sleepygirl 32k snowieV3

egirl voice inflation

I'm not gonna lie blyuv, your last 3 models have sounded like the exact same person lol

🤣 🤣 🤣

sounds like merged voices to me

yeahhh

i have so many female voices at this point

i can just merge them and make infinite

i just dont post them bc people dont credit anymore

or try reselling them

i merge different ones, idk why they sound the same at the end

but i merge and then train them at rvc

if you merge similar voices will sound kinda the same

rvc has too many limits honestly i think is cause of that

if you don't have a merge folder with hundreds of different merges are you really a realtime enjoyer

let's be honest here

this is a 3 merged models i made and then remake them on rvc again with a inference and new dataset

but i think the rvc has too many pitch tones limites that they sound the same

ah that's why

you should just post the merge

instead of making a new model of the inference of the merge

cuz it sounds worse that way

yeah i know

but im not giving like good models or mergeds xD

so i just post what i recycle

like this one

ah, fair enough

i heard the gonna release a new " good " pretrain

But like, what kinda voices are you merging. Like 3 egirl models together, or are you merging different kinda voices?

this one is 3 differents vtubers

yeahhhh

that explains it

it sounds good on singing at least owo

merging gives the best dividends when you're merging voices that are pretty distinct from one another, but then again I am assuming what kinda vtubers you're using as data xD

haha to be honest i just pick random good quality vtuber voices

dont even know the names i just listen to the sound of the voices lol

lmao

based

and pick 3 or 4 make dataset make model and merge them

the lazy part is just cleaning the audio

yeah it's pretty tedious

Titan, but it's still just a general pretrain so it's not going to be huge or anything. ~~~~I am finetuning a pretrain for English female voices in particular~~

like idk how it's going to turn out, but my hope is that it gives better results for choosing a specialty for that data.

rvc makes me feel so fustrated ;-; i hope they release a new tech soon

what's wrong with it

i mean its good but once you get used to it, you realized it has a lot of limitations

play around with diff-svc

and not good quality parts

way better for singing

and overall vocal range

tbh i just use it for realtime

i dont care about the singing or inference lol

realtime has its limitations

since rvc models are trained on voice

any non voice sounds will fuck up

like coughing, laughing, sneeze and more

what limitations are you talking about in terms of realtime?

prob those i mentioned

the pitch and tone is not the same as the dataset

also tone change

like whispering

Well I can manage laughing with realtime, but not like full on belly laughs

more like giggles

when you train a model it doesnt detected all the sound of the voice even if you boost all the dynamics

dont expect the model to translate emotions if you talk act girly

some guys use it talking like a boy

it becomes just a guy with a girl voice

i know, im not saying is bad but is not the same as dataset

its never the same as dataset

of course

it wont ever be 1:1

yeah, that is a problem

yeah thats why i say it has limitations

we call that... a tomboy

i wouldnt say its a technology limitation

it does what is supposed to do

it doesnt convert a TTS to a human voice

nor someone with no emotion to something with emotion

But like, I only use merges anyways so I don't really care about it lining up with the source audio 100%

yeah im not talking about the emotion part

as long as it sounds like I want it to after the fact

but yes the pitch tone and quality

its never the same as the dataset

like if you clone your own voice

it wont sound the same

xD

similar and accurate yes

but not the same

rcv gives you its best guess after you train

basically lol

and to get back to this, the solution is not to act like something you're not, but to sculpt a voice towards your particular speaking style 🤔

by merging different types of speakers

into a ratio that just vibes with your speech patterns

~~unless you think every woman is a bubbly valley girl lmao~~

yeah i mean all im saying is even if you clone your own voice it wont sound the same, not the talking part and the voice acting, i say the pitch and tone it wouldnever be the same, that is the limitation im refering too

~~ I was talking to MrModz~~

I get what'cha mean.

101 comments damn XD

You should try training all the datasets together instead of merging

gives better results imo

it'd be super tedious to test different ratios of voices that way

yea

but better sounding results

example?

https://discord.com/channels/1159260121998827560/1216461330127523910

thats a model combining 3 voices in the training phase

i tried that too but idk why i get better results with merge maybe my voice or idk

on inference sounds really good , but on realtime is other story

well do the voices you combine sound similar?

i dont remember, but i know i did that experiment xD

yeah I don't think it's better than merges 100%

Like you can prob get similar results with both methods

but it's like way more of a pain in the ass having to train a new model each time

i agree but just from what ive done training them has gotten me better results

oh yeah blyuv, have you messed around with using plugins on your models to make them sound more realistic

like what plugins

I see you in the voice chat here alot so I know you use real time quite a bit, but idk if you're interested in making it sound more like a real mic

i tried to use some vst

but idk why i prefer more virgin audio xD

Equalization on the voice to quite the robotic parts of the voice, and make other parts louder, convulations on your model can imitate the sound of a bad mic more, and like bit crushers to make it sound less unnaturally clear

yeah that's fair enough, I just use them on the model while I'm playing games or whatever

How would i go about doing this?

makes it so no-one can tell its ai

Well first you need something like Elgato wave-link or voice meter to apply plugins to your virtual cable

I have voice meter

also mares do you know how to change samplerate on wokada on client side?

i have my windows on 44khz my mic at 44khz and everything on 44khz but when i output a audio on okada its 48k

idk, I don't use okada

mine output to 44khz just fine on go_realtime_gui.bat

Well you get all these plugins then

I usually save models from others under "Model (by Author)"

and for the EQ what you really wanna do is basically soften the spectrum on the sides here, since a lot of the unnatural clearness on voices I have noticed comes from like, 50khz to 200khz.

do i need to install another thing thats not voice meter? bec there is no plugins menu that i can find. unless im dumb

I think you need voice meter banana

i do have that

But if you have an elgato mic just use wave link imo

Yeah I don't use that, so idk. I just know other people have used plugins on it.

i dont

But yeah here's an example of what it sounds like with all of those plugins on a voice.

I usually don't have the bit crushing on it from cymatics origin, but I had that on too just for examples sake.

Like most voice things compress your voice enough where you prob don't need the bitcrushing with a good voice

Tbh you should make some guides about how to clean datasets.

Huh?

I don't have any great secrets, I only train models on data that already has zero background noise lol

I mean, a updated guide about dataset cleanup and plugin usage for cleanup.

🐢 👍

teamongus

And specifically these plugins were for real time, but you're right people could probably apply some of them to inferenced audio