Permutation invariant training of deep models for speaker-independent multi-talker speech separation | Synapse