SVM-Based Text-Independent Speaker Verification Using Derivative Kernel in the Reference GMM Space

This paper proposes a new SVM-based method for text-independent speaker verification using derivative kernel in the reference Gaussian mixture model (GMM) space. The model for speaker utilizes the power of SVM and GMM, reference GMM used first, and then SVM followed. Using the reference GMM, not only clusters and compacts the speech, but also distinguishes the reference speaker and imposters. Then, derivative kernel combines SVM and GMM effectively. Experiments on text-independent speaker verification on NIST SRE 2001 dataset show that the equal error rate (EER) of the new method is reduced to 6.51% from 9.88%.