End-to-End 3-D Sound Source Localization from the Raw Waveform Based on Stereo Microphone Array

The problem of performance degradation in current sound source localization algorithms under reverberant and noisy environments remains a critical challenge. Consequently, this paper introduces a novel approach to estimate the 3-D position of sound sources directly from raw audio signals using an artificial neural network (ANN), which improves the performance of sound source localization algorithms under reverberant and noisy environments. Instead of relying on handcrafted features, raw audio signals recorded by a tetrahedral stereo microphone array are fed directly into the ANN. This design eliminates spatial symmetry issues found in 2-D microphone arrays and enhances 3-D localization accuracy. Inspired by human auditory systems, a convolutional layer is added after the input layer to simulate frequency analysis to search localization cues in different frequency bands. Furthermore, the proposed algorithm incorporates residual connections (RC) and squeeze-and-excitation (SE: an attention mechanisms). Residual connections introduce raw features into deeper network layers to prevent localized information loss caused by excessive network depth, while also enabling improved model training stability. The attention mechanism dynamically adjusts weights across and within channels, suppressing interference while enhancing localization-critical cues, thereby playing a pivotal role in boosting the algorithm's reverberation and noise resistance. Experimental results demonstrate significant improvements: in semi-anechoic chambers, the method reduces localization errors by 0.2 m and increases accuracy by 10%; in conference rooms, errors decrease by 0.26 m with a 21% accuracy gain. These outcomes conclusively validate the effectiveness of the proposed approach in enhancing robustness against reverberation and noise in sound source localization systems.