Robot-Guided Data Collection for Deep Learning-Based 3D Sound Localization

Abstract

This study presents a robust 3D sound localization method using three microphones, based on a deep residual convolutional neural network architecture with 13.6 million parameters. The model was trained on a high-accuracy dataset generated by an industrial robot with a sound source, where the robot followed a randomly generated trajectory during data collection. This novel approach accurately determines the 3D spatial position of the sound source with only three simultaneously recorded microphone signals, which offer the advantage of capturing more detailed spatial information, including depth and height, compared to traditional stereo setups. The trained model achieved a Train R2 of 0.99 and a Validation R2 of 0.96 in predicting the 3D spatial environment of the sound source. © 2025 IEEE.