✨│ai-help

[leaden yacht]

well

not that new

but I'm kinda new to training the models

[low shard]

still seems training good

[fading summit]

so i dont know the ins and out o thse things

[leaden yacht]

Is that because its a not so extreme curve up?

[low shard]

yea its not that extreme, let it train more so you can see more if its training or overtraining

[leaden yacht]

alright

btw one quick question

I js saw that it reached about 20k, multiplied so I couldve chosen 40k, but I chose 32k

is that bad?

and does it have a big effect on the models quality

[fading summit]

i just realized when it asked to put a preview of the model i put a vocal of a regular song without the ai

[low shard]

well, it would have been better to choose 40k, as it should have a bit more of quality being an higher rate so that contains more info

[leaden yacht]

ppl will think the model is so good 💀

[leaden yacht]

Should I retrain at 40k if the model isn't good enough in the end?

[low shard]

you could try yea

[leaden yacht]

On the page this is mentioned

but these are my graphs

so is it overtraining?

[low shard]

yea that could be

at 6k of loss/g/total

[leaden yacht]

I already thought that too

This is the exact point there

Do I js go back and get the closest one I can find?

this one?

[stark kettle]

where can i download this voice changer?

[leaden yacht]

I used this one https://github.com/w-okada/voice-changer

dont know if there are any others

[stark kettle]

ty

[leaden yacht]

yw

[leaden yacht]

@low shard

[low shard]

try the s6000 as it seems more closer

[low shard]

-rt

[azure marsh]

This interaction has expired, use the command /guides realtime if you wish to see it again.

[low shard]

first guide

also for wokada ask in #🔍│help-w-okada or #1192011222023950368

[stark kettle]

oke, ty

[leaden yacht]

Did that

It sounds a bit breathy tbh

[graceful obsidian]

you should also look at few other metrics

mel, fm, kl and total g

just send me the tensor file if you want

Imma tell you where to look

[stark kettle]

breh, how to download on github?..

[graceful obsidian]

and highlight few zones to pay attention to

[leaden yacht]

g total was shown on there

[graceful obsidian]

I need the file, screens won't do

[leaden yacht]

how do I do that?

[graceful obsidian]

your model's dir should have a file that's .0 in extension

or either way, is named tfevents

[leaden yacht]

the thing in the logs folder?

this?

[graceful obsidian]

if it's in your model's dir, then ye

[leaden yacht]

mhm it is

[graceful obsidian]

then that's the one

[leaden yacht]

its too big for dms, ill send quickly here and del

js say it when u downloaded

@graceful obsidian

did u get it?

[graceful obsidian]

nope

[leaden yacht]

its sending

[graceful obsidian]

alr, let's see

[leaden yacht]

js download and tell me once ur done

got it?

[graceful obsidian]

I did

[leaden yacht]

alr

The lowest point it had was 5.934k steps

[graceful obsidian]

no idea why you'd be so wary of tfevents files

bot anyway

[leaden yacht]

in g/total

[leaden yacht]

Forgot who, I think shad? No idea but someone asked us to if we share files, delete them as soon as the other gets them

[graceful obsidian]

I mean yea but that's just graphs

nobody will do anything useful with em lmao

[leaden yacht]

hold on why did you tell me to use 0 smoothing

[graceful obsidian]

For your visual ease rn I use smoothing

now, orange you avoid

green tells you something's already going on

as there's a rise, downfall and rise + flatlining

[leaden yacht]

so rise, down and then another rise and flatlining is probably overtraining?

[graceful obsidian]

here's in other form

another thing

[leaden yacht]

but where did it start overtraining?

[graceful obsidian]

FM doesn't really improve over the training

[leaden yacht]

FM?

[graceful obsidian]

esp past 2k

now gonna do side by side metrics to show you what to be warned of

[leaden yacht]

???

[graceful obsidian]

🤦‍♂️

did you actually take your time in reading of whatever I send you

[clear snow]

what is epoches?

[leaden yacht]

I already told you I didn't understand most of it

[graceful obsidian]

it has the literal explanations

of what they are

you can't get it any easier than that, explanation wise so, please have a read

there's just no other way, even if you might not like walls of text

@leaden yacht here

I just don't quite know the hell's with fm

maybe mangio has reversed fm calculation? up vs down? but that'd be kinda pointless

might be it's the set itself that fm doesn't really improve but degenerates

[leaden yacht]

is the fm bad then or smth?

[graceful obsidian]

cause the loss is increasing

rather than going down ( bigger picture wise )

[leaden yacht]

ah

btw I might have found a solution to a huge problem

https://github.com/1AdityaX/RealtimeOvertrainingDetection

[graceful obsidian]

I was always against automating of things that require human evaluation

sooo can't really say much about it

if it's doing conclusions and comparing graphs, using logic, I trust my brain

not machine

[leaden yacht]

but if it goes up, down, up and then stays flat its overtraining?

[graceful obsidian]

It's not universal, it's usually that way
but in reality, overtraining has a lot of shapes and forms

[wispy burrow]

If that is the same thing Applio uses, it's pretty bad...

[graceful obsidian]

can be G flat then down
D up n stable

G up and D down, stable

can be idk.. thunder like shape for D and flatlining above avg loss
G would be stuck in flatline

lots of possibilities

I can type out ckpts I'd consider if I was you

[leaden yacht]

sure

but if its alright I really wanna know where it started to overtrain

[graceful obsidian]

past 7k or rather, inbetween 6-7k

blue and past that

[leaden yacht]

I saw that in here too,

[graceful obsidian]

yea

[leaden yacht]

js before 7k mark

[graceful obsidian]

that's right

[leaden yacht]

so I used the s6750 model

[graceful obsidian]

now, you say 6750

[leaden yacht]

But in the dataset the talking was too breathy I think

[graceful obsidian]

and here is an example of why log interval matters

you have no metrics for step 6750 logged

closest to that is the one on the ss

but that specific " epoch " doesn't exist

it can be anything inbetween " inside the neural network "

so you won't ever know what the s6750 is, loss wise

[leaden yacht]

this is what I have btw

[graceful obsidian]

aha

not only that, you save every 10th epoch

that's even worse for accurate models

[leaden yacht]

I know

Someone gave me the advice to do so

[graceful obsidian]

You see, epoch, say, 10th could be hella different

to 11th

like, entirely different

and smaller the batch_size, more extreme the difference potential

[leaden yacht]

ohhh

[graceful obsidian]

because again, higher batch_size = more samples shown to ai at each epoch

[leaden yacht]

so I should also turn the batch size back to 8 instead of 12?

[graceful obsidian]

so that means

more " avg " results and estimation

[graceful obsidian]

that'd worsen it

[leaden yacht]

or the opposite

ah

[graceful obsidian]

I'd stick to 12 if you already have it working

save every single epoch

sync the log

[leaden yacht]

damn, every single one?

[graceful obsidian]

but then, mangio won't allow for the third

I mean yea, every single one

scenario where epoch 68 is the best one and you save every, idk, 19th

[leaden yacht]

btw I am planning to do that with another dataset then, cuz this one is pretty bad tbh

[graceful obsidian]

what u gonna do?

[leaden yacht]

too much breathy talking

[leaden yacht]

Then I have to backtrack too much

which could change it a lot

ohhh

[graceful obsidian]

well, then differently

good would be 12 epoch, but you save every 8th

you get epoch 8 epoch 16 epoch 24

but there's no 12th

and there is no way to get it back / extract it from G/D files

[leaden yacht]

alright

So I need to save more

[graceful obsidian]

hold up, first gonna recommend you the ckpt to try

[leaden yacht]

But I don't wanna save THAT often, can I js use 5?

[graceful obsidian]

s5934

[leaden yacht]

youre gonna recommend me I the ckpt to try?

[graceful obsidian]

that's sadly not gonna help

[leaden yacht]

thats the lowest point

yeah

[graceful obsidian]

ye, that has the best fm, that is, if you had the actual epoch

which you don't

[leaden yacht]

so saving every epoch is pretty smart then yeah

[graceful obsidian]

my bad, and the best mel would be

5221k is the best mel

in normal conditions, I'd fuse such models

2 of em. Mel and Fm

[leaden yacht]

the best fm? I thought it was only on the g/total thingy as the lowest point

Wait is higher or lower better for the fm?

[graceful obsidian]

nope, it's aligned with fm in this case

fm with G

( doesn't always occur, where metrics align

total G is a sumup so, decent mel but best FM

which you then merge with best Mel

[leaden yacht]

but which one do I follow? total g or fm?

I think total g then right?

[graceful obsidian]

tbf. FM

total G = Fm, Mel, KL

Going by Kl is almost never a good idea as sadly, those are the most tricky

so you have Fm and Mel left

[leaden yacht]

Mel is the pitch I read

[graceful obsidian]

Mel at best total G isn't the best one

it's fidelity / clarity ish

[leaden yacht]

thats what I read at least

[leaden yacht]

ah alright

[graceful obsidian]

Mel is really mel spectrogram metric

such as these

this is a spectrogram in mel scaling

you can easily test the thing by doing an inference on best FM and best Mel separately

and notice the subtle and extreme changes in clarity

[leaden yacht]

what point was best mel?

[graceful obsidian]

then pay attention to " fullness of voice " and accuracy, with fm

5221k

[leaden yacht]

best fm was 5934 which I dont have tho

[graceful obsidian]

yea, that's unfortunately

log interval and saving frequency } issues

[leaden yacht]

so should I use 6000 and 5250?

[graceful obsidian]

you can try but it won't be the same

those are just different totally unrelated epochs

[leaden yacht]

rip

[rare goblet]

Ayo? @leaden yacht level 11 !!!

[graceful obsidian]

;/

that's how it is

[leaden yacht]

alright ill do it with another model later on

but

question

[graceful obsidian]

yea?

[leaden yacht]

this is a 40k sample rate right?

[graceful obsidian]

mostly 18khz

some peaks at 20

you multiple it times 2

[leaden yacht]

mostly but ur supposed to pick the highest right?

[graceful obsidian]

reason why it shows the frequencies divided in two is due to " Nyquist-Theorem "

if you're curious on that, do some research I think

[graceful obsidian]

well, by the avg response from there

you can safely assume 32 to 36 maybe 37khz

[leaden yacht]

so I should always pick the one most seen and multiply that times 2?

so if its 36k I should use 32k right

[graceful obsidian]

here's ur avg freq response range

[leaden yacht]

mhm

[graceful obsidian]

in specs like that, you take the khz and double it

so, total 24 becomes 48

total 22.05 becomes 44.1

[leaden yacht]

I knowwww

[graceful obsidian]

but the file / meta sample rate =/= the fresponse range

[leaden yacht]

but do I pick the highest point? or the one most seen?

[graceful obsidian]

avg

you estimate the avg response, not really the samplerate on its own

[leaden yacht]

is there an app that can calculate that for me? Spek only shows it a bit

[graceful obsidian]

let's put it that way

press P

few times in spek to change colors

pick those most vibrant for your case

[graceful obsidian]

for instance, this one

file is 44khz but the response is kinda around 40

there

see the artifacts?

and color vanishing

[leaden yacht]

where do u see the response?

[leaden yacht]

mhm

[graceful obsidian]

this is the response

from 0 hz to the max of your file

[leaden yacht]

This is like 20.1 or smth then?

[graceful obsidian]

lower than that

[exotic fractal]

I have a ryzen 7 5700x, should it still be running slow?

[graceful obsidian]

because you saw the peaks reaching 20khz or so

yet that's just peaks / some parts

the " body " itself is mostly around 17-18khz

[leaden yacht]

but the peaks are the max

[leaden yacht]

your words

[graceful obsidian]

they are max, yeah, but they are not an avg

[leaden yacht]

I know

[leaden yacht]

but you js said the response is 0 to max

[graceful obsidian]

all in all, you'll still do 48khz training
or either resample the audio to 32khz

[graceful obsidian]

you asked about what is the response

and so I told you

response in literal way

and I talk about avg frequency response

[leaden yacht]

ohhh

[graceful obsidian]

which in your file floats around 17-18khz

I'll tell you more but that in a bit, going shop

[leaden yacht]

so the average shown of the whole thing is the avg freq response?

[graceful obsidian]

red

is the avg

[leaden yacht]

so is this 32k?

[graceful obsidian]

that is just info valuable for you

not quite ai

it tells you which pretrain you should rather go for

orrrr to which samplerate you should resample it

brb

[leaden yacht]

alr

[leaden yacht]

The avg for me is about I think 17kHz, times 2 so I think I need to use 32k sample rate

[frosty wolf]

hey shad i dmed you some stuff

[floral token]

Why sometimes when inferring does the pitch of the model go up? I'll refresh and then it'll work fine

[graceful obsidian]

yuh

seems like it

[frosty wolf]

@graceful obsidian good news and bad news..!

good news - rvc opened

bad news...

• audios still fucked
• crashed when i clicked "start audio conversion"

yeah okay my pc is just incapable of running any of these voice changers

only explanation

[sand iris]

https://tenor.com/view/masive-skill-issue-skill-issue-yakudza-gif-14004060172467418456

embed failure

[frosty wolf]

message still got across

:(

im kidding idc

im just bummed out that it doesnt work

[sand iris]

what did it crash with though

[frosty wolf]

just stopped responding

[sand iris]

surely there's something in logs

[frosty wolf]

also when opened, my headphones' audio bugs out big-time and any audio that i try to play thru them is super loud and distorted

[wind cove]

whats a good live voice changer

[graceful obsidian]

well shit

[wind cove]

which start am I supposed to use 😭

theres 3 bruh

also they all just open in python and the read me files are in Japanese

[rare goblet]

Ayo? @wind cove level 1 !!!

[wind cove]

nvm ill figure it out

[wintry fjord]

---

JSONDecodeError Traceback (most recent call last)
<ipython-input-25-929c96c23701> in <cell line: 31>()
31 if os.path.exists(config_path):
32 # File exists, proceed with creation of creds and client
---> 33 creds = Credentials.from_service_account_file(config_path, scopes=scope)
34 client = gspread.authorize(creds)
35 else:

5 frames
/usr/lib/python3.10/json/decoder.py in raw_decode(self, s, idx)
353 obj, end = self.scan_once(s, idx)
354 except StopIteration as err:
--> 355 raise JSONDecodeError("Expecting value", s, err.value) from None
356 return obj, end

JSONDecodeError: Expecting value: line 1 column 1 (char 0)

it just shows this error in [ https://colab.research.google.com/drive/1Gj6UTf2gicndUW_tVheVhTXIIYpFTYc7?usp=sharing#scrollTo=OVQoLQJXS7WX ] second step no matter what i put in there

[pastel oak]

Outdated colab. Type -colab in this channel to get a list of the new ones

[wintry fjord]

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• easyGUI, by rejects Google Colab
• 🆕 UVR5 NO UI for Google Colab, by Eddy Google Colab

How to use?

🤗 Hugginface Spaces

• Applio, by IA Hispano Huggingface Spaces
• Ilaria RVC, by thestingerx Hugginface Spaces
• RVC-HFv2, by r3gm Huggingface Spaces
• AICoverGen, by r3gm Huggingface Spaces
• Advanced RVC Inference, by r3gm Huggingface Spaces
• RVC v2 Huggingface version, by Clebersla Huggingface Spaces

More commands: -audio, -local, -overtrain, -realtime, -rvc, -help

[exotic cedar]

Is there any way to convert a 40k model to 48k for merging purposes?

[brittle wing]

how to make an rcv

[frosty wolf]

good news for now (bad news is yet to be tested)

i did something wrong with installation

and am in the process of trying again

PERHAPS it will work

im not gonna get my hopes up

[graceful obsidian]

Let's hope for the best

will check up on here later

got some gaming session with the boys rn

[frosty wolf]

W

[frosty wolf]

its not going well

i think i gotta wait until i can get a new pc

cause this rinky dink laptop is temporary

[pastel oak]

we got the voice changer working but somethings up with his sound devices ngl

i mean u could have a look at it code if you wanted to but ive never experienced this before

think bigshot will just get a new pc and do it on there

[midnight fern]

How do i make a ai model ios

on phone

[pastel oak]

i think rvc disconnected works on mobile in general
https://docs.aihub.wtf/rvc/cloud/rvc-disconnected/

[pastel oak]

@graceful obsidian sum frequencies above 20khz, would you personally train 48 or 40. want some of your thought process

[odd shale]

I would highly suggest 40k, not 48k.

Also, codename is rn playing videogames with his friends.

[nimble flax]

what are the best settings for the rvc voice changer ? it keeps sounding so fake

[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

[leaden yacht]

You should ask that in #🔍│help-w-okada , and I think you should set the wait time/delay (ms) to a higher number, it will take a little longer but will give you a better voice output

That usually makes the voice sound a lot better

[graceful obsidian]

40, ye, like Leo mentioned

if it's just " slightly " above 40, then it's not worth the bad convergence risk

if it was like 42~ maybe 43khz
or 41 with dominating " bright sibilant " zones reaching sub 44-46khz ( ye, such cases can happen ) then I'd go for 48

tl;dr:

• tragic / some game-sourced or heavily compressed audio of which freq. response avg. at 28 to 35khz = train 32k models
  ( sometimes 48k works too. For instance, in visual-novel-sourced audio )

• yt / aac / ac3 / opus / vorbis sourced audio - 40k models ( or in really terrible cases, go for 32k models )

• 44.1 files ( some uvr / isolation cases ) where freq. response range averages at 38-41khz = train 40k

• 44.1 files ( lossy codecs; opus or some vorbis ) where freq. response range averages at 41-43 ( or plain 44.1 ) khz = train 48k

• 48 files ( lossy codecs; opus or some vorbis ) where freq. response range averages at 41 to straight 48 = train 48k

• lossless pure 48khz and 44.1khz files? = train 48k models

[thin stump]

@peak tusk I got a list index error when trying to use the Titan pretrain on RVC disconnected. I'm assuming it's not compatible for 32k sample rate yet?

kk, but ov2 still hasn't been fixed either

with the same list index error

No worries, I can wait a bit longer if it's not an easy fix #✦│chat message

[pallid thicket]

Is it possible to change the sample rate of a model?

[fading lodge]

so when im using the voice changer, everything works fine but untill someone is super loud it playsback and sounds so wierd, anyone help?

[pastel oak]

your mic picks up your headphones

if you use wokada move s.treshold further to the right, maybe enable sup2

[fading lodge]

my thresh is at 0.001

[woeful spruce]

Hi, i tried searching around a bit before asking this since it could have been asked already, but i didnt find something that could help me.

Is there a collab or offline rvc gui that could let me select multiple audio files? i have a lot of audio files which i want a model to use, but with all the stuff i've already used its all just, select an audio file, convert, and so on

[fading lodge]

i tried the sup2 thing already, dosent work

[pastel oak]

then lower your headphones volume or move it further away from your mic

or add noise suppression to your mic so it doesnt pick up everything

not much else to do

[pastel oak]

im pretty sure you can dump a folder path and it does everything. not sure which one but it exists

[woeful spruce]

mmm... do you remember if it was one listed in the -colab command?

[brittle wing]

hi what is the comand for search ai models

[nimble field]

what rvc you recommend that can be used on google colab?

i mean, rvc gui

[woeful spruce]

Applio has a batch convert, will see if it works

[brittle wing]

why the search application doesnt work?

[rare goblet]

Ayo? @brittle wing level 1 !!!

[brittle wing]

it says The application did not respond

Whenever I run go-realtime-gui I get this error:

C:\Users\Toby\Desktop\RVC\RVC1006Nvidia>runtime\python.exe guiv1.py
Traceback (most recent call last):
File "C:\Users\Toby\Desktop\RVC\RVC1006Nvidia\gui_v1.py", line 59, in
import torch
File "C:\Users\Toby\Desktop\RVC\RVC1006Nvidia\runtime\lib\site-packages\torch_init.py", line 122, in
raise err
OSError: [WinError 1114] A dynamic link library (DLL) initialization routine failed. Error loading "C:\Users\Toby\Desktop\RVC\RVC1006Nvidia\runtime\lib\site-packages\torch\lib\cublas64_11.dll" or one of its dependencies.

I have CUDA 11.8 installed with Pytorch 2.3.0. I've added environmental variables for CUDA bin and libnvvp. I've also tried reinstalling my drivers, Python and Cuda.

[fading lodge]

it worked ty

[brittle wing]

Hi, how can i use the voice changer in google colab?

[violet heron]

Real time or covers

[brittle wing]

real time i guess I just want to congratulate a friend on her birthday

[brittle wing]

.

[violet heron]

-realtime

[azure marsh]

This interaction has expired, use the command /guides realtime if you wish to see it again.

[brittle wing]

ty so much

[magic lance]

guys wich one ?

og u cant do pics

Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'dio': improved speech but slower extraction; 'harvest': better quality but slower extraction):

wich one do i choose: pm, harvest, dio, crepe, mangio crepe, rmvpe

[junior halo]

For 9 minutes of data and a batch size of 8, how many epochs would you guys recommend?

[brittle wing]

im lost

if anyone see's this how do i do step 2?

i run it with the correct link and it loads for a second

then some code appears for a second then goes away

then i get a red ! beside the button

@odd shale can you help?

[rare goblet]

Ayo? @brittle wing level 1 !!!

[finite galleon]

Thats very old and outdated

[finite galleon]

99.99% of the time rmvpe will do the job the best

Just use rmvpe and if that doesnt give good results try mangio crepe and if that doesnt give good results either its probably an audio input or model issue

[violet heron]

Your using RVC-GUI which is very outdated

[magic lance]

its my first time trying to make a model

[rare goblet]

Ayo? @magic lance level 1 !!!

[finite galleon]

Rvc gui doesnt have rmvpe tho

[violet heron]

Download original RVC or a fork like applio and use RVMPE

[magic lance]

i was following a yt guide

[finite galleon]

Yt guides are always outdated

[magic lance]

its mangio rvc

is that bad ?

[violet heron]

You can’t make models using RVC-GUI

[finite galleon]

Just follow this guide https://docs.aihub.wtf/

[violet heron]

You downloaded a outdated version then

Original RVC > Mangio RVC

[magic lance]

is applio better then ?

[finite galleon]

Yeah mangio havent updated since last year

[finite galleon]

Yep

[magic lance]

ill try it then

thank you <3

[finite galleon]

yw

[magic lance]

where do i click on the github page to dopwnload, im lost T-T

nvm sorry i got it SORRY

the what ?

[queen kraken]

are there any other RVC besides mangio rvc that support mangio crepe?

[brittle wing]

Whenever I run go-realtime-gui I get this error:

C:\Users\Toby\Desktop\RVC\RVC1006Nvidia>runtime\python.exe guiv1.py
Traceback (most recent call last):
File "C:\Users\Toby\Desktop\RVC\RVC1006Nvidia\gui_v1.py", line 59, in
import torch
File "C:\Users\Toby\Desktop\RVC\RVC1006Nvidia\runtime\lib\site-packages\torch_init.py", line 122, in
raise err
OSError: [WinError 1114] A dynamic link library (DLL) initialization routine failed. Error loading "C:\Users\Toby\Desktop\RVC\RVC1006Nvidia\runtime\lib\site-packages\torch\lib\cublas64_11.dll" or one of its dependencies.

I have CUDA 11.8 installed with Pytorch 2.3.0. I've added environmental variables for CUDA bin and libnvvp. I've also tried reinstalling my drivers, Python and Cuda.

[fading lodge]

how can i find the right settings for my voice

[magic lance]

its taking soo long to extract features

[fading lodge]

i can help if u dont know how to setup

[torpid wasp]

Why there's no public URL in easy GUI? Only local

[rare goblet]

Ayo? @torpid wasp level 1 !!!

[desert niche]

I need help with gpt sovits but Idk what's the right channel to get help for it

oh wait nvm I got it working now

[magic lance]

guys i got me a error on the pitch extraction algorithm

please help

[brittle wing]

Whenever I run go-realtime-gui I get this error:

C:\Users\Toby\Desktop\RVC\RVC1006Nvidia>runtime\python.exe guiv1.py
Traceback (most recent call last):
File "C:\Users\Toby\Desktop\RVC\RVC1006Nvidia\gui_v1.py", line 59, in
import torch
File "C:\Users\Toby\Desktop\RVC\RVC1006Nvidia\runtime\lib\site-packages\torch_init.py", line 122, in
raise err
OSError: [WinError 1114] A dynamic link library (DLL) initialization routine failed. Error loading "C:\Users\Toby\Desktop\RVC\RVC1006Nvidia\runtime\lib\site-packages\torch\lib\cublas64_11.dll" or one of its dependencies.

I have CUDA 11.8 installed with Pytorch 2.3.0. I've added environmental variables for CUDA bin and libnvvp. I've also tried reinstalling my drivers, Python and Cuda.

anywho

[brittle wing]

You have to follow that statement up with a contribution to the conversation

i was asking someone to follow me over here

from genereal

general*

Oh okay

But if you can help I would love that

cuz idk whats wrong

Did you ask ChatGPT?

no

Try that

[final palm]

dude sends his problem twice then tells someone to use chat-gpt to solve an issue

[brittle wing]

lol

[brittle wing]

Ok and?

[final palm]

https://tenor.com/view/ok-and-ok-and-whole-sqaud-gif-22248092

[brittle wing]

I have a Bachelors degree in Computer Science and you're a literal child

[final palm]

aw lame it didn’t embed

[final palm]

then figure this shit out yourself instead of asking other literal children😭

like you don’t gotta flex a degree on a 16 year old 💀

[brittle wing]

Whatever you say little man

[final palm]

nah you got it bro 🙏

[junior halo]

Help!! I am attempting to use the Hina Mod Google Colab and it's not giving me a gradio URL. It's still loading. It only gave me the local URL.

What do I use now?

Ilaria is doing the same thing. :(

Is Gradio down???

Could you elaborate?

[violet heron]

Tells person to ask ChatGPT to solve a problem for a program that it 1. Wouldn’t know about since GPT only knows January 22 and older, and 2. Would most likely get wrong since GPT likes to pull things out of its robotic-ass.

I’m no nerd but your degree sure is showing 😅

[brittle wing]

OK retard

Didn't ask for your opinion

[rare goblet]

Ayo? @brittle wing level 3 !!!

[proven hill]

no, they are june 2023

[brittle wing]

OK whatever

That still doesn't fix my issue

[hot lagoon]

is it me or that is olny showing Running on local URL no pubile url in colab

[blissful scaffold]

i dont get any public url either

came in to ask

[hot lagoon]

and what did they say?

[rare goblet]

Ayo? @brittle wing level 9 !!!

[hot lagoon]

there is no ngrok token what you mean

is both links hold on

https://colab.research.google.com/github/iahispano/applio/blob/master/assets/Applio.ipynb

https://colab.research.google.com/drive/16LkwvFZeudTpUOsE_6bMjOq2qkxFo8Hr?usp=sharing

[blissful scaffold]

you dont need token for 2 hours

wdym

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• easyGUI, by rejects Google Colab
• 🆕 UVR5 NO UI for Google Colab, by Eddy Google Colab

How to use?

🤗 Hugginface Spaces

• Applio, by IA Hispano Huggingface Spaces
• Ilaria RVC, by thestingerx Hugginface Spaces
• RVC-HFv2, by r3gm Huggingface Spaces
• AICoverGen, by r3gm Huggingface Spaces
• Advanced RVC Inference, by r3gm Huggingface Spaces
• RVC v2 Huggingface version, by Clebersla Huggingface Spaces

More commands: -audio, -local, -overtrain, -realtime, -rvc, -help

[noble compass]

someone can help me solving my problem?

i cant send any screenshot

btw

[nocturne mural]

possibly it has to do with an internal problem of the gradio

[hot lagoon]

where is this at?

[nocturne mural]

https://status.gradio.app

[blissful scaffold]

what does that mean

is it like

just wait untill it works again?

[nocturne mural]

yes...

[hot lagoon]

how long as it been

[blissful scaffold]

okay thanks

is it hard using applio without ui?

[nocturne mural]

I don't know, maybe less than 24 hours?

[nocturne mural]

Not so much, just that it will take a while to start because I forgot to put the precompiled

I think I'll put it up later

[blissful scaffold]

you made it?

[hot lagoon]

How to use the applio without ui for infer cuz is say model_name but my useing a huging face link to dowanld a modle

[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

[velvet latch]

im trying to train a model but this is happening anyone know how to fix this ?

[nocturne mural]

yes...

[thorn abyss]

try installing dependencies or rvc again

[nocturne mural]

if you don't know the name of the model when you download it, look in the colab files in the applio/logs/ folder.

[velvet latch]

alr ill try reinstalling it

[stable violet]

can i get help

i cant send

photo

[junior halo]

It's reassuring to see I am not the only one with that issue regarding Gradio.

[stable violet]

no witch chat is for the voice mod thingy

bc im having a problem

[dusty lava]

no public url fix?

[thorn abyss]

did it worked?

[torpid wasp]

Can i use easyGUI again now?

[dawn nebula]

we dont use voicemod here

https://rentry.co/VoiceChangerGuide

[barren vault]

is this better than the rvc i downloaded few months ago

[dawn nebula]

okada is based on rvc, it only has better ui

i think

[twin pollen]

is there more detailed training guide?

[hot veldt]

can someone help me? im having trouble with a link to a voice model, it saying there was an error with the zip

[turbid meteor]

where do i download rvc again? the one with directML for amd gpu

[dawn nebula]

send me the link

[dawn nebula]

https://docs.aihub.wtf/

[hot veldt]

https://huggingface.co/TJKAI/Wonka-Timothee-Chalamet/resolve/main/Wonka-Timothee.zip?download=true its this one

[dawn nebula]

which repo u using

[lavish sand]

How to make AI Cover with google colab ? It still not working

[clever geode]

i need help

i want to clean audio, like remove static noise of microphone from background

so how do i do it using UVR?

is it possible to do it with it?

[gusty kestrel]

@loud flint, I have found 2 results that match your search!

sonic2022_mario

Sonic (Jason Griffith And Roger Craig Smith) , Voicemy.ai Trained

• Uploaded: <t:1697372369:d>
• Server: Unknown
• Likes: 0

Epochs

300

Technology

RVC

Algorithm

Rmvpe

Requested by rings2006wilson · Apr 27, 2024, 8:18 AM

An error occurred

Sorry, I could not find models that match your search "junichi kanemaru"

An error occurred

Sorry, I could not find models that match your search "sonic (junichi kanemaru"

[pastel oak]

the long explanation:

#✨│ai-help message

I cant answer anything detailed on that. You can wait for codename and ask him questions yourself

Dont gorget to check the drive link

forget*

[graceful obsidian]

Kl is just a part of it

in reality, you almost never pick ckpts based on em

I'd prioritize FM

then add Mel to it

and if there's a situation you have few " best " fms or few " best " mels

pick ones that are relatively lowest kl wise

pretty much

alternatively, you can fuse Fm+Mel hybrids with best kls ( say in 75:25 proportions )

sometimes

total is a mix of all 3

but issue of total G ckpts is that they are an avg best

they don't prioritize accuracy nor fidelity

so I use total G as sorta indicators to where more or less I focus on graph

to find other good-metric ckpts

I can help with graph but, do you have it synced

log interval matching epochs' steps?

I meant like, did you manually input the steps into config file

during 2nd training run ( actual one

Here's why it matters:
#✨│ai-help message

that part of convo

oh

for that check the #✨│ai-help message

last section

oh, wrong one

thought they linked this one:
#✨│ai-help message

here's all you'd need

ye, last one's about syncing, pretty easy to do

Yup, all the times you train models

Just, yeah. You'd have to train again

current graphs of yours have different logging points

those needed were never logged and are lost permanently

Unfortunately, yea

what's the set size?

dataset
but info on batch_size used appreciated too

so 14-18min ~ zone

can you go as high as 16 for batches?

or not quite

using colab or something?

happened every time or just once

musta been a bug or some freezing

cause they grant T4s for acceleration

they can handle even 35-45 mins at batch 16 ( tested

but ye, if not 16 then at least 14 I'd try

but I'd rather go for 16 cause it's faster than 14, 12, 17, 20 etc

[rare goblet]

Ayo? @graceful obsidian level 24 !!!

[graceful obsidian]

and in ur case, most likely better for ur set

what ver you using of rvc?

mainline, applio or

neat

nah, don't have to yeet all

just those files mentioned on the ss

epochs ( in weights ) and
in model's directory ( in logs ) G, D, tfevents file, eval folder, train log file and config file

but before that

check your model's steps

first epoch in weights folder

what's the "s" value?

you'd check it after doing a test run on 16 batches

that's fine, you'd have to do a test run anyways

hold up

Pretty much:

1. You yeet all that's related to your unsynced models ( what you have rn ) - you keep only the dataset used in first step in the ui ( preprocessing )

2. You start the training as you normally would;

• preprocessing
• feature extraction
• index training
• model training; save every single epoch ( saving freq at 1 ) batch 16 ( in your case - recommended ) as for checkboxes: yes, no, yes

you train until you get 1 epoch

then stop

as a step 7

8. you note the steps value

.pth models in weights and whole model folder from logs

ye, that's all there is to the model

it's experiment folder ( in logs ) and it's ckpts so, .pth weights

[graceful obsidian]

will be in: rvc folder / assets / weights / yourmodel_e1_s10.pth for instance

( hence why you train til first epoch )

ye

if lost just follow this ^

that's the exact same procedure

Np man, after you done with training just @ me

gonna help with the tensorboard

Nope

🗿

250+ models / prototypes and rvc mechanics-studying be like

meaning, you can trust me ¯_(ツ)_/¯

cause all info is tested in field, not guesses

[edgy fjord]

what colab can i use on phone ?

[low shard]

check our docs https://docs.aihub.wtf you can find them here
also check #📰│dev-updates message cus there could be some issues rn

[edgy fjord]

Ok thx

Could not create share link. Please check your internet connection or our status page
do I have to wait for Share API to be online again?

[rare goblet]

Ayo? @edgy fjord level 1 !!!

[graceful obsidian]

I've no idea honestly
Not quite handy with Ngrok or such tunnelings ( neither I use em

Unfortunately nope @brittle wing

Personally don't use colab or kaggle

Saturn. Have a fully working notebook and fork for that but ya see

they're whitelisting people / account registrations in a way

and the free hours per month are even worse now

but generally, you won't even make acc sadly, atm

[torpid wasp]

Uh mind if i ask how do i move the file "frpc_linux_amd64_v0.2" using Android device?

EasyGUI won't give public URL or there's something i didn't know

[low shard]

You can't

Its a problem of gradio api, its down

Read #📰│dev-updates

[torpid wasp]

Ohh so i have to wait for it to be fix?

Sorry

[low shard]

Wait for it to be back up again

Dww

[torpid wasp]

How do i know if it's fixed? Announcement?

[low shard]

On #📰│dev-updates they will say when its fixed, btw in it they also putted a link to check gradio status to see if it's down or not

[torpid wasp]

Thanks!

[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

[nimble pawn]

@graceful obsidian

hey u free atm?

[graceful obsidian]

sup?

gonna be eating and then processing my dataset and wut, need something?

[nimble pawn]

yee

[nimble pawn]

just wondering if we can hop in vc rq to optimize some settings

[opal lagoon]

JSONDecodeError Traceback (most recent call last)
<ipython-input-11-8114ca135b8c> in <cell line: 31>()
31 if os.path.exists(config_path):
32 # File exists, proceed with creation of creds and client
---> 33 creds = Credentials.from_service_account_file(config_path, scopes=scope)
34 client = gspread.authorize(creds)
35 else:

5 frames
/usr/lib/python3.10/json/decoder.py in raw_decode(self, s, idx)
353 obj, end = self.scan_once(s, idx)
354 except StopIteration as err:
--> 355 raise JSONDecodeError("Expecting value", s, err.value) from None
356 return obj, end

what is this? when I try to paste the link in google collab and click on start

this happen

[pastel oak]

outdated colab, type -colab in this channel to get the newer ones

[opal lagoon]

thanks

[rare goblet]

Ayo? @opal lagoon level 1 !!!

[opal lagoon]

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• easyGUI, by rejects Google Colab
• 🆕 UVR5 NO UI for Google Colab, by Eddy Google Colab

How to use?

🤗 Hugginface Spaces

• Applio, by IA Hispano Huggingface Spaces
• Ilaria RVC, by thestingerx Hugginface Spaces
• RVC-HFv2, by r3gm Huggingface Spaces
• AICoverGen, by r3gm Huggingface Spaces
• Advanced RVC Inference, by r3gm Huggingface Spaces
• RVC v2 Huggingface version, by Clebersla Huggingface Spaces

More commands: -audio, -local, -overtrain, -realtime, -rvc, -help

[molten iris]

Hello why does my rvc use my cpu instead of my gpu even though I installed the amd version pls?

[pastel oak]

RVC or wokada

[leaden yacht]

Has this model trained properly?

[pastel oak]

I assume you mean wokada cause it has inconsistent cpu usage

• Make sure you selected your gpu. Open task manager - performances to check for the number

• Make sure you exported your voice models to onnx. Thsi is a MUST for wokada amd users. You upload voice model, then theres a button to export to onnx, then you reupload that onnx as PTH voice model file

• dont move s.threshold all the way to the left or right, that causes 100% cpu usage

[pastel oak]

definitely finished at around 18k for the best average

[leaden yacht]

Yeah I thought so too

[pastel oak]

or wait maybe even 7k

[leaden yacht]

I can safely delete every save above 20k right?

[pastel oak]

you should try both

[leaden yacht]

I saved every epoch, as recommended by code which gives me the perfect spots

[leaden yacht]

u mean the deepest point?

Yeah I was gonna try that one too

[pastel oak]

ye

[leaden yacht]

although at that moment other things like mel were still pretty high

[molten iris]

rvc

[leaden yacht]

is it safe to delete every epoch above 20k?

[pastel oak]

send screenshot and tell me your gpu name

[pastel oak]

thats wokada not rvc

[molten iris]

mb

[pastel oak]

dont do s.thrshold 0.001 move it bit to the left

anything else is fine

amd wokada just sucks, your cpu usage should be lile 30-60%. its normal

if i was you id download the RVC voice changer instead. You want links? it runs better, less delay, better performance

[nimble pawn]

^i agree

[leaden yacht]

Which one do I pick? or do I need to try them all

(Thats the lowest point at 18k)

[molten iris]

ok thank you but is it normal that my gpu works less than my cpu?

[pastel oak]

for amd wokada, yes its normal

[leaden yacht]

can I have the link?

[pastel oak]

youre nvidia right

[leaden yacht]

yeah

4060

[molten iris]

no can I have it?

[pastel oak]

https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/RVC1006AMD_Intel.7z

Download this

Then I recommend you download the (old) dev version as theres a quick trick to reduce some more delay. To do this, heres a quick step by step:
https://huggingface.co/Shadicti/rvc-old-dev/tree/main

Open the zip, open the first folder called Retrieval-based-Voice-Conversion-WebUI-dev and drag all the files into your RVC folder. If it asks you to replace any files, say yes. If it does not ask you to replace any files, you probably dragged the folder Retrieval-based-Voice-Conversion-WebUI-dev in it, but you should drag the 8 folders and 26 data files directly to the RVC1006 folder.

Then you run go-realtime-gui-dml.bat to start the voice changer

On there, make sure to change the default index thats there, just upload any random custom index to prevent crashes

Select input and output of the same driver type (MME) at the end. Is important
Sample length is delay (you should start with ~0.20-0.25 and see if you can go up or down), ignore harvest algorithm, do max fade length and extra you could do full
Choose rmvpe

[leaden yacht]

Is that not working good with the rvc voice changer or smth?

[pastel oak]

Dont download the first link that i just sent, ill grab the nvidia version

[leaden yacht]

alr

doesnt it say rvc-"old"-dev?

[pastel oak]

Wait you already have Mainline installed so you only have to do the dev patch i linked

[leaden yacht]

and which one is the right one?

[leaden yacht]

mainline?

[pastel oak]

you have rvc1006 installed

right

[leaden yacht]

no

mangio

[pastel oak]

Ahhhhh

https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/RVC1006Nvidia.7z

download this

then do the rvc old dev thing i wrote

[molten iris]

TY 🙏

[brittle wing]

if anyone is able to help me rq. I'm not sure what I'm doing wrong but im getting a error when i click convert

[pastel oak]

when you got rvc1006nvidia and the dev, you run go-realtime-gui without the dml to start it

[leaden yacht]

alr so no dml

[pastel oak]

remove the spaces from the audi file, make it one word only

maybe that fixes it already

[leaden yacht]

do I still need to download the rvc old dev thing?

[pastel oak]

I recommend it yes

[leaden yacht]

and put that in my voice changer folder or smth?

the one u sent?

[pastel oak]

yea you paste the files into the RVC1006Nvidia folder

[leaden yacht]

alr thx

[pastel oak]

I wrote a guide to reduce more delay, and it depends on that dev thing so yes

[brittle wing]

ok

i'll try it

[leaden yacht]

will do that in a sec but first, could it be that they rounded it up? As I haven't seen one with 1879

[pastel oak]

just use that yes

[leaden yacht]

alr thx

[pastel oak]

youll rarely get the EXACT step number

[leaden yacht]

I mean I save every epoch

[pastel oak]

then its rounded up

[leaden yacht]

yeah probably

cuz there isnt nothing around it

and is 606 epochs good?

[brittle wing]

im getting the same issue. What else could it be? if you dont mid me asking

[pastel oak]

I dont use colab so i rlly dont know :/ maybe restarting helps, else try other voice models to see if the issue persists

or any error codes displayed?

[brittle wing]

AttributeError: 'NoneType' object has no attribute 'tobytes'
NameError: name 'cpt' is not defined
these are the errors im getting inside colab

i guess i'll try again

[graceful obsidian]

as in?

then perhaps 14 or 12 for batch if 16 is a nono

as for how to modify files on kaggle ( json), no clue
you should ask people here who use it

closer to 16, the better in ur case

as for editing, just an idea but, you can ask gpt to generate you some cell code for rewriting files
if you want

[graceful obsidian]

yea

Not an issue if you know what you're doing and actively checking tensorboard

I've been working in that manner for over half a year now

having max 4-5 gb for a training, maybeeeee 6

to maximize space, get rid of no f0 v2 pretrains, also v1 pretrains ( all of em, not just no f0 )

[honest whale]

So uh, what is going on here guys??

it's weird, every single voice model I did and it kept instantly gave me this error

[leaden yacht]

can you send the download link rq?

wanna test smth

[honest whale]

download link?

as in the model?

[honest whale]

.

[leaden yacht]

yeah

[honest whale]

https://huggingface.co/SUP3RMASS1VE/Chester-Bennington-0v2/resolve/main/Chester Bennington 0v2.zip?download=true

[leaden yacht]

I had this yesterday with someone else too I think

[honest whale]

JSONDecodeError: Expecting value: line 1 column 1 (char 0)

[leaden yacht]

hm no your download link is working though

[honest whale]

on EVERY SINGLE MODELS

i've tried

[leaden yacht]

maybe try another collab?

[honest whale]

there's another?

[leaden yacht]

many more

[honest whale]

i only know of this one google colab

[leaden yacht]

https://colab.research.google.com/github/iahispano/applio/blob/master/assets/Applio.ipynb
Applio

https://colab.research.google.com/drive/16LkwvFZeudTpUOsE_6bMjOq2qkxFo8Hr?usp=sharing
Ilaria

Are the ones I recommend

Ilaria is pretty popular atm, and I've used applio myself in the past

[honest whale]

Ilaria and Applio is just name, yes? Everything should work fine?

[leaden yacht]

yeah

those r just the names of the forks

but it should work as normal

[graceful obsidian]

aside few gimmicks here n there, or utilities like tts or such
it's all more or less the same

[leaden yacht]

where do I submit my trained model?

oh btw @graceful obsidian , thanks for the tips

I saved at every epoch and it really helped

I finally got my model now

[honest whale]

im using ilaria

so

downloaded done

I just press start? and then uh, what?

[leaden yacht]

uh I think you wait until theres a link

and then open that link

could take a few mins idk

[peak osprey]

Greetings!

[leaden yacht]

sup

[peak osprey]

I think i can't send any pictures

By the way, how can i get RVC version 2?

[leaden yacht]

RVC v2 Is usually built into any forks/mainline

If u think urs doesn't or wanna check, you could go to the train tab, and if it shows RVC v2 ur good to go

like that

u dont need to do anything else in there if u js wanna check

[low shard]

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

[honest whale]

Ohhhh

[peak osprey]

Oh well ummm, i am using google collab the whole time 🐸

How do i get that?

[leaden yacht]

so still ilaria and applio like I sent

[rare goblet]

Ayo? @leaden yacht level 12 !!!

[leaden yacht]

which collab?

[peak osprey]

https://colab.research.google.com/drive/1Gj6UTf2gicndUW_tVheVhTXIIYpFTYc7?usp=sharing#scrollTo=Y13Eh9r_g8f-

That

Colab*

[rare goblet]

Ayo? @peak osprey level 1 !!!

[leaden yacht]

I would use one of these as they are supported

but yours already has rvc v2 so ur still good to go

[peak osprey]

I'll use yours

[honest whale]

TY!

[leaden yacht]

alright

[leaden yacht]

?

[honest whale]

i mean thank you

wow

[leaden yacht]

Yeah I know dw

[honest whale]

this works pretty fast

even though uh

damn

[leaden yacht]

I take it the collabs worked?

[honest whale]

I need to know how to set up Chester's vocal to like

not be messed up with this song's vocals

[graceful obsidian]

You're confusing it a little bit

[peak osprey]

I went to appolio, where can i find the v2 choice?

[leaden yacht]

Yeah i know, I shouldve said it a lil differently

[graceful obsidian]

RVC is how the whole thing is called, V2 because newest pretrains are in their version 2
as for mainline, it means the main / original repository
forks are things based on the mainline, mirrored ones but modified

[leaden yacht]

ah

[honest whale]

where does one go and talk about stuff like 'would this song work with this model's vocal' and stuff?

[graceful obsidian]

all good, just worth mentioning

dw

[leaden yacht]

Uhh I think here

[peak osprey]

Oh!
Didn't get anything

[leaden yacht]

btw code where do I submit my model?

[leaden yacht]

Cuz after following ur tips everything worked out pretty well

and I got the exact epoch I needed

[honest whale]

so

okay

https://www.youtube.com/watch?v=TvSYkh0-Ntg

I'm tyring to put Chester's vocal onto this song

[leaden yacht]

did you remove the vocals?

[honest whale]

problems is the high keys, most models I've tried seems to just

work out

like at the normal voice, and yes, I separate the backing vocals and main voca

[leaden yacht]

so what you could do, is change the semitone to like -4 or smth like that

[honest whale]

but at high keys, it starts to just malfunctions

[graceful obsidian]

what you don't get

[leaden yacht]

that could change it a little bit

[graceful obsidian]

Neat!

glad it worked out

[leaden yacht]

his voice will be deeper though, so if its not the wanted effect you can always change it around

[honest whale]

was there ever a model that's actually made for metal screaming here?

[graceful obsidian]

Hmmm, perhaps on weights.gg
or voice models channel

[honest whale]

now that I think about it

I've seen a Corey model and even that one struggles

[peak osprey]

The whole thing

My question is, i am in appolio, how can i find V2 choice?

[leaden yacht]

no I meant like to get the voice maker role etc

in this server

[graceful obsidian]

O, model maker chat should have something pinned about that

alternatively, you could ask @red kayak

cause I don't quite remember the procedure

[graceful obsidian]

It should be in the training tab

[leaden yacht]

ah alr

[graceful obsidian]

tho, lemme check on that

Hmmm, seems like the applio is now something else

I haven't used it personally but, might be they've incorporated so-vits into it? vits gpt thingy from rvc's devs? ( that's a tts )
or just named it that way " vits based " rather than " retrieval based " (( tho, rvc itself uses some of vits things too

[red kayak]

no they just have tts

alternatively they might have added gpt sovits support

but i doubt it heh

[graceful obsidian]

ah ye, so that's just rvc + vits gpt and microsoft based tts stuff yeye I recall that one

incorporating tts pipelines into rvc ye

[red kayak]

mhm

consider trying out gpt sovits btw

[graceful obsidian]

Ye I do indeed plan to, just waiting for finetune availability

for japanese and english

unless they've already added it

[red kayak]

the support chines and english

[graceful obsidian]

As for this
generally " pitch issues " or " screaming / non screaming " stuff etc
it is related to how well the model is traind + what kind of voice data the training was exposed to

[red kayak]

maybe japanes too

[graceful obsidian]

yeee, sadly
the finetuning is for eng and chinese just yet

tho training iirc is jp, eng, chinese and maybe some other

[red kayak]

yeah

[graceful obsidian]

For example
If the model was trained on rather soft / chill singing or speech, the model might not be able to replicate screaming / screamo type of sounds etc

especially if index is used ( or is used with higher ratio values )

Another thing is, the tonal / pitch range

if the model was, say, made on speech samples and they were monotone ( lacked pitch variations + at different volumes ) they won't have that high of a range - ofc, they won't " tear ", they just might not sound as ' accurate '

If the model is tearing, it might mean few things:

• Dirty acapella ( for example; harmonies, backing vocals etc. )
• overcompressed dataset ( compromised dynamic range : pitch relationship )
• overtrained / undertrained model
• or simply poor quality dataset at specific phonemes which your acapella at given pitch or volume might need

Am well aware it might be confusing for some ^ so feel free to ask about anything in particular or any of those related things, Imma simplify it

[leaden yacht]

what do I put as license on huggingface?

to upload the model?

[graceful obsidian]

Hmmm, not sure but that's cause I don't share my models so.
Maybe check someone else's models and see what they've added in there?

or check rvc's repo for any mentions of that

[leaden yacht]

alr

[odd shale]

openrail

[leaden yacht]

thx

[odd shale]

https://docs.aihub.wtf/essentials/voice-models/

Scroll down and you'll find a guide about how to upload models to HF

[leaden yacht]

alright thx

https://docs.aihub.wtf/extra/model-maker-role/
Shouldn't I use this?

[odd shale]

That's the guide about how to upload your model to the voice models channel.

First you need to upload your model to your HF account.

[leaden yacht]

oh yeah

Do I have to attach an image in the get model maker /submit command?

[violet heron]

I guess

[leaden yacht]

its a narrator of a game, and he doesn't show up

hes js a voice

do I js put fanart or smth? 💀

[violet heron]

Show promo art for the game

[leaden yacht]

smth like this?

[violet heron]

Yeah

[leaden yacht]

alr

[unborn viper]

I’m using applio and it says 2000 seconds… is that even normal?

[violet heron]

Under what?

[peak osprey]

Can i use these RVC and make ai covers on mobile?

[graceful obsidian]

after you do the sync ( know the steps ) you gotta purge all cause if any from the actual training is of similar name, unknown conflicts could arise

• generally, always best to remove em all
• the test training is just to know the steps
• after syncing, you just train from scratch ( same settings - cause different batch changes steps )

[peak osprey]

Sir

[rare goblet]

Ayo? @peak osprey level 2 !!!

[upper furnace]

Sorry, I have a question, how do I put titan on rvcv2 disconeted? (because I don't understand how to start it)

[unborn viper]

I clicked on convert and it’s still going on, it’s been 3000 seconds it says.

[rare goblet]

Ayo? @unborn viper level 2 !!!

[violet heron]

How long is your audio file

And what’s your GPU

[unborn viper]

Well it’s bout 5 minutes, 4.8MB to be exact. I use a GTX1650

[earnest viper]

any python library for only inference with files from weights.gg? not a webui

[violet heron]

Shouldn’t take that long for inference

[unborn viper]

Should I try again?

[violet heron]

Yes try restarting

don’t close any terminal windows

[unborn viper]

What’s a terminal window

[leaden yacht]

would my 4060 8gb vram, with 12 batch per gpu setting, be able to train with an hour long dataset?

[violet heron]

idk if it applies to Applio but when you open it there’s a black window with text that opens with it

[unborn viper]

There isn’t any sort of thing like that

Is there any alternatives?

[leaden yacht]

yeah you can

you could use a collab on a mobile browser

same as other devices js another screen

[fleet knoll]

no vol help meh

[violet heron]

To applio

[unborn viper]

Yeah

[violet heron]

Original RVC

[violet heron]

It’s also slightly faster too

[leaden yacht]

local or cloud?

[unborn viper]

Can i get a link to thar?

[violet heron]

I would try restarting applio first and trying again

Make sure you selected your model and audio

If not

Here’s the link

-local

[azure marsh]

🖥️ Local stuff

• 🍏 Applio, by IA Hispano GitHub
• Mangio-RVC-Fork, by Mangio621 Huggingface
• RVC Studio, by SayanoAI Huggingface
• AICoverGen, by SociallyIneptWeeb GitHub
• Replay, by Replay Team Website
• Original RVC, by the RVC-Project team GitHub
• GPT-SoVITS, by RVC-Boss GitHub

Need some help? 🤔

You can find more info on the #1159513888199540817 channel. If you can't find your answer, feel free to ask for help in #✨│ai-help. Credits to Faze Masta and Antasma for compiling these links.

More commands: -audio, -colabs, -overtrain, -realtime, -rvc, -help

[leaden yacht]

wait @unborn viper u have a gtx 1650 right?

[unborn viper]

Yes

[leaden yacht]

if the original rvc doesn't work, I recommend trying mangio

I used to have a gtx 1650 too

almost nothing but mangio worked

and if you wanna train you also have to change a bit of code, I debugged the code with someone and found something that was keeping gtx 1650 users from training and giving them errors

[unborn viper]

Well I’ll try applio again and if it dosent work then I’ll try that

[leaden yacht]

alright

I recommend trying mainline first though

[peak osprey]

Then why it says error here?
I wanna make TF2 heavy ):

[leaden yacht]

smth wrong with the collab I think as it says connection errored out

you didn't close the collab tab did you?

[unborn viper]

I closed applio it literally approximated 4000 seconds

That’s wild 💀

[leaden yacht]

nah

if its approximated you should js wait

[unborn viper]

How do I do this mangio thing

[leaden yacht]

for me it approximated 5000 secs and still it was only 15 secs after all

you could just wait a bit

[unborn viper]

Or wait

Nvm I wanna try the original rvc

Then I’ll do mangio

So I think im supposed to download the Nvidia GPU version I think

[leaden yacht]

https://github.com/Mangio621/Mangio-RVC-Fork
Go over to releases, then download either INSTALL_Mangio-RVC-v23.7.0_INFER.bat or INSTALL_Mangio-RVC-v23.7.0_INFER_TRAIN.bat

The infer_TRAIN.bat allows you to train your own model if you want to, so if you wanna train make sure to download that one

[peak osprey]

No, that's it

[unborn viper]

Imma try both

I’m downloading them rn

[leaden yacht]

The .bat files will download a .rar, 7z, tar.gz (forgot which one), but either way you need to extract that and then you need to run go-web.bat

[leaden yacht]

both the batch files?

[peak osprey]

I will restart the RVC again

[leaden yacht]

or both mainline and mangio?

[unborn viper]

I’m trying both mangio and this original RVC one

[leaden yacht]

alr

[unborn viper]

I’ll wait till it downloads

[leaden yacht]

are you gonna train models too?

[unborn viper]

What happens when I train models?

[leaden yacht]

Training rvc models is js making a model

the ones you use are made by others

[unborn viper]

Oh

[violet heron]

Mangio is way more bloated

And slower

[leaden yacht]

mangio also is the one that worked best with gtx 1650

[violet heron]

Have you tried orginal rvc