✨│ai-help

[tame mica]

hmm

you could try training that dataset

and mixing both of them i suppose

[fast scarab]

Hi everyone, do these pre-trained models work well for French, or are they mainly optimized for English?

[livid plaza]

hi guys, i have a question about the rvc. the problem is that even by setting monitor to none I can still hear myself. I've also installed the virtual cable, etc. still i can hear myself

[knotty moth]

!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.

[livid plaza]

oh okay sorry

[hallow thistle]

Most of pre-trained models are trained on English audio, but I think they also do work with any audio language as well.

[hallow thistle]

Please specify more of the context next time. Don't ask with none of such context.

You can ask something like "what do I need to install and use RVC in my PC?" for example.

[rich flume]

how to remove the electrical noise of the rvc?

[low shard]

wrong channel, monitor i a wokada option, also do not follow yt tuts

RVC is not wokada

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

@livid plaza tell:

• your pc gpu
• what guide have you followed
• what you want to do
  in #🔍│help-w-okada

[low shard]

be sure the dataset you're training is high quality and use the tensorboard

[hollow idol]

For a very low diverse dataset (same prosody, different sentences), is it worth to expand by using ones-shot tts tools to increase dataset size and by picking the samples that most reflect the original voice?

My dataset is 30 minutes long now (it was 2h more or less), I've seen that by reducing the size it improved the results (to reach to a decent model, I would have to reach 300 epoches, now I can do this with 150 ~ 200)

But it seems the bottleneck is the dataset itself, changing pretrain does little effect for now

my idea was to use only 15min of the original dataset and 15 min from synthetic generated samples

[crude flame]

synthetic data isnt the best for rvc, it would be better to mess with the batch size or learning rate

[hollow idol]

I'm using 8 batch size, I'll attempt to use 4. For learning rate, I'll check more into

[crude flame]

you can try setting the generators learning rate to 5e-5

note that this will require you you to train the model a bit more

[hollow idol]

Good to know, I'm gonna use this one to try again. I'm trying to keep expectations for the model improvement around at max. 200 or 250 epoches (even for a 30 min dataset), if I don't see the graphs decreasing on tensorboard after 10 epoches I just keep the best epoch around from the last lowest point reached.

Thanks for the tips!!

[sinful ridge]

It's not like I want to, but my current time is about 4-5 seconds from activation to audio playback start, with around 4 different voices coming into a queue. I can spare a couple of seconds for tts of one sentence if I can find something to replace edge-tts.

[robust palm]

Need sm help💔

What ts mean how do I check the queue if it changed

[jovial leaf]

Are there any hobbyist voice actors here in the server?

[knotty cobalt]

I'm running RVC on my macbook m4 2024. I got it somewhat working, and the WebUI is popping up. But, whenever I try to do anything, my tasks load in the queue indefinitely without completion. Is this a common bug? Anyone have any info on how to fix it? (https://imgur.com/a/oySN48M <- screenshot of the problem because I can't send images in this channel for some reason)

[simple ore]

it is a known issue with mac

[knotty cobalt]

Thank you. v0 gave me the followign troubleshooting steps to try. Are there any well-known fixes that I'm missing, other than these?

1. Disable MPS Acceleration (most effective fix):

2. Force CPU Mode (if the above doesn't work):

3. Check if PyTorch is using MPS correctly:

4. Reinstall PyTorch (if needed):

5. Clear cache (if models might be corrupted):

6. Make the fix permanent by adding to your shell profile:

[knotty moth]

you can join into a whole for each speaker

[sinful ridge]

That's not an issue, I just don't know the fastest method to get the cloned voices that I want.

[muted musk]

I'm facing the issue issue, I have a Macbook (M3), I thought it was an issue with the voice files and tried a couple more and still getting the same problem.

[brittle wing]

broken?

[knotty moth]

then consider a zero shot TTS solution like fishspeech, etc. though dont expect it too expressive

[sinful ridge]

So the thing I was already going to try. okay.

[muted musk]

I'm using the commit mentioned here and now it is working fine without additional changes, give it a try.
https://github.com/IAHispano/Applio/issues/869#issuecomment-2725189180

[knotty cobalt]

Thank you so much!!

[knotty moth]

idk what kind of sus cloudfront link ur trying to open

[brittle wing]

huggingface

[knotty moth]

https://tenor.com/view/alicia-keys-no-one-gif-11087924

[jovial leaf]

😭

[hallow thistle]

Voice actor for what? I'm an artist who draws art, my guy.

[sterile hinge]

any site I can find a database isolated vocals of songs?

[unborn sigil]

is the linked applio on the wiki working for anyone?

[turbid rampart]

is there a guide for using the rvc ai cover maker colab i havent done this in a yearr

[grim bay]

What is the best extraction model for making AI cover right now??

I'm using UVR btw

[hallow thistle]

You could've say what is the best UVR5 model for extracting audio to use vocals audio with AI cover.

[grim bay]

Yeah that

[grim bay]

How can I use these models? It's not in the UVR I installed from the UVR official website

[knotty moth]

refer to this guide and use ctrl+F to search what you need https://docs.google.com/document/d/17fjNvJzj8ZGSer7c7OFe_CNfUKbAxEh_OBv94ZdRG5c/edit?usp=sharing

[grim bay]

Btw, should I use this instead of the app one?
https://github.com/Eddycrack864/UVR5-UI

[knotty moth]

different consoles/PC for the same games you want to play

[grim bay]

Yeah so now I'm confused... Could you give me a simple tutorial to just extract AI voices?

[knotty moth]

I can't guide you more as that would be completing your task. I have shown you the guide like above so you should try suggestions within it. This ensures you understand about the vocal extraction, what separation models do, and how well they could perform for your cases.

[hallow thistle]

!howtoask

[patent trellis]

How To Troubleshoot

__**GIVE CONTEXT.**__ 📝

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

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

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

__**BE PRODUCTIVE.**__ 🤝

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

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

[hallow thistle]

If I give you my own very simple tutorial, you'd think it's too complicated.

[pastel oak]

Extracting ai voices doesnt make sense in this context, do u mean Extracting voice from an instrumental?
If uvr is too complicated for you, go to the page mvsep.com , make a free account to skip 95% of the queues, use the latest bs roformer model it shows. Upload your audio and wait

[harsh ravine]

hey

[grim bay]

Yeah. I was bad at describing anything

[harsh ravine]

how do i create my own ai singing voice???

[hallow thistle]

What is your PC GPU? Applio the RVC can run on PC locally.

Asking "how to do AI cover" just that is one of the bad questions I've ever seen. The "AI singing voice", you either mean by AI cover.

[grizzled rapids]

Hey i'm kinda new to this training thing. I got a quick question. I downloaded a .pth file from huggingface, and i wanna use it on W-Okada for realtime voice conversion. It sounds quite cool, but i want to do some fine tuning because emotions like laughing, anger, sadness sound really weird with this voice model. I downloaded RVC WebGui, and i tried to load the model in there but it says "error" because the model apparently was trained on different sizes etc., what can i do to fix this or how can i do finetuning with this model? I already recorded .wav files to load etc., but its just not working. If anyone can help me i'll donate them 20 dollars.

Just DM me if you can help, i will donate you even 50 dollar if you can successfully help me.

🙂

[knotty moth]

• do you want to use rvc or realtime voice changer? the non-realtime one has better quality for pre-recorded audios
• most models struggle on those kind of emotions, but better quality dataset and pretrain might help reducing artifacting on that case
• you should read the model description carefully, it might have been trained using refinegan which is different from the default hifigan that most RVC applications and voice changer support.

[tame mica]

bruh

lowkey

yr jst better off using the actual website atp

[pastel oak]

You cant change a model you downloaded

make your own

Genuiene laughing is impossible at this current stage, emotions need to be overly expressed

[robust palm]

I’m using the app it’s way better

[tight linden]

Hello. I'm trying to make a good ai voice for elsa with a bit of emotion. I am on a mac though. And idk where to really go from.

[fleet night]

Heres the docs for training models or making covers with them https://docs.aihub.gg/rvc/local/mainline/

[tight linden]

I am on a mac though? does this work with that?

[fleet night]

Oh right,

[tame mica]

macs can only train on cloud

[fleet night]

i should know more

[tight linden]

which would be best for what I want to do.

[tight linden]

so how do I get to the cloud is it rvc?

[tame mica]

cloud's like training on another machine from a service

i could probably explain better but eh

it can train other ai's not jst rvc

-kaggle

[karmic olive]

📘 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

[tame mica]

kaggle's the best option since it has 30 hours free 2x gpu

but you need to use a phone number

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

[tight linden]

ok ty. I'm trying to do an elsa ai voice with some emotion to be firm

would that work well?

I want it to say a specific word

[knotty moth]

SCRFilms made some good elsa voice model (I bet you mean the disney frozen one)

[tight linden]

yes I meant that one. I can't train it though.

I need it to say a certain word that i'm going for

and the text to speech I don't think would sound well.

[knotty moth]

macs aren't viable for training

[knotty moth]

may depend on the accent compatibility

[tight linden]

Yea.

I can't seem to download it either.

Download
Download your Creation and additional outputs

Output

it says this and I click done but it doesn't download

could I download it to my mac or that wouldn't work?

just realized theres a queue....

is there any way to not do that

[tame mica]

what

it shouldnt have any queue

what are you using

[tight linden]

a ai voice model?

text to speech?

bc I want it to say something

I've been in the queue for 16 minutes

[tame mica]

???

doesnt answer my question

what service

[tight linden]

https://www.weights.com/models/clsyvi3lm00oj47ck6lss9r1h

weights.gg

I waited 21 minutes and it sounds awful omg. whyyyyyyy

[fleet night]

i figured

[knotty moth]

how awful is it? show few seconds of the inferred sample

[tight linden]

I think I deleted it.

[fleet night]

well anyways i think i made a good enough model

hope it gets accepted

[vale osprey]

-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

[dusk sphinx]

-kaggle

[karmic olive]

📘 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

[river trout]

https://imgdb.in/mwDY

Can anyone tell me which graph is better?

[simple ore]

fm going up is not good

[fathom geyser]

Hi. I'm having this error when using RVC Mainline Colab. I did everything to the letter and I still get an error. What did I do wrong? Can anyone help me fix this please?

[simple ore]

use other colabs from #📰│dev-updates

[plucky socket]

i can't get applio voice blender to work at all, is it possible to instead blend a voice by training data from two different singers? has anybody tried that?

[analog obsidian]

to merge models it is necessary that both have been trained on the same sample rate

[harsh ravine]

hey

how do i create my own ai singing voice?

[odd shale]

So, you want to make a singing model of your own voice or make a model of any singer?

[harsh ravine]

model of my own voice

[odd shale]

In that case, if you wanna make a singing model of your own voice, record at least 20, 30 or 40 mins of yourself singing (with at least a decent mic)

[harsh ravine]

i got the recordings bro

[odd shale]

Did you make sure these recordings are clean and got decent/good quality?

[harsh ravine]

yup i recorded it on a studio mic

[odd shale]

Nice, in that case go ahead and read this guide.

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

It will teach you how to train models.

[harsh ravine]

is it hard to do

[odd shale]

Nope.

It's not hard.

[harsh ravine]

any easier ones bro

[odd shale]

Easier ones what?

Guides?

[harsh ravine]

yeah

or like site thingy

[odd shale]

Nope, that's the only guide site we got. (afaik)

Sorry.

[harsh ravine]

damn man thank you though

[odd shale]

You're welcome bud, but i would suggest you to try and read the guide whenever you want.

[brittle wing]

anyone can help me

my colab is disconnecting and show this error

[brittle wing]

Runtime disconnected
Your runtime has been disconnected due to code execution that is not allowed at the no-cost level. Colab supports millions of users and prioritizes interactive programming sessions by prohibiting certain types of usage, as described in the FAQ. If you believe this message was sent in error, please submit a dispute. Please include relevant information about the context of your usage.
Your compute unit balance is 0. Buy more

To connect to a new runtime, click the connect button below.

[torn seal]

so i see one of my favorite model makers created a 2pac model. but he says the index is not required. but i think it is. doesnt that hold all the persona of the voice or accent

?

[idle crypt]

Hi ! I am having trouble with Google Collab No UI at the first step of trainng:

this is the error:

Starting preprocess with 2 processes...
0% 0/1 [00:02<?, ?it/s]
Traceback (most recent call last):
File "/content/Applio/rvc/train/preprocess/preprocess.py", line 269, in <module>
preprocess_training_set(
File "/content/Applio/rvc/train/preprocess/preprocess.py", line 241, in preprocess_training_set
audio_length.append(future.result())
^^^^^^^^^^^^^^^
File "/usr/lib/python3.11/concurrent/futures/_base.py", line 449, in result
return self.__get_result()
^^^^^^^^^^^^^^^^^^^
File "/usr/lib/python3.11/concurrent/futures/_base.py", line 401, in __get_result
raise self._exception
concurrent.futures.process.BrokenProcessPool: A process in the process pool was terminated abruptly while the future was running or pending.

anyone knows what is it?

[odd shale]

Yep, you can use the index.

[torn seal]

i always use the index files that come with the models, i think he is confused tho. i think he thinks you don't need it. but i think it is needed to make it sound more accurate

[torn seal]

sounds like its your dataset

[idle crypt]

Should i re upload it? I have a 24 min truncated audio recorded in a studio

[torn seal]

you don't need that much audio but you can

did you cut them up?

[crude flame]

The index contains the voice's accent and speech manor

It is very optional, and yes it may make the model sound more accurate

[torn seal]

also did you use clean vocals without doubling

[idle crypt]

I did the truncate process on audacity, I've done several models this way I dont know why I'm having this problem

[crude flame]

but if you infer the model on something that talks like the person that can also make it sound more accurate

[crude flame]

did you run feature extraction?

[torn seal]

thats what i thought. man if he only knew. the qaulity is there with his models but it dont sound like 2pac, it sounds like his accent with 2pac's voice if that makes sense.

[idle crypt]

bus feature extraction is the second step... I'm on the "pre-process" data fase

[torn seal]

i would say check your dataset files then reupload them

[crude flame]

is your dataset in the right folder?

[torn seal]

might be how you routed the folder

did you try to use a zip?

if you zip them sometimes it don't work.

[idle crypt]

Yes right folder and everyhing. Thw .wav file in 44.1kHz 16bit is inside the folder

[crude flame]

the file has to be a .wav or .flac it cant be in a zip

[idle crypt]

I'll try to zip it

[crude flame]

no no no dont zip it

[torn seal]

yea try that

[idle crypt]

oooh so I dont?

[crude flame]

dont

[torn seal]

why?

zip works too

[crude flame]

nope

[torn seal]

yea maybe your right.

i would start over again

reupload again

might be the notebook is down

[idle crypt]

ok I'll try that

Be right back with updates lol

[torn seal]

haha

also it will stop if you don't baby sit it

for a dataset like that i would go 1000 epoch

[crude flame]

this is what happened to me when i try to preprocess a zip

[torn seal]

aww i see

some note books are better than others. some can do it some cant

[knotty moth]

no, unzip it

[torn seal]

yea unzip it

i kinda feel like the old rvc did a better job on my models

[idle crypt]

I'm starting all over again

I hope it works?

[crude flame]

wdym?

its only had improvements

[torn seal]

well on the pre trained thingy mobober i forgot what its called but there was an option that it used for deeper voices

[knotty moth]

it depends on the pretrain you're using, besides the dataset and cleaning method used

[torn seal]

yea it was an old pretrain that did a swell job

[crude flame]

look in https://discord.com/channels/1159260121998827560/1235952130855010365 you might find it there

[torn seal]

its been a min since i trained, but i remember i had good results with an older pretrain

[crude flame]

unless you are talking about the og one

[idle crypt]

should I install the new numpy 1.26.4 version before training?

[knotty moth]

the default pretrain is still good, but those made in 2024 may be abit problematic

[analog obsidian]

klm 4 hifigan final and original pretrain are good

[torn seal]

il try that

[idle crypt]

It's not working

[torn seal]

is it in the any notebooks ? like a disconnected one

it could be your dataset or the note book

[crude flame]

look here https://docs.aihub.gg/rvc/cloud/applio-no-ui-colab/ to make sure you are doing everything correct

[knotty moth]

the colab notebooks may still have issues #📰│dev-updates message

[idle crypt]

oh so maybe is an error on the notebook itself?

[torn seal]

i was unaware of any issues i know it stopped working for a while then it came back on.

[idle crypt]

Yes all correct, is my 8th model lol... never had any issues until right now

[torn seal]

yea sounds like the notebook is either done for or is down for a min

what notebook are you using

[idle crypt]

Applio_NoUi from the documentation of applio.org

[torn seal]

hmm. you might want to look into finding another notebook

thats what i would do

[idle crypt]

How? I'm kinda new on the whole notebook/Google Colab/AI stuff

Or should I try on another cloud based platform?

[torn seal]

you can just type in the search in the discord of this channel

or ask around some folks have some

you can try kaggle but i have note used it yet

i know google plays games with people training

they dont like it.

[idle crypt]

maybe using another google account?

lol

[torn seal]

haha

you can try

im about to make some chicken sandwhiches mayne

[idle crypt]

I'll see if that works lol

[torn seal]

yea i requested you to be my pal

so let me know what happens

i been so lazy at training and getting things going with my music. i made a few 2pac songs last year. but dammit it takes up alot of time after you have not done it in a while to relearn everything

[idle crypt]

yes! I'm new to this and it's a lot of information

[torn seal]

yea it takes a min to get it all figured out. but now a dayz its much easier because everything is one click. when it first came out i didn't have that luxery.

[idle crypt]

true that man

[fathom geyser]

I'm getting this error when trying to load UVR5 UI. Can anyone help me please?

[idle crypt]

I think the notebook is having issuesbecause I can't infer...

[fathom geyser]

But how can I fix this?

[idle crypt]

No idea man, I'm having usues as well and I have no idea how to fix them

[idle crypt]

Not working... I'll try tomorrow

[fathom geyser]

Don’t worry about it. I’ll try with Google Colab

Thank you for your reply anyway

[viscid moss]

how tf u install UVR5 UI?

[viscid moss]

Ok, so i've no clue what did u do, but maybe u can try the precompiled version

#📰│dev-updates message

[neon berry]

can someone generate some voice msgs for me? rlly confused on how this works

[knotty moth]

there are many TTS options, including kokoro TTS
https://huggingface.co/spaces/hexgrad/Kokoro-TTS

and then you can use some RVC model to infer the TTS output

[neon berry]

im confused on this part

[glacial pollen]

simplifying it..

inferencing = using a given input / audio / sample / voice ( however you interpret it ) on a model ( in this case rvc )

[neon berry]

and how do i do that😭

[hallow thistle]

If you still have no idea what to do, then I don't know what to tell you either.

!howtoask

[patent trellis]

How To Troubleshoot

__**GIVE CONTEXT.**__ 📝

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

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

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

__**BE PRODUCTIVE.**__ 🤝

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

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

[hallow thistle]

You can't expect anyone to do everything for you. You have hands, you can do thing by yourself.

[knotty moth]

you should first understand about RVC here: https://docs.aihub.gg/essentials/whats-rvc/

[hallow thistle]

[odd ledge]

Hi, I tried the path here in order and it epochs am I on the right track, just now when I did one click traning, it gave an error and I tried again and it started to epoch

https://www.upload.ee/image/17868327/resim_2025-03-18_123851515.png

[knotty moth]

try Applio, it has several improvements and features in training

[odd ledge]

I don't know how to use it

[knotty moth]

I should have shown you the guide https://docs.aihub.gg/rvc/local/applio/

[odd ledge]

You don't know anything about this?

[knotty moth]

you mean the issue you're having?

the one click feature is not recommended for the possible issues

that's why it's not included in rvc forks including applio

[hallow thistle]

Who doesn't know about that?

You asked like if a mod/helper doesn't know what Applio even is and how to use it.

[odd ledge]

I don't know because

[fathom geyser]

How can I add voice models in UVR5 UI on Colab??

[thorn fossil]

does anyone have a way of making a voice into a text to speech?

[low shard]

There are different Text To Speech (TTS) AIs:

GPT So Vits: RVC isn't as good as GPT So Vits for tts, but gpt so vits (few shot tts, which means needs just a lil training for models) can't use rvc models (and viceversa), and its only limited to: english, chinese & japanese, if you wanna check gpt so vits instead, read https://docs.ai-hub.wtf/tts/gpt-sovits/

Freemium 11labs: Easy way to do TTS is https://elevenlabs.io/, you can't use RVC model on this but its a mostly premium easy way for good quality TTS

FishSpeech: FishSpeech is a 0 shot (no explicit training needed) TTS, if you got a good pc you can use it locally else use their site

You can check TTS in our tts index

With RVC Models:

RVC is natively for Speech To Speech, but forks such as ilaria rvc mainline & applio have built in tts (using Microsoft Edge TTS to make a generated tts audio, which i suggest you to choose a tts model that is the same gender and language of the rvc model you wanna use, and then convert it with rvc)

If you wanna do tts locally with RVC Voice Models (if you got a good pc):

• You can get Applio in our docs
• While Ilaria RVC Mainline here (no guide as of right now)

If you don't got a good pc you can do tts with RVC Voice Models on cloud:

• Ilaria RVC Zero (Running on A100 GPU, free fasted rvc on cloud) and the guide

• Use Applio UI Colab (with google colab T4 free daily limit gpu)

• You could try another tts from our tts index and use the output as an input in rvc

[olive hill]

Anyway to bypass this?

[hallow thistle]

Unfortunately, no. You'll have to pay for ngrok or wait for another month to continue using it.

[olive hill]

anything alternative?

[analog obsidian]

delete your account and make a new one with the same email

sounds silly but it used to work for me before lol

then ofc use the auth code of the new acc

[olive hill]

hahah interesting

[olive hill]

Absolutely worked

[analog obsidian]

🔥

[viscid moss]

bro what

[latent kettle]

404 solution not found

[fathom geyser]

Sorry for the noob question, but idk where is the weights directory of the UVR5 UI

And how to download models from Hugging Face

Honestly, I'm not very familiar with this specific Colab

That's why I'm a bit confused and I need help

[viscid moss]

Models are automatically downloaded

[hallow thistle]

Unless a UVR5 Colab notebook doesn't provide any model to use or you want more other models to use with.

[fathom geyser]

But that means I can't load my own models?

Even the ones that I download from Hugging Face?

[hallow thistle]

I don't know. You should be able to use any other UVR5 model in UVR5 Colab notebook, although you'll have to do some code a bit there.

[fathom geyser]

Hmm... I'm actually new to this. So Idk how to write some code

[viscid moss]

UVR5 UI isn't for voice conversion, it's for getting instrumentals, vocals, reverb, and more from songs/audio files

[fathom geyser]

And what tool do you recommend to me?

[hallow thistle]

If you mistaken UVR5 for an AI voice changer like RVC, it's not. UVR5 is an AI audio stem separator.

[fathom geyser]

Ok

I didn't know. Sorry

[hallow thistle]

It's fine. If you actually mean by a program that can do voice changing, Applio the RVC is what I'd recommend.

[fathom geyser]

Ok, I'll use it. Thank you

[hallow thistle]

You can learn more about Applio there. https://docs.applio.org/applio

[odd shale]

Actually another alternative is creating various ngrok acccounts lol

A mí me sigue funcionando hasta el día de hoy este truco

[brittle wing]

Yo, I'm using Codename Fork, but the problem is that the interface in TensorBoard doesn't look the same. Which one should I follow to get the same results as g/total?

[idle crypt]

Can I install the Applio Repository on Hugginface spaces and run it there?

[vast forge]

i havent made a voice model in a year and i forgot how to can someone explain

[vale osprey]

-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

[idle crypt]

is anyone having trouble with applio on colab while training or infering?

[flint solar]

it should b working

[brittle wing]

Yo, I have a question: is it possible to separate voices from the same video? Thanks in advance for your answer.

[hallow thistle]

What does it mean to separate voices from the same video? The file must be in an audio format (mp3, wav) before going separation process.

[brittle wing]

How exactly can I separate voices from a video? Which software or tool do you recommend?

[knotty moth]

follow this guide https://docs.aihub.gg/rvc/resources/dataset-isolation/#vocal-isolation--cleaning

[grim bay]

Between Mangio and Applio, which one should I use for Local?

[pastel oak]

Applio and its not even close

[hallow thistle]

Mangio as an RVC program has no update as of now. Applio is better.

[mint prism]

uhhhh i keep getting this error on applio

An error occurred extracting the index: need at least one array to concatenate
If you are running this code in a virtual environment, make sure you have enough GPU available to generate the Index file.

?

[heavy arrow]

it detected overtraining below 100 epochs. is this normal or is it just because i had a really high quality dataset

[knotty moth]

not recommended to use "overtraining detector", just observe tensorboard graph and do testing on several checkpoints

[heavy arrow]

aii thanks

[simple ore]

you did not preprocess audio (no slices of acceptable length)

[mint prism]

it was too short

sorry

[simple ore]

too long

use built-in audio cutting

[mint prism]

what

no

original sample:

working sample:

it was just too short

smh

[simple ore]

are you trying to train on 0.5s of audio?

🤦‍♂️

[mint prism]

y e s

but now it works

uwu

i dont know what hte model will be tbh

[ornate dagger]

hello how to text to speech with my model pls

[plain fulcrum]

💔

[mint prism]

trust me

my model be peak

just wait

[civic meteor]

besides https://github.com/facebookresearch/svoice is there any other easier to use speaker separation

[mint prism]

how do I recontinue training again?

[vagrant island]

Hello
can someone help me with my RVC model. I trained it through RVC2025, I got the .pth files. and want to use it as a realtime voice changer voice.ai but it requires some zip file. could someone help me with this? would be thankfull for some tips

[knotty moth]

omfg still using that garbage voice.ai?

[vagrant island]

trying rn. any better options?

[knotty moth]

for better voice changer, go to #🔍│help-w-okada to get one in the pinned guide

[vagrant island]

thx

[hallow thistle]

Don't waste your time trying the shitass Voice.ai. Use W-Okada instead, and go to #🔍│help-w-okada for more information.

[vagrant island]

thanks

[mint prism]

So like what's a pretrained model

And what does the language thing do

Japanese korean

[tame mica]

ggrks

[mint prism]

Wise words

[gusty valley]

this only for rvc? https://discord.com/channels/1159260121998827560/1198095259293450341

[tame mica]

???

Yes

[mint prism]

what does this mean 🥺

I ALREADY DID

5 TIMES

[grim bay]

Anyone knows why I can't download?
I'm using Applio Local

[mint prism]

I keep getting these buzzy noises at this specific range, yall know whats up?

I put it as a random noise file

just to hear the uhh frequency

[simple ore]

what do you expect from training on 0.5s sample?

it can only reproduce this exact sample and nothing else

[mint prism]

yeah but why the voise

noise

[simple ore]

maybe use a different browser

[mint prism]

the sample doesnt have that noise

[simple ore]

not edge

[mint prism]

I mean I expect it to HAVE A BAD VOICE

but not an unreasonable magic sound

[analog obsidian]

well thats not how rvc works

needs data to be able to clone stuff

[mint prism]

so do I just

loop it more

is what ur saying

[analog obsidian]

different data

[mint prism]

dont make me use FL studio's piano roll to give RVC every pitch that I will accept

because I will do it

and it will sound bad

[analog obsidian]

like actual real other data, not the same sample pitched down, etc

[mint prism]

aw

welp

f

but yeah till where does the noise come from

do other models have this noise?

[analog obsidian]

when they lack generalization, yes

good trained models, not

[mint prism]

do you have like

any reason whatsoever

as to why

[crude flame]

.5 second dataset

[mint prism]

no like thats the ISSUE

[analog obsidian]

rvc tries to generate frequencies
ur dataset don't have those frequencies
rvc ded

[mint prism]

I meant like WHY does it just get random noise

[mint prism]

seee? this is the answer I wanted

thank you shark person

[analog obsidian]

🦈 🔥

[mint prism]

I will now

pitch it different

and make it work

because I AM STUBBORN

[analog obsidian]

[mint prism]

we will get

sans

ONE DAY

ONE DAY YOU WILL SING MY BOY

[azure patio]

guys ive lost track of rvcs but which is the easiest one to use to make covers with the models i have made?

[mint prism]

applio

best

[crude flame]

applio 🍏

[mint prism]

IF NOT GREEN

IT BEAN

[azure patio]

any link? 😭

[crude flame]

https://applio.org

[candid elbow]

Can anyone recommend some quality voice model with a deeper voice? Kinda like morgan freeman

[mint prism]

would this make a good dataset /j

[sage oasis]

can it not be used for like okada

cuz i wanna use it in okada

[fallow veldt]

guys is hina mod rvc the only one that works with youtube links and does the vocal separations by itself?

[grim bay]

From what I just tested, Hina RVC is not working atm so there's an alternative
https://colab.research.google.com/github/Eddycrack864/RVC-AI-Cover-Maker-UI/blob/main/assets/RVCAICoverMakerUI.ipynb

[tropic mesa]

He does anyone know if i can transfer Mangio RVC onto an external hard drive after i already have it installed? It takes up so much space but i also don't want to break it.

[old badge]

Hiii

I need some help

my applio is not working, at all

[mint prism]

how to continue training a model you have already started training?

[vast forge]

whats the link for the rvc application that lets you make the ai cover

[serene hull]

hello everyone, can someone reccomend me model which I can use in UVR

I don't really understand which one I need to choose, need something that don't need much time, remove simple music, and have a normal quality

[viscid moss]

https://docs.aihub.gg/rvc/resources/dataset-isolation/#the-best-models-for-uvr-are

[viral mason]

what did I do wrong man :(

this is my applio space

[gentle hollow]

Yo is there any way to do models on phone except for weight

[viscid moss]

Colab

[gentle hollow]

Can sent link

[viscid moss]

-colab

[karmic olive]

Suggestions for @gentle hollow

📒 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

[viscid moss]

Colabs like Applio, RVC Disconnected and RVC Mainline can help u with that

Here's our guides btw

-rvc

[karmic olive]

📚 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

[gentle hollow]

Thx

[willow nebula]

Hello, how do I train a voice? sorry just new here

[knotty moth]

please refer to the guide https://docs.aihub.gg/rvc/resources/training/

[brittle wing]

-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

[turbid trail]

-realtime

[karmic olive]

Interaction has expired, use the command again for a new interaction.

☁️ (Online) W-Okada - Colab Version

• How to use Original W-Okada's Voice Changer Google Colab
• Modified W-Okada's Voice Changer Google Colab
• Original W-Okada's Voice Changer Kaggle
• Modified W-Okada's Voice Changer Kaggle

⚠️ Notes

• Colab free plan GPUs tipically works for about 4 hours each day
• Kaggle restricts GPU usage to 30 hours per week
• These options may not work on mobile devices due to the lack of a Voice Audio Cable (VAC)

[mint prism]

Guys is rtx 5090 good enough for training

With applip

[low shard]

#✦│chat message please don't waste helpers time

[mint prism]

Sorry

[low shard]

It's fine, just don't do it again

[mint prism]

Yes sir

[hallow thistle]

If you're really really serious about that, let me answer to you. There's no stable version of Applio that can work with NVIDIA GeForce RTX 50 series GPU as of now. The version that actually works with this GPU is in development and experimental.

[mint prism]

So I waste 1600000 dollar for nothing

Wow..thank you

[hallow thistle]

And make sure you don't play around something you have no idea what it is for.

[simple ore]

Well, for 5000 series just need to install nightly torch and all

nothing experimental applio-wise

[low shard]

He wasn't serious

[hallow thistle]

That's what I was saying.

[simple ore]

https://github.com/IAHispano/Applio/issues/969#issuecomment-2738392518

[mint prism]

That was about rx 9000, not rtx 5000

[hallow thistle]

I did explain things just to confuse her head.

[grim bay]

Does anyone changing these settings? And what are these really do?

[simple ore]

search feature 0 to 1, pitch

the rest - no, dont bother

[hallow thistle]

Pitch changes the audio pitch. That's all I know.

[simple ore]

volume envelope is broken, I think

[grim bay]

So what does search feature do?

[simple ore]

it controls the index influence

it is blending features of source audio and the model

prononciation, accent

[grim bay]

So in general, it makes the output smoother?

[hallow thistle]

Voice model accent stores in index file, which is alongside a pth file.

[simple ore]

lets say you use a french voice model and english audio

if you use index 1.0 the inferred audio would sound like a french person speaking english

[mint prism]

What if sans was speaking ERenise

Wouldn't it be better for that type of model to have 0 index

[simple ore]

with index 0.0 the output will be english with just the new voice

well, still may have minor quirks, but not as pronounced

[knotty moth]

pitch is mostly the one you may need to tune, depending on if male to female or vice versa

[mint prism]

Which for singing would be the most preferable approach for exact replication?

[simple ore]

experiment, my dude

it only takes a few seconds to hear the difference

[mint prism]

Oh no I know this

I'm just asking for anyone reading

the 3 people here lurking I see you

[hallow thistle]

[grim bay]

Also may I ask, I have the really old model from 2023, does it still works on those new tools like Applio or those new pitch extraction algorithm?

[simple ore]

you may need to use the same pitch extraction for inference.. I've explained before why it may not be a good idea

also depends on whether it was trained with pitch guidance or not

[grim bay]

How can I know if the model was trained with pitch guidance or not?

[simple ore]

generally I would not advice using rmvpe pitch extraction for a model trained with crepe

[grim bay]

Should I use hybrid[rmvpe+fcpe] pitch extraction for a model trained with rmvpe?

Yeah I was outdated

[simple ore]

just rmvpe

[hallow thistle]

A voice model trained with rmvpe should always be rmvpe, no need to use a hybrid pitch extracting algorithm for this one.

[simple ore]

run python, import torch, load the model, print the model

at the end there is model information

[grim bay]

And what are these things? Especially the split audio thing, should I turn on this?

[simple ore]

it is for inferring audio >10min

depends on VRAM size

on 4090 can probably do an hour without splitting

[grim bay]

Ohhh
So there's no need to turn on any of these if I do a 3 minutes song?

[knotty moth]

for long audios, mostly better do manual split or use the split audio feature

before that, make sure to do denoise & noise gate first

[analog obsidian]

btw @simple ore have you tried comparing crepe vs crepe with custom hop length?
i feel the default crepe is better despite being the same thing

crepe with custom hop sounds out of tone sometimes

(64)

[simple ore]

you can run that f0 view to see how it does

[knotty moth]

#🔥│model-maker-chat message

there's 128 vs 64

[analog obsidian]

thats the crepe with custom hop

[simple ore]

why those charts do not look like he spectrogram?

[analog obsidian]

there's other that u cannot change the hop length

so 160 hop iirc

and i feel that crepe (without custom hop) is slightly better

[simple ore]

anyway, no, I did not bother with crepe at all

thing is, i'm not sure how rvc handles custom hop

rmvpe is hop is 160

so 1 second is 16000, 1 hop is 1/100 of a second

so for 1 second you get 100 values

feature extraction is 320, so for 1 second you get 50 values

now, if you use hop 128 or 64, you get 125 or 250 values

i cant imagine it would align properly

[analog obsidian]

i feel crepe with custom hop should be replaced with the original crepe

but i gotta ask claude for that

💔

[simple ore]

well, it has to align features and pitch somehow

so either features are getting padded or pitch getting truncated

none of it is good

[midnight cloak]

who has a bluudud voice model

ive been looking for a long time

[knotty moth]

commission in #1191429836321849435 for more chance of getting response

[brittle wing]

Hi, what is this cell for and how do I use it?

The phase Fixer cell in your tweaked colab

What kind of files do I use it on

[clever dove]

Why does the step 2 doesn't work

[glacial pollen]

This actually isn't the case because

the model is essentially the exact same crepe model according to mangio

it is the " crepe full "

It just has an override of hop

( Unless the rvc uses some other crepe model? in that case.. yea but I don't think so

[neat peak]

Guys does anyone knows which is better for pitch extraction RMVPE or crepe for Arabic voice?

It's urgent if you know the answer please reply:)

[analog obsidian]

can you try in your old fork of mainline and compare inference results? i swear crepe (not mangio) results were better than mangio

[glacial pollen]

Actually, my fork always used mangio crepe

because the foundation I kept on using ( prior to moving to applio

[neat peak]

I meant in Training process

[glacial pollen]

was based on mangio build

that I merged with mainline

[analog obsidian]

yes but it had the option to inference crepe (without custom hop)

[glacial pollen]

( in case you talk about og forks of mine

[glacial pollen]

then that was the og crepe

[analog obsidian]

and i feel my inference results were better with og crepe

[glacial pollen]

try on mangio crepe ( using 160 hop

should more or less align with og

see if that changes anything

[analog obsidian]

oki

[merry spindle]

heyo, i havent used ai since it first became a thing (2023) and was wanting to use it again, there was this voice i wanted to use (microsoft sam) but it doesnt have an index file, instead it has a ckpt file, ilaria RVC doesn't accept ckpt files so is there any way i could fix this?

dunno if this is just a stupid moment from me but arent pythons and index files the way to use ai voices

[simple ore]

ckpt is probably for so-vits

[merry spindle]

huh?

[simple ore]

rvc-based models do not use it

[merry spindle]

oh, could you tell me how to use them?

[simple ore]

dig out the so-vits-svc from the grave

reanimate by cursing and spitting

and figuring the set of ancient libraries that works

[merry spindle]

can you elaborate further

[simple ore]

all I'm going to say, it is dead

and buried

[merry spindle]

alright

[simple ore]

we have google colab breaking monthly and you expect 2-year old softwave to work somehow

[low shard]

Monthly? More like weekly 😭

[bleak nymph]

guys pls help quickly

im training with rvc but i have no idea what im doing

it's at the 11th epoch

how do i stop and save this epoch, and then continue training it from this epoch later on?

like say im done training for today and i wanna continue tomorrow

how can i make it so it'll continue from epoch 11

[zenith wren]

https://cdn.discordapp.com/attachments/1352370785251102793/1352385635176022088/image.png?ex=67ddd2c5&is=67dc8145&hm=9f407df2ab783db3188cac2c2d3505883580c0b79687259e43712ad784eed454&

how to fix that

anyone help please

[pastel oak]

increase chunk to 140

and ask in #🔍│help-w-okada next time

[zenith wren]

ok thanks

[bleak nymph]

my model is messed up 😭

[dusty bramble]

yo?💀

[bleak nymph]

yeah 😭

[dusty bramble]

where other shark guy he prob knows what happened😭

[bleak nymph]

my real voice for comparison

[bleak nymph]

who

[dusty bramble]

this dude^

[bleak nymph]

o

[dusty bramble]

it does trust

[bleak nymph]

😭

better than the last 4 attempts id say 😭🙏

joking those weren't as bad

[dusty bramble]

did u change the pitch for the cover?

[bleak nymph]

yeah i made it -12 semitones

an octave down

to match it a lil better

ill try it without adjusting it then if you think that's the problem

[dusty bramble]

well i don't think it will save wtv happened there then😭

[bleak nymph]

true 💀

this is why i went for a longer dataset originally

also

[dusty bramble]

idk man, did u train it rvc 1 or 2?

[analog obsidian]

[bleak nymph]

to use the model i went to inferencing voice and chose "isling.pth" (the model name i chose) BUT there's an option to manually select it

[analog obsidian]

nah that happens because he didnt sliced his dataset properly

[bleak nymph]

shall i try manually selecting it

[bleak nymph]

wym slice it

[analog obsidian]

so rvc trained nothing

[dusty bramble]

oh😭

[bleak nymph]

did it not?? 😭

[analog obsidian]

nuh-uh

what are u using for training???

[bleak nymph]

RVC

[analog obsidian]

mainline, applio???

[bleak nymph]

v2

[bleak nymph]

idk

[analog obsidian]

ok first do me a favor and uninstall whatever u installed

[bleak nymph]

RVC1006NVIDIA

[analog obsidian]

and install this https://github.com/IAHispano/Applio

[bleak nymph]

that's the folder

[analog obsidian]

mainline

[bleak nymph]

oh ok

[bleak nymph]

right

shall i still uninstall it

[analog obsidian]

yes because mainline is outdated asf

2 years without any update

[bleak nymph]

oh 😭

[analog obsidian]

do this

download the applio zip

and uncompress it in a non onedrive folder

C:\Users/yourname

[bleak nymph]

download zip? can't i just go to releases or is that the same thing

[analog obsidian]

no

download the zi

[bleak nymph]

oh ok

[analog obsidian]

release sends you to the compiled version which is old

[bleak nymph]

ahhhh

right

ill just do to my desktop

[analog obsidian]

is your desktop in a onedrive folder?

[bleak nymph]

no

screw one drive

all my homies hate one drive

[analog obsidian]

ok just to be sure right click a folder and click properties

as long it doesnt says onedrive

ur fine

[bleak nymph]

it does not say one drive

[analog obsidian]

pog

[bleak nymph]

why can't it be in onedrive

[analog obsidian]

python shit

has a stroke or something

anyways

run-install.bat <--- run that

and wait

[bleak nymph]

btw i put my dataset in the root of the rvc thing

is that ok

not the root

[analog obsidian]

yes it doesnt matter

[bleak nymph]

i made a folder

ok

[analog obsidian]

i'll teach what to do later

is ez

[bleak nymph]

nice

also ty for being patient with me 😭

[analog obsidian]

its okay

the process is extremely simple

whats hard about rvc models is getting good source audio

training them takes a few clicks

and 1% brain power

[bleak nymph]

nicee

i kinda went in blind before

in rvc

[analog obsidian]

well, at least for training small models
for more complicated stuff like finding the appropriate hyperparameters... ehh that'll require your entire brain

but for us mortals is simple when we don't care about the complicated stuff

[bleak nymph]

i installed it and started training without knowing what anything means

ayy

it's installed

[analog obsidian]

yay

now

do u have audacity?

[bleak nymph]

yes

[analog obsidian]

do you have spek?

[bleak nymph]

audacity my beloved

[bleak nymph]

i have spectro

same thing basically

[analog obsidian]

show me the spectogram of your dataset

[bleak nymph]

oh 😭

it aint good

192 kbps mp3

[analog obsidian]

ive trained compressed asf audio from games and they sound fine, a bit robotic due to compression, but decent

[bleak nymph]

ah, nice

[analog obsidian]

okay its noisy asf

time to denoise it

[bleak nymph]

ooo

[analog obsidian]

https://colab.research.google.com/drive/1IC6Q1hLF55_tK6mhky0SWYKGVF9T5WsY?usp=drive_link

[bleak nymph]

i'm good at filtering actually

i use uvr and uvronline

mvsep too

[analog obsidian]

be sure to use denoise aggr by aufr

save as wav

[brittle wing]

Do you need help?

[analog obsidian]

no

[brittle wing]

Ok

[bleak nymph]

it's not really that noisy, it's mostly keyboard clicks and mic noises

[analog obsidian]

spectogram is telling me its that noisy

🦈

[bleak nymph]

lmao

[analog obsidian]

so time to remove that ugly noise

the colab is ez to use

[dusty bramble]

arguably if he wants to be accurate to his mic and his voice, wouldn't it better to keep the noisy to copy his mic quality too?💀

[analog obsidian]

true

but noise disturbs the learning process

takes more epochs than usual to get something good

if he doesn't care about a long training process i guess its fine

[bleak nymph]

lmao i don't really care but shorter is better ig

[dusty bramble]

no matter what is going to be faster than his 2 hour dataset attempt😭 🙏

[bleak nymph]

like as long as it dont take over a day

[bleak nymph]

😭😭

[analog obsidian]

30 minute noisy set took me 3 hours*

[bleak nymph]

woah

[dusty bramble]

big oof

[bleak nymph]

oh 30 mins?

[bleak nymph]

i cut mine down to 18 mins

[analog obsidian]

should be in around 2 hours

[bleak nymph]

i don't understand tesnorboard graphs

[bleak nymph]

will it still be accurate tho

[analog obsidian]

graph going down = pog, victory royale, great
graph going up = cringe, ass, eww

[bleak nymph]

oh ok

ty

[analog obsidian]

graph going up for over 15-30 minutes = dying

[bleak nymph]

ok colab is taking ages

imma denoise it on uvronline

[analog obsidian]

i mean rvc technically works with noisy audio

and the increased epochs arent that much

is like 40 more?

[bleak nymph]

ok

ill just do noisy audio

[analog obsidian]

okay great

[bleak nymph]

uvronline supports up 13 mins anyway

[analog obsidian]

now time to open audacity

[bleak nymph]

yup

now what

[analog obsidian]

open your audio and

[analog obsidian]

first step

[bleak nymph]

it's already mono

i exported it as mono because my microphone is mono

[analog obsidian]

second step

[bleak nymph]

plus way more storage space saving

[analog obsidian]

be sure to select the whole audio

while doing that

[brittle wing]

Do you need help?

[bleak nymph]

nope am good

ty

[bleak nymph]

ok done

silence is gone

12 mins 30 now

is that still ok?

[analog obsidian]

now about this next step im not sure how good audacity's resample is compared to soxr(applio's) uhh
ok for this time we will use applio's resampler

so

don't touch sample rate

but encoding set it to 32-bit float

and exports

[analog obsidian]

small but might work

[bleak nymph]

that seems a bit overkill

[analog obsidian]

not realistic results but decent

[bleak nymph]

oh

i can make it longer

real quick

[analog obsidian]

you've said your speech is monotone

for realism rvc needs diverse data

[bleak nymph]

yeah but when i listened back it wasn't that monotone

i was playing a game with my friend so i was quite expressive

[analog obsidian]

if you add more data be sure that is the same quality as your dataset

30 minutes is enough

[bleak nymph]

oh ok

[analog obsidian]

also speech models suck at singing

[bleak nymph]

i suck at singing

🔥

[analog obsidian]

so a realistic speech model will sound ass while singing

[dusty bramble]

i mean, they don't suck, but it doesn't sound like how u would while singing

[bleak nymph]

i tried training model on me singing but it sounded ass

it didn't sound like how i sing

it sounded like the original

and uh

yeah the original but worse

[analog obsidian]

singing in a speech model will cause the model to not sound like the person since the model lacks singing range

[bleak nymph]

i understand what u mean

but what if my real voice lacks singing range

ppl do tell me i sound like robot when i sing

[analog obsidian]

the model will be limited to your singing range

rvc doesnt do magic

[bleak nymph]

true

[bleak nymph]

actually when i trained with weights

it was far far far far wider than my real range

[analog obsidian]

well this varies by models
a veeery expressive speech set can do singing just fine

[analog obsidian]

yes at that point rvc was using the pretrain's knowledge and not your dataset

[bleak nymph]

no

i mean on weights.gg

[analog obsidian]

pretrain is a thing that is used when training models

a thing that has knowledge about the human voice

rvc uses that thing to train your dataset

aka finetuning

[bleak nymph]

anyway i truncated silence of whole 2 hr datset and then i cut it shorter to 30 minutes

and exported as 32 bit float 44.1khz

[analog obsidian]

did you used my settings right? thats important

[bleak nymph]

it was already (44.1khz)

[bleak nymph]

yes

[analog obsidian]

nice

now you can finally open applio

[bleak nymph]

ok

[analog obsidian]

save your dataset in a random folder or whatever, you know the deal

now run-applio.bat <- run this

[dusty bramble]

he is almost there 😭 🙏

[bleak nymph]

😭

i ran it

[analog obsidian]

ok wait

[bleak nymph]

um

my desktop wallpaper changed

[analog obsidian]

🦈

[bleak nymph]

it shows this what does this mean

[analog obsidian]

lmaoo

[bleak nymph]

💀

[dusty bramble]

looking like a d1 virus scam

[bleak nymph]

lmao it's installing

[dusty bramble]

💀

[analog obsidian]

oops forgot to remove the ransom from the code

[dusty bramble]

[bleak nymph]

ayyy

[dusty bramble]

his crypto wallet

[bleak nymph]

it opened in web browser

i dont use crypto

[analog obsidian]

great now relax and follow my instructions

they're ez

and apply for any model

[bleak nymph]

ok

i wasnt stressed

[bleak nymph]

oh ok

[analog obsidian]

training tab and be sure cpu cores is set to 1

(i don't trust applio's multicore usage )

[bleak nymph]

lmfao

[knotty moth]

[bleak nymph]

i wanna use my gpu to train

[analog obsidian]

yes

it'll do it

calm

[bleak nymph]

oh ok

[analog obsidian]

now here

[bleak nymph]

bro has a 4060

[analog obsidian]

EXTRMELY IMPORTANT
use these exact settings

[bleak nymph]

oke

[dusty bramble]

why is bro judging

[bleak nymph]

I'M NOT

[analog obsidian]

😭 im from a third world country man ffs

[bleak nymph]

I'M JUST SAYING IT'S REALLY COOL

i have a 3070 but i got it during the GPU crisis

so it was like quintiple the msrp

😭😭

anyway

[analog obsidian]

just in case you dont know how to copy the address of the dataset

right click the name of the folder

[simple ore]

cutting 10-hour long files you may want to keep it under 4

but it is fine otherwise

[analog obsidian]

yeah for me the speeds are basically the same

[dusty bramble]

i got my 4070 super this December, all my friends were saying I got cooked, become the 5070 got announced with "4090" performance
but it turns out a 4070 super out performs it in some instances

[analog obsidian]

after you preprocess the dataset tell me

don't rush things

🦈 🤙

[bleak nymph]

ok

[analog obsidian]

now after you did it, go to the logs folder

its on applio's folder

(obviously)

and go to your model's logs

then go to sliced_audios

[bleak nymph]

oke

[analog obsidian]

see if every slice is exactly 3s

[bleak nymph]

whats the difference

[analog obsidian]

16k is used for training the index

[bleak nymph]

oh ok

[analog obsidian]

and also for f0 estimation

[bleak nymph]

yup 3 seconds

[analog obsidian]

perfect

[analog obsidian]

exact settings

and extract features

rmvpe is fast, it'll take a couple of seconds

[bleak nymph]

oke

[analog obsidian]

after feature extraction use these exact settings
(saving every epoch is more accurate, you can delete the useless epochs after training)

every has to match

batch size is more complex than just "8" but no point to explain what batch size is for now

[bleak nymph]

okie

[analog obsidian]

first generate index

[bleak nymph]

it matches

okee

is generating

[analog obsidian]

and after the index is done you can finally click "start training" be sure that fresh training is enabled

[bleak nymph]

wait so quick question, if i wanna stop training and i wanna continue training from latest epoch saved the next day

[analog obsidian]

you can now use run-tensorboard. bat <--- very important, tells you when to stop the training as we don't want to train the full 500 epochs

[bleak nymph]

do i need to change any settings

[analog obsidian]

nope

[bleak nymph]

what abt the fresh thing

[analog obsidian]

just don't use fresh training

when resuming

[bleak nymph]

ok

it'll automatically start from latest?

[analog obsidian]

yeah

[bleak nymph]

it's half 12 am for me rn

[analog obsidian]

remember to not use fresh training when resuming, and you'll be fine

[bleak nymph]

yeye i clicked start training

[analog obsidian]

batch size has to be 8 as well

[crude flame]

imo index is so useless i dont understand why we require it for applications 🗿

[bleak nymph]

yup

AH MOYAI

[analog obsidian]

kinda makes things better sometimes

[dusty bramble]

i use index🗿

[crude flame]

somtimes

[analog obsidian]

second 0:18 sounds better with index 0.5

[bleak nymph]

opened tensorboard

is a lot of stuff

[dusty bramble]

matter of fact this is annoying me, it resets back to .5 everytime

[analog obsidian]

avg loss g/total 50

only graph that matters for you now

[crude flame]

just a bit better

[analog obsidian]

(everything matters but i dont want to complicate things for now)

for now u can kinda ignore the graphs since the training just begun

watch it when you're around 50epochs

ideally everything should be going down

[bleak nymph]

oh alr

[analog obsidian]

(fm most of the time doesnt go down, thats a rvc problem devs are trying to figure out why)

loss_avg_50

[analog obsidian]

the g/total one matters

ignore the others that are under "loss"

only the loss_avg_50 matters

[bleak nymph]

this?

[analog obsidian]

yup

[bleak nymph]

alr

so basically it should keep going down

and i stop training when it go up?

[analog obsidian]

for over 15 minutes or 30 minutes

just wait

[dusty bramble]

im realizing how lucky i been with my datasets cuz i never did all this stuff, but my models been rlly good so

[analog obsidian]

because sometimes it goes up randomly on early epochs

[bleak nymph]

ah