Yeah

Got it. Idk when I'll be able to get it completed, but it looks like Imma have to restart the entire process and gather the clearest voice clips from absolutely every episode of the show

Also, should I start at 100 epochs and increase by 100 until I get a perfectly curved line?

How is it possible to start at 100 epochs?

That's what it let me do

*101 epochs

you mean start with the limit set to 100 epochs? i'd say there's no need, since you can just stop whenever and it'll save the progress up until then

Idk if anybody can help, but I got the uberduck github to run locally on my machine. Training models seems to work fine. However synthesizing with the notebook "Uberduck Tacotron 2 Synthesis but awesome" is giving me an error. At first I thought I just needed to enable the use_gpu node but it gives the same error.

Hmm, @hecko, what is your input on this?

looks like I'm getting some warnings when I start to train:

don't worry. because: 1. you are using base model, and you need to ignore some layers to train on base model 2. it's first epoch so there is no model in training folder. it's set to not ignore layers if there is a model in your training folder.

Hi, can someone give me tips about training my first tacotron2 model?

What tips specifically?

First, my model uses the portuguese language

i have a dataset of about 396 wav files from a tv show

Mostly with 3 sec

honestly i have no idea

i give the model 700 epochs and got a loss of 0.13

duplicate them and merge duplicate ones to create longer sentences

first question: is that amount of data enough to an non english model?

with that training i find some good results on phrases contained in the dataset, but unintelligible outputs in other phrases and even in slices of phrases contained in the dataset

Why 700 epochs?

after 0.13 of loss the model gets really slow to train, i tried to change learning rate to 0.005 and batch size to 12 and got some speed up

not necessarily duplicate, just merging is fine too

is my model overfitting ?

i saw your norwegian method on huggingface i merged mozilla back then (merging) i thought your approach is better (duplicating and merging duplicates)

eh

so it is better to give the model long phrases instead of short?

Yeah

actually variety would be better