Speaker verification based on SVM and total variability

Summary form only given. The total variability factor extractor followed by the probability linear discriminant analysis (PLDA) has been the state of art algorithm in text-independent speaker verification. In this paper, we use the Support Vector Machine (SVM) to replace PLDA. The low dimensional i-vectors of the total variability system are used as the inputs of the SVM, and the cosine kernel function is adopted to achieve better discrimination. The proposed method can achieve comparative performance with the PLDA system. Furthermore, the score fusion of SVM and PLDA can obtain even better results. The experiments are conducted on the female part of interview section of the NIST 2012 core test corpus. Compared with the best results of single system, the detection cost function (DCF) of the fusion system can be reduced by 25.1 % and 25.2% for common condition 1 and common condition 3, relatively.