SVM Based Speaker Verification and Gender Dependent NAP Variability Compensation

In recent years, Support Vector Machine is used in many application areas and has shown dramatic achievement. In this paper, we apply it to a text-independent speaker verification task using the NIST 2001 Speaker Recognition database. Starting from a baseline based on Gaussian mixture models, we use the state-of-the-art GMM supervector and SVM to improve the performance. We alter several kernels and find out the linear kernel yields the best performance. Finally, the latest compensation method nuisance attribute projection (NAP) is examined, and the gender-dependent NAP shows more reduction than gender-independent NAP in equal error rate.