audiosronnx-mossformergan

ONNX export of the MossFormerGAN_SE_16K generator from ClearerVoice-Studio (Alibaba, Apache-2.0), for the mossformergan engine in audiosronnx.

MossFormerGAN pairs the MossFormer attention backbone with a metric-GAN objective and holds the highest published PESQ of any model surveyed for this library — 3.47 on VoiceBank+DEMAND. It predicts a masked magnitude plus an additive complex residual, so it corrects phase rather than reusing the noisy phase.

spec[1, 2, 401, 201]  ->  enhanced[1, 2, 401, 201]     (real, imag)

Only the generator is exported; the discriminator supplies the training loss.

Measured SNR gain on speech at 19 / 11 / 5 dB input: +2.6 / +8.3 / +11.6 dB.

Fixed window

MossFormer's group attention reshapes the sequence into fixed-size groups, and that reshape captures the traced length — a dynamic-length graph runs at its trace size and fails elsewhere, even with constant folding disabled. The graph therefore takes a fixed 401-frame window (~2.5 s at 16 kHz, hop 100); audiosronnx slides it with a crossfaded overlap, as upstream's own decode segments long audio.

Tracing at 1601 frames (upstream's 10 s decode window) exhausted memory during export, which is why the window is 401.

Export notes

Two operators needed rewriting, each bit-for-bit equivalent:

• torch.complex has no ONNX operator, but angle(complex(re, im)) is atan2(im, re).
• torch.eye(dtype=bool) exports to EyeLike(bool), for which onnxruntime has no kernel; an arange equality builds the same identity mask from supported ops.

License

Apache-2.0, inherited from ClearerVoice-Studio.