About a minute per track, locally
A four-minute song is usually done in 30–60 seconds. Nothing queues behind other people, because the work happens on your CPU, not a shared server.
Runs in your browser
Drop in a recording and get back an editable .mid file — every note, chord and rhythm pulled out of the audio. No install, no signup, and nothing uploaded: the note detection runs on your own machine.
Drop a track in — note detection starts when you press Generate
Drag and drop your audio file here
or
MP3, WAV, OGG, FLAC and M4A, up to 50 MB
Choose an audio file to enable conversion.
Preview and download your .mid
No MIDI file yet
Pick a track on the left, then generate. The .mid file lands here.
What is MP3 to MIDI conversion?
Converting MP3 to MIDI means analysing a performance and writing down what was played as data instead of sound. The output is a .mid file listing each note: its pitch, when it starts, how long it holds, and how hard it was struck. MIDI is the common language every DAW and notation program reads, so once the notes exist you can rewrite them freely.
The hard part has always been the listening. Older converters struggled the moment more than one instrument played at once, which is most music. This tool uses Basic Pitch, a polyphonic transcription model trained on a broad set of multi-instrument recordings, and it runs it locally in your browser rather than on a queue you have to wait in. After conversion, open the MIDI editor to fix timing and velocity before you export.
A four-minute song is usually done in 30–60 seconds. Nothing queues behind other people, because the work happens on your CPU, not a shared server.
Each note carries the four things a DAW needs: which note it is, when it starts, how long it lasts, and how hard it was played.
Polyphonic material arrives as stacked notes rather than a single melodic line, so sustained pads and voiced chords survive the trip.
Detected timing is written into the file so the project lines up when you drop it into a session at the tempo you set.
There is no upload step. The file is decoded and analysed in the tab you already have open, which is what makes unreleased material safe to run through it.
A plain MIDI file, not a proprietary project format. Anything that has opened a .mid since 1983 will open this one.
Accuracy
There is a riff you half know. Playing it back at 0.75× and hunting for the right key is one way to spend an evening. Running it through a converter and reading the result off a piano roll is a different way to spend four minutes.
Music teachers use transcription this way in lessons: pull the notes out of a recording, hand the student the score, talk about what the player actually did rather than what everyone remembers hearing.
Where it helps most
For music production
Typical uses
Most production work starts from something you already heard. Getting that something into your session as editable notes beats replaying a part by ear, and it keeps whatever the original player did with timing — which is usually the part worth keeping.
Because the output is plain MIDI, the conversion is a starting point rather than a finished thing. Edit timing and velocity in the browser, then export and drop the .mid into your DAW — quantise it, thin out the doubled notes, keep three of them and write your own fourth.
How it works
MP3, WAV, OGG, FLAC or M4A, up to 50 MB. Uncompressed formats tend to give the detector slightly more to work with.
The model walks the waveform a couple of seconds at a time, deciding frame by frame which of the 88 piano keys are sounding.
Check the preview, then open the notes in the MIDI editor and fix what needs fixing — move a stray note, re-time a chord, drop the ones the detector imagined.
Export from the editor and drop the .mid into your session. Rearranging and re-instrumenting happen there — the file is just notes.
FAQ
Your file is decoded to a mono signal at 22,050 Hz, then handed to Basic Pitch — a neural network released by Spotify under the Apache-2.0 licence that predicts, for every ~11 ms frame, which pitches are sounding, where notes begin, and how the pitch bends. Those predictions are cleaned up into discrete note events and written into a MIDI file.
On a single clean instrument — piano, guitar, bass, a solo vocal line — the result is usually very close to playable and needs light tidying. On a dense full-band mix, expect to do real editing work: the model only hears one mono signal, so a kick drum and a bass note an octave apart can fight over the same slot. Reverb-heavy or heavily compressed masters are the hardest case.
It is free, with no account and no credit counter. There is no catch because there is no server doing the work — the model runs in your browser, so a thousand conversions cost the same as none. The trade-off is that it uses your CPU and battery, and older laptops will take longer.
MP3, WAV, OGG, FLAC and M4A, up to 50 MB each, and as many files as you like. WAV and FLAC often transcribe slightly better than MP3 because the attack transients the detector depends on have not been smeared by compression.
No, and that is the point. MIDI carries notes, not audio. You get the pitches and the rhythm of the performance, then decide in your DAW whether that line becomes a piano, a string section or a synth. This is exactly why a guitar take can be replayed as anything else.
No. There is no upload endpoint in this app — nothing is sent anywhere. The model weights live on this site rather than a CDN, and they are fetched by the tab only when you press Generate, so the decoding and the inference both happen on your machine and the audio never leaves it.
Because a stereo mix is folded down to a single mono signal before analysis. Instruments that share a pitch range compete for the same note slot, and the model has no way to tell a snare from a piano stab. Splitting the track into stems first — even roughly — and converting each stem on its own gives markedly better results.