KLM - BeatzForge V2

[arctic crystal]

This is Pretrain model for Drum Boys.

SeoulStreamingStation, RegalHyperus, Miqus, Poopmaster

BeatzForge V2

Sorry for the delay. It took a long time to find and remove all the bass guitar sounds from over 500 files. I have removed all the bass guitar sounds in the sub-bass range corresponding to silence and the range corresponding to Tom. Some tests showed no major issues, but further testing is still required.

Model Link 32k (only)
G - https://huggingface.co/SeoulStreamingStation/KLM_BEATMASTER/resolve/main/G_BeatzForge_V2_32k.pth?download=true
D - https://huggingface.co/SeoulStreamingStation/KLM_BEATMASTER/resolve/main/D_BeatzForge_V2_32k.pth?download=true

BeatzForge Test Model Link
https://huggingface.co/SeoulStreamingStation/KLM_BEATMASTER/resolve/main/BeatzForge_Test_Model.pth?download=true

[viscid heron]

@mystic lintel 's dream becomes real

[arctic sleet]

Literally

[viscid heron]

I was going to do this pretrain before but i got AFK ahh problems with Kaggle

[pale dagger]

[viscid heron]

[pale dagger]

[arctic crystal]

ew.. you guys again!! lol

[viscid heron]

we love chatting

[arctic crystal]

[viscid heron]

i wish i could do pretrains without AFK problems lol

[pale dagger]

Everybody ping @mystic lintel

[viscid heron]

(kaggle/paperspace)

[arctic crystal]

btw, how do I make sample audio with this... lol

[viscid heron]

@mystic lintel

[viscid heron]

with drums from random songs

like how hyperus does

separate a song that has drums using mvsep and demucs 4 ht, and use the drum separation as a inference base audio

[pale dagger]

PD: Seoul, you'll may need Demucs to isolate drums of any model.

[arctic crystal]

err.......

[viscid heron]

beatboxing sounds

[arctic crystal]

lul

oh beatboxing!

[viscid heron]

i wish i could do a fnf chromatic pretrain (2 hours of fnf character chromatics if its possible)

for fnf models lol

could u do a comparison of a drum model with the rvc og pretrains and klm beatmaster? if you can ofc

[valid breach]

This is a pretrain of all time

[viscid heron]

RegalHyperus' dream

[arctic crystal]

omg.. you guys opened another chat room here lol

I'm out! good luck!

[arctic sleet]

Wow thats a huge difference

[viscid heron]

the drum pretrain gives a amazing improvement lol

essential for regalhyperus

[arctic crystal]

LetS gO Drum Boys!

[pale dagger]

Everyone ping Hyperus

[arctic crystal]

[spiral palm]

Ayo? @viscid heron level 52 !!!

[arctic crystal]

Thanks!! 52 lev!! congrat!!

[mystic lintel]

HAHAHAHAHAHA

[mystic lintel]

TFW you summon RegalHyperus/Hyperus18 himself

[mystic lintel]

Next drum model: Lone Digger
I'm using this pretrain

[arctic crystal]

hope you like this. actually I forgot "rim shot" in this model. but still it will be better than normal 'voice model'. I will update "rim shot" soon

[mystic lintel]

Sure!

[mystic lintel]

It's here

Lone Digger drum model just released

[gusty sentinel]

BUN BUN CHA!

[arctic crystal]

poi~!

[gusty sentinel]

why the 'tss' sounds like that?

[spiral palm]

Ayo? @gusty sentinel level 41 !!!

[arctic crystal]

I have no idea. lol I never made drum sound with RVC before

[gusty sentinel]

haven't tried and seen if anyone could make some violin model and do vocal to violin cover well

[mossy finch]

it's just how the GANs work, it can't really randomize the higher frequencies sometimes

[willow viper]

lusbert would touch u for this

[mossy finch]

test batch 10, 400 epochs. I think this proves that pretrains are just for noise arrangements and harmonic details. The model could have improve if the loss/g/fm rises up

surprisingly, this pretrain did upscale the audio, the dataset has 20khz cutoff

[willow viper]

youll have competition

[arctic crystal]

[spiral palm]

Ayo? @arctic crystal level 33 !!!

[arctic crystal]

Beware! competitor! now we have very strong supporters!

[mystic lintel]

Oops, embed fail

[viscid heron]

I wish that too lol

[mystic lintel]

Same here

[viscid heron]

For what?

[mystic lintel]

I use Run Selection on Colab

[arctic crystal]

Guy, we have some bad new here.
First, I'll share three samples. I'm not sure why, but I noticed that the sound gradually disappears as the training progresses. Specifically, the bass part starts clipping more and more, though I don't know the exact reason. However, it's clear that the issue becomes more pronounced as the training continues.

epoch 50

epoch 70

Epoch 100

[mystic lintel]

It's overtraining

[arctic crystal]

it's like 5~6 hours of dataset. overtraining at 70 epochs?

KLM - BeatzForge

[arctic crystal]

Uploaded!

[willow viper]

do you want my drum data?

i did oopsie and i dont want to train again😭

[arctic crystal]

Yeah, let's just combine all of them. for the drum bois. I'll put your name on credit.

[willow viper]

alr ill dm u it

u can delete mainP files i send it before to poopmaster so probably they are in dataset

[mystic lintel]

Then there's this that keeps popping up
We need to fix this

[arctic crystal]

can you send me audio sample?

[arctic crystal]

Thx!

[viscid heron]

Seems to be too sensitive to noise

[pale dagger]

This one or KLM4?

[arctic crystal]

this one

fixing now~

[pale dagger]

And KLM4 Test2 is still sensitive to noise too?

[arctic crystal]

no

[pale dagger]

Really?

[arctic crystal]

it's not reacting with whitenoise anymore.

yupyup

[pale dagger]

It got between 70-80 hours of korean/english/japanese audio, right?

[arctic crystal]

yup as your feedback. we included lots of mid-quality audios with some noisy stuff.

there won't be any instances of human voices sound popping out in the white noise.

[pale dagger]

Have you tried comparing Titan results with results of your KLM4 results?

[arctic crystal]

sort of?

[pale dagger]

I mean, comparing a model trained with TITAN pretrain with one made with your pretrain.

[arctic crystal]

Hmm, to reiterate,yes we always conduct tests. however, we don't directly mention specific pre-trained models made by others while testing. We don't want to create a competitive atmosphere by comparing our models with those made by others, as they have all put in their best efforts. We aim to solve problems based on feedback from users or testers like you, rather than updating based on comparisons with specific models.

[mossy finch]

I still don't get the term "sensitive to noise" is, pretty much any audio has noise, sibilance is a type of random noise, noise between harmonics are also noise. Like what's the point? haha

imo tho, still having noise free dataset are a must, because how it works is like rvc will base it off the noise found in the pretrain. Like if the gt dataset has weird noises, the output model will also have weird noises. But when it's trained on a clean pretrain, higher chance that the weird noise found in the gt dataset will not be included, because the pretrain doesn't have weird noise in the first place.

there's a reason why the OG pretrain will come out noisy, because the source material that was used for the pretrain was also bad in the first place

[viscid heron]

If i use Titan as finetune when i create a pretrain, results are better or worse?

[arctic crystal]

you're right. that's why having clean inference samples is much more advantageous in terms of quality. however, sometimes, in some samples with severe white noise, certain voices tend to react to this noise. At first, I also thought there wouldn't be any problems as long as we had clean sources and inference samples, but there were cases where the model inferred human voices from low-frequency noises that are inaudible to the human ear. more importantly, while people pay a lot of attention to the dataset when creating models, they often neglect to properly manage the inference samples. so, a kind of trick is used where we introduce "inaudible low frequencies" to counteract these low-frequency noises.

[spiral palm]

Ayo? @arctic crystal level 34 !!!

[pale dagger]

So SCR, do you mean we should try using normal-quality datasets on HQ pretrains?

[mossy finch]

I'd say yes, and much better if HQ datasets on HQ pretrain

[pale dagger]

Can you someday make some tests?

[mossy finch]

like what test?

[pale dagger]

Good/normal quality dataset in OG pretrain and the same dataset in an HQ pretrain

[mossy finch]

OG and KLM? soon

[viscid heron]

i'll retrain this with titan, nvm, this sounds buggy af

[arctic sleet]

Dang that is bad

[viscid heron]

I'm kinda frustrated

That's why i don't make drum models sometimes

[arctic sleet]

I usually like to use the OG pretrains for drums or non voice models

[viscid heron]

https://tenor.com/view/ishowspeed-try-not-to-laugh-gif-7682731162751353849

[arctic sleet]

I listened to like half of this and when I pause it my hearing sounds like its bitcruhed but not really lol

[viscid heron]

i hope that for beatzforge v2 this issue gets fixed

i dont think that this is the adecuate moment to use the pt

[mystic lintel]

I just silence that bass using Ctrl-L in Audacity
I even sampled it in a multisource YTPMV soundfont

[mystic lintel]

I was already aware of this problem, but from my drum models, the audio was pretty much salvageable

[arctic crystal]

now I'm fixing this problem. so yes do not use this model until v2

[mystic lintel]

Any updates?

[viscid heron]

Nope, i guess

[arctic crystal]

KLM - BeatzForge V2

[arctic crystal]

sorry for delay! try version 2 and let me know if you have any issues!

[mystic lintel]

Next drum model:
HEYYEYAAEYAAAEYAEYAA

[viscid heron]

model trained with drums from a live rehearsal

[viscid heron]

510 epochs, batch size 4, beatzforge pretrain

[arctic crystal]

hew, does it works?

[viscid heron]

ofc

[viscid heron]

i wonder why this happened? maybe guitar bass on dataset?

[mossy finch]

here is a seen speaker test, audio with and without offset in timing. Problem arise when rvc can't fully recreate the waveform, it creates phase issues and weird delay effect, kinda sound like splicing frequency when changing audio pitch

[arctic crystal]

to put it very simply, the amplitude and frequency of the bass and drums are almost the same or identical. The electronic sounds often included in the drums have the same frequency as the bass. As a result, the model learned the sounds of the bass and electronic instruments, which were inferred instead of the drum sounds.

[arctic crystal]

interesting.

[viscid heron]

@arctic crystal KLM - StringzForge next? (guitar pretrain)

[mystic lintel]

Or maybe RepliRiff

[arctic crystal]

nope, we dont have any plan to make new ptm

[mystic lintel]

It's 32K man

[cursive forge]

OK, Thanks For The Tip

[viscid heron]

loren85 alr took the idea

[gloomy sleet]

i think this is the same for all the other pretrains

[pale cedar]

I'm kinda wondering if I should redo my minecraft skeleton with this pretrain

[tropic coral]

I'm gonna try out this Pretrain

[viscid heron]

no freaking way that og pretrain makes drums sound like that

[arctic crystal]

Pff-hu-ha! Ssu-hu-bu-ta! pff-mh-pa!

[mystic lintel]

WELCOME BACK

[arctic crystal]

eh? I've always been here, just not chatting much

[pale cedar]

[wary seal]

scream mj drums o-o

[viscid heron]

yeah