A Joint Identification-Separation Technique for Single Channel Speech Separation

We present a generalized approach to speaker dependent model-based single channel speech separation techniques in which a priori knowledge of the underlying speakers is used to separate speech signals. For this purpose, we add an identification stage by which we first identify the underlying speakers in the mixture and then use the identified speakers' model to separate speech signals. The proposed technique not only preserves the advantages of model-based speaker dependent single channel speech separation algorithms (i.e. high separability) but also is able to separate the speech signals of an unlimited number of speakers given the speakers' models (i.e. generality). Evaluation results conducted on a database consisting of 100 mixed speech signals with target-to-interference ratios (TIR) ranging -9 dB to +9 dB show significant performance improvements over those techniques which use a single model for all speakers