https://uberduck.ai/ logo
Powered by
# machine-learning
  • h
    i think it's good but i've been wrong before
  • h
    also it's not a spectrogram
  • h
    it's a graph of what characters the ai is looking at (↑ axis) at what time (→ axis)
  • h
    here it seems to just be reading the sentence faster than in the training audio
  • h
    what's important is that it read it
  • a
    @User of
  • a
    8ok
  • a
    *ok
  • a
    sorry
  • a
    so when I get this again should i stop the model asap?
  • u
    well not really
  • u
    Sometimes training for longer will show the same graph but sound different
  • a
    ok, how will i know when to stop the training?
  • h
    you can look at the loss
  • h
    i've heard 0.1 and below is good
  • h
    but also you can make copies every so often and check which you like best
  • a
    when do you recommend i stop the training in mins? I'm worried about overtraining it
  • h
    ¯\\\_(ツ)\_/¯
  • u
    There's no real specific time
  • u
    How many samples are you training on
  • a
    154
  • a
    varying length, the actor uses a different voice in the first thirty or so
  • a
    and he is shouting in 3 of them
  • a
    does that affect anything?
  • u
    Then maybe go for like 100-200 epochs (close to 200 might be overtraining)
  • a
    i have been doing 150
  • a
    All the times I've trined it it comes out extremely robotic and demonic and it does not read too much
  • a
    but i made only a couple baby steps of progress in v5
  • a
    is 170 ok?
  • a
    @User still there?
1...474475476...1068Latest