Modal expansion-based data generation approach for deep learning-enabled sound source localization in a small enclosure

Abstract

Accurately locating sound-emitting objects in small and confined spaces is an important but challenging topic within the field of Sound Source Localization (SSL). Most traditional SSL methods are physics-based, lacking the ability and accuracy in dealing with noisy and reverberant environments. Recently, deep learning-based approaches have emerged, but they typically require large amounts of training datasets and reliable data generation tools. To address these needs, methods for generating SSL datasets, such as Image Source Method (ISM), have been developed, which are capable of modeling large acoustic spaces with moderate reverberations. However, in small confined acoustic spaces, audio signals generated by these methods may fail to capture the dominant features of sound fields due to strong modal behaviors. In this work, we investigate SSL in small spaces by employing Modal Expansion (ME) method to generate training dataset. The general workflow is established first, applicable to a range of similar problems with common modal-dominating features. To validate the method, we choose a representative shoebox model with rigid-walls. The sound field in the enclosure, specifically the Frequency Response Functions (FRFs), are calculated using the proposed method, numerical simulations, and compared with actual experiments. The response functions that correlate the spatial relationships between any receiver and source positions within the enclosure are then transformed into Impulse Response Functions (IRFs) for comprehensive dataset generation. To evaluate the effectiveness of the proposed method, we conduct a series of SSL experiments to prove the capabilities of the proposed dataset generation tools. A neural network is trained, and its prediction accuracy is assessed with extensive validation datasets. This work proposes a promising deep learning method for sound source localization in small spaces. Our related code is available at https://github.com/Devin-Pi/modal-expansion-for-ssl.