https://uberduck.ai/ logo
Powered by
# machine-learning
  • a
    no worries if i had some time to make it, i'll be continuing on it
  • a
    even tho i think i still forgor trying out the local training tho
  • a
    but uh... @User i might need your help on that tho
  • c
    i remember making a summary of what you have to do to setup a conda environment but the closest i could find is like
  • c
    Copy code
    [12:32] buster: conda create -n tacotron2 python=3.7
conda activate tacotron2
i think the only packages i installed through conda is pytorch (you can get the command to install that on  pytorch.org)
  • c
    How much training would be enough to have a somewhat decent result with Mellotron?
  • h
    well last time i checked mellotron training was broken and only took into account the first few files
  • h
    maybe it's gotten better since but i've my doubts
  • c
    I'm using the Nvidia one
  • c
    Is it broken too?
  • h
    iunno
  • c
    is fastpitch multispeaker
  • h
    the demo models on the website are i've heard
  • d
    Is it normal i tend to get worse results with Hi-Fi finetuning?
  • d
    I did about 6,000 steps
  • j
    Did out here my first model and the voice worked kinda well, but a bit robotic... But for some reason when I put English words for the AI to say, he doesn't say anything. But when I turned off the dictionary option, he just said it normally.
  • j
    Is this normal for y'all as well?
  • j
    Thank you
  • j
    hmmm not really, how do I do that?
  • j
    tacotron2
  • j
    Thanks a lot! I'll try that
  • j
    Btw, what does this val_filelist box means? I just have one list here saved which is all of the dialogues
  • t
    Should I send my notebook file along with the instructions in the Read me so I can get a better idea of what I'm supposed to be doing?
  • j
    Thanks!
  • h
    train_filelist is what wav files itll use to train the voices
  • h
    so basically all the wavs you transcriped
  • h
    val_filelist is what the ai uses to validate that the model will sound like the original source and will not overfit (in theory)
  • j
    Thanks for the explain!
  • j
    I can see here it is recommended to reach more than 5000 steps so I can finish training. How much usually it takes?
  • c
    It depends on how much wavs he has
1...712713714...1068Latest