Convolutional neural network trained with synthetic pseudo-images for detecting an acoustic source

Abstract

We report a detection method of an acoustic source by the convolutional neural network (CNN) that utilizes analytic predictions of sound radiation. The analytic predictions with various source conditions are implemented to effectively collect a large-annotated training dataset, allowing straightforward utilizations of the CNN in the acoustic domain. The data conversion from the synthetic audio signals into the pseudo-images is presented to secure compatibility with actual audio signals in terms of the direction of angle (DOA) estimation. Our source localization network fully trained with the synthetic pseudo-images were verified with various source conditions in a semi-reverberant room. The verifications demonstrate remarkable robustness and noise resistance to estimate the DOA regardless of source conditions. Moreover, considering our network has implemented a small number of short-time audio signals (i.e., three audio signals for 0.1 s), the proposed algorithm can be a breakthrough in a real-time tracking of the acoustic source by hybridizing analytical and data-driven approaches.