Vocal Separator — browser-local separation Reviewed 2026-11-12. The model is downloaded only after an explicit user action. Audio, filenames and results are not sent to the model host or inference service. Ultimate Vocal Remover (UVR), and its developers Anjok07 and aufr33 — MIT. Primary license/model attribution statement: https://github.com/Anjok07/ultimatevocalremovergui/blob/master/README.md#license The README describes UVR core developers as training the bundled models except Demucs v3/v4, and asks application developers to credit UVR and its developers when using their models under MIT. We use only the UVR team's MDX-Net model below, not Demucs, Kim, viperx or BS-RoFormer weights. The maintainers have not supplied written confirmation for this individual file. Training-data rights are unverified. Model licensing never grants rights to the user's source audio. Model: UVR-MDX-NET-Voc_FT.onnx Official release: https://github.com/TRvlvr/model_repo/releases/download/all_public_uvr_models/UVR-MDX-NET-Voc_FT.onnx Bytes: 66762490 SHA-256: 534b2070fcc7df514b13ef660dc8cbb328679c2374d04354a5c42bb14ecce111 Every download and cache read must pass both byte-length and full SHA-256 checks. The official release URL is the provenance source. Browser delivery uses this CORS-compatible fixed-revision mirror, independently downloaded and compared byte-for-byte with the official file (same full SHA-256 and length): https://huggingface.co/Blane187/all_public_uvr_models/resolve/fddec39677560e41e3194f24a9e4c4cd32ef0e83/UVR-MDX-NET-Voc_FT.onnx The mirror is a delivery location, not an independent license grant. MDX-Net architecture: Kuielab & Woosung Choi. https://github.com/kuielab/mdx-net Preprocessing reference (UVR): lib_v5/tfc_tdf_v3.py and separate.py. Model-specific parameters in models/MDX_Net_Models/model_data/model_data.json: UVR tail-MD5 77d07b2667ddf05b9e3175941b4454a0; n_fft=7680, hop=1024, dim_f=3072, dim_t=256, compensate=1.021, primary stem=Vocals. Stereo complex input/output: [1,4,3072,256], real/imag per channel. Periodic Hann, centered reflect-padded STFT inside zero-padded chunks, first three bins zeroed; context trimmed, squared-window normalized inverse STFT, instrumental=mix-vocals. This app implements its own DSP and WAV encoding. It uses no ffmpeg.wasm or @ffmpeg/core. The server's unrelated existing practice renderer is separate. MIT license notice for UVR and its developers (Anjok07, aufr33): Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. Runtime: ONNX Runtime Web 1.31.1, Microsoft Corporation, MIT. See onnxruntime-LICENSE.txt and onnxruntime-ThirdPartyNotices.txt. https://github.com/microsoft/onnxruntime/tree/v1.31.1 FFT: fft.js 4.0.4, Copyright Fedor Indutny, 2017, MIT; see fft-LICENSE.txt. Our 7680-point transform composes fft.js's 512-point transform with radix 15. Historical note: the previous local prototype used Spleeter / sherpa-onnx weights. It is no longer used by either free or paid browser-local separation. Retained Spleeter/sherpa license files describe that historical version, not this pipeline. Cloud Spleeter attribution is at /notices/cloud-separation/NOTICE.txt. No universal separation quality or native-phone compatibility is promised.