✨│ai-help

[pastel oak]

Whats ur gpu and send screenshot of settings

Send screenshot

[feral marsh]

you want a screenshot of my settings, yea?

[pastel oak]

Yes

[pastel oak]

Chunk controls delay
Select 80 for example

[feral marsh]

ima try that!

god damnit

thank you.....

that fixed it.

[pastel oak]

easy

[tight tiger]

4060 ti

and my extra is like 13100 and my chunk is 2400

[brave garnet]

RVC Colabs and Spaces

⠀

Google Colabs

⠀
AICoverGen-WebUI
Useful for making quick covers, by Hina.

AICoverGen-NoWebUI
Useful for making covers, doesn't include a UI, by Ardha, by Eddy, Hina and Gdr.

RVC Disconnected
To train new voice models, by Kit Lemonfoot.

EasyGUI
The OG interface, by Rejects.
⠀

Made with love by Ilaria 💖

[brave garnet]

RVC Colabs and Spaces

⠀

Google Colabs

⠀
AICoverGen-WebUI
Useful for making quick covers, by Hina.

AICoverGen-NoWebUI
Useful for making covers, doesn't include a UI, by Ardha, by Eddy, Hina and Gdr.

RVC Disconnected
To train new voice models, by Kit Lemonfoot.

EasyGUI
The OG interface, by Rejects.
⠀

Made with love by Ilaria 💖

[sick wraith]

Yeah I can't get the voices to sound like the people they should. I downloaded Robocop and can't get b the voice to sound right

[odd shale]

If you're using W-Okada, maybe you could try with different settings and also reading the docs below.

-realtime

[azure marsh]

This interaction has expired, use the command /guides realtime if you wish to see it again.

[odd shale]

Check Deiteris' one.

[brave garnet]

RVC Colabs and Spaces

⠀

Google Colabs

⠀
AICoverGen-WebUI
Useful for making quick covers, by Hina.

AICoverGen-NoWebUI
Useful for making covers, doesn't include a UI, by Ardha, by Eddy, Hina and Gdr.

RVC Disconnected
To train new voice models, by Kit Lemonfoot.

EasyGUI
The OG interface, by Rejects.
⠀

Made with love by Ilaria 💖

[wild vale]

Okay so im confused right now.

The reason noone could answer my questions were because my questions didnt make sense?

That is very confusing.

Its not like I am speaking a different language

The Buffering rises above the threshold of 512 to 1000, and the res goes to 4k plus which makes my voice robotic and hard to hear.

[rare goblet]

Ayo? @wild vale level 2 !!!

[wild vale]

What part of that is hard?

Do I edit the chunk or something or am I missing something?

When I use a voice in discord chat its fine

But when I use it in game or a heavy game
The Chunk goes from
buf:512
res: 12-128

to

buf: 3x the normal
res: Goes to places higher than 2.7k

[simple ore]

since there's no way to prioritize RVC over the game, yeah, that's gonna happen

get another GPU I guess

[pastel oak]

Chunk could be too low for your gpu, id need to know the ms number instead of the 2400 number

Else download v1
https://rentry.co/voicechangerguide

[pastel oak]

I dont know who youre flaming here but youre in the wrong channel first of all but ok

You might be running into 100% GPU issues, so you have a few options to try:

• reduce your ingame quality and cap fps to just above your monitors refresh rate
• increase chunk and reduce extra for less gpu and cpu load from the voice changer
• if that didnt work, try out the fork. Has very little resources used and runs better: https://rentry.co/ForkVoiceChangerGuide

If all fails and it turns out youre playing a very high end game that goes to 99% gpu usage either way, then upgrade gpu or get multi pc setup

[odd valve]

anyone know the end of some words get cut out when u say a sentence with rvc

it doesnt cut out, but the ai gets weird and almost ignores the last few letters if you know what i mean

unless you really put emphasis, which just kind of makes it sound unrealistic

[pastel oak]

If your threshold is too high it might not pick up if you get quieter at the end of the sentence below the threshold

Move threshold/n gate to the left if its on the right

[odd valve]

i have it on the lowest thing

[pastel oak]

Oh ur on rvcs voice changer

Send full screenshot

[odd valve]

like i said it doesnt "cut out"

it just ignores some last letters

most of the time

[pastel oak]

Put extra to 2.70

[odd valve]

what does the extra do?

thats soo much better

[pastel oak]

Extra is both the voice model quality and controls the length of a consistent tone, like if you hold a tone aaaaa you can hold it up to 2.7 before the voice breaks. And in this case, 2.7 is considered the max setting (most models struggle to go above this number, but some are capable of it)

In rvcs gui it does more damage than benefits to go above 2.7 from my testing

[odd valve]

i see, thanks though it fixed basically everything

[brave garnet]

RVC Colabs and Spaces

⠀

Local Forks 🖥️

⠀
Mainline RVC
Original project, suggested for advanced users,
by the RVC-Project team.

Applio
Simplified, suggested for all, by the Applio team.

RVC Studio
Simplified, suggested for all, by SayanoAI.

Mangio-RVC
Simplified, may not be supported anymore, by Mangio621.

AICoverGen
Simple yet great way to make covers, by SociallyIneptWeeb.

Replay
From the greators of weights.gg, excellent product for everyone.
⠀

Made with love by Ilaria 💖

[next plinth]

How to use this AI 🥹

I'm newbie and i don't understand anything 😭

[low shard]

which ai ?

and whats ur pc gpu

[next plinth]

I want to try Genshin rvc model by HuggingFace but the web i click on doesn't look like the old one 🥹

[low shard]

whats ur pc gpu?

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

Also, are u looking for ai covers or realtime

[next plinth]

Is "Intel(R) UHD Graphics" a GPU? I don't understand 🥹💔

[rare goblet]

Ayo? @next plinth level 1 !!!

[low shard]

its the integrated graphics, which is bad, u cant dont do it locally (on ur pc) but can use cloud (remote good pc)

are you looking for ai covers or realtime for calls

[next plinth]

Ai covers 🥹

[low shard]

use ilaria rvc zero, a zerogpu (A100 paid by Ilaria) huggingface (biggest ai platform) space (service they offer to try ai), its the fastest way

[brave garnet]

Ilaria RVC: CLICK HERE 🤗

Guide on how to use it: CLICK HERE 📝

Don't forget to thank Ilaria if you find it useful! 💖

[next plinth]

O-okay

Thank you

[sick wraith]

What do I need to do to get the voices to sound right in voice changer? They always sound off? Am I supposed to be tweaking the voices based on the sound file or voice I'm using?

[odd shale]

Probably some models you're using aren't meant to be used on W-Okada.

[sick wraith]

Ah ok. Are rvc models not universal?

[odd shale]

Some voices can work for everything, some don't.

It also depends (i think) how the author made the model. (dataset cleanup and length)

It can also depend on your settings and your own voice.

[wild vale]

I have an
Processor Intel(R) Core(TM) i7-10700K CPU @ 3.80GHz 3.80 GHz
Installed RAM 32.0 GB
Device ID 60F634EE-B521-4FCF-A554-CDCB5FDC830E
Product ID 00330-80000-00000-AA684
System type 64-bit operating system, x64-based processor
Pen and touch No pen or touch input is available for this display
RTX 3060Ti

[wild vale]

Whats the right channel then

[analog obsidian]

Singing models are very bad at speech
While speech models are mid/bad at singing

and models not sounding like the original voice are undertrained or the dataset had timbre issues

[odd shale]

Y algunas veces no es tan buena idea mezclar canto y dialogo en un mismo dataset verdad?

Porque soy fiel creyente que es mejor hacer 2 modelos distintos de la misma persona/personaje dependiendo del proposito que uno le quiera dar.

[analog obsidian]

Es mejor hacer dos modelos distintos y no mezclar canto con dialogo

[odd shale]

Factos.

Es lo mismo que evito hacer.

Pero no mucha gente sabe de esto.

[pastel oak]

#🔍│help-w-okada

[pastel oak]

Ok, can you tell me what game youre running and what settings on wokada, send screenshot

[sick wraith]

where does ilaria rvc store it's models?

i took a secondary model and placed it in it's root folder where model.pth is but it's not detecting it

[low shard]

ilaria rvc zero?

[brittle wing]

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• 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
• Modified W-Okada's Voice Changer, Google Colab
• 🆕 FaceFusion UI, by Nick088 Google Colab
• 🆕 FaceFusion NO UI, by Nick088 Google Colab
• 🆕 EasyGUI, by Rejekts 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, -overtraining, -realtime, -rvc, -help

[rare goblet]

Ayo? @brittle wing level 6 !!!

[ornate hawk]

Which option should I choose for a pre-trained model if the dataset is at 44,100 Hz? Titan only supports 32k, 40k, and 48k

[pastel oak]

First check if its truly 44,1k hz because often times the waves show something different. You can check this with a program called "spek"

This is a debated topic: You can use 40k because you do not have those ranges from 44.1 - 48k in your dataset, so the model could get inaccurate. Some say use 48k because you wont hear the difference anyway and get more out of it

Imo it doesnt matter, i would probably still use 40k

[radiant loom]

Does anyone know how to remove the robotic sound at the end of words?

[low shard]

@jagged hawk im pinging gu in the right channel, whats ur pc gpu?

[jagged hawk]

4070 ti

[low shard]

alright good enough

the docs are temporary down so lemme send u the temp ones rq

[jagged hawk]

Ok, thanks

[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, same quality tho
• Mainline: The original RVC

[jagged hawk]

Oh nice

I'm instaling Applio rn

Is it intuitive?

Or u recomend watching a tutorial?

[low shard]

The guide i sent is already a written guide

there is no updated video tutorial

[jagged hawk]

Ohhh

[rare goblet]

Ayo? @jagged hawk level 2 !!!

[jagged hawk]

It's a link

I didn't realize 😅

Ty! ur the best

[simple ore]

you can always just read the docs

[wise lark]

How to watch voice model?

[low shard]

wtf, mind u refreshing the site?

[wise lark]

I refreshed site and I restarted discord

but I still can't see it

[nocturne mural]

I've experienced something similar. It might be because you still have an active search in the top section. Try clearing it and maybe that will fix it.

[sly sluice]

python trainset_preprocess_pipeline_print.py "/content/dataset/EVRAART" 40000 2 "/content/Mangio-RVC-Fork/logs/EVRAART" 1
python3: can't open file '/content/Mangio-RVC-Fork/trainset_preprocess_pipeline_print.py': [Errno 2] No such file or directory
python extract_f0_print.py "/content/Mangio-RVC-Fork/logs/EVRAART" 2 rmvpe 64
python3: can't open file '/content/Mangio-RVC-Fork/extract_f0_print.py': [Errno 2] No such file or directory
python extract_feature_print.py "device" 1 0 0 "/content/Mangio-RVC-Fork/logs/EVRAART" v2
python3: can't open file '/content/Mangio-RVC-Fork/extract_feature_print.py': [Errno 2] No such file or directory

i have the same error as this guy, except i did put .wav files, and that i already tried installing depencies again

[brittle wing]

How much crepe hop length in inference and training

[sly sluice]

i'm thinking about deleting the whole mango-RVC-fork folder to install it again is this a good idea?

[analog obsidian]

mangio rvc fork is very outdated and most of the dependencies aren't compatible to each other anymore
for training use mainline (the original rvc) or applio (fork of mainline)
mainline has faster ui and in some cases, training is faster than applio
applio has slower ui but some claim they have better training speed there

both options give the same result in terms of model quality, etc

[analog obsidian]

64 for both

[sly sluice]

k good thanks

[rare goblet]

Ayo? @sly sluice level 1 !!!

[brittle wing]

Wasn't it 64x2=128?

[analog obsidian]

128 is too innacurate when i tried training with that value (the model ended up having more voice cracks)

[brittle wing]

Okie but I'm asking about inference rn...and yes you're correct

[analog obsidian]

for inference try 64 or 32

[brittle wing]

I remember training a model with 128 hop length and it sounded bad

[hidden dew]

where do i find the saved models while training?

i cant find them

nevermind

i found them

lol

[crude flame]

Bruh why am i getting this error now, applio was working just fine the other day

ive already tried reinstalling the newest complied version and last versions pre compiled and it still gave me that error

[analog obsidian]

have u tried updating ur gpu drivers?

[crude flame]

yup

[nocturne mural]

.\env\python.exe -m pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 --upgrade --index-url https://download.pytorch.org/whl/cu121

try reinstalling the torch dependencies again

[crude flame]

still same error

[nocturne mural]

DeadCat

[analog obsidian]

what if u just delete python and everything related to it lols, thats what i do when shit stops working

[noble vortex]

ive been trying to use rvc webui for model training, but when i click on one-click training, the output information box has been stuck on 'processing data' for the past hour. any suggestions?

[twilit kernel]

Hi did anyone manage to download Mangio on Mac?

[low shard]

mangio is outdated, its better to use applio or mainline, however, mac can only inference (use models) locally (on ur pc), i would suggest to use cloud (remote good pc)

[twilit kernel]

How can I cloud for RVC?

Can I download Applio on mac?

[low shard]

for inference ilaria rvc zero or applio colab

For rvc training cloud you can choose between:

• Google Colabs (4 hours of daily gpu for free, not much hours, but easy to use):
  • Applio (ui)
  • Mainline (UI)
  • RVCDISCONNECTED (no 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)
  • Applio by Shirou (UI, no guide as of right now)
• Lightning.ai (Kinda hard, needs login, no issue with web uis or anything, but only free 15 credits monthly):
  • Mainline (UI, No guide as of right now)
  • Applio (UI, No guide as of right now)

[twilit kernel]

Doees it work on Mac?

[rare goblet]

Ayo? @twilit kernel level 1 !!!

[low shard]

if u wanna do it locally, theres a mac installation: https://docs.applio.org/getting-started/installation

But i Would HIGHLY suggest to just use cloud

[low shard]

yes, its cloud, it doesnt run on ur pc, it runs on a remote good pc

[twilit kernel]

And it makes good results?

[low shard]

yes, rvc is the best Speech To Speech program

its used by like 90% of ai covers

[twilit kernel]

Thanks I'll try

[low shard]

yw

[twilit kernel]

the cloud version works as good as localy?

[low shard]

the performance of the local one depends on ur pc, and cloud will run better than your mac

in terms of quality: yes, its just the same program

[twilit kernel]

even if I have m1 pro?

[low shard]

yea, an A100 (The ZeroGPU used by Ilaria RVC Zero) is way better than that

[twilit kernel]

thanks

[low shard]

those use AI Specialized gpus, like T4, A100, etc

Yw

[low shard]

for training id suggest Kaggle as it gives the most gpu time so u wont lose ur work

For inference i would suggest ilaria rvc zero

[feral plank]

Is there a way I can do this on mobile

[low shard]

Inference (use models) or train (make models)

[feral plank]

i don’t get it sorry

[low shard]

are you trying to use models for pre-recorded audios like ai covers, or make models?

or are you trying to use modles in realtime for voice changing in calls?

[feral plank]

use models

[low shard]

for pre-recorded audios right?

[feral plank]

yeah, i’m trying to use a donald duck voice for some audios i found but the files won’t work, unless it’s not for mobile to do that

[rare goblet]

Ayo? @feral plank level 1 !!!

[low shard]

idk what ur using, but this is RVC Technology

You could technically do it locally on ur phone but its on CPU so slow and not suggested

Its way better u use cloud (remote good pc), use ilaria rvc zero

[radiant loom]

how many epochs should i train my IA model?

[rare goblet]

Ayo? @radiant loom level 1 !!!

[radiant loom]

I have 9:44 min of audio

[keen pollen]

Hey, i downloaded my voice model and i cant find files, pls dm me

[low shard]

there isn't a right amount of epochs, see https://docs.ai-hub.wtf/rvc/resources/epochs--tensorboard/

[keen pollen]

Can you dm me?

[low shard]

its better to ask here than dms

[keen pollen]

I cant send pic here

[rare goblet]

Ayo? @keen pollen level 1 !!!

[low shard]

how did u train it? u sure u downloaded the .pth and .index?

[keen pollen]

Nvm

now i can xD

[keen pollen]

I downloaded folder

[low shard]

Oh, sorry but i cant help much about local

[radiant loom]

tysn

[low shard]

I don't do things locally, i use cloud

I would suggest using a more updated version like mainline or applio tho

[low shard]

maybe this could help https://docs.ai-hub.wtf/rvc/local/mangio/#15-gather-models-files

[radiant loom]

@low shard can u help me installing this app?

idk what im doing wrong

[low shard]

are you doing it locally or using google colab?

[radiant loom]

locally

or idk

really

[low shard]

did you download mangio rvc on ur pc?

[radiant loom]

i use mangio rvc

ye

[low shard]

first, whats ur pc gpu?

[radiant loom]

2060

im cooked?

[simple ore]

you ran tensorboard without any logs present?

[radiant loom]

idkl what are u talking abt

im noob

[sullen jungle]

Ive got these 2 models, both v2 and 40k, but it keeps saying they are different versions

[rare goblet]

Ayo? @sullen jungle level 2 !!!

[brittle wing]

How do I make my ai vocals sound more realistic?Like what kind of lowpass/highpass filter or settings do I use

https://vm.tiktok.com/ZGdJph7aE/
Like this video

[frail plank]

you guys probably get this all the time

but how do you make voice models

[brittle wing]

Rvc

Chat is there a website version for it frfr

[brittle wing]

could anyone help im not hearing any outputs for the voice changer (rvc google colab) but I am for just regulaur in discord

[prisma carbon]

how do i use a voice model 😦

[brittle wing]

What system are you using

[rare goblet]

Ayo? @brittle wing level 1 !!!

[keen stratus]

how to make these settings:
Epoch: 620
Steps: 9000+
Pretrain: Snowie V3

in voice changer there is only: Gain, pitch, index, chunk and extra

[fading bone]

when your using applio and you finish training a model. how do i export the pth and the index file to my downloads folder or google drive

[spice owl]

-local

[azure marsh]

🖥️ Local stuff

• 🍏 Applio, by IA Hispano GitHub
• Mangio-RVC-Fork, by Mangio621 Huggingface
• RVC Studio, by SayanoAI Huggingface
• AICoverGen, by SociallyIneptWeeb GitHub
• Replay, by Replay Team Website
• Original RVC, by the RVC-Project team GitHub
• GPT-SoVITS, by RVC-Boss GitHub

Credits to Faze Masta and Antasma for compiling these links.

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

[golden walrus]

guys, can i ask why i tried to train but it only process 1 step/epoch ?

[simple ore]

you provided unsupported audio, you did not split audio, extract features did nothing, you're training on two mute files

[golden walrus]

but i splitted it

[simple ore]

go to logs/yourmodel and see what files are there

[golden walrus]

xd i stopped that one

btw, about pretrained

should i use these with high steps or those with low point

[simple ore]

use default pretrain

[golden walrus]

i mean, default pretrain don't support my language so i tried to make my own pretrain xd

[simple ore]

you can not make a good pretrain from scratch

[golden walrus]

so is there anyway to make 1

[simple ore]

at least not without using some magical way the original one was made

I doubt the original pretrain had russian language, yet a model trained with it for just 20 steps does fine

[golden walrus]

i tried to make one with vietnamese

[simple ore]

it may take some extra source data to shape it up, but it is better to use an existing pretrain as a base

instead of trying to do it all from the scratch with 100+ hours of audio

[golden walrus]

ah so, base pretrain + train one with my desired language, then use these D and G to train another voice i want ?

[simple ore]

i mean... you can do that too

but I mean use pretrain with 30-60 min of audio in your desired language and voice

if it ends up not good enought, use D/G from it + more audio

[golden walrus]

i got 2 hours of audio

oh okay, i got it

[simple ore]

you can always buid up on top of existing model

it simply adjusts weights

even just doing 5-10 epochs on top of default pretrain you should hear your trained voice, maybe not perfectly speaking certain syllables, but close enough and training longer should fix that

[golden walrus]

-rc

-rvc

[azure marsh]

Documentation

AI HUB Docs

https://docs.aihub.wtf/

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

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

[molten relic]

Hey! I’m sorry I’m new to this AI stuff, where do I start to start learning to create a voice model? I apologize if this is an inconvenience to some, I’m very new and I’m just really want to learn! Very appreciate any help would be awesome! Thanks anyone that responses!

[low shard]

whats ur pc gpu?

[molten relic]

I’m in bed right now so unsure but how does a GPU effect training?

[low shard]

You could run the ai locally (on ur pc), meaning it runs on ur pc

like the same way u need a good gpu for games, ai takes alot of computing

especially training

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

[simple ore]

if you have a push-cart you cant move a mountain, that requires a quarry dump truck

AI training requires a tremendeous amount of number crunching with specialized hardware, you can't do it on a cheap laptop wit intergrated GPU

[simple ore]

woosh

[rare goblet]

Ayo? @simple ore level 15 !!!

[runic schooner]

i need help

[low shard]

!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.

[runic schooner]

I’m having trouble finding the right file to download it my cpu is amd ryzen 5 3600 6-core processor

[low shard]

to download what?

are you looking for realtime voice changing for calls, use models on pre recorded audios or make models

and also, i need the gpu

[runic schooner]

Rtx 2060

[rare goblet]

Ayo? @runic schooner level 1 !!!

[low shard]

alr, so what are u looking for?

[scenic arch]

is it just me or are the aihub docs down?

[low shard]

yes, check #📰│dev-updates message

[pastel oak]

Bro does NOT wanna answer questions 😭

3 times in a row lmao

[low shard]

did he also do that to u too?

😭

[pastel oak]

Nah nah just seeing u ask him for what he needs 3 times with no answer is funny

[low shard]

😭

[simple ore]

HE'S A CATFISHER I TELL YOU

[low shard]

what

[simple ore]

Hiding his nasty desires

not answering questions

naughty naughty

[pastel oak]

SMH

EXPOSE HIM

[low shard]

REAL

[low shard]

BAN HIM FOR CATFISHING

[molten relic]

Hey nick! / anyone that reads this for this purpose i have a NVIDIA Geforce RTX 3050 Ti

[knotty moth]

laptop 4 GB variant? nope, not recommended for training, only inference

[molten relic]

32 GB Installed Ram

[molten relic]

Thanks for responding fast

[knotty moth]

no

system ram is irrelevant

https://tenor.com/view/facepalm-gif-4576513125411549651

[molten relic]

I am unsure what you are asking my apologies, im really new to some stuff like this. I dont want for you to waste your time though! I tottaly understand being busy or if im to new to understand a lot!

[knotty moth]

I'm also busy playing around with SD, but not an excuse to not respond to you

[molten relic]

Ah thats okay!

Let me know what you need from me and ill figure it out and give it to you!

[low shard]

the memory of the gpu

its suggested to have 8 or more gb of memory of gpu aka vram for training

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

you should also be able to find the gpu memory

Also is it a desktop or laptop ?

[molten relic]

12 GB

[knotty moth]

u prob mean desktop 3060

[low shard]

12gb vram should be good

[molten relic]

Techically a Desktop but i shoved some parts from a few laptops nto it

[low shard]

alr, rtx 3060 12gb is good

for both training and inference

[molten relic]

Let me know where to start and im ready, sir nick!

and alisa!

[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, same quality tho
• Mainline: The original RVC

[molten relic]

Downloading now

[tepid atlas]

---

JSONDecodeError Traceback (most recent call last)
<ipython-input-6-75abb3770c40> in <cell line: 31>()
31 if os.path.exists(config_path):
32 # File exists, proceed with creation of creds and client
---> 33 creds = Credentials.from_service_account_file(config_path, scopes=scope)
34 client = gspread.authorize(creds)
35 else:

5 frames
/usr/lib/python3.10/json/decoder.py in raw_decode(self, s, idx)
353 obj, end = self.scan_once(s, idx)
354 except StopIteration as err:
--> 355 raise JSONDecodeError("Expecting value", s, err.value) from None
356 return obj, end

JSONDecodeError: Expecting value: line 1 column 1 (char 0)

uh help

[low shard]

Don't use outdated colabs

What's ur PC GPU?

Yt tuts are outdated

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

[tepid atlas]

Intel(R) UHD Graphics 630 and AMD Radeon RX 6600 XT

problem 2: python3: can't open file '/content/infer-web.py': [Errno 2] No such file or directory

[low shard]

you are using an outdated colab

[low shard]

You could be able to use RVC Locally (on ur pc) via: https://docs.applio.org/getting-started/installation#amd-gpu-support-windows with the AMD GPU

Google Colab is a Cloud Computing service (remote good PC), so used for weak PC

Your pc should be able to handle it

Btw, you are looking for using models for pre-recorded audios, or making models, or using models in realtime for voice changing in calls/games?

[low shard]

@tepid atlas bc the one i sent is applio, an rvc fork (modified version) for making and using models for pre-recorded audios

for realtime voice changing for calls theres another program

[tepid atlas]

okay

[tepid atlas]

both pre-recorded audios and making models

[low shard]

yea then stick with Applio locally

[tepid atlas]

okay

[rare goblet]

Ayo? @tepid atlas level 1 !!!

[tepid atlas]

[rugged solar]

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• 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
• Modified W-Okada's Voice Changer, Google Colab
• 🆕 FaceFusion UI, by Nick088 Google Colab
• 🆕 FaceFusion NO UI, by Nick088 Google Colab
• 🆕 EasyGUI, by Rejekts 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, -overtraining, -realtime, -rvc, -help

[rancid heron]

What happens if you run inference with a different pitch extraction method than the model was trained on?

[simple ore]

does not matter

it is just a method, the range of values is about the same

[brittle wing]

Error

[delicate oak]

Ilaria rvc doesn't work anymore ?

[polar plaza]

Can someone please tell me why this keeps happening

[delicate oak]

Same

[knotty moth]

but other zerogpu spaces worked on you?

[molten relic]

I have a problem that showed up when i started the Train Model Button

~~

Any ideas or anything would be highly apperciative! Thanks so far for everyones thats helped me so far! I am just sorta bad at this stuff lol

[pastel oak]

Put the RVC1006 folder somewhere outside of OneDrive

Buuut tbh I dont see any error messages unless I'm blind

Epoch 1 started training, how long did you wait before sending this txt file?

[molten relic]

About a half a hour ish

I let it do its thing and made lunch came back and it came to that

Nothing was moving so I was worried

[pastel oak]

Would wait for someone else to comment on it then, but would still move it out of OneDrive just in case

[brittle wing]

Hey guys i was adding a new voice model but have a CKPT file what do i do with this?

[rare goblet]

Ayo? @brittle wing level 3 !!!

[brittle wing]

I got the path file and CKPT but no index, what do i do? 🤔

[low shard]

check: https://docs.google.com/document/d/1YbXcLFPaGjhOdG5NFkK3QrucCEpHZBwFUxkeMO8aB18/edit#heading=h.mfxmgfqiaevu

[low shard]

check https://docs.google.com/document/d/1YbXcLFPaGjhOdG5NFkK3QrucCEpHZBwFUxkeMO8aB18/edit#heading=h.mfxmgfqiaevu

[low shard]

its a limitation of the inference time that it takes for converting in ZeroGPU huggingface spaces

i explained it above better in the guide

@proven hill can't u put the limit back to 300s (5 min) instead of 1 min on Ilaria RVC Zero?

[molten relic]

Hey Nick any ideas for my issue?

[molten relic]

I just want another opinion on it as it could be just the OneDrive thing but just checking as they are unsure

[low shard]

I don't do local sorry

[timid olive]

In some parts of the songs, there are backing vocals; should they be separated?

You can listen to the sample sound below; it is only the vocal.

This example is from a single song.

Do you think I should separate them? Some songs sound different to me.

[low shard]

Yea

[rare goblet]

Ayo? @low shard level 108 !!!

[low shard]

With HP KARAOKE 6 of UVR

[timid olive]

First, I processed it through this model, then through the other model, and finally used this model to delete the reverb.

Do you think I did it correctly? Will it be of good quality? Also, some parts sound different to me; for example, the audio.

It sounds like the same song but with a different sound. Will it cause problems in training?

I’m curious about this.

[molten relic]

what do you do then?

[rare goblet]

Ayo? @molten relic level 4 !!!

[knotty moth]

I have successfully inferred ~3 mins audio, though the GPU task aborted issue also sometimes occured, but I think it should be still less demanding than even Flux1.Dev generation, etc

[molten relic]

~~

Slightly differnt then last time but still no generation, not in a onedrive this time ago, anyone have simular problems or know how to fix?

[simple ore]

says not in a onedrive, yet "C:\Users\storm\OneDrive"

but nothing on the is an 'error'

[molten relic]

Huhhhhh? I switched the data set onto a external hard drive

[simple ore]

logs and everything else is still on onedrive?

[knotty moth]

uninstall onedrive

[molten relic]

Son of a nugget I just got this and in bed I’ll resume tomarrow.

Would anything bad happen if I do so?

[knotty moth]

what

if you don't need that bloatware, why not

[molten relic]

I don’t know I thought was it something I needed

[fleet cargo]

what should i use

for text to speech?

not okada i suppose?

@knotty moth

[low shard]

The limitation isn't about your audio file length, it's about how much time it takes to inference that audio

If it takes more than 1 min to inference that audio, it gives task GPU aborted

[low shard]

i cant listen to it rn

[low shard]

I do Cloud, i got a bad pc so i use a remote good pc

[low shard]

wokada is for speech to speech, if u want realtime text to speech, u can look at https://docs.google.com/document/d/12hCYJqNCFl6jWKoVvCxtwt2V6nSoilgi5La8dkZa1KY/edit#heading=h.xweoq2pdv4uj or use the tts client https://github.com/w-okada/ttsclient (but cant really help for the second one)

[molten relic]

What’s cloud?

[low shard]

remote good pc, basically i run theprogram in a cloud computing service like google colab, kaggle or lightning.ai, instead of running them on my bad pc

[molten relic]

Interesting

[low shard]

its way better to use it locally on a good pc but i got integrated graphics

[dusk tulip]

.

[pastel oak]

.

[jaunty shale]

https://docs.aihub.wtf/ doesn't work for me. Trying to find RVC Disconnected guide so I can re-learn stuff again.

does anyone have link?

[pastel oak]

#📰│dev-updates

[jaunty shale]

thank you so much! I appreciate it

[lucid cove]

could anyone help me i got screenshots

[languid lotus]

i need help with rvc can anyone help me

[jagged hawk]

@low shard i finally created a voice model, how can i upload it in the channel voice-models?

[low shard]

!howtoask

[patent trellis]

How To Troubleshoot

__**GIVE CONTEXT.**__ 📝

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

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

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

__**BE PRODUCTIVE.**__ 🤝

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

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

[low shard]

check https://docs.ai-hub.wtf/extra/model-maker-role/

[noble dawn]

Anyone can help me train on local ? Rq

If they free

[quasi lynx]

Whatchu need?

[noble dawn]

It’s giving me an error whenever I click train

Idk why

When I got a 3060Ti gpu nevida

😭

[quasi lynx]

Might help if you send the error

[noble dawn]

Yes

I got u ima dm u

Ty

Let me screen shot rq

[remote trellis]

Hi

When I log in with new docs, the main menu opens but when I click on the app or any other button, it doesn't work. Why?

@timid olive

[pastel oak]

its not fixed yet

[remote trellis]

So how do I make an AI cover with the applio link?

@pastel oak

[pastel oak]

Are you not talking about the docs guide? the normal applio stuff should still work

[remote trellis]

Not

Can you send me the collab link?

For aı cover

[pastel oak]

https://colab.research.google.com/github/iahispano/applio/blob/master/assets/Applio.ipynb

[remote trellis]

Thanks

@pastel oak @pastel oak

Look, now I opened the "applio" and I couldn't download the sound I wanted. I paste the sound I wanted into this doenload, it says it downloaded in 1 second but it doesn't download, why? Sound model: https://applio.org/models?id=1218683186431660072

[rare goblet]

Ayo? @remote trellis level 2 !!!

[pastel oak]

I dont know applio colab

[remote trellis]

This

[pastel oak]

thats not what i meant

i cant help with applio colab

[remote trellis]

Or

[pastel oak]

try ilaria rvc zero

[brave garnet]

Ilaria RVC: CLICK HERE 🤗

Guide on how to use it: CLICK HERE 📝

Don't forget to thank Ilaria if you find it useful! 💖

[pastel oak]

download model from applio and manual upload it

[remote trellis]

So where can I find the hugging face of this applio voice model?

[remote trellis]

didnt worl

Worl

Work

[pastel oak]

Did you even read what i said

Modeli applio'dan BİLGİSAYARINA indir
Ilaria RVC Zero'yu aç
"Model Loader" sekmesine git
.pth ve .index dosyasını yükle

[remote trellis]

im mobile

@pastel oak

[pastel oak]

Dont know

[brave garnet]

RVC Colabs and Spaces

⠀

Local Forks 🖥️

⠀
Mainline RVC
Original project, suggested for advanced users,
by the RVC-Project team.

Applio
Simplified, suggested for all, by the Applio team.

RVC Studio
Simplified, suggested for all, by SayanoAI.

Mangio-RVC
Simplified, may not be supported anymore, by Mangio621.

AICoverGen
Simple yet great way to make covers, by SociallyIneptWeeb.

Replay
From the greators of weights.gg, excellent product for everyone.
⠀

Made with love by Ilaria 💖

[next wharf]

what is the difference between FCPE and rmvpe?

[low shard]

Fcpe is faster, rmvpe is better quality

[next wharf]

thanks

[low shard]

First of all. What's ur PC GPU?

[low shard]

its fixed now btw

[low shard]

fixed it https://github.com/AIHubCentral/docs/pull/7#event-14585577515

[pastel oak]

Boss

[low shard]

there were 256 broke hyperlinks

just fixed it with search and replace all yk lol

[stark wadi]

I can't get applio to install on mac. It keeps saying that a java runtime wasn't found. Anyone know how to fix this?

[rare goblet]

Ayo? @stark wadi level 4 !!!

[brave garnet]

Voicechanger Settings (Okada)

⠀

Settings for Nvidia GPUs

F0 Det.: rmvpe (suggested for all series)

RTX 40-series: 80-96 chunk | +16384 extra
RTX 30-series: 96-112 chunk | +16384 extra
RTX 20-series: 112-128 chunk | +16384 extra
GTX 16-series: 128-192 chunk | +8192 extra
GTX 10-series: 128-192 chunk | +8192 extra

Advanced Settings

Protocol : Sio or Rest
Crossfade: 4096 start 0.2 end 0.8
Trancate: 300
Silencefront: Off
Protect: 0.5
RVC Quality: Low

⠀

Made with love by Ilaria 💖

[brave garnet]

Voicechanger Settings (Okada)

⠀

Settings for AMD GPUs

Don't forget that your models needs to be converted in ONNX!

F0 Det.: rmvpe_onnx (suggested for all series)

7xxx XT cards: 112-128 chunk | +16384 extra
6xxx XT cards: 128-192 chunk | +16384 extra
5xxx XT cards: 192-256 chunk | +8192 extra

RX 580: 192-256 chunk | +8192 extra
RX 570: 192-256 chunk | +8192 extra
RX 560: 256-384 chunk | +8192 extra

Advanced Settings

Protocol : Sio or Rest
Crossfade: 4096 start 0.2 end 0.8
Trancate: 300
Silencefront: Off
Protect: 0.5
RVC Quality: Low

⠀

Made with love by Ilaria 💖

[languid lotus]

what is the best model pre train please awnser fast

[pastel oak]

Original

[languid lotus]

ok thank you

[patent quarry]

How do I transfer to ONNX?

[mighty vortex]

why can't I mount drive on the colab?

[low shard]

What Google colab are you using and what's the issue specifically

[mighty vortex]

i'm using disconnected, the error is "credential propagation was unsuccessful"

[low shard]

Can you send the Google colab link?

You need to be sure also to always allow the Google drive when you get the popup

[mighty vortex]

https://colab.research.google.com/drive/1XIPCP9ken63S7M6b5ui1b36Cs17sP-NS#scrollTo=ZodNcumpg-JM

[rare goblet]

Ayo? @mighty vortex level 1 !!!

[mighty vortex]

I do allow

[low shard]

Yeah that seems the right colab

Try re running it and give it access again

[mighty vortex]

i've done that multiple times already

[prisma grove]

do you guys know how long would it take to train an rvc model locally

compared to training on colab

[pastel oak]

Depends on gpu

[prisma grove]

rtx 2080ti

[pastel oak]

Thats probably faster than the colab gpu

[prisma grove]

colab's got t4 tho

[low shard]

lemme check it myself rq

[pastel oak]

Tesla T4 is worse than 2080ti

[prisma grove]

oh

I didn't know that

cool

[low shard]

just tried it myself, i ran the cell, it asks me permission, then i choose google account and allow everything

works fine no issue

be sure to not modify its permissions

[prisma grove]

btw what should be the dataset zip structure?
something like
| name of the dataset.zip
| | speaker0
| | | audio files
?

or just
| name of the dataset.zip
| | audio files

or should I not even zip it for local training? 😵‍💫

[pastel oak]

Dont zip, put in a random folder and select the folder path in rvc

[mighty vortex]

how many epochs should you train a model with a 3 minute dataset?

[pastel oak]

Everything is measured by tensorboard graphs

[mighty vortex]

but is there a general amount?

[pastel oak]

No, every run, every dataset is unique

[prisma grove]

I found this though? https://www.desmos.com/calculator/yeqx4dmcfm?lang=pl

[rare goblet]

Ayo? @prisma grove level 2 !!!

[prisma grove]

honestly I'm confused now, the calculator says that 20 minute dataset is about 100 epochs

but that's very low

what does the loss stuff mean?

and why is each epoch taking so long to train?

is 1 epoch per minute the normal speed? I haven't used rvc in a while

and I don't think I've ever trained a model locally

[tropic plover]

Like Shad said, u have to rely on Tensorboard to determine when to stop and which epoch to pick

[prisma grove]

where is tensorboard

[tropic plover]

Found this here some time ago
https://drive.google.com/drive/folders/1o1ZZuUHQ6MuclA6B-AtQZlHlC5Uf34pb

[prisma grove]

I don't understand anything in that tutorial

I'm just gonna test all 7 models (700 epochs, saving every 100) and see which one sound the best

what's more concerning to me is how long it takes to train every epoch

what is the normal speed?

[tropic plover]

It all depends on ur GPU and batch size, dataset size, etc. My 4060 on 4 batch size, 30min dataset is taking abt 2:00 per epoch

[low shard]

Idk about local but check https://docs.ai-hub.wtf/rvc/resources/epochs-tensorboard/

[stark wadi]

does anyone know how to install on mac m3 pro?

[pastel oak]

Go to the docs guide Nick linked, download the file, put it in the same rvc folder where you launch it, run it, thats all

It opens another webui

[simple ore]

saving every 100 epochs is silly

go every 10

[prisma grove]

I don't have that much free space on my drive

[simple ore]

unless you got hours and hours of sample audio files, using 700 epochs is crazy

[prisma grove]

then why are people doing 700 epochs for 5 minute datasets 😵‍💫

[simple ore]

they are stupid

or they follow a stupid guide

running 700 epochs on 5 minute file is trying to squeese a gallon of juice from one lemon

you can get all you can from 5 minute file in 20-50 epochs

[prisma grove]

I have 18:50 long dataset

as in 18 minutes 50 seconds

[simple ore]

should be under 200 epochs at most

again, use tensorboard to check

[prisma grove]

yeah I got it now

[pastel oak]

You can delete the first 50 epochs safely while saving every 10 epochs, then delete as you go if you hit another OT point etc

[prisma grove]

another OT point?

shouldn't I stop the training when it starts to OT?

also it's saving every 50 epochs anyway lol

I think the interval is limited to 50

[serene horizon]

I have tried using this but it says “You cannot currently connect to a GPU due to usage limits in Colab” for the last two days.

What can I do?

[low shard]

mac can only inference (use models) locally (on ur pc) not train (make models)

i dont have a mac but u could check https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/blob/main/docs/en/README.en.md

tbh id suggest to just use cloud (remote good pc)

[simple ore]

wait until tomorrow?

or whenever the limit expires

[analog obsidian]

its 2024 and some people still believes 1000 epochs are better despite not being true at all lmao
just use tensorboard and select the lowest point in your g/total graph, for a 18 minute dataset your model should be done in less than 200 epochs

[serene horizon]

No idea on how long the limit is there?

[low shard]

you can:

• Use an alt google account
• Use kaggle which gives more gpu time but its harder
• Wait until tomorrow
• Pay for colab pro

btw, whats ur pc gpu?

[simple ore]

the free usage period goes down the more you use it

and slowly resets back when you dont

[knotty moth]

imagine the graph keeps going down on above 1000 epochs for some short dataset

[low shard]

which is why kaggle is on top

lightning.ai is also cool but lower limits (as in gpu time) so

[analog obsidian]

batch 16 in 4 minute datasets be like

[serene horizon]

I tried registering another Google account but Google doesn’t allow me.

[low shard]

You mean like it asks u for phone verification when u make another google acc?

[serene horizon]

Yeah.

[low shard]

also its WAY BETTER to use Kaggle

[low shard]

you can use the Phone Gmail app to make another google acc without phone verification

ofc u need a phone tho

just like open the app, make a new acc and it will do it without needing any verification

but i suggest u way better to just use kaggle, its a bit harder and needs just a single phone verification but gives u 30 hours weekly (yes they refresh)

its WAYYY better than google colab

and u dont have the risk to losing ur stuff for randomly getting disconnected as 30 hours are alot for free

[serene horizon]

I actually did register another account on my phone, but later, when I went to log in on my laptop, it asked me to verify the account with a phone number! 🤦‍♂️

[analog obsidian]

kaggle is a bit buggy but works good when decides not to randomly end the session

[rare goblet]

Ayo? @serene horizon level 3 !!!

[low shard]

since when does it randomly end the session? Never happened to me

be sure to use encryption

[low shard]

wtf very weird

u should be able to login on ur pc of the same acc made on ur phone without phone verification

at this point i suggest u to use kaggle or wait

don’t u have even just 1 phone number ?

[analog obsidian]

it does for me lols, everything set up correctly then randomly decides to stop the session when im downloading the dependencies
i fix this by creating another version of the notebook

[low shard]

skill issue ngl

never happened to me

nor heard it from others

[analog obsidian]

[low shard]

[rugged solar]

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• 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
• Modified W-Okada's Voice Changer, Google Colab
• 🆕 FaceFusion UI, by Nick088 Google Colab
• 🆕 FaceFusion NO UI, by Nick088 Google Colab
• 🆕 EasyGUI, by Rejekts 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, -overtraining, -realtime, -rvc, -help

[serene horizon]

I do, but i already used it so it won’t accept it.

[low shard]

dw, kaggle is a different service than google colab

use kaggle, u will be able to use that phone number

As you dont got a good PC, its better you use cloud (remote good pc) for training an RVC Voice Model:

• Google Colabs (4 hours of daily gpu for free, not much hours, but easy to use):
  • Applio (ui)
  • Mainline (UI)
  • RVCDISCONNECTED (no 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)
  • Applio by Shirou (UI, no guide as of right now)
• Lightning.ai (Kinda hard, needs login, no issue with web uis or anything, but only free 15 credits monthly):
  • Mainline (UI, No guide as of right now)
  • Applio (UI, No guide as of right now)

here i sent all the cloud ways, use mainline kaggle

[serene horizon]

Okay, I’ll have a look and see if I can work it out!

[low shard]

i sent u the guide with the kaggle too

click on mainline from kaggle

[serene horizon]

Thanks. I’ll give it a go now.

[low shard]

yw

[languid lotus]

can some one tell me if this model is good or not and how to know from this pic

its still in 150 ecpoh

18 minuts of training data

[prisma grove]

is this OT?

[rare goblet]

Ayo? @prisma grove level 3 !!!

[simple ore]

what batch size?

[languid lotus]

what about this

[simple ore]

it tells nothing without the metrics

[languid lotus]

what are they and how do i get them

[simple ore]

run tensorboard

see scalars

[simple ore]

use smoothing, like 0.975 or a bit more

metrics loss - d/total, g/total, fm, mel, kl

[rare goblet]

Ayo? @languid lotus level 2 !!!

[languid lotus]

what do they say

[simple ore]

not grad

[languid lotus]

what do they mean

[simple ore]

just terible

[languid lotus]

what

wdym

how can i make it better

[simple ore]

what is the size of ther training set?

[languid lotus]

18 minutes

[simple ore]

you seem to have some terrible quality then

[languid lotus]

do you want me to send you the drive link for the audio

im making a billie eilish one

https://drive.google.com/file/d/1tQgSnHJ1isRzrTfRQX4newovWE5YJcHG/view?usp=sharing

bro i used very high quality wdym

@simple ore

[simple ore]

did you cut out all the silence?

[languid lotus]

yes using audacity

[simple ore]

that and there's some slight weird band over 16khz

[languid lotus]

i dont know how to do all these stuff

[simple ore]

the model needs to have silence gaps in order to learn a separation

[languid lotus]

so i dont trim silince

??

[simple ore]

you dont need to

unless you plan to replace Eminem voice in his rap

[languid lotus]

so if i send you the raw voice will you make me a training data

just to know what ive been doing wrong

[simple ore]

just take your original interview

[serene horizon]

Why does it say in the guide to truncate silence then?

[simple ore]

before you cut the silence gaps out

[languid lotus]

ill try

[simple ore]

the training specifically inject a couple of mute audios for the model to train how to reproduce silence

but it is only 2 3sec files

[languid lotus]

yeah another question do i need to make my voices cut or whole

[simple ore]

you need to have only the voice of the target person

[languid lotus]

if i do what app do i do it with

[simple ore]

obviously

[languid lotus]

no like one audio file or multipile

[serene horizon]

Is this just for Applio? What if I use it with another RVC trainer?

[simple ore]

here's a model i'm testing right now, there's no weird jumps or craziness

and it is only 10 minute set

[simple ore]

RVC mainline does the same

[languid lotus]

and what pre train do i use original or titan or ov2rsuper

[rare goblet]

Ayo? @languid lotus level 3 !!!

[simple ore]

I think without the silence gaps the audio becomes too complex to learn

I use original

[serene horizon]

But not RVC disconnect in Colab?

[simple ore]

probably the same.. all the base code should be very similar

[serene horizon]

Okay.

[simple ore]

but I've seen some projects where they do not include silence for some reason

[languid lotus]

bro if i want to make a rvc model of 21 but his voice is bad when i uvr look

@simple ore

[simple ore]

yeah, has some echo and another voice blended in

not good

[languid lotus]

any tips to make it better

[simple ore]

you cant unbake a cake

[languid lotus]

thats true unfortunately

[serene horizon]

I got to the training point, but when I click train, it says “error.” 🤦‍♂️

But in Kaggle, I see it working.

So weird.

[prisma grove]

5

I left it at default

it's at epoch 266

should I stop it or no

[simple ore]

5???

look at the other charts

[prisma grove]

you mean this right?

[simple ore]

that's not right

[low shard]

show the error in the kaggle output

[simple ore]

it has very little to do with GPU memory size

[prisma grove]

which chart

[simple ore]

last page, with fm, mel, kl

[serene horizon]

Where would I find that?

[simple ore]

fm metric is weird and high

[prisma grove]

wdym

hold on let me check what the model sounds like rn

[low shard]

in the kaggle site, where it shows the output of the cell u are running

[simple ore]

the metric is not going down/not stabilizing

but you can check how it sounds

[serene horizon]

There’s a lot of text there.

It’s currently telling me what epoch it’s on.

[simple ore]

g/total is kinda high, but that's probably of the batch size

[low shard]

there should be an 'traceback' part

show a screenshot

[analog obsidian]

what causes fm to go up anyways? i've never seen a fm graph that goes down lol

[prisma grove]

sounds pretty normal?

[simple ore]

depending on the learning ratio, the model may overshoot the target, so fm goes up and down

[prisma grove]

I had models that sound like static so

[simple ore]

large variation of data in the set may result in fm going up and down

[analog obsidian]

do you think is good to change the learning rate everytime we train a dataset? last time we spoke about this we conclude is better to not change the default lr (which is 1e-4 i believe)

[serene horizon]

First, I’m going to try from the start again. Maybe I missed something. I

[simple ore]

you can probbaly set it to 5e-5 (half of default), the training may take longer

[prisma grove]

?

set what

[analog obsidian]

this could prevent fm to overfit faster?

[simple ore]

possibly... also using FP32 may prevent it too

[analog obsidian]

thanks! i'll try it

[prisma grove]

this is what the 150e ckpt sounds like

this is all with no index because I'm too lazy to grab it

[simple ore]

technically the metrics should go down or at least stabilize due to the learning rate automatically adjusting down

[analog obsidian]

yeah all of the metrics go down EXCEPT fm for some reason

[prisma grove]

so what does that mean 😭

[analog obsidian]

ive always believe this happens because the dataset is small but im pretty sure fm goes up even in big datasets

[simple ore]

neshi'zzzzzz'te

that's what I hear

at 7 seconds

[prisma grove]

nen ne shite

it's genkotsu yama no tanuki san

[simple ore]

this part

[prisma grove]

I mean yes

[analog obsidian]

sibilances are artifacting

[prisma grove]

the sh sounds weird

tf is a sibilance

[analog obsidian]

SH ch K sounds

and S ofc

u can decrease the artifacting by de-essing the dataset

[prisma grove]

it's already mostly de-essed

also it doesn't do that at 50e

or 250e

[analog obsidian]

because fm is not overfitted

[prisma grove]

what does it mean

what is fm

and what does overfitted mean

[analog obsidian]

type fm in the tensorboard, is a metric

[analog obsidian]

model confuses and starts to learn the same patterns over and over again

[prisma grove]

this

what about it

[prisma grove]

so overtraining

[analog obsidian]

similar

[prisma grove]

not the same?

[analog obsidian]

nop

[analog obsidian]

it started to fluctuate here

so any epoch in that zone have a big chance of having broken S sounds

e50 probably is before that so is not doing it

[prisma grove]

isn't ot determined by the loss/g/total metric thing

[analog obsidian]

every metric does something
g total is your average of:
fm
mel
kl

g total raising means your model start to degrade and overtrain

[prisma grove]

it's not raising

[analog obsidian]

fm going up means model overfitted the dataset features like sibilances

[rare goblet]

Ayo? @prisma grove level 4 !!!

[analog obsidian]

yeah its fine thats what i was talking about
fm metric usually overfits very fast

[prisma grove]

so what am I supposed to do

[analog obsidian]

choose the lowest point in g/total

if u don't have the exact epoch choose the closest

[prisma grove]

it's still decreasing wdym

the lowest point is the latest epoch

[analog obsidian]

show me g/total smoothing 0 and ignore scalars off

[analog obsidian]

try that epoch

that is your lowest point

[prisma grove]

how do I know which one that is

[analog obsidian]

hover your cursor in that low point

[polar plaza]

Bruh

[prisma grove]

and?

[analog obsidian]

check the step number

then find the epoch that is that step number

[prisma grove]

14.4

[analog obsidian]

or the most closest to that step number

[analog obsidian]

so you need epoch step number 14.400k

[simple ore]

checked my logs

[prisma grove]

so that's epoch 150 😐

[analog obsidian]

this is default lr fp16?

[prisma grove]

the one you said sounds bad

[analog obsidian]

yep that is your correct epoch and has broken sibilance

[simple ore]

i dont remember, that was a small test set

[analog obsidian]

it doesn't sound bad to me

[prisma grove]

https://tenor.com/view/math-calculate-confusing-figure-out-gif-6237717

the epoch that's the best is the most broken one ??

[analog obsidian]

basically

[prisma grove]

tf you mean

[analog obsidian]

is not bad, just the sibilance are artifacting

[prisma grove]

and that makes it worse

no?

[analog obsidian]

nah, artifacting happens randomly

dont worry

is a usuable model

[prisma grove]

I don't want it to be usable I want it to be good

[analog obsidian]

hmm try this epoch and see if the sibilance are artifacting

or if they got better

[prisma grove]

that's step 8.8k, the closest I have is 9400

e100

[analog obsidian]

try that

[analog obsidian]

even in smaller datasets (5 minutes) fm should go down alongside g/total?

[analog obsidian]

way better

[simple ore]

dunno about 5 minutes, that's barely enough

[prisma grove]

so the calculator thing was right

?

[simple ore]

my small set was like 12 min?

the result was not good anyway

[analog obsidian]

yea im aware, i don't have a 10 minute dataset rn lol

[prisma grove]

https://www.desmos.com/calculator/yeqx4dmcfm?lang=pl

it says for ~20 minute dataset you do ~100 epochs

[analog obsidian]

don't use this, just use tensorboard, more easy

[prisma grove]

yeah easy af man

[analog obsidian]

smoothing 0 and scalars off helps u choosing low points

[simple ore]

well, 20 min / 100 epochs is about right, +50 maybe

[analog obsidian]

like what we did now

[simple ore]

all depends on the content

[prisma grove]

BUT 150 WAS BROKEN

[analog obsidian]

yeah epoch 100 sounds fine to me

[prisma grove]

[simple ore]

you used batch 5 instead of 4 like a weirdo

[analog obsidian]

wasnt batch 5? lol

[simple ore]

fixed

[analog obsidian]

i notice the breathings are robotic in every epoch, probably the dataset lacked breaths

[prisma grove]

THAT'S THE DEFAULT

[prisma grove]

YES BECAUSE IT'S MIKU

MIKU IS A ROBOT

FFS

[analog obsidian]

miku has breaths

[prisma grove]

WHICH ARE ROBOTIC

[analog obsidian]

it was br1 in vocaloid iirc

[simple ore]

and it is not right

[analog obsidian]

nono, miku breath samples in vocaloid sounds like that

she is one of the voicebanks that has broken breathing samples

[simple ore]

i mean batch being 5 as default

[analog obsidian]

ah yeah

[prisma grove]

why does it even matter? what does batch size even mean?

[simple ore]

how many random sets of samples it trains in parallel

[analog obsidian]

this is how rvc is gonna learn ur dataset
a bad batch size can lead to unstable training and potentially bad outcomes

[prisma grove]

so it has to be 4?

[simple ore]

for under 1hour use 4

[analog obsidian]

yeah use 4

[simple ore]

there's no speed benefit in using more

[prisma grove]

who makes 1hr datasets

[analog obsidian]

good question

[prisma grove]

you can make decent voices with 5 minutes, I thought my 19 minutes was overkill

[analog obsidian]

welp model quality is tied with dataset quality, so a 5 minute high quality dataset is gonna sound high quality, just unnatural compared to bigger datasets

[prisma grove]

miku is not gonna sound natural ever 😭

[analog obsidian]

miku should never sound natural

[prisma grove]

wdym by unnatural anyway

like with the s's?

[low shard]

alright

[analog obsidian]

yeah and that smaller datasets are always gonna sound like a rvc model rather than a real human

[prisma grove]

what do you consider a small dataset

[low shard]

refresh it

[prisma grove]

under 10 minutes? under 30? under an hour 😭 ??

[analog obsidian]

under 30

it starts to get realistic at over 30 minutes

[prisma grove]

how would that affect miku

[analog obsidian]

if u add more miku samples, she's just gonna sound like her exported vocaloid vocals

like no one is gonna tell is rvc

[prisma grove]

so worse?

[analog obsidian]

nope, good
you can't make her realistic with more minutes, you can only make her sound if the vocals were made in vocaloid rather than rvc

(which is why people prefer to just use vocaloid and not rvc)

for miku

[prisma grove]

I want to make realistic miku

[simple ore]

hmm... i wonder

gimme a sec

[prisma grove]

cause it's also a matter of tuning, note transitions and all

[analog obsidian]

technically she can sound more realistic because with more minutes rvc can do more precise pitch changes

she's not gonna sound like a human but rather like a very well tuned vocaloid exported

[prisma grove]

hm

[rare goblet]

Ayo? @prisma grove level 5 !!!

[prisma grove]

so how much should I do

30 minutes?

[analog obsidian]

30 yeah

[prisma grove]

okay then

I'll try

[analog obsidian]

your miku samples are her singing random vsqx files?

[simple ore]

is there a source of this Miki voice?

[analog obsidian]

she cost real money (unless.... yw)

[simple ore]

i just need 30 seconds

[analog obsidian]

i dont have her installed srry, but if u want any audio then check this https://www.youtube.com/watch?v=swqbfMh467A

[prisma grove]

random vsq/vsqx/vpr with different voicebanks (v3, english, v4, chinese) and also stuff ripped from project diva megamix, de-reverb'ed if needed

[analog obsidian]

only use her japanese one

[prisma grove]

y

[simple ore]

gime me a wav

or mp3

[analog obsidian]

use actual vocaloid exports and nothing isolated

[prisma grove]

I can't do tuning

so I wanted to include other people's tuning too

[analog obsidian]

don't worry just let the default tuning

anything that comes with the vsqx

[prisma grove]

but why only japanese

[analog obsidian]

she'll be able to sing any language in rvc despite being trained only in japanese

[prisma grove]

no

[analog obsidian]

yep

[prisma grove]

I mean yes but some sounds are gonna sound wrong

like the english "r"

[analog obsidian]

you can do 2 rvc models, one with her japanese model and the other with the english one

[prisma grove]

what 4

[analog obsidian]

bootleg miku

[prisma grove]

why can't I just stuff it together

[analog obsidian]

rvc learns better if there's consistency

[simple ore]

quality is a bad because I used @prisma grove's song

[analog obsidian]

so if every sample is in japanese is better

than japanese + english + chinese

[prisma grove]

it is consistent because it's still miku

[analog obsidian]

yes but pronunciation changes

its up to you anyways

[prisma grove]

wouldn't that be good? make it able to pronounce more stuff accurately?

[analog obsidian]

yes but is also going to get confused sometimes

and pronunciation might be worse

[prisma grove]

how would it get confused?

it's gonna pronounce stuff like the source audio does

[analog obsidian]

is not going to know correctly which pronunciation will use

[prisma grove]

if the source audio pronounces it right then it will too, no?

[analog obsidian]

hit or miss

but higher chance of better pronunciation if all of the dataset has the same language

and that the inferenced audio is also the same langauge present in the dataset

[prisma grove]

well it's still 90% japanese

[analog obsidian]

so with this info you know the model is better at japanese than the rest

and also the model has a bias of japanese pronunciation

[prisma grove]

doesn't rvc work by just picking whatever's the closest

I don't get it

[analog obsidian]

index 0 can help avoiding this bias

but not always

[simple ore]

index blends original features from the audio and features of the voice model

by mapping original to voice model

[analog obsidian]

if she has a bias of prononucing "la" like "ra" then she has a 99% chance of doing this even with index 0

[simple ore]

um.. no

[analog obsidian]

index just forces that to always happen

[simple ore]

that's not how the index works

[analog obsidian]

doesnt force the dataset samples over the audio features?

tbh no one explained what index is

[prisma grove]

oh my god

if the model is 90% japanese

[analog obsidian]

bro just use your model if u like it

[prisma grove]

then if you convert an audio of someone saying the word "Carrot", it's not gonna suddenly say "カロット"

[analog obsidian]

already sounds good for me

[analog obsidian]

ofc no

try inferencing more audio and see if u like the results

at the end of the day what matters is if you like the model

[simple ore]

again, that's now how index works.... it take source audio feature, tried to look up something close enough from the voice model features

then it blends original and voice in selected ratio

[analog obsidian]

finally someone explains what index is lols

[simple ore]

english audio + french speaker at 0 index has minimal accent, the accent comes in full force when you use index 1

[analog obsidian]

most of the stuff i learned was from trial and error

[simple ore]

#🔊│ai-development message

there you can check

I made a test

[analog obsidian]

yea same i noticed this too

[analog obsidian]

i see makes sense now
sucks that when we are starting doing rvc models there's no info about anything in the internet

how are we gonna know what the metrics are? you go to the official rvc github site and there's nothing that tells u what even g total is

😭 its like only a couple of people actually know how this thing actually works

[prisma grove]

this is the 90% japanese 100e model

it's pronouncing it better than I can with my shitty polish accent lol

[simple ore]

dude./.. give me some good wav you're using for training

[analog obsidian]

xD

[prisma grove]

like, you want a wav of miku's singing converted through miku rvc?

[analog obsidian]

he wants a sample of your dataset

any wav

an audio that u used for training