State-of-the-art sequence kernels for SVM speaker verification

We present a comparative study of three State-of-the-art SVM speaker verification systems based on sequence kernels: the Generalized Linear Discriminant Sequence (GLDS) kernel, the GMM-supervectors sequence kernel and the feature space normalized sequence (FSNS) kernel. We also compare these three SVM systems to the conventional generative UBM-GMM. We carry out experiments on NISTpsila2005 SRE evaluation set. The results show that the FSNS system significantly outperforms the GLDS one, and that the GMM-supervectors system outperforms all the others. They also show that the fusion of the FSNS and the GMM-supervectors systems leads to the best performances.