https://uberduck.ai/ logo
Join Discord
Powered by
# tacotron-2-support
  • h
    1a: cleaner datasets 1b: bigger datasets 1b: smaller datasets with a more consistent tone of voice
  • g
    But like the Ugly Sonic voice model, Ugly Sonic doesn’t say a lot of lines so what if I wanted to make a voice of a character that doesn’t say much in the show/movie?
  • h
    well 1a still applies
  • h
    at the very least try both with the ultimate vocal remover preset and
  • g
    So if I wanted to make a voice but the audio has a lot of background music/noise do I just use a online background noise/music remover?
  • h
    yes
  • h
    specifically these two have good separation quality
  • h
    most others use spleeter which is outdated and garbage
  • g
    Okay I’ll put them in my notes
  • h
    additionally lower the batch size and learning rate
  • h
    as low as 1 and 0.0001 if you wish
  • h
    and if the model sucks try switching the base from aitch to crust and training again (with a new filename)
  • g
    I have asked mega b how to use crust but he didn’t answer (probably because we didn’t friend each other but yk)
  • h
    crust is just a different base model, nothing special
  • g
    Do I have to use different notebooks or?
  • g
    My main ones (it’s not crust I don’t think) is the ones made by grimsalad so uhm yea
  • h
    grimsalad?!
  • g
    Yeah?
  • h
    never heard of it
  • h
    does it show graphs when it's training
  • g
    I don’t really know a lot about graphs so I don’t know what they look like
  • h
    does it show an image of an orange line being mostly diagonal
  • h
  • g
    Yea, MEL Spectrograms?
  • g
    Mhm
  • h
    attention graph actually
  • h
    idk why everyone keeps calling these spectrograms
  • h
    but anyway
  • h
    this means you're using a version of the old notebook
  • h
    we have a whole new system colloquially called "pipeline"
12345...158Latest