https://uberduck.ai/ logo
Powered by
# machine-learning
  • k
    I use special software
  • h
    ah
  • k
    currently use a Scottish tts
  • h
    well there are apps that let you put in text and they'll speak with tacotron for example https://github.com/lokkelvin2/tacotron2-tts-GUI
  • k
    I use grid3
  • h
    oh mmm
  • k
    I type with eye tracking
  • h
    welp
  • k
    I can do it but running out of screen space
  • k
    haha
  • k
    it works on twitch already
  • k
    I just want to use it
  • h
    well like i said there's unfortunately no app to do that it could theoretically be made, but all the programmers here are busy with other things
  • k
    I have a questionj and I will shut up
  • k
    if I get someone. can crossbreed priscila be a package?
  • k
    that is the last question
  • h
    not unless someone decides to make it
  • k
    ok ty! I'll just have to settle for twitch
  • j
    Checkpoint worked. Finally got it to 249 epochs. Sadly it doesn't really sound any better. Either it needs way more training or the data set is just shit...
  • h
    you could try training with this notebook instead, it's a bit of a mess but it uses a thing called arpabet (pronunciation_dictionary in the synthesis notebook) to make pronunciation better https://colab.research.google.com/drive/1h9NMKterow9s8ggC87s8wM6iQaD0T8Lh just remember to tell a site moderator to enable arpabet after you submit
  • j
    Does batch size make a difference in the training? I have it set to the default 18, but am wondering if changing it would get better results...
  • h
    if you have very few wavs you should set it so that the amount can be divided cleanly by the batch size
  • h
    because if you have e.g. 50 wavs at batch size 18 then it makes 2 batches of 18 and then throws away the remaining 14
  • h
    if it can't divide cleanly then you can copy and paste the entire list.txt so there's e.g. 15 copies and then set batch size to also 15
  • h
    outside of that i think bigger batch size leads to more stable training and thus better results, but i haven't tested it and some people disagree
  • j
    Maybe that's what's happening, then. This is pulling from a dataset of 269 files, so a bigger dataset might yield better results.
  • h
    269 is pretty large for uberduck standards
  • h
    i've trained on a dataset of 6 minutes' worth of wavs and it turned out very well in my opinion but that's probably thanks to arpabet
  • j
    bigger batch size is what I meant. Dataset's already big, as you said.
  • h
    just be careful not to make it too big or it'll crash from lack of memory
1...768769770...1068Latest