Exploiting support vector machines in hidden Markov models for speaker verification