uddompang/kamvox-voice-separator

KamVox voice separator (MDX-Net ONNX for the browser)

Voice / background separation models from Ultimate Vocal Remover (MDX-Net), repackaged for KamVox Studio, which runs them in the browser with onnxruntime-web (WebAssembly or WebGPU). Nothing is retrained: the weights are the original ones.

id	model	predicts	files	n_fft	compensate
kim-vocal-2	Kim Vocal 2	voice	kim-vocal-2.onnx (66.8 MB), kim-vocal-2.fp16.onnx (33.4 MB)	7680	1.009
inst-hq-3	UVR-MDX-NET Inst HQ 3	background	inst-hq-3.onnx (66.8 MB), inst-hq-3.fp16.onnx (33.4 MB)	6144	1.022

All models: hop 1024, dim_f 3072, trained on 256-frame chunks at 44.1 kHz stereo. The other stem is the mix minus the prediction.

What changed from the original files

• Free time axis. Input and output are [batch, 4, 3072, frames] instead of [batch, 4, 3072, 256]; frames must be a multiple of 32. Only the declared shapes changed (the time axis goes through convolutions only). At 256 frames the output equals the original's.
• fp16 copies (*.fp16.onnx, float32 inputs and outputs) for a smaller download.

config.json lists the settings and the sha256 and size of every file. tools/separator-export/build.py in the KamVox repository rebuilds this folder from the original files and checks it against them.

Input and output

input / output: float32 [batch, 4, 3072, frames] = (left real, left imaginary, right real, right imaginary) of a torch.stft-style spectrogram (Hann window, periodic, center=True, reflect padding, bins 0..2 zeroed, bins above 3072 dropped). Multiply the inverse STFT of the output by compensate. KamVox does the STFT, chunking and overlap-add in Rust (crates/kamsub/src/separate.rs), following python-audio-separator's MDX inference.

Credits and licence

• Kim Vocal 2 (Kim_Vocal_2.onnx): Kimberley Jensen (KimberleyJSN), released with Ultimate Vocal Remover.
• UVR-MDX-NET Inst HQ 3 (UVR-MDX-NET-Inst_HQ_3.onnx): Anjok07 and aufr33, Ultimate Vocal Remover.

Ultimate Vocal Remover is MIT-licensed. Please credit the model authors above when you use these files.

Repository: uddompang/kamvox-voice-separator.