Discrete expected likelihood kernel for SVM-based speaker verification

The construction of kernel functions to handle sequences of speech feature vectors is crucial in using support vector machine (SVM) for speaker verification. Previous studies have reported the idea of representing speech signals as sequences of discrete acoustic or phonotactic events. This paper introduces a class of SVM kernels derived based on the expected likelihood measure between the probability distributions of discrete event sequences. We investigate and compare the effectiveness of three expected likelihood kernels using the universal background model (UBM) as the discrete event detector. Experiments conducted on the NIST 2006 speaker verification task indicate that the proposed kernel outperforms the popular rank-normalized kernel.