Prefect the pronounciation

[colafly]

I'm training the model to speak Chinese with Taiwanese accent. I've recorded my own voice for 3 hours and the quality is better, however it got a bit off when generate longer sentences. I found a larger dataset here and was trying to train with it but the quality become far off.

How can I improve the naturalness of the generated speech? would the public dataset useful or i have to record more of my own voice over a even longer period of time to get better result?