Exploiting Reverberation Fingerprints for Neural Network-Based Acoustic Emitter Localization

• Published in: 2024 10th International Conference on Control, Decision and Information Technologies (CoDIT) pp. 546 - 550
• Main Authors: Apolinario, Jose A., Ramos, Antonio L. L., Fernandes, Rigel P., Duarte, Julio Cesar, de Sousa, Marcelo N.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.07.2024
• Subjects: Feature extraction; Feedforward neural networks; Fingerprint recognition; Location awareness; Noise measurement; Reverberation; Software; Software measurement; Time difference of arrival; Training
• ISSN: 2576-3555
• DOI: 10.1109/CoDIT62066.2024.10708420

This paper addresses the problem of localizing a sound source in strong reverberating and conceivably noisy environments and its relevant potential applications, e.g., target location in interior room acoustic environments, in- and outdoor navigation, and robotics. Available audio emitter location techniques, especially those based exclusively on Time Difference of Arrival, do not perform well in unfriendly environments like those discussed herein. Recent works propose using neural networks to solve the emitter location problem but do not adequately address the issue of severe reverberation. In this work, we propose a solution that uses a two-layer feedforward neural network supported by signal processing techniques to extract features to train the neural network. Indeed, using features incorporating information from the reverberation fingerprint of the environment to train the network results in an efficient system featuring reduced training time and relatively low mean squared error values.