Intra-class variation reduction of speaker representation in disentanglement framework.