Single Sound Source Localization Using Convolutional Neural Networks Trained With Spiral Source

Abstract

This paper proposes a sound source localization method which is based on deep learning. A convolutional neural network (CNN) model is developed in this method. Unlike traditional sound source localization methods, the proposed CNN can directly estimate the three-dimensional coordinates of a single sound source through learning. The proposed CNN positioning method is divided into two steps. The first step is to prepare trainable data and train the neural network. Features required for positioning were extracted from the training data using CNN. The second step is to input the data to be tested to the neural network. In this paper, the feature extracted by CNN is GCC-PHAT, which is an important feature of sound source localization. In terms of the sound source, a new type of spiral sound source is adopted for training and test. This kind of sound source includes most information about the elevation angle and azimuth angle of a sound source, which makes CNN learn more comprehensively. The experimental results show that the proposed CNN localization method has higher accuracy than the traditional localization method, and also demonstrates the performance of the proposed method.