not minutes of training, I meant total minutes of audio in the dataset

60 samples and about 1 hour of training.

Is a good combo

Oh

I got it. Thanks

To get a decent voice, the validation loss should be at under 0.1... but I usually train it until it got to validation loss of 0.06-0.05...

I don't look at the validation loss, I have made 2 models that are very good with about 1 hour of training. I'm sure that the validation loss is like 0.102 or something from those models. I really just test the models while training. When I think it sounds good, I save a copy of the model and let it keep on training. I repeat doing this for 3-4 times or for when I think the model is good.

why do suddenly all people here only talk about validation loss?

idk

look at me, I can overfit data

argh

🤦‍♂️

overfitting happened to me 3 times

with the dream singing model

I gave up

pushing a low validation loss on tiny datasets (everything less than 1h is tiny) will almost always lead to overfitting

t r u e

and I see people here over and over again training way longer than it would be appropriate

0.05 validation loss is way too low for me tbh

60 samples, 1 hour training, bam

and then the voice sounds way worse than with not pushing low val loss

Huh... I guess I should consider that, apologies...

0.000001 validation loss

😛

u dont have to apoligize lol

Maybe like 100 samples and more training could do with a better voice

but if you want a great voice, 60 samplés and 1 húor of training is gud

@User If you want an example of one of my 60 sample models, #835647711305793551 and scroll up

I think mostly 1h is already too much