efficient learning algorithm and evaluated it on NIST-SRE data. The results show that this new algorithm not only outperforms both the original LM-GMM and the traditional GMM, but also outperforms state-of-the-art discriminative methods such as GMM-supervectors SVM.

Related Publications: [9] , [10] , [11] , [12] , [8] .