https://uberduck.ai/ logo
Powered by
# machine-learning
  • e
  • e
    Pog hifigan2 dropped
  • e
  • m
    Yo sounds great
  • h
    when can we use it :3c
  • u
    when tacotron3 comes out
  • s
    Erin Esurance failed attempt
  • s
    Guess I’ll add more WAV files tomorrow
  • h
    I just realized my learning rate was messed up
  • h
    😩
  • a
    y'know, I never tested what's the minimum amount of data needed in order for something to sound good.
  • a
    Since I know 7 hours of atenborough was enough for it to be perfect, what's the lowest amount needed?
  • a
    I was initially planning to do 30 minutes, 45 minutes, and an hour of audio. But just 30 minutes of data is enough for it to sound perfect
  • a
    The Aerith dataset never sounded that good, even with over an hour of audio. And that's probably because Aerith wavs have her acting in every wav so that probably threw off talknet
  • a
    May as well test 10 minutes and 20 minutes of audio and see if that
  • a
    see if that's enough
  • u
    for talknet honestly as long as its like over 15 wavs a million and one other factors dictate whether the model will sound good
  • u
    as for tacotron
  • u
    my acemillion model was trained with 15 wavs and it sounds impeccable
  • a
    yea, i'm starting to think what's most important is make sure every wav sounds similar in speech and tone.
  • u
    i guess a neutral speaking tone dataset with as many phonemes as possible
  • u
    is what makes a good dataset for both archetypes
  • a
    yea, i can agree to that. That's the same conclusion from the examples I've seen so far
  • a
    but, gotta test it to make sure. I wonder how small the number of wavs I can decrease it down to and have it still sound perfect
  • a
    with less wavs, I need to decrease batch size right? 109 flies. So prediction trainer should decrease 64 -> 20. And spectogram should decrease 32 -> 8
  • a
    otherwise too large a batch size means a lot of files get discarded while training, right?
  • h
    i think that's the sequel to the other hifi-gan, the denoising one instead of the spectrogram→audio one
  • e
    Yeah I figured that out too. Though it seems pretty good at cleaning audio
  • h
    in your case i think batch size 18 would be ideal, making 6 batches while discarding 1 wav
  • u
    What happens when you use multiple voice providers to voice a single model? Has anyone ever tested it?
1...791792793...1068Latest