✨│ai-help

[odd shale]

That Mel Roformer drums model is not available publicly.

I mean, it's files. So no.

Sorry.

[obtuse glen]

ok

is this bad?

its still at 0%

[cyan cobalt]

thanks man works great, do you have any substitute for uvr as well?

[odd shale]

Which batch size you used and which was your dataset's length?

[odd shale]

Yeah, my friend Eddy did make an UVR Colab.

#📰│dev-updates message

[cyan cobalt]

thanks

[odd shale]

You're welcome bud.

[knotty moth]

• BS roformer 1053 in UVR beta (drum+bass, other)
• demucs4_ft (drum, bass, vocals, other)
• some good models in x-minus (https://uvronline.app/) but it needs pro subscription

[eternal yoke]

what are the best vocal remove app

[odd shale]

MVSEP.

https://mvsep.com/en/home

[eternal yoke]

does it got limit like 2 times a day or...?

[odd shale]

Nope, all u gotta do is register.

[eternal yoke]

alr thanks a lot

[rare goblet]

Ayo? @eternal yoke level 1 !!!

[obtuse glen]

how do i use bs roformer

oh i think i found an exe

[obtuse glen]

how long is it gonna be stuck on 14%?

i forgot to select gpu

nvm

great.

[knotty moth]

note that 1296 and 1297 are usual voc-inst models, and 1053 is the drum+bass model as I said before

[obtuse glen]

it was for vocals and intrumentals

[obtuse glen]

ohh

me dumb

completely ignored it by accident

[knotty moth]

but anyway you should do separate vocals & inst first, then use 1053 on the inst stem

[obtuse glen]

it doesnt have any vocals

[obtuse glen]

why is it doing this

sometimes it cuts in and out

[graceful patrol]

can someone help me im new to this

when i download the file it how do i upload

[proper shale]

what are you using

epochs really vary, but for batch size, I'd say go high, 12+

lower will kinda be unstable really

idk how much VRAM you have so

[eternal yoke]

i use mangio rvc is it still good

[proper shale]

if you're using T4, go for 16 batch

yesn't

it might be more unstable and might not do well... but eh, it's all testing... if you wanna test that it's all up to you

[proper shale]

yeah i guess

go for 250 (or 500 if it's fast enough) and check for overtraining... if not, keep going

if the training is fast

cause it's a 1 hour dataset 😭 can take a while

let's hope it ain't the dataset

[eternal yoke]

does those thing affect the ai sound

[knotty moth]

also check mel & kl as well

g/mel & kl should be there

[reef trellis]

@pastel oak My gpu is a 11th Gen Intel(R) Core(TM) i3-1125G4 @ 2.00GHz

[proper shale]

you're welcome :)

[proper shale]

yes

[proper shale]

that's your cpu, btw

[reef trellis]

@proper shale Intel(R) UHD graphics

[rare goblet]

Ayo? @reef trellis level 1 !!!

[fallen panther]

hi guys, in model training the number decreased to 6.8 but now nothing changes, can u help?

[proper shale]

yeah, that isn't gonna run okada

[proper shale]

what number are you referring to? give us more details

[reef trellis]

Oh ok ,thanks '^^

[proper shale]

You're welcome... sad it didn't work out for ya

[fallen panther]

Now I'm training my model, when I press the button, loading starts, but it doesn't reach the end and disappears

btw previously the final number decreased, but now it doesn’t

[proper shale]

Check your command prompt

[fallen panther]

I don't see any errors there :< or do I need something else?

[rare goblet]

Ayo? @fallen panther level 1 !!!

[fallen panther]

like that

[proper shale]

that's not a good sign

how did feature extraction go

and preprocessing as well

[fallen panther]

i see "ffmpeg error" in the output information of the 2a step, is this the problem? before that everything was fine..

[proper shale]

okayokay

how's your path to dataset look like?

[eternal yoke]

why did the ai sound like it got throat problems any tips

[surreal schooner]

How long does it take RVC-GUI to unpack from zip? Is it a very long time process?

[fallen panther]

"C:\Users\Grish\Downloads\RVC-beta\RVC-beta0717\ihiwk" this ig?

[violet heron]

Are you using RMVPE

[pastel oak]

if youre a guy and using a female model, you should increase the tune to maybe 10-12 and try around. else idk what you mean

[eternal yoke]

i use mangio

[violet heron]

Make sure to use RMVPE as the pitch extraction method/f0

[brittle wing]

anybody know where i can find a GOOD google colab or just online version of rvc?

i cant install rvc itself

[wild vapor]

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• EasyGUI, by rejects Google Colab
• UVR5 NO UI for Google Colab, by Eddy Google Colab

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

[cosmic holly]

anyone have crazy lag when using rvc with vb audio virtual cable?

it's super jarbled and hard to hear but sounds fine if I'm listening to it using my headpohnes

but if i put it through the cable input it sounds terrible

[brittle wing]

thank you 🙏

[violet heron]

Use VAC Lite instead

[quasi maple]

whats the best

-realtime

[azure marsh]

This interaction has expired, use the command /guides realtime if you wish to see it again.

[pastel oak]

https://rentry.co/RVCorWokada

but pls make sure you actually got a gpu that is not like 10+ years old

[modern surge]

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• EasyGUI, by rejects Google Colab
• UVR5 NO UI for Google Colab, by Eddy Google Colab

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

[eternal yoke]

Is it an app?

[pastel oak]

a setting

[empty raptor]

I wanna use UVR locally but I am iffy about space, power efficiency

Can you calm my nerves about it

[rain iris]

hi, i would like an up-to-date video tutorial on how to use the ai voices

last time i did anything with them was 2 years ago

i forgot everything

last time i made an ai cover was when the google colab thing still existed

[rare goblet]

Ayo? @rain iris level 1 !!!

[rain iris]

stfu

[tidal trail]

https://docs.aihub.wtf/essentials/how-to-make-ai-cover/

[rare goblet]

Ayo? @tidal trail level 10 !!!

[rain iris]

thanks

now i will be able to make a heavy x medic fanfic

[proper shale]

honestly id just re-export your dataset in audacity as a wav i think the audio is screwing up

[violet heron]

No it’s a option

It’s like crepe, harvest, pm

[odd shale]

I wouldn't suggest you watching videos, since these tend to be outdated.

Instead read the docs.

[rain iris]

already did and i successfully made a voice cover

[main lichen]

soo how do u cough withiut the ai tweaking

[violet heron]

Don’t get sick

[main lichen]

LMAOO

[cobalt pilot]

I don't know if it's intentional or not but Ilaria RVC isn't there anymore

[misty elk]

my download custom pretrain thing in the appolio colab keeps saying

---

HTTPError Traceback (most recent call last)
<ipython-input-6-962fec43dc94> in <cell line: 44>()
44 for url in pretrained_urls:
45 filename = os.path.join(output_directory, os.path.basename(url))
---> 46 urllib.request.urlretrieve(url, filename)

6 frames
/usr/lib/python3.10/urllib/request.py in http_error_default(self, req, fp, code, msg, hdrs)
641 class HTTPDefaultErrorHandler(BaseHandler):
642 def http_error_default(self, req, fp, code, msg, hdrs):
--> 643 raise HTTPError(req.full_url, code, msg, hdrs, fp)
644
645 class HTTPRedirectHandler(BaseHandler):

HTTPError: HTTP Error 401: Unauthorized

how do i fix it

[tidal trail]

change the custom pretrained cell code

[proven hill]

#📰│dev-updates

[proven hill]

https://docs.google.com/document/d/1YbXcLFPaGjhOdG5NFkK3QrucCEpHZBwFUxkeMO8aB18/edit

porcodio <3

[junior halo]

How big of a dataset is too big? Right now, I am sifting through a 4 hour Twitch stream. (don't judge my source it is my best option rn)

[placid holly]

1 hour (59.9 minutes), more than that it will not start train

sometime i just reduce them into 55 minute

[thin stump]

you would be lucky to have 30 minutes because most of the lines get removed during cleaning. Depends on your twitch vod source too

[junior halo]

YouTube....
If you're curious, I am creating a voice model of Gabriel from Ultrakill. I am sourcing from the VA's Ultrakill Twitch VOD from YT where he speaks like the character almost the entire time. I would use the original voice lines from the game, but the echo is messing with the quality. Assuming all of the voice models of Gabriel uses the game audio files, it is precisely why they all sound like shit. It's the echo and reverb that are messing with the output. The VOD, however, only uses reverb. I know how to remove reverb. I just can't remove echo. I've decided to make a hopefully higher quality voice model of him using this exact source.

[thin stump]

We do have a tool call Dialogue Isolate for echo. I'm assuming you're using RX11 already? I'm doing a twitch vod/youtube myself and I can't bring myself to review the whole entire audio so it's split into parts. At least 1 hour and 45 minutes -> currently bring brought down to whatever length because of unremovable mic noises

oh yeah, that echo is easily removable if you use that. I looked at one of Gabriel's video

[junior halo]

I see. I'm just gonna go through the entire thing and knock out similar phrases.
And no, I am not familiar with RX11. I am using the deverb plugin for FLS.

[thin stump]

I'll just plug this https://rentry.co/RVC-dataset-RX11#modules-overview. We have entire guide if you exhausted all your options on fl studios

[junior halo]

Ah, neat. I'm curious how this works.

I assume it's not a DAW plugin?

[thin stump]

It's a "Paid" DAW like audacity. It's been used by model masters for a long time and there's currently no alternative workflow thats being used to make models

it just gives you plugins, which you can also use on FL

[junior halo]

By "paid," you mean it's got a free trial or something?

[thin stump]

free trial?

[faint steeple]

does anyone know why voice ai keeps crashing everytime i try to speak with my voice changer?

[mossy coyote]

ok your/mother

[knotty moth]

you can bypass minibatch k-means by removing the code part if you wish

[knotty moth]

there's no noticable diminishing return until about 2 hours ig, besides the amount of time and effort you gotta have, but 10-30 mins (without silences) is usually enough to have variations

[graceful patrol]

when i download a girl voice model it doesnt sound good at all

[violet heron]

Are you using RMVPE

are you sure you aren’t using a troll model

Make sure your using a good tune

[edgy fjord]

I can retrain a voice?

[pastel oak]

with dataset yes

[edgy fjord]

I only have the .pth and .index file

[warm oriole]

https://cdn.discordapp.com/attachments/1173555844965740545/1258313633146867764/image.png?ex=66879764&is=668645e4&hm=b90d04a9970f4fcc9ca8e9506de87d9e5815e2cff73b9d8e3c4a6d3725c9e1cc&

guys, while i`m downloading model at Google Colab it says JSONDecodeError no matter what model i use. What am i supposed to do to fix it?

[rare goblet]

Ayo? @warm oriole level 1 !!!

[pastel oak]

then no

[pastel oak]

probably outdated colab

use -colab in this channel for working ones

[edgy fjord]

Ok thx

[rare goblet]

Ayo? @edgy fjord level 2 !!!

[warm oriole]

thx i`ll try

[pastel oak]

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• EasyGUI, by rejects Google Colab
• UVR5 NO UI for Google Colab, by Eddy Google Colab

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

[mild quiver]

any way to apply voice to a recording?

[pastel oak]

its called inferencing, check the last pinned message here and open Ilaria Zero + the guide

[mild quiver]

ty..i'll check

[calm crown]

Guys, does someone knows why some of voice models breathing sound sounds like a robot?

[glacial pollen]

1. lack of breathing samples in set / damaged breathing due to denoise

2. Crappy models / over or undertrained models

[quaint dock]

hey guys, is Applio safe?

sorry for the stupid question

[pastel oak]

Yes

We use it

[quaint dock]

alright, thanks!

[patent pasture]

I trained model for 500 epochs and found out that 1 epoch = 7 steps. 3927 / 7 = 561. How?

[glacial pollen]

a form of rounding.

provided you use a manual sync approach

tho, that should not happen

never in my training career encountered such issues

[patent pasture]

maybe RVC Disconnected bug

[glacial pollen]

Cause you did sync it on your own, correct?

or is disconnected using " auto syncing " bullshit

[proper shale]

most people don't

[proper shale]

and no you can insert your own value of steps per epoch to sync manually

but it's Mangio RVC and I've heard it ain't accurate when it comes to tensorboard logs

[glacial pollen]

if it's Mangio then ye, sadly.
But it's not because of the mangio itself but rather related to what kalo added, auto syncing. it is hardcoded.
so any " manual inputs " are overwritten

[glacial pollen]

@red kayak Hmmm, I think Imma update my fork - specifically, move from mainline base to fumiama's base

noticed some changes here n there, esp for training pipeline itself

after I'm done, will need some testers. you up for it maybe? ( will probs add some different gradient clipping in case of fp16 hence why I'd need testers as I have only so many sets on my disposition

[unkempt flame]

Hello, I want to use V1's voice from ultrakill and I don't know how to use it, can somene explain me how to sue this AI concepts?

[wispy lodge]

Time to use mixed precision and/or bfloat16

[knotty moth]

see the model checkpoint file names with suffix _eXXX_sXXX (epoch and step, respectively), or you can see it in the checkpoint saving process, also too few steps per epoch may be "inaccurate"
btw you should set 0.987 smoothing

[glacial pollen]

bfloat16 already tested. Mixed precision is already a go-2 for fp16 training ( AMP afteral )

bfloat might be a lil more stable than fp16 cause better hparam compatibility but sacrifices a lil bit of performance ( model quality wise ) but it's too small to notice for casual users and there aren't any gains

fp16 has it's own fair share of problems; exploding gradients, NaNs at the start so, awful training initialization stage and is more prone to single-mode collapses

fp32 is obviously the best choice for quality models, sadly it's naturally the slowest ( unless.. well, you can use tf32 but let's be real, not a lot of people here can use it.

[tame mica]

sue??? lol

[unkempt flame]

use*

My bad

[tame mica]

uhh barely any1 uses v1 models anymore but it should still be useable if i'm not wrong

-docs

[azure marsh]

Documentation

AI HUB Docs

https://docs.aihub.wtf/

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

• How to use RVC Mainline Colab by Cauthess
• Full AI Voice Model Training Guide (local) by Christopher Villanueva
• AICoverGen Colab Guide by Eddy (Spanish Helper)
• Create a model with RVC disconnected (colab) by Angetyde

[unkempt flame]

And can you help me knowing where or how to use this models? pease, I just- dont know lmao

[rare goblet]

Ayo? @unkempt flame level 1 !!!

[tame mica]

there's tuts on how to use the model in the docs

[red kayak]

yes d

do

thats how ull get the latest stuff

rvc boss isnt interested in updated rvc really

:p

[unkempt flame]

thanks a lot

[red kayak]

so let the community do its thing

[pale sorrel]

is anyone else having problems training with applio?

[subtle phoenix]

Is the RVC v2 architecture multilingual/language neutral?

[analog obsidian]

You can train any language and its gonna work just fine

[subtle phoenix]

Thanks!

[rare goblet]

Ayo? @subtle phoenix level 1 !!!

[glacial pollen]

yup, heard the stuff and read everything

Just concerned a lil about my own integrations and 'older' ported mechanics but we'll see.

but guess people do be taking inspirations from me lol

[red kayak]

it shouldnt be that hard

[glacial pollen]

you see, a lot of stuff I had in was slightly rewritten mangio code

to have it compatible with new structures

fumiama further did split a lot of stuff even further and changed some integrations but ye, should be fine

just wonder why they disabled the vhq sox

hm

[analog obsidian]

Fumiama saving rvc development

[glacial pollen]

nonsense

[red kayak]

what is that

[glacial pollen]

sox resampler

[red kayak]

ahh

[glacial pollen]

set to very high quality

it's commented out for whatever reason

I personally do use it

other than that, envelopes

so most of stuff Imma keep in

but ye, can't wait to test everything out

[red kayak]

well if it causes less aliasing

then good

[glacial pollen]

sox vhq is the best algo you can get really

[analog obsidian]

is fp32 training possible in fumiama rvc?

[glacial pollen]

it's possible in every rvc

[analog obsidian]

yay

[red kayak]

if u change the .json file u are going to use it'll always work

and then also the config.json too

[glacial pollen]

ye, cause it's set once per machine

from then onwards, only " inuse " ones are used

so if you change it on your own, that's what it is kept

[analog obsidian]

oh

[glacial pollen]

I still am not sure what was causing optimizers with adaptive mechanics to not work properly

so hopefully it's somewhat fixed in fumiama's revision

Really want to utilizer Ranger ngl

or AdamW with a lil lookahead support n gradient clipping for fp16

[red kayak]

those things basically help with stability u say?

[glacial pollen]

yea, gradients won't explode as they're clipped

and so, NaNs also

and that's actually a huge deal

[red kayak]

mm yeah

[glacial pollen]

cause sometimes you get NaN during epoch 0 / 1

[red kayak]

gradients also wont explode?

[glacial pollen]

so, during initialization

and that has it's consequences later on

[red kayak]

like from 100 it goes ti 1925

huge spike up

gradient boom

[glacial pollen]

that on it's own can happen in fp32 too

so that's just a matter of param updates

or re-sitting

[red kayak]

yeah but will the optimizers help

[glacial pollen]

but it won't cause overflows or underflows

so it's out of range

[glacial pollen]

that's a huge "?"

like with everything

[red kayak]

lmao

drop me the repo

[glacial pollen]

Lookahead is meant to be looking ahead with small weights

compared to normal ones

and sorta guide the convergence

but then, yea

a lot of adaptive stuff I've tried before

won't cooperate

which is a shame, cause even simple stuff as warmup or lr warmup was out of options

[red kayak]

ig a new gan can help ><

[glacial pollen]

I think it's more bound to how the schedulers were written in there

huge mess of a code ngl

a lot of stuff being " hardcoded "

and taken directly from hparams ( from file

rather than dynamically accessed n stored

iirc, I saw some changes related to that so, having big hopes

[glacial pollen]

Yea, it's not finished yet

there's no eta

[red kayak]

ye its meant for the community to cntribute too

[glacial pollen]

oh, wait

you meant the optimizers?

or what repo

[red kayak]

repo

[glacial pollen]

yea but, what repo

[red kayak]

funima

[glacial pollen]

oh, that

[red kayak]

mhm

[glacial pollen]

https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/tree/main

[red kayak]

yuppie

[glacial pollen]

mh mh

[glad zealot]

@brittle wingimma make when im bored enough

[red kayak]

it happens cuz of the learning decay, it has happened once to me b4

[fading rivet]

what is crepe hop

what is it doing

[night pier]

😎 👌🏿

[proper shale]

basically the length it will use to check for pitch variation. lower means it'll get more but it might crack, higher will get less but will be faster and less prone to cracking

if you're using rmvpe it doesn't change anything

[glacial pollen]

learning rate decay ( scheduling ) does not affect steps whatsoever

steps are directly related to the batch, dataset and so the data loader

[red kayak]

it doesnt?

[glacial pollen]

it does not

yes

[red kayak]

the more u know

[glacial pollen]

steps are directly related to batches and amount of data really

yup

but again, still can't get more " into it " cause the data loader works internally so

[red kayak]

i'll ask gpt for more info abt the parameters then

[glacial pollen]

Good idea

[stark wadi]

-overtraining

[azure marsh]

Overtraining

You can detect if a model is overtraining if the TensorBoard graph starts to rise and never comes back down. An overtrained model will sound robotic, muffled, and won't be able to articulate words well.

Check these resources to learn more about this topic

• Epochs & TensorBoard from AI HUB Docs
• TensorBoard from 🍏 Applio Docs

[brittle wing]

Does anyone know any tips to make the ai sound less glitchy and make it not voice break a lot?

[maiden raft]

-uvr

[azure marsh]

Ultimate Vocal Remover

One of the best free and open source vocal and instrumental isolation tool.

Tip: Use /guides UVR to learn how to use it.

Ultimate Vocal Remover

This guide shows how to separate the vocals and the instrumental of songs using UVR locally
Link: https://docs.aihub.wtf/vocal-cloning-guide/vocal-and-music-separation/uvr

Credits: Nick088

Ilaria UVR

Link: https://docs.google.com/document/d/1M3rEj5RPSE60f17RdM9MaIKKkS2bLrj-dwwuCUOdRaY/edit?usp=drivesdk

Credits: Angetyde

[maiden raft]

what happened to the uvr guide

[brittle wing]

how to install rvc models? theres many files

i used this
Aliyah Ortega (Irmã da Jenna Ortega) (Influencer)
https://huggingface.co/Pablao0948/Aliyah_Ortega_RMVPE/resolve/main/Aliyah Ortega.zip

Author: walker0384_28183 (@Pachwco)
Created: November 5, 2023 2:52 AM
Type: RVC
Algorithm: Rmvpe
Epochs: 300

[proper shale]

where do you want to use it on? because there's a lot of tools tbh

id just use weights.gg if you don't wanna handle more complex stuff, involves no installing shit and is simple af to use

https://weights.gg/

[brittle wing]

well i downloaded https://github.com/w-okada/voice-changer/, im not sure i came from yt vid but it js said ab index and pth files

[proper shale]

oh i see

just unzip and use the .index and .pth files

ignore the others

[brittle wing]

theres just pth file no index one tho, i was gonna send a pic but it seems like i aint allowed here

[rare goblet]

Ayo? @brittle wing level 1 !!!

[proper shale]

just use the pth then

[rare goblet]

Ayo? @proper shale level 109 !!!

[proper shale]

dw

[brittle wing]

alright, thank you so much!

[proper shale]

index is optional after all

[proper shale]

you're welcome!

[brittle wing]

Sorry i got another question, how does widgets.gg work or how to use it? specifically with discord

[proper shale]

weights gg is for covers n stuff. they host basically all our models and a bit more

to use okada anywhere else, you need a virtual audio cable

check https://rentry.co/VoiceChangerGuide virtual audio cable section

[brittle wing]

alright tysm

[jagged coral]

guys, what can i do, if some sounds are kind of loopping? i already increased chunk, still have this problem

im using radeon rx5700xt graphic card

[shut fog]

-audio

[azure marsh]

📚 Audio Guides

• 🆕 Creating Datasets for RVC using iZotope RX11, by Cauthess
• Perfecting Audio Isolation on Low-End Rigs, by Litsa The Dancer and Faze Masta
• Gathering and Isolating Audio, by SCRFilms :snowflake:
• Vocal Mixing Tutorial, by Roomie

🛠️ Tools

Audio Separation/Isolation

• https://mvsep.com/
• https://multimedia.easeus.com/vocal-remover/
• MDX-NET Colab

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

[gray crow]

Hey guys, anyone knows if RVC will run on 3.10.14 since im getting this message even tho i always used 3.10.12 and never had that before or just straight up ignored it lol.

2024-07-04 17:45:28,316:WARNING - WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
    PyTorch 2.3.0+cu121 with CUDA 1201 (you have 2.0.1+cu118)
    Python  3.10.14 (you have 3.10.12)
  Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)

[open mountain]

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• EasyGUI, by rejects Google Colab
• UVR5 NO UI for Google Colab, by Eddy Google Colab

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

[pastel oak]

Ye dont think it runs with 3.10.14

should use 10.12

[open mountain]

-hf

[azure marsh]

<:huggingface:1179800228946268270> Hugginface Spaces

• Ilaria RVC Zero, by thestingerx Huggingface Spaces
• RVC⚡ZERO, by r3gm Huggingface Spaces
• Applio, by IA Hispano Huggingface Spaces
• AICoverGen, by r3gm Huggingface Spaces
• Advanced RVC Inference, by r3gm Huggingface Spaces
• RVC v2 Huggingface version, by Clebersla Huggingface Spaces

More commands: -audio, -colabs, -kaggle, -local, -overtraining, -realtime, -rvc, -help

[gray crow]

thank ser

[maiden raft]

oh, sorry I just needed the uvr guide

[brittle wing]

Oh I thought u were replying to me

So sorry

[keen hound]

does anyone know how to add downloaded vocies into rvc?

[pastel oak]

local or colab

[keen hound]

local

[pastel oak]

assets -> weights

put it there

[keen hound]

thank you

[faint cloud]

anyone know where to put your path and index files in applio?

nvm figured it out

[ornate gulch]

hi

[rare goblet]

Ayo? @ornate gulch level 7 !!!

[edgy lion]

what is the recommended amount of epochs to train for a audio sample that is only like a minute long?

[crystal sky]

where can i find some audios that i can just dump into applio to test them.

[proper shale]

varies a lot, but I'd say 500/1000 epochs on 4 batch is where you'll get good results probably... ofc check the graphs for overtraining

[proper shale]

just generate stuff on ElevenLabs or test with memes... idk

oh and Applio has their TTS as well .... if that works

[empty raptor]

Can i set mvsep to not to have long ass file names

[ashen cipher]

i have a good voice model, but when i try it out, it sounds like shit...

[rare goblet]

Ayo? @ashen cipher level 1 !!!

[ashen cipher]

what settings do i need?

[proper shale]

in what aspect specifically

[ashen cipher]

making songs

[proper shale]

pitch? pronunciation?

[ashen cipher]

pronunciation

[proper shale]

mess with "Search feature ratio" in case the accent ain't similar to the person

if that ain't enough it's probably the audio you're using

[ashen cipher]

ok thanks

[proper shale]

you're welcome, lmk how that goes

[ashen cipher]

it sounds a bit better, but still not so good. It's probably because of my audio input file. I extract audio from songs and try to replace the vocals with ai vocals, but I think the vocals have a bad quality

[violet heron]

Make sure you use wavs instead of MP3s

[ashen cipher]

i use flac

is this good?

[proven hill]

whats the source of the flac

[violet heron]

Yes, just make sure you are getting them from a source where it’s loseless

YouTube will be lossy

[brittle wing]

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• EasyGUI, by rejects Google Colab
• UVR5 NO UI for Google Colab, by Eddy Google Colab

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

[brittle wing]

-audio

[azure marsh]

📚 Audio Guides

• 🆕 Creating Datasets for RVC using iZotope RX11, by Cauthess
• Perfecting Audio Isolation on Low-End Rigs, by Litsa The Dancer and Faze Masta
• Gathering and Isolating Audio, by SCRFilms :snowflake:
• Vocal Mixing Tutorial, by Roomie

🛠️ Tools

Audio Separation/Isolation

• https://mvsep.com/
• https://multimedia.easeus.com/vocal-remover/
• MDX-NET Colab

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

[junior halo]

It should be okay if I just feed the RVC trainingt Colab an hour long wav file, right? Or should I break it up?

[agile apex]

so im trying to download my ai covers but when I click download it redirects me to a cloudflare 404 page does anyone know how to fix this

[low shard]

What tool do you use

[agile apex]

weights ai cover tool

[rare goblet]

Ayo? @agile apex level 1 !!!

[low shard]

The site? Mmm @willow lichen have any idea?

[brittle wing]

-uvr

[azure marsh]

Ultimate Vocal Remover

One of the best free and open source vocal and instrumental isolation tool.

Tip: Use /guides UVR to learn how to use it.

[agile apex]

yes the website but my first few times I used the discord bot to make the covers

[small iron]

Yo how do i make my own voice rvc

On mobile

[low shard]

Mmm, the site seems working fine, i pinged the creator he may knows more

[low shard]

You mean making an rvc voice model?

[small iron]

Yes

Im completely lost here

[low shard]

Check https://docs.aihub.wtf/essentials/how-to-make-voice-models/, use the cloud ways such as rvcdisconnected or applio

[small iron]

Does it work on mobile

[low shard]

The cloud ways yes, its google colab, you are using a google pc gpu to run it so not locally on your phone

[small iron]

Btw can i sing when i train my own voice

Or have a complete high quality

Of a conversation

[low shard]

Wdym with sing? Like using it later for ai covers?

[small iron]

Datebase

Data*

[low shard]

I think you mean dataset, yea in the rvc model you can add CLEANED singing vaw files

Be sure that the audio is cleaned and only vocals

[small iron]

It gave me pc orders

Which im mobile

Isnt there a mobile version of this

[low shard]

Pc orders? Where? You sure u are using the cloud google colabs?

[low shard]

The google colabs can be used completely fine on phones

[small iron]

In Datasets

"Press CTRL + A to select the whole audio."

[low shard]

Could you send a screenshot of where you see this? Or a link of what part of the guide u see this

[shy hemlock]

I am training a model of my voice, and was just wondering what the best training settings I can use, mainly the amount of epochs and batch size... i got a lot of audio 7 files in approx 2hrs and 50mins in wav.

[small iron]

How do i do this

Ob mobile

On mobile

[hasty canopy]

Hello i was wondering if all the voice models were safe or only the popular ones

[agile apex]

I would send an image but it won't let me

[small iron]

Ho how do i make a zip file

In mobile

??

How do i clean my database with using phone apps

I dont have pc

[odd shale]

What do you mean with "safe"?

No model should contain viruses.

[agile apex]

@low shard sorry to ping you but I made a new cover and it let me download it pls don't ban me

[low shard]

You can make a folder on your phone and zip it, you can just go to whatever storage manager your phone has built in, and for example for mine i hold the folder and click the 3 dots to compress it to zip

[rare goblet]

Ayo? @low shard level 94 !!!

[hasty canopy]

I mean no risk

[hasty canopy]

ok nice

[odd shale]

You can use any model you want, it's fine.

[odd shale]

Do you mean cleaning datasets in mobile?

[small iron]

Is there any app where it can clean my vocals

[hasty canopy]

ok cause I wanted to try some models in my native language but nobody tested them it seems so idk if it was safe

[small iron]

To remove all disturbing noises and

Stuff

[odd shale]

You can use UVR on PC.

[low shard]

Dw abt ban, also i meant to say that's a problem of the weights.gg site so i dont really know how to help you with that sorry, that's why i @ the creator of it

[odd shale]

Which is your native language?

[small iron]

Im in mobile

[hasty canopy]

french 🥖

[rare goblet]

Ayo? @hasty canopy level 1 !!!

[odd shale]

In that case.. Emm, i'm not sure. Maybe there are some french models you can use.

[low shard]

Not app, sites, theres mvsep, cloud uvr https://docs.aihub.wtf/rvc/resources/vocal-isolation/ and also uvr ui #📰│dev-updates message

[hasty canopy]

I mean as long as there's no virus it's good for me

[small iron]

Can uvr make my audio without any disturbing noises only the vocal of my own noise

[odd shale]

Well, as i've said no model should contain viruses.

[agile apex]

its ok maybe it was just an error or the file got corrupted

[hasty canopy]

In that case it's good

lmao

thx

[agile apex]

but it lets me download new covers

[odd shale]

You're welcome bud.

[proven hill]

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

[low shard]

All models are safe, for example on hugging face they are checked to see if there's anything bad

[low shard]

Weird then

[agile apex]

bye

[small iron]

How do i do the Noise Gate step

Which requires pc

[potent helm]

how do I continue training the same model in the same colab I started training it in if I got the Error
Connection errored out error

[knotty moth]

use Renegate free vst plugin for that purpose

[small iron]

Yo your here for a moment

[knotty moth]

huh?

[sand beacon]

i need help

can someone send goof yt

good

[errant osprey]

anyone have any tips or tricks for getting an AI voice to work well with natural inflections and pitch changes while talking (realtime), assuming thats even possible

[opal depot]

Would an rx 580 be fine for running rvc?

[craggy sorrel]

does anyone know how to deal with this error i paperspace?"root@nzdhhlx08h:/notebooks/Mangio-RVC-Fork# make run-ui
python infer-web.py --paperspace --pycmd python
Traceback (most recent call last):
File "/notebooks/Mangio-RVC-Fork/infer-web.py", line 26, in <module>
import faiss
ModuleNotFoundError: No module named 'faiss'
make: *** [Makefile:59: run-ui] Error 1
root@nzdhhlx08h:/notebooks/Mangio-RVC-Fork# "

[shy yacht]

Does anyone know where A RVC model of ajitani hifumi?

[pastel oak]

path could have issues if theres spaces or OneDrive or special chatacters

[pastel oak]

You cant train voice models with amd gpu on windows
You can inference voices
Realtime voice changer will probably have a 1 sec delay maybe slightly more

[pastel oak]

no yt

https://docs.aihub.wtf/

[inland jasper]

hey so, it might be a stupid question but i'm trying to make my first voice model for the first time and i trained my model and i can't find it, where does it go to?

[hard belfry]

Good day everyone! Can somebody please advise, there used to be a desktop app, within which you could convert audio using voice models you downloaded and even search for other voice models withing the software. Had a nice black interface, please advise if it's not up anymore of if you know which one was it, struggling with finding it online

[pastel oak]

if you trained local its in assets -> weights

colab maybe sonewhere similar i dont know much about it

[glacial pollen]

you most likely mean old rvc gui or something of that sort

if that's the case, it's heavily outdated and nobody uses it / shouldn't be using

tho wait.. I read it again, hmmm.. it had a search for models built in.. hmm

Applio is the only thing I can think of

it has built in search

this one:
https://www.google.com/url?sa=t&source=web&rct=j&opi=89978449&url=https://github.com/IAHispano/Applio&ved=2ahUKEwiuod78zo-HAxXOVPEDHco9CHcQjjh6BAgQEAE&usg=AOvVaw1Y2ajvvkBWdvNWTnscb5GU

[inland jasper]

thanks! i also wanted help on how to remove voice models to not waste space incase it is terrible

[pastel oak]

if you saved G and D paths for every x epoch. then go into logs -> yourmodelname and delete any d and g epochs you dont need

else if you saved every x epoch for models theyre also in assets - weights

[inland jasper]

thank you so much man really appreaciate it 🙏

[rare goblet]

Ayo? @inland jasper level 1 !!!

[low shard]

You can search rvc ai voice model:

• #1175430844685484042
• #🔍│find-models
• https://weights.gg/
• https://huggingface.co/models (but watch out cus in hugging face there arent only rvc ai voice models)
• https://applio.org/models
• https://voice-models.com/

if there isnt one, you can:

• #1159289738314919936
• #1191429836321849435
• make it yourself with our docs guides https://docs.aihub.wtf/

You can't convert the voice model, even if not natively, you can use tts with rvc models, some 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)

NOTE: RVC isn't as good as GPT So Vits for tts, but gpt so vits can't use rvc models (and viceversa), and its only limited to: english, chinese & japanese

If you wanna do tts locally (if you got a good pc):

• You can find Ilaria RVC Mainline in #📰│dev-updates message
• While Applio in our docs

If you don't got a good pc you can do it online:

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

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

• if you don't wanna use edge tts, you could try another tts ai in https://discord.com/channels/1159260121998827560/1212368971127590922 and use the output as an input in rvc

[subtle phoenix]

When finetuning or training a pretrain from scratch, the large dataset doesn't have to all of the same speaker right? I checked out the pretrains guide but still need some clarification on this point. Thanks!

[hard belfry]

will see if that's it now, tysm anyway!!!

[analog stream]

Hi, I just started using this today and have been messing around with various voice models with the intention of recording some stuff, everything was working as expected, however all of a sudden in my recording the voice isn't changing, I have tried to remove the application and such but am unable to fix it, does anyone know how to fix this?

[patent pasture]

I guess Ilаria RVC Zero isn't confidential

[pastel oak]

Wdym

[pastel oak]

Which application

[pastel oak]

It does not have to be of the same speaker correct

[patent pasture]

when you load your model, dO other users see it too?

[pastel oak]

Not sure

[subtle phoenix]

Thanks

[brittle wing]

@viscid moss can you add bs roformer to Karafan Improved?

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• EasyGUI, by rejects Google Colab
• UVR5 NO UI for Google Colab, by Eddy Google Colab

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

[analog stream]

Voice-changer-native-client or realtime voice changer client v1,5,3,18a onnxgpu-cuda is the application

[fresh thistle]

Sometimes, the output comes out choppy and laggy and even with trying out different chunk sizes, I don't get any desired result
Until I restart and spend more time, and then it works? but once I restart the program and try again, it's still choppy so I'm not sure why

[idle hedge]

is there a way to download illaria rvc to run local on my machine on mac?

[odd shale]

I'm not sure if Ilaria Mainline is supported on mac.

[idle hedge]

are there any alternatives to it then?

[odd shale]

Just download Applio or original RVC.

[idle hedge]

could you by any chance drop a link? :)

[odd shale]

-rvc

[azure marsh]

Suggestions for @idle hedge

Documentation

AI HUB Docs

https://docs.aihub.wtf/

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

• How to use RVC Mainline Colab by Cauthess
• Full AI Voice Model Training Guide (local) by Christopher Villanueva
• AICoverGen Colab Guide by Eddy (Spanish Helper)
• Create a model with RVC disconnected (colab) by Angetyde

[knotty moth]

your gpu may not be strong enough

[idle hedge]

sorry, dont want to be a bother, but i only see the windows and linux installation guides, nothing for macs

[fresh thistle]

Like Intel(R) Iris(R) XE Graphics isn't going to cut it? :"")

[rare goblet]

Ayo? @fresh thistle level 1 !!!

[odd shale]

I'm not sure if Applio is supported on mac.

You can just use Applio Colab, all you need is a google account.

[idle hedge]

ah man, i kind of want to have all these things locally, maybe ill find some other option, thank you for your help thought i appreciate it

[rare goblet]

Ayo? @idle hedge level 2 !!!

[odd shale]

You're welcome bud.

[tidal trail]

solo para inferencia

aparte que la instalacion es un infierno

[viscid moss]

Currently, KaraFan gets a lower SDR compared to the BS models. Roforer could be added (it would be somewhat difficult) but the SDR will be slightly lower than using the model directly in UVR5 NO UI or UVR5 UI. So... why not use those?

[brittle wing]

For a reason cause it gives good results.I wanna ensemble hq4 and bs roformer, Along with other MDX models and I don't have a computer

[tawny tinsel]

Hi, I have changed both of my desktop input and dscord input to Dubbling Virtual Devices but when I tested on discord sound and voice, it couldnot detect anything

[brittle wing]

Other people are asking for that too.

[proper shale]

it should also be okada's output

[brittle wing]

@viscid moss

[proper shale]

The input output settings should be like this in W-Okada:

Input - Your mic
Output - CABLE Input (line 1 if you use VAC, or whatever your virtual cable output is)

And in Discord:

Input - CABLE Output (same thing as okada output as well)
Output: Your headphones

basically

[tawny tinsel]

Let me try

[viscid moss]

But you can use UVR5 NO UI or UVR5 UI to separate them using Colab or HF. The ensemble is not ready yet but you can separate several audios at once with one model and then separate all those audios again with another model and so on.

[brittle wing]

This is not the case other people are also asking for bs roformer on karafan I have done that already its not the same.So you're planning on adding ensemble on your UVR no UI COLAB?When

Karafan is ensemble

@viscid moss

[viscid moss]

The problem is that KaraFan has not been updated in 8 months, the results are quite good but it uses the UVR5 code from 8 months ago, which will require more effort to implement the new models (Roformers).

KaraFan is a good tool but it is not receiving updates, for now I can only tell you that I will review it but I cannot promise that I will fix everything completely.

I know that KaraFan makes an ensemble with UVR5 models, that's what I'm working on for UVR5 UI and UVR5 NO UI.

[brittle wing]

Will you add an ensemble cell to the UVR colab you've made?

[viscid moss]

It's what I'm working on, along with other projects

[brittle wing]

That's useful

When will it be added soon

So you have to update the code first ah yes I get it

[proper shale]

give @viscid moss some time ik she can pull off cool shit

https://cdn.discordapp.com/attachments/1254252587973083187/1254428982133395496/image0.gif

[viscid moss]

Soon, after the implementation of audio separation using YT and batch separation for UVR5 UI

[brittle wing]

Cat explotano

[brittle wing]

In a month or days?

[viscid moss]

month

[brittle wing]

Next month

Okay

[simple hedge]

A question, what page do you use to make audios with AI???

with rvc

[crystal sky]

you input audio

and it changes the audio

there is a text to speech button too

but that just uses a basic text to speech

[simple hedge]

I asked what program they use

Why does that appear to me?

[odd shale]

Which colab are you using?

[brittle wing]

Use RVC DISCONNECTED

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• EasyGUI, by rejects Google Colab
• UVR5 NO UI for Google Colab, by Eddy Google Colab

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

[brittle wing]

anyways what does the allign inputs feature do in uvr v5?

[low shard]

outdated colab

use the inference colabs u can find in our docs https://docs.aihub.wtf

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• EasyGUI, by rejects Google Colab
• UVR5 NO UI for Google Colab, by Eddy Google Colab

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

[low shard]

i think he wants to do inference

not training

[brittle wing]

Well then he should use these colabs provided

[odd shale]

He can easily use Applio Colab

[worthy lantern]

hmm which collab best for using public? 🤔

[snow rose]

How to make text to speech fit song?

[azure marsh]

🍏 Applio-RVC-Fork

Enhanced RVC Variant: Optimized Performance Through Modifications, Built upon Mangio-RVC-Fork.

Documentation

📚 All-In-One English documentation

https://docs.aihub.wtf/

Full AI Voice Model Training Guide (Local)

Link: YouTube
Credits: Christopher Villanueva

Model training with Mainline RVC

Link: Rentry
credits: Raven (ravencutie21)

AICoverGen Colab Guide

Link: Google Docs
Credits: Eddy (Spanish Helper)

Create a model with RVC disconnected (colab)

Link: Google Docs
Credits: Angetyde

[nimble blaze]

The fun part is going to be figuring all of this out, in terms of knowledge of any of this I'm at like...0.5/10 lol.
I've used Applio/have a model downloaded that I want to convert to TTS. It's just figuring all this out. And this could be used for an AI assistant that reads ChatGPT/etc in real time right?

[rare goblet]

Ayo? @nimble blaze level 3 !!!

[worthy lantern]

have someone a good google collab to use public url?

[spice thorn]

how can i install hina rvc on my pc?

[spice thorn]

https://colab.research.google.com/github/hinabl/AICoverGen-Colab/blob/main/Hina_Mod_AICoverGen_colab.ipynb#scrollTo=NEglTq6Ya9d0

[worthy lantern]

not working 🤔

[dawn tusk]

oops, I need help, I've already downloaded the voice change app, but when I press it to start, the voice doesn't come out, can anyone help me?

[west cosmos]

I don't mean to comment over the person above (sorry) but I just wanted to ask if anybody has any experience doing batch conversions on Hugging Face using RVC V2? I do a lot of conversions but I cannot work out how the batch converter works - I can never find the files to download once they're done

[rare goblet]

Ayo? @west cosmos level 1 !!!

[spice thorn]

I'm on this idk

Try use ngrok

Not gradio

[worthy lantern]

do u separate vocals and instrumentals on this site or no

[spice thorn]

Yes

I do

[proper shale]

You'd probably need to use Colab for that since you can't see temporary files on HF iirc

[west cosmos]

Right, understood. Thank you!

[brittle wing]

rvc cli stopped working for some voice models for no reason

do models need replacing every now and then?

it'll generate but then it won't make a wav file

[analog current]

guys, do you know anything about applio app? is it better than other apps that use RVC?

[glacial pollen]

lol, and yet you use " ignore outliers "

that's not the approach

mark it off, then evaluate

set the smoothing to 0.70

or less

100 ( smoothing at 1 ) is for general trending recognition

[rare goblet]

Ayo? @brittle wing level 4 !!!

[proper shale]

looks like a collapse

you should be fine using the most recent ones though, probably

[violet heron]

there’s no reason to use over RVC

[brittle wing]

Use Harmonify Colab

[slow dirge]

hello when i use rvc a message appear and i cant use rvc someone can help me pls ?Error opening Stream: Illegal combination of I/O devices [PaErrorCode -9993]

[naive shoal]

yo tiny tim!!

That's precicely what happens

[thin stump]

yes! That's the only reference I had from your conversation so i'm kinda glad I can pull from that

[naive shoal]

Yea and I found it fascinating discovering it. It's making me want to start expiramenting with those multi-models

To see what would happen if I trained a base model on tiny's baritone then did another model on his falsetto

I dont know how that works though at all

[thin stump]

It's all AI, like if you change the distribution of a particular sample then the accent and quality varies

[naive shoal]

Wouldnt that be good though?
Because no matter what his resonance and accent between falsetto and normal singing is very differet

And RVC seems to just kinda mash them together between those ranges

[thin stump]

I tested a model with monotone speaking and including expressive sample improved the accent by a lot. Variety is needed so yes

[naive shoal]

On top of that, the baritone voice can't ever belt out anything higher up in his range because its trying to mash into the quiet falsetto

[naive shoal]

But too much variaty seems to be a huge problem too. I just dont think it understands the tonal difference between headvoice (quiet) and chestvoice (booming)

Tiny can hit the same notes in chestvoice as he can headvoice

But the AI tries to force everything past a certain limit into headvoice

Or some weird inbetween

[thin stump]

Yup that's just the limitations with the current RVC model (Hifigan) I have no clue about the Bigvan, evagan stuff

[naive shoal]

What are those two? ive never heard of them

[thin stump]

#📰│dev-updates message

v3 soon, no guarantees it will improve from v2 but you never know

[naive shoal]

Oooooo V3!!!

heh it doesnt even look like itll do anything

Why the pretrains?

[thin stump]

actually ignore what I said

[thin stump]

well maybe I was onto something because the pretrains have to be trained on hifigan? So if we switched then it would be way better

[naive shoal]

Whats wrong with hifigan

Also whats hifigan

[thin stump]

that, I would have to research more into it for the problems. It's the generator and discriminator that makes up RVC rn

[thin stump]

https://rentry.co/RVC-dataset-RX11#dereverbing-and-de-echo-the-audio

[naive shoal]

Im not sure what we'd move to though if we stopped using it

[thin stump]

You can search up BigVgan or evagan in the server logs. It's fun to read up on stuff that probably wouldn't happen soon

i spelled it wrong

[naive shoal]

My only worry is that with things advancing my GPU will become too old to use

Its already going on 6 years old

It was topend at the time but now itsj ust old

It does the same things as the modern RTX cards but is just slower

But if they introduce anything that minimally requires more VRAM ill have issues

[thin stump]

oh, it will definitely increase. I don't know where I read that discussion from

[naive shoal]

(i got the RTX 2080super card)

[thin stump]

maybe in model-maker chat

[naive shoal]

Here's a tiny clip of tiny singing https://youtu.be/8B2qxVbKgV8?si=5T6ehCItdoubLRLZ&t=7

"never hit your grandma with a shovel - it leaves a bad impression on her mind!" XD

Or this https://youtube.com/clip/Ugkx2ilCNaRW8FHfxgnqN0ALVejkS8TH3cBg?si=paIlL7VxutWitcaw

[thin stump]

Never knew he was an actor too, huh cool

[naive shoal]

Oh yea he did all sorts of stuff. He was even a clown in a certain horror film

[thin stump]

holy shit its the joker

[naive shoal]

He's the perfect mix of comedy and scary

And goofy

He was plain goofy he couldn't be scary

But his looks could be

This is him when he was young

[thin stump]

I wonder if that was heavily inspired for Joaquin Pheonix since he's a more grounded guy

[naive shoal]

I wonder that too
Tiny was very grounded though

He had his acts but if you listen to him talking he's a very real guy

Catholic and very oldschooly conservative XD

[oblique bay]

Hello, I'm having issues with discord where the virtual audio (or the rvc voice changer) sounds very glitchy. However when I listen to the actual audio of the virutual audio it sounds perfectly fine on playback on my system. I've turned off all noise cancelation and it still sounds glitchy, I suspect it being the transfer over to virtual microphone to discord playback.

[violet heron]

Use VAC Lite instead of vb audio

[oblique bay]

Got it, let me try it out. Thanks

[rare goblet]

Ayo? @oblique bay level 1 !!!

[oblique bay]

This fixed my problem, thanks!

[violet heron]

np

[oblique bay]

is there any free alternatives?

[violet heron]

I just said vac lite

VAC Lite is free

Make sure you download the lite version not trial

[oblique bay]

I see, I downloaded the VAC trial one

[cloud jetty]

I noticed a lot of the RVC models are singing models - do they still work for speaking?

[violet heron]

Yes they should

[cloud jetty]

It is better to use a model that was trained on talking for talking vs singing?

[violet heron]

Singing models can have a better range sometimes

Try both and see which one works better

[cloud jetty]

good to know! Thank you

[junior halo]

How do I resume a training session?

(non local RVC)

I've started over countless times because I was inactive for too long.

I need someone to train this for me. Please.

[brittle wing]

can someone help me with the ai voice?

i need help

im trying to use the google collabs thing

when i click on a model

it says

Frequent errors occur. Please check if the model of the framework being targeted is loaded.

[misty elk]

--colab

shoot

[brittle wing]

can anyone hop into vc with me to help me with starting the training

everythings set up and i have my audio ready fully

[knotty moth]

lite is free and enough, wont spawn some kind of annoying popups tho

[fair onyx]

Hi. Is there any way to use RVC with an amd gpu (rx 6600). I want to use inference locally, but I don't need to train my own models.

[rare goblet]

Ayo? @fair onyx level 1 !!!

[knotty moth]

https://rentry.co/RVCRealtimeGuide
perhaps, despite not mentioning offline inference

[violet heron]

-local

[azure marsh]

Suggestions for @fair onyx

🖥️ Local stuff

• 🍏 Applio, by IA Hispano GitHub
• Mangio-RVC-Fork, by Mangio621 Huggingface
• RVC Studio, by SayanoAI Huggingface
• AICoverGen, by SociallyIneptWeeb GitHub
• Replay, by Replay Team Website
• Original RVC, by the RVC-Project team GitHub
• GPT-SoVITS, by RVC-Boss GitHub

Credits to Faze Masta and Antasma for compiling these links.

More commands: -audio, -colabs, -overtrain, -realtime, -rvc, -help

[violet heron]

It will use your cpu

[misty elk]

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• EasyGUI, by rejects Google Colab
• UVR5 NO UI for Google Colab, by Eddy Google Colab

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

[fair prism]

what do yalls use for tts?

[dull dagger]

how can i save my weights.gg outputs

[oblique bay]

Yes it is sufficient enough

[agile apex]

I made heavy from TF2 sing rainbow connection from the muppets movie and it kind of sounds like he's actually singing it

[wispy pawn]

how can i fix no interface i running right now

[rare goblet]

Ayo? @wispy pawn level 1 !!!

[wispy pawn]

each time i load a local url in google colab it say that

[golden karma]

-guide

[azure marsh]

Documentation

AI HUB Docs

https://docs.aihub.wtf/

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

• How to use RVC Mainline Colab by Cauthess
• Full AI Voice Model Training Guide (local) by Christopher Villanueva
• AICoverGen Colab Guide by Eddy (Spanish Helper)
• Create a model with RVC disconnected (colab) by Angetyde

[golden karma]

-spaces

[azure marsh]

<:huggingface:1179800228946268270> Hugginface Spaces

• Ilaria RVC Zero, by thestingerx Huggingface Spaces
• RVC⚡ZERO, by r3gm Huggingface Spaces
• Applio, by IA Hispano Huggingface Spaces
• AICoverGen, by r3gm Huggingface Spaces
• Advanced RVC Inference, by r3gm Huggingface Spaces
• RVC v2 Huggingface version, by Clebersla Huggingface Spaces

More commands: -audio, -colabs, -kaggle, -local, -overtraining, -realtime, -rvc, -help

[golden karma]

-rvc

[azure marsh]

Documentation

AI HUB Docs

https://docs.aihub.wtf/

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

• How to use RVC Mainline Colab by Cauthess
• Full AI Voice Model Training Guide (local) by Christopher Villanueva
• AICoverGen Colab Guide by Eddy (Spanish Helper)
• Create a model with RVC disconnected (colab) by Angetyde

[golden karma]

-audio

[azure marsh]

📚 Audio Guides

• 🆕 Creating Datasets for RVC using iZotope RX11, by Cauthess
• Perfecting Audio Isolation on Low-End Rigs, by Litsa The Dancer and Faze Masta
• Gathering and Isolating Audio, by SCRFilms :snowflake:
• Vocal Mixing Tutorial, by Roomie

🛠️ Tools

Audio Separation/Isolation

• https://mvsep.com/
• https://multimedia.easeus.com/vocal-remover/
• MDX-NET Colab

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

[modest abyss]

Any help with this? (I tried searching the server but couldn't find anything)
I made sure to reinstall Python just in case, delete and redownload the zip, moved the folder to my main drive instead of my 2ndary, nothing

[rare goblet]

Ayo? @modest abyss level 2 !!!

[modest abyss]

It's been awhile since I've done this (last time I tried making a model was when Google Colab was still the top option lol)

[brittle wing]

I downloaded RVC and want to use this voice https://discord.com/channels/1159260121998827560/1258842728813559909 but when I talk, I can't hear myself and the voice isn't being detected.

HELP

[dawn dew]

using RVC GUI for the first time. How long does it usually take to convert the voice after clicking convert button?

[modest abyss]

Depending which GUI you're using and how long the audio is. For me the average song takes around 30 minutes, but all my stuff is lower end so it may be faster for you

[dawn dew]

ok. I have a low end laptop, no gpu. Model is Taylor Swift Red Era v2, the song file is an acapella of "Empire" by Of Monsters and Men. I set 'crepe' as the F0 method. crepe hope is at 128

felt like it's been 15 to 20 minutes so far and it's still converting.

[wispy lodge]

May take a very long time on CPU

Especially with crepe f0

[dawn dew]

what settings do you recommend for low enders?

[rare goblet]

Ayo? @dawn dew level 1 !!!

[wispy lodge]

Not sure about rvc-gui, but at least rmvpe as f0 would speed up the process a bit

[dawn dew]

ok it just finished

it seemed like 40 - 45 minutes. The next one I do I'm going to time it.

[exotic mural]

i finsihed training i think, but my assets/weights folder has no added_(smth).index file

i only see the .pth

[knotty moth]

click "train feature index"

[exotic mural]

tysm

[sick canopy]

i have a question what does the segment and overlap do?

[brittle wing]

Is there anyway to make the AI voicr sync with your laughing without it sounding like wind

Or is it just not possible

[proper shale]

Apparently laughing is kinda not possible with RVC rn...

But I'm not sure if that's a limitation or just the lack of laughter in datasets

[chrome orbit]

when i launch the tensorboard in applio rvc it just keeps reloading, is there something else i need to install for it?

[brittle wing]

Do you think it would be possible in 2 years time

Where gpu's like the RTX 5090 already being released

[proper shale]

Well... only the future shall tell

[brittle wing]

Nvidia is pretty much carrying AI

[tidal trail]

the link is above

[glacial pollen]

This has nothing to do with gpus and generally there's a big misunderstanding here in terms of laughing or such sounds in rvc or models I see

Laughing compared to speech / singing is more airy, it contains more " noise " type of data compared to phonetics, so, formants and so on

base models rvc uses are done on vctk dataset which barely if even had laughing-like pattern samples in it

rvc is just not accustomed to handling such well

And so, it is not bound to tech or gpus, it is just how the base models were made

Laughing is definitely do-able but not in all cases. It's a lil more tricky than that

Oh yea, not to mention contentvec rvc uses for " recognition " of what is the content of input audio ( for inference ) - I believe - has some harsh time or issues detecting such " unusual " sounds like that

[pastel oak]

i cough, laugh and moan daddy bulldozer on rvc 🔥

[glacial pollen]

So really, it all comes down to many variables.
Contentvec, base models' dataset used and obviously, quality of models

[glacial pollen]

Yea, it is do-able like I said, but it's def on a harder difficulty spectrum compared to speech or singing

[analog obsidian]

So in theory if someone add laughs to contentvec and a pretrain with laughs, would be possible to get perfect laughs in inference?

[glacial pollen]

There is no such a thing as perfect X in rvc

or AI

[analog obsidian]

[glacial pollen]

there always be artifacts or irregularities

maybe if some mid-way step was to be added

such as diffusion denoising or whatever

then maybe it'd be very very close to what we'd call perfect

I just feel like people expect too much from tech that's been with us for barely few years

[analog obsidian]

ye ai is new tech

[glacial pollen]

not really AI, but voice cloning and reconstruction in a way we do it

normally, in fact, you'd want to literally train from scratch your own voice or a dataset

not using any " universal " pretrains

but having your own end-to-end finetuned

so that's also to be added to the mixture

People expect way too much from something that was meant to just work for broader audience

[analog obsidian]

O yea i remember someone said our models are actually all finetunes

[glacial pollen]

That's right

that's the reason we use pretrained bases

our models are just " snapshots "

of finetuned generator and discriminator networks - kinda, simplifying it.

which were previously trained from scratch on vctk dataset

What I think could work is doing a 2 run training but I haven't tried it myself yet.

but that'd require you to have a rather big dataset for your model

first training is the fine-tune training, you'd then take your G and D from model's folder, and use them ( provided you can adapt hparams properly ) and train again

on same voice / person but different set - and that'd be your main training but let's be real, not a lot of people here can do it

[analog obsidian]

This would require a lot of audio hours?

[glacial pollen]

Not really

1st run you'd do like normal, but preferably 10-25 mins

2nd run would be 15-35

but I think 10-15 could work again

just each run has different set ( or different section of the set provided the consistency is kept )

but if I had to be honest, first run should be as good as you can get it to be

'll be your foundation for the main model

[analog obsidian]

This sounds interesting, uhm, i believe this would be the best thing to do if you’re recording a dataset yourself

[glacial pollen]

Definitely

[crude flame]

Imma try this and see what I can get. Sounds very interesting

[glacial pollen]

Go for it

tho when I mention that 2nd run should be a different set ( in ideal conditions ) I mean it

as the first run ( first model, so, your base ), provided you legit trained it to the best of your possibilities, will be already accustomed to the voice

so the risk overtraining is drastically increased

[crude flame]

Different in length and what's in it? Or just length

[glacial pollen]

content of it

[distant turtle]

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• EasyGUI, by rejects Google Colab
• UVR5 NO UI for Google Colab, by Eddy Google Colab

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

[glacial pollen]

the voice should match but the content should be, ideally, new

or at least, much more diverse than the base

[analog obsidian]

Oh i believe this would also work in characters that have hours worth of talking

[glacial pollen]

Quite possibly

[crude flame]

Ight, will post my results here when it's done 🙂

[glacial pollen]

Neat

[knotty moth]

worth to try whenever I finish preparing my dataset

[glacial pollen]

yey. More testers the better.

[chrome orbit]

Im getting an error: 'FeatureInput' Object has no attribute 'device' when trying to extract features

whats the fix for this?

[glacial pollen]

are you choosing rmvpe_gpu during the step?

and foremostly, what's your gpu

[chrome orbit]

i chose crepe and have a 2060ti

[glacial pollen]

gpu's detected?
mentions running half precision / fp16 in the log?

[chrome orbit]

says running fp32

[analog obsidian]

just in case, did you tried this on applio?

[chrome orbit]

yea its applio rvc

[analog obsidian]

yeah crepe is bugged on applio

doesnt work atm

[chrome orbit]

ahh

[analog obsidian]

use codename's fork if u want to use mangio

[chrome orbit]

ight is mangio better than mainline

[glacial pollen]

There:
https://github.com/codename0og/codename-rvc-fork

Mangio is just crepe but allows you for adjustable hop

[chrome orbit]

ight cool

[analog obsidian]

nono, what are u using is called mangio-crepe

[glacial pollen]

You'll find more about it in the fork's faq tab

• some more technical info

if that's within your interest

[analog obsidian]

applio renamed it to just "crepe" for simplicity

[glacial pollen]

btw, I doubt Imma fully move to " new " fumiama's base

[analog obsidian]

yes pls

[glacial pollen]

the hash / encryption and such, is a pain in the ass

[analog obsidian]

aww

[glacial pollen]

will be all from there

excluding hashing

to me it's a nonsense and bs

[knotty moth]

it worked fine on 3.0.9/3.1.1 and the cpu threads set to 2

[glacial pollen]

it tries to read info from models despite most of em made on other rvc

and some of that, such as author or few others

aren't present

[analog obsidian]

yeah its bugged only in the recent versions

[glacial pollen]

and if they're not, it spills errors

and I can't make it cooperate with older mangio's fork code structure, the hash

[analog obsidian]

hey as long you can add some of the new changes of fumiama mainline is fine

[glacial pollen]

So most of the changes, including training part and few others, will be added

but Imma stick to the core of what I have

[analog obsidian]

ye codename has them

[analog obsidian]

there u go xD

[glacial pollen]

Yea, made quite a lot of these back in the day

was tired of constantly repeating myself lol

yeah sure, yw

[glacial pollen]

tho if I had to be honest with you

aside of some qol changes, code restructure and few differences here n there, there's not that much of new stuff

[analog obsidian]

sad

[glacial pollen]

so with that being said, gonna stick to new fumiama's structure

but most of what I deem necessary, in the core, will be kept as it is

[analog obsidian]

its fine

[glacial pollen]

✨

ofc, new features and all of that will be actively ported

[analog obsidian]

yay

[glacial pollen]

btw, there we go

example of why fp16 sometimes sucks

NaNs

[analog obsidian]

after fp32 gave me better results i stopped using fp16 haha

also i noticed that im not getting those NaNs at the start

after switching

totally worth waiting a bit more

[glacial pollen]

Yep, exactly what I meant

[tame mica]

codename rvc

[glacial pollen]

ye, I have my fork lol

[tame mica]

ah gl

[pallid patio]

Good evening

I have a doubt, what a "Strong Version" of a model is? Like, its just better vocals or smth?

[glacial pollen]

wdym " strong version "

there's no such a thing

[pallid patio]

In https://huggingface.co there's some versions "strong". Like Shawn Mendes, for example.

[glacial pollen]

Perhaps it's meant to represent the style of a particular voice

so like, strong vocal, soft vocal etc or maybe speech

cause as long you're referring to rvc models, there's no official or community accepted / agreed terms as " strong " version of model or such

so it must most likely refer to the model's speech / singing style it was trained on

[brittle wing]

Does anyone know where I can find that typical AI voice that’s used for Instagram philosophy quotes/ideologies?

[snow hazel]

How do I remove harmonies in UVR?

[brittle wing]

Or something like that, i serarch and try but got nothing

[rare goblet]

Ayo? @worn river level 1 !!!

[snow hazel]

How do I remove harmonies in UVR?

[modest abyss]

Could I get help with this?
Fixed the previous problem, but now this happens, and then crashes shortly after when trying to load Ilaria RVC Mainline

[split parcel]

I had a question, how can I make a voice model?

[pastel oak]

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

[pallid patio]

It makes sense, thanks so much for your help ❤️

[glacial pollen]

yw : D

[misty elk]

-colab

[azure marsh]

☁️ Google Colabs

• Applio, by IA Hispano Google Colab
• RVC Disconnected, by Kit Lemonfoot Google Colab
• RVC Mainline, by Hina Google Colab
• AICoverGen-WebUI, modded by Hina Google Colab
• AICoverGen-NoWebUI [English], by Ardha, fixed by Eddy, Hina and Gdr Google Colab
• AICoverGen-NoWebUI [Spanish], credits to Eddy, Hina and Gdr for translating and fixing Google Colab
• Ilaria RVC, by thestingerx Google Colab
• EasyGUI, by rejects Google Colab
• UVR5 NO UI for Google Colab, by Eddy Google Colab

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

[modest abyss]

I got it to work but MAN is this slow
Maybe I did something wrong but I'm only getting 1 epoch every 6 minutes, meaning it would take me 5 whole days to train this model!

[shut abyss]

I'm trying to use the voice changer but it says I don't have a gpu, is it because i have a radeon gpu?

I dont think it should matter too much since i have a pretty decent cpu

but my gpu is way better

[proper shale]

I think you may have installed the wrong version for your GPU

https://rentry.org/VoiceChangerGuide

Get the AMD version

[shut abyss]

ty ❤️

[proper shale]

You're welcome! Let me know if you face any issues :)

[shut abyss]

For the gpu options, should i choose gpu 0?

[rare goblet]

Ayo? @shut abyss level 1 !!!

[proper shale]

Most likely is your GPU, but check Task Manager > Performance tab, it should have the number there

[shut abyss]

ohhh, thank you

[proper shale]

You're welcome

[modest abyss]

Do you know how to help my issue?
I got the RVC Mainline and am trying to train, but it's taking forever, nearly 6 minutes per epoch. Do you know of any way to fix that?

[proper shale]

What's your GPU, current batch_size in training and how long is your dataset?

[modest abyss]

GPU is NVIDIA GeForce GTX 1060 6GB, current batch_size is 6, and my dataset is a little under 7 minutes long (was going to make it longer but wanted to use the shorter version for a test thinking it would go faster)

[glacial pollen]

Verify whether you see " half precision " or " is half " during rvc opening ( in log )

[proper shale]

You could probably cache dataset

[glacial pollen]

if it doesn't say anything of that sort / says fp32 / full precision
or anything like that

Won't do

[proper shale]

Tbh the 10xx cards are slow as heck sadly

Someone said it was because of the FP stuff 😭

GDR iirc

[glacial pollen]

That too but caching additionally shouldn't be done on less than 12 gig vram cards

[proper shale]

Ah

[glacial pollen]

esp for above 7-8 mins sets

aside, the gains aren't even that big tbh

[rare goblet]

Ayo? @glacial pollen level 17 !!!

[modest abyss]

I don't see that anywhere no

[glacial pollen]

oh yea, then that'd make sense

You're training in FP32 mode ( single precision / full ) better quality models but performance ( speed wise for training ) takes a hit

[wispy lodge]

1000 and 1600 series have horrible fp16 performance anyway. It shouldn't be used on them and I think rvc mainline disables fp16 for them

[glacial pollen]

From what I can read, gpus from that ti-

yea, was about to say

[modest abyss]

How do I change it? The last time I trained a model was about a year ago when it was as simple as importing the sound you wanted, pressing a button, and have a pretty accurate model in an hour

[glacial pollen]

it's already in fp32 mode

If you do not see half precision or anything like that in the log ( which you do not ) then it fallbacks to fp32, which is the case for you

tl;dr.: Just bite through it ~ it's worth it.

[shut abyss]

is it better to have the sensitivity threshold higher or lower?

[proper shale]

see what works better for u really

[modest abyss]

Guess I'll wait here for about a week for this to process then

[proper shale]

if it doesnt cut away your words

thats fine

[shut abyss]

I have it at max right now

[glacial pollen]

what's the amount of time per epoch for ya

[shut abyss]

but sometimes it will pick up noises of me like sighing etc

[glacial pollen]

if it's not 10-15 or 20 minutes, you won't wait weeks lol

chille

[proper shale]

enable sup2 as well

[modest abyss]

About 6 minutes, with the last one taking 7:20 and the previous ones taking about 5:30 each

[glacial pollen]

Then you're fine

fp32 for me on rtx 3060 ( 12 gig ) - 10 min set - 16 batch size takes 4-6 mins

(( but then, I am using cuda memory fallback so ))

But still, you're fine

[modest abyss]

Alright (the reason I say about a week is because I did the math and it would take at minimum 4 whole days to process the model at its current rate)

[rare goblet]

Ayo? @modest abyss level 3 !!!

[glacial pollen]

I mean, you're not blindly " Imma train for 1000 epochs to have the best model 😄 ! "

right?

[modest abyss]

I wasn't making that decision blindly, that's the settings I used back then when, like I said, it was as simple as just throwing an audio file at it and having a pretty damn good model in an hour

[glacial pollen]

Then it is blindly " calculated "

unless you're doing an exact same dataset, same parameters

[shut abyss]

is there any optimization for response time i can do while maintaining quality

[glacial pollen]

Each dataset is unique, each model is unique, each needs different hparams, each has own training curve and initialization
and each requires own tensorboard supervision

[shut abyss]

because its taking about 7000ms right now

[modest abyss]

It isn't the same exact dataset but it's the same length as most of my other datasets that only took an hour back then

[glacial pollen]

Like I said, this is not how it works my dude

RVC or Machine learning is not linear

you can't train X1 with similar data length and expect X2 to follow the pattern

in fact, in rvc, 1 slight hiccup in initialization ( during model's initial stage of training ) or a taking away of few secs from the set can drastically change the outcome

But anyways, you do you.
All I can say, your time per epoch is fine for fp32 and your hardware

[modest abyss]

Maybe I'm just getting confused then because of those older methods which seemed quite linear (not on the inside I know, but on the outside it was simple to calculate how long it would take)