Knowledge Distillation from Discriminative Model to Generative Model with Parallel Architecture for Speech Enhancement

Generative speech enhancement methods, especially diffusion methods, have recently gained attention. However, current generative methods face several unresolved issues, such as lagging performance compared to discriminative methods on conventional enhancement metrics and slower inference speeds. In order to address these challenges, we propose a knowledge distillation approach from discriminative to generative models, and we design a parallel architecture with a latent vector fusion module to leverage their respective strengths. We conduct experiments and ablation studies on the LibriMix dataset to validate its effectiveness. The results demonstrate that our approach substantially improves performance over baseline systems in terms of speech quality and inference speed.