Probabilistic sequence kernel for speaker recognition