https://uberduck.ai/ logo
Powered by
# machine-learning
  • b
    or inference I believe is the word, not vocoder
  • h
    0.1 is decent i've heard
  • b
    gotcha, yea
  • c
    validation loss is what youre looking for but theres also the alighment graph you could take a look at
  • b
    I might ditch this mozilla tts/coqui
  • b
    I tried to mess around with a custom dataset, about 500 clips
  • b
    didn't go too hot on the tacotron 2
  • b
    I've seen a couple juypter notebooks, ill take a look at those. Thanks much for the help
  • h
    yeah 500 clips is not enough for from-scratch training
  • b
    its about 300mb tho, the clips are somewhat longer
  • h
    the only maybe-from-scratch one i know of on uberduck is the jeopardy one, which had several hours of data
  • h
    outside of uberduck i think delta's twilight sparkle with probably 4 hours
  • b
    well, actually question
  • b
    so the "from scratch"
  • b
    do you mean training the tacotron, and then the melgan too
  • h
    i'm pretending the vocoder doesn't exist
  • b
    ahh
  • h
    from scratch is when you train an entirely fresh tacotron that has to learn how to english
  • h
    the usual is finetuning, aka taking the already trained ljspeech tacotron and training it some more
  • b
    is there another method to train tacotron ontop of a the ljs?
  • h
    the #841437191073955920 do it by default
  • b
    gotcha
  • b
    does it just take a checkpoint, and start feeding it the new data?
  • b
    or is it some higher end, messing with pytorch and stuff
  • h
    yeah just takes the checkpoint
  • b
    gotcha, ill give that a go instead
  • b
    damn, thanks for the help
  • b
    I know you get a bunch of dumb asses coming though asking the same types of questions
  • b
    Glad I could be a dumbass
  • h
    this one's new actually
1...515516517...1068Latest