Mix Directive dereverb (ONNX)
A dereverberation mask network trained by Mix Directive. The Mix Directive desktop app uses it for the Learned method of Reduce Reverb in the Spectral Editor, running locally with ONNX Runtime on the CPU.
Licence: proprietary, all rights reserved (see LICENSE). The weights are published here only so the
app can download them.
Model
- A temporal convolution network: 2 × 9 dilated depthwise-separable residual blocks (256 channels, dilations 1–256 frames, receptive field ≈ 10.9 s), about 1.5 M parameters, float32, ONNX opset 17.
- Input
logmag: float32[1, frames, 513],log(|STFT| + 1e-5)of the RMS-normalised mono signal (48 kHz, n_fft 1024, hop 256, periodic Hann window, centred with reflect padding, as intorch.stft). - Output
mask: float32[1, frames, 513]in [0, 1]. Multiply the complex STFT by it (phase kept) and invert. The mask keeps the direct sound and the first 50 ms of the room, and removes the late reverberation.
Evaluation
On held-out singers (VocalSet) and speakers (VCTK) that were not used in training, with generated plug-in-style reverbs, it removes 4–5 dB more reverb error than the input (about +4.4 dB for singers and +5.1 dB for speakers, against +0.6 dB for the app's classical method). It leaves dry voices and chords within ±0.05 dB, and steady pure tones are lowered by about 0.2 dB on average.
Training data
Trained on pairs made by convolving dry recordings with generated room and plug-in-style reverb responses. The dry recordings came from:
- VocalSet — Wilkins, J., Seetharaman, P., Wahl, A., & Pardo, B. (2018). VocalSet: A Singing Voice Dataset. ISMIR. https://zenodo.org/record/1442513 — licensed CC BY 4.0.
- CSTR VCTK Corpus 0.92 — Yamagishi, J., Veaux, C., & MacDonald, K. (2019). CSTR VCTK Corpus: English Multi-speaker Corpus for CSTR Voice Cloning Toolkit (version 0.92). University of Edinburgh. https://doi.org/10.7488/ds/2645 — licensed CC BY 4.0.
- Generated tones and dry instrument stems that Mix Directive is licensed to use.
No recordings from these datasets are distributed here.
Files
| File | Size | SHA-256 |
|---|---|---|
model.onnx |
5,963,876 bytes | db0d2df647e946751ca80613f3909d2bc3aa83bafc6566b2d604834911bbea37 |