| title: MuScriptor Music Transcription | |
| emoji: 🎵 | |
| colorFrom: red | |
| colorTo: blue | |
| sdk: gradio | |
| sdk_version: 6.20.0 | |
| app_file: app.py | |
| short_description: Convert music audio to MIDI with MuScriptor | |
| python_version: "3.12" | |
| startup_duration_timeout: 30m | |
| # MuScriptor Music Transcription | |
| This Space demonstrates [MuScriptor](https://huggingface.co/MuScriptor/muscriptor-large), a ~1.3B parameter multi-instrument automatic music transcription (AMT) model developed by Mirelo x Kyutai. | |
| Upload a music recording (any genre, multiple instruments) and the model will transcribe it into a downloadable MIDI file with per-note onset, offset, pitch, and instrument information. | |
| ## How it works | |
| 1. Upload an audio file (WAV, MP3, FLAC, etc.) or record from microphone | |
| 2. Optionally select which instruments are present to improve accuracy | |
| 3. Click "Transcribe" to play back / visualize the result and download the MIDI file | |
| ## Model details | |
| - **Architecture**: Decoder-only Transformer (dim=1536, 24 heads, 48 layers) | |
| - **Input**: 16 kHz mono audio → mel-spectrogram (512 mel bins) | |
| - **Output**: MIDI-like token sequence using the MT3_FULL_PLUS taxonomy (36 instrument subgroups) | |
| - **License**: CC-BY-NC 4.0 (weights) / MIT (code) |