Applio trained models seem off..

[meager elm]

Hello, first time chatting here.

Ive been making AI voices with Mainline RVC for a while.

I have a bit of a concern because I have this issue where I have a hard time making models in Applio.

What I mean by that, is...

I usually keep everything in default if I dont know what they do. what I do know, is the 40k and 48k sampling rate, and all the Pitch extraction algorithms, as well as Embedder Models. <ainline RVC let me make beautiful models, but as soon as I came to Applio, the audio sounds crashing down.

Both trained on a 5090, 40k sampling rate, same pitch extraction models.

Ill just attach 2 audios right here:

[mainline rvc]

[molten cradle]

there is no 40k and 48k pretrains for refinegan, only 32k and spin v2

[meager elm]

[Applio]

[meager elm]

oh I know its just selection screenshot sorry xD

Im clearly doing something terribly wrong with the same exact dataset

and Im quite confused

[molten cradle]

enable "split audio" in the inference settings

[molten cradle]

you are training from scratch

without a pretrain, thats why it sounds like that

if you want to train 40k and 48k don't use refinegan

use hifigan

and contentvec

[meager elm]

can I bring the RVC model that I have trained before to here?

[molten cradle]

not needed, restart applio and use what i said, its what mainline uses

hifigan, contentvec, and train

it's going to work

[meager elm]

oki Ill try thank you, ill be back in a bit

[molten cradle]

also NEVER disable this setting! always leave this enabled

[meager elm]

what I have on currently

[molten cradle]

disable overtraining detector is not accurate

the rest is fine

[meager elm]

oki

also

this is the folder of where all the audio is right

the same thing as mainline?

[molten cradle]

yup, copy and paste the location of where the audios are located (dataset)

[meager elm]

do I check dataset creator?

[molten cradle]

not needed, it's going to work if you manually paste the location of the dataset

[molten cradle]

this fine also

[meager elm]

this is when ur making the tts/song using the model ig

done

[molten cradle]

what does the cmd window say?

[meager elm]

index algorithm?

[molten cradle]

yes enable that when you're doing voice conversion, split audio makes it work like mainline

[molten cradle]

everything looks fine

[molten cradle]

faiss

or leave it in auto

[meager elm]

oki

this should be enough to test

[molten cradle]

if you click train model the cmd window should say something like:
loaded pretrained G: f040k.pth

if it says that, all is good

[molten cradle]

batch 24 is too high tho

[meager elm]

I think I did it well, at least

too high?

[molten cradle]

yes batch size its too high

use 4 instead

[meager elm]

whats the difference

between 24 and 4?

[molten cradle]

hard to explain in simple words, but its just gonna sound better

[meager elm]

how do I stop this and make it into a 4

[molten cradle]

click stop training

[meager elm]

well I did but doesnt it like

[molten cradle]

if its stuck just close the cmd window

[meager elm]

if I have like

different batch size

doesnt it need all the other steps again?

or does it start right from 0

[molten cradle]

you need to start from 0

and always click "fresh training" if you change batch size

very important

[meager elm]

ok so

fresh training is on so

I can just change the batch size

then press start training again?

[molten cradle]

yea, as long fresh training is on

[meager elm]

yk whats crazy

batch 24 cmd line says

0.02s per epoch kekw

thats crazy

2s

[molten cradle]

its a tiny dataset with a huge batch

[meager elm]

mb

[molten cradle]

expected xD

[meager elm]

its not that tiny its like

probs 10 minutes or so

ehh maybe not 10

like 8

max

[molten cradle]

but the cmd window said u got 2 minutes

[meager elm]

just started with batch 4

[meager elm]

?? maybe I selected the wrong folder

wth

wait hold on

[molten cradle]

yea

[meager elm]

k I found out there was this one file where

my friend was talking for a hot minute

then he just stop speaking

for like 5 mins

kekw

it probs cleaned that up

I didnt even realize it

welp Ill see if this model is good 😄

welp I dont think its gonna take that long

[meager elm]

done, let me test

better, but still a bit lacking from mainline model Ive made previously. any suggestions to make it better?