✨│ai-help

[restive kiln]

A bit ocnfusing, dunno what im looking at

[wicked fractal]

could someone probably help me out. I try to set up a local claude code with a free model on VSC but im getting an error if i type on the terminal; claude . Im surely missing a step im not a coder im just new in this space

i just watched a video and did everything that the guy said but i think im missing some basic stuff

[wicked bane]

ok

[wicked fractal]

wow how helpfull

[wicked bane]

Thanks.

[proven hill]

basically you need to import

[wicked fractal]

import what

[proven hill]

hope this helps!!

[wicked fractal]

bro is trolling my life

[proven hill]

your life?

[wicked fractal]

yea

[proven hill]

why

[wicked fractal]

cause i needed help and u troll me

[proven hill]

im not trolling

[wicked fractal]

what do i need to import i cant read ur mind btw

chat gpt said i should install an ai extension so that vs code can connect with openrouter

[proven hill]

yea

[wicked fractal]

can i probably send u a video its just 5 min but im surely missing smthing

and maybe i can get ur opinion on it aswell maybe u know better options

[low shard]

W8ights isn’t a real server, this looks like a scam, don’t trust it, weights closed and there’s no official continuation

many staffers from there started bomb promoting here and got banned, i think i need to make an announcement on this later

rvc is retrieval-based-voice-conversion, not realtime voice changer

elaborate your pc gpu, your pc os, what are you trying to do and any tutorial link you’re using

[low shard]

that looks like an old version of wokada

[proven hill]

can i get an invite?

[low shard]

please don’t troll users

[low shard]

invites can’t be shared in the server, and its a scam

[proven hill]

i wanna know what type

[restive kiln]

• RX 9060XT
• 25H2 Windows
• Deiteris Fork

[gentle granite]

not really but the filters on other generators are too restrictive

[reef monolith]

Hi, guys I am trying for 1 week now to prompt a UGC charakter who is talking words out without mistake in german but it doesnt matter what I do it dont work out. Did someone here face the problem and can help me pleaseee.

[teal portal]

OK thanks

[teal portal]

I can DM it to you

[low shard]

Could you DM me too the invitation? We are going to investigate in the scam

[low shard]

anything NSFW related isn't allowed

[desert notch]

hello guys i need a help

is there anybody free

[low shard]

This is a General AI Discord Server, please elaborate:

• your pc gpu
• your pc os
• what are you trying to do: TTS, AI Covers, E Girl Trolling / Catfishing or Roleplay
• the tutorial link used

[desert notch]

i was trying to find out some script to video generating

ai

as i use voice tts model for text to speach

[proven hill]

sure, thank you

[desert notch]

so i just need some thing which will create or gen the video or it will atleast help me

[mellow oasis]

Hello, I'm trying to start making videos on TikTok, with the character of Makise Kurisu, but I don't know where to find a voice of hers to put on the character, does anyone know how to find voices of anime characters for free? I have seen many people on TikTok who have them and don't pay anything

[viral mason]

You can check here https://discord.com/channels/1159260121998827560/1175430844685484042 or here
https://voice-models.com

But if you don't find anything you like you can make one yourself, it's free and easy

[dusky island]

Small question i want to isolate main lead vocals from a recording, what model can be used best to remove crowds and backing vocals in UVR?

[viral mason]

I don't use local uvr but MVSEP does the same thing as it and I personally use this model

it's pretty good from what I've seen

[dusky island]

Ah okay, yeah im already in the cue atm XD

[viral mason]

mainly for bg vocals but I believe it works on crowds too

[dusky island]

okay how do you know which models are the correct ones to use, because i have to wait in line behind 231 other people. so you know if i know which one to do i wont waste time figuring it all out

is there somesort of guide that tells you what each model is for specifically or is it just kind of trying and see what works really?

in other words whats really the learning curve when it comes to picking models?

[viral mason]

make sure u make an account to lessen the wait time

it's free

anyway for vocal separation in general I use this

and reverb I use both these two, stereo seperation first then room reverb

[dusky island]

thanks, although the wait time only lestens when you buy premium it seems

[viral mason]

dang, I just used the seperation tool and swear it was only a few ppl using it at the time

I haven't paid

[dusky island]

yeah mostly in the morning here in europe there are only 1-20 online

but thank you so much for the help!

[viral mason]

one last thing, for echo I use this one

tho if it starts removing breaths I use deecho-dereverb

[daring sail]

I have a problem with the image, I use flux2 but in the place where I used inpainting I have a loss of quality, the picture in that place looks a little blurry, how can I fix this?

[feral crater]

How would you do it in order?

To get clean vocals as much as possible?

[viral mason]

For cleaning a song I use the bs reformer vocal separation, karaoke (background vocals) stereo reverb, room reverb, then finally de-echo

@feral crater

[gusty merlin]

anyone have like a bunch of rapper rvc? willing to pay

[viral mason]

Payed commissions aren't allowed here, but you can ask in here https://discord.com/channels/1159260121998827560/1159289738314919936 or make a model yourself, it's free

If you're just looking for a model then look here
https://discord.com/channels/1159260121998827560/1175430844685484042 or here https://voice-models.com

[feral crater]

Thank you

[viral mason]

You're very welcome

[mint snow]

anyone got a website creator?

the best one (cheapest)

[leaden iris]

When training a voice model, is it better to include or exclude vocalizations that aren't words or phrases? like grunts, screams, mouthy noises and such?

[viral mason]

I personally keep them, it's best to make sure there is more speach tho than those souds

[leaden iris]

Fair fair, I figured as much. I just wasn't sure if it'd enhance or mess up the dataset

Ty! ^^

[viral mason]

ur welcome! ^^

[viral mason]

<@&1159293140440723499>

[last wolf]

Hello y’all,

I’m planning to make a personal FitzyVA model (mainly for Cyn/Cynessa and Mel LARPing, mainly the latter though), and I was wondering, which program is the better option for expressing emotions when using RVC, (regardless of the dataset used?)

Personally, I think Vonovox is really good and has significantly lower latency. Meanwhile, Applio is also a solid option and works as more of an all-in-one program, though it does have some flaws in certain areas. But even then. It's really good. And from what I was able to gather. Vonovox is superior for real-time emotional expression. I'm not too sure though. Cuz I did test out both and got mixed results

So which one is better for expressing emotions and stuff, not just basic talking?

[analog obsidian]

none because the model doesn't learn expressions in the training, it learns linguistics, pitch and spectograms
the outputs can sound a bit monotonous because the RVC core was primarily designed for TTS, although it did come with an option for voice conversion inference (without using F0, so it’s also quite monotonous)
rvc is a kind of hack that removed all of the tts stuff and improved voice conversion by adding f0, an self-supervised embedder (contentvec/hubert) and replaced the hifigan vocoder with nsf hifigan, although in older versions it is possible to use the standard hifigan, without f0 (it's worse)

tldr; rvc is kinda like an enhanced tts, but with the typical 2022 old tts flaws (not lying, the rvc core is that old), like not being able to be expressive enough
more modern TTS can do pretty good and realistic expressions because they learned them in their training, that tech did not existed when rvc was made/was too heavy to run in realtime
realtime clients like vonovox, applio, wokada etc can't fix architectural issues like these

https://media.discordapp.net/stickers/1222246464932020264.gif

[scarlet thistle]

Did you get banned

I saw you joined and aren't there anymore

[viral mason]

that isn't a helper, ping the helper role for help lol

anyway what voice changer are u using, what gpu do u have (nvidia or AMD) and what are u using it for?

wdym?

[last wolf]

Alright, much appreciated

[viral mason]

I have a cyn model

one without effects so it sounds just like the va

it's not public here but

yknow

[last wolf]

thats kinda cool actually, i like when people include the actual va speaing for their voice models instead of just relying on the character clips

[viral mason]

well she did basically make all her voice lines she recorded public on her yt

plus the advertisements for merch

[last wolf]

yeah, thats actually smart

can extend the few lines cyn has in the show

[viral mason]

ye

[viral mason]

if you'd like I can send my current Cyn voice to u

[last wolf]

if youre okay with it, sure

oh wait i got dms off

add me rq

[viral mason]

yuh!

I don't mind sharing my models

[leaden iris]

I just finished training a model with applio and I'm getting an index file the size of 500mb, every index i've seen from others was 50mb, anyone know what's up?

[analog obsidian]

most likely you used the faiss algorithm to train your index
move/delete the index file, go to the training tab again, put the name of your model, scroll down until you find the training options, click "advanced settings", in the index algorithm select "kmeans", click train index

now the index will be smaller at the cost of slightly less accuracy

[leaden iris]

Ooh I see, I'm fine with a slight loss, I just don't want a model file to be paired with an index 10x it's size haha

Tyvm!

Huh.. generating an index with either Faiss or KMeans both results in a 500mb index file

[analog obsidian]

huh that shouldn't happen lol

[leaden iris]

Could it be because I have a huuge dataset?

[analog obsidian]

i mean yeah thats the reason, kmeans should give you a very tiny index file, i have trained index files of 60 hours datasets and with kmeans they're still very small lol

[leaden iris]

The dataset is about 50mb total but the amount of individual files is eheh.. 1875

[analog obsidian]

im preprocessing some 3 hour dataset i have, let me see if i get a tiny index

[analog obsidian]

hold on, did you deleted the previous index, right?

https://media.discordapp.net/stickers/1222246464932020264.gif

[leaden iris]

Mhm yeah, both times i deleted the index, swapped the algorithm, then generated

[analog obsidian]

waaait ur right, this thing is generating huge indexes

im looking the code 1 sec

[analog obsidian]

yep it's bugged

i fixed it

i love applio

when i clicked kmeans it was generating a simple faiss file

https://media.discordapp.net/stickers/1222246464932020264.gif

and now i got a 30mb file

bruh

[analog obsidian]

rvc > train > process

no need to restart the gui, after you replace the file click kmeans again and train the index

[leaden iris]

Oh huh that's silly, fair enough, thanks a ton!

[analog obsidian]

hope that works for u

worked for me

[leaden iris]

yep! it generated a 30mb file this time, it works!

[analog obsidian]

yay

https://klipy.com/gifs/cat-leaf-1--k01KQ981TYH2A9PFCS1WJY88SNA

[gentle granite]

it's not nsfw my guy

I just need less filters

[low shard]

Wdym?

[low shard]

#1175430844685484042

[low shard]

*paid rvc voice model commissions AREN'T allowed

[low shard]

What?

[proven hill]

i left

[plucky crown]

Hi

I7 13th gen 13700
Msi mobo mortar b760m
32gb DDR4
3090 GIGABYTE OC 24gb
2x1tb kingston mv3 ssd
1000w PSU

This will be the pc i will buy
Is it good

Is it enough need help asapp

Helpp

[simple ore]

big dataset with unique sounds - big index

[simple ore]

no

50mb index is a model created with like 5 min audio

big chonky model

[analog obsidian]

yea but kmeans is not working currently in applio, i had to fix it lol

[simple ore]

so Blaise fucked up something extra lol

i dunno why even 'Auto' is needed

just index_algorithm == "KMeans" or big_npy.shape[0] > 2e5

okay, made the fix

anyway, my point stands - big dataset, big index and it is actually good

[analog obsidian]

No one said otherwise; I said it wasn't normal to have a 500 MB index when selecting kmeans.

https://media.discordapp.net/stickers/1222246464932020264.gif

glad it's fixed now

[loud rivet]

Hey.. everyone. I wanted to ask what is the best app for RVC conversions? I use replay.. and sometimes the voice just cracks up somewhere.

And I dont have the GPU required for some apps that I saw on YouTube and stuff.

So if anyone can help me out here.. I will be great!!

[low shard]

what's your pc gpu and os?

[loud rivet]

It's a Microsoft pc

I dont really know the GPU

Because I am a noob... and I only ever used hugging face and uhh replay

Yea it's a Microsoft pc

You can google it.. because I cant upload the pic here..

[viral mason]

Open task manager and click the performance tab

U can check the gpu there

[loud rivet]

Ok one sec

[viral mason]

If it's Intel you can't really run anything good locally

[loud rivet]

Wait sorry ppl I cant find it one moment

[viral mason]

I'd show u how but I'm not on my pc rn I'm still trying to wake up

[loud rivet]

Oh.. sorry I woke you up

[low shard]

is it running windows 11 or 10?

if so, search task manager, go to the performance tab and check the GPU component

or is it a Mac or a Linux PC?

[loud rivet]

Ok found it

[loud rivet]

11

[viral mason]

Lol did that fire wolf guy get banned

[low shard]

alright, check what I told you

[low shard]

NSFW isn't allowed at all

[loud rivet]

K

[viral mason]

I think you misread that question

[loud rivet]

So it says utilization is 1% and shared GPU memory is 0.8/7.9 GB.

While the memory is 0.5/7.9 GB

So...what now..?

[viral mason]

?

[low shard]

It was about uncensoring anime bath scenes and looking for an adult video generation AI 😭

[low shard]

you should tell us the GPU name

[loud rivet]

Uhhhmmmm....

One moment again

How do I tell the name..?

I dont get it.. sorry

[low shard]

!give-media-perms 3h @loud rivet

send a screenshot

[loud rivet]

Ok..

[viral mason]

No one reads the rules 😔

[loud rivet]

Soo true!! Whenever I request for a model they dm. And ask for payment before making the model.

[low shard]

You have got an Integrated Intel GPU PC, it's not good for Local AI, that's why Replay was slow and having issues for you too btw,
It runs on your hardware

[loud rivet]

Oh.. I didnt know that.

So silly of me

So what should I do..?

[low shard]

It would be better if you use Cloud (Remote PC) alternatives with a limited Free GPU time

it would be faster but ofcourse not unlimited

what do you want to do? You could use the suggested cloud, or try to run on your PC CPU which will be slow ofc

[loud rivet]

What do u mean?

[loud rivet]

..I dont have patience.. astgfirullah...

[low shard]

you either still try Applio (RVC Fork) on your PC CPU, which will be unlimited but slow, or use the fast and more suggested Cloud alternatives, which ofcourse won't be unlimited for free becuase you're using someone else GPU

[low shard]

you just want the easiest fastest thing you can get, so like the easiest AI Cover Maker for cloud?

[loud rivet]

Oh.. what do u recommend sir?

[loud rivet]

Oh... what's that...??

[viral mason]

Good thing I don't take model requests

[low shard]

that's not allowed, if you see someone do that, report them by dming @vernal bone

[loud rivet]

Ohhhhhhhh I see..m

Got it I keep that in mind from now on!

[low shard]

you just want the easy fastest way to make AI Covers right? I'm not sure what do you mean

[loud rivet]

Yes

[low shard]

there's a cloud program that automatically separates vocals and instrumentals like Weights did, try the cloud easy version: https://docs.aihub.gg/rvc/cloud/aicovermaker-cloud/#google-colab

read the guide and let me know

[loud rivet]

Ok! Thank you for ur time and patience with me!

https://tenor.com/view/my-hero-academia-boku-no-hero-academia-anime-mha-bnha-gif-17532278

[low shard]

you're welcome

[loud rivet]

oh and sorry to bother again mister. i wanted to ask the link that u shared me. does it also allow music mp3 files to be uploaded? because i make AI music. and want to make JJK characters sing them..

[warped laurel]

can anyone help me in ai

i always wanted to do something i do not know what to

i always wanted to make money with ai

pls anyone help

dm me

and message ai so i can understand

[low shard]

this server isn't really about "1 click easy free money with AI", there isn't anything that will garuantee you free easy money in 2 seconds

[plucky crown]

Help the seller sent me this for the pc
I7 13th gen 13700
Msi mobo mortar b760m
32gb DDR4
3090 GIGABYTE OC 24gb
2x1tb kingston mv3 ssd
1000w PSU

Should i get it?

[viral mason]

What kind of PC is this? The 3090 is good if it's the RTX 3090

Is it a desktop or a laptop

[plucky crown]

Pc

[viral mason]

Intel for the CPU is kinda just ok, the other stuff I'm unfamiliar with as I'm not super into computers

How much will it cost?

[plucky crown]

Its 901usd

Equivalent in dollars for the whole pc

This is the 3090 gigabyte brand , its my first time with gpu pc so i need help

[viral mason]

Depends on what u want to do with it

[plucky crown]

Ai and vidoe editing and fl studio

[viral mason]

Could you expand on what u mean by ai, there's tons of stuff like video, images, rvc, ect

[plucky crown]

Rvc

[viral mason]

Video editing and fl studio can be done for sure on a 3090 just fine

[plucky crown]

I only have 900 usd

I can add the 1 usd so i could buy it

Should i get it or nah i would regret for the rvc voice ai in this server

[viral mason]

Vonovox should run just fine on a 30 series, I have a 5070ti and it's also pretty good there

You want real-time right?

[plucky crown]

Yes

[viral mason]

Yea the PC you're buying should work just fine with everything you want to do

[plucky crown]

5070ti should i get it instead of 3090? But 3090 has higher vram wouldnt that be best for ai voice training

Since 5070ti is newer

[viral mason]

Hmm

I mean I train on cloud bc I'm lazy so I dunno how it is with training voices locally but I'm sure it works just fine

But I definitely think if you can afford it to get a 50 series Nvidia gpu over 30 series

[plucky crown]

But i also want a text ai or train ai they said u need higher vram for that

[viral mason]

Hm

[plucky crown]

I can only afford a 3090 for higher vram

[viral mason]

I'd ask Lyery as I'm not as smart with PCs and rvc as him

Might get better insight from him

[plucky crown]

@analog obsidian

We needHelpp @analog obsidian

[viral mason]

Just ping once

[analog obsidian]

you want it for both realtime and training? hmm, the 5070ti is super fast and should give you good performance in realtime

it's true that you need more vram if you want to train models, but in rvc you don't really want to go above batch size 8 (about 5gb in fp16 mode, 7gb in fp32)

realtime inference uses 2gb of vram iirc

hmm hard decision... the 3090 might be useful in cases where you play unoptimized games that uses lots of vram

[plucky crown]

Yes also the 3090 price for the whole pc the owner/seller is able to get it to 901 usd so i could buy it right now

[analog obsidian]

also, seems like faster gpus do get sligthly less delay, so in theory the 5070ti is technically better for realtime, but yea, the vram

also is way faster than the 3090 for training models

[plucky crown]

Its my first seeing in marketplace a whole 3090 pc less than 1000usd so should i get it, i only have 900usd the seller lowered it to match my savings

[analog obsidian]

more vram helps getting less stutters too

[plucky crown]

Yes

[analog obsidian]

yea tbh the 3090 is a good option despite being older just because it can provide a more smooth experience with not that much stutters

[plucky crown]

Should i get it then i wont regret it for real time voice/ the ai voice thingy in this server @analog obsidian ?

[analog obsidian]

im not 100% sure, because the 5070ti is indeed faster and will give you better fps

[plucky crown]

I cant afford a 5070ti pc i only have 900usd for a whole pc

[analog obsidian]

ah then the 3090 is enough dw

rvc was made in 2023

late 2022

[plucky crown]

So 3090 can handle it even if its older?

[analog obsidian]

yup

[plucky crown]

Thanks

[plucky crown]

Is there anything else i should ask other than this

[analog obsidian]

thats enough for rvc

[worldly plank]

whats the directory / name of this program again

[viral mason]

what program

[leaden iris]

Is it better to have separate audio files, or combine them all into single file for voice model training? Or does it not matter?

Does it matter if it's tons of files? Like hundreds or thousands?

[mighty hinge]

Hi, i make videos for youtube, bit i want to ad the voices in applio, how can i use the TTS? Someone have the link to use it?

[left totem]

The best ai modal, hi ive tried to find the best modal overall, but i dont know what i should use i dont like cloude at all ive used it for 5 mouths is there any other?

[hallow thistle]

-rvc

[patent trellis]

📚 RVC Documentations

AI HUB Docs

https://docs.aihub.gg

🍏 Applio Docs

https://docs.applio.org

[hallow thistle]

Likely a realtime voice changer.

[viral mason]

cloud?

[hallow thistle]

NVIDIA GeForce RTX 3090 is already enough, potentially a bit overkill, for simple RVC inference, but for voice model training, there's an advantage. RTX 5070 Ti is way newer, but RTX 3090 has more VRAM. Kingston NV3 is a budget SSD, uses QLC NAND which gives more capacity but slower than "TLC NAND", so if you were to do heavily files moving, either Kingston KC3000, WD Black SN850X or SN7100 might be better. Ensure that motherboard's BIOS is up to date. Should you get that one? It's your choice still if you have that budget.

Another concern is that one power supply. That power supply is stated 1000w, but always check its brand, model and health because you might not expect it to smelly or turn off suddenly during runs. The DDR4 RAM speed is not stated, but probably a bit slower than DDR5 anyway.

[low shard]

This is a general ai discord server, what program?

[low shard]

there’s a tts tab, do you already have applio? If not, what’s your pc gpu and os?

[daring sail]

Hi, I have a problem with Flux2 Klein. I'm trying to use inpainting to remove text, but in the areas where I applied it, there's a blur issue, it looks kind of smeared. The second problem is that in those areas the color changes compared to the surrounding environment. Is there any way to fix this?

[young mango]

Hello i can help you to fixed this

[daring sail]

Well, how can this be fixed?

[young mango]

Let talk in dm

[hallow thistle]

Why direct message?

[young mango]

Yeah that is a common thing with inpainting the blur usually comes from denoise being too high so it over rebuilds the area instead of blending it and the color shift happens because the model is reinterpreting lighting in that masked region

Try lowering the denoise so it keeps more original detail keep the mask tighter around the text and do it in two passes first remove the text then a lighter pass to match texture and color

[daring sail]

If I change the prompt, won't it improve the situation?

[daring sail]

Do you mean lowering denoise on ksampler? I reduced it from 1.00 to 0.75 and now it doesn't remove anything, it just changes the color in that area.

[median crag]

yo does anyone have a working rvc colab? i cant find one

[lavish lantern]

sorry to bother anyone but is there a RVC of sash lilac from freedom planet to download anywhere ive been searching for so long and what is here are sadly gone since weight sadly had its farewell and applio is giving me 404 thing

[viral mason]

have you tried applio on kaggle? it works fine for me

[lavish lantern]

im new to this sadly and not the smartest there is i know of applio but not kaggle im just looking for a working download for sash lilac to put on applio

[viral mason]

not really sure what sash lilac is but kaggle works well, it's a cloud option that can use applio

[lavish lantern]

oh my bad i didnt mention this along side my first text she is a character on a game similar to sonic also i took the time to look up kaggle i know what it is now ill keep it in mind all im looking for is a working downlod since all the rest ive tried to look for dont work anymore mainly use for tts on applio

[viral mason]

ah I see, idk at all about getting applio working locally but I know how to use the interface and train models ect

best bet would be looking here

-rvc

[patent trellis]

📚 RVC Documentations

AI HUB Docs

https://docs.aihub.gg

🍏 Applio Docs

https://docs.applio.org

[viral mason]

it should be in the applio docs

[viscid minnow]

Any good, realistic models you’d recommend?

[viral mason]

not really sure tbh, just check around https://discord.com/channels/1159260121998827560/1175430844685484042 or https://voice-models.com

[viscid minnow]

[viral mason]

[lavish lantern]

i see i was hoping someone can share theirs before the it cant be downloaded anymore happened

[viral mason]

wdym?

[lavish lantern]

i found older downloads here but it all lead to 404 not found or just dead links for sash lilac so i was wondering if anyone have it they can share it with to me but its understandable if no one have it she isnt a very well known character and so is her game

[viral mason]

is this the character you're looking for?

I found a 13 minute voice line video of her that could be used to train a model

[lavish lantern]

yes but her voice to use on tts

[lavish lantern]

ill see how i can make one on my own if my brain can handle it if and if its easy and if not ill have to wait for someone to make it or can share the old one thanks for your time for helping me out and sorry i didnt mention the clear idea of who she was and is and from

[viral mason]

I have a video I made on Kaggle's applio if you'd like to try that

https://www.kaggle.com

[viral mason]

I did skip making an acc and verifying but you do need to verify by phone or one of the other options it gives or the notebooks do not function

[lavish lantern]

Oh cool thanks ill keep it in mind for now

[mint snow]

AI that creates websites

and i need one for domain

[twin snow]

DISBOARD: The Public Server List

Hi! I am a bot for DISBOARD (https://disboard.org)

COMMAND LIST

/help: This!
/bump: Bump this server.
/page: Get a link to the server's DISBOARD page.
/invite [channel]: Set invite to this channel. If [channel] is specified, create an invite for that channel. (Admin only.)

How do I add my server to DISBOARD?

1. Login on DISBOARD website,
2. Go to Dashboard,
3. Click "Add New Server"
   Fill out your server info and save it. You will be redirected to Discord's authorization screen. If not, click the "Add Bot" button on the server edit page.

Need help? Join the support server.

[viral mason]

<@&1159293140440723499> hacked account

[viral mason]

<@&1159293140440723499> promoting nsfw

[hallow thistle]

I once used Adobe Dreamweaver to create a HTML page.

[hallow thistle]

Check out my Neco-Arc model in #1175430844685484042.

[kind ether]

Following the RVC Applio tutorial on the ai hub docs,
I have a RX 6500 XT
Shader ISA: GFX10.3 (gfx1034)
Do i follow the top one or bottom one?

If its the bottom one do i follow this or the other gpus one?

[kind ether]

Nevermind, checked the repo under the "Other GPUs" tab, it is in there

[wintry scaffold]

hey guys

im new to this ai stuff

i tried running llama 3.1 8b version on ollama locally on my system

i said "hi"

10 seconds later i heard my pc fucking screaming

then it shut down

what do i do?

Full GPU Name: NVIDIA GeForce GT 610
Operating System: windows 11

[wind path]

Hello! I just came here on one question: I have Projects (folder organization) on the ChatGPT Free plan, I also live in Poland 🇵🇱 so I don't have ads (yet, lucky me).. I do have more features, and I wonder - Should I move to Go or Plus? I usually run not a lot msgs per day, but I do give some kind of context (not a lot, but some details, I'm not an advanced user) and the weakest model is still enough for me I guess? I'm also using uBlock, I wonder if it works with ChatGPT lol

[wind path]

don't even try to run LLMs

This GPU is so weak bro

[hallow thistle]

This GPU is too old to run any AI locally.

[low shard]

yeah that might be too old to run locally

not sure if it the GPU would even get detected, probably running on the CPU with RAM (and you probably have 8gb of RAM)

unless you try extremely small models, it's best you try cloud

[wind path]

in free time, pls answer my question 😊 It's important to me

[low shard]

uBlock Origin (and lite) probably should work fine

If you're not running into any any issues with your current free tier, it might be best to keep using it

You should decide on it depending on your needs, because if you're doing just small things and it works fine, there isn't really a need

But if you needed more powerful models, for like coding or complex math tasks, or are running out of free messages, you might want to try other services like Gemini, Claude, or subscribe to chatgpt

Like for example, you wouldn't need a 5K super powerful pc if you're just going to use it for Microsoft Word, if you get the analogy

its your money and your needs, so it's best you take the decision valuating what you want to do, instead of blinding just buying a 2nd car you might not even use because your 1st car works fine

btw I would suggest to learn how to prompt engineer it better, giving it more clear instructions and info can help alot for results, especially on complex tasks

[wintry scaffold]

I'm thinking about a PC upgrade

below 100,000 rupees tho ($1,053)

what should i get?

what specs should i upgrade

[hallow thistle]

If you mean to upgrade a GPU in your current PC, there's NVIDIA GeForce RTX 3060, but there's another better workaround. With $1000 budget, you might rather expect a complete newer mid-tier PC with more recent specs.

[wind path]

4th paragraph: "But if you needed more powerful models, for like coding or complex math tasks.." - math is not for me, but I learn to code, often just asking questions about interesting settings to tweak in my PC or what tool to learn, that will suit my needs.. BUT I rather want to achieve some assistance for decision-making, don't expect to AI be my doer

I'm not a programmer tho, I don't work with ChatGPT, it's rather a set of hobbies, not PRO work, so just chilling out.. want to improve myself

[wind path]

Sounds interesting - I recently did ask my GPT for this topic, and he got me pretty satisfied with it.. He gave me that example, if you would like to rate this as a real person.. I want to objective feedback, another perspective 😊

NOT
What headphones to buy?

BETTER
I'm looking for headphones for work (long sessions, comfort, and good sound quality). What are some reasonable options for a mid-range budget, and what are the real differences?

===
I suffer due to ChatGPT prices in PL 😭 The Plus plan is 20$, it's going to be 72PLN by the currency calculations, but it says, like, 100PLN (which is 25$ in this price, and currency converter says 91PLN).. It's weird..

[wintry scaffold]

okay

[low shard]

you could get an rtx 3060 12gb vram or rtx 4060 ti 16gb vram

[wintry scaffold]

thx

[karmic trellis]

guys whats this. this happens when i change index rate and when i click on start in vonovox beta

[viral mason]

Rupees like, from the legend of Zelda?

[inland carbon]

Hi. I'm trying to run gemma4:26b through Continue on VSC

I can't add it to models tho no matter what's in the config

[hardy yew]

.\runtime\python.exe -m pip install faiss-cpu

run this command in vonovox directory

[vale harbor]

Does anyone know which version this is? I reformatted my computer and can't find it.

[viral mason]

This is Wokada tg fork

But what gpu do u have, there's an Nvidia version and AMD version

I feel like those tik toks where they go "what kind of fruit is this"

[vale harbor]

nvidia

[viral mason]

You should try out Vonovox, it's much better than tg fork

I'll get the links rq

https://huggingface.co/dr87/vonovox/resolve/c8034f5f6d50648a8109bb4f847182362e2b779b/Vonovox_beta_17_11.zip

https://software.muzychenko.net/freeware/vac470lite.zip

[vale harbor]

[viral mason]

[vale harbor]

but I liked the one I had before better

[viral mason]

U dont wanna just, try it out first?

[vale harbor]

I don't know how to use it—there are so many options, and I have dyslexia

[viral mason]

It's ok the only options you ever need to touch is the pitch

And block size (controls the delay and quality)

[vale harbor]

[viral mason]

I know it looks like a lot but most of that you don't need to even touch

Btw make sure to import the index

Right underneath the embedder

[vale harbor]

I don't have an index on a model I like

[viral mason]

O ok

So u do have a virtual audio cable right?

Like VB cable or vac lite

[vale harbor]

virtual cable

[viral mason]

Yah, it lets u use the voice changer on discord or in games

[vale harbor]

[viral mason]

It's that second link I sent here in case u didn't have one ^^

[vale harbor]

What should I do?

[viral mason]

Download the second link, extract, then run setup64.exe, and then install driver

[vale harbor]

I already have the first one I mentioned—the virtual cable

[viral mason]

Oh

Nvm then

So for your output in vonovox select your virtual cable

[visual dome]

hi guys im new just wondering how do i use a voice changer

[vale harbor]

[vale harbor]

where

[viral mason]

Where it says output device, it's underneath the one that says audio device settings

[viral mason]

Hey! What gpu do u have (Nvidia or AMD) and what do u plan on using it for?

[vale harbor]

[viral mason]

It's ok take your time finding it

[vale harbor]

now what

[viral mason]

Send a screenshot rq so I can see what changed

[visual dome]

https://discord.com/channels/1159260121998827560/1472273212459974749

i wanna use this just in discord voice calls

[viral mason]

Nice

The important part tho is knowing your gpu, if it's Nvidia you can download Vonovox, if it's AMD you need wokada tg fork

[vale harbor]

It won't let me upload the index, but on the other one it did

I didn't like it

[viral mason]

This seems like it's fine now, if the model is a female and your voice is deep/are a guy try from pitch 3-12 and opposite of you're a girl

And if you're using a female voice and are a female keep at 0 same other way around

[vale harbor]

uh?

[viral mason]

Ok so if you are a guy and using female voice pitch the voice up until it sounds good

If you're girl and using a guy voice pitch down until it sounds good

If guy using guy voice keep at 0

If girl using girl voice keep at 0

[viral mason]

No need to change this it is good at default

[vale harbor]

AEAEAEAEA

Can you send me the link to the first one I mentioned?

[viral mason]

:(

ok

you'll be missing out on the best voice changer but it's ok

[vale harbor]

link pls

[viral mason]

-rt

[patent trellis]

🔊 Realtime RVC

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

• Vonovox

A Realtime Voice Changer with similar performance to Wokada Tg-Develop Fork, with extra features, but it supports only Nvidia GPUs on Windows 10/11 unlike other options that have wider support. and without cloud options

• Local Guide

• Wokada Tg-Develop Fork

A personal fork (modified version) of Wokada Deiteris Fork, it just adds some Quality of Life improvements to it like supporting Spin Embedder and Audio Effects. Don't expect too much about it since the creator made it originally as a personal project.

• Local Guide
• Cloud Guide(s)

• Applio Realtime

A Realtime Voice Changer with similar performance to Vonovox & Wokada Tg-Develop Fork, with extra features.

• Local Guide

• Wokada Deiteris Fork

Deiteris' fork (modified version) of wokada that doesn't get updates anymore.

• Local Guide
• Cloud Guide(s)

⛔ Outdated/Discouraged

These options are not recommended for use.

• Original Wokada

Not suggested, older versions in youtube tuts are even way worse. GUIDE

• RVC GUI Mainline Realtime

The program is worse compared to the ones above, and much less updated. GUIDE

[viral mason]

https://github.com/tg-develop/voice-changer/releases/download/b2397/voice-changer-windows-amd64-cuda.zip.001

https://github.com/tg-develop/voice-changer/releases/download/b2397/voice-changer-windows-amd64-cuda.zip.002

u need these two, download both, and extract the first one, then put the second one in the folder made by the first one

[vale harbor]

That's not it; I already downloaded it.

[viral mason]

wdym

[viral mason]

this is the one you want

that's what I just sent

[vale harbor]

it's different

[vale harbor]

[viral mason]

this one is very old

you don't want that one

I didn't send u that one :(

[vale harbor]

I downloaded it from your link

[viral mason]

did u have anything else running when u opened the one I sent?

[vale harbor]

no

[viral mason]

hm

[vale harbor]

[viral mason]

I am confused now

[vale harbor]

now what

[viral mason]

close anything related to the voice changer then try one more time opening it

I may need to help you in a voice call if this happens again

that might help

[vale harbor]

TY

[viral mason]

you're welcome!

do u need any more help?

[sly elk]

how do u use the text to speech models?

[south wyvern]

Im not sure where to post this question, but it's regarding a1111. Its the program i started with and know it well. I had an nvidia 3060ti (8gb) and it was slow but it worked. Ive since upgraded to a 5060ti (16gb), was so excited to jump back into a1111, but then i found misery and pain, lol.

Is there any way to get it to run on a 50xx series GPU? Ive found some discussion online but its all waaaaay outdated, seems like everyone just stopped trying maybe?!? I currently have a driver that has cuda 13.2 and can find absolutely NOTHING that helps... is there help? Is a1111 just dead?

thats a very fine website, unfortunately I dont feel like digging through endless stuff that seems irrelevant to me in hopes that maybe I might find something that maybe is what I need. I want to ask a person and have an answer. Im just old fdashioned that way

[viral mason]

what is a1111

[south wyvern]

automatic 1111

its an webui, an older one

[sly elk]

how do i find models that are less heavy to run?

[viral mason]

what does this mean

[mint snow]

noted

[white shadow]

Whenever i try starting my notebook it gives this error in kaggle.
It does it everytime. I've tried making a new copy, making a new notebook, restarting my browser

nothings working mannn

[viral mason]

I got this too

[white shadow]

ok good, its not just me

[cold rover]

I dont know which program to run for the voice changer... im kinda dumb

[viral mason]

what gpu do u have (Nvidia or AMD) and what do u want to use it for? just curous

[elder garden]

how to use ai voice changer without using gpu?

[hallow thistle]

Why?

[elder garden]

my gpu is low end

[hallow thistle]

What would you use the voice changer for?

[elder garden]

i want to sound like kanye west

[hallow thistle]

What model?

[hallow thistle]

Not really recommended to run a voice changer with only CPU. Try run W-Okada voice changer on Kaggle instead. https://docs.aihub.gg/realtime-voice-changer/cloud/tg-develops-w-okada-fork-cloud/#kaggle

[hallow thistle]

NVIDIA GeForce RTX 4090 or RTX 5070 Ti? You should set processing unit to your "GPU", extra time to 2.7 s and pitch extraction as rmvpe. Also, output and monitor audio devices look swapped but they'd still work.

[vale harbor]

5070 ti 16 GB

[hallow thistle]

[vale harbor]

[crimson geyser]

where can i install voice changer nvidia

hi

[gaunt saddle]

i got a question if i train a model using a song that got adlibs will the model come out well and will it know how to work with songs that got adlibs'

[low shard]

most models in #1175430844685484042 are RVC models, are you trying to use those for TTS?

[foggy dome]

hii i want some advice how to find the perfect rvc model for me. like how do ppl find the natural ones that will fit them

[hallow thistle]

AMD Radeon RX 5700 XT or NVIDIA GeForce RTX?

[quasi pagoda]

I like Nvidia

A gpu without cuda is paperweight

[viral mason]

@waxen kayak

-rvc

[patent trellis]

📚 RVC Documentations

AI HUB Docs

https://docs.aihub.gg

🍏 Applio Docs

https://docs.applio.org

[waxen kayak]

ok

[viral mason]

Check the applio docs to see how applio works, there are many different options like using it locally on your pc or on the cloud with Kaggle or Google colab

[waxen kayak]

it's a little bit difficult to seetup but thx

[viral mason]

No problem! If you need help ping the helpers role

[white wedge]

"Error: Pretrained model sample rate (40000 Hz) does not match dataset audio sample rate (32000 Hz)." both the dataset and pretrained model are 32k

[proven hill]

are you 100% sure?

[white wedge]

yep

[proven hill]

link to the pretrain?

[white wedge]

im trying to use this pretrain model (32k sr) and my dataset is also 32k sr but like
https://discord.com/channels/1159260121998827560/1492203850747216083

idk and its not only for that pretrain model ive tried other 32k ones and they all give the same error

[viral mason]

if u chose 40k in applio when using a 32k pretrain u should choose 32k instead

[proven hill]

did you select 32k in the interface? what are you using? applio?

[white wedge]

yes and yes

[feral sierra]

i have this too, but only if i try to create a new notebook. existing one works just fine somehow

[viral mason]

I can't get it to work either, this is killing me

[white wedge]

actually wait

i might have a solution 1s

not for that foor my problem

[feral sierra]

okay now i can't even run my older notebooks anymore :')

[viral mason]

https://tenor.com/view/there-is-no-hope-there-is-no-hope-gif-13610811744268717441

[main turret]

So is anyone using https://multinex.ai/memq?

[proven hill]

obvious sponsorship

[leaden iris]

so i'm very much a beginner when it comes to training voice models, are there any like biig tips or things i should be doing when training? like certain settings or things i can do before or after to improve the quality of models? currently i'm using applio and using essentially default settings

[untold marten]

Heya, I understood that chunk size is used to control the delay of the voice resulted from the rvc app and higher means as well better results. Is it supposed to give such a big difference the OS (windows vs linux) and the developer (deiteris vs Tg)?

• On windows, I need to set the chunksize to 192ms at minimum to have a stable perf and set extra to 2.7s
• On linux, I can go as low as 24ms altho it's running tight (sometimes goes yellow and lag spikes in voice). Also 2.7s extra

I also tested different delays of chunksize on linux from 50-100-150-200.... they all sound mostly the same in quality from my perspective. So does it help if I run it at around the same delay like I do on windows or ? e.g 50 vs 192 or 100 vs 192? I mean they do seem to sound the same but can't be too sure and said to ask as well even tho I understand that I use 2 different versions of rvc from 2 different people but wanted to know if it gives any benefit running on the same delay even tho I could go lower.

[feral sierra]

this guide might be helpful if you havent read it yet
https://docs.aihub.gg/rvc/local/applio/#training

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

[feral sierra]

i'm glad I'm not the only one with this issue haha

hope it will work again soon, i like kaggle more than lightning ai

[viral mason]

I cannot even use lightning ai it's too confusing

and colab is dookie with only 4 hours max

[feral sierra]

yeah kaggle is definitely more user friendly and simpler, but lightning ai has very very powerful gpus

their verification process is weird though

[quasi condor]

if u learn it :p

[odd shale]

Probably it's your case because i can still run my old applio notebooks on kaggle

[quasi condor]

its confusing at first but you'll get the hang of it

[odd shale]

But either way, @simple ore maybe you should check this out, tested it myself to check if that error was real and it is.

[quasi condor]

its user friendly but still slow

same speed on colab even tho its using two t4 gpus

[viral mason]

meh I dislike how there isn't even a stop button anywhere clearly labeled for dumb people like me

[quasi condor]

uh

there is tho o - o

[simple ore]

what exactly does not work?

[quasi condor]

if u click the cpu and gpu stats there's a button where can put applio to sleep

[feral sierra]

i could run them untouched earlier today, but as soon as i changed something i get permanent unknown errors trying to start the session

its really strange

[odd shale]

@white shadow claims that when trying to import a new applio kaggle notebook and running it, this error pops up preventing it's startup

[feral sierra]

yeah i get the same popup

[viral mason]

thatr was never explained to me

hmpf

[simple ore]

when does the error happen?

[viral mason]

kaggle has an obvious power button to turn it off

[odd shale]

Some seconds after clicking the power button

Tested it myself to see if i can recreate the error and it seems it's true.

[simple ore]

tf is power button

[feral sierra]

start session

[simple ore]

????

tf you guys are smoking

[simple ore]

install runs

[viral mason]

that lol

[simple ore]

Why are you clicking it

instead of clicking install |>

[quasi condor]

oh yeah that error happened to me a week ago i never told anyone about this error

[viral mason]

no matter what I press in kaggle I get that same error off it not working

start does it, that button as well

run all

[simple ore]

just run install first, when it is finished run the other cell

[odd shale]

That error.. Never happened to me on my current notebooks.

[viral mason]

someone needs to call mr Applio

[simple ore]

dunno, I just tried and it is fine

create a new notebook, then import applio/main/ kaggle thing

[viral mason]

?

[quasi condor]

@viral mason

here's how u turn it off

[viral mason]

that's too confusing it should be always on the front without any extra steps

ty tho

[quasi condor]

wdym its literally simple

[quasi condor]

ur like that one lizard on the movie hoppers

[viral mason]

not sure what that means but I still get the error when making a new notebook

[quasi condor]

https://tenor.com/view/waving-joy-lizard-hi-there-well-hello-there-gif-15622233555234689915

[viral mason]

I love this guy

[simple ore]

okay, that happens when you select

maybe there's a space limit or somethng

[viral mason]

so I just do nothing and let my model die if something bad happens

[odd shale]

Just switch to variables and files after properly starting the notebook xd

No errors popped after switching

[viral mason]

o

how odd

[simple ore]

nothing to do with Applio

[odd shale]

Already figured it prolly wasn't something related to applio

[simple ore]

make a support ticket or something

[odd shale]

Sorry if i looked like a dummy noobies

[odd shale]

I mean, i only wanted to inform that issue in case of anything

Myself i never had issues with that

[viral mason]

I figured how to get it working ty Leo

[simple ore]

fixed now

@viral mason

[simple ore]

fixed

[vale harbor]

How do I configure vonvox?

[viral mason]

how so?

[vale harbor]

[viral mason]

The main thing needed to change is pitch and block size

[vale harbor]

[dire light]

i followed the guide perfectly and the thing doesn't work (using lightning cloud w-okada)

[edgy violet]

Anyone got the TTS Ivy voice?

[leaden iris]

oh i meant like tips for things to improve outputs
i already know how to use the program itself, ive previously trained about half a dozen models on rvc webui
i was wondering if there are any like checklist type steps that you should always do or any extra things you can do to make it better

[analog obsidian]

you can try merging epochs of different training runs
for example, you train the same dataset two times; one with batch size 4 and another with batch size 8, then you merge both of them and see if the quality increases

besides that, you could also try using different batch sizes for the same dataset, maybe batch size 8 is going to work better for that specific one than lets say 4

also,ive seen some use lower lr for the discriminator but honestly everytime i change the default lr things don't sound as good as with 1e-4

perhaps a different pretrain can help too https://discord.com/channels/1159260121998827560/1235952130855010365

[leaden iris]

ooh noted noted, tyvm! ill definitely be trying those
i did see pretrains and that, is it just you get a custom pretrain and tell it to use it when training and it can potentially change the result? i've no clue about them

[analog obsidian]

pretrains are RVC models trained using many hours of audio
each model learned to reconstruct a spectrogram in a different way; for example, the original pretrain learned to reconstruct the spectrogram based on the VCTK dataset combined with pitch shifting.
when you load the original pretrain, what you’re essentially doing is continuing the pretrain’s training with a new dataset.
so if the pretrain has a good understanding of the human voice/spectogram, by continuing its training, your model will inherit all that previously learned knowledge.

well, when you load a pretrain different from the original, you’re continuing the training of another model that learned to reconstruct spectrograms in a different way, you get the idea

https://media.discordapp.net/stickers/1222246464932020264.gif

you can enable the usage of custom pretrains by enabling this

but first you have to download a pretrain G and D files (generator and discriminator)
and place them in rvc > models > pretraineds > custom

[analog obsidian]

and then you just select the previously downloaded g and d

u can try these since they're the most recent

everything has to match, like, for example, legacy core 1.6 requires you to select 32k, hifigan and contentvec in the training tab, otherwise it wont work

[leaden iris]

huh, that is really friggin cool actually! thank you for being so informative with everything

[analog obsidian]

ah and about the lower lr for the discriminator
if you open: rvc > train > train.py
you'll find these
with them you can either decrease the lr of increase the lr of both the gen and the disc
in the past people were experimenting by decreasing the d_lr_coeff to something like 0.5 or 0.7

lr is how fast the model learns

the discriminator is what forces the model to produce realistic audio

i personally don't like to change neither of them (1.0 is fine imo)

[analog obsidian]

sometimes it gets too strong and prevents the generator to produce better quality results, thats why sometimes decreasing how fast it learns can produce better results

i know everything i said is confusing but they're actual ways to potentially improve a model lol

one thing to keep in mind is that rvc doesn't make miracles, if a dataset sounds bad, is going to sound bad always after training it

[obtuse summit]

hi

i need help making an rvc, i have the latest version of applio, i have windows 11 home, 32 gb of ddr4 ram, a ryzen5 5600x and an nvidia rtx 5060 from MSI, im trying to make an rvc for caine from tadc, since in the rvc list there isn't an italian version of it. my dataset is 1:20 minutes long but i don't know what settings to put, i modified it in audacity to make it sound better and exportted it in 48k

[proven hill]

ciao ti posso aiutare

[obtuse summit]

asp che sto pulendo il dataset

ok ho fatto

[obtuse summit]

che impostazioni dovrei mettere su applio

[proven hill]

usa il 32k

[obtuse summit]

è già su 32k gemini mi ha consigliato di usare refineGan

[proven hill]

è abbastanza inutile, se non sbaglio poi avrai problemi con l’inferenza SE NON ERRO

[obtuse summit]

c'è scritto che ha più quality

epoch faccio 300 e salvo ogni 30?

[proven hill]

sì okay ma puoi usarlo solo su Applio

[proven hill]

direi ogni 50

[obtuse summit]

io ho solo applio per fare rvc

[proven hill]

e allora va bene

[obtuse summit]

io ho iniziato a trainarlo, spero che si senta bene, il mio primo rvc, quello di N si sentiva male anche dopo 5 tentativi diversi

[simple root]

Any 1 has suggestion on running AI channel

which AI is best for Animation / frictional Character

<@&1159293204038955078>

[delicate cradle]

@quasi condor hi ^-^

i don know how to get the full image :P

[wraith flume]

what is the software called for these voice changers

[proven hill]

what gpu do you have?

[wraith flume]

3050

[wraith flume]

ik its cheekz but its a laptop

[proven hill]

nah its good

https://docs.aihub.gg/realtime-voice-changer/local/vonovox/

follow this guide

[wraith flume]

k ty

[wraith flume]

its a confusing installation, do u think W-okada would be better

[proven hill]

no.
vonvox is simpler, whats confusing to you?

[wraith flume]

the vc cable thingy and stuff

[proven hill]

its easy just install and restart your pc, sadly its the only way

[viscid minnow]

bruh wtf

mxm

[wraith flume]

i watched this tut now it makes no sense at all

[proven hill]

wdym?

[wraith flume]

i watched a tutorial video

[wraith flume]

what is a 16inch virtule cabe

[proven hill]

you should only follow the text guide from the docs

[wraith flume]

mi confondo tantissimo, non so perche🤦‍♂️

[proven hill]

ah ma sei italiano

scrivimi in dm

[wraith flume]

poquito

[viral mason]

bro was so confused they switched languages

[proven hill]

AHAHAHAHAAHAHAHAAH

[whole mason]

italiana su ai hub?

crazy

[viral mason]

I'm making the model btw

if you're still interested

[boreal creek]

Can I upload and rvc model I made?

[viral mason]

you have to first go through this and if your model is good you'll get accepted to post in voice models https://discord.com/channels/1159260121998827560/1453581938538315847

[boreal creek]

idk what it means by vocoder, i just trained it through rvc, i did it a year ago so im not entirely sure what it means

asks like if i used hifigan or refinegan

[viral mason]

you used Hifigan if it was a year ago

refinegan is somewhat new

[boreal creek]

alright i made a submission, was i supposed to upload the model somewhere for someone to test or

what exactly

oh

i got a message

im not sure what embedder i used lol

is the embedder model architecture version?

[viral mason]

embedder is most likely contentvec

[boreal creek]

need to upload the model to huggingface?

[viral mason]

yea you should upload it to huggingface

make a sample as well

[boreal creek]

idk why but it says it should be a valid huggingface link

i gave it the link to my huggingface upload

https://huggingface.co/Grimden/Daniel_Of_Mayfair this is what i uploaded

[viral mason]

the model is not uploaded properly, pth and index should be in one zip file

[boreal creek]

lol

[boreal creek]

still gives me this error Model link should be a valid Weights.gg or Huggingface Model Link!

[boreal creek]

i uploaded it as a zip file...

[viral mason]

then do this

[boreal creek]

aight

it worked

thank you for helping

[viral mason]

you're welcome!

[viral mason]

Like I said in chat, if ur gpu is Nvidia download Vonovox and if it's AMD download tg fork

Butt

Why don't u wanna say

You have no need to be embarrassed of why u want to use the voice changer to be w girl

No judgement here, unless it's to troll

Very judgemental there

So you're not gonna troll right?

Good

https://huggingface.co/dr87/vonovox/resolve/c8034f5f6d50648a8109bb4f847182362e2b779b/Vonovox_beta_17_11.zip

https://software.muzychenko.net/freeware/vac470lite.zip

Here ya go

Was looking for the links

Virtual audio cable, connects the voice changer to games or discord

Nobody can hear it without that

You're welcome

[viral mason]

Nah, for the virtual audio cable just run setup64 then install driver

then Vonovox just run start

[hallow thistle]

start.bat

[plucky crown]

Helppp

So i got the 3090 pc

Its in windows 11

Which is the voice softwares in the server more comfortable with? Windows 11 or 10??

[hallow thistle]

I use Windows 11.

[low shard]

This is a General AI Discord Server, please elaborate:

• what are you trying to do: TTS, AI Covers, E Girl Trolling / Catfishing or Roleplay
• the tutorial link used

[plucky crown]

Real time voice change and voice training

[verbal fractal]

hey is the git safe?
https://github.com/w-okada/voice-changer
https://huggingface.co/wok000/vcclient000/tree/main

[low shard]

it's outdated, did you come from some old youtube tutorials?
This is a General AI Discord Server, please elaborate:

• your pc gpu
• your pc os
• what are you trying to do: TTS, AI Covers, E Girl Trolling / Catfishing or Roleplay
• the tutorial link used

[verbal fractal]

ah ok because i grad it then followed YouTube tutorial telling me to go to a different place and then I saw the warning at the top

[low shard]

the link is for the original wokada, it's legit but it's really outdated and not suggested, the warning on https://docs.aihub.gg is related to people trying to impersonate us and not about the tools, or are you talking about another warning?

it would be best you reply to the questions so i can help you get better tools

[low shard]

could you send the tutorial link you're using?

[verbal fractal]

https://www.youtube.com/watch?v=qiRpxHEoZZ4 was this one

That I'm pretty sure has the legit way of doing it the other one I just Google searched to find

[low shard]

it's kind of old, and not even related to the link you sent

the link you sent was about the original wokada, not wokada deiteris fork which the video talks about, i think you just clicked the first link

[low shard]

please answer those questions so I can help you better

and wokada deiteris fork isn't suggested either, it's more "archived"

[verbal fractal]

Yeah like I followed a Google search to get to this one https://github.com/w-okada/voice-changer
Then the video sent me to this one

https://docs.aihub.gg/realtime-voice-...

[low shard]

please elaborate:

• your pc gpu
• your pc os
• what are you trying to do: TTS, AI Covers, E Girl Trolling / Catfishing or Roleplay

those are needed, there isn't a one singular best app that works for everyone

[verbal fractal]

It's ok now that I know it's safe and I didn't screw up somehow

4060it
win10
Roleplay/dnd stuff

[low shard]

alr i replied here https://discord.com/channels/1159260121998827560/1500451925810872330

[low shard]

are you like looking to do roleplay or trolling as an e girl or eboy?

[plucky crown]

Doing all the voices for music production/music cover

[verbal fractal]

Roleplay for dnd stuff

[low shard]

check https://docs.aihub.gg/realtime-voice-changer/local/vonovox/ and https://docs.aihub.gg/rvc/local/applio/

[plucky crown]

thats two , do i pick one or both is needed

[final bough]

Hi need some help how do i stop hearing my self from the voice changer

[low shard]

This is a General AI Discord Server, please elaborate:

• your pc gpu
• your pc os
• what are you trying to do: TTS, AI Covers, E Girl Trolling / Catfishing or Roleplay
• the tutorial link used

[low shard]

Applio is an all in one but specialized in RVC Inference and Training, you could use it for realtime too but you could also experiment Vonovox which is specialized for Realtime

[proven hill]

fr

[dire sluice]

its not working for me

[low shard]

What's not working? This is a General AI Discord Server, please elaborate:

• your pc gpu
• your pc os
• what are you trying to do: TTS, AI Covers, E Girl Trolling / Catfishing or Roleplay
• the tutorial link used

[boreal creek]

isnt applio garbage for training?

i consistently heard people say it produces terrible models

[low shard]

Applio is an up to date Fork of RVC, the original mainline RVC is abandoned since years

Perhaps you confused with an issue that there was on the descriminator of an older version of Applio, or the Overtraining Detector which didn't work well

[boreal creek]

just relating what i heard from people

never tried with applio just because what i heard

[hardy yew]

what did you train with instead then? mainline RVC?

I wonder where you heard about applio being terrible because that seems just wrong to me

[boreal creek]

i used something else

[boreal creek]

might have heard it from here but im unsure

[low shard]

Applio is more suggested, did you use something off youtube tuts?

[boreal creek]

no

[boreal creek]

if you say so

[low shard]

what then?

[whole mason]

FR

[minor bramble]

why is the voice changer so laggy

[edgy minnow]

What the audio cable im supposed to use? I just reinstalled windows and I do not remember.

[low shard]

This is a General AI Discord Server, please elaborate:

• your pc gpu
• your pc os
• what are you trying to do: TTS, AI Covers, E Girl Trolling / Catfishing or Roleplay
• the tutorial link used

[low shard]

This is a General AI Discord Server, please elaborate:

• your pc gpu
• your pc os like Windows 11 or 10
• what are you trying to do: TTS, AI Covers, E Girl Trolling / Catfishing or Roleplay
• the tutorial link used

[minor bramble]

i found out the issue and fixed it thanks though.

[low shard]

be sure you're not using one of those outdated yt tus suggesting vb cable for windows

[bronze socket]

Good morning. In rvc as anyone had the issue of noise suppression 2 stopping all audio output? My second PC is offline now and for whatever reason i can't get it to output voice

I had it working a few days ago but idk what its doing now

[bronze socket]

It def works normally when I have suppression 2 disabled

[proven hill]

whats your voice changer and what gpu do you have?

[bronze socket]

Mmvcserversio. Realtime voice changer client for cuda. Nvidia 3070 GPU

[proven hill]

you did not tell me the name tho, i suppose this is the original w-okada right?

[bronze socket]

I'm sorry I'm really not sure

Its the browser based one

[proven hill]

you can either download vonovox (better for nvidia gpus) or tg developed fork (more like okada)

-rt

[patent trellis]

🔊 Realtime RVC

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

• Vonovox

A Realtime Voice Changer with similar performance to Wokada Tg-Develop Fork, with extra features, but it supports only Nvidia GPUs on Windows 10/11 unlike other options that have wider support. and without cloud options

• Local Guide

• Wokada Tg-Develop Fork

A personal fork (modified version) of Wokada Deiteris Fork, it just adds some Quality of Life improvements to it like supporting Spin Embedder and Audio Effects. Don't expect too much about it since the creator made it originally as a personal project.

• Local Guide
• Cloud Guide(s)

• Applio Realtime

A Realtime Voice Changer with similar performance to Vonovox & Wokada Tg-Develop Fork, with extra features.

• Local Guide

• Wokada Deiteris Fork

Deiteris' fork (modified version) of wokada that doesn't get updates anymore.

• Local Guide
• Cloud Guide(s)

⛔ Outdated/Discouraged

These options are not recommended for use.

• Original Wokada

Not suggested, older versions in youtube tuts are even way worse. GUIDE

• RVC GUI Mainline Realtime

The program is worse compared to the ones above, and much less updated. GUIDE

[bronze socket]

OK Ill try them both! Thank you so so much for the help!

[proven hill]

no problem!

[lilac peak]

guys what does "epochs" mean

[viral mason]

it's just how many times the ai went over the audio used to make a model before it sounded good to whoever was making it :D

[lilac peak]

ooooh thanks a lot!

[viral mason]

you're welcome!

[proven hill]

like in real life?

(what are you using? whats the gpu you got?)

please respond to the question i sent and provide me a screenshot of the program soni can guide you

yea dm me the screenshot

[viral mason]

that is extremely outdated

[fringe heron]

This isn’t a basic usage question, i would like to share my idea of why converting a speaker e.g. (A) most of the time it deosnt sound really that good no matter the dataset, ive read some docs and it doesnt seem they cover this that well, if i missed something i am sorry, please dont hate me 🙏

I’ve been digging into the RVC training code and experimenting with different pretrained models (like Titan), and I keep hitting the same wall

B to B conversion (same speaker) sounds fantastic (where b is target)
A to B conversion (my voice to target) sounds glitchy, unnatural, and full of artifacts in many cases, even when my voice is similar to the target.

I think I’ve figured out why (might be very wrong)

if i understood right, the rvc's generator consists of four main parts, Textencoder takes phonemes + pitch and outputs a content latent, posteriorencoder + flow encodes the real mel spectrogram into a posterior latent then the flow transforms it to match the prior from the text encoder, decoder takes the latent and the speaker embedding to produce audio.
During training, all these blocks are updated together using the same optimizer

when you fine‑tune a generic pretrained model on a single target speaker B, every part of the generator adapts to that speaker s data including the textencoder.
Even though the textencoder only sees phonemes, it learns to produce a latent representation that’s subtly tailored to B’s vocal tract. the flow and the decoder expect that latent to decode into B’s voice. The encoder is never exposed to other speakers during fine‑tuning, so it becomes speaker‑biased.

The result
B to B works great because the encoder s B bias
A to B breaks your voice A goes through an encoder that outputs a B biased latent, creating a mismatch with the content, and the decoder struggles to produce clean audio

The solution I’m planning
Pretrain the generator on a large multi‑speaker dataset roughly 60 hours of my own voice (speaker a) and 50 hours from 15–16 speakers forcing the encoder to extract only content

then freeze the textencoder and finetune
i still havent tried this and i am curious of thoughts of people that are probably more experienced than me and maybe already tried this.

And why isnt freezing the text encoder a native option?

[viral mason]

the only pretrains I'd use currently are Legacy core 1.5 and 1.6, and the new PABP one although it is a bit new and could be noisy

[fringe heron]

i mean, the point is not the pretrained itself

[viral mason]

titan, ov2, rin3 are very old and cause harmonic distortions

[fringe heron]

its the way you fine tune it

[fringe heron]

iv tried almost all of them at this point

[viral mason]

freezing the text encoder isn't native as it could cause issues to casual users who don't know much on rvc training

how do you clean your datasets btw and what are the usual lengths?

[analog obsidian]

freezing te is not a bad idea, finetuning is a bit agressive in rvc and you want to preserve a lot of things from the pretrain

i have tried freezing it before but i didn't notice improvements tho

[fringe heron]

usually already almost studio quality, from 5 to 30 hours, ive tried many sizes

[fringe heron]

oh

[viral mason]

I mean do you remove room reverb, echo if there is any, ect

[fringe heron]

yea if needed but thats rare

[viral mason]

ah kk

[fringe heron]

as almost always they need no preprocessing

[analog obsidian]

multispeaker models have natural timbre leak sadly

og pretrain and every other leaks

[fringe heron]

tried with a generic pretrained?

[analog obsidian]

yes, with the original pretrain

[fringe heron]

my idea was to from scratch train a multi speaker pretrained that has 60% of its dataset my voice

to see if on my would sound less ai fart due to latent mismatch

(my first big explenation msg might be wrong so correct me if i say somethign wrong )

[analog obsidian]

it wont work as you expect, the other speakers are going to affect the timbre of your voice in the model

even if your voice is dominant

it still sees the other speakers

[fringe heron]

but if you divide speakers by speaker id?

like is that bad

the leakege

[analog obsidian]

its quite bad sometimes yeah

when you finetune the pretrain that gets fixed since every speaker in the pretrain gets override by that singular speaker

but the actual pretrain itself leaks

[fringe heron]

and using only my voice wouldnt be such a better idea either i think

from scratch

[analog obsidian]

thats what im actually trying, training a singular speaker (ljspeech) from scratch, then finetune other speakers

but i havent finished the training bc i dont have money lol

[fringe heron]

but see what i think is that then if you do A to A alone the latent leakes like crazy no?, to make the dec a easier job

[analog obsidian]

with that experiment i confirmed that yeah without any other sids, ljspeech doesnt leaks

[odd flower]

Hi, so elevenlabs using the same RVC for speech to speech? as its S2S is amazing with the same dataset.

[analog obsidian]

hmm i haven't tried this yet

like, finetuning the same speaker found in the pretrain?

[fringe heron]

no

like

if i train from scratch a model with lets say, 70 hours of my speech then the latent is mostlikely leaking my timbre so that the dec has a easier time, so when i finetune it on a speaker b to get a better A to B then the dec has to fight the enc

[viral mason]

I don't think so no, if the models cannot be used in applio or in realtime it's most likely something different

[fringe heron]

never tried this so i am not sure

but its a possibility in my head

[analog obsidian]

well i tried finetuning my undertrained ljspeech pretrain with a voice that sounds very different to her and that model had no leakage

but i haven't tried it with a fully trained model from 0

[fringe heron]

okay so no leakege

but i assume lj is different from your voice right

[analog obsidian]

im not using nsf-hifigan as the decoder/vocoder tho, and we found that vocoders do have an impact in leakage

[fringe heron]

always used hifigan

[analog obsidian]

yea i finetuned some random anime character that doesnt even speak english, no leak

[fringe heron]

becouse my goal is to get the model actually reliable on my voice specificly, even if you train with 0 leakege a model then if the vocie of the target isnt similar to yours you wil lalways get problems

sure feading you voice can may not leake but actually being preceptually good and realiable

[analog obsidian]

true hmmm

[fringe heron]

thats why thought about my plan of training a single speaker or multi speaker from sctratch so the text enc is good on my voice and also is hopefully invariant

so then i freeze it and fine tune the rest

the flow and dec are forced to map my voice to target

[analog obsidian]

you should try your first plan, even if the multispeaker have some leaks, it still knows your timbre

[fringe heron]

yea

also you said that vocoders have their impacts, what do you suggest?

i tried refinegan but never really chould say if its really better than the others

[analog obsidian]

keep using nsf hifigan, i tried training refinegan pretrains and they all suck, doesn't sound natural

[fringe heron]

okay

thanks

[analog obsidian]

no problem and good luck, i really never managed to ever train a functional nsf hifigan from scratch, the model always got stuck at some point where it refused to do any changes to the quality

so 1m steps and 1.5m basically sound the same

[fringe heron]

how much time it took to get there?

[analog obsidian]

4 days

[fringe heron]

i feel your pain

[analog obsidian]

the original pretrain was trained for about 2.5m steps

[fringe heron]

on a single gpu i imagine

[analog obsidian]

yes iirc with a 4080

and batch 16

[fringe heron]

okay

[fringe heron]

solid

[analog obsidian]

maybe you should try finetuning the original pretrain first and see if your new "finetuned" pretrain + freeze TextEncoder helps

[fringe heron]

Sure thats an experiment that i can do, even thought i doubt it will be reliable on my specific voice

[proven hill]

italian

[fringe heron]

yes

[proven hill]

thats it youre banned

[fringe heron]

🙏

[analog obsidian]

aah but train your pretrain dataset using the og pretrain as base, thats what i tried to say (my english is terrible lol)

https://discord.com/channels/1159260121998827560/1235952130855010365 nothing here is trained from 0, everything is a finetune lol

[fringe heron]

so you are saying like, take my 60 hours dataset of me and finetune a pretrained?

[analog obsidian]

exactly

[fringe heron]

i dont think that would help

maybe with a low lr

but that would make the enc biased to my voice wich is good but then the enc might start leaking my timbre in the latent as that is the easiest path to reconstruct my voice

so the dec fights enc anyway

[analog obsidian]

[fringe heron]

it would need to undo my timbre then

maybe you can do single speaker pretrianed if you augment data?

[fringe heron]

like heavly augments that puts pressure on the enc to keep only content

[analog obsidian]

never tried that, idk

[fringe heron]

i mean it makes sense

i dont know add like formant, eq and pitch shift (must pay attention on that tho)

and also change sids if you augment

and with those experiments bye bye 50kW

[proven hill]

@analog obsidian have you ever tried my pretrain?

[fringe heron]

you did the italian 32khz no?

[proven hill]

yes

[fringe heron]

i used it

[proven hill]

how was it

be honest

i know it wasnt good

[fringe heron]

its alright

depends thought

i had some voices that where kinda trash

but never had leakege

honestly the main issue is still always the same and its that on my voice is not reliable

[proven hill]

why?

[fringe heron]

not becouse it makes a wrogn timbre but becouse of latent mismatch

[fringe heron]

if you havent read here i wrote my theory of why when you finetune a pretrained to have A to B it almost always deosnt sound reliable or glitches out with some words/ sounds unnatural most of the time

holy i cant write right

sorry for eventual mispelled words

[mild ridge]

Heya. im new to this hub . i tried researching first before asking here. using gpu 4070ti/windows 11. my question : i have decent experience using rvc, trained my own models and used the old wokada for the longest. and after a decent amount of tries i hit a plateau and im looking to improve my model. right now im trying to use the applio to train my model instead of rvc v2. which are the best options to pick from sample rate/ vocoder/embedder. custom pretrain etc.

[cloud abyss]

[gpu_processing] cv2.cuda.GpuMat exists but missing: createGaussianFilter, resize, cvtColor – falling back to CPU.

[proven hill]

mhh couldnt that be a model problem?

well yes me of the past, hence why the pretrain theory

[fringe heron]

in what way?

like

[proven hill]

maybe it wasnt trained on clean data(?)

[fringe heron]

oh

no

[proven hill]

?

[fringe heron]

i use studio quality

[proven hill]

oh well

[fringe heron]

tried even datasets also similar to my voice, similar acent and all

never fixed this issue

[cloud abyss]

deeplivecam error, its using my cpu instead of rtx 4070, i used to use this and get 30 fps now i get 3

[fringe heron]

no idea

[viral mason]

what is deeplivecam?

[fringe heron]

indeed

[proven hill]

hes a catfisher

[cloud abyss]

wtf

[viral mason]

pretty obvious

[cloud abyss]

my gf is in my pfp

calm down

i just like being elon musk

[fringe heron]

even if he is what even is deeplivecam 😭

[cloud abyss]

face swap live

[proven hill]

a thing to live render people faces, deepfake category

[fringe heron]

i see

cant help thought, i know nothing

[analog obsidian]

no

[proven hill]

do it

[fringe heron]

@analog obsidian for the pretrained multispeaker i might have an idea to no leak and still train to be very good at a specific voice, lets say i have 60 hours of voice A and idk 50 of other 15 speaker, since my goal is to have a pretrained that is very good at voice A but also invariat, maybe if we modify the batches to load idk, 1000 clips of A 1000 of B and so on for all speakers even if speaker F has like 3 hours compared to A since the batch has the same amount of each if the model tries to learn specific speaker it would hurt the other speakers and since gradients have same weight might work better.

[analog obsidian]

hmm sounds like a good idea to me, but never tried this so im not sure if it's going to reduce the leakage, but in theory, it should like you say

[fringe heron]

Yea same, would need to be tested

[bronze socket]

Oh my goodness your my hero, this new vonovox is so much better quality. Thank you again

[proven hill]

no problem!

[stone jolt]

[VCClient] wait web server...340 http host wouldnt load

[viral mason]

that's extremely old what gpu do u have? (Nvidia or AMD) and what do u plan on using it for?

[gray dagger]

im trying to train a model using replay (i have a 3060 12gb and im using windows 11) and i get this error

Error creating model: CUDA error: unspecified launch failure Search for cudaErrorLaunchFailure' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information. CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile with TORCH_USE_CUDA_DSA` to enable device-side assertions.

how would i fix this

i have auto epochs turned on

replay still gives the same error with auto epochs turned off

im on version 8.7.2