the ljspeech dataset (which was used to train the base model that the notebook finetunes with your voice) doesn't do clean sentence boundaries either so that's not a big deal i'd imagine

gotcha

ah, but i'd imagine i'd also need to manually edit the files to give data on which one of them is talking/singing

you can treat talking and singing as separate characters

right yeah, but. like i said, these particular men come as a package deal

if ya get rhett, ya get link

hmm

the format is always both of them in the same video talking to each other

and occasionally talking OVER each other, which could also be problematic

i'd just ignore those

yeah fair

you don't need to scoop up every single possible clip

got it. how long is a good bare minimum? like, would a single season of gmm be good enough?

like, a season lasts a year, and the episodes are about 15 minutes long and come out like...at LEAST weekly i think

so thats about 780 minutes of audio

maybe round down to like 700 because of the unusable parts idk

that's probably more data than any tacotron voice has ever used

chills has an hour of data cookie masterson 100 minutes twilight sparkle might have 4 hours but idk the base model has 24 hours and you're worrying whether 30 is enough

I mean the Alex Trebek thing has over 6 hours of data so yeah, 30 will definitely be fine

wait chills only has an HOUR of data?? that's wild because his model is fuckin. Uncanny

anyway 700 minutes isn't 30, it's closer to 12 hours

and assuming rhett and link spend a roughly equal amount of time talking (which they do as far as i can tell), that's still about 6 hours for both of em

so...damn, yeah that might be overkill

hell, if chills has an hour, i might be able to get away with just using 8 episodes

which would give roughly an hour for rhett and link each

emotion filtering for /mlp/ datasets is now existent in my notebook https://colab.research.google.com/drive/1xnFT5gXY5CIdsyG0YZbCulXdJksJR4ei

also please someone check if loading files from google drive works, for me it fails unless i've already downloaded the file and i wanna know if it's global or just me

its going to take a long time to train lol

could be worth it though

hmm maybe it could be possible to host the model elsewhere and have the notebook upload there instead, then have 2 or more people tag-teaming ther colabs to minimize downtime from running out of gpu