Applying HiFi-GAN to long sequences

[bryanhpchiang]

Here's a basic question: so HiFi-GAN is trained with a segment size of 8192 by default (8192 waveform samples), which corresponds to around 8192 / hop_size Mel frames.

What happens when we want to turn more than 8192 / hop_size Mel frames into audio? Is the generator applied repeatedly? Thanks!

[XuesongYang]

I'm not entirely certain about your question, but you could opt to specify a longer audio segment or longer mel frames during the inference stage.

[Ziyi6]

@bryanhpchiang Did you find out something regarding the segment size?