Determined Source Separation Using the Sparsity of Impulse Responses

In this paper, we propose an over-determined sound source separation method considering the sparsity of impulse responses. Conventional methods, including independent low-rank matrix analysis (ILRMA), have mainly focused on design of realistic sound generation models, but the separation performance is sometimes not improved due to the incorrectness of the generation models and convergence to some poor local minimum. In the proposed method, we utilize a prior information on the mixing process, i.e., the sparsity of impulse responses, to determine the demixing matrices. Numerical experiments using publicly available impulse responses demonstrate that the proposed method based on ILRMA with supervised bases can robustly obtain better results compared to the standard and the supervised ILRMAs.