Learning the decision function for speaker verification