Robust real-time audio signal enhancement increasingly relies on multichannel microphone arrays for signal acquisition. Sophisticated beamforming algorithms have been developed to maximize the benefit of multiple microphones. With the recent success of deep learning models created for audio signal processing, the task of Neural Beamforming remains an open research topic. This paper presents a Neural Beamformer architecture capable of performing spatial beamforming with microphones randomly distributed over very large areas, even in negative signal-to-noise ratio environments with multiple noise sources and reverberation. The proposed method combines adaptive, nonlinear filtering and the computation of spatial relations with state-of-the-art mask estimation networks. The resulting End-to-End network architecture is fully differentiable and provides excellent signal separation performance. Combining a small number of principal building blocks, the method is capable of low-latency, domain-specific signal enhancement even in challenging environments.
