Unsupervised Speech Enhancement Using Dynamical Variational Autoencoders

Dynamical variational autoencoders (DVAEs) are a class of deep generative models with latent variables, dedicated to model time series high-dimensional data. DVAEs can be considered as extensions the autoencoder (VAE) that include temporal dependencies between successive observed and/or vectors. Previous work has shown interest using over VAE for speech spectrograms modeling. Independently, been successfully applied enhancement in noise, an unsupervised noise-agnostic set-up requires neither noise samples nor noisy at training time, but only clean signals. In this paper, we extend these works DVAE-based single-channel enhancement, hence exploiting both signals representation learning and dynamics We propose algorithm combines DVAE prior pre-trained on based nonnegative matrix factorization, derive expectation-maximization (VEM) perform enhancement. The is presented most general formulation then three specific illustrate versatility framework. Experimental results show proposed approach outperforms its VAE-based counterpart, well several supervised noise-dependent baselines, especially when type unseen during training.