I wasn't happy with how long my dataset was training while still being wobbly, so I tried a different voice, a nice clear narrator voice. But it's even worse. I first tried my narrator on default settings, and the result in the synthesis notebook, after over 600 epochs, was totally garbled. I tried again with crush instead of aitch, and the result is even worse. I'm running it again with aitch, with a change to the transcription. Some of the lines were out of order. and now that I've started training again, I see that there's was a blank line at the end. But other than that, I'm out of ideas. Do I need to make the wavs shorter? Some of them are almost 20 s, just below the limit for the auto transcription notebook.

(and there are 25 wav files)

added it up and its 4:43 minutes of data

Good to know thanks

how much should I train for?

with the pipeline trainer?

should I continue if I do, how much?

how does this look

That looks very good.

https://colab.research.google.com/github/justinjohn0306/FakeYou-Tacotron2-Notebook/blob/main/Latest_Tacotron2_and_HiFi_GAN.ipynb Just to confirm, this is strictly for Legacy Tacotron2, right?

The Pipeline notebook appears to support Legacy, however, the results came out all glitchy

how long does it take for tacotron voices submitted to uberduck to be approved to appear on https://app.uberduck.ai/speak?

Depends on how long a site admin sees it.

According to them, it's a mess because there is no way of sorting/testing them.

It basically gets buried among bad/blacklisted submissions

Why does my model keep coming out as an American woman

It’s a British dude

Makes no sense

batch size > data

What?

Oh is that an issue

Ok

So I should put batch size to lower amount than the amount of files??

Also how come I got it to look good in the training thing with perfect line. But in the synthesizer it says max decoder steps and then produces a jumbled mess

This is from the synthesizer

And this is the training

Here’s another from synthesizing it makes no sense

undertrained

Is 100 epoch too little?