that's straight up godlike

Single speaker lol.

god

i mean i have seen a larger dataset but still that's a ton

there are another 17 speakers i am trying to work on, just, alot of data is involved and stuck with a rather limited bandwidth haha.

they will be all individualized datasets, all trained using tacotron2_statedict and hifigan UNIVERSAL_V1_g. Wouldn't mind tinkering around with a few others once i have things sorted out lol.

going good so far

with only 21 lines

this just started

Going really well

after only a few minutes

fluctuating

There is a thing called "overfitting", which is the process of an AI voice model getting to be over-trained with the same audio data too many times. As a consequence of that process, overfit Tacotron 2 models have difficulty pronouncing words good.

How many iterations should I do for 60 clips

It's taking longer to get a good graph for now

do it until validation loss below 0.15 or 0.09

still high, its at 0.30

0.29 now

is this any good

almost under 0.15

alr thanks i have 343 sound files and that added up with 343 is 686 so i used 686 epochs instead

final

No!

ima just see what happens

Reminder that this exists

though sometimes stopping even earlier can be beneficial

my version of the notebook saves a backup every few dozen epochs in case i end up wanting to go back, with the regular notebook you can just copy the file in drive every so often

How so?

because the more you train the more general english knowledge is lost

Fine, but training for 686 epochs is going to cause the model to be overfit. Just use the amount of epochs we tell you to use, 200 or 250.