Design, implementation and evaluation of an acoustic source localization system using Deep Learning techniques

Abstract

This Master Thesis presents a novel approach for indoor acoustic source localization using microphone arrays, based on a Convolutional Neural Network (CNN) that we call the ASLNet. It directly estimates the three-dimensional position of a single acoustic source using as inputs the raw audio signals from a set of microphones. We use supervised learning methods to train our network end-to-end. The amount of labeled training data available for this problem is however small. This Thesis presents a training strategy based on two steps that mitigates this problem. We first train our network using semi-synthetic data generated from close talk speech recordings and a mathematical model for signal propagation from the source to the microphones. The amount of semi-synthetic data can be virtually as large as needed. We then fine tune the resulting network using a small amount of real data. Our experimental results, evaluated on a publicly available dataset recorded in a real room, show that this approach is able to improve existing localization methods based on SRP-PHAT strategies and also those presented in very recent proposals based on Convolutional Recurrent Neural Networks (CRNN). In addition, our experiments show that the performance of the ASLNet does not show a relevant dependency on the speaker’s gender, nor on the size of the signal window being used. This work also investigates methods to improve the generalization properties of our network using only semi-synthetic data for training. This is a highly important objective due to the cost of labelling localization data. We proceed by including specific effects in the input signals to force the network to be insensitive to multipath, high noise and distortion likely to be present in real scenarios. We obtain promising results with this strategy although they still lack behind strategies based on fine-tuning.