✨│ai-help

[analog obsidian]

but thats fine

rvc gets confused sometimes

[bleak nymph]

so if it goes up for more than 15 - 30 mins then i stop it

[analog obsidian]

it should start to raise forever in the 200 epoch zone

yeah

[bleak nymph]

alr

[analog obsidian]

but if you aren't sure just tell me

[bleak nymph]

alright tyy

u two have been a great help

[dusty bramble]

all the credit goes to him, atp i been just watching

[analog obsidian]

after you noticed the g/total rises forever you choose the lowest point BEFORE the rising

thats your best model

[bleak nymph]

i see

also can someone describe what an epoch is in simple terms, i wanna explain it to my friend but i don't understand it well enough to explain it

[analog obsidian]

epochs are snapshots of the model

[bleak nymph]

snapshots?

like minecraft snapshots

[analog obsidian]

yeah

rvc models are numbers and calculations

funny numbers and calculations

[crude flame]

thats actually a pretty good way to put it

[analog obsidian]

rvc saves these calculations in epochs

[bleak nymph]

lmaoo

[analog obsidian]

there are bad calculations and good ones

the bad calculations are when the graph raises

and the good ones are when it goes down

in one point of the training rvc gets very smart

[analog obsidian]

and vomits a very good epoch

(the lowest point in the graph)

and that epoch stores the very good calculations and predictions

[dawn copper]

Hey all! So, I'm trying to use Applio for the first time. I followed all the instructions, buuut.. When I try to do some TTS using a voice model, I get an error. My GPU is an RX 9070, which I'm assuming is the problem tbh.. So far, I've had no real luck in using it for playing with AI things. Anyway, here's the error, if it matters.

rocBLAS error: Could not initialize Tensile host:
regex_error(error_backref): The expression contained an invalid back reference.```

[crude flame]

zluda doesnt support 9000 series yet as far as i know

[dawn copper]

Annnd that explains it! I didn't know it was zluda as a whole that didn't support them. I'll be looking forward to it getting some support.

Thank you.

[bleak nymph]

i thought u said rtx 9070 and thought u were a time traveller for a sec

😭

[dawn copper]

I wish!! But no. I probably should have said "AMD 9070".

[bleak nymph]

lmao nah i understood

i was considering getting amd but decided nvidia

glad i did because literally everything python requires nvidia graphics card 💀

@analog obsidian uh the graph hasn't changed

a bit

[analog obsidian]

o i forgot to tell you something

[bleak nymph]

o

[analog obsidian]

click that

[bleak nymph]

WOAHHH

[analog obsidian]

XD

[bleak nymph]

things changed

it's a flat line now 😭

nah it's still beginning basically

i gotta wait

[analog obsidian]

click the third box

or something more ez

just press f5

[analog obsidian]

no joking btw, this does indeed updates the graphs xD

[bleak nymph]

oh yea

f5 just refreshes tab

anyway is late now

[simple ore]

see here https://github.com/IAHispano/Applio/issues/999#issuecomment-2736368448

for 9070

[bleak nymph]

i think imma continue it tmrw

[analog obsidian]

for resuming the training put the model's name
don't extract nor preprocess again (important)

put the same batch size and save epoch frequency

[bleak nymph]

is at 37 rn

alr

[analog obsidian]

and don't use fresh training

if done correctly the graphs are gonna be fine

and they will continue logging as usual

[bleak nymph]

alr

[analog obsidian]

🦈

[bleak nymph]

also what time is it for u rn

if i were to guess, 8pm

the previous day

am i correct

[analog obsidian]

10:09pm here

🦈

[bleak nymph]

ahh

nice

1:09am for me

am tired

honk shoo or whatever

💤

[dawn copper]

Ooo, thank you! I'll try this out later tonight, or tomorrow.

[simple ore]

you may need to amend the patch zluda .bat to download the latest version

i dont have AMD card in my PC any more, so I cant test the latest zluda and the patch it requires

rvc\lib\zluda.py

[dawn copper]

I think I understand. I'm downloading Rocm 6.2 now. Once that's done, I'll start trying zluda patches. Though, I'm getting pulled away for now.

[simple ore]

there are some hijacks in that file, but they may not be requred with latest zluda

[crude flame]

The latest version works fine for me and I didn't change any code

I got a 6700xt

[simple ore]

zluda 3.9.1?

[crude flame]

Yeah

[simple ore]

I think Lee made some changes like enabling CUDNN and other stuff

so there's no need to disable them

In order to enable cuDNN acceleration on supported device, you should download and unpack HIP SDK extension upon your existing HIP SDK 6.2 installation.

HIP SDK extension: DOWNLOAD

[thick summit]

is there any google collab active?

been trying since yesterday

tensorboard doesnt work

[simple ore]

make sure numpy==1.26.4 gets installed

make a new code cell and pip install it

[dawn copper]

So, peeking in at the zluda patch... And forgive my ignorance ahead of time!
But in the zluda patch, it has a line that says
curl -s -L then it points to a github Zluda release. I'll update that link to https://github.com/lshqqytiger/ZLUDA/releases/download/rel.ae0540beb129ffd140226ce956b386619b38f84c/ZLUDA-nightly-windows-rocm6-amd64.zip
? Which that's the newest Zluda release.

And I'll leave the stuff like copy zluda\cublas.dll env\Lib\site-packages\torch\lib\cublas64_11.dll /y alone, I assume? (There are 3 lines similar to that one, all with different .dll files.)

[simple ore]

you can download the .zip manually

unzip it to zluda folder

and just use the part of .bat that copies the .dlls into site packages

[dawn copper]

Ah, ok. Thank you!

[tardy cedar]

does anyone use software like voicemeter [i think] to modify the mic settings before it gets sent to RVC? i use a usb mic and theres no noise cancelation or anything on it to change settings

[simple ore]

with nvidia you can use broadcast app

mic -> app -> voice changer

[tardy cedar]

yeah thats true, i tried it and dident really get better results than whats in rvc. do you use that personally?

[simple ore]

it is amazing

[tardy cedar]

alright bet ill run it again. i really do appreciate the advice

[simple ore]

https://x.com/thegunrun/status/1252789873699745792?lang=en

theGunrun (@theGunrun) on X

Nvidia's RTX Voice tech looks AMAZING. I can finally use a desk mic again with my mechanical keyboard!

@Barnacules's test with the fan and hammer is so impressive

https://t.co/ES7ELMiMQS

X · Apr 22, 2020, 2:42 AM

that was 5 years ago

[tardy cedar]

thats wild lmao, i used it for a while on normal streams but i thought it made the audio quality rough compared to just the noise cancelation in obs

[knotty moth]

@bleak nymph fyi fresh training is to start over training, deleting the G & D files and tfevents file (tensorboard log)

by not selecting that option, leaving those files that means training will resume

[bleak nymph]

wait what

ohh

i understand

[simple ore]

fresh training = i want to clean up the training atttempt and start over

it does not delete sliced audios and other things, just D/G and events

[bleak nymph]

yea

[simple ore]

if you have D/G files in the folder the training will resume as long as the number of epoch is larger than saved

[bleak nymph]

can’t wait to continue it tmrw

[tardy cedar]

good call about nvidia broadcast, its def better than what we have in obs

[grim bay]

So is there a way to separate a part of the voices in this reverb audio? I need to get that since it was in the lyrics

Same thing with this too

[tame mica]

no

[grim bay]

What about I choose another de-reverb model?

Does that change anything?

[simple ore]

it is hard to unbake a cake into a flour and sugar

[grim bay]

But what if I use flour and sugar but different techique to bake a cake?

[simple ore]

find original stems, master them differently, sure

[grim bay]

So this one works but I cannot remember if UVR-DeEcho-DeReverb used to be the best model for dereverb

[warped thistle]

Is a 3060ti 8GB TI LHR good enough to have a realistic sounding voice? I've gotten some models alright but ALL models give me static during long sentences.

I'm using MMVC

[hallow thistle]

For W-Okada, go to #🔍│help-w-okada. This channel here #✨│ai-help is about RVC programs, not W-Okada.

[warped thistle]

I thought RVC meant realtime voice change

my bad

[hallow thistle]

In this context, RVC doesn't refer as realtime voice changer. RVC is another program that used to do AI voice cover.

[vapid mantle]

If I change the batch size to 3 instead of 4, will the quality be better?

@hallow thistle

Do you have any information about this subject?

[simple ore]

less batch size, more instability in the training process

[vapid mantle]

Duration is not important, will the quality increase?

[simple ore]

depends on the set

small batch = smaller steps, may get stuck in a local minima

very large batch = big steps, may overshoot the target

it is all about experimenting

[vapid mantle]

ok

[viscid moss]

https://github.com/Eddycrack864/UVR5-UI/blob/main/info/docs.md#best-models

[light pelican]

Hey which rvc i should use for training voice? I have RTX 3060 12gb

should i go for mainline or applio pls?

[hallow thistle]

Applio.

[light pelican]

Is it better quality?

[hallow thistle]

More like it's easier to use, but it should give better quality at the same time.

[light pelican]

Ok ty👍

[grim bay]

May I ask about UVR5-UI on local?
Do I have to download the model manually?

[viscid moss]

If u re using UVR5 u need to import the new models manually, But on UVR5 UI models are downloaded automatically, pls refer to this thread
https://discord.com/channels/1159260121998827560/1343297771104501871

[grim bay]

Also, when I use De-Reverb by anvuew, it detects some parts of the lyrics as reverb and separate it. Do you know what model should I use then?

[viscid moss]

Hmmm try the Sucial ones, and maybe the anvuew mono de-reverb. Ur songs are kinda hard to separate

[knotty moth]

because it is not centered and gets rid of

[grim bay]

Similar result when using the Sucial ones

[viscid moss]

Have u already separate the lead vocals and backvocals?

[grim bay]

I don't know, I just separate the vocal and music first then de-reverb

Btw, when I use the UVR-DeEcho-DeReverb model, this works okay but idk if this model is the best option

[viscid moss]

Use the karaoke model first and then the dereveb model

[viscid moss]

Well, that model is kinda old

[knotty moth]

sucial ones are worse, and most lead vocals & speech under normal mics are centered so it should be easy to separate

the remaining one might be room reverb which can be removed using dereverb mono or RX 11 dialogue isolate

[knotty moth]

it still works for mono one or backing vocals but has 17.5khz cutoff

[viscid moss]

I see

[knotty moth]

uvr de-echo can also work quite more aggressive

though note that it basically noise gates weak signals

[grim bay]

The Karaoke model took the same parts too

[viscid moss]

Damn

[grim bay]

So any solutions for this now?

[viscid moss]

As far I know, nope

[grim bay]

Damn, guess I have to use UVR-DeEcho-DeReverb and then De-Reverb by anvuew

[simple ore]

and then end up with junk

but give it a try

[heavy arrow]

is this clean enough for conversion?

it takes a min for my local installation to start and i dont got enough resources to run it with UVR at the same time so im fishing for opinions before i try

[simple ore]

inferring that file?

garbage in, garbage out

using it as a training set? double garbage

it is a repeated nonsense

[viscid moss]

(-)*(-) = +

[viscid moss]

If u don't have a good pc for UVR use it on Colab or HF Spaces

[heavy arrow]

not using it as training or anything, literally just putting annoying orange voices over it to mock it
i get that its bad. thats why im putting shitty ai over a shit song. MYOB and answer my question next time instead of scrolling a help channel to start an argument

[heavy arrow]

wait... he's mathematically

[viscid moss]

XD

[heavy arrow]

alr thanks ive been avoiding colabs because most of them are kinda broken ATM but my computers just barely good enough to use ultimate-rvc to convert songs in abt 40sec

[viscid moss]

Well, u can also use HF Spaces it's easier than Colab.
Max audio lenght is 20 mins btw, and it takes like 1 min to separate

-hf

[karmic olive]

🤗 Huggingface Spaces

• UVR5 UI, by Eddy and Ilaria Huggingface Spaces
• Ilaria RVC Zero, by thestingerx Huggingface Spaces
• RVC⚡ZERO, by r3gm Huggingface Spaces
• Applio, by IA Hispano Huggingface Spaces

More commands: -audio, - colabs, -kaggle, -local, -rvc, /help

[heavy arrow]

sweeet, tysm ^^

[viscid moss]

Ur welcome

[bleak nymph]

hi guys

i've been training for a few hrs

my model is at 355 epoch

but the tensorboard line is still flat

idk what that means

it's still improving a lot? or no

doesn't seem tobe going up

[bleak nymph]

it's set to finish at 500 epochs so i'm wondering if it's not enough if i need to train it more than that

[analog obsidian]

reduce smoothing to 0.6 and turn on ignore outliers

[bleak nymph]

how do i do that

[analog obsidian]

just find it

im not on my house rn lol

[bleak nymph]

o

[bleak nymph]

on tensorboard? 😭

im not good at it

OH ON THE SIDE

i'm dumb

sorry

[bleak nymph]

both are already set to that

[analog obsidian]

restart the tensorboard site

and press the third square button thats below the graph

wait

hold on

you should be looking the scalars tab

go there

gray graphs = not good

[analog obsidian]

scalar graphs = great

[bleak nymph]

oh 😭

this one right

it's still improving ig

395 epoch almost 400

it seems if it continues at this rate, even 500 epoch will not be enough (i think)

[analog obsidian]

its not improving lmao

[latent cypress]

so the model stop improving after 10k?

[analog obsidian]

around there

i cant tell

need a more smooth graph

anything above 15k is rising and overtraining

[latent cypress]

oh i see

[bleak nymph]

it isn’t?

[bleak nymph]

oh 😭

wait i thought lower was better

actually, isn't this the best one?

idk how this really works

[analog obsidian]

reduce smoothing to 0 and turn off ignore outliers

choose the lowest point before 20k

[bleak nymph]

i stopped the training for now

[analog obsidian]

also show me the graph bc im not sure when it started to rise

[bleak nymph]

should i open it back up

[bleak nymph]

oh alr

[analog obsidian]

started to die around 15-20k but i want to be sure

[bleak nymph]

which graph should i show u

[analog obsidian]

avg 50 g/total

[analog obsidian]

smoothing to 0
disable ignore outliers

click third box

[bleak nymph]

3rd box?

ah

here

[analog obsidian]

you still havent disabled ignore outliers bro

[bleak nymph]

wait it's cut off 😭

[bleak nymph]

i did

[analog obsidian]

no???

[bleak nymph]

this one

right

[analog obsidian]

in the scalars tab bro

ignore the cringe gray graphs

[bleak nymph]

i aint the sharpest tool in the shed

[analog obsidian]

choose the lowest point in the 10k zone

you know which one

pretty obvious

[bleak nymph]

the bright colour or the faded colour

does that matter

u mean tthis point, right?

[analog obsidian]

yuh

[bleak nymph]

alr

how do i choose it

it's 10.6k

[analog obsidian]

find the epoch closer to that number

[bleak nymph]

this one, right?

[analog obsidian]

yup

[bleak nymph]

alr sick

[latent cypress]

bruh, you're such a noob here compared to you in the audio separation server 😭

[analog obsidian]

its always the next epoch after that value just in case
if its 12.4k you choose 12420 ish, never one prior like 13800

[bleak nymph]

LMFAO so true 😭🙏

we all start from somewhere tbf

[faint elk]

iis there a guide somewhere on how to get rvc the newest version running? im a bit confused

[latent cypress]

real

[bleak nymph]

lmao

the model sounds

interesting

[simple ore]

check other losses - fm and mel

and gradients

[analog obsidian]

well its a noisy dataset

also like noobies said
every losses actually matters

not just that one exactly

[simple ore]

G loss may go down like that, but if g norm is in 50000+ range, it is likely messed up

then there are things like this

disc is getting better, so fm loss is slowly going up and norm g also going up because of that

but as long as it is not a sharp rise from 2 to 10 it is fine

[analog obsidian]

is possible for a small model reach that range tho?

[simple ore]

I saw that with finetunes when I used a very different dataset comparing to the pretrain.. .like singing dataset vs speaking pretrain

[analog obsidian]

i wonder if someday this will get fixed

idk how seoul managed to keep fm going down

[simple ore]

training a voice on his own hifigan klm4.9 pretrain

I bet

[analog obsidian]

loool yea

[simple ore]

lemme check that

[analog obsidian]

🦈 👍

[analog obsidian]

👀

[simple ore]

next test is default pretrain

[analog obsidian]

nice

[simple ore]

gonna take about an hour to finish this

just doing 100 epochs

[analog obsidian]

its fine, fm raises pretty quickly

[bleak nymph]

um

chat

my model trained on me speaking only works better on singing then it does on actual speaking

[analog obsidian]

xD

[bleak nymph]

😭

i tried putting in a file of speaking but it sounded nothing like me because it copied the range of the original speaker (way higher than me) so i set it to -12st and it became so raspy and robotic and cursed

it never usually does this

[analog obsidian]

well your dataset is noisy and uncleaned

[bleak nymph]

that's not the issue i dont think

[analog obsidian]

pitch too low

increase it

[bleak nymph]

really? usually an octave works fine but alr then

[analog obsidian]

🦈

[bleak nymph]

thing is i'm a bass he's a tenor

soooo

yeah

i tried -6 still weird and dont sound like me

[analog obsidian]

try 0

[bleak nymph]

i did try 0

it was way too high

[analog obsidian]

-3 then

[analog obsidian]

the thing that happen at 0:05 is when the dataset lacks vocal fry

[bleak nymph]

vocal fry? i don't know techniques that much

[analog obsidian]

aka raspy voice

[bleak nymph]

u mean like raspiness right

ah

my voice is usually rough and raspy which is the interesting part

[analog obsidian]

idk how you sound sooo

and i dont know the original audio either

[bleak nymph]

oh right 😭

that's true

[analog obsidian]

🦈

[bleak nymph]

for reference ig

i just got off an hour long vc lol

[analog obsidian]

and the audio used for inference?

[bleak nymph]

it couldn't be more different

imma try smth that everybody knows

[analog obsidian]

that voice is way more raspier than yours

[bleak nymph]

fitness gram pacer test lol

[bleak nymph]

yeah true

in interviews it's a lot lot more

when singing, not raspy at all i don't think

[bleak nymph]

woah

[analog obsidian]

you just need to have raspiness in the set to inference raspy voices

[bleak nymph]

yeah seems like it

[analog obsidian]

also voice similarity depends in various complex factors

[bleak nymph]

do accents play a big part in it?

i have a weird blend of both american and british

[bleak nymph]

not in here tho

[analog obsidian]

consistency matters more

[bleak nymph]

so a consistent accent?

[analog obsidian]

consistent voice

[bleak nymph]

ah

like volume

[analog obsidian]

so not mimicking someone's else voice

[bleak nymph]

well i wasn't mimicking anyones voice in my dataset

thing is i'm a complete noobie at singing too so i copy other people's voices for some reason 😭 it's so weird and idk why i do that

i feel like if i added singing to the dataset it'd be worse

[analog obsidian]

you were laughing a lot in the dataset?

[bleak nymph]

i don't think so?

i was playing games tho so possibly

it has a bit of shouting

[analog obsidian]

but i think what affects the most is when the set is too monotone

[bleak nymph]

i feel that's good tho, more variety

[analog obsidian]

that usually causes the voice to not sound exactly like the original one

[analog obsidian]

I noticed this when training game character models, if I infer voices very different from the character's, the result will be weird (deepl translate XD)

[bleak nymph]

that's interesting, i would assume that monotone voices are easier to make ai's out of than expressive voices, right? but no apparently not

[analog obsidian]

monotone voices are easier to train

[bleak nymph]

oh

[analog obsidian]

but the least accurate ones

[bleak nymph]

yeah that's what i meant

i would assume it'd be the more accurate ones but nah

the apple one is ridiculously accurate though it's crazy

too bad it's just text to speech

[analog obsidian]

try to infer voices closer to yours

[bleak nymph]

what do you mean by that

[analog obsidian]

don't try to inference voices too high or too deep

[bleak nymph]

yeah no it wasn't

it was a regular VC i was in and i recorded myself in audacity

i didn't sing or go higher than i usually go

or lower

[analog obsidian]

my model always sound like the original person regardless of what i try to inference because the dataset has every pitch he can do

[bleak nymph]

yeah but every pitch i can do is not a lot

at all

[dusty bramble]

[analog obsidian]

skill issue

[dusty bramble]

how the model went?

[bleak nymph]

i think that's why it's harder to train ai's on me

[bleak nymph]

still testing

[analog obsidian]

good when he inference monotone voices
bad when he tries something too different from his

due to his dataset lacking pitches

[bleak nymph]

i haven't really tried inferencing monotone voices

i tried a song and it wasn't bad

[analog obsidian]

also known as monotone voice

[bleak nymph]

true 😭

that's the full thing 😭

[dusty bramble]

gotta get a dataset of u playing a fun game💀

[bleak nymph]

LMAO

[bleak nymph]

it was

it was a dataset of me talking to my friend and playing lots of games

[dusty bramble]

do karaoke

[analog obsidian]

no need to go crazy in the set in order to have near perfect voice similarity
just have recordings of your whole voice range

[bleak nymph]

yeah nahhhhh

[bleak nymph]

but it was my whole vocal range

[analog obsidian]

my model was trained using regular conversation, nothing crazy

[bleak nymph]

me straining my voice isn't a part of my vocal range

i strain when i sing

i don't strain when i talk

so the dataset was me talking comfortably and i wasn't straining

[analog obsidian]

speech models arent good at inferencing singing i already told u that

[bleak nymph]

then why'd it work better on singing then talking?

imma try different speaking audio

fitness gram pacer test

[analog obsidian]

thats what u believe

also the original pretrain was trained using very monotone speech

[bleak nymph]

oh

[analog obsidian]

is not a pretrain made for singing

lacks high pitches

[bleak nymph]

hm

my highest note i've hit was in the 3rd octave

that's quite low

is it not

i think i just need to get better and more expressive

def more of a me fault than ai fault

something happened

[analog obsidian]

i remember that error but i cant remember what causes that

[bleak nymph]

rip 😭

[bleak nymph]

it was too long i think

i cut it down and it work

[analog obsidian]

O

yeah

it was actually that

now i remember

thanks

🦈

[bleak nymph]

yeah this is uhh weird

too high or too low?

or smth else

[analog obsidian]

maybe too high

[dawn copper]

Ok, so. I did all this (properly, I think?)
I can now run Applio. The only error I got when starting it, was

INFO: Could not find files for the given pattern(s).```
Which is odd, but I assume not problematic.

I'm trying to run an AMD 9070 with Applio. Noobies was trying to help me get it set up yesterday, which I'm grateful for.

Buuut I do have a problem. Which is that, when I try to use inference, or TTS, pretty much nothing happens. TTS gave me a driver crash every time I try it. My GPU usage doesn't move on either. On inference, my CPU will use about 15-20%. The Applio cmd window says "Compilation is in progress. Please wait...", so I waited for 15 minutes, and nothing seemed to happen. I've tried different voice models and audio clips. Just for the sake of testing it, I cut an audio clip down to 2 seconds. That's the one I waited 15 minutes on, with no result.

[simple ore]

you did not read the installation note for amd

[dawn copper]

Oh! I did read that, but I thought that was with the first startup of Applio. So the first time I started it up using the run-applio-amd.bat, I just went afk for about an hour, and came back. My apologies!

[simple ore]

no, it only does it when you run inference and traning

inference has a small subset of computing tasks, but anyway.. should take ~20 min

index is a lookup table that translates voice features.. like english audio 'the' into german voice model's 'ze'

use index search > 0

[dawn copper]

Okay. I Got the first Zluda thing done. Now it's just taking a couple of seconds on inference and TTS. However, audio is completely blank. On TTS, there's just a Pop! sound at the start, and end. On inference, it's purely silent the entire time. I opened it in Audicity to double-check, and yeah, it's just a flat line.

Sorry for being a pain. I am trying to search if there's a solution to my problems, before I ask.

[graceful latch]

why cant i send pic here idk how to describe the problem

[simple ore]

inference empty too?

close applio, edit app.py and comment out line

import rvc.lib.zluda

with #

then try again

[dawn copper]

The place where I input audio is working. But the output (Bottom right of Convert), has no audio being output.

I did as you said, so now app.py has

# import rvc.lib.zluda```

Now, as soon as I click Convert, the Applio command prompt gives:
```  File "C:\Applio\env\lib\site-packages\torch\functional.py", line 665, in stft
    return _VF.stft(input, n_fft, hop_length, win_length, window,  # type: ignore[attr-defined]
RuntimeError: cuFFT error: CUFFT_INTERNAL_ERROR```

Oops, that's only the last line. It starts with
An error occurred during audio conversion: cuFFT error: CUFFT_INTERNAL_ERROR
And then it shows a total of 8 files. I can paste it all in here, if you like. And again, thank you for trying to help me with this! I do appreciate it.

[simple ore]

okay, revert back

[elfin geode]

I want to ask, I want to dub my own animation using rvc for female voice. If I use rvc will I get into trouble like copyright. How do I know which rvc model doesn't have that problem?

[simple ore]

it could be the model crashing with fp16

the empty audio i mean

[dawn copper]

I'm trying it in fp32, using Edge now instead of Firefox... Juuust in case that matters!
Annd it just finished. Same thing. No audio.

And I did restart Applio entirely, after changing to fp32.

[simple ore]

this was inference fix for fp16

https://github.com/IAHispano/Applio/commit/837b945c69d3c057e5de452ef5af7d252ca8293b

rvc/infer/infer.py

[dawn copper]

Is rvc/infer/infer.py The only part I need? Or should I be using everything in that page?

[simple ore]

just that for inference

[dawn copper]

I did try to use just the infer.py changes (By opening infer.py, and then copy/pasting everything from the new one on github. If there was a better way to do this, that just shows how ignorant in all this I am!)

Annd I get yet another error. It ends with ModuleNotFoundError: No module named 'soxr'
But starts with Traceback (most recent call last):, then lists 12 files. Last file being C:\Applio\rvc\infer\inferp.y

[simple ore]

env\python -m pip install soxr

from applio folder

[dawn copper]

It is one thing after another with me, I'm so sorry!
RuntimeError: expected scalar type Half but found Float

I tried on fp16 and fp32. Same thing either way.

[simple ore]

where?

only means you missed some .float()

hm

or perhaps need to also change lib/predictors

[dawn copper]

  File "C:\Applio\rvc\infer\infer.py", line 602, in convert_audio
    audio_opt = self.vc.pipeline(
-------
  File "C:\Applio\env\lib\site-packages\torch\nn\modules\conv.py", line 306, in _conv_forward
    return F.conv1d(input, weight, bias, self.stride,
RuntimeError: expected scalar type Half but found Float```

I cut out a *bunch* in between, so it'd fit here.

There were like 10 or so other files.

[simple ore]

okay, change rvc/infer/pipeline

and those two lib/predictors

or alternarively, just clone the whole repo

[dawn copper]

... Ok, forgive my noobishness 😛 (Yet again)
I'll open cmd in Applio's folder, and do
git clone https://github.com/IAHispano/Applio/commit/837b945c69d3c057e5de452ef5af7d252ca8293b
Correct? I don't wanna mess things up.

(With that github link, being the one you sent a bit ago)

I have to step away for 30 mins or so. Once again, thank you so much for taking your time to help me!

[graceful latch]

can someone help me make a stupif ai song cover i needed it for a video but i tried half an hour yesterday trying to make one with rvc and it always shows some error

[simple ore]

no, whoe applio

https://github.com/IAHispano/Applio/archive/refs/heads/main.zip

[simple ore]

nobody can help if you just say "always shows some error"

[graceful latch]

wdymm?? i know nothing about this and i cant send image here😭😭 i have no clue

[simple ore]

we do not posess an ouja board that would tell us what error you saw

no telepathy either

[graceful latch]

well uhm sorry for bothering then i actually have no clue how to say the error

[simple ore]

by providing a screenshot?

[graceful latch]

i dont have img perm

if i had i wouldve sent it the first message here💔💔

[simple ore]

post it to imgur.com or wherever

[graceful latch]

okay okay i will thanks 😭😭

[dawn copper]

Surprise, surprise!.. It's me again >.>
I backed up my Appolio folder first. Then downloaded the file you linked here, updated my current install with this new main.zip.
And I'm greeted with a new error, sadly.

  File "C:\Applio\env\lib\site-packages\gradio\queueing.py", line 624, in process_events
    response = await route_utils.call_process_api(
--- (lots more files here, ending with:)
ModuleNotFoundError: No module named 'rvc.lib.algorithm.generators.hifigan_mrf'; 'rvc.lib.algorithm.generators' is not a package```

I went ahead and tried to do `env\python -m pip install`, along with the two names of the things listed. Since this error seemed the same as the `soxr` one from before. But both time I got an error saying `Could not find a version that satisfies the requirement rvc.lib.algorithm.generators.hifigan_mrf (from versions: none)`

I ran the error through ChatGPT (for better or worse?), and it said to put a blank file called __init__.py in rvc/lib/algorithm/generators/ (assuming the hifigan_mrf.py file exists there, which it does)

That got rid of the error, but I'm back at step 1: No audio comes out from anything I do in Applio.

[simple ore]

of course that's not a solution

the file should be there

[dawn copper]

Yup! The file is there. But as soon as I click the Convert button in Applio, it gives me the error
ModuleNotFoundError: No module named 'rvc.lib.algorithm.generators.hifigan_mrf'; 'rvc.lib.algorithm.generators' is not a package

Well, it has the whole

  File "C:\Applio\env\lib\site-packages\gradio\queueing.py", line 624, in process_events
    response = await route_utils.call_process_api(
--- (lots more files here)```
Stuff above that.

[simple ore]

that's on you.. i've just installed that main.zip, and everything works

so perhaps you did not unzip things properly, did not move all files over

[knotty moth]

clone from the repo as fresh installation, not doing in-place update

[dawn copper]

Just to double check, I redownloaded it. Extracted using 7zip. Opened the newly downloaded Applio-main folder, where I then see all the files (Such as run-applio.bat).
I select all files, then go to my main install folder of Applio (C:\Applio), and move all the files there. It asks to overwrite, I click yes. All the files move over. Then I get that error again.

[knotty moth]

also the main repo is the most recent compared to the precompiled one

[dawn copper]

I'll probably have to go through all my previous steps again, to fix all the issues I've been having, from step one. I will try this though.

[simple ore]

well, the only thing you need to do is run the install and then replace torch and patch zluda

[dawn copper]

That's what I tried the first time around, and now we're here! I have no luck with this, it seems. I'll try it all again, but probably tomorrow.
Thank you for all the help though. Sorry if I've been a pain! I've been trying, but am waaay out of my element with all this.

[hallow thistle]

[knotty moth]

try asking the voice owner for permission

[brittle wing]

Hi, just wanted to ask about how people uploading their models have those vocal only files to show them off? Does anyone know what website they use to produce those?

[hallow thistle]

You mean by a voice model, right? Most of model makers here use an RVC program to train voice model.

[brittle wing]

Yeah. For a project im doing I need the vocals standalone so i can edit them properly myself, so I was wondering what program they use

Or programs*

[lunar quartz]

Guys i'm having Trouble making the VC client Work. The only think that works is Beatrice T^T Also can someone send me or atleast message me a Tutorial on Downloading the VC Client Properly? I'm Strugging so hard 😭 Also i have Python and Pytorch and Cuda Installed Just tell me if Python and Cuda and Pytorch ain't needed but Overall the question is I want help installing the VC Client Properly T^T. Also I'm using a PnP Speaker and Mic not a Cable Jack or something

I just want some help T^T

Also Please?

I'm gonna wait.

[hallow thistle]

For W-Okada, go to #🔍│help-w-okada. This channel #✨│ai-help is about RVC programs, not realtime voice changer.

[lunar quartz]

Thanks!

[knotty moth]

it's model files (pth & index) to show, not the audio as dataset

[faint elk]

is there a tutorial how to record something?
LIke a song or something with the realtime voice changer
I tried setting up virtual audio cable but im confused about how to route it and where to

[analog obsidian]

rvc stands for retrieval-based-voice-conversion and not realtime voice changer
rvc is its own thing, w-okada allows the usage of rvc models in realtime

anyways, you have to use rvc to clone pre-recorded audio

https://huggingface.co/IAHispano/Applio/blob/main/Compiled/Windows/ApplioV3.2.8-bugfix.zip

place your model in the logs folder

then for conversion follow the instructions of the gui

[fading sentinel]

whats rvc?

os there a channel for music creation?

[analog obsidian]

ai-based voice cloning software

does not stand for realtime voice changer

but retrieval based voice conversion

[fading sentinel]

i see

so this discord is things like that mostly? was hoping it was big on music

[analog obsidian]

the discord is about doing rvc models (training)
and using them (inference, realtime inference)

[fading sentinel]

ok thank you

[knotty moth]

please go to #🔍│help-w-okada for voice changer and read the pinned guide there

but for recording, it's recommended to use applio rvc on prerecorded audios

[brittle wing]

What is this cell for & how do I use it?

The phase Fixer cell

There are options like "transfer magnitude" "transfer phase"?

[knotty moth]

what r u talking about?

[brittle wing]

On what kind of files

https://colab.research.google.com/drive/1IC6Q1hLF55_tK6mhky0SWYKGVF9T5WsY?usp=drive_link&authuser=1#scrollTo=ScA4L7gmQEjM

[analog obsidian]

phase fixer is for instrumentals

[knotty moth]

it's not related to the rvc I was referring

[brittle wing]

Okay but how do I use that cell

On what kind of files

[knotty moth]

it's just an additional tool, read the notes at the bottom for what the use case

[brittle wing]

Uhm

It's for noise removal okay what is "transfer magnitude" for?

[knotty moth]

go read in wikipedia to understand about magnitude & phase in fast fourier transform

[brittle wing]

Okay

[knotty moth]

well not the right way to ping me cuz I thought you referred that topic with yours that is completely different

[brittle wing]

Hmmm

[faint elk]

where's can I read the guide about applio rvc? I didn't find anything in pinned
https://applio.org/

is this the correct site?

[low shard]

elaborate:

• your pc gpu
• what you want to do

[faint elk]

i use a 3080ti
all i want to do is record the voice changer really into audio files so i can use it to do stuff

though i use prerecorded audio

[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

Wokada = uses RVC for realtime inference

so, you need RVC or wokada?

also your gpu is good

btw, Applio is an RVC Fork (modified version)

[faint elk]

i see
UHm well
I'm not sure what i need?
I just have pre recorded audio clips which i need to turn into a funny voice..

I already have the models
they are rvc models or something

[low shard]

I'm guessing RVC, wokada is a realtime version for VCs basically

what's your OS?

[faint elk]

i use windows

[low shard]

alright you can just follow https://docs.aihub.gg/ which gives RVC info, like using and downloading it

this is our documentation

applio also got their own documentation https://docs.applio.org/applio

[faint elk]

Thanks so much!

[low shard]

lmk for any issues

also, our docs ofc have applio info

[hallow thistle]

RVC and W-Okada are two different programs. In this context, RVC isn't an abbreviation of realtime voice changer.

[simple willow]

where do i find the pitch

[pastel oak]

Transpose?

[simple willow]

maybe?

there is this extended pitch slider

so I assume there is an option for this like volume

[pastel oak]

what program

[simple willow]

it's general settings in voice ai

[pastel oak]

We dont offer support for voice ai

paywalled garb

[simple willow]

oh XD

HAHA

https://tenor.com/view/funny-hahaha-gif-22505124

[pastel oak]

We can suggest u open source options

whats ur gpu

[simple willow]

rtx 2060s

[pastel oak]

https://rentry.co/ForkVoiceChangerGuide#download-nvidia

any other questions in #🔍│help-w-okada

[simple willow]

thanks man big help

[grim glen]

how do i get to the google collab for appolio

[brittle wing]

Okay, thanks for clarifying

[low shard]

first of all, tell:

• your pc gpu
• what you want to do

[hard shadow]

I'm getting an error on Applio that says "No module named 'tqdm.auto'". Any help, please?

[low shard]

elaborate:

• your pc gpu
• what u want to do
• what guide link are u following

[cedar canopy]

which fork/guide am i supposed to install/follow to run local inference on apple silicon?

[simple ore]

latest applio would use CPU for inference

[cedar canopy]

ah that sucks

[simple ore]

MPS torch is not there yet

[cedar canopy]

i was under the impression that torch added metal support

thank you!!

[simple ore]

https://docs.google.com/document/d/10ZyhkaJGdoa_M7QVUmlNSe94brAYG1BusKS4EHTrtwc/edit?pli=1&tab=t.0

it only works in a single thread last time we checked

[cedar canopy]

for how much they push the AI capabilities of their arm chips you really cannot run much on them besides LLMs

at least UVR is fast on it

[simple ore]

opm_num_threads = 1, otherwise it just hangs after 1st inference

apple has its own AI model, it is tricky to convert RVC to it

Core ML

[brittle wing]

Anyone have any idea why AttributeError: 'NoneType' object has no attribute 'pipeline'
Is showing up every time I try to synthesize on Applio?

[low shard]

huh wasn't it there but then removed bc of macs overheating?

[low shard]

elarboate:

• your pc gpu
• what guide link are u using
• what u want to do
• a screenshot ofthe error

[simple ore]

you get that error when you did not select a voice model

[simple ore]

it just buggy, hangs up after 1st inference

cpu works fine

since I dont have mac I cant do anything about it

[cedar canopy]

im guessing any overheating reports would come from macbook air owners since those are fanless 🤔

[low shard]

and that guide fixes that issue?

I checked it fast and it seems enabling mps

[simple ore]

I think it was made before we nuked it 🙂

so 3.2.7 should work?

technically it can be enabled back

rvc/configs/config.py

but I dont recommend

[simple willow]

how do i increase the volume, I am using this atm

volume increase nowhere to be found

[low shard]

wrong channel again

use #🔍│help-w-okada

[low shard]

[simple willow]

oh

[broken vortex]

can someone tell me why the uploadbutton while uploading models doest work

@low shard

[bleak nymph]

i'm kinda curious to see what it'd sound like if i made an ai of me talking and singing for 3 hrs 😭

it's 3 hr long audio file of me

insane

probably would be better than my current one

[knotty moth]

it loads the entire audio to memory like what UVR does

unless the audio is split at most 30 mins each

depending on ram capacity

[bleak nymph]

ohhh interesting

i mean i have 32 gigs of ram 😭

quite a lot

[tight linden]

is applio good for mac if I want to use ai voice model to say a specific word?

[knotty moth]

it may run in cpu mode, or in mps (rather experimental, only for apple silicon M series)

[tight linden]

Ok.

[hallow thistle]

Intel or Apple Silicon Mac? Applio will sure work with only CPU.

[knotty moth]

it's still plenty enough for most tasks but splitting is always a good thing to do for streams/podcasts. I usually did it around 5-15 minutes each or per session.

[bleak turtle]

where do you download the voice changer (not models)

[knotty moth]

in #🔍│help-w-okada

[bleak turtle]

thanks

[broken tulip]

Not sure if this is the right place and I don't have the software... but what software do I need to mess around with songs so that I can convert one voice to another?

[low shard]

are u still having issues? elaborate the program and the issue deeply

[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

Wokada = uses RVC for realtime inference

I'm guessing you want RVC

[broken tulip]

Ahh... thank you.

And is there a specfic website to grab this software?

[low shard]

what's your pc gpu?

[broken tulip]

Nvida, I think.

[low shard]

that's just the brand, nvidia made a lot of gpus

[broken tulip]

Oh...

[low shard]

some nvidia gpus could be too old, some others are good, some others are faster

I can't know if your gpu is good enough by just the brand

[broken tulip]

GeForce RTX 3060 Ti

[low shard]

there's no such thing as the rtx 1060 ti, the rtx come out after the 20 series, before that, it was gtx

ohh 3060

nice

[broken tulip]

I made a typo!

[low shard]

yeah dw it's fine lol

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

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

I would suggest you to use applio

[broken tulip]

Thank you.

[low shard]

yw and lmk for issues

[broken tulip]

I am blind as a bat. So I'm on the huggingface.co site, and I don't see the download button on the right hand side. Do I need to sign up?

[low shard]

nope you don't need to sign up, open the link on the guide, go to windows (guessing you're on windows), select the only .zip file, then do as said in the pic

[broken tulip]

Doesn't look like that on my end, but I'm gonna guess it's the ApplioV3.2.8-bugfix.zip

[low shard]

yep, u gotta click that

[broken tulip]

Thank you. I got it now. 🙂

[broken tulip]

I found a setting I like, but how do I make it lose the autotune in the voice on some of the higher notes?

[low shard]

it might depend on how the model was trained

[broken tulip]

Oh.

[low shard]

you can either try to lower the index (which contains the accent), or use another one

[analog obsidian]

looks like its a model problem

or using crepe with a high hop length can cause that too

[broken tulip]

Or I suck at this so far.

XD

[knotty moth]

autotuned voices could feel less natural

[broken tulip]

fair...

[tight crane]

does anyone know how many epochs i should set for like 5mins of audio?

[broken vortex]

no its fine

[dusty rune]

im just tryna get epoches for a 25 second dataset

the guy im doing barely talks

[livid cosmos]

So I am using RVC v2 Disconnected, I also have done Feature Extraction and got this error? What did I done wrong?

/content/Mangio-RVC-Fork/logs/Kilixa_v1/3_feature768

Exception Traceback (most recent call last)
<ipython-input-38-86df4775937e> in <cell line: 0>()
29 listdir_res = list(os.listdir(feature_dir))
30 if len(listdir_res) == 0:
---> 31 raise Exception("No features exist for this model yet. Did you run Feature Extraction?")
32
33 try:

Exception: No features exist for this model yet. Did you run Feature Extraction?

[jovial leaf]

For RVC fandubbing projects

[livid cosmos]

Problem solved

[low shard]

RVC v2 is outdated and will NEVER come back, do NOT use it

you shouldnt use it at all

it's based on mangio fork, an rvc fork abandoned since 2023

its slower than mainline and other rvc forks

and is more unstable

[low shard]

@livid cosmos #📰│dev-updates message , do not use rvc disconnected

whats your pc gpu and what do you want to do

[livid cosmos]

Gigabyte GeForce RTX 3060 Gaming 12GB, and I want to train a voice model for myself

[hallow thistle]

It's crazy to think that people using RVC v2 Disconnected in 2025, even if there are the Mainline and Applio available.

[hallow thistle]

That GPU should good enough for voice model training, with that VRAM.

Just give the locally program a try.

[livid cosmos]

Which is the best way to do it locally?

[low shard]

you don't even need to use google colab, google colab is a cloud (remote good pc service) for people with a bad pc

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

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

use Applio

locally

[hallow thistle]

Applio is easier to use. It has web UI.

[livid cosmos]

I got used to it before I didn't have a good GPU/PC xd

[hallow thistle]

[livid cosmos]

But yeah, I will try with Applio then

[low shard]

you gotta always adapt to updates 🔥

[tame mica]

does these look fine

ive ignored any tensorboard talk since it makes my head burst so idfk atp

[lilac gust]

@austere nexus ma'am what the hell r u doing here

[austere nexus]

LMAO WHY CAN'T I

[lilac gust]

Why did you tag my msg that's like 16 years ago and LMAO at me 😭😭💀

[austere nexus]

it's the most recent one you sent 😭

[simple ore]

for only 2k steps it is fine

[tame mica]

she sounds like you put the saturator on max in fl but whatever

gonna let it cook fo a bit

[compact saffron]

yo how can I change the effect of a voice model? i.e. i wanna sound to 25% like juice, but still recognize my voice, dm me if you can help

[nocturne mica]

Can I find here some modules with russian accent? Because using an English one cause some issues

[low shard]

you can't like partially apply the voice, but you can adjust the index rate

also if you're talking about realtime voice changer, that's wokada, use #🔍│help-w-okada

be sure to not use ty tuts

[low shard]

u mean models?

You can search rvc ai voice models at:

• #1175430844685484042
• In #🔍│find-models , Do /find with @earnest musk
• https://weights.com/ (login required)
• https://huggingface.co/models (but watch out cus in hugging face there arent only rvc ai voice models)
• https://voice-models.com/
• https://thevoicemodels.com/ (for Turkish Models, login required with discord and level 2 on their server)

if there isnt one, you can:

• #1159289738314919936
• #1191429836321849435
• make it yourself with our docs guides https://docs.aihub.gg/essentials/how-to-make-voice-models/

[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

[eager rivet]

yo

what software should i use to convert vocals

i have a rtx 4060

[low shard]

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

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

I gave you written guides, read them and lmk

[eager rivet]

alright what do i do when im n this screen

[zinc herald]

who is good with making ai songs (like fake songs)

i need help

[eager rivet]

maybe help instead of being a dickhead

[low shard]

which one are u using

[eager rivet]

RVCNVIDIA1600

idk i downloaded it a long while ago

[low shard]

tbh I would suggest you to use applio since its more updated

there is a written guide

[eager rivet]

i have applio

had it for a while

[low shard]

be sure u got the latest version

[eager rivet]

for some reason whenever i do the run applio bat thing nothing opens in my browser

[low shard]

delete the one you got, read the guide I sent, and download it from there

[eager rivet]

alright Ill get back to you shortly

this the download right @low shard

[low shard]

yes

[candid meteor]

can anyone help me? i download a model in Applio the model does not show up

nevermind

[brittle wing]

-colab

[karmic olive]

📒 Google Colab Notebooks

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

ℹ️ Note

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

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

[eager rivet]

@low shard

i opened the runapplio bat file

and nothing happens

[opal quarry]

Hola alguien tiene link de rvc?

[low shard]

This server is English only

Also elaborate ur PC GPU and what u want to do

[low shard]

U sure u extracted it before

[eager rivet]

yea

also in my c drive

[low shard]

Go on the top bar, type CMD and enter, then type run-applio.bat and show the output

[eager rivet]

alright it worked

[eager rivet]

whenever i upload my index i get this error

[low shard]

Did u upload the pth first, wait for it to upload, then the index?

Also share which index file download link

[unique rock]

Can someone help me with this?

---

FileExistsError Traceback (most recent call last)
<ipython-input-12-e49906b94ca3> in <cell line: 0>()
17 if os.path.exists(time_):
18 shutil.copy(time_, time__)
---> 19 shutil.copytree(source_path, destination_path)
20 print("Model backup loaded successfully.")

2 frames
/usr/lib/python3.11/os.py in makedirs(name, mode, exist_ok)

FileExistsError: [Errno 17] File exists: '/content/program_ml/logs/Ahyeon

[eager rivet]

dms

[low shard]

There's no need, u can share the download link here

[low shard]

Elaborate:

• ur PC GPU
• what u want to do
• what are u using
• what did u do

[unique rock]

Applio colab

[low shard]

You need to elaborate everything I said

Not just one single part

Also there isn't only 1 applio colab, there's the UI version and the no UI one

[eternal oasis]

Hello, its possible to make a release pre-compiled of Applio included the latest torchaudio for RTX 5000 users ?

[low shard]

How to (unofficially) use Applio for RTX 50 serie cards

Follow to download it as said it in https://docs.aihub.gg/rvc/local/applio/

After you extracted the precompiled, go to the path in Windows explorer, write "CMD" and press enter, then in CMD write env\python -m pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

If you get any already satisfied requirement issue, run env\python -m pip uninstall torch torchvision torchaudio then the command said above

[unique rock]

I use Applio UI Collab and when I run the Load Auto Backup cell I got this:

[eternal oasis]

Ok thank you Nick

[low shard]

Yw and lmk

[low shard]

Also, what's your pc GPU? And tell what you did step by step, you might have skipped a step

[unique rock]

I installed the latest version of Applio in queue, then mounted the Drive, then connected the backup, loaded the model and got that error, and the files and name were correct (for retraining) and finally downloaded the custom pretrain to run Applio finally but I can't do it because of that error

[eager rivet]

https://drive.google.com/drive/folders/1G2Xxd_80u1MFRpd95be68UwQt6ZOlEbL?usp=sharing

[knotty moth]

he was asking for your gpu, if you have at least 3060 or 8 gb card, it's recommended to train locally

[unique rock]

I use mobile haha it worked fine until now

[simple ore]

just put pth and index into /logs/mymodel

[hallow thistle]

That still doesn't answer what he asked for, but close enough to where you don't even have a PC.

[unique rock]

I've been training on my cell phone for two months. I always made my models on my phone. It never gave me any mistakes, only on that model.

[hallow thistle]

[knotty moth]

then explain the issue and error message

[indigo wolf]

where do i change voice for singer?

[tame mica]

it sounds so fucked at the higher pitches but who gaf

[grim bay]

How can I fix this?

[simple ore]

use an up to date colab

[low shard]

@viscid moss check this out

[viscid moss]

Hmmmm

Weird

-colab

[patent trellis]

📒 Google Colab Notebooks

Google Colab is a Cloud (Remote Good PC) Service. While the Free plan provides up to 12 hours of daily usage, the GPU is typically available for only about 4 hours each day on average.

• **Applio**

by IA Hispano
Google Colab

• **RVC Mainline**

by Hina
Google Colab

• **UVR5 NO UI**

by Eddy
Google Colab

• **UVR5 UI**

by Eddy
Google Colab

• **Wokada Deiteris Fork**

by Deiteris & Hina
Google Colab

• **Hina's Modified Original Wokada**

Google Colab

• **RVC-AI-Cover-Maker-WebUI

by Shiro & Eddy
Google Colab

• **FaceFusion UI**

by Nick088
Google Colab

• **FaceFusion NO UI**

by Nick088
Google Colab

• **Music Source Separation Training (Inference)**

by Jarredou & Makidanye
Google Colab

[karmic olive]

📒 Google Colab Notebooks

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

ℹ️ Note

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

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

[viscid moss]

Fixed

[low shard]

🔥

Lemme guess , uv?

[viscid moss]

Nah, some dependency was broken so i remove it.

[low shard]

oh lol

[hot lagoon]

RVC-AI-Cover-Maker-UI

kaggle

only happen when trying to switch the Dereverb Model

Same with switching the Deeecho one

[low shard]

@viscid moss

[verbal oasis]

Hi, does anyone have a link to EasyGUI google colab with the UI interface?

[low shard]

EasyGUI is outdated since months, don’t use it

the creator is busy irl

what’s your pc gpu and what do you want to do?

[verbal oasis]

uh really? what can I use?
I've used that a while back to train 3 voices from audio and it was pretty easy to use

[verbal oasis]

I want to run on colabs not local possibly

[low shard]

ai progresses fast, things change alot, never stuck with a program for many many years

[low shard]

I’m asking because local is actually better than cloud (remote good pc), you won’t hage a random limit of max 4 hours daily of gpu

however that depends on how good is your gpu

i mean if you really don’t wanna tell it’s fine, i’m just trying to help out

[verbal oasis]

yeah, better on cloud if there's any option

[low shard]

btw i would still need to know what you need to do, so i know what colab to give you

there’s colab only for inference, others only for training, others that are more automated for ai covers, others for both training and inference yk

[verbal oasis]

ah yeah, I want to train a voice model from an audio recording so that it gives me back a pth and a index file

in order to use the model in applio

I need only training

[low shard]

Train (make) RVC Models on cloud:

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

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

3. Be sure to know about the tensorboard

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

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

You can use either:

• Weights.com: Easiest Possible Ever Automatic
• Ilaria RVC Zero: Fastest free on cloud
• Applio UI Colab: RVC Fork with some extra features like TTS
• RVC AI Cover Maker UI: Automatically Separates the Vocals and Instrumentals, converts the voice and mixes them back

[knotty moth]

you should try another working colab, i.e. applio
https://colab.research.google.com/github/iahispano/applio/blob/master/assets/Applio.ipynb

[low shard]

i also gave kaggles and lightning ai links

which are other cloud platforms

[verbal oasis]

can I train in applio too?

[low shard]

for example kaggle gives more time and better gpu than colab but its harder to setup

[low shard]

yeah, tbh i would suggest the kaggle version of applio for training

[verbal oasis]

kaggle is an alternative of colabs right?

[low shard]

Colab: random daily gpu of max 4 hours
Kaggle: 30 hours of free gpu weekly granted, with better gpus, but needs phone number and is harder

Both are owned by google

[low shard]

yes, they are both cloud (remote good pc)

[verbal oasis]

thanks, I'll try for now to train on applio on colab

have you ever tried to do so?

[knotty moth]

much better

I'm on the kaggle one

[verbal oasis]

is it complex to train from audio?

[low shard]

i mean u would just need to follow the guide

[verbal oasis]

sure I will

[verbal oasis]

From your experience any recommendation for settings to get a better result?

[simple ore]

the secret of good result is trying a lot and failing a lot until you figure what works best for your specific dataset

[viscid moss]

Hmmmm I'll try to fix it

[finite pumice]

guys

how can i use the huggingface models in rvc

can someone give like a tutorial please

@long forge

@acoustic scarab

on pinikio

[viscid moss]

I'm pretty sure that u are using an outdated version or u are using the outdated link from Shiromiya

Here's the new link:
https://www.kaggle.com/code/eddycrack864/rvc-ai-cover-maker-ui

-kaggle

[karmic olive]

📘 Kaggle Notebooks

• Applio Notebook, by Vidal Kaggle
• Applio Notebook, by Shirou Kaggle
• Music Source Separation, by Shirou Kaggle
• UVR5 NO UI, by Eddy Kaggle
• Original W-Okada's Voice Changer, Kaggle
• Modified W-Okada's Voice Changer, Kaggle
• 🆕 UVR5 UI, by Eddy, ArisDev & Nick088 Kaggle
• 🆕 RVC AI Cover Maker UI, by Shirou & ArisDev Kaggle
• 📖 How to use RVC Mainline on Kaggle by Cauthess

Note: Kaggle limits GPU usage to 30 hours per week.

More commands: -colabs, -hf, -local, -rvc, /help

[patent trellis]

📘 Kaggle Notebooks

Kaggle is a Cloud (Remote Good PC) Service that offers 30 hours of GPU weekly, but needs a phone number verification

• **Applio Notebook**

by Vidal Kaggle

• **Hina Mod Original Wokada**

Kaggle

• **Wokada Deiteris Fork**

by Hina & Deiteris
Kaggle

• **UVR5 UI**

by Eddy, ArisDev & Nick088 Kaggle

• **UVR5 NO UI**

by Eddy Kaggle

• **RVC AI Cover Maker UI**

by Shirou & ArisDev Kaggle

• **Music Source Separation**

by Shirou Kaggle

[viscid moss]

@low shard Kaggle links needs to be updated

[low shard]

the sapphire ones are the updated ones, I'm working on it

oh right u took ownership of rvc ai cover maker

gonna fix it soon

[brittle wing]

What are some good websites to find rvc voice models?

[low shard]

You can search rvc ai voice models at:

• #1175430844685484042
• In #🔍│find-models , Do /find with @earnest musk
• https://weights.com/ (login required)
• https://huggingface.co/models (but watch out cus in hugging face there arent only rvc ai voice models)
• https://voice-models.com/
• https://thevoicemodels.com/ (for Turkish Models, login required with discord and level 2 on their server)

if there isnt one, you can:

• #1159289738314919936
• #1191429836321849435
• make it yourself with our docs guides https://docs.aihub.gg/essentials/how-to-make-voice-models/

[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

[near axle]

Can anyone make a new RVC V2 api on replicate?

[low shard]

I don't think anyone is going to do that, we mostly use free stuff, like colab, kaggle, lightning.ai and ofc hosting things locally

[near axle]

Is there a video tutorial on lightling.ai about RVC V2?

[low shard]

youtube tutorials are old, don't use them at all

the only updated guides are written ones

by the way, what's your pc gpu and what do you want to do?

[near axle]

I have a intel gpu nothing else

[low shard]

oh, I'm guessing an integrated intel gpu in your laptop?

[near axle]

Yea

[low shard]

well, yeah you could run it locally but would be slow as hell, cloud (remote good pc) is way better in your case

just a reminder:
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

Wokada = uses RVC for realtime inference

I can give you cloud links, but I need to know what program you need to give you them

[hot lagoon]

I use this one of kaggle and it stile dose the same thing

[viscid moss]

Hmmmm

I tested with normal deecho and bs roformer dereverb, lemme try that ones

cause that aggressive is too aggressive and bs roformer de-reverb it's better than MDX23C

[hot lagoon]

it only happens when I try to switch it to someing like bs roformer de-reverb or change the Deeecho Model to Aggressive

The normal settings works fine. It’s just that somehow switching it causes the error.

[viscid moss]

wait, you have the problem even with the BS Roformer Dereverb?

[hot lagoon]

yes change to BS Roformer Dereverb? dose it

and change Deeecho Model to Aggressive

https://tenor.com/view/sad-mario-walking-in-rain-mario-sad-mario-walking-gif-7966847823215482947

[viscid moss]

ok, that's weird cause i change it to bs roformer de reverb and it's working

[hot lagoon]

🧐

[viscid moss]

lemme try aggressive

[hot lagoon]

weird

[viscid moss]

got an error but not the same error as u

mdx23c dereverb seems to not being working

ok found, if u choose MDX23C dereverb it gets bugged so if u cange to roformer again it will no work

Yep, MDX23C dereverb and Deecho Agg isn't working. The rest of the models are working

[hot lagoon]

Yeah I rest and chose just roformer for dereveb and its working

Just Deecho agg and the other Derever is not working

https://tenor.com/view/neco-arc-gif-25900194

[viscid moss]

Ye, for now use it like that. I'll improve it when the CoverMaker rework is ready

[spiral osprey]

@low shard Hey Nick,could you look at the discord private dm please?

[analog verge]

?colab

-colabs

[karmic olive]

📒 Google Colab Notebooks

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

ℹ️ Note

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

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

[knotty moth]

the newer anvuew mel is better

though it doesnt remove delay echoes

[viscid moss]

ye but haven't added that model yet

[knotty moth]

https://tenor.com/view/just-do-it-shia-la-beouf-do-it-gif-4531935

[viscid moss]

It will come with the rework of CoverMaker

[vestal helm]

what pretrain should i use for weeknd

dataset is

very clean

[muted portal]

hello beautifull people, if there is anyone out there that knows how to create ai voice models for rvc and has a bit of time to show me how, a dm or mention would be very greatly apreciated <3

[pastel oak]

https://docs.aihub.gg/essentials/how-to-make-voice-models/

[atomic sparrow]

if you training the same model in diferent days , you need to execute the "EXTRACT FEATURES " section in the 2nd or 3rd time? in google colab

[rich moss]

Hey everyone, I’m running into an issue with TensorBoard. Here’s the error I’m getting:

tensorboard_venv already exists, skipping creation and activation...
Launching TensorBoard...
'tensorboard' n’est pas reconnu en tant que commande interne
ou externe, un programme exécutable ou un fichier de commandes.
Keeping the command prompt open...
Appuyez sur une touche pour continuer...

[low shard]

Alr

[low shard]

What google colab link? What's your PC GPU?

I really hope your not using rvc disconnected or other old yt tuts

[hallow thistle]

What RVC program are you using? Can you give the name of it?

[rich moss]

RVC1006Nvidia https://docs.aihub.gg/rvc/local/mainline/

[hallow thistle]

And what is your PC GPU?

Because the latest update for this RVC program is from 2023. There's a better one available.

[rich moss]

rtx 3060 12gb

[rich moss]

where pls

[hallow thistle]

There's Applio. The UI overall looks similar to "mainline" RVC, but it's easier to use, and it has TTS built in it. https://docs.applio.org/applio

[rich moss]

ty

and btw how many minutes of audio data do you recommend I collect? Also, where can I find good quality datasets pls?

[rich moss]

?

[hallow thistle]

30 - 60 minutes of combined dataset audio is enough, if you mean by to train a voice model. I'm not sure where the good source to find audio dataset is, it can be from some YouTube videos containing spoken audio.

[rich moss]

bcs I want to make a model to use with Okada, so it needs to be the highest quality possible. Do you think 60 minutes is enough?

[hallow thistle]

Yes, but the audio quality is also what makes a voice model to sound good.

While I can only give some basic ideas on how to train a voice model, some more advanced things go for some other helpers instead. If you have any problem training a voice model, you can go to #📑│making-models.

[knotty moth]

but not from different sources

being shorter but stick on one of them is better

[knotty moth]

compromise the quantity before consistency

5 minutes of dataset from a single fine quality source is better than longer one using multiple sources or worse consistency

[median island]

I've tried a couple times to make a cover, but I keep getting a runtime error saying I don't have enough memory when I know I do. Is there something I'm doing wrong?

[reef sonnet]

[Voice Changer] MMVC_SocketIOApp initializing... done. how do i fix this

[median island]

I don't know why it keeps happening

[simple ore]

you may want to start with explaining what's your GPU, what error message you're getting

[median island]

I don't have a GPU, it reverts to CPU and the message is
RuntimeError: [enforce fail at C:\actions-runner_work\pytorch\pytorch\builder\windows\pytorch\c10\core\impl\alloc_cpu.cpp:72] data. DefaultCPUAllocator: not enough memory: you tried to allocate 379509352 bytes.

[knotty moth]

which application are u using?

[knotty moth]

ye you should use cpu mode

[median island]

I am using CPU

[median island]

RVC Gui, or do you mean something different?

[knotty moth]

it's too old