Transcribing the drum part is the easy half

Every drummer knows the evening. You love a song, you want to play it, and you spend three hours with the record instead of the kit: four bars, rewind, four bars again, arguing with yourself about whether that is an open hi-hat or a ride. By the time the part is written down you are too tired to practise it.

So the obvious wish is: feed the song to a computer, get the notes back. Several tools do this now, with real machine learning behind them, and they work reasonably well.

And then they hand you a PDF.

That is the moment I kept getting stuck on. The file is correct, and it is useless in the way a recipe is useless when you are hungry. You still open it next to your kit, find your place, start the record, pause it, count bars in your head, rewind again.

The transcription is not the product. Playing the song is. That is the whole idea behind what I have been building, and everything else follows from it.

A score you play, not a score you read

The transcription opens as a page you use while drumming, with the record playing inside it. A cursor runs through the notes in time with the music. That single change removes most of the friction: you never lose your place, and you never guess which bar you are in.

Repeated bars are folded — ×19 instead of nineteen identical lines — so the whole song fits on one screen and you can see its shape before playing a note: intro, verse, the bar where the fill lands, the chorus that comes back three times. The seek bar is the same idea taken to its end. It is not a generic progress bar but a map of the song, one block per bar, coloured by pattern. You do not scrub to somewhere in the middle; you jump to the second chorus because you can see it. Hold Shift and click, and that section loops until you own it.

The same song at your level

Here is the feature I use most, and the one no PDF can have.

A drum part is not one thing. There is the groove that carries the song, and there is everything the drummer added on top: ghost notes, sixteenth-note kick syncopation, ride bell accents. A beginner who tries to play the full transcription of a song they love usually fails at all of it at once, and concludes they are not ready for the song.

They are ready. They are just not ready for that version of it.

The slider rewrites the part for your level, from as-recorded down to a bare pulse. So a beginner plays the simple version against the real song today, and adds detail as their hands catch up, without ever switching to some other, easier song.

That is what a teacher does in a lesson. It is strange that transcription tools do not.

It does not have to be perfect, because you can fix it

Every automatic transcription is wrong somewhere. Mine gets kick and snare almost always right, hi-hat usually, and argues with itself about cymbals and toms — the same places where two human transcribers of the same song disagree with each other.

Most tools deal with this by not mentioning it. I would rather deal with it by letting you correct what you hear. Every song also opens as a grid — one row per drum — and a wrong hit is one click from being right. The synthesised drums play your corrected version, so you can check by ear immediately.

When you send those corrections back, they become training material. Which flips the usual relationship: the tool does not have to be perfect on day one, it has to be close and honest, and it gets better because people use it. Every drummer who fixes a bar teaches it something a synthetic dataset never could.

The metronome breathes

One small thing that gives away whether a practice tool was made by someone who plays: the click.

A metronome set to the average BPM of a recording drifts away from it within a minute, because human drummers speed up in choruses and lean back in verses. Ours follows the record's own pulse — it comes from the same beat grid the notation uses — so it stays with the music all the way through, and the accented beat always lands on the bar line you are reading.

What it cannot do

Because the fastest way to lose a drummer's trust is to oversell:

• In a dense chorus it cannot reliably tell an open hi-hat from a ride. Neither can two human transcribers of the same song — I checked, and Songsterr and a professional chart of one song disagree on exactly this.
• Toms are the weakest of the five voices.
• Ghost notes and fast kick syncopation come and go.
• It reads everything in 4/4. Odd time will confuse the bar lines.

For the record, on a held-out set of ten songs with human-made charts that the model never saw during training, it scores slightly above the strongest published open model. But I would rather you judged it on a song you know cold than on my number — the about page has the full table if you want it.