https://uberduck.ai/ logo
Join Discord
Powered by
# machine-learning
  • k
    This is another example of poor quality AI voice
  • k
    What factor makes AI generated audio sound so horrible?
  • h
    probably the vocoder
  • h
    -or nevermind you used talknet which trained the vocoder
  • h
    so the other aspect might be that the notebook tries to upsample the audio from 22050hz to 32000hz which does add some high-pitched hissing
  • h
    if you convert it back to 22050 it should be fine
  • j

    https://www.youtube.com/watch?v=fZMiD8sDzzg▾

  • r
  • u
    oh my god does this mean they're, in some semblance of the word, alive?
  • r
    Until I can shove a rat in a Pi, I say no (not entirely, at least)
  • a
    Is it normal?
  • h
    it is Not normal
  • h
    you seem to have very very few wav·s given that each epoch only has 1 iteration
  • h
    i recommend that you set the batch size to 1, and lower the learning rate if you can
  • a
    I understood. But I don't need two wav audios anymore, because I'm already making 30 wav audios here. A guy told me that to create a perfect voice you have to have 30 wav audios to form 10 minutes
  • a
    but thanks anyway
  • h
    what's your batch size
  • a
    Currently, I set the batch size to 2
  • h
    hm
  • h
    do you only have 2 wav·s then
  • a
    Ah, do you want to know how much batch size I can put with the 30 wavs?
  • h
    no like
  • h
    you said batch size 2
  • h
    and this says 1 iteration per epoch
  • a
    I think I can 500
  • h
    no
  • h
    that's not how batch size works
  • h
    if you have let's say batch size 8
  • h
    that means your dataset is split into batches of 8 wav·s
  • h
    and if you have batch size 8 and 30 wav·s, then there will be 3 batches of 8 and the last 6 will be thrown away because they don't make a full batch
1...101710181019...1068Latest