#what are the main reasons why the model goes wrong?

1 messages · Page 1 of 1 (latest)

rose drift
#

One message removed from a suspended account.

sinful hatch
#

This is def an overprocessed dataset

#

alot of the silences are where you get the robotic noise too

image.png
#

Can I ask for some samples from your dataset?

#

You should be targetting the most unprocessed, natural voice possible. It sounds to me like you heavily overproccessed with plugins, but i cant be sure

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

serene hearth
#

make sure you're using the same embedder as what the model was trained with

#

it's not always as the default (contentvec)

rose drift
#

One message removed from a suspended account.

serene hearth
#

I haven't seen anything with jp hubert good enough

rose drift
#

One message removed from a suspended account.

serene hearth
#

the old pretrain using it isn't proven to be that good

#

if you were training the model, I'd recommend KLM 4.9 with ofc default contentvec

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

rose drift
rose drift
sinful hatch
#

Are you including silence in your dataset? That's what I would recommend avoiding first. RVC gets really pissy with silence

#

with your completed dataset, try these steps. This is "audio labelling" basically splitting every phrase into chunks, to remove silence, and also to help RVC not cut off phrases. Try retraining your set with audio exported like this

sinful hatch
#

i like to think silence like this is what's causing it to fuck

image.png
sinful hatch
rose drift
#

One message removed from a suspended account.

sinful hatch
#

I ignore everything in the advanced section for preprocess

rose drift
#

One message removed from a suspended account.

sinful hatch
#

your other settings are okay. I use contentvec for english, idk how jp-hubert does, never touched it

sinful hatch
rose drift
rose drift
#

One message removed from a suspended account.

sinful hatch
#

o

rose drift
sinful hatch
#

i highly highly suggest trying this and see if it improves your model

#

im curious too

rose drift
#

One message removed from a suspended account.

sinful hatch
#

oh and in your export menu, use these

image.png
rose drift
#

One message removed from a suspended account.

sinful hatch
#

yea being dynamic in the dataset helps it learn how to act

#

especailly for realtime. you never know

rose drift
#

One message removed from a suspended account.

sinful hatch
rose drift
sinful hatch
#

if you were to select 48k it kinda just makes shit up

#

any youtube/UVR dataset should prob be 32k

#

idk what exactly what youtube shits out. also verify its actually 44.1k, with a spectogram

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

sinful hatch
#

if its all mismatched, id just resample to the lowest i have and move on. really, the different sample rates dont sound different in reality unless youre being a nerd

#

just make sure it doesnt make up shit in the high end

rose drift
#

One message removed from a suspended account.

sinful hatch
rose drift
sinful hatch
#

i mean, yes to a point i guess

#

dont have extremely enraged screaming in the dataset

#

but its good to have excited tone, sad tone, ect

#

yea its good to be sorta selective and target one style of voice. but if youre using realtime, youll be forcing the model into alot of situations

rose drift
#

One message removed from a suspended account.

sinful hatch
#

what are you planning to use it for btw? realtime or inferring files with applio?

rose drift
sinful hatch
#

if youre concerned about getting good emotion, you can train separate models to target specific emotions. but that's not very good in realtime, cant switch models quick

#

RVC is pretty good at being dynamic tho, dont overthink. Some of the louder clips you shared are good to cut tho

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

rose drift
serene hearth
rose drift
rose drift
serene hearth
rose drift
#

One message removed from a suspended account.

rose drift
#

One message removed from a suspended account.

serene hearth
rose drift
#

One message removed from a suspended account.

serene hearth
#

it seems in lossy quality tho

#

like mostly

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

serene hearth
rose drift
#

One message removed from a suspended account.

serene hearth
rose drift
rose drift
serene hearth
rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

serene hearth
sinful hatch
#

oh yea copyright bad dont post that lol

rose drift
#

One message removed from a suspended account.

sinful hatch
rose drift
#

One message removed from a suspended account.

sinful hatch
#

There's enough guides and stuff that I'm sure new LLMs can explain RVC alright. Depends if they scraped aihub.wtf or rvc githubs lol

rose drift
#

One message removed from a suspended account.

rose drift
serene hearth
#

btw I once had some voice from a vn by Aniplex exe
not sure about it but I never share but the trained voice model

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

rose drift
rose drift
#

One message removed from a suspended account.

rose drift
#

One message removed from a suspended account.

sinful hatch
#

i dont trust gpt

sinful hatch
rose drift
rose drift
#

One message removed from a suspended account.

sinful hatch
rose drift
#

One message removed from a suspended account.

sinful hatch
#

that llm is thinking in the context of other ML stuff

sinful hatch
rose drift
#

One message removed from a suspended account.

sinful hatch
#

idk if mixing into one track is required but makes it much easier to see

rose drift
#

One message removed from a suspended account.

sinful hatch
#

like, youll get 200 5 second .flacs kinda thing

#

the labels are targetting anything with -42db or higher, essentially truncating the audio. But we don't want to actually truncate, this method preserves the natural beginnings and ends of phrases

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

sinful hatch
#

thats fine

#

just export it with the audio label method and itll be nice for RVC

rose drift
#

One message removed from a suspended account.

rose drift
sinful hatch
#

40 mins is nearing the limit of dataset size. Most people dont go over 1hr for sure

sinful hatch
#

saves space and same quality

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

sinful hatch
#

You could try that with your next dataset iteration. I'd just label what you already have and see how that goes before anything else tho

rose drift
sinful hatch
#

ye

#

if you dont like how that turns out, you can try experimenting with dataset changes and stuff

#

btw, youre using tensorboard to see your training progress right?

rose drift
#

One message removed from a suspended account.

sinful hatch
#

looks like that

#

good

rose drift
#

One message removed from a suspended account.

sinful hatch
#

yea just dont cook the model

rose drift
#

One message removed from a suspended account.

sinful hatch
#

set it to maximum epoch and just train it until its definately overtraining. It saves every 50 epoch (or whatever you set), so you can always grab older weights

#

dont overthink tensorboard too much, people get lost in the sauce. The 4 graphs pinned by default in Applio's tensorboard are sorta the most important

rose drift
sinful hatch
#

ngl i dont even have applio on this pc so i cant check lmao

rose drift
#

One message removed from a suspended account.

rose drift
sinful hatch
#

if you notice weird behavior, maybe ask about it. this is an example of mode collapse which usually isnt good

image.png
rose drift
#

One message removed from a suspended account.

sinful hatch
#

your G/D totals are most important.

btw, GPT can help you understand alot about tensorboard, if gpt is your flavor. This is a tool used widely by ML nerds

rose drift
rose drift
serene hearth
#

the sharp dips mean it logs when learning mute files in the last batch

#

it's kinda misleading to call it "mode collapse"

sinful hatch
sinful hatch
#

usually mode collapse

serene hearth
#

it usually means a condition where either G or D loss collapses to near zero, hindering the model improvement

sinful hatch
serene hearth
#

which means no enough improvement made to the finetuned model

sinful hatch
#

this isnt refinegan tho

#

and thats an edge case

#

im just trying to help them understand, dont want to overload with a bunch of jargon and information. I appreciate you sharing the experience but i dont think its very related here

serene hearth
#

the normal mainline rvc and applio shouldn't have mode collapse issue

sinful hatch
#

theres years of chats here they can research if they wanna get deeper. but for now, we're focusing on making sure the audio is preprocessed okay

serene hearth
#

afaik I have discussed it before with codename & noobies cat_wtf

rose drift
sinful hatch
#

tldr ask if you think the graph looks fucky

rose drift
#

One message removed from a suspended account.

sinful hatch
#

usually its ok but sometimes it fucks

rose drift
serene hearth
#

sometimes it could have unexpected results

#

that's what our fellow staff engineers have been figuring it out

rose drift
serene hearth
rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

sinful hatch
#

yos ping me if needed im like halfway active

#

and yos use your ears theyre the best

rose drift
#

One message removed from a suspended account.

void bay
#

Man

#

Uhh.. I've already made the best Kurisu you can get so ( imho )
Demo if you're interested:
https://www.youtube.com/watch?v=YVh1o_glnYA

YouTube
Codename;0
Kurisu Makise - Tsuki ga Michiru by Tsukuyomi / 月詠み 「 AI Co...

¬ Yo ~ I'm back... for now at least, lol.
I feel like I've put my whole self into this one so, hopefully you like it.
If you have any requests, lemme know!

Also, big announcement; I've been working on Violet model ( from Violet Evergarden ) so far 3-4 episodes isolated but I am doing the 4th samples rework, was too aggressive with de-reverbin...

▶ Play video
#

It's a private model of mine and the first one I ever made, so I treat her like my dearest and most precious gem she is, but in the same time..
I can understand the pain with samples you went through.
( uh.. I've spent like 2-3 months perfecting her and even the dataset alone is a legit hell lol. )
If you want, I can share the model with ya ~

Orrrrr you can keep on training your own model yeah, no issues with that,
but from experience I can already tell you she's really hard to train;
Vorbis compression and it's lack of consistency is one thing,
but finding the right hyperparms is another.

few tips;

  • carefully inspect samples you cleaned / concatenated ( perhaps ? )
    some have different frequency range response, some not.
    ( some are ' phone calls ', some are the assistant kurisu, so filtered / with effects. )
  • if you got around 8-12 mins of audio ( +/- due to silence trimming ) then you're good,
    but f it's more than that then either:
    A) you used both games ( and just one is right, don't remember which had better audio. )
    B) you included too many of ' bad samples '.

In any case. if you wanted the model, just write me a dm.

#

Gonna leave this one in here too if you wanna hear some raw recording
( I'm heading to sleep right now and will be on.. ig,in 6-8 hours if anything. )

ps. I Read a bit more in the chat and " mode collapses " aren't true mode collapses in this case, those are just places where silence or mutes were dominating / encountered.
ps.2. If you use rvc, refrain and use applio or even better if you went with my fork ( og rvc's logging is totally skewed and should not be trusted. )

rose drift
void bay
#

oh yeah

rose drift
#

One message removed from a suspended account.

void bay
#

yeah.. well, at some point I thought that perhaps I should make it priv cause back then people rarely credited

#

oof

#

might make an exception for you

rose drift
#

One message removed from a suspended account.

void bay
#

Actually, in that case try to make a good one on your own first and if it turned out too hard or with issues, I can share

#

kinda don't wanna downplay your work

void bay
#

took me 3 months to train, more or less

rose drift
void bay
#

but then, at the same time, I was also learning rvc

void bay
#

that's all there is

#

but now, it has a crucial update and even better logs

#

the adversarial loss for g was missing

#

and it was just total G

#
  • most crucial part here is the way of logging
#

mine's per epoch, applio's every 25 steps

#

which can be biased for some models and isn't " choosing epoch easily " friendly

#

( sorry for formatting btw. on phone rn

rose drift
void bay
#

logging*

#

it means, tensorboard metric saving

#

( well, model stats )

rose drift
rose drift
void bay
#

True

void bay
rose drift
#

One message removed from a suspended account.

rose drift
#

One message removed from a suspended account.

void bay
#

must be the choice of the steins gate

rose drift
#

One message removed from a suspended account.

void bay
#
  • I recognize the naming
rose drift
void bay
#

of the files

#

lol, all good

rose drift
#

One message removed from a suspended account.

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
# rose drift One message removed from a suspended account.

Actually, you can hear a lot of Kurisu in Asami if you used to watch some of her older streams but that aside
now, as for samples, from what I remember, there's quite a lot of inconsistency so I myself personally, decided to only use and trust steins;gate 0's samples

rose drift
#

One message removed from a suspended account.

void bay
#

the " static noise generator voice model " is something you can get in both applio and fork, it is not reserved to my thing but it depends on the conditions

rose drift
void bay
#

as for " the fork's for advanced people " it's a disclaimer / safety-check for myself so when some.. less caring newbies ( who can't bother to spend 5-10 mins reading ) " experiment " with switches or things they have no clue, I won't have to explain everything like they're 5

rose drift
void bay
#

around 12 mins

#

as for denoising, well.. noise-profiling is a delicate thing and needs time

#

you have to obtain all possible safe traces / zones one by one, bit by bit
and get that " full spectrum " noise

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

That part entirely depends on your preferences

#

if a model is full of ' very sharp sibilants ', it's just that the model gonna inherit it
( just like mine did )

#

but I quite like it and have no issues eq'ing if needed
( but then again, that's preferences. )

#

Anyhow, an important catch is, consistency

#

All of the sources you have most likely use different recording chamber

#

different mic

#

and so on

rose drift
void bay
#

so compression method and much more gonna vary, even proximity from mic has a lot of influence on the color

#

that's why, quality is better over quantity
( consistency.. consistency.. and consistency. rvc really hates lack of it )

rose drift
void bay
#

It's always good to verify the knowledge you gain in actual irl scenarios

#

I'd say

rose drift
void bay
#

cause a lot of things in docs and so on, are subjective interpretations

#

and afaik ( can be wrong on that. ) not everything is 1:1 with technical terms

void bay
rose drift
native pewter
#

docs are a bit outdated

rose drift
#

One message removed from a suspended account.

void bay
#

that's more or less how I have my set

#

but yeah, what's been compressed, stays that way

#

cannot be avoided

void bay
#

In any case, there's no right or wrong mh mh
just one's methods

rose drift
#

One message removed from a suspended account.

void bay
#

here's a reference point for you

#

oh

#

wait lol

#

one sec

#

wrong sr

#

mb

rose drift
#

One message removed from a suspended account.

rose drift
void bay
#

here

rose drift
void bay
#

Well, yeah

rose drift
# void bay

One message removed from a suspended account.

void bay
#

confuses rvc?

#

there is no such a thing

#

it all gets mixed up and suffled in the first place ( in the data loader or train loader

rose drift
# void bay

One message removed from a suspended account.

void bay
#

The reason you want to concatenate and silence-truncate, is because you want to yeet the silence completely

#

and then, get even and full tip-top samples

void bay
rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

mmm.. so, what exactly do you want to know

rose drift
void bay
#

other than " don't mix up too many sources ( divergent sources, that is. ) there's no wrong way to approach it

rose drift
#

One message removed from a suspended account.

rose drift
void bay
#

so let's sum up some facts

#

what batch size you tried
and your current set's length

#

and whether you train from-scratch or not

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
rose drift
#

One message removed from a suspended account.

void bay
#

if not then you're good

#

because that's the way it is, one cannot use differnt vocoders without a pretrained model like that

#

yet I gotta say, by default it has all the things set the same way applio does

#

one sec

rose drift
void bay
#

gonna show you the ui

void bay
#

just Noobies or someone else turned it off / yeeted it from the ui

#

so others stop using refine or whatever and then cry of it not working ( because of not reading carefully

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

Nah, that's fine, it is about those:
A/B/A/B format;

" I wanna use refine gan "

" There is no pretrains for it tho "

" where do I get them ? "

" .... "

rose drift
#

One message removed from a suspended account.

void bay
#

yup

#

that kind of people, so dw

rose drift
#

One message removed from a suspended account.

void bay
#

( + I believe my translations / explanations are more user-friendly or correct in actual terms, if you asked me

rose drift
void bay
#

tl;dr, the only thing that is different from applio is " double-update strategy "

#

but you can turn it off and get an exact 1:1 behavior as observed in applio, yeah

#

( because all the start / default settings are just as in applio

rose drift
void bay
#

the thing with from-scratch is

rose drift
# void bay

One message removed from a suspended account.

void bay
#

you need either A) really really lots of quality and diverge audio ( if single speaker )

#

or B) A lot of speakers ( if you intend to fine-tune it on imperfect or limited sets (( like Kurisu's

rose drift
#

One message removed from a suspended account.

void bay
#

I am afraid that she just won't click that well with "specialized-model" approach

#

Butttt, you can still try

#

you should use a batch size of 16 and if that gave you bad results.. perhaps 8 ( and only if all failed, 4 )

rose drift
#

One message removed from a suspended account.

void bay
#

for instance, VCTK dataset based pretrains ( original ones )

#

used 4 gpus * batch_size 4

#

so, the global batch becomes 16

#

so that's your reference point

rose drift
void bay
#

as in, multi-gpu setup

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

oh, you don't?

#

wait

void bay
#

you said you do

rose drift
#

One message removed from a suspended account.

void bay
#

🤨

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

from scratch = no pretrains used

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

well then, in that case

rose drift
#

One message removed from a suspended account.

void bay
#

batch size 8 to 12, perhaps 14 or 16
( for a set of around 9-13, maybe 15 mins )
is what worked for me

but you have hours of data

rose drift
void bay
#

I'd probs cap it at around 25 to 35 minutes

#

and then attempt either batch size 4, 8 or 16
.. but given she's quite emotional / not monotone, I'd perhaps be closer to picking anywhere from 8 to 12 for batch, maybe 16 if other attempts failed

#

that's just how it is, it is not something you can predict or calculate

#

teste tests and tests 😛

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

rose drift
rose drift
#

One message removed from a suspended account.

rose drift
#

One message removed from a suspended account.

void bay
#

I myself cap the max within 20-35 region, but I believe up to 1 hour could work too

#

more than that and you might be getting diminishing returns

void bay
rose drift
#

One message removed from a suspended account.

void bay
#

oh, I mean yeah

#

because refinegan has no pretrain

#

Unless you go and find one, that's what's you gonna get " the static noise "

#

because you're effectively starting the training from absolute 0, no knowledge in generator how to reconstruct the audio and no knowledge in discriminator how to judge and spot

#

btw

rose drift
#

One message removed from a suspended account.

void bay
#

I thought they do

void bay
#

because there is no " override "

#

to switch the vocoder ( which you'd tick )

#

that's all the secret really

rose drift
rose drift
void bay
#

yeah I get that part, but in the same time, LLM and transformers are different from neural vocoders and so, hifigan like that

void bay
#

can explain if you want

rose drift
void bay
#

well, there's nothing new in here, " terms " wise

#

unless you encountered something new to you?

#

if so, lemme know and I'll explain it the best I can

rose drift
void bay
#

but yea, dw too much about fork / applio dilemma

#

just go with what you find more fitting for your goals

rose drift
void bay
#

oh yeah, this convo made me realize one thing
I should mention in the ui that " non-default vocoders " need pretrains explicitly

void bay
#

oh, you wanna start again ye? with the model

rose drift
#

One message removed from a suspended account.

native pewter
rose drift
rose drift
rose drift
#

One message removed from a suspended account.

void bay
#

yeah sure, we can try to think of some workflow

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

alrrr

btw gonna be back in 10-15 mins

rose drift
#

One message removed from a suspended account.

void bay
#

ye, back

#

soz, took a lil longer

#

@rose drift

rose drift
rose drift
void bay
#

I heard you want the project for nsfw purpose

rose drift
#

One message removed from a suspended account.

void bay
#

And I have to make it clear right away, if that's the case I can't and won't support it ( nsfw models in general too, don't get me wrong on that ~

rose drift
#

One message removed from a suspended account.

void bay
#

Alr, then we all good

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

You can continue

rose drift
#

One message removed from a suspended account.

void bay
#

oh, nah

#

been listening to asmr since 2013 or so, so we good

rose drift
#

One message removed from a suspended account.

void bay
#

yeah, in that case hmm...

#

I think it can be quite hard

#

yet, it is not impossible

#

You'll need KLM pretrains

#

Actually, I believe I used to want her asmr too
but didn't work too well sadly ( Back then we had no klm and such, obv

rose drift
#

One message removed from a suspended account.

void bay
#

bruh

#

well, then that is indeed a sub-set of nsfw if you account for groans and such

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

but in general view, groaning and " arousing " sounds are nsfw

#

But I get it if you need waifuu chems, let's put it like that.

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

I mean, I have no issues with what you wrote specifically

#

just simply claryfing that generally having " kisses " and " groans " in the same chain, is quite nsfw

#

But still, you'll need KLM or such specialized pretrains

#

og pretrains can't handle such whispery and ' misc ' content like that well tbf

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

well I only said that because most of " groan, kisses " like asmr quite often end up having " ear licking, moaning, ' yosh yosh good boi ' " kind of content, so it's usually safe to assume nsfw.

anyways, guess our opinion just differs, that's fine.

#

Now, like I said, you're going to need klm pretrains or such

#

I believe there's no other than KLM that could potentially support such ' extras' that well

#

Have you heard of those klm ones?

rose drift
#

One message removed from a suspended account.

void bay
#

Yup so, as of right now? I'd recommend klm 4.9
( at least until we get ' new gen ' pretrains that use new embedders n such. )

void bay
rose drift
rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

And with that ^ gonna test the new embedder

native pewter
rose drift
void bay
#

klm is quite potent in a lot of things but yeah, non-og pretrains is def a must

#

and I also hope for the new embedder to be helpful in lots of things

void bay
#

but doing what I recommended gonna increase the chances of a success

#

so as always in ml, you should give it a try

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

well, a good model is always a big W

rose drift
void bay
#

+, if you do have a good set, there's always hope in future

void bay
#

but generally, if it's such " vocalized " stuff like ' groans ' and such

#

I am fairly sure she could somewhat be fine

#

because I know she does make certain type of sounds / vocal frying, at times.. ( at least in sg0 n dramas, don't remember much details about the rest

#

so again, you should def try and hope for the best

rose drift
#

One message removed from a suspended account.

rose drift
#

One message removed from a suspended account.

void bay
#

some time-lines are crazy

#

as hell

rose drift
#

One message removed from a suspended account.

void bay
#

xd

#

true tho

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

Don't stress man lol, I already get what you mean

#

you good

rose drift
void bay
#

Oh yea, these

#

nah, she's fine with such

#

I promise you ( or well, used to be for me during rt voice changer performance testing iirc

#

so you should be fine, mostly

rose drift
#

One message removed from a suspended account.

void bay
#

I'd say, as long they're not dominating the dataset, yes, it wouldn't hurt

rose drift
#

One message removed from a suspended account.

void bay
#

key idea here is that they should occur " naturally "

rose drift
#

One message removed from a suspended account.

void bay
#

so, not without context / randomly injected inbetween sentences or words

#

yeye

rose drift
void bay
rose drift
#

One message removed from a suspended account.

void bay
#

and if some are " loose "

#

also good

#

as in, there was :

okame: blahblah
kurisu: groan
okabe: blahblah
kurisu: groan

so you'd have: groan, groan in ur set, after concatenation ( voice-lines chronologically wise

#

that is fine too

#

ps. Always export as 32 bit float

rose drift
#

One message removed from a suspended account.

rose drift
void bay
#

only time you don't really have to is if you don't touch the volume / dynamics

void bay
#

I remember the Lol one

rose drift
#

One message removed from a suspended account.

void bay
#

really liked processing such
( am quite sensitive to such, asmr wise

#

xd

rose drift
#

One message removed from a suspended account.

void bay
#

ps. Spectral de-noise is your best friend

rose drift
#

One message removed from a suspended account.

void bay
#

avoid AI denoisers by all means
( they can damage stuff, esp her samples as they are compressed to begin with

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

yup

#

also one sec, I can check if I still have some old kurisu noise profiles

rose drift
#

One message removed from a suspended account.

void bay
#

yup, seems like I do

#

just not sure which one are those
( I had one for dramas, one for vn

#

so, gonna send these and you can try it out, perhaps

#

( also got one for that nasty line in upper spectrum

#

" fking line "

#

lol

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

You know wut, you kinda motivated me to try and make her V2 variant

rose drift
#

One message removed from a suspended account.

native pewter
void bay
#

Kurisu arc 2025, soon ™️

void bay
#

btw, if you still have sources for your things, perhaps you could share em in dm ( also, dm cause discord is touchy
I believe I lost all my .txt from back then when switching drives a while ago

#

and it'd be helpful lol

rose drift
#

One message removed from a suspended account.

rose drift
void bay
#

I might have heard it but won't give my hand for it

#

so quite likely I haven't found that one, or missed

rose drift
#

One message removed from a suspended account.

void bay
#

oh, then that's probs why

rose drift
#

One message removed from a suspended account.

void bay
#

I assumed app's gonna have LQ audio

#

and decided to skip it just in case

#

welp

rose drift
#

One message removed from a suspended account.

void bay
#

low quality

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

lq, hq

rose drift
#

One message removed from a suspended account.

void bay
#

tho ye, how's samples from the app?
48khz or 44.1?

#

or 32/36khz ~ ( some devs decide to go this route for whatever the heck reason

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

Oh yeaaa

#

that's actually promising

#

yey

rose drift
#

One message removed from a suspended account.

void bay
#

but then, I'd have to compare the spectrums later
( The " distribution " of frequencies, on avg - cause if there's an inconsistency like that in a dataset... well rip the model's performance )

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

#

One message removed from a suspended account.

rose drift
#

One message removed from a suspended account.

void bay
#

Audio generally is complex

#

I'd even go as far as saying, it is way more complex than tokens or semantics ( even tho I know shit about llm's from technical standpoint

#

cause for instance, the problem of phase reconstruction in ML is still a huge issue and not perfect

rose drift
#

One message removed from a suspended account.

void bay
#

yuh

#

old times, dead past

#

but more or less, yes

rose drift
#

One message removed from a suspended account.

rose drift
void bay
#

Oh, well..
I mean, aside of music making, I have experience in utau and vocaloid

rose drift
#

One message removed from a suspended account.

#

One message removed from a suspended account.

void bay
#

ml / ai / hifigan came later
( And Kurisu was my test-subject )

#

I wanted to make an " amadeus " assistant

#

first goal was to make a model.. then I got to know about rvc

rose drift
void bay
#

lol

rose drift
#

One message removed from a suspended account.

void bay
#

and lost the passion for further work ( learning llm n shit

#

but I might attempt it one day

rose drift
void bay
#

and here we are, yea

rose drift
#

One message removed from a suspended account.

rose drift
void bay
#

Anyhow, what's your plan now?

rose drift
void bay
#

oh, nahh, I can do just fine with the sources ( I like that part of the work, own commitment

void bay
rose drift
#

One message removed from a suspended account.

rose drift
void bay
void bay
#

websites you used, torrents etc etc, whatever you have
( hence mentioned dm, as discord's very itchy about such

rose drift
void bay
#

mh mh ✨

rose drift
#

One message removed from a suspended account.

sinful hatch
#

😭 this is codename's grail model I didn't even know

void bay
#

oof

#

shhhh

#

I remember how I used to compete with 1sky

#

iirc

sinful hatch
#

Trust Cody he might know what hes talking abt lmfao

void bay
#

I mean, nono, I did cut in

#

I can provide feedback n stuff, but yea

sinful hatch
#

Esp since you've handled this dataset yourself lol

void bay
#

welp

tardy stream
#

This gonna be one of the longest conversating threads I've ever seen on Discord.

sinful hatch
serene hearth
#

real

void bay
#

been a while ~