https://uberduck.ai/ logo
Join Discord
Powered by
# machine-learning
  • a
    Seriously? So how do I make a character voice more cartoon-like?
  • h
    this didn't train at all
  • h
    if you're using legacy this might be because you set the batch size to more than what your amount of wav·s is
  • a
    I thought the lot size was simply measuring the size of the model. But is there a way to change the voice through this option?
  • h
    batch size is how many wav·s it trains on at once
  • h
    if it's more than what you have then it can't make a full batch so it just doesn't train
  • a
    I edited only 3 wav audios in audacity and then put it in colab along with the transcript
  • a
    in the filelists folder
  • h
    then you need to set the batch size to 3 or lower
  • h
    preferably even 1
  • h
    and also lower the learning rate to prevent it from going haywire, to something like 1e-4
  • a
    okay, thanks
  • a
    Which options should I put "1e-4": 
    hparams.learning_rate
    or 
    hparams.min_learning_rate
    ?
  • h
    the former
  • a
    Like this?
  • h
    i said the former
  • h
    as in the first
  • a
    okay. I changed hparams.learning_rate to 1e-4.
  • a
    is it what I have to change?
  • h
    yes
  • a
    okay. in the second I don't need to change, right?
  • h
    correct
  • a
    okay, thank you
  • a
    I even managed to make a similar voice, but it's very robotic and repetitive, is there a way to change the voice so it's not too robotic and repetitive?
  • a
    I'm learning yet
  • a
    xD
  • r
    Sounds undertrained? What was the loss for that epoch? Idk how big ur dataset is cause the voice is there.
  • a
    The epoch was 200
  • r
    The loss, like validation loss? Different sized datsets need a general loss value to be considered "passable" Epochs aren't exactly the thing to go entirely off of
  • a
    I don't remember, I was playing game while waiting for the model to be ready
1...100010011002...1068Latest