Set seed to -1 for random generation.
Turn off at early stage might offer better results
separate language tokens
No mel condition
Use accent grl condition
Use prosody encoder