search

sp-uhh / storm

StoRM: A Diffusion-based Stochastic Regeneration Model for Speech Enhancement and Dereverberation
MIT License
164 stars 22 forks source link

Pretrained checkpoints have lower results than article results #13

Closed CZ-2128 closed 9 months ago

CZ-2128 commented 9 months ago

Hi, I'm learning a lot from your ideas.I downloaded the Voicebank/DEMAND checkpoint and got the following results: pesq=2.88 stoi=0.87 si-sdr=18.5 Is it the difference in enhancement settings that's causing this? Current settings: mode : storm, corrector : ald, corrector-steps : 1, snr : 0.5, N : 50

TianyuCao commented 1 month ago

Hi, I got exactly the same results as yours. Did you finally figure it out?

CZ-2128 commented 1 month ago

Hi, I got exactly the same results as yours. Did you finally figure it out?

Hi, you can reduce sampling step N to 20-30.

TianyuCao commented 1 month ago

Hi, I got exactly the same results as yours. Did you finally figure it out?

Hi, you can reduce sampling step N to 20-30.

Sorry, I just saw your message. Do you mean if we reduce the inference steps from 50 (by default) to 20 or 30, then we can get higher results like what was shown in the paper? Thanks for your clarifications in advance.

Btw, did you have any chance to use their other pretrained checkpoint trained on WSJ0+Chime3? I also got lower results than article results, but tried to figure it out.