← All projects
EPFL / VITA Laboratory2025 — now

Machine learning for audio restoration

Research engineering in generative modelling and evaluation for audio restoration.

Machine learning for audio restoration

At EPFL’s VITA Laboratory, I work on machine-learning methods for restoring degraded music recordings. The research combines generative modelling, controlled degradation experiments and evaluation of the restored signal.

View OpenRestor on GitHub ↗

Role

Research engineer

Laboratory

VITA Laboratory, EPFL

Research focus

Generative audio restoration

Related project

OpenRestor

01

Research problem

Archival recordings can combine noise, distortion, restricted bandwidth, reverberation and compression. A clean reference is often unavailable. Restoration therefore involves estimating missing or damaged information while limiting changes to the performance itself.

02

Method development

My work investigates generative models for transforming degraded music recordings. Controlled degradations applied to clean recordings provide paired examples for development and comparison. A central question is how to improve clarity while retaining timing, timbre and interpretation.

03

Evaluation

Evaluation considers reconstruction quality, perceptual measures and listening. Numerical scores support controlled comparisons; listening helps identify artifacts and changes to musical detail that an aggregate metric may not capture.

04

OpenRestor

I also work on OpenRestor, which provides shared data preparation, degradation and scoring tools. Its dedicated project page describes the benchmark and links to the public repository.

05

Research direction

An ongoing objective is to support expert assessment of restoration candidates. This includes examining how model behaviour can be compared and controlled, rather than treating a generated output as an automatically accepted final recording.