What actually happens to the file
Upload an audio file and three things run in order. A separation model pulls the drums out of the mix, because a transcription model listening to a full arrangement mistakes guitars for cymbals. A detection model then finds every hit and labels it as kick, snare, hi-hat, tom or cymbal. Finally the hits are placed on a beat grid derived from the recording itself, so the timing follows the drummer rather than a fixed metronome.
You get the result as a page you can play, plus drums.mid to download. A song takes a minute or two.
What is in the MIDI file
One track of percussion on channel 10, using the General MIDI drum map, so it will trigger sensible sounds in any DAW or sampler without remapping:
- Kick — note 36, acoustic bass drum
- Snare — note 38, acoustic snare
- Hi-hat — note 42, closed hi-hat
- Tom — note 47, mid tom
- Cymbal — note 49, crash
Velocity carries information. It is not a constant — each note's velocity is the model's confidence at that hit, which correlates with how hard the drum was struck. Quiet ghost notes arrive quiet. If you want a flat programmed feel, select all and normalise; if you want the human dynamics, leave it alone.
The tempo map comes from the recording, so a track that breathes stays in sync when you play it against the original audio.
This is a draft, not a capture. Machine transcription is wrong somewhere in every song. Kick and snare are reliable, hi-hat is good, toms are the weakest voice, and an open hi-hat in a dense chorus is often written as a ride or the other way round. Hits are placed on a sixteenth-note grid, so a swung or triplet feel is squared off.
Budget five minutes in your editor. That is still far less than programming the part from scratch, which is the honest case for doing it this way.
Once it is in your DAW
A few things worth doing in the first minutes, in the order that saves the most time:
- Check the tempo before anything else. If the grid looks half or double what you expected, fix that first — every later edit depends on it.
- Listen to the chorus, not the verse. Dense sections are where a transcription slips, and where you will find the cymbal that should have been a hi-hat.
- Split the cymbal lane by hand. We write one crash note; if the part rides through a section, move those hits to your ride.
- Decide about ghost notes. Some will be missing and some will be invented. For a programmed part they rarely matter; for a realistic one, add them back by feel.
If you also want it written down
The same transcription exports as MusicXML, which opens in MuseScore, Guitar Pro or Dorico with proper drum voicing — hands with up-stems, kick with down-stems, x-noteheads on hi-hat and cymbals — so you can keep editing it as notation rather than as a piano roll. If you are not sure how to read what comes out, the guide to drum notation covers the whole stave in about ten minutes.
Try it on something. The example below is a full song with the score, the grid and the downloads in place, so you can look at the output before uploading anything of your own.
Or upload a track — .mp3 or .wav, up to 50 MB. Your file stays on a private link and is never published anywhere.