We propose a novel deep learning model, which supports permutation invarianttraining (PIT), for speaker independent multi-talker speech separation,commonly known as the cocktail-party problem. Different from most of the priorarts that treat speech separation as a multi-class regression problem and thedeep clustering technique that considers it a segmentation (or clustering)problem, our model optimizes for the separation regression error, ignoring theorder of mixing sources. This strategy cleverly solves the long-lasting labelpermutation problem that has prevented progress on deep learning basedtechniques for speech separation. Experiments on the equal-energy mixing setupof a Danish corpus confirms the effectiveness of PIT. We believe improvementsbuilt upon PIT can eventually solve the cocktail-party problem and enablereal-world adoption of, e.g., automatic meeting transcription and multi-partyhuman-computer interaction, where overlapping speech is common.
展开▼
