How accurate is it?
Accurate enough to be a first pass you correct, not a finished chart. Kick and snare are reliable, hi-hat is good, toms are the weakest voice, and telling an open hi-hat from a ride in a dense chorus is unreliable — a distinction human transcribers argue about too.
On a holdout of ten songs with human-made charts the model never saw during training, it scores micro-F1 0.952 against 0.937 for ADTOF, the strongest published open model, on the same material. That benchmark runs on clean drum stems; on a dense real-world mix the figure is lower, around 0.92. Ten songs is a small sample, so read the lead as "no worse than the reference" rather than a decisive win.
What happens to the music I upload?
Your file and its results sit on a private link for seven days and are then deleted. Nothing is published, indexed or shared, and result pages are closed to search engines.
The one exception is explicit: if you press Fixes in the player, your corrections and that song are sent to us and used to improve the model. That only happens when you press the button.
Can I correct mistakes?
Yes, and it is the point rather than an afterthought. Every song opens as a step grid as well as notation — one row per drum, one column per sixteenth. Click a cell to add or remove a hit, and the synthesised drums play your corrected version so you can check by ear. Corrections stay in your browser until you choose to send them.
What formats can I upload and download?
Upload MP3 or WAV, up to 50 MB. Download MIDI for a DAW, MusicXML for MuseScore, Guitar Pro or Dorico, and a plain-text grid. The score and the grid are also playable in the browser without downloading anything.
Does it handle odd time signatures?
No. Everything is read in 4/4, so 5/4 or 7/8 confuses the bar lines. The hits themselves are still detected correctly — it is the bar grouping that goes wrong, which makes the score hard to read even when the notes are right. This is the limitation we would most like to remove, and hearing which songs broke it helps us prioritise.
What kind of recordings work best?
Anything with a real kit: studio, live, or a phone recording from a rehearsal room. Dense, heavily compressed mixes are harder, and so is music where cymbals wash over everything. Programmed electronic drums often work, but the labels can be odd — a machine-made sound need not resemble the acoustic drum it stands for.
Can I use a YouTube link instead of a file?
Not yet. Right now you upload an audio file.
Do I need to read music?
No. Every transcription also opens as a step grid, with the cursor running through it in time with the recording, and plenty of people use only that. If you would like to learn the notation, the guide to reading a drum stave covers the whole thing in about ten minutes.
What was the model trained on?
Two sources: STAR Drums, a synthetic dataset released under BSD-3, and the note timings of publicly available rhythm-game charts aligned to their audio. No charts or audio are redistributed — the training material stays on our machines and the weights are our own.
Training from scratch was not a preference. The strongest published open model is licensed CC BY-NC-SA, which permits non-commercial use only, so nothing could be built on it.
Does it work on a phone?
Yes. The score reflows to the screen width, and looping a bar works by pressing and holding it instead of shift-clicking. There is no app to install.
Still the best answer to most of these: look at a real transcription and judge for yourself.
Something not answered here? Write to support@drumanalyzer.com — questions that come up twice end up on this page.