Muse
What it is
Muse is a free desktop application that turns recorded or uploaded audio into MIDI, with separate solo-piano and multi-instrument modes. In the piano workflow, it separates piano from a mix, transcribes the notes, and divides them between left and right hands. The multi-instrument mode transcribes up to 13 tracks in one pass without source separation. Audio can be recorded with a microphone or supplied as MP3, WAV, OGG, or FLAC files. Built-in playback uses a grand piano synth and provides per-track mute and solo controls. Muse runs locally, and its maker says audio stays on the user’s machine. GPU acceleration supports CUDA and Apple Silicon MPS. Builds are listed for Windows, macOS on Apple Silicon and Intel, and Linux Debian/Ubuntu and AppImage. Setup requires Python 3.10 or later, and the first run downloads about 3 GB of machine-learning models. The maker estimates multi-instrument CPU inference at roughly two minutes per 30 seconds of audio; a GPU makes a significant difference. Piano transcription can lose accuracy on synth piano, electric piano, or sounds unlike a concert grand. MIDI is available now; MusicXML export is on the roadmap.
For and against
For
- Transcribes up to 13 tracks in one pass.
- Accepts MP3, WAV, OGG, and FLAC.
- Runs locally; maker says audio stays on-device.
- Supports CUDA and Apple Silicon MPS acceleration.
- Free and MIT-licensed.
Against
- First run downloads about 3 GB of models.
- Requires Python 3.10 or later.
- CPU processing can take roughly two minutes per 30 seconds.
- MusicXML export is on the roadmap, not available now.
The verdict
Muse offers distinct piano and multi-instrument transcription workflows with local processing and several audio input formats. Account for the model download and performance limits, and note that MIDI—not MusicXML—is currently available.
Who it suits
Muse suits musicians or creators who want to turn recordings into MIDI, especially solo piano or multi-instrument material. It may suit users who can accommodate the initial model download and, for faster multi-instrument processing, GPU acceleration.
Plans and prices
Free
MIDI output · local desktop use · models download on first run
Read 15 Oct 2026 · github.com
Muse
Open-source · MIDI output
Read 4 Oct 2026 · github.com
On the record
- Audio input
- Users can record through a microphone or import MP3, WAV, OGG, and FLAC files.github.com · read 15 Oct 2026
- Community signal
- The GitHub repository shows 6 stars, 2 forks, and 0 watchers.github.com · read 15 Oct 2026
- Demucs model limitation
- An independent HTDemucs model card identifies six-stem htdemucs_6s separation as experimental for guitar and limited for piano quality.huggingface.co · read 15 Oct 2026
- Demucs research
- The Hybrid Transformer Demucs paper reports 9.20 dB SDR on MUSDB with extra training data.arxiv.org · read 15 Oct 2026
- Developer package
- The repository's Node package is named audio2sheets, exposes an audio2sheets CLI, and declares Node >=18.github.com · read 15 Oct 2026
- Downloads
- Builds are listed for Apple-Silicon and Intel macOS, Windows x64, Debian/Ubuntu, and Linux AppImage.github.com · read 15 Oct 2026
- Extensibility
- The source tree contains a Tauri desktop app, SolidJS frontend, Rust backend, Python pipeline, Node CLI/library, and automated tests.github.com · read 15 Oct 2026
- GPU support
- CUDA and Apple Silicon MPS acceleration are supported.github.com · read 15 Oct 2026
- Independent comparison context
- AlternativeTo lists NeuralNote and Basic Pitch as free open-source audio-to-MIDI alternatives and AnthemScore as a paid alternative.alternativeto.net · read 15 Oct 2026
- Issue activity
- The repository's open-issues view currently shows zero open issues.github.com · read 15 Oct 2026
- Licence
- Muse is released under the MIT License with copyright attributed to Muse Contributors in 2025.github.com · read 15 Oct 2026
- License
- The repository identifies Muse as open-source and provides an MIT license.github.com · read 4 Oct 2026
- Limit: CPU speed
- The README estimates about two minutes of CPU inference for each 30 seconds of multi-instrument audio.github.com · read 15 Oct 2026
- Limit: instrument timbre
- The maker says piano results worsen for synths, electric pianos, and sounds unlike concert grands.github.com · read 15 Oct 2026
- Limit: model size
- Muse downloads roughly 3 GB of machine-learning models on first launch.github.com · read 15 Oct 2026
- Limit: MusicXML
- MusicXML export is listed as future work; MIDI is the currently available output.github.com · read 15 Oct 2026
- Local processing
- The native Tauri desktop application is designed to keep audio on the user's machine.github.com · read 15 Oct 2026
- Maker identity
- A public GitHub gist profile identifies mqtik as Iván Matías and lists Argentina as the location.gist.github.com · read 15 Oct 2026
