Speaker Verification Using Sequence Discriminant
Vincent Wan, Steve J. Renals · 2005
This paper presents a text-independent speaker veri- fication system using support vector machines (SVMs) with score- space kernels. Score-space kernels generalize Fisher kernels and are based on underlying generative models such as Gaussian mix- ture models (GMMs). This approach provides direct discrimina- tion between whole sequences, in contrast with the frame-level ap- proaches at the heart of most current systems. The resultant SVMs have a very high dimensionality since it is related to the number of parameters in the underlying generative model. To address prob- lems that arise in the resultant optimization we introduce a tech- nique called spherical normalization that preconditions the Hes- sian matrix. We have performed speaker verification experiments using the PolyVar database. The SVM system presented here re- duces the relative error rates by 34% compared to a GMM likeli- hood ratio system.