https://uberduck.ai/ logo
Join Discord
Powered by
# tacotron-2-support
  • m

    Mortam

    12/10/2022, 2:11 AM
    what if its just a minor part、like 10% or so
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:11 AM
    the problem will still exist
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:11 AM
    just will happen a lot less
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:11 AM
    though some eqing on the seaman lines should blend them better
  • m

    Mortam

    12/10/2022, 2:12 AM
    this shit will be a nightmare to train
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:14 AM
    it wont be a nightmare
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:14 AM
    unless you Hate tediousness
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:14 AM
    in which consider the next week or two a personal hell
  • m

    Mortam

    12/10/2022, 2:15 AM
    the transcription part is great、so that's good at least
  • m

    Mortam

    12/10/2022, 2:15 AM
    was paranoid starting out thinking i was doing it wrong so i checked a lot of lists and the one on github for examples
  • m

    Mortam

    12/10/2022, 2:16 AM
    being careful spelling out numbers and splitting compound words into two
  • m

    Mortam

    12/10/2022, 2:16 AM
    like 「boat house」 instead of 「boathouse 」
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:17 AM
    i dont see a reason to do either of those things
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:17 AM
    numbers get autoconverted to plaintext by our notebooks
  • m

    Mortam

    12/10/2022, 2:18 AM
    oh that's nice
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:18 AM
    and compound words i dont recall any instance of them being an issue
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:18 AM
    in fact they would probably help because models ive noticed can be kinda dumdum doodoo when it comes to those
  • m

    Mortam

    12/10/2022, 2:18 AM
    i did run it for a bit on just 1 hour worth of audio until my computing units ran out and it got to the random prompt i gave it on just 60 epochs
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:18 AM
    but then again this is Base Model level data
  • m

    Mortam

    12/10/2022, 2:19 AM
    obviously robotic but it got the voice down even with mixed audio from seaman in
  • m

    Mortam

    12/10/2022, 2:19 AM
    > Zach, HALT the flow, of time.
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:20 AM
    ew is that not hifigan
  • m

    Mortam

    12/10/2022, 2:20 AM
    no idea
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:20 AM
    that sounds like something the pipeline notebook would spit out while training rather than a usage of the end result
  • m

    Mortam

    12/10/2022, 2:20 AM
    yeah
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:20 AM
    it would sound better with hifigan
  • m

    Mortam

    12/10/2022, 2:20 AM
    computing units ran out and that was the last one
  • m

    Mortam

    12/10/2022, 2:21 AM
    will probably get a friend with a card to train it for me when its finished or something
  • m

    Mortam

    12/10/2022, 2:21 AM
    cant be bothered babysitting it in colab with multiple accounts
  • u

    (Dawn) Will Draw Fictional Women

    12/10/2022, 2:21 AM
    oh absolutely not especially with colab pro being a scam and drive being whiny with rate limits or oh i dunno
1...858687...158Latest