Mask-based Beamforming Using Complex-valued Neural Network for Recognition of Spatial Target Speech

We propose a novel method of mask-based beam-forming using a complex-valued neural network (CVNN) for recognition of spatial target speech. Conventional neural network-supported mask-based beamformers enhance speech signals regardless of direction of arrival. In contrast, our method enhances target sounds (i.e., speech signals from a set of specified target directions) and suppresses non-target sounds (i.e., speech signals from other directions and noise signals). The CVNN for mask estimation is trained to distinguish between target and non-target sounds. Our method is evaluated in simulated environments and on real-recorded signals. Experimental results show that our method can enhance target sounds and suppress non-target sounds. Our method can configure a beamformer with two target directions using a two-channel microphone array. This demonstrates that the number of target directions is free from the constraints of conventional generalized sidelobe canceller methods, in which the number of directions must be less than the number of microphones.