https://uberduck.ai/ logo
Join Discord
Powered by
# machine-learning
  • h
    iiiiiit's subtle
  • u
    i was about to test this with a low data voice of mine but then i realized most of them are pipeline
  • h
    and?
  • u
    the plaintext one would be autoconverted
  • u
    redundant central
  • h
    i mean this one was autoconverted too
  • u
    i thought chills was legacy?
  • h
    was gonna remove it but didn't for i guess clarity
  • h
    yeah but arpabet
  • h
    in fact he was the arpa base
  • u
    oh pure arpa?
  • u
    i thought it was a mixed model
  • h
    not sure
  • u
    i thought you were comparing plaintext to the two arpa strings
  • h
    point is he has arpa on and that means everything gets autoconverted
  • u
    ah
  • u
    do i have any mixed models up on the site i wonder…
  • u
    i cant check because profile voice lists still broken
  • m
    Good look, I need to be able to deviate away from a word's normal verbal stress though (which is what ARPAbet identifies.) This is because I'm creating music out of the text, not just speech, so sometimes the word will deviate away from its normal pronunciation, and sometimes we need to ignore certain small words (like "today") altogether.
  • m
    I think I found the answer though, I'm surprised a search of "SSML" (and "Speech Synthesis Markup Language") of this discord srvr came up nil though.
  • m
    For posterity, I'm trying AWS Polly to do this. I tried Google's Text-To-Speech but their authentication system is so unnecessarily complciated to me: https://docs.aws.amazon.com/polly/latest/dg/supportedtags.html
  • m
    hecko do you work for uberduck? You're always on top of this stuff, thank you for that
  • h
    not really work, just moderate things
  • r
    Isn't there like, no commonly accepted standard? SSML is made for this application, but not everyone supports it. Arpabet is made only for American English which is not helpful
  • r
    IPA idek what's the opinion on that
  • h
    ssml is supported by all the big players really
  • h
    microsoft, amazon, google,
  • h
    for phonemes there's also a universal system, ipa, but idk how much that is supported
  • h
    i know amazon supports it, and uberduck has an ipa symbol set but apparently there are issues with getting it to work
  • r
    Ik pepe have treied to train IPA but I haven't heard much from other than saying they would
1...102710281029...1068Latest