Accurate separation of overlapping sound sources in reverberant environments remains a critical challenge for spatial hearing devices and acoustic scene analysis. Most existing source separation approaches do not explicitly incorporate distance cues, limiting their performance in natural acoustic environments where reverberation and near-far effects significantly alter the observed signals. In this work, we propose a distance-aware source separation framework that exploits multichannel spatial cues to improve separation in room-acoustic conditions. Using simulated room impulse responses with varying source-to-microphone distances, we extract spatial features such as interchannel time and level differences, direct-to-reverberant ratios (DRR), and multi-resolution power envelopes. These features are integrated into a deep learning model trained to disentangle and reconstruct individual source components from mixed recordings. We evaluate our method on synthetic mixtures generated using image-source modeling across a variety of room sizes, source arrangements, and reverberation times. This study highlights the potential of incorporating spatial and distance cues for more robust source separation and lays the groundwork for future applications in hearing aids, teleconferencing, and spatial audio systems.
Rameshbabu et al. (Wed,) studied this question.