https://uberduck.ai/ logo
Join Discord
Powered by
# machine-learning
  • o

    OctolingVladTTS

    12/15/2022, 3:01 PM
    I want to update the voice of merasmus and bombinomicon, someone asks me a favor about how I train the voices with the new and better tacotron2, pipeline Tacotron 2 model, since the legacy seems somewhat obsolete to me
  • h

    hecko

    12/15/2022, 5:10 PM
    the pipeline notebook has instructions built in try following them, and if that's not enough then ask me for help
  • z

    zwf

    12/15/2022, 5:13 PM
    This is pretty wild that it works as well as it does https://www.riffusion.com/about
  • p

    PixPrucer

    12/15/2022, 5:25 PM
    Oh that's super cool I knew it was possible earlier considering dance diffusion was a thing but this is next level
  • p

    PixPrucer

    12/15/2022, 5:26 PM
    I wonder if you can fine-tune the model to fit some other artist's music
  • p

    PixPrucer

    12/15/2022, 5:27 PM
    (looks at self)
  • h

    hecko

    12/15/2022, 5:29 PM
    @Pikachu ✓
  • p

    PixPrucer

    12/15/2022, 5:34 PM
    Dream of an AI imagining my music coming true closer and closer
  • h

    hecko

    12/15/2022, 5:41 PM
    i mean it's literally just stable diffusion so i don't see why not
  • p

    PixPrucer

    12/15/2022, 5:44 PM
    Oh it is?
  • p

    PixPrucer

    12/15/2022, 5:44 PM
    Neat
  • p

    PixPrucer

    12/15/2022, 5:44 PM
    But then I have like No idea where to start
  • h

    hecko

    12/15/2022, 5:45 PM
    well first you have to wait for huggingface to stop being dead
  • h

    hecko

    12/15/2022, 5:45 PM
    get this model https://huggingface.co/riffusion/riffusion-model-v1
  • h

    hecko

    12/15/2022, 5:45 PM
    get spectrograms, not sure how but there's probably gonna be code soon
  • h

    hecko

    12/15/2022, 5:45 PM
    and then just dreambooth it
  • p

    PixPrucer

    12/15/2022, 5:45 PM
    So true
  • p

    PixPrucer

    12/15/2022, 5:45 PM
    I'll give it a shot once all the resources are up
  • p

    PixPrucer

    12/15/2022, 5:46 PM
    I'm a colab bitch and I'm afraid to train elsewhere
  • o

    OctolingVladTTS

    12/15/2022, 6:00 PM
    ok, I ask you when necessary
  • o

    OctolingVladTTS

    12/15/2022, 6:01 PM
    in DM
  • p

    Pikachu ✓

    12/15/2022, 6:29 PM
    I still think the dance diffusion thing is better, have you heard demos of the 44 second model?
  • p

    PixPrucer

    12/15/2022, 6:30 PM
    Oh I didn't yet
  • h

    hecko

    12/15/2022, 6:31 PM
    but does it accept text prompts
  • p

    Pikachu ✓

    12/15/2022, 6:31 PM
    These are about a month or two or three old
  • p

    Pikachu ✓

    12/15/2022, 6:31 PM
    Okay ready your seatbelt
  • p

    Pikachu ✓

    12/15/2022, 6:32 PM
    Harmonai is working on something like that
  • p

    Pikachu ✓

    12/15/2022, 6:32 PM
    You can steer it though, they're experimenting with it
  • p

    PixPrucer

    12/15/2022, 6:32 PM
    Not them making NCS their primary dataset
  • h

    hecko

    12/15/2022, 6:32 PM
    don't worry they're gonna use kevin macleod too
1...102010211022...1068Latest