✨│ai-help

[plain drift]

guys im having this issue

[astral jungle]

It doesn't work

[brittle wing]

is there a free website to train models? i dont want to use kits.ai

[azure marsh]

RVC Guides (How to Make AI Cover)

Documentation

🇺🇸 English (main)

https://docs.aihub.wtf/

​

Translation by country

🇧🇷 Brasil (PT-BR)

https://docs.aihub.wtf/v/brazil

🇫🇷 France

https://docs.aihub.wtf/v/france

🇩🇪 Germany

https://docs.aihub.wtf/v/germany

🇮🇹 Italy

https://docs.aihub.wtf/v/italy

🇯🇵 Japan

https://docs.aihub.wtf/v/japan

🇰🇷 Korea

https://docs.aihub.wtf/v/korea

🇵🇱 Poland

https://docs.aihub.wtf/v/poland

🇷🇺 Russia

https://docs.aihub.wtf/v/russia

[grand orbit]

guys

what to do with the mc its lagging so much

[rugged shuttle]

idk abt minecraft

[grand orbit]

not minecraft

[rare goblet]

Ayo? @grand orbit level 1 !!!

[grand orbit]

i mean its lagging the vc in discord and everywhere

like the sound not coming correctly and ms is so high

[tepid nimbus]

@low shard Sorry, didn't notice I'd accidently deleted the guide while deleting some other documents, SMH.

[low shard]

dont worry, anyways i tried to improve it, there was a part that told the viewers to go view some guides to isolate vocals but they werent working because the guides were my italian ones, so i adjusted the font and putted all english guides for isolation of audios

angetyde should just delete his original post of ilaria rvc, btw dont worry abt this its already fixed and pubblished

[tepid nimbus]

That's great. I didn't actually think of that since I already had some prior experience, good catch.

[daring flower]

what it this RVC

[worn echo]

How can I make laughing sound normal and not glitchy?

[rare goblet]

Ayo? @worn echo level 2 !!!

[plain drift]

hello I am having out of memory error on my rvc cmd

any optimised code or solution to this i got 4gb gpu and 16gs ram

[sand iris]

training on a 4GB gpu won't do unfortunately

[tranquil cliff]

hi guys....a new much more simplified colab will be released soon with added files and other files that download automatically

[astral jungle]

It worked

[astral perch]

Does a mixed dataset with singing and speaking result into a better or worse voice model

[glossy coyote]

-colab

[azure marsh]

Google Colabs

RVC

• AI-Cover-Gen-No-UI (inference only)
• RVC v2 Disconnected (training only)

🌐 Translated colabs

• PT-BR: -colabs_br

Audio Separation/Isolation

• MDX Net Colab
• UVR arch colab 5.0.2/c1.04
• KaraFan colab

[glossy coyote]

-rvc

[azure marsh]

RVC Guides (How to Make AI Cover)

Documentation

🇺🇸 English (main)

https://docs.aihub.wtf/

​

Translation by country

🇧🇷 Brasil (PT-BR)

https://docs.aihub.wtf/v/brazil

🇫🇷 France

https://docs.aihub.wtf/v/france

🇩🇪 Germany

https://docs.aihub.wtf/v/germany

🇮🇹 Italy

https://docs.aihub.wtf/v/italy

🇯🇵 Japan

https://docs.aihub.wtf/v/japan

🇰🇷 Korea

https://docs.aihub.wtf/v/korea

🇵🇱 Poland

https://docs.aihub.wtf/v/poland

🇷🇺 Russia

https://docs.aihub.wtf/v/russia

[viscid moss]

[glossy coyote]

oh my

[viscid moss]

what colab is that?

[glossy coyote]

CoverGen_No_UI.ipynb

i'll try using the huggingface one instead :/

[viscid moss]

try mi fixed version
https://colab.research.google.com/drive/1qezaBvCn4eSG1U2L4a64tEHTynbIu8eN?usp=sharing

Is in Spanish but is the same

[viscid moss]

works too but it's too slow

[glossy coyote]

ooooo thanks! ill try this too

[glossy coyote]

do i have to edit the dir_name or just leave it?
And for song_input, do I upload the audio to my google drive and copy & paste the path?;o

[viscid moss]

change the dir_name and.... works with a YT link

If you want to use a drive file, you must mount the cloud

create a code cell and paste this:

from google.colab import drive
drive.mount('/content/drive')

[glossy coyote]

oo okie thanks!

[rare goblet]

Ayo? @glossy coyote level 4 !!!

[glossy coyote]

I have this error for some reason D:
BadZipFile: File is not a zip file

[eternal finch]

hey i need some help

when i made a private RVC space for me in hugging face.... the screen is just blank white , no interface or buttons are visible.... i followed the group guide but this found myself stuck in middle due to this

[rare goblet]

Ayo? @eternal finch level 1 !!!

[eternal finch]

will appreciate if anyone can help

a screenshot for reference

[brittle wing]

Question

Does RVC split the audio if it's very long?

Nevermind, It does 👍

[brittle wing]

How do i find out how much to train my model, im using Mangio local RVC2 training, how many ephocs? and gpu blocks, ect, how do i find out that number

[quartz sky]

does anyone have a fork of AOD? the original github repository is down

[tawny hollow]

Hello people, is this possible to convert an audio file directly and how ?

[sleek vortex]

Is it normal that it takes only 1-2 seconds for one e-poch ? (in rvc v2 disconnected)

[tawdry ruin]

sorry, it's currently being reworked in #🔊│ai-development message. hopefully it'll be back soon

[fallow gull]

https://tenor.com/view/thanos-memoji-gif-23490017

fuck

[quartz sky]

i would still like to use the old release version, if thats possible...

may i ask him directly? or if there is an archive available i would like to use it

[tawdry ruin]

ig you can ping him yeah. the repository is not available anymore and I'm personally not aware of another way to download it soo

[quartz sky]

@brittle wing hello, do you have the old release version of AOD?

[rare goblet]

Ayo? @quartz sky level 1 !!!

[quartz sky]

why did you private the repository?

[agile oxide]

Why is my Res so high with RVC?

[frosty night]

someone help a mod out

[quartz sky]

i understand, but if you still have the old release version id like to use it

[agile oxide]

Trying to use the RVC in discord, but it keeps lagging out, are there any better ways to use it?

[quartz sky]

you deleted your repository,.,?

oh no >.>

[proper shale]

It's either your GPU or the wrong settings.

[agile oxide]

I'm pretty sure my GPU should be fine for it, are there any known optimal settings?

[proper shale]

What's your GPU btw

[agile oxide]

3080

[rare goblet]

Ayo? @agile oxide level 1 !!!

[proper shale]

So I can kinda give you a good chunk sise

Oh

I think 92 is good for you

[agile oxide]

i only have a 96 option

[proper shale]

i cant type lmao forgive me for it

oh that's it

And then RVMPE as your f0

it's the fastest really

[agile oxide]

what do you mean by this? sorry

[quartz sky]

what do you mean bat version? the one in releases or source

[agile oxide]

oh

found it

[frosty night]

someone help me w this

cant run easy gui

never tried before so idk

[proper shale]

f0 is basically just pitch detector

[proper shale]

could be that ur running from onedrive? idk

[agile oxide]

Is there also a better like mic system to use? I'm just using one I found online.

[frosty night]

oh

idk cause it says cannot import name I18nAuro from i18n

[quartz sky]

if you installed it in a virtual environment you should activate it first

[frosty night]

i tried reinstalling i18n

[frosty night]

what

[proper shale]

try moving it out of onedrive

[quartz sky]

unfortunate, ill be looking forward to it then...

[agile oxide]

Any ideas why it's still super choppy or laggy in VC?

[quartz sky]

im sure people here have a copy of it somewhere.. right?

[proper shale]

Have you tried like, 256 chunk size? Maybe that'll fix it

Sometimes low chunk size can bring choppiness

[agile oxide]

1 second

[proper shale]

It will have a considerable amount of delay tho, but hey, it won't be as choppy

Try out anything from like, 128 to 320, and see what works best for you in performance and quality

[agile oxide]

I'm looking at a video rn, and he's using 40 chunk size and it sounds perfectly fine.

[proper shale]

That's probably a 4090 then

[agile oxide]

Yeah, I think so.

[proper shale]

Either way, maybe a bigger chunk size will solve the issue

[brittle wing]

How do I find an AI better than the search thing in search-models

[quartz sky]

just go to weights.gg

[agile oxide]

@proper shale Here are some settings I have rn

[proper shale]

decrease that extra size

[agile oxide]

to what?

[quartz sky]

recommend lower extra size

[proper shale]

to like 4096

[agile oxide]

Yeah, idk it's still cutting out.

[quartz sky]

theres certain chunk sizes that dont work well with certain extra values

[rare goblet]

Ayo? @quartz sky level 2 !!!

[quartz sky]

idk why but its like that for me at least

[agile oxide]

I swapped to 96 with 4096

Are there some softwares that could help with it?

[quartz sky]

do 192 chunk, 32k or 65k

or 128

[agile oxide]

Still bad

[rare goblet]

Ayo? @agile oxide level 2 !!!

[proper shale]

that extra is way too much

[quartz sky]

it works pretty well for me

[proper shale]

Go for client audio tho

[quartz sky]

it picks up what youre saying better

[radiant laurel]

ok what is going on

[quartz sky]

i use this

[agile oxide]

I don't know it just keeps cutting still.

[quartz sky]

turn off

press stop

and start again and see if it makes any difference

[agile oxide]

i've tried

[sand iris]

are you running a game at the same time ?

[agile oxide]

minecraft in the back but it's not super intensive

[sand iris]

is the framerate unlocked ?

[agile oxide]

yeah it is

[sand iris]

Lock and see

[quartz sky]

could be just lacking memory honestly

[radiant laurel]

any help please

[agile oxide]

Still not working yeah, no idea.

[quartz sky]

do you have index on?

[stoic cloud]

hi

[proper shale]

...nothing?

its literally normal

[brittle wing]

guys help

[radiant laurel]

is that normal with mangio?

[stoic cloud]

If you have not pressed anything, yes

[proper shale]

yeah it's

it's normal

there's no traceback stuff

so you're good to go?

[radiant laurel]

I hope

[brittle wing]

how do i get the tensorboard

[quartz sky]

ahaha-

[agile oxide]

i did not no

[quartz sky]

#🔊│ai-development message

[quartz sky]

he deleted the repo with no backups so theres nothing rn...

[quartz sky]

then- idk what is wrong with it

[proper shale]

I think the all in one document has a little tutorial on it

#1159513888199540817

[brittle wing]

hmmm

i just want to know

how much to train it

and where does it save itself

[quartz sky]

the tutorial only tells to download the release so

[brittle wing]

is this ok?

[proper shale]

thats something tensorboard will "tell" you

[brittle wing]

that's why i wanted it

[brittle wing]

but is this good? @proper shale

[proper shale]

decrease that total epochs count because you will definitely not do that in 1 session

[brittle wing]

so like how much

[proper shale]

otherwise it's... ok

[brittle wing]

300?

[proper shale]

yeah that should be okay

[brittle wing]

i want it to be really good

[proper shale]

(hopefully)

[brittle wing]

and batch size per GPU??

4?

[proper shale]

if ur dataset is good, then even at 50 epochs it'll already sound decent

[proper shale]

idk, how much vram you have

[brittle wing]

16gb

[proper shale]

try uh

8

8 batch_size seems to be versatile so

[brittle wing]

its like

3332mhz

so like a really good vram

really fast

[proper shale]

no i meant

your gpu vram

[brittle wing]

oh

[proper shale]

this is RAM

[brittle wing]

hold on then

14gb

[proper shale]

nah you got 6 dedicated

yea 4 is fine

[brittle wing]

ALR

alr*

which options do i check?

[graceful obsidian]

check uncheck check rest seems like unecessary automated stuff

[proper shale]

^

[graceful obsidian]

and that mode collapse method won't save mode-collapse-prone datasets, will only delay it occuring

[brittle wing]

yo wtf

[rare goblet]

Ayo? @brittle wing level 5 !!!

[brittle wing]

its so fast

[graceful obsidian]

mode collapse is caused by either:
A) Dataset is too small + Dataset/samples to Silence ratio is higher towards silence.

B) There's way too much silence ( or even contaminated silence which doesn't get discarded ) in the dataset

[brittle wing]

bro the epochs literally count

its like

55

56

57

[graceful obsidian]

also, if you want best results, depending on the dataset, you should commission someone

[frosty night]

what do i install applio or normal rvc

[brittle wing]

so fucking fast

[graceful obsidian]

batch_size 4 is usually never used unless in some special circumstances

[brittle wing]

why tho?

[graceful obsidian]

lemme explain you briefly what it is

it takes your dataset and divides it by 4

then you get groups

and that many groups ( of samples 0 are being used during training

[frosty night]

-rvc

[azure marsh]

RVC Guides (How to Make AI Cover)

Documentation

🇺🇸 English (main)

https://docs.aihub.wtf/

​

Translation by country

🇧🇷 Brasil (PT-BR)

https://docs.aihub.wtf/v/brazil

🇫🇷 France

https://docs.aihub.wtf/v/france

🇩🇪 Germany

https://docs.aihub.wtf/v/germany

🇮🇹 Italy

https://docs.aihub.wtf/v/italy

🇯🇵 Japan

https://docs.aihub.wtf/v/japan

🇰🇷 Korea

https://docs.aihub.wtf/v/korea

🇵🇱 Poland

https://docs.aihub.wtf/v/poland

🇷🇺 Russia

https://docs.aihub.wtf/v/russia

[graceful obsidian]

sometimes the groups are too small for that one specific dataset

to get a good " avg / weighted " results for one dataset / voice

[brittle wing]

oh i understand you

[graceful obsidian]

on avg, 16 is go-2

if no good results are present, only then you either go lower; 14, 12, 10, 8
or add more data ( 1 min, then another 1 min and so on
it's a gradual process

[brittle wing]

ohhh

btw

one thing just doesent feel right..

the epochs are going too fast

[graceful obsidian]

what's your dataset size?

[brittle wing]

i started training 2 mins ago

[graceful obsidian]

if it's below 1-2 mins

[brittle wing]

and its already on 200

its going too fast

it just doesent feel right..

[graceful obsidian]

what's your dataset's length approx ?

[brittle wing]

30 mins

[graceful obsidian]

oh yeah, then that's def something off

[proper shale]

ooooh no

[brittle wing]

wdym

whats wrong

[graceful obsidian]

can you go into your model's folder

check the 0 gt folder in there

[brittle wing]

how

[graceful obsidian]

( it contains .wav sliced samples used for training )

[brittle wing]

can i stream it to you??

[graceful obsidian]

RVC's folder

and there's assets I think

or logs, and in there, your models' folders

not sure how newest rvc or so handles it

[proper shale]

its prolly training 1 epoch per step

[brittle wing]

https://discord.com/channels/1159260121998827560/1159273424053608498

[graceful obsidian]

gimme a sec

@brittle wing so

first

go into ur rvc's folder

now assets

hmmmm

weird structure, wait go back

ah cause that's applio

well, go root folder

it should be somewhere there

hmmm, I've no idea where applio stores models lmao

ye but those are pretrains

those are the " base " models

Got better idea, search for ur model's name in root folder

it's the main rvc's dir, well, applio's

in other words

a, there we go

check 0 gt folder

hmmm, samples are in there, so that's not the case then

[graceful obsidian]

@brittle wing I'd first check if it works

in applio

ah, so it's for training only?

a no, nevermind
ye, inference, you wanna test it there

Well, gonna move to unlimited's chat, I look like schizo here, hm 🤔
nevermind

@brittle wing check dm as well, provided a stress test audio

@brittle wing no that's not a rule

You should only train for as long tensorboard " tells you to "

because AI models have 2 states

overtrained and undertrained

nono, If you want it locally, I can give you something

nah, don't trust those misunderstood information
each model and dataset is individual

wait I'll just get a mic hooked up ig, brb

[heady fable]

what do i do if it sounds good in audacity but when i listen to it in an mp3 it sounds low quality

[rare goblet]

Ayo? @heady fable level 1 !!!

[proper shale]

export to wav

[heady fable]

thanks

[radiant laurel]

how long does it normally take with 45 mins of voice lines?

[woeful depot]

anyone can help? i trained rvc locally, but i dont have g & d pth file? it does appear the paths when i first open the gui, then i clicked on extracting features it disappears

should i reinstall my rvc?

[cursive shore]

I don't know which channel to use, so I'll copy here what I posted to #🔍│help-w-okada .

I'm getting really high delay after a few seconds of live voice-changing, as in every couple of seconds the delay grows by a second. I'm also getting a "stretching" sound as the output cuts in and out. I have the AI set to low-ish settings and an Nvidia 1660 Super. I'm not sure if I have the GPU set up correctly or how I can do that.
Is my GPU too weak for okada or is there some other problem?

[woeful depot]

or what way do you guys usually train?

i cant stand the fan sound

[proper shale]

Apparently 16 series of cards have problems with W-Okada. I suggest opening an issue on GitHub

[proper shale]

they should be in ur logs

[woeful depot]

the paths disappear, i cant enter the path, it will show errors

[proper shale]

wym by that?

gimme screenshots or video of it

[frosty night]

still doesnt wortkj

work

someone help

[woeful depot]

its chinese if you dont mind, i first opened the gui the paths appear, and then i extract voice features the paths disappear

and i enter manually it shows errors

so theres two blank spaces for these two path

I first trained locally and inference, the sound is terrible

[proper shale]

Oh those are the pretrained paths

[woeful depot]

yea i dont have that

[proper shale]

Huh

[woeful depot]

oh so i need to download the infer train ver?

[rare goblet]

Ayo? @woeful depot level 2 !!!

[woeful depot]

i never train locally before

[frosty night]

i modified the script but now it says

its literally missing files

tf am i supposed to do

literally had to copy my_utils from an old rvc version for it to actually stop giving an earlier error

[marble lantern]

rvc

hi..last colab link pls ?

[zenith cairn]

so... what's the difference between the "f0 methods" (dio, pm, crepe, harvest, etc.)

[frosty night]

just use rmvpe or mangio-crepe

i recommend rmvpe

mangio-crepe is less rough

[zenith cairn]

is there any documentation on the diffrences between them?

[frosty night]

idk how to explain it

the other methods r just older

[zenith cairn]

ah ok

[frosty night]

nobody uses em anymore

[rough peak]

how do u use a model

[proper shale]

#1159513888199540817

[buoyant saddle]

uh, just a quick question
Do the perfect amount of epochs for a dataset depend on the quality of the audio, or the time of the audio?

If the time, is there some sort of chart for what the perfect amount of epochs per minute or second?

[proper shale]

both

only the graph "tells" you how long you'll train for

[buoyant saddle]

oh, its because I usually use a dataset at about 3 minutes and 10 seconds at 600 epochs

[buoyant saddle]

the graph "tells" me is sort of a lie. I know I have to use trial and error, but it takes me like 2 hours to make a model and train it at 600 epochs, so no thanks.

[proper shale]

...wha

look, trial and error is good for like

batch_size

the graph does give you important info you should see to make the decision to keep training or not

[buoyant saddle]

how am I supposed to know when it is overtrained? The graph either flucuates randomly, or I am too zoomed out to see and noticible things

[proper shale]

because in the worst case scenario you can get a overtrained model, but most of the time you keep training it's still undertrained

[proper shale]

-overtrain

[azure marsh]

Hey, @buoyant saddle!

👇 Here are some resources to help you identify if your model is overtraining

How do I know if my model is overtaining?

All-In-One Guide on how to make a good model

This guide explains how the D and G files works and much more: https://rentry.org/RVC_making-models

Credits: LUSBERT

Automated Overtraining Detection (AOD)

Will be available soon in #1159513888199540817

Credits: grvyscale

[proper shale]

this might help

but basically

if g goes up and doesn't come back down after a good while of training, then it's most likely overtraining.

[buoyant saddle]

thank you

[rare goblet]

Ayo? @zenith cairn level 2 !!!

[proper shale]

np!

the model's zip link might be down?

[azure marsh]

RVC Guides (How to Make AI Cover)

Documentation

🇺🇸 English (main)

https://docs.aihub.wtf/

​

Translation by country

🇧🇷 Brasil (PT-BR)

https://docs.aihub.wtf/v/brazil

🇫🇷 France

https://docs.aihub.wtf/v/france

🇩🇪 Germany

https://docs.aihub.wtf/v/germany

🇮🇹 Italy

https://docs.aihub.wtf/v/italy

🇯🇵 Japan

https://docs.aihub.wtf/v/japan

🇰🇷 Korea

https://docs.aihub.wtf/v/korea

🇵🇱 Poland

https://docs.aihub.wtf/v/poland

🇷🇺 Russia

https://docs.aihub.wtf/v/russia

🎃 Spooky contest is active til 31st, check the announcement channel for further details.

[zenith cairn]

generally how long does ilaria rvc take?

[proper shale]

depends on the song's duration

[zenith cairn]

it's 4mins

[proper shale]

since it uses cpu it could take like, 400 secs for a 4 min file

[zenith cairn]

it's been going for 600

guess I have a weak cpu

[graceful obsidian]

if it's fluctuating too much then it's either ur dataset or batch_size being too small

[vapid gust]

what’s a good batch size

is there anyway to calculate that?

[high night]

16

[vapid gust]

For any dataset?

[high night]

Ask codename

But overall it is good

16

[graceful obsidian]

you can't calculate it, each dataset and so the model, is unique in that matter

as a rule of thumb, start with 16 and only decrease if the model despite good graphs, sounds bad or not as good
( provided it's not because of your dataset ) or if simply, graphs are too chaotic

so, 16 -> 14/12 -> 10/8

but typically, it's either 16, 12, or 8 going below 8 is super rare and most of the time not required

tl;dr, more than 5-7 mins of hq audio? try 16 or 12 if 16 failed
less than 5 mins? 16 and if that fails, try 8 or 6

overal, you want your graphs not too zigzaggy but neither too flat

[vapid gust]

I’ve got hq datasets only

maybe 5 mins

[graceful obsidian]

are they legit hq or " subjectively hq "

there's a big difference

[vapid gust]

yes it’s studio recordings

[graceful obsidian]

legit studio or stems

[vapid gust]

legit studio

[graceful obsidian]

then rest assured, you should freely try 16

tho, if you can provide a short sample

I can evaluate it for you spectrally

[vapid gust]

wav I assume

[graceful obsidian]

and do " rvc processing " simulation

yea, wave preferably

all raw

[vapid gust]

I’ll send it in dms is that fine

[graceful obsidian]

naturally

[vapid gust]

alr

[graceful obsidian]

I'll share visual results here tho

( screens )

[vapid gust]

all good

[graceful obsidian]

so it's informative for every1

[vapid gust]

yh

[graceful obsidian]

alr, lemme see

[vapid gust]

say less

[graceful obsidian]

first thing would be, some voices have assymetrical waveforms, phasing issues
sometimes it's the mic

relative to the middle

getting it right helps with compression / normalization steps
also helps the loudness / dynamics a lot

another thing would be mic thumps

rvc does handle a bit of that ( incl dc offset issues ) up til around 75hz

some exceed that and could be audible as bassy pops in models ( not always tho but worth noting, esp for low voices )

Other than that, my congrats

you're actually the first person here that stayed true to their claims of the audio being HQ

check passed

suitable for 48khz training ( training will be a bit more sensitive butttt, extra 4.1khz handling is worth it )

Definitely use batch 16 and only attempt 14 or 12 if the model doesn't come out any good

Now, I'd be a lil hypocritical of myself to " estimate " stuff but, from my experience, model shouldn't exceed 300-400 epochs zone

[high night]

Dann

Damn*

This guy out here

[graceful obsidian]

it is super hq and 5 mins so, pay attention to tensorboard and don't trust ckpts past 300 or 400

[high night]

Yeah same model shouldn't exceed 300-400

That's what I always do

[graceful obsidian]

it's more so that sometimes overtraining looks tricky on graphs

[high night]

Over training is weird

Thanks to rod

[graceful obsidian]

and ckpts that are 20-30% past the estimated threshold

[high night]

It helps

[graceful obsidian]

sometimes look like they aren't overfit

overtraining is just, well, lemme do a simulation

[high night]

Is audio engineer a job? It must be

[graceful obsidian]

quite a common ( if to, well, surrealistically assume there's no mode collapses ) scenario

zigzaggy due to either a bit lower than should be batch_size or just the nature of the set / or processing

but you can visualize the trending

so you'd want to pick the " somewhat stable " zone on the graph

and test those ckpts
key elements are:

• sibilants ( whether they glitch or not )
• volume / dynamics issues
• glitchy ( cutting / stuttering ) plosives
• lack of clarity or graininess, esp on breathing

but then, those graphs above are that " perfect scenario "

now, a bit more realistic with mode collapses

you wanna " ignore " collapses, thooo if you are suspicious, check them out too ( but don't get fooled )
and see the trending

again, these are example graphs. Each model has different scenario

this is a flat-lining graph
( too different from each other samples, bad processing, too much noise, not enough data compared to batch_size used or simply too much data )

an example of high / smooth batch_size scenario:

example of low / small batch_size scenario:

reason is, higher the batch_size, more samples are used per " update " of the model's parameters, having smoother and more full / more rich estimation

( values are for just demonstration, do not reflect how it's in reality )

higher batch_size = smoother
smaller batch_size = more random

decision depends on the situation. Sometimes one voice is so specific it requires less examples to show the model / AI / RVC sometimes needs more

[vapid gust]

what do u suggest then

are the audio files I sent good?

[dusk tendon]

Ok so I got myself a nice AMD GPU and someone here told me to use Applio so I got that set up and it doesn't give me any errors but it doesn't show the GPU index selection are in the "train" tab and when ever I train a model, it works just fine but when I try to use it with w-okada's RVC and it sounds nothing like the dataset I trained it on, and yes, w-okada's RVC is working for I tested a model from here and it worked fine.

[vapid gust]

nvm I read everything

[graceful obsidian]

Yes, they're perfect

Also, timestamp for me, for future

tensorboard 101

[vapid gust]

I’ve got a lot more like that

these are all drill songs so they tend to leak a lot

[vapid gust]

what does this mean 😭

[graceful obsidian]

#✨│ai-help message
@mighty parcel try to get it pinned or ask someone if you can please

this should def help majority of people
in terms of graphs, over and undertraining

[vapid gust]

yea thanks

[lament eagle]

so in order to make it good, you have to do a little editing like using compressor and EQ or smth for the dataset

[vapid gust]

are you still going to be online for a while v

im going to make that model soon

[graceful obsidian]

so I can redirect people to this

pretty much

[graceful obsidian]

compressing the peaks

becauseeeee

rvc won't care about peaks, it'll normalize all to -2 dB or so

risk of clipping
and lack of consistency in samples

[lament eagle]

yeahhhh that looks bad

[graceful obsidian]

best is to compress peaks + normalize it manually to -3 dB

and turn off rvc's normalization ( but with that I can't help, people use newer forks, mine's based on older variant )

until I update my stuff, that's that

[lament eagle]

There is no option to turn off the RVC's normalization

sadly

[graceful obsidian]

so in this case:

1. clean the audio
2. compress
3. normalize to -4 or -5 dB + denoise ( as noise is amplified )
4. feed it rvc ( it'll hopefully and safely get it to -2 / -3

[graceful obsidian]

yea it's done manually

in preprocessing script

but lots of stuff changed and so did the file structure in newer stuff

once I am done with upgrading, will possibly share the file if one wanted
( in 3-4 days, waiting for saturn's compute hours to refresh )

hmmm...

gonna condense all info into a google doc overview lel

[lament eagle]

Yeah you should do that

[graceful obsidian]

discord screns tho, too lazy for full guide

one day tho, for sure

but it def won't be noob friendly

sadly

some hardcore audio stuff lmao

[lament eagle]

Yeah

I don't understand like 70% of what you're saying 😭

[graceful obsidian]

hmmmm, if you have any questions, ask right away

oh

[lament eagle]

but honestly, better than nothing

[lament eagle]

Okay I have question

[graceful obsidian]

I can try to simplify it, just say or specify what exactly you gotta know

if something's unclear, remind me

[lament eagle]

Does target_sample_rate causes pitch too

like

[graceful obsidian]

target as in, resampling in rvc?

or you mean pretrains' samplerate

[lament eagle]

these are the same training I did, just with different sample rate, it sounds bad in 40k

wait\

fuck i think that's the wornfg

[graceful obsidian]

that is because

your dataset is approx 16/18 khz

so anywhere from 32 to 36 maybe 38

uvr output based?

[lament eagle]

THis is the REAL 40k

[lament eagle]

Yeah

[graceful obsidian]

yeah so here's the thing

[lament eagle]

I basically use UVR

[graceful obsidian]

rvc's samplerate isn't for resampling really

or upsampling

it is just to match ur dataset

because it'll handle that specific or approx sample rate the best

by training, say, 48khz model on 35khz data

there's 13khz of frequencies gone

rvc won't create it

only smooth it out a lil bit, maybe tiny tiny bit of upsampling

but that's that

anything below 37/38khz is 32khz pretrains / model

stuff that are above 40, so, standard 44.1 are both 40 or 48 applicable

[old wagon]

ok uhh imma import the question i asked in general cus am peepeepoopoocaca stoopid

[rare goblet]

Ayo? @old wagon level 1 !!!

[graceful obsidian]

i'd go for 48 pretrains in that situation

as 48 will handle those .4.1khz that 40 can't and missing .3.9khz dw about it

[old wagon]

how do i separate vocals if they are the same person
for example yoru ni kakeru
the background vocals are on different keys, but they are the same person

[graceful obsidian]

this is a 44/48 fusion model for instance ( 70% of data is 44.1 )

https://www.youtube.com/watch?v=YVh1o_glnYA

made on 48 pretrains

[graceful obsidian]

that's what we call harmonies

either manually in rx ( not recommended for unexperienced people )

or via UVR's VR arch models
bve ones ( iirc backing vocal excluded )

but prior to that, prepare mdx 23c's output

[lament eagle]

I used 6HP UVR

[devout haven]

are the models here for voice changers

i stupid idk how to use dis stuff

[graceful obsidian]

models are working on w-okada / voice changer and rvc

[dusk tendon]

anyone know why

[graceful obsidian]

use official rvc

it supports directml fully

( as long you don't have 4gb card, it should go fine, at least with index )

if applio doesn't work for you, that is

[old wagon]

i usually use online vocal removers cus my pc sux (cant handle uvr)

[devout haven]

can someone send me a video link or something on how to use/apply the voice

[rare goblet]

Ayo? @devout haven level 1 !!!

[dusk tendon]

are you talking about https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI

[rare goblet]

Ayo? @dusk tendon level 1 !!!

[devout haven]

i just installed the files and idk what to do next

[graceful obsidian]

https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/RVC1006AMD_Intel.7z

amd / intel variant, yes, official
can be glitchy ( or at least was on my end ) but worth trying

[old wagon]

oh okay

[dusk tendon]

I'll give it a try

[graceful obsidian]

I can screenshare for you the process if you wish

if adding voices ( models ) to rvc and using them ( doing inferences ) is all you need to know

[devout haven]

ok

[graceful obsidian]

gimme a sec

max 5

[dusk tendon]

So does this work on windows or do I needa use a linux distro?

[graceful obsidian]

win

[dusk tendon]

alright

[graceful obsidian]

https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/RVC1006Nvidia.7z

[dusk tendon]

(as I refered to but ima try and train it and see if it says "GPU not detected, falling back to CPU " again

ah

alr

[vapid gust]

@graceful obsidian What’s the minimum dataset you think should be used if you have studio stems

[graceful obsidian]

hq? reasonably I'd go for 5-6 mins

for optimal results, 10 mins

minimum perhaps 3-4

[vapid gust]

also should I need to remove silence?

[brittle wing]

5 days of raw studio acapellas 64bit/384khz are needed per model

[graceful obsidian]

trim it so it's at least 0.3-0.5s between each phrase

and make sure it's absl. mute, so enveloped preferably

[vapid gust]

any software that will do it for me

• manual trimming

[graceful obsidian]

perhaps noise gating

I think kalo did a guide on that one specifically

https://docs.google.com/document/d/13ebnzmeEBc6uzYCMt-QVFQk-whVrK4zw8k7_Lw3Bv_A/edit#heading=h.axhflt6fdph2

[vapid gust]

when I mean like removing silence I mean like so the words are together correct?

[graceful obsidian]

there's a step for that

[vapid gust]

is that what ur on too

[graceful obsidian]

however, I'd always keep it manual for full control

[vapid gust]

yea it’s just a pain to do in fl studio

[rare goblet]

Ayo? @vapid gust level 9 !!!

[vapid gust]

ye to remove the green part

[graceful obsidian]

so it's not grainy ( 2nd silence zone )

envelope actually

having it - infinity dB

[vapid gust]

so basically removing any silence

simplified

[graceful obsidian]

so it's not like this

but this

[river gate]

uh-

heLP-

Temporary folder already found. Wiping...
Archive:  /content/drive/MyDrive/rvcDisconnected/zipfile.zip
   creating: /content/temp_dataset/zipfile/ariTestUnit/
  inflating: /content/temp_dataset/zipfile/ariTestUnit/samp1.wav  
  inflating: /content/temp_dataset/zipfile/ariTestUnit/samp2.wav  
  inflating: /content/temp_dataset/zipfile/ariTestUnit/samp3.wav  
  inflating: /content/temp_dataset/zipfile/ariTestUnit/samp4.wav  
  inflating: /content/temp_dataset/zipfile/ariTestUnit/samp5.wav  
  inflating: /content/temp_dataset/zipfile/ariTestUnit/samp6.wav  
Sanitizing...
Dataset Type: Multispeaker
---------------------------------------------------------------------------
FileNotFoundError                         Traceback (most recent call last)
/usr/lib/python3.10/shutil.py in move(src, dst, copy_function)
    815     try:
--> 816         os.rename(src, real_dst)
    817     except OSError:

FileNotFoundError: [Errno 2] No such file or directory: '/content/temp_dataset/zipfile/ariTestUnit' -> '/content/dataset/ariTestUnit'

During handling of the above exception, another exception occurred:

FileNotFoundError                         Traceback (most recent call last)
3 frames
/usr/lib/python3.10/shutil.py in copyfile(src, dst, follow_symlinks)
    252         os.symlink(os.readlink(src), dst)
    253     else:
--> 254         with open(src, 'rb') as fsrc:
    255             try:
    256                 with open(dst, 'wb') as fdst:

FileNotFoundError: [Errno 2] No such file or directory: '/content/temp_dataset/zipfile/ariTestUnit'```

[rare goblet]

Ayo? @river gate level 1 !!!

[vapid gust]

alright got it

thanks

[graceful obsidian]

but then, if you wanna go automated which I don't recommend as it's less accurate

kalo's guide will help

https://docs.google.com/document/d/13ebnzmeEBc6uzYCMt-QVFQk-whVrK4zw8k7_Lw3Bv_A/edit#heading=h.axhflt6fdph2
reminder link

[vapid gust]

kk

and also one more thing

[graceful obsidian]

ye?

[vapid gust]

knowing that I got studio stems

what would be the best settings here

I use rvc disconnected bcuz my pcs shit

[graceful obsidian]

leave it as it is

[river gate]

[vapid gust]

Snm

bet

[graceful obsidian]

in this case it's already all set, ye

[vapid gust]

crepe hop length?

[graceful obsidian]

keep 64

[vapid gust]

what happens if I lower it

[graceful obsidian]

it is a

well

rms detection window

smaller = more accurate but also sensitive

[vapid gust]

in robloz terms?

[graceful obsidian]

going below 64 is not recommended because

in other words, keep it 64 because higher is for worse audios and older methods

[vapid gust]

ah ok

[graceful obsidian]

and lower is for highly vibrato / wobbly voices ( and for older methods

[vapid gust]

alr alr

training on colab got way harder

I gotta use Firefox for tensorbord

pain

[graceful obsidian]

that's why I made my port for saturn

colab's been crap since good 3-4 months

or so

[vapid gust]

is urs easier?

[graceful obsidian]

depends, but it's private due to prev abuses

[vapid gust]

ah

alr I’ll jus continue w rvc disconnected

[graceful obsidian]

that's the only way as of now

[vapid gust]

ye

I know how to use tensorbord properly cuz u taught me like a wile back

[graceful obsidian]

https://www.python.org/ftp/python/3.10.6/python-3.10.6-amd64.exe

[river gate]

......................................................i aint gonna say anything

[graceful obsidian]

so, what happened?

[river gate]

@graceful obsidian

[graceful obsidian]

what are you actually doing?

extracting the dataset or wut

[river gate]

Preprocessing

Specifically the Load Dataset part

[graceful obsidian]

weird stuff going on

[rare goblet]

Ayo? @graceful obsidian level 14 !!!

[graceful obsidian]

well, I'm clueless on that specific error

you should perhaps ask disconnected's maintainers

or try to restart it all + reupload the zip

[river gate]

okay i restarted it all, reuploaded the zip, andddddd same error

[sullen marsh]

does anyone know if there's a site where i can use elevenlabs but enter in my api key to bypass that vpn detection thing

[rare goblet]

Ayo? @sullen marsh level 1 !!!

[astral jungle]

Heloo

[main patio]

Is the Google Collab for making AI covers usable now?

[wooden lotus]

Is there any quick tensorboard to view my train logs
I don't want to run the whole colab again

[violet heron]

Don’t use colab for interference

Anything else should work

Colab should only be used to train

Use spaces

[lament eagle]

It's been 2 hours holy shit

is it really this slow?

[wooden lotus]

Idk

[vapid gust]

@graceful obsidian \

[graceful obsidian]

mode collapses

[uneven elm]

is splitting of the acapella necessary? what tool can i use

[vapid gust]

huh

[graceful obsidian]

ignore them for now

you can see the shadow behind smoothed graph

[vapid gust]

its like 3.5 minutes

[graceful obsidian]

it has low low low spikes

[vapid gust]

i js checked

[vapid gust]

yeah

meaning?

[graceful obsidian]

3.5 could be potentially not enough

but u never know

[vapid gust]

ah

[graceful obsidian]

those are mode collapses

[vapid gust]

oh

[brittle wing]

how can i load a model?

[graceful obsidian]

ignore them unless final model is bad regardless of which ckpt u test

[vapid gust]

ive got a good accurate model of 2 minutes

once

if u would like to see

[graceful obsidian]

actually I'll be heading to sleep

[vapid gust]

alright

[graceful obsidian]

but ye I know, sometimes it's possible to get em decent

just try to experiment and test em around

[vapid gust]

yeah okok

thx

[graceful obsidian]

model .pth goes into rvc folder/assets/weights

index goes into rvc folder/logs/here or rvc folder/logs/Make_ur_model_folder/here
( up to u to pick the scheme for this one )

I go to sleep. Take care y'all n gluck

[brittle wing]

♥

[wooden lotus]

I used rvc disconnected and it finished training,how do I save the model

[coarse sun]

how do i get it to work on a steam deck?

[rare goblet]

Ayo? @coarse sun level 1 !!!

[severe juniper]

Can anyone send ROD zip, plz? Cuz github link doesn't work.

[spare saffron]

How exactly do you convert pre-existing pth files to ONNX?

[tired rapids]

@daring verge

just like that?

[daring verge]

your supposed to put the name of your zipfile

[tired rapids]

I have 5 audio recordings

[daring verge]

umm, you're supposed to zip file your recordings together into 1

[tired rapids]

oooohh

[tired rapids]

So I zipped it into one

should I just put zip

like this?

[daring verge]

"zip.zip" the same as your file name

[tired rapids]

so like this?

[daring verge]

yes

now run it

[tired rapids]

this?

[daring verge]

yeah like that

[tired rapids]

do I run training after that?

[daring verge]

do you know how epoch works?

[tired rapids]

I used to do the old one with gradio

I just used epochs around 250 and it all worked fine

[daring verge]

alright i see

you can run it

[tired rapids]

only the training right?

[daring verge]

yep

[rare goblet]

Ayo? @daring verge level 24 !!!

[tired rapids]

okayy, thank you so much

[low shard]

btw for epochs, in pinned messages theres a tabel with the suggested epochs based on the dataset lenght, its not like a rule to follow but js the suggested ones,it could ofc depend tho

[tired rapids]

honestly, I don't know what epochs really do. I just put 250 because it is what I usually do and all works out fine

[low shard]

its like how much the ai will train with the dataset, but the amount should change based on the dataset tho,depends

[tired rapids]

I'm done training it

but how do I download it?

[daring verge]

run the export model cell

and also index training

[tired rapids]

I've ran both, what now?

[rare goblet]

Ayo? @tired rapids level 3 !!!

[tired rapids]

oh

thankss

I found it

[potent root]

hi

anyone needing help with smth rn?

[tired rapids]

I made a voice model, but I don't know how to make another one and it keeps giving me errors

[daring verge]

you don't need to run clone respiratories if you already run the first time

since, its already downloaded to your google drive, you just need to run install dependencies

[grand orbit]

guys im having in issue with the voice like they can notice that im using a voice changer i dont know is it because of the models or there is a way to make sound more real

[daring verge]

try using a different voice model

[grand orbit]

any recommended once or pupolar model , i have try more than 5 until now i had the same issue with it

[tired rapids]

I did that, but it won't work

for training

[daring verge]

did you run preprocess?

[tired rapids]

yes

[daring verge]

its missing some files, try running preprocess again

[tired rapids]

I re did everything again, but it still won't work

[daring verge]

lemme see the load dataset section

[daring verge]

ok now i see the problem

delete the folder and Mangio-RVC-Fork folder and run the clone respiratories cell again

[tired rapids]

what folder?

the rvcDisconnected?

in my drive?

[daring verge]

no not that

open Show file browser on your colab

[rare goblet]

Ayo? @ancient jewel level 1 !!!

[tired rapids]

I did the manage sessions and terminate sessions

is that ok?

okay, thankss.

[bold salmon]

Hey guys can anyone give me link to Ai cover

[glad zealot]

what browser are you using btw

[tired rapids]

opera gx

[glad zealot]

noice thanks

[lime dome]

can someone help me with the colab

idk how to do

[low shard]

which one?

[lime dome]

.

[glad zealot]

what kind of help ya need?

[low shard]

i would suggest u to use hugging face instead, its slower but theres no risk and isnt limitless

https://discord.com/channels/1159260121998827560/1167844377046040769 this is better than the AICoverGen Colab

you cant get banned or disconnected with this

[lime dome]

okay ill try it when im free

[low shard]

i would suggest the hugging face version than the colab man, just so people dont have to risk account or get disconnected, the aicovergen is not really that stable

[glad zealot]

seems stable to me

never got a crash or any error since i started using it

[low shard]

but it has limits bc of the colab gpu

the hugging face one can be used all day without having to worry abt ban or anything

[glad zealot]

ye thats why theres a warning on the colab one

[lime dome]

wait guys

is these colabs real-time voice changer?

or cover

[glad zealot]

thats cover

[lime dome]

oh

[glad zealot]

https://colab.research.google.com/github/hinabl/voice-changer-colab/blob/master/Hina_Modified_Realtime_Voice_Changer_on_Colab.ipynb

this is for the voice changer colab

[lime dome]

is it real-time

[glad zealot]

yup

well more like 1-ish sec of delay

[lime dome]

okay ill tag u when im trying to do it

because idk anything about it

[glad zealot]

imma go eat...

https://rentry.co/VoiceChangerGuide#gpu-chart-for-known-working-chunkextra will give you settings that should work 99% of the time.

If you are on AMD or INTEL ARC make sure your voice models are in ONNX format, refer to https://rentry.co/W-Okada-FAQ#the-default-voices-are-fine-but-ones-i-upload-arent it should guide you through it.

[lime dome]

same

[glad zealot]

ignore the 2nd part

[lime dome]

ok thx

[neat tendon]

Hi I am facing an issue where my virtual audio cables have been interchanged , like the input is in the output section and the output is in the input section . I tried reinstalling the drivers, didn't work! Anyone got something?

[glad zealot]

that is correct tho

in audio inputs it you need to put the cable output

[neat tendon]

Got it thanks

[vapid gust]

Is it possible to resume training after closing the tab?

Im on rvc disconected

[glad zealot]

yes but i personally havent tried it

[vapid gust]

how do I do it

[lime dome]

@glad zealot

[glad zealot]

mew

[lime dome]

i even dont understand this

[glad zealot]

lewl

[lime dome]

😭

[glad zealot]

are you runing it rn?

[lime dome]

i mean idk what to do

[ember crow]

DAMN IT

[lime dome]

i only need to copy the settings on the 1. link u sent or do something with colab?

[glad zealot]

yup

mostly ye

[lime dome]

okay but

what i will do with colab

[glad zealot]

just run the cells

and on this part you can just ignore it if you want

[lime dome]

without doing anything?

[glad zealot]

on this part you need to put your ngrok token

[lime dome]

only that?

[glad zealot]

yup

[lime dome]

okay

and

i will run every cell?

[glad zealot]

ye

[lime dome]

okay

[lime dome]

what this do?

[rare goblet]

Ayo? @lime dome level 3 !!!

[glad zealot]

it downloads a model from huggingface

[lime dome]

to my voicechanger?

[glad zealot]

ya

[lime dome]

oh

okay

[glad zealot]

if you have the model download on your computer you can also just ignore that and upload it later

[lime dome]

should i change the settings on my voice changer program?

[glad zealot]

yup

depends on your voice

[lime dome]

before or after?

[vapid gust]

fatal: destination path 'Mangio-RVC-Tweaks' already exists and is not an empty directory.

---

OSError Traceback (most recent call last)

<ipython-input-20-bfb8a4688ddf> in <cell line: 13>()
11 get_ipython().system('git clone -b pr-optimization --single-branch https://github.com/alexlnkp/Mangio-RVC-Tweaks.git')
12 #Rename to keep backwards compatibility with old variants of Disconnected
---> 13 os.rename("/content/Mangio-RVC-Tweaks", "/content/Mangio-RVC-Fork")
14 get_ipython().system('git clone https://github.com/maxrmorrison/torchcrepe.git')
15 get_ipython().system('mv torchcrepe/torchcrepe Mangio-RVC-Fork/')

OSError: [Errno 39] Directory not empty: '/content/Mangio-RVC-Tweaks' -> '/content/Mangio-RVC-Fork'

i got this error

[glad zealot]

after

[lime dome]

i will run cells before opening voice changer or after?

[glad zealot]

mean that the mangio-rvc-fork folder already exists

[glad zealot]

you cant open the voice changer without running the cells

just run the cells till the last one

[lime dome]

but i have it downloaded

[lime dome]

last one is ngrok one?

[glad zealot]

ye

[vapid gust]

i dont have anytrhing like that

[lime dome]

@glad zealot

[glad zealot]

on your pc?

[lime dome]

yes

[glad zealot]

then why do you need the colab one?

[lime dome]

it doesnt work

[glad zealot]

then just ignore your local

[lime dome]

wdym

[glad zealot]

i have no idea how to fix local ngl so ye

cant really help with local stuf

[lime dome]

other mods said ''ur pc is old to run it''

''u can try colab''

[glad zealot]

ye

on colab just run literally all the cells needed that are not optional

[lime dome]

okay

[glad zealot]

thats literally it

[vapid gust]

fatal: destination path 'Mangio-RVC-Tweaks' already exists and is not an empty directory.

---

OSError Traceback (most recent call last)

<ipython-input-28-bfb8a4688ddf> in <cell line: 13>()
11 get_ipython().system('git clone -b pr-optimization --single-branch https://github.com/alexlnkp/Mangio-RVC-Tweaks.git')
12 #Rename to keep backwards compatibility with old variants of Disconnected
---> 13 os.rename("/content/Mangio-RVC-Tweaks", "/content/Mangio-RVC-Fork")
14 get_ipython().system('git clone https://github.com/maxrmorrison/torchcrepe.git')
15 get_ipython().system('mv torchcrepe/torchcrepe Mangio-RVC-Fork/')

OSError: [Errno 39] Directory not empty: '/content/Mangio-RVC-Tweaks' -> '/content/Mangio-RVC-Fork'

i literally dont have anything like this

in my drive

[glad zealot]

OSError: [Errno 39] Directory not empty: '/content/Mangio-RVC-Tweaks' -> '/content/Mangio-RVC-Fork'

[vapid gust]

i dont have anything like that in my drive

my drive is empty

i only got my datsaset in it

[glad zealot]

its on the colab storage

[vapid gust]

it says i cannot delete it because it is not empty

[glad zealot]

add a new code cell

[vapid gust]

how

[glad zealot]

!rm -rf /content/Mangio-RVC-Tweaks```

[glad zealot]

hover betweeen 2 cells

[vapid gust]

got it

thanx

---

FileNotFoundError Traceback (most recent call last)

<ipython-input-36-10dc9864c428> in <cell line: 5>()
4
5 if not os.path.isdir("csvdb/"):
----> 6 os.makedirs("csvdb")
7 frmnt, stp = open("csvdb/formanting.csv", "w", newline=""), open("csvdb/stop.csv", "w", newline="")
8 csv_writer = csv.writer(frmnt, delimiter=",")

/usr/lib/python3.10/os.py in makedirs(name, mode, exist_ok)
223 return
224 try:
--> 225 mkdir(name, mode)
226 except OSError:
227 # Cannot rely on checking for EEXIST, since the operating system

FileNotFoundError: [Errno 2] No such file or directory: 'csvdb'

SETUP CSVBD

[lime dome]

okay what i will do now

[glad zealot]

just end it and start from the beggining

you need to move to specific locations if you continue it

[lime dome]

@glad zealot

[vapid gust]

Updating and installing system packages...
Installing build-essential...
Installing python3-dev...
Installing ffmpeg...
Installing aria2...
Updating and installing pip packages...

---

CalledProcessError Traceback (most recent call last)

<ipython-input-38-46706f3aac76> in <cell line: 12>()
10
11 print("Updating and installing pip packages...")
---> 12 subprocess.check_call(['pip', 'install', '--upgrade'] + pip_packages)
13
14 print('Packages up to date.')

/usr/lib/python3.10/subprocess.py in check_call(*popenargs, **kwargs)
367 if cmd is None:
368 cmd = popenargs[0]
--> 369 raise CalledProcessError(retcode, cmd)
370 return 0
371

CalledProcessError: Command '['pip', 'install', '--upgrade', 'pip', 'setuptools', 'wheel', 'httpx==0.23.0', 'faiss-gpu', 'fairseq', 'ffmpeg', 'ffmpeg-python', 'praat-parselmouth', 'pyworld', 'numpy==1.23.5', 'numba==0.56.4', 'librosa==0.9.2', 'gdown', 'onnxruntime']' returned non-zero exit status 1.

[glad zealot]

it should give you a link in the last cell

[vapid gust]

Updating and installing system packages...
Installing build-essential...
Installing python3-dev...
Installing ffmpeg...
Installing aria2...
Updating and installing pip packages...

---

CalledProcessError Traceback (most recent call last)

<ipython-input-40-46706f3aac76> in <cell line: 12>()
10
11 print("Updating and installing pip packages...")
---> 12 subprocess.check_call(['pip', 'install', '--upgrade'] + pip_packages)
13
14 print('Packages up to date.')

/usr/lib/python3.10/subprocess.py in check_call(*popenargs, **kwargs)
367 if cmd is None:
368 cmd = popenargs[0]
--> 369 raise CalledProcessError(retcode, cmd)
370 return 0
371

CalledProcessError: Command '['pip', 'install', '--upgrade', 'pip', 'setuptools', 'wheel', 'httpx==0.23.0', 'faiss-gpu', 'fairseq', 'ffmpeg', 'ffmpeg-python', 'praat-parselmouth', 'pyworld', 'numpy==1.23.5', 'numba==0.56.4', 'librosa==0.9.2', 'gdown', 'onnxruntime']' returned non-zero exit status 1.

[lime dome]

i didnt run the last cell yet

should i?

[glad zealot]

yes

[lime dome]

okay

[vapid gust]

Updating and installing system packages...
Installing build-essential...
Installing python3-dev...
Installing ffmpeg...
Installing aria2...
Updating and installing pip packages...

---

CalledProcessError Traceback (most recent call last)

<ipython-input-40-46706f3aac76> in <cell line: 12>()
10
11 print("Updating and installing pip packages...")
---> 12 subprocess.check_call(['pip', 'install', '--upgrade'] + pip_packages)
13
14 print('Packages up to date.')

/usr/lib/python3.10/subprocess.py in check_call(*popenargs, **kwargs)
367 if cmd is None:
368 cmd = popenargs[0]
--> 369 raise CalledProcessError(retcode, cmd)
370 return 0
371

CalledProcessError: Command '['pip', 'install', '--upgrade', 'pip', 'setuptools', 'wheel', 'httpx==0.23.0', 'faiss-gpu', 'fairseq', 'ffmpeg', 'ffmpeg-python', 'praat-parselmouth', 'pyworld', 'numpy==1.23.5', 'numba==0.56.4', 'librosa==0.9.2', 'gdown', 'onnxruntime']' returned non-zero exit status 1.

dependies

the first

[glad zealot]

i needs to be !pip

can you give me the link for that?

[lime dome]

@glad zealot i did it yay

should i make the same settings at the other site?

[glad zealot]

for now i guess? thats raven's settings

[lime dome]

okay ill try

thank you

[lime dome]

it doesnt change my voice..

i tried it

[glad zealot]

can you send your settings?

this is my setting

[dire vapor]

how to set low batch in applio

[lime dome]

okay

i refreshed page by accident

i need to open it again

1 minute

[lime dome]

ill try

[rare goblet]

Ayo? @lime dome level 4 !!!

[lime dome]

what did u do the f0 det?

[glad zealot]

rmpve

[lime dome]

okay lemme try

thanks

it worked

it has a bit delay but

it works

but

its

uhh

sound comes and goes

ok i fixed it

[vapid gust]

@graceful obsidian

[graceful obsidian]

yea, you should learn how to check these out

I can't really be going around and doing it for people, like I include all the info on such things everywhere on the server

aaaand, am currently busy with new saturn port

[vapid gust]

yeah ik

I can already tell it’s not overtraining

I trained this once last night but it closed for some reason so I lost all that progress

but I went to 500 epochs and the graph was still heading down with never a spike up

[graceful obsidian]

because you have it smoothed out

set smoothing to 0 or 0.2 and you'll see more

[vapid gust]

how muhc u want me to keep it at

[graceful obsidian]

than that

[vapid gust]

0.2

[graceful obsidian]

turn on ignore outliers and middle icon / button

under the graph

[graceful obsidian]

you gotta select that middle icon under the graph

as I said

it's for scaling

[vapid gust]

there

at 295 epochs

[graceful obsidian]

aside of the sort of flat-lining behaviour

normally you'd, most likely, search around here

[vapid gust]

soo

[graceful obsidian]

buttttt

it is not synced

with steps so

the graph's almost unusable

[vapid gust]

what do i do

does that mean the model is gonna be bad?

[graceful obsidian]

no, it means you can't effectively evaluate the ckpts based on that graph

[vapid gust]

oh

ok

[graceful obsidian]

maybe ask @ lusbert

[vapid gust]

im gonna train another 30+ minute

[graceful obsidian]

on how to sync the graph

[vapid gust]

studio stems

[graceful obsidian]

without it, you can forget about tensor

as for me, I am too busy to explain it atm, that is

[vapid gust]

alr

ill just wait for the rise

then

[graceful obsidian]

well, as i said, it doesn't matter, rise or not

whatever you'll see is inaccurate

[vapid gust]

o

[graceful obsidian]

could be from 20 to 60% inaccurate

without syncing