Speaker Verification based on Single Channel Speech Separation

In multi-speaker scenarios, speech processing tasks like speaker identification and recognition are susceptible to noise overlapped voices. As the voices a complicated mixture of signals, target extraction method from this is good front end solution for further understanding classifying. The quality separation can be assessed by ratio or subjective scoring also accuracy downstream identification. order make model more adapted complex overlapping research investigates incorporate with voiceprint task. This paper proposes feature-scale single channel network connected back verification MFCCT feature, so indicates datasets prepared synthesizing Voxceleb1 data, used training testing. results show that using an objective evaluation effectively improve overall performance, as optimized significantly reduced error rate verification.