Speech Enhancement Using Dynamical Variational AutoEncoder

This research focuses on dealing with speech enhancement via a generative model. Many other solutions, which are trained with some fixed kinds of interference or noises, need help when extracting speech from the mixture with a strange noise. We use a class of generative models called Dynamical Variational AutoEncoder (DVAE), which combines generative and temporal models to analyze the speech signal. This class of models makes attention to speech signal behavior, then extracts and enhances the speech. Moreover, we design a new architecture in the DVAE class named Bi-RVAE, which is more straightforward than the other models but gains good results. Experimental results show that DVAE class, including our proposed design, achieves a high-quality recovered speech. This class could enhance the speech signal before passing it into the central processing models.