or hm i could get a billion vocals from this http://beta.ccmixter.org/ and then train a general singing model for people to finetune ...or more realistically get someone else to finetune it because repeatedly running a colab for a week doesn't sound that fun

what is that

free stems for indie songs

woah

nice

now that i think about it this could be used to train vocal-remover in reverse, giving it song-vocal pairs instead of song-instrumental

but alas

why do they call acapellas `pell`s http://beta.ccmixter.org/pells

a cap**pell**a

yeah ik but Why

it's the internet, you're not gonna run out of letters

in fact the non-beta site says

A Cappellas

(though oddly still puts

pells

in the link)

the internet's all about shorthand

that's deadpool btw

yes

hmm would anyone in the universe be interested in a singing model of a *non-*celebrity *non-*character voice just some random singer

@User do you own a Jetson nano?

i don't

the person asks about inference on a jetson nano, training was just proposed as a solution and here's thing blah words

my gpu has a mere 4gb of dedotated wam

noice!

Is this normal? Epoch 9..

yes, thats really good

the more diagonal, the better

Even epoch 0 was good

Well somewhat broken but fine

how many audio files are training with?

271

What validation loss do i stop at?

they say to stop when it goes below 0.15