sp-uhh / storm

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

bad results for all the pre-trained model #8

Closed danhuixie closed 1 year ago

danhuixie commented 1 year ago

Hello, I have tried to process a small piece of noisy wave, but all the pre-trained model output bad results like white-noise. input wave's spectrum

1

output wave's spectrum

2

so , how to infer correctly?

jmlemercier commented 1 year ago

Dear danhuixie, Thanks for reporting, there was indeed a typo that came in the code with my last commit, an artifact from some old code causing the score network output to be (wrongly) divided twice by the noise level sigma. This has been corrected with the latest commit: I ran some dummy tests and it should work fine. Please pull and let me know if that works on your side Best Jim