Machine learning for audio restoration
Research on generative audio models for restoring archival recordings.
Overview
I develop and evaluate machine-learning methods for audio restoration. The work focuses on making archival sound more usable while preserving the characteristics that matter in a recording.
Focus
Audio restoration with generative models.Approach
Controlled transformations and expert listening.Context
Research engineering at EPFL VITA Laboratory.01
Research problem
Historical recordings rarely come with a clean reference. Noise, distortion, limited bandwidth, reverberation and compression can overlap, which makes restoration an underdetermined problem rather than a simple denoising task.
02
Approach
The work adapts generative music models to transform degraded audio into a clearer signal while retaining the original timing, interpretation and instrumental identity. Training relies on clean recordings paired with controlled synthetic degradations.
03
Evaluation
Automatic metrics are useful during training, but they do not fully describe musical quality or fidelity. Listening remains central for detecting artifacts, excessive smoothing and changes to timbre or ambience.
04
Long-term direction
The objective is an interactive workflow in which the system proposes restoration candidates and an expert compares them, adjusts the transformation and keeps control of the final result.