https://uberduck.ai/ logo
Powered by
# machine-learning
  • r
    Which annoys me greatly
  • c
    piece of modern art
  • u
    Lmao how did that happen...
  • c
    i dont know and quite frankly i never want to
  • c
    im scared
  • u
    Maybe it'll clear up with more training
  • u
    perhapss
  • c
    why are you posting this here
  • p
    Idk
  • s
    Please stop posting the Garfield pr0n post
  • c
    please help me its devoured my family and its coming for me next
  • w
    POV: sonar looking at the marianas trench
  • w
    Side thought but would it help the training to double the number of samples by duplicating each audio file once?
  • h
    only in special cases
  • h
    e.g. if you have not a lot of wavs and want them to not be trained on in the exact same order every epoch, you can copy the lines and randomize the order
  • h
    but in general it's not really useful
  • w
    Yeah I have 141 unique samples but the problem is that it's SHODAN's voice from System Shock 2
  • w
    The distortion and pitch shifting seems to have hindered training quite a bit
  • w
    I was curious if using duplicate files would give the model more "floor space" so to speak
  • m
    i would avoid having any duplicate files or files that say the same thing in the exact same way
  • h
    it loops over the dataset anyway
  • w
    Fair enough. There are some very similar lines so I'll whack those and see how it goes with some param changes
  • w
    Thanks for the advice
  • m
    as long as its not the exact same line, add it
  • w
    Ah, got it.
  • d
    Why does Tacotron seem to have a base frequency somewhere around 80hz?
  • d
    Probably....if your graph looks like a fish finder, you probably want to redo the entire thing.
  • w
    That looks neat
  • w
    The second Epoch looks even weirder
  • u
    Okay what's with all these people getting such weirdass graphs
1...894895896...1068Latest