MicGraphNet: Microphone graph network for cross-correlation-based time delay estimation for accurate sound source localization

Sound source localization (SSL) has gained widespread attention due to its critical role in various applications. As a representative SSL algorithm, the time difference of arrival (TDOA) method estimates the direction of arrival (DOA) of the sound source by calculating the cross-correlation (CC) of measured signals from multiple microphones. For TDOA-based DOA estimation, noisy environments and finite sampling rates complicate precise DOA measurements, while inaccuracies in microphone array (MA) configurations further challenge SSL accuracy. To address these challenges, we propose the microphone graph network (MicGraphNet), a novel deep learning framework leveraging graph neural networks (GNNs) for TDOA-based SSL. MicGraphNet takes advantage of the inductive bias between microphone-array configurations and CC features, enabling efficient learning that compensates for noise- and reverberation-induced errors and thereby improves DOA accuracy. MicGraphNet also directly regresses continuous TDOAs rather than relying on grid-based CC peaks, which mitigates quantization errors and enhances precision. Finally, the model incorporates a position-uncertainty injector, enhancing robustness to practical sensor placement inaccuracies. Experimental validations conducted in free-field, anechoic chamber, reverberant rooms, as well as conversational speech scenarios, demonstrate that MicGraphNet consistently improves DOA accuracy across multiple tetrahedral sub-array configurations. In particular, average DOA error reductions of approximately 14°, 3°, 8°, and 4° were obtained in these respective environments, confirming the model's effectiveness in real-world applications.