Ight

gonna see if cd-i impa trains well with only 13 lines

well that's promising

epoch 14

oh

default female voice, not enough

mine is doing this weird zoom in thing where the encoder and decoder timestep is smaller again

How is this?

looks good so far

stonks!

I just got a gpu upgrade with 11gb, I'd train it for you if I have time to set it up It's only a 1080ti so idk how long itd take, no machine learning cores for fp16 optimization

If you prep all the datasets for me and point me in vaguely the right direction I'll do it

I wanna mess with a multispeaker singing model ;) But tbh, isnt melgan already trained via that? I've also heard multispeaker models only really benefit the vocoder, not really the synthesizer.

this is how my new character sounds 150 epochs in, i dont know if it sounds like the character or not

My model just went from .15 to .25 :anger :

if it goes up drastically like that, then i think you should stop it

Ok

Ill wait for the line to fix up a lil

This is same when I was trying to train tiger's voice from an american tail

ooh! that'd be quite epic as for dataset oop turns out there's this mega-bundle http://multitrack.eecs.qmul.ac.uk/ (unfortunately not very well tagged but somewhat well -ish) training is just "put the full songs in one folder and the instrumentals (in our case vocals) in another and run train.py"

as for singing model i was thinking less multi-speaker and more pretrained, though i certainly wouldn't mind if the former was possible

Have any data on how far you'd want to have it trained? Sample outputs on similar datasets?

Wanna make sure it's repeatable and proven to work before investing so much electricity into it

in that case you might not wanna do it

it's just an off-the-cuff "hey what if"

https://github.com/Anjok07/ultimatevocalremovergui Heres some pretrained remover models

i know and have used this, but afaict it's tuned more for instrumentals than vocals

Ahh

But it works?

and i can't find a way to improve the latter within the thing itself