https://uberduck.ai/ logo
Join Discord
Powered by
# machine-learning
  • m
    What's d-id?
  • m
  • z
    Yep
  • m
    Does uberduck have any speak-portrait products, even in dev? I'd be really interested in using one of your guys' solution, landing page doesn't list any though
  • h
    open-source software
  • h
    though perhaps the more relevant term would be "self-hosted"
  • m
    thank you sir, I got makeittalk working yesterday but it takes too long (30 seconds, which is not bad overall, but for an end user is a lifetime)
  • m
    Sorry, didn't do direct reply: Does uberduck have any speak-portrait products, even in dev? I'd be really interested in using one of your guys' solution, landing page doesn't list any though
  • h
    the pricing page does advertise "clone your face or voice" but i haven't heard what tech it uses or if it's even active
  • z
    We don't, sorry.
  • m
    All good thank you sir!
  • j
    You can set it up locally as well
  • h
    What is state of the art for swapping a face in a video, including the original hair possible? From photo possible?
  • p

    https://youtu.be/7mUr-8h60kM▾

  • p
    So there's this singing AI code library called NNSVS
  • u
    Nnsvs my beloved
  • r
  • u
  • u
    quick digging brings this up
  • p
    is vall-e kind of like tortoise architecturally?
  • h
    at a glance the answer seems to be no tortoise used diffusion and clip and stuff, whereas vall-e is like a language model but trained on encoded audio so like stable diffusion vs gpt/vqgan
  • h
    or you could call it reverse whisper maybe
  • u
    yell
  • j
    ow
  • w
    @hecko this sounds like almost exactly what your trying to put together
  • w
  • h
    pretty much, but they probably used regular human speech and i want character voices
  • h
    there's actually this thing called tortoise tts that was made 8 months ago, trained on 50k hours, and it's probably the best regular human speech tts i've ever heard
  • h
    plus it accepts a speaker embedding so i can mix voices together, whereas vall-e takes in sample audio directly
  • w
    That's interesting i'll look at that
1...103210331034...1068Latest