The word it's struggling with is "reality" which is in the cmu dictionary, so it should be able to convert it to arpabet and get the phonemes from there?

I don’t know honestly I’m new to all this.

Ok, I've just let it train a bit longer and it can say "reality" now.

👍

I'm seeing

Warning! Reached max decoder steps

after every epoch. What effect could this have? Anything I can do to try and resolve it?

that means the network hasn't learned to stop at the end of a sentence yet

it's usually gone after some epochs, but if you've been training for quite a while then: - check if you have punctuation at the end of sentences - trim any excess silence from the end of wavs (but leave a fraction of a second in so it doesn't cut off abruptly) - the usual: more data and/or lower learning rate

ok, thanks. it's just started so I'll watch to see if it goes away.

is there anyway to pick up on training? Like to use one of your own models as a pretrained one?

yeah just put it in the

warm_start_name

thing

and also set

ignore_layers

to

["speaker_embedding.weight", "spkr_lin.weight", "spkr_lin.bias"]

when you start training using the base so the voice mapping gets cleaned

- I'd do this guy if it wasn't for the damn music in the background that could throw the model off.

i tried removing it with uvr and mmmm

Hmmm, this one file is so clear it might just be enough.

Ran into an issue, files are okay.

ah nevermind I see it. so I just have to get rid of the lines with no text.

LOL tried to make a better Strong Bad one, and testing resulted in this hilarious fail.

And now a nearly black graph like this denotes a model that is "Thunderdinked"

can you make him tell me to go fuck yourself

damn

a friend is training an english vb transcribed in japanese phonemes, along with a bunch of other japanese speakers for the eng vb to learn pronunciation off of it's going really well!

i don't really know how to know what to change the learning rate to, what should i try changing it to?

i'm gonna try changing it from 5e-4 to 2e-4 based on some stuff i found in the chat history...

wait nvm it was already at 1e-3??

been training for over four hours now... it's a multispeaker english model with one japanese speaker transcribed in arpabet, with the intention of training crosslingually this is the output from tensorboard in the pipeline notebook so it's really crunchy ofc but it's going well! i'm super excited. this is a sample of the japanese speaker "that quick beige fox jumped in the air loudly over the thin dog fence"

1e-3 is 0.001 which is 2 times 5e-4 = 0.0005

it's confusing yeah

@Gosmokeless28 hi! is there any way you could provide the model that's used on uberduck for Spongebob? I understand if not--I do have a roughly trained model using the speaking of AI dataset (using offline talknet)

Oh ty... i'm so bad at that

I ended up changing it to 1e-4 1e-5 gave me an error