PFGM++ Combined with Stochastic Regeneration for Speech Enhancement

Diffusion models have been applied in speech enhancement due to its capability to learn complex data distributions. However, the extended Poisson flow generative model (PFGM++) outperforms the diffusion models in terms of robustness. In this work, we introduce PFGM++ to speech enhancement, and SR-PFGM++, which samples using ordinary differential equation (ODE), is proposed by combining the stochastic regeneration model (StoRM) with PFGM++. The testing results on the VoiceBank-DEMAND dataset show that SR-PFGM++ achieves a higher performance with fewer sampling steps compared with StoRM. We also performed a mismatch test on the TIMIT+NOISE92 dataset and the results show the strong generalization capability of SR-PFGM++.