✨│ai-help

[low shard]

Is RefineGAN not stable enough yet?

[simple ore]

well, it was great in small scale tests, but large tests not so much

[low shard]

will applio keep it in the future or possibly remove it forever
https://cdn.discordapp.com/emojis/1225582606062456893.webp?size=48

[simple ore]

i'm testing some changes

but that means yet another retrain

at this point it is no longer "refine" gan

[jaunty iris]

so thats why i couldnt get it to work 😭

i was going insane earlier i reinstalled like 7 times

[analog obsidian]

sad to see refinegan go
but will you continue working on it?

[old sigil]

hey all

[simple ore]

you can re-enable it by editing tabs/train/train.py

[jaunty iris]

Oh ok

[simple ore]

change this back to True

[old sigil]

i am new in rvc and I want help to understand how to setup it on either local or collab?
can somebody guide me please

[simple ore]

I would not say it is gone gone

[jaunty iris]

Thank u

[simple ore]

but it is no longer does what the original paper did

my current version is using interpolation and a parallel resblock, so it is more like hifigan that has solved the problem with horizontal lines at 4, 8 and 12KHz

[old sigil]

hey could you help me a little bit in giving a rough overview

?

[simple ore]

???

[old sigil]

I want to setup rvc and train a model with my voice

can you give me a rough idea on how to do this ?

[simple ore]

why dont you look at https://docs.aihub.gg/essentials/how-to-make-voice-models/

[old sigil]

thanks!

[kindred eagle]

The orignal one made by rvc boss

[low shard]

tha'ts commonly called mainline/original

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

You can:

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

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

I would suggest you applio for more updates

also, check the docs https://docs.aihub.gg for more info on vocal separation

[fluid topaz]

I know rx but which one is vx? Kinda lost

[low shard]

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

[glacial pollen]

one sec

waves' Clarity vx - dereverb pro mono

( Reminder; it is for mono audio. So you take one channel and work on it (( as you should anyway )) )

[fluid topaz]

Oh cool thanks

[glacial pollen]

✨

[old sigil]

I have a mac m2 air
and I want to try voice changer real time

[low shard]

wrong channel then, rvc isn't for that

RVC = Retrieval-based-Voice-Conversion, the best Speech To Speech AI Models (on v2), Inferences (use models) pre-recorded audio (ai covers) and train (make) models

Wokada = uses RVC for realtime inference

always elaborate your requestes when asking, people can't know your pc nor what u want to do

@old sigil go to #🔍│help-w-okada

[old sigil]

How about this speech to speech conversions?
this has some latency?

[low shard]

rvc is just on pre-recorded audios, it's not meant for realtime, you need to use only wokada if you want realtime with rvc models

those are 2 different program names

rvc does not equal to "realtime voice changer"

[old sigil]

so what I am getting is, for rvc
If i have pre-recorded audios and i could train a model out of it
can I use it for speech to speech conversion?

[low shard]

there is a go realtime for the original/mainline rvc, but it's way worse optimized than original wokada, and way worse than wokada deiteris fork

[low shard]

RVC is a STS AI, you just said you wanted realtime voice changer for calls, not on pre-recorded audios

[old sigil]

ya but I was curious in knowing how rvc works

but Yeah!
thanks for your help

I'll have a look on wokada for my usecase

[low shard]

yeah it's speech to speech

[brittle wing]

Yo guys,
I’m tweaking the Batch Size setting and not sure what to pick. 4 = better accuracy but slower, 8 = faster but "standard".
Does 4 actually improve sound quality, or is 8 just as good?
Would appreciate a simple explanation!

[oak sandal]

i used to do up to 20 batch size

i think theres barely any difference between 4 and 8 so if you have the power to spare just do 8

[oak sandal]

if we are still on RVC v2 training on mangio-crepe that is

[low shard]

check https://docs.aihub.gg/rvc/resources/training/#batch-size

it depends on your gpu and dataset lenght

[low shard]

till 20 seems too much

[low shard]

you don't use the mangio fork still, right?

because that's old since 2023

[oak sandal]

do you think i would care with an SLI of gpu's that are worth like an used BMW?

once you get to cloud training you never go back

[oak sandal]

Applio

also yea i'd rather use the old mangio fork even in 2025

Still crepe and still rmvpe, i see we dont have a new extraction method yet after all this time

the "embedder" is new to me tho

iirc it was by default on chinese hubert

[analog obsidian]

if your dataset is below 30 minutes, use batch size 4
if your dataset is above 30 minutes use batch size 8

[oak sandal]

now we got.. contentvec?

[low shard]

you don't need multiple gpus to train RVC models

[oak sandal]

ofc you don't, you can manage aswell with a tiny lil 3060 12gb

[analog obsidian]

no

[oak sandal]

but when your dataset is 2 hours long

[analog obsidian]

rvc always used contentvec

hubert = contentvec

same thing

[low shard]

no, local training is better but it depends on your setup, cloud isn't as much stable and you can see that in #📰│dev-updates , especially google colab for training kinda sucks

[oak sandal]

not google colab

actual cloud servers with RTX server gpu's

lmao

[low shard]

mangio rvc fork is pretty old and not maintained, I don't get why you would say that when Applio got more improvements

[oak sandal]

i used colab only like 3 or 4 times when RVC wasnt a thing and SVC is all we had

[analog obsidian]

mangio fork uses extremely outdated code, everything its wrong there

[low shard]

^^^

[analog obsidian]

even the logging is bugged

[oak sandal]

i made great models even with the outdated code and everything being wrong + bugged logging in 2023

i seriously doubt there was much improvement since then

[analog obsidian]

because it wasnt outdated in 2023

mangio stopped receiving updates around that time

[low shard]

there's other enhancements in the code, it's not just that

[analog obsidian]

it was on pair with mainline back then

now its even behind mainline

and mainline is also extremely outdated

[oak sandal]

i would love hearing audio difference between 2023 mangio-crepe and nowadays RMVPE on applio or whatever yall prefer now 👀

[analog obsidian]

rmvpe its just a f0 estimation

is not a quality thing

[oak sandal]

it had pitch issues

[low shard]

you can even train on worse, it's just not suggested for not wasting time

[analog obsidian]

rmvpe has been always the same

🥹

[oak sandal]

not RMVPE itself

the implementation

it was a mess

[oak sandal]

yea and thats why i always used Runpod or other paid services

[unkempt sapphire]

does sup1 even work

only sup2 is good

and it sounds so ass with sup2 on

[low shard]

wrong channel, use #🔍│help-w-okada

[unkempt sapphire]

oops

[oak sandal]

bro i legit checked on Weights.gg, my old model trained on mangio's fork is still magnitudes better than the newer models

[low shard]

RVC = Retrieval-based-Voice-Conversion, the best Speech To Speech AI Models (on v2), Inferences (use models) pre-recorded audio (ai covers) and train (make) models

Wokada = uses RVC for realtime inference

[oak sandal]

😭

most of the work is separating the vocals, not using the right version of RVC

im p sure i wouldn't be able to replicate that level of clean vocals again, kim vocal 1 was goated back then

[low shard]

kim vocal is pretty old, there's better models now

[low shard]

you can't know how good the person trained the model nor how

[oak sandal]

read what i wrote above

[oak sandal]

.

i used multiple models on FLACs files, then manually fixed issues in FL Studio + izotope rx

most people just download random mp3 128kbps acapellas from youtube and be done with it

every other people that tried to make an "updated" version of my model failed miserably, the audio that gets generated is a mess, both in sound quality and voice similiarity

[crude flame]

what is your model?

[oak sandal]

Sfera Ebbasta

this one

[crude flame]

oh a singer

no wonder people mess it up

[oak sandal]

then how come i didn't mess it up?

🤔

[crude flame]

you prob used studio sessions

[oak sandal]

no

[crude flame]

you tried cleaning the dataset

[oak sandal]

no studio session, all manual labour and AI

from the song themselves

i just downloaded the whole album in FLAC quality

then one by one, minute by minute, cleaned the dataset

[analog obsidian]

its not

[crude flame]

oh

wow

[analog obsidian]

[low shard]

also using an updated version helps, applio got more enhancements like the benchmark flag #🔊│ai-development message which is turned on the main branch

but yeah ofcourse it depends much on how people clean the dataset

[oak sandal]

ok so, after 2 years the advancements were... better GUI and slight code optimization?

what happened to that one dude who was developing RVC "v3"

[low shard]

rvc boss left rvc to rot, he's working on gpt so vits and recently released the v3

[low shard]

there are also newer things, like the refinegan vocoder on the main branch being experimental rn

idk why ur so attached to mangio fork

[oak sandal]

i think last time they were trying vocos

iirc that was the name

[oak sandal]

i'm not, we just didn't get any decent improvement since 2023

atleast for RVC

[low shard]

we did get some improvements with applio, it's just the original dev now works on other projects than RVC, so it's not an RVC v3 but more experimenting

rvc boss works basically mostly on https://github.com/RVC-Boss/GPT-SoVITS

[oak sandal]

which is probably best for speaking only

[low shard]

ofcourse it's not such an entire new type of architecture, but i would rather use code that gets updated and improvements rather one that doesn't get any at all

[oak sandal]

i dont think we are gonna see any decent or sizable improvements in the next year

maybe 2027

[glacial pollen]

It always was contentvec, not hubert

As for f0 extractors.. yeah, there's nothing really better as of now, afaik

As for vocos, it is not worth it.
Quite tricky to get properly working and tbf, potential phase reconstruction issues aren't worth it, similarly stft and I believe istft vocoders

[simple ore]

most generators are only useful for mel to wav reconstruction, not encoded latents to wav

hifigan uses NN filters to predict the waveform from the encoded latents

vocos does not have enough capacity to predict

[tough fiber]

-kaggle

[karmic olive]

Suggestions for @distant turtle

📘 Kaggle Notebooks

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

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

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

[low shard]

I hope we get them asap

[unique rock]

Can someone tell me why I get this in Applio Voice Blend?

model_blender(model_name, pth_path_1, pth_path_2, ratio)
ValueError: too many values to unpack (expected 2)

I tried to blend a 200 epoch and 48k sample rate model with a 270 epoch model the same sample rate

[simple ore]

if both models were made in Applio and both at the same sampling rate, then there should be no issues with merging

there may be an issue if the models came from different sources

[icy vessel]

what pretrain does the applio colab use by default ?

[low shard]

The original/OG pretrain

[icy vessel]

ok thanks

[deft flare]

Hi everyone I hope I'm texting the right chat.. I have had to cancel kits.ai due to being ridiculously expensive.. anyone knows good complete walkthrough tutorial how to make a model from scratch? kits would sort everything for me so I feel kinda lost

[low shard]

Kits.ai just uses RVC but in a simplified User Interface

First of all what's your PC GPU

[deft flare]

macbook pro m1

[low shard]

Oh, macs ain't really good for AI

[deft flare]

oh?

i'll bite the bullet i guess

[low shard]

Train (make) RVC Models on cloud:

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

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

3. Be sure to know about the tensorboard

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

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

You can use either:

• Weights.com: Easiest Possible Ever Automatic
• Ilaria RVC Zero: Fastest free on cloud
• Applio (ui)

[deft flare]

so I want to train after this musician.. perhaps the whole discography not faring more than 30-40 minutes tota

[low shard]

If you want, you could try locally (runs on your Mac), https://docs.applio.org/applio/getting-started/installation , but no one has ever reported training a model successfully on Mac since it's kinda slow

[deft flare]

I actually downloaded this earlier this evening but it's all so overwhelming

i have uvr5 as well just don't know whether there's an ubiquitous setting to extract clearest lead vocals

if you could guide me step by step mate, I'd paypal you or something if that suits you

[low shard]

I'm not even sure if that supports Mac at all

[deft flare]

uvr5 seems to work when I open it haha

[low shard]

Hopefully it runs on MPS using the integrated m1 pro chip rather than the CPU

Else it's gonna be even slower

[deft flare]

right

(pretends to understand)

[low shard]

@viscid moss hey sorry to disturb, do you remember if UVR supported MPS for Macs, or does your version of it support it?

[deft flare]

ur so kind thank u ❤️

[low shard]

I was asking the other staffer if he remembers if that program supports it, or if his own version supports it since he made a separate own version of UVR by the way

[deft flare]

yeah I gotchu, I was just moved how there's actually someone being understanding towards complete n00b

lost kind of kindness on the internet for most part

[low shard]

Unfortunately unless you spend like 4k dollars, macs are pretty shit for AI

And even if you spend that much and get the most powerful Mac made for user consumption, the issue is not many AI programs support Mac at all

[deft flare]

when I was experimenting with google collab stuff it seemed to do decent like?

[low shard]

Yeah that's because it doesn't run on your Mac, it runs on a cloud server (remote good PC) that use Nvidia GPUs

Nvidia GPUs are the best in terms of performance and support in AI field

[deft flare]

Ahh

[low shard]

I also sent you cloud links before

Did you install applio locally, or did you just use a Google colab? And this for UVR too?

[deft flare]

both locally

[low shard]

Applio should have Mac support, hopefully uvr does too

Tbh, I don't know how much I could suggest you to use them locally, At this point cloud would be faster than your Mac, but the only issue is limited GPU time

[deft flare]

I can leave it overnight or w/e not to worry about that really ❤️ guidance is pivotal for me

[low shard]

Be sure to watch out if it overheats or anything though

I would personally suggest you to use cloud for faster processing, but your choice

I don't know how much time it could take to be honest, not sure if it's going to be overnight or more or less, so I can't guarantee you much on that

[viscid moss]

I think it has MPS support, cause they made releases for Mac. UVR5 UI doesn't have Mac Support (There's no installation files + idk if works correctly cause I don't have a Mac to test) but audio-separator (UVR5 UI core) have MPS support.

UVR5 UI probably works with MPS but I haven't tested it because I don't have a way to test it.

So I recommend just use UVR5 for mac

[low shard]

https://github.com/Nick088Official/Minecraft_Skin_Generator/blob/main/Scripts%2Fminecraft-skins-sdxl.py#L137 maybe this can help

[low shard]

Alright thx for letting me know

[viscid moss]

that's already built in, just need to test xd. I'll try in a while using a VM maybe, to make the installation works at least atm I'm busy with irl work

[low shard]

Alright goodluck

[rapid spade]

is there a guide on how to download the software needed for ai stuff

[dire juniper]

anyone know why i cant upload a rvc model to my voice changer ?

[karmic flax]

i get alot of errors when installing RCV with the TroubleChute one line command. and im to stupid to install it manually. i used RVC on my old pc (win10) before and now on win11 everything just wont work. it openes the website but it wont ever finish converting

[hallow thistle]

Which RVC program are you using?

[karmic flax]

Retrieval-based-Voice-Conversion-WebUI

if that wasnt the correct answer im sorry lol im not rlly deep into software stuff..

[hallow thistle]

https://cdn.discordapp.com/emojis/872008154257432636.webp?size=48

It would be better to use Applio the RVC instead.

What is your PC GPU?

[karmic flax]

RTX 5080

[hallow thistle]

Damn. Unfortunately, there's no known stable version of Applio for this specific GPU. But I think you can use Applio with CPU instead.

[brittle wing]

i want a model that sounds realistic please, i dont mind the size of the file

[karmic flax]

hm i have the 9800x3D im sure it will be enough right?

[brittle wing]

smh bros flexing

[karmic flax]

no im wondering about stuff cuz i dont have a clue lol

[brittle wing]

that's like one of the most powerful if not the most poweful cpu rn

ik cuz i have one myself

[karmic flax]

yeah for gaming but for AI stuff idk?

[brittle wing]

youre thinking its like nvidia vs amd but its different from gpus

[karmic flax]

i just heard for productivity stuff core count is more important but ig not

[hallow thistle]

I'm not sure about this one, but yeah just like RTX 50 series, there doesn't seem to be any version of Applio made for this specific AMD GPU.

[brittle wing]

what do you want to do with it

[hallow thistle]

Are you saying you hate people flexing their stuffs or something?

[brittle wing]

no?

i flex my 7900xtx sometimes

flexing is ok as long it dose not include lying imo

[hallow thistle]

Imagine thinking "my laptop GPU is Intel HD Graphics 3000" is a lie.

[brittle wing]

yea imagine

[simple ore]

9800 is cpu

[brittle wing]

imagine imagining stuff that isn't true and just assuming it over the conversation

[simple ore]

at the moment there's no proper torch build for 5000 series

but you can install something for applio

[karmic flax]

so the issue is my gpu? dang lol

[hallow thistle]

I've mistaken 9xxx for AMD RX GPU.

[simple ore]

you can download applio compiled version, then update torch to cu128

manually

[karmic flax]

would the best bet be to use Applio? i have it installed but i dont rlly get it haha

[simple ore]

actually not compiled, clone the repo

then run the installer

hopefully it does not error out

then update torch

[karmic flax]

"just do it" already confused lol

[simple ore]

https://huggingface.co/w-e-w/torch-2.6.0-cu128.nv

env\python -m pip install that_wheel_file

https://huggingface.co/w-e-w/torch-2.6.0-cu128.nv/resolve/main/torch-2.6.0+cu128.nv-cp310-cp310-win_amd64.whl

[hallow thistle]

Applio is the only way to get converted audio done very fast with proper GPU. With RTX 50 GPU, you'll have to do some code a bit.

[karmic flax]

saddies. so there is no way for me lol

[hallow thistle]

You can use any other RVC program, but all of them will only use your PC CPU because neither of them have complied for RTX 50.

[karmic flax]

i mean i wouldnt care what it uses as long if it just works

[hallow thistle]

That's all good.

[karmic flax]

i wont be training my own models and stuff. i just need stuff to be converted into voices via pretrained models

[simple ore]

no code, just replace torch version with update that supports 5000 series

[karmic flax]

but hooow

u linked me a file i cant even do anything with lol

cuz tf is a .whl file

[simple ore]

clone the repository

run installer

[karmic flax]

thats the first thing im failing on lol

[simple ore]

once that is done run the command I provided from command prompt

download zip

[karmic flax]

step one completed successfully lol

then ig run-install.bat

then running ur command in a terminal with admin perms

module "env" couldnt be loaded

[karmic flax]

[karmic flax]

it just throws me an error when trying that

[karmic flax]

but how do i even get there

[simple ore]

open command prompt

[karmic flax]

yeah

[simple ore]

you need to download the wheel into the same folder

[karmic flax]

okay did that

dang i did it. im a hacker. lets hope it works lol

greatly appreciate the help tho 🫂

[simple ore]

FCPE wont work until there's a updated torchaudio

but hopefully the rest would

[karmic flax]

hm now it doesnt start up anymore lol

[simple ore]

give it a bit

[karmic flax]

it throws an error that it doesnt find smt

[karmic flax]

ig "env/lib/site-packages/torchaudio/lib/libtorchaudio.pyd" wasnt found lol

ig i need to do the same with libtorchaudio but idk what version

[carmine shuttle]

Guys is it normal for a voice model in zip that i heavy 259MB ?

[karmic flax]

id still appreciate help with getting any voice conversation to work on my gpu.

[brittle wing]

Anyone know the python version requirement if I want to run this version of UVR5 locally?

https://huggingface.co/spaces/TheStinger/UVR5_UI/tree/main

I cloned the repo and tried to install the requirements with the lastest python 3.13.2 but failed:

ERROR: Could not find a version that satisfies the requirement torch<2.5,>=2.3 (from audio-separator) (from versions: 2.5.0, 2.5.1, 2.6.0)
ERROR: No matching distribution found for torch<2.5,>=2.3

[hallow thistle]

Do not install any Python program related using the very most recent version of Python. Use a version of Python like Python 3.10.x or Python 3.11.x.

[simple ore]

can you show the full error stack?

[simple ore]

you may need to downgrade python to 3.10

[knotty moth]

there's only 9070 & XT

[knotty moth]

the alternative is Anjok's UVR with latest patch, and ZFTurbo's MSST repository that works on python 3.10-11

[glass igloo]

Hello. I've encountered an issue while creating an index file. I have approximately 7 hours of audio for voice training, and the training process went smoothly. However, when trying to create the index file, the process stops at around 1,300 files out of 48,000, and then an index file is generated, which is only 30 megabytes in size. When using this index file, the voice often converts with artifacts. In other models I've trained on 10-20 minutes of data, the index files weigh 120+ megabytes. What should I do in this situation?

[simple ore]

with 4000+ sliced segments the index creation attempts to narrow down the dataset, but it only goes so far before it detects that it gets no improvement

you can always run inference without an index and check if that comes okay

[glass igloo]

Without the index is fine, but the voice is not as similar as with the index.

[simple ore]

the voice comes from the model

index is just an accent

7 hours of audio is too much for a finetune and not enough for training from scratch

[flint glade]

guys

can anyone help me

on the voice changer

[glass igloo]

What is the best length of the dataset to use? If I shorten the dataset, is it better to continue training the same model or start training again?

[brittle wing]

Thanks I now use 3.10.16 and it works flawlessly

[flint glade]

i downlaoded it but when i will open it is doesnt open

[brittle wing]

Thanks it works

[flint glade]

can you help me

[brittle wing]

Can you be more specific about why you cannot open it, is it because of the dependencies?

[flint glade]

i have downloaded the voice changer and the guy on the video says open it

i open it but it doesnt open

[low shard]

Don't follow YouTube tuts ever at all for RVC and Wokada,those are old

First of all, you want a realtime voice changer for calls?

[flint glade]

can you say me a good voicechanger i will make a girl voice

[flint glade]

yes and for games

[low shard]

Wrong channel then

[flint glade]

okay

for

discord

i will

[low shard]

RVC doesn't mean realtime voice changer

RVC = Retrieval-based-Voice-Conversion, the best Speech To Speech AI Models (on v2), Inferences (use models) pre-recorded audio (ai covers) and train (make) models

Wokada = uses RVC for realtime inference

Tell your PC GPU in #🔍│help-w-okada

[flint glade]

ok

[simple ore]

30-60 min, content varity > quantity, 3 hours of mumbling is worse than 15 minutes of speaking and singing

new model, obviously

[knotty moth]

that's unless you force select faiss method (which could take longer and produce an index file in around 2 gigs) instead of leaving auto in the index training

[empty parrot]

Yo, i didn’t know index is for accent. I always thought .pth always needs it to work. Is there a good documentation to learn some of this? Id like to understand general idea how it works but not the too deep explanation for like engineers.

[hallow thistle]

A pth file always required to work, of course it is.

An index file is a file that stores voice accent and used for specific voice model. It can be achived during voice training process.

[brittle wing]

Yo guys,
I still don’t get how Batch Size really works.
Does 4 actually improve sound quality, or is it just a performance thing?

[odd shale]

Yep, it will improve it depending on your dataset.

(I'm talking about quality)

It will mostly be useful if you have a short dataset below 5-4 mins.

If your dataset is around 10+ mins, go for 8 on batch

[analog obsidian]

batch size = number of files trained at once
bs 4 trains 4 files at once
bs 8 trains 8 files at once
for simplicity:
less than 30 mins of audio = use batch 4
more than 30 mins of audio = use batch 8

too high in a small dataset might hurt the model ability to generate new audio

[odd shale]

Espera Lyery

Si es buena idea usar batch 4 con cualquier dataset que vaya debajo de 30 minutos?

Esperate leí mal XD

[analog obsidian]

si el dataset es menor a 30 mins usa batch size 4
si el dataset es mayor a 30 mins usa batch size 8

[brittle wing]

Got it! So for less than 30 mins, Batch 4 is best, and for more than 30 mins, Batch 8 is better. Just to be sure, using Batch 4 on a longer dataset wouldn’t really improve quality, right?

[analog obsidian]

nope
it could actually make it worse, the model would be stuck in a bad place and never leaving it

so the model would sound weird

[brittle wing]

Ohh I see! So using Batch 4 on a long dataset could actually trap the model in a bad state? Thanks for the clarification!

[odd shale]

Yep, for that reason i mostly recommend using 8 on batch for any dataset beyond 10-20 mins.

[analog obsidian]

yesss
technically any batch size works but the results depends heavily on the dataset
for making things simple just stick to what i said before

there are times where bs 4 gives better results than 8 and viceversa

[odd shale]

Yup, it's matter of testing

[brittle wing]

Thanks a lot!

[earnest stone]

how do you make rvc?\

[low shard]

elaborate:

• ur pc gpu
• what do you mean? did you mean make an rvc model?

[karmic flax]

im sure he gone already.

[low shard]

welp, youre right 😭

People need to understand that helpers might be busy sometimes and they can't reply in 2 mins

Do you need any help?

[karmic flax]

yeah..

[karmic flax]

well. i got help before but im sure he was also busy. i couldnt solve my issue but im also veeery nooby when it comes to software so yeah

from what i understood is that torch/torchlibaudio doesnt have working versions for the 50series nvidia cards? and id need to update manually to a nightly version of both

buutt yeah

[low shard]

elaborate:

• your pc gpu
• what guide/download link are you using
• the issue specifically
• what do you want to do

[karmic flax]

RTX5080
Got told Applio is best.
Not getting it to work in 50series cards
Just convert existing audio using models into other voices

[low shard]

Thank you for replying

I checked that you meant this chat #✨│ai-help message , unfortunately this is related to the rtx 50 serie needing a new pytorch version, meaning there's no precompiled version, and you need to do it via source
I don't have a 50 serie, but I can try helping you out

[karmic flax]

Would greatly appreciate that

if u have the patience to handle me haha

[low shard]

you're running on windows 11, right?

[karmic flax]

Yes

[low shard]

try going back to where you had that libtorchaudio missing issue, open CMD, run env\python -m pip install https://download.pytorch.org/whl/nightly/cu128/torchaudio-2.6.0.dev20250308%2Bcu128-cp310-cp310-win_amd64.whl

then, try running applio

@karmic flax btw be aware of the missing ROPs and melted connectors for the 50 serie, many 50 serie gpus are having that issues

you should prob check out if it's happening to you too

[karmic flax]

yeah. I luckily dont have missing rops c: and my gpu is powerlimited to use 288watts. the 12v connector is rated for 600watts

[low shard]

alright, hopefully nvidia does something soon about this

[low shard]

also let me know about this

[karmic flax]

hmm i have errors but i cant dm u and i dont have perms to upload a pic

[low shard]

!give-media-perms 1h @karmic flax

[low shard]

can you try uploading the pic now?

weird, @karmic flax can you try to run the same command #✨│ai-help message , but add after install, add --force-reinstall then leave everything else as it was

[karmic flax]

already confused

[low shard]

go to cmd again, as you did before, then run env\python -m pip install --force-reinstall torch-2.6.0+cu128.nv-cp310-cp310-win_amd64.whl

[karmic flax]

so i needa get that file first

[low shard]

oh I thought you downloaded #✨│ai-help message

well, you can run env\python -m pip install https://huggingface.co/w-e-w/torch-2.6.0-cu128.nv/resolve/main/torch-2.6.0+cu128.nv-cp310-cp310-win_amd64.whl

seems like you don't need to force-reinstall, because you didn't install it in the first place

[karmic flax]

i was looking for it but i couldnt find the correct version

[low shard]

could you try the command I just told you?

[karmic flax]

yeah its downloading rn

its doing something at least lol

[low shard]

nice, try running applio after it's done

[karmic flax]

i mean red is always bad right. but ill try running it lol

[low shard]

can you run #✨│ai-help message and tell me the output?

[low shard]

shit

what if you try
env\python -m pip install https://download.pytorch.org/whl/nightly/cu128/torchaudio-2.6.0.dev20250306%2Bcu128-cp310-cp310-win_amd64.whl

this build is earlier than the one I sent you before

[low shard]

my last guess would be running: env\python -m pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

see if that works

[karmic flax]

well dang. it worked

thank uuu aloooot 🫂

[low shard]

@simple ore you might find this helpful btw

[low shard]

you're welcome

[simple ore]

yeah, they've fixed the nightly build 3 days ago

[violet badger]

Sorry I have not trained any models before. How might I do such a thing for rvc? (https://huggingface.co/spaces/TheStinger/Ilaria_RVC)

[low shard]

IlariaRVC isn't for training models, it's only for inference

what's your pc gpu

[violet badger]

3070

[low shard]

yeah you wouldn't even need to use Ilaria RVC, it's a cloud (remote good pc) service, but your pc is good enough

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

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

[violet badger]

Oh Alright, how might I train a voice though

oh

[low shard]

lmk

[violet badger]

WHich one would be easier in your opinion

[low shard]

Applio

[violet badger]

Wasnt Ai Hispano the people who made Applio hacked/

or is that fixed

[low shard]

that happened long time ago in october 2024

it's all completely fixed and safe

[violet badger]

Alright, which one would I download? I cant find the download button shown and im not sure which one I should click

[low shard]

Compiled > Windows > The download button

[violet badger]

Thank you

[low shard]

yw

[brittle wing]

Yo quick question about my RVC training. I’m at 27k steps now. Loss kept dropping but has been super slow since 10k-12k.

[analog obsidian]

hear the epochs and compare them

if its overtrained the model is going to sound robotic

while in epochs that arent overtrained they're going to sound just fine

[brittle wing]

Thanks! I'll compare and pick the best one.

[analog obsidian]

if you use spek its easier to spot overtraining, there'll be missing frequencies in the result

[brittle wing]

Oh okay, thanks! I'll check with Spek to see if there are missing frequencies. Appreciate the tip

[low shard]

#📰│dev-updates message

[knotty moth]

missing frequencies? it is when the spectrogram cutoff is lower than the target sample rate

[analog obsidian]

no like sorry i explained bad
i meant to say to check for missing harmonics lmao

overtrained models generate noise instead of harmonics at some point

🦈

[knotty moth]

imo the detail will be crisper to match the dataset but it loses some pretrain ability (that could lead to some possible robotic sounds)

[analog obsidian]

yuh

[small vortex]

Best anime latina egirl mommy wifey voice model?

[hallow thistle]

Y'all be popping out and asking for E-girl voice model just to troll and cat the damn fish someone.

[small vortex]

I actually did this to see if you would still answer but damn

https://tenor.com/view/spongebob-get-a-job-soup-spongebob-squarepants-find-a-job-gif-19390154

[hallow thistle]

https://cdn.discordapp.com/emojis/1016021305545461950.webp?size=48

[small vortex]

Like everything you do is just answering people who asks for e-girl models

lol

just actually searched your messages lol

[tame mica]

mf predicted what he would say

lowkey ts funny i fw u doe

[hallow thistle]

Like if asking anything would make you more money though.

Shit. You acting like if I do this everyday huh.

[turbid root]

Hi! How do you know if the model is overtrained?

[knotty moth]

this kind of thing somehow irritates me

[strong shadow]

Hey does, anyone know how i can get multiple (5) .VOB files and combine them into 1 whole video? No one seems to. I've tried clipchamp but there's a 1second gap/pause thats a mess.

[glacial pollen]

Not a channel for that

[long gazelle]

@low shard

[low shard]

Oh nvm this seems an error on HuggingFace side, you could try again tomorrow

[long gazelle]

ah okay

also could you explain the license thing

[brittle wing]

Anyone know the required python version to run https://huggingface.co/spaces/TheStinger/Ilaria_RVC/tree/main

I use python 3.10.16 and run into errors while installing requirements:

Building wheels for collected packages: omegaconf, samplerate, srt, antlr4-python3-runtime
  Building wheel for omegaconf (pyproject.toml) ... done
  Created wheel for omegaconf: filename=omegaconf-2.0.6-py3-none-any.whl size=36882 sha256=0b988ea25770e060c1ad0bde20dfbd7da84924e620f08626d4507bff6e337ece
  Stored in directory: /tmp/pip-ephem-wheel-cache-mgtyolse/wheels/ee/67/d9/a68a521e487bb78d6599d3a157f5bb01d0760c689a9c2ac78f
  Building wheel for samplerate (pyproject.toml) ... error
  error: subprocess-exited-with-error
  
  × Building wheel for samplerate (pyproject.toml) did not run successfully.
  │ exit code: 1
  ╰─> [4 lines of output]
      running bdist_wheel
      running build
      running build_ext
      error: [Errno 2] No such file or directory: 'cmake'
      [end of output]
  
  note: This error originates from a subprocess, and is likely not a problem with pip.
  ERROR: Failed building wheel for samplerate
  Building wheel for srt (setup.py) ... done
  Created wheel for srt: filename=srt-3.5.3-py3-none-any.whl size=22483 sha256=2ca8125c77c760943695358d90630f7dbf1a8fc14d0c479a94cc8bbaa9d08d93
  Stored in directory: /home/yui/.cache/pip/wheels/d7/31/a1/18e1e7e8bfdafd19e6803d7eb919b563dd11de380e4304e332
  Building wheel for antlr4-python3-runtime (setup.py) ... done
  Created wheel for antlr4-python3-runtime: filename=antlr4_python3_runtime-4.8-py3-none-any.whl size=141246 sha256=5175ca3614fd8bc7ba3b90590c4920a09065e395cd7a667f3d53289ec8ed5974
  Stored in directory: /home/yui/.cache/pip/wheels/a7/20/bd/e1477d664f22d99989fd28ee1a43d6633dddb5cb9e801350d5
Successfully built omegaconf srt antlr4-python3-runtime
Failed to build samplerate
ERROR: Failed to build installable wheels for some pyproject.toml based projects (samplerate)

Also if there's a better alternative please let me know

[low quail]

so a question, haven't touched ai voices in a while...
how do you make a cover of a song again? colab doesn't work for me anymore, whenever I try to load a model it gives me not found

[hallow thistle]

The better alternative to RVC is Applio.

[brittle wing]

Thanks

[low shard]

It's the license for usage of your model

[long gazelle]

ah oaky

[low shard]

Tell:

• your PC GPU
• the google colab link you're using
• the error

[low shard]

Are you trying to run it locally? It's not meant to run locally

Also what's your PC GPU and what do you want to do

AMD moment

What's your PC GPU

[tulip cloak]

@low shard hey is mailine back, the colab one

[low shard]

Maybe @simple ore knows, he used to have AMD till he recently switched to finally Nvidia

[low shard]

Nope, Hina is busy and won't be fixed for at least another week, use Applio or something else meanwhile, also you can just look at #📰│dev-updates

[tulip cloak]

Yeah I saw it, was just making sure

[low shard]

Well the batch size does depend also on the dataset length, but 8 shouldn't make your PC crash at all

It's fine

[outer isle]

I’m going to create (Mostly) Blackiana model with Apollo, is there a voice file?

[low shard]

Tensorboard to check how the model goes

[low shard]

You need to find yourself the dataset

[hallow thistle]

You use Tensorboard to track the process, to make sure the model won't be too overtrained or undertrained.

Oh, so that's what Tensorboard looked like if run locally.

[low shard]

https://docs.aihub.gg/rvc/resources/training/#usage-guide

[brittle wing]

Just a Laptop RTX 4050, I want to do this since running it online has restrictions

[low shard]

Ehh you can inference but will be limited on the training

[brittle wing]

IK, I don't train models. But huggin face has limitations even you are just inferencing isn't it?

[low shard]

Yeah that's because you're using an expensive PC GPU, google colab and Kaggle exist too btw

[brittle wing]

oh the using the CPU won't trigger the limit, thanks

[low shard]

CPU is over 10 times slower

Are you sure it's using your GPU

[hallow thistle]

If your GPU percent goes high in Task Manager, it's definitely working.

[analog obsidian]

gpu and batch size?

also dataset size matters

its normal
Amd is just very slow

150 epochs if every slice is 3s

if the slices are different lengths
around 200 epochs

ok
yes, around 200 epochs

as long the cmd is open, its fine

500 epochs is too much for small models, even when using the automated slice

ai is random so u cant just predict when its going to sound fine

is always a good practice to compare the epochs
at some point the model will naturally overtrain, which makes final epochs sound very robotic

in that case your final model would be any epoch before overtraining

[simple ore]

read the note at the bottom of the AMD installation instructions

[low shard]

AMD moment

[simple ore]

45min set on 6700xt was taking ~4min/epoch

yeah, it was an overnight training to 200e

[low shard]

tbh I thought it was way faster, wasn't zluda optimized?

[simple ore]

it is faster then colab

not as fast as kaggle

[low shard]

yeah T4x2

yeah 30 hours of GPU weekly

way better than google colab that has random daily gpu with a max of 4 hours daily

[simple ore]

use scalars tab

[upbeat chasm]

Does enyone know why when I import rvc to voice.ai app every voice sounds almost the same

[low shard]

voice.ai sucks, don't use it at all

@upbeat chasm you want realtime voice changer for calls? tell your pc gpu in #🔍│help-w-okada

I mean wokada ones are fine, they are open source

[upbeat chasm]

Then what free service should I use

[low shard]

you should use wokada, either locally or on cloud

as I said, tell your pc gpu in #🔍│help-w-okada

you can also use it on google colab and kaggle, yes

rvc realtime from mainline/original rvc is pretty old

[upbeat chasm]

I want voice change Just for mayself to use not to use on calls

[simple ore]

expand losses, or avg_50 if you have that

[low shard]

so, inference (use models) on pre-recorded audios?

still, tell your pc gpu

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

[simple ore]

does not look good at all

depending on the model

[upbeat chasm]

I know

[simple ore]

flat mel and g loss means the model is close to the dataset in the type of the data it has

generally both have to go down

[low shard]

@upbeat chasm #🔍│help-w-okada message since you got a 3060 laptop, and don't need to use the models in realtime for games/calls

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

You can:

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

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

[simple ore]

in at least the first 5k steps both mel loss and g loss have to go down

if they dont there's something wrong with the dataset or something else

too big of the batch size o something

g total

depends on the dataset size/batch size

What's your batch size?

too much for 15min

use 4

[blazing solar]

-colab

[karmic olive]

📒 Google Colab Notebooks

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

ℹ️ Note

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

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

[simple ore]

and now your model exploded

because you're training with FP16

you can stop now, it is dead

[rough parrot]

! C:\Users\user\Downloads\vcclient_win_cuda_2.0.76-beta.zip: C:\Users\user\AppData\Local\Temp\Rar$EXa23844.2909.rartemp\dist\main\web_front\assets\
yall know the reason for this?

[simple ore]

unzip the file lol

[rough parrot]

alr

[dim nacelle]

Sorry, I want to know how to control the speed of the voice, and can I produce the vocie and srt together? Thank you!

[low shard]

are you following a youtube tutorial?

the file name seems to be the latest version of Original Wokada (Wokada is used for using RVC models in realtime for calls/games)
But the wokada deiteris fork is way better than the original wokada in performance and quality

@rough parrot tell your pc gpu in #🔍│help-w-okada , this is also the wrong channel since RVC and Wokada aren't the same program

[coral viper]

What is the step to use voice models in voice.ai app?

[quasi condor]

literally don't use voice ai it sucks

[coral viper]

Why is that?

Then what should I use?

[quasi condor]

if u wanna use realtime voice changer use w-okada

or if u wanna make a cover use applio

[coral viper]

Is this free?

Where can I download?

[quasi condor]

first what's ur gpu?

[coral viper]

What's a GPU?

Where can I see my GPU?

[quasi condor]

right click ur taskbar and click task manager

[coral viper]

And then?

[quasi condor]

there should be gpu 0

click that and tell me what's ur gpu

[coral viper]

I don't quite understand this

Can you just send me the pic of the step?

[quasi condor]

@low shard u can take it from here

[coral viper]

I'm not smart

Now you speak in my language

RTX 3050, I guess

[dim nacelle]

Could Applio 3.1.1 control the speed of ai voice? and can I produce the vocie and srt file together?

[coral viper]

That's what you mean right?

Or the processor?

And then what? How to instal Applio

Ok

[low shard]

don't use voice.ai

[low shard]

are you looking for ai covers or realtime for calls

[coral viper]

No, for song cover

But I'll try another fun I guess

[low shard]

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

You can:

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

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

[coral viper]

Download compiled version right?

Kinda sus

Seriously please

On it

Why is that?

ApplioV3.2.8-bugfix.zip right?

Oh

Can we request a model? Is there a payment?

After download Applio zip, what else?

And then?

Why is it take so long to extract?

[analog fossil]

Where can I get rvc v2 voice models?

[coral viper]

@primal barn I didn't find batfild

[low shard]

You can search rvc ai voice models at:

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

if there isnt one, you can:

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

[earnest musk]

:wave: @low shard, How can I help?

Available Commands:
• @weights find <query> or /find <query> - Search for RVC Voice Models
• /create - Create an AI Cover
• /image - Generate an Image

[coral viper]

Oh

@primal barn The screen is black

Then what?

It's just dark with "Applio" on the left up corner

Yes I guess

Wait, is it require internet?

its like this

Oh yes, now it open in browser

Now what?

In browser?

Yes, and then what?

Why not in the program itself instead?

Yeah

So it's online then?

Oh cool

@primal barn Now how to put the model?

[stray raven]

hi guys i see support came for the 50 series https://github.com/IllIlIlIllIl/voice-changer/releases/tag/b2335

i have a 5080 can i get help installing the voice changer please? for some reason i cant work it out, i had a diff pc with a weaker gpu and it worked good there

[low shard]

wrong channel

RVC = Retrieval-based-Voice-Conversion, the best Speech To Speech AI Models (on v2), Inferences (use models) pre-recorded audio (ai covers) and train (make) models

Wokada = uses RVC for realtime inference

use #🔍│help-w-okada

[stray raven]

my bad

[coral viper]

How to put the model?

@primal barn So I must move the model file into that folder?

Just the pth? Not the index too?

Put the index and pth alongside the reference and the mute?

folders

Aight aight. So after I put the model ...

Then I put the song or audio I want there

And then just convert?

How to do that?

Dang, it sounds goofy

@primal barn It sounds goofy. And some part there's like a glitchy sound

Can I fix it to sound more smoothly?

[lime wadi]

hi i tried to download RVC but when i start it it gives me an error saying that the 'thinker' module is missing does anyone know how to fix this?

[coral viper]

Both

[glacial pollen]

This is quite likely due to harmonies / vocal layering / and echo / harsh reverb smearing the f0 traces too much

[coral viper]

The song. Cuz I don't know how to separate

[quasi condor]

jesus

u could've just separated it on uvr5 or mvsep

[coral viper]

Are they paid?

[quasi condor]

using melband

[quasi condor]

hell no

[coral viper]

Why melband?

[quasi condor]

its good for extracting vocals

[coral viper]

How?

I hate watching youtube tutorial

[quasi condor]

the tutorials on youtube are outdated

[coral viper]

You said that I should pay for it

[glacial pollen]

You wanna get uvr as others said, and also get fv4 model ( gabox's melband roformer voc fv4

Imho, currently the best one

Also, don't use the link they gave you, it won't have support for newest newest models

https://github.com/Anjok07/ultimatevocalremovergui/releases/download/v5.6/UVR_1_15_25_22_30_BETA_full.exe

first this one

[glacial pollen]

and then you wanna patch the uvr with:
https://github.com/TRvlvr/model_repo/releases/download/uvr_update_patches/UVR_Patch_1_21_25_2_28_BETA_small_rofo.exe

oh

there

Once you dona all, you wanna get a model I'll link

[coral viper]

I don't understand how to patch

[glacial pollen]

those are installs my man

install one, install the 2nd one

No manual work involved in that part

Once you install both in the order I specified, head to:
https://huggingface.co/GaboxR67/MelBandRoformers/tree/main/melbandroformers/vocals

and download this:

It's the fv4 model for isolation I mentioned

[coral viper]

Using 2 the files from above?

[glacial pollen]

yes, first the first link, then the 2nd

first is the base ish, 2nd is the patch

[coral viper]

Then what's thi

[glacial pollen]

It is a model for separation

which you'll use in uvr

the fv4 model I mentioned

you'll have to download it

• get my config file:

Now, focus.
Once you install the uvr, you'll have to find it's folder and then, put the model ( fv4 ) in here:

so, uvr's folder / models / mdx_net_models

( yes, you'll just cut n paste the model file in there )

[glacial pollen]

As for the config file ( the yaml file )
Once you open the uvr and select the fv4 model:

[coral viper]

I don't understand. Aren't the models are just pth?

[glacial pollen]

.pth is just a format of models used in pytorch

pth = pytorch

.ckpt is another format meaning a checkpoint

don't worry about that part

[coral viper]

Oh god, this is too much for my 12 y.o brain

[glacial pollen]

well.. If you wanna hop into vc, I might help you if you screenshare

Other than that, you gotta follow the text instructions

[coral viper]

That's just make things worse

I'll just follow your tutorial from what you sent above

If I stuck, I'll just ping you

[glacial pollen]

sounds good

[glacial pollen]

anyway, back to this.
Once you select the fv4, you'll be prompted to configure it or select the config, something along that

like this

In the model type you'll have to select " mel band roformer " ( just not the v2 variant )

[coral viper]

What should I do with yaml?

[glacial pollen]

that part is for the first box

the " select model param "

there'll be an option to open / use the yaml config

and you'll use the one I sent you

Lastly, you'll be clicking ok or was it apply for all windows

And that's that from adding custom models

Now as for configuring the uvr for usage

( do it after properly setting up fv4 model

you could also do this:

so, setting the wav type to 32 bit float

( in case you'll be using uvr for making datasets / samples for model training, else you can keep it as 16 bit )

Oh yea, in here you can use 11 or 16, I'd recommend 16 tho

Now... if you need an authentic guide for uvr, models n stuff..
https://docs.google.com/document/d/17fjNvJzj8ZGSer7c7OFe_CNfUKbAxEh_OBv94ZdRG5c/edit?tab=t.0
But I warn you, it's a messy spaghetti 👀

[coral viper]

What's this? Wdym spaghetti?

[glacial pollen]

Open the website up and you'll understand what I mean : )

as for what it is, well, exactly what it says

Documentary, just docs on uvr and models

[coral viper]

Aren't the steps you sent above before enough?

[glacial pollen]

I mean ye, it's enough, but if you wanna know more about everything, that's the place

[coral viper]

Oh ok

[glacial pollen]

mh mh

[coral viper]

I'm here. Which one I should click?

[glacial pollen]

I mean

I already wrote there which ones

Did you actually read my msgs

[coral viper]

I did, but it confusing, looks like not in order

[glacial pollen]

👀 because someone cut into my msgs making it more confusing than it should be.

Model type: mel-band roformer

model param box: from yaml config / config file or something along that

[coral viper]

In Select Model Param, which one I should click?

Or is that where I put yaml?

After I download yaml file, where I should put it?

In what folder?

[glacial pollen]

send ss

[glacial pollen]

the model param box, open it and show an ss of it

pro tip, shift+win+s lets you select the area to make a screenshot out of / crop

then ctrl+v in here

well, taking you long

Don't want to sound rude but, can't be staying here for a whole day 🙄 got stuff to do

[coral viper]

When I do that, it keeps back to before I click Select Model Param

[glacial pollen]

install new yaml

[coral viper]

Ok sorry. You don't have to reply it now

[glacial pollen]

this is what you always use when first adding new model

so, any custom models

[coral viper]

Ok

[glacial pollen]

then ye, set the type as mel-band roformer ( again, not v2 ) and confirm all

and that's all

[coral viper]

What yaml I should install?

[glacial pollen]

One I provided

[coral viper]

Oh ok

[glacial pollen]

this

[coral viper]

and then?

[glacial pollen]

Save config

then confirm

[coral viper]

Confirm where?

There is no confirm button

[glacial pollen]

Unless the window disappeared then that's fine

Either way, once you close all the windows ( if any ) you'll have the main ui

[coral viper]

Save config and close?

[glacial pollen]

ye

[coral viper]

Then how about the model type?

[glacial pollen]

......

[glacial pollen]

Why so many people lately have such attention issues

[coral viper]

Wait wait

It's 3 AM here

[coral viper]

Done. Amd then?

[glacial pollen]

got the main ui?

[coral viper]

This?

[glacial pollen]

yes

now

mdx-net, and the box below, fv4 model

[coral viper]

On it

[glacial pollen]

as for 1 and 2

you can drag n drop your music / song onto 1 area

and you can also drag some empty folder ( for instance, if you wanted it for outputs ) into 2 area

much quicker than doing it the other way

[coral viper]

Shall I must turn all of this in order?

[glacial pollen]

well no

[coral viper]

Good

@glacial pollen Oh, anyway. What version is it again?

5.6.1 right?

[glacial pollen]

of the uvr?

it's 5.6 but beta variant

just name it 5.6_beta if you want

[coral viper]

Yes

The patch you said to install was installed by me with the same way like the full version one

Is it right?

[glacial pollen]

why asking

something doesn't work?

both uvr and patch is installed the same way as anything else that is an installer

[coral viper]

Make sure that I didn't missed steps

[glacial pollen]

you'll know if you did all right once u run an isolation

If you get no errors, means you did all right

[coral viper]

It said no errors in green

[glacial pollen]

ye, then it works

[coral viper]

But why it take so long?

Is it because I want to make it in flac?

[glacial pollen]

Uhhhh, your gpu?

Aside, I never said it'll be fast

it takes time, esp at overlap 11 or 16

[coral viper]

30

[glacial pollen]

1-3 mins is pretty normal

for most songs

for instance, 3~ ish min song for me, takes 1.2 to 1.5 mins or something ( rtx 3060

so there's nothing wrong about it

[coral viper]

Aight, I take it maybe because I want it in flac

@glacial pollen So, the result will be vocal and instrument?

[knotty moth]

overlap 8 or lower is faster but 16 has optimal quality consistency (let's consider it like ultra quality in game graphics settings)

[glacial pollen]

No, vocal

fv4 is a vocal model, haven't tested it in instru mode so can't promise anything

and if you need a model that yeets the backing vocals, try this:

it is downloaded from this section

you download it, then hit refresh list and done. It'll appear in the model list

[flint solar]

Does applio Kaggle use the latest code?

[glacial pollen]

Not sure

I don't maintain kaggle or colabs so, can't help sadly

[analog obsidian]

by default no but you can easily change it to use the main branch instead

[knotty moth]

there is also melroformer karaoke by aufr33, but that can be also a decent alternative

[glacial pollen]

Oh, will have to check it out

[coral viper]

No I just want to separate vocal and instrument

And then put the vocal in Applio

And then put the ai covered vocal to instrumental back together again

[glacial pollen]

karaoke 2 is super nice and, tbf not sure how it compares to classic bve ( been ages

but kara 2 has an awful noise

so, maybe the one u pointed out is a lil better

[flint solar]

What abt the applio colab

[glacial pollen]

For instru there are probs better models

https://docs.google.com/document/d/17fjNvJzj8ZGSer7c7OFe_CNfUKbAxEh_OBv94ZdRG5c/edit?tab=t.0

check the docs

[coral viper]

Wdym?

[glacial pollen]

ctrl+f and type in instru or instrumental

[glacial pollen]

Exactly what I said. For instrumentals there are propably better models

because fv4

is a voc type model

it was made to handle vocals, mainly ( or rather, is mostly good at vocals and I can't promise anything on instrumentals' quality. Haven't tried it for that

[coral viper]

I thought the character model

[glacial pollen]

🤔

[coral viper]

Voice model I mean

[knotty moth]

even after you separate using a vocal model like gabox fv4, you should check if it contains some backing vocals/harmonies

[glacial pollen]

prev-gen models were rather universal ish, mostly
Nowadays tho, more and more models are being made with a specific specialization in mind

Instrumentals, vocals, stems, sfx, backing vocals / harmonies

and some diverge into having specific properties, such as fullness of vocals, or noise, bleedless etc

In this case, fv4 is primarily a voice / vocal model

If you need more details and explanations / overviews, the doc I sent you has all of that.
All you need to do is search for a keyword ( ctrl + f ) and have some read

[broken urchin]

can someone help me

inference doesnt work on Applio

i put in my audio and convert it and it just gives out a blank voice recording

[glacial pollen]

send ss of the console log

Also, which applio, newest one? or precompiled package

[broken urchin]

newest applio

[glacial pollen]

firstly, try to change the mp3s name

no spaces

use _s as alternative

[knotty moth]

check if the file really exist in that directory, or if not sure convert it to wav first

[broken urchin]

yeah the file is in the folder

its there

[knotty moth]

tbh it should still work on files with spaces and somewhat kanji chars

[glacial pollen]

well, supposedly ye
yet most times issues of a " file doesn't exist " sort happen, it's either the naming, corrupted file, path or lack of ffmpeg

[simple ore]

likely it is not .mp3

it may be just .wav renamed to mp3

[glacial pollen]

Can you check on some other random audio first?

gotta exclude if the file itself's not an issue

[broken urchin]

i will try

[coral viper]

Where can I get UVR-MDX-NET Karaoke 2?

[broken urchin]

i tried on different audio

still the same thing

just gives out a blank recording

also same error in the console

i also tried on different format

[glacial pollen]

I showed it on screenshot

it is download from within the uvr

[glacial pollen]

Check if you have ffmpeg file in ur applio ( folder

these

[broken urchin]

i have both of those

[glacial pollen]

Also, again, which applio you running

[broken urchin]

the one from the server

from ai hub website

[coral viper]

No. In my UVR5 only UVR-MDX-NET Inst HQ 5

[broken urchin]

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

[glacial pollen]

well

[broken urchin]

this one

[glacial pollen]

lemme rephrase. Did you download a zip

aka, unpack n run

[broken urchin]

yeah it was a zip file

[glacial pollen]

It is there, you gotta scroll down

( doesn't show in my case cause I already have it downloaded, the karaoke 2

[glacial pollen]

well, that's all weird

Normally people don't encounter such issues

nowadays at least

[broken urchin]

so what do i do lmao

[glacial pollen]

well, I can propose checking my fork maybe
if you're up for it

but that'll result in some downloading 🤔

cause like, having ffmpeg, checking the path / naming, validating on other files
that's pretty much all there is to diagnosing the issue

[broken urchin]

is that the only way?

[glacial pollen]

Those audio related problems are pretty obscure

and unclear

[glacial pollen]

Imo the easiest

as it has to work 100%, my fork, no other way

that'd indicate an issue with something else, other than applio itself at least

Well, it'd end up on downloading stuff anyways

cause you'd have to get normal applio from repo ( not the package

but while we're at it, imma recommend my fork cause why not ¯_(ツ)_/¯

Else I'm out of ideas

[broken urchin]

so i cant really fix this

[broken urchin]

its something in my pc

thats causing my issue

its not applio

[coral viper]

I did scroll, but it reached limit and no more showing

[glacial pollen]

man man man

What is written there 🙂

[glacial pollen]

Because you show me the main ui, not the settings section

[glacial pollen]

I have a hard time imagining what it'd be

Applio works within it's own environment

it's independent from ur pc

Unless you're not using it that way 🤔 ( for whatever reason

you have env folder in applio?

[broken urchin]

yes i have env

[glacial pollen]

Try this

for infer ^

put it in assets/audios

[broken urchin]

alright hold up

what language is that i dont understand anything

[glacial pollen]

That is surprising 🤔

considering ur anime pfp

It's japanese, the language

Anyways, did it work?

or nah

[broken urchin]

yeah it worked

[glacial pollen]

Then that's the case

[glacial pollen]

It's either the path u screwed up

or the naming

[broken urchin]

it gave out a recording actually

[glacial pollen]

as I said before

oh

recording?

wdym

It didn't do the inference? did output the input file?

[broken urchin]

yeah it did the inference

[glacial pollen]

well then, it works

so again

[coral viper]

So which one I should use to yeet the background vocals?

Karaoke 2 or Inst HQ 4?

[glacial pollen]

^

[low shard]

-# I mean not everyone who watches anime knows japanese though

[glacial pollen]

Show me how you previously would input the path to your audio

In here

[broken urchin]

i would just record a voice message of my voice in voice recorder and just put it in the audios folder and thats it

[glacial pollen]

well

[knotty moth]

still that ffmpeg error in the recording file? like I said try converting to wav in audacity

[glacial pollen]

then name it differently

try "test"

for the name

[broken urchin]

ok

[glacial pollen]

if that fails, then those mp3s are being screwed up in some way

or applio from precompiled had issues with mp3s? ( doubt it but Nothing surprises me anymore

[broken urchin]

ill try to rename it to test then

[coral viper]

@glacial pollen sir

[broken urchin]

yeah i just tried them out on my voice recordings again but doesnt work lol

[glacial pollen]

try hq4

karaoke 2 is for backing vocals

fv4 is for vocals

hq4 for instrus or really any other the docs recommend or mention

[broken urchin]

also the files are different for some reason

[glacial pollen]

bruh

I mean, it's just metadata so

you can always use ffmpeg to convert it, maybe could help

you got ffmpeg installed and added to path, on ur pc?

[broken urchin]

nah

i dont even know what that is honestly hahaha

[glacial pollen]

or idk, take the file to the same location you have the ffmpeg in

[broken urchin]

i just have it in the applio folder

[glacial pollen]

then open up the cmd in there

then enter and you'll get cmd

now, having both ffmpeg and ur mp3 there

type in the cmd:
ffmpeg.exe -i test.mp3 test.wav

the " -i " means input, that's your file right after it

and then, there's output, in this case test but as wave