What's wrong?

i can't access that because it's your private copy

please send me the dataset

wait a minute

the notebook won't help me

i need the dataset

dataset?

yes

the audio

oh ok

I'll send the folder

hm

so i see you have a 30.mp3 file

which might be the issue

but you also have 8.wav which is stereo instead of mono

and 30.wav is stereo too

but 30-2.wav is mono and good

Ah yes. it's just that I forgot to convert when I isolated the voice and removed the background music

convert to wav file

okay, thank you

which one should i keep my eye on? this or the other?

(the other)

i think they are the same :)

Hey guys! Pretty new to this stuff. I've been training on a pretty sizeable dataset (approx 750 files, avg time maybe 4-5 seconds each) for quite a few hours now. I don't really know what the graphs are supposed to look like. But I noticed there's an audio tab on TensorBoard so I refresh and check that every once in a while. It's about 75 epochs in and the audio still sounds extremely demonic. The one that says "AudioSample/train" also sounds really demonic. Is this normal? Intuitively the train file would be the original file and so it should sound normal, is that not the case? And if this is all normal, do I just need to let it train for a while longer? Thanks in advance!

Yes it's normal, that's just the audio quality of the preview clips.

Ok cool, thanks. So, what's a good way to know when it's trained enough?

Personally, I use what I call "one size fits all" parameters when it comes to training Uberduck Pipeline Tacotron 2 models. Here are my recommendations: Epochs: 200 Batch Size: 8 Learning Rate Decay: ☑️

Ok so you think I should let it run until it reaches 200 epochs?