LDMSE: Low Computational Cost Generative Diffusion Model for Speech Enhancement

Recently, a generative model called diffusion model has attracted attention. Compared to GANs, it can be trained stably but has a high computational cost in the generation stage. This paper proposes a method called Low computational cost Generative Diffusion Model for Speech Enhancement (LDMSE). It reduces its computational cost with comparable quality by compressing speech signals to a latent space using an autoencoder and removing noise in the latent space using the diffusion model. In our evaluation using VOICBANK-DEMAND and WSJ0-CHiME3 datasets, the proposed method reduced the generation time by more than 35% without any degradation in speech quality.