Fibonacci array-based temporal-spatial localization with neural networks

Abstract

This paper proposes a novel Sound Source Localization (SSL) algorithm based on neural networks in the time domain space. Building upon previous research [Tang J, Sun X, Yan L, et al. Sound source localization method based time-domain signal feature using deep learning. Appl Acoust 2023;213:109626] that leverages neural network techniques for sound source localization, our methodology diverges from conventional grid-based approaches by circumventing spatial resolution limitations inherent to meshing through direct prediction of target coordinates via a regression method. We employ the Fibonacci Sphere Algorithm (FSA) to ensure a uniform distribution of microphone array elements, enhancing the array’s response consistency to sound sources from various directions. Our comprehensive model simulates a spatial SSL system within a 10-m spherical space. Experimental investigations have substantiated that the proposed neural network architecture demonstrates exceptional localization precision, as evidenced by the Mean Absolute Errors (MAE) obtained, which are 0.268, 0.304, and 0.287, correspondingly, when applied to 64-, 32-, and 16-element array configurations. Furthermore, our experiments demonstrate the strong generalization capability of the trained models, maintaining satisfactory performance even with element losses ranging from 5 % to 30 %. These findings highlight the potential of neural networks in SSL applications and provide valuable insights for future research and development in this field.