OpenRestor
A reproducible benchmark for musical audio restoration.
Overview
My work on OpenRestor at EPFL concerns the experimental infrastructure for comparing music-restoration systems: preparing data, constructing degraded inputs and evaluating restored outputs.
Explore the code and documentation ↗Context
Research at EPFL VITA Laboratory.Purpose
Comparable restoration experiments.Source code
Public Python toolkit on GitHub.01
Experimental objective
Comparable restoration experiments require explicit data, degradation conditions and scoring protocols. OpenRestor brings those stages into a reproducible workflow.
02
Dataset preparation
The pipeline audits, indexes, segments and verifies recordings. SonicMaster supplies the main training data; SDD and MUSDB18-HQ provide separate transfer-evaluation material.
03
Controlled degradation
The degradation registry defines 24 canonical classes. Configurations and manifests record how degraded examples are generated from clean audio.
04
Scoring and validation
Validation checks restored outputs against a submission contract. Scoring tools compute reconstruction and perceptual measures under shared conditions.
05
Research output
The public repository contains the implementation, configurations and workflow documentation. These artifacts make experimental conditions inspectable and repeatable; restoration-model results should be assessed separately from the benchmark infrastructure.