SVM Kernel Combining System for Speaker Verification

Authors

Charles-Alban Deledalle

Type

techreport

Year

2008

Number

0757

Keywords

Gaussian Mixture Models, Support Vector Machines, Kernel functions, Score fusion, Kernel combining

Abstract

The best speaker verification systems are based on score combination of several approaches. Support Vector Machines (SVM) give very hopeful results. Thus, combining these methods could be very efficient. In our approach, we propose a new combination method for speaker verification systems based on SVM methods. This one performs a linear combination of several kernel functions in order to produce a new kernel function. In this combination, the weights are speaker dependent, by opposition of score combination approach for which the weight are universal. The idea is to adapt the combination weights for each speaker in order to take the advantage of the best kernel. In our experimentcombinations are performed on several kernel functions: the GLDS kernel, linear and Gaussian GMM supervector kernels. The method can use every kernel functions with no modification. The experiments are done on the NIST-SRE 2005 and 2006 (all trials) database.