Welp.

I learned that you can actually overcook a speech synthesis model.

This is 25 minutes

?

and this is 1 hour.

it's called overfitting

The heck is that

a symptom of small datasets mostly

Oh.

just don't train too much

But what made no sense was that it was starting to do this at the 35 minute mark

i assume your dataset is quite small

It's only 35 minutes.

I was trying to go for an hour, but I didn't have enough time due to university.

The reason why I was hoping for it to sound better was because the dataset is better in comparison to the original dataset's bad splicing and bad transcription.

And last but not least, I can never figure out that perfect golden spot in training datasets.

IMHO the sweet spot is between 100 and 200 epochs

But I'm overfitting, and I'm at 64 epochs

after 1h you're at 64 epochs?

Yes.

on a v100?

of large is your dataset?

I had to lower the batch size from 32 to 24

I stopped it because it was getting worse.

I assume the ends of your files aren't good

No, they're fine.

the alignment breaks down at long sequences

weird...

do you want to share your dataset?

I want to take a look