https://uberduck.ai/ logo
Powered by
# machine-learning
  • n
    L was so easy I have no idea why Orpheus isn't taking well
  • m
  • m
    this is seems nice
  • j
    I've tried it
  • u
    This seems like an interesting way to train voice models — it has tacotron 2 as an option too https://espnet.github.io/espnet/index.html
  • u
    It has fully implemented japanese and mandarin support, which is something that we don't currently have It uses the pinyin g2p for mandarin, and jaconv and openjtalk for japanese I wonder if openjtalk could be implemented as a cleaner for the sort of tacotron 2 training used here
  • u
    A sort of japanese text cleaner could possibly be created with pyopenjtalk https://qiita.com/nishiha/items/6e2a2ddaafe03fa7f924
  • u
    The biggest problem after that would be pitch accent. I think it would learn it wrong with phonemes only, and that would need to be taken into consideration with japanese text cleaners
  • l
    so like what happenes when you overtrain your tacotron thing? would it like glitch out when you do synthesis?
  • h
    yeah it'll forget how to read correctly and fail at anything that isn't in the training data especially if your dataset is small
  • t
    Yeah I overtrained Taliyah and she couldn't even say a wav from her dataset without stroke
  • m
    If you've got a working model and you want to improve it, what would you try first? More data, adding ARPAbet to the dataset or doing fine-tuning with HiFiGAN?
  • h
    i train all my models with arpabet nowadays so that's a null question for me if i think the prosody's wack then more data if i think the sound itself is wack then hifi-gan
  • m
    thank you! to use arpabet do you just add the lines to your transcription file? i don't see any parameters in the training notebook to specify you're using arpabet.
  • h
    it's done in a separate notebook
  • h
    #841437191073955920
  • m
    oh i don't think I've see that one
  • h
    however this one only does ~20% arpabet
  • h
  • h
    or train on pipeline once that's supported better
  • m
    thank you for sharing!
  • h
    glad to help
  • h
    might as well post my version of the pipeline notebook too https://colab.research.google.com/drive/1jF-Otw2_ssEcus4ISaIZu3QDmtifUvyY
  • h
    i've made it as plug-and-play as i could imagine but suggestions are always welcome
  • m
    oh, amazing! thank you
  • h
    np
  • h
    if it breaks it's probably my fault, send me screenshots of any errors you encounter
  • u
    that doesnt mean an overtrain that means an undertrain
  • h
    could be either
  • h
    or it could be excessive silence at the end s the ai doesn't learn when to press the "i'm done" button
1...829830831...1068Latest