Multi-Channel Speech Source Separation And Dereverberation With Sequential Integration Of Determined And Underdetermined Models

In this paper, we propose a joint multi-channel speech source separation and dereverberation method in which multiple speech sources and late reverberation are separated in an unsupervised manner. The proposed method jointly optimizes an auto-regressive (AR) model based speech dereverberation and a time-varying multi-channel Wiener filtering (MWF) based speech source separation. So as to increase separation and dereverberation performance and to overcome the inter-frequency permutation problem, the proposed method adopts a sequential parameter optimization strategy. At first, the parameter is updated based on a determined model, and the permutation problem can be solved based on the non-negative matrix factorization. The determined model reduces reverberation by only the AR model and residual reverberation remains. Inspired by the fact that a parameter of a determined model can be converted into a parameter of a underdetermined model, the proposed method regards residual reverberation as an additional source and reduces residual reverberation with the converted parameter based on the underdetermined model. We further propose additional update of the parameter based on the underdetermined model. Experimental results show that the proposed method outperforms the conventional method based on only the determined model and the proposed additional parameter update based on the underdetermined model is effective.