"Music-grade" voice removal in the browser: HTDemucs through ONNX Runtime Web.

- The model, htdemucs_embedded.onnx (181 MB), is HTDemucs by Meta Platforms, Inc.
  (https://github.com/facebookresearch/demucs, MIT), exported to ONNX by demucs-web (timcsy, MIT) and
  mirrored byte-identically at https://huggingface.co/ptasior/pizu-audio-models. It is not in this
  repository: Booth serves it from the server's data directory (/models/), the demo fetches it from
  the mirror.
- lib/: the demucs-web pre/post-processing (STFT in, mask and iSTFT out), MIT, see lib/LICENSE.
- ort/: ONNX Runtime Web 1.31.1 (Microsoft, MIT, see ort/LICENSE), the WebGPU build and its wasm.

Nothing leaves the browser: the sound is separated on the viewer's own GPU.