Estimation of sound source direction of arrival map using convolutional neural network and cross-correlation in frequency bands

Abstract

In this paper, we present an evaluation of the usage of a convolutional neural network (CNN) for the estimation of the sound source direction of arrival (DoA) map. Crosscorrelations in different frequency bands, calculated for pairs of microphones were used as input features. We propose a technique for generating data for the CNN training, a means of presenting the direction of arrival information for an arbitrary number of sound sources and a viable CNN architecture. In our proposed approach for sound DoA estimation, there is no need for prior knowledge of the number of the active sound sources nor the properties of their signals. We present the results of the evaluation of the two distinct CNN architectures.