✨│ai-help

[proper shale]

I DIDN'T

[snow rose]

Yeah i plan to watch it

[graceful obsidian]

clicks are fixed in there, it's quite op

[proper shale]

Wonder what it can do with Michael's snaps 🤔

[modern oxide]

it should be much more goated for synth datasets

[graceful obsidian]

if to ensemble it with mdx23c, It'd prob kill the stuff

but ye, ft bsroformer is currently world's best isolation

[modern oxide]

so for like any speech with music and sfx over it

[snow rose]

i recommend watching new parasyte the grey, somehow netflix made it good

[red kayak]

how is fm supposed to look like?

i need an answer to that lmao

[graceful obsidian]

for instance this ( 4th prototype so far ) on 6 mins

ft bsroformer sourced

[graceful obsidian]

in what way?

[red kayak]

like on the graphs, how is it supposed to look like, is it supposed to be going down up?

[modern oxide]

i hope anjok adds the ft version in uvr soon

[red kayak]

cuz like mine always went sky high

[modern oxide]

and that mysterious demudder thingy

[graceful obsidian]

Oh, yes, all metrics are meant to be going down

even D in ideal scenario should be fairly balanced, both going down, G and D

[pastel oak]

did someone say a new bsroformer, where can we use it

[graceful obsidian]

with slight G bias

[proper shale]

fm always went up for me too

[modern oxide]

mvsep

[proper shale]

lol

[modern oxide]

and x-minus (i think?)

[red kayak]

yeah i really dunno why

[graceful obsidian]

if it goes up that'd be the balance is out of place

in such cases, try to verify the gradient normalization

check if G is skyrocketting

and D decreasing

[red kayak]

nope ive had it happen when they were both fairly balanced too

[pastel oak]

oh the ver 2024.04 under bs roformer?

[red kayak]

g and d being ideal per se

[modern oxide]

yep

[graceful obsidian]

if you have the log file I can take a look

would tell a bit more if I could see it ye

[red kayak]

i train on kaggle, but i will train a model now so i will ping u here

[graceful obsidian]

Alrighty

[red kayak]

https://cdn.discordapp.com/attachments/1159290096458149938/1215230831597985802/20240307-0930-18.4734275.mp4?ex=662b74e8&is=662a2368&hm=bb13871e5b72b2a0113ac0ee57f35b96a0661b6857805fe27a5bd2d5e757e6fb&

also the label splitting i was talking about

[graceful obsidian]

ye tho, what's the purpose of this step ^ ?

[nova zealot]

Hi there, just saying that the new applio no ui is unusable. First I get an a100 GPU which I never get, then after installing it prompts me to restart the session and I lose the GPU. And after it gives errors of not being able to decode the audio file (ffmpeg error), good waste of a GPU.

[graceful obsidian]

curious

[red kayak]

to export all the audio labels and have it split up already

and it usually helps rvc not mess up as much

[graceful obsidian]

try one to time to do side by side training

rvc split and your method split

curious on how rvc reacts to lack of contextual-overlap

[red kayak]

okay i will train on both peices of audio

and i will let u know

[graceful obsidian]

O7

[red kayak]

just let rvc split on its own right

no truncating and just

such*

[graceful obsidian]

uhhh

if possible ^

but I imagine it to be difficult if you're not doing it from the very start as you build the set

cause truncating doesn't care if it's inter-phoneme or inter-sample gap

[red kayak]

i'll try to replicate that

[graceful obsidian]

tho, if you have quite a bit of data, effect of gaps isn't as devastating really

model's biasing towards silence mainly when presented with smaller sets + higher batches or like, when encountering difficult patterns to learn, then it cheats

[red kayak]

rvc has the threshold set to -40 db and everything is well de noised and noise gated so splitting wont be an issue for it

[graceful obsidian]

that's not the noisegate that's the threshold for what to treat as "cuttable" silence

[red kayak]

yes

that what im saying

[graceful obsidian]

but that thing doesn't quite work as fully intended, tested the set with wider gaps / normal and with small gaps
result is that the wider / normal gaped sets have more silence in there

at the beginning and ending

like, at times 0.3 to 0.6 sec

but if you pack it all tightly, you feed it more contextual voice data rather than silence

[red kayak]

hmm okay then i'll do as u say and compare both

might take me sometime tho

[graceful obsidian]

Take your time and remember, might have worked for the way I process my sets, might not for you

afteral, we all shaped our own methods right 👀

like I was focusing on handling smaller sets or those prone to collapses vs those 15/ 20+ mins

[red kayak]

yush but i still like trying stuff out so yeah

[graceful obsidian]

mainly 5 to 8 mins

yuhhhh

[red kayak]

i legit havent gotten a mode collapse in a hot minute

generator is always okay with it

[graceful obsidian]

yup, bigger sets have less to almost none of such issues ( maybe max 1 or 3 collapses per 2-4 training hours and even that, rarely

sadly that's not so easy for smaller sets

[red kayak]

i did one with 6 mins didnt really get one

[red kayak]

i needa see

i'll also train like 2 mins of audio

with labeling

and i'll also show u those

i use ngrok so ull have access to tboard

[graceful obsidian]

for instance, this mostly happens when doing standard sets with no gap control

oh, this one

so ^

1st one is actually when you pack it tightly

[red kayak]

i see i see

[graceful obsidian]

and while those silence zones might not be an issue for generally bigger sets

for smaller? it's devastating

[red kayak]

is 2 mins enough u think

maybe i should try one

[graceful obsidian]

like so ^

but then when you scale it with right batch ( kinda negate effects of small set

[graceful obsidian]

hmmmm, you can try

but to replicate it, you gotta use batch at 16, 14 or 12

[red kayak]

ohh thats what u want

i was going to use 4

yeah sure

[graceful obsidian]

ohhhh, then that's why

[red kayak]

ill put that to the test

[graceful obsidian]

bigger batch trainings have more data to calculate more consistent kl

's why I don't go below 12 as minimum

also feature matching is more averaged too

[red kayak]

well kl can still be consistent with 8

but we;ll see

i need to test all of those things out

[graceful obsidian]

it's still half less with 16 cause

gt 100 samples

100 / 16 = 6~ groups, each having 16 samples fed to the network at each gradients update
100 / 8 = 12/13~ groups but each having just 8 samples fed

so while there's less groups, each provides more averaged gradients / params update

hence why graphs tend to be less noisy and more uniform

[red kayak]

also how labels look

same audio

[graceful obsidian]

kinda like that

[red kayak]

yes smoothed out

[graceful obsidian]

ohhh

[graceful obsidian]

that too, but too much data lack of convergence, too little, same

is why I recommend doing range-finding trainings:
4, 8, 12, 16
then comparing 2 best sessions

and say, 12 was the most stable that was going down but could be better ( deeper in tendency )
while 8 was really good at going low but too noisy and would sometimes back-track itself
you'd then want to narrow it to 9-10-11 batch

kinda like that but then, it's tedious and not everyon can afford time for that

[red kayak]

powers of 2 are usually best

[graceful obsidian]

speed wise for sure cause memory utilization

[wanton comet]

Hello friends, I need a model of a Russian teenager of 16-17 years old, a help desk

[graceful obsidian]

but while for one 10 min set 16 might work, some other 10 min might benefit from 7 or 13 batch due to more complex patterns and spacial characteristics etc

and needs more randomization ( lower batch ) to more efficiently capture the patterns + compare it with totally unrelated yet close enough ones

[wanton comet]

Hey bro, can you help me?

[graceful obsidian]

well, for models maybe try weights.gg website

you could find something

yet remember, if you need it for real-time voice changer, model being " russian " won't be enough, you also gotta do the accent as well, at least to some degree
if it's for covers / speech generation from input, any good russian model will do, provided you'll use high index ratio - if model allows you to go that high, that is

[clever shadow]

guys when i press save it says file not found what is the problem?

[graceful obsidian]

save? where?

[clever shadow]

when i press upload custom voice

I press save it says file not found

[graceful obsidian]

oh you probs using some custom forks or something (hf applio? iliaria rvc?), in that case I can't help

[proper shale]

im thinking its either W-Okada or voice.ai

[clever shadow]

voice ai

[rare goblet]

Ayo? @clever shadow level 1 !!!

[proper shale]

yeah don't use that

[graceful obsidian]

ohhh right, maybe w-okada

[proper shale]

use w-okada instead

https://rentry.co/VoiceChangerGuide

[graceful obsidian]

w-okada or rvc's native voice changer

[proper shale]

free and open source

[graceful obsidian]

both of em are tons better than that website

[proper shale]

yeah

[graceful obsidian]

native has fcpe tho, so that's a plus

[proper shale]

doesnt W-Okada have that tho? .18a must have added it

[clever shadow]

from where i can download w okada

[graceful obsidian]

oh, then maybe, perhaps

[proper shale]

read the thing, has download links there

[graceful obsidian]

can't tell as I switched from dml w-okada to native

kinda easier on artifacts and speed for me

and that is on pytorch not onnx

oh ye I remember why I switched, cause I use pytorch rather than onnx

kinda can't bite through accuracy issues with onnx so

[clever shadow]

where I didn't find it

[graceful obsidian]

what gpu you have?

nvidia? amd?

[clever shadow]

amd

[graceful obsidian]

older 4/5xx series or

5-6-7xxx

[clever shadow]

what does that mean

[proper shale]

just scroll down for 1 or 2 seconds, there's a download link for Nvidia, intel/amd and mac

[graceful obsidian]

@clever shadow
https://huggingface.co/wok000/vcclient000/blob/main/MMVCServerSIO_win_onnxdirectML-cuda_v.1.5.3.18a.zip

it's this one

[clever shadow]

?

this one

[proper shale]

also, please, read the guide, you definitely need it cuz you'll have to do onnx converting on models

[proper shale]

yeah

[graceful obsidian]

ye that's the one

18a

if you'll have performance issues or stutters despite config testing:

[clever shadow]

amd 5700 xt

[red kayak]

do i really need to set saving freq for this to work

[graceful obsidian]

yes

wait

you weren't doing it?

[red kayak]

i do it

not now though

[graceful obsidian]

👁️👄 👁️

[red kayak]

cuz it isnt a com model

so im just

setting it as 5

so that i can let it bake

and not worry about drive overfilling

[graceful obsidian]

if you wanna replicate my environment so to have fair comparison

I'd set it to 1

[clever shadow]

should i press this

[rare goblet]

Ayo? @clever shadow level 2 !!!

[clever shadow]

?

[graceful obsidian]

cause collapses can occur at any time

[red kayak]

okay

[graceful obsidian]

there's that " download " button

upper left corner

or that you showed too, ye

[clever shadow]

k tysm

[graceful obsidian]

but ye, was saying that, if performance will suck despite testing settings, try rvc's native voice changer too

Might be similar to my case where native won over w-okada

also note: w-okada needs you to convert the model to onnx, rvc's native doesn't
onnx models are a lil less accurate in few things but, you most likely won't notice ( just mentioning it in case

[clever shadow]

from where i can convert it

[proper shale]

thats what the guide shows how to do

[clever shadow]

k

[graceful obsidian]

^

[clever shadow]

sry i didn't read it

xd

ill read it now

[proper shale]

hm... native works out of the box like that for AMD stuff?

[graceful obsidian]

yes

[proper shale]

damn, pretty cool

[graceful obsidian]

there's the

just, it needs repo's " native " libs n such

[clever shadow]

which one is better okada or native

and easier for me

[graceful obsidian]

as my fork doesn't rely on that, gonna check what's needed and diffuse it within my repo

[proper shale]

okada is easier overall

so yeah

[graceful obsidian]

that's how native looks

main pro is that it doesn't require you to use onnx and has fcpe

w-okada's dml release iirc as I have 18a too, doesn't have it

and is full pytorch and for some magical reason lmao, cuts less
funny isn't it?

I just feel like native has few more sliders to play with

like overhead buffer ( unless wokada has it too and I forgot

[floral token]

Still taking an insane amount of time @graceful obsidian . Currently training the second epoch and the first one took over 7 minutes.

[graceful obsidian]

can you quickly go into model's folder

check the config file?

does it say true next to fp16?

@floral token

[floral token]

fp16_run false

[graceful obsidian]

aha there's the issue

yours runs in single precision

that's weird

@floral token a simple one but

did you:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117

this exact one?

or did you:

pip install torch torchvision torchaudio

even some older gpus should support nvidia amp and mixed precision hmmmm

[floral token]

which folder could I check?

can't say I remember which I installed

[vestal sedge]

where can i find an e girl arabicc sound

[vestal sedge]

is that better or the voice changer cilent

[graceful obsidian]

I'd try both to be fair, see which you like better / which works better for you

but if you asked from technical point of view, I'd say native
( but then, I am an accuracy freak so

@floral token

try this

or maybe in cmd:

nvcc --version

can't tell for sure as I don't have nvidia really

[vestal sedge]

should i download it

[graceful obsidian]

oh, nono, that's for charm

oh, my bad lol, tagged wrong person

anyway, if you have nvidia, I'd just probably use w-okada as it's more documented

if it's amd uhhh, maybe native?

just try stuff around

[vestal sedge]

oh okk

[vestal sedge]

wdymm

[graceful obsidian]

it was a msg for an user called charm

[vestal sedge]

oki

[floral token]

I'll try a driver update. Not sure I have the patience to install another client after taking so long to get back to what I was familiar with

[graceful obsidian]

you can keep the client

[floral token]

I had no problems, trained several models a year ago that were great. Now seems completely different

[graceful obsidian]

if it's bugged imma just send you a config

to repaste

where fp16 is forced

[floral token]

ah to just replace my current config file

[rare goblet]

Ayo? @floral token level 3 !!!

[graceful obsidian]

not that one, there's one which switches between fp16 and fp32, it also controls sample length limit based on vram

Oh wait a second

[floral token]

right that's just my model's config file

[graceful obsidian]

when did you download the mainline?

[graceful obsidian]

nope, that's different

[floral token]

set it up yesterday

[graceful obsidian]

models configs are taken actively from other place

there are templates in configs/v2 v1 and ones that are in use in configs/inuse/v2 v1

those from inuse/.... are copied to your model's dir

so if init of rvc forces change in "inuse"

they quite possibly switch the dir's one too, and even if not, the training pipeline is taking info from inuse

tl;dr, if not to hardcode fp16 use in you, you gotta fix the driver ( if it's the case, that is )

[floral token]

okay, my config file under configs folder shows 10/20/2023 modified date

[graceful obsidian]

a

well, just get my fork

def something bugged

you can re-use some files from mainline to not waste net

https://github.com/codename0og/hifi_testing_new

you'll only have to copy from yours to mine these:

and:

pip install -r sox.txt

that's sox resampler for better audio resampling

And if that won't help, means it's the driver / gpu

but hey, you'll get features mainline doesn't have so, it's a win situation lol

Alternatively, you can try this:
hardcoded fp16 for ur rtx

[clever shadow]

guys when i press start i can't hear my voice why

I change the output to my mic and still I can't hear my self

[graceful obsidian]

screenshot of your settings?

I/o section
so, input, output

[rare goblet]

Ayo? @graceful obsidian level 20 !!!

[floral token]

giving it a go now @graceful obsidian thanks so much

[graceful obsidian]

Sure thing man

[graceful obsidian]

take a look at the input output

you have it set to none

[clever shadow]

yes

tysm

[graceful obsidian]

this one's probs not what you need

[clever shadow]

wdym

[graceful obsidian]

Initially, you set it in the wrong section

is what I meant

[clever shadow]

should i remove it

or ill leave it

[graceful obsidian]

remove

[clever shadow]

how

[graceful obsidian]

there's no dropdown menu?

I mean, well, nevermind.
Just ignore it

[clever shadow]

k

[floral token]

this one might be more obvious but can't see what I'm missing @graceful obsidian

[graceful obsidian]

did you paste the .txt in there?

@floral token

[floral token]

sorry, which txt file? @graceful obsidian

[queen coyote]

how am i supposed to change the singer of a song? can yall help

[clever shadow]

I didn't find the mic of voice ai in discord why?

[graceful obsidian]

cause it won't work like that

you need virtual cable or such, but I can't help with that

others know more than me about re-routing audio that way

[graceful obsidian]

you won't need that one unless you have my fork

sox is utilized only in my fork, mainline uses resampy or such

[floral token]

gotcha. I'll try it this way

[graceful obsidian]

the thing you need is uhhh

testing fresh rvc from repo

cause I believe main issue for you are the outdated files, perhaps

tl;dr:

1. test if fp16 works for you ( when you run rvc, console should say something about half precision or fp16 ) using fresh / newest rvc from repo ( not zip / prepackaged versions! )
2. if that doesn't help ^ It must mean the gpu or driver is the fault here

[finite pond]

Is it meant to take like over an hour to try and do japanese vocals?

i was mainly curious as to how it would work

[floral token]

right you said just to replace the hubert_inputs and the rmvpe_inputs only right? @graceful obsidian

[lusty knoll]

hello

[clever shadow]

when i close the app how to open it again from start http?

[lusty knoll]

i get jsondecode error on rvc v2 colab, im using an rvc model, idk what to do

[clever shadow]

and should i leave the cmd or its fine if i close it

[graceful obsidian]

nono look.
rvc repos don't contain models like rmvpe, hubert, pretrains and so on

you gotta get the repo's rvc ( it's not packaged, there's no pretrains and models )

from your older rvc, you gotta copy over what I showed on the ss

so:

[floral token]

I already had all of those files is my confusion

[graceful obsidian]

and pretrained_v2 and v1 have f0G48k, f0D48k, etc, each for their own

when you dl the repo

[floral token]

My folder name is RVC1006Nvidia. Sorry if that's not actually Mainline

[graceful obsidian]

you just get blank rvc

which is what you gotta get

because pre-packaged rvc, namely, rvc1006 and so on

aren't being updated

see ?

[floral token]

yes for sure

[graceful obsidian]

whereas the repo itself, indeed, gets updated

but repo doesn't contain heavier files

namely, models ( hubert, rmvpe etc ) and pretrains

which you get from older rvc full releases, so in your case, 1006

[graceful obsidian]

these

you wanna paste em into repo ( clean ) rvc

[clever shadow]

what should i put this to

[graceful obsidian]

whatever the table showed as working for your gpu

[clever shadow]

this app is using to much cpu

[graceful obsidian]

it should be using gpu tho, not cpu

[clever shadow]

my gpu is rx 5700 xt

[clever shadow]

cpu is 80%

[graceful obsidian]

meaning, you don't use gpu hm

I suppose, restart w-okada

[clever shadow]

I restart it

already

[rustic aurora]

is there overtraining ?

@red kayak

i cant decide it

[clever shadow]

i change cunks and extra to

but I can't hear my self now but cpu and gpu is good

around 20%

[red kayak]

pretty much yeah

it doesnt look too good either

[rustic aurora]

which point very much good

[red kayak]

select the lowest point from here and u should be good

[rustic aurora]

so lowest point

what it say "value" attribute

do you know

what is that mean

for example in the graph lowest point maybe 5k step but 40 value. Other hand, a little bit high point maybe 7k is 35 value

why

Can you enlighten me

[stuck spoke]

hi, I am having trouble with the RVC GUI

downloading any model always fails and I cant figure out why

[rustic aurora]

@red kayak

[rustic aurora]

where do you download

from where

[stuck spoke]

https://colab.research.google.com/drive/1Gj6UTf2gicndUW_tVheVhTXIIYpFTYc7?usp=sharing#scrollTo=OVQoLQJXS7WX

[red kayak]

from ur weights folder

[rustic aurora]

but my colab worktime is stopped

[rustic aurora]

i know

how can i continue

to training

from where i left off

[clever shadow]

can anyone help me here

[low shard]

outdated colab

use only the updated ones in our docs https://docs.aihub.wtf

[stuck spoke]

aaaah

[rare goblet]

Ayo? @stuck spoke level 1 !!!

[stuck spoke]

thanks m8!

[low shard]

yw

[finite pond]

so i'm using huggingface and it normally doesn't take ANYWHERE near this long but its taking literal hours just to do something as little as a 4 minute song

[violet heron]

Did you duplicate the space

[finite pond]

yes

[violet heron]

Screenshot?

Fully scroll up with HuggingFace toolbar

[violet heron]

Did you just create it

[finite pond]

no

[violet heron]

Try restarting it

3 dots, restart this space

If not try Ilaria RVC instead

[finite pond]

I just made a new one

[rare goblet]

Ayo? @finite pond level 2 !!!

[odd shale]

Hello Tangoville

[zinc anchor]

Akitai?

Is he even alive?

Bro just dipped and never came back

[graceful obsidian]

yea it's

supposedly he had plans but something didn't work out and he got depressed even more (?)

after a while, weird statuses happened and I kinda stopped liking his attitude

simply unfriended him a while ago but ye, he was and most likely still is alive

@zinc anchor

[zinc anchor]

I think he was an actual engineer

[red kayak]

he chillin i think dw

[graceful obsidian]

as in?

non machine learning related engineer you mean, right?

[zinc anchor]

Yeah I think so

[graceful obsidian]

hooo, interesting actually

wait what, oh, you meant lusbert, I thought you misspelled akimaya

lmao

[half cove]

Lusbert did left ai hub and now iran blocked discord so rip

I have to talk to bro through google chat

[graceful obsidian]

that means, I didn't know lus was an engineer

[half cove]

He's not

[graceful obsidian]

well, then I am confused

who's akitai

if it ain't bert's another nickname or akiyama miss-spell

man, seems I missed out too much

[pulsar breach]

hi guys! sorry to bother, I just got in, do you know where to find the tutorial on how to make ai covers for begginers? 😦

[graceful obsidian]

uhh, wait

-rvc

[azure marsh]

Documentation

📚 All-In-One English documentation

https://docs.aihub.wtf/

❔ English FAQ

https://docs.aihub.wtf/faq

Credits: AI HUB FRANCE

✨ More guides

Full AI Voice Model Training Guide (Local)

Link: YouTube
Credits: Christopher Villanueva

Model training with Mainline RVC

Link: Rentry
credits: Raven (ravencutie21)

AICoverGen Colab Guide

Link: Google Docs
Credits: Eddy (Spanish Helper)

Create a model with RVC disconnected (colab)

Link: Google Docs
Credits: Angetyde

How To Make an AI Cover With Ilaria RVC

Link: Rentry
Credits: 👽 Julia (ailen2091)

[graceful obsidian]

oh, there ^

[pulsar breach]

thank you so much!

[graceful obsidian]

Sure, and for more advanced things later later, in future
check this:
#✨│ai-help message

[half cove]

Links r old

[graceful obsidian]

oh

[half cove]

We gotta fix that

[graceful obsidian]

fffffffff

[graceful obsidian]

so if links are outdated or whatever then
you just gotta first figure out your gpu

nvidia or amd?

[half cove]

Leme get on pc rq

[graceful obsidian]

orrr ye, you can handle that
I'd perhaps just help with mainline or my fork
not that well aware of applio and other forks' differences

[half cove]

https://docs.aihub.wtf/essentials/how-to-make-ai-cover/

this should link you to the tutorial that's right for you

[pulsar breach]

omg i have no idea 😭 i think this is too advanced for me LMAO my friend send me the link and told me: there, u r gonna find tutorials

[rare goblet]

Ayo? @pulsar breach level 1 !!!

[pulsar breach]

thank youu

[graceful obsidian]

dw, all will be clearer once you get a better handle of it

[pulsar breach]

ur too kind tysm

[graceful obsidian]

suree, mods, helpers, lots of people to help if you'll have issues so dw

[prisma locust]

the google colab ai songs maker is not working what should i do

[violet heron]

How is not working and which colab are you using

[azure marsh]

RVC Guides (How to Make AI Cover)

Documentation

🇺🇸 English (main)

https://docs.aihub.wtf/

​

Translation by country

🇧🇷 Brasil (PT-BR)

https://docs.aihub.wtf/v/brazil

🇪🇸 Spain

https://docs.aihub.wtf/v/spain/

🇫🇷 France

https://docs.aihub.wtf/v/france

🇩🇪 Germany

https://docs.aihub.wtf/v/germany

🇯🇵 Japan

https://docs.aihub.wtf/v/japan

🇰🇷 Korea

https://docs.aihub.wtf/v/korea

🇵🇱 Poland

https://docs.aihub.wtf/v/poland

🇷🇺 Russia

https://docs.aihub.wtf/v/russia

🇮🇹 Italy

https://docs.aihub.wtf/v/italy

[placid talon]

What do the pretrained model variants offer for each one

and which is the "best"

[odd shale]

There's no best pretrain (still)

I would recommend you to just use the original one.

[odd shale]

tangoville

hello jimboville

[placid talon]

what do the others provide tho

[odd shale]

Others allow you to train models with shorter datasets.

But results may vary. (sometimes you can get better results with ones, worse with others)

I would recommend you to use original one if your dataset is near 10 mins and it's decently clean.

It's matter of testing.

There's no best pretrain.

[odd shale]

Use Applio No UI Colab.

Tho make sure to have an instrumental and a clean/dry acapella with no backing vocals or harmonies for inference.

#📰│dev-updates message

[placid talon]

abt 5 mins rn, but this stuff is fairly new so idk how to work them

ive been using original for a bit now still, just was wondering. thanks

[odd shale]

You're welcome.

[frank shuttle]

Does anyome know if it is possible to make a free ai voice model with no dataset?

Like if I wanted to make a voice for a person that does not exist

[odd shale]

Nope.

You can make any kind of model.

But you obviously need a dataset.

The only way to make an original voice is by merging at least 3 similar voices in a dataset.

[frank shuttle]

Can you merge pre existing models?

[odd shale]

Yeah.

[frank shuttle]

Oh that should work

do you know how?

[odd shale]

Nope, but i think you must put the paths of both .pth's of the models you wanna merge

[frank shuttle]

Put them where

[odd shale]

ckpt Processing tab

[graceful obsidian]

#✨│ai-help message

merging voices ^

@frank shuttle

[odd shale]

Hai Code.

[graceful obsidian]

mhm. hai hai ~

[odd shale]

Hawaryu

(How r u)

[graceful obsidian]

• redacted -

[frank shuttle]

wow

I understand

[graceful obsidian]

fyi, am decent enough as long I don't have to talk with certain individuals

[frank shuttle]

the fact that there is a picture

[graceful obsidian]

well, you're meant to read it

[frank shuttle]

yeah but like

there are so many

words

[graceful obsidian]

there's " ckpt fusion " section

[frank shuttle]

Can't I use ckpt processing in Mangio?

[graceful obsidian]

in a tl;dr format, you just want to put 1 model in slot A, other in slot B
each box has to have a full path to the .pth

[frank shuttle]

Ok

[graceful obsidian]

you can

That's universal for any fork

[frank shuttle]

right

[graceful obsidian]

just make sure to have the voices quite similar in timbre otherwise it might sound pretty bad

notes:

• always have v2 checkbox checked ( nowadays probs nobody is doing v1 ) when fusing
• 40 or 48, you gotta match that one box too

[frank shuttle]

My mind feels like a bomb that is about to explode trying to figure this out

[graceful obsidian]

Go ahead, ask
Imma simplify it

[frank shuttle]

Ok

first of all

how can I do this on the cloud

[graceful obsidian]

Depending on which service you use
I haven't been here for a while and don't quite know which are working, which are not
but, generally, when you have rvc or fork running, you wanna put the models ( .pth ones ) into rvc main folder > assets > weights > here

or if it's old-rvc based fork / notebook
then there's just weights folder in the rvc folder ( so, no assets )

[frank shuttle]

Deadass I feel like I am a dummy surrounded by a ton of geniuses

[graceful obsidian]

hmmm, then let's start this way
what service ( cloud ) you use?

Google colab? hugging face? uhhh, what else

[frank shuttle]

Colab

[rare goblet]

Ayo? @frank shuttle level 3 !!!

[graceful obsidian]

so you should have the directory tree view on the left

see it?

[frank shuttle]

uh give me a sec

[graceful obsidian]

will be here when you click the folder icon

[frank shuttle]

oh

ok

[graceful obsidian]

in there you should either see the rvc folder or whatever they call it nowadays for cloak purposes

[frank shuttle]

Do I need to run anything first?

[graceful obsidian]

In a short, yes

[frank shuttle]

And does it work with Ilaria

[graceful obsidian]

Hmmm.. maybe?
I haven't used any of those " newer gen " notebooks

also, I am not sure if ilaria or cover-gen ones have ckpt fusion tab

they're ( iirc ) focused on just generation rather than utilities ( such as ckpt fusing )

[frank shuttle]

right

so probably not

[graceful obsidian]

Do you have pc?

[frank shuttle]

I have an old laptop that does not have much storage left

[graceful obsidian]

can't afford around 50 to 120 ish mb?

[frank shuttle]

that is why I want to hold off on doing it locally unless it is a lass resort

yeah

[graceful obsidian]

Hmmmm.. I could do the fusion locally for you if you had the models and just share it on gdrive

if that's fine with you ofc

[frank shuttle]

I don't even know what models I am using yet

[graceful obsidian]

Then my suggestion is, find any models you'd want to fuse and just lemme know, until you wanna handle it all yourself
but then again, can't promise all forks / notebooks have the ckpt fusion

[frank shuttle]

ok

[graceful obsidian]

but just in case
once you install ( so usually init step in notebooks ) the rvc
you'll have the rvc ( insert potential cloaked name ) folder

in there you should have: assets/weights/ <put your .pth models here>

then you'd go into ckpt fusion tab and insert your dominant ( model you think should be the lead ) model into the slot A and secondary one in the slot B
.
ex.

Model A: path/to/rvc_folder/assets/weights/your_model_A.pth
Model B: path/to/rvc_folder/assets/weights/your_model_B.pth

you'll also have the " weight for model A " slider or something like that, it operates from 0.01 to 1.00 ( 100 )

It roughly means, how much of Model A you want to leave in the A + B mixture ( final hybridized model )

I recommend doing 50/50 ( for slider ), so, setting it at 0.5

for version pick " v2 " ( a checkbox ) and for "40k" or "48k" select ones that correspond with your models
(( whenever you dl a model, they should be properly named or tagged with their samplerate, so, 40khz, 48khz etc ))

the thing about " pitch something ", again a checkbox, ( Sorry I ain't giving full names. Been a while I used something else than my fork and I have my custom namings ), leave it at " yes "

@frank shuttle Hope all's clear so far. Best of luck ~ ✨

[frank shuttle]

So

much

info

[graceful obsidian]

take your time, have a read

[frank shuttle]

lol

thx

[graceful obsidian]

or if you prefer other format, just wait for other folks giving you links to guides or stuff

but ye, yw, gluck

[frank shuttle]

thx

[frosty wolf]

since w okada doesnt work no matter what i do, what is rvcs native whatchmacallit

[graceful obsidian]

Essentially, it's native to rvc realtime voice changer

do you have rvc on your pc?

[frosty wolf]

no

[graceful obsidian]

and do you have time rn to get stuff done n ready or not quite?

[frosty wolf]

i got all day brother

[graceful obsidian]

Alright, in that case

take this

name it rvc or something, put it on ur fastest drive ( C? if ssd )

[frosty wolf]

yeah its C

[graceful obsidian]

ye so, unpack it to C to some folder, name it rvc for convenience

yuh

[frosty wolf]

so like in my downloads or what

[graceful obsidian]

can be just root of c

root, ya know

so it's directly on C:\rvc

[frosty wolf]

yep

done

[graceful obsidian]

alr, now you'll have to download one thing

gimme a sec, gonna confirm which one is it

[frosty wolf]

lay it on me

[graceful obsidian]

https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/hubert_base.pt?download=true
hubert base

you'll put that base.pt in here:

so, rvc folder of yours / assets/ hubert ( the hubert_inputs will be there already, or should at least ~ if not, lemme know

[frosty wolf]

okey dokey

where?

rvc folder?

[graceful obsidian]

oh, uhhh
take stuff from rvc ( one on the ss ) and move it to the main " RVC "

rest you can yeet

just the insides ~

[frosty wolf]

so go into the "rvc" folder and move its contents into the "RVC" folder?

[graceful obsidian]

ye

all for the sake of paths clarity

those can be tricky at times, in some cases

oh wait

[graceful obsidian]

bruh

[frosty wolf]

like that..?

[graceful obsidian]

my mistake, this package

I keep forgetting they added developer section

[frosty wolf]

convenient name

replace files in destination?

[graceful obsidian]

ye

[frosty wolf]

OH SHOOT

quite a bit of files

[graceful obsidian]

ye, that's the whole rvc ( aside some bigger files needed for training )

[frosty wolf]

what do i do with this (if anything)

[rare goblet]

Ayo? @frosty wolf level 3 !!!

[graceful obsidian]

you can yeet

just remember about the hubert .pt thingy

[frosty wolf]

put it in hubert_inputs folder?

thats the only thing in here with "hubert" in its name

[graceful obsidian]

will be here

[frosty wolf]

oh wait im dumb lol

[graceful obsidian]

the dir/ assets / hubert

[frosty wolf]

assets

hubert base is now in hubert folder

[graceful obsidian]

alr, now you gotta wait for my upload of " runtime "

[frosty wolf]

alrighty

[graceful obsidian]

4 more mins or so

[pastel oak]

I get that manual installation is better, but wouldnt it be easier to just send the binary release, paste the dev code inside and call it a day

[graceful obsidian]

it'll be approx 700 mb

[graceful obsidian]

the issue is

their pre-packaged releases aren't aligned with the newest changes and fixes

there's tons of bullcrap done recently and poorly documented too

for instance, recent config rewriting changes and stuff

Hence why I told him to get the newest mainline + imma share tested in field runtime

cause I can't promise the package's ver is compatible with fcpe realtime and stuff

Just for life's ease sake ✨

[frosty wolf]

is this voice changer anywhere close/better/worse than okada

[graceful obsidian]

they have the same principle

main difference ( and reason why I personally switched )

is because I have amd

amd needs onnx

[frosty wolf]

so its roughly the same but different layout?

[graceful obsidian]

and I don't like onnx as compared to pytorch, they're worse in accuracy

[pastel oak]

Thats fine, just think the average user really does not care about allathat and is fine with rmvpe + the old dev version + some wasapi or asio routing

[graceful obsidian]

and native rocks in terms of speed, pytorch and that is on amd

[graceful obsidian]

yeee I am aware
it is more so I can debug it easier

[pastel oak]

its not as bloated as wokada. better performance, and especially better for amd gpus

[graceful obsidian]

if I know what they work with 1:1

[frosty wolf]

hope this one isnt as buggy and weird as the other

[graceful obsidian]

esp if it's files I tested in field / keep updated myself

[frosty wolf]

i got nvidia gtx 1650

on a laptop

[graceful obsidian]

will be fine

I use the voice changer in pytorch "mode" ( nvidia optimised naturally ) on old 4 gig amd, rx 570 👀

so dw

[frosty wolf]

pytorch was a pain in the butt to install

[graceful obsidian]

yea, can imagine

esp for older 1xxx series

[frosty wolf]

hype for this summer so i can get a job and purchase an actually decent pc ⁉️

[graceful obsidian]

haha

https://drive.google.com/file/d/1xY4DM4D-dwqLIzpelZKfYD2CMYu305Ng/view?usp=sharing
alr, here

[frosty wolf]

hooray

[graceful obsidian]

unpack the "runtime" folder from the archive straight to the rvc folder

so, rvc dir/runtime/

[frosty wolf]

👍

[graceful obsidian]

also, a question, you have python installed, right?

[frosty wolf]

yes

[graceful obsidian]

3.10 version ?

or 3.12

[frosty wolf]

lemme look

[graceful obsidian]

in cmd:

py

will print the ver

[frosty wolf]

i did pip --version

and it says 3.12

[graceful obsidian]

hmmm.. well, let's hope it'll work 🤞

lmao

[frosty wolf]

is 3.10 better or something?

i just got the latest

[graceful obsidian]

might be just my luck with people that have python issues but, 3.12 seems to be often times problematic

I'd say, for rvc purposes, 3.10.6 is the most stable and reliable

[frosty wolf]

might have to get

[graceful obsidian]

but let's stay optimistic for 3.12, if that won't budge then we'll look further into it

[frosty wolf]

if this doesnt work i'll grab that up

[graceful obsidian]

ye

[pastel oak]

would try to get 3.10 100%

[graceful obsidian]

my mannnn

3.10 gang

[frosty wolf]

well that'd be a hassle and i'll just wait to see if this doesnt work

holy cow this runtime thing is huge

its taking a lot to extract it

[graceful obsidian]

ye, unfortunately

it's just, kinda own python runtime with libraries

[graceful obsidian]

as much as I was able to get my fork to run without it

[frosty wolf]

57k items?!

[graceful obsidian]

the voice changer not quite

until I figure out what so-darn-custom it has

[graceful obsidian]

ye

scripts, libraries, configs

it also contains stuff one would use for training-ready rvc

[frosty wolf]

hm sounds good to me

[graceful obsidian]

hmm.. might actually isolate the voice changer from rvc in few days

perhaps the needed libs too for smaller dl

[frosty wolf]

i hope this vc is good lol im going through a lot for this

spent all day trying to fix okada

(to no avail)

[graceful obsidian]

everything is better than something that's broken afteral

it has few extras too

[frosty wolf]

facts

oo

[graceful obsidian]

for instance, pre-render buffering

[frosty wolf]

haha.... whats that

[graceful obsidian]

sure, adds delay but can enhance stability

you kinda, infer ( generate ) speech in " memory "

which naturally causes delay, but as you have some overhead, yea

[frosty wolf]

memory?

ram?

[graceful obsidian]

ye, ram

that's the memory
disk I would just call storage

[frosty wolf]

so it uses my ram to generate the voices?

sorry im confused

[graceful obsidian]

nono
in simple words

rather than have it literally realtime ( as long you want to use that feature ofc )
you have it delay it by whatever you set, say, 2 seconds

those 2 seconds it first fully generates the speech

kinda yt buffering

[frosty wolf]

OHH

so like better speech, but more delay

i see

[graceful obsidian]

where buffered part is what you hear

and what buffers is your speech realtime

can help for weaker hardware to keep stability

etc

[frosty wolf]

nice 👍

its still going

[rare goblet]

Ayo? @frosty wolf level 4 !!!

[graceful obsidian]

yea.. and with that, I can safely say 7zip's superior to winrar

in terms of compression algos

lol

[frosty wolf]

46%

damn this is a lot

[graceful obsidian]

yea.. all for those few or so libraries that are needed but can't be isolated just yet 🤡

[edgy jolt]

where do i put the voice on the app

[graceful obsidian]

rvc locally on your pc, right?

[edgy jolt]

umm

whats rvc

[graceful obsidian]

as in, on your pc and not some notebooks, colabs et-

ah

then what app you talk about?

🤔

[edgy jolt]

voice.ai

[pastel oak]

heads up people dont know that wokada is wokada so they ask in here thinking rvc stands for real voice changer

[graceful obsidian]

oh, well, this is a channel for rvc so

oh, lmao fr?

[pastel oak]

We do not offer support for voice.ai on this server and would recommend not to use it

[graceful obsidian]

^

voice.ai is trash

and rvc means retrieval voice conversion not really realtime voice conversion

[edgy jolt]

oh ok sorry

i got to know this server by voice.ai

on a tiktok

lma

[graceful obsidian]

much better if you do it locally on your own pc if you have a chance / or possibility - hardware wise

[frosty wolf]

i tried voice.ai but its essentially pay-to-win

[graceful obsidian]

ye, and trashy

[frosty wolf]

yeah its pretty dog water

[edgy jolt]

yep

[graceful obsidian]

only legit " rvc " like services that are paid I can recommend is really elevenlabs

they're just gods in terms of style-replication for speech generation / tts, but ye clears throat

[frosty wolf]

STILL GOING

[graceful obsidian]

mada madaaaa

[frosty wolf]

81% man this is taking far longer than i thought

[frank shuttle]

I am trying to figure out what singers sound like Ariana Grande

[graceful obsidian]

you tell me

would probs be 1-3 mins on my end

[frosty wolf]

perhaps i should switch off of winrar

[graceful obsidian]

can't

and shouldn't

[frosty wolf]

no?

[graceful obsidian]

well, I just wouldn't

[frosty wolf]

okay

to each their own

[graceful obsidian]

or if you wanna be risking the biscuit 👀

[frosty wolf]

ITS DONEEE

[graceful obsidian]

yo, there we go

now, part 2

[frosty wolf]

oh boy

[graceful obsidian]

you gotta install the reclist

this one:

requirements-win-for-realtime_vc_gui

pro tip

cmd in the address bar

and viola. No need of cd or contextual menus or whatever

[frosty wolf]

this?

[graceful obsidian]

yes

pip install -r requirements-win-for-realtime_vc_gui.txt

[frosty wolf]

alrighty

where should i paste that

[graceful obsidian]

do this ( then hit enter ofc )

and paste in the cmd / console

[frosty wolf]

am i dumb

[graceful obsidian]

that's uhh

did you

[graceful obsidian]

in the rvc's folder?

like shown on the ss

[pastel oak]

@jovial pollen you might wanna read up on this conversation since you wanted to try out the new rvc update. good luck catching up lol

[frosty wolf]

the cmd folder??

im so lost

i gotta be slow

[graceful obsidian]

when you're in rvc folder, you have the address bar above

you wanna click there and just type in cmd and only that, then enter

[frosty wolf]

ah thanks

[graceful obsidian]

it'll open the cmd which is already looking in the rvc folder ye

[frosty wolf]

got it

[graceful obsidian]

neat

[graceful obsidian]

now

[frosty wolf]

paste > enter

[graceful obsidian]

ye

[frosty wolf]

thanks for being patient with my stupid brain

anyway its installing

[graceful obsidian]

's fine. we all start somewhere

anddddd, back in the day I'd help so many people at once, daily to the point I decided to quit some server lol or was it going perma-offline-status?

don't remember

[frosty wolf]

youre cool

[graceful obsidian]

you too, you actually take time in doing things

but some ( pardon whoever reads this ) really impatient tiktok-era kids have difficulties typing in 1 command properly

[frosty wolf]

when i wanna do something im gonna do it thats all

[graceful obsidian]

good attitude ye

[rare goblet]

Ayo? @graceful obsidian level 21 !!!

[graceful obsidian]

oh, right

[frosty wolf]

this isnt looking good

[graceful obsidian]

oh shit

that one's a pain in the ass uhhh

[frosty wolf]

oops doxed my name

eh idc

what do i do

[graceful obsidian]

hold on, lemme think

[frosty wolf]

what could have possibly have gone wrong 😭

[graceful obsidian]

uhhhh, well ye, wanna hear bad or good news first?

[frosty wolf]

alright lets hear the worst of it

[graceful obsidian]

you'll have to get c++ build tools to get one specific thingy for the rest to push through ( building wheel as shown in the log, which failed )
and that is a pain in the ass. Finding it, downloading ( quite heavy in size )

all that ^ to use cuda itself

[frosty wolf]

alright as long as you know how i can do that

good news?

[odd shale]

Search it on the internet.

[graceful obsidian]

Good news is, you can use it in DML mode for now

change the reclist to this one: requirements-win-for-realtime_vc_gui-dml

[frosty wolf]

how good/bad is that compared to if i got c++

[graceful obsidian]

pip install -r requirements-win-for-realtime_vc_gui-dml

well, tbh? probably similar or same performance

given your gpu

[frosty wolf]

alright cool

[graceful obsidian]

but in any case, if you were to ultimately decide to try cuda as well
that'd have to be tomorrow

[odd shale]

He should install C++ Build tools?

https://visualstudio.microsoft.com/es/visual-cpp-build-tools/

Bc i found it.

[frosty wolf]

rename the text file to this?

[graceful obsidian]

pip install -r requirements-win-for-realtime_vc_gui-dml.txt

nah, just this ^

[frosty wolf]

alright

[graceful obsidian]

as there's few reclists in rvc folder

[frosty wolf]

installing

[graceful obsidian]

this one is for dml - direct ml
difference between direct ml and cuda is that direct ml is for all devices that support directx 12

cuda is for nvidia only

[frosty wolf]

oh...

...oh

yikes

[graceful obsidian]

wut

[graceful obsidian]

huh, maybe we can just ignore the pyworld

you'll use fcpe anyway

[frosty wolf]

uh alright

not sure what that means

[odd shale]

Fcpe is optimized for realtime right?

[graceful obsidian]

yes

[odd shale]

Nice, ig he can use that.

[graceful obsidian]

torch / crepe / rmvpe like but faster

ye, it's the go2

[graceful obsidian]

Alright so, we can actually use cuda if pyworld is all it is about

edit the first reclist

one without -dml

[frosty wolf]

okay

[graceful obsidian]

it's a text file

[rare goblet]

Ayo? @frosty wolf level 5 !!!

[frosty wolf]

so open that sucker up?

[graceful obsidian]

you gonna yeet this

in a notepad

[frosty wolf]

hey my name is purple my favorite

[graceful obsidian]

then save and ye

pip install -r requirements-win-for-realtime_vc_gui.txt

[frosty wolf]

so find that line and remove it

[graceful obsidian]

should do the trick if pyworld's the messy thingy

yuh

[frosty wolf]

okay

[graceful obsidian]

oh, nevermind

[frosty wolf]

okay im installing after backspacing pyworld out of the requirements-win-for-realtime_vc_gui.txt

what??

[graceful obsidian]

let's see how it goes

[frosty wolf]

okay

huh.

why cant it just work properly

[graceful obsidian]

ah yes, so build tools that is

[frosty wolf]

so i gotta cop that

[odd shale]

You'll must download Build Tools.

[graceful obsidian]

ye, guess can't be helped

[frosty wolf]

alright...

[odd shale]

https://visualstudio.microsoft.com/es/visual-cpp-build-tools/

[odd shale]

Alright code, i gave him the Build Tools link.

Guide him thru the installation.

🙂

[graceful obsidian]

ok ✨

[frosty wolf]

im guessing free download or this is the wrong page

[graceful obsidian]

https://visualstudio.microsoft.com/es/visual-cpp-build-tools/

this one

big ✨ descargar ✨ lol

[frosty wolf]

i hate spanish class

[graceful obsidian]

stuff I learned about spanish within the last 5-7 years?

[odd shale]

Descargar means Download hermano.

[graceful obsidian]

xxxxx descargar gratis.
kek

[frosty wolf]

i assumed

[odd shale]

XD

[graceful obsidian]

¯_(ツ)_/¯

[odd shale]

Download free no virus 100% real no password

[frosty wolf]

installing visual studio shite

[odd shale]

Are you sure you don't wanna apply for helper?

(Bc that's what you're doing)

[graceful obsidian]

I feel more as audio / model engineer or specialist than helper

aaaand I only help when I feel like it or am bored ( like rn

[frosty wolf]

you sure are HELPING me out a lot

[odd shale]

Then apply for Engineer role.

[graceful obsidian]

individual components, I believe

[graceful obsidian]

nah, doesn't feel special

what can I say, I like feeling unique ✨

[frosty wolf]

what do i select

[graceful obsidian]

search for visual c++ 14

[frosty wolf]

alright

[graceful obsidian]

or anything that resembles it from the name

[graceful obsidian]

don't remember. Last time I got that crap was like a year ago or so

hmmmmmm

try 2014

or 14.0
or visual redist
or visual basic

[frosty wolf]

ok

you mean 2015?

[graceful obsidian]

perhaps?
it's for sure version 14.0

and c++

so I guess ye, visual c v140

msvc v140

[frosty wolf]

this?

[graceful obsidian]

ye

[frosty wolf]

alright

if this doesnt work i swear 😭

[graceful obsidian]

after you done, just re-add pyworld to the reclist (( standard one, not the dml one

( positioning doesn't matter, can be at the end

[frosty wolf]

okay

im more determined to make this work than i am for school

probably not the best thing

[graceful obsidian]

🗿

[frosty wolf]

real 💀

[cosmic shadow]

hi, I have problems with the voice change, it doesn't work, the voice doesn't sound, only my normal voice in the passthru

[odd shale]

🐢 👍

[frosty wolf]

it would seem its complete

dml one or nah

[graceful obsidian]

without

[frosty wolf]

okay

command

[graceful obsidian]

pip install -r requirements-win-for-realtime_vc_gui.txt

[frosty wolf]

ty

[graceful obsidian]

there

btw, the format is really simple for installing reclists ( which are, requirement lists )

[frosty wolf]

alright

please work for the love of god 🙏

[graceful obsidian]

pip - packages manager
install - cause you install
-r means " requirements / reclist - you get it"
then you provide the list.txt

[frosty wolf]

there should be absolutely no reason to fail, unless i got the wrong c++ or smth