✨│ai-help

[crude flame]

or minecraft building stuff

[analog obsidian]

if ur dataset is above 2 hours, try using batch size 32

but 16 works too

i trained mine using batch 24

and 2x d passes

L1 Mel Loss

[winter dew]

ooo ok

ill keep that in mind as well

do you have any more tips I should know? about to sleep

[analog obsidian]

models can only do things similar to what they have in their dataset

so remember ur model is only gonna be able to do things similar to what are u hearing

[winter dew]

i see

even with a larger dataset?

[analog obsidian]

yup, they just sound very robotic while trying to do very different stuff

i can show u

[winter dew]

ohhhhh

[analog obsidian]

extreme example ^

[winter dew]

yea I get you

[analog obsidian]

now u kinda understand why most models sound robotic hehe

the more u fill those gaps, the less robotic results u get

[winter dew]

yup I understand that now

[analog obsidian]

use applio F0-spin branch because that allows u to disable multi-scale mel

L1 mel is more natural (in my opinion)

[winter dew]

sounds good

screenshotting all of this for later lol

[analog obsidian]

my settings ^

applio > rvc > train > train.pý

[winter dew]

okay

do you any good guides for it?

bc im gonna have to watch a guide prob to even get started lmao

[crude flame]

there is this https://docs.aihub.gg/rvc/resources/training/ and https://docs.aihub.gg/rvc/resources/dataset-isolation/ but they arent advanced

[analog obsidian]

these works but ye this is a bit more advanced

[winter dew]

ah

okay

[analog obsidian]

https://colab.research.google.com/github/jarredou/Music-Source-Separation-Training-Colab-Inference/blob/main/Music_Source_Separation_Training_(Colab_Inference).ipynb

this is what i use for removing background music

[crude flame]

voc_fv4 the goat

[analog obsidian]

oh yes don't use gaming streams

only just chatting or whatever

[winter dew]

ooooo

ok

yea that’s helpful

[analog obsidian]

for removing room reverb i use this

[winter dew]

unfortunately I really gotta sleep but you guys are the goats

ty

[analog obsidian]

dw just ask later

im usually here most of the time

[crude flame]

the heck is this?

[analog obsidian]

de-reverb

[winter dew]

sounds good thanks

[crude flame]

yea but like

yk

[analog obsidian]

its good

its not ai based tho

[crude flame]

did you 🏴‍☠️ or

[analog obsidian]

Acon DeVerberate 3

"buy" it

is very decent imo

my set had a pretty loud reverb

[crude flame]

i pretty much never have reverb in my sets

so idk if ill use it

dialogue isolate is all i need 😎

[analog obsidian]

[quasi iris]

I've got a mac Intel that I'm running the mac deiteris w-okada file on - I've managed to quarantine the files so I can open the application, but it crashes after a few minutes of loading - could someone help me with this?

[knotty moth]

intel mac is too old and unsupported on most AI workloads

consider the cloud alternative https://docs.aihub.gg/rvc-voice-changer/cloud/w-okada-kaggle/

[quasi iris]

Thanks! I'll give it a whirl

I'm a little lost on what to do after hitting copy and edit in kaggle - is there a video or anything I could watch?

[knotty moth]

I don't think anyone has made an updated & reliable video tutorial, only the above guide

[quasi iris]

Unfortunate - I'll see if I can figure it out and ask more oof

[tough elbow]

helloo, can someone tell me how can i download w-okada voice or is there better options?

🙏

[sturdy vigil]

How to download w okada

[sullen lion]

its fixed on dev

now im not at 100% with the results im after (character with heavy style influence to the point that i can prompt the character in different outfits at full weight and still keep the art style on offshoot models like wai)

but im at like 70%

as opposed to like 45% before (gens were still very booru outside of base illy)

one thing im noticing in all the other models and idk how much it even rly affects me

but kohya gets rid of the noise offset field when i select multires noise

and all the models i look at off civit has a 0.1 noise offset

is it possible for me to get the noise offset AND the multires noise params like these models?

i wanna be at like near 100% parity cus i want these same kind of results

[elder coral]

litsa whyy

[knotty moth]

https://tenor.com/view/huh-cat-gif-26460616

[sand bison]

?????

[sturdy vigil]

Sorry but they're not lying

[elder coral]

why rejoin on may 12 of thisyear

[sturdy vigil]

who

[elder coral]

litsa

how do i know the epochs of the model from my weights

i have to submit

-colab

[patent trellis]

📒 Google Colab Notebooks

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

• **Applio**

by IA Hispano
Google Colab

• **RVC Mainline**

by Hina
Google Colab

• **UVR5 NO UI**

by Eddy
Google Colab

• **UVR5 UI**

by Eddy
Google Colab

• **Wokada Deiteris Fork**

by Deiteris & Hina
Google Colab

• **Hina's Modified Original Wokada**

Google Colab

• **RVC-AI-Cover-Maker-WebUI**

by Shiro & Eddy
Google Colab

• **FaceFusion UI**

by Nick088
Google Colab

• **FaceFusion NO UI**

by Nick088
Google Colab

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

by Jarredou & Makidanye
Google Colab

[sturdy vigil]

who's lista

[sturdy vigil]

I don't know

[elder coral]

you don't know? that means you're new here

[knotty moth]

were you trying to submit for https://discord.com/channels/1159260121998827560/1305527335646269440 ?

[elder coral]

yes

[sturdy vigil]

hey

[knotty moth]

I suppose there should be other QC staff members that could respond on

perhaps wait till them online

[pastel oak]

Gpu too old

Use kaggle or upgrade

[elder coral]

qc i need help

how do i know the epoch of my weights voice model

[viral mason]

I don't think u can check, I've made some on weights to test, btw don't use weights for model making it's bad

[low shard]

there are different ones, choose based on the things i said

[stark zephyr]

what do i do after i install the realtime voice changer?

[low shard]

elaborate:

• your pc gpu
• what you want to do
• what tut link are you using

[wet lantern]

when i install the voice it says u cant download that type why

[stark zephyr]

my gpu is an rx 6600 i wanna use voice changer ig and no tutorial link i just got the disocord link Duckus's youtube vid

[low shard]

elaborate:

• your pc gpu
• what you want to do
• what tut link are you using

[low shard]

youtube tutorials are outdated asf

forget everything you got from it

[stark zephyr]

okayy

[low shard]

-realtime

[patent trellis]

💻 Local Realtime RVC

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

• Wokada Deiteris Fork

Most suggested. GUIDE

• Original Wokada

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

• **RVC GUI Mainline Realtime**

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

[low shard]

read 1st lin

[stark zephyr]

i alredy downloaded it

[low shard]

no, the one from youtube tutorial is a year old outdated version

as i said

[stark zephyr]

ik

[low shard]

forget that duckus tutorial even existed

[stark zephyr]

i downloaded this 1 alredy

[low shard]

yeah you shouldn't use it

[low shard]

so you got wokada deiteris fork or the one from youtube?

[stark zephyr]

i downloaded the yt one first and then deleted as it wasnt working and downloaded the wokada one

[low shard]

@winter iron slurs arent allowed.

[winter iron]

Okay

[low shard]

show a screenshot

!give-media-perms 1h @stark zephyr

[stark zephyr]

this 1?

i extracted alredy

[low shard]

click it, and then open MMVCServerSIO.exe

[stark zephyr]

done

[wet lantern]

3060 ti
i want to downolad voices from the server
idk i saw a vid says join this server and download the voice u want but when i download it says u cant download that type

[low shard]

then show a screenshot

[low shard]

you shouldnt use video tutorials, they are outdated

[low shard]

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

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

which do you need

[low shard]

click enter a few times then wait some seconds

[stark zephyr]

don3

done

[wet lantern]

is that program?

[stark zephyr]

Its doing its thing

ig

[wet lantern]

i use voice ai

[low shard]

whats happening

[stark zephyr]

when do i know when its finished?

[low shard]

and you shouldnt

[low shard]

which do you need?

[low shard]

show a screenshot

[wet lantern]

Is it a live sound or just a clip?

cuz i want a live sound

[stark zephyr]

@low shard its stuck here waht do i do lol

[low shard]

i explained you the differences, which do you need? realtime or pre-recorded audios?

[low shard]

yeah its downloading internal files

holy shit ur internet is slow asf

[stark zephyr]

ik

im playing roblox

val

and calling wiht someone

at the same time

[low shard]

just be safe to not fuck it up so hard that it doesnt download

[stark zephyr]

okayy

this ius taking awhile

[wet lantern]

if realtime means live voice ye realtime

[low shard]

realtime means using the changed voice in discord vc or games for example

alright

delete everything you got off youtube and voice.ai

-realtime

[patent trellis]

💻 Local Realtime RVC

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

• Wokada Deiteris Fork

Most suggested. GUIDE

• Original Wokada

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

• **RVC GUI Mainline Realtime**

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

[low shard]

get wokada deiteris fork, read the 1st link

[wet lantern]

How do I install it?

i installed it but it has many things

[cedar quest]

Did those postmodern jukebox ai covers get DMCAed by the YouTube channel (frank Sinatra covers for example ) or are they generally okay with people uploading AI covers of their covers as long as their channel is credited? Because last time I had a bunch of PMJ Brittany Murphy covers they never got takedown at all , I just deleted them from my main YouTube in case the channel turned against me so I wouldn’t lose my main YT content that was non AI.

I can’t find any info about this online about PMJs stance on the ai covers and whether they issued takedowns

Obviously I’m not famous enough to have dialogue with someone from PMJ to ask for permission to have Brittany covers of their music on my channel.

I was going to upload them to an alternative YT channel and not my
Main.

[wet lantern]

i installed the app but how i use it or open it

[stark zephyr]

it didnt move at all

should i re install?

[brittle wing]

i have a question

.......................

will somone answer

[craggy bough]

@brittle wing ^

oh wait

[viral mason]

ask

[brittle wing]

I

wanna know if w okada would work well with a 3050, 4050 and a 4060 all with 16gb ram

[viral mason]

I could help u set it up

[brittle wing]

i domnt have it yet son injust wanted to know

[viral mason]

idk how well it would run tho

[brittle wing]

would it work?

what specs do i need for it to wpork dceently

[viral mason]

I'm not sure but I use this gpu and it works well

[brittle wing]

idk much about gpus is that better than what i said

[hasty stump]

Hi, i wanted to use the ai voice changer thing to make a song, i had downloaded it a while back but i unfortunately forgot how and what do i need to install and properly work it out, anyone can help me?

[viral mason]

I am not entirely sure

but again I can help u download it and see if it runs well

[craggy bough]

those are all more than good enough

[viral mason]

are you trying to use an ai voice over existing vocals or do u want the voice changer to sing using the ai?

[hasty stump]

sing using ai, i wanna record my raw vocals and then use that voice changer to make it sound like juicewrld singing

(purely for my purpose not for profit ofc)

[viral mason]

I reccomend using this, you don't need the voice changer for doing that
https://www.weights.com

but if you do want the voice changer u can talk to me in dms about it

[hasty stump]

well its free right?

[viral mason]

Your gpus should be good, SussyBoi69 said they are good enough so just send a dm when u want to download it all

[viral mason]

the voice changer and weights is free yea

[hasty stump]

ohk should i test the app u sent me above and then if i wanna try out the voice changer later dm you?

[viral mason]

well u can still dm me if u want help with the weights app

[hasty stump]

ohk I'll lyk if i encounter any difficulties

[viral mason]

@brittle wing u still interested?

[low shard]

did you read the guide? you need to get the nvidia version

[low shard]

id just suggest to not post ai covers at all, companies do strikes once on a while and you could find yourself with 3 strikes in a day (ban) lol

i got 2 strikes in less than half a day once

[low shard]

what part of the guide didnt you understand?

[low shard]

be sure to not waste your internet and retry

[viral mason]

some people are slow, you have to show them with video,

like me

words don't do much if they don't read

[little timber]

so basically im looking for a realtime live voicechanger with a smaller delay and more realistic is there any good alternatives?

[tall cedar]

Can anyone here help me with the rcv voice

i tried to download it but got told i had 2 pay

[alpine sable]

would anyone happen to know what was used for this voiceover? https://x.com/gochionsol/status/1916526942037193001

[flint anvil]

ive been experimenting with a lot of different voice models and the realism on some is real hit or miss (robotic sounding, artifacts, etc.), what should i be looking for in voice models in #1175430844685484042? i usually filter by rmvpe, english, and rvc
i already have chunk size at 74ms and extra at 2.7, f0 extractor rmvpe, with force fp32 and a dedicated gpu on deiteris' fork (so i can up the settings but the models themselves just have artifacts and stuff)
so how can i find good models or how can i up the quality of the realtime voice changer? i'm just looking for a normal talking model not singing

[crude flame]

other than listening to the model samples there is nothing you can do to tell which one has better quality without downloading them

[flint anvil]

what about things i should look for like epochs/pretrain/etc.?

[crude flame]

pretrains dont change much
epochs dont indicate quality

f0 method you should either look for rmvpe or fcpe

pretty much everyone uses rmvpe so you dont really have to filter for it

[sand bison]

Well, I choose the one from Google Collab or I don't know if that's what you're referring to.

[sand bison]

what is kaggle?

[pastel oak]

Like google colab except it works

[sand bison]

or ok and where can I find it

[pastel oak]

https://docs.aihub.gg/rvc-voice-changer/cloud/w-okada-kaggle/

[sand thunder]

yo i execute the setup but i do nothing after ?

[sand bison]

There is no other option. The truth is that using Kaggle is more tedious. I prefer Google Collabs.

[pastel oak]

Kaggle more sustainable in the long run, Colabs literally do not work at all for wokada

[sand bison]

o ok thanks

[tall cedar]

So i got it to work but it dont work in vrchat anyone who can help

[low shard]

elaborate

!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.

[crimson kiln]

Hey everyone, quick question. Does stability matrix support .onnx models?

[stark zephyr]

@low shard

It worked I think.

I left it overnight.

[timber merlin]

Can someone please help me with okoda rvc? I recently upgraded my gpu to amd and i’m stuck on the cuda version, i don’t know how to switch to another version

The reason i want to switch to the other version is because next to the “gpu” processing option it only shows the CPU and overworks it to the point the voice changer stops working

I had help set this voice changer up a long time ago so i don’t know much or if the gpu change is the cause for the problems im facing with it, could be that or its an outdated version

[winter dew]

what exactly does it mean to merge voices?

i heard you can make it sound more realistic

[gilded robin]

in vonovox is there a way to add more effects? more vst plugins in the same line?
or do i just have to keep using things like reaper for that

[silent stratus]

if my dataset if over 2 hours should I just train from scratch

[winter dew]

also does anyone know why it feels like the outputted sound feels like its talking faster than the original?

it feels like im talking to fast after its converted

[analog obsidian]

no

2 hours is too small for a pretrain from scratch

[timber merlin]

anyone please?

[silent stratus]

gochu thanks

[timber merlin]

Its only letting me use my cpu so im certain its the cuda version being the problem so I followed the guide and downloaded the files but i dont know what to do next, where to put them/extract

[vapid dirge]

can you hear the changed voice after?

[timber merlin]

I do but super like corrupted, late, and laggy

And sometimes it comes through sometimes not

In the guide it says its not recommended for cpu to do the work so yeah

[wet lantern]

umm idk english too much but ill try to understand

[silent stratus]

would 8 batch size be fine

sorry for the questions i havent trained a model close to this big ever

or should i go 16

[analog obsidian]

i would use 16 but 8 works too

[silent stratus]

okay thanks

[winter burrow]

is wokada still the best real time voice changer? its been a minute since ive used it

[lethal depot]

I’ve been trying to find the best place to clone voices I’ve been using Kits.AI and it’s pretty good but all they want is to take your money for things that shouldn’t even be charged for. What is the best way to clone a singer that sounds really good?

And I don’t mean Weights I’m sorry but it’s just not good in my opinion

[odd isle]

انت عربي؟

[soft stratus]

How i use Applio RVC to train my voice

[latent kettle]

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

[soft stratus]

But the news is that I don't find upload my voice

[latent kettle]

https://docs.aihub.gg/rvc/local/applio/#training

[latent kettle]

Do you want to inference or train a voice model?

[full drum]

what do i put here if i wanna use the voice changer on discord ( i already have virtual cable downloaded)

[soft stratus]

Train voice model with my voice

[full drum]

euhh i cant send images but i mean for audio input and output

[latent kettle]

Go to your discord settings select mic input to virtual cable

[full drum]

what about the input and output for the actual voicechanger

[latent kettle]

https://docs.aihub.gg/rvc/local/applio/#training

[full drum]

its still on default idk which one to change

[latent kettle]

Input your mic
Output > virtual cable

[full drum]

thanks its work

[soft stratus]

Can it work for android

[latent kettle]

Read all the 5 steps 1. Pre-processing

[full drum]

i think you need a pretty beefy device to run rvc

[latent kettle]

You mean w-okada ?

[soft stratus]

What is that

[latent kettle]

Real time voice changer

You said "can it work for Android " what are you talking about?

@soft stratus

[soft stratus]

I mean upload my voice file to Applio works in android

[latent kettle]

You are training on Android? I mean in cloud like Google colab or kaggle

[soft stratus]

Google colab

[latent kettle]

#📰│dev-updates message

Yes it works but check it

I mean you directly have to upload your dataset file to a folder named "Dataset"

[fair hearth]

i dont have access to server mentioned in voyage contest

https://discord.com/channels/1159260121998827560/1374356047497527317 i dont have access what is this

#🏆│vc-leaderboard message i dont see vc anytime run in this server tho no one join

[rotund cairn]

can anyone help I use okada voice changer but my voice lags a bit and mumbles a lot

[craggy bough]

thats the old channel, this is the new one https://discord.com/channels/1159260121998827560/1380902127366574151

[vapid dirge]

does anyone actually have good success with w-okada?

i feel like i havent seen anyone with it working well

[knotty moth]

maybe it's just you, please explain your problem

!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.

[vapid dirge]

does yours work well

im genuinely curious

cause ive been trying and it just doesnt sound realistic

[knotty moth]

maybe try another model that sounds better, or check out this guide

https://docs.aihub.gg/rvc-voice-changer/realism/

[vapid dirge]

do you have one that works? and ive already checked that out

although i havent done it completely

im a little lost in connecting light host to the voice meter

[knotty moth]

there are some example models included in this guide: https://docs.aihub.gg/rvc-voice-changer/local/deiteris-w-okada-fork/#models-to-try

otherwise try searching in #1175430844685484042 or weights.com

[vapid dirge]

k ill try those models too

i just wanna see someones that works

[knotty moth]

it may depend on what kind of your voice on the mic

[vapid dirge]

i got a blue yeti is that good?

[knotty moth]

so try several models and find which of them work

also don't forget to adjust pitch as needed

[vapid dirge]

i do

[timber merlin]

Can i get some support too? My problem i believe is that im on the cuda (nividia) version when i should be on another since i upgraded to amd, because i noticed during the processing options for gpu it only shows my cpu and overloads it, i downloaded this link but i dont know what the next steps are, can i get a step by step on what to do after i download it in winrar?

(Dm me pls on what to do if a step by step guide is against the rules here, i’m worried i’ll mess up for the fourth time)

[simple ore]

voice changer client demo = you are using original w-okada, most likely cuda version of it. Not the fork the guide tells about.

[timber merlin]

so should i delete it along side with everything on the MMVCServerSIO with it?

[knotty moth]

definitely wrong version, should get the AMD one as you showed above

[timber merlin]

okay i downloaded that link idk what to do next...

and what should i do with the old one?

is it fine to keep?

[simple ore]

You can delete the old one.

[timber merlin]

everything in the file? or just the start_http?

[golden walrus]

Guys. Can i ask about what is spin 7 12 ? And how can i use it

[simple ore]

the while voice changer folder

[simple ore]

you can use it as a custom embedder.. or you can get an experimental branch of applio for that

[timber merlin]

okay im deleting everything in it

[knotty moth]

still wrong version, the b2332 one should have only the exe file to run directly, like this

[timber merlin]

well i just deleted the folder so now i got this one which i assume is the correct one?

[golden walrus]

Ah, i use Codename's applio and it only have spin. I don't know if 7 12 is much different from normal spin

[simple ore]

it should be 7-12 one

[timber merlin]

(i dont mind sharing my screen in help vc to speed it up the process)

[golden walrus]

And i have no idea if i can pair pretrain from Seoul Streaming Station for spin, the KLM4. Because i got some cracking issues with KLM6

[golden walrus]

Ah they are the same ? Dang it

i thought they are different

[simple ore]

I'm not sure how he included spin, I don't see his rvc/lib/utils.py downloading it

[golden walrus]

i have no idea, but it is in the fork somehow

[simple ore]

so I guess it just uses wharver is in rvc/models/embedders folder

[golden walrus]

do i have to download it again ?

i mean, it's here

[simple ore]

certutil -hashfile pytorch_model.bin MD5

for comparison

[golden walrus]

i don't get this, too advance for me

[simple ore]

it is a command that calculates the checksum of the file, so you can find which version you're actually got

since the names are the same

open command line in the spin folder and run it

[golden walrus]

ahhhhhhhhh

so 7-12

ahhhhhhhh thank you so much

[timber merlin]

i got the voice changer but why is it a web version? is there a none-bowser version?

[simple ore]

it just uses browser UI

[timber merlin]

okay ill test it

[simple ore]

if you are running it locally, change to server

for direct access to hardware without the browser limitations

[timber merlin]

oh okay

uhhh

oh ok

it switched

which one of these should i use? the WDM?

or mme? idk the difference

[golden walrus]

ah, Noobie sir, do you know if refineGAN can be used in realtime yet ? i read in KLM5 article, and someone said it wasn't suitable for realtime

[empty sundial]

help applio says no api found 😦

[golden walrus]

ah nvm, i will wait for SSS release KLM6v3

[simple ore]

in Vonovox maybe, but there are no models

[simple ore]

what version of applio?

the only api there is core.py command line

[golden walrus]

hmmmmm, so for now there is no pretrain model for refineGAN ?

ah

klm6 use hifi

[simple ore]

there is, even for spin.. but is it mostly suitable for speaking

singing is an issue because it renders harmonics too well, so they end up mirroring like crazy

[golden walrus]

like, it is not working well when singing ? because i don't remember any models can sing

[stark zephyr]

How to get the RVC to work in discord and games?

[simple ore]

there are artifacts that you may hear

[golden walrus]

ah, understood, so i should stick to hifi-gan for now xd

[lucid stratus]

hello. I recently got a new gpu and wanted to ask about gaming and w-okada.
With the 5070 ti it sounds quite good without gaming, but when i play games such as hunt showdown it starts to sounds a bit worse.

Is the gpu still not good enough or can i adjust settings and it will improve? I don't have that much knowledge about it. (as in reducing ingame graphics and maybe increasing the delay from 0.4sec to 1sec?

[simple ore]

hifigan mirrors too 🙂

[golden walrus]

i knew itttttttt

but what if i don't really sing

[simple ore]

when you play a game that in addition to 3d also uses upscalers like DLSS, there's a big competition for the computing resources against W-Okada

[simple ore]

then either should be fine

[golden walrus]

i just wonder if it can process breathy or scream

it's based on my data too right ?

[lucid stratus]

i see. so if i dont activate DLSS and maybe reduce the graphics it might bring better results? Thanks 🙂

[simple ore]

add fps limiter if the game allows, that may help

but 5070ti is decent

unfortunately there's no way to prioritize the gpu resources

[timber merlin]

i got it to work finally!!!

thanks for the help everyone 💙

though is there a way to increase/decrease chunks?

because its locked for me

also i cant click on sup1 or sup2

oh nvm i just had to turn it off im stupid lol

[elder coral]

does this model have artifact problems

-colab

[patent trellis]

📒 Google Colab Notebooks

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

• **Applio**

by IA Hispano
Google Colab

• **RVC Mainline**

by Hina
Google Colab

• **UVR5 NO UI**

by Eddy
Google Colab

• **UVR5 UI**

by Eddy
Google Colab

• **Wokada Deiteris Fork**

by Deiteris & Hina
Google Colab

• **Hina's Modified Original Wokada**

Google Colab

• **RVC-AI-Cover-Maker-WebUI**

by Shiro & Eddy
Google Colab

• **FaceFusion UI**

by Nick088
Google Colab

• **FaceFusion NO UI**

by Nick088
Google Colab

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

by Jarredou & Makidanye
Google Colab

[signal crater]

guys, which tool should i use for below use case?
lets say i have a video of a person (a movie scene)
i want it to say something some custom dialogue
which tool i should use?

[low shard]

gpu: ur rx 6600
chunk: 128
extra: 2.7

did you install vac lite?

[timber merlin]

1 small issue, i can't mess with the noise supression

[stark zephyr]

I did.

It works in discord now.

[low shard]

set input to microphone
output to line 1
monitor to headphones

also you can optionally use force fp32 mode in advanced settings for better quality at the cost of some delay

[low shard]

show a whole screenshot of all ur settings

[stark zephyr]

with these settings

it sounds so bad tho

[low shard]

show a whole screenshot of all ur settings rn

[low shard]

^

[timber merlin]

yes sir!!!

[low shard]

set extra to 2.7
chunk to 200ms

the reason why you cant use noise/echo suppression is because you're in server mode

client = can use noise/echo suppression, but can have more delay and is easier

server = harder to use, less delay and can't use noise/echo suppression

if you need noise/echo suppression, you need to use server or use a 3rd party tool for noise suppression

also, you can optionally turn force fp32 mode on in advanced settings for a bit more delay and a bit more quality

[stark zephyr]

It sounds good enough for me.

[low shard]

set extra to 2.7, chunk to 128

ALWAYS check the triangle when youre changing settings on AMD

the triangle will be your life saver for AMD

[stark zephyr]

Which triangle

[stark zephyr]

oh

right

i cant make it 128.

[low shard]

hover your mouse over it, it will tell you wha tto do

[low shard]

you need to click stop, change the settings, it doesnt matter if they arent as accurate just close, then check the triangle

like it doesnt matter if u set it to 127 or 129 instead

[stark zephyr]

Best I can do is 136...

[low shard]

its fine

[stark zephyr]

Okay.

Now im switching back to the gpu.

No triangle anymore

Is that good?

[low shard]

also, if you click advanced settings, then set force fp 32 mode on, you can get more quality with a bit more delay

[low shard]

yes, always check the triangle whenever you change settings, its good when its not there anymore

[stark zephyr]

okay

.

This sounds way worst?

Its like cracking up.

[low shard]

try out other models

[stark zephyr]

Okay.

[low shard]

check https://docs.aihub.gg/rvc-voice-changer/local/deiteris-w-okada-fork/#crackle-fix

[stark zephyr]

Alright thanks.

[low shard]

also reminder that the quality depends much on the model, if the model sucks, quality will

its always better to try many models

[timber merlin]

oooh okay thank you, should i adjust the ms if im gonna play a game like vrchat? or it doesnt matter? (also i cant set it to 200ms only 210.7ms is the closest because i can't just type it by the looks of it)

[low shard]

should i adjust the ms if im gonna play a game like vrchat?
yes, just remember to set the chunk an higher value slightly than the perf value

also i cant set it to 200ms only 210.7ms is the closest because i can't just type it by the looks of it
yeah you cant type it unfortunately, but it doesn't really matter dw, just as long as its close and not that you put like 560 instead

[low shard]

so, do you need any other help?

[stark zephyr]

Nah thats it

Tysm.

[timber merlin]

so ingames the higher i set the chunk value the better? will it increase the delay in the voice if i increase it?

[low shard]

Yw

[low shard]

chunk basically controls the delay

if you put a lower value than the perf value, it will start lagging basically (because the perf value is the ms of performance ur gpu is doing)

so, in games you have to make it control an higher delay so it doesnt lag out

[timber merlin]

ahh okay so it'll have probably a larger delay but it wont lag out or be inaudible

[low shard]

exactly

else it will lag because you're forcing your gpu to do a less delay than what it can do while in game

[timber merlin]

does using quest 3 pcvr mode also affect it?

[low shard]

its suggested to play at the lowest settings possible btw

[low shard]

not sure, i dont have a vr

vr games might be a bit more intensive tho?

[timber merlin]

low settings in terms of the games or the voice changer?

[low shard]

game

[timber merlin]

ahh okay

thanks a lot for this info, i was struggling yesterday

trying to understand a lot of it

[low shard]

yw, need any other help?

[timber merlin]

i think that's all ill be messing with it and will try the client side instead of the server

[low shard]

btw the great majority of settings is either explained in https://docs.aihub.gg/rvc-voice-changer/local/deiteris-w-okada-fork/ or when you hover to the settings name and the settings has like "....." under the name (sorry idk how to explain it lol)

[low shard]

that means more delay, tho u can surpress noise

[timber merlin]

i'll see if its acceptable amount of delay when i test it in vr, also i noticed the audio in client is quieter than the server?

like i hear myself but at lower volume

[low shard]

try increasing the mon volume

[timber merlin]

it didnt change anything

[low shard]

can you try increasing the volume of your headphones?

[timber merlin]

i mean i hear myself its just quieter when i switch to client i dont think its a headphones issue

but its fine its not so low its inaudible, its still fine, i just noticed it

oh!

what fixes is it is the out and in (even tho both is 100 client and server but in client its quieter so it should be increased in my case)

all good thank you again for the help that's all

[fickle minnow]

when i download a voice it doesnt sound good at all and its buggin

[hallow thistle]

!howtoask

[patent trellis]

How To Troubleshoot

__**GIVE CONTEXT.**__ 📝

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

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

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

__**BE PRODUCTIVE.**__ 🤝

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

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

[hallow thistle]

Which W-Okada or RVC program are you trying to use? And what is your PC GPU?

[fickle minnow]

5700xt directml 1.5.3.18

[hallow thistle]

Download the "better" W-Okada DirectML from this guide instead. Yours is old and outdated. https://docs.aihub.gg/rvc-voice-changer/local/deiteris-w-okada-fork/#download-amd-intel-and-cpu-on-windows

[fickle minnow]

ill try and let you know thanks

[low shard]

1.5.3.18 is an over year old version of original wokada lmao

delete everything you got off video tutorials, they are old

the version that namari gave you is the latest wokada deiteris fork, which is the best

if you have vb audio cable, uninstall it

[fickle minnow]

why?

[hallow thistle]

VB-Cable gives random issues for Windows users when trying to use it with W-Okada, as its settings kinda complicated to get it work properly.

[fickle minnow]

but i need that if i wanna make it work on discord ill just try

[hallow thistle]

https://docs.aihub.gg/rvc-voice-changer/local/deiteris-w-okada-fork/#virtual-audio-cable

There's an alternative to that one. It's Virtual Audio Cable lite, works out of the box. It doesn't have always be VB-Cable.

[fickle minnow]

alright thanks, i deleted vb cable but it still appears in my output settings

might have to restart my pc

[fickle minnow]

it works perfectly fine thats crazy

[latent kettle]

@simple ore sorry to bother you, but is it possible for a normal user to train a custom pre-trained model. If yes, please elaborate how.

[simple ore]

like.. have money to buy 4090 or rent something better

find 200+ hours of various audio from different speakers, should be same quality

[latent kettle]

Yeah it could be possible

[simple ore]

make up to 109 files

just train from scratch

[hallow thistle]

Pre-trained RVC model? I think it sounds possible, but you would need to gather a lot of audio datasets to make one, more than for a typical normal voice model.

[latent kettle]

Will it give any good results if I fine tune my models with it ?

[simple ore]

55h set on 4070 runs ~30min/epoch, you need to get like 300-600 epochs

pretrain is a base

like a blank sheet cake you buy from the store to write 'happy birthay' on it

[latent kettle]

It will take like months on my system

[hallow thistle]

Will it?

[latent kettle]

I got 4060

[hallow thistle]

[latent kettle]

Maybe possible on kaggle?

[simple ore]

if you pay for something better than 2x 10-year old T4s

but there are cheaper options around

[idle osprey]

how to make a model not sound robotic

[light latch]

Does anyone have the Google Collab link to make the covers?

[marsh acorn]

Yo can anyone come in VC support to check if my Overwatch Winston voice changer is working ?

just ping me if you can

[latent kettle]

Is it a good idea to train a custom pre trained? Like will it improve the overall quality?

[simple ore]

Dont think so... a pretrain can help to expand the range of the model a bit, but default hifigan kinda sucks

you risk of replacing the pretrain data with just your finetune if you over-do the train

and a bigger pretrain dataset does not seem all that better than smaller

sure, having some singing data and dynamic range helps, but...

[humble ermine]

how do i package my rvc model using a command here ?

[viscid moss]

HF is down btw

[analog obsidian]

im training a spin pretrain, if everything goes well im going to share it in #1235952130855010365
but hifigan only, not refinegan
speech only, no singing

[forest vector]

Im getting problems getting the RTX 5090 to run on deiteris' optimized W-Okada Fork Voice Changer.
I installed both versions, the one for the 50-series doesnt work, and gives "voice not selected" error messages.
tested the 50-series with a 4080M card and getting the same error.

downloaded the non 50-series version and it worked for CPU.

[crude flame]

Have you tried selecting a voice?

[forest vector]

yes, it is selected, made sure to upload a pic on the voicemodel to make sure

[crude flame]

Did you click the voice?

[forest vector]

yes, multiple times

[crude flame]

Did you try switching to a different voice then back to that voice

[forest vector]

im certain it is selected, I know how to work with the non-50 version

[forest vector]

I didnt, but I uploaded another Voice to the same slot

[solar sinew]

How do I train a model?

[golden walrus]

i don't think any current real time stuff can sing tbh

but gambatteeeeeeeeee

i can only wish you the best

[analog obsidian]

singing is hard

for ai

[golden walrus]

but we will get there eventually

[analog obsidian]

maybe maybe refinegan + a singing only dataset

[golden walrus]

somehow i can't use KLM 4. or 5. or 6. correctly

it keeps losing voice and having lots of cracking

[analog obsidian]

you're trying to infer singing with speech datasets?

[golden walrus]

no lah, i don't have singing, only speech data

like, i can't use it for real time for whatever the reason

[analog obsidian]

welp you answer urself why your models cant infer singing

[golden walrus]

i mean, not singing, just talking normally

and it has cracking and lose voice mid sentence

[analog obsidian]

how big was the dataset?

[golden walrus]

1 hour of pure speech

[analog obsidian]

could be your mic

[golden walrus]

but with normal pretrain, i can speak normally

same dataset

[analog obsidian]

ahhh

ye klm is not that great for speech

[golden walrus]

[analog obsidian]

imo just stay with the og pretrain

[golden walrus]

but but

in the pretrain

i heard those samples was pretty neat

[analog obsidian]

it was ok for me but i noticed speech sounds unnatural

the og pretrain is great

only issue is that it cannot sing because the pretrain was trained with speech

[golden walrus]

i can't sing irl so

i don't expect AI to sing well

[analog obsidian]

multi scale mel loss seems to increase vocal range but it made my models not resemble the original voice too much

but the true solution would be training a pure singing pretrain, then finetuning singing
(refinegan may be a better choice for this)

the spin embedder seems to be better at handling noise so in theory breathing should be better

but i havent compared that yet

[crude flame]

have you tried wavlm at all?

[analog obsidian]

nope, did someone here tried it?

[crude flame]

i dont think so

[austere hollow]

how do i unlink my discord from weights

i got a new acc

[sullen lion]

yknow what i should ask the general question of what are some recommended settings for training cartoon style loras for usage on non-base illustrious checkpoints

im like 85% there but i wanna be closer to 90-95%

i see plenty of loras that supposedly accomplish this but trying to follow their training params has been only Okay

if i have to change other settings in my genner (i use automatic1111) to get closer i can do that too

ideally i wanna get the style to remain when i do gens on wai-nsfw (what a lot of the models i see use to prove style rigidity)

[crude flame]

could be bad tagging

but i wouldnt know im so new to training loras 😔

if someone could actually answer that question that would be very helpful to him and me

i mean ig you can also try a lycoris

like a locon or a glora

[elder coral]

how do i do this qc

the demo audio

[austere hollow]

send it in drive format

i did that

[knotty moth]

if you're not satisfied with booru-like thing, try pony or flux

[sullen lion]

i share an instance with my friend but he doesnt seem privy to go back to pony

trust me if it was up to me id run crying back to pony training on that was a blessing

but i KNOW its possible on illustrious

i SEE models that do it all the time

https://civitai.com/models/1037358/star-butterfly-svtfoe

[gilded robin]

hi theres an issue with vonovox and flexasio setup i dont know why if anyone could help

Input Device: FlexASIO
Output Device: FlexASIO
Configured Sample Rate: 48000
Error creating audio stream: Error opening Stream: Invalid number of channels [PaErrorCode -9998]
Critical error in start_vc: Error opening Stream: Invalid number of channels [PaErrorCode -9998]
Traceback (most recent call last):
  File "gui\\gui.py", line 1229, in gui.gui.GUI.start_vc
  File "gui\\gui.py", line 1270, in gui.gui.GUI.initialize_voice_conversion
  File "gui\\gui.py", line 1324, in gui.gui.GUI.start_stream
  File "core\\audio\\audio_processors.py", line 797, in core.audio.audio_processors.AudioDeviceManager.create_stream
  File "core\\audio\\audio_processors.py", line 787, in core.audio.audio_processors.AudioDeviceManager.create_stream
  File "C:\Users\legen\Desktop\Vonovox-1.4.5\runtime\Lib\site-packages\sounddevice.py", line 1825, in __init__
    _StreamBase.__init__(self, kind='duplex', wrap_callback='array',
  File "C:\Users\legen\Desktop\Vonovox-1.4.5\runtime\Lib\site-packages\sounddevice.py", line 909, in __init__
    _check(_lib.Pa_OpenStream(self._ptr, iparameters, oparameters,
  File "C:\Users\legen\Desktop\Vonovox-1.4.5\runtime\Lib\site-packages\sounddevice.py", line 2796, in _check
    raise PortAudioError(errormsg, err)
sounddevice.PortAudioError: Error opening Stream: Invalid number of channels [PaErrorCode -9998]```

shows up in the cmd

[elder coral]

ok

[sand bison]

So, are the only two ways to create AI models with Applio and Kaggle? I used a Colab that said it was called RVC v2. Disconnected - Colab. Sorry for the inconvenience.

[stark zephyr]

hi how do i fix this

I clicked enter and it doesnt pull out the RVC client on my chrome

[unreal linden]

How much epoch should i train if i have 15 min voice with clear sound.

[peak path]

1. should i use mono or a stereo wav file for my dataset files?
2. is it possible to use flac instead of wav for dataset files? which one is better?

[analog obsidian]

1. mono, but if you forgot to convert the dataset to mono it doesn't matter, applio/rvc will automatically convert the dataset to mono
2. flac works too

[pastel oak]

No answer to that, use tensorboard to track progress

[pastel oak]

What version are you using

Client uses MME, a sound device, by default which is extremely outdated
Server lets you choose which audio type you choose. You can use mme there too but obviously no point when wasapi is newer. Wasapi has less delay (faster audio processing) to name one benefit

[pastel oak]

For you its best to use kaggle yes. Colabs dont rlly work atm because of an update Google had since they dont like the use of colabs like this in the first place unless im misremembering

Applio is not a category like colab and kaggle, you would use applio inside kaggle you get me

[stark zephyr]

Fixed alredy.

[unreal linden]

How to use it?

[pastel oak]

https://docs.aihub.gg/rvc/resources/training/#tensorboard

?

Youre describing something different

One has to be ticked by default

[gilded robin]

but it works on w-okada just fine just not on vonovox, is there a way to contact who made vonovox?

[knotty moth]

adjust chunk according to the gpu capability

[gilded robin]

whats your extra & chunk & gpu?

[pastel oak]

@shy spruce

Gpu goes idle maybe. Try lowering chunk a little bit more like 150, if it doesnt fix, then run the "force gpu clocks.bat" file outside of the mmvc folder

[pastel oak]

For everrything

[small jacinth]

hey yall
do any of you use AI in your projects?
i want to integrate AI into my project
but i dont know how to do it for free

[pastel oak]

48k is fine

Higher frequencies getting picked up

[pastel oak]

Which doesnt matter cause wokada downscales to 32k anyway

But your mic inputs everything into wolada to its fullest range this way

[pastel oak]

Means your mic picks up your headphones output and loops it
Move mic further away from your headphones, lower volule on your headphones, move in. Sens. Further to the right

[broken urchin]

hello

when running deiteris fork voice changer i was wondering if i can change the name of the run file from MMVCServerSIO to anything else and the voice changer folder name to something else

if i change the name of these files will the voice changer still be able to run?

[paper locust]

not sure if heres where i should ask, but im having this issue with the voice changer, when i use it i have it set so it inputs my mic, outputs VAC, and then in discord inputs VAC and outputs my headset, what happens is when other people talk loud enough, the voice changer for some reason picks up their voices slightly and then replays it, but my mic normally doesnt do that so idk what im doing wrong

[paper locust]

hm, my headphones do that but they are old so the system or something for it probably broke

arctis 7p+

the weird issue is that ive done things like this before but ive never had this specific issue happen, my mic normally doesnt pick up noises from my computer, but for some reason the program does

[knotty moth]

turn down volume of voice chat/system sound and use decent noise suppression

if you have someone irl, tell them to be quiet

[paper locust]

and i cant make it audio loop on itself by screaming loud either

which is the weirdest part

[paper locust]

hmm, i can try that, krisp noise suppression completely kills off the voice changer

so im forced to use standard

so i guess i can noise supression on the voice changer itself alongside discord standard noise suppression

[pastel oak]

Might aswell try before asking it takes 10 seconds lol

[shrewd verge]

does anyone know how to create your own voices for the VCClient? there's probably a whole load of tutorials out there im just looking for a point in the right direction

i have a sample for the voice all ready to go i just dont know how to use it lol

[latent kettle]

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

[shrewd verge]

much appreciated

[shy spruce]

there is a channel setting in flexasio when you set the devices. make sure both devices have the same number of channels, they look mismatched

[median monolith]

when making a voice model of a character that doesnt have much dialogue (for example: a character wich only has 1 minute of unique dialogue in total or maybe even only a 1 second audio clip), how long should the dataset recommendably/ideally be? should it be only the unique instance(s) of voice/dialogue or should it be an audio of said dialogue/voice repeated until you have an audio file with certain duration? if the second case, I wonder what duration should be enough.

[analog obsidian]

1. 10 minutes minimum
2. no, don't repeat the audio, its going to give very bad results

[cosmic vigil]

#ask Index Rate what for ?

[analog obsidian]

is the accent of the model
index 0 uses your accent

values higher than 0 will begin to blend the model's accent with your accent

[cosmic vigil]

finally the best simple answer that can i understand with lower iq, thank you 🙏

i think for ideal real time voice on games, using higher index will be better 🤭

[median monolith]

well, in that case, the character im trying to make a model only has like 1 minute of usable unique dialogue... Im not expecting it to be a SpongeBob or Michael Jackson quality type of model, but still, I wanted to know, would it then just be "better" for me to use only this 1 minute long audio of him "singing" than repeat it like 5 or 10 times? (to atleast make it less mediocre) its what I understood.

[analog obsidian]

no that wont help at all

the model requires actual 10 minutes of diverse data

no the same thing over and over again

technically you can train anything, even something below 10 mins

but don't expect good results

[median monolith]

so yeah, i guess i will just use the 1 minute audio then... theres not much I can do ¯_(ツ)_/¯

[median monolith]

yeah, again, not expecting a super high quality result

[analog obsidian]

sounds interesting, what if you try it?
im currently training my hifi pretrain, breaths are still robotic :<

[median monolith]

now im thinking about the same thing, but instead of different volume, it would be different pitch and "time".

[analog obsidian]

gpt told me about different volumes, pitch and time being viable as data augmentation

i dont think there have been tests of data augmentation here

adding more breaths technically may also improve them

[crude flame]

hifigan needs natural audio right? If so then thats why we never used data augmentation

[analog obsidian]

hifi/refine can clone whatever it's in the dataset

[crude flame]

according to gpt it may perform poorly

[analog obsidian]

ye seems like gpt changed its mind since last time i asked about data augmentation in vits

[crude flame]

i mean we can still try but i doubt anything good will come out

[analog obsidian]

i think this to work the augmentated data should be shorter than the original data

like adding 5 minutes of augmentated data to a 10 min set?

but I'm not sure, someone should try it

[crude flame]

i can try it rq

[analog obsidian]

try what noobies said
duplicate the dataset but with lower volume

i can try after this pretrain learns how to reproduce breaths

[crude flame]

👍

how much volume lowering do i do?

[analog obsidian]

good question

-3db?

[crude flame]

ill try -3, -6, and -10

[wheat egret]

i just joined to try a specific ai voice made by MartinFLL and i wanted to know how can i use it? do i have to host an ai on a local device?

[crude flame]

i love my 5060 ti

[analog obsidian]

[wheat egret]

sorry to bother you do you know where can i have informations about getting started?

[crude flame]

if you want to make models follow this:https://docs.aihub.gg/essentials/how-to-make-voice-models/

If you want to make a cover follow this: https://docs.aihub.gg/essentials/how-to-make-ai-cover/

[wheat egret]

i suppose for TTS i should look at the "how to make ai cover" page ?

i don't think there is TTS

[crude flame]

what tts do you want?

we have a list of them here: https://docs.aihub.gg/tts/tts-tools/ (minus chatterbox, i havent added it yet)

[wheat egret]

https://huggingface.co/MartinFLL/ai-voices/blob/main/cassie-20epochs-rvcv2.zip

[crude flame]

thats a rvc model you cant use it for tts

[wheat egret]

alright

[crude flame]

unless you make some audio with a diff tts then infer with rvc

[wheat egret]

oh alright

i see

[wheat egret]

Sorry for the dumb question what's an RVC ?

[crude flame]

Retrieval-Based Voice Conversion is a voice cloning ai

[wheat egret]

thank you

[crude flame]

^

[wheat egret]

Alright well i'll check the ai cover thing and if i have some troubles could i ping you back?

[crude flame]

sure, you can ping me or just drop your question here

[wheat egret]

ok

thank you for your attention

[gilded robin]

ive genuinely tried everything idk, (1,2) (0,1) (0,0) (1,1) (2,2) input output channels on vonovox & flexasio, i checked the sound settings to make sure what amt of channels they have aswell,
it worked for the first few times on 0 channels for flexasio & vonovox but just stops randomly, it works like genuinely 1/10 times without reason.
if it helps when i tried to re-open setup.bat it worked for 1 single time but then stopped right after i clicked stop and start again

[brittle wing]

-colab

[patent trellis]

📒 Google Colab Notebooks

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

• **Applio**

by IA Hispano
Google Colab

• **RVC Mainline**

by Hina
Google Colab

• **UVR5 NO UI**

by Eddy
Google Colab

• **UVR5 UI**

by Eddy
Google Colab

• **Wokada Deiteris Fork**

by Deiteris & Hina
Google Colab

• **Hina's Modified Original Wokada**

Google Colab

• **RVC-AI-Cover-Maker-WebUI**

by Shiro & Eddy
Google Colab

• **FaceFusion UI**

by Nick088
Google Colab

• **FaceFusion NO UI**

by Nick088
Google Colab

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

by Jarredou & Makidanye
Google Colab

[crude flame]

bam done (ish)

these are all bs4 10 min of real data and 5 min of quieter repeated audio all at 100 epochs

weird thing is that they all had different step counts

idk why

i changed nothing

[analog obsidian]

they sound almost the same for me xD

@crude flame try this

[vagrant flare]

Hey does anyone know a really realistic image generator?

I heard about midjourney but i want to be sure it will be really reaslitic before i spend money

[median monolith]

Thats strange..., I did everything the guide for Applio Colab (cloud) said, but when I reach the part where I have to press the "start training" button, it seems to be not working for some reason is what it seems.

[crude flame]

heres the audio noobies sent

i dont really got any audio with much volume difference

[lean rover]

Hi, could someone help me solve this error?
Looks like there's a bit of a problem.
unknwon message

If you clear the information being managed by this app, it may be recoverable.

Initialize

Reload without initialize

Error
unhandledrejection
no error stack
Error: Could not load Voice Focus estimator.
Error: Could not load Voice Focus estimator.
at http://127.0.0.1:18888/index.js:2:1042547
at Generator.throw (<anonymous>)
at s (http://127.0.0.1:18888/index.js:2:1039349)

[lean rover]

and how do I do that?

Yesterday when I used it everything was normal, and today it won't let me open the application.

[wooden girder]

Does anyone know why when I change protocol to rest my pref stays 0 but if it's on sio I can see it jump and change accordingly when speaking for wokada?

[gilded robin]

i have this really good model that i like a lot but i dont like the voice that much, is there a way to fix it? merging?
is there a guide to merging?

or a way to make it softer?

[fluid lion]

what are some of the most convincing/best sounding voice models out right now?

[gilded robin]

for?

[fluid lion]

RVC

unless there a better software

[gilded robin]

ye but like girl u mean?

[fluid lion]

if it sounds the most "realistic" than sure, im trying to see how good it sounds now, i tested it a year ago or so

like live voice to voice

[gilded robin]

there are always more and more settings to finetune if you want it as realistic as possible

such as EQ, bitcrush etc

a year ago or so also wasnt rmvpe instead it was crepe iirc?

[fluid lion]

i dont remember

i am useing RVC

[gilded robin]

dm

[peak shadow]

For some reason, RVC in Google Colab does not generate the necessary files for saving AI voice.

[peak path]

i use this colab
https://colab.research.google.com/github/iahispano/applio/blob/master/assets/Applio_NoUI.ipynb

is applio/rvc convert flac to wav before training?
i want to know if rvc support flac natively or convert it to wav before training.

[wheat egret]

how can i use these two .INDEX files for RVC?

it's not documented on https://docs.aihub.gg/essentials/how-to-make-ai-cover/

they only have the "added" file

[peak path]

i use UVR 5 in order to remove background noises and extract human voices.
i tried a lot of models in
https://colab.research.google.com/github/Eddycrack864/UVR5-NO-UI/blob/main/UVR5_NO_UI.ipynb
but i can't find a model that able to remove birds sound.
any suggestion to remove birds sound?

BS-Roformer and Mel Band Roformer
MDX23C
MDX-NET
VR ARCH
Demucs

[wheat egret]

ok

[severe hawk]

I'm using Kaggle with Applio and I have finished training my thing, but when I press "Restart Applio" it continues training (I can't stop it no matter what I press) and there is no added_ file. How do I fix this?

[severe hawk]

great, now my ngrok ended

smh

ima try a different method then

I was seeing that response to people having the same problem as me, but I couldn't find the Generate Index button for the life of me...

Ohhhh, that one

alright, thank you so much

[hallow thistle]

This is W-Okada, not RVC. RVC is another different program.

[latent kettle]

How to add mel reformer models in UVR GUI

[scenic atlas]

hey when i try to write something to chat gpt he just doesnt answer me or gives me an error does any one know how i can fix this ?

[toxic ingot]

Is this normal with Taco2?
https://files.catbox.moe/uh8omt.png

[low shard]

you cant fix it https://status.openai.com/

the servers are down

either wait or use something else like gemini

[scenic atlas]

Ok ty

[low shard]

be sure to not watch video tutorials for rvc

1. what's ur pc gpu?
2. what do u want to do?
3. what are u using?

[low shard]

that's wokada deiteris fork, not RVC

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

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

what's ur pc gpu? what do u want to do?

[plain orchid]

can please someone help me with image generation either chatgpt or midjourney

i want to use a certain style but can't figure out how to make it match

[frigid stone]

I got Applio from the compiled Windows ApplioV3.2.9.zip on HuggingFace but the run-tensorboard.bat isn't working
It throws the error ImportError: cannot import name 'notf' from 'tensorboard.compat' (D:\ApplioV3.2.9\env\lib\site-packages\tensorboard\compat\__init__.py)

I might've fixed it by moving to the C: drive, testing by moving it back to my D: drive

[steady thorn]

hello
i need some help
im looking to create an ai influencer ( im trying and im overwhelmed ) with comfyui, can someone help me with a workflow that could run on my laptop? low vram ( 6gb vram ) rtx 3060 laptop
text to image workflow + lora stacker

[frigid stone]

Welp, Tensorboard is now open
But it still throws those errors

I guess they weren't important

There're 5 or 6 of these in a row

This is the second time unzipping it, not sure how I can unzip it more properly

11

[prisma kettle]

Got this error in Replay and went to the website, is there any way to use the ones it suggests instead?

Oh boy

What does this mean? I'm on a 5070 Ti

https://www.reddit.com/r/weights/comments/1ki6n3e/replay_ai_5000_series_support/

...suffering from success man, wtf

[frigid stone]

Bizarre, not sure what I can change

[prisma kettle]

Ooooooh yeah that's right

I never installed it back in April, thx for the reminder

I forget what I was trying to even do

Well now is not a good time for me to install torch but back in April that's what you said to do lol

Oh it was UVR that I was trying to get to work

Is there a torch installation guide handy

This is what you said back in April but I'm a dummy and don't want to mess up an install

I can just run that in cmd?

Don't think I have any old versions of torch since I never installed the nightly build

Sweet thx

Wdym by activate sorry

I am a noob

The rest of it meaning the other lines above?

Alright thank you

I'll start with this

[frigid stone]

Re-extracted Applio again for the third time, didn't change anything
Not sure how to debug this

Going to try with 3.2.8-bugfix

[prisma kettle]

@simple ore Do you know what this means?

Restarting PC lol, had to install Python

[frigid stone]

The website does open, I assume it works but I haven't checked if a graph shows up yet

3.2.8-bugfix has no error messages

[prisma kettle]

Ok python installed and restarted PC

Nooo I'm still getting this error.

Now "The system cannot find the path specified."

I am so out of my depth sorry

Will this also make it work for Replay? Both are broken rn bc I don't have torch

What is that sorry

What link are you looking for

Oh I just got it from here
https://www.weights.com/replay

this is cool

Oh this is the error I got earlier

Idk I'm just going with what weights gives me

the weights one still gets updates

Anyway, I'm in the UVR folder in cmd now, so you're saying I should install torch there?

My folder is just called Ultimate Vocal Remover, unless I'm in the wrong one

That's weird, I just installed it and rebooted

I'm in the AppData/Local/Programs/Ultimate Vocal Remover folder, but it seems like that's not where I should be

I did

I'm computer noob I use exes

it's what I do

Oh maybe tha tis what i did

I thought I did that actually but it's been a bit

Nah I ran the exe

Why can't I use the exe

Alr

Thanks

I do have a torch folder in this Ultimate Vocal Remover folder

this is all that's in there tho

Can I install torch for Replay?

Hmm

I just got it from their website and then here

https://github.com/Anjok07/ultimatevocalremovergui/releases/tag/v5.6

That's the version that the official website links to 🤷

Thank you

Did Anjok abandon UVR

Thanks for helping me sorry I don't know what I am doing at all

I'm installing it through the bat now

The website said to run the bat

It's already going

Oh

Shit

Lol

I moved too fast

I'm dumb

Idk if I should just let this play out or what

Idk how to uninstall any of what I just ran

Ok

Unzipped into a C: drive folder

Don't think I have ffmpeg on this pc so I will install

But I can't rn

thank you for bearing with me

I have to get back to work but I’ll send you a screenshot, it just may not be until later or tomorrow. Can I ping you when I’ve got it

[median monolith]

(Trying to train a voice model on Applio collab) I genuiely do not know what im doing wrong. yes, im not using the "GPU" thing since its time limit has ended T-T, and the video quality is dogshit because I didnt want the file size to be heavy.

[latent kettle]

Because you are not using gpu

[median monolith]

good point, but I also did the same process when I still had it

i may be wrong, but apparently the problem is with the audio itself somehow? since the log says "no wav file found" and "not enough data present in the training set", but idk how, it has more than a second of duration, its put correctly in my drive folder, and put its path in the dataset path box. the log also states an error regarding an "attempting to register factory for plugin" thing. not sure if this matters or not, but still pointing it out.

[median monolith]

also, when that time limit ends, what should I do? just wait for it to reset/recover or something? or am I actually screwed?

[latent kettle]

what is your datasets's length?

[median monolith]

1:03 minutes

[latent kettle]

thats all you can get ?

[median monolith]

yep, the character im doing a model only has that ammount of "unique" usable audio

[latent kettle]

if you have all the model files in your drive, you can continue training after 24hrs

[median monolith]

and reapeating this 1 minute source audio until i have a 10 minute one doesnt seem to be a good idea

[latent kettle]

naah,

just reduce batch size to 2

also try to train on kaggle. kaggle provides you 28 hours of runtime per week

you can either utilze it in one day or 2 days or in a week

[median monolith]

if you say so. I just put it to 4 since the docs state that "if your dataset is short (around 2 minutes or less) put it in 4".

[median monolith]

not sure wich one, but there they are

[latent kettle]

everything including g and d checkpoints logs eventfiles etc

[latent kettle]

yeah it can be put on 4 if your dataset is super high studio quality. actually you can experiment with it,

[median monolith]

alright, so when a day passes, and have all these necessary files, what should I do then to continue my training?

[median monolith]

also yes I will try this if Applio doesnt end well in the end

[latent kettle]

just put the same values. like name sample rate etc.. and uncheck fresh traning then put same batch size and traning parameters. and start traning

[latent kettle]

i mean dont use applio on colab, use it on kaggle

[median monolith]

yea sorry, i should have said "if Applio Colab doesnt end well in the end"

[paper bloom]

hey is there a way to have a better accent on wokada

like sometimes it wont pronounce the L

[frigid stone]

I fixed the error in the end by uninstalling tensorflow and tensorboard using pip in the conda environment then reinstalling them

In the codename fork for Applio, there's no g/total graph on tensorboard, there's only a generator_total graph
These are the same thing, right?

[median monolith]

thats strange, theres no accelerator option to change

[frigid stone]

Thanks!

[light latch]

Does anyone have the Google Collab link to make the ai covers?

[median monolith]

do I need to say I did everything the docs said and I still get stuck :´)

[median monolith]

im pretty sure I did but alright, m a y b e

oh wait... I was supposed to replace the literal word "token" on the thing with mine... now I get it, thanks for help :>

[median monolith]

Sadly, the exact same problem I had on colab still happens on kaggle. Yes I set the batch size to 2 and I didnt let the site open for 28 hours

[flint anvil]

ok chat i want to start making models cuz i just cant find rly good models with realistic qualities, where do i start? like how do i set everything up (im familiar with programming just tell me what to download and start with if you can)
assuming i can train locally which i will likely do

[next mist]

any idea when RVC 3.0 will be released?

[crude flame]

i mean we technically have a "community v3" not a official v3

this "community v3" has a new embedder and thats what rvc-boss wanted for his v3 but he never got around to it

[dark ginkgo]

I cant seem to use my flux on forge ui since upgrading to my rtx 5060ti

[modern surge]

-colab

[patent trellis]

📒 Google Colab Notebooks

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

• **Applio**

by IA Hispano
Google Colab

• **RVC Mainline**

by Hina
Google Colab

• **UVR5 NO UI**

by Eddy
Google Colab

• **UVR5 UI**

by Eddy
Google Colab

• **Wokada Deiteris Fork**

by Deiteris & Hina
Google Colab

• **Hina's Modified Original Wokada**

Google Colab

• **RVC-AI-Cover-Maker-WebUI**

by Shiro & Eddy
Google Colab

• **FaceFusion UI**

by Nick088
Google Colab

• **FaceFusion NO UI**

by Nick088
Google Colab

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

by Jarredou & Makidanye
Google Colab

[opaque sparrow]

How can I train using my own voice recordings to always output some other voice model? like this one https://www.weights.com/models/clroz1aic012sjmfug54yft0u
I have a 9800x3d rt7900xtx
if I train it locally how long would it take to get a really good model for Deiteris' W Okada Fork?

[opaque sparrow]

what's confusing about it?

is my esl english confusing? I want to train a voice model on my voice to the target model, without messing with the tune, index, pitch in the UI, or having to speak at a certain volume. If it's not possible I don't know

[opaque sparrow]

Ahhh, I see, it's ever explained anywhere, RVC does not need to be trained on your voice specifically, it turns any audio input to the target voice

Grok:
While not required, there are scenarios where including your voice in training could be considered:
Custom Fine-Tuning (Optional): If the model doesn’t sound natural with your voice (e.g., due to extreme pitch differences), you could fine-tune the model by adding a small dataset of your voice to improve compatibility. This involves:
Recording 5–10 minutes of your voice.

Fine-tuning the existing model using RVC-WebUI with your voice data to adjust the model’s mapping for your specific vocal range.

This is advanced and rarely needed for general use.

Improving Robustness: If you have a unique accent or speech pattern, fine-tuning with your voice can help the model handle your input better, but this is typically unnecessary for pre-trained models designed for broad compatibility.

[jovial kraken]

How can I run rvc in kaggle?

[latent kettle]

https://docs.aihub.gg/rvc/cloud/applio-kaggle/#applio-kaggle

@simple ore Is it over fitting? I mean it increased in last 100 epochs. The blue circle represents 200 epochs and the final point is 300 epochs. There is no loss

[latent kettle]

How to look at them?

More loss = good quality?

Okay. But how to analyze it ?

I see

I stopped training

Mb

The dataset was very poor and I didn't expected good results. So ya, that's fine

[runic kraken]

what setting refers to making voice more understandable?

[dry sable]

need a bit help here
for some reason i cna't paste images
https://colab.research.google.com/github/w-okada/voice-changer/blob/master/Realtime_Voice_Changer_on_Colab.ipynb#scrollTo=86wTFmqsNMnD the default way is giving some errors

Installing pre-dependencies...
ERROR: Could not find a version that satisfies the requirement faiss-gpu (from versions: none)
ERROR: No matching distribution found for faiss-gpu
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 261.0/261.0 kB 7.1 MB/s eta 0:00:00
Preparing metadata (pyproject.toml) ... done
Building wheel for pyworld (pyproject.toml) ... done
Installing dependencies from requirements.txt...
ERROR: Ignored the following versions that require a different python version: 1.21.2 Requires-Python >=3.7,<3.11; 1.21.3 Requires-Python >=3.7,<3.11; 1.21.4 Requires-Python >=3.7,<3.11; 1.21.5 Requires-Python >=3.7,<3.11; 1.21.6 Requires-Python >=3.7,<3.11
ERROR: Could not find a version that satisfies the requirement onnxruntime-gpu==1.13.1 (from versions: 1.15.0, 1.15.1, 1.16.0, 1.16.1, 1.16.2, 1.16.3, 1.17.0, 1.17.1, 1.18.0, 1.18.1, 1.19.0, 1.19.2, 1.20.0, 1.20.1, 1.20.2, 1.21.0, 1.21.1, 1.22.0)
ERROR: No matching distribution found for onnxruntime-gpu==1.13.1
Successfully installed all packages!

is there any way to fix it?

[hot lagoon]

Why we doing loss_avg_50 charts now then just gtotal?

[dark ginkgo]

what command should I use? Everytime in the past (and currently) involving command line for python does not install

pip install --pre torch==2.8.0.dev20250605+cu128 torchvision==0.23.0.dev20250605+cu128 torchaudio==2.8.0.dev20250605+cu128 --index-url https://download.pytorch.org/whl/nightly/cu128

command I tried to run

also should I update my pip?

I have cuda 12.9, what version of Cuda do I need then?

or do I need to somehow update the ai softwares? Wasnt they supposed to auto update?

Before I do this, can anyone verify if this would work?

google gemini seems to want me to install the 12.1 version instead of 12.8 since it believes it's forward compatible apparently and stable. Like it recommend to not use the nightly build due to "stability issues"

pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

but I did get this after pushing it for cu128

[prisma kettle]

@simple ore Here's the inside of the Replay folder in AppData

https://www.weights.com/replay got it from here

[granite python]

guys,i have [Voice Changer] Pipeline is not initialized.
how do i fix that

[brittle wing]

could someone help me with setting recommendations? for the clearest least robotic sound?

[viral mason]

what do you use to clean datasets?

[brittle wing]

i have 0 clue what that means !

can i send u a screenshot of my current settings?

[viral mason]

of course! my dms are open

[dark ginkgo]

still not working

I updated to the one you said

[pine zealot]

hi guys, can anyone tell me where can i upload the voice models i downloaded?

i have no idea how i use these models

[jaunty shale]

I use Kaggle mainline and in logs I accidentaly removed folder named "mute"

should I be worried?

[latent kettle]

Put that in logs folder

[pine zealot]

sorry i mean if theres any app or something i should download

i have the models but i dont have any app or program to run it

[naive furnace]

Hi, is there any AI RVC that works with 5080/5090? Tried install Applio, OpenVoice, codename-rvc, spend half of the day trying to bypass python compatibility with those gpu's and nothing, it doesnt work.

[dark ginkgo]

so how would I do this? like in the folder where forge is?

[hearty viper]

hey, is there a good voice changer i could use without a good gpu?

colab gives me an error when i try upload a voice model

[dark ginkgo]

yeah

just type run pip install?

or type what you had ealrier

#✨│ai-help message

already done that before

I did that then did what you told me to ealier

it installs for system wide

so all the other ai software works now but not forge and not comyfui

is there not an updated version of forge? if automatic1111 is working, not forge or comfyui,

flux work with it now?

wait how? I got forge becuase at the time, automatic 1111 is incompatibel