https://uberduck.ai/ logo
Join Discord
Powered by
# machine-learning
  • t
    Why some people are trying to make the AI debate a left vs. right thing (anti vs. pro respectively)? Not every conservative is pro AI. Matter of fact, Tucker Carlson is anti-AI and Hasan Piker is pro-AI:

    https://www.youtube.com/watch?v=6ve1W1EuSVg▾

  • m
    Because dumb.
  • u
    i dont want a breadtuber on this side of the argument 🤢
  • t
    It's just usually I see it the other way around
  • t
    Anti-AI people pinning pro-AI people as mainly right-wing tech guys who like NFTs and crypto
  • t
    I think the real issue that is corporations are cheapskates and will screw over workers for higher profit margins. The tech itself isn't the problem
  • t
  • h
    here's what i made with my Firey model
    download-2
  • k
    how many datasets is required to train a really good voice model?
  • h
    You mean wavs, it’s usually 19
  • k
    19 wavs file to make a model synthesize proper speeches? its not gonna become garbled?
  • k
    how is that possible? im guessing it is done by modifying existing models?
  • h
    i'd say like an hour for real good quality
  • h
    pretty much, we have what's called a base model that was trained on a bunch of voices so it can quickly adapt to new ones
  • k
    where might i find these existing models?
  • h
    uberduck pipeline training notebook, code of cell 4
  • h
    what do you want them for though
  • k
    i expect there are models for different language so i would like to choose the correct one
  • h
    the notebook will do that for you automatically
  • h
    some languages don't have specific bases though, but it's possible to use an english base if you have enough data
  • k
    so i could use english base but train arabic?
  • k
    or korean or whatever
  • r
    Yeah
  • r
    Forcing the model to speak a different language you would need enough data
  • k
    for how long should you train an english model? is there some kind of measure for that, the more dataset the longer you should train or something like that?
  • h
    opinions vary i've seen people train models for thousands of epochs personally i think there's not much difference beyond 200 with default learning rate, and i sometimes like to stop earlier than most people
  • k
    what constitutes a good dataset? as in the more clear the sentence the better or is it fine to have sounds such as "uhhhh"
  • r
    As long as you can accuratly transcribe it, and there's little to no big noise, I would consider it good. Having "uh" in it doesn't harm a model as far as I've seen.
  • r
    Laughter - probably wouldn't recommend in most cases, if not all
  • t
    Well shit, could that be why I can't get a good result outta training my Mario dataset?
1...105210531054...1068Latest