Sound source localization for source inside a structure using Ac-CycleGAN

Abstract

We propose a sound source localization (SSL) method called Ac-CycleGAN, which estimates the position of the sound source inside a structure using the frequency spectrum of the accelerometers (FSAs) observed on the exterior of the structure. Accurately localizing sound sources is crucial for noise mitigation in the development of automobiles, machinery, and home appliances. However, SSL inside a structure from its exterior has its limitations, representing a significant gap in reducing product noise levels. To solve this challenge, the Ac-CycleGAN learns under unpaired data conditions using a small amount of real-environment data and a large amount of simulated data. The Ac-CycleGAN generator contributes to the bidirectional transformation of FSAs across both domains. The discriminator of the Ac-CycleGAN model distinguishes between the transformed and the actual data, while simultaneously predicting the location of the sound source. The proposed model improved SSL performance with an increase in real data and achieves an accuracy exceeding 90% when trained with 80% of the real data (12.5% of the simulation data). Furthermore, despite the imperfections in the domain transformation process by the Ac-CycleGAN generator, it becomes apparent that the discriminator selectively utilizes only the features with a small transformation error to SSL.