A New Reduced-Interference Source Separation Method Based on a Complementary Combination of Masking Algorithm and Mixing Matrix Estimation

摘要

Array signal processing, as a versatile approach, can be used in both source separation and source localization applications. In the realm of the multi-channel blind source separation, time-frequency masking methods and data-dependent beamforming algorithms are commonly used, and the proposed approach in this paper employs a complementary combination of the two methods. Most mask-based approaches suffer from a basic problem: the use of a real-valued mask is not justified in some time-frequency blocks. Thus, we use both mask and spatial information of sources simultaneously to reduce interference and distortion in the separated sources. Moreover, a novel mixing matrix estimation is introduced in this paper that employs information extracted from binary/fuzzy clustering of a high-performance feature set. To employ the complementary combination in source separation approach, a criterion is used that splits time-frequency points into two groups. For the first group, masking method is used for separation, whereas for the second one, mixing matrix is employed. Experimental results show that the proposed source separation method is very promising in anechoic conditions, as well as in the low-reverberant ones. In comparison with other masking-based approaches, especially for the anechoic case, the results get much closer to the best theoretically achievable signal to distortion ratio.