✨│ai-help

[brittle wing]

wait is there no good notebooks rn?

to train on

[junior gull]

do you have the link to both I might jsut need to check them both out and see which works better for me?

and really appreaciate the feedback!

[glacial pollen]

No idea, I don't maintain any so possibly wouldn't know

[brittle wing]

alr thanks

[glacial pollen]

Sure!
For applio:
https://huggingface.co/IAHispano/Applio/blob/main/Compiled/Windows/ApplioV3.2.9.zip
( precompiled, use just run-applio

For the fork:
https://github.com/codename0og/codename-rvc-fork-3/releases/download/3.1.5-rev1/Codename-RVC-Fork-V3.1.5-rev1.zip
( you'd need to run run-install first, then run-fork )

Gluck ~
( and again, fork's meant for advanced users so, not gonna go through the hassle of explaining each thing one by one (( best you'll get is the from-ui descriptions

[junior gull]

Thank you again!

[glacial pollen]

Np! Gluck

[upbeat carbon]

seed vc

[glacial pollen]

nevermind then

[brittle wing]

yall know what could be the problem the voice changer doesn't send any audio to the audio cable but it works in monitoring

[glacial pollen]

-howtoask

[patent trellis]

How To Troubleshoot

__**GIVE CONTEXT.**__ 📝

• Don't simply mention your issue, like "my rvc is not working".
• Describe the step you are on, what you're trying to do, the RVC you're using, a screenshot, etc.
• The more context, the better.

__**BE POLITE.**__ <:matsuripray:1159685390156967936>

• Don't be desperate. You can ping a Helper, but if they ignore, they aren't available/don't know the answer.
• It's okay if you're frustrated, but don't take it into this server.
• Don't DM without prior consent.

__**BE PRODUCTIVE.**__ 🤝

• Don't ask for every little instruction. Put your own effort & test things by yourself.
• Don't ask to ask.
• Check if your answer is a Google search away/on our guides website.

If you've changed your mind about your message, feel free to delete/edit it.

[signal spire]

Does anyone here knows about replay.ai? I recently download it. Currently I am training my voice, does it usually takes too much time to finish?

[tulip cloak]

Just wait, someone will get to you soon. Everyone sleeping

[signal spire]

Aight, man. Thanks for your help

[craggy bough]

what specs do you have

[signal spire]

I'm using
ryzen 5, 4500
16gb ram gr4
B450m assrock
GTX 1060 6GB GDDR5

Is is too low? That why it take too much time to covert my voice?

[craggy bough]

yeah thats gonna take quite a while

[signal spire]

I see. I been here waiting for about 3hrs now and it still in 55 epochs

[craggy bough]

i wouldnt bother doing any training with anything below a 3000 series

[signal spire]

How many epochs does it usually take to finish?, my file that I sent is only 3mns long

[signal spire]

I see. Bob in yt said the same thing hahaha

[craggy bough]

not sure tbh

i never tried

someone on weights is stuck at 1100 epochs

[signal spire]

No way hahaha. He's already 1100 and still can't proceed. I'm kinda hesitant now to wait

[brittle wing]

like voice changer litterally just stopped working 1 day and never worked again i tried reinstalling the virtual cables and still didint fix it even reinstalled the whole voichanger and still doesn't work so idk what to do more

[hallow thistle]

Can you send the full screenshot of your W-Okada?

[knotty moth]

RTX 20-series should be okay

[lucid creek]

hello there there way use GPT-SoVITS on cloud { kaggle} without ui

[brittle wing]

the terminal or site?

[hallow thistle]

The website (GUI).

[brittle wing]

alr give me 2mins

[junior gull]

finally got Applio up and running...mostly....but the UI anytime click anything "connection timed out"
Also missing hubert_base.onnx which I can't find ANYWHERE or convert becuase of fairseq issues with windows 😵‍💫
it seems like no matter which I chose....there is some singular file that is like chasing a unicorn....so many dead links and repos where people say to download from.
Is there another way to get it? or convert it? Sorry if these are dumb questions...I am still very new to this >_<

[golden walrus]

Did you turn off the cmd? The black board thingy

[junior gull]

def no no . i know that much 😅

i had to resintall allthe modules like 50x....eachtime i isntalled one...it updated a handful of others to incompaitlbe versions.....like dominos spreading out lol....eventually got them all back to the right versions and webUI launched, but i cant initiate any tasks or it just gives connection timed our errors

i'm getting to the point where i will jsut throw money at someone to give me a working copy of those files 😑 just so i can move on lol

[golden walrus]

I had the same time like you

Still have these issues

But you were able to open the webUI right?

I don't know much but most of the time, those issues will pop up in these cmd

[knotty moth]

that means closing the program

[junior gull]

i know that if i kill the cmd promt that is running it ...the webui will die. RVC was the same way

but it was still going

RVC v2 worked great...until the very last step...because i didnt' have a hifigan.pth file. And in Applio the only error I can identify atm is poitning to needing a missing resourcer hubert_base.onnx and the ui crashing. 😓
Is anyone willing to DM those files?

its so weird that they seem to be so critical...and i see lots of posts online of people looking for them....but they dont seem to BE anywhere

😵‍💫

[golden walrus]

Wait, you use applio to train right ?

[junior gull]

i tried both. RVC i was able to almsot fully train wiht but it got stuck on the index training part because of hte missing file hifigan.pth.
applio i cant get to even go that far becuase of missing hubert_base.onnx and the UI constantly timing out almost instnatly sometiems.

[golden walrus]

https://huggingface.co/IAHispano/Applio/tree/main/Compiled/Windows

download 3,2,9

[knotty moth]

show the console window

[junior gull]

#✨│ai-help message

Yeap thats the same link from earlier

kk sec

[golden walrus]

Oh, it is called console window

[knotty moth]

if you have closed it, you'd have to run the program again

[junior gull]

a

sec i dont know why it was all zoomed in

there we go

and also before but not present atm the error mentioning the missing resource file hubert_base.onnx

[knotty moth]

try moving to another directory path without spaces

[junior gull]

i'll have to give that a try this weekend maybe.....the idea of reuibiding everything a 5th time makes my head spin right now lol

[simple ore]

you're using some outdated af version of applio

that you also installed from requirements instead of downloading the compiled version

so now you have a bunch of incompatible libraries such as gradio vs pydantic

a stupidly long path does not help either

if you got this whole thing as a package from somewhere

it is not up to date

[winged ember]

:loss_disc=3.788, loss_gen=3.481, loss_fm=10.017,loss_mel=20.140, loss_kl=1.865

what does these means?

with some colab there's like a built in chart where we can read n understand easily but I'm not sure how to read these when training locally.

[simple ore]

what are you using to train locally?

mainline does not come with tensorboard, but you can install it manually

[winged ember]

Ah yes tensorboard that's the name haha, Im trying to train voice models. I'm not sure if I can just read it from that or is it better to have tensorboard? If so can you point me to where or how i can install it Thanks! btw

[lucid creek]

guys how i know if model is overtrained in cloud {kaggle}

[simple ore]

you need to tell me what are you using to train a voice model

mainline rvc, mangio or some other outdated stuff

[lucid creek]

hello anyone help

[winged ember]

Im sorry I misread you last message, Im using mainline i believe

https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI

this is where i got it from

[winged ember]

I think you need to have tensorboard but idk how to use it on kaggle

[lucid creek]

yes but how i read it 🙂

[winged ember]

you can read when your model has enough training by looking at if the bar keeps goin down. At some point it will not go down and may go up. That's the point of overtraining from what I understand

[simple ore]

mainline has environment in realtime folder I believe, so realtime\python -m pip install tensorboard

then realtime\python tensorboard --logdir=c:\path\to\where\logs\are

and I beleive the mainline requires some editing to config.json in the model's folder or somewhere else to set the logging frequency

[winged ember]

alright, I will try this thanks!

[lucid creek]

thx dude

[brittle wing]

y'all i been having this issue on kaggle since yesterday is kaggle broken still?

Traceback (most recent call last):
File "/kaggle/working/program_ml/app.py", line 1, in <module>
import gradio as gr
ModuleNotFoundError: No module named 'gradio'

i ran everything as it should be ran but this happened

[devout furnace]

What is the best voice changer?

[dry owl]

Hey! i have the rtx 5080 what ms should i set the chunk at? for the 3090 it says 72 ms chunk + 2.7s extra what should the settings be for this gpu while gaming?

[simple ore]

remove highlighted part

also in install cell

and re-run the pretrain load line

[brittle wing]

okay thank you

[simple ore]

!python core.py "prerequisites" --models "True" --exe "True" --pretraineds_hifigan "True"

or make it a new cell

[brittle wing]

okay so delete this and create a new cell with that new code u sent

?

[simple ore]

just create a new cell

[brittle wing]

ok

[brittle wing]

where should this cell be?

after the installation or before?

[dense drift]

Why is it taking so long?

[simple ore]

if you dont have anything installed yet, it is easier to remove that highlighted code from both cells

if you have it instaleld already, make the cell anywhere

does not matter

[brittle wing]

okay thank you

i'll try it rn and lyk if i have any more issues

@simple ore

whats the issue now

[simple ore]

same issue -requirements install happening in a wrong place

[brittle wing]

hmm but i did copy the code that u gave me

look

[simple ore]

did you run 'setup ngrok' cell at all?

[brittle wing]

yes

[brittle wing]

i ran it again with the other method and its happening again

[simple ore]

delete yours, use this

[brittle wing]

okay

@simple ore

[simple ore]

run every cell

[brittle wing]

again?

[simple ore]

duh

[brittle wing]

lol

@simple ore lfggg it worked

[brittle wing]

@simple ore tensorboard isnt detecting the logs so its not letting me check the scalar graph and such to know when its gonna overtrain

[sick moth]

hi

[latent kettle]

Hlo

[latent kettle]

The same thing happened with one more user. @simple ore

[simple ore]

no idea, what's in /kaggle/working/program_ml?

[brittle wing]

where can i check that

[simple ore]

i have no idea, i dont use kaggle

should just be the content of applio including logs

the folder tensorboard checks

[brittle wing]

okay

okay i found it

there's a file called

run-tensorboard.bat

[simple ore]

no

that's for local

i mean there should be logs folder

in the logs folder thre's your model folder, inside is eval, inside eval is events.out file

[brittle wing]

okay yeah

that file is there

[simple ore]

so refresh tensorboard then

[brittle wing]

it's not working even if i refresh it

[brittle wing]

yes i know but still it doesnt show anything

[simple ore]

how big is events file?

[simple ore]

🤷‍♂️

all looks good

[brittle wing]

do u want the tensorboard link

so u can check?

[simple ore]

okay

well, delete that

in the start cell you can try changing the line to %tensorboard --logdir /kaggle/working/program_ml/logs --port 8077

[brittle wing]

shittt so i gotta start over?

[simple ore]

the training should resume from the last save

if you stop the cell, make the fix and restart

[brittle wing]

ahh yea thats true

[simple ore]

this is weird

it does cd into /kaggle/working/program_ml, so anything that starts after should be from that local folder

tensorboard and app.py

app.py works as expected

so weird tensorboard does not

[brittle wing]

so will this fix it or not?

[simple ore]

i dunno

[brittle wing]

cause i swear it happened to me before but it didnt always happen

[bleak nymph]

so

i recently learned not too long ago that my voice technically doesn’t have a voice type

reason is because despite me being low (f#2 - f#3) i’m NOT a bass

because to be a bass your low note has to be at least an e2

so for the people who have no idea what the fuck i just said - my vocal range is very narrow

[bleak nymph]

and for higher notes i strain my vocals

reason i’m saying all this is because my question is

[analog obsidian]

bro singing-support is at the next door

[bleak nymph]

does this affect the voice model?

[bleak nymph]

😭 wasn’t about that

[bleak nymph]

.

[analog obsidian]

the model will only be able to do what you can do

it's not that hard to tell

[bleak nymph]

basically does me straining my high voice affect the ai model negatively?

[analog obsidian]

if your voice sounds very breathy then yes

[bleak nymph]

then how come when i put some vocals into the ai of me, it can actually hit those higher notes that real me will be no where near

[bleak nymph]

ah i see

[analog obsidian]

try the klm pretrain since it can generalize high notes

[bleak nymph]

yeah i tend to go breathy the higher i go but i’ve been tryna train myself not to do that

[bleak nymph]

oh yeah abt pretrains, how tf do i use em

[bleak nymph]

ah ty

[analog obsidian]

https://discord.com/channels/1159260121998827560/1339155300720054316

place g and d of the pretrain here

[bleak nymph]

so i also remember u telling me that the main thing of a dataset is to have my whole vocal range

[analog obsidian]

do a singing session of 30 minutes

and train that

[bleak nymph]

what if i sing a few song at the lowest of my range, the regular sitting point of my range and the highest point of my range, sould that help?

[bleak nymph]

oh 30 mins 😭

so sing an entire EP is what ur tryna tell me 😭

orrr i can set my playlist to shuffle and try singing along to it

[analog obsidian]

i mean u can just record different song clips then merge them in one big file

[bleak nymph]

while recording

[bleak nymph]

yea i did that once

to make a me singing dataset

but the model sounded crap

and not like me

so far the best one was the one of me talking really expressively (with a ton of laughing)

but i think i deleted that file

it was also the same dataset that just got stuck

according to you

the model just got stuck

at it

[analog obsidian]

just be sure your voice is consistent

not sudden timbre changes

avoid excessive denoising

if you have to denoise just do it right

[bleak nymph]

i don't denoise it

my mic stand make funny reverb noise

and i think it funny if the ai has it too

[analog obsidian]

welp i told u last time rvc kinda hates reverb

two biggest things rvc hates:
Raspy voices
room reverb

[bleak nymph]

nah because the vocal model actually worked

[bleak nymph]

oh

raspy voices?? why??

[analog obsidian]

because it's going to try training that regardless

it cant say no lol

[bleak nymph]

yeah but like the reverb nosies actually sounded like noise from my mic stand

irl

[analog obsidian]

f0s are ass at estimating the pitch with those type of voices

[bleak nymph]

in the ai

[bleak nymph]

oh yeah

pitch correction software always get confused at raspy voices

maybe that's why

[analog obsidian]

mhm try not sounding raspy in your recordings

[bleak nymph]

wait hold on

what's rasp again

cause i keep getting it confused with vocal fry

[analog obsidian]

f0s have a hard time inferencing voices like that

[bleak nymph]

ah right

my voice don't have that type of like rasp

wait is that ai 😭

it has a weird sound that is like similar to ai voice

[analog obsidian]

is a real guy lol

[bleak nymph]

wow

[analog obsidian]

lord i saw his yt videos

[bleak nymph]

the sounds at the end of each line especially the breathing sounds ai to me

[analog obsidian]

rvc brainrot

[bleak nymph]

lmfao

[junior gull]

Oh? I just used the link that posted abovr

https://huggingface.co/IAHispano/Applio/blob/main/Compiled/Windows/ApplioV3.2.9.zip

is this link bad? Or did i screw up unpacking it?
😓

[simple ore]

that one should be fine, unzip with 7-zip

into C:\Applio or some shorter path

[idle stag]

this is my second time training a model after a few months, I should use 32k as the sampling rate in this case right?

[junior gull]

And with another user pointing out an issue with my directory, I think that could be the issue why it’s not finding what it needs… so I tried to install them myself. I made it worse. It looks like. 🙃 thank you for pointing this out. I will try this again!

[mellow ermine]

How do I generate images with AI or train one? Is there a Google Colab for that?

[grim rain]

how to use rvc do u have a link or some ?

[odd shale]

You can either install RVC locally our use it in cloud services like Kaggle or Colab

-rvc

[patent trellis]

📚 RVC Documentations

AI HUB Docs

https://docs.aihub.gg

🍏 Applio Docs

https://docs.applio.org

[odd shale]

Read the docs above for further info.

[steep prairie]

heey

[old bridge]

Why can't i hear myself with the voice changer

[safe echo]

hey guys, is there anyway to get perf down? (im guessing this is "lag"?)
i keep reaching 140+ which in terms is dimishing voice quality

[analog obsidian]

change f0 to fcpe

if you're still lagging increase the chunk

[safe echo]

thank you! this decreaed it dramatically!

[analog obsidian]

fcpe has slightly worse accuracy than rmvpe and breaths may behave strange

[safe echo]

you're an actual god because i just realised this also.

its picking up waaay clearer

[analog obsidian]

is more sensitive to noise yep

i havent tried rmvpe models using fcpe as f0 often, so i have no idea how well it behaves

[wide olive]

what is the process for turning my audio sample into a .pth file?

And I can't seem to find any free software to do it

[analog obsidian]

https://docs.aihub.gg/essentials/how-to-make-voice-models/
https://docs.applio.org/applio/beginners/first-model

[wide olive]

ty

I'm basically working with a dozen or so short samples of my rpg character going "uh" and "um" and different action sounds.

Will that be enough?

[analog obsidian]

no, the dataset has to be natural speech or singing to give good results, don't expect good results with huh ah sounds only

minimum 10 minutes
recommended 30 minutes up to 1 hour max

[wide olive]

ahhhhh

[knotty moth]

there is already some mc villager model

though not too ideal as a serious model

[golden walrus]

Guys, i have one question, have anyone tried Seoul Streaming Station's pretrained for realtime voice changer? 😦 i used their pretrained and cloned my model for about 50 epoch, the result is pretty robotics somehow. Do i need to push it to 250 and see more result?

[glacial pollen]

You should use TensorBoard to evaluate the model's performance, rather than assuming a fixed number of epochs to train it for.

Other than that, until Noobies merges a specific change in applio main, you can try my fork as there's a rather deal-breaking change + tensorboard logging aligned with that change

[golden walrus]

i'm using ur fork, but i'm too dumb to know how to use tensorboard lah

[glacial pollen]

ohhh

well, then uhm

Update is up

you should update to 3.1.6-rev1

[glacial pollen]

and as for tensorboard, is it about running it or using? ( Reading the metrics

[golden walrus]

like, somehow it only has 4 instead of lots in applio

ah, i have issue in reading, i'm looking for a guide how to read these

but i still don't get it

oh 3.1.6 is up ? i'm still using 3.1.1

let download it

[glacial pollen]

Because logs on my fork and applio are not the same

the loggings differ
( especially now, starting with 3.1.6-rev1

one sec

[golden walrus]

but i swear to god yesterday i saw the learning rate chart

[glacial pollen]

that's how it looks rn

the " total generator " loss is not the main one you wanna monitor anymore

and the " total discriminator " -> " discriminator_adv

[golden walrus]

so i want to look at g and d right ?

[glacial pollen]

tl;dr:

total generator: describes metrics; mel, fm, kl

generator_adv: describes adversarial performance of generator
discriminator_adv: same as ^ but for discriminator

so in reality, the only change is that now you don't watch for total generator, but generator_adv

[glacial pollen]

more or less, adv for g and d + fm

if fm stays rather high ish, ( 11-9, maybe 8 loss region ) then that's fine, model is learning

if fm goes too low, it means discriminator got too strong ( no point training for longer

But then, soon ( hopefully ) docs are gonna get updated... and if not, well, guess I'll have to do my own on the metrics

Anyways. If you got lost, ask right away. Gonna simplify it more

[golden walrus]

thank you so much, kind soul

[glacial pollen]

✨

If you have access to #🔊│ai-development

( ai-testing ) dr87 wrote quite a lot of useful information ( on how to understand the losses, more or less

[golden walrus]

ye, i read some from him

[glacial pollen]

that should help anyone really to get a better grasp

[glacial pollen]

yup

Now, to straighten it up.. just in case..
the double update thingy does exist in applio, but in the spin branch

[golden walrus]

i swear to god dude is a walking research center to me

[glacial pollen]

lol

just take your time

[golden walrus]

i will try to use spin and use it in his vonovox

i have no idea if index file is necessary, because i don't see the option to use it in vonovox

[glacial pollen]

well... generally index is rather something one should, imo, avoid

in voice changers, that is

if the model is / was trained poorly or perhaps, if not on a lot of ( diverse ) data aaaand, it will happen you don't " fit " within it's known indexed feature / ' phonetic range '

glitches could occur or some funky / quirky pronunciation

yeah... simplifying it; if you don't have to / don't have some legit reasons, rather avoid index in rt voice changers

[golden walrus]

that explains why my bot can't spell "ng" for whatever the reason

[glacial pollen]

More or less, yes

[visual thunder]

could i use an ai voice model using clownfish?

[golden walrus]

[glacial pollen]

Tomorrow gonna do tests on wavlm and..
if all goes well, gonna add that to ui n stuff

[golden walrus]

thank you, thank you so much

[glacial pollen]

ye np man

[glacial pollen]

oh whoops- actually, gonna be in few days lol

first need to test-train a pretrain ig
( unless someone's gonna be faster than me 🤔 so I can test it sooner

[golden walrus]

but wavlm is an embedder right ?

[glacial pollen]

yes.. to put it simple
contentvec is the default one we always used
then contentvec based spin happened ( trash )
then tests on hubert-based spin ( few attempts
and now, lastly, wavlm-based spin ( supposedly the best and so I think too

[golden walrus]

gpt once told me to use whisper ppg for embedder

i have to finetune it for Vietnamese

dr87 said my pc succ, don't do it

[proven fractal]

Sorry for the interruption, but where would AI covers be put?

In terms of channels

[glacial pollen]

They aren't allowed here because of copyright strikes from discord

[proven fractal]

Oh ok

[glacial pollen]

( trust me, you better do not. I already got 1 strike oof

[proven fractal]

Sounds good to me

[glacial pollen]

✨

[glacial pollen]

guess you'd have to consult dr in terms of training such

He's def the spec in that particular delicate matter

but then I'd trust his judgement if he said ur pc sucks

those trainings are resources-hungry, and I mean very hungry

[golden walrus]

also the spin embedder work in vonovox, just got some robotic voice, but i don't know why

[glacial pollen]

well, it can be the model or can be embedder

[golden walrus]

i trust him too, he knows way more than me

[glacial pollen]

I am still only in 60% informed on the embedders, still gotta catch up ( was busy with my own work

but generally, you should wait for wavlm

If our expectations are met.. it's gonna be a game changer

[golden walrus]

okay, i shall wait for wavlm then

knowledge in this convo is way more than i tried to research on my own

[glacial pollen]

those who seek wisdom, gonna get wisdom.
~ probs doge, in another universe

oh yea, in case of issues with new double-update method, just uncheck it in the ui

this:

then it'll behave 1:1 as before ( except the logging. Now is the accurate one

[golden walrus]

[glacial pollen]

and in case you wonder why the logging got changed in the first place

[golden walrus]

oh if it sits in one place then i turn this on ?

i find it weird in applio cuz sometime the graph don't even move. i tried to turn on and off but nothing happened

[glacial pollen]

I'd say, if no matter what you can't train a good model with it

then try again but with it off

[glacial pollen]

that is because..

Applio does averaging over 25 steps

So, say, if you get 15 steps per epoch

you'll get logging in: 1 epoch ( 15 steps ) + 10 steps from the next

my logging however does an average of each epoch's steps

so yea

( if I got you right, that's the reason. )

[golden walrus]

so small data set will lead to that problem

i see

[glacial pollen]

( tl;dr, mine averages epoch's loss by: average all steps' losses vs over fixed steps count

[glacial pollen]

it's not really a problem but

a design flaw I'd say ( if you wanna look at it like that

Both methods have their pros and cons

But then, I mitigate it by having avg every 5 epochs, as additional metrics

( Yet.. am considering if I shouldn't actually make it avg every 3rd epoch

Anyway. Dw, it's a normal behavior in applio ~

[golden walrus]

i will use ur fork

[glacial pollen]

✨
( remember, there's no good or bad. Just preference lol

[visual thunder]

GUYX HOW DO I ACTUALLY USE THE VOICE MODELS

[golden walrus]

i need to try again with Seoul Steaming Station's pretrain, it sounds good on sample but somehow broken in rt voice changer

[visual thunder]

PLS TELL ME

GUYS*

[golden walrus]

click on edit

add model

[glacial pollen]

First of all, patience is key

and 2

-howtoask

[patent trellis]

How To Troubleshoot

__**GIVE CONTEXT.**__ 📝

• Don't simply mention your issue, like "my rvc is not working".
• Describe the step you are on, what you're trying to do, the RVC you're using, a screenshot, etc.
• The more context, the better.

__**BE POLITE.**__ <:matsuripray:1159685390156967936>

• Don't be desperate. You can ping a Helper, but if they ignore, they aren't available/don't know the answer.
• It's okay if you're frustrated, but don't take it into this server.
• Don't DM without prior consent.

__**BE PRODUCTIVE.**__ 🤝

• Don't ask for every little instruction. Put your own effort & test things by yourself.
• Don't ask to ask.
• Check if your answer is a Google search away/on our guides website.

If you've changed your mind about your message, feel free to delete/edit it.

[glacial pollen]

familiarize yourself with it

[visual thunder]

UM OK

[glacial pollen]

considering It's 8 am and I am tired.. just gonna leave this ^
( going to sleep soon so, even if I wanted to help. too tired

[golden walrus]

i might go with Rigel finetune again if things keep sound so robotic

go to sleep lah

[glacial pollen]

few tests more and yeah.. will have to lel

[visual thunder]

i still dont know how to use them😔

[glacial pollen]

man..

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

[bright basalt]

is the cuda 2.0.78 beta the most updated version?

[golden walrus]

good night and thank you

[glacial pollen]

gnight

[golden walrus]

https://docs.aihub.gg/

[visual thunder]

no but like as a voice changer

[golden walrus]

i mean which part you get stucked

[golden walrus]

this one ?

[visual thunder]

um idk what that is😭

how do i get that

[golden walrus]

[bright basalt]

???

[golden walrus]

i don't know dude 😦

[bright basalt]

😱

[golden walrus]

you should take ur time and look at the doc above

slowly. or hop on youtube

[bright basalt]

idk cuz I'm having an issue where any RVC models I upload it just doesn't work

[visual thunder]

ok which doc tho

[golden walrus]

this one

[visual thunder]

oh ok thanks

[low shard]

of what?

this server isn’t only about rvc and wokada

[bright basalt]

RVC

[low shard]

it’s a general ai server

[golden walrus]

check the voice changer tab

[low shard]

RVC = Retrieval-based-Voice-Conversion, the best Speech To Speech AI Models (on v2), Inferences (use models) pre-recorded audio (ai covers) and train (make) models. Technically, Mainline RVC does have a go-realtime.bat (aka RVC-GUI), but it's pretty messy and outdated so it's extremely not suggested for realtime.

Wokada = uses RVC for realtime inference. There's 2 main versions, Original made by Wok, and the most suggested one is Deiteris Fork (modified version)

are you sure you aren’t talking about original wokada?

[bright basalt]

i have no clue

[low shard]

it looks like ur getting it from a video tutorial, which u shouldn’t

i explained you above the difference

read it for more understanding

video tutorials get outdated easily, never use them for rvc or wokada

[bright basalt]

Yeah, I think I'm seeing the problem now.

[low shard]

@bright basalt what’s ur pc gpu and what do u want to do?

[bright basalt]

My old version stopped working so I tried getting the updated version.

[low shard]

yeah ur deffo talking about original wokada ig, never use video tutorials for wokada

what’s ur pc gpu and what do u want to do?

[bright basalt]

Lemme look rq

[low shard]

You can check your pc gpu on Windows via:
ctrl+shift+esc (task manager) -> Performance tab -> GPU

[bright basalt]

using the NVIDIA GeForce RTX 5070

[low shard]

alr, and u want a realtime voice changer for calls?

[bright basalt]

Yeah sure

[low shard]

alright, then what you need is wokada deiteris fork for rtx 50 serie support and better performance

you were using original wokada which is worse and doesn’t even support ur gpu

-realtime

[patent trellis]

💻 Local Realtime RVC

Guides for Programs that use RVC Models in Realtime for Calls/Games

• Wokada Deiteris Fork

Most suggested. GUIDE

• Original Wokada

ONLY the latest alpha comes close to the Deiteris Fork performance, older versions in youtube tuts are way worse. GUIDE

• **RVC GUI Mainline Realtime**

Unavailable, the guide is outdated and the program is worse compared to the ones above, and much less updated

[low shard]

1st link, read it and lmk

[bright basalt]

Alr I'll take a look

[low shard]

be sure to elaborate next time u ask for help tho, this isn’t an rvc server

[low shard]

alr lmk

[bright basalt]

Alr

[bright basalt]

So, what I did was I downloaded the updated version for MMVCServerSIO for my RTX 5000 series. This would be the correct step?

[mellow ermine]

?

[mystic tangle]

why is there no output or input devices showing up for client on W okada modded

server works though

nevermind i figured it out

[low shard]

did u also download vac lite and uninstalled the old programs u had?

[low shard]

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

[hallow thistle]

Make sure your PC has any microphone plugged in and enabled, and Virtual Audio Cable program presents. Once you have all these ready, the browser may ask you for microphone permission.

[low shard]

they said they figured it out

[idle stag]

what pretrained should I use for an english speech dataset? (must support 32k and 40k sampling rate options)

[hallow thistle]

Which RVC program?

[idle stag]

I am using applio rn for training

[blazing wharf]

I need help with setting up virtual cable for discord and games

[hallow thistle]

!howtoask

[patent trellis]

How To Troubleshoot

__**GIVE CONTEXT.**__ 📝

• Don't simply mention your issue, like "my rvc is not working".
• Describe the step you are on, what you're trying to do, the RVC you're using, a screenshot, etc.
• The more context, the better.

__**BE POLITE.**__ <:matsuripray:1159685390156967936>

• Don't be desperate. You can ping a Helper, but if they ignore, they aren't available/don't know the answer.
• It's okay if you're frustrated, but don't take it into this server.
• Don't DM without prior consent.

__**BE PRODUCTIVE.**__ 🤝

• Don't ask for every little instruction. Put your own effort & test things by yourself.
• Don't ask to ask.
• Check if your answer is a Google search away/on our guides website.

If you've changed your mind about your message, feel free to delete/edit it.

[low shard]

what’s ur pc gpu? What u want to do? What tut link are u following?

[hallow thistle]

Are you trying to use VB-Cable instead of Virtual Audio Cable lite? Which W-Okada version are you using? And what is your PC GPU?

[blazing wharf]

wait,I am sending ss

[low shard]

elaborate too

don’t just send the screenshot

explain all that i asked u

[blazing wharf]

ok

[hallow thistle]

No need to hop into my direct message for that.

[low shard]

@blazing wharf don’t dm without consent

[blazing wharf]

so I can hear my converted voice in the web but discord is not

[low shard]

answer my questions too

[low shard]

^

the more you elaborate, the easier you will get help, else it will be harder to help ya

!give-media-perms 1h @blazing wharf

now u can also send screenshots

[blazing wharf]

Vb-cable, deiteris' W-okada,gpu - rtx3050

[low shard]

uninstall vb cable

[blazing wharf]

thank you

[low shard]

it’s broken for windows

it gives issues

[hallow thistle]

[low shard]

many users reported that

[low shard]

get vac lite from the wokada deiteris fork tutorial

-realtime

[patent trellis]

💻 Local Realtime RVC

Guides for Programs that use RVC Models in Realtime for Calls/Games

• Wokada Deiteris Fork

Most suggested. GUIDE

• Original Wokada

ONLY the latest alpha comes close to the Deiteris Fork performance, older versions in youtube tuts are way worse. GUIDE

• **RVC GUI Mainline Realtime**

Unavailable, the guide is outdated and the program is worse compared to the ones above, and much less updated

[low shard]

1st link, in the virtual audio cable step

[blazing wharf]

ok

[low shard]

yup, forget about vb audio cable

[hallow thistle]

Use Virtual Audio Cable lite instead.

[low shard]

u can uninstall it

[hallow thistle]

After you install Virtual Audio Cable, make sure you set your main speaker and microphone to have only green tick. Otherwise either one program would use that Line 1 instead.

To hear what W-Okada output, you can set monitor to your speaker/headphone on W-Okada.

[blazing wharf]

default communication device? where should I assign this ?

[hallow thistle]

Right click.

[blazing wharf]

should I assign it to Line 1 in both playback and recording ?

[hallow thistle]

Shh. If the device doesn't show default device in menu, it means the device is already set as default device.

When did I say you should set Line 1 as default device in both Playback and Recording? I said to set your main speaker and microphone as default devices.

[blazing wharf]

it's showing like this now

[hallow thistle]

Mate, what are you doing? Why did you disable Line 1 in Recording?

[blazing wharf]

Enabling it ,make it look like this

[hallow thistle]

How can I explain this for you to understand?

[idle stag]

what should I use? I am using applio locally (3.2.9) for training.. or at least could someone tell me weather I can use a pretrain that uses RefineGAN?

[hallow thistle]

Do as what I did in my screenshots. Simply. I won't be repeat the same step for another time.

[blazing wharf]

it's working !!!

😭

[hallow thistle]

[blazing wharf]

sorry

😔

[pastel oak]

Original is still the best

[idle stag]

where is the download for it? is it just the titan/ov2 pretrain?

[pastel oak]

None of the customs. Just the original it comes pre selected

[idle stag]

do I just click train then with custom pretrained turned off?

[simple ore]

by defaut, as long as you dont uncheck [x] Pretrained and do not check [ ] Custom Pretrained, you're using the original pretrain set

[idle stag]

alright thanks a lot man

[bright basalt]

I don't think I did the vac lite, and I'm pretty sure I uninstalled the old version.

[low shard]

downlaod vac lite

[bright basalt]

Oh ok Yes I actually do have it

[low shard]

want me to check ur settings or have any issues?

[bright basalt]

elaborate

[low shard]

what?

I just asked you if u want me to check your settings or if you had any issues

[bright basalt]

actually nvm I do have an issue

[wheat tapir]

Does anyone know if you can use different pre-trained models for training in Replay? Replay works well for me but I am learning about pre-trains now and I heard KLM4.9 or KLM4.1 is good for singing and I want to use that one in Replay because I feel like Replay auto-epoch is such a nice addition. If I were to train in Applio or something more difficult/manual I would not know how many epoch's to choose etc. Does anyone either know another training tool not too heavy on CPU/GPU just like Replay which has an auto-epoch feature or know how I can do this in Replay? 🙂 Please let me know if you know this, I have been searching but can't find the answer anywhere 😦 thank you! 🙂
(I also may need to add that pre-trained models are new for me and I don't complétely understand them so not sure if those can be used in Replay.) any help appreciated

[bright basalt]

Yeah im having an issue extracting the MMVCServerSIO zip

[low shard]

extract it with 7-zip/winrar or whatever

[bright basalt]

I can't just right click it and press extract all?

[low shard]

it's better to not use windows default extraction method for better speed and compatibility

[bright basalt]

alr ill try extracting it the other way soon and lyk if i need more help

[low shard]

alr

[mellow ermine]

i want to generate funny pics / meme etc

my specs i5-14600KF Nvidia GeForce rtx 4070

[low shard]

To generate images for free (text2img), either:

• An easy and good ways with weighs.gg are:
  • Use /image with @earnest musk in https://discord.com/channels/1159260121998827560/1202754985255764060
  • Create an image on their site https://www.weights.gg/ (which you can also use LoRAs, Low-Rank Adaptations, basically a small trained additional model to adjust your generation)
• Use Open Source Models like stable diffusion & flux that could be a bit **harder **but good, you can check tools like Automatic1111 and ComfyUI

[earnest musk]

:wave: @low shard, How can I help?

Available Commands:
• @weights find <query> or /find <query> - Search for RVC Voice Models
• /create - Create an AI Cover
• /image - Generate an Image

[naive palm]

im kinda dumb how do you select a voice model?

[ashen plank]

is there any way to change the sample/bit rate in UVR UI? i can only chose the format

[molten fog]

wahts the best site / app to rip music from without losing any quality?

[stiff goblet]

Is log interval synced in Applio fork ? @glacial pollen

[glacial pollen]

yes it is

it does averaging over N steps of your epoch, per epoch

so each avg metric's log point is an average of all steps' losses ( from a given epoch )

Tho, I am remaking the logging system rn

You'll now have " avg 50 " instead of previous avg_5

so

• avg epoch like always
• avg_5 gets replaced by avg 50 ( same as in applio

[stiff goblet]

Ok thank you. I was confused that it logged every 50 steps in the graph and thought maybe it was not synced. so I wanted to ask

like step 50-100-150-200 ..

[glacial pollen]

check how many steps you get per epoch

        # Calculate the avg epoch loss:
        if global_step % len(train_loader) == 0: # At each epoch completion:
            avg_epoch_loss = epoch_loss_tensor / num_batches_in_epoch

            # Dictionary for losses:
            scalar_dict = {
            "loss_avg/discriminator_adv": avg_epoch_loss[0],
            "loss_avg/generator_adv": avg_epoch_loss[1],
            "loss_avg/generator_total": avg_epoch_loss[2],
            "loss_avg/fm": avg_epoch_loss[3],
            "loss_avg/mel": avg_epoch_loss[4],
            "loss_avg/kl": avg_epoch_loss[5],
            "learning_rate/lr_d": lr_d,
            "learning_rate/lr_g": lr_g,
            }

[stiff goblet]

How do I check it? In console ? Or like in file name etc

[glacial pollen]

console

[stiff goblet]

I guess each epoch takes 14 steps

[glacial pollen]

But for those curious:
It collects losses from each metric availble at every step for N steps ( in here, N is amount of steps your epochs have )
then does the average of summed loss / n_steps

[glacial pollen]

🤔

but this is not my fork

[stiff goblet]

Is something wrong with that training ?

[glacial pollen]

well no, see ^ ?

[stiff goblet]

I'm using Kaggle with an Applio fork

[glacial pollen]

you ain't using my fork

you use Applio

Applio is not a fork

lol

if you were to use my fork you'd see " mel similarity " reporting in console

[stiff goblet]

How can I use it ? Available in Kaggle

Or just locally

[glacial pollen]

Not sure if there's any kaggle made for it

You'd have to ask around

other than that, it's local only for now ( or collab if you can switch repo link and stuff

[ocean pelican]

can anyone pls help me set up rvc?

[vestal helm]

#1159513888199540817 should have everything already

[stiff goblet]

Can't do it locally, unfortunately. this is how Applio graphs looks like

[ocean pelican]

i looked into it but i didnt find any help to solve my issue

[vestal helm]

What’s the issue

[glacial pollen]

yeah, that's not mine

Unfortunately that's what it is for Applio
( if you hoped for pin-point logging

[stiff goblet]

Uhmm that's sad. So graphs are not accurate ? @glacial pollen

[ocean pelican]

i hear like a crackling sound, huge delay in sound and apps and voice doesnt work

[vestal helm]

mm

are you using an audio interface by chance?

[ocean pelican]

wdym by that

like steelseries sonar?

now crackling stopped but the playback i hear is in sequences idk hot to tell u

its like playing a game in 5 fps

[glacial pollen]

it's not that they are not accurate

it is more so they're misaligned ( if we look at it from: " hmmm.. will this epoch I choose rn have good metrics ? " standpoint )

@simple ore this is why I kinda wanna keep avg loss, cause some people ( including me ) want alignment Ig?

Gonna just add a switch in the ui to turn it off and only leave avg_50 ( for pretrains n shit you guys mentioned

[stiff goblet]

I consider here as the best right ? With 0,5 smoothing

[glacial pollen]

well, yes and no

after recent discoveries, this is not an actual ( and only ) metric we should focus on

but ig, you can reference that point yeah

check the steps count and test each epoch that's near that steps count

tl;dr, noticing anything new?

take a guess

[simple ore]

dont worry too much about it not being aligned to the epoch size

[stiff goblet]

I'll reference it for now. What confuses me is that these graphs log every 50 steps but I have 14 steps between each epoch.

[glacial pollen]

Noobies idea, more universal

[simple ore]

if you have 14 steps/epoch you either have a very small dataset, used too high batch size

either way you have a different problem 🙂

[stiff goblet]

14 minutes dataset , 8 batch size

[simple ore]

there you have it

[stiff goblet]

too high?

[simple ore]

there should be like 300 slices in sliced audio, yes?

300/8 does not make 14, so something is wrong

[stiff goblet]

let me check, 333 slices

[simple ore]

300 / reasonably decent batch 4 = 75 steps/epoch

[glacial pollen]

def shouldn't

I on 7 mins and batch 8, get 20 steps iirc

[stiff goblet]

I have no idea.

[simple ore]

re-check the batch size.. you may have use like 24

[stiff goblet]

I'm pretty sure I've chosen 8 batch size. Could 2x T4 have triggered this? I'm using Kaggle.

[simple ore]

yes, batch splits 2-way, then each gpu uses 8

so you're using total 16

[stiff goblet]

Lmao

That's why then.

But still 333/16 is not equal to 14

[simple ore]

good question.. no idea why

[stiff goblet]

Using Colab. I set batch size to 4. 333/4 = 83.25 and each epoch takes 87 steps this time. Applio is kinda buggy or something ? I don't know.. It's still close though @simple ore @glacial pollen

[glacial pollen]

wdym?
it is still applio

[stiff goblet]

Yeah , still Applio. @simple ore said step count = slices/batch size

[glacial pollen]

unless you did something extra?

but then, that'd round up to 83 I guess, so it should be 83

not 87

or, well, even 84

[stiff goblet]

I didn't do something extra. After preprocessing and feature extraction I started the training, as usual.

[glacial pollen]

idk idk, that's hella weird

well either way, few steps this or other way shouldn't matter in the first place ( at least in ur situation
you still have to test epochs in the end right? right. so I suppose, you can ignore it

as long you don't see nonsense on your graphs, you should be fine

[stiff goblet]

Yeah, I guess 3-4 steps difference won't make a difference. But I think Applio is kinda buggy.

[glacial pollen]

If you believe so, create an issue on github

[stiff goblet]

Because I didn't do anything wrong, I just uploaded the dataset and that's it lol.

[glacial pollen]

well and here's the thing

did you actually preprocess it right and truncated all the silence

[knotty moth]

the number of mute files does count, but at least it shows progress bar each epoch unlike the mainline

[stiff goblet]

Applio does it automatically. I guess we find out the problem. The silent files seem to be the main cause of this imbalance.

[glacial pollen]

No it does not truncate

🙂

[knotty moth]

which bug do you refer? as said above, create an issue on github

[glacial pollen]

• if you did not denoise the set, the garbage auto-slicing doesn't even slice right

[stiff goblet]

Oh my bad. I'm already truncating silence using Audacity

[glacial pollen]

because it is based on the silence regions

[glacial pollen]

yeah do that and retry, see if it improves your situation

[stiff goblet]

this bug @knotty moth

[knotty moth]

https://tenor.com/view/linus-tech-tips-ltt-bug-gif-19900847

[stiff goblet]

Should I keep it in 2 ? Silence training files

[solemn pagoda]

do you know how to make ai?

[craggy bough]

no advertising

[solemn pagoda]

what am i advertising?

[stiff goblet]

lmao

[solemn pagoda]

sussyboi69 is hall patrol

[craggy bough]

your app

[solemn pagoda]

i asked the chat who can make ai for me

[craggy bough]

yep that counts as advertising

[solemn pagoda]

do u get paid to do this

[craggy bough]

you sent that message in 4 channels

(counts as spam as well)

[solemn pagoda]

here let me help you

go outside look up and grab a grip

[knotty moth]

I dont see it

but yea posting in multiple chans is spamming

[craggy bough]

#🔹│message-logs message

[knotty moth]

I don't think he intends to promote something

[craggy bough]

isnt that basically the same as the trickhouse guy "looking for a partner"

[knotty moth]

ah mb

but I don't think that kind of request is allowed

sorry we shouldn't have talked about it here, but well it ends here

[simple ore]

if you used automatic slicing, not all slices are equal length

so it adds some batches to compensate

[stiff goblet]

yes , i used automatic slicing. Does it affect the model quality adversely ? I think it would be better if I choose the simple method.

[simple ore]

not really, it is just slower and uses more memory, but in most cases it makes same slices

[stiff goblet]

Alright

[stiff goblet]

What are these batches ? Silence etc

To make slices same length right

[simple ore]

just adding the same file more than once to another batch

when there's not enough slices of specific size to make a full batch

not silences

[crystal canopy]

my voice changer lags my pc a bit, its there anything I can adjust to help it?

[lunar creek]

anyone know the best ai image generator?

[hallow thistle]

There are plently of AI image programs available to use. I use Stable Diffusion.

[hallow thistle]

!howtoask

[patent trellis]

How To Troubleshoot

__**GIVE CONTEXT.**__ 📝

• Don't simply mention your issue, like "my rvc is not working".
• Describe the step you are on, what you're trying to do, the RVC you're using, a screenshot, etc.
• The more context, the better.

__**BE POLITE.**__ <:matsuripray:1159685390156967936>

• Don't be desperate. You can ping a Helper, but if they ignore, they aren't available/don't know the answer.
• It's okay if you're frustrated, but don't take it into this server.
• Don't DM without prior consent.

__**BE PRODUCTIVE.**__ 🤝

• Don't ask for every little instruction. Put your own effort & test things by yourself.
• Don't ask to ask.
• Check if your answer is a Google search away/on our guides website.

If you've changed your mind about your message, feel free to delete/edit it.

[hallow thistle]

Which W-Okada are you using? And what is your PC GPU?

[brittle wing]

!help

[tawdry gull]

Great News!

Unlock the world of LunaBotPrime, absolutely free! Dive into the realm of premium music quality, enjoy it to the fullest. LunaBotPrime - where the music never stops, and the magic continues!

You can invite LunaBotPrime with this link

[glacial briar]

​No Category:
  help Shows this message

Type !help command for more info on a command.
You can also type !help category for more info on a category.

[tawdry gull]

LunaBot 🌙#9997

LunaBot 🌙 is the perfect music bot! Feature rich with high quality music! And Custom Playlist

You can start listening music by just joinning a voice channel and typing: /play [song name or link] (Remove brackets).
We support only Spotify, soundcloud, bandcamp and more!

To view more help on a specific command or category, run
/help <command> or /help <category>

Important Links:
Support
Premium
Invite

Command Categories:
🎶: Music
💰: Premium
⚙️: Utility
📕: Admin

Select A Page From Dropdown Menu Below

LunaBot · May 17, 2025, 4:16 AM

[hallow thistle]

To use other Discord bots unrelated to AI support channels, go to #🤖│bots.

[hallow thistle]

"Looking for a partner", I thought he meant a girl from this server that wanted to be his girlfriend.

[junior gull]

Got Applio running....just dumped it in C: and also learned I should only use 7zip to extract it....winzip breaks the prepackaged virtual enviorment and I wasted an entgire day on that

I'm still very keen to finish the RVCv2 inference training though on the 1000epoch model I trained and made an index for already....but i still need a pre-converted hubert_base.onnx ...but the orignal author doesn't appear to be hosting the file anymore
https://huggingface.co/lj1995
I can't seem to get it manually converted anymore either using hubert_base.pt via the python script I attached.
Is anyone who has successfully trained real time voice models with this file (hubert_base.onnx ) that is willing to share that file to me in DM please rather than tell me to use a different app/method? I'm going to need this file for a few more things eventually too.
Thank you in advanced!

[glacial pollen]

rvc's / applio's " hubert " ?

I mean..

https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main

you can try to fiddle with it

Cause if you need it for RtVc then that's 100% the one you want

[simple ore]

you are confused. huggingface is a github repo, it has history, if there was ever such file there you can find it

but there was not

[junior gull]

not exactly but i think i found part of the problem....never knew or part of any instructions anywhere that i had to run "dlmmodels" and use aria2.exe from the tools folder....i've been combingin trhough every file for hours....and found that and read its code.....it was so obvious when i read it....it auto downlods the needed files.... first attempt from hugging face then backup mirrors if that fails. ALL THE HUGGING FACE attempts failed. but all the mirrors worked thankfully

I just finished donwloading them a few minutes ago

and am going to try to get things working again

so far it LOOKS like i have all the assests i need....fingers crossed...i'm goign to start over and retrain from sratch to be safe

[idle stag]

When preparing a dataset, a big one (30 mins+) would it be better to add a variety of tones? like asmr, high pitch, lower pitch, speaking in normal voice of the person's voice or should I just stick to one type of tone? like ONLY asmr or ONLY high pitch audios or ONLY lower pitch etc

[simple ore]

if you want a model that does a lot of things, you may want to add some variety

[idle stag]

alright thanks I will make the majority of the dataset in normal speech and add some other tones too

[glacial pollen]

just make sure to balance it out

if extras you intend to use aren't well contained within pretrained models, having thrown-off balance in ur set might make the model biased towards a given subset of sounds
( or might diverge ~

In any case, gluck ~ ✨

[idle stag]

I am using orignal pretrain rn I think it should be fine

[glacial pollen]

oh well

I do wish you best best of luck then

( in any case, I head off to sleep now so, might respond tomorrow if anything ~ )

[potent bone]

whats the difference between client and server in new w-okada

[simple ore]

presumably you can run a server on a remote host with gpu and use a browser on, say, a phone

client takes audio, sends it to server to process

server options uses devices available on the host directly

[junior gull]

I was wondering about that too! Probalby only useful on rooted android phone though?

[knotty moth]

the client mode only uses MME which has more latency than WASAPI and ASIO being only available in server mode

[junior gull]

oh wow...with teh propper assets....
1000 epoch is taking like 5 minutes vs 3.9 hours last time lol

[dry owl]

?

[simple ore]

more like "with only 2 mute files" lol

hubert_base is only used to extract features.. which I assume it failed, so you have an empty train

[knotty moth]

there must be some issue where the preprocessing & feature weren't done properly so it doesn't learn it at all

[odd isle]

how do i make my own voice models

[golden walrus]

wait a min

https://docs.aihub.gg/rvc/local/codename_fork/

there is a guide on these

https://docs.aihub.gg/rvc/resources/dataset-isolation/

you should read about this first

to have good data set

then use it to train your voice models

[golden walrus]

with this

[bleak torrent]

Is RVC Colab no longer usable?

[viral mason]

I have no idea how to use the voice blender (this is in the applio kaggle space)

[viral mason]

I can help u make voice models

if you want to use kaggle applio

[golden walrus]

oh, please do. I have lots of issues with my model. People said i sound like a kid.

Do you know if i need the same person voice but lots of up and down, singing to make it able to have a realistic voice ?

[viral mason]

Having a bigger range will most definetly help

[golden walrus]

I need it for realtime by the way

[viral mason]

perfect, I use realtime all the time for trolling

[golden walrus]

So 1 hour of data of the same person, but saying different stuffs and vocal range will help right?

So ASMR is a no no?

[viral mason]

yes

[golden walrus]

[viral mason]

u kinda don't need 1 hour

but if u want sure why not

[golden walrus]

How long should it be

Because none of embedder support my mother tongue

also pretrain. I need Rigel finetune to able to handle 48k

[viral mason]

I don't really know how much data is best for a model but nothing less that 10 minutes I'd say

[viral mason]

32k is usually best..

[golden walrus]

Wait

32k is better ?

[viral mason]

not always

but in most situations I've seen it is

[golden walrus]

So train in 32k, and realtime tools will upscale it to 48k?

Or

[viral mason]

uhh

no?

[golden walrus]

[viral mason]

why must it be 48k?

[golden walrus]

Because Vonovox has the output of 48k

[tame mica]

if all goes wrong why not formant shift the dataset

[golden walrus]

I thought i need to train it to 48k to ultilize it

Let me google how formant shift work

[viral mason]

idk what Vonovox is 😭

[golden walrus]

Oh, dr87 made it. A tool like wokada

But somehow my pc don't let me run fork version or og version

So i run vonovox instead

it has less delay too

[viral mason]

that's cool

[viral mason]

no nvidia?

[golden walrus]

Nvidia

But it's broken and i can't figure it out why

You used to work fineeeeee

Maybe it got jealous cuz i use Vono instead

[viral mason]

rip

anyways I'm up for the task of helping u make the model

dms are open when ready!

https://tenor.com/view/ena-joel-g-ena-bbq-dream-bbq-announcement-gif-16993356939999711983

[golden walrus]

Thank yoi

[viral mason]

No problem

[golden walrus]

Now i need to find a clean data set of Vietnamese peeps

Have a great day, people

[analog obsidian]

no

[analog obsidian]

yes

[viral mason]

neat, I didn't know that it did that

[analog obsidian]

yes, in 32k datasets

[golden walrus]

knowledge acquired

[analog obsidian]

for 16khz cutoff sets sure, for 20khz + and above not

[analog obsidian]

you kinda need 1 hour

[analog obsidian]

this pretrain is very bad and undertrained
use og pretrain

[analog obsidian]

dont use singing for realtime models, speech only

[golden walrus]

okay, let me train again

I mean i do speech only

But there are words that the model can't spell

i don't know what is the cause of it

[analog obsidian]

this is embedding limitations

cvec is trained mostly on english

[golden walrus]

Wait. So it is not pretrain problem?

[analog obsidian]

no

but a pretrain in vietnamese may also help

for better results

[golden walrus]

Ye, i picked Rigel because it has language that spell similar to Vietnamese

The quality of the audio is not as good as og pretrain

[analog obsidian]

welp i told u why

[golden walrus]

Ye, embedder

[analog obsidian]

no?

[golden walrus]

[analog obsidian]

i said rigel gave u worse results because is undertrained

just stick with the og pretrain

[golden walrus]

But you told me it's embedding limit which cause misspelled

Ohhhhhh

Sorry, my brain is a bit laggy

[analog obsidian]

that will improve the results quality wise but not the pronunciation

[viral mason]

Once you've gotten everything figured out dms are open for me to help

[analog obsidian]

u can try using the index file

for realtime (or any model actually) aim for 30mins to 1 hour max

[golden walrus]

i appreciate that, kind soul

[viral mason]

Yummy index file

[analog obsidian]

dont use asmr

results are weird

rvc cant handle breathy voices well

[viral mason]

In one of my Ena models there was a singular line with whispering and it didn't do too bad

[analog obsidian]

and og pretrain cant whisper so

[analog obsidian]

a singular audio is not gonna destroy a model

[golden walrus]

Now i kinda know what to do

Vietnamese pretrain. Here i come

[crude flame]

some asmr is fine

[analog obsidian]

yea some
better be safe and use something that yw is gonna work

[crude flame]

its been a while since ive made a asmr model

might need to make a return

its been 3 months

wow

[viral mason]

no, women scare me

[latent kettle]

During training, my pc went to hibernation. I checked logs it's still training but in gui it's showing error. What to do now

[potent bone]

has anyone ever told you that ur voice sounds robotic when using the voice changer?

[golden walrus]

Was

If sound robotic

I suggest you take a look at tensorboard

Or check your data

[latent kettle]

And there is no model.pth file over there. There is only G, d and other stuff

@viscid moss sorry to ping you but I need help. Please help me

[viscid moss]

In that case u should stop the training cuz u can't save the .pth file

[lusty quartz]

I need help related to ai

[hallow thistle]

!howtoask

[patent trellis]

How To Troubleshoot

__**GIVE CONTEXT.**__ 📝

• Don't simply mention your issue, like "my rvc is not working".
• Describe the step you are on, what you're trying to do, the RVC you're using, a screenshot, etc.
• The more context, the better.

__**BE POLITE.**__ <:matsuripray:1159685390156967936>

• Don't be desperate. You can ping a Helper, but if they ignore, they aren't available/don't know the answer.
• It's okay if you're frustrated, but don't take it into this server.
• Don't DM without prior consent.

__**BE PRODUCTIVE.**__ 🤝

• Don't ask for every little instruction. Put your own effort & test things by yourself.
• Don't ask to ask.
• Check if your answer is a Google search away/on our guides website.

If you've changed your mind about your message, feel free to delete/edit it.

[latent kettle]

I stopped

What to do next

I have all g and d files event files and other stuff

Should I Train the model again? Or I can re launch training process

[lusty quartz]

I need someone who can help to build my ai startup

[hallow thistle]

Which AI program? Is it about AI voice, LLM or anything else?

[viscid moss]

What things do u have? If only have G and D u need to train again

[latent kettle]

Lemme show you

[lusty quartz]

Something else it's in an untapped market

[viscid moss]

disable that hibernation thing btw

[hallow thistle]

Um, that doesn't tell which AI you're looking for.

Also, no need to hop into my direct message. If nothing is personal, better not.

[latent kettle]

Yes I will do it. I just re installed the windows and forgot about it

[viscid moss]

there is no pth file inside those folders?

u have everything but not the pth files

[latent kettle]

😭😭

Why this happened

[viscid moss]

cuz hibernation

[latent kettle]

So there is no fix ?

Just restart the whole training?

[viscid moss]

ya

[lusty quartz]

Sorry for that namari, it's related to automation

[viscid moss]

i think so

[latent kettle]

Lemme train it again.

[hallow thistle]

[lusty quartz]

?

[latent kettle]

--Wasted--

[hallow thistle]

Sorry, but if you don't tell a specific AI program, like the name of it, I can't help with that.

You can say something like LLM or Python for example, so to see if that I can help about.

[simple ore]

you can resume, strange to have events.out in the same folder

they've were moved to eval in latest update

[lusty quartz]

Python, playwright, ocr, etc

[lusty quartz]

?

[lusty quartz]

I'm using this

I need a tutor

[hallow thistle]

Yeah, I've never heard of this program. I only heard of Python.

[lusty quartz]

Ok, but thats not my question

Can someone else help me in this I need a tutor

??

[hallow thistle]

No, no, you didn't ask me like that. I asked you which AI program you looking for, and I just sent a screenshot of Google of the program I've never heard of.

[lusty quartz]

Is their someone who can help me in this, if yes then please msg

??

[lusty quartz]

Are u their?

No one is replying I need help please someone guide

[hallow thistle]

No, I don't own this program. Make sure you ask the right question.

I read a bit of the guide, and the first thing I found is Visual Studio Code. Not sure what npm, yarn and pnpm are.

[lusty quartz]

Have u installed vs code?

[hallow thistle]

Fortunately, I have installed Visual Studio Code before.

[lusty quartz]

Wdym by right question did I hurt u if yes then sorry

[lusty quartz]

Great

[hallow thistle]

In Extensions tab on VS Code, search Playwright, there's one that says "Playwright Test for VSCode" by Microsoft, you click on that and then click install.

[lusty quartz]

??

I know that

Can anyone tell me how to redirect to the page from were came to this, as I'm complete newbie

[hallow thistle]

This is their guide about using Playwright on VS Code. https://playwright.dev/docs/getting-started-vscode

[lusty quartz]

Why u r telling me this, do u wanna really help me or wanna make fun of me, is their anyone else who can help please help

[covert kelp]

i have quick question

how do i train a model

[hallow thistle]

I don't know what you think. What you say confuse me a lot, so I can explain what I understand.

[hallow thistle]

Which AI program are you looking for? Is it RVC or Stable Diffusion? And what is your PC GPU?

[covert kelp]

is rvc also RTX 3060 (i think its enough)

[hallow thistle]

-rvc

[patent trellis]

📚 RVC Documentations

AI HUB Docs

https://docs.aihub.gg

🍏 Applio Docs

https://docs.applio.org

[hallow thistle]

Use Applio.

[covert kelp]

sure

also why dot

[lusty quartz]

Ok got it namari, I need help in browser automation n rendering results

??

[hallow thistle]

Mate, to install Playwright in your browser without the need of VS Code, I think you should install either npm, yarn or pnpm program so to run command "install Playwright". I won't install either of these, since my laptop storage is full now.

You can ask others if they know about Playwright or search the program on YouTube. As much as I know about AIs, it's not like I would know everything.

[latent kettle]

I'm training a model but it's not saving the actual model files

Model.pth

There are g and d files, event files and bunch of other files

[simple ore]

you did not select the option to save the model

[stiff goblet]

When training in Kaggle, should I set batch size as 2 for using batch size of 4 ? Because Kaggle using 2 x Tesla T4 right?

[latent kettle]

What? Where it is ?

You mean save every checkpoint ?

[analog obsidian]

yes

[simple ore]

save every weight

[latent kettle]

But it will save each and every epoch ?

Like from 1 to 100

??

[simple ore]

This setting enables you to save the weights of the model at the conclusion of each epoch.
Save Every Weights

so it saves every epoch you selected, 10 by default

[latent kettle]

I see. So this was the mistake 😭

I restarted training 3-4 times

[magic elm]

! C:\Users\frime\Downloads\voice-changer-windows-nvidia-b2332.zip: Cannot create C:\Users\frime\Downloads\voice-changer-windows-nvidia-b2332\voice-changer-windows-nvidia-b2332\MMVCServerSIO\MMVCServerSIO.exe
Access is denied.
When trying to extract the file

[simple ore]

shitty anti-virus detected

[magic elm]

Got it thanks

[uneven stone]

Hello everyone, can someone please help me? I would like to have a short voice clip from Fatman Scoop or DMX, about 10 seconds long. I'm having trouble making the voice sound right. If anyone can help, please get in touch with me. Thank you very much

[glacial pollen]

right click on the zip and " unlock "

Whenever you see such an option available, I advice you to do so.
else windows might flag it by mistake ( also as Noobies said about the anti-vir ~ some crappy ones do whatever they please

[lucid creek]

its good ?

[naive palm]

how do i put it like towards my output? so other players can hear my voice changed

[simple ore]

use charts from SCALARS tab

and less smoothing if you're using avg_50 charts

[stiff goblet]

If I am not mistaken, it is necessary to use a smoothing value of 0.5 and lower for avg charts.

[lucid creek]

its 0.6

[stiff goblet]

Training still seems to be going on. There is no significant increase in the g/loss chart.