Reigningwet/wet-stems-hd

Wet Stems HD (HT-Demucs for the browser)

The model behind HD stem split in Wet Studio: splits a song into drums, bass, other and vocals, in the browser on WebGPU with onnxruntime-web.

• Model: HT-Demucs v4 (Hybrid Transformer Demucs) by Simon Rouard, Francisco Massa and Alexandre Défossez, Meta AI, from facebookresearch/demucs. MIT License.
• ONNX export: StemSplitio/htdemucs-onnx (MIT).
• This file: constant-folded for the fixed input shape and stored as float16 (see make_model.py), so it loads on WebGPU and is 174 MB. Output matches the export it came from to ~90 dB.

Input mix: float32 [1, 2, 343980] (7.8 s of 44.1 kHz stereo). Output stems: float32 [1, 4, 2, 343980], in the order drums, bass, other, vocals. Longer audio: 7.8 s windows overlapping by a quarter, cross-faded.

SHA-256: 319e5c46b4959c91976b7e3eb18f99a38dffb73a8e6976aa35fc40552c24b085

@inproceedings{rouard2022hybrid,
  title={Hybrid Transformers for Music Source Separation},
  author={Rouard, Simon and Massa, Francisco and D{\'e}fossez, Alexandre},
  booktitle={ICASSP 23},
  year={2023}
}