Source Signal Capture in Acoustic Sensor Networks based on Robust Beamforming and Source-Related Cluster Estimation

When multiple sound sources are simultaneously active in an acoustic environment with randomly placed arrays, the identification of the most suitable array becomes crucial for the extraction of a high-quality source signal. This paper investigates a method that uses filter-and-sum beamformers with fixed equiangular look directions across several uniform circular microphone arrays to generate a variety of output signals. These signals are further processed to form signal clusters using deep neural network-based representations and to assign specific beams to sources. To evaluate this method, the signal-to-interference ratio (SIR) and automatic speech recognition on the beamformer signals is employed. Compared to the output of a randomly selected beamformer signal and a previously proposed method, the proposed method substantially improves SIR and reduces word error rates (WERs).