README.md

metadata

title: MuScriptor Studio
emoji: 🎼
colorFrom: gray
colorTo: indigo
sdk: gradio
sdk_version: 6.10.0
app_file: app.py
short_description: Multi-instrument audio transcription to editable MIDI
startup_duration_timeout: 1h

MuScriptor Studio

A custom ZeroGPU interface for MuScriptor, the multi-instrument audio-to-MIDI model from Kyutai and Mirelo.

The app runs the 307M-parameter medium checkpoint, streams transcription progress after each 5-second audio window, renders an interactive color-coded piano roll, and exports both the full arrangement and isolated instrument MIDI files.

Three one-click real-music examples are included: Mozart's Magic Flute overture with symphony orchestra, a live modern performance by the Harry Mitchell Quartet, and a double-tracked pop-rock guitar performance. Each example uses verified instrument guidance selected from live quality tests. Their Creative Commons provenance is documented in examples/README.md. Example results are cached lazily after first use so startup never consumes ZeroGPU quota.

The checkpoint is gated under CC BY-NC 4.0. This Space requires an HF_TOKEN secret whose owner has accepted access to MuScriptor/muscriptor-medium.

API

The Gradio endpoint is /transcribe. It accepts an audio file, optional exact instrument group names, a sampling toggle, temperature, and a beam width from 1–4. It returns the combined MIDI, all isolated track files, the browser visualization payload, and a machine-readable manifest.