because almost any amount of noise if harmful

why use bad data if more than enough good data exists?

True I guess to my untrained ear I couldn't tell that noisy was even bad/harmful in the first place

but why use bad data if more than enough good data exists?

For example something like is labeled Noisy

it was something that I picked up when I started playing around with the first Sunset Shimmer voice by Jeremythepeanuts. it sounded like having a stroke every single time.

you realize there is no voice in that?

it's just onomatopoeia

probably that's the reason it was labeled that way

I could easily cut out the Noisy that's not a problem

strongly modulated voice, not ideal to use

but why use bad data if more than enough good data exists?

?

?

?

I'm asking for why these labels are not good enough

I can cut them out

just answer pi Holos, they won't stop until you do

I'm not the guy who labeled it

I just want the justification

So i guess modulation is also not good

interesting

I'd assume when you do the same task over and over again you also get really sensitive to minor disturbances

and most of the mlp dataset was labeled by a single person iirc

Got you yeah ok cutting them out is pretty straight forward so regardless if they sound ok i'll cut them

train on whatever subset you want lol

To my untrained ear the noisy ones did not seem far too different than the regular neutral ones

but why not just use the subset labeled as clean?

I mean what is to gain from that?

Doesn't sound impressive.