DPARNet-RSE: Toward Angular Region-Customizable Speech Extraction

Abstract

Most existing angular region-wise speech extraction methods face two critical limitations: inflexibility when handling different region boundaries, and performance degradation due to the varying numbers of speakers within the target regions. To address these issues, we adapt our recently proposed DPARNet, a lightweight dual-path attention and recurrent network for speech separation, into DPARNet-RSE, to perform angular region-customizable speech extraction. The key innovations include: (1) a boundary-conditioned attention module that encodes target boundaries into dynamic queries for robust region modeling; (2) a curriculum learning-based training approach that stabilizes convergence by progressively introducing data diversity; (3) a silence probability prediction module that directly triggers silent outputs when no target speaker is detected, effectively reducing speech and noise residuals in zero-target cases. The experimental results demonstrate its superior performance, robustness, generalization capability, and scalability in complex scenarios.