https://uberduck.ai/ logo
Join Discord
Powered by
# machine-learning
  • h
    or
  • h
    there's spectrogram from waveform
  • h
    but image from spectrogram is a todo
  • p
    Todo?
  • h
    to do
  • p
    How long is each image
  • m
    super cool...
  • m
    i think i could totally make a vocoder model for music
  • m
    the demo uses griffinlm
  • m
    @hecko do you know any datasets that is just music wavs?
  • m
    like 5 second clips
  • h
    pre-spliced no
  • h
    i do know that kevin macleod is willing to share his stuff for ml projects
  • m
    well i think im just going to pretrain a vocoder model so no need to splice anything
  • m
    thanks for the tip
  • h
    there's musdb18 but that's kinda small and more for stem separation
  • u
    I'm willing to share my music for ml purposes too if you'd like
  • u
    There's only about 1 hour worth though
  • u
    And some of it is garbage
  • u
    Cuz it's old
  • j
  • h
    -oh and it requires contact too
  • h
  • g
    This is the coolest StableDiffusion model I've ever seen.
  • z
    I think someone should see if that method works for speech. I'd like us to do it if we can make the bandwidth
  • z
    pretty experimental tho
  • h
    welllll - stable diffusion as text to speech? probably not, since it can't even assign colors to objects properly - as a description-based voice shifter? maybe!
  • p
    Isn't that what diffsvc has going on
  • h
    not description-based though
  • h
    but idk where one would get described voice clips
1...102310241025...1068Latest