not sure where to put this but i feel this could be a good pretraining dataset? maybe? not sure https://github.com/CheyneyComputerScience/CREMA-D

found in this doc https://docs.google.com/document/d/1oZpCCFJFcgmPkvso8ecNPmMo2mPLLGt4M-8KDslEJ1U/edit

was expecting it

almost done with my first model

Would somebody mind explaining the encoder and decoder graph to me?

For taco what is a good large amount of clips to use and what is the good validation loss to get to.

@Waltman13 (semi-active) No I want MLP Starlight Glimmer voice to come out first!

Patience, my friend.

patience is the key

I finished training Vagan model in Tacotron 2 last night, but his voice sounds low or deeper.

time to have fun with segmentated style transfer

AI issues tonight?

How did you do this?

How do I get the best result with 100 wavs? (Batch size & epoch)

this is already making me angry

Update: the training has gotten better after a few epoches

Generating Mels 0% 0/102 [00:00<?, ?it/s] /content/tacotron2/utils.py:14: WavFileWarning: Chunk (non-data) not understood, skipping it. sampling_rate, data = read(full_path) --------------------------------------------------------------------------- RuntimeError Traceback (most recent call last) in () 1 if generate_mels: ----> 2 create_mels() 3 frames /content/tacotron2/stft.py in transform(self, input_data) 82 83 # similar to librosa, reflect-pad the input ---> 84 input_data = input_data.view(num_batches, 1, num_samples) 85 input_data = F.pad( 86 input_data.unsqueeze(1), RuntimeError: shape '[1, 1, 30665]' is invalid for input of size 61330

Any fix for this?

Your WAV files arent 22KHZ

thanks for the help

what is the lowest possible variation loss so far the lowest I have had was .0005 with 30 clips of audio and 4 hours of iterations

damn having that few of data and training for that long, it definitely has overfitted already

no thats normal

bad graphs at very early training is expected at least from my experience

yeah i think i made no models so far that had alignment in the first epochs

i remember shadow took 13 to start something

what does overfitted mean?

"do not too much or it will be burnt training is like an oven" - woody 2021

so how much training show I go for?

until loss is more or less at 0.15 and the alignment looks yellow