https://uberduck.ai/ logo
Powered by
# machine-learning
  • c
    if he has around 200, 5k is more than enough
  • c
    If he has 300 or more, you can train it for longer
  • j
    alright ty
  • c
    is waveglow multispeaker
  • h
    i mean technically you can yeet multiple speakers at it,
  • c
    are there any tradebacks on doing that
  • h
    probably yeah
  • h
    but if your datasets are very small it might be better? iunno
  • c
    oh yeah i forgot to mention this but earlier i looked into the issues tab on the waveglow repo and theres people whove tried multispeaker before
  • c
    i think the guy who works on it said that the waveglow on mellotron/flowtron were single speaker
  • c
    so i have no idea if thats pona or ike
  • j
    Im pretty satisfied with the results. Without synthesis he talks quite fast but thats a problem with the source
  • u
    it is it stops training
  • h

    https://cdn.discordapp.com/emojis/680469243996930133.png▾

  • c
    ‼️
  • h
    i mean i've heard about the default uberduck hifi-gan being trained on like 9 speakers
  • p
    I was wondering about the reason of all audios required to be stereo and 22050Hz, is it because of processing requirements, I mean they would have more data, but going from mono to stereo doesn't sound too complicated, does anyone know why this is?
  • u
    afaik increasing sample rate almost quadruples training time and requires modifying tacotron2
  • j
    Yeah it did tbh, it doesn't sound perfect but it sounds really good and less robotic, I already sent the model to Uberduck and they already approved and turned it public if you want to test
  • j
    Maybe it needs some more training but I'm already satisfied with that for a first model
  • d
    Anyone know where reference file list is supposed to go and what format it's supposed to be in for controllabletalknet?
  • u
    in drive
  • u
    and same format as tacotron more or less
  • d
    Found it thanks
  • d
    Any tips on making a singing model?
  • d
    Should it all be about the same tone of voice?
  • d
    Should I include parts of songs where words a prolonged?
  • d
    are*
  • y
    Is there a Speech to text script that's good enough to help knock down manually populating the text file for tacotron?
  • b
1...713714715...1068Latest