✨│ai-help

[latent kettle]

i didn't understand that. Just tell me it matters or not. Is it good for model? Should I continue training

[simple ore]

you can discount the "raising" fm

it is not really raising, not with the average numbers like that

[latent kettle]

Okay. So I think I shouldn't look at FM ?

[simple ore]

when the average is mostly flat like that, or slowly creeping up, (+1.0-1.5) it is nothing to worry about

[latent kettle]

Emm. Lemme send you something

is is good ??

@simple ore

Should I worry about it ?

Other losses are still going down 📉

[analog obsidian]

the fm graph fluctuates a lot, dont worry too much about it

as long the rest are going down you should be fine

for g/total u should look if the graph is not too noisy
and also it should not be too flat

[latent kettle]

i think it is looking fine ??

Smoothing 0.999

[analog obsidian]

looks fine to me

[latent kettle]

I'm worried because I want to submit it for model maker role.

[analog obsidian]

how big is your dataset? and which batch size you used for training?
for submissions we aren't too harsh when we review models, we only ask for them to be functional without major problems, no need to be perfect

[latent kettle]

In future I also want to get model master

Batch size is 6
Dataset length (in minutes) 23
Total trained epochs 449/500

[analog obsidian]

sure! with enough practice you can reach this, finetuning is not that hard, what matters the most is the dataset

best approach to this is not to make it more harder and confusing

keep things simple for finetuning

[analog obsidian]

as long your g/total does not look too noisy or too flat, this setting is fine

[worn river]

#1319621721950785586 message

can we train models in portuguese?

[knotty moth]

it cannot be used in mainline rvc nor Applio public release

as it said it is for experimental purpose

[patent musk]

Hi, can someone tell me where to download the freaking GUI and package for the latest rvc inferance? I can't find it

[latent kettle]

As a noob, can I ask what is fine-tuning

[cinder grotto]

Hello question I have a model that I want to continue training but when I put it and give it start apple I only get that and I can not continue training my model.

[steel pivot]

!help

[dull iron]

Wally Commands

---

-# The prefix for commands is !

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

[tawdry gull]

LunaBot 🌙#9997

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

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

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

Important Links:
Support
Premium
Invite

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

Select A Page From Dropdown Menu Below

LunaBot · Dec 20, 2024, 2:22 PM

[analog obsidian]

training with a pretrain = finetuning
in order to train without a pretrain you need more than 44 hours of audio
that is for advanced users/dev, so always train with a pretrain

[brittle wing]

-guides

[azure marsh]

📚 Documentation

AI HUB Docs

https://docs.ai-hub.wtf

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

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

[marsh coral]

quick question, is it possible to use the rvcc to change already existing audio? like a youtube video or something? Or does it have to be real-time speaker?

[latent kettle]

you mean custom pretrain??

[latent kettle]

yes you can use

[analog obsidian]

using a pretrain in general, the original or custom pretrains

[marsh coral]

Thanks for the quick response! how would I do that?

[analog obsidian]

rvc can be used locally

-rvc

[azure marsh]

📚 Documentation

AI HUB Docs

https://docs.ai-hub.wtf

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

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

[analog obsidian]

also rvc does not stand for realtime voice conversion

[latent kettle]

RVC is not realtime voice changer. there is different thing for realtime

[latent kettle]

most of the beginers think it is realtime

[analog obsidian]

w-okada is just rvc inference in realtime

[marsh coral]

like here?

[latent kettle]

so i use orignal pretrain. that means im fine tuning?

[analog obsidian]

no, don't use w-okada, use actual rvc

w-okada is not rvc

[latent kettle]

its wokada not rvc

[marsh coral]

merci

[analog obsidian]

yup

[latent kettle]

should i submit my model now ??

i think 500 epochs are enough

[analog obsidian]

https://huggingface.co/IAHispano/Applio/tree/main/Compiled/Windows download this

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

and here a tutorial on how to convert audio files to your model's voice

pretty easy and quick

place your model in the logs folder of applio

[marsh coral]

awesome! Thanks

[latent kettle]

@analog obsidian

[analog obsidian]

if you like how it sounds, sure

[latent kettle]

why there is a little a very little noise in every model. do rvc produce it ??

[flint solar]

yes

[analog obsidian]

if the model was trained with noise and the original inference audio has noise, the model will generalize to it

normal behavior

[latent kettle]

is it possible to remove that noise

[flint solar]

how bad is the noise

[latent kettle]

i normally use UVR to process my data and i use best modles for isolating the vocals

[analog obsidian]

if the model was trained without noise you can actually use the index file to remove some of the noise

but if the model was trained with noise, nothing that you can do

[latent kettle]

can i send you a sample ??

[flint solar]

denoise the output

[analog obsidian]

noisy models will add noise to the inference output it despite the input being denoised

[latent kettle]

lemme send a sample from dataset and an output

[analog obsidian]

u can try it by converting the og pretrain to a small .pth file and inference a clean audio
the result will have noise regardless if the input audio is denoised

[latent kettle]

please stay here ill be back in a minute

[flint solar]

yeah its normal

[analog obsidian]

oki

having a bit of noise is not problematic anyways @latent kettle

a model being able to clone noise is good when inferencing noisy audio

too clean might cause the model to add weird sounds instead of noise

[latent kettle]

@analog obsidian @flint solar

[flint solar]

too much shit going on with this model

[latent kettle]

what to do now ?

i use MDX23CInstVoc HQ

[flint solar]

u made this model?

[latent kettle]

to isolate vocals

[latent kettle]

yes 😭

[flint solar]

u gotta retrain

[latent kettle]

maybe..

[flint solar]

u using the wrong models to clean ur vox

[latent kettle]

guide said.. thats why i used it

[flint solar]

-audio

[azure marsh]

Suggestions for @latent kettle

📚 Audio Guides & Tools

• Creating Datasets for RVC using iZotope RX11, by Cauthess
• Gathering and Isolating Audio, by SCRFilms ❄
• Instrumental and vocal & stems separation & mastering guide, by deton24
• Vocal Mixing Tutorial, by Roomie
• https://mvsep.com/

More commands: -colab, -uvr, -karafan, -overtrain, /help

[flint solar]

u most likely followed an outdated guide

[analog obsidian]

yup has noise

[latent kettle]

is it possible to remove it ??

[analog obsidian]

if the model was trained with noise, no

[latent kettle]

should i send a sample of my dataset ??

[analog obsidian]

sure send it here

ah wait

[flint solar]

@latent kettle why are u training on 48khz

[analog obsidian]

nono don't send dataset samples here

only model output

and again don't be too afraid of noise, is not really a bad thing

[latent kettle]

i trained on 32Khz

[flint solar]

i dont think u did

[latent kettle]

only one sample 15 sec.

[latent kettle]

im sure with it

[flint solar]

the output u sent isnt 32khz

[analog obsidian]

use spek instead, move your sample to spek and show the spectogram

[analog obsidian]

😭

[latent kettle]

okay

[analog obsidian]

i don't see noise

[analog obsidian]

rvc converts it to 32k dw

as long he selects 32k training}

for inference if you use applio you have to enable this option

[latent kettle]

im sure i selected 32Khz and im using Applio

[analog obsidian]

if the slicer and add effects options are enabled, applio will convert it to 32k

so is fine

[knotty moth]

that seems not fine

[latent kettle]

ya i did it

[latent kettle]

whats wrong with it

[analog obsidian]

very compressed

because is mp3

[latent kettle]

i used songs and then i put itto UVR 5 and selected FLAC

[analog obsidian]

yes but you converted it to mp3

so u compressed it

[latent kettle]

i used FLAC For Traning

[latent kettle]

This is FLAC

[analog obsidian]

you sent an mp3 file

but is your actual dataset in .flac?

[latent kettle]

yes..

[analog obsidian]

converting a mp3 to flac is not gonna remove the compression

hmm, as long your dataset was always .flac and never were .mp3 converted to .flac everything should be fine

training on compressed audio just makes the model a bit lower quality

[knotty moth]

the source is still mp3 anyway, a pig with lipstick will be still a pig

[analog obsidian]

should not cause more problems besides this but im not 100% sure

[latent kettle]

so basically from starting. i downloaded an audio from an XYZ website in MP3 320kbps and then i processed it into UVR and it gaved me FLAC. and then I put That FLAC into FL Studio And Just Cut bad Sounding Portions. After That i Export in In FLAC

[analog obsidian]

ok thats bad

what u did, you compressed the audio heavily (.mp3) then convert it the same heavily compressed audio to .flac

you did not increased the quality doing that

the same compression is in the flac

[latent kettle]

idk what to do.. i thought FLAC is Good. Because It Is LossLess 😭

[analog obsidian]

its good but you didn't downloaded a .flac, you downloaded an mp3 file then converted it to flac

[latent kettle]

how do i download FLAC?

[analog obsidian]

for yt videos use yt-dlp.exe or cobalt tools

[latent kettle]

cobalt tools dont have FLAC maybe. but it have WAV

[analog obsidian]

you don't need neither of them from youtube

you need .opus

[latent kettle]

is it good ?

[analog obsidian]

less compressed audio yeah

after that you convert the .opus to .wav

[latent kettle]

wav is good or not ?

[analog obsidian]

and now you have an audio that is not heavily compressed

[analog obsidian]

yup

convert the .opus to .wav

[latent kettle]

what if i download wav from cobalt tools?

is it same thing ?

[analog obsidian]

comes with a bit more compression than .opus

thats how youtube works

choose best audio quality

this downloads .opus or .webm

then you convert it to .wav

[latent kettle]

okay. now what about noise ?

[analog obsidian]

noise is not a problem unless is very loud

[latent kettle]

i also use De-EchoReverb to process my dataset

[knotty moth]

again it's just a lipstick for the pig

[latent kettle]

why there is noise in every model which i train 😭

[analog obsidian]

select split audio in applio

in inference

and also you can clean your inference audios, removing the noise from them

[latent kettle]

okay but my qustion is how do i improve my dataset

[analog obsidian]

why are you so scared of noise

that doesnt damage the model

the original pretrain was trained with very noisy audio

[latent kettle]

1 mistake>: i have used compressed audio. 2nd ? how do i process my dataset

[latent kettle]

he said i have to retrain

[analog obsidian]

separation models also adds noise to the outputs (yea a bit ironic i know)

but again is not a bad thing 😭

[latent kettle]

okay so my model is good to get Moddel Maker role 🥺

[long dirge]

how to make w okada work on whjatsapp

[latent kettle]

maybe by using virtual cable and using whatsapp web

[long dirge]

but on settings u cnat change mic

[analog obsidian]

you can't use w-okada on whatsapp

[long dirge]

sad

[analog obsidian]

sure! try submitting it

believe me, reviews are not harsh on new model makers

[latent kettle]

then idk sorry. maybe you can't

[knotty moth]

~~but you can use chatgpt on whatsapp https://www.techspot.com/news/106040-beyond-smartphones-chatgpt-now-available-landlines-whatsapp.html~~

[long dirge]

😭

[analog obsidian]

lmao

[flint solar]

potentially the worst de echo model out there

[long dirge]

can soemone gimme a realistic girl voice model for trolling 😭

[latent kettle]

which one do you use ?

[analog obsidian]

#🔍│help-w-okada
use this channel for w-okada things

[flint solar]

just use the normal de echo

[analog obsidian]

this channel is for rvc, not w-okada

[long dirge]

i to bored to switch

[analog obsidian]

there are realistic girl models for trolling there 🙂

[flint solar]

get yo ass outta here

[long dirge]

alralr

[long dirge]

cap

[latent kettle]

i use Vr Arch UVR-DeEco-Dereverb

is it normal ?

[analog obsidian]

you're not supposed to use deecho-dereverb

you are supposed to first use dereverb

then de-echo

[flint solar]

use UVR-De-Echo-Normal

[analog obsidian]

uvr-deEcho not deecho-dereverb

[long dirge]

WHAT RVC USES https://www.youtube.com/channel/UCPlDFENHg4_HX8alladwHhg

[crude flame]

depends on who you get Evilplans

[analog obsidian]

😭

[flint solar]

mantrax

[marsh coral]

getting a:

RuntimeError: cuDNN error: CUDNN_STATUS_NOT_SUPPORTED. This error may appear if you passed in a non-contiguous input.

Can anyone help?

[azure marsh]

Hey, gringochileno! Please use the command !howtoask to increase your chance of getting help by structuring your question in a way others can understand better. Also make sure you're asking in the right help channel:

• General RVC help: #✨│ai-help
• W-Okada / Realtime RVC: #🔍│help-w-okada
• AI image related: #🔍│help-ai-art

[analog obsidian]

amd gpu?

[marsh coral]

I don't know, sorry 😭

[analog obsidian]

[flint solar]

check?

[latent kettle]

also what should i use to isolate voclas

[flint solar]

bs roformer

[flint solar]

https://mvsep.com/en

here

[latent kettle]

is it avaiable on uvr ?

[flint solar]

i dont think it has the newer version

[latent kettle]

i want to use it locally

[latent kettle]

i have a decent GPU

[analog obsidian]

tldr; mel roformer kim bas curtiz

[flint solar]

never use dat bs

[analog obsidian]

big beta 5e is more clear but adds too much noise 😭

[knotty moth]

even worse than Apollo/Lew enhancer

[analog obsidian]

worst case sceneario he should use the compressed audio

[latent kettle]

for isolation ?

[analog obsidian]

rvc already does audio upscaling

no need to do it twice

[analog obsidian]

yup

[latent kettle]

is it on UVR 5

[analog obsidian]

bas curtiz no

[flint solar]

but mel roformer sdr is lower than bs roformer

[latent kettle]

i dont want to use any web

[analog obsidian]

bs roformer has muddy vocals and no one uses it anymore in the audio separation discord

[latent kettle]

is there any local alternative?

[analog obsidian]

i personally use mel for my models

[analog obsidian]

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

free colab

[flint solar]

i acc never used melr 😭

[analog obsidian]

is more clear

[latent kettle]

no colab. the actual problem in uploading and downloading speed with limited internet. i dont have a wifi at home

[analog obsidian]

i repeat: rvc already upscales your audio

[flint solar]

we get dat u got a powerful ass gpu bro js use mvsep

[latent kettle]

but

[analog obsidian]

yup, hold on

[marsh coral]

Intel (R) UHD Graphics

[flint solar]

u cant do shit with dat sadly

[analog obsidian]

https://github.com/ZFTurbo/Music-Source-Separation-Training/ https://www.youtube.com/watch?v=M8JKFeN7HfU

[analog obsidian]

sorry i don't know if its possible to inference in integrated gpus

[crude flame]

ive used it and it sounds fine

[lavish lintel]

Congratulations Razer by Weights!

Your Grotle is now level 29!

[crude flame]

https://tenor.com/view/allmight-bnha-mha-meme-fist-up-gif-13588517052470008833

[knotty moth]

note that if the dataset has lower cutoff than the target sample rate, the model will learn those missing frequencies instead of actually "upscaling" it

[analog obsidian]

im actually using it for my models now, i dont mind the noise, because again noise does not damage models, guys

[flint solar]

which version should i use

[crude flame]

im actually making my first singing model with it

[analog obsidian]

unwa's big beta 5e

[crude flame]

yes its been a year and i havent made a singing model

[analog obsidian]

is the more clear, legit sounds like an actual raw sample sometimes and you forget its isolated lol

[crude flame]

fr

[flint solar]

noted.

[latent kettle]

can you please DM Me All stuff ?

[analog obsidian]

sure! give it a try, keep in mind big beta 5e is noisy!

[marsh coral]

what about a NVIDIA GEForce GTX 1050 with Max-Q?

[analog obsidian]

i don't know if that can run cuda applications (applio)

sorry

[marsh coral]

all good

[analog obsidian]

if it can run cuda applications it can do inference

[flint solar]

u think ver 2024.10 would perform better?

[latent kettle]

can you DM Me All ??

[analog obsidian]

no, watch the tutorial

[analog obsidian]

has less noise than big beta5e but not as clear like it

[latent kettle]

i will watch just DM the links and names of models and other things

[analog obsidian]

but more clear than bs roformer at least

[knotty moth]

fine for inference, but not for training

[latent kettle]

if you are free yes you can

im confused now what to do 😭

[latent kettle]

i'm worrid i even don't no how to make a good dataset 😫 how will i get model master

💔

please help me

[analog obsidian]

this is your first time making a model, relax

[latent kettle]

and i wasted my 3hrs btw

[analog obsidian]

i waste more hours because my datasets are all 1 hour long or more

its normal

[latent kettle]

do i stop using UVR

[analog obsidian]

use the script i sent you

follow the tutorial

or

i can send you an more updated version of uvr

[flint solar]

i think i should start use mel de reverb too

[analog obsidian]

https://github.com/Anjok07/ultimatevocalremovergui/releases/download/v5.6/UVR_12_8_24_23_30_BETA_rofo_full_install.exe

this is the latest uvr update

delete your old version and install this one

[analog obsidian]

[flint solar]

which one tho

[latent kettle]

Ultimate Vocal Remover v5.6.0 i have this version currntly

[scarlet cedar]

sietangaingu

Server Avatar

[weak cipher]

No

[analog obsidian]

idk, maybe try sucial v2 dereverb

[shell holly]

Hi, I’m trying to train an AI voice model but ran into this error:
C:\ia nvidia\RVC1006Nvidia/logs/testtest
load model(s) from assets/hubert/hubert_base.pt
move model to cuda
no-feature-todo

Any idea what’s causing it?

[crude flame]

anvuew v2 is better

[analog obsidian]

thanks i have no idea about de-reverb models 😭

[knotty moth]

I didn't see the improvement over v1

[analog obsidian]

noted, im going to remember this in case i need a dereverb

[knotty moth]

anyway both Sucial's and anvuew's are for stereo reverb

[crude flame]

First is sucial V2 second is anvuew V2

sucial leaves some reverb in ( not much )

cant really hear a difference tho

so

[analog obsidian]

[flint solar]

I used anvuew mel dereverb

[dim jewel]

Hi guys, I have a question regarding dataset.
Let's say I have singer. They have 1 song that I want to make cover off, which would be the same song, with the same singer, but with different lyrics. Is there any point in training model on the other singer's songs, if model would be used for this one only particular song?

[glacial pollen]

if it's " 1 time use " model, for that one given song you use in your dataset, no, no point.

In this scenario, generalization to unseen data ( aka. Model's capability to adapt to songs / content it wasn't exposed to throughout the training ) losses the meaning

pretty much

[dim jewel]

What if it's the same scenario, but the whole album?

[glacial pollen]

Hmmm.. yea, same rule applies

If you intend to use the model on the data it was exposed to ( Again, during training), having it " full of variety " ( the dataset ) kind of losses the meaning

[silk hearth]

/create

wow that worked

[dim jewel]

But wouldn't it still needed to learn variety to adapt to changed lyrics?

[glacial pollen]

Oh yea, if you mention changed lyrics, in that case you want the dataset to contain ( more or less ) phonetics and / or pitch variations of a given phonetic / word

But it isn't 100% a strict rule

All comes down really to how well you model generalizes ( It's ability to adapt to stuff it did not see during training )
And generalization is a matter of: Good training, properly picked batch_size ( smaller promotes better generalization, typically ) and naturally, dataset's diversity

But let's not make it any huge or extreme deal really

[dim jewel]

As always, thank you

[glacial pollen]

Glad to help

[fair pelican]

yo i need help with setting up the voice changer to work in discord, the output virtual cable is in input and the input virtual cable is in output how do i fix it help

[valid stream]

i put my chunk at 640 but its still like 3 seconds delayed how do we fix it

[glacial pollen]

#🔍│help-w-okada is for voice changers

Both of you

[low shard]

#🤖│bots

[low shard]

#🔍│help-w-okada

[low shard]

#🔍│help-w-okada

[weary pond]

hello I'm trying to set up a two PC RVC, i followed this guide https://rentry.co/VoiceChangerGuide#opening-on-multi-pc-setups but for some reason the IPs only loads in the PC with RVC installed and not the secondary PC where i want to connect at, anyone can help?

[low shard]

That's Wokada not RVC

Wokada is the program to use RVC (Retrieval-based-Voice-Conversion, Speech To Speech) Models in realtime for calls

There's the fork (modified version), the deiteris fork which has better performance

Wrong help channel, use #🔍│help-w-okada

[weary pond]

my bad

[low shard]

Dw

[unique rock]

what is this for?
cache_all_training_sets

[simple ore]

cache training data on GPU, provided small performance improvement, as long the as the dataset is not too big to fit into vram

@unique rock

[humble river]

anyone on rn

does not even mention its an error

just says this

2024-12-20 21:56:06 | INFO | fairseq.tasks.hubert_pretraining | HubertPretrainingTask Config {'_name': 'hubert_pretraining', 'data': 'metadata', 'fine_tuning': False, 'labels': ['km'], 'label_dir': 'label', 'label_rate': 50.0, 'sample_rate': 16000, 'normalize': False, 'enable_padding': False, 'max_keep_size': None, 'max_sample_size': 250000, 'min_sample_size': 32000, 'single_target': False, 'random_crop': True, 'pad_audio': False}
2024-12-20 21:56:06 | INFO | fairseq.models.hubert.hubert | HubertModel Config: {'_name': 'hubert', 'label_rate': 50.0, 'extractor_mode': default, 'encoder_layers': 12, 'encoder_embed_dim': 768, 'encoder_ffn_embed_dim': 3072, 'encoder_attention_heads': 12, 'activation_fn': gelu, 'layer_type': transformer, 'dropout': 0.1, 'attention_dropout': 0.1, 'activation_dropout': 0.0, 'encoder_layerdrop': 0.05, 'dropout_input': 0.1, 'dropout_features': 0.1, 'final_dim': 256, 'untie_final_proj': True, 'layer_norm_first': False, 'conv_feature_layers': '[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2', 'conv_bias': False, 'logit_temp': 0.1, 'target_glu': False, 'feature_grad_mult': 0.1, 'mask_length': 10, 'mask_prob': 0.8, 'mask_selection': static, 'mask_other': 0.0, 'no_mask_overlap': False, 'mask_min_space': 1, 'mask_channel_length': 10, 'mask_channel_prob': 0.0, 'mask_channel_selection': static, 'mask_channel_other': 0.0, 'no_mask_channel_overlap': False, 'mask_channel_min_space': 1, 'conv_pos': 128, 'conv_pos_groups': 16, 'latent_temp': [2.0, 0.5, 0.999995], 'skip_masked': False, 'skip_nomask': False, 'checkpoint_activations': False, 'required_seq_len_multiple': 2, 'depthwise_conv_kernel_size': 31, 'attn_type': '', 'pos_enc_type': 'abs', 'fp16': False}
2024-12-20 21:56:10 | INFO | infer.modules.vc.pipeline | Loading rmvpe model,assets/rmvpe/rmvpe.pt

then at 5.3 secs the gradio thing says error

[simple ore]

@craggy wyvern this is the right channel to ask your question

the answer is - depends on what you're trying to install. Likely some outdated build.

[hallow thistle]

-gui

[azure marsh]

https://cdn.discordapp.com/attachments/1122285248844144733/1203460490475343953/caption.gif?ex=65d12cec&is=65beb7ec&hm=bd2fb8d010006dd7c6e3c1c67d3ae846fd1478e1a3124c544c31b43086fe54aa&

[hallow thistle]

A batch file for launching up webui went missing from this most recent OG RVC GUI, but instead got this batch file for launching up the "realtime" RVC pre-installed. Of course, I don't think it gonna work well.

[low shard]

Wrong channel, use #🔍│help-w-okada

[hallow thistle]

The first time I clicked and run this batch file back in 2023, I was like surprised by how this GUI looked. Then later I found out it wasn't meant for "audio conversion" thing, but rather the "realtime" one.

I'm surprised by how people still think RVC has another realtime program. Could it be this thing?

[glacial pollen]

well, this is the " in-built rvc's real-time voice changer " I always mention

not sure if there's any other tho

[brave ermine]

hello

i got an error while generating index

[azure marsh]

Hey, muhamet! Please use the command !howtoask to increase your chance of getting help by structuring your question in a way others can understand better. Also make sure you're asking in the right help channel:

• General RVC help: #✨│ai-help
• W-Okada / Realtime RVC: #🔍│help-w-okada
• AI image related: #🔍│help-ai-art

[brave ermine]

im using local applio and it says v2 extracted file is not found

anyone know what i am doing wrong

nah i found it i just didnt press this button how silly lol

[humble river]

anyone know fix

[humble river]

at 5.3 sec it says eroor

3070ti 13700k btw

[humble river]

you on rn?

[glacial pollen]

ye

[humble river]

oh

ok

[glacial pollen]

what's up then

[humble river]

do you know a solution for this

no error in the logs

just that its loading rvmpe

[glacial pollen]

Well, 2 things

[humble river]

alr

[glacial pollen]

1. You should provide logs using " > ", like so:

asdfghjkl

and 2... what release / fork you use? ( of rvc / applio )

This is by far the most important part / info you didn't provide I believe

[humble river]

ok

[humble river]

sorry ill do that again

[glacial pollen]

Actually nah

[humble river]

i installed it yesterday

[glacial pollen]

Just tell me what release you use

where does it come from

applio github? or official rvc one

[humble river]

offical

sry not that good eng

[glacial pollen]

Yea, tbf, instead of trying to debug og rvc which can at times be problematic or headaching
I'd actually go for Applio

[glacial pollen]

It's fine

[humble river]

ty

[glacial pollen]

Applio is easier to run and generally to get running

[humble river]

does it have the same quality

[glacial pollen]

More or less yes, it's with everything that's based off of rvc really

[humble river]

oh

do u have the link

wanna make sure i download the right one

or the user who made it

will do

[glacial pollen]

You're a newbie right?

[humble river]

yes

[glacial pollen]

If so, I recommend this one
https://github.com/codename0og/codename-rvc-fork-3/archive/refs/heads/main.zip

Has some nicer descriptions / easier to get descriptions I added

[humble river]

alr

[humble river]

ty

[glacial pollen]

overall, you'll run 2 things

[humble river]

pip install and run app

[glacial pollen]

1st goes the install .bat and then run .bat files

[humble river]

oh

[glacial pollen]

just, simply run it with no " run as admin " or whatsoever

[humble river]

ok

[glacial pollen]

In any case, read the repository's description, if you had some troubles
https://github.com/codename0og/codename-rvc-fork-3/blob/main/README.md

[humble river]

ok tysm

[glacial pollen]

yw man

( ps. Make sure to unpack the fork folder / the folder from archive to C drive / os drive directly, if you can afford some space on your drive )

Like so

[humble river]

i got 2tb im alg

[glacial pollen]

Neat, in that case best of luck~

[humble river]

ty

YOO UI IS FIRE

[idle osprey]

nword

nword

nword

nword

[humble river]

@glacial pollen where do models go?

just directly in models

or

[glacial pollen]

logs folder and in there, per-model folder

[humble river]

ok

[glacial pollen]

This is also the case when you train a model, .pth models appear in there ( index too )

[humble river]

ty

[humble river]

thats not for me lol

but ty

[glacial pollen]

yea just saying in case

[humble river]

that can possibly help

there is no per model folder

should i make it

[glacial pollen]

Oh I meant

[humble river]

so creare folder

[glacial pollen]

per-model as in, each model ( pth and index ) gets a folder

[humble river]

OHHH

[glacial pollen]

per is like " for "

[humble river]

so should i make one with the pth and index

[glacial pollen]

Yea, applio searches up for models in logs location

[humble river]

oh i didnt grab that in context ;p;

[glacial pollen]

and organizing it in folders makes it easier

that's all there is to it

[humble river]

ah

alr ty

[jaunty talon]

One message removed from a suspended account.

[low shard]

could u retry?

i just tried myself and the bot works #🤖│bots message

[jaunty talon]

One message removed from a suspended account.

[dense drift]

Please help

[tame mica]

?

what are you trying to do

[lavish lintel]

Congratulations kar@shin padoru 🎄!

Your Dewott is now level 34!

New move!

Your Dewott can now learn Aqua Jet!

‎

‎

‎

‎

[dense drift]

Add a song

To create an ai song

[glacial pollen]

That is not rvc tho ( I mean yea, but weights manages it

[dense drift]

Soo where should i ask for help

[glacial pollen]

So like, if yt dls are disabled

you gotta dl the vocals on your own

Not sure how weights manages it tho

Whether they isolate ( the vocals ) or not

Well.. You'd have to dl the vocals / song, if it's a song then isolate / separete the vocals using mvsep or uvr and only then, upload the vocals to weights

[dense drift]

There is no option to upload

[glacial pollen]

Idk then man
they say to upload it in afterall

I ain't associated with weights nor I use their services so, I possibly couldn't know how it's done in there

[dense drift]

Ohh alright

[glacial pollen]

Alternatively, screenshot the whole ui

[dense drift]

But thanks for helping

[glacial pollen]

and show me how it looks

[dense drift]

Its fine

[glacial pollen]

alr then

[torpid prairie]

can i use a voice model with only pth file

[low shard]

yes, in rvc context, the pth is the actual voice, the added index is the accent

may not sound the best though

[torpid prairie]

alr

[severe sand]

Does anyone have experience getting "sh", "ch" and "sch" sounds to be pronounced properly? like in the German "Ich"
Everything I try, every model completely fails at those sounds even if they exist in the dataset

Unfortunately its pretty important since we use them for music but I just can't find any solution

[crude flame]

How long is your dataset?

[severe sand]

about 15-20 minutes of talking and singing on average

[crude flame]

if possible you could try making it longer

[severe sand]

nvm the main model I'm looking at is trained on 32 minutes of talking and singing

mostly singing, I'd think thats way more than enough, especially considering the models are completely unable to pronounce "Ich" like its not even really close

they always turn it into "isch" if they pronounce it at all

[crude flame]

so it has "ich" in the dataset and in the inference audio?

[severe sand]

yes

it seems a bit like most of the time rvc even thinks that the ch is supposed to be breath noise and supresses it entirely

[crude flame]

how noisy is your set? little bit is fine but if its loud enough it could make a difference

[severe sand]

No noise, studio environment

not sure if you can pronounce the german ch but do you think you could find a clip were its being pronounced correctly? I can try making one showcasing the issue

[crude flame]

welp, you can blame vctk for not having any german in it making that "ich" suck

[severe sand]

whats vctk?

[crude flame]

the dataset the default pretrains were trained on

it sucks

[severe sand]

ah I see, yea I was worrying that this might be the issue

I've tested some other pretrained models but they seemed mostly terrible

so I didn't even keep those models

[crude flame]

if you want you can grab several hours of german speech and make a small little finetuned pretrain for german

and hope it makes it better

[analog obsidian]

original pretrain was trained in english only

[severe sand]

Anything specific I have to do for that? Like maybe I have to somehow reduce the learning rate or something?

[crude flame]

nope

just train it like a normal model

[severe sand]

And do you mean using the pretrained models and finetuning it into a better one or make a new model from scratch

I do kinda hope that this is the issue now, else I will put a bunch of time into that without fixing it xD

I will try some other pretrained models first, this one could be interesting

[crude flame]

sorry a bug scared the life out of me, but yeah make a pretrain using a pretrain

[severe sand]

And you are dead sure that those kind of noises are results of bad pretraining? Because I feel like even english noises like "shark" are kinda struggling with the sh

[crude flame]

dont use rigel

[lavish lintel]

Congratulations Razer by Weights!

Your Grotle is now level 31!

[crude flame]

it sucks

[severe sand]

welp damn, now where am I gonna find hours of german. It probably has to contain male and female voices right? Since I will need both

[crude flame]

you done need both but it will be better with both

[severe sand]

Okay I will try around with that a bit, thanks

Oh another question, do I have to fear overtraining when creating a pretrained model? And a general idea of how many epochs I will need for multiple hours of data? I assume it will take quite a while even for 50-100

[crude flame]

its the same as training a model, so overtraining is bad and there is no certain amount of epochs

[severe sand]

okay thats much easier than I assumed then, good to know

[low shard]

https://tenor.com/view/yikes-gif-21831957

[carmine hearth]

Hey guys, is there anything written about audacity dataset cutting settings? I'm looking for it for people using RVC Mainline or RVC Disconnected (that's me!), but my inexperienced searching skills have yet to find it...
I'm using machine translation. Sorry if it's hard to understand!

[placid heath]

i just had a question, to make songs with ai, like to rap as different things, what is recommended? like settings and everything. im new to this so please be patient with me!

[simple ore]

ideally you want an evenly cut audio, some overlap between segments

3-5 seconds, no more than that and overlap 0.3-0.5s

[carmine hearth]

Thank you for your help, I will try that.

[mental spade]

I'm having issues it's just not comign through anything I can hear it on the client thats it
https://i.imgur.com/EJEuzG3.png
https://www.youtube.com/watch?v=IS_SPQVv5iY Was watching this

Think I fixed it

[hallow thistle]

-colab

[azure marsh]

📒 Google Colab Notebooks

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], by Eddy, Hina and Gdr Google Colab
• UVR5 NO UI, by Eddy Google Colab
• UVR5 UI, by Eddy Google Colab
• Modified W-Okada's Voice Changer, Google Colab
• 🆕 FaceFusion UI, by Nick088 Google Colab
• 🆕 FaceFusion NO UI, by Nick088 Google Colab
• 🆕 EasyGUI, by Rejekts Google Colab

ℹ️ Note

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

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

[hallow thistle]

Tutorial videos on YouTube are outdated.

[knotty moth]

also they could have video editing magic, so it'd look too good but not in reality

[low shard]

YouTube tutorials are SEVERLY OUTDATED, do NOT use them

Go to #🔍│help-w-okada and tell ur PC GPU

This is the wrong channel, and you're using old software

[low shard]

You want to make ai covers? What's ur PC GPU

[unique rock]

Can someone explain to me the Tensorboard graphs? Following the Applio guide I only learned to see the total loss g, but they say that I should not only take into account that graph but also more.

[low shard]

https://docs.ai-hub.wtf/rvc/resources/epochs-tensorboard/

There's a more advanced guide here

[simple walrus]

For rvc-gui I create a voice model for Ai Cover. When I use the voice model I create in any song, whether male or female, no matter if I make the pitch negative or positive, the voice I use in the song is not the same as the voice model I created in the song. Why?

[low shard]

https://github.com/Tiger14n/RVC-GUI

Do u mean this one

It's super OUTDATED

What's ur PC GPU

There's a better program for realtime voice changer

[hallow thistle]

-gui

[azure marsh]

https://cdn.discordapp.com/attachments/1122285248844144733/1203460490475343953/caption.gif?ex=65d12cec&is=65beb7ec&hm=bd2fb8d010006dd7c6e3c1c67d3ae846fd1478e1a3124c544c31b43086fe54aa&

[low shard]

@simple walrus You looking for ai covers or realtime voice changer?

[simple walrus]

@low shard I want to use my own voice in the rvc-gui but when I use it the voice does not change to my own voice in the song

[unique rock]

Can you tell me which is the best pre-train?

[low shard]

don't use rvc gui, it's outdated

sooo, just do ai covers

or realtime voice changer for calls?

[low shard]

there isn't

it depends

[hallow thistle]

RVC is the audio conversion program, while W-Okada is the realtime voice conversion program that uses RVC voice model.

[low shard]

^^^

[simple walrus]

@low shard There is no alternative program can you recommend me an ai cover program where I can install on pc without connecting to websites

[hallow thistle]

The "RVC-GUI" can refer to the OG RVC GUI program, which has been long outdated.

-rvc

[azure marsh]

📚 Documentation

AI HUB Docs

https://docs.ai-hub.wtf

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

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

[low shard]

when you say RVC-GUI, do you meanhttps://github.com/Tiger14n/RVC-GUI/blob/main/README.md

if so, that's OUTDATED, DON'T FOLLOW YT TUTS

Tell me your pc gpu

[hallow thistle]

You still didn't answer us about what GPU your PC has.

Applio is a recently developed fork program of RVC GUI, one of the only RVC forks AI Hub by Weights recommended.

[simple walrus]

My notebook cpu is Intel® HD Graphics 500

[hallow thistle]

Is it only GPU 0? If so, that's mean your laptop doesn't have a dedicated GPU.

[low shard]

that's not a cpu, that's integrated GPU, it's really bad and slow

The program won't even run on integrated gpu, it will run on your cpu

making it very very slow

It's suggested to use Cloud, your CPU is SLOW

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.gg: Partnered with AI Hub, lets u do them easily but u may be in a queue
  • Applio Colab: max 4 hours, not granted, of GPU

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

if you really want to do it on cpu slow, you can locally via applio

[lavish lintel]

Congratulations Nick088 [ITA/ENG] by Weights!

Your Charizard is now level 76!

[low shard]

but It's NOT suggested

[simple walrus]

Applio I used it but it is a bit complicated it connects to website is confusing

Thxx for Suggestions I will try

[steel forge]

applio hands down the best

imo

[hallow thistle]

Applio doesn't connect to the "internet", it hosts locally on your PC localhost port. Unless you set Gradio to share to around the world.

[steel forge]

yeah that too

runs without an internet connection but requires a network

all RVC GUIs will be web-based. Unless someone wants to cobble together a modern RVC-GUI

[low shard]

it actually does connect to the internet

they use Edge TTS API

[steel forge]

but once you download the pretrains and stuff you can infer no internet right

no tts

[low shard]

if u use tts u need internet

without it yea u don't

[steel forge]

bet]

based applio

[hallow thistle]

An internet connection is needed to download voice model online.

[low shard]

You probably used the Google colab, ofc that one does

The local applio doesn't unless you do TTS

There are different versions of the same program, cloud and local

[hallow thistle]

-colab

[azure marsh]

📒 Google Colab Notebooks

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], by Eddy, Hina and Gdr Google Colab
• UVR5 NO UI, by Eddy Google Colab
• UVR5 UI, by Eddy Google Colab
• Modified W-Okada's Voice Changer, Google Colab
• 🆕 FaceFusion UI, by Nick088 Google Colab
• 🆕 FaceFusion NO UI, by Nick088 Google Colab
• 🆕 EasyGUI, by Rejekts Google Colab

ℹ️ Note

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

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

[low shard]

If you really need it offline, you can install Applio locally but will be extremely slow

Else I'd suggest to use weights.gg which is on cloud

[hallow thistle]

Google Colab is a cloud service.

[simple ore]

you did not read the advanced section?

[unique rock]

just did it

[placid stone]

hi, I have a question, I can only switch between GPU0, CPU, GPU1, GPU2 and GPU3 in the voice changer and my voice changer is also lagging a lot. can someone help me?

[azure marsh]

Hey, Jan.! Please use the command !howtoask to increase your chance of getting help by structuring your question in a way others can understand better. Also make sure you're asking in the right help channel:

• General RVC help: #✨│ai-help
• W-Okada / Realtime RVC: #🔍│help-w-okada
• AI image related: #🔍│help-ai-art

[hallow thistle]

For W-Okada, go to #🔍│help-w-okada. #✨│ai-help is about RVC the audio conversion.

[sage heath]

my voice changer is capturing my pc voice too and its eco what do i need to do

[hallow thistle]

For W-Okada, go to #🔍│help-w-okada.

[sage heath]

what?

[hallow thistle]

Please read my earlier message above.

There's no way your PC has four GPUs at once. Unless you've downloaded the old OG W-Okada, which can be tricky to cause its GUI to show more GPU than one, and each GPU could all be picking up CPU.

[low shard]

you're using the wrong channel, RVC is NOT Wokada, use #🔍│help-w-okada

[left crow]

I have a question I wanna make ai covers but someone told me to use rvc is it only pc or is it on mobile too?

[low shard]

It's technically for both

What's ur pc gpu?

[left crow]

I have hp laptop but my mom uses it for work so I use mobile

And what's a GPU im very dumb in this kind of things lol

[low shard]

GPU = Graphics Processing Unit
The component used for every heavy task

[low shard]

yea laptops aren't good either for that

[hallow thistle]

It is possible to run RVC entirely on smartphone, but it won't be as fast as desktop.

[left crow]

Ohh

It's fine I only wanted to make lads ai covers

But I don't know the website lol

There's a lot of website idk what to choose

[hallow thistle]

Weights.gg is a website that can do AI cover for free.

[left crow]

Ohh I use that but idk some part of the song breaks but it's good ig

Maybe it depends on the song lol

[low shard]

that depends on the model and song

you can't do much about it

Weights.gg uses RVC but in a easier way for users

Other sites use RVC too, but they make you pay for it

[charred drum]

bro is marketing weights.gg

[low shard]

Weights.gg doesn't make u pay 70 dollars monthly like kits.ai atleast

[charred drum]

[opal kelp]

-colab

[azure marsh]

📒 Google Colab Notebooks

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], by Eddy, Hina and Gdr Google Colab
• UVR5 NO UI, by Eddy Google Colab
• UVR5 UI, by Eddy Google Colab
• Modified W-Okada's Voice Changer, Google Colab
• 🆕 FaceFusion UI, by Nick088 Google Colab
• 🆕 FaceFusion NO UI, by Nick088 Google Colab
• 🆕 EasyGUI, by Rejekts Google Colab

ℹ️ Note

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

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

[alpine bough]

hi, i downloaded the ngrok file, but cannot open it, can u help me?

cannot find or open /Users/batokaraevzafarbek/Downloads/ngrok-v3-stable-darwin-arm64.zip, /Users/batokaraevzafarbek/Downloads/ngrok-v3-stable-darwin-arm64.zip.zip or /Users/batokaraevzafarbek/Downloads/ngrok-v3-stable-darwin-arm64.zip.ZIP.

i tried through collab, but there is a 403

по русски можно

-rules

rules

@acoustic scarab

[acoustic scarab]

#🌏│русский

https://media.discordapp.net/attachments/1026837636024901663/1076879572886364241/7qdsjRATzn8.gif

[alpine bough]

помоги

пэжэ

[low shard]

speak english here, or use the channel that @acoustic scarab said

[low shard]

what are you trying to do?

[alpine bough]

so i was trying to set up through the collab

and firstly i was getting 403 errors so i tried using ngrok

and getting erroro there too

[alpine bough]

should i use help-w-okada channel>

?

[low shard]

yes, please show screenshot and elaborate more on #🔍│help-w-okada , also tell me what colab & tutorial you're using (send the link)

you shouldn't download ngrok,that gets downloaded on google colab not ur pc

[upper tusk]

Hello there, I have a question, I would like to try and make an RVC model based on sherry birkin from RE 2 remake, but I have only maybe 2-3 minutes of dialogue from her that is usable, is that enough to work with or is it just not going to work ?

[brittle wing]

Hewwo! I was wondering if anyone can help me to get my voice onto discord and games?

just struggling a lil, got the virtual audio cable installed etc, just dunno how to do it

[brittle wing]

also, when playing some games, the ping and total MS skyrockets, that normal?

https://i.ibb.co/W6wNxvm/chrome-f-Hk-PJ8-Zssi.png

[low shard]

Wrong channel, use #🔍│help-w-okada

RVC ≠ Wokada

[tranquil raven]

!help

[dull iron]

Wally Commands

---

-# The prefix for commands is !

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

[woeful canyon]

Available Commands

!ping - Check bot latency
!help - Show all commands
!status - Show bot status

[tranquil raven]

Where is the link for actual rvc?

!help

[dull iron]

Wally Commands

---

-# The prefix for commands is !

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

[tranquil raven]

I'm tryna find the link for this version of rvc

[hallow thistle]

This is the realtime .py for RVC, which can be found inside the folder of the most recent original RVC GUI. However, the OG RVC GUI has been long outdated, and the realtime RVC won't be working as well as the fork W-Okada.

https://cdn.discordapp.com/attachments/1159290139609137264/1320055226409160774/image.png

I'm not sure why you're looking for this specific RVC version when there's a recently developed real-time voice conversion program available, which it works better than that.

If you're looking for the best real-time program out there, let me know at #🔍│help-w-okada.

[open stag]

• colab

[azure marsh]

📒 Google Colab Notebooks

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], by Eddy, Hina and Gdr Google Colab
• UVR5 NO UI, by Eddy Google Colab
• UVR5 UI, by Eddy Google Colab
• Modified W-Okada's Voice Changer, Google Colab
• 🆕 FaceFusion UI, by Nick088 Google Colab
• 🆕 FaceFusion NO UI, by Nick088 Google Colab
• 🆕 EasyGUI, by Rejekts Google Colab

ℹ️ Note

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

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

[hot sonnet]

!howtoask

[patent trellis]

How To Troubleshoot

__**GIVE CONTEXT.**__ 📝

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

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

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

__**BE PRODUCTIVE.**__ 🤝

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

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

[hot sonnet]

what is the best download for amd :(

[low shard]

this one is very old and sucks, check the fork i was saying in #🔍│help-w-okada

[low shard]

for what

[hot sonnet]

to be honest I don't know, because i thought okada was rvc but i think rvc is okada

[low shard]

😭

what do u want to do

[hot sonnet]

idk

the difference

[low shard]

wdym u don't know

[hot sonnet]

i just want a smooth voice changer

idk ive used this before

[low shard]

yea u can just say that dw

let's go to -> #🔍│help-w-okada

[hot sonnet]

ok ty

[tough stone]

my audio is around 3 minutes long and its separated into like 10 shorter audios, is THIS normal?

Also my epochs are 250 and save time 30

rmvpe_gpu, no other settings touched

rtx 2060 S

its been around 20 minutes, no changes/updates in the console as well

[latent kettle]

Are you training a model? Show me the console screen

[tough stone]

Nvm its solved, I just had to uninstall Python

[flint solar]

what?

😭

[tough stone]

idk bro

It caused some issues

most important thing it works

[supple salmon]

Any idea what I should go about this? :o

I believe the page has been taken down

[low shard]

btw if u tryna download a model for wokada, u can just continue in #🔍│help-w-okada

link me the discord post where u got that model

[supple salmon]

Dunno how

[supple salmon]

Also I thought it was related since rvc thingy

[low shard]

yea but u using it for wokada, so maybe let's go to #🔍│help-w-okada

[molten fog]

is this option any good?

assuming not i havent heard any talk of it

[pallid ocean]

Is the applio-rvc colab ver. working for anyone? I seem tobe getting an error related to circular imports upon training the model (index seems to be training without any problem)

[dire vapor]

how to use rvc v3 in applio colab or other colab?

[silent stratus]

depends on if you cleaned your data well or not

[low shard]

Rvc v3 doesn't exist

Officially

Since like 2 years

There's some unofficial forks that are experimental like @glacial pollen 's fork but idk if there's a colab

[low shard]

Show a screenshot

[silent stratus]

i had that error

[silent stratus]

i had that error and idk what i did spicifically but i kept switchiung google accounts and it eventually worked

[dire vapor]

I said Codename fork

[low shard]

anyways i pingd codename, he prolly knows

[low shard]

wtf never seen that before

looks like smt isn't installed properly

you told noobies about it?

[silent stratus]

yeah i did it seems like a new issue that hasnt happened before and a lot of ppl are getting it

noobies dosent even know why its happening

[simple ore]

i said you may need an update to colab

Vidal did fix some torch imports

you can just manually re-install torch

!pip uninstall torch torchvision torchaudio -y
!pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 --upgrade --index-url https://download.pytorch.org/whl/cu121

[silent stratus]

okay thanks

[low shard]

u should tell that to vidal or push the update if u can

coloab moment

[molten fog]

if my dataset has like multiple singing tones (calmer singing tones and then more almost "yelling" singing tones where the artist puts more emotion in their voice), is that what this is talking about when it says "diverse"? shoudl i use a bigger batch size?

[azure marsh]

Dataset

A set of audio files compressed into a .zip file, used by RVC for voice training. The quality & length of the dataset are the biggest determining factors of the final quality of the model.

[analog obsidian]

diverse as different expressions, words, tone, etc

for example a non diverse dataset would be one having monotone speech, with repeated words in between speech, similar sentences, basically 0 variety

if your singing dataset 90% of the time is singing in the same tone then is not diverse enough

for batch sizes you can try 4-8
4 for small datasets (10 minutes and below)
and 8 for 30 minutes and above

choosing batch sizes is more complicated than that but for training models these are the most used values in rvc

[molten fog]

should i just do 8 my dataset is 14 minutes

it is pretty diverse in terms of tone

btw

[analog obsidian]

14 minutes is not big, so you only got the diversity

[molten fog]

so should i meet in the middle at like 6

its longer than 10 minutes but not thirty minutes long

[analog obsidian]

do 8 if you want the model to be cooked fast and quick
or do 4 to squeeze everything from the dataset aka having better generalization

up to you

[molten fog]

ill do 4 since i prefer accuracy

i have no problem wating as im training locally and im patient

waiting*

[analog obsidian]

sure batch size 4 will work fine in this case

as long the graphs are not extremely noisy

[molten fog]

for avg running loss do i train one epoch get like 20% of what the total steps were and then input it and start training again

[analog obsidian]

yup

follow codename's suggestions

[molten fog]

do i need it enabled to 0 when i train the first epoch or do i enable it after and resume training

[analog obsidian]

both works

[silent stratus]

is this overtraining i cant tell plus the overtraining detector isnt going off

but it seems to be going up

[analog obsidian]

yes

[silent stratus]

okay thanks

[clear oasis]

How do I train a model? I'm just starting out

[low shard]

what's ur pc gpu?

[brittle wing]

Yes seems like

10k is the Lowest point mhm

[low shard]

Open "extracting vocals from songs"

[flat yoke]

ModuleNotFoundError                       Traceback (most recent call last)
<ipython-input-3-bd2dc64d26a0> in <cell line: 2>()
      1 #@title Save the model
----> 2 from mega import Mega
      3 import os
      4 import shutil
      5 from urllib.parse import urlparse

ModuleNotFoundError: No module named 'mega'

---------------------------------------------------------------------------
NOTE: If your import is failing due to a missing package, you can
manually install dependencies using either !pip or !apt.

To view examples of installing some common dependencies, click the
"Open Examples" button below.
---------------------------------------------------------------------------

can someone tell me why i get this error?

im trying to save this model:

https://huggingface.co/meaisae/NCTMODELS/resolve/main/taeyongrap-modelrvc2_rmvpe.zip?download=true

nvm

[low shard]

Be sure to not use Ilaria rvc Google colab

[flat yoke]

which one should i use

[low shard]

Tell me what you want to do

And if possible, your PC GPU

(I'm asking your PC GPU bc I seen people with actual good PCs using cloud

[flat yoke]

rtx 3050

[flat yoke]

just use some models for short voice lines

[low shard]

Is it the 4gb laptop one

[flat yoke]

no

[low shard]

6gb?

[flat yoke]

wait yeah it is the 4gb one

[low shard]

Yeah nvm

[knotty moth]

the 4gb one is laptop

[low shard]

• 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.gg: Partnered with AI Hub, lets u do them easily but u may be in a queue
  • Applio Colab: max 4 hours, not granted, of GPU

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

Ilaria rvc zero and weights.gg are prob your best options

[flat yoke]

yeah using weights rn

wait

are they using the same models tho

?

[low shard]

All models on ai hub are synced on weights.gg

So unless you're comparing 2 different models, they prob are the same

They still are RVC v2 ofc

[oak edge]

hi, to get best quality in rvc model should batch size be higher or lower? (I'm using colab free, applio, and T4 gpu)

[knotty moth]

batch 8 generally, or 4 for short/less diverse dataset
also fp32 may give better "quality" (precision) and gradient stability than fp16, but double the vram usage and slower. batch 8 fp32 would fit the T4 vram tho, or better the kaggle notebook with 2x T4 gpu.

[oak edge]

i got a 20minute single file datasset also what is fp

[knotty moth]

floating point precision for the model params

[oak edge]

i mean where to configure it

[knotty moth]

go to Settings tab in Applio

I'd recommend the latest version 3.2.8 bugfix

[oak edge]

ohh see it

fp32 always better than fp16?

[knotty moth]

it seems too technical to explain quality-wise, except the time and vram usage difference, you can just try it by yourself

[oak edge]

okkkk

thanks a lot

[mental spade]

So it seems now virtual cable died?

got it to work last night

Computer crash nwo nothing

[surreal spruce]

-colab

[azure marsh]

📒 Google Colab Notebooks

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], by Eddy, Hina and Gdr Google Colab
• UVR5 NO UI, by Eddy Google Colab
• UVR5 UI, by Eddy Google Colab
• Modified W-Okada's Voice Changer, Google Colab
• FaceFusion UI, by Nick088 Google Colab
• FaceFusion NO UI, by Nick088 Google Colab
• EasyGUI, by Rejekts Google Colab

ℹ️ Note

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

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

[lunar quartz]

how can i download rvc? any link

[low shard]

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

[lunar quartz]

radeon rx580 and i wanna do covers

[low shard]

Your AMD GPU is good enough to do inference (use models) locally (on ur pc), you won't be able to train (make models) but use them

You can:

• Locally (runs on your pc so the speed depends on that, you will have to set it up with the guides):
  • Applio (AMD Windows) : A fork of RVC with some extra features like Applio TTS, kinda faster and simpler but same quality tho
  • Mainline (AMD Linux/Windows) : 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.gg: Partnered with AI Hub, lets u do them easily but u may be in a queue
  • Applio Colab: max 4 hours, not granted, of GPU

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

[low shard]

RVC is not Wokada
elaborate in #🔍│help-w-okada

[fathom reef]

What is the most high quality pretrain type?

[foggy cedar]

What is vocoder and checkpointing?

I just saw it.

[simple ore]

you saw it, but you did not read the notes? 🙂

[simple walrus]

I have a question: How can I change words in a song? Is there a free application or program that can do this? What methods are there?

and I think this suggested ai cover website is great, it does what it should and it is free https://www.weights.gg/de

[light pelican]

I'm at this step, but it’s not working. Can someone help me, please?
https://docs.ai-hub.wtf/rvc/resources/epochs-tensorboard/#tensorboard:~:text=%23-,In the left panel%3A,-Activate Ignore outliers

[simple ore]

are there actual logs?

[foggy cedar]

I only saw this in Applio no ui, so I didn't see a note about it.

[simple ore]

vocoder is a new generator (MRF HiFiGAN and RegineGAN), no pretrains for those yet

Checkpointing - save vram at cost of slower training, can use larger batch sizes

[low shard]

@craggy stratusthis is the right channel for ai covers

• 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.gg: Partnered with AI Hub, lets u do them easily but u may be in a queue
  • Applio Colab: max 4 hours, not granted, of GPU

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

You could also do it on phone cpu locally, but it will be harder and slow asf, not suggested

[craggy stratus]

the voice i want to achieve (AI cover of yukari i found on tt) and i just want to kinda figure out what settings to set, cause when i tried to do it, the voice was off

[lavish spruce]

enhypen

[latent kettle]

38 epochs dataset length about 40 minuts. batch size 8. D loss is going up. do i stop traning ?? and change the batch Size ??

[pure tangle]

how to download model

[latent kettle]

#✦│chat message

[simple ore]

you may be ~100 epochs short of the target

or 200

[latent kettle]

i set to 250

[simple ore]

keep going

[latent kettle]

so should i keep traning ??

thank you sir

[simple ore]

as long as you have a good clean set with variety of content it will be more than 38 epochs

[latent kettle]

sorry what ??

[simple ore]

i mean it takes longer than 38 epoch to exctact everything useful from a 40 min set

provided the dataset is of a good quality

[latent kettle]

okay. thank you again

it started increasing again

[simple ore]

3.62 to 3.63 is not the increase that you should worry about

3 to 36, yes

[latent kettle]

okay.. i see

[onyx crater]

Guys im an ekitten now

ill get them eboys for money

[latent kettle]

Are there any symbols of overfitting ??

D loss is going down but G loss seems little bit increasing @simple ore can you please help me ??

[low shard]

That's not RVC

RVC = Retrieval-based-Voice-Conversion

It's used only for inference on pre recorded audios and training modela

You have wokada instead

Did you download this from a YouTube video

You downloaded an old version then

Go in #🔍│help-w-okada I can help u

[latent kettle]

@low shard can you help me

if it is overfitting

[low shard]

Mmm maybe let it train a little more

It doesn't seem really increasing much

[latent kettle]

This picture is on 175 and now it's 200

[glacial pollen]

You won't see any true overtraining for a long while

especially given you're not using averaged loss

[glacial pollen]

yet you'll be able to recognize it once it happens. it's quite apparent when it happens
Differs too much from normal graphs, the behavior

[charred portal]

Hi, I was wondering. Does anyone know any tool to train AI models using samples that would support multilanguage? I want to use it for AI voice changer.

[low shard]

Does anyone know any tool to train AI models using samples that would support multilanguage?
RVC (Retrieval-based-Voice-Conversion) is the tool used to train the best Speech to Speech models
But you can't make a model support EVERY SINGLE LANGUAGE, else you would have to train the voice of that guy/girl of him speaking EVERY SINGLE LANGUAGE EXISTING IN THE WORLD

However, you can train an english model and technically use it for other languages too, in rvc context, the model is made of a pth (actual voice) and added index (the accent), you could use it in the voice changer lowering the index ratio so it doesn't have the accent it was trained on

I want to use it for AI voice changer.
Be sure you're using Wokada Deiteris Fork and not some YouTube tut one

[charred portal]

I only want Czech Language

Speaker is also Czech

[low shard]

Then you just need to train it

Is it your first time

[charred portal]

And would you recommend some software that I can run locally on my computer? I used to use Mangio RVC, but it didn't do much results

[low shard]

Mangio RVC is an outdated fork

what's ur pc gpu first?

[charred portal]

RTX 4060

[low shard]

how much vram?

and is it laptop?

[knotty moth]

should be the same 8 gb

[charred portal]

It is Laptop

Only 8GB

[low shard]

eh, usually laptop gpus aren't the best

could be doable

For Locally (runs on ur pc):

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

Train 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)
  • RVCDISCONNECTED (no ui)
• Kaggles (a bit harder to use and needs phone number but gives 30 hours weekly of better gpus):
  • Mainline (UI)
  • Applio by Vidal (UI)
  • Applio by Shirou (UI, no guide as of right now)
• Lightning.ai (Kinda hard, needs login, no issue with web uis or anything, but only free 15 credits monthly):
  • 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.gg which ofc uses RVC

Maybe you could try locally first

@simple ore did u ever hear someone train on a 4060 8gb laptop?

[simple ore]

You can, 8GB bs4 is doable

can be more with checkpointing turned on in the new build

[knotty moth]

either bs 8 fp16 or bs 4 fp32, or turn on checkpointing

[charred portal]

When I generated it before it generated a whopping 150 .pts files, is that correct or am I doing something wrong?

[brazen skiff]

Help, does anyone know what frequency I should train at if my dataset has a frequency of 44hz, should I train it at 40 or 48hz?

[knotty moth]

go 40k or inspect the spectrogram using spek or audacity

[glacial pollen]

48khz

those 4khz are actually important for sibilants and fidelity, regardless
40khz model would dumpen the clarity so, 48 is the way to go for. Luckily 48 works with 44.1khz audio fairly fine if you're careful with training and the data itself
( Tho ensure it's truly 44.1khz (( frequency spectrum itself )) )

[latent kettle]

I got model maker, now it's time for model master, can someone please guide me ?

[woven anvil]

Hi there RVC beginner here,
Does anyone have a good tutorial that explains how RVC/Gradio works?
Can't find a tutorial I understand on YouTube.

[latent kettle]

Do you want to learn principle and programming of Rvc ?

[woven anvil]

No I just want to create a voice model for myself.

But it's awfully complicated : S

[latent kettle]

Which GPU Do you have

[woven anvil]

Oh one sec I'll check.

[knotty moth]

it's even harder to "master" it

[latent kettle]

I got model maker, now it's time for model master, can someone please guide me ?

[woven anvil]

Nvidia Geforce GTX 1660 Super and then I also have a Quadro P4000.

[knotty moth]

I'm not a master yet

[latent kettle]

Do you want to be ?

[woven anvil]

I believe you 😅

[latent kettle]

And vram ?

[knotty moth]

you're not even litsa

[woven anvil]

The PC with the Quadro has 32 GB and this one has 16GB

[latent kettle]

VRAM, video memory, not RAM

[woven anvil]

Oh my bad.

[latent kettle]

What does it mean 🤔

[knotty moth]

at least 8 GB rtx card recommended

[woven anvil]

Introducing The GeForce GTX 1660 SUPER
Making it SUPER is the addition of 14 Gbps GDDR6 VRAM, which boosts peak memory bandwidth to 336 GB/s (a 75% improvement over the GeForce GTX 1660's 8 Gbps 192 GB/s GDDR5 VRAM).

Quadro should be 8GB

[knotty moth]

so yours is GTX 1660 super or which quadro?

[woven anvil]

I have two PC's, one has Quadro the other the super.

[woven anvil]

One I'm using right now has the super.

[latent kettle]

I think gtx 1660 super has 4GB vram

[knotty moth]

which quadro, I said? quadro turing/newer?

[latent kettle]

So p4000 will de a good option

[latent kettle]

P4000

[woven anvil]

Quadro P4000 that's all I know : S

Okay so I need to use the other Desktop to do this.

[latent kettle]

P4000 will be good for training I guess. Because it have 8GB of VRAM

[knotty moth]

it's not as optimized as RTX cards, but fp32 performance should be fine
in this case, you should go batch size 4 and fp32 (usually defaulted for non-RTX cards)

[woven anvil]

I have no idea what you mean by all that, are these the things I need to enable/select in Gradio?

[knotty moth]

then read the guide here https://docs.ai-hub.wtf/essentials/how-to-make-voice-models/

[latent kettle]

Can you please tell. Me what did you said?

[knotty moth]

https://tenor.com/view/cat-annoyed-eyeing-cat-meme-cat-funny-gif-14867130448325066479

[woven anvil]

Isn't there a YouTube that has a tutorial on it?
I don't understand the terms, for example how do I convert my voice sample into a pth file?

[latent kettle]

Ahh, tell me....

[knotty moth]

the existing videos are too old and misleading

[latent kettle]

Everything is mentioned in Guide

Read carefully

[knotty moth]

I don't own that pandora's box

[woven anvil]

I know, but I don't know what there telling me in the steps, there are lots of terms and file types I'm completely unfamiliar with.

[latent kettle]

I didn't understand. Anyways NVM

[latent kettle]

You can ask that terms here

[woven anvil]

How do I turn my audio sample into a .PTH file?

Says I have to drop it into the weight folder.

I only have wav type right now.

[knotty moth]

rename the extension to .pth
there's no choice but to read the guide carefully

[woven anvil]

Dear gods...

[latent kettle]

I'll assist you

So let's begin

1st you need a dataset, "dataset is the audio or samples of your character's voice " From me, recommended length is minimum 10 minutes maximum 30 minutes.

[woven anvil]

I have done that, I have almost 20 minutes of clean audio with lots of variation in pitch and emotions etc.

In WAV format.

[latent kettle]

To prepare a good dataset, you need to remove background music, reverb eco and noise by using UVR 5

Also cut the silence and pauses from dataset

[woven anvil]

Was like half way with training a model with Google Colab before it completely broke and became unusable 🤭

[woven anvil]

Done that.

[woven anvil]

This as well.

[latent kettle]

Using colab is not recommended. Instead you can use kaggle

[woven anvil]

Can't I do it with gradio?
Was kinda happy I got it working 🥲

Colab is completely unusable so even if I wanted too I couldn't get it to work again.

[latent kettle]

Kaggle is a cloud computing platform just like colab. You can use RVC on Kaggle

[latent kettle]

Kaggle is stable and provides 30 hours of GPU runtime

[woven anvil]

Okay so I have to install that first then?
And if it's based on the cloud, then I could use this PC with the super instead?

Is Kaggle free to use?

[latent kettle]

Firstly tell me, do you want to train on locally, or on cloud ☁️?

[woven anvil]

Doesn't matter aslong as I get a voice model at the end 😬