Improve training quality / synthesis

[Mixomo]

Hi, I see that the PRIORS datasets are synthetic. I plan to replace them with more natural sounding datasets (either synthetic and/or human). This will help for a better and more natural pronunciation and intonation in voice synthesis?

I also plan to make an improvement on Arpabet, since it has a problem with the double R ("RR"), in Spanish, and does not read accents.
Is it possible to improve the dictionary?

I have a mod I made some time ago for Tacotron for the Spanish language, but I don't know if it can be implemented here. I leave the link to the files below. Thanks!

https://drive.google.com/file/d/19AGqgfWiMc8MYHuH_705phCbm8-GGIDa/view?usp=drive_link