✨│ai-help

[sand iris]

What

[sand iris]

there's a guide at https://rentry.co/VoiceChangerGuide

[stray lantern]

i have duplicated space, i got the voice model i want, i put a mp3 as my track, but it still says error? (huggingspace)

[sand iris]

But you should have a .index file with your .pth model

you can keep it empty if there somehow isn't one

[twin quiver]

how do i listen

to my voice

[stray lantern]

@small wolf

[small wolf]

what is the error

[stray lantern]

it just says error

im stuck

i literally have everything

[plush compass]

How much training do you do on a 16-minute long dataset? (RVC2)

[small wolf]

it's still uploading the mp3

how big is it?

Once I get my server working I have an idea I'd like to try... trying to train the base RVC model from scratch but adding my voice to the dataset

[stray lantern]

OH

how do i know when its done?

even after that it says error

[small wolf]

¯_(ツ)_/¯

[small wolf]

the space might be broken, not sure

[rare goblet]

Ayo? @small wolf level 4 !!!

[small wolf]

try using colab

[small wolf]

since the RVC base model is actually trained only on VCTK...

or maybe I'll just mess with my own multi-shot architecture ¯_(ツ)_/¯

[tardy flint]

this isnt for rvc but im trying to split the backround vocals from a song and it just wont work can someone please help me. pwetty pwease 🥺

[small wolf]

UVR?

[mortal plinth]

help me pls, What's the problem?
[VCClient] wait web server...230 http://127.0.0.1:18888/

[small wolf]

How does RVC work, like how can it learn a target voice without any source voice as input?
surely at least the base model was trained with some kind of source voice input at some point?

like during training, what is it even using as a source voice to derrive loss from???

[median surge]

im a totally new to this whole rvc thing, but i can not get to the part where i can put my voice in, i just wanna make a voice model of my voice and make christmas covers, but everytime i download the rvc file from github, there is no go-web.bat file and if there is one(which there has been since i tried different ones) it says "server not found" and "press any key" and if i press a key it just ends the window, Please help i have windows computer, am i dumb or something like that? am i missing something?

[proper shale]

Oooh I think I know why it screwed up.

You should've downloaded it from Huggingface and not GitHub

Therefore you only got the code and not the usable application

[modern oxide]

rvc is built on a hubert with a huge dataset (i dunno how many, prolly a couple hundred hours at least) which uses pretrained models, that are trained on 50 hours dataset

[proper shale]

^

[small wolf]

interesting... so do you know if there is a way to fine-tune the base RVC model or even a trained RVC model but using audio pairs instead of just target audio files? (like older parallel voice cloning models)

[modern oxide]

uhh, i dunno actually, but i think pretrained models are kinda doing that. Like hubert is just a huge mess of different harmonies on which we can build sound models, and pretrains usually help rvc to understand what should the sound sound like. our ptretrains are trained on voices, so rvc can easily understand the voice of a human being even without enormous amounts of data.

[small wolf]

yeah... but problem:
RVC models consistently turn out horrible when using my voice

(Yes I've tried pitch shifting, all the advanced parameters, etc etc)

ugh, at this point I might just make my own model from scratch 🤣

[modern oxide]

hmm, can you send an example?

because i've made multiple voice models of my own voice, and they all sounded fine

[small wolf]

probs in like ~10 mins
away from main PC atm

notably weights.gg sounds somewhat better
also I'm using applio, if that impacts performance (it probs shouldn't)

[small wolf]

I meant when using my voice to run inference on a model, ie: Cave Johnson

it just sounds a lot worse for most models than it should

[modern oxide]

nah, all rvc's have the same bases

[proper shale]

yg

[small wolf]

yeahhhh :/

[proper shale]

yh

so the thing is, maybe the dataset wasn't that good?

[modern oxide]

maybe you have a bad mic?

[small wolf]

possibly? Just using standard laptop mic...
any preprocessing I could do to my input audio to try to improve?

[small wolf]

nah because I've had a friend use it and it sounds spot on lol

[modern oxide]

oh, yeah that's problematic xd

[small wolf]

XD

[modern oxide]

to get a good inferencing you also need to provide an rvc with a good audio input

so it's better to buy a microphone and use that

[median surge]

oh thanks ill try that

[rare goblet]

Ayo? @median surge level 1 !!!

[modern oxide]

• have a good volume settings, so that it won't pick up much noise background and will pick up your voice

[small wolf]

fair...
but what exactly makes sound like... "good"? (opposite of an audiophile here)

[small wolf]

I can guarantee there's no background noise in the recording lol

complete silence here

[modern oxide]

and + you yourself should have a good voice with correct pronounciation of words, and have a good intonation control

[small wolf]

what's intonation control? (yes I tried googling it)

ngl I'm kinda leaning towards just mangling together two HuBERT models in an autoencoder config atm...

[modern oxide]

it won't help

and it will take a lot of time

[proper shale]

little to no noise (check the spectrogram for that 💀)and no silences between words/phrases (you can do this after recording, in Audacity)

[small wolf]

I mean, it could, given matching audio samples from my mic and desired output it could probably overcome the noise

[proper shale]

Anything else just works

[small wolf]

right... lemme check

[modern oxide]

our russian community has already tried to implement a russian hubert base with pretrains for a month

still no good results

[small wolf]

how's uhhh... how's it going?

:/

how big is their dataset atm?

also what are they training on?

[modern oxide]

it's not as easy as training a regular model

[small wolf]

obv lol

[rare goblet]

Ayo? @small wolf level 5 !!!

[small wolf]

finally

[modern oxide]

1300 hours of russian speech for hubert and around 300 for a pretrained model

but

yeah

you won't be able to make your own pretrain

or hubert

it's easier to just buy a microphone and practise your voice, because the program works completely fine

[median surge]

ok so itried it now but it didnt work, i downloaded from hugging face, and put the file into the windows file(system) and when i oppened it with an extention"breezip" and it still doesnt work. im sorry please help

[small wolf]

ok... so maybe training HuBERT from scratch not good idea LOL

but surely there must be a way to finetune existing model with audio pairs?

[small wolf]

fair

[modern oxide]

nah, probably there's no way

[small wolf]

makes no sense tho
like how can RVC even calculate loss when training?
what is the input that it's getting when calculating loss

[modern oxide]

like any other machine learning stuff

[proper shale]

just use this: https://huggingface.co/MangioRVC/Mangio-RVC-Huggingface/tree/main

[small wolf]

well yeah but from what I know loss calculation basically takes some of the dataset (or from the test split), puts it into the AI and compares it with the expected result

[proper shale]

download the second one if u wanna train

[small wolf]

except the RVC dataset consists only of the target voice, no input so to speek

[proper shale]

probably uses the pretrains to compare? hmmm i actually should research about this lol

[small wolf]

lol

[modern oxide]

discriminator takes content from given data, shows it to generator, generator tries to fool the discriminator trying to generate stuff based on the dataset, it does that till discriminator believes it and then uppers it's expectations, and they do that till they reach some immovable point of learning

[small wolf]

makes me wonder if I could use
matching audio pairs

to calculate loss
[that's my endgame here it's just somehow training or finetuning using my voice as an input...]

[small wolf]

what is "content"?

[modern oxide]

part of the given dataset

[small wolf]

ok I see a problem here

if the dataset consists only of the target voice
then it's just learning to turn target voice into target voice

right?

then I assume it "just works" because of the base model, which was trained with the other source voices

[proper shale]

i think it uses part of the pretrain

would make a bit more sense that way

[small wolf]

either way I suppose my goal now would be to figure out how to change this... get it to calculate loss using some other source...

[small wolf]

ye

[modern oxide]

uhh, no, it tries to generate something resembling that data, remember i told you that rvc is built on a hubert which is basically a mess of harmonies

[small wolf]

generate...
from nothing?

[modern oxide]

hubert is like a dough, the dataset is like a recipe to bake a pancake

[small wolf]

dataset is the pancake

[small wolf]

like loss is basically calculated by inferring the model as-is?
RIGHT?

I feel like I've misunderstood something fundamental here...

[proper shale]

what i meant was, it's trying to fool discriminator with pretrain samples

that'd make more sense

[small wolf]

ye if it's basically running inference with the audio samples from the VCTK datseet that would make more sense idek

traditionally GANs just input random data for training is this... also the case for RVC?

[stray lantern]

Hello, this TikToker makes his covers with RVC disconnect, what is rvc disconnect?

And anyone know how to use a polio

[latent kettle]

RVC disconnect is a colab Notebook

[proper shale]

RVC Disconnected is for training voice models.

There's a guide in #1159513888199540817 if you wanna use it.

[stray lantern]

Can you get models from there?

[proper shale]

You make your own models there.

[proper shale]

There are also guides for this too.

[latent kettle]

You can train your own models here by uploading suitable dataset

[stray lantern]

Cool!

[latent kettle]

To train your own models..You have to prepare a detaset!

[stray lantern]

Btw, I have made my first cover, and I feel like it could sound better, how do I make the voice stay as the same tone?

As in like the movie/show that it’s from?

I used replay btw

I might try applio

What do you guys use?

[latent kettle]

To make a Ai voice cover. i use RVC

[proper shale]

yeah both replay and applio use rvc

[proper shale]

there is a transpose/pitch change option in there

[latent kettle]

You can also do that with any DAW software(s) . U can change the pitch!

[proper shale]

just increase or decrease the pitch according to the voice you wanna replicate and the voice of the song

[stray lantern]

In replay?

[proper shale]

yeah

there should be, at least

[stray lantern]

Alright

If it doesn’t have it I’ll try applio

[stray lantern]

But there’s many variations of rvc, like google colab, huggingfacd rvc

[proper shale]

exactly

they all use the same base but are all different forks

of the same thing

[latent kettle]

Yes ... Use anyone. U like

[stray lantern]

Do all of them have the same sounding quality?

[proper shale]

...yeah?

they're literally all the same under the hood

[latent kettle]

I think it depends on your model.

[median surge]

im so sorry if this is a dumb question but if i use rvc disconnected to make voice model what is a dataset?

[stray lantern]

That is true, but I meant like how it would just convert in general

[stray lantern]

Thanks!

[latent kettle]

Its the voice data.. audio files

[proper shale]

np

[stray lantern]

Just glad I’m not using one that lessens the quality

[median surge]

hahahah thanks!

[small wolf]

idea:
synthetic dataset

[latent kettle]

But there is something which you have to remember

[willow raft]

I'm trying to use model extraction from checkpoint processing but I'm getting an error code, does anyone know what the error is or how to fix?

  File "C:\Users\Public\Mangio-RVC-v23.7.0\train\process_ckpt.py", line 64, in extract_small_model
    ckpt = torch.load(path, map_location="cpu")
  File "C:\Users\Public\Mangio-RVC-v23.7.0\runtime\lib\site-packages\torch\serialization.py", line 791, in load
    with _open_file_like(f, 'rb') as opened_file:
  File "C:\Users\Public\Mangio-RVC-v23.7.0\runtime\lib\site-packages\torch\serialization.py", line 271, in _open_file_like
    return _open_file(name_or_buffer, mode)
  File "C:\Users\Public\Mangio-RVC-v23.7.0\runtime\lib\site-packages\torch\serialization.py", line 252, in __init__
    super().__init__(open(name, mode))
OSError: [Errno 22] Invalid argument: '"C:\\Users\\Public\\Mangio-RVC-v23.7.0\\logs\\Depeche-Mode\\D_2333333.pth"'```

[tardy flint]

huh

[small wolf]

https://docs.aihub.wtf/vocal-cloning-guide/vocal-and-music-separation/uvr

[proper shale]

used the wrong file? D and G are training files

not the usable pths

[willow raft]

where would i find the right file? the website only says "Model extraction (enter the path of the large file model under the ‘logs’ folder). " and i don't know what file it's refering to

[proper shale]

wait is this a sovits model?

do you not have a pth named "depechemode.pth"

or something similar

[willow raft]

i don't

[proper shale]

oh no should use the one in weights

[willow raft]

ohh

i found it, thank you

[proper shale]

np

any other question?

[tardy flint]

thanks ill give it a try

[willow raft]

none for now, thanks for the help

[proper shale]

np! good luck ;) hope the model's great

[willow raft]

actually sorry to bother again, it's still giving me the same error, i tried bothing the topmost file and the bottom one

[proper shale]

wait what the hell is this

i have never used this

[willow raft]

checkpoint processing tab because i didn't finish doing all of the epochs

[proper shale]

but uh

[willow raft]

i think i'm supposed to have a .ckpt file somewhere

[proper shale]

you already have all of the models in between the epoch you stopped on

so I don't see why it's useful

[willow raft]

how do i turn them into a usable model? i don't have any index files

[proper shale]

the pths in weights are already usable models

u can go back in the train tab to uh

train the index

[median surge]

yeah?

[latent kettle]

It should be in .wav format
It should be of minimum 10 minutes
It should be clear with music and background noises!
There should not be any pause(s)

[willow raft]

You have saved my life, I understand how all of this stuff works now 🙏 thanks for your patience, the model needs a bit more training but now I know where to get all the files I need and where they go

[median surge]

ok but how no pauses?

[latent kettle]

Pause meaning. When someone speaks or sings .. he/she may Stop for a couple of seconds. Which will increase your dataset length without any need. your Main motive should be to cover maximum data in a short length.

[latent kettle]

Cut it by using DAW such as (audacity, fl studio)

[latent kettle]

Also remember there should not be any other speakers in your dataset !

[severe rapids]

Is it normal to get triangles?

[finite galleon]

No thats a mode collapse

Perhaps your dataset has too many silences?

[severe rapids]

Maybe, I'll cut more silences from it then

Should I restart training from 0 or just continue it?

[rare goblet]

Ayo? @severe rapids level 2 !!!

[proud elbow]

probably your training session gets interrupted and resumed? mode collapses are supposed to be hard dip on <15 value, and silence aren't always only cause of collapses, it can be if dataset length is too short, etc.

[severe rapids]

The dataset is 10 minutes long

there's like a 1-2 second silences between sentences

[proud elbow]

you can do noise gate & truncate silence using Audacity tho

[severe rapids]

Can you tell me where's the truncating?

Audio is clean, there's no noises so I don't think I need noise gate

[severe rapids]

why would it get interrupted?

INFO:saturn:Train Epoch: 57 [0%]
INFO:saturn:[3080, 9.927757679628145e-05]
INFO:saturn:loss_disc=nan, loss_gen=nan, loss_fm=nan,loss_mel=21.914, loss_kl=7.541
INFO:saturn:====> Epoch: 57 [2023-11-30 15:21:35] | (0:01:20.316749)
INFO:saturn:Train Epoch: 58 [2%]
INFO:saturn:[3136, 9.926516709918191e-05]
INFO:saturn:loss_disc=nan, loss_gen=nan, loss_fm=nan,loss_mel=21.711, loss_kl=8.367
INFO:saturn:====> Epoch: 58 [2023-11-30 15:22:59] | (0:01:24.328599)
INFO:saturn:Train Epoch: 59 [4%]
INFO:saturn:[3192, 9.92527589532945e-05]
INFO:saturn:loss_disc=nan, loss_gen=nan, loss_fm=nan,loss_mel=21.252, loss_kl=8.694
INFO:saturn:====> Epoch: 59 [2023-11-30 15:24:24] | (0:01:24.975206)
INFO:saturn:Train Epoch: 60 [5%]
INFO:saturn:[3248, 9.924035235842533e-05]
INFO:saturn:loss_disc=nan, loss_gen=nan, loss_fm=nan,loss_mel=22.173, loss_kl=7.824

Console logs

[severe rapids]

alright

[proud elbow]

which RVC are you using? neither of my latest original RVC, mangio and applio have that issue

[severe rapids]

mangio

[proud elbow]

I remember having all nan values issue when before upgrading GPU from 1660 super to 4070 and before updating the driver, but now it works completely fine

[severe rapids]

I have a 1060 lol

[proud elbow]

how about try original RVC or applio, the former uses a bit less VRAM for the same settings and batch size tho

[severe rapids]

Okay, i'll try

[sand iris]

gtx 1xxx issue

Fix is turning off "half precision" or "fp16_run". Unsure if you'd be able to train on a 6GB card because of that though

@severe rapids

I don't recall how you're supposed to do it though

[severe rapids]

In stable diffusion I believe you use --no-half-precision or something similiar

[proud elbow]

does it also affect on 4 gb GPUs (e.g. 1650)?

[severe rapids]

Will probably be the same

[sand iris]

Any gtx 1xxx series cards so yes

[sand iris]

Here you gotta edit one of the config files by hand before doing anything related to training the model

Or straight up edit all of them at once, but I can't check where the files are atm

[severe rapids]

I think I found it

I setted is_half to false, will try now

[sand iris]

I believe it'll be reset according to settings defined in other .json files

Also that might not even be the correct toggle but unsure

[azure marsh]

RVC Guides (How to Make AI Cover)

Documentation

🇺🇸 English (main)

https://docs.aihub.wtf/

​

Translation by country

🇧🇷 Brasil (PT-BR)

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

🇪🇸 Spain

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

🇫🇷 France

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

🇩🇪 Germany

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

🇯🇵 Japan

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

🇰🇷 Korea

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

🇵🇱 Poland

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

🇷🇺 Russia

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

🇮🇹 Italy

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

[severe rapids]

It's been working fine for the past 10 minutes, I guess I'll just start over with the new dataset and see how it goes

[short linden]

So how do you import models from weights.gg to huggingface?

[proper shale]

u can upload them to your drive and copy that share link (make sure everyone w link can see else it won't work) and it should work

[severe rapids]

nvm, it's at it again....

[proper shale]

is this g/total? what the hell

[severe rapids]

yeah

[rare goblet]

Ayo? @severe rapids level 3 !!!

[glad zealot]

wtf

is that

[short linden]

Great! So which area of the discord do I upload it to?

[proper shale]

wait is it ur model?

[short linden]

No. I downloaded it from weight.gg and uploaded it to my Google Drive.

[proper shale]

no need to upload it here then

[short linden]

Except the model isn't available in the voice-models area of this discord.

[proper shale]

You could ask the model maker to upload it then

[hot violet]

why i do get this in RVC V2 Disconnected

[severe rapids]

Basically, I have a 1060 and 10xx series don't support half precision.
Just now I found a way to turn it off, but I had to sacrfice 3 batch size down to 1 (maybe 2 might work, I'll try now).
Now it looks like it works fine, yet instead of a minute, it takes 2 per epoch

[hot violet]

nvm, i got the solution

[distant turtle]

the program has been updated, it's the first time I've used it, I want to use it for TTS, but it's changed, does anyone know how to start it?

[proper shale]

oof

[proper shale]

I think installation failed, because the name "applio tts" has a space.

[distant turtle]

no applio tts I created it I did the installation inside

[proper shale]

• Applio is not a TTS - RVC needs sample audio to work

[proper shale]

Either way, remove the space on the name.

[distant turtle]

Done! but how do I start the program?

[proper shale]

I guess you should try reinstalling Applio, delete the files on the tts folder and then try again

Might work now

[distant turtle]

I'll try, thank you very much!

[proper shale]

Np! ;)

[severe rapids]

@proper shale Is it suppose to flatline that soon? It's been training only for 1h, 34 epochs

[gaunt ruin]

does Ilaria still work?

[mellow roost]

are mac (silicone) users cursed with RVC? only trying to do inference as colab seems better for training then doing it local on mac.

is W-Okada the way to go in this case?

[molten pecan]

Yeah

Why are you asking?

Having some trouble?

[violet heron]

Yes

[proper shale]

for inferencing, no.

[proper shale]

hmmm yeah, sadly

depending on batch_size

[mellow roost]

hmm so there is a mac silicone version on the w-okada page. its not good/working?
i may just try it and see. if it works are there any drawbacks from using the webUI inference? maybe less tweakability?

[proper shale]

it's lower quality because it's made for realtime.

therefore, an inferencing fork would be better, like mangi

mangio

[mellow roost]

i see, makes sense yea. ill check that out thx!

[proper shale]

Alrighty, any more questions?

[mellow roost]

no sir

[severe rapids]

How good is it for the first model?
How can I improve it?
Should include more of a different pitch in the dataset?

[languid swift]

im having error

[rare goblet]

Ayo? @languid swift level 1 !!!

[distant turtle]

it gives me this error, with go-applio-manager.bat I activated the resolution of the problem but it remains the same, I even reinstalled everything and updated the app, does anyone know the problem?, I have amd, intel as the processor,

[ancient sinew]

[VCClient] waiting for the web server...200 http://127.0.0.1:18888 /
[VCClient] waiting for the web server...210 http://127.0.0.1:18888/
Backtracking (last call):
File "MMVCServerSIO.py ", line 258, in <module>
 File "subprocess.py ", line 1209, File in standby
mode "subprocess.py ", line 1506, in _wait
KeyboardInterrupt
[10884] The 'MMVCServerSIO' script could not be executed due to an unhandled exception!
Wrap the package filling [Y(yes)/N(no)]?```
help :(

[rare goblet]

Ayo? @ancient sinew level 1 !!!

[plucky geyser]

is it known whether or not setting is_half to false in the configs actually makes everything run in full precision or just the training?

training a model locally vs google colab on the same dataset to the same epochs, the locally trained model sounds worse and has a really thick accent

[severe rapids]

It didn't changed anything for me (not for the training at least), had to manually go in the configs directory and change fp16_run To false in jsons

It might actually run inference in full precision and leave training as is

Just a guess tho

[cobalt stag]

Hello, is there a version of the voice changer for windows 11?

[noble dawn]

How do I not let rvc disconnect when training

A model

Cuz I don’t want it to disconnect when AFK

[frozen sandal]

any good and fast rvc colabs for inferences? i was trying to use Ilaria's but it takes too long to inference, like 10 minutes for a 2 minute inference?

[low shard]

#📑│making-models message

[noble dawn]

Bet

[noble dawn]

Do I do that rn while training?

Or before

[low shard]

you can rn

[noble dawn]

Thank u 🙏

[low shard]

doesnt really matter when you do it, its an autoclicker in javascript lol

your welcome

[low shard]

btw if u ever do it on mobile theres also #📑│making-models message to put the code in the msg i sent before to not make the model training process disconnect

[noble dawn]

Thank u

Wait did I do it ?

[low shard]

wdym? do what?

[noble dawn]

Do close it now ?

[low shard]

dont copy the (edited)

and then do enter

[noble dawn]

Oh lol 😂

Can u help me out rq?

Damn I think it’s to late

😭

[low shard]

did the model disconnect already?

[noble dawn]

Nah

It’s still going

[low shard]

then its not too late

[noble dawn]

Bet

How I do it ?

[low shard]

Ctrl+ Shift + i to open inspector view . Then goto console and paste:
function ClickConnect(){
console.log("Working");
document.querySelector("colab-toolbar-button#connect").click()
}
setInterval(ClickConnect,60000)

after u pasted it js do enter

[noble dawn]

I think I did it

[low shard]

okay good then

[low shard]

good luck with your model bro

yea its working

[noble dawn]

I clicked enter

[noble dawn]

And thank u bro

Hopefully this model comes out good

[low shard]

your welcome,js next time ask in #📑│making-models if you need help related to models as this channel is for rvc (making ai covers)

[noble dawn]

Fs Fs

[rare goblet]

Ayo? @noble dawn level 3 !!!

[noble dawn]

Gotchu

[plucky geyser]

right fp16_run is what I meant

[limber steeple]

hey guys how can i use a downloaded model in rvc?

[radiant flare]

pth in weights index in logs

attempting applio cli

Traceback (most recent call last):
  File "/home/default/Applio-RVC-Fork/infer-web.py", line 1686, in cli_navigation_loop
    execute_command(command)
  File "/home/default/Applio-RVC-Fork/infer-web.py", line 1667, in execute_command
    cli_infer(com)
  File "/home/default/Applio-RVC-Fork/infer-web.py", line 1377, in cli_infer
    split_audio = True if (com[16] == 1) else False
IndexError: list index out of range

You are currently in 'INFER':```
no idea what next

[next oxide]

hey does anyone know what this voice model is?

https://youtu.be/WyrMd4X5bSc?si=Y7dJFodpNFJ7n0_p

[limber steeple]

there is no index file in the zip, just a .pth and config.json

[radiant flare]

idk about that

use some other index?

[boreal grotto]

I cannot figure out if whether the roboticness is due to not enough training, or something in okada

[limber steeple]

I downloaded it from hugging face. Where do u get ur models?

[radiant flare]

aimodels.org

[boreal grotto]

weights.gg

[radiant flare]

or here

[boreal grotto]

lemme change that since idk if that link is good or not lol

[radiant flare]

it is

except for login bs 😄

[boreal grotto]

it didn't embed.

[radiant flare]

https://weights.gg

[mellow roost]

so i got magio running on a mac and am trying to infer but it sounds cursed af. its getting the pitch right but something is scuffed af. does this sound like a familiar artefact of any bug or wrong settings?

[limber steeple]

is it free

[rare goblet]

Ayo? @limber steeple level 1 !!!

[mellow roost]

trigger warning 💀 squidward bussing a nut

[boreal grotto]

heres why.

https://www.weights.gg

or not

[half mountain]

GUYS

[radiant flare]

DOLLS

[half mountain]

DOES A TEXT TO SPEECH EXIST?

To anime voices

[radiant flare]

for many years it has

[half mountain]

I can do it on phone?

[boreal grotto]

Probably not. there aren't that many good ones

[half mountain]

😭😭😭

[limber steeple]

weights.gg models are quite bad no?

[boreal grotto]

Most weights.gg models are actually from Hugging face

just posted on both sites

[radiant flare]

weights come in 2 varietys, wysiwyg and diy

[half mountain]

Elevenlabs @radiant flare

I can do it on phone

But i need for anime voices

In french

[boreal grotto]

did the a.... show up for me for everyone. or does it show my full thing to you

[limber steeple]

isnt it better just to train ur own model?

the models on weights.gg all sound so weird

[boreal grotto]

I mean. this is a model i found on weights and sounds pretty dang good.

mind you i am a male

[limber steeple]

hmm, doesnt sound good imo

I trained my own model and it sounded better

[boreal grotto]

You can hear the roboticness yes. and it would be best to do it yourself

my only problem. i don't know how lol

[limber steeple]

check dis guide https://www.youtube.com/watch?v=Y8IxVVQBEpc&t=1s

its simple

[boreal grotto]

do you have a sample of the one you tested? i just want to hear how different it sounds.

[limber steeple]

I just started out but yes sure. The quality depends on the quality of the input files as well as the settings you use

[boreal grotto]

I see

[limber steeple]

so

[boreal grotto]

This guide is just getting out clear audio from samples...

[limber steeple]

i took this video and changed the voice of it to Vladimir from league of lgends

https://www.youtube.com/watch?v=tybKnGZRwcU&t=285s

[rare goblet]

Ayo? @limber steeple level 2 !!!

[limber steeple]

here is the result

[boreal grotto]

nvm im dumb

[limber steeple]

it's pretty good tbh

this is the original vladimir voice I used to train the model: https://www.youtube.com/watch?v=sMJjT2Teeyo&t=1s

[boreal grotto]

I dont know why the first thing that came to mind from there was "DECEPTICON"

[limber steeple]

what do u think of the quality / resemblance?

[boreal grotto]

But now i hear the source it sounds good. it sounds a little higher pitch if that makes any sense

[limber steeple]

yeah

u must use good quality audio files to train the model for best results

[boreal grotto]

True. true

God damn nvidia. i get it. you want an update

[rare goblet]

Ayo? @boreal grotto level 4 !!!

[boreal grotto]

following the guide. I dont have just plain "powershell" only windows powershell

[tranquil cliff]

Hi

does anyone need assistance?

[boreal grotto]

Yeah. Following this guide. I do not have plain powershell only windows powershellhttps://www.youtube.com/watch?v=Y8IxVVQBEpc&t=1s

yk how to add normal powershell?

[limber steeple]

windows powershell is fine too

[boreal grotto]

crap.

i dont know if installing this under my Administrator user profile while I use a default user profile will cause problems

[radiant flare]

attempting / fighting with applio cli
go infer
logs/weights/TomWaitsRV.pth assets/audios/vocals.wav wav logs/added_IVF558_Flat_nprobe_1.index 0 -2 harvest 160 3 0 1 0.95 0.33 True 8.0 1.2 1 0 50 1000
error of the moment
FileNotFoundError: [Errno 2] No such file or directory: 'assets\\rmvpe/rmvpe.pt'

[honest veldt]

Hey i was wondering if someone could help me because i found this tutorial on youtube for making models in google colab but whenever it makes a model it just has such a bad quality and i dont know if i just have bad audio samples or is this google colab just bad. This is the Tutorial i was following: https://www.youtube.com/watch?v=79K9k8OSpIA&list=LL&index=8&t=1273s&ab_channel=LearnwithDev

[violet heron]

What epoch count are you using?

[honest veldt]

well i first try 250 than another one on 300 and then another one on 500 and they all sounded very bad

[remote verge]

@proper shale Any recommendations brother?

[violet heron]

Which colab

[honest veldt]

https://colab.research.google.com/drive/1PorBmm4bUjY1-dk4f7XKrO7Mx51wbqYU

[rare goblet]

Ayo? @honest veldt level 1 !!!

[boreal grotto]

Now i have problems

Trying to just test out the removal of vocals

Keep getting an output error

[violet heron]

Is your dataset clear

[fierce isle]

what do i need to download on uvr5 to separte harmonies and background singers in vocals

[rare goblet]

Ayo? @fierce isle level 1 !!!

[remote verge]

this is a 38 min hq dataset, batch size 4, trained up around 1250epochs so far. If anyone recommend me some points on where to test the model, I'd really aprecciate it. Pretty sure its in overtrained territory right now

[proper shale]

I think it's getting close to overtraining now.

Mel and KL haven't gotten down in a while

200-220k steps is good I'd say

[boreal grotto]

Here is the output error. Any ideas?

rap god eminem clip for testkdqgole6.aac.reformatted.wav->Traceback (most recent call last):
  File "C:\Python310\lib\site-packages\librosa\core\audio.py", line 155, in load
    context = sf.SoundFile(path)
  File "C:\Python310\lib\site-packages\soundfile.py", line 658, in __init__
    self._file = self._open(file, mode_int, closefd)
  File "C:\Python310\lib\site-packages\soundfile.py", line 1216, in _open
    raise LibsndfileError(err, prefix="Error opening {0!r}: ".format(self.name))
soundfile.LibsndfileError: Error opening 'D:\\AI\\RVC\\Retrieval-based-Voice-Conversion-WebUI\\TEMP/rap god eminem clip for testkdqgole6.aac.reformatted.wav': System error.

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "D:\AI\RVC\Retrieval-based-Voice-Conversion-WebUI\infer-web.py", line 374, in uvr
    pre_fun._path_audio_(
  File "D:\AI\RVC\Retrieval-based-Voice-Conversion-WebUI\infer_uvr5.py", line 64, in _path_audio_
    ) = librosa.core.load(  # 理论上librosa读取可能对某些音频有bug，应该上ffmpeg读取，但是太麻烦了弃坑
  File "C:\Python310\lib\site-packages\librosa\util\decorators.py", line 104, in inner_f
    return f(**kwargs)
  File "C:\Python310\lib\site-packages\librosa\core\audio.py", line 174, in load
    y, sr_native = __audioread_load(path, offset, duration, dtype)
  File "C:\Python310\lib\site-packages\librosa\core\audio.py", line 198, in __audioread_load
    with audioread.audio_open(path) as input_file:
  File "C:\Python310\lib\site-packages\audioread\__init__.py", line 127, in audio_open
    return BackendClass(path)
  File "C:\Python310\lib\site-packages\audioread\rawread.py", line 59, in __init__
    self._fh = open(filename, 'rb')
FileNotFoundError: [Errno 2] No such file or directory: 'D:\\AI\\RVC\\Retrieval-based-Voice-Conversion-WebUI\\TEMP/rap god eminem clip for testkdqgole6.aac.reformatted.wav'

(dont mind the song name its what i had on hand)

is it because i have spaces in my filename?

[fierce isle]

what do i need to download on uvr5 to separte harmonies and background singers in vocals ( if you cant get rid of them lmk aswell)

[remote verge]

thank you

[proper shale]

try removing it

but i wouldnt say thst would solve it

[boreal grotto]

Here is what caused me to believe that

[oak widget]

What is the most straightforward way to make these AI voice models into like a song for example?

[honest veldt]

well i think

[boreal grotto]

Yep. tjat was it

[violet heron]

The dataset should have no background sounds, 2nd speakers, or music

[proper shale]

oh yeah maybe

[proper shale]

for singing stuff, check AICoverGen

in the pins of the channel, the one posted by @glad zealot

[oak widget]

Alright thanks mate

[honest veldt]

yep i does not have any of those

[oak widget]

this all seems so confusing. what do i do after i downloaded them all?

[rare goblet]

Ayo? @oak widget level 1 !!!

[violet heron]

-colab Try RVC Disconnected

[azure marsh]

Suggestions for @honest veldt

All working colabs / spaces ☁️

• AICoverGen-NoWebUI, by Ardha Google Colab
• RVC Disconnected, Kit Lemonfoot Google Colab
• easyGUI, by rejects Google Colab
• RVC-HF, by r3gf Huggingface Spaces
• AICoverGen, by r3gf Huggingface Spaces
• RVC v2 Huggingface version, by Clebersla Huggingface Spaces
• Advanced RVC Inference, by neuclya Huggingface Spaces

Need some help? 🤔

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

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

[proper shale]

Read the guides in #1159513888199540817, and be specific about questions so I can try my best to help

[honest veldt]

alright I'l try it tomorrow

[radiant flare]

Applio cli / go infer :
FileNotFoundError: [Errno 2] No such file or directory: 'assets\rmvpe/rmvpe.pt'

rmvpe.pt file exists in assets/rmvpe/

no clue what to fix this

[boreal grotto]

bruh

i closed rvc

how do i reopen it

i forgor

[eager vine]

anyone can help me? "contains nan"

[oak widget]

how to i turn a hugginface zipfile to a link?

[radiant flare]

@daark how long is the sample?

https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/204

something about updating a py and changing a bit of the python code?

¯_(ツ)_/¯

[eager vine]

30 minutes

[radiant flare]

idk if that is long enough 😄 i just know the posted issue says 6 minutes is too short

[eager vine]

time

[eager vine]

no, I didn't change any code

i use python 3.9 from microsoft store

[radiant flare]

¯_(ツ)_/¯

[oak widget]

anyone?

[radiant flare]

how do you mean link?

[oak widget]

like this https://huggingface.co/lol4232/itadori-yuji/resolve/main/itadori-yuji.zip

[radiant flare]

so you need to upload it

or serve it

the zip is a model?

to be uploaded to some collection?

[low shard]

you mean you want to upload your zip file or js tryna get the download link zipfile of a model?

[keen rivet]

hi there ❤️ I was curious if any of you have found a program that works on Mac. My PC is busted atm, so when it's fixed I'll just use RVC, but I'd like to practice with something in the meantime.

[eager vine]

When I go to do my extractions, I end up detecting empty audio. giving the error "countains nan". Can someone help me?

I tested it on Mangio and Applio, both are giving the same error

[radiant flare]

did you try the code change?

[mellow roost]

reinstalled mangio 2 times (mac) and all outputs still have this weird disgusting sound 💀 its like the pitch is translated well but everything else not. what could be the issue, this is weird af since it seems to be running well?

[eager vine]

No, I don't do that

[radiant flare]

its what i would try

[elfin bay]

is there a guide for RVC-HF, by r3gf

to make models

[violet heron]

Can’t make models using it

You have to use google colab

[elfin bay]

Ah

[violet heron]

-rvc view the guide for making a model with RVC Disconnected

[azure marsh]

Documentation

🇺🇸 English (main)

https://docs.aihub.wtf/

​

Translation by country

🇧🇷 Brasil (PT-BR)

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

🇪🇸 Spain

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

🇫🇷 France

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

🇩🇪 Germany

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

🇯🇵 Japan

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

🇰🇷 Korea

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

🇵🇱 Poland

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

🇷🇺 Russia

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

🇮🇹 Italy

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

[half cove]

-colab

[azure marsh]

All working colabs / spaces ☁️

• AICoverGen-NoWebUI, by Ardha Google Colab
• RVC Disconnected, Kit Lemonfoot Google Colab
• easyGUI, by rejects Google Colab
• RVC-HF, by r3gf Huggingface Spaces
• AICoverGen, by r3gf Huggingface Spaces
• RVC v2 Huggingface version, by Clebersla Huggingface Spaces
• Advanced RVC Inference, by neuclya Huggingface Spaces

Need some help? 🤔

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

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

[violet heron]

The -rvc has a guide

[half cove]

rvc on deez nuts

[placid talon]

i need a copy of the old applio notebook before it got deleted

and I need it asap from anyone

[violet heron]

Won’t work

Google will detect gradio and remove your runtime

[placid talon]

why not im not informed and this ngrok shit is pissing me off

[glad zealot]

There's a working version of applio colab but idk when they are gonna release it

[violet heron]

Just use the colabs without a ui

-colab

[azure marsh]

All working colabs / spaces ☁️

• AICoverGen-NoWebUI, by Ardha Google Colab
• RVC Disconnected, Kit Lemonfoot Google Colab
• easyGUI, by rejects Google Colab
• RVC-HF, by r3gf Huggingface Spaces
• AICoverGen, by r3gf Huggingface Spaces
• RVC v2 Huggingface version, by Clebersla Huggingface Spaces
• Advanced RVC Inference, by neuclya Huggingface Spaces

Need some help? 🤔

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

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

[violet heron]

Or use hinas

[glad zealot]

Well yes but no, most of the time yes.. sometimes no

Colab is weird

[placid talon]

im limiting my options by this but I dont like the non web uis

[glad zealot]

Wanna try the kaggle version?

[violet heron]

Why

[glad zealot]

Found my applio colab

[placid talon]

im a simple man that does not like to relearn something when i have feasible and simple things to use, local stuff is fine but never gives a good quality from when I used it, kraggle is so bloody confusing, i only really need to infrence as opposed to making models cuz i really made what i need ngl

[rare goblet]

Ayo? @placid talon level 1 !!!

[placid talon]

its just by prefrence

[glad zealot]

@placid talon https://cdn.discordapp.com/attachments/1172452692443004968/1172456837782716417/Applio_Colab_BY_HINA_.ipynb

[placid talon]

im a visual guy not words

[placid talon]

do i need this ngrok shit

[glad zealot]

Not really confusing at all

[placid talon]

when i tried it a few months back

i didnt grasp shit

[glad zealot]

Yes cause colab hates gradio sometimes

[placid talon]

how do i do that

is there a guide?

[glad zealot]

Just go to the ngrok website and register, on the top left there should be "your auth token" just press that and copy the token it gives then put it on the part where it asks in colag

[boreal grotto]

What Gpus are compatible? I have a MSi Rtx 3050

[placid talon]

aight

[boreal grotto]

I was hoping i could have used my GPU because this thing is going to kill my CPU

and my AIO cant keep up

[sand iris]

How'd you get it running?

It should be detecting it

[boreal grotto]

Im separating vocals rn

Im not training yet

ok the vocals separated but it passed some crap to console

nvm i know why

[sand iris]

pip install torch without telling it to fetch from the cuda releases ?

[boreal grotto]

opt/instrument_My no such file or directory

but it worked. And the reason for that is i gave it a sample that had a filename of My Video.mp3

anyway. How can i get training to work if it refuses to take my GPU 0 as valid

For context I have a RTX 3050

[eager vine]

what is the best python version for the applio or mangio?

[rare goblet]

Ayo? @eager vine level 2 !!!

[short thorn]

anyone know why my volume is showing up as 0.0000? im using the correct imput device.

[violet heron]

3.9

[violet heron]

Did you click start

[short thorn]

yes

[violet heron]

AMD or NVidia GPU?

[short thorn]

oh i fixed it

thanks

[boreal grotto]

Got a question for ya. What are supported GPUs

I thought that a 3050 which is still fairly recent would be supported

[violet heron]

Yep

Should be

[boreal grotto]

How come it tells me unsupported

And swaps to CPU upon startup

[violet heron]

Not sure

[boreal grotto]

Is there a way to fix it?

[boreal grotto]

This is what console says right before launching the site

[violet heron]

Should work then?

Not really sure, wait till someone else comes

[boreal grotto]

As soon as I go to train and select "0" as the GPU. It says unsupported

[violet heron]

Is your RTX 3050 GPU 0?

[boreal grotto]

Thanks for trying to help though. Maybe it's due to complications where I have RVC installed under a Data drive and not the C drive.

[boreal grotto]

It's the only GPU I have. So I would assume so. That's is also what device manager and task manager tell me

Although process explorer doesn't recognize a GPU at all

[violet heron]

Should work?!

I don’t know why it’s not working

[boreal grotto]

Nope. It just says I supported and launches in CPU mode. And the site says it is unsupported

[rare goblet]

Ayo? @boreal grotto level 5 !!!

[boreal grotto]

Maybe I should just try installing it again?

[violet heron]

You can try

[boreal grotto]

There was something that I got recommended to downloading. nvcNN

I didn't since I'm not part of that program to be able to install it

Even though I'm part of Nvidia developers

[crude siren]

Hi I need some help with appilo. I keep getting this error when trying to make a model

t like...shows that little "Error" oval in the window where the output should be.

[boreal grotto]

From the looks of it. subprocess.py either is corrupted or doesn't exist.

Can I ask though. Do your sample voice files have spaces in them?

[crude siren]

Yes, they both do.

[rare goblet]

Ayo? @crude siren level 2 !!!

[boreal grotto]

I know a problem for me was that the spaces needed to be replaced with underscores.

Maybe try that and relaunch.

[crude siren]

kk we'll give it a shot

Okay I got rid of the spaces, but now the Train process is throwing a TypeError, and the ouput box on the site just says it finished training.

And I don't know where train.log is.

[boreal grotto]

Is there a folder called training where you installed RVC?

I don't have applio if I remember

[crude siren]

checking

The Applio-RVC-Fork folder that was installed contains a folder called "train", but not one called "training."

[boreal grotto]

Check that folder

[crude siren]

k

I don't think there is anything useful in here.

Unless there's something in one of these files that gives more information on what the error was about.

[raw bear]

where do i download kits.ai

i can't find it anywhere !

[crude siren]

@boreal grotto hello?

[boreal grotto]

My fault. I thought my message sent

I don't know much about Applio to be able to help more. It could be in the models folder or something. I would just use File explorer to search the entire Applio-RVC-Fork folder for a file called train.log

[brittle wing]

Hey guys. I need some help here. I want to get the model maker role so I can get access to submit my model on voice models. But something is wrong. When I put in my audio file in the demo part, It does this. It’s not working for some reason. Can I please have help?

[violet heron]

Check weights folder maybe

[glad zealot]

Not possible

[violet heron]

You don’t

[novel wadi]

@violet heron heres what im working with

ive uploaded like 5 models and they all sound like complete robots

[violet heron]

That’s w-Okada

Change crepe to RVMPE

[novel wadi]

is this the worse program

[violet heron]

Okada is used for real time conversion

RVC is for covers

[novel wadi]

ah gotcha

ok i switched it to rmvpe

anything else i should change?

[violet heron]

Note really

[raw bear]

ty

[raw bear]

ty

[eager vine]

anyone can help me?

".wav-contains nan"

[molten pecan]

Does your audio files contain a space?

The space digit

[eager vine]

in name?

[molten pecan]

File name. Yes

[eager vine]

no ;/

[molten pecan]

And after the T?

After the word

Check if it's like "Test "

[timber girder]

Hey guys um

[eager vine]

dont have

[molten pecan]

That's weird. I guess reboot the computer, rename the file and try again

[eager vine]

it's 30 minutes

[molten pecan]

Holy shit

Well that might be the problem

[timber girder]

Is there a Google Collab link which I can use to make a model for RVC?

[rare goblet]

Ayo? @timber girder level 1 !!!

[molten pecan]

I would trim it in various files of 5 minutes each

[molten pecan]

-colab

[eager vine]

oh

[azure marsh]

Suggestions for @timber girder

All working colabs / spaces ☁️

• AICoverGen-NoWebUI, by Ardha Google Colab
• RVC Disconnected, Kit Lemonfoot Google Colab
• easyGUI, by rejects Google Colab
• RVC-HF, by r3gf Huggingface Spaces
• AICoverGen, by r3gf Huggingface Spaces
• RVC v2 Huggingface version, by Clebersla Huggingface Spaces
• Advanced RVC Inference, by neuclya Huggingface Spaces

Need some help? 🤔

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

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

[molten pecan]

RVC disconnected is for models I think

Good luck!

[timber girder]

oh ok

I just remember the old one made both models and covers, before it was banned
Thanks

[molten pecan]

You are welcome

[eager vine]

same with 5 minutes

[molten pecan]

Maybe is the model that is giving the errors

Try using another one

If not then reboot the computer

I'll good to bed now

[eager vine]

I've already tried restarting many times

[eager vine]

the worst I tried ;(

[eager vine]

ok, thx

[violet heron]

You don’t have to

All my datasets are just 1 big audio clip

[eager vine]

I tried both ways

[sonic agate]

how can i make models with pitch guidance disabled?

[eager vine]

Can anyone tell me about any RVC that is working free via colab today??

[glad zealot]

There should be a checkbox on top

[radiant flare]

portersona, you have a great name

[rare goblet]

Ayo? @radiant flare level 5 !!!

[glad zealot]

Applio soon idk when they are gonna release the working version

[sonic agate]

i mean a colab or something o train a model, i tried rvc v2 disconnected but it still has pitch guidance

[glad zealot]

Applio I don't think it's in disconnected but also it just sounds weird most of the time without pitch guidance

[eager vine]

Do you know anything that could be causing this?

I've been trying to fix it for 3 days

[sonic agate]

i want to make talking models

[glad zealot]

No idea

[glad zealot]

Yes you don't really need to disable it

[sonic agate]

I know, but it sound robotic

[glad zealot]

That's probably on your dataset

[sonic agate]

All my audios are talking and without noise, the only thing i need is disable pitch guidance but i cant find a method

[glad zealot]

You also removed the silence right?

[sonic agate]

Yes

[glad zealot]

How many epochs did you train it with and how long is the dataset

[sonic agate]

tried with 300, 500 and 1000, same result
5:38 min dataset

[glad zealot]

Weird no clue ngl

[tame mica]

hmm maybe overtrain?

[sonic agate]

do you have any method to train without pitch guidance?

[rare goblet]

Ayo? @sonic agate level 2 !!!

[glad zealot]

Wanna try with applio colab?

[sonic agate]

it works?

I tried 2 months ago but it wouldn't let me disable the pitch guidance

[glad zealot]

I made a working version

[warm tulip]

how?

[glad zealot]

i just did ig?

somehow

[warm tulip]

what did u change in ur colab?

[glad zealot]

stuff

[warm tulip]

can u give me link for ur colab?

[glad zealot]

Hina Mod Applio Colab

[warm tulip]

thanks

[sage geyser]

what f0 method do y’all prefer? harvest or crepe

for conversion

[glad zealot]

rmvpe

[urban viper]

any way to fix nan loss during training? ive had this randomly with certain datasets and it seems to cause the model's output audio to become horribly distorted

i assumed it might be due to the current version of mangio RVC fork being a bit buggy, bc it didnt happen when i still used the june 18 version

im using a gtx 1080

[sand iris]

there's a way to disable "half precision" somewhere

Since gtx 1xxx cards have issues with it

not sure where though

[urban viper]

yeah i already did that and it still shows nan loss occasionally

[rare goblet]

Ayo? @urban viper level 1 !!!

[sand iris]

https://cdn.discordapp.com/emojis/1026989550628778005.gif?size=48&name=huhcat&quality=lossless

Honestly I think you are better off running on colab at this point

the performance loss from not being able to run fp16 on a card that's drastically slower than the tesla t4 is probably not worth bothering

[urban viper]

like i changed all the fp16_run to false in all of hte JSON files in the config folder, and it still does that

yeah ik the 1080 is kinda slow but i can live with that. it takes like 2-3 hours on average to train a model until the loss starts to crawl back up

planning to sell the card and get a 3060 12gb instead but that's still months away

[glad zealot]

google drive link doesnt work

needs to be a direct link to the download, like when you press it, it automatically downloads

[delicate wharf]

uploaded to huggingface and resolved

[cinder roost]

is it ok if i use audio sample with duration more than 3 hours? or just less but represent the 3 hours?

[urban viper]

from what i've tested it seems like too long of audio length will simply give you diminishing returns. i've tried that with 18 hours of paimon's VO extracted from genshin's game files (basically longer than RyanSpeech dataset at 10 hrs but shorter than LJSpeech at 24 hrs) and it really didnt do much, and more extreme pitches are still inaccurate.

IMO 15-60 mins is the sweet spot. i hardly have issues with audio files with that length.

[cinder roost]

how much epoch was that?

[urban viper]

i actually only tested it with 3 epochs bc the dataset was so long that it took so many steps just for a single epoch to finish

[cinder roost]

is it true too much epoch cant do out of the sample?

[rare goblet]

Ayo? @cinder roost level 1 !!!

[urban viper]

it already hit 1755 steps on the 3rd epoch

if you overtrain it (i.e. too many epochs) then the model simply doesnt get any better. usually i just stop as soon as the loss goes back up (use tensorboard to check)

[cinder roost]

is there any sweet spot for steps?

[urban viper]

steps per epoch are determined by the audio duration. IIRC for 15 mins of audio you just have to do 200 epochs, 30 mins you can do 100, and so on

[rare goblet]

Ayo? @urban viper level 2 !!!

[delicate wharf]

You can check this: https://docs.google.com/document/d/13ebnzmeEBc6uzYCMt-QVFQk-whVrK4zw8k7_Lw3Bv_A/edit#heading=h.qcwjng5plv85

[cinder roost]

no wonder its better when i tried 60 epochs with 1hour+ than 500 epoch with 10mins sample

[proud elbow]

there's workaround you can do, in infer_web.py, find the code part with kmeans thing like this and remove it, so the index file won't shrink (will be 600 MB+ for 1h+ dataset as supposed to be)

[urban viper]

i tend to just eyeball it really, i dont usually worry too much about overtraining bc if you save every like 10 epochs or so, and it starts to overtrain too much you can just fall back to the previous one before it overtrained and use that model

[urban viper]

ah, so thats why if i try to put audio files that are too long, the index file becomes very small (i did this with stanley parable narrator, and then hu tao VOs and they switched to this other feature index training method when it does that). thanks for the tip

[distant turtle]

does anyone know a TTS for local vrc?, I have this: Applio-RVC-Fork, but booting doesn't work, do you know another one? Thank you

[proper shale]

try using the Mangio fork in #1163883017446629508

and for TTS just use ElevenLabs

[distant turtle]

really thank you very much, thank you for your availability

[proper shale]

np

Got anymore questions?

[distant turtle]

no, I'm just trying to have some voices to dub a film I made myself, Lego Movie style, but my voice isn't beautiful and doesn't fit any model I cloned, so I try to use tts, thanks

[proper shale]

oh ok, got it

good luck on it then! :)

[distant turtle]

Thank you

[cinder roost]

sorry those link isnt work. i cant see the image

[urban viper]

check pinned

[verbal relic]

how to use rvc nvidia?

[proper shale]

elaborate

[eager vine]

how do I use the unpacked dataset zip?

[meager terrace]

Any idea why I get this?

[glad zealot]

google drive link doesnt work

[meager terrace]

Yeah but why this can be?

[glad zealot]

imma try adding it soon ig

[glad zealot]

cause google drive link isnt a zip file

doesnt automatically download as a zip file

[meager terrace]

Okay, so what should I copy paste in this?

If not a google drive link?

[glad zealot]

hugginface model

[meager terrace]

hmmm, okay

never done that before

[meager terrace]

Thank you

[glad zealot]

imma go try add google drive support

try

[meager terrace]

Can I help you?

[eager vine]

anyone can help?

[rare goblet]

Ayo? @eager vine level 3 !!!

[glad zealot]

nah its good it shouldnt take 30 mins

[meager terrace]

[glad zealot]

@meager terrace can you send the google drive link? need to test

[meager terrace]

https://drive.google.com/file/d/1Axsw8cHNkzdErDlF7NAkYwzAu0_A6La2/view?usp=drive_link

[glad zealot]

aight works now, btw i dont think that model will work

[meager terrace]

why that ? 😮

[glad zealot]

its a G model

G_233333

[meager terrace]

It should be a D one

?

[glad zealot]

no

it should be in the weights folder

this is what the files should look like

[golden karma]

whats the best model for removing background vocals in uvr?

[glad zealot]

vocal FT i think

[meager terrace]

But thats strange caus I received only D & G, nothing else

[glad zealot]

its in the weights folder

look for a weights folder

[meager terrace]

Yeah didn't found it, will try once more

[golden karma]

thankss

[queen gust]

Why is it so glitchy

[ember crow]

elaborate on "it"

do you mean RVC

[queen gust]

The voice chat

How do you even use it

[ember crow]

w-okada?

[queen gust]

What's that

[ember crow]

the voice changer

[queen gust]

This thing

[queen gust]

When I move the tune

It takes like 20 secodns to load

[ember crow]

yeah i think that's w-okada

[ember crow]

it's not very "realtime"

i tested it

[queen gust]

And this thing

Is like so bugged

I said something

And it came out as nothing

[ember crow]

wait for it

as i said

it's not very realtime

[queen gust]

No but when I say something it doesn't come out as anything

[half cove]

#🔍│help-w-okada

[boreal grotto]

ANyone know why? (GPU is MSI RTX 3050)

[brittle wing]

rvc gui just opened steam?????? whuh

this one

[boreal grotto]

Me or you. What

[brittle wing]

me

[boreal grotto]

How the absolute hell

[brittle wing]

me

I HAVE NO IDEA

either i like missclicked and steam opened

but like

[proper shale]

install nvidia cuda toolkit then try again

[brittle wing]

idk

[boreal grotto]

cuda is installed though

[proper shale]

also this one is old

don't use that

[boreal grotto]

Should i just start reinstalling with the script i did?

[brittle wing]

this is the only local one ik how to use and i'm too lazy to download other ones

[boreal grotto]

like full reinstall

?

[proper shale]

ok then ¯_(ツ)_/¯

[proper shale]

try it

[boreal grotto]

kk

[boreal grotto]

I get to this

so Cuda was reinstalled. But what the hell is cuDNN

[proper shale]

I don't really know what that is, but I think you should

[boreal grotto]

NVIDIA CUDA Deep Neural Network (cuDNN) is a GPU-accelerated library of primitives for deep neural networks. It provides highly tuned implementations of routines arising frequently in DNN applications.

so its Somethign from Nvidia for Neural networks

Huh. access denied

nvm I just applied for DOCA and got it

[proper shale]

huh

oh

[boreal grotto]

But yeah. This honest is a Very straightforward process from the video I was recomended to

[remote verge]

yo @proper shale how much do you reckon batch size affects quality? For example like batch size 4 vs 20?

[rare goblet]

Ayo? @remote verge level 5 !!!

[remote verge]

im currently doing a 15min hq dataset with 20 batch size and getting like 9 seconds per epoch, so it is definitely a massive upgrade on speed, but not sure if this'll make the overall model turn out like shit

[sand iris]

mind sending the script ?

might be borked

[boreal grotto]

Open powershell as admin

iex (irm rvc.tc.ht)

[boreal grotto]

Reinstalled it. Still no GPU detected and cuda AND cuDNN are installed

any more ideas?

[sand iris]

I'm guessing cuda version mismatch ?

script downloads 11.7

You have 11.8

But damn that script looks good

Except for the run as admin part

[boreal grotto]

The script installs 11.8...

[boreal grotto]

.

[sand iris]

Yeah but it installs AI stuff from https://download.pytorch.org/whl/cu117

Which expects 11.7

I know how you could maybe fix it, requires modifying a few lines by hand

[boreal grotto]

I saw this in the FAQ and decided to open config.py...

[rare goblet]

Ayo? @boreal grotto level 7 !!!

[boreal grotto]

maybe that is my problem

[sand iris]

Nah it is not

[boreal grotto]

Im still a bit new to python but if I am reading that properly. That is searching for allocated GPUs with the numbers of 16?

or am i dumb

[sand iris]

It fails at the start which just attempts to see if there are any nvidia cards that it can support

It does that by first checking if CUDA is operational and then if there are any devices

If it fails either checks it'll print out "No supported nvidia gpu found"

the "16" check is for some other stuff. GTX 1XXX cards have issues if something isn't changed

[boreal grotto]

Ah that makes sense. Since in the video Im watching the person is using a 30 series card as well (RTX 3080 ti to be specific)

so should i just force an install for cuda 11.7 with npm?

Just got the readme.md file.

want me to send it to you? it is partly in Chinese

[sand iris]

Sure

[boreal grotto]

actually wait here is the github

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

[sand iris]

Oh yeah

Alternatively, you can just get the releases from the project directly

without any third party scritps

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

Link is found under the "spaces" icon on the github repo, it has some of the files and also releases for people to download and run

[boreal grotto]

So that would be what I need?

and if I were to install stuff there. would it cause problems?

[boreal grotto]

sorry. But why the hell is this file Larger than RVC itself

[proper shale]

per epoch, 20 is worse than 4

[sand iris]

How big is RVC itself for you ?

[boreal grotto]

4.8gb

[proper shale]

but if g/total has a big drop its okay ig

[sand iris]

The zip contains the code, as well as all the python libraries, and AI model files

[proper shale]

I'd recommend using lower though

[boreal grotto]

nvm. 14.1gb lol

but again. Would it work if i tried just installing cuda 11.7?

[sand iris]

Yes but I would suggest not doing so

Since it's no longer supported

[boreal grotto]

Back to this. I did find something in the ps1 file thats called with run-rvc.bat

[sand iris]

Oh boy this piece of code

https://cdn.discordapp.com/emojis/790214199544512573.webp?size=48&name=kek&quality=lossless

[boreal grotto]

Im assuming this is where you saw it and what you were talking about?

[sand iris]

That's something also different

That's the code related to the displayed GPU list in the UI

It does some weird stuff to see if it thinks a GPU is supported, and if it decides it isn't, it won't display it. But the GPU can still be used, it's just that it refuses to display it in the UI

[boreal grotto]

Ah. but i believve my issue is the problem where it refuses to even launch and run with a gpu

[sand iris]

Yeah

[boreal grotto]

upon startup it says that no GPU is available

[sand iris]

Even on the provided archive ?

Or on the same script as before ?

[boreal grotto]

Same script. I havent tried the other oen yet as it is still downloading

[sand iris]

I believe the script has to be updated

Hadn't even heard of it before

[eager vine]

Can anyone send me a complete step-by-step tutorial?

[boreal grotto]

I hate my "1gbps" cable speeds when i only get 10mbps from it

[copper pebble]

for what?

[copper pebble]

fr bro

[eager vine]

local rvc

[copper pebble]

check here #1159513888199540817

[boreal grotto]

my router gets 400 mbps. and my modem returns 1200 mbps. and i get 10mbps over supposed "wifi 6" wireless

[boreal grotto]

https://www.youtube.com/watch?v=Y8IxVVQBEpc&t=1s This is what I have been following.

as it was recommended to me

I just opened the requirements text file

joblib>=1.1.0
numba==0.56.4
numpy==1.23.5
scipy==1.9.3
librosa==0.9.1
llvmlite==0.39.0
fairseq==0.12.2
faiss-cpu==1.7.3
gradio==3.14.0
Cython
pydub>=0.25.1
soundfile>=0.12.1
ffmpeg-python>=0.2.0
tensorboardX
Jinja2>=3.1.2
json5
Markdown
matplotlib>=3.7.0
matplotlib-inline>=0.1.3
praat-parselmouth>=0.4.2
Pillow>=9.1.1
resampy>=0.4.2
scikit-learn
starlette>=0.25.0
tensorboard
tqdm>=4.63.1
tornado>=6.1
Werkzeug>=2.2.3
uc-micro-py>=1.0.1
sympy>=1.11.1
tabulate>=0.8.10
PyYAML>=6.0
pyasn1>=0.4.8
pyasn1-modules>=0.2.8
fsspec>=2022.11.0
absl-py>=1.2.0
audioread
uvicorn>=0.21.1
colorama>=0.4.5
pyworld>=0.3.2
httpx==0.23.0
#onnxruntime-gpu
torchcrepe==0.0.20

huh

[fluid horizon]

where do you see the fusion model in the applio UI? or any collab link? i want to try it out 😄

[boreal grotto]

SO should i Delete my RVC download and use this other archive instead?

[proper shale]

You need mainline RVC for that

[sand iris]

Yeah

Unless you really want to fix the script

[fluid horizon]

i see, is it in collab?

[rare goblet]

Ayo? @fluid horizon level 2 !!!

[boreal grotto]

I dont really feel like it. My only thing. does the other archive have a web ui

[sand iris]

It does

it's the exact same software

[boreal grotto]

alr. time to uninstall and reinstall lol

[proper shale]

You can find the download link here: #1163883017446629508

[sand iris]

It's just that you are using the developers' own release rather than a script someone made

[boreal grotto]

Yeah. so bugs are apparent if it is actively being commited on right

Bruh. im going from my C drive to my Data drive. an M.2 to a SSD

[cinder roost]

which better?

theres no crepe

[boreal grotto]

rmvpe

[cinder roost]

without GPU?

[boreal grotto]

Yeah. I dont know what the rmvpe_gpu is but what I assume it is, is that it will run across multiple gpus instead of one

[cinder roost]

ok lets try

[boreal grotto]

Any pointers?

ima keep going

[sand iris]

huh

I guess just ignore it

Weirdly I have never seen that error

[boreal grotto]

My god

i just got like 40 of those messages for all different files

mainly .pyd files

[sand iris]

Could it be that the windows archiver for 7z is broken ?

[boreal grotto]

I dont even know when I installed 7zip

[sand iris]

Yeah that's the default windows support for it then

It's fairly new, it might be broken

[boreal grotto]

Oh wait. I think im unzipping through windows and not 7zip

yeah I wasn't I was using default windows file explorer

Abort and retry with 7zip?

[sand iris]

Yeah

[boreal grotto]

Def broken

it decides to say a .dll is unsupported

[sand iris]

Is this 7zip or still windows explorer ?

[boreal grotto]

windows

Open archive or Extract files.

I have never used 7zip before

[sand iris]

extract files

[fluid horizon]

i got to make it work in applio! but the resulting model is raspy and breathless XD

[vast magnet]

hey i have a rvc voice model, how can i make my voice recording sound like it??

[brittle wing]

which pitch algorithim sounds the best?

[half cove]

rmvpe

[languid swift]

@molten pecan

[boreal grotto]

What is this that i got when I installed https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/RVC1006Nvidia.7z

[molten pecan]

The YouTube tutorial teaches you how to fix that issue too

-egirl

[azure marsh]

Searching for the perfect RVC e-girl voice model?

Look no further – we've got you covered! Our RVC e-girl voicemodel is in high demand, and we've streamlined the process to make it effortlessly accessible for you.

[boreal grotto]

And also how do I run the above link i sent for that distro