Speaker Verification with SVM and DGMM

The Gaussian mixture model (GMM) is the main technique used in speaker recognition systems. However, in tasks other than speaker recognition, GMM is often outperformed by modern classifiers, such as support vector machines (SVM). This work seeks a better understanding of the reasons for discriminative classifiers not being as successful in speaker recognition as in other applications. This is done by comparing GMM and a novel technique called discriminative GMM, which is similar to SVM in many aspects. Simulation results using the IME corpus show that both SVM and DGMM can improve the performance compared to GMM, and indicate that a proper model selection is essential to make them competitive in speaker verification.