Exploring Generative Models for (Audio) Source Separation

Abstract: This talk will delve into the application of generative models for source separation in the audio domain. By focusing on the use of autoregressive models and diffusion models, we will examine techniques for isolating individual components in complex compositions. We will first examine the use of VQ-VAE autoregressive models for weakly-supervised source separation through the Bayesian formulation in LASS.
Additionally, we will investigate the application of denoising diffusion models in constructing a diffusion-based generative model for music synthesis and source separation. This model is capable of performing separation and imputation on partial input tracks, making it a valuable tool for a wide range of applications.