Sound source localization based on three-element microphone array and RepMobileViT model

Abstract

Traditional microphone array-based sound source localization methods typically begin with signal processing, where the source position is regarded as an estimated continuous value in a certain space. However, in some practical scenarios, such as conference rooms, source localization only needs to focus on some predefined areas, thus can be transformed into a deep learning classification problem. To solve the requirements of small size microphone arrays while maintaining the accuracy of source localization in reverberant environments, a feature extraction method based on sound intensity and an efficient lightweight model are proposed for the direction of arrival (DOA) estimation using a three-element microphone array. Firstly, two types of sound intensity feature spectrograms are designed to effectively extract sound source azimuth information under reverberant conditions. Secondly, the RepViT architecture is introduced into the MobileViT network to construct a new lightweight network, which estimates the source position with smaller parameters and faster speed. Finally, the sound intensity spectrogram is fed into the proposed model to estimate the DOA. Compared to existing methods, the proposed method exhibits more efficient and accurate performance in localizing with small arrays under simulated and real environments.