---
license: cc-by-nc-4.0
tags:
  - audio-to-audio
  - mastering
  - onnx
library_name: onnx
---

# ITO-Master (white-box) — ONNX export for in-browser mastering

ONNX export of the **white-box** ITO-Master music-mastering model, for running the parameter
predictor in the browser with ONNX Runtime Web. Used by the
[audio-master](https://huggingface.co/spaces/kramp/audio-master) Space.

- `fxencoder.onnx` — reference audio `[1,2,N]` (44.1 kHz stereo) → style embedding `[1,2048]`
- `mastering_tcn.onnx` — input audio `[1,2,N]` + embedding `[1,2048]` → **46** normalized
  parameters `[1,46]` in `[0,1]`
- `config.json` — the exact parameter order and per-parameter min/max for de-normalization
  (`value = norm*(max-min)+min`). The differentiable EQ / distortion / multiband-comp / gain /
  imager / limiter chain is re-implemented natively (Web Audio); only the learned predictor is
  exported here.

Exported with a Hugging Face Job (`torch.onnx.export`, opset 17), validated against PyTorch
(max abs error < 1e-3) with `onnxruntime`.

## Attribution & license

Model & weights: **ITO-Master**, Koo et al., Sony Research —
[github.com/SonyResearch/ITO-Master](https://github.com/SonyResearch/ITO-Master),
[jhtonyKoo/ITO-Master](https://huggingface.co/spaces/jhtonyKoo/ITO-Master).
Licensed **CC BY-NC 4.0 (non-commercial)**; this derivative export inherits the same license.