✨│ai-help

[coral viper]

Wdym "for"?

For get rid of it or for left it?

[broken urchin]

it says no such file or directory

in the cmd

[glacial pollen]

Either getting rid of it or extracting just that

[glacial pollen]

oh, ye, do ffmpeg.exe

my bad

[coral viper]

What

[glacial pollen]

Either getting rid of it ( backing vocals ) and leaving just main / lead vocals or extracting just that ( Only backing vocals )

[glacial pollen]

well

send me the file

can be dm

There is something wrong with the file ( or you didn't put it in there?, next to ffmpeg

[coral viper]

How can it be a gambling?

[glacial pollen]

man

see this?

depending on the model, you always have boxes to check

instru, vocals, other, etc etc

[broken urchin]

i think its something wrong with the files

[coral viper]

Alright, then

[glacial pollen]

Now, depending on what the model does

[broken urchin]

maybe i should use something different to record voice messages

[glacial pollen]

the 2nd option, in this case instrumental only

in case of, say, bve model

it doesn't output instrumental but just the backing vocals

[knotty cobalt]

Do you guys know where I could find a large number of CC0 free-use voice speaking samples with high quality recordingings? I tried Mozilla's Common Voice but each speaker only has like 5-ish minutes and the recording quality sucks.

[glacial pollen]

Ye, send it over dm
I'll analyze the file if you want and tell you how to fix it

Alternatively, just change the recorder

[broken urchin]

what recorder should i use

[glacial pollen]

Idk man

just use wo-mic

use ur phone as a mic

https://wolicheng.com/womic/

[broken urchin]

what??

i have a microphone

[glacial pollen]

oh, wait, by voice messages you don't mean phone-related?

[broken urchin]

no

on pc

voice recording on pc

[glacial pollen]

then what is it, you record discord or such?

[broken urchin]

no i just use the voice recorder app

[glacial pollen]

anyway, obs or audacity ig?
Whatever you use, fucks the files up

[broken urchin]

i use this

[glacial pollen]

stock-apps ah moment

Use audacity

https://www.audacityteam.org

[broken urchin]

ok i will thanks

[glacial pollen]

Either way, rest is up to you to figure out

recorder fucks up the files, that's that

I gotta go for now so, gluck

[broken urchin]

alright

[knotty moth]

any audio editors/DAW are better than the built in windows, and it should allow saving in lossless wav format

[coral viper]

@glacial pollenWhy

[glacial pollen]

Makindanyee hopefully explains you what's what n why - I am busy atm

[knotty moth]

• try converting the input file to wav first if not sure
• if you have only igpu, I'd recommend to uncheck "GPU conversion"

[rugged hinge]

HomiethePossum told me to ask you whats wrong with 20000 epochs on a model when it's says that amount on config.json. I don't see anything wrong with it. If it said it on config.json then it should be correct

[novel arch]

I mean it’s correct, it’s just not ideal. You always want to use tensorboard to look out for overtraining on your models. Depending on your dataset length overtraining can happen quicker or slower, but you need to watch out for it. If you need a tutorial for Applio or Tensorboard, someone else will be able to help.

[rugged hinge]

How do I put a model on tensorboard

[novel arch]

@formal wind will help he just msged me

[formal wind]

Wait huh.

[novel arch]

yeah you said you’d help him

[rugged hinge]

Ok. I'll wait for an answer

[formal wind]

Haz has free time he should be able to help

[rugged hinge]

OK. I'll wait for an answer from him

[knotty moth]

srsly is it 20k epochs or steps?

[novel arch]

epoches

[formal wind]

https://discord.com/channels/1159260121998827560/1348468990162239641

[rugged hinge]

Look

[formal wind]

Thatll answer it

[novel arch]

I have no time my friend

[rugged hinge]

Then who does?

[knotty moth]

tensorboard pls

[rugged hinge]

Then who does

[novel arch]

@viral mason they have time

[formal wind]

YEP

He has time lol

[rugged hinge]

I never used it before. I'm new to voice model uploading and making so I need help

[knotty moth]

and the log interval seems cursed

[novel arch]

I mean you could just look at the tutorials?

[rugged hinge]

I'll still wait for an answer

[viral mason]

wha

[rugged hinge]

Where do I find tensorboard tutorials

[novel arch]

@viral mason help a brother out

[viral mason]

sure

[novel arch]

good luck

[rugged hinge]

I'm still waiting for an answer

[novel arch]

worm give him an answer

[rugged hinge]

You heard him

[viral mason]

uhhhhh

[rugged hinge]

Wayta?

[viral mason]

@red kayak

[novel arch]

don’t uhhhh worm you know exactly where they are

[viral mason]

idk any tutorials

[novel arch]

you are the tutorial

[formal wind]

Help him out worm 🤣

[novel arch]

please worm

[rugged hinge]

You just don't want to say

[novel arch]

EXACTLY right

he’s gatekeepking them from you

[viral mason]

???

[rugged hinge]

You heard the man

[viral mason]

I'm genuienly confused

[novel arch]

worm just give him the tutorials

[rugged hinge]

You heard the man

[rugged hinge]

Oh please

[novel arch]

@viral mason I mean come on you can’t have all the tutorials to yourself? help a brother out.

[rugged hinge]

Great. Now my 3 models were removed for no reason the 3rd time. I didn't do anything wrong. I don't know what's going on

[novel arch]

damn

[viral mason]

do u clean your datasets?

[rugged hinge]

I don't know who keeps doing it. But I know it's not me

[rugged hinge]

Seriously, I didn't nothing wrong

[rugged hinge]

I don't know who keeps doing it. But I know it's not me

[rugged hinge]

Also I reported and blocked you Your_Local_Worm. Because you offended me with that Bonnie gif

[viral mason]

what???

[viral mason]

L

[novel arch]

that’s fair enough

[rugged hinge]

I'm telling the truth. I did nothing wrong

[formal wind]

LOL Worm

I dunno what you did to him but he madd

[rugged hinge]

I had to do that. Cause that was unnecessary

[viral mason]

sent that bonnie gif

[viral mason]

https://tenor.com/view/sisyphus-rock-mp4-transparent-color-gif-11114772717737196968

[rugged hinge]

I'm gonna upload them one last time for tonight. But if it happens again, I'll post them on Uberduck model showcase instead

[novel arch]

worm don’t anger him more

[rugged hinge]

Yes. And that was so unnecessary

[rugged hinge]

You heard the man

Worm. I don't appreciate your accusing about me doing something wrong which I didn't

[novel arch]

https://tenor.com/view/markiplier-bald-bald-markiplier-funny-lightning-gif-17736098350794360848

[rugged hinge]

Great now I got banned for nothing. I didn't put any instrumentals on the demos. I wasn't against the guidelines at all. I did nothing wrong. There wasn't a single answer. So banning me was a huge mistake

[novel arch]

I think it was the 20,000 epochs

[formal wind]

You posted a low quality model

They take them down

[rugged hinge]

I don't see that saying anywhere on the guidelines

[crude flame]

you didnt get banned you had a role removed

it was removed for repeated low quality models

[rugged hinge]

I don't see that saying anywhere on the guidelines

[novel arch]

I mean when you think about it, it’s kinda pointless posting a model that’s low quality because then nobody will use it?

[rugged hinge]

I'll try again and I'll try harder

[crude flame]

Just because you passed the qc test doesnt mean you are allowed to post bad quality models

[novel arch]

well you have to redeem yourself to even get your role back before you can post

[rugged hinge]

Is there a way I can get the role back

[novel arch]

make a good model, become model maker.

[crude flame]

apply for model maker again

then dont post bad models

so it doesnt get removed again

[viral mason]

I didn't lol

[rugged hinge]

How can I make them good?

[knotty moth]

well I think u shouldnt post meme models using few second dataset, or try repeating it to at least around 2 mins, even that wouldnt be ideal enough

[crude flame]

clean them and train them properly

[formal wind]

Oh razer while you're here. I wanted to bring up my application for model master. It's been about 3 days and have gotten a response yet. Is that normal

[rugged hinge]

Where's a tutorial for that?

[steel forge]

yes

[rugged hinge]

Will UVR Denoise on X minus work?

[crude flame]

yea, our docs

https://docs.aihub.gg

[crude flame]

i dont do those

[rugged hinge]

Will UVR Denoise on X minus work?

[crude flame]

i think litsa does them

[crude flame]

depends

[formal wind]

Bet thx

[rugged hinge]

I'l try harder

[novel arch]

that’s the spirit

[knotty moth]

and smarter

[rugged hinge]

Where do I find the channel for apply again?

[crude flame]

https://discord.com/channels/1159260121998827560/1305527335646269440

[viral mason]

bro is not making it

[novel arch]

don’t anger him again

[viral mason]

[spring gate]

That wasn't nice for you to say to him. Just for that I'll report and block you as well

[formal wind]

I cant tell if youre jokin

[spring gate]

I really am gonna do it cause that wasn't nice

[steel forge]

no jokes allowed on ai hub

[spring gate]

I'm not joking. I'm serious. I'm gonna do it

There. Reporting and blocking done

[formal wind]

He wasn't even being mean that guy was literally just being not to be rude here. A little dumb

[spring gate]

Either way, That was unnecessary

[novel arch]

it’s not that deep

[spring gate]

I'm sure that guy will get better at model making. You just gotta make practice and all you have to do is keep trying. Practice makes perfect

[novel arch]

then you msg him and teach him?

[spring gate]

You hear that, @rugged hinge?

[spring gate]

I can try

[novel arch]

go for it

[rugged hinge]

While i try to figure myself out, I'm gonna post my models on the Uberduck model showcase instead in the meantime

[spring gate]

I already did it and gave him some advice

[novel arch]

good on you my friend

[spring gate]

Thanks

[spring gate]

Good idea

[viral mason]

👍

[novel arch]

https://tenor.com/view/yes-yas-gif-19093781

[viral mason]

https://tenor.com/view/we-are-so-cooked-gif-11199807318849761981

[glacial pollen]

wtf happened here
y'all gotta chill

[analog obsidian]

🦈

[hallow thistle]

Imagine blaming model checkers for rejecting your voice model lol.

!howtoask

[patent trellis]

How To Troubleshoot

__**GIVE CONTEXT.**__ 📝

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

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

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

__**BE PRODUCTIVE.**__ 🤝

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

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

[hallow thistle]

I can give you at least some basic advices on how to make a voice model to sound good: sample rate, audio dataset quality for example. But I don't have a permission to accept a voice model to pass to #1175430844685484042. Some advanced things go to engineers and model checkers. If you think you know things better than anyone else here, you can try think for yourself what you gonna do with your way.

[tame mica]

wtf happened im couhging the fuck up

[knotty moth]

someone yapped about a person with some problematic model

[tame mica]

😭

[low shard]

basically it can help speeding up training by using the VRAM, but will cause a big vram usage

I'm not even sure if it would work with AMD, but 8gb or more vram is probably suggested

I didn't try that option honestly

not sure if it gives any difference in quality since I never tried that option

[analog obsidian]

only for small datasets of 10 minutes
also if you use more vram than your gpu has you'll be using system ram (which is painfully slower)

[analog obsidian]

nop

4060

[hallow thistle]

Intel HD Graphics 3000

Yes, that's my laptop only GPU.

[analog obsidian]

dont use overtraining detector, its innacurate and bad
ive asked noobies to remove it but he told me that 'you can just ignore it'

yesss

first one

u want to undertrain your model?

🦈

[low shard]

I never tried it, but can't it be adjusted ?

[analog obsidian]

just pay someone to do the model for you i guess

[analog obsidian]

nop

[analog obsidian]

i mean yes but still the thing is bad lol

idk better ask him 🦈

i would do it for free bc u have the dataset
tho im not going to clean it and you cannot submit that model for your model maker role application

remember that also if the model sounds bad dont blame on me, i'll be using your exact dataset, no cleaning whatsoever

👀

removing reverb, removing noise
any dirt noise/sound that is not a voice

artifacting such as clicks, etc

did you truncated the silence of your dataset?

or kept the natural silences intact

[full rover]

hey just looking for like a simple answer since im not the brightest but what website can i find beatrice v2 models on

[analog obsidian]

deleting the silence like this

no but like, you said you removed every pause

i need natural pauses

yes thats bad

coz rvc needs natural pauses

otherwise model results are meh

yes bc after that i do truncate the excessive silence and i keep the natural pauses

you have the original one?

without the pauses removed

ok so i can train that but

possible issues:
model will not be able to inference silence
longer training time

im looking the spectogram rn

to be fair i've trained models that have similar pauses to that

yes, rvc also injects silence specifically to help datasets like yours

to help them learn what silence is

okay i'll be training ur set

yes, the model will be able to generalize more

30 minutes is enough for good generalization

hold on
which separation model are you using?

also

wait

your dataset has harmonies

i cant train that bruh

only mono audio

no harmonies, chorus, whatever

minute 14 in your dataset

harmonies

yuh

no overlapping in the dataset

only 1 voice

🦈 👍

too old

use gabox fv4

download the uvr beta roformer version https://github.com/Anjok07/ultimatevocalremovergui/releases/download/v5.6/UVR_1_15_25_22_30_BETA_full.exe

after installing that exe, install the patch (DON'T INSTALL THE PATCH BEFORE INSTALLING THE MAIN VERSION)
https://github.com/TRvlvr/model_repo/releases/download/uvr_update_patches/UVR_Patch_1_21_25_2_28_BETA_small_rofo.exe

uninstall the old one

before installing this

after u install all you do this
(click install model)

and you follow the steps

pretty easy

https://huggingface.co/GaboxR67/MelBandRoformers/tree/main/melbandroformers/vocals

download voc_fv4.ckpt

and voc_gabox.yaml

uvr will ask for these when installing the model

[analog obsidian]

^

with these settings

u can decrease the overlap to 8 or 4 if you dont care losing some audio information

16 gives better SDR outputs

[knotty moth]

the speed

[analog obsidian]

but its veeery slow

yea

u can speed up by decreasing to 8 or 4 but u got less sdr

but hold on

we have a colab for dis

no need to install all of this locally

if u have google colab premium u can make this very fast by using an a100

🦈

[knotty moth]

the spectrogram consistency

[analog obsidian]

uhhh

[knotty moth]

pointless, the colab T4 is still not too slow for overlap 16, it's around half the audio duration

[analog obsidian]

amd moment

[analog obsidian]

technically true

1 sec

he has a colab actually

with all of this

https://colab.research.google.com/drive/1IC6Q1hLF55_tK6mhky0SWYKGVF9T5WsY?usp=drive_link

everything is self explanatory

just read the cells

yess

so your output dont have noise introduced by 24bit wav

codec things

🦈

[analog obsidian]

^ do this

after that you can clean your dataset, remove noise, harmonies, chorus, reverb, echo, whatever

what i need is only his voice alone

mel karaoke model instead of gabox fv4

is not perfect and gives muddy outputs

its up to you

tho if the audio is very muddy rvc will make the model sound robotic

[knotty moth]

karaoke after fv4

[analog obsidian]

O

i thought mel karaoke worked bad on already isolated audio

[knotty moth]

using karaoke directly on mixture can work, less bleed but muddier

[analog obsidian]

fv4 first

then mel karaoke

[analog obsidian]

thx 🔥

[analog obsidian]

which model you recommend for deverb?

after mel karaoke you have to use a dereverb model

yuh

yess

🔥

[low shard]

wdym with ai cover device?

also tell your pc gpu

[analog obsidian]

in case u forgot the settings just check this

works for every model you'll be using

[knotty moth]

that's too cheap for commissioning to a model master

also they would only accept through paypal etc. instead of such in game currency/tradable items

even that's too low

[umbral patrol]

!colab

!help

[dull iron]

Wally Commands

---

-# The prefix for commands is !

Select a category from the menu down below to view all related commands

[tawdry gull]

LunaBot 🌙#9997

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

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

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

Important Links:
Support
Premium
Invite

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

Select A Page From Dropdown Menu Below

LunaBot · Mar 10, 2025, 3:26 PM

[umbral patrol]

whats the command

i forgot

[tame mica]

-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

[umbral patrol]

thank you 👑

[low shard]

reminder to also check #📰│dev-updates

[analog obsidian]

its on the same colab

btw i forgot to mention, dont remove breaths from the dataset

keep them

[analog obsidian]

yea its not perfect sadly

u just remove these

manually by silencing it

(don't delete, SILENCE it)

deleting it causes pops

silencing it does not give pops

[analog obsidian]

rvc hates pops

[topaz sierra]

would i have better performance using internet than my 1650 yo

[viral mason]

this probably isn't the right channel but been having issues with UVR5 UI huggingface space with it just not wanting to work

[analog obsidian]

yes but u can easily avoid them by just silencing the parts you want to delete

audacity literally has a silencer

hold on i found something

silencing on audacity gives pops

[simple ore]

@analog obsidian a sharp cut is only visible in the voiced parts of audio

and it is all across the spectrum

[knotty moth]

enable feathering in izotope RX to enable crossfade while doing silence

[analog obsidian]

forgot to mention that

but

hes doing it on audacity

[analog obsidian]

i would prefer to have a pop free dataset

[knotty moth]

and yea pressing delete could still cause popping

[simple ore]

yes, you should not cut in the middle of

but the training happens on a small random slice of a segment and that can be anything

[analog obsidian]

what i shown it was actually an example of something you shouldn't do

[analog obsidian]

what are your settings for feathering?

🥹

tldr of this is to avoid having pops in your set

rx11

yup

yes also very important

you have to disable audacity's dither

adds noise too

🦈

nice

yess nice

no pops

you also can see the spectogram in audacity

pops would look like that

🦈

right click the track and click spectogram

use that if you're unsure that you fixed the pop

🦈 🔥

[crude flame]

do you need the "funds" for rx11?

[knotty moth]

32 bit float shouldnt have dithering at all

[analog obsidian]

better to have it off regardless

for rvc i highly reccomend using rx

👀

[crude flame]

hi

[placid holly]

After i Update the Applio Today (pulling it from Github), where tf is selecting training mode, like Hifi-gan & refinegan??

[crude flame]

i can make a "donation" for rx if you want

1 min per epoch

6700xt is faster then a 6650xt

[analog obsidian]

refinegan got deleted because it was having too much problems

[crude flame]

well it is

6700xt is barely faster then colabs t4

just sell both the cards and get a nvidia

[placid holly]

i see, but should just delete refinegan option only, not all option

[crude flame]

hes fine

more like 200-250

if youre going to do ai its so worth

idk

im not on linux

[analog obsidian]

mrf-hifigan at that state wasn't better than hifigan either

[placid holly]

so what it default mode to?

i can't select like RVC or Applio like Hifi-gan

[knotty moth]

refinegan has advantage on absence of spectrogram mirroring but unfortunately it still suffers more robotic sounds and some other problems

[analog obsidian]

hifigan

[crude flame]

ask me in october (ill be on linux then)

yep

[analog obsidian]

by default the model should work in every rvc client (w-okada, mainline)

[placid holly]

thank you

[crude flame]

so would you like me to "donate" the "funds" for rx?

sent you the "funds"

[analog obsidian]

🦈

[knotty moth]

if u still stick on the amd card, I'd recommend running under rocm in linux, and 7000 series ones or 9070/XT are better in rocm support and optimization

[analog obsidian]

just buy nvidia

[knotty moth]

windows also getting worse, and linux gaming is getting better thanks to valve

[crude flame]

valorant 🕯️

thats also stopping my switch 😔

[knotty moth]

im not really a fan of multiplayer games that inevitably use anticheat that lacks linux support

[crude flame]

and the 60gb of rvc datasets and stuff

[analog obsidian]

i hate this fr

[crude flame]

😔

[analog obsidian]

[crude flame]

nope

i have a single 64gb usb drive

dont have one

no, i give it away

[knotty moth]

I have made 400 gb free space after purging my old model datasets having done already

[crude flame]

wild

[analog obsidian]

lol

[crude flame]

my fault 😔

?

i didnt even know thats what this is

[knotty moth]

isnt it in mac/ios?

[crude flame]

this is some theme i liked

[analog obsidian]

ill remove it when you give me the cleaned dataset

30 mins to 1 hour max for a model

time to remove the harmonies i guess

just silence it and remove any pops

no need to delete the whole clip, you can apply silence to only the parts where you hear harmony

then do the thing to fix the pops

use rx

idk how to fix pops on audacity

yea remove those

[glacial pollen]

for pops, rx can be used or if you're desperate, heavy eq on the lower end

[analog obsidian]

just don't leave the pops and no need to go for over 1 hour

this is fine

[glacial pollen]

30-40 is more than enough

[analog obsidian]

yess

[glacial pollen]

sometimes it might even be bad to go for too much of data, if it's too divergent or repetitive

[analog obsidian]

🦈

a properly cleaned dataset sounds great so dont worry

delete the first one
second one is fine

only that

just delete everything that is very damaged bruh

🦈 delete that small part

or well silencing it

[olive hill]

Anybody knows why its not loading the audio

[low shard]

holy shiet you're using a very very old version of applio

[olive hill]

Whats the latest version?

[low shard]

what's ur pc gpu and what u tryna do

[olive hill]

GPU NVIDIA GeForce GTX 1050 Ti

i am just converting a vocals to ai version, not training

it was working till yesterday, dont know what happened today

[analog obsidian]

yes

[olive hill]

Please help

[low shard]

were you using a local downloaded version or a cloud huggingface space?

[olive hill]

Local

[low shard]

yeah u prob download some old youtube shit

[low shard]

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

You can:

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

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

[olive hill]

which is latest?

Applio?

[low shard]

all the ones I sent are updated, you can choose based on your needs

[olive hill]

Thanks mate, appreciate it

[low shard]

the other ones are more updated than mainline though

[low shard]

you're welcome

[olive hill]

but those are cloud based right?

[low shard]

both of those have a cloud and local version

mainline colab is broken right now though #📰│dev-updates

[olive hill]

i dont wanna train any model, just wanna use for conversion, which do u suggest?

[low shard]

are you going to make ai covers or just convert files?

[olive hill]

Both

More of Conversion only

[low shard]

Applio locally if u want to wait, or applio colab ui if u can deal with the gpu time and don't want to wait more time for conversions

[olive hill]

gotcha

thanks

Installation process is same as old vers?

[low shard]

I'm not sure how old version installation was, but it's pretty easy, you just gotta read the guide

[olive hill]

Cool Cool

?

[analog obsidian]

save it as flac

yup

dw i'll resample it later

its fine just upload it somewhere else

[zinc thorn]

Guys, I installed the Snowie v3 pretrain model and before I even started the training it ended with error. Any solutions?

It works with default pretrain model, but when i switch it to Snowie one - it brokes

[simple ore]

you probably did not download it completely so the file is corrupted

[zinc thorn]

download completely what?

[simple ore]

the pretrain

[zinc thorn]

I use bult-in download pre-train feature

[zinc thorn]

Where i can download pretrain model completely?

[simple ore]

https://huggingface.co/MUSTAR

[zinc thorn]

Oh ok i just downloaded it from another source

[simple ore]

if thats what you really want

I think the issue is that because @scenic gale moved the files to another place, you get a huggingface error html trying to download 40 and 48k models

[broken urchin]

can someone help me

Voice blender isnt working in Applio

gives an error

[simple ore]

different sample rate? or the sample rate label (40000 vs 40k)

[broken urchin]

yeah they are different sample rates

32k and 40k

[simple ore]

cant blend that

[broken urchin]

i must do a 40k and 40k?

[simple ore]

yeah, only similar models can be blended that means same sample rate

[broken urchin]

ok

[glacial pollen]

Btw, the slider's description is a bit misleading
It controls the weight / influence of model A ( left one ).
In a short, 70% / 0.7 on slider means the model A gonna be dominant in 70%

No idea why would anyone declare the function as " the side you move the slider to, has more influence " ish

[broken urchin]

i just tried to blend a 32k model with 32k model but still get the same error

[glacial pollen]

Yea, it's a lil broken

drag n drop

the models

then it'll work

[simple ore]

then it is probably the 32000 vs 32k label

[glacial pollen]

Nah, it's just broken ( the picker, I mean, the path box

same thing on my fork

Gotta drag n drop the files

[broken urchin]

yeah i dragged and dropped the files

[glacial pollen]

huh

[broken urchin]

but isnt that the same thing

[simple ore]

when it compares values, it is not

[glacial pollen]

^

[simple ore]

put the models in some folder

[glacial pollen]

gotta edit the dict

[simple ore]

use a path

[analog obsidian]

O
today i learned... THIS

[simple ore]

manually provide the path C:\models\model1.pth, C:\models\model2.pth

[glacial pollen]

:' )

Yea, whoever does the translations, is kinda goofy

[broken urchin]

ok

yeah still just gives an error

then i have to make similar models and try

[glacial pollen]

Try to restart applio

and try again

If that doesn't help, get my fork ( just copy the env folder to it )

[simple ore]

you can do this

[glacial pollen]

As long you use the repo's one and not precompiled, it'll work

[simple ore]

edit this file in notepad

rvc\train\process\model_blender.py

[glacial pollen]

@broken urchin

[simple ore]

and delete lines

that's the check for sample rate

[broken urchin]

wait

how do i do that

[broken urchin]

im using the applio from ai hub docs website

[glacial pollen]

yea, the zip one then

[broken urchin]

thats what i use

[glacial pollen]

Anyways, just do what noobies recommended

[broken urchin]

where is that?

[glacial pollen]

you know

Applio folder and from there extend it to the path noobies provided

That's the rvc folder

[broken urchin]

ok i found it

and know

now *

i edit it with notepad+ ?

[simple ore]

yes

[broken urchin]

and i just delete this line

ok i did it

ill try now

[simple ore]

restart applio

[broken urchin]

ok

again i got an error

@simple ore

[simple ore]

send links to the models

[broken urchin]

in dm?

[simple ore]

whichever way

[broken urchin]

ok i will

[simple ore]

hm

[broken urchin]

whats wrong?

[simple ore]

maybe it is my issue

regardless, the models you're trying to merge are different

they are 40 and 48k

the ones you sent me

[broken urchin]

ok

i will try on similar models

[viral mason]

chat help

kaggle applio broken

[scenic gale]

Hello. I think no. Haven't touched those files for some quite time

[trim wave]

Hello, does anyone know what software use to have the same voice as him (text to speech). The voice actor is called Michael McConnohie (HunterXHunter narrator) and go check the link to see what i'm trying to have pls (go at 15 seconds to hear the voice i want). https://www.tiktok.com/@opiumomen/video/7476499628079320366?is_from_webapp=1&sender_device=pc

[glacial pollen]

Well, the workflow would be to first create an " rvc model " ( using applio )

NOTE: RVC models originally are for voice to voice conversions ( v2v )
( The way it works is: You input target vocals / acapella ( idk, some rapper's vocals extracted from some track, and the model you trained on some voice, say, mickey mouse, gonna sing that in style of the original vocals. )

Then, once you have the model, the model would be used for real-time voice changing / conversion ( Software for that we call " W-okada " )

Alternatively, there is a chance someone already made a model with the voice you desire.
Worth checking: #🔍│find-models #1175430844685484042 and www.weights.com website for models

.
So these two are for, respectively, covers and real-time voice changing

Now, if your interest is in TTS voice cloning then there's a few solutions around, yet I personally would recommend " gpt-sovits "

There's other things as well. Xtts, tortoise, style-tts, etc etc

But I believe gpt-sovits is the most straightforward and easy to do

[trim wave]

Thank you a lot, and yeah my goal is to use TTS because i will use the voice in Video and i think the audio quality is better in that mode than in covers or real time voice changing right?

[glacial pollen]

In that case, gpt-sovits yes

However, if I may add something from myself

A really nice workflow with polishing the tts inputs is to use the cloned voice's output as input for rvc ( and rvc using the same voice model )

That way you'll boost up the quality

Here's an example:

Gpt-sovits due to gpt part of it ( you can think of the gpt component as the language / context understanding element )

learns how to speak in a manner and tone, style and so on, of the speaker

however there are some gimmicks and quirks in the voice cloning part itself

[trim wave]

so

by doing that

u have more emotions?

in the voice of the character

its like his personnality

was represented

[glacial pollen]

essentially, gpt-sovits:

1. clones the voice
2. does learn to more or less reproduce the style, emotions and manners of the speaker

however getting the quality decent is a lil difficult and has implications

rvc however, being a voice to voice, uses that tts's output to enhance the quality
( provided the rvc model is done on the same voice / samples as your gpt-sovits model )

Pretty much. tl;dr, gpt-sovits would be your input audio source and rvc your refining / finisher / polisher tool

[trim wave]

yeah i'm trying to understand how it works

[glacial pollen]

Ofc, it involves learning 2 programs, but the end results are worth it

This is the invite for gpt-sovits repository's discord: dnrgs5GHfG

[harsh ravine]

hey how do i find voice models

[trim wave]

thank you

[glacial pollen]

@harsh ravine

Search there

[glacial pollen]

Yea, np man

That's the exact workflow I follow rn

[harsh ravine]

i did but i dont know how to activate the search

[knotty moth]

I don't remember which version but the applio I was using worked on blending two 32k models, as I manually entered both the model path

[glacial pollen]

Perhaps, ye

tho what I meant is, manual-path-wise blending seems to be jammed compared to drag n drop

[formal wind]

Chat is this overtraining or not i'm so confused

it stays roughly still for a while then goes down

[zenith zealot]

I need help, which version of rvc should I download, I am using amd, I am also on windows 11. Should I download the vcclient_win_std_2.0.76-beta.zip? I want the latest release but the one that works for what im using

[glacial pollen]

Inaccurate graph, sorry my dude

Not sure why you're still running old applio

[formal wind]

Old applio?

[glacial pollen]

yes

Depending on whether you go for Noobies approach or mine, Applio is now using new method for logging the metrics

As you can see, there's average loss incorporated

[glacial pollen]

the prepackaged / pre-compiled zip one is still not updated I believe

[formal wind]

I use kaggle

One of the applios

I can send a link

[glacial pollen]

Oh, Kaggle

[formal wind]

Yeah lol

[glacial pollen]

well

that changes it a lot

[formal wind]

Good or bad change

[glacial pollen]

Either way...
( examples of how it could look like (( ofc, there's way more patterns than that, but you get the idea

graphs are inaccurate butttt, if you really wanna know then no

doesn't seem like overtraining to me

The change would be way too abnormal compared to the rest of graph

and from there it could " supposedly " get better or hover around flat ish with occasional drops

Orrrr, could flat out and go up

Great aid is also looking at discriminator really

If it's loss ( for discriminator ) seems to be abnormally going up and then getting stable + dropping or dropping right away, then you know gen fails hard

[formal wind]

I feel more confused than when I started

[glacial pollen]

clears throat just refer to this

if you see no patterns of that sort, keep on training

Here's some real life example

You see the dip right?

between 4 and 5k

then there's a valley like and flat ish region ( 5.5k to 6+ k )

There ^

Until you get such patterns or those I demonstrated before, just keep on training.
In the end, you'll be ( hopefully ) testing lots of ckpts anyways to find the best fit ( From those around the " dip " on graph

[formal wind]

Yeah listening is the most annoying part

[glacial pollen]

Depending on how you look at it 👀

Better to make absolute sure it's that one than worry later and regretting, having that thought in the back of your head it coulda been better lol

ig

If you're struggling with graphs I think catching up on how Generator and Discriminator synergy works would help you

Most videos are pretty basic and max 10-15 mins ( some even demonstrate the loss dynamics

That's how I started initially anyway

[formal wind]

There is way too much info for my brain to be processing

I might die right here right now

[glacial pollen]

well, just look at the pictures, that's all

rest comes with time

I leave such information so others can benefit from it too ( those more advanced

or might be you gonna come back to it some time later yk

[formal wind]

I'll try my best to understand this

[glacial pollen]

For now, just focus on spotting the dip and then a dramatic change in ur graphs

[formal wind]

Its at 1k epochs and still looks like the image I showed you basically. It feels wrong

[glacial pollen]

1k epochs..

[formal wind]

Yeah my exact thoughts

...

[glacial pollen]

Even pretrained / base models don't do that

uhhh

what is ur dataset's size and batch size

to begin with

cause like, 7k steps for 1k epochs 👀 crazy

[formal wind]

3:05
batch size 4

[glacial pollen]

oh man

ughh

That.. is most certainly an overkill

[formal wind]

Stopp Im boutta crash out

[glacial pollen]

aaaand that changes the whole situation, lemme look at the graph again

[formal wind]

Ill send you the current one

[glacial pollen]

before that

can you do few things

[formal wind]

Yep

[glacial pollen]

set the smoothing to around that, ignore outliers, resize the D and G total graphs

also, rescale with the button tagged with number 2

for both G and D

and if ur graph's bottom is cut

[formal wind]

Tensor board is TWEAKING 😭

[glacial pollen]

well dang uhhh, can you send me the tf file?

Might take a look at it, seems messy as hell

[formal wind]

tf file?

[glacial pollen]

tensorboard file containing ur graphs

it is located in model's folder ( in logs

gonna be called tfevents something something

also, it should have a size that exceeds 1kb

if it's not directly in model's folder, then in eval folder ( don't remember how old applio builds handled it

[formal wind]

theres 2 tfevents

[glacial pollen]

Did you resume the training at any point?

[formal wind]

Nope

[glacial pollen]

both have size above 1kb?

[formal wind]

both 88b

[glacial pollen]

Try to find one that weights a bit

unless you meant 88k kb ( 88mb

[formal wind]

It just says 88 B

[glacial pollen]

the heck

ss it and show here

[formal wind]

Oh ok well theres 2 tfevent files in eval and in the models path

one is 103 mb

[glacial pollen]

then the heavy one is the one

other is empty and useless ( newer applio builds don't generate that one

[formal wind]

Is there a newer version of applio on kaggle?

[glacial pollen]

Not sure if someone done any updates

[formal wind]

Wait so do you still want the file, I've got the download link for it copied

[glacial pollen]

ye, can check it for you

[formal wind]

https://huggingface.co/Homiebear/Blahblahblah/resolve/main/events.out.tfevents.1741665421.2285d1bf6124.508.0?download=true

[glacial pollen]

Alr, gimme a bit. First gonna help one dude in dm installing applio

[formal wind]

Thats alright. Take your time sorry for me bein' braindead 😭

[glacial pollen]

Alr, lemme take a look

[knotty moth]

like said above, the avg graphs are more accurate and the avg g might have saturated or turned up earlier

[formal wind]

jeez I thought I knew enough about training models yet I have no clue what suturated means in this context

[glacial pollen]

Welp

so

[formal wind]

Jesus

[glacial pollen]

that's around where it ends ( hopefully )

because..

lemme showcase it graphically

wait

See how at first it goes rather " uniformly ish " down?

and then, the more it goes, more " flat like " it becomes, even if at first glimpse it's descending

Now, we can't be 100% sure because these graphs are inaccurate

yet my best bet would be to.. check this region

Which is this

[formal wind]

I was boutta ask that thx

Lemme listen to them all

[glacial pollen]

Feels bad man

It's so darn messy that even I'd probs give up

not only the silence nodes

but so many of those points you can barely read it lol

Why won't you train locally btw

[formal wind]

My gpu doesnt work with it

It just has so many issues

[glacial pollen]

amd or rtx?

[formal wind]

Its an AMD Radeon RX 6600

[glacial pollen]

right

anyway.. If you asked me

I'd check those

Unfortunately those ultra deep dips are a hit or miss

so yea, no other choice but to listen to 'em all

And if you wanna know why old graphs are inaccurate

You know how an epoch has steps right?

[formal wind]

yeah

[glacial pollen]

Old graphs long only the last step

from a given epoch

at each " node " in the graph

New logging averages the result over all steps from a given epoch ( at least in my approach, noobies handles it differently

[formal wind]

So assuming by your response before, training locally wouldnt work

[glacial pollen]

I can't say for sure as I never tried playing with rocm nor attempted training when I still had AMD

Noobies def knows more than me about it

as he used amd before

[formal wind]

Oof

[glacial pollen]

or, well, zluda

I think there's a part on it in the guide? in docs

https://docs.applio.org/applio/getting-started/installation

should be, ye

search for zluda in there

[knotty moth]

the normal g graph have sharp dips because it logs batches that learn a mute file

[glacial pollen]

^

Or so happened the last step ended up on the mute file

In any case, a hit or miss

[formal wind]

yeah Imma give up on that model

Ive never had an issue like that so Imma try reinstalling applio

Wait you said don't download the Compiled Version earlier right?>

[glacial pollen]

Depending on how you wanna go about it, there's Applio from repo ( not the zip ) or my fork

They're in-line when it comes to core functionality but some things are changed in my fork

For instance the way it logs, the ui and stuff ( and some more

Noobies' approach is based upon averaging every N steps

mine's done on all steps from a given epoch basis

[formal wind]

Well whats best in your opinion

[glacial pollen]

looks like so

Well, there's no winner but

As I'm the owner of my fork, Imma recommend just that

[formal wind]

Lol

[glacial pollen]

Anyway, whichever pick you make, it'll be fine

but I think my logging's a bit more safe ( subjectively

[formal wind]

Where do I get my grubby hands on your fork

[glacial pollen]

https://github.com/codename0og/codename-rvc-fork-3/releases/tag/v3.0.4

[formal wind]

https://tenor.com/view/ecospace-compostable-cutlery-save-the-planet-eco-gif-15402505

[glacial pollen]

There haven't been any zluda changes ( afaik ) for a while and so, all is in-line. Meaning it has to work ( if it's meant to work ) else, you'd have to ask Noobies for some help

[formal wind]

So does the installation for normal applio match up with yours, or is it a different process?

[knotty moth]

(why not both)

[glacial pollen]

Becaue it doesn't reflect exact epoch's performance

but more averaged / long-term one

having fixed 50 steps just isn't ideal imo ( when you want a per-epoch-monitoring strategy

Hence why it depends on what ones prefers really

I like to epoch-scope so that's the approach I took

Matter of how you approach it tbf

[knotty moth]

well yea per epoch avg seems to make more sense

[glacial pollen]

it's far from perfect for sure

but still tons better than what it originally was

Other than per epoch

I also incorporated per 5 epochs avg so there's also longer-tendency checkup

[formal wind]

What does it mean by:
"Using command line from the Applio folder run"

Does it mean run a command prompt from inside the folder and paste the line of code in

If so it doesnt work

[glacial pollen]

I have a feeling they meant Applio folder's runtime

propably?

[formal wind]

Wdym runtime

[glacial pollen]

like, environment

env

[knotty moth]

something like env\python.exe app.py --open

[glacial pollen]

well, show me which guide you read and where exactly you're stuck

( gimme a link

[formal wind]

https://docs.applio.org/applio/getting-started/installation#install-applio

"Install Applio"

2nd step

[glacial pollen]

Which route you should go?

a, b or c

[formal wind]

Wdym

[glacial pollen]

oh

wait

[formal wind]

B

Since it corresponds with my gpu

[glacial pollen]

ye then that basically means:

you open up a cmd in applio, and execute these

just like so

Gonna open up cmd right there

[formal wind]

yeah I was doing that

it wasnt working properly

its fixed now though

[glacial pollen]

o

[knotty moth]

I'm curious if there's eventually someone with 9070/XT willing to test

[glacial pollen]

Good point

[olive hill]

How to load models, locally which folder?

[glacial pollen]

easiest to just make a folder for each model

in logs

or pile em in 1

[olive hill]

Should i rename them as "weights"?

[glacial pollen]

Then hit refresh

[glacial pollen]

nope

name doesn't matter

but for best sorting, I'd make a folder for each model, having there pth and index file

[olive hill]

Okay

[knotty moth]

would feel impressed if they find if it could match or outperform 7900 XTX

[glacial pollen]

ikr

But then, I'm unfortunately still team Nvidia

That is, until amd finally takes AI into own hands and make shit work as it should, natively on windows

[olive hill]

Which to choose for best output?

Hybrid?

[glacial pollen]

no, rmvpe

hybrid can be quirky at times

[knotty moth]

yea eventually esp for linux users

[glacial pollen]

yup

[olive hill]

whats fcpe?

[glacial pollen]

Generally, embedder depends on what model was trained on

but in reality, I'd say 99 or 98% of models are contentvec

[olive hill]

Oh gotcha

[glacial pollen]

as for rmvpe, fcpe etc, it's pitch extraction method

fcpe.. I haven't tested it much but it's probs faster or more lightweight than rmvpe?

In any case, rmvpe is the way to go as it's the most robust one

[olive hill]

thanks mate

[glacial pollen]

Np, cheers

[knotty moth]

a newest pitch extractor being more lightweight than rmvpe (or can just say it as "bootleg")

[olive hill]

but is it good then rmvpe?

[formal wind]

What does it mean by "edit the run-applio-amd.bat file and change the value from "0" to "1"." (I'm like so frickin' sorry for the hassle)

[knotty moth]

rmvpe is better for most cases, except fcpe might crack less on doubled pitches just like crepe

[olive hill]

ah gotcha

[glacial pollen]

first

It's assumed your primary AMD GPU has index 0. If your iGPU is listed first under 'Display Adapters' in Device Manager, edit the run-applio-amd.bat file and change the value from "0" to "1".

Is it 0 or 1

Or rather, Imma ask, do you have any igpu?

[formal wind]

Wtf is an igpu

[glacial pollen]

ur cpu having one?

It is integrated gpu within your cpu ( as long yours have one

[formal wind]

How do I know the answer to that question.

[glacial pollen]

What's ur cpu

[formal wind]

Intel(R) Core(TM) i7-10700F CPU @ 2.90GHz

[glacial pollen]

Well, you typically know if you know ur cpu's details

F series do not have igpu

[knotty moth]

check task manager if it is labeled GPU 0 or 1

[formal wind]

its 0

[glacial pollen]

In that case, you don't have to edit anything

The whole idea is to change 0 to 1 ( as that's your main gpu ( indexed as '1' ) in cases where you have igpu, which is seen as 0

[formal wind]

Ill lyk if I have anymore issues

So sorry!

[glacial pollen]

👀 Worry not, hope all works for you

[formal wind]

Thank you Mr. 0🙏

[glacial pollen]

✨

[formal wind]

I have to hand it to you it, your fork looks alot cleaner than regualr applio

The box where I have to put the path to the dataset. Do I put the path to the dataset thats in my file explorer cause I keep getting error's

[glacial pollen]

What does the error say?

[glacial pollen]

Also thanks. Am very proud of that theme lol, my baby

[formal wind]

Lol

[glacial pollen]

oh, that one's weird 🤔 haven't seen it yet lmao, the error

[formal wind]

A new one... Why am I cursed

[glacial pollen]

Anyway, try to put the dataset in:

assets/datasets/

so:
assets/datasets/my_pog_model/sample.wav or multiple .wav files

[glacial pollen]

Btw, for errors it's better to check the console

gradio likes to simplify / obscure lots of things

[formal wind]

Well the issue turned out to be that the console crashed

Lol

The error is fixed but its pretty late so I might get some shut eye

[knotty moth]

it sometimes pop out in applio kaggle due to probably ngrok's "connection instability"

[glacial pollen]

Makes sense, in that case

[glacial pollen]

Ye, get some rest man

[formal wind]

Tysm for tonight! Both of ya!

[glacial pollen]

Np man, have a nice one, Gnight n sleep well

[trim wave]

@glacial pollen Hello again, sorry for bothering you, do you know by any chance how to use the voice I talked to you yesterday (Michael McConnohie) and use it in ElevenLabs (TeamtoSpeech)?

I don't find any options to import the voice tbh

But someone told me he cloned it

DK how

[glacial pollen]

I think ( can be wrong ) that you need subscription to clone voices ( as in, custom )

iirc, free users are both limited in total usage time and limited to a specific library of voices

[trim wave]

Oh

Lemme check

[orchid orbit]

I am looking to run some of these voices locally for annoucements-- can I use coqui with these models?

[trim wave]

How do i install gpt_sovits? and can i write text in the app and use the cloned voice in it?

@glacial pollen

[glacial pollen]

You gotta follow the instructions on repo

https://www.youtube.com/watch?v=haw7glBmIbA

alternatively ^

yet as for specific values, configurations and all of that.. it's something you have to figure on your own

There's almost lack of guidelines everywhere

[trim wave]

i'm checking

I have like two versions for the software : the one from github ("GPT-SoVITS-main" and the one from the guide online "GPT-SoVITS-v2-240821"). Is there a specific one to download? @glacial pollen

[glacial pollen]

you can grab the v3 release

from releases

[trim wave]

in repo?

[glacial pollen]

yes

[trim wave]

lemme check

[trim wave]

where do you find it here?

[glacial pollen]

oh?

[trim wave]

wdym ohhhhh

[glacial pollen]

https://github.com/RVC-Boss/GPT-SoVITS/releases

[trim wave]

oh

[glacial pollen]

🤨

[trim wave]

mb bro

[glacial pollen]

smh

[trim wave]

i was testing you

all according to my plan

dw dw

[glacial pollen]

According to keikaku 😌

yosh

[low shard]

RVC is STS, not TTS, so nope unless you firstly make an audio with TTS then use it as an input in RVC

[trim wave]

how do i check env.py please i forgot @glacial pollen

[glacial pollen]

Did you download the zip

[trim wave]

yup

i opened cmd

in the runtime repo

and then

im blocked

[glacial pollen]

Tho, why you even doing that