Ov2Super (RMVPE) A High-Quality fine-tune of the original V2 pretrain

[iron abyss]

Main Link: https://huggingface.co/ORVC/Ov2Super
Fix for mainline and mangio forks Made by the infamous Poopmaster: https://huggingface.co/poiqazwsx/Ov2Super32kfix/tree/main

First ever made community pretrain which support 32k and 40k sample rates, trained on a precisely chosen clean speech and singing dataset, with bright and emotional voices.

Models made with Ov2Super:
#1229087854244921344 message
#1228600059180875837 message
#1227672326209011842 message
#1219337845471187024 message
#1218543866618318869 message
#1217467273246146560 message
#1213068920219766824 message

Ov2.1Super coming soon, which will support all sample rates and will have even better quality.

[nocturne kayak]

best pretrain on the market yall

[iron abyss]

lmao, thanks xd

tho Ov2.1Super... It'll be waaaaay better

[worthy vale]

Is this the same pretrains from the google disconnected or a different one? Now being super??

@iron abyss

[iron abyss]

yep

[weak phoenix]

Titan

Titan and OV2, no other ones.

[nocturne shell]

Both are only for English datasets right?

[weak phoenix]

Nop.

It should work on almost any language.

[nocturne shell]

Ohh

I see

Thanks

[nocturne kayak]

https://tenor.com/view/well-agree-disagree-gif-19083826

[weak phoenix]

☝️

🐢 : XD

[toxic swallow]

if i see one more post involving pre-train drama you're all banned

🤣

nsrs

[weak phoenix]

👍

pretrain station?

[glass tinsel]

👀 I'm going to wait for it

[iron abyss]

dataset 7 minutes, pretty noisy, 6 batch size, same settings, ORVC made with ov2, the other one with regular one, model by @ivory wolf

[rapid mist]

help this is only 100 epoch and I just pretrained it...and it's kinda fine!!! damm...

[zinc ore]

TITAN and OV2 besto pretrains, you can't change my mind

[weak phoenix]

Titan.

[zinc ore]

Titan and Ov2.

[weak phoenix]

Titan and OV2

[zinc ore]

Yeah, both are rly good pretrains

[worthy vale]

titan better tho

titan coming wit a v2 soon

with a larger dataset

[nocturne shell]

when will titan large be released

[weak phoenix]

No ETA still.

[nocturne shell]

Okeyy

[zinc ore]

yo @iron abyss any ETA of Ov2.1?

[iron abyss]

Cancelled project

[zinc ore]

[iron abyss]

@barren thistle

[barren thistle]

Oh yeah that is definitely better

How much data was it trained on? the pretrain

[zinc ore]

bad thing about the pretrain is that it dosen't support 48k

and idk what to do with that pretrain if im going to use it with a 48k dataset

[iron abyss]

48k is useless imo, cus rvc's gan just can't reach that pure sound of hq audio. hifigan is pretty limiting and creates too many artifacts that just make 48k simply not that much better than 40k

[zinc ore]

if my dataset reaches 24 khz cut-off?

[iron abyss]

my recommendation is to just go with 32k or 40k, even if you're using og pretrains

[sleek trench]

yo is ov2.1 out

[zinc ore]

ov2.1 is private

[sleek trench]

oh, is titan large out?

[zinc ore]

not yet

or blaise gave up

[sleek trench]

those the best pretrains like ov2 for 5 min and under and titan for over 5 min but they starting to feel outdated ykwim

[zinc ore]

original pretrain is better

[sleek trench]

I feel like it's a lil robotic

well on smaller datasets

[sleek trench]

range wise ye

[iron abyss]

It is not, it's trained on 50 hours of emotionless speech, while titan and ov2 are fine tunes that are trained on emotional dataset with wide variety of voices, which suits general use case better than original pretrains

[zinc ore]

titan has low end noise and distorted harmonics, ov2 has too much mode collapse, while original pretrain has no issues, i've trained a lot of models with original pretrain and they're not bad

i don't want to make a long ass discussion of why original pretrain is better, i'll just say that original pretrain is the most stable rvc pretrain. simple as that.

[sleek trench]

u tryna train some data I trained with titan and compare

[iron abyss]

I trained over 400 models trying multiple pretrains and know what I'm talking about

[zinc ore]

no, it's details i've noticed from other users' opinions and spectogram viewing

[sleek trench]

ohh

[iron abyss]

It's too sensitive to noise, has a very bleak understanding of bright and compressed voices, which make training take longer and voices sounding worse

[zinc ore]

i'm not glazing the og pretrain but original pretrain at least doesn't have mode collapse and distorted harmonics unlike ov2

[iron abyss]

Mode collapses don't hurt the model

[zinc ore]

oh my god

mode collapse RUINS your model

[sleek trench]

I guess it all comes down to personal preference and what you prioritize in ur model

[iron abyss]

In fact it doesn't affect the model in any way, we discussed that with litsa

[zinc ore]

if you choose mode collapse weights of your model, it will sound bad

[zinc ore]

oh really?

so mode collapse it's just visual or what

[iron abyss]

Yeah, what it does is it basically skips an epoch

[zinc ore]

ok. i'll repeat this LAST TIME. original pretrains are stable and have no issues unlike other custom pretrains.

goodbye, thank you for participating everyone

[smoky lintel]

several of my models made in early 2024 used ov2 and didn't notice such distorted harmonics

[sleek trench]

when is ov2.1 not going to be private

and will it be effective on larger datasets

[iron abyss]

Yeah no, they're stable at making noisy models

That's it

[weak phoenix]

This discussion again?

[iron abyss]

Any fine tune with a clean dataset is better than og pretrains

[zinc ore]

you silly billy hilichurl, this is because the og pretrain was trained on a vctk dataset which is SUPER NOISY. i'd tell you more but i'm sick of it already.

[iron abyss]

That's why I recommend to use titan, cus it has a very clean dataset that is almost as long as the original pretrains

[zinc ore]

yea, john ai hub is kinda dumb tbh

john doesn't accept that og pretrains are more stable than other custom pretrains

[iron abyss]

Yes, then why the fuck are you arguing that it's better, if it's fucking noisy

[weak phoenix]

Watch your tone Mantrax.

[zinc ore]

yea but i'm feeling kinda mad.

my apologies for the insults.

[weak phoenix]

Entonces porque sigues discutiendo con ranita Simpl?

[sleek trench]

this is kinda funny to watch lmao

[zinc ore]

motherfucker, let's stop this shit for once

[zinc ore]

leo and i are getting sick of it

please stop.

[weak phoenix]

It was necessary to use insults?

[zinc ore]

sorry. when i'm mad i can't control myself sometimes.

[zinc ore]

tho "sick of it" isn't a insult

it's like getting tired of x thing

you alr know

[sleek trench]

i feel like it just depends on your preference, both ov2 and the og pretrain have their pros and cons

[weak phoenix]

I'm not referring to that.

[zinc ore]

oh.

[weak phoenix]

I was referring to "mf" word.

[zinc ore]

john was also using insults.

[iron abyss]

The only one getting sick of it is you, I was testing and making pretrains for over a year, and clearly understand how they work more than you do, I won't sit here and just agree with ridiculous claims

[zinc ore]

i don't give a... well i don't care about this anymore

[iron abyss]

It works good for everything, not sure if I'll make it public

[weak phoenix]

Song name?

[sleek trench]

okay bc im rlly interested in trying it out if ov2 is that good as it is

[zinc ore]

it's on the image, ig

[smoky lintel]

https://open.spotify.com/track/5bU4KX47KqtDKKaLM4QCzh?si=d94598eab29543d1

[weak phoenix]

i got no spoti account. 😭

[sleek trench]

ov2 is my go to when i use a smaller dataset

[smoky lintel]

you can use a spotify downloader

[zinc ore]

song is pretty neat

(hearing on yt, no spotify cuz i got yt music)

[zinc ore]

search the song on yt :3

[iron abyss]

Ov2.1 works surprisingly well with low frequencies, and overall gives better and cleaner results than any pretrains I've tested so far, but eh, not sure if I'll make it publicly available, probably will keep it to myself until I come up with something even better (if I will, cus I'm kind of tired of testing and testing)

[sleek trench]

okay sounds good, im excited to see where u go with it

[weak phoenix]

🐢: eats a lettuce

[sleek trench]

did you know an artist with 4 million monthly listeners uses a model trauined with ov2 in his songs

https://open.spotify.com/artist/0lZ2oIRdMsQfFEPmOa3YyL?si=4JxpvTWJRy-kxU9PWeiF3g

[iron abyss]

Lmao

Nice

Can't open Spotify though cus it's blocked in Russia

Maybe he has yt?

[sleek trench]

yeah

[iron abyss]

Link pls (or name) 🙏

[sleek trench]

https://www.youtube.com/watch?v=Lk7tMF5pTSw&list=OLAK5uy_lAhbwC6W8hoQ7CMVlgAOKawngdTYp-au0

he uses it in most of his songs

[iron abyss]

How do you know that?

[sleek trench]

u can hear its the same model plus i know the guy

[iron abyss]

Nice, glad the pretrain found a good use

[sleek trench]

hes starting to slowly use his real voice in his music and move away from ai

[sleek trench]

fr

[sleek trench]

you should train ov2 with refinegan