https://uberduck.ai/ logo
Powered by
# machine-learning
  • g
    the graph kind of looked like 90% right from the example in the colab so i assume that it only really fully(?) trained on four samples in my dataset where a younger version of me said "1935" and "2018"
  • a
    is the arpabet checkbox checked?
  • g
    where
  • g
    the training colab or the usage colab?
  • a
    are using talknet or tacotron 2 for training?
  • g
    tacotron 2
  • a
    the new one, or the older one?
  • a
    cause i had the same problem and that's how i fixed mine
  • a
    otherwise, i have no idea
  • g
    the one that's linked on the gitbook
  • g
    This is G1 Starscream at 160 epochs.
  • j
    Converted the transcript to arpabet and now it's complaining again. Does it not like the transcript, or is it something else?
  • u
    I don't think i've ever trained with my transcript in arpabet I've always had my transcript in plain english I think that's what i'm supposed to do, anyways...
  • u
    look for blank lines
  • j
    No blank lines that I can see
  • j
    Figured it out. One line had an extra | for some reason.
  • j
    Still doesn't like batch size 30 for some reason, so guess I'm sticking with 15.
  • j
    Trained it to .08 val loss with arpabet. It still sounds really robotic, but there's definitely a huge difference already in pronunciation.
  • j
    Before vs after arpabet
  • j
    It still falls apart a bit with commas, but replacing commas with periods yields markably better results.
  • j
    I think I'm going to comb through the wavs for any that have rustling and try to denoise them, then train a brand new model using arpabet and the denoised wavs. Hopefully that'll fix both the pronunciation and the robot-ness in one go.
  • a
    it's really quiet too
  • t
    Okay I have an idea
  • t
    Because S1 tells me who speaks what at a given time, I can use that to pick out voice lines for each character
  • t
    And then I can use a STT program to transcribe it in Japanese
  • a
    Alright, so here's a comparison with Aerith between uberduck's Talknet vs Tacotron2+torchmoji+arpabet. Both of those were found in the #841437191073955920 room. Trained to 200 epochs for both of them. It reached .15 val loss for tacotron 2. I don't remember what val loss reached for talknet.
  • a
    Taco results = You get a new result every time. Some of the results can sound really really good. Others are horrible. It's impossible to get the audio to "act" they way you want the lines to be read though. Taco has limited uses since you can't control how the lines are read.
  • u
    .15 val loss for taco is still really early
  • a
    Talknet results = can sound good or bad. Has more uses since you can control how the lines are read.
  • a
    is it? how much should I try for? Someone said not to train too much
1...780781782...1068Latest