Superdirective non-linear beamforming with deep neural network

Abstract

Beamforming has been one of important issues in acoustic signal processing, since it can achieve signal enhancement and sound source localization. In general, traditional beamformers are designed by an analytical approach or an adaptive approach. It is, however, difficult to properly optimize the beamformers under the complicated acoustical scene. An alternative non-linear beamforming can be substituted for the linear beamforming. In this study, a flexible framework for optimizing the beamformer is introduced based on a deep neural network. Capturing acoustic signals using a microphone array is regarded as spatial sampling, so that annoying grating lobes appear in beam-pattern when the relationship between the wavelength and the microphone spacing does not satisfy the sampling theorem. The proposed method achieves sub-band beamforming using the non-uniform microphone array with eight nesting microphones, which are carefully designed not to cause spatial aliasing. Feasibility of the proposed method has been confirmed by computer simulation. The proposed non-linear beamformer could successfully achieve superdirectivity compared with conventional beamformers.