Introduction of switching mechanism into mask-based beamformer for distortionless blind source separation in underdetermined scenarios

The minimum variance distortionless response (MVDR) beamformer is widely used for multichannel signal enhancement owing to its ability to suppress interferers while preserving the desired signal. However, its application in real-world scenarios faces two major challenges: the need for the prior spatial characteristics and performance degradation in underdetermined scenarios. To address the former, the mask-based MVDR (MB-MVDR) beamformer framework has been proposed, in which time-frequency (TF) masks are used to estimate the spatial characteristics required for performing the MVDR beamformer. To address the latter, the TF-bin-wise switching (TFS) beamformer has been introduced, which combines multiple MVDR beamformers to selectively suppress dominant interferers in each TF bin. In this study, we propose a novel method for blind source separation in underdetermined scenarios: the mask-based switching MVDR beamformer, which introduces TFS frameworks into the MB-MVDR. First, TF masks are estimated from the observed signals using the complex angular central Gaussian mixture model, requiring no prior training data. These masks are then used to compute the spatial covariance matrix of each source, which are subsequently utilized in the TFS beamformers. Numerical experiments demonstrate that the proposed method improves the signal-to-distortion ratio compared with directly applying the estimated masks.