search

Text-to-Audio / Make-An-Audio

PyTorch Implementation of Make-An-Audio (ICML'23) with a Text-to-Audio Generative Model
MIT License
750 stars 110 forks source link

confusing things in CLAPScore evaluation #14

Open PeiwenSun2000 opened 2 months ago

PeiwenSun2000 commented 2 months ago

I have try to use the CLAPScore to eval my own model after generation. however, I find two confusing things.

  1. CLAPScore is different, even if I maintain the generated audio unchanged. Where did the random factor come from? How can I get a fixed test result.
  2. My CLAPScore is higher (better) than the GROUND TRUTH. Although I use more data to train this model, is it possible?