RefineGanVCTKV1

[remote prairie]

HF page: https://huggingface.co/SimplCup/RefineGanVCTKV1
RefineGan model trained for Applio RVC on VCTK-Corpus v0.92 (mic1 variant) dataset.

• 177 epochs
• 1180590 steps
• dataset length: 35 hours
• sample rate: 44.1k

To use this model you'll need to download the latest version of Applio (https://github.com/IAHispano/Applio/tree/main) This is a first version of such model, so expect inconsistencies in training small models. Future improvements are coming with the next versions.
This is my first attempt at training a new pretrain for RefineGan vocoder. The model is severely undertrained so to make a good sounding model you'll need a pretty big dataset, minimum 10 minutes but the more the better (i personally recommend at least 20 minutes of clean speech) and train the models to minimum 400 epochs.

How to use:

• download the weights and put them in pretrained_custom folder.
• in "Training" tab change from RVC to Applio, then choose 44.1k sample and RefineGan vocoder.
• In "Training" window open advanced settings, enable custom pretraineds, and choose the G and D weights.

[patent chasm]

yahoo :3

[remote prairie]

if you've trained a model on this pretrain please post the example of your model it will help me and engineers to understand how it behaves better

[patent chasm]

pls get some samples in here im too lazy

👉👈

[remote prairie]

examples of a small model trained on the pretrain, dataset length 5 minutes (not enough), epoch count - 250 (also not enough), minimum dataset length and amount of epochs is listed in the post.

[patent chasm]

served as per usual

[vast wharf]

thats really good

is there a refinegan colab

[remote prairie]

i don't know, maybe applio should support it, but you'll need to add these pretrains yourself each time in colab

[vast wharf]

yeah thats not a big deal

thanks for making a good pretrain for it that isnt japanese🙏 😭

[remote prairie]

i mean, it's good enough, but not the best yet, wait for v2, i'll train it more and this time properly

[cunning violet]

I'll do this for that.

[vast wharf]

bet

[opal gale]

could I have the original samples you used so I have more speaking samples for my models?

[remote prairie]

you can use your own samples, it doesn't really matter

[opal gale]

no I mean the audios without the model over it bc idk where to find them

besides the singing one I have that song on my pc

[remote prairie]

yes, you can just use your own samples, i don't need the exact ones i used for my examples

use balling or smth lmao

[opal gale]

I'm just asking for the audios 😭

[remote prairie]

bro i have 0.5 mbps upload speed leave me alone

it will take me 3 years to upload the samples i use

[opal gale]

you're good man can u possibly send the yt links to the audios if possible

if they have any

[remote prairie]

two of them are just ludwigs mogul mail videos untouched and cut, the singing one is laufey from the start

ludwig is quite expressive so his recordings are pretty good for checking the models quality

[opal gale]

ahh oki thx man👍

[remote prairie]

np

[opal gale]

do u remember which videos they were from?

[remote prairie]

just use any, he's expressive in all of them

[broken depot]

how many days did it take to train this pretrain?

[remote prairie]

5 days

[broken depot]

jesus

[cunning violet]

V2?

[remote prairie]

in training

[cunning violet]

While on training, I will do titan pretrain first

For my new model

[opal gale]

[static rune]

KLM is also good esp for singer models

[vast wharf]

gah dayum

[remote prairie]

next one will possibly take 10

[vast wharf]

https://tenor.com/view/skeleton-gif-13148928981517710530

[remote prairie]

that's pretty fast for a singular consumer grade gpu

[vast wharf]

true

u train locally?

[remote prairie]

yes

[vast wharf]

what gpu do you have

[remote prairie]

4070s

[vast wharf]

luckyy

[remote prairie]

i've worked for it

[vast wharf]

well good job on the pretrains you make they are all super good

[remote prairie]

thanks

[vast wharf]

10 epochs sounds like a fart gun but im trusting the process😭

[patent chasm]

it needs lots of training