Robust Sound Localization of Sound Sources using Deep Convolution Network

This paper describes a sound source localization (SSL) approach based on Deep Convolutional Neural Network (CNN). Conventional SSL methods, either based on the time delay of arrival (TDOA) or Subspace such as Multiple Signal Classification (MUSIC), often fail in scenarios where real-time localization is required. To resolve this limitation, in recent years, researchers have been using feature-based neural networks to locate sounds. However, this type of network is not ideal and suffers from the same limitation as feature extraction must be done prior. In this paper, we propose a non-feature based learning approach to bypass the need for feature extraction through the use of convolutional layers to directly learn the features responsible for the non-linear mapping between the raw signals from each microphone in the microphone array to their respective directions and locations. Advantages of this method over conventional feature-based learning approaches are that the computational cost is reduced, and it no longer suffers from low accuracy arises from the features derived from traditional methods. Primarily experimental results have shown promising accuracy on locating sound source as compared to conventional methods and feature-based neural networks.