audiosronnx-voicefixer

ONNX exports of VoiceFixer (Liu et al., MIT) for the voicefixer engine in audiosronnx.

VoiceFixer is general speech restoration at 44.1 kHz. Where most engines target one defect, this one is trained to undo noise, reverberation, clipping and bandwidth loss together. Two stages, the same shape as the sidon engine: a ResUNet predicts a clean log-mel and a TFGAN vocoder resynthesises the waveform.

analysis : mel_orig[1, 1, 501, 128]  ->  mel[1, 1, 501, 128]
vocoder  : mel[1, 1, 501, 128]       ->  wav[1, 1, ~221k]

File	Purpose	Size
analysis.onnx	ResUNet mel predictor	282 MB
vocoder.onnx	TFGAN vocoder	133 MB
mel_filters.npy	1025x128 htk mel filterbank for the numpy front-end	0.5 MB

The STFT (n_fft 2048, hop 441), the mel projection and the 10**clip(x, max=5) log inverse all run in numpy inside audiosronnx.

Verified against torch on real speech: analysis correlation 1.00000000, the pair end-to-end 0.99999996 (71.4 dB SNR).

Notes

• Fixed window. The analysis graph is length-specialised, so it takes a fixed 5 s (501-frame) window; audiosronnx slides it with a crossfade. Upstream's own restore_inmem segments long audio too.
• sp is genuinely unused. The module signature is forward(sp, mel_orig), but zeroing or scaling sp changes the output by exactly zero, so the exported graph takes only mel_orig. That is correct, not a folded constant.
• Measure parity on speech. The same graph reads 6.8e-02 against random noise and 2.3e-04 against speech. Noise is out of domain for a speech restoration model and drives it into a regime where float error through a 70M-parameter ResUNet amplifies.
• Generative. It resynthesises rather than filters, so restored detail is invented.

License

MIT, inherited from the upstream VoiceFixer release.