Performance evaluation of a neural network-based beamformer with a small-scale microphone array

Beamforming has been one of the most important techniques in noise reduction under adverse acoustic environments. Traditional beamformers were designed in analytical and adaptive schemes. On the other hands, recent neural network-based beamformers achieve the great success in signal enhancement. The authors have also proposed a broadband neural network-based beamformer, which can deal with wide-band speech signals. The performance of the proposed beamformer was confirmed in computer simulation with an 8-ch equally-spaced linear microphone array, of which the spacing between neighbouring microphones was 10 cm, respectively. In this paper, the performance of the proposed beamformer is evaluated with a hand-held small-scale microphone array, which is developed with miniature MEMS microphones. It is confirmed that the nonlinear beamformer could sharpen the main lobe even with the small-scale microphone array, although an analytical linear beamformer, that is, delay-and-sum beamformer, could not achieve signal enhancement. It is considered that the nonlinear neural networkbased beamformer could enhance minute differences in amplitude and phase among multi-channel observations. The performance evaluation is also carried out under a real environment.