MuScriptor
muscriptor.kyutai.org
What it is
MuScriptor turns audio recordings into MIDI and sheet music, with downloadable MIDI files, sheet-music PDFs, and guitar-tab PDFs from its hosted service. The web demo accepts MP3, WAV, FLAC, OGG, M4A, AIFF, and Opus files. It can transcribe voice, piano, guitars, bass, strings, brass, woodwinds, drums, and other instruments, and users can specify which instruments they expect to improve transcription accuracy. Guitar and bass parts can include tablature alongside standard notation, but the system does not produce chord symbols. MusicXML output can be edited in MuseScore, Sibelius, or Guitar Pro. Users can self-host the web interface, and transcription is also available through a CLI and Python API. Local use requires Hugging Face authentication and acceptance of the model's CC BY-NC 4.0 license. Hosted audio is processed on the server and not retained afterwards; local use keeps audio on the user's machine. The service says it performs better on acoustic and guitar music than electronic music, and that results are not perfect. The hosted demo is free; the code is MIT-licensed, while model weights carry the non-commercial license.
Spec sheet
- What it does
- MuScriptor converts recordings into multi-instrument MIDI and printable sheet music.Read 14 Oct 2026 · muscriptor.kyutai.org
- Purpose
- MuScriptor transcribes recordings into MIDI and sheet music.Read 30 Sep 2026 · github.com
- Platforms
- The repository documents Linux, macOS with Apple Silicon, Windows GPU use and Intel macOS support.Read 30 Sep 2026 · github.com
- Integrations
- MusicXML output can be edited in MuseScore, Sibelius or Guitar Pro.Read 30 Sep 2026 · muscriptor.kyutai.org
- Support
- Bug reports and feedback can be sent to [email protected] or submitted as GitHub issues.Read 30 Sep 2026 · muscriptor.kyutai.org
- Headquarters
- Paris, FranceRead 28 Sep 2026 · muscriptor.kyutai.org
- Accuracy limitation
- The service says it works better on acoustic and guitar music than electronic music and is not perfect.Read 30 Sep 2026 · muscriptor.kyutai.org
- API detail
- The Python API supports instrument filtering, sampling, temperature, beam search, batching and prelude forcing.Read 14 Oct 2026 · pypi.org
- Audio formats
- The web demo accepts MP3, WAV, FLAC, OGG, M4A, AIFF and Opus files.Read 30 Sep 2026 · muscriptor.kyutai.org
- Browser interface
- The hosted site runs as a JavaScript web transcriber with no installation described on its home page.Read 14 Oct 2026 · muscriptor.kyutai.org
- CLI
- The package provides a command-line transcribe command that writes MIDI, JSON, JSONL or sheet outputs.Read 14 Oct 2026 · pypi.org
- Code licence
- The inference and application code is released under the MIT licence.Read 14 Oct 2026 · github.com
- Crunchbase facts
- Crunchbase lists Kyutai as an active Paris nonprofit with 11–50 employees and founders Laurent Mazare and Neil Zeghidour.Read 14 Oct 2026 · crunchbase.com
- Documentation depth
- Documentation includes installation, model selection, CLI examples, Python usage, server operation, output formats and development setup.Read 14 Oct 2026 · github.com
- Full-mix transcription
- The model is designed to transcribe complete mixes containing several simultaneous instruments.Read 14 Oct 2026 · mirelo.ai
- Hardware guidance
- Small is recommended for CPU-only machines, medium is the default trade-off, and large is intended for GPU use; Apple Silicon uses Metal automatically.Read 14 Oct 2026 · github.com
- Independent technical coverage
- A Japanese independent article reports that MuScriptor handles full mixes, explains its training approach and compares local, browser and Mirelo Studio use.Read 14 Oct 2026 · note.com
- Independent usage report
- A Russian video review demonstrates local installation, memory use, MIDI export and opening results in MIDI editors.Read 14 Oct 2026 · rutube.ru
- Input formats
- The local package accepts WAV and other formats supported by libsndfile, including MP3, FLAC, OGG and M4A.Read 14 Oct 2026 · huggingface.co
- Instrument detection
- It detects and labels parts such as voice, drums, bass and synth lead.Read 14 Oct 2026 · mirelo.ai
- Instrument targeting
- Users can specify expected instruments to improve transcription accuracy.Read 30 Sep 2026 · muscriptor.kyutai.org
- Instruments
- It can transcribe voice, piano, guitars, bass, strings, brass, woodwinds, drums and other instruments.Read 30 Sep 2026 · muscriptor.kyutai.org
- Kyutai company
- LinkedIn lists Kyutai as a Paris nonprofit founded in 2023 with 11–50 employees.Read 14 Oct 2026 · linkedin.com
- Kyutai funding
- Kyutai says it is funded by Iliad Group, CMA CGM Group and Schmidt Sciences.Read 14 Oct 2026 · kyutai.org
- Launch news
- Kyutai and Mirelo announced the open model and free audio-to-MIDI product on 9 July 2026.Read 14 Oct 2026 · mirelo.ai
- Limit: local performance
- A reproducibility report found Large inference on an RTX 4090 Docker setup over six times slower than manual hosted-demo timings.Read 14 Oct 2026 · github.com
- Limit: MusicXML
- A Windows user opened a bug stating that generated MusicXML differed substantially from the source score.Read 14 Oct 2026 · github.com
- Limit: transcription errors
- A reported case had sustained blocks, overlapping instruments, wrong instrument labels and a slowdown toward the end of a song.Read 14 Oct 2026 · github.com
- Limit: velocity
- A user reported that every note was exported with velocity 100, losing loudness information.Read 14 Oct 2026 · github.com
- Local deployment
- The same web UI can be self-hosted with `uvx muscriptor serve`, and transcription is available through a CLI and Python API.Read 30 Sep 2026 · github.com
- Local requirements
- Local use requires Hugging Face authentication and acceptance of the CC BY-NC 4.0 model license.Read 30 Sep 2026 · github.com
- Local server API
- The self-hosted server exposes POST /transcribe and streams note events with Server-Sent Events.Read 14 Oct 2026 · github.com
- Mirelo legal entity
- Mirelo's product pages identify the company as Mirelo AI GmbH.Read 14 Oct 2026 · mirelo.ai
- Model architecture
- The released models use a decoder-only transformer that predicts transcription tokens from five-second audio chunks.Read 14 Oct 2026 · mirelo.ai
- Model variants
- Small, medium and large checkpoints contain about 103 million, 307 million and 1.4 billion parameters respectively.Read 14 Oct 2026 · github.com
- MuseScore dependency
- PDF sheet engraving requires MuseScore 4 or newer installed separately.Read 14 Oct 2026 · github.com
- Musical context
- The associated application can detect chords, key, tempo, meter, beats and downbeats.Read 14 Oct 2026 · mirelo.ai
- Operating systems
- Documented local commands cover Linux, Apple-Silicon macOS, Intel macOS and Windows.Read 14 Oct 2026 · github.com
- Outputs
- The hosted service provides MIDI, sheet-music PDFs and guitar-tab PDFs for download.Read 30 Sep 2026 · muscriptor.kyutai.org
- Praise: timing
- A GitHub user reported that note timing and lengths were good in testing.Read 14 Oct 2026 · github.com
- Python package
- MuScriptor is installable from PyPI with uv or pip and exposes a TranscriptionModel class.Read 14 Oct 2026 · pypi.org
- Release status
- PyPI labels the package Beta and lists version 0.3.0 released on 5 August 2026.Read 14 Oct 2026 · pypi.org
- Research publication
- The MuScriptor paper was submitted to arXiv on 9 July 2026 and revised on 3 August 2026.Read 14 Oct 2026 · arxiv.org
- Security and privacy
- Uploaded audio is transcribed on the server and not kept afterwards; running locally keeps audio on the user's machine.Read 30 Sep 2026 · muscriptor.kyutai.org
- Self-hosting
- The repository documents running the same web UI locally with uvx muscriptor serve.Read 14 Oct 2026 · github.com
- Separate tracks
- Each detected instrument is returned as its own editable MIDI track.Read 14 Oct 2026 · mirelo.ai
- Sheet music output
- The tool can export MusicXML, full scores, individual PDF parts and tablature for supported fretted instruments.Read 14 Oct 2026 · mirelo.ai
- System requirements
- The package requires Python 3.10 or newer; Intel Macs should use Python 3.10–3.12.Read 14 Oct 2026 · pypi.org
- Tabs and notation
- Fretted instruments such as guitar and bass produce tablature alongside standard notation, and the system does not write chord symbols.Read 30 Sep 2026 · muscriptor.kyutai.org
- Training data
- The model was trained on 170,000 songs spanning classical music to heavy metal.Read 30 Sep 2026 · muscriptor.kyutai.org
- Usage restriction
- Users must have the necessary rights to music they submit and to generated outputs, and the model is provided without warranty.Read 14 Oct 2026 · huggingface.co
- Usage scale
- Kyutai said in a LinkedIn update that its server was transcribing hundreds of songs per day two months after release.Read 14 Oct 2026 · linkedin.com
- Weights licence
- The Hugging Face model weights use CC BY-NC 4.0 and are restricted to non-commercial use.Read 14 Oct 2026 · huggingface.co
Plans and prices
Strong points and limits
Strong points
- Hosted demo is free.
- Accepts seven listed audio formats.
- Produces MIDI, sheet music, and guitar tabs.
- Offers a CLI and Python API.
- Local use keeps audio on the user's machine.
Limits
- Transcription is not perfect.
- Results are better on acoustic and guitar music than electronic music.
- Chord symbols are not produced.
- Local use requires authentication and license acceptance.
Who it suits
MuScriptor suits musicians who want to turn recordings into editable notation or MIDI, particularly for acoustic and guitar music. Local deployment may suit users who want audio to stay on their own machine and can meet the model access requirements.
Verdict
MuScriptor offers several output formats and ways to use its transcription system, including local deployment. Expect limitations with electronic material and chord-symbol notation.
Where Specifiction places it
- Audio-to-MIDI converters#7 of 438.7
- Music Transcription Software#7 of 318.7
- AI Music Transcription Software#7 of 298.6