RefGanVCTKV2

[forest crest]

HF: https://huggingface.co/SimplCup/RefGanVCTKV2

This is a RefineGan RVC pretrain for training small models.

Epoch count: 341 (40k), 341+34 (32k)

Number os steps: 2600125 (40k), 2859375 (32k)

Total dataset length: 68 hours

Sample rate: 40k, 32k

Batch size: 12

To use this model u need to download Applio latest build in the main branch, and either replace the pretrains in rvc/models/pretraineds/pretraineds_v2 or put them in pretraineds_custom and then choose the weights in the ui.

[quasi pecan]

rip the electric bill

https://tenor.com/view/christmas-vacation-electric-gif-10040448

why is there a robotic noise on hifi gan?

[flat blade]

it has to do with the dataset of the pretraining; both models were trained on different pretrains hence why tou can see that both models struggle in different areas

for example if u look at hifigans output of bbnos you will hear that the esses tended to be softer and nicer, you can hear the difference between the to at 0:25

refinegan poorly reproduces whereas hifigan clearly does a little better and vice versa

[quasi pecan]

i see

does it have a singing speaker?

[forest crest]

almost half of the dataset contains singing voices

[quasi pecan]

les gooo

[forest crest]

32k version is up

[quasi pecan]

is there 44k or is it still training?

[forest crest]

Training

[earnest oyster]

Does Refine GAN sound better than HiFi

[forest crest]

Sounds more stable, but depends on a situation

[earnest oyster]

What do you mean stable?

Like it can handle more emotional speech? Without breaking?

[forest crest]

Less random artifacting

[earnest oyster]

Ohh

[quasi pecan]

welp

[quasi pecan]

very gud for a 1 H O U R dataset

[lunar wave]

Hmmmm

While KLM5 RefineGAN not working properly for me (My Datasets), this VCTK v2 working properly, nice and Thank You

10 Minute Dataset, Batch 4 RMVPE

[forest crest]

Glad it works well

[earnest oyster]

This pretrain is super good

[quasi pecan]

me training again cuz i ran out of credits on lightning ai

good thing the g and d file was saved

[ripe fjord]

why is your model so much smaller than the KLM5 refgan? Did they upload the wrong file?

[forest crest]

no i just removed the optimizer and converted them to fp16 with noobieses code

[round storm]

need a little help, where is the refgan vocoder option? using main branch

[flat blade]

since its unstable atm the devs hid it in the main branch

[quasi pecan]

e

[ripe fjord]

is it due to inadequate pretrained model or architecture?

[flat blade]

A bit of both, no real improvements with it

[cold lake]

If you need refinegan available, use my fork
otherwise as others said, it's hidden in the mainline due to unstability and other architecture issues we've no solution for atm

[round storm]

thanks!