https://uberduck.ai/ logo
Powered by
# machine-learning
  • u
    ok john arpa
  • c
  • u
    (he is referring to the dataset)
  • c
    big big chungus big chungus big chungus
  • h
    you could ask mega b to add it to #903688200973873222 (the international channel)
  • u
  • u
    I should probably add a link to the tutorial from within the notebook Then i'll ask
  • c
    the dedicated 日本語 channel on your keyboard
  • m
    I need a Japanese base tacotron2 model then I can add
  • u
    It's in the notebook
  • u
    Well The gdown command for it is
  • c
    wonder how multilingual is going
  • u
    One sec
  • u
    It's integrated into the notebook already as warm start but here's the link to it https://github.com/Haruqa/tacotron2/releases
  • m
    Thanks
  • u
    And here's the notebook + tutorial i made It's my first repo on github so it might be formatted weird, sorry https://github.com/Mildemelwe/Japanese-Tacotron-2-notebook
  • z
    keep digging you'll discover it one day
  • h
    i did a small experiment, turns out just yeeting data in 2 languages works semi-decently
  • h
    though google did uh > incorporating an adversarial loss term to encourage the model to disentangle its representation of speaker identity (which is perfectly correlated with language in the training data) from the speech content
  • u
    Sorry I figured it out
  • h
    and apparently that helps and idk how do
  • z
    ooh nice you got any samples?
  • h
    from:hecko#8977 in:pl-polski has:sound
  • h
    i only used about an hour of data per language though
  • h
    and probably didn't train it to completion
  • h
    but it does feel like if someone with a good gpu were to train on h (english) + witcher 3 (polish) it'd turn out well
  • z
    to confirm from those samples that I understand correctly
  • z
    you're training on multiple languages, and then doing inference with a speaker A using a language that speaker A doesn't have in their data?
  • h
    mhm
  • z
    nice
1...882883884...1068Latest