MuScriptor

MuScriptor is an open-weight, multi-instrument music transcription model that converts complex, multi-instrument audio recordings into symbolic MIDI representations. Developed by Kyutai and Mirelo, the model leverages a decoder-only Transformer architecture trained on a large-scale dataset of 170,000 real-world songs across diverse genres.

MuScriptor is an open-weight, multi-instrument music transcription model that converts complex, multi-instrument audio recordings into symbolic MIDI representations. Developed by Kyutai and Mirelo, the model leverages a decoder-only Transformer architecture trained on a large-scale dataset of 170,000 real-world songs across diverse genres.

• 01Multi-instrument transcription from full audio mixes
• 02Instrument conditioning to selectively transcribe specific instruments
• 03Three-stage training pipeline (synthetic pre-training, real-world fine-tuning, and reinforcement learning)
• 04Three scalable model variants: small (103M), medium (307M), and large (1.4B)
• 05Includes CLI and FastAPI-based web UI for real-time live piano roll visualization
• 06Outputs standard MIDI files for compatibility with DAWs and music analysis software