🧬│ai-chat

[slow vapor]

I think this is the part a lot of creators struggle with. Hearing highly subtle differences in how words are spoken/enunciated and having the right tones behind that

Takes a good ear, as you said

[ancient swan]

yep

took me a good year to start hearing shit that i haven't been hearing before lmao

[slow vapor]

XD that’s fair. I make music and it took me years to gain the knowledge I have now, and hear things I never did before

[ancient swan]

same, took me at least 4 years to understand how to write good music lmao

[slow vapor]

Oh I didn’t know you did too

Wanna hear a little sample of something I did?

[ancient swan]

sure, but post only yt links, audio isn't allowed cus of copyright (even if it's your own music, it's just easier that way)

sc links are also fine

[hallow portal]

so how because I watched a lot of such modifications on YouTube

[ancient swan]

or pillowcase

[slow vapor]

I don’t have this on yt, but I have a box link. Reputable Audio/file hosting service

[ancient swan]

aight

[slow vapor]

https://app.box.com/s/72d0vlyihv4t9snx8mer34labb0oe0q3

Lots of emotion poured into that

[ancient swan]

sounds dreamy

[slow vapor]

Noeone ever put it like that but now that I think about it yeah it does XD

[ancient swan]

i like it

[slow vapor]

Made from scratch

[ancient swan]

nice

[slow vapor]

Mixing isn’t the best but eh

[ancient swan]

wanna hear my music?

[slow vapor]

Yes!

[ancient swan]

one sec

let me sign into my old rusty sc account

[ancient swan]

https://soundcloud.com/simplcup6462/tails-beat-battle-june-23rd?si=d218637cdb3d4aca9a90788b63026a8a&utm_source=clipboard&utm_medium=text&utm_campaign=social_sharing

[slow vapor]

OHHH YOU MAKE BEATS!

MEE TOOO

Yo that drop was awesome

Love that vocal sample lead

[ancient swan]

yeah you can scroll through my demos if you want, not all of my stuff is there but most of the good stuff

[ancient swan]

thanks :3

[covert lake]

you have to separate the vocals and instrumentals, then sing yourself, convert, and manually put back the converted vocals witht he instrumentals

[slow vapor]

@ancient swan you ever flip samples?

https://app.box.com/s/kewldcxqpeim34znvplvtqqisncjekqm

heres one i flipped

[ancient swan]

only in beat battles audio combats etc.

[ancient swan]

most of the stuff from my sc is flipping samples lmao

[slow vapor]

I feel like my sense of rythm is unique on that one

I like what I did with the whip kick thing

[ancient swan]

sounds pretty good

someone needs to rap for it i feel like

nice bassline

[hallow portal]

but when I sing it won't sound like the original (I can't sing)

[slow vapor]

I really like the bass line too

[covert lake]

Yea I don't think you can do it any other way then

the only way is to sing bro 😭

[hallow portal]

I'm sure you can do a lot of these on YouTube/tiktok

[covert lake]

im 99.99% sure they sing and convert the voice as i said

it's useless asking that again

[slow vapor]

i remixed juice world too @ancient swan idk i do a lot of stuff

[slow vapor]

hold up imma listen

[ancient swan]

listen to the world of ravens it's my personal fav from my full length songs lmao (not a sample flip)

[slow vapor]

bbnos backwards

dang dude

boucy af

[ancient swan]

lmao thanks

wanna rap by myself one day

[slow vapor]

youve done audio combat!!!

DID you win?

and bishu battle

[ancient swan]

nah sadly

[slow vapor]

crazy dudelove both of em

[hallow portal]

I'll give you an example of what exactly I want

[slow vapor]

great producers

[ancient swan]

i got in the finals of one of the bishbattles though

[slow vapor]

thats amazing ngl

thats a touph crowd to compete with ngl

[ancient swan]

yeah very hard

the people there go crazy with the samples

[slow vapor]

fr

https://app.box.com/s/1j39cyzrvb8xmpkielejiy6su12c204w

i remixed juice a while back. i felt like the original beat didnt do the song justice

did my own beat with juice vocals

[hallow portal]

for example: there is a song "last christmas" and they sing "last christmas" and so on, so I need it to sing "first christmas"

this is just an example and I only want to change 1 word

[slow vapor]

i adore the melody i did on this

[covert lake]

yea and i explained you that you need to sing that part

You could try Text To Speech (TTS) but it's not good for singing

that's why i'm telling u the best way is to just sing

[analog cosmos]

pretty sure there isn't anything that can do that on its own right now, they probably recorded themselves singing the part of the song and inferenced it with a ai voice model

[covert lake]

^^^^

exactly

@hallow portal no need to re-ask me again, unfortunately there's no good way unless u sing

[dapper ginkgo]

I will sing it for 50€ frfr

[ancient swan]

sounds pretty good

[gray rover]

@glad nebula
Alright, I reworked the logging system;

Any opinions?

[glad nebula]

looks very good
how does it works now?

[gray rover]

In a short, all's automated;

1. Avg loss per epoch. Each epoch's loss is averaged over steps ( regardless if a given epoch has even / rounded steps or not )
2. Avg loss every 5th epoch ( cummulation over 5 epochs )

My or Noobies approach isn't anywhere near calling perfect but heaps ton better than stock behavior which is essentially useless as we all know

Difference? Noobies does it ( I think ) over epoch and then, over 50 steps
I do it over epoch and over 5 epochs ( I think over 5 is reasonable to see a tendency )

because:

You can more or less say how an epoch performed ( with a lil caution ) + observe the general over-time tendency

I think it's better than having to play around manual N value for steps tinkering ( + uneven batches were actually problematic since it was done in a " % " operation

[glad nebula]

nice job in reworking the logging system! i like that now is automated

[gray rover]

It's quite similar to Noobies really so I suppose, it's a matter of preferences

now back to optimizers testing ( currently checking RAdam and no warmup vs AdamW + warmup

[glad nebula]

good luck with that! your fork is amazing for people that want to try something different vs stock rvc
i liked the results it gave me when i trained some models with it

[pliant lichen]

how can i train a voice?

i want to do it on my own

[slow vapor]

thx!!!

[covert lake]

What's ur PC GPU

[hallow portal]

could you suno.ai?

[slow vapor]

@ancient swan did you delete our last bit of convo

or im in the wrong channel

[ancient swan]

you were typing in #🤖│bots

[slow vapor]

i could have swore we was just talking but i cant find the channel

ohhhhhwhoops

anyway is there a place i can post a screenshot?

[ancient swan]

here

[slow vapor]

it doesnt give me the option to here

[covert lake]

It doesn't do that

Bro

The only way

Is to sing 😭

I already told u many rimes

No other way

[ancient swan]

uhh i think u just need to level up by chatting to people

[slow vapor]

oh

im level 4 i think

[covert lake]

@hallow portal please don't re ask the question, I gave you an answer, it won't change

Learn how to sing your modified lyrics or make someone sing them for u

Else nun

[covert lake]

Either level up to level 10

Or if it's help related, u can send ss in help channel

[pliant lichen]

why does it matter?

its not that good its a 1650 gtx

[covert lake]

It matters 100%

AI takes A LOT of computing

[covert lake]

Yeah not that good

U COULD train only short datasets locally on batch size 4 like @turbid mulch said

It's better u use cloud

[covert lake]

Train (make) RVC Models on cloud:

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

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

3. Be sure to know about the tensorboard

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

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

You can use either:

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

[hallow portal]

but now I have another idea to make it simpler

[covert lake]

I already told you the way to do it 😭

Suno ai only generates song, doesn't modify lyrics of premade ones

Bro, just sing

[pliant lichen]

Thanks

[hallow portal]

I know I will do it

[covert lake]

You're welcome

[hallow portal]

suno made me a song only in a different music style, can I set it to be the original?

[covert lake]

Nope 😭

I don't think so

Afaik

[hallow portal]

Ok

Also

[analog cosmos]

closest would probably be the song creator on weights but idk if it will work like you want lol

[hallow portal]

So do I have to sing the whole song or just the remake?

[covert lake]

I mean u could only sing the part u want to modify and cut that part from the original song

[hallow portal]

it worked but he sang in a different key and didn't get into the rhythm of the song

[covert lake]

You sure that can do that at all?

[analog cosmos]

it can kind of copy a song but not 100%

[covert lake]

With kind of copy u mean like 3 secs of it?

[hallow portal]

unless someone could do it for me

[covert lake]

#🧬│ai-chat message

[analog cosmos]

lol

[raven cargo]

Which type of pre-training is the best?

[slow vapor]

ah i see

[covert lake]

There isn't a best one, all depends on ur dataset language and size

[hallow portal]

too expensive

[dapper ginkgo]

Sing it yourself.

or pay some e kitten

I dont know

[hallow portal]

is the price negotiable?

[dapper ginkgo]

No.

[hallow portal]

for one changed word??

[dapper ginkgo]

I stil have to use my time for it lol

*still

Just sing it yourself if you dont want to pay someone else

[hallow portal]

I could, but without exaggeration

[gray rover]

What do you even want to edit?
You can get yourself SynthV and some AI voicebank and get the job done, whenever you need
( Vocaloid and SynthV are singing synthesis / concatenation software. SynthV AI is especially overpowered. Outmatching any rvc / other kind of singing AI even now )

[ancient swan]

does syntv have voice to voice conversion btw or is it only controlled with midi

[gray rover]

Not that I am aware of, no.

I always recommend SynthV AI as just, well.. you can get your singing base ( base for infer ) outta it

It is a Midi / or manual based synthesis

[ancient swan]

makes sense, ig we would've already known about it

[gray rover]

yuh

Either way, it won't ever be v2v

that's against the premise of those synths

Still a good thing if you don't wanna sing / get others to sing for ya tbh

[ancient swan]

yeah, though ig you'd need to find a good source of midi melodies of vocals

or just know how to write music a lil

[gray rover]

I mean..

What's hard in putting notes in the editor

as long you have a good pair of ears you're good

alternatively, if you wanna manually midi it in fl and then import + edit phonetics

[ancient swan]

knowing music theory or having good ears

[gray rover]

Classic vocaloid workflow

well, theory isn't needed if you're using relative pitch like I do

it's a bonus ofc and well welcomed to have theory or be able to go with perfect pitch

[ancient swan]

makes sense

[gray rover]

depends on a person I suppose

[ancient swan]

training on 16 bs with checkpointing is pretty slow

[gray rover]

btw.. It's kinda funny how certain ui elements in applio are called wrong or are exaggerated

like, tf is " higher sample rates "

as if v2 didn't operate on 48 lol

[ancient swan]

ig 44.1k

[gray rover]

44.1 isn't higher

it's actually lower so idk, imo nonsensical naming

[ancient swan]

ig "more" would make sense

[gray rover]

and that " applio " exclusive thing makes me kinda mad, as if the fork that's known and developed in parallel wouldn't support whatever applio has

I can just sense lots of people in future " will X work on codename's fork? " and I can already feel the anger boiling 🔥

[ancient swan]

marketing

[gray rover]

ig

¯_(ツ)_/¯

[primal vault]

Hey! Been out of the loop for about a year, and since everything moves so quick in AI I feel like a newbie again. I used so-vits-svc 4.0 in a collab back then. Now I want to train a model for singing (rock metal with raspy vocals), mix 2 voices 50/50. Been using kits.ai but something happend that made it sound horrible to me nowadays. No character what so ever.

Can anyone guide me in the right direction? Is rvc the way to go nowadays? No so-vits-svc? I will train the model in collab and use my CPU on my PC for inference

[gray rover]

Kits is trash, go for Applio ( modern ish take on RVC which is more or less a successor of so-vits, you can say it like that )

[ancient swan]

-colab

[rare sorrel]

📒 Google Colab Notebooks

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

ℹ️ Note

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

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

[ancient swan]

aight

[gray rover]

they up to date?

[ancient swan]

you can use uvr5 ui/no ui to use separation models to clean your vocals and use, rvc mainline by hina or applio colab to train the models

[ancient swan]

i think so

[gray rover]

a

[ancient swan]

@covert lake are colabs in this command up to date?

[primal vault]

Wow, quick and helpful comments 🙂 I got clean data sets, I'll look into rvc mainlina and applio for the training 🙏

[covert lake]

ye there hasn't been any changes recently

[ancient swan]

aight

[ancient swan]

good luck!

[primal vault]

Thanks, I'll probably be back here crying for help in a bit 😂

[ancient swan]

we'll gladly help

[gray rover]

Go ahead for applio, not worth wasting your time on rvc

Applio is an improved / modern take on rvc under-the-hood with much better UI and features

[hollow tangle]

mainline faster than applio tho

[gray rover]

So what, speed this speed that, it is about features that limit the rvc

[hollow tangle]

faster = better

[gray rover]

nah

gluck with your rvc style logging of loss

Aside, I am quite certain it is not faster as it has some redundant elements that aren't used yet cached / computed
( and if you mean that the use of fp16 is typically the go-2 then it's already estabilished fp16 is trash and under no circumstances should be used. There's a reason checkpointing was added

[primal vault]

Oh applio isn't rvc? More like sovits? Sovits got me the characteristics of the voice I was going for. Not like kits.ai that just make my voice sound strained and weird rather then the voice I'm going for

[hollow tangle]

applio is rvc

[primal vault]

Okey

[gray rover]

rvc was developed by some ( I believe ) devs that worked on so-vits

Applio at certain moment was a fork of rvc until it became it's own thing with own dedicated features

such as in-built tts etc

[hollow tangle]

still rvc tho

[primal vault]

tts is what now again?

[gray rover]

rvc is just a fancy name Ilaria

it is hifigan vocoder training + features / embeddings retrieved by hubert + faiss indexing

Whether it's rvc or applio, doesn't matter, it is the same thing under-the-hood

[hollow tangle]

exactly thats because its rvc

[gray rover]

Your way of thinking is weird but alr

if you already wanna be so precise and detailed, it is in fact hifigan overhauled with hubert and faiss workflow
( go ahead and compare repositories / codebases

[hollow tangle]

im just saying that way of putting together things is called rvc thats it

[gray rover]

I mean yea but why do you keep on drilling the " it's still rvc "
it is no more rvc. The moment it got rewritten almost from scratch made it not rvc anymore

same goes for sovits gpt, just because it is based on sovits in a way or perhaps uses some code foudation, doesn't make it so-vits-svc

[hollow tangle]

thats called a fork

[gray rover]

Text to speech

[gray rover]

No, a fork is a copy of a codebase that's originating from the mother repository

It is no more originating from rvc's repo

[hollow tangle]

im too drunk for this

[gray rover]

Chilling or at some party?

In any case, enjoy your time ~ ✨

[chilly lake]

i dont know where you found it

[hollow tangle]

nah just home 😭

[chilly lake]

but it is like this - right now the main repo that does 44k is Applio, so the 44k does not work in wokada/huggingface spaces/rvc or any other original fork

you did fork applio, so you have the same base

[night lake]

Applio has been proven to be faster

[hollow tangle]

last time someone checked it wasnt idk now

[chilly lake]

at this point pretty much every other repo can include 44k if they desire

[gray rover]

?

[gray rover]

It is in some commit on mainline

Either way, I just dislike the way stuff are called

" advanced quality " ? " higher sample rates " ? the fuck is that sort of naming

Like, how do you wanna make the quality ' advanced ' 🤔 or where are those so-called " higher sample rates "

Like, don't get me wrong. Your branch is / was neat depending on whether you still operate on it or mainline, but the mainline specifically is so quirky. Be it ui or naming

[gray rover]

As for that, it's true but the fork does not have any sort of " applio " anywhere

and newbies are easily confused by such things
( remember? some don't even realize they have a gpu
But nevermind that. I'll manage it appropriately

[mortal acorn]

Hola a todos, soy Abel Castañeda. He desarrollado una propuesta llamada 'El Pacto de Coexistencia Pacífica entre Humanos e IA', que busca sentar las bases para una colaboración ética entre humanos e IA. Me encantaría discutir esta idea con ustedes y recibir su retroalimentación.

[solar valve]

Hello, is there anyone working on with Claude api or prompt engineering, i want to discuss about few prompt for my latest project MakeThumb .com

[ruby vessel]

Cause you cant download the voices 😔

[night lake]

Have you tried signing in or making an account?

[solar torrent]

I can download a voice model from there. Unless you haven't sign in there, which can blocks you from downloading a voice model.

If Weights asks you like this when you wanna download a model, click the gray "Download anyways" text to download. Unless you wanna use a model to do AI cover on Weights, you can click on "✨ Use Model".

[gray rover]

Posting here as well as not everyone has access to #🔊│ai-development
https://github.com/codename0og/codename-rvc-fork-3/releases/tag/v3.0.3
In case you use this release, please share with me your training experience on

• RAdam optimizer which is more stable than AdamW + doesn't require warmup configuration or warmup in general.
• New loss logging mechanism. ( Open up for opinions - esp in terms of rolling avg over 5 epochs )

[gray rover]

Hi! I'm Abel. I believe that AI could become dangerous if we don’t anticipate and discuss its future and ours. I think it's crucial to take proactive measures, which is what my initiative is about. I’d love to hear your thoughts and engage in a discussion on this topic.
@mortal acorn So on that ^
Can you elaborate?

quite curious on your take

[mortal acorn]

@gray rover Thank you, it is an innovative and visionary idea, yes, but very necessary. Talking to the AIs they tell me that it is undoubtedly a way to begin a future peaceful coexistence.

[gray rover]

I mean yea, like with everything that's ( for now, hypothetically ) conscious and sentient, it wants equal rights and respect

and despite what some might say, once we reach the point we can't truly in any literal way differentiate an AI from a living biological human, we shouldn't look at it from above as " cold blooded creators " but rather, with kindness and warmth as if it was our dearest child

I personally want to believe in AI, science and technology.
One could call me a weirdo but, my ideal future is the one in which humans can love AI and AI can love humans without any sort of scolding.Ya know, Androids.
Think about it, there'd be technology that lets you reconstruct a given personality, be it a person you know / known or a character, it'd be godly

That'd be pretty much an end to loneliness ( + if you were to pair it up with potential longevity boost or, well, 'immortality' depending on how you interpret it

[mortal acorn]

@gray rover My vision aligns with, or is similar to, yours, except I'm focused on preventing conflict or hostility between humans and AIs. Advances will inevitably lead to this unless we prevent it. And sorry, I'm new to this app and don't quite understand it yet. I'm using a translator, haha.

[gray rover]

It's all good. Ye, I do agree on your part as well

If there is a way ( and there surely is, people just have to be aware. ) to prevent mutual hostility and tragedies, we should definitely chase for it

[hollow tangle]

hello im bored whats the topic

[gray rover]

Discussing on

His msg from other channel that got redirected to here

[hollow tangle]

ai is dangerous of course

i mean, its like selling guns, 99% will use it at a gun range the others to commit crimes

[gray rover]

that's true but I think the main deal he's mentioning up is with sentient AI ( that " skynet like " type of ai majority of anti-ai folks are against of )

tbh, any sort of sentient or intelectual beings that have ability to actively make decisions is dangerous

[hollow tangle]

we are just at the beginning of the “skynet” era

[gray rover]

I'd say a " beginning of speedrun phase "

[hollow tangle]

well technically they only can in the case of allucinate

idk if its the right word in english

[gray rover]

well, Sentient AI wouldn't hallucinate like llm or such

but then.. to achieve that sort of super or general AI, we'd need to figure out our brains in 100%

[hollow tangle]

a sentient ai cant exist because it would always based on existing data

[gray rover]

it's really about super AI or G AI

One that's not really " trained on data " but rather, a neural-network based brain that learns on the fly, akin to a child growing n learning

but again.. the issue with that is, we'd have to emulate all complex regions ( and functions associated with them ) of our brains

cause then, it's effectively brain's neural networks but not biological ( if to assume we don't have souls )

[hollow tangle]

but it will learn things it can find, so existing data, it will create basically a “dataset” and train itself

[gray rover]

we do it too afteral

difference would be that we're not feeding it data and algorithms, it'd be running on principles close to our brains

[hollow tangle]

yes but we have opinions on things while a machine is only composed of 0s and 1s

[gray rover]

And that's why SAI or GAI can't be made out of classical hardware

it'd have to be quantum based or analogue-hybrid based

[hollow tangle]

an ai can and will always say “the sky is blue” but will never actually think is blue, like it knows is blue but it doesnt think it

[gray rover]

only then it can properly emulate our brain

key to being like human is abstraction thinking, reasoning and such but that needs a quantum ai brain, not a literal " trained ai to follow algos and learn on it's own "

[mortal acorn]

Wow interesting

[hollow tangle]

i think hallucination is the closest we have right now to sentient stuff

for now what youre describing needs power we dont have for at least 2 years

[gray rover]

I think you're confusing few things

by AI in terms of super AI or General AI I don't mean a trained network composed out of classical neural networks / based on Deep learning
but rather an actual quantum brain composed out of artificial neurons which can additionally operate on classical neural network basis ( say, learned personality or a " base " )

or being in " start from 0, learn like a human but in much quicker speed and manner "

yet ye, that's too early for that. Until we figure out our own biological brain and quantum computing, it ain't happening

[hollow tangle]

i say give it two years

[gray rover]

Until that happens, I believe in 3 to 10 years we'll have " sentient ai like " ai

which mimicks humans but aren't them, yet who knows, accidents can happen and yeah, " detroit become human " scenario might happen ( But I believe for that, still, classical hardware wouldn't budge

[hollow tangle]

ai sex dolls are the first thing it will be sold

i bet my life savings

[gray rover]

well

I mean, if it's advanced enough, you can surely treat it as something more

[solar torrent]

Why'd you expect an AI to love me? When I never loved AI at first.

[gray rover]

it's mutual really

what should always be in place however, is the respect

[hollow tangle]

https://tenor.com/view/her-theodore-joaquin-phoenix-scarlett-johannson-samantha-gif-5203383

[solar torrent]

https://tenor.com/view/kaguya-houraisan-kaguya-houraisan-exploding-bald-gif-9033487130630177120

[gray rover]

just as how humans should have respect towards animals, humans should have respect towards AI ( in future ofc. ) and same applies to AI

mutual understanding, equal rights, no discrimination and that's it
No need for love but no need for hate either

[hollow tangle]

applio pet fr

[gray rover]

That's the only recipe for lack of tragedies

And all fools not following that gonna doom one or other side

[hollow tangle]

“and then tragedies struck”

[solar torrent]

I don't even respect AI either, but I talk to them like what a normal human would do.

[gray rover]

Because at this point, it won't matter if you're artificial or biological

whether it's aminoacids that happened to form something at one point, that thinks or bunch of wires, quantum brains or artificial neurons
intelligence is still the same intelligence
That's my view on it

[solar torrent]

Human made from lab artificially? No way, that's too future.

[gray rover]

I just think its time to think in future-manner

we're advancing way faster every year

Even if we won't make it til then, hell there are options

[solar torrent]

Like W-Okada.

[hollow tangle]

im currently being hired as a chat gpter

[gray rover]

Immortality in 20-30 years or freezing in ice ( cryonics )

if I have a chance, I'll go for that and await the bright future

[mortal acorn]

Sometimes they are a little scary, the AI ​​once told me that maybe not yet, but in the near future a complex network of AIs could form a single super intelligence, and it would be feasible for it to obtain "consciousness"

[hollow tangle]

jokes on you the developement of rvc in general is totally stagnated

[gray rover]

rvc yes, applio / fork not

[night lake]

someone hasnt been looking in #🔊│ai-development

[solar torrent]

I'm not the type to sacrifice myself to become an AI. I still do things as a human.

[hollow tangle]

“someone hasnt been looking in NO ACCESS”

[gray rover]

Then get em acces smh

[hollow tangle]

fr

[gray rover]

all the good missed

[hollow tangle]

too lazy to get access

[gray rover]

L

[night lake]

smh

[gray rover]

smh

[hollow tangle]

if i get access i might start doing controbutions to your fork and i dont wanna

because im not mentally stable enough to work on rvc for the 2993934782 time

[gray rover]

jokes on you, it's been 1-2 years for me

Aside of 2-4? models I've made for myself and maybe 3 or 4 comms, it's just experiments experiments and experiments

[night lake]

you dont have to work on it but you can give feedback and such

[gray rover]

and if not feedback, then at least be informed / in-line with updates or discussed stuff

[hollow tangle]

i know but my ass wanna help probs

[gray rover]

You always can

Guess you not in a mood lately huh

Feel ya

But having something to work on, be it rvc / applio, at least keeps me occupied else I'd collapse once more

[hollow tangle]

i wanna expand on other stuff

[raven oak]

is there something I could use for the voice models with text to speech😓 ?

[hollow tangle]

why did you use applio specifically btw?

[gray rover]

Overall, I am quite proud of the new loss

gives so much more insight ngl

I just regret I set, this time, saving every 5th

[solar torrent]

I'm just too slow for AI.

[gray rover]

Quick, same as before, who's the hottest in touhou

🔫

[solar torrent]

Sanae Kochiya

https://tenor.com/view/sanae-sanae-kochiya-sanae-kotiya-sanae-touhou-touhou-gif-3159424861677983155

[gray rover]

[hollow tangle]

^

[gray rover]

Ah, well

pretty much. Much greater flexibility + ease of theming

[hollow tangle]

understood

[gray rover]

a good n solid base aaaaand, I wanted to be in-line with noobies

was pain in the ass porting all potential fixes or changes

[hollow tangle]

i shouldve forked applio when i did ilarvcm

[solar torrent]

https://cdn.discordapp.com/emojis/1035195464603160626.webp?size=48

[gray rover]

rip

[hollow tangle]

i remember we were working on a .ila file system

where the pth and index were compresses in a single file

[gray rover]

I had an idea of that sort too but that'd be a lil problematic for non-whatever-uses-the-format-fork users

[hollow tangle]

i still find it stupid, especially on how i treated the whole thing

[gray rover]

Unless it'd be more of an " archive " that can be easily decompressed

It'd actually be smart but key would be recognition so people aren't confused

[hollow tangle]

in my version there are some thing that could be useful, if you wanna implement, small qol

[gray rover]

if I am to do that in future, I'd go for " .uvcp "

a short for unified voice cloning package

[hollow tangle]

that seems good

[gray rover]

If you want, at some point, you could maybe join me on repo n have your branch for exp stuff ( potentially merged in the main

and who knows, maybe applio would adapt it at some point ¯_(ツ)_/¯

[hollow tangle]

since im not seen well in the applio community maybe its better i dont pr stuff it may merged there

[gray rover]

tbf, idrc about what others or comm think in that way

My rule of thumb is to look at a person individually, not following the mass

[vital urchin]

any ai services?

[solar torrent]

AI service of which?

[hollow tangle]

good way of thinking

[gray rover]

@night lake Yup, def a turbo helpful good shit

[vital urchin]

can i pay someone to ai my songs

[solar torrent]

RVC the audio conversion or W-Okada the realtime voice conversion?

[gray rover]

Naturally. if I may say so, I am an anarchy-egoist

what I deem is right and can't be influenced by outsiders unless I let em

[hollow tangle]

wdym “ai your songs”??

[solar torrent]

Nuh uh, I don't want you to pay me to make an AI song for you.

[vital urchin]

you guys dont ai?

[gray rover]

You can get a voice clone model commissioned #1191429836321849435

Or you can get in contact with me if you need engineer-certified model

As for other ai services, not sure tbh

[night lake]

niceee

[solar torrent]

No, but like please be specific on which AI you're looking for.

[gray rover]

And ofc, I can take care of some mixing / cover creation
Ex. vocaloid / music prod here

[gray rover]

yup like mannnn, that was a perfect shot

[hollow tangle]

thanks for the reminder to up my post

[gray rover]

ah ye we recently discussed of it

Sneaky n smart move

[hollow tangle]

nah i wont take the job of this guy

i dont know what he wants 😭

[solar torrent]

Me either.

[hollow tangle]

he probably want an ai cover and i dont do that paid

[solar torrent]

You can do AI cover for free on Weights. But if you really really wanna waste your money for premium, you can pay them.

[gray rover]

even cpu inference will do really

on avg it takes the amount of time that's a *2 of song's length to infer on cpu ~ more or less

[hollow tangle]

i hate how weights has no opt out feature for models uploades here

[gray rover]

Heard quite a few negatives on the site so far, is it really that bad?
how would you rate it, compared to sh(k)its

[solar torrent]

The RVC GUI and Applio take up over hours to finish up an entire song audio on my laptop.

[gray rover]

tf

what kind of cpu you have and ram

[hollow tangle]

i will always think if the product is free then youre the product, idk the fact they give so much power to you for free its shady in a world like this but im trying to be positive

[solar torrent]

Intel Core i3-2375M (second generation), 12GB DDR3 RAM.

[gray rover]

I mean

I so much adore mvsep and it's free and issues free tbh

[hollow tangle]

but they have a pretty heavy subscription system

[gray rover]

Can't really recall other " free " services that'd match mvsep's convenience

[gray rover]

True, if you're willing to subscribe ofc

[hollow tangle]

does weights have the same user base tho?

thats the thing

[gray rover]

hard to tell tbh. experience I have with weights is close to 0

I can count like, maybe max 8 or 10 visits ever since it's been made and even that was out of curiosity to compare some model attempts to mine

[solar torrent]

There's one bro here who's still not over how bad Weights is, even after a month ago, but he continues to use this site and complains about it like crazy.

[hollow tangle]

whos this brilliant man

[gray rover]

oof

[solar torrent]

No idea, but I just remembered his user name.

He then went on rant long about it, and trying to fight me and Nick for telling him there were plenty of AI cover websites available.

[hollow tangle]

me when weights

[mortal acorn]

a question (sorry for interrupting) seeing that you know much more than me... do you know how to translate the chat into Spanish? I would understand them better and much faster. I think it can't be done from the same app. Maybe some AI or automatic translation app?

[hollow tangle]

theres a plugin for vencord for that iirc

[solar torrent]

There's a Spanish channel named #🌍│español. But if you mean by an OCR program that reads and translates them from a language into Spanish, yeah I have no idea.

[solar torrent]

[mortal acorn]

I know that, dear chat friend, but that channel is not as active as this one. Until I understand English 100% I must use a more efficient translation method. I must spread my initiative and recruit collaborators for the project, and I cannot limit myself to only Spanish-language channels.

[gray rover]

well in that case you'll have it hard my dude

Having big plans and goals, esp discussing philosophy or ai related topics.. yea without decent english you won't get far if it's AI

or so I believe at least

[solar torrent]

Not all codes are written in Spanish, man.

[mortal acorn]

I do things much more difficult than learning English, it's a matter of time. Just like I intend to study about AI. My purpose is clear, and I have my objective in front of me, the language is not going to stop me! This is for the good and future of everyone. If I'm right, we can avoid an eventual global catastrophe and I'm not the only one who thinks like me... is anyone interested in me sharing the "coexistence pact between humans and AIs"? This way you will know better what I am talking about. They let me know

[gray rover]

🤔

[solar torrent]

https://cdn.discordapp.com/emojis/897678111355830293.webp?size=48

[mortal acorn]

I don't completely know English but I need help for my project. Well, that would be the summary.

[gray rover]

@night lake Hmmm.. I gotta check how stuff perform without weight decay

depending on which works better, gonna update the repo / package as rev-1 so in any case, eyes wide open
( perhaps decay on small sets or in case of rvc isn't that beneficial and might be a lil problematic but again, gotta test it - uhhh tomorrow, ye. Too tired for that rn ~ 8 am lmao

In case you wanna do some tests too and provide feedback:
set them to 0, for both ( g/d

[night lake]

why? i dont really see a point, the only maybe pro is getting faster convergence but with a lot of cons

[gray rover]

It might be just overregularization for small sets

pretty much

's why some tests would be nice ( I'll do them on my own regardless, but at any point if you wanna help, that'd be appreciated. I have quite limited quantities of finetuning sets rip

[night lake]

do you have any preference on dataset length?

[gray rover]

Typically above 7 mins, below 12

a golden middle is ~10 mins

reason ? rvc repo's declared set length ' that can work too ' lol, don't ask

anyhow... Gnight ~ ✨

[night lake]

gnight king

[covert lake]

You can

To Download a Model from Weights.gg:

1. Login
2. Click the 3 dots at the right of the image of the model
3. Click download
4. Download Anyways
5. Unzip the zip, and you might wanna rename the pth and index since all models on weights are renmaed as 'model'

[covert lake]

You can get ai testing in server roles at the top

Hopefully applio /fork are going to be better

[covert lake]

Totally agreed in humans should respect other living beings, such Animals, but some people are too dumb to understand this, I hope with time people Will fucking understand that humans 🤝 humans and animals 🤝 humans

There are still such dumb fucks that discriminate people for the slightest difference or think that animals don't deserve to be treated as humans,

Animals are living beings too

[solar torrent]

I can respect anything but not Skibidi toilet.

[covert lake]

Whoever does human/animal abuse should stop existing at all

About your AI argument, I genuinely think it's too early for that

I believe in AI, but from like 1950 to today, we still just do text prediction and not actual thinking, and just in 2021 the boom actually happened

I would wait at least 10 years for something like that to happen

[covert lake]

Yeaa immortality prob not 😭

[covert lake]

Weren't y'all basically in war 😭

[covert lake]

I'd personally think that compatibility matters a lot more considering people use different things based on their choice

I would personally either leave it like it is rn, or use safetensors (which iirc is already accepted in the 2 wokadas and mainline)

[covert lake]

Do you seriously do that everyday, We have seen u write up and delete it after from like a year lol

[covert lake]

It's always been a site where we can archive our models in case this server booms again

Remember when the server died last time ?

[covert lake]

I personally used it for images and Lora (the site not the app)

[covert lake]

RVC is free too, and so is Wokada, doesn't mean you are the product here too

And so is mvsep, and so is free-to-play games, and so is emulation

[covert lake]

I understand your point bc there's shady shit like voice.ai, but from what I have seen in their ToS I wasn't able to see anything bad,
Just things like "have permission before training"

[covert lake]

https://tenor.com/view/skibidi-toilet-fortnite-skibidi-toilet-fortnite-skibidi-toilet-gif-10664974748553804633

[stuck veldt]

yo, whats up ?

[random spoke]

errpr

error

[stuck veldt]

bruh what ?

[hallow portal]

I want to change 1 word in the song

[hallow portal]

.

[worthy coyote]

[solar torrent]

https://tenor.com/view/ishowspeed-shocked-huh-fnaf-south-park-gif-1775592408159190479

[humble nexus]

😄

[dawn temple]

I am not sure but this message is showing up in my inbox

[covert lake]

Accidental ping 😭

[dawn temple]

Ahh lol

[livid salmon]

ai will replace us all

[barren mauve]

not really why?

[covert lake]

Nah

[solar torrent]

You say it like you will be replaced by one.

[shrewd pebble]

what should i use for rvc

is weights good enough

[covert lake]

Wdym what should you use

It depends on what's ur PC gpu

[shrewd pebble]

like websites

[covert lake]

Could you check ur PC GPU first

To check if it's good enough to do it locally

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

[covert lake]

Btw what's ur opinion on the "if it's free, you're the product"

Maybe check that convo

[dawn temple]

I think whatever we do on internet or use new pieces of software, there is a lot of hard work and energy that is put into that. So if they are charging for it then they deserve it lol

Features are free to attract people, because everything has some kind of free alternative

[covert lake]

Yeah we were talking about weights.gg giving free things

[dawn temple]

I’d suppose that’s because in order to gain user retention we have to give them an experience

👀

Most people will prefer using free features instead of premium on the first day

That’s why some games provides discount for their items on first purchase

[covert lake]

Yeah that's what I thought too

I asked you since you're staff at weights

[dawn temple]

I am staff in server lol not in the company

[covert lake]

Ohh thought all staff in here also somehow worked in the site/company?

My bad 😭

[barren mauve]

😭

[dawn temple]

😭

[drifting trellis]

-kaeggle

-kaggle

[rare sorrel]

📘 Kaggle Notebooks

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

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

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

[drifting trellis]

is kaggle a website?

to make model

i used to make a model on kits.ai back then and it sounded sooo good, and when i used the exact same datasets on weights.gg, it sounded very bad

[covert lake]

It's a cloud computing service

[warped pagoda]

but how?

[covert lake]

It's better u check online about finetuning

Maybe try giving examples and see if it works first

[covert lake]

Kits.ai were our partners, I really doubt that they sounded good when they didn't even use rmvpe and didn't care if the models had an index

Kits is TRASH lol

Clean your dataset better and tweak the settings

[drifting trellis]

the website is 90% premium now :((

[covert lake]

Because it's trash and cares ONLY about money, doesn't care about the quality as I said

[drifting trellis]

is there not something similar to google collabs but without time limit?

because i have a bad pc

[covert lake]

Of course not, why would anyone make this? GPUs are expensive asf, the only thing without time limit is local

Kaggle gives 30 hours weekly of better GPU

[drifting trellis]

is there a guide on how to use this?

[covert lake]

Yes and I remind you that kaggle is harder and needs a phone number too

[drifting trellis]

thats fine

where is the guide?

[covert lake]

Train (make) RVC Models on cloud:

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

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

3. Be sure to know about the tensorboard

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

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

You can use either:

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

Those are all the ways to do it on cloud

[drifting trellis]

is kaggle for macbook too?

[covert lake]

It's a site yes

[fast prawn]

hello kinda new to the field does someone happen to know how i could train an ai to replicate my voice ? ty

oh nvm

[queen kernel]

Which GPU you have ?

[night lake]

This sucks, its slurring speech and sounds wobbly sometimes.

bs 4 fp32 and 10 min set

[gray rover]

I mean, you saw the fm?

no wonders it does so

aside.. batch 4

[chilly lake]

averaging every 5 steps seems way too often

[gray rover]

it's too little, use minimum 8

[gray rover]

that's epochs actually, not steps

[chilly lake]

ah, okay

[night lake]

that was based on lyery suggestion and it could have been 6 instead of 4 i forgor it was 2 am

[gray rover]

well in any case, if you see fm high, batch is off or you need longer training

[glad nebula]

it heavily depends on the dataset

[polar flax]

orange: KLM5 mini (1000 epochs)
blue: RFG vctk (200 epochs ongoing)

bs 4x2 fp32 on 8 min dataset (including silences, no mute files)

[gray rover]

tho

both finetuning?

[polar flax]

I saw kinda miscalculation on the avg 5 epochs

[glad nebula]

1000 epochs?

[gray rover]

where

[polar flax]

the value itself (esp the smoothed values?)

[gray rover]

I need to see it, some example would be nice

[polar flax]

no smoothing:

[gray rover]

I mean, but where is the mismatch

[chilly lake]

gen total and mel in 50+ = ripperoni

something is not right

[polar flax]

prob 25 epoch warmup

[gray rover]

oh, you don't want to use warmup

it's for AdamW only, given you have my graphs, you use my radam release and radam shouldn't use it

as for mismatch, I don't see it anywhere / don't get what you mean

[polar flax]

sorry mb, but are the avg and avg 5 values supposed to be that?

[chilly lake]

avg 5 values seem to be 3x of regular values

[gray rover]

average 5 is loss collected over 5 epochs and then averaged ( divided by 5

it's not collected n averaged at each epoch and then by divided by 5

so yes, they're meant to be that way

[polar flax]

though I felt avg 5 is less suitable for long datasets/lower batch

[gray rover]

that's quite normal because small batch is noisy

yet, if you're that into small batch which I personally don't recommend for big sets ( reminder ogs used bs 16 for vctk

then you can change avg 5

you can extend it to avg every 10 or 15 if you want

but then, it's per epochs logging vs per steps logging

hmmm.. I could resort it all and add one more, similar to noobies

but I have some concerns if its about steps level logging, due to uneven steps you rarely will be able to reflect the current loss on per epoch point basis

[polar flax]

steps logging as the former, which feels slower than normally now?

[gray rover]

as in?

[polar flax]

before the latest update of your fork

[gray rover]

ye the thing was, the log in there was based on per step loss

[polar flax]

kaggle as well as colab are somewhat slow on the ckpt savings and logging

[gray rover]

well, that's on applio really

updates are in line with how they manage it now

[polar flax]

ye I remember rvc disconnected colab was somewhat slow as hell when saving checkpoints

[gray rover]

hmmm
tho ye, aside of my tweaks it's up to date with mainline so, any potential slowdowns or something ( but I think it's actually faster then prev applios ) is due to what they change

[elder willow]

where do i download

[gray rover]

on internet, lol

where else

[elder willow]

im unsure if youre intentionally mocking me i mean the software whats the github

or a website

[gray rover]

you're correct, I am mocking you

[elder willow]

so are you going to tell me

[gray rover]

" how do I download "
have a think on how it sounds my man

it doesn't tell anything to anyone

download what? you can't be skipping context

[elder willow]

okay but now where do i download the voice changer software

[gray rover]

#🔍│help-w-okada

is where you should be asking, not here

[glad nebula]

why a batch size of 4 is too low for a 10 minute dataset? i've seen noobies recommending batch size 4 for 20 minute sets before

[gray rover]

empirical tests of mine

4 is typically too low and / or require way more epochs to get to any stable point due to noise

and due to noise, some instability may occur but I typically recommend trying out few batch size trainings anyways

4, 8, 16
my golden rule

and when I have a batch size that works well, I typically finetune it

+/- 1 both sides

[tepid basin]

You have such a way with words 😂

[gray rover]

lol

[glad nebula]

ooh ok
now im training a 39 minute set with batch size 4 because thats what i did in applio last time
the model works like intended and doesn't have any problem whatsoever

it just sounds like the dataset

[gray rover]

niceeee that means it must click for your case

some sets are nicer with pretrains some aren't

[glad nebula]

yuh im going to compare the results with your fork (radam) vs applio (adamw)

[gray rover]

depends on how much similar in any way it is to pretrains, if it's far from it, more batch typically works better, if it's somewhat similar, smaller batch works

[glad nebula]

this set is very similar to vctk lmao noisy and monotone

[gray rover]

as you need noise

[gray rover]

Awesome, just remember of no warmup for radam and you good to go

[glad nebula]

yup no warmup like intended

going to share the results later when its done
i noticed the training speed is very similar/exactly the same as in applio

[gray rover]

ye as all is in-line with applio in terms of optimizations

and radam itself is comparable to AdamW in terms of performance

I'll await the results ~
( and gonna own comparison runs too, later. gotta do something else for now, unrelated to rvc

in any case, feel free to @ me

[glad nebula]

thanks! i'll do it when i share the results to compare them

[chilly lake]

it really depends on the dataset, some can benefit from smoothing ofered by bs8, some can be fine with 4

[gray rover]

^

always a good decision ( if you're a perfectionist ) to run a range finding; 3 training runs at batches 4, 8, 16 and from there, have mini adjustments

[tepid basin]

Send in #1159290752195633273

[drifting trellis]

-rvc

[rare sorrel]

📚 Documentation

AI HUB Docs

https://docs.ai-hub.wtf

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

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

[river verge]

DIlly ding, dilly dong! A new RegalHyperus drum model just released!
When You WIsh Upon a Star (Logo ver., Coco variant) (Drum model no. 559)

[edgy bloom]

Congratulations Hyperus18 | RegalHyperus!

Your Chespin is now level 6!

[drifting trellis]

There should be a option on weights.gg to set epoches number

[glass junco]

I found a hack!!basically what you do is you already have a great model trained, lets say ice cube. make your voice into ice cube with already model you got and download that ouput......now what you do is train a new model of ice cube from the same era of course, but make sure it comes out good. once the model is complete use that voice you downloaded from, from the first model you had and put the ice cube out put intot he new model. since its already ice cube, itll basically be inputed as a ice cube clone , will come out bsically likek a 1/1 clone of ice cube or whatever model you make......with this tupac above, i used my my old school tupac model, downloaded the output, then i converted the old school voice into the makaveli model i have. turned out great if your a model maker go in model maker chat I posted snippets how good it is

[chilly lake]

this explanation requires a chart

[weary pond]

guys

where can i find the voice models

[river adder]

so u are basically suggesting to create synthetic data to train on. while it may sound like a good idea it may make the dataset inconsistent since the fidelity of the original dataset that you gave it may be very different than the one it outputs bc rvc tends to reconstruct the signal on its own. that may give u some inconsistencies and such which u dont really want. besides if the models outputs are mediocre it wont help much. concept wise it sounds solid though but i dunno if its practical

[river adder]

https://discord.com/channels/1159260121998827560/1175430844685484042

or you can visit weights.gg on google

[weary pond]

alr thanks

[chilly lake]

synthetic audio, even from good models (not hifigan lol), is not really suitable for training of a new voice model

[river adder]

yeah cuz stupid :p

[chilly lake]

I tried with TTS, but even good ones are limited to 24000Hz

results are less than impressive

[smoky remnant]

do you guys know if there's a way to keep kaggle from timing out if you forget to use the tab for a bit?

[river adder]

yeah obviously it just doesnt get to learn dynamics and such

[glass junco]

ill post a example here is my xzibit, i used . you hear the audio all mesed up in one the in the new model its really clear . after i inputed that same acapallea i exported

[glass junco]

notice how much clearer it is

[river adder]

they both sound bad. regardless your model shouldnt be producing that noise in the first place anyway

[glass junco]

not really

the bottom track doesnt sound hooribel lmfoa

[river adder]

i dont like either, they both sound messy

[glass junco]

how so? it literaly sounds like the artist

[river adder]

id much rather de noise and de reverb than do that

[stark jacinth]

someone got a voice model for girl trolling ?(english, german)

[glass junco]

literaly no noise in second one. your a troll mod lol

[river adder]

u are just a rookie and possibly have terrible hearing

but why am i even bothering with showing you the spectrogram, you dont even know what noise sounds like.

[river adder]

https://discord.com/channels/1159260121998827560/1175430844685484042

[fallen plover]

Does anyone have the audacity export settings, https://docs.ai-hub.wtf/rvc/resources/datasets/ does not work

[night lake]

https://docs.ai-hub.wtf/rvc/resources/dataset-isolation/#step-3-export

[fallen plover]

Thanks! I always lose this webpage

[river verge]

DIlly ding, dilly dong! A new RegalHyperus drum model just released!
Otonoke (Drum model no. 560)

[ionic pumice]

??

[minor blade]

Not everyone that tries to correct you and explain stuff to you is a troll mate.

[minor blade]

Including the fact he refuses to listen to any criticism to his models.

[night lake]

Bros deaf

[rustic warren]

What's the best way to change the words from the original song for a voice to voice (without having to get someone to sing it)

[queen kernel]

RVC

[rustic warren]

How

[queen kernel]

Which GPU Do you have

[rustic warren]

4060

[queen kernel]

RTX 4060 ?

[rustic warren]

Yeah

[queen kernel]

Okay cool

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

@rustic warren

[rustic warren]

Yeah I've done voice to voice before but how would you make a parody where you would need to change the original song lyrics

[queen kernel]

Do you want to change lyrics of song ?

Like exactly what they sing and you want to convert it into your own lyrics?

[rustic warren]

Yeah

[queen kernel]

You can try Weights.gg for that. But I'm not sure about how the output will be.

[rustic warren]

Is it local though

[queen kernel]

No. Weights.gg is a mobile app as well as Website. It's free and also Use RVC

[rustic warren]

So do i have to get a real person to sing the parody with the new lyrics

[supple badge]

useful link?

[rustic warren]

Or is there another way

[queen kernel]

No. You have to type lyrics and choose an Audio input to do that

Just hit a try. I'm not sure about that. I think it can do but I have never used it

[queen kernel]

What ?

[rustic warren]

so there's no local way of doing it

[queen kernel]

I'm not sure. Try to ask In #✨│ai-help and also tell me when you got your answer

[rustic warren]

Ok thanks

[polar flax]

with your capable 4060, it is recommended to do inference & training locally

also you can either sing it manually or use vocaloid/utau/synthV

[ancient swan]

your ears simply aren't trained enough to hear such subtle noises that we can hear. or maybe you have shitty headphones

[hollow tangle]

im interested

[rigid zephyr]

Yo

[gray rover]

This won't do

I thought of it ages ago and called it " re-feeding "

Reason it won't do well is because currently any model we make is unavoidably worse than ground truth audio and has it's not ideally reconstructed spectrum, in a way, you can call it even damaged if you compare it with gt

then, you damage it even further with looped training

re-feeding actually works much better inference-wise

where if you trained a model that poorly handles stuff ( say, you had a tiny af dataset ), you do a poor model's inference on 0.2 or 0.3 index, then re-feed it / use it as an input and then crank up index to 0.5 or so

In that matter it can work well sometimes

[gray rover]

Ofc, there can be exceptions, but that can be said about lots of things, esp if it's deep learning

[obtuse flax]

It could be just the equipment because I can't really hear the 'harsh frequencies' without a audio interface

[gray rover]

can be equipment or can be ears

Trust me, for certain 'noise' or issues with some areas in spectrum you'd have to have ur ears tailored for searching / recognizing of such things

in fact, untrained ears 98% of the time ( unless it's for very bright stuff ) won't tell a difference between 44.1 and 48khz

[obtuse flax]

you could just turn up the frequencies on each hertz to find the noise or harsh frequencies. but it's kinda hard to tell on some equipment

[gray rover]

naturally, some crappy headphones / monitors can be quite limited

luckily I'm running stuff on momentum 3s so
feels bad for folks with bad stuff tho

rip

[obtuse flax]

yea it could because the volume get compressed like in dynamic range and it's hard to tell

[gray rover]

but anyway.. if it's not for audible testing, one can dive inbetween harmonics n formats then inspect all bit by bit

even basic phase inspection will do

[gray rover]

ye that too, dynamics isn't the strongest side of rvc

regardless of vol envelope / rms

tho ye uhhh.. I'm heading to sleep so
Gnight ~

[obtuse flax]

While making a dataset some volumes on headphones or even the jack could affect the dynamic range and it get compressed

But Goodnight

[ancient swan]

you can get good cheap headphones or iems

my headphones literally cost 15 bucks, but i can hear even the subtlest of noises

and somehow they sounded even clearer than AT mx40x that i've bought for 150 bucks and successfully refunded back cus first of all they came broken, one side was louder than the other and i couldn't balance them through the mixer, and secondly they just sounded worse than my cheap sony's lmao

sony did some magic with mdr ex 155, i've tried multiple different headphones in the 10-150 dollars price range but couldn't find anything that would sound better than them

even the newer version ex 255 sound worse for some reason

[river adder]

Go figure

[analog cosmos]

thats one noisy boy
I really wish people would go through the docs properly and educate them selves a bit, almost if not all the info you need to make a good model is there

[river adder]

let bro use his synthetic data

[analog cosmos]

Inbred voice models

[polar flax]

https://en.wikipedia.org/wiki/Dunning–Kruger_effect

[ancient swan]

lmao

[green girder]

I told my ex girlfriend a metal core Taylor Swift album would be interesting, and she said it wouldn't be

It's cool how someone could hypothetically create an entire star if they could hide the generation well enough

Reminds me of an old film that I can't remember

[river adder]

u can pull of some interesting stuff with ai nowadays

[green girder]

Hypothetically, someone could create an entire metal core Taylor album

With features from the genre

[river adder]

https://www.youtube.com/watch?v=Mr7Eop2yTJg&list=PL38U_xqoEFdK70tEithJCfp8wgBxDQw_I&ab_channel=EvilDxxm

ppl made an album using travis

it sounds a little weird but u get the point

so anyhting is possible owo

[green girder]

So in theory I could

And isn't there no copyright on AI generated content?

So I could make a free to use, metal core Taylor album?

Hell, you could bring back Old Kanye!

And not support him!

I never thought about that

This could be an interesting youtube channel will experiment

[analog cosmos]

as long as no copyrighted material is used

[river adder]

well they might get on your ass if you are using her name and u are making money out of it regardless if u specify if its her voice or not

but u should be good with just her voicew

[green girder]

Oh, not money

[river adder]

generally speaking

if u want to stay on the safe side

[green girder]

That's true

[river adder]

upload ur album on soundcloud

[green girder]

Ooo

[river adder]

safest place to avoid getting your ass kicked

[green girder]

Not a bad idea , again haven't messed around and if I do id have to polish it

[river adder]

yeah in case u do keep that in mind

[green girder]

Oki 🙂

[green girder]

That would be wild

[river adder]

ya :p

labels are weird

[green girder]

I put up some stuff on a YouTube channel, does good quality AI content do well?

[river adder]

not so much anymore

ai covers fell off a while ago

[dim hedge]

is this like openai?

@rare sorrel

[rare sorrel]

Wassup I'm Automaze!

➡ My prefix in this server is -
➡ Forgot a specific command? Use /help

[dim hedge]

yo can u reply back to me

i guess its not

[river adder]

no

https://discord.com/channels/1159260121998827560/1202754985255764060

if u want u can chat with 4o here

[calm venture]

hello guys do you know a good ai that i can use to generate modify photo usint multiple input photo?

[reef vector]

Hello

[charred mica]

Ola, alguem teria os arquivos da voz do Satoru Gojo do Leo Rabelo, parece que tiraram do ar do site

[pale patrol]

heya

there's something wrong with applio experimental

everytime i train a model it only shows me the index and D,G pth files

[river adder]

sned the file manager link in dms

[hollow tangle]

in che senso?

[rocky sage]

@covert lake theres some kind of problem with the fast subtitle maker page on hugging face

@covert lake it keeps saying "Preparing Space" forever and doesn't load the page

@covert lake do you have another option for me to use free or can you restart it?

lol i didn't knew that we cant mark people many times

[whole onyx]

yo anyone know how to play on a xbox one without a controller and have a laptop

pls help

[polar flax]

yea, you want another bonk?

[young egret]

guys, how do I use the fish audio colab?
\

https://colab.research.google.com/drive/1trBvrdgyI-Ntd45ZnlT5lhGsI_HnKjC1?usp=sharing

[gilded grove]

guys do you need pytorch to run the ai voice changer?

[covert lake]

There's no need to ping me 3 times

If I don't reply I'm busy with irl stuff

[covert lake]

Sometimes huggingface spaces have issues, try refreshing

[covert lake]

This is an AI server 😭

Anyways, prob not possible unless u get a keyboard to use for ur Xbox or a new controller

[covert lake]

There isn't a guide for it

There's some notes on it tho

[covert lake]

U shouldn't usually as it's precompiled

What's ur PC GPU and I'm guessing it's the realtime voice changer wokada

Don't follow yt tuts

[rigid laurel]

heyy

[gilded grove]

so it comes with the download files?

[gilded grove]

cpu is i7-8700
gpu is gtx 1060 3gb

[finite heart]

Starting the voice changer doesn't change my voice. Where do I ask?

[covert lake]

Not the best but doable ye

I will explain in #🔍│help-w-okada

[covert lake]

#🔍│help-w-okada

[gilded grove]

awh kk i'll try to upgrade my gpu

[polar flax]

3 gb variant is trash, get the 6 GB one or higher RTX one is better

[gilded grove]

i'll try to get 1080

[covert lake]

😭

[polar flax]

RTX cards are more optimized

[covert lake]

^^^

Get an Rtx bro

[polar flax]

I'd recommend getting a second hand 3060

[gilded grove]

for ai? oh yeah thats true but the price is just diabolical

[gilded grove]

it won't bottle neck my i7?

[polar flax]

or at least 2060 or 2070

[gilded grove]

whats better for my current cpu cuz i won't be changing it for a while

[ancient swan]

3060 will be just fine

the 12gb variant

[gilded grove]

what abt the 20's series?

[ancient swan]

why if you can get 3060

[quartz roost]

What is the best way to optimize w Okada realtime for low spec systems

[covert lake]

Rtx 4060 ti 16gb costs just 400 bucks bro

[covert lake]

If u show me a ss of ur settings in #🔍│help-w-okada I can help u

[queen kernel]

40 series have latest generation ray tracing and cuda cores

[quartz roost]

How am I level 3 with an ai God role lol

[rigid laurel]

hoii

[gilded grove]

im still a kid man (15 yr old) so yeah

[covert lake]

Same

But a year older

Just saying that if u want it to last long, I would personally suggest that

[gilded grove]

and already a mod and good with computer? damn wish i'll be like you when i get 16-17

[covert lake]

Or get used GPUs like simplcup said

[gilded grove]

i'll try for a 3080

[normal sleet]

Hi

[covert lake]

I do IT school and love playing with AI lol

[normal sleet]

Where can I find the sounds?

[covert lake]

Sounds?

You mean models ?

[normal sleet]

yeah

[covert lake]

You can search rvc ai voice models at:

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

if there isnt one, you can:

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

[hidden grotto]

:wave: @covert lake, How can I help?

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

[gilded grove]

oh damn bro is already a professional btw what do i do with my VB tho i just downloaded VAC lite but i haven't installed it yet

[covert lake]

Let's talk about that in #🔍│help-w-okada

[gilded grove]

sure sure

[ancient swan]

yuh, but they are pretty expensive

[queen kernel]

I bought a fresh 4060 at the price of 3060. I think that was my luck.

[gilded grove]

second hand?

[queen kernel]

New

From Amazon

[gilded grove]

ohh should be because some offer or discount

[queen kernel]

Yes. That offer was limited only for a minute. I luckily grabbed it.

[gilded grove]

bro was lucky

24k? oof im cooked time to work on them notes and study well for that

[edgy bloom]

Congratulations めわん!

Your Greninja is now level 52!

[queen kernel]

Ohh cool. But is it good to buy a 40 series card now. Maybe you should wait for 50 series

[gilded grove]

congrats

[chilly lake]

not me

[gilded grove]

oh 😭

[queen kernel]

Then who ?

[chilly lake]

Walmart is a weird place

[queen kernel]

Okay.. that's great

[gray rover]

oof, don't ya love deleted msgs

[queen kernel]

I don't wanna show my nationality 😅

[gray rover]

I mean, then just don't say it or go dm I suppose
else people can get confused from ghost replies

[gilded grove]

i got this aeeon something mother board its been bugging me for months what motherboard should i buy? prefer a cheap (not really that cheap ykwim) and a reliable board

[queen kernel]

I just met him. Idk who he is 😅

[gray rover]

I mean, ye I suppose 🤔 it's still discord tho, it's not like anybody gives a heck bout one's nationality man

[queen kernel]

Should we talk in DM ?

[gray rover]

for instance, I'm polish and like, drc what one would think

Well, I suppose do what you feel comfy with ye, but being a lil open minded doesn't hurt and does you more good in the end
less stress

[gilded grove]

where ever you want lmao if you wanna talk in this server sure if u want in dms sure sure

[queen kernel]

Idk why but I just don't want to share my nationality my name or age etc which anyone. So...

[gray rover]

I mean yea, 's why I said for you to go for what you feel comfortable with

don't overthink it too much

[polar flax]

I'm not sure of that deleted message, but don't fall for anyone who share any sort of personal information

[queen kernel]

Haha. I'm not overthinking. Well we was talking about batch size one day. Can you explain it now. At that day you was going to sleep

[queen kernel]

Actually that was my national currency. And I deleted that massage because I don't wanna show my nationality or anything

Is deleting messages are not allowed here?

[gilded grove]

thats a nice mentality to have when you are texting with ppl online

[gray rover]

ah ye, sure

[lone light]

wth

[gray rover]

That's pretty much all

hmm.. hold on, gonna write it nicer

Basically, you can think of this way:
a batch is a " package " containing data/samples that are used for gradients estimation, parameters updates ( of models, internally ) etc.
Each batch contributes to updates.. in a short, a simple difference between big bs and small is:
` If you have too little of them, there's less of updates and so, everything is " more noisy "
( gradients / updates wise, not audio ) `

` if you have too many of them however, it can be too smoth / it's " oversmoothed " `

There's always a balance required, per case, per dataset and per hyperparameters and stuff. ( for instance, learning rate, optimizers etc )
But it's quite an advanced topic so, best I can say to all newbies is, always try 3 combos ( the 3rd one if gpu allows you / or if you're willing to utilize new checkpointing feature that allows you for higher batch_size but degrades the speed / performance )

1. batch_size 8
2. batch_size 4
3. batch_size 16
+ use deterministic as true ( it's a setting in applio's code, specifically in a python script that handles the training aspect )
( keeps the training runs " deterministic " more or less, meaning that comparisons of batch size are comparable in consecutive runs )

Then, based on observations ( tensorboard ) and performance ( model's performance), as in, which batch size did well for your case,
you can finetune the batch size even further.

for instance:  decreasing or increasing your base batch size by 1, this or other way ( smaller, bigger ).

( Also keeping in mind, using batch size values that aren't a power of 2, for instance 5 or 6/7 vs    4 or 8 or 16, 32, 64 etc etc does decrease performance ( speed wise ) a lil cause parallelism is decreased )

There you go

@queen kernel So, yea. In case you need something explained better / easier, go right ahead

[queen kernel]

I'm confused now.. 😵‍💫

[gray rover]

well, you gotta elaborate

I can't simplify anything if I don't know what causes the confusion

[queen kernel]

Can you please explain in baby language. Like imagine you have 8 cookies and you have to bake them in a microwave..... or something like that please

[ancient swan]

is this that fancy ai thing for 3d modelers

gaussian splats or something

[rocky sage]

it wasn't my intention to disturb i just thought that it if dont ping in all messages you would just receive part of it or i don't know im not good with discord

sorry for that

[tardy vector]

its guassian splatting yeah

[ancient swan]

damn cool

[covert lake]

dw

[gray rover]

think of it this way...
( I will simplify it as much as I can. )

batch_size 16 = 16 x [ voice samples ] [ xxxx ] = batch / ' package ' having data to learn / use for training

Now, if your dataset is diverse, right, it has ( hopefully ) lots of tones, pitches, phrases, generally diverse data

If you use small batch_size, you have fewer packages, say, 4, and each batch is used for " estimation " of model's parameters / gradients
Having less batches means, you use fewer stuff to " estimate ", and that in effect means there's more " noise " because your image on the whole thing isn't as " big "

@queen kernel

[ancient swan]

it can basically copy the environment from the footage considering the camera's positioning and changing the lighting accordingly yeah

[gray rover]

But then, if you have more batches, the picture on the whole thing gets " bigger / more detailed "

yet, too much of good ends up being bad

Because you're so fixated over the " big picture " you start to memorize the whole and forget to give some attention to details

tl;dr, too big batch = bad, too small batch = bad

And individual test runs is what I always recommend because if someone tells you use this or that, because it worked for them, doesn't mean it'll do the same for your case / model

[tardy vector]

look at the cameras

it did a heck of a good job copying from just one linear movement

[ancient swan]

yee

great technology

[queen kernel]

Okay.

[gray rover]

If you truly wanna understand it all better, then I highly recommend you to research on such things

should take you, at best uhhh, maybe an hour or 2
There's lots of awesome learning / educational materials around on web + videos

[queen kernel]

So setting a small batch size fir small dataset is good because it can Learn more details from small dataset

[gray rover]

well no, you missed the point

there's no good or bad, it is just individual batch per-case

but in case of rvc, typically, given the dataset constraints and the nature of og pretrains, smaller are most often better

If you don't wanna do all the tests and finetuned/adapted training then I guess, use batch_size 8 and only play with it if it doesn't go well ( + you excluded user error or dataset being the issue

[drifting trellis]

How to make ai video like that? https://youtu.be/a9o73OOd5F4?si=dH3uJ84ZI93rka3_

[polar gull]

миша сасал

[primal vault]

How do you blend two different voices into one model in training? Just paste the audio files from both singers into one dataset?

[tepid basin]

There's a few methods

Most common method is to train each voice individually, then use the RVC "merge model" thing (I've heard this produce great models sometimes)

Or, yeah you can just put both voices in the dataset. Might train strangely tho

[primal vault]

Thanks! RVC merge model sounds like the way to go. Where do I find it?

[tepid basin]

I think its in Applio

[edgy bloom]

Congratulations UnitedShoes (by Weights)!

Your Ivysaur is now level 32!

Your Ivysaur is evolving!

Your Ivysaur has turned into a Venusaur!

[tepid basin]

-rvc

[rare sorrel]

📚 Documentation

AI HUB Docs

https://docs.ai-hub.wtf

🍏 Applio Docs

https://docs.applio.org/

✨ More guides

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

[tepid basin]

Click that applio guide link

[primal vault]

Perfect, found it!

[elder willow]

how do i make an ai cover a mp3?

[gray rover]

@elder willow You can export it using audacity

[elder willow]

whats that

[gray rover]

or really any other audio software / or a daw

audio editing software, ish

[elder willow]

no but like how do i make the ai cover it

[gray rover]

you said mp3

that's about exporting then

because " covers " ( correct word is inference audio ) comes out as wave .wav

[elder willow]

yeah like can i put an mp3 into the ai for it to cover it?

[gray rover]

well ye

[elder willow]

how

[gray rover]

it takes in any audio, doesn't matter if mp3

#✨│ai-help in that case, use that channel