https://uberduck.ai/ logo
Powered by
# machine-learning
  • h
    it sorta works
  • e
    I see that's all I was looking for, evidence that the approach should work
  • h
    the main issue is when the kick drum hits the vocal ducks
  • h
    which tbh would need a dataset of pop music that has 500db kick drums
  • h
    which yain't gonna find on indie sites
  • e
    Tho if the model isnt designed for keeping vocals, might behave weirdly
  • h
    yeah it's only trained on mix-instrumental pairs and then i guess subtracts them to get the vocals
  • e
    Have you looked into the paper to see if they tried preserving only vocals?
  • h
    have not
  • h
  • h
    some of the models promise cleaner vocals
  • h
    so uhh nevermind everything i said earlier there might already be a way to do good vocals
  • e
    Ye lemme know and lemme know if you want it finetuned more
  • e
    Finetuning will be way faster, their models are probably trained on like 2-3x 2080tis or 2 3090s
  • e
    Or actual online gpu servers
  • h
    that'd require data to finetune on
  • h
    which they probably have more of
  • h
    > You can download the model pack here links to source code download hmm
  • e
    Maybe it's in the source code?
  • h
    doubt
  • h
    -actually wait v5 is what the #841437191073955920 notebook uses so worst-case i can use that with the vocal model
  • h
    i'd rather have it on my computer though so i'll keep looking
  • h
  • h
    given they can't provide the dataset i'm guessing they're using leaked stems from copyrighted songs
  • s
  • h
    i mean i'm not complaining, as long as it can filter out 500db kick drums nicely
  • h
    ...hm am i gonna have to install another copy of python because v5 asks for a different exact version of torch and co
  • e
    How about demuc?
  • h
    or y'know i can finally figure out how virtualenv works
  • e
    O and use a virtual environment for yea
1...389390391...1068Latest