the results and training loss

sp-uhh / sgmse

Score-based Generative Models (Diffusion Models) for Speech Enhancement and Dereverberation

MIT License

491 stars 74 forks source link

the results and training loss #42

Closed DingGitTemp closed 8 months ago

DingGitTemp commented 8 months ago

I attempted to train and test the model on the voicebank-demand dataset, but the results were not satisfactory. The enhanced speech couldn't be recognized as human voice . Are there any parameters that need to be reset?Additionally, during the training process, the loss of the training set consistently remained around 700. Is this normal?

cobalamin commented 8 months ago

Hi, did you downsample your version of the VB-DMD dataset to 16 kHz? The model is by default designed for 16 kHz.

DingGitTemp commented 8 months ago

The issue has been resolved; it turns out I hadn't downsampled the data to 16K. Thank you very much for your response; this truly is a remarkable piece of work.

cobalamin commented 8 months ago

Thanks! Happy to hear that it works now.