https://uberduck.ai/ logo
Powered by
# machine-learning
  • t
    Of course I have to deal with collecting and transcribing Japanese audio for a given character
  • t
    Fortunately, S2 of the show has Japanese subtitles so I can probably work with that.
  • t
    And hopefully it'll give me enough to attempt my first model
  • h
    that's specifically for speech recognition though, so punctuation and signal cleanliness and sampling rate are yeeten out the window plus it takes data from random sources so there's not even a cc license (though i'd argue a license isn't enough for ethical voice cloning) it could still be interesting to use, but personally i'd rather get a hundred character voices than a million random person voices
  • h
    speaking of i should really get around to doing the paladins voices if i got a good forced aligner i could even automate the non-speech removal
  • s
    Doug Dimmadome with 15 WAV files.
  • s
    Some of them being tongue twisters.
  • h
    bru that model spittin fax
  • s
    Someone should make a music beat out of it.
  • s
    This is the only failed model I’m not deleting off my Drive.
  • s
    Because of how god tier it is.
  • z
    Doug Dimmadome do be beatboxing
  • s
    I shall train him on TalkNet once I do a functional model with more WAV files.
  • h
    Also I remembered that I can't connect to GPU backend so I put my Video Game Agent Mark dataset as a request for training: https://discord.com/channels/768215836665446480/843619158834020393/940392246199611412
  • c
    anyone know why model dont learn attention other than "make sure your wavs align with the text"
  • c
    there is slight line but it feels like it never gets close to good alignment
  • s
    Do you think the reason why Doug Dimmadome ended up like this is because of me using 15 WAV files or because I transcribed long sentences involving tongue twisters?
  • s
    (Every WAV is under 15 secs, by the way)
  • u
    both
  • u
    btw what's your pfp it seems familiar
  • c
    MMI in Tacotron 2?
  • e
    There are some Japanese speech corpora out there, you'd have to use a phonetic language to make the Japanese training easier tho (hiragana and romaji)
  • e
    Tho, there's I think a good tacotron2 pre-trained model out there
  • h
    pose is from sonic adventure
  • s
    Pinky Dinky Doo doing the Sonic Adventure cover pose
  • t
    That's called fanart
  • m
    FNAF Jack O Chica voice... she isn't the greatest with long sentences but other than that i think its okay... maybe?
  • u
    jack o chica tacotron got hifigan like
  • u
  • h
    i hate this now
1...735736737...1068Latest