Skip to content
Euphona

Measured 22 September 2026

What we measured

Restoration results on a public-domain recording, with the method written out so you can run it yourself. Two of the four results argue against us and are here for the same reason the others are.

Why these numbers can be checked

Restoration is the one capability that needs no human to compare against. Take a clean file, inflict a defect you control, repair it, and measure how far the result sits from the original. The right answer is exact, so the test is falsifiable rather than a matter of taste.

The second number is the one an engineer actually asks for. The same chain is run on undamaged audio, to measure what it does to a file that was already fine. A denoiser that recovers 8 dB and takes 3 dB out of a clean take is not a denoiser worth having, and only measuring both reveals it.

Source
Chopin, Mazurka Op. 7 no. 4 — Musopen, released under CC0 1.0, via the Internet Archive. Public domain recording of a public-domain composition, so both rights are clear and you may use the same file.
Excerpt
22 seconds, 44.1 kHz stereo, set to 0.7 peak for headroom.
Harness
The same defect functions and signal-to-error metric the engine is held to in its own quality suite, pointed at a file whose licence can be stated.
The results
Restoration measurements by stage: recovery, share of removed material that was music, and effect on undamaged audio.
StageRecoveryRemoved that was musicEffect on clean audio
Hum removal+10.31 dB0.0%−111.1 dB
Click repair+7.70 dB0.0%−48.1 dB
Noise reduction+1.86 dB2.9%−40.3 dB
Declipping+0.50 dB15.9%−49.9 dB

Recovery above zero means the repair removed more than it added. The last column is how far the chain moves audio that had nothing wrong with it — more negative is better, and −111 dB is silence.

Stage by stage, including the weak ones

Hum removal

95% of what it removed sits within 2 Hz of 50, 100, 150 and 200 Hz — mains and its harmonics, and no music. Run on a clean file it removes nothing measurable: it looks for mains, finds none, and places no notch.

Click repair

92% of what it removed was the injected impulses, with no programme in the residue. It does flag transients on clean piano, and alters such a file by 48 dB down — small, but not nothing.

Noise reduction

The modest recovery is deliberate: a harder setting measured worse than the noisy file against ground truth, because it removed more signal than noise. This is also the stage that disturbs clean audio most.

Declipping

It removed about as much music as clipping, for very little gain. Solo piano is close to the worst case — only 446 samples ever reached the rails — but on this material the repair is not discriminating.
What this does not claim

Nothing here compares Euphona to a human engineer. Every figure is a distance from a known-clean file, not a judgement against someone’s work, and we do not publish a parity or superiority claim because we have not run the listening panels that would support one.

Nor is one recording a general result. These numbers describe solo piano at 44.1 kHz with defects we injected. Different material moves them — declipping in particular is close to its worst case here, because piano reaches the rails for only a few hundred samples.

The honest summary is narrower than a headline: on this material, hum removal and click repair took the defect and left the music, noise reduction traded a little of both, and declipping did not earn its place.
Run it yourself
  1. 01

    Take the same file

    Download the Musopen Chopin collection from the Internet Archive. It is CC0, so nothing about re-using it is in question.

  2. 02

    Break it in a way you control

    Add hum, clicks, hiss or clipping to a copy. Keep the original — it is the reference that makes the answer exact.

  3. 03

    Repair it

    Run the damaged copy through restoration. The free tier gives ten minutes of the full engine and needs no card.

  4. 04

    Null the result

    Subtract the repair from the damaged file. What is left is exactly what was removed — listen for music in it, which is the test that matters.