✨│ai-help

[unkempt vale]

This share link expires in 72 hours. For free permanent hosting and GPU upgrades, run gradio deploy from Terminal to deploy to Spaces (https://huggingface.co/spaces)
Timer: 00:05:43 2025-01-31 13:00:06.283284121 [E:onnxruntime:Default, provider_bridge_ort.cc:1836 TryGetProviderInfo_CUDA] /onnxruntime_src/onnxruntime/core/session/provider_bridge_ort.cc:1511 onnxruntime::Provider& onnxruntime::ProviderLibrary::Get() [ONNXRuntimeError] : 1 : FAIL : Failed to load library libonnxruntime_providers_cuda.so with error: libcudnn.so.8: cannot open shared object file: No such file or directory

2025-01-31 13:00:06.283321319 [W:onnxruntime:Default, onnxruntime_pybind_state.cc:921 CreateExecutionProviderInstance] Failed to create CUDAExecutionProvider. Please reference https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html#requirementsto ensure all dependencies are met.
Timer: 00:06:51

Unfortunatelly it seems the same 😥

[glad zealot]

pain

lemme try again

[glad zealot]

actually did you reload the colab first?

[unkempt vale]

I stopped running the code and restarted it

[glad zealot]

try reloading the colab then try again

[unkempt vale]

Ok. I stopped the code and closed the colab tab at browser

Then I'm gonna open it and run again

[low shard]

Alex left the server many months ago if not a year

Kit Lemon foot works on the rvc disconnected colab rather than the mangio code itself, the colab is using mangio rvc

You can see if the colab is fixed in #📰│dev-updates

But yeah applio is more updated

[unkempt vale]

I don't know how to appreciate you! It fixed.

[brittle wing]

What's going on?

-guides

[karmic olive]

📚 Documentation

AI HUB Docs

https://docs.ai-hub.wtf

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

• How to use RVC Mainline Colab by Cauthess
• AICoverGen Colab Guide by Eddy (Spanish Helper)
• Create a model with RVC disconnected (colab) by Angetyde

[craggy aurora]

Hi guys. What RVC interface do you use? I try to use RVC-GUI-pkg and i get really bad resultats, idk why

[knotty moth]

-gui

[karmic olive]

https://cdn.discordapp.com/attachments/1122285248844144733/1203460490475343953/caption.gif?ex=65d12cec&is=65beb7ec&hm=bd2fb8d010006dd7c6e3c1c67d3ae846fd1478e1a3124c544c31b43086fe54aa&

[low shard]

Rvc GUI is outdated

Don't follow yt tuts

What's ur PC GPU?

[craggy aurora]

GTX 1650

[low shard]

Your Nvidia GPU is good enough to do inference (use models) locally (on ur pc), not the best to train (make models) even if still possible

You can:

• Locally (runs on your pc so the speed depends on that, you will have to set it up with the guides):
  • Applio: A fork of RVC with some extra features like Applio TTS, kinda faster and simpler but same quality tho
  • Mainline: The original RVC
• Cloud (remote good pc, easier and faster than ur PC but it's limited):
  • Ilaria RVC Zero: fastest and simplest that you can get for free
  • Weights.gg: Partnered with AI Hub, lets u do them easily but u may be in a queue
  • Applio Colab: max 4 hours daily, not granted, of GPU

Easiest possible (automatically separates vocals & instrumentals) : weights.gg
easiest cloud: Ilaria rvc zero
easiest local: Applio

[craggy aurora]

Thank you homie, ill try it

[low shard]

Yw brother, let me know

[brittle wing]

-colab

[karmic olive]

📒 Google Colab Notebooks

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• Hina's Mod AICoverGen WebUI, by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], by Eddy, Hina and Gdr Google Colab
• UVR5 NO UI, by Eddy Google Colab
• UVR5 UI, by Eddy Google Colab
• Hina's Modified Original W-Okada's Realtime Voice Changer, Google Colab
• FaceFusion UI, by Nick088 Google Colab
• FaceFusion NO UI, by Nick088 Google Colab
• EasyGUI, by Rejekts Google Colab
• 🆕 Music Source Separation Training (Inference), by Jarredou & Makidanye Google Colab

ℹ️ Note

While the Colab free plan provides up to 12 hours of daily usage, the GPU is typically available for only about 4 hours each day on average.

More commands: -audio, -hf, -local, -kaggle, -realtime, -rvc, /help

[wanton valley]

@low shard there is nothing here, i press train index and it just process some seconds and nothing happens

the model just worked fine? but index is not

[low shard]

Show the output in the google colab output

[wanton valley]

nothing at all lol, but i can show you feature extraction, dont know if it works

[low shard]

I mean what shows inside of the google colab cell output, not the UI output

[wanton valley]

here?

like i said, nothing at all

also, feature extraction:

['infer/modules/train/extract/extract_f0_rmvpe.py', '2', '0', '0', '/content/training/logs/RitaRVC', 'True']
['infer/modules/train/extract/extract_f0_rmvpe.py', '2', '1', '0', '/content/training/logs/RitaRVC', 'True']
no-f0-todo
no-f0-todo
['infer/modules/train/extract_feature_print.py', 'cuda:0', '1', '0', '0', '/content/training/logs/RitaRVC', 'v2']
/content/training/logs/RitaRVC
load model(s) from assets/hubert/hubert_base.pt
move model to cuda
no-feature-todo

[low shard]

No feature Todo means the feature extraction didn't go well

Show a screenshot of what's ur dataset folder like

And what did u use to make ur dataset

[wanton valley]

start preprocess
['infer/modules/train/preprocess.py', 'datasets', '40000', '1', '/content/training/logs/RitaRVC', 'False', '3.0']
datasets/place-audio-here.txt->Traceback (most recent call last):
File "/content/training/infer/lib/audio.py", line 37, in load_audio
ffmpeg.input(file, threads=0)
File "/content/training/.venv/lib/python3.10/site-packages/ffmpeg/_run.py", line 325, in run
raise Error('ffmpeg', out, err)
ffmpeg._run.Error: ffmpeg error (see stderr output for detail)

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "/content/training/infer/modules/train/preprocess.py", line 87, in pipeline
audio = load_audio(path, self.sr)
File "/content/training/infer/lib/audio.py", line 42, in load_audio
raise RuntimeError(f"Failed to load audio: {e}")
RuntimeError: Failed to load audio: ffmpeg error (see stderr output for detail)

end preprocess

like this?

[low shard]

Show what's inside the datasets folder

[wanton valley]

well would you look at that, though it would be more automatic like old collab

but wait, how did it make my model if feature extraction or index didnt work? are they separated?

i.. already uploaded the audio there, in zip or in wav and still giving me the same error, do i need to restart the collab or..?

[tame mica]

"path to the dataset"

not the folder name

search for it in the colab file explorer

and copy the path

[brittle wing]

-guides

[karmic olive]

📚 Documentation

AI HUB Docs

https://docs.ai-hub.wtf

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

• How to use RVC Mainline Colab by Cauthess
• AICoverGen Colab Guide by Eddy (Spanish Helper)
• Create a model with RVC disconnected (colab) by Angetyde

[placid pond]

is there a way 2 make it less delayed

[low shard]

Wrong channel, use #🔍│help-w-okada and elaborate more saying what tut link u followed, a SS of ur settings + what's ur PC gpu

[low shard]

U need to put the path to the dataset in the colab explorer

[wanton valley]

alright, i dont know what i did that feature extraction finally worked and i could take the index, but the model was broken or something so i took the pth of my previous model and added this new index, from what i seen it works fine

[placid pond]

shiet okey mb

[low shard]

Be sure to use the right path from the colab explorer next time

[wanton valley]

https://tenor.com/view/kowalski-noted-penguins-penguins-of-madagascar-gif-1306663699157096677

fack

[crude flame]

https://tenor.com/view/embed-no-no-embed-megamind-megamind-meme-gif-25261934

[wanton valley]

alright

[quasi condor]

https://tenor.com/view/kowalski-noted-penguins-penguins-of-madagascar-gif-1306663699157096677

yus

[formal wind]

Yo how do I install a custom pretrain that isnt on the list on rvc mainline?

[hexed shard]

I know nothing about this topic, but I think this might be the place to ask? I have a really stupid idea where I make a voice changer but for it to straight up replace my voice with a sound effect once. So like, if I say "hello guys whats going on" it'll pick up the "hello" since it's the first word spoken and replace it with a roar, and then do it again after the effect is over if I speak again. I'm really tired rn so I don't know if I explained it well but I hope this makes sense. I don't know how I'd go about this if anybody could help

[rocky maple]

seems to work fine now - someone a few hours ago had the same issue and it got fixed :D thanks again

[glad zealot]

Local or colab? On colab there's instructions on there that actually says how to do it right above the list

[formal wind]

I dont see any sort of instructions.

[glad zealot]

It's basically the same thing as the theme loader

You go to hugging face/ #1235952130855010365 , find that part that has <creator name>/<pretrain name,>

[formal wind]

So something like that for example

[glad zealot]

Yup

Note that for titan models it might take longer than the others to download just because of how many files it has

[formal wind]

Alrighty tysm!

[surreal sleet]

Hello guy

RVC v2
RMVPE
480 epochs
OV2

What is Epochs ? on the software ?

[wheat plover]

rvcDisconnected is broken?

[fervent rover]

Yeah is Broken

Anyway I was about the say that I’m trying Applio Ui, I will update you guys to know how was is it going

Wish Me Good Luck👍🏻

[simple ore]

you're trying Ui version

[fervent rover]

Yeah well the true is it’s not working because I needed more GPU so yeah

[thin edge]

Is there a place to make another Rvc that is friendly to newcomers? Weight is very easy but nothing can be set, RVC DC is broken.

[fervent rover]

Well nothing man, I tried some of them that I find on YouTube but it says your data must be preprocessing even though I done that, so yeah I had no choice but the wait for it

[thin edge]

I'll stick to weight then, after all I don't find any noticeable difference between Weight RVC and RVC made by people outside of weight.

[knotty moth]

dont think they will fix that colab anymore

[fervent rover]

Oh that sucks😢

Also is there going to be anything new of RVC that is like RVC Disconnected?

Can the owner just make a new one of it?

[wheat plover]

F in the chat boys

[fervent rover]

F

[brittle wing]

How do I resume training on Applio

[knotty moth]

sorry I get bored on this kind of question, but refer to this: https://docs.aihub.gg/rvc/cloud/applio-colab/#5-resuming

[brittle wing]

Thx

I meant Applio no ui

[knotty moth]

https://docs.aihub.gg/rvc/cloud/applio-no-ui-colab/#5-resuming-training
well this is enough

[brittle wing]

Thx

[brittle wing]

There are resuming & backup cells

Does the sample rate have to do with the audio frequency?

[simple ore]

that's what the sample rate is

[brittle wing]

So I have to resample every each data sample

To use a certain pretrain?

[brittle wing]

How is it possible for a 30 minutes dataset to overtrain at 40 epochs with ov2 super?

[raw stirrup]

change 1e-4 to 5e-5 in config

will help balance ut out 1e-4 is probably to strong for it at first

[wheat plover]

what's a good alternative?

[raw stirrup]

local

[hallow thistle]

[low shard]

its just an info on how the model was trained

[low shard]

yes, #📰│dev-updates

[low shard]

applio or weights

[low shard]

youtube tuts are old

[brittle wing]

Again, how do I resume training on Applio no UI colab?

[low shard]

https://docs.aihub.gg/rvc/cloud/applio-no-ui-colab/#5-resuming-training

[low shard]

what’s ur pc gpu?

[knotty moth]

I have answered it above

[low shard]

Oh 😭

[brittle wing]

There are resuming/backup cells

[brittle wing]

Is it normal for a 30 min dataset to overtrain at 40 epochs with ov2super?

[raw stirrup]

turn youre training setting down abit if its at 1e-4 change it to 5e-4 or 5e-5

[brittle wing]

?

[fervent rover]

Yeah

[brittle wing]

Overtraining?

Looks like OT

[brittle wing]

40 mins of data overtraining at 50epochs with ov2super?

[open stag]

• colab

[karmic olive]

📒 Google Colab Notebooks

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• Hina's Mod AICoverGen WebUI, by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], by Eddy, Hina and Gdr Google Colab
• UVR5 NO UI, by Eddy Google Colab
• UVR5 UI, by Eddy Google Colab
• Hina's Modified Original W-Okada's Realtime Voice Changer, Google Colab
• FaceFusion UI, by Nick088 Google Colab
• FaceFusion NO UI, by Nick088 Google Colab
• EasyGUI, by Rejekts Google Colab
• 🆕 Music Source Separation Training (Inference), by Jarredou & Makidanye Google Colab

ℹ️ Note

While the Colab free plan provides up to 12 hours of daily usage, the GPU is typically available for only about 4 hours each day on average.

More commands: -audio, -hf, -local, -kaggle, -realtime, -rvc, /help

[brittle wing]

yo dumb question

so i got this private model i trained fusing my homies voice with mr. krabs, i DO NOT wanna share it publicly bc that would be cringe, but hina's new rvc just WILL not take gdrive links.

isnt there a certain way you gotta do gdrive links now so rvc will take it?

its usually "ends in usp=sharing" or something but there was like, some way around that right?

[hallow thistle]

To request someone to do the voice model for you, go do one in #1159289738314919936

[raw stirrup]

your data is either bad or you need to lower the training like i said earliler in config from 1e-4 to 5e-4

what hz are u training the model 32 40 or 48?

[brittle wing]

The dataset sample rate is 40k, so it's 40k.

[brittle wing]

How do I lower training?wdym

[raw stirrup]

open the 40k config file and change 1e-4 to 5e-4

download it edit save changes then re upload

unless youre training locally

[brittle wing]

????

Whaaat

[raw stirrup]

are u training locally or google collab

looks like no one ever bothered to explain how the collab works and the importance of changing the config file to ajust training to prevent overfit

1e-4 is quite aggressive to start with and usually causes over fitting

some models work fine some u need to lower the training to 5e-4

and ajust as u go along

[brittle wing]

Applio no UI Colab.

[raw stirrup]

u need to go in the applio folder in collab and find config folder download the 40k file and edit it

save it and re upload

{
"train": {
"log_interval": 200,
"seed": 1234,
"epochs": 20000,
"learning_rate": 1e-4,
"betas": [0.8, 0.99],
"eps": 1e-9,
"batch_size": 4,
"fp16_run": true,
"lr_decay": 0.999875,
"segment_size": 12800,
"init_lr_ratio": 1,
"warmup_epochs": 0,
"c_mel": 45,
"c_kl": 1.0

it will look like that when u open it

change the learning ratge to 5-e4

[simple ore]

5e-4... crrraaayzy

[brittle wing]

What is this...

[simple ore]

5e-5 maybe

[brittle wing]

Is it possible for a 40 minutes dataset to overtrain at 50 epochs with ov2super?

[simple ore]

no

default learning rate is 0.0001

[brittle wing]

"Small epochs "

[simple ore]

if you really want it to learn slower, you can change to 0.00005

[brittle wing]

Then what's the problem

[simple ore]

but not to 0.0005

[brittle wing]

Where's that

[raw stirrup]

dont keep it at 1e-4

[simple ore]

nowhere, it is not the problem

[raw stirrup]

thats where youre going wrong

[brittle wing]

Where do I set the learning rate in applio no UI colab?

[simple ore]

do not touch it

[brittle wing]

Then where's the problem

[raw stirrup]

if he keeps it at 1e-4 it will just keep overfitting

why tell him that?

[simple ore]

dude, you're suggesting 5x more value for him

who's crazy here

[brittle wing]

She*

[brittle wing]

Where's the problem really...

[simple ore]

here's a model with 40min audio I trained a while ago

it was like 180 epoch

[raw stirrup]

5e-5 is what i meant

less aggresive

[brittle wing]

With pretrain or without?

[simple ore]

it is nowhere close to overtraining

[brittle wing]

Pretrain or not

[simple ore]

of course

a small bump on the chart is not overtraining

this would

[raw stirrup]

this one is not really a small bum,p

top right pic

[brittle wing]

ThI KNOW

[simple ore]

it is 0.5 value difference

jfc

it is not +50, not even +5

chill the f out and keep training

[brittle wing]

Me?

THIS is the most recent status

[simple ore]

you're looking at the old unaveraged totals, those charts are stupid

[brittle wing]

Of the model. 100 epochs

[brittle wing]

I'm using colab-

[simple ore]

so? clone the main repo

[brittle wing]

How

I'm on a phone

[simple ore]

remove the branch

[brittle wing]

No such cell in the colab.

Do I add it

[brittle wing]

?

[brittle wing]

What does it do

[simple ore]

clone clones the code from the github repository to colab

[brittle wing]

I've already ran it w the installation cell.

[simple ore]

specifically 3.2.8-bugfix branch

[brittle wing]

What are you trying to say

It's the same cell code.

[simple ore]

i'm trying to say, if you remove --branch 3.2.8-bugfix, it will clone the main repository

[brittle wing]

Do I have to edit it?

[simple ore]

yes

[brittle wing]

And...?Will I hv the avg graph in my tensorboard

[simple ore]

yes

with the new training

[brittle wing]

Uh thx why nobody told me that earlier

[surreal beacon]

Guys, sorry to interrupt the conversation, but can someone help me??

Lately, Google Colabs haven't been working very well, at least for me...
I’d like to know if anyone has a Google Colab link that generates good AI covers with backing vocals!!! I can only access the one from Applio, but it doesn’t load backing vocals, so the audio ends up being just one layer, and the sound quality is terrible.

If anyone has it, please send me the link. Thank you so much 😭

[simple ore]

it is 'use on your own risk'

[simple ore]

why dont you use something like audacity to master multiple tracks?

[brittle wing]

I have one but it's very slow what I actually do is filter out both lead&BV then I infer them and mix them in bandlab and add sfx

[brittle wing]

You can infer bvs in applio

[surreal beacon]

How do I enable backing vocals in Applio? Can you help me?

[brittle wing]

Just infer them
..

[surreal beacon]

Is Audacity an AI cover Colab??

[brittle wing]

Just infer them

[brittle wing]

No it's audio program

[simple ore]

it is audio mastering application

[brittle wing]

I suggest bandlab since it's easier and has so many sfx & presets

Reverb, bitcrush, chorus

[brittle wing]

Why did you say that the old graphs are stupid?Are they deceiving

Or outdated

I guess

[surreal beacon]

(Sorry for the delay in responding to the messages, it's because I'm Brazilian, so I need to translate.)

But anyway, can I activate the background backing vocals directly in Applio? I’d like to know how.

[brittle wing]

You just extract them w uvr bve or Mel roformer karaoke... And infer them lol

[carmine hearth]

Are you looking for a colab like aicovergen?

[surreal beacon]

Oh my God, this seems so difficult. 😢 Is this done in the Gradio link generated by the Colab (on the website) or in the Colab itself?

[simple ore]

average losses are calculated each step and it is a running total, old losses are a value from a random batch

[surreal beacon]

Well... yes, if it's easier.

[brittle wing]

So they're outdated graphs?

[brittle wing]

It only converts lead vocals.

No

[brittle wing]

On MVSEP

[surreal beacon]

Oh, ok

Or guys, if it's easier, can you just send me a link to an AI cover generator that includes backing vocals in the final audio?

[brittle wing]

https://colab.research.google.com/github/ShiromiyaG/RVC-AI-Cover-Maker-UI/blob/main/assets/RVCAICoverMakerUI.ipynb

It's slow!

But it's what you're looking for.

[surreal beacon]

TYSM!!

[brittle wing]

It's slow and a pain to use

I prefer applio/harmonify

@simple ore is this the graph?

[brittle wing]

It's slow...

[surreal beacon]

Ahh

Do I necessarily need to use NGROK here? What is that?

[brittle wing]

Idk anymore

[surreal beacon]

Is this where the harmonization is done in Applio? (I mean the backing vocals?)

[brittle wing]

Idk I said learn by yourself

[surreal beacon]

Never mind, I’ll just use Applio. Lol

[brittle wing]

https://docs.google.com/document/d/1dW9t5UxO6vga_dIg4chfnf6veaYYFXXyTKdQkDAirz4/edit?usp=drivesdk

[brittle wing]

https://docs.google.com/document/d/1tOwL7P-NRLVwc27R-ThkyxJSpL2XIMceWAC_6HWXHtc/edit?usp=drivesdk

[brittle wing]

Bro
.... there's a start without ngrok cell

[surreal beacon]

I clicked it, but it didn’t generate the Gradio link LMAO.

[brittle wing]

???

[surreal beacon]

Wait, I’ll send a photo of what it generates when I click.

[brittle wing]

Did you press the 'Install' cell first?

[surreal beacon]

Yes men

[brittle wing]

And?

What happened then

[surreal beacon]

.

[brittle wing]

Send then

[surreal beacon]

It’s not generating the Gradio link.

[brittle wing]

Did you install?

[surreal beacon]

Yes

[brittle wing]

Oh you're on a computer

[surreal beacon]

Is this for mobile?

[low shard]

that's old and it doesn't work anymore

[brittle wing]

Then there might be an issue with the Colab wait I'm testing out myself.

[low shard]

the entire github project is archived

always check #📰│dev-updates

[low shard]

i mean you have to separate the back vocals yourself

or you can use https://weights.gg that does that automatically iirc

[surreal beacon]

Ok, thank you

[brittle wing]

So my method inferring both bvs&lvs separately

Yes

And mixing them in a program!

[low shard]

that colab is broken

[brittle wing]

Same error...look... it's a colab issue.

[brittle wing]

Yes

[surreal beacon]

Ah

[brittle wing]

Mhm

Yes

[surreal beacon]

Does this one work correctly?

[brittle wing]

https://colab.research.google.com/drive/1X8YR4Ruv7zzY8YAMPTfC7hkxqT_d4Q5d
This one is precious & still works 👀

[low shard]

yes

[brittle wing]

Just asking...is using a hybrid pitch extraction a good thing?

[low shard]

that's harmonify, an old fork made by Eempostor, not sure if that works since he's not active in the rvc community anymore

[low shard]

it's just better to use rmvpe

not sure how good hybrids are

[brittle wing]

I liked them

[brittle wing]

It still does.

[brittle wing]

it would be VERY pog if one of yall could answer this for me

[low shard]

when was last time you tested it?

[brittle wing]

Last month...whattt

This one too 😦

[low shard]

which hina rvc ?

[brittle wing]

It was my favorite

[low shard]

the google change has been made 2 weeks ago

[low shard]

welp its outdated

[brittle wing]

the official one

[brittle wing]

It doesn't work?I have to remove it from my doc/list then

see i aint used this stuff in a while, and it turns out i was using a discontinued one by an indonesian guy

[low shard]

what are you exactly using and what's your pc gpu?

[brittle wing]

colab only

[low shard]

can you send the colab link u are using

[low shard]

ohh you're that guy

you're using https://colab.research.google.com/github/hinabl/AICoverGen-Colab/blob/main/Hina_Mod_AICoverGen_colab.ipynb right?

[brittle wing]

yus

does this just not accept drive links or am i pasting them wrong?

i remember a year or 2 ago you had to like, change how the url looks to get it to work

[brittle wing]

What change.

[low shard]

I will check it in a sec, it's installing on colab rn

[low shard]

#📰│dev-updates message

this is why colabs aren't working

you gotta check always #📰│dev-updates

it affects every single existing colab

[brittle wing]

Was there a certain change in a code or something...

[low shard]

i sent you the link to the message that explains it, read it

[brittle wing]

It's a code change, so I guessed correctly

good to know google colab is still kinda doo doo water

[low shard]

google updated colab's python's version

so kinda

[brittle wing]

at least its not constantly disconnecting ppl like it was the last time i used it?

[low shard]

huh?

[brittle wing]

Yes that's what I mean.

[brittle wing]

yeah last time i stopped messing with it was bc it was like "only pro users can use it without disconnecting 5 minutes in"

[low shard]

I think you meant when web uis were banned and it said disallowed code since it was only for pro users, so we encrypted the code

you can still get disconnected randomly if you reach your max gpu time, which is random daily

[brittle wing]

if i had a nickel for every time i wanted to use rvc to make a meme or two and it started getting all fucky, id have 2 nickels its just weird it happened twice

[brittle wing]

yeah the pro user thing

that

[low shard]

yeah it doesn't support gdrive links

but you can just reupload the model on pixeldrain, or go on the Upload Model tab and upload the zip file of the model

[low shard]

https://tenor.com/view/i-have-two-nickels-nickels-doofenshmirtz-phineas-and-ferb-dr-doofenshmirtz-gif-26627596

[brittle wing]

oh thanks

[low shard]

you're welcome

[wheat plover]

since rvcDisconnected is down, can someone recomend another training colab?

[low shard]

Check #📰│dev-updates

Ecc disconnected got fixed FOR NOW, but it's discontinued and will never be maintained again after this update

You can see the list of fixed colabs a in dev updates

Btw what's ur PC gpu

[wheat plover]

rtx 4060

[low shard]

Why do you use colab then

You can do it locally

[wheat plover]

sometimes the model is just to big to do it locally

• I use it to play games

[low shard]

I mean it should work fine, the only limit would be the 8gb VRAM in some cases but it works fine for most things

But yeah if u play while training then that's a problem

Honestly local is better since you don't got random disconnected for GPU time limit, but your choice

[low shard]

Train (make) RVC Models on cloud:

1. Prepare the Dataset
2. Setup RVC:
   Choose a cloud way to use RVC,

• Google Colabs (4 hours of daily gpu for free, not much hours, but easy to use):
  • Applio (ui)
  • Mainline (UI)
• Kaggles (a bit harder to use and needs phone number but gives 30 hours weekly of better gpus):
  • Mainline (UI)
  • Applio by Vidal (UI, no guide as of right now)
  • Applio by Shirou (UI, no guide as of right now)
• Lightning.ai (Kinda hard, needs login, no issue with web uis or anything, but only free 15 credits monthly):
  • Applio (UI)
  • Mainline (UI, No guide as of right now)

3. Be sure to know about the tensorboard

Google Colab = Easier but risk of getting disconnected
Kaggle = Harder but way more gpu time
If you are looking for the easiest way and for free, is using https://weights.gg which ofc uses RVC

RVC Inference (use models) on pre-recorded audio on Cloud

You can use either:

• Weights.gg: Easiest Possible Ever Automatic
• Ilaria RVC Zero: Fastest free on cloud
• Applio (ui)

You can also see all fixed colab a in #📰│dev-updates

[wheat plover]

15gb of vram for bigger batch sizes is to good to pass

[low shard]

I personally suggest Applio on Kaggle or colab

Kaggle is better since it gives 30 hours weekly of better GPu, but it's harder and needs phone number

[brittle wing]

Is it possible for a 40 minutes dataset to overtrain at 50 epochs with ov2super?

Really

[odd shale]

Welp, all will depend on your dataset, training settings and the very same voice you're wanting to train.

Some voice can converge earlier than expected

[potent urchin]

heyo idk if this is the right channel but yea

I'm trying to setup AICoverGen on local and I'm running into a few errors when I click the generate button:

1. when separating vocals, I get this error message:
   However, the vocals and instrumentals are successfully generated

Failed to create CUDAExecutionProvider. Require cuDNN 9.* and CUDA 12.*. Please install all dependencies as mentioned in the GPU requirements page ...

2. After a certain amount of time, the process terminates with 1805 killed

anyone know what I'm doing wrong?

[sand path]

i have a problem, normally i know all the basics and so on to train a model, but i tried to make a model yesterday again (i made already over 30 models by myself) and i have a 3.30h long dataset. and i wanted to make a model, and got the problem with, INFO:torch.nn.parallel.distributed:Reducer buckets have been rebuilt in this iteration.

then i tried to reduce the dataset in hope it would work, but it was stuck on 1 epoch and that same issue was there again.

[low shard]

holy shit 3.30 hours

[sand path]

yup

only talking

[low shard]

reminder that you need to prioritize quality over quantity a bit

like be sure all of those are high quality

[sand path]

yea ik but it was all the same quality

[low shard]

also usually it doesn't change much over 1 hour

[sand path]

it was all from talking streams

of the same streamer always same quality

[sand path]

hmm okey so it should be under 1h idealy

[low shard]

anyways, have you tried waiting sometime?

[sand path]

yep..

i slept

woke up

the same

[low shard]

damn

[sand path]

i really wanted it to work so

[low shard]

Maybe try reducing the dataset to 1 hour and retry, (be sure to save up a back up tho in case the lenght isn't the issue), I honestly haven't seen this issue before

[sand path]

i did already

i tried it with 45 mins

but wait ill try again

[low shard]

your gpu drivers are updated right?

[sand path]

maybeee some wonders will happen

[sand path]

yep

but im probably also using an old version of mangio rvc

maybe thats the issue? but it worked till yesterday without problems

[low shard]

tell me where you downloaded your rvc

mangio rvc is super old, it's abandoned since 2023

it's fr mangio rvc?

[sand path]

yep the version is 23.7.0

so its prbly from july 2023

[low shard]

yeah

it's old, delete it

[sand path]

pretty sure it was on the first ai hub server before they switched to a new server

[low shard]

we even removed that from our docs, it's too old

yeah it was there before

but then the dev stopped working on it

[sand path]

ooooouhhhh...

[low shard]

it's full of bugs and slower than updated forks

[sand path]

so whats the best one at the moment, because there are so many now

[low shard]

As you got a good PC, you can use RVC locally, you can choose between:

• Applio: A fork of RVC with some extra features like Applio TTS, kinda faster and simpler but same quality tho
• Mainline: The original RVC

I would suggest Applio, you said you got an rtx 4070 so it's good

[sand path]

okey

wait

i think i got applio over pinokio

but ill download it nevertheless

[low shard]

that should be updated iirc, can't be sure since I used the normal precompiled installation

[lone nest]

What exactly does the torch installation on RVC disconnected fix?

[low shard]

torch is a pip package
there was a dependency issue
you shouldn't use rvc disconnected
even if it's fixed for now, it won't ever be maintained again, it uses the old rvc mangio which is slower compared to mainline and applio, and the dev is busy, check #📰│dev-updates

rvc disconnected will not be supported anymore

oh you were the guy who helped for the first python fix lol

[lone nest]

Yeah I'm wondering what exactly broke because there were no errors when the first fix came

[low shard]

okay so basically, the fix worked at first

then, rvc disconnected was based on alex's mangio rvc tweaks (a better fork of the mangio rvc fork), so kit had to switch to the older and slower mangio rvc
but that caused dependency issue since it was even older

the fix seems to be downgrading torch

well, rip rvc disconnected

was the first training colab that I used too

[lone nest]

Huh

Why switch to older mangio if the newer one worked?

Also does it mean the models trained between the first and the new fix are worse?

[sand path]

my applio is now

in german

how tf do i change this

im on opera and usually dont have a translator

[simple ore]

in settings menu

it you can change it from automatic to other

[sand path]

ahaa

thx i didnt see it

my bad

[lone nest]

Oh I get it

Well time to adapt to new environment thanks for everything RVC Disconnected and Kit 🫡

[low shard]

so basically
mangio621 made the mangio rvc which is a fork of mainline

alex which was an engineer who left many months ago, made a fork of mangio, adding some tweaks, called mangio rvc tweaks

since alex deleted forever the github repository, rvc disconnected nor anyone else can use it, so he's forced to use the original mangio fork

[low shard]

yea time to adapt to applio or mainline

[low shard]

btw operagx is known for spyware

[sand path]

WHAT

noooooo

nuh uh

[low shard]

well it does 83 request included geolocalization on startup, and the company was known for having a previous browser that had a backdoor

all articles linked inside that rentry

but hey, your choice not mine

[sand path]

I have to adapt to everything new, first okada, then applio NOW MY BROWSER

[sand path]

no no i get you

well i found the firefox opera gx theme

ig thats good enough

well time to switch to firefox again...

even if i hate firefox the most

[low shard]

I mean I'm not forcing you to do that, I'm just warning you to not 100% trust a company who launched a kids smartwatch that sent unknown data to their servers

[sand path]

no i totally understand you

i just thought if so many people are using it that its good

[low shard]

I mean chrome is fine too, except that you might have issues in the future for adblockers since they are gonna push a new extension manifest v3

which is also the reaosn I'm prolly gonna switch to firefox in the future if I can't use adblockers

[sand path]

yea no ig im gonna install firefox with the opera gx theme and the other tweaks that are shown in the text

[low shard]

today gotta be a big change for you 😭

[sand path]

thanks for telling me tho

[sand path]

at this point i could just make a new install of everything

[low shard]

that doesn't always apply to everything

there's prob many people using voice.ai letting that thing stay 24/ in background and utilize their power for distributed training

[sand path]

true

[low shard]

operagx has a good marketing strategy to look 'gamer cool' to appeal to younger audience

you prob saw other companies doing the same acting 'skibidi' on Twitter or Tiktok

[sand path]

now i dont wanna imagine what would be if i also would upgrade to windows 11

cuz im staying on win10 as long as i can

[sand path]

yea

i mean to be fair , seeing the admin oiling up himself on tiktok was a reason

what should i say

[sand path]

yea i already got that information as a big ass pop up on my screen 2 times

but i hate windows 11

i dont like the look of it

its tooo simple

and as far as i know there arent huge performance upgrades or smth else

[low shard]

as a win11 user it isn't the worst thing that could ever happen, the main issues are resources hungry and recall (recall can be turned off tho)

[sand path]

whats recall

[low shard]

https://support.microsoft.com/en-us/windows/retrace-your-steps-with-recall-aa03f8a0-a78b-4b3e-b0a1-2eb8ac48701c

basically a features that does screenshots of your usage of your PC like today so that you can check everything you did

[sand path]

awwhh hell nah

yeeaaa if im upgrading thats getting turned off imediatly , if i dont forget

oh nvm firefox isnt that bad anymore, its really nice looking

[low shard]

basically:
does snapshots of what u are doing every tot intervall of time

uses AI to analyze the content and make a roadmap of what you did

It's absolutely useless imo

from what Microsoft said it's all stored locally, but still why the fuck would I spend 25gb of storage for this

[sand path]

yeaaa nahhh

[low shard]

Idk if this was added for people who have issues remembering things but for the majority of users it's really not worth it

[sand path]

so basically dementia helper

[low shard]

kinda ig

some people switched to linux after w11

[sand path]

yea no linux isnt for me

in no dimension

except it would be something like steamOS

[crude flame]

what really?

[low shard]

you mean too complex or not that good for gaming?

[crude flame]

thats mad impressive if so

[sand path]

but even SteamOS is "bad"

[sand path]

too complex

[low shard]

i have seen some people saying that on socials like reddit

[sand path]

not bad its complex yes

[low shard]

i can't 100% put my hand on that statement but some really really hate win11

[sand path]

yea i understand tehm

them*

cuz ive got a steamdeck and its running on linux

as soon as im in desktop mode on it

IM REALLY lost

[low shard]

It's that bad? Btw we should talk more about this in #✦│chat since this is an help channel

[sand path]

true

[torn edge]

i wonder what a model trained off multiple different voices sounds like

[simple ore]

like an abomination

[hallow thistle]

Trying to get around with Linux, but you have no idea how to use it despite using Windows for many years.

[analog obsidian]

You can use easy distros

[hallow thistle]

No. Debian and Android x86 are Linux distros I used.

[brittle wing]

What do you mean by training settings?How do I set it up to reach more epochs

[brittle wing]

But my models overtrain 50-100 epochs idk why

[wheat pewter]

Any advice on making these vcs use less of my cpu

[brittle wing]

How to set up dataset with RVC mainline

-guides

[karmic olive]

📚 Documentation

AI HUB Docs

https://docs.ai-hub.wtf

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

• How to use RVC Mainline Colab by Cauthess
• AICoverGen Colab Guide by Eddy (Spanish Helper)
• Create a model with RVC disconnected (colab) by Angetyde

[brittle wing]

-colab

[karmic olive]

📒 Google Colab Notebooks

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• Hina's Mod AICoverGen WebUI, by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], by Eddy, Hina and Gdr Google Colab
• UVR5 NO UI, by Eddy Google Colab
• UVR5 UI, by Eddy Google Colab
• Hina's Modified Original W-Okada's Realtime Voice Changer, Google Colab
• FaceFusion UI, by Nick088 Google Colab
• FaceFusion NO UI, by Nick088 Google Colab
• EasyGUI, by Rejekts Google Colab
• 🆕 Music Source Separation Training (Inference), by Jarredou & Makidanye Google Colab

ℹ️ Note

While the Colab free plan provides up to 12 hours of daily usage, the GPU is typically available for only about 4 hours each day on average.

More commands: -audio, -hf, -local, -kaggle, -realtime, -rvc, /help

[brittle wing]

-rvc

[karmic olive]

📚 Documentation

AI HUB Docs

https://docs.ai-hub.wtf

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

• How to use RVC Mainline Colab by Cauthess
• AICoverGen Colab Guide by Eddy (Spanish Helper)
• Create a model with RVC disconnected (colab) by Angetyde

[brittle wing]

RVC Mainline reminds of the crepe+persistent storage colab.

The first one I used for training back in 2023

[warm hamlet]

we can train on colab now? my bad I'm OOTL

it's been a while, I was last here when RVC was getting banned off of colab for the webui I think?

is Applio still the best?

[knotty moth]

linux gaming is soon getting better on steamOS and arch linux than the problematic win11 (esp the 24H2)
AMD and intel gpu driver on linux are also getting better

[tawny raptor]

this overtraining?

20 minutes clean audio, 500 epochs

computer will finish in 5.5 hours

not sure if this if this is what its supposed to look like at 30 epochs

[brittle wing]

My datasets are Denoised I use the most aggressive denoise model

[tawny raptor]

ah nvm I just need to let it run longer

[tawny raptor]

I am not understanding this:

at the end of the graph its rising

does that mean the voice cloning is essentially untraining itself?

[thorn carbon]

overtraining

that lowest point is the best result you'll get from that training

by the way does my progress reset if i refresh the applio page?

cause i feel like it's stuck

[brittle wing]

Is possible for a model to train faster with a pretrain?

[thorn carbon]

🤷‍♀️

oh no applio crashed

guys does applio keep going from where i last left it if i start training the same audio with the same settings?

okay yeah it does

[low shard]

Tell me ur PC GPU, CPU, and a screenshot of ur wokada in #🔍│help-w-okada since this is the wrong channel

I'm guessing with VCS you mean realtime voice changet

[low shard]

Yeah since much time, we encrypted the code so it doesn't get detected

But tbh Kaggle offers more GPU time and better GPUs,just harder and requires phone number

What's ur PC GPU btw

[simple ore]

how did you manage to do that?

[lone nest]

What does this mean?

Starting preprocess with 2 processes...
0it [00:00, ?it/s]
Preprocess completed in 0.01 seconds on 00:00:00 seconds of audio.
Backup Complete: 1 new, 0 updated, 0 deleted.
Files are up to date.```Still getting this on preprocessing ![skullsob](https://cdn.discordapp.com/emojis/1159372531992645662.webp?size=128 "skullsob")

[simple ore]

you did not provide a path to the dataset

not .zip

process effects is a butterfly filter + normalization, with a prepared set you dont need tha

[lone nest]

I did

The path to a .flac

Even on noUI

[azure patio]

which site should i use to make covers with my trained models?

[simple ore]

Does a guide tell you to provide a file path or a folder path?

does noUI has .wav as default input?

or does it have a folder?

[low shard]

what's ur pc gpu first ?

[thorn carbon]

i clicked the x button

Guys is it normal that my epochs are taking around 1 hour per epoch? Or can i do something to make it more optimised?

[simple ore]

depends on GPU and dataset

[thorn carbon]

alright

[simple ore]

seems that you're training on CPU

[thorn carbon]

oh

how do i set that?

oh i see it

yeah i'm on cpu

would it be better if i chose gpu? I have a gforce gtx 1660 super

6gb

[simple ore]

it will be 10-20x faster

as long as you dont use a crazy high batch size

[thorn carbon]

oh wow

okay i will do that

thanks for the help

will it use my gpu now?

or do i make it "o"

like this

[knotty moth]

GTX 1660 super is still capable but can only use full fp32 precision without tensor cores

[thorn carbon]

i have no idea what the second half of that sentence means

[thorn carbon]

?

ooh i see the settings for training now

do i turn this on?

okay i got it

[knotty moth]

google is your best friend, bud

[thorn carbon]

right

so should i turn on tensor cores?

[knotty moth]

by getting a decent RTX gpu

[thorn carbon]

oh

nah that's out of my reach i'll just stick with my thing

[simple ore]

it must match the gpu, so 0 (zero), not lowercase O

do not cache, you only have 6GB vram

[thorn carbon]

oh i thought that was a lowercase O

[thorn carbon]

oops

[wheat pewter]

I’m extremely new to all this, I got a 7800xt gpu and a 5800x cpu, my friend helped me setup everything so I’m not sure what a lot of the things mean

[low shard]

alright, send me a screenshot of your wokada in #🔍│help-w-okada since I'm guessing you want a realtime voice changer

[wheat pewter]

Alright will do

[wheat pewter]

It says you are only accepting messages from friends

Cause i cant post pictures or screenshots anywhere in the server for some reason

[knotty moth]

!give-media-perms 30m @wheat pewter

[wheat pewter]

Ty

[tender walrus]

Anyone know how to improve the dry burned out high notes like this clip? I can't get any luck with high notes and sustained high notes, it always ends up sounding like the speaker got cursed by evil pharaoh magic

[low shard]

no need to dm me

[odd shale]

Nope, there's no way to improve very high notes on hifigan-trained models. (RVC's OG pretrain limitations)

I'm not sure if you were also testing your model with a sample and that sample got backing vocals/harmonies you didn't remove

[tender walrus]

It's just the last one I did, it does this on all files including clear single voice clips with no backing or reverb

[odd shale]

Welp, it's very likely that the voice breaks on very high notes due to hifigan's limitations

There's sadly no way to fix this.

[simple ore]

the fix is to adjust the pitch in the inferred audio a bit

[odd shale]

Oh that's right. That can be a alternative.

[carmine ether]

Is voicesona legit and a good replacement for other ai voice tools

[thorn carbon]

guys i keep getting this weird question mark what does it mean?

[low shard]

Never heard of that before

For realtime voice changers we suggest Wokada which is open source and free

[carmine ether]

See, im trying to get into vocaloids n stuff and voicesona got recommended. But on the first look it just seems to be too good to be true. This is why im asking

[tawny raptor]

Think its overtraining? Development has gotten stale after 4k-5k steps

[low shard]

Vocaloids are legit and actually a good thing, it's used to make songs tho not voice clone or realtime voice changers

[tawny raptor]

I got it actually, its been stale for a long while now I know for sure

any of you RVC-Project users out there

do you know how to stop training?

[quick fern]

Does anyone know why suddenly none of my voice models can pronounce l's or r's anymore. Certain words that I never had a problem with aren't able to be said with the VC anymore

[brittle pulsar]

@round fog

[rough cairn]

I've created a ChatGPT bot that provides complete RVC support, including development. You can find it in the "Explore ChatGPT" store under the name "RVC Assistant."

[novel wadi]

WHAT DOES THIS MEAN

it does it everytime i start up

[low shard]

You're using an outdated original wkada and also the wrong channel, tell ur PC GPU in #🔍│help-w-okada

[unkempt sapphire]

are RTX cores used for the ai voice changer?

[mossy reef]

any recomendation for cloud okada

[upbeat breach]

i just reload without blahblah

[lone nest]

Oh the guide said file path but now I'm trying it seems to work with folder path

[lone nest]

What's a butterfly filter btw? I can't find anything about it

[iron cobalt]

why do I get one step every epoch when training my own voice model? I'm trying to train

[lone nest]

Starting preprocess with 2 processes...
  0% 0/1 [00:57<?, ?it/s]
Traceback (most recent call last):
  File "/content/Applio/rvc/train/preprocess/preprocess.py", line 269, in <module>
    preprocess_training_set(
  File "/content/Applio/rvc/train/preprocess/preprocess.py", line 241, in preprocess_training_set
    audio_length.append(future.result())
                        ^^^^^^^^^^^^^^^
  File "/usr/lib/python3.11/concurrent/futures/_base.py", line 449, in result
    return self.__get_result()
           ^^^^^^^^^^^^^^^^^^^
  File "/usr/lib/python3.11/concurrent/futures/_base.py", line 401, in __get_result
    raise self._exception
concurrent.futures.process.BrokenProcessPool: A process in the process pool was terminated abruptly while the future was running or pending.
```New error ![boohooh](https://cdn.discordapp.com/emojis/1176674698629750975.webp?size=128 "boohooh")

[low shard]

first tell ur pc gpu in #🔍│help-w-okada , this is the wrong channel

[lone nest]

Seems to be dataset size problem, it works on a smaller dataset

[lone nest]

Hm Applio is a bit slower than Disconnected, what was ~1:15/epoch on Disconnected is ~1:45/epoch on Applio

[raw stirrup]

who fixed rvcdisconnected?

i was about to work on it but its fixed!#

[low shard]

Kit Lemon foot #📰│dev-updates

[low shard]

Don't even bother to fix it, it's at the end of life

It uses old mangio rvc

It's slower

It won't ever get maintained again

[raw stirrup]

i use rvc mainline local anyway always predered it over applio etc

[low shard]

It has no pros against mainline rvc and applio

[raw stirrup]

prefered*

[low shard]

Both mainline and applio are better than rvc disconnected

[raw stirrup]

i might change the code to make it mainline instead of mangio

[tame mica]

i dont think thats how it works

[raw stirrup]

im gunna re do all the code and make it rvc mainline

dont worry

i never bothered to look before to realise it was mangio

[raw stirrup]

looks like their banning youre mainline rvc colab notebook Runtime disconnected
Your runtime has been disconnected due to executing code that is disallowed in our free of charge tier. Colab subsidizes millions of users and prioritizes interactive programming sessions while disallowing some types of usage as outlined in the FAQ. If you believe this message is in error, file an appeal. Please include any relevant context about your usage.

this is the notebook for mainline that u guys suggested but they wont let me use iit lol

[low shard]

which notebook are u using? send link

if u are modifying rvc disconnected, well it gets detected u are putting mainline

hina's mainline rvc is encrypted

[raw stirrup]

im using youre rvc mainline colab

[raw stirrup]

https://colab.research.google.com/github/hinabl/RVC-Online/blob/main/Mainline_Colab_Full.ipynb

i set the auth token etc and went to click the link for local gateway and it gave error and then i got the message it was code thats not allowed and disconnected me

[low shard]

@glad zealot

[low shard]

could u also show a ss

[glad zealot]

pain

not in the ban list

¯_(ツ)_/¯

[raw stirrup]

yeah 2 min

[young halo]

What does this mean

[raw stirrup]

weird it loads now but it keeps crashing when trying to load the dataset audio to train, ive tried flac and wav format

[brittle wing]

hey makaveli

can u help me with smh

my voice changer is not detecting my voice when i test it in discord playback voice

[raw stirrup]

i dont use the voice changer sorry dont know much about it,

[brittle wing]

oh ok

[cunning isle]

please tell me how to get a model

how to download

[knotty moth]

as said #✨│ai-help message

[lavish lintel]

Congratulations chibi hater!

Your Gengar is now level 37!

[dire atlas]

anyone know why the recording while using a voice changer is super choppy and jumpy, actually ineligible. i wish i could attach a photo of the audio waves lol

[tardy copper]

how could i use rvc gui with an AMD gpu?

[low shard]

which google colab link are u using

[low shard]

rvc gui is old dont watch yt tuts

what are u looking for

[knotty moth]

again read this before posting here #✨│ai-help message

[tardy copper]

hm?

[low shard]

what's ur pc gpu and what do u want to do

[young halo]

HinaMod

[tardy copper]

got an rx 550 4gb i just wanna make some simple songs using models

[simple ore]

just no

too weak

CPU inference wont be much worse lol

[young halo]

I didn't send the link directly bcs I'm not sure if they're allowed here or not 🧍🏻‍♂️

[tame mica]

why wouldnt they be

[young halo]

I don't know, I've seen a lot of servers not allowing links so i just wanted to make sure

[tame mica]

girlypop thats a colab link it aint harmful unless you mess up bad 🧍‍♂️

[young halo]

Sorry, I'm just paranoid 🧍🏻‍♂️

[raw stirrup]

can someone please explain these 2 settings please is it better for me to lower them for better timbre and tone for 2pac or make the settings higher

"c_mel": 45,
"c_kl": 1.0

c_mel im told if its lowered it will produce better timbre

but needed more info on the 2 settings

[simple ore]

they are loss multipliers

c_mel is for loss/mel

dont touch kl

higher the vaue is, more the generator gets punished for the mismatch vs real data

[raw stirrup]

oh so neither of them affect the accurateness of the model

[low shard]

which? send link

[simple ore]

mel loss is a difference between mel spectrograms for original audio and generated audio

so it is the accurateness of the model

[raw stirrup]

so lower it dont make it higher?

and whats value should i not go below

[simple ore]

higher multipler, more punishment for the mismatch

[raw stirrup]

more punishment but has better chance of model accuracy ?

[simple ore]

maybe yes, maybe no

[low shard]

well
Your AMD GPU is kind of good enough to do inference (use models) locally (on ur pc), you won't be able to train (make models) but use them

You can:

• Locally (runs on your pc so the speed depends on that, you will have to set it up with the guides):
  • Applio (AMD Windows) : A fork of RVC with some extra features like Applio TTS, kinda faster and simpler but same quality tho
  • Mainline (AMD Linux/Windows) : The original RVC
• Cloud (remote good pc, easier and faster than ur PC but it's limited):
  • Ilaria RVC Zero: fastest and simplest that you can get for free
  • Weights.gg: Partnered with AI Hub, lets u do them easily but u may be in a queue
  • Applio Colab: max 4 hours daily, not granted, of GPU

Easiest possible (automatically separates vocals & instrumentals) : weights.gg
easiest cloud: Ilaria rvc zero
easiest local: Applio

[simple ore]

punish too much and it will start cheating

[raw stirrup]

im going based of internet

[low shard]

tbh just use cloud

[raw stirrup]

Lmao

[raw stirrup]

this is what i was guided to do since im training 2pac models 3️⃣ Adjust Training Settings
Lower KL Divergence (c_kl): Reduce c_kl (e.g., 1.0 → 0.5) to prevent excessive generalization.
Higher Batch Size (-bs): If Colab allows, increase batch size (6 → 8) for better stability.
Use F0 (-f0 1): Keeps pitch accuracy, making it more voice-like instead of robotic.
Pretrained Model: Fine-tune on a good base model (e.g., one trained on multiple rappers).

[simple ore]

that's nonsensical advice

kl loss is ~1

it barely has any effect comparing to fm loss 10+ and mel loss 20+

kl loss affects the encoders for phonemes and spectrogram

[raw stirrup]

this is my loss_fm=8.531,

and

loss_mel=14.108,

roughly in these ranges stable

[simple ore]

the advice above is nonsensical

pitch guidance should always be on

higher batch size not always leads to better results, highly dependable on the dataset and its size

[patent pasture]

МММ ёпта

[crude flame]

where on earth did you get this from?

[meager comet]

how do i download a model at a specific step on rvc applio noui

[young halo]

Here!
https://colab.research.google.com/github/hinabl/AICoverGen-Colab/blob/main/Hina_Mod_AICoverGen_colab.ipynb#scrollTo=NEglTq6Ya9d0

[meager comet]

lmao

[red kayak]

chatgpt

[raw stirrup]

github

someone replied to an issue simmilar to mine and that was their as a response

but it turned out lowering c_mel actually is better my loss_mel went from 15 16 to 10 and under

and fm_loss under 7

experimenting is always useful

[simple ore]

yes, it simply lowers the number

It did not get any better

loss_mel = fn_mel_loss(real, fake) * c_mel

so you see a lower number and the generator gets a feeback "the difference is not that bad"

instead of "dude, what the hell you're making?"

it is like looking at the thermometer and multiplying the number by 2, then waking outside and wondering why it is so cold

[raw stirrup]

if its fake numbers and useless why is it implemented into the graph to be viewed as useful info

?

confusing

i was also watching the tensorboard it did get much better

from the graph

[meager comet]

does anyone know how to do this

[raw stirrup]

go to where the weights folder is and download the weights u see

[meager comet]

no weights folder

what are the g and d files

don't they have something to do with pretraining

[raw stirrup]

assets folder?

its either in assets/weights

or weights and youre missing it

[meager comet]

no pth files in assets

[raw stirrup]

there has to be a weights folder for it to even save models

is there a weights folder in the assets folder?

[meager comet]

i don't know how this one works but it saved to the large g and d files

and exported the final model only

[raw stirrup]

im not familiar with applio as i use rvc mainline but usually its always weights folder g and d files are just from pretrains

basically the index

[meager comet]

it's saying it saved to the g and d files

i don't know what that means

[raw stirrup]

do u see logs folder

[meager comet]

yes

that's where they are

[raw stirrup]

ok u found the weights?

[meager comet]

no the g and d files

[raw stirrup]

screenshot the thing where it says it saved to g and d

and send it here

when it says this INFO:makaveli_v5:Saving model and optimizer state at epoch 140 to ./logs/makaveli_v5/G_2333333.pth
INFO:makaveli_v5:Saving model and optimizer state at epoch 140 to ./logs/makaveli_v5/D_2333333.pth

thats just the checkpoint

not youre actual model

[meager comet]

so what is it saving

i'm confused

[raw stirrup]

u need to find the weights folder

thats a checkpoint

whats youre model name tesT?

[meager comet]

so what exactly is it saving

[meager comet]

yes

[raw stirrup]

theres nothing in the test folder other than d and g file?

[meager comet]

correct

and the final model

[raw stirrup]

did u make sure to enable save at each checkpoint

a model

[warped thistle]

Is there a best all-round model to sound like an e-girl in english?

[meager comet]

i am assuming this is it

[meager comet]

so what is the checkpoint for i'm pretty confused

[raw stirrup]

the checkpoint is to save youre progress of the model

its basically acting like u havent set a save every so many epoch model

for me it saves every 20 epoch

[meager comet]

why isn't there an automatic feature that takes the model with the lowest loss/g/mel and saves that specific model only

[raw stirrup]

once im past 200 epoch i set it to every 50 or 60 epoch

[meager comet]

that seems like it would make the most sense

[raw stirrup]

its applio

i cant stand it

[meager comet]

is there something that does it automatically?

[raw stirrup]

youll have to ask a applio expert lol

[meager comet]

a fork?

[analog obsidian]

Because lowest mel or lowest g total is not always the best sounding epoch

[meager comet]

what is then

how do you measure the accurateness of a model

[analog obsidian]

Random

lowest mel or g total is a starting point to find it

[meager comet]

i mean i don't know

the lowest g/mel in here has like half the loss of the final model so

i'd imagine it's probably a lot better

[raw stirrup]

do u use the tensorboard? to monitor youre model loss and overfitting

[meager comet]

yes

[raw stirrup]

then when u train in the tensor board type g/total loss and that graph will show u how youre model is doing

[meager comet]

yeah i know

i just want to be able to get the step with the least g/total and save only that model

but there is just no option to do that

and my model's loss spikes at the end

[raw stirrup]

applio is different im a proud mainline user

smooth it out

to 0.99

and resend

[meager comet]

it is down sloping

[crude flame]

use the avg graphs they are more accurate

[meager comet]

accurate to what?

[crude flame]

more accurate to how the actual epoch preformed

current graphs are just logging the ending of the epoch

[meager comet]

isn't how the actual epoch performed just that point

[crude flame]

not everything else

[meager comet]

i'm confused

what is the point for if the average is what it is actually doing

[analog obsidian]

lowest mel is the clearest epoch which is fine and most of the time its good enough, but results are random, so like sometimes lowest mel has an issue other epochs dont have

[analog obsidian]

Avg loss are more accurate graphs

the values in the old graphs are innacurate bc it only logs the last step of an epoch
avg losses fixes that

[crude flame]

here is an image codename made to help, if you scroll up there is more info

[simple ore]

loss value is a measured difference between the original audio and the generated audio

mel loss is a difference calculated using 'mel spectrogram' method

the difference is given to the generator and it tells it how far it is from the original

there are 4 loss values - fm, kl, mel, gen, each contributing to the adjustments the generator makes in order to perform better

by multiplying individual losses by 'weights', you tell the generator which difference is more important than others

[thorn abyss]

yo, i was wondering cs i never tought about this, does loudness matter while making a dataset? for example, i got a vocal of a rapper and he starts mumbling, or rapping in a low voice, should i include this on the dataset or only keep the actual loud voice?

[simple ore]

loudness does matter, a dataset with screaming, normal speech, whispering will do better than a dataset with a monotone voice

[hazy cradle]

hi everyone, outside of chunk, what settings can I change if the voices sound rather "robotic"?

[simple ore]

👉 #🔍│help-w-okada

[hazy cradle]

a h, thanks

[errant wyvern]

I downloaded some voices with the ending npy, do you know what these files are?

[young halo]

Can someone please try this model out for me? It's not fully trained yet, i just wanna know how it sounds so far, but the colab that i use for covers is not working
https://huggingface.co/ivoryprincemilo/WILLOW_COPINGMECHANISM_era/resolve/main/willowcopver1.zip?download=true

[simple ore]

those are not voices or voices for something other than RVC/Okada

[glad ravine]

any good way to get the output of okada into a phone call? My phone is android

[sand pond]

hey so it works but its not coming thru my virtual cable, but when i go to the rvc it said to check cmnd center for help

[ancient current]

is there any tutorial on the exact steps you need to follow to get RVC working? everytime i try to use it, it always bugs out and i dont know why

im always stuck on the screen where you have to run the base64 code n all that stuff

[steep cobalt]

Can I use this ai version that I just uploaded without having to download it separately? It's my first time using it, and it's quite difficult.

Hello everyone I'm trying to use this vocal ai for the first time, is there a place where it's easy to explain?

[young halo]

What do you want to do exactly? A model or a cover?

[steep cobalt]

I'm a composer. I'm trying to turn a file I recorded with my voice into someone else's

[young halo]

Ok, so it's a cover then

Do you have the voice model for that voice?

The one you want to turn your own into

[steep cobalt]

I want to use the kpop idol voice in ai hub.

[hot lagoon]

My not sure where to put my Authtoken

[young halo]

Which idol is it?

[steep cobalt]

nct mark!

[young halo]

Got it! Do you have the link to his voice model? If not, i can search it up for you

[steep cobalt]

#1175584660730945577 message

I think it's this!

[young halo]

Yeah that one, ok

So

Let's try with this colab
https://colab.research.google.com/github/hinabl/AICoverGen-Colab/blob/main/Hina_Mod_AICoverGen_colab.ipynb#scrollTo=NEglTq6Ya9d0

Run all the cells

[hot lagoon]

any help?

[young halo]

Idk anything about ngrok, sorry 😔

Maybe there's something out there on the AI Hub documents

[steep cobalt]

I tried to use this in CoLab yesterday, but I've finished inserting the mark model. But I tried to route the song and convert it to my file, but I don't know where the changed file is.

[nocturne mural]

Replace token

In the cell is commented the link where you should get your ngrok token, which is obtained by registering in their page.

[harsh nebula]

Hello! I need help, please. When I try to work with colab NO UI to make a cover, I receibe this error after pasteing the link of the song: Exception: 'NoneType' object has no attribute 'setdefault'. What is the cause? I never had this problem before. Thanks

[young halo]

This Is happening to me as well...

I guess it's a problem with the collab itself

You might have to try this one
https://colab.research.google.com/github/iahispano/applio/blob/master/assets/Applio.ipynb

For this one you put acapella only

[wide perch]

As soon as I begin audio conversion, the entire process freezes and the command prompt is empty

Other people I talked to had this same issue

Anyone know how to fix it?

[wide perch]

Nvm fixed it

[faint pasture]

how to select only Japanese language?

[low shard]

like in #1175430844685484042 ?

use the japanese tag

[brittle wing]

Lead vocals*

[young halo]

On which colab?

[brazen spear]

How to fix this?? (Sorry for bad eng)

[simple ore]

probably Cuda Toolkit is missing

https://developer.nvidia.com/cuda-12-1-0-download-archive

[brazen spear]

Thx man

It the same man i install it already

[simple ore]

https://github.com/lucasg/Dependencies/releases/tag/v1.11.1 - download x64 release.zip, unzip, run dependenciesgui.exe

open the file in question

it shows which dependencies are missing

this is normal

so most like you're also missing VC Runtime

https://aka.ms/vs/17/release/vc_redist.x64.exe

[brazen spear]

Thx man

I will try

Im here so what now

[simple ore]

and you need to actually locate the flagged missing dll in the list

[low shard]

@proven hill u might also wanna check this since it's ur project

[knotty moth]

typical social media users who are less tech savvy

[brazen spear]

Sorry i have no intelligent on this kind of stuff

I will printscreen now

[simple ore]

find what is flagged by

[brazen spear]

here

[simple ore]

collapse

resize to see the full name

scroll down to see others with red warning

[brazen spear]

i see only one red warning

[simple ore]

I did not say collapse the entire thing

I said collapse the second level

and resize

[simple ore]

expand

[brazen spear]

i think the missing is C:\pinokio\api\Ilaria-RVC-Zero.git\app\env\Lib\site-packages\torch\lib\c10.dll

[simple ore]

hm

well, you're using pinokio, so it is on you

[brazen spear]

yep

[low shard]

maybe wait for Ilaria's response

[brazen spear]

ok

[simple ore]

or you can just download Applio, unzip and you're good to go 🙂

[hallow thistle]

Are you using Ilaria RVC from Hugging Face? Because the Hugging Face one, the author didn't make it to run locally.

[low shard]

she did actually

#📰│dev-updates message

[hallow thistle]

Oh, so mine is still an old one, which runs off from venv out of pre-installed Python.

[random dock]

I joined this server today.

I'm Japanese, so I'm not good at English.

I'd like someone to tell me how to use this RVC voice changer.

[low shard]

realtime voice changer for calls?

Also, what's your PC GPU?

[random dock]

I want to use a voice changer for phone calls.

Can I download and use the voice of Son Goku, for example?

My GPU is NVIDIA GeForce RTX3060

[low shard]

alright,

RVC is for Inference (use models) on pre-recorded audios, and training (making) models

Wokada is for RVC in realtime for calls

so the right app for you is wokada, not RVC

use #🔍│help-w-okada

[random dock]

Can you send me a video of the steps to installation?

[proven hill]

whats the issue

[low shard]

there is no video

all of them are outdated

[karmic oxide]

i dont think rvc has even changed enough to the point where you can say that. only thing that really changed was pushing applio which i dont know why you guys did that, you literally said yeah, lets have something that works for literally everything and lock it down to its one ecosystem where the devs want the end users to fix their code.

[low shard]

He can't install Ilaria rvc via pinokio

[proven hill]

cuda is missing afaik

[low shard]

@brazen spear said he downloaded cuda

[proven hill]

restart pc

[low shard]

First of all he was talking about Wokada, that has nothing to do about Applio

Second of all, yes the code changed, the Wokada deiteris fork gives better performance with less power usage, and better quality, proof: #🔍│help-w-okada message

3rd of all, even for RVC the code changed too, Mangio RVC is way slower than Applio, Applio is way more optimized

I feel like you're saying this just because you don't like that rvc disconnected doesn't work anymore, but you have to realize that program needs to update to be better, you can't always use the same abandoned program forever and hope for best performance and no issues

We are pushing mainline rvc, applio and deiteris fork because they are more updated and have optimized performance

[karmic oxide]

nick u do know who i am right

[low shard]

Never heard of u before

[karmic oxide]

ok check the original okada guide

and since you want to push to okada now ill take apart the entire fork guide right now in real time

[low shard]

Damn u changed name again?

[karmic oxide]

lets start with the known working settings page

the entire purpose of known working settings is not to use 100% of your gpu so people cant even do simple things like opening a browser without it lagging

which every single one you guys have on that page is doing exactly that

[low shard]

Ohh ur that person

[karmic oxide]

we then are pushing that its better than a "beta version" which was performing worse than the original okada prebuilt