Kyutai and Mirelo released MuScriptor, a 1.3B-parameter model that turns a finished song into separate MIDI notes for its instruments. The team trained it first on 1.45 million synthetic MIDI files, then on 170,000 real recordings spanning 11,000 hours, and finally on 300 hand-checked transcriptions. On the authors’ 372-track test set, the full model scored 60.4 for note starts, 73.3 for note duration, 49.0 for note endings, 50.2 for drums, and 48.2 for multi-instrument accuracy. These are F1 scores, which balance correct detections against missed and invented notes. The cited YourMT3+ baseline scored 32.52, 45.54, 17.79, 41.4, and 21.9 on the same measures.
A musician can use the command-line tool or local web interface to turn a mix into editable MIDI, inspect an arrangement, or isolate a practice part without uploading audio to a commercial service. The code is MIT-licensed, but the model weights require an account and use a non-commercial licence, so this is not an unrestricted production tool. It also misses dynamics, struggles with dense mixes and rare instruments, and cannot represent overlapping notes with the same pitch and instrument. When the authors kept those overlapping notes in the test, multi-instrument F1 fell from 48.2 to 42.0. No independent reproduction is available yet.
Read More: Stable Audio 3 generates music; MuScriptor tries to recover the notes.
Sources:
- MuScriptor paper
- MuScriptor code and local setup
- MuScriptor large model card and licence
- MuScriptor browser demo
Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service
Reuse
Citation
@misc{kabui2026,
author = {{Kabui, Charles}},
title = {MuScriptor {Turns} {Full} {Music} {Mixes} {Into} {Editable}
{MIDI}},
date = {2026-07-20},
url = {https://toknow.ai/posts/muscriptor-multi-instrument-music-transcription/},
langid = {en-GB}
}