https://uberduck.ai/ logo
Powered by
# machine-learning
  • c
    Yes
  • c
    ok i guess i have dementia it is not talknet
  • c
    I even tried removing superres thinking it would get better
  • c
    Didn't changed
  • m
    how much wavs are in your dataset?
  • c
    Around 250
  • m
    how long did you train hifi gan?
  • c
    But I don't think that's the problem, like i've said, even with the same generator, the results are much better in Uberduck
  • c
    2k in one of the models, 10k in another
  • m
    have you tried using other hi-fi gan models?
  • c
    Yes, they all suffer from the same issue
  • m
    could you DM me the download link for the generators?
  • m
    or send it here
  • c
    what does uberduck actually do when the mel is put through the vocoder
  • c
    is it just pulled directly from the hifigan repo
  • c
    ()
  • m
    im sure that + other optimizations
  • c
    what kind of Optimisations
  • m
    i think adding threading for api requests
  • c
    update: the model does not want to learn attention
  • c
    look at my lawyer dawg im goin to jail 😭 😭
  • c
    what the hell. so thats why my attention sucks
  • c
    (this is ground truth spectrogram)
  • c
    anyone who knows torch or numpy do you guys know why this happens
  • c
    because that is NOT supposed to happen. the range isnt supposed to be from -9 to -11.5
  • b
    What are you training?
  • c
    finger-player/wile-e-coyote
  • c
    and i just found out that some melspecs wont generate properly
  • b
    Is that tacotron? Sorry, I don’t recognize the graphs
  • c
    yes this is tacotron
1...676677678...1068Latest