https://uberduck.ai/ logo
Powered by
# machine-learning
  • p
    i'm guessing it would take longer
  • h
    i've tried but it didn't work for some reason
  • h
    the pitch ended up twice as high as it should be
  • h
    somehow setting it to 22khz × √2 made the pitch roughly correct but the audio was √2 times slower
  • h
    maybe i need to train the model from scratch iunno
  • p
    that's something i've been trying, training from scratch, but I can't disable the warm start
  • p
    i remember hearing that you can't train a 44khz model from the pretrained 22khz one, so you would need a pretrained 44khz model or trained from scratch
  • p
    do you know how to train from scratch?
  • h
    no idea but i'll try
  • h
    iinm 44khz does work on tacotron2
  • h
    based on ppp research anyway
  • p
    interesting, do you know how training a 44khz model would work?
  • h
    id assume training from scratch but im not sure how to that
  • h
    since the 22khz pretrained model will likely interfere
  • p
    yeah training from scratch is probably the way to go, unless there is a pretrained 44khz model
  • p
    no clue where one would be found
  • p
    Any luck?
  • h
    haven't started yet
  • p
    oh, okay
  • u
    Espnet ♡
  • u
    hi what is a dataset
  • u
    and a val filelist
  • h
    dataset is basically a collection of files (in uberducks case, a collection of audio files and a transcript of them)
  • u
    ohhh ok
  • u
    so all my wav files in a zip folder
  • u
    and the list
  • h
    yup, thats basically a dataset
  • u
    cool thx
  • u
    'what abt a val filelist?
  • h
    oh
1...838839840...1068Latest