Want ya guys attack a spawn?

"Buncha guys Attackers Spawn"

ooh

odd

I'd recommend only using Wavs that is longer than 2 seconds in this instance

although it is highly advised to use files longer than 1.5 seconds

Or just bigger lengthen the short lengthed audio files

ill try to pick some long files along the way but i need as much as possible considering that theres prob not 100 in the files

You can try combining audio files

oh gosh so many of these are like .5 seconds

this might not work out

i need more like these

everywhere at the end of AI

Hey, I just want to confirm, all we need to do to submit a model composed of a zip file with, like,

foo.wav

optimally around 15s long and

foo.txt

with a plain-text transcript of it, repeated at least fifty times? Does it help if we make more than fifty? Can these models be reused on future models to improve performance? What can we do to make the models less prone to that weird janky sound you get if you put in more than a sentence or two? (Also, this random guy on Reddit says that tacotron2 is outdated? https://old.reddit.com/r/VocalSynthesis/comments/gv1aub/anyone_have_any_good_resources_on_the_tacotron2/fsm9jy9/)

???

Oh! Derp. Thanks! Time to go cut up an audiobook or podcast or something!

oh yeah 👍

amazing people over here: https://discord.gg/e6EWZaRK have a lot of expirience with making models via audio books

check em out!'

yes they are awesome

I tried everything to fix Woody, reducing samples from 52 to 44, fixing the transcript, but little changed. I got one decent attempt, but it still stutters when giving it longer words. So fuck it, I'm gonna save my sanity by doing a Jim Hanks version instead. Cause I can't let this stroked woody stay on the site for any longer. I may come back with the Tom with a separate model, but for now, Jim is taking over.

why would reducing samples fix woody

tacotron2 isn't state of the art anymore

but it works 🤷

does normalising all the wavs make your dataset better?

or does consistent volume not make a difference

pretty much.

ok