https://uberduck.ai/ logo
Powered by
# machine-learning
  • t
    what do you mean by convert?
  • u
    you can bruteforce japanese to english voices by just
  • u
    transcribing the best possible english approximations
  • u
    i.e. try to hear as many english words in the japanese as possible, if nothing is close then try to spell it out using english pronunciation rules
  • u
    though i really would not recommend it at all it is such a fucking pain
  • t
    if its a pain then im this guy
  • t
    ouch
  • h
    it's a polish-specific thing, they have 
    ą
    (pronounced like hon hon hon) but japanese doesn't so the synthesis bit turns it to 
    on
  • h
    wouldn't work with english i'd imagine
  • t
    im transcribing some of the japanese R-s as L-s and some of the Z-s as TH (which is how it works in romaji)
  • t
    is there just like
  • t
    no way to get the right one now
  • h
    oh you're using an outdated notebook
  • h
    the one in #841437191073955920 doesn't depend on gpu at all
  • t
    o...
  • h
    but yeah colab free tier only gives out k80·s nowadays
  • h
    yeah i trained jp before
  • h
    tho it was only meant for jp so it couldn't do eng at all
  • h
    i used haruqa's method which is just romaji close together
  • h
    like koNnichiwa.
  • h
    hiragana works as well but don't b use kanji
  • t
    hey where do i put the .txt file with the transcriptions in the training notebook
  • t
    oh wait i see
  • t
    i only have 90 wavs how do i let it know that
  • u
    by not listing a 91st file in the transcript??
  • u
    like wh
  • u
    i wonder if 30 files is enough for a model.
  • j
    How do i use /voice suggest?
  • u
    since the model i wish to someday train has 30 files
  • u
    is 30 files enough or do i need to dupe a few files
1...639640641...1068Latest