https://uberduck.ai/ logo
Powered by
# machine-learning
  • c
    i think embedding has to do with symbol set since it complains about size mismatch if you change the symbol set
  • t
    yeah i included all of them too but don't know if any layer does something useful
  • u
    Since it's a warm start you need to ignore certain info so it doesn't train while including too much info from the old bank
  • c
    cant look at the word 
    bank
    without thinking about utau
  • u
    Oh i'm just so used to using the word bank as opposed to model Concatenative synthesis got me
  • h
    all these are just lookup tables really speaker embedding turns speaker id into neural embedding turns characters into neural spkr_lin is uh,, not sure actually
  • t
    this is sample inference:
  • t
    and this is how it should sound like
  • z
  • c
    something tells me this won't go too well
  • l
    use crust notebook
  • c
    haven't used it before
  • c
    any link to it
  • l
    Ok one sec.
  • l
  • c
    first I think I'll finish training this version
  • c
    then try the crust notebook
  • c
    Training now
  • c
    How do I check if the model looks good or not
  • l
    if its like a mountain like graph its good
  • l
    like a sharp hill
  • c
    it's a Bliss-style hill now
  • c
    so I'll wait a bit
  • l

    https://cdn.discordapp.com/attachments/820353681567907901/854402123306369045/download_2.png▾

  • l
    heres an example
  • c
    I meant for the Crust notebook
  • l
    ohhh
  • c
    which I am now using
  • l
    so theirs obviously the graph and when you go onto the audio section theres this little preview on the bottom and it says the quick fox jumped loudly over the thin dog fence
  • d
    Read #answers, @Lexi (delulu posts on the daily).
1...921922923...1068Latest