A Novel Data Description Kernel Based on One-Class SVM for Speaker Verification
Yufeng Shen, Yingchun Yang · 2007
In this paper we develop a novel data description kernel based on one-class SVM (OCSVM-DD kernel) used for text-independent SVM speaker verification. The basic idea of the new kernel is to combine the data description model OCSVM with SVM discriminant classifier. Utterances are firstly mapped to the normal vector of the separating hyperplane in OCSVM model. Then a SVM classifier with linear kernel is applied on those mapped vectors. Experiments results on NIST 2001 SRE database show that the performance of our new kernel is superior to generalized linear discriminative sequence (GLDS) kernel and comparative with UBM-MAP-GMM method.