sorry i lost the dataset

Didn't take long for the AI to sort this out.

And two graphs later

Trying to find more lines, got like a minute's worth so far

Got 37 of 'em now, retraining

It sounds like shit. Voice is from gta 5 tracey 159 wav files ~6120 steps 600 epochs

which hifigan are you using?

what is crust? and how do I use it to train models?

Crust is a model trained by [@]Pikachu ✓#5870 Idk much else

I find that a smaller batch size takes longer but usually results in a better model.

training usually starts from a base model that can already speak english (though with a different voice) so it doesn't have to relearn it crust is one of those base models, trained on 168 speakers at once (whereas before we had 53 or just 1), which makes training on it apparently require less data and less time it's the default in my version of the pipeline notebook https://colab.research.google.com/drive/1jF-Otw2_ssEcus4ISaIZu3QDmtifUvyY

though apparently there's a special procedure to have good training uhh set learning rate to 1e-3 (0.001) and train until the graph in

Images

is clean and diagonal, then stop and resume at 5e-4 (0.0005) for 50 more epochs, then stop and resume at 1e-5 (0.00001)??? and keep training for however long you want

From Uberduck tacotron2 inference notebook I don't know which version is

I suggest you try training using the Pipeline colab

I have used that sorry i mixed up

Remote speaking

based off this clip

Mama

oooo

is there a way to view training in tensorflow live during training? or do I have to stop and run a cell for it.

are you talking about this screen?

yes

run tensorboard first

then run the training process

refresh using this or you can set reload data setting in the tensorboard settings

Hmmm, someone changed how the graph displays.

thank you

you're welcome