Neural Full-Rank Spatial Covariance Analysis for Blind Source Separation

This paper describes aneural blind source separation (BSS) method based on amortized variational inference (AVI) of a non-linear generative model mixture signals. A classical statistical approach to BSS is fit linear that consists spatial and models representing the inter-channel covariances power spectral densities sources, respectively. Although autoencoder (VAE) has successfully been used as with latent features, it should be pretrained from sufficient amount isolated Our method, in contrast, enables VAE-based trained only Specifically, we introduce neural mixture-to-feature directly infers features observed integrate feature-to-mixture consisting full-rank model. All are optimized jointly such likelihood for training mixtures maximized framework AVI. Once optimized, can estimating sources included unseen The experimental results show proposed outperformed state-of-the-art methods was comparable supervised learning sourcemodel.