Billie Eilish (RVC2) KLM HifiGAN (50-100ep) (Rmvpe)

[tawny hare]

50 ep

100 ep

[raw vortex]

32bit?

are these from sessions or isolations

[lofty light]

Model Ready ✅

This model has been synced with Weights and is ready to use for free!

[tawny hare]

isolations

[raw vortex]

what did u train on

48k sample rate?

i have a feeling this undertrained and not trained on the right sample rate

[tawny hare]

32k

[raw vortex]

did you train it beyond 100e?

[tawny hare]

no, 100

[raw vortex]

train it more

how long is the dataset

[tawny hare]

No, the model is in its exact point.

[raw vortex]

^

[tawny hare]

40min

[raw vortex]

train it more lol

[tawny hare]

the graph started to rise, I know why I say that.

[unborn hornet]

Its more than just that

[tawny hare]

I'm sorry, I don't like my models to be noisy or robotic. I always look for clear points. For now, this model is finished. I'll update it later.

[unborn hornet]

So whos gonna tell this kind soul. I don't wanna

[tawny hare]

Friend, working on Billie's voice is no small feat. It takes time to get her voice perfectly clean, and on top of that, her voice whispered. It's not just any neutral voice, that's why there are currently noisy models here, without capturing the whisper of the voice, which can't be fixed even with a preset FX. So far, I've tried this model and I can say that it's good. Separating her vowels is too hard, and it became perfect when searching for her whisper. 🦕

[unborn hornet]

🤷‍♂️

[raw vortex]

LMAO

😭

so how do you know if its overtrained if you didnt continue training it ?

thats kinda like

how you figure out its overtrained

ur supposed to let it train for a while and look at the graph for points

not say "its overtraining cause it started rising"

that's not how you detect overtraining lol

and with a dataset of 40 mins 100 epoch is fairly low

[unborn hornet]

Thank youu! 🙏

[raw vortex]

the dataset is not fully trained

you gotta contiune to train to find out if its "overtrained"

[tawny hare]

Bro, I told you that the graph stopped going down to 150 epochs, it started to sound distorted

[raw vortex]

show graphs

[unborn hornet]

Hold up, (Bro).
Sometimes (NOT ALL THE TIME) But sometimes the quality can start sounding better later on after its gotten a little worse

[tawny hare]

Haha, I don't have them now, I closed the colab 5 hours ago

[tawny hare]

Yeah, that's not my case.

[unborn hornet]

You dont know that because you stopped training at 150! 😭

[raw vortex]

^

moral of story keep training

otherwise youd be leaving a bunch of quality off the model

[tawny hare]

I test the models every 5p times and as soon as it sounds bad or noisy I leave it, if you don't like it then don't try to do anything else, if I leave it there it's because I know what I'm doing

[raw vortex]

now it does sound ok at 100 epoch but you can tell its not fully trained

sounds very robotic

accents aren't there yet

[tawny hare]

I swear I know what robotic actually sounds like so that's my problem then, right?

[unborn hornet]

I mean yeah it is

[stray remnant]

f0 method in title and epoch count in title pls

[tawny hare]

clear

[tawny hare]

If you say it, well, that's good.

Billie Eilish (RVC2) KLM HifiGAN (50-100ep)

[raw vortex]

f0 means RVMPE or crepe

what did u use

[unborn hornet]

Something tells me he lied when he said he knows what hes doing 😭

[tawny hare]

Please go talk somewhere else, you've filled the post and I don't like it.