https://uberduck.ai/ logo
Powered by
# machine-learning
  • b
    Huaaaa......That's really fackin cool
  • z
    so it has the pitch and duration controls, but it's trained on graphemes which (I think) is why it's not as controllable as I'd like it to be
  • z
    but the latent space navigation is really fun, you can mix all their voices and get totally new rappers
  • h
    once i get home i'm absolutely gonna attempt to train a multispeaker model
  • h
    or at least attach sliders to an existing one
  • b
    Is the colab for that available in #841437191073955920 or can it be found somewhere else?
  • b
    I know that I see Cookies in a certain location
  • z
    I haven't made a colab for this but I'll try to do that soon
  • b
    Coolio, thanks zwf
  • z
    sure thing! message me if you have any more feature ideas along these lines
  • h
    Fast pitch iirc does not use relative positional embedding right?
  • h
    Wouldn’t it be more suitable to use FastSpeech 2?
  • j
    250 epoches on yoshi using 99 voice clips :|
  • j
    guess I just gotta keep adding
  • j
    after all it took cartman 220+ voice clips to start sounding coherent
  • h
    in my experience training on less epochs is enough and won't cause this to happen
  • h
    unless you had this earlier too in which case oh
  • f
    Mr. Krabs (by Dylan Wong)
  • f
    Wong nailed it
  • f
    he may be wong but he ain't wrong
  • j
    250 yoshi epoches with 113 clips
  • j
    a liiiittle better but still needs more
  • h
    please train for less time
  • h
    if you train for a long time with this few samples it'll start obsessing over them at the cost of general english knowledge
  • h
    with elec man good results were had on 100 epochs on 8mb of voice clips (don't remember the count sorry)
  • d
    Damn right he isn’t wrong he did good
  • f
    jacob your voices are robotic like shit man :p
  • f
    they sound like if they were at 2% battery
  • p
    I'm new to talk net what does train_file list and val_ file list mean
  • h
    When you train a model you have to seperate around 95% of your data into a training file list and 5% into a Val file list. The reason for the Val list is to test that your training data is accurate
1...505506507...1068Latest