MP3 to MIDI: piano transcription

The result isn't a raw list of detected notes but a cleaned-up piano arrangement in a standard MIDI file:

• a Right hand track: the melody (the top voice) plus one to three chord tones within an octave;
• a Left hand track: the bass line, following harmony changes or the original accompaniment's rhythm;
• a control track with the detected tempo (BPM), time signature (4/4 or 3/4) and key.

Notes are aligned to the beat, so the file opens as readable notation in a score editor and plays back in time. The page shows a piano roll (blue right hand, green left hand), lets you play both hands or each hand alone, and plays the original recording for comparison.

The recognizer is trained on piano sound, so the best results come from:

• solo piano: your own playing, a concert recording, a piano cover;
• an isolated piano track (stem) from a song;
• clean audio with little noise, reverb or distortion. A digital piano or keyboard with a piano sound works well too.

In a full band or vocal song, drums, voice and guitar can sneak extra notes into the result and the bass gets less reliable. Separate the piano first with a stem-separation tool, then transcribe that. Other instruments (guitar, violin, voice) are not transcribed reliably.

1. Note recognition. A neural network (the open-source high-resolution piano transcription model by Kong et al., released by ByteDance under the CC BY 4.0 license, adapted to run in the browser) finds which keys were struck, when, how long and how hard.
2. Cleanup. Very quiet, very short or uncertain notes, such as overtones or accidental brushes, are dropped.
3. Tempo and bars. Beat, a sixteenth-note grid and the downbeats are worked out from the note onsets.
4. Arrangement. Melody and chord tones go to the right hand, bass notes to the left, kept within a comfortable hand span.

The model (21.7 MB) downloads the first time and is cached by your browser. Recognition runs on your own device and takes roughly as long as the section itself on a typical laptop, longer on a phone. Your audio is never uploaded.

The arrangement updates instantly when you change a setting; there's no need to recognize again.

• Density: Simple (melody plus one chord tone, good for beginners), Normal (melody plus two), Rich (up to three chord tones).
• Tempo: automatic tempo detection sometimes lands on half or double the real tempo (a 70 BPM ballad read as 140). Use Half or Double, or set the BPM manually.
• Time signature: automatic, or force 4/4 or 3/4 (waltzes, minuets).
• Octave bass: doubles the bass in octaves where the original does; off, the left hand plays single bass notes.

MIDI isn't audio but a list of instructions: which note, when, how long, how loud. That makes it easy to edit and play with any sound.

• Score editors such as MuseScore show two staves, treble and bass clef, with tempo and time signature.
• Piano learning apps with falling notes treat the two tracks as separate hands, so you can practice hands separately.
• DAWs (GarageBand, FL Studio, Logic, Ableton Live) let you edit each track or swap the instrument.

Transcribing commercial recordings for personal study is common; sharing or selling transcriptions of copyrighted music may need permission.