AI-generated piano to MIDI
Music generators hand you a finished audio file. That is the point of them, and it is also the problem: you cannot change one chord, you cannot swap the piano, and you cannot print a note of it. Transcription gives the piece back to you as something editable.
Why you would want the MIDI at all
A generated track is a photograph of a performance that never happened. Once it exists, you are stuck with it. Turn it into MIDI and four things become possible that were not before:
- Re-voice it. Drive a real sampled piano — or a Rhodes, or strings — instead of whatever the generator's rendering sounded like.
- Edit the music itself. Change a chord, cut eight bars, transpose, fix the one note that was wrong.
- Take it to notation. A MIDI file is the entry point to a score. An MP3 is not.
- Keep the phrasing. Because velocity and the sustain pedal come across, you keep the performance shape instead of re-entering the notes flat by hand.
Where generated piano is easy
In some ways it is the friendliest material a transcriber ever sees. There is no room, no audience, no traffic outside the window, no microphone placed badly. The instrument sits in front of you at a constant level from first bar to last, and nothing else is playing over it. Compared to a phone recording of an upright in a kitchen, it is a clean read.
Where it gets hard, honestly
| Heavy compression | Generated audio is usually delivered loud and flat. Compression removes exactly the dynamic differences that velocity is measured from, so the transcription's dynamics are often narrower than a live recording of the same music would give. |
|---|---|
| Soft, felted textures | The muted, close-mic'd "felt piano" sound that generators produce so readily has a quiet attack by design. Quiet attacks are harder to locate precisely than bright ones, so onsets can drift. |
| Pedal that was never pressed | Nobody played this. The long ringing tails you hear may be reverb rather than a damper lifting. What comes back is a reading of the sustain you can hear, which is the honest answer, but it is a reconstruction of an effect rather than a record of a foot. |
| Rendering artefacts | Generators sometimes produce smeared or doubled notes that no piano could play. Transcription will faithfully write down what it hears, artefacts included. |
What to do before uploading
- Use the highest quality export the generator offers. WAV over MP3 if you have the choice. You are not gaining fidelity for listening, you are giving the transcription more to read.
- Use a solo piano track, not a full arrangement. If drums, pads or vocals are in the mix, the transcription tries to turn them into piano notes.
- Trim the fade-in and fade-out. Generated tracks often start and end on a ramp, and a ramp confuses the reading of how hard the first notes were struck.
- Do not normalise or compress it further. Every processing step that flattens the dynamics costs you velocity detail.
One thing that is not a technical question
Whether you may use, release or sell a piece a generator made for you is decided by that generator's terms and by your plan with them, not by the fact that you transcribed it. Transcribing does not create rights you did not already have. Check the terms of the service that produced the audio before you publish anything.
How to judge the result
Listen to the transcription against the original rather than looking at it. If the phrasing survives — the melody still sits above the accompaniment, the quiet bars are still quiet — the velocities came through and the file is worth editing. If it sounds mechanical, the source was probably compressed hard, and there is nothing any transcriber can do to recover what was flattened before it arrived. The velocity guide explains what to look for.
Transcribe a recording