---
license: cc-by-nc-4.0
library_name: onnxruntime
base_model: MuScriptor/muscriptor-medium
tags:
  - muscriptor
  - music-transcription
  - automatic-music-transcription
  - audio-to-midi
  - onnx
  - onnxruntime
  - int4
  - hqq
---

# MuScriptor Medium — ONNX Runtime FP16 and W4A16

This is an **unofficial derivative ONNX conversion** of
[`MuScriptor/muscriptor-medium`](https://huggingface.co/MuScriptor/muscriptor-medium),
developed by Mirelo × Kyutai. It contains an FP16 export and a smaller W4A16
export for ONNX Runtime. This repository is not affiliated with or endorsed by
the MuScriptor authors.

The conversion uses upstream revision
`f32236969308476e01fd3aae67357de5feb05a2d`. The source checkpoint used for the
export has SHA-256
`ac80adbdf85d87231735fd948af7013441c0afced316c4e9067fd5d8a7fb97ec`.

## Variants

| Directory | Transformer backbone | Embeddings | LM head | Activations | Weight files |
|---|---:|---:|---:|---:|---:|
| `onnx/fp16` | FP16 | FP16 | FP16 | FP16 | 613,150,680 bytes |
| `onnx/w4a16` | HQQ UINT4 | FP16 | FP16 | FP16 | 169,649,866 bytes |

The W4A16 graph quantizes only the 96 constant-weight MatMul operations in the
24-layer Transformer backbone. It uses ONNX Runtime `MatMulNBits`, HQQ
quantization, and block size 128. The token embedding, all other embedding
tables, conditioner, and LM head deliberately remain FP16.

`decoder.onnx.data` is external tensor data and **must remain next to**
`onnx/w4a16/decoder.onnx`.

## ONNX interface

The conditioner consumes a 5-second, 16 kHz mono log-mel spectrogram plus
instrument and dataset IDs. The decoder exposes an autoregressive interface
with explicit `past_key`/`past_value` inputs and `present_key`/`present_value`
outputs. Log-mel extraction is intentionally outside the ONNX graph.

## Quick smoke test

Download this public repository:

```bash
hf download happyme531/muscriptor-medium-onnx --local-dir muscriptor-medium-onnx
cd muscriptor-medium-onnx
uv venv
source .venv/bin/activate
uv pip install -r runtime/requirements.txt
python runtime/test_onnxruntime.py --model-dir onnx/w4a16 --provider cpu
```

Use `--audio path/to/audio.wav` to replace the generated 440 Hz test tone. For
CUDA, install a matching `onnxruntime-gpu` package instead of `onnxruntime`,
and pass `--provider cuda`. PyTorch is optional; if installed, the helper
imports it first so its CUDA/cuDNN libraries can be preloaded.

The included script validates conditioner/decoder execution and greedy token
generation. It is a low-level ONNX Runtime smoke test, not a complete
audio-to-MIDI application.

## Validation

Both models pass `onnx.checker.check_model`. The FP16 and W4A16 exports were
executed with ONNX Runtime on CUDA and x86-64 CPU; on a real audio sample their
first eight greedy tokens matched. ARM64 execution was not validated and
depends on whether the target ONNX Runtime build/provider implements the
`MatMulNBits` graph used by the W4A16 decoder.

| File | SHA-256 |
|---|---|
| `onnx/fp16/conditioner.onnx` | `cc94f58ba8023f339b4b8eba06a3f3e7623079c8234fa1be075612ee96efcf81` |
| `onnx/fp16/decoder.onnx` | `6832149920078feb4562c1c211a0dcac52a4ea10d7b715459e2c6fd0aee83b25` |
| `onnx/w4a16/conditioner.onnx` | `cc94f58ba8023f339b4b8eba06a3f3e7623079c8234fa1be075612ee96efcf81` |
| `onnx/w4a16/decoder.onnx` | `2d3b8a35c353b581982a1567eea1629ce70d4b899de1d020f4fa9cb85757943e` |
| `onnx/w4a16/decoder.onnx.data` | `8d92ff62bed61814593222a9fd8fcb731f0aa83cbd56591c56cc845b6e16b61d` |

## License, attribution, and conditions

The model weights and these derivative ONNX weights are distributed under
[Creative Commons Attribution-NonCommercial 4.0 International](LICENSE).
Commercial use is not permitted under that license. You must also comply with
the upstream supplemental conditions summarized below, including
having all necessary rights to any audio you transcribe and complying with all
applicable laws.

> **Summary of upstream supplemental conditions of use:** MuScriptor and any generated
> content by MuScriptor are provided as is without any warranty of any kind,
> including but not limited to any warranty of non-infringement. Use of
> MuScriptor and its output must comply with all applicable laws and must not
> result in, involve, or facilitate any illegal or unauthorized activity.
> Prohibited uses include, without limitation, inputting music files and
> transcribing them to MIDI/music sheet without having all the necessary
> rights, including intellectual property rights, under applicable laws.
> Accordingly, users undertake and warrant that they have all necessary rights,
> including intellectual property rights, in connection with their use of
> MuScriptor and its output. Mirelo and Kyutai disclaim liability for
> non-compliant use; users shall indemnify, defend, and hold them harmless from
> claims, damages, losses, liabilities, and expenses arising from failure to
> comply with CC BY-NC 4.0 and/or these specific conditions. The exact original
> wording is preserved in `UPSTREAM_MODEL_CARD.md`.

See [NOTICE.md](NOTICE.md) for attribution and
[UPSTREAM_MODEL_CARD.md](UPSTREAM_MODEL_CARD.md) for the original model card,
intended uses, limitations, supplemental conditions, author list, and citation.
The small runtime helpers are covered by [their MIT license](runtime/LICENSE).