search

sp-uhh / sgmse

Score-based Generative Models (Diffusion Models) for Speech Enhancement and Dereverberation
MIT License
514 stars 76 forks source link

log show? #56

Open a897456 opened 1 week ago

a897456 commented 1 week ago

图片

Excuse me. in the above figure, has the PESQ curve reached a point where it's hard to improve?（In the training process represented by the green curve, I cancelled the calculations of STOI and si_sdr.）

Yangliusha commented 1 week ago

Hi, May I know whether you can train your model to get similar results as shown in the paper? Do you use the default settings as given in the codes?

a897456 commented 1 week ago

Hi, May I know whether you can train your model to get similar results as shown in the paper? Do you use the default settings as given in the codes?

Yes, the model has not been changed in any way. I trained it twice. As can be seen from the two PESQ lines, the only difference between them is the dataset. However, I did not achieve the metrics in the paper, and the PESQ could only reach 3 and then stalled.

libai2345 commented 5 days ago

What dataset you're using？

a897456 commented 4 days ago

What dataset you're using？

Voicebank+DEMAND

libai2345 commented 2 days ago

Why is my training so much worse than yours?