Furina EN - Genshin Impact 4.2

Furina EN - Genshin Impact 4.2 (RVC2) (140epoch) (47460steps) (RMVPE)

*this model might be incomplete. The index is only 30mb which appears to be incorrect.

Will be retrained if i can fix the issue

tf, 47 k steps on 140 epoch??? how long is your dataset?

2hrs

why so long

Because more voice lines is more consistent than less?

Training time and index size is not much of a concern.

well, you don't need that much data anymore thanks to pretrain v2

Works fine for my models so far. I don’t see any drop in quality.

you'd need that much data if we didn't have pretrained models, but now we can use like 5-10 minutes and the model will come out as good as it will with 2 hours dataset

Also makes speaking other languages better

does the dataset have her speaking different languages?

No, but just having more samples of words should theoretically be better no?

no it doesn't work like that sadly

there is like minimum length that will not work for a model

but 5-10 minutes will work just fine

I’ve had 5-10 minute models and they definitely are not as good as when I got more voice lines.

it depends more on the quality

The quality is basically as original as they can

if you can get 5-10 minutes pure audio with wide range of voice, a little bit of laughs, and like 20-30% singing then the model will come out very good

I remove not useful battle, very shot lines and ones with weird cutscene eq/reverb stuff

my friend Mustar was doing the same thing, harvesting 7 hours datasets to make a model, i was like "<:kittypawbite:1167394009887539200> "

Well depends. Ripped from game lines are essentially as perfect quality as it gets

Might not make a difference if you are ripping lines from other premixed media sources

Because the quality it’s gonna be only so high

It doesn’t take me much more time to rip 5 minutes vs all of it

then i even more don't see the reason to harvest more data than 5 minutes, if there is a way to get a studio quality audio

https://discord.com/channels/1159260121998827560/1164022394218889329

this one was one i uploaded with a difference

the 7m one sounded less accurate than actual voice lines

can't tell which one is which tbh, both sound great, i hear that they have differences, but mostly the same

1.3 one is the 7m one. It did not have her croaky voice

on 0:50 the 1.3 one sounds even better, it's not breathy as the first one and has decent voice timbre

It’s a little less noticeable with the instrumental.

if you gonna use the model only for songs then you can even make 3 minute model for it to work

Ok but, I like more

you do you, i'm just advising, pretrains make it much easier and give possibility to train good models even with short datasets, gl with future models <:AIHC_Heart:1163807521010630758>

no worries. The other 7 genshin models have all been quite nice with most of them being 1hr+. Also like the for fun paimon model with 16h dataset

Oh paimon was never archived by weight.gg it seems

8gb index was a fun hard drive killer https://huggingface.co/Dolyfin/RVC2Models/blob/main/PaimonEN4.0_e110_s295130.zip

daym 16 hours

hmm <:52634615_730845017316044_3880792:1159397351782354974>

i think you can even try to train a model without the pretrain

Oh is that a thing?

because you harvest such enormous amount of data you might not even need the pretrains

Or just v1?

nah

you can try to train one model without any pretrain and see if it'll improve the quality even more

so like changing or removing these?

to do that you just leave these empty

I'll do that one day

Paimon took like. 6 days

Probably more if its from a scratch model

mention me once you do that, very interesting to see what will come out of it <:aismug:1159365471368400948>

Yeah will do

Why does it sound like a gremlin? How do I make it sound like Furina

When its instrumental and vocals and its a little high in pitch furina starts tweakin

Thes signings i gave it is in whoel anotber languege sorry yall

Do you like furinas vocals on the srah models i like the way it goes with her so well

Check for the tweaking when she signs high on 13

English

😃👍⭐️⭐️⭐️⭐️⭐️

The fact that I made her sing Le Festin (from Ratatouille)

Try it with a Zara Larsson song, trust me

What pitch is recommended?

Neutral

The fact I just made her sing Still Dream from Rise of the Guardians (don't mind me of this)

ES TAN HERMOSA

Hey

英文声库？真正的外国人做的是吧（

Hi