A neural network based microphone array approach to grid-less noise source localization

Deep learning and Neural Networks strategies have become very popular in the last year as tools for image and data processing. As for acoustics, neural network-based approaches have been typically used to recognize audio patterns or features or to spatially localize a single emitting source like a speaker. More recently, some authors used deep learning to localize multiple-sources exploiting the grid-based approach typical of sound source localization methods or to filter/improve acoustic maps obtained by more traditional techniques like conventional beamforming. This paper wants to propose the use of artificial neural networks (ANNs) for localizing and quantifying multiple sound sources in a grid-less way. The approach uses the microphones Cross-Spectral-Matrix (CSM) as input to the network and provides as output both the location and strength of sources contributing to the acoustic field. The grid-less strategy targets improving spatial resolution and computational efficiency. The proposed solution is discussed on simulated data for assessing its accuracy and sensitivity. Preliminary investigations on real data are also reported.