Hi Edson — I tried again a couple of weeks ago with the Coqui.AI TTS library (this is the former Mozilla CommonVoice team). I have a CoLab notebook here — https://colab.research.google.com/drive/1-ii1rvA6mciae1NHqZTZwFG2I-WOBq5c?usp=sharing — but when loss stopped decreasing, the model was still not usable.
Audio samples: https://soundcloud.com/ndoiron