Harmonic noise in MB Melgan if not trimming silence in training

[fervmsanas]

Hi, I got some harmonic noise at silence segments (top plot) if I set the silence trimming option to False for MB Melgan training. It does not happen if I set this the trimming to True (bottom plot) and I keep the same training conditions. I have been trying different configurations and this is the only case when I get this issue. By including the silence in training I was expecting to reduce some slight noises appearing sometimes at these segments. Any thoughts or experience on this?, thank you in advance.

[kan-bayashi]

This is an interesting observations. I have never seen this kind of phenomenon but in my experiences trimming the silence part as much as possible leads the improvement of the quality. Not directly related to this issue, but hifigan discriminator is very strong so I’m curious the combination mb melgan G + hifigan D.

[ZDisket]

@kan-bayashi

but hifigan discriminator is very strong so I’m curious the combination mb melgan G + hifigan D.

I've been using that combination for the past few months and it seems to help a lot for making the generator finetuneable and increasing multispeaker performance. With hifigan discriminator and a learning rate decrease I can adapt mb-melgan generator (this requires a pretrained mb-melgan with hifigan d, both models to load) within 15 to 30k steps to new voice with at least 30 minutes of data. This is my config, I can share samples.

[yt605155624]

mark