A Multitask Learning Approach For Sound Source Tracking With Icosahedral Convolutional Neural Networks

Many different modern applications require systems that are able to track sound sources. To this end, recent methodologies combine microphone array signal processing with deep learning. In this paper, a new technique for neural network training using multitask learning is proposed. To illustrate the new approach, a state-of-the-art architecture for source direction of arrival (DOA) tracking based on icosahedral convolutional neural networks (icoCNN) was taken as a starting point for the novel training method. In the proposed strategy, a deconvolutional layer for hexagonal grid icosahedral maps was employed in an autoencoder pre-training stage, in which SNR and T60 classification networks are trained in parallel. The AutoEncoder+icoCNN (AEicoCNN) combination was able to achieve a 6% reduction in root mean square angular error when compared to the regular icoCNN training.