Kyutai's MuScriptor Now Turns Any Song Into Printable Sheet Music

Kyutai's open-source music transcription model just picked up a feature that turns it from a MIDI-only tool into something a musician might actually print out and read. MuScriptor, the multi-instrument transcription model built with Mirelo AI, can now generate sheet music PDFs and guitar tabs directly from an audio file, in addition to editable MusicXML.

Under the hood, the pipeline still produces MIDI first. The team then converts that MIDI into engraved notation using MuseScore with some custom preprocessing to clean up the output. That means the same transcription that gives you a piano roll can now hand you a printable score or a fretboard-aware tab for the guitar part.

What the update actually adds

The transcription core is unchanged. What is new is the export path on top of it:

• Sheet music PDFs rendered from the transcribed MIDI
• Editable MusicXML you can open in any notation editor
• Guitar tabs with standard string and fret notation
• The existing MIDI output for producers and DAW workflows

You can try it in the browser at muscriptor.kyutai.org, or self-host from the GitHub repo. Local install is a one-liner with uv add muscriptor or pip install muscriptor, and the web UI runs with uvx muscriptor serve.

Where it shines and where it breaks

The team is upfront about the failure mode. Sheet music generation works best when the underlying music has a steady rhythm played to a metronome, which covers most rock, pop, and electronic tracks. Classical pieces played in rubato, where tempo stretches and contracts expressively, can produce what the authors describe as terrifying results. That is a MIDI-to-notation quantization problem more than a transcription problem, but the effect on the output is the same.

The transcription model itself is the piece doing the heavy lifting. The paper describes MuScriptor as a decoder-only Transformer that reads a mel-spectrogram of a short audio segment, then autoregressively predicts MIDI-like tokens for pitch, timing, and instrument, following the MT3 tokenization scheme. Transcription becomes a language modeling task, with notes standing in for words.

Why the base model matters here

Sheet music is only as good as the notes underneath it, and MuScriptor's numbers on multi-instrument mixes are what make this feature useful rather than novelty. Training combined 1.45 million synthetic MIDI files with roughly 170,000 real recordings totalling more than 11,000 hours, then reinforcement learning post-training. The 1.4B large variant reportedly hits Multi F1 of 48.2 versus 21.9 for the YourMT3+ baseline. Fine-tuning on real audio made a large difference: training on real data improved transcription metrics by roughly 20 percentage points over a version pre-trained on synthetic audio alone.

Three model sizes are published on Hugging Face, and you can pass the size keyword directly to the CLI or Python API:

You can also condition on which instruments are present. The model supports around 36 instruments including piano, guitar, drums, and bass, and you can pass a list like ["acoustic_piano", "drums"] to keep the output focused.

Practical use cases

The sheet music addition changes who the tool is for. MIDI is useful in a DAW, but a PDF or a tab is what you hand to a bandmate learning a cover, or what a student uses to practice a solo by ear. The concrete workflows that open up:

1. Extract a printable lead sheet from a rehearsal recording
2. Generate guitar tabs from a track without an existing transcription online
3. Produce MusicXML to import into Sibelius, Dorico, or MuseScore for cleanup
4. Get per-instrument parts from a full band mix in one pass

The licensing catch

The inference code is MIT licensed, but the weights are not commercially usable off the shelf. Weights ship under CC BY-NC 4.0, so anyone hoping to drop MuScriptor into a paid product will need a different arrangement, since the inference code is MIT but the model itself is not commercially usable off the shelf. For personal projects, research, and internal tooling, that is fine. For a commercial transcription SaaS, it means a conversation with Kyutai before shipping.