Watch your notes play.
Tavim writes your playing out and lets you watch it: the notes roll by as they sound, and the score lights up note by note, ready to share as a video.
See it on real pieces in the library. Read more: A piano with no samples and You recorded it on your phone. Here is how to hear it properly..
How it works
Drop it in.
A phone video, a voice memo, any file. Sign in once and Tavim gets going.
Tavim listens. Hard.
It hears every part, writes down every note, and irons out the wobbly bits. Your timing stays.
Press play. Grin.
Flip between your take and Tavim’s. Tweak anything by typing what you want.
99.4% of the notes Tavim writes are right.
Held-out real performances, whole pieces, scored against a Disklavier’s own record of what was played. A note counts when the key is right and it lands within 50 ms. Precision 99.4, recall 98.8, F1 99.1.
What Tavim does
- A piano video, written out. Paste a YouTube link: Tavim hears the piano, writes it as a score you can read and edit, and plays it back as it was played.
- Your piano recording, as MIDI. Drop a recording, even from a phone in a noisy room: Tavim writes the notes and gives you MIDI for your DAW, or aligned to the recording, beside the score.
- Watch your notes play. Tavim writes your playing out and lets you watch it: the notes roll by as they sound, and the score lights up note by note, ready to share as a video.
Questions
Is it truly the best transcriber?
We think so, and we would rather show you than tell you. On the standard test, the latest published research system scores 97.3. Tavim scores 99.1 on the same kind of recordings, and 99.4% of the notes it writes are right: the right key, at the right moment. Those are real piano performances it has never heard before, played start to finish.
How we got there: two listeners instead of one. Our own model, trained on a big mix of real performances (studio, concert and phone), does the main job. A second, independent transcriber listens to the same take. A note only the second one heard is added only if our own model also shows a spark of it at that spot. Two independent ears agreeing is the whole trick. It wins back about 4 in 10 of the quiet notes the main model missed, and about 95 of every 100 notes it adds are real.
It never changes a note the main model heard, and every note it adds is tagged, so it can be checked. We are not claiming perfection. We are claiming we measure it hard, and keep going after the last 0.6%.
How do you even measure “best”?
Every note is a tiny test with two questions: is it the right key, and did it start at the right moment? We compare Tavim’s notes with the true ones, recorded by sensors under the keys of a Disklavier piano, so nobody is guessing what was played.
Humans do not land on the millisecond, and neither do microphones, so the test allows 50 milliseconds either way, a twentieth of a second. Inside that window, each true note is paired with at most one of ours. That is what keeps polyphony honest: a six-note chord is six separate questions, so dropping one voice, doubling one, or smearing two into one all cost points.
Then two numbers. Precision: of the notes we wrote, how many were real. Recall: of the real notes, how many we found. The score everyone quotes, F1, blends them, so you cannot buy a high mark by writing everything, or by writing only the safe notes. On top of that, beats get 70 ms, and note endings get 50 ms or 20% of the note’s length, whichever is larger.
We run it on whole pieces, not hand-picked excerpts, on performances the model never trained on, with the same rules the research papers use so the numbers can sit next to theirs. And before a number goes public, we re-score a known case and check it still matches. If it does not, we assume the scorer is wrong, not the music.
Will it make me sound like someone else?
No. Same player, same timing: Tavim plays back what you played. A pianist’s style is there only if you choose one, and you can flip back to your original any time.
My recording is terrible.
Phone, video or audio, any format, or even a score or MIDI file. Terrible is a perfectly good place to start.
Do I have to sign up first?
Drop the file first if you like. Tavim gets going once you’re in, with Google, Apple or an email link. No password.
Do you train your AI on my playing?
Only if you leave the box ticked, and you can turn it off in Account whenever you like. Tavim works the same either way.
Piano only?
Yes, piano, on purpose. Tavim separates everything in the take, voice and room included, and lets you change the piano, the voice or the instrument.