Modelpublic

TigreGotico/audiosronnx-mossformer2

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes

Model Card

audiosronnx-mossformer2

ONNX export of the MossFormer2_SE_48K mask predictor from ClearerVoice-Studio (Alibaba, Apache-2.0), for the mossformer2 denoise engine in audiosronnx.

MossFormer2 is a 55 M-parameter hybrid transformer + recurrent mask predictor and the strongest fullband (48 kHz) denoiser in the library. It reports PESQ 3.16 / STOI 0.95 / SI-SDR 19.38 on VoiceBank+DEMAND.

The graph predicts a spectral mask from log-mel features; the fbank front-end, the mask application and the STFT/ISTFT all run in numpy inside audiosronnx:

fbank[1, T, 180]  ->  mask[1, T, 961]

The 180 channels are 60 Kaldi mel bins plus their first- and second-order deltas.

File	Rate	Size
`mossformer2_48k.onnx`	48 kHz	229 MB

Export note

This graph must be exported with `do_constant_folding=False`. With folding enabled the model is only correct at the length it was traced at — error grows with the distance from that length (max abs err 1.4 at T=246 and 4.4 at T=400 for a trace at T=200), while a same-length check still looks clean at 1e-4. With folding disabled, parity holds across lengths (~1e-4 from T=50 to T=1000). End-to-end against the upstream pipeline: correlation 0.99999999, 71.6 dB SNR.

License

Apache-2.0, inherited from ClearerVoice-Studio.