UVR-MDX-NET-Voc_FT model attribution and provenance

Model: UVR-MDX-NET-Voc_FT.onnx (unchanged weights, served as uvr-vocals.onnx)
Model developers: Ultimate Vocal Remover (UVR), Anjok07 and Aufr33.
Original MDX-Net architecture: Kuielab and Woosung Choi.
Source: https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/UVR-MDX-NET-Voc_FT.onnx
SHA-256: 534b2070fcc7df514b13ef660dc8cbb328679c2374d04354a5c42bb14ecce111
Size: 66762490 bytes

The current official UVR README asks third-party application developers using
UVR's models to honor the MIT license and credit UVR and its developers. That
model-use statement is the basis for distributing this UVR-branded model here.
Statement: https://github.com/Anjok07/ultimatevocalremovergui#license
README source pinned to: 5517e0cf0d1acd16a1618eeedec596957523f9e1

The README's LICENSE link is absent from the current repository tree. The
accompanying LICENSE-UVR-MIT.txt preserves the complete original MIT text and
copyright notice from the official repository commit
0fd6b751dbbe9da9c96660d0b4ec2b07108727b7. It is identified as the text source,
not represented as a file in the current tree.
https://github.com/Anjok07/ultimatevocalremovergui/blob/0fd6b751dbbe9da9c96660d0b4ec2b07108727b7/LICENSE

The original MDX-Net project is MIT licensed; its original notice is retained
separately in LICENSE-MDX-NET-MIT.txt.
https://github.com/kuielab/mdx-net-submission

Official inference sources:
https://github.com/Anjok07/ultimatevocalremovergui/blob/5517e0cf0d1acd16a1618eeedec596957523f9e1/separate.py
https://github.com/Anjok07/ultimatevocalremovergui/blob/5517e0cf0d1acd16a1618eeedec596957523f9e1/lib_v5/tfc_tdf_v3.py
https://github.com/TRvlvr/application_data/blob/3826b05b570dbd4fbedbc807758803b35348ba1b/mdx_model_data/model_data.json
The metadata is selected using UVR's MD5 of the final 10000*1024 bytes:
77d07b2667ddf05b9e3175941b4454a0 (not the complete-file MD5).

This browser application applies those published STFT and model parameters in
JavaScript and ONNX Runtime Web. It does not retrain or alter model weights.
The model estimates singing vocals; instrumental audio is mixture minus vocals.
Separation artifacts and retained instrument sounds remain possible. It does
not distinguish a lead singer from backing singers.