Generative Models for Inverse Problems in Audio

Inverse problems deal with estimating an unobserved signal from degraded observations produced by a forward process. They are inherently difficult because the inverse mapping is often ill-posed: solutions may be non-unique, unstable, or not exist at all. In audio, examples include stem separation, audio enhancement, denoising, and timbre transfer.

Recent approaches tackle these challenges by learning data-driven priors with generative models. In particular, diffusion models and flow matching offer a principled framework for modelling complex audio distributions and for solving inverse problems through iterative or continuous transformations.

Available as a Bachelor’s or Master’s thesis, research internship, or project.

Prerequisites

• Solid knowledge of (statistical) signal processing
• Basic knowledge of machine learning, in particular deep learning and generative methods
• Interest in audio and music applications

Contact

Cristóbal Andrade (cristobal.andrade@fau.de), Prof. Sebastian J. Schlecht. Also listed at LMS.