An open benchmark, including the part that goes against us
Every transcription service says it is accurate. Almost none of them publish a number you can check. Here is what we measured, how we measured it, and the one result that suggests Damper has a bug.
What this measures, and what it does not
Read this part before the table, because it decides what the numbers are worth.
These figures describe what is inside the delivered MIDI file: how many notes it contains, how many sustain-pedal events, how varied the velocities are. Two people opening the same file get the same numbers. That much is fact.
None of it measures accuracy. To say a transcription is more faithful you need a ground truth — the same performance captured simultaneously as audio and as MIDI, from an instrument that records both. We do not have that material, so we do not make that claim, and you should be sceptical of anyone who makes it without showing you their reference set.
More notes is not better; it may mean invented notes. More pedal events is not better; it may mean pedal that was never pressed. That last point turns out to matter below.
Method
- Six solo-piano recordings from public archives, chosen before any result was seen, covering slow and fast, sparse and dense, modern studio sound and a noisy 78 rpm transfer.
- The first 30 seconds of each, cut once with the same command, so both systems received a byte-identical file.
- Damper's output and Eldoraudio's free 30-second preview, both taken as delivered, with no hand correction.
- Measured on 18 August 2026.
The numbers
| Recording | Notes Damper / Eldoraudio | Pedal events Damper / Eldoraudio | Distinct velocities Damper / Eldoraudio |
|---|---|---|---|
| Bach, Aria from the Goldberg Variations | 82 / 81 | 28 / 12 | 36 / 33 |
| Bach, Variatio 4 | 183 / 187 | 54 / 0 | 36 / 43 |
| Chopin, Nocturne Op. 48 No. 1 | 84 / 82 | 22 / 24 | 45 / 38 |
| Chopin, Scherzo Op. 39 | 356 / 328 | 58 / 48 | 66 / 64 |
| Satie, Gymnopédie No. 1 | 58 / 58 | 26 / 24 | 32 / 29 |
| Beethoven, Moonlight (78 rpm transfer) | 122 / 125 | 26 / 30 | 40 / 48 |
| Total | 885 / 861 | 214 / 138 | — |
What we read into it
On note counts the two are close enough to call level. 885 against 861 across six pieces, with each ahead on three. Nothing here supports either service claiming to find more of the music.
On velocity both produce a genuine spread. Dozens of distinct values on every piece, and neither is consistently wider. Whatever else is true, neither is writing flat, invented dynamics.
On the pedal there is one striking difference, and it does not flatter us. On Bach's Variatio 4 Damper wrote 54 pedal events and Eldoraudio wrote none. That looks like a win until you consider the piece: a fast contrapuntal variation, the kind of Bach a pianist plays with almost no pedal at all. Damper produced a pedal change roughly every 1.1 seconds there. We think that is more likely to be our error than their omission, and we are investigating it rather than presenting it as an advantage.
What would make this benchmark actually conclusive
Recordings from a reproducing piano, where the same performance exists as both audio and MIDI. Feed the audio to each transcriber, compare each result against the instrument's own MIDI, and you get real precision, recall and onset error instead of counts. Two such datasets exist publicly, and getting them in a usable form is the next step.
Until then this page says what it can support and no more: on the properties we can measure, the two systems are comparable. If you want a reason to choose Damper, it is not that we found a hidden note the other one missed — it is that you can hear your own recording transcribed in full before paying anything, and you pay for one file rather than a subscription.
Corrections welcome
If you can reproduce this and get different numbers, or you have ground-truth material that would sharpen it, write to the address shown on your Stripe receipt. A benchmark nobody can challenge is marketing, not measurement.
Transcribe a recording