DMR V1 | Mangio-Crepe / English / A Fine Tuned Pretrain For Mommy Voices ~

[silver tusk]

DMR Pretrain V1

DMR Download D: https://huggingface.co/Razer112/DMR_Pretrain/resolve/main/D_DMR-V1.pth?download=true
DMR Download G: https://huggingface.co/Razer112/DMR_Pretrain/resolve/main/G_DMR-V1.pth?download=true

The DMR pretrain aims to improve e-girl voices and deep male voices.

Dataset Info

• Length: 11.3 Hours
• Total Voices: 22
  • 16 Female voices
  • 6 Male voices

Model details

This was trained on Applio with a 4060 TI 16gb on a batch size of 12 and is trained with Mangio Crepe at a 32 hop length with a sample rate of 32k.

Samples:

https://discord.com/channels/1159260121998827560/1255616722480926860
https://discord.com/channels/1159260121998827560/1254834021368598608
https://discord.com/channels/1159260121998827560/1256387296677072988

Special Thanks

I would like to give a BIG thanks to @vestal dawn for helping me finish the dataset for DMR he was a massive help ❤️

How to use:

1. Download G and D pth files from Hugging face
2. Place it in the pretrained_v2\custom pretrained folder
3. Paste the name of the pretrain into the pretrains box
4. Youre ready to make cringe voices!

[echo roost]

First

[unreal plank]

Does this pretrain work well with whispering?

[silver tusk]

no

no pretrain can make whispering work

its an rvc issue

[oblique socket]

rmvpe could improve but mangio crepe is great tho

[silver tusk]

Crepe is better with softer voices

[oblique socket]

That's great

[stone garnet]

I am afraid of this pretrain

[echo roost]

pretrainophobia

[stone garnet]

Womanphobia

[echo roost]

Um actually, its Gynophobia.

[solemn mirage]

I tried using it but RVC v2 Disconnected gave me an Error

[silver tusk]

What error?

[solemn mirage]

I can't remeber sadly but i tried to use custom with both links liked to the G and D
And it wasn't able to use that so i had to switch back to original to be able to train my Voice earlier

[silver tusk]

Maybe you accidentally pasted the G link into D

[solemn mirage]

Hmm i made sure that i didn't and looked 3-4 times before i was proceeding

[silver tusk]

Yeah I just tried it and it works

[solemn mirage]

I used Google Collab RVC v2 Disconnected

[silver tusk]

I just tried it on there

[solemn mirage]

Hmm idk now

I putted links there maybe that was false

But its fine

[oblique socket]

Is it possible to do pretrains on cloud?

I always get errors on Kaggle and Colab for being AFK

[silver tusk]

Like to train them?

[queen crag]

Ayo? @silver tusk level 21 !!!

[oblique socket]

Yep

[silver tusk]

You can

[oblique socket]

How?

[silver tusk]

Just train them like any other model

I suggest kaggle bec it has more GPU hours then colab

[oblique socket]

U can use around 6 hours i guess

or 3

[silver tusk]

On kaggle I get 30

[oblique socket]

30 per week

[silver tusk]

Yeah

[oblique socket]

But i tried once and i got disconnected due to AFK

[silver tusk]

After how long?

[oblique socket]

3 hours, i kinda dont remember (i was training it on multinotebook)

[silver tusk]

Huh i ran kaggle for like 4 hours and didn't get afked

[oblique socket]

Whoa

Maybe i'll try one day

because i'm planning to do a castillan spanish pretrain

[oblique socket]

Oooh

Yeah

I was training a Sonic model for 3 hours and i didn't get afked

But i'm sure that a average pretrain (50 epochs + 32k rate + 2hour data) should take more than 6 hours i guess

[silver tusk]

Cool, how many hours do you got?

[oblique socket]

I'm planning to do 2 hours of castillan spanish speech

50 epochs, batch size 16 and 32k sample rate support

[silver tusk]

That's good for a start but if you really want to hear a difference you need at least 10 hours

[oblique socket]

Maybe if i get a rtx 3090 or a 4060 i could do longer hours of dataset and train locally

[silver tusk]

Training a pretrain is like training a model so that means it can overtrain/ undertrain

[oblique socket]

I'll try with 50 epochs as a start

If it undertrains i'll increase the epochs a bit or reduce batch size

Tho i wonder if people use YT to source datasets for a pretrain

[silver tusk]

I do

[oblique socket]

Maybe if i can find hours of castillan spanish podcasts on youtube i'll use it as dataset source

[silver tusk]

That would be good

Just be sure it isn't super compressed

Also variety is needed so get like ~30 minutes of audio per person

And get a ton of people

[oblique socket]

30 minutes from juanita, pablo, etc.

I'll use random voices, as long they sound HQ

[open stag]

Btw, added this pretrain to the index in the pretrains guide !

also, what is this pretrain finetuned on ? on the Original Pretrain?

[silver tusk]

Yes, this is finetuned on the og pretrain

[open stag]

Alright

[silver tusk]

The sample you linked for my pretrain leads to this model https://discord.com/channels/1159260121998827560/1254196145492856976 instead of an sample of my pretrain

[open stag]

Ah my bad, might got confused as i added also another pretrain

[open stag]

Just fixed it

[silver tusk]

nw

[silver tusk]

thanks 🙂

[open stag]

yw

[foggy pumice]

dominant mommy roleplay

[silver tusk]

Im in shambles. I did some testing of my pretrain and i found out that it adds noise to models but doesnt distort harmonics

DMR:

OG:

[normal falcon]

i wonder what kind of dark magic rvc-boss did to the og pretrain, because the vctk dataset is noisy af

and rmvpe didnt existed when he made it

[silver tusk]

did he use crepe?

[silver tusk]

if only rvc wasnt abandoned 😭

[normal falcon]

i don't think so, looking through mainline updates it seems crepe got added very briefly as a pitch extraction method for a couple of weeks/months, then got replaced by rmvpe
v2 pretrains were already a thing tho
edit: the first part of this message is true, but crepe was never in a public release, the second part is false, rmvpe was already a thing when v2 pretrains were made

[silver tusk]

huh

time to use every pitch extraction to see if they make a difference

[normal falcon]

Oh wait
rmvpe was already a thing when v2 pretrains got released

since they deleted crepe off existence even before releasing v2
there's a high change the og pretrains v2 use rmvpe xD

[stone garnet]

I wanna know the best one

[normal falcon]

mangio for high quality datasets that are properly cleaned, mangio hates any type of noise residual
rmvpe for everything else, can also work in high quality datasets

rmvpe sometimes adds a metallic sound to the model
mangio is less pitch accurate than rmvpe

[stone garnet]

Thx 👍

[marble prawn]

Can you even imagine a pretrain for non voices

[stone garnet]

there should be one

OV2 is the closest bc it works best from my experience with non voice models

[main tangle]

BeatzForge already exists

[oblique socket]

but its only for drums

[oblique socket]

Probably hifigan limitations.

Leo told me that hifigan is limited for custom pretrains

[unreal plank]

Is it because they added KL loss? The original paper didn't have that I think.

[zealous kraken]

no idea, gotta wait for anyone implementing bigvgan v1/v2

probably till RTX 5090 is released

https://tenor.com/view/mr-bean-mrbean-bean-mr-bean-holiday-mr-bean-holiday-movie-gif-11625313334009911234

[unreal plank]

Why not jump directly to Eva gan, which claims to be an upgrade over bigvgan

[oblique socket]

eva gan?

[zealous kraken]

I mean bigvgan v2, which is actually better than evagan and even vsan

[oblique socket]

rtx 5090? u cant just use 4x or 8x h100 to train bigvgan v2?

[zealous kraken]

what

[oblique socket]

like, h100 is faster than rtx 4090 iirc

[zealous kraken]

the price is also much "faster"

[oblique socket]

i imagine that the dataset like dis:

[oblique socket]

a price can't be fast

[zealous kraken]

comparing owning 4090 to the H100 rental price is unfair

[oblique socket]

finally found this gif