SR-PFGM++ Based Consistency Model for Speech Enhancement

Diffusion models and the extended Poisson flow generative model (PFGM++) have been applied to speech enhancement. They are sampled via stochastic differential equation (SDE) or ordinary differential equation (ODE), but usually require a large number of sampling steps. Hence, we introduce the consistency models, which allow for high-quality one-step generation with non-adversarial training. Specifically, based on our previous work, SR-PFGM++ (PFGM++ combined with stochastic regeneration) is distilled to train consistency model, resulting in the proposed Consistency Model for speech enhancement. Test results on the VoiceBank-DEMAND dataset show that the proposed model significantly reduces the inference time relative to SR-PFGM++ while maintaining comparable performance. Besides, mismatch test results on the TIMIT+NOISE92 dataset demonstrate the generalization ability of the proposed model.