Log-Likelihood Kernels Based on Adapted GMMs for Speaker Verification

Kernels in a SVM-based text-independent speaker verification system determine the performance. One of the main difficulties in designing a kernel arises from the unequal length of cepstral vector sequences. To simplify the above problem, time information is discarded and each speaker is presumed to have a unique probability density distribution. Gaussian mixture models (GMMs) are often used to estimate the probability density distribution from the train cepstral vector sequence. The methods of constructing SVM kernels by adapted GMMs become an open and key question in a GMM-SVM system. In this paper, we introduce a novel way of measuring the similarity between adapted GMMs and propose a log-likelihood kernel. We demonstrate that the presented kernel has an excellent performance on the National Institute of Standards and Technology (NIST) speaker recognition evaluation (SRE) 2008 tel-tel English corpus.