Neural Network-based Broadband Beamformer with Less Distortion

摘要

Beamforming has been one of the important issues in the field of multi-channel signal processing including acoustic signal processing. A wide variety of beamformers have been proposed for each application. In general, acoustic beamforming deals with broadband signals such as speech signals compared to narrowband beamforming for antenna array and radar applications. Recently, neural network-based non-linear beamformers become popular but have a problem that causes an annoying non-linear distortion on the output signal. In the case of speech enhancement, it is a serious problem because our auditory system is highly sensitive to artificial non-linear distortion on speech signals. This paper proposes to solve the problem with the relaxed dual cost functions in the neural network-based beamformer for speech enhancement. The primary cost function aims at sharpening the beam-pattern, and the second cost function is introduced to achieve decreasing speech distortion. Those cost functions are alternatively used for optimizing the beam-pattern in the frequency range of speech signals. The feasibility of the proposed method is confirmed by carrying out a listening test.