https://uberduck.ai/ logo
Powered by
# machine-learning
  • l
    lemme copy and paste them
  • l
  • c
    redoing
  • c
    redoing training
  • c
    I think I have only 30 minutes to train though
  • l
    is this nice for my first model i've posted to uberduck i used crust notebook
  • c
    @mega b I keep getting a traceback on the synthesizer
  • c
    I trained my model so what should I do
  • c
    and apologies for the ping
  • c
  • l
    I'm confused so when i put in the mitochondria is the powerhouse of the cell in arpa bet and not in arpa bet it doesnt sound the same for some reaosn?
  • t
    what to do if loss stuck at 0.30 ?
  • c
    How do I add audio
  • c
    the folder doesn't exist
  • c
    nevermind
  • m
    was feeling good today, and decided to work on something i was meaning to do a while ago, and training it right now.... only 6 epochs in but beautiful graphs
  • c
    this Qubo voice was also on the backburner because of slacking off
  • c
    but now I've made it
  • l
    Is this like low quality model or just eh
  • b
    Do higher pitch prediction training values result in better vocoding? Or?
  • h
    pitch and duration prediction are ignored in uberduck
  • h
    hifi-gan is the vocoder but the loss doesn't really mean anything because it's 2 models fighting against each other
  • b
    Gotcha, so I'm kinda referring to an offline version (TalknetOffline)
  • h
    whereas with spectrogram it just compares to the correct data and the loss is the difference
  • b
    Aside from manually shifting the pitch of the reference audio, is there any other way to improve the "tracking" of the pitch? I guess having it in the same timbre or whatever
  • h
    uhh you could remake it in something like synth v
  • h
    i've done that once and it turned out nicely i think
  • h
    assuming it's singing
  • b
    Ooouuu.....Synth V is actually not a bad idea at all
  • b
    Some of the demos are terrifiyingly real
1...923924925...1068Latest