Open-Unmix UMX-HQ vocals (ONNX)
An ONNX export of the vocals network from Open-Unmix
UMX-HQ (pretrained weights: Zenodo 10.5281/zenodo.3370489,
vocals-b62c91ce.pth, MIT). The Mix Directive desktop app uses it to isolate vocals from a full mix
before aligning lyrics, running locally with ONNX Runtime on the CPU. No PyTorch is required.
Changes from the original
This is a modified redistribution under the original MIT licence (see LICENSE, © 2019 Inria):
- Only the vocals target's network is included, converted to ONNX (opset 17).
- The graph is traced through a thin wrapper that repeats
OpenUnmix.forwardwith a dynamic frame axis. The original reads the frame count as a Python int, which tracing would freeze. Batch size is fixed at 1 and stereo input is fixed. - The STFT/iSTFT and Wiener filtering are not part of the graph. The app applies the output as a ratio mask instead.
The weights are unchanged (float32). The ONNX output matches PyTorch to within 5e-6.
Files
| File | Size | SHA-256 |
|---|---|---|
model.onnx | 35,623,512 bytes | f27742bb52b24cb039614dc76d0764e279df3eb2a9d6daa82c6daee3369ec747 |
export_umxhq_vocals_onnx.py | — | Script that produced model.onnx |
Usage
- Input
magnitude: float32[1, 2, 2049, frames]. This is the magnitude STFT of a 44.1 kHz stereo signal (n_fft 4096, hop 1024, periodic Hann window, centred with reflect padding, as intorch.stft). - Output
vocals_magnitude: the estimated vocal magnitude, the same shape as the input. Divide it by the input magnitude, clip to [0, 1], multiply the complex mix STFT by that mask, then invert. - The network is a bidirectional LSTM, so give it a few seconds of context either side when processing in chunks.
Evaluation
This was tested informally for lyric alignment on five songs with hand-corrected word timings. Words aligned within 0.5 s rose from 77.6% on the unseparated full mix to 85.5% after separation, compared with 87.4% from the original vocal stems. Separation took 1–2 s per song on a laptop CPU.
Training data
The upstream authors trained UMX-HQ on MUSDB18-HQ and publish the weights under the MIT licence. See the Open-Unmix repository and the Zenodo record.
Citation
Stöter, F.-R., Uhlich, S., Liutkus, A., & Mitsufuji, Y. (2019). Open-Unmix – A Reference Implementation for Music Source Separation. Journal of Open Source Software, 4(41), 1667. https://doi.org/10.21105/joss.01667