From what I know, dear @CookiePPP, sometimes the models sound bad because of the vocoder. Have you made any newer vocoders that worked with codedump, thanks.
(ps. If you have better models than the available ones in infer script, can you send those as well)