search

yxlu-0102 / MP-SENet

MP-SENet: A Speech Enhancement Model with Parallel Denoising of Magnitude and Phase Spectra
MIT License
267 stars 40 forks source link

Question about phase-domain loss employment #16

Open skylarkie opened 6 months ago

skylarkie commented 6 months ago

Thank you very much for fantastic work and code release!

I tried your anti-wrapping loss and other phase-domain loss (e.g. L1(clean_phase, est_phase)) on training other networks in a naive way, simply adding it to original losses with a scaling. However, it's very likely to result in that the model parameters ends up nan. Therefore, I'm wondering if you are used to face similar situation and solve this problem perfectly. If so, could you please share some experiences or tricks?

Thank you!

yxlu-0102 commented 6 months ago

In our subsequent work, we also encountered this issue You can try incorporating a bias when calculating amplitude and phase:

mag = torch.sqrt(stft_spec.pow(2).sum(-1)+(1e-9)) pha = torch.atan2(stft_spec[:, :, :, 1], stft_spec[:, :, :, 0] + (1e-5))