SS task model is much more slower then SE why?

[MonolithFoundation]

The SS task model is significantly slower than the SE model. Why is this the case?

When compared with the same mossformer-16k.

Also, for the SE task, mossformer-GAN is far superior to mossformer-se. However, it is much slower. Is there any way to have a model that is both fast and accurate?

(The better here means that GAN maintains the voice style without much distortion, but mossformer-se is very distorted.)

[alibabasglab]

Hi, thank you for your feedback. SS model is time-domain basis, the window size is quite small. That means it needs to process more frames compared to the time-frequency dimain se model.

[alibabasglab]

MossformerGAN is processing on both magnitude and phases, and trained using GAN. So it focuses more on perception, while mossformer-se is masking basis, and trained using mse losses

[MonolithFoundation]

MossFormerGAN is a good more for noisy reduction, (as I noticed SS is hard to use since it's slow), but MossFormerGAN still slow on CPU.

Will there any further plan to make both model fast while also keeping the ability to denoise?

[alibabasglab]

Thank you for your feedback, we will be looking through ways to improve the inference speed.