D-SKD Audio Demo

Abstract

Multi-channel speech enhancement based on deep learning has achieved significant progress. While keeping low complexity, lightweight dual-channel models exploit spatial cues to outperform single-channel models on edge devices. However, their performance often degrades for closely-spaced sources due to unconstrained spatial reliance. To mitigate this, we propose a Dynamic Spatial-aware Knowledge Distillation (D-SKD) framework that enhances spatial discrimination during training. Specifically, D-SKD introduces a dynamic arbitration into the distillation process with a spatially invariant single-channel teacher, enabling the dual-channel model to adaptively regulate spatial reliance without explicit angular supervision. Experiments on dual-channel GTCRN and other models show that D-SKD consistently improves performance for closely-spaced sources while preserving quality elsewhere. Requiring no additional inference overhead or auxiliary inputs, it is well suited for lightweight deployment.

This page is for research demonstration purposes only.

Framework Overview

Overall architecture of the D-SKD framework — **Figure 1.** Overall architecture of D-SKD. (a) Training stage with teacher-student distillation via DAM. (b) Inference stage with only the student branch.

Audio Demo

PESQ, eSTOI, and SI-SDR are computed against the clean speech in each sample. SC-GTCRN is the single-channel teacher, DC-GTCRN is the dual-channel model trained without D-SKD, and Proposed is the DC-GTCRN student trained with D-SKD.