search

shivammehta25 / Matcha-TTS

[ICASSP 2024] 🍵 Matcha-TTS: A fast TTS architecture with conditional flow matching
https://shivammehta25.github.io/Matcha-TTS/
MIT License
716 stars 87 forks source link

how did you choose sigma_min? #68

Closed seastar105 closed 6 months ago

seastar105 commented 6 months ago

in p-flow tts paper, they choose sigma_min as 0.01, and here use 0.0001

i'm also applying flow matching for tts based on p-flow. is there any reason to choose sigma_min as 1e-4?

shivammehta25 commented 6 months ago

Actually, I did not do a hyperparameter tuning of it, you can try both :) In theory, since we want the Gaussian to sit at individual data points and then learn that vector field I tried a small sigma which worked just fine. But if you do find some differences let me know :)

Regards, Shivam