SVM speaker verification based on NAP sequence kernels

For the sake of solving the problem of variable-length feature vectors and channel impact which existed in SVM speaker verification, a novel kernel function based on GMM supervector, called NAP mapping KL divergence linear kernel function, was proposed in this paper. This kernel could enable SVM to classify on whole audio sequences, and also had the benefit that channel subspace, which cause variability, could be removed in kernel space. By doing so, the classification performances of SVM was improved excellently. Our simulation experiment results demonstrated the effectiveness of the new kernel.