Impact Noise Suppression with MWF and Low-Rank Approximation for Binaural Hearing Aids

In this paper, we propose a speech enhancement method with suppressing impact noise for binaural hearing aids. Recently, speech enhancement using deep learning has been proposed, however it is difficult to mount it to hearing aids with limitations such as size and weight. Therefore, we adapt the multi-channel Wiener filter (MWF). The MWF is widely used to suppress stationary noise while preserving acoustic spatial information by using the spatial correlation matrix, but it cannot suppress impact noise with fast power change in short time. We propose to use low-rank approximation for improving the estimation accuracy of spatial correlation matrix of impact noise. Simulation results show that the proposed method suppresses various kind of impact noise while keeping acoustic spatial information.