Restoration can make historical audio sound cleaner.
Did it stay closer to the original, or drift from it?
A research benchmark for ML restoration of degraded historical audio, built around one question: signal-fidelity metrics and content preservation don’t automatically move together. This page reports what a first real test found — every number below is generated from a checked-in result file, never hand-typed.
Headline finding — STANDARD tier, speaker-held-out
The discriminative ML denoiser was trained on 6 speakers and evaluated on 6 it had never heard. Checked per speaker, at the two most severe degradation tiers:
of held-out speaker×tier cases: ML denoiser beats classical Wiener filtering on SI-SDR (signal fidelity)
of those same cases: the ML denoiser also increases word-error rate more than Wiener does — it’s less faithful to the words actually said
N = 12 held-out speaker×tier combinations. Real and consistent across every held-out speaker tried, but too small a sample for a bootstrapped effect size — see research/claims_registry.md row C6.
Signal fidelity by tier (SI-SDR, dB, higher is better)
| Tier | Passthrough | Spectral gate | Wiener | Median declick | ML denoiser |
|---|---|---|---|---|---|
| LIGHT | 12.3 | 12.2 | 12.3 | 11.5 | 12.2 |
| MODERATE | 6.6 | 6.7 | 6.7 | 6.4 | 7.0 |
| SEVERE | 2.1 | 2.7 | 2.4 | 2.0 | 3.4 |
| EXTREME | -4.2 | -2.8 | -3.6 | -4.2 | 0.1 |
Held-out eval_pool speakers only (n = 24 clips). Source: results/standard/benchmark.parquet.
What this site does not yet cover
- No generative-restoration method exists yet — only classical baselines and one discriminative ML denoiser.
- No human listening-study data exists anywhere in this project.
- No fidelity-vs-quality frontier, figures, or paper yet.
Full status: docs/HANDOFF.md. Source: github.com/Gariyuuu/soundrevive.