https://uberduck.ai/ logo
Powered by
# machine-learning
  • u
    banned from pixel
  • u
    (you are still allowed to be hy)
  • h
    here's a thing that might be useful for those who want to clone their own voices https://github.com/MycroftAI/mimic-recording-studio
  • h
    idk how well it works though, i personally use shtooka recorder with some python post-processing
  • h
    ...actually heck this requires post-processing too so not that useful
  • b
    how did you fix it?
  • b
    Im also having that issue
  • u
    clone it to your drive
  • u
    and it says theirs a new version at the top
  • u
    use that instead
  • b
    tysm
  • b
    Im getting this issue did you run into this at all?
  • u
    Yeah
  • u
    I just refreshed the page and ran a new instance
  • b
    k thx
  • b
    Yes its working perfectly now thank you
  • b
    Time to do a tonne more training ill be back in 2 weeks lol
  • u
    it takes like 40ish mins to get 0.15
  • u
    apparently i gets so much better at 0.08
  • u
    so yeah
  • u
    if you want to wait the lil bit extra, be my guest lol
  • u
    honestly 0.09 is where id even call it functional
  • u
    0.15 is just way too stroke prone
  • u
    yeah
  • u
    i trained tutorial unit to 0.14 and it worked great on my test sentences, until it got released and it has a stroke 9/10 times
  • b
    How many files do you guys use for a good model and what batch size and epochs do you use? I've got 75 files with 15 batch size and 500 epochs
  • g
    With that many voice clips, I recommend training with the batch size of 14 for 150 epochs.
  • c
    i think woody woodpecker was bs14lr5e-4 and there was 97 wavs and it sounded fine at like 200 epochs or something
  • b
    Got it down to 0.041 before it disconected me for inactivity
  • g
    That should be more than enough.
1...724725726...1068Latest