Sample Generation Rate for WaverRNN

I’m doing a training experimentation via Tacotron1 + WaveRNN configuration and with LJSpeech database I got pretty good results (359.3 KB)

Currently sample generation rate in batched mode is about 7 kHz for a Tesla GPU, and it corresponds to about 0.35 of realtime speed. I wonder if this generation rate can be further improved.

Best Regards.