even if it wasn't it definitely beats "can you make [obscure cartoon voice] please"

Damn, that doesn't sound fun

last year I spent a lot of time reading into tacotron, but never had a gpu

specifically cartoon because i don't mind doing game voices, they're often pre-clipped and the transcription i sorta enjoy even

pre clipped and subtitled/transcribed too

sometimes

if the game happens to be easily piratable

though even then i usually give the transcripts a once-over because they sometimes don't match the speech exactly

aaand because i like making the punctuation match the speech

ahh, makes sense

alright, small last question. I have a 3090, should I try and refactoring the juypter notebooks to train on my local linux pc, or just use google colab

with the recent k80 plague you might as well try

k80 plauge?

ever since the introduction of colab pro+ free users are more likely to get k80s than t4s

google ran out of t4·s so everyone's getting k80·s which are 2.5× slower

damn

I think my 3090 is better than the T4 as well

yea, ill just try and refactor

alright, thanks again, might miss a reply, diving into the deep end

🤿

wait I can just open it on my pc with juypter notebook

with a small dataset (2 hours), is it better to train tacotron2/hifi-gan on 22khz instead of 44khz? in terms of audio quality / overfitting. or it doesnt make that much difference?

that's a small dataset?

2 hours is probably more than enough for 44khz, if ponies are anything to go by

though i don't think zwf would allow a 44khz model

That is probably better than the P100

is there a tutorial for the vocal remover collab?

yes there is actually