check if one of your audios are in 22050 then

or 32-bit

I literally was converting all of them to 22050, 16 bit, mono

and rechecked rn

nothing odd

What does increasing or decreasing the batch size in the training notebooks do? I know increasing it will require more RAM, but does it like make it faster or better?

So I'm just curious now

What would be a good network for training on like...non-speech audio?

Because like I can't find a colab for WaveNet itself (well i DID find one but it's "barely functional")

But I did find a solid one for TacoTron

And I THINK OpenAI jukebox uses WaveNet too

So like...are jukebox and tacotron the same network underneath, or has tacotron been optimized for speech in some way?

This guy's just using Torch I guess????

more batch size means smoother training because it'll only change the neurons to improve performance on all audios in the batch, meaning the more wavs the less it fumbles around and it's probably faster too but idk

for one tacotron doesn't generate audio directly, it generates a spectrogram (specifically mel spectrogram which is optimized for speech) that is then fed into hifi-gan to turn into audio

for two it, y'know, takes text as input

goooot it

i've been vaguely thinking about giving it an hour of abc music to see if it can adjust

let us know the results!!

what audio do you wanna generate btw

eh screw it

?

part of me wants to keep it secret

but fuck that

i hate copyright law so much

ok. its uh

drum samples

ooh

which i was TRYING to keep secret bc I Really Need Money and i was wondering if this was some kinda get-rich-quick scheme

but