cursedhelm/cubaser-muscriptor-medium-onnx

Cubaser MuScriptor Medium ONNX

Browser-targeted FP16 ONNX conversion of MuScriptor Medium for Cubaser. It performs multi-instrument music transcription entirely in a WebGPU-capable browser.

This repository is a format conversion, not a newly trained checkpoint. The original MuScriptor model was developed by Kyutai and Mirelo. The conversion splits the conditioner and autoregressive decoder into ONNX graphs with external data and supplies Cubaser's browser manifest and vocabulary.

Files

• conditioner.onnx and conditioner.onnx.data: audio/instrument conditioner.
• model.onnx and model.onnx.data: 24-layer autoregressive decoder.
• manifest.json: Cubaser WebGPU runtime metadata and instrument mapping.
• vocab.json: MuScriptor event vocabulary.
• config.json and source-config.json: architecture metadata.

The large external-data payloads are intended to be fetched through Hugging Face's resolve URLs. Cubaser retains a chunked Neocities mirror as a runtime fallback.

Requirements

The converted model uses FP16 and requires WebGPU shader-f16, currently best supported by a hardware-accelerated Chromium browser with a capable GPU.

License and attribution

The converted weights remain licensed under CC BY-NC 4.0. They are for non-commercial use. MuScriptor code is available from the official repository.

If you use the model, cite:

@misc{rouard2026muscriptoropenmodelmultiinstrument,
  title={MuScriptor: An Open Model for Multi-Instrument Music Transcription},
  author={Simon Rouard and Michael Krause and Axel Roebel and Carl-Johann Simon-Gabriel and Alexandre Défossez},
  year={2026},
  eprint={2607.08168},
  archivePrefix={arXiv},
  primaryClass={cs.SD},
  url={https://arxiv.org/abs/2607.08168}
}