Young Nudy / EA Monster Era (200 Epochs) / TITAN / RVMPE

[thick lily]

Not sure about this model , i inferenced using content vec and it made it sound wierd, results may vary and sound better **

This model works for rapping / sing may not work since most of it was rapping.

📊 Dataset: 15:00

🔄 Extraction: RVMPE

🎵 Sample-rate used: 40K

⚠️ SINCE IM ON MAC, THE DOWNLOAD LINK WILL NOT WORK (ZIP) SO I AM GOING TO BE UPLOADING THE PATH / INDEX SEPERATELY WITHOUT ZIPS! ⚠️

Index: https://huggingface.co/bankplayz/ai_moel/resolve/main/added_IVF1283_Flat_nprobe_1_youngNudy_v2.index?download=true

Model Path: https://huggingface.co/bankplayz/ai_moel/resolve/main/youngNudy.pth?download=true

if you use this model credit me and just let me know what you think !

SAMPLES DOWN BELOW!

[shut apex]

Rate the model

To rate this model, either use /rate or click the buttons below.

[thick lily]

am i tripping or does my model have a mic bleed ??

lmao

can someone lmk

sounds like there some clipping or like bleeding in the ai

@deep shoal yo bro sorry for tha ping but i need ya help

does this model have a bleed ? or am i tripping

i hear like a mic bleed

when barely any of the dataset has any bleed at all

[deep shoal]

u hear like a click sound? hmm did you applied mouth-declicker to your dataset?
besides that, if the dataset had no clipping/clicks sounds, it could be the pretrain, as titan is known for adding exactly the sound you're describing (like a click/clipping sound)
make sure your dataset doesn't have clipping or clicks
if you're 100% sure your dataset lacks any clip or clicks, is titan

[thick lily]

i mean theres barley, but i dont think it would be clicking , its more of a clipping sound, like a bleed when the ai gets like strong or loud if yk what i mean

and no i forgot to add the mouth declicker

but it didnt need it

more than like 80 percent of the dataset has any clipping at all

[deep shoal]

even a small portion of clipping can cause rvc to have a bias towards it

[thick lily]

you think if i use the defualt pretrain it might be less harsh? because it is pretty hard the clipping in most of the samples i provided lmao

[deep shoal]

yes, original pretrain lacks any issues titan has

is also less robust

(less harsh)

[thick lily]

but isnt titan a finetuned of the original adding more varitey and like better data to it? or thats not how it works

[deep shoal]

also make sure your dataset is indeed 40k because if is not in 40k, rvc will try to generate the missing frequency (this translates as artifacts)

[deep shoal]

is intended to help smaller datasets (this is why the hugging face compares it to ov2)

[thick lily]

yeah checked the spectragram on spek and it hits 20-21k the average

so 40k

[deep shoal]

oh fine

[deep shoal]

is not really how it works, pretrains only guide the sound during the finetuning process, it tells rvc "this is how this sound should works"

is a bit hard to explain

[thick lily]

damn now i gotta retrain with the normal pretrain 😭 this took me 2 and a half hours to do im cooked

[deep shoal]

when you're training your dataset you're also making a finetune

titan is exactly that, a regular rvc model used as a pretrain hehe

[thick lily]

im afraid the clipping might happen again tho and might waste my time ngl

[deep shoal]

yeah, is more safer to delete the clipped audio

or try to fix it in izotope rx/any other editing software

[thick lily]

but its tough because im looking through the audio and its like barely any clipping

like very minimal where i can barely hear it

and its 15mins of dataset

i would have to look carefully thru

[deep shoal]

rvc does a bit of audio upscaling, so the sounds that were barely audible, will be easily audible after training

[thick lily]

any way on izotope to tell if theres clipping? or like clicking

should i just use the demouth clicking tool thingy

[deep shoal]

yes, use this, open your dataset and click that

things that rvc hates:
dc-offset
clipped samples

[thick lily]

gotcha apprecaite it

ill juss use the normal pretrain

next time

[deep shoal]

good choice, believe me, is the best pretrain we have

because it doesn't damage our datasets at all

it just works like intended, so if your model has some issues after training is done, you can be 100% sure is your dataset and not the pretrain

unlike if you were using titan for example, sometimes is your dataset, sometimes is titan

[thick lily]

guess ive been lied too why was i told its better than / is it usually better in some cases?

or default pretrain good for everything

[deep shoal]

default pretrain good for everything

[deep shoal]

titan has better gradients, this translates in more stable training, and also is more robust
basically helps smaller datasets sound more clear

but we found about the noise issue just recently

[thick lily]

dam well apprecaite the help imma train my next model with the default pretrain

[thick lily]

the clipping ?

or just noise

in general

if ykwim

[deep shoal]

the clipping here seems to be your dataset and titan at the same time
the "noise" titan adds is like a clicking sound

similar to clipping

here u can notice it more easily
(This audio is not mine, seoul made it, is the titan pretrain voice)

around 0:04-0:05

[fringe ledge]

Omfg why are you all spammin here. You guys have private dms for a reason 😭

[thick lily]

ehh because why not

[deep shoal]

lols ok
tldr; original pretrain for all, if you notice issues, is your dataset xD (avoid clipping audio, clicks, and is going to work fine)
also i recommend #📑│making-models in case you need dataset related help, known rvc model makers are here, is the best place to get help for training