I'm trying to train a pipeline model, but step 4 keeps stopping abruptly after a couple of seconds of running

do the audio files for it have to be 22050 kz mono?

Oh yeah, I think that was the issue

Yes

I'm trying to use a pipeline tacotron on the inferencing notebook and I'm getting this error

Don't use finalized model

oh

thanks

it still shows the error

this is what i put

Thought I could show the test of Mr. Lunt I made

I'm actually wondering now. Is pipeline tacotron 2 better than legacy tacotron 2?

Pipeline TacoTron2 is pretty good. I used it for my Pokemon Masters EX models, and Dorothy the Dinosaur.

I won’t probably use pipeline because it seems a bit more harder than legacy (I use grimsalad’s tacotron2)

im curious as to what grimsalad's tt2 has compared to nvidia's and if it even works with the website

Pipeline is actually pretty easy. It's just the training that's the hard and time consuming part. Training isn't really that hard, but you need to know what to do.

Yea but why is everyone using pipeline? Almost every single voice I’ve seen has a symbol set

Does it make the voice sound more accurate when you make one?

It depends. Especially depends on how it's being trained. Take a look at my models and mess around with them and see how good you think they are. My Uberduck username is MasterRuby.

Alright

Remember that training with Pipeline is done when you have characters with very few lines or audio clips. My Pokemon Masters models have a little over 1 minute of dialogue when combining all audio clips together, while my Dorothy model only has like 46 seconds of dialogue all together. I trained the Pokemon Masters voices using Crust, which is part of Pipeline. The Pokemon datasets have 12-16 wavs. And I trained Dorothy with AITCH, which is also part of Pipeline. Dorothy has 5 wavs.

So if I train 12-16 wavs, it'll talk accurately?

Like I said, it depends. There are times it can be good, while other times it can be bad. Just make sure your wavs are about 5-12 seconds long. If they are longer than that, then that's better.

Gotcha

I am pretty used to training legacy models but I could try to make a model in pipeline just to see how good it is

AITCH is trained on more voices than Crust

@TheRoyalRuby2000 I just tested one of your model and what the crap? ROFL

Oh yeah, the Brock model does stroke after he says a sentence. But he can at least say the sentence decently. Dorothy is another model that strokes after she says a sentence. But like Brock, she can say the sentence decently.

undertrained

learn to learning rate decay