https://uberduck.ai/ logo
Powered by
# machine-learning
  • h
    Nvm I was confused
  • u
    this is definitely something to post in machine-learning totally
  • u
    hella offensive post
  • h
    Ok how do I add new text?
  • h
    Uhhh was I doing this right?
  • u
    idec about that it just dont belong here
  • h
    My text I want her to say is “My Chemical Romance” but I previously had her say “hello”
  • b
    when it fully generates a text, a new textbox appears under it
  • b
    oh also the inference notebook is annoying because to test a voice you have to edit the code to add it
  • h
    i don't know an exact number but over a dozen seconds it starts to get wacky plus the longer the samples are the less can fit in a batch
  • h
    yeah
  • h
    though i'd personally have some variety so it still knows how to do shorter and longer lines
  • h
    uhh 15 i'd say
  • h
    for small datasets i think it's important to keep it as close to a clean divisor as possible
  • h
    because if you set it to let's say 10 then only the first 40 wavs would be taken and the last 5 ignored because they don't fit into a batch
  • h
    at least i think
  • h
    haven't tested it scientifically
  • h
    yeah
  • j
  • j
    A must try
  • j
    Improve your dataset
  • i
    i would wait until loss 0.15 to see if it aligns until then
  • u
    its the bruno porowznik dataset right?
  • u
    any attempts to declip them? cause knowing bruno im sure all of them clip
  • h
    i didn't declip cbt guy and he was fine in the notebook but atrocious on uberduck.ai
  • c
    wonder if rx8 has tools to fix hard clipping
  • u
    it does have a declipper
  • c
    (and i dont mean sounds going above 0db i mean sounds hard clipped to 0db)
  • u
    so good on lj hifigan but bad on trebek hifigan?
  • b
    I always just use normalize in audacity to fix clipping
1...667668669...1068Latest