https://uberduck.ai/ logo
Powered by
# machine-learning
  • b
    Yeah that's what I was going for, just wanted to give it a really decent run
  • b
    It actually went reasonably well, there's a little bit of a tinny sound in the voice and sometimes a bit of a vvv sorta sound, maybe just needs more training?
  • u
    ehhh
  • u
    sounds like it’s either a hifigan thing or a dataset noise thing
  • b
    How do I do additional training to the model from where I left off? Is that a thing?
  • u
    same notebook same dataset same parameters
  • u
    make sure the output folder is where the original checkpoint is so it can warm start
  • b
    cool thanks
  • s
    Is there a way to train models on mobile?
  • h
    the training notebook used to just work on mobile but nowadays it just removes all the code for some reason so you'll have to open the site in desktop mode
  • h
    also personally i would recommend my version which makes it easier to load datasets (though make sure it's properly formatted or it'll crash and burn) https://colab.research.google.com/drive/1xnFT5gXY5CIdsyG0YZbCulXdJksJR4ei
  • s
    Dus that work on mobile?
  • h
    yeah
  • h
    but like i said your dataset needs to be formatted exactly correctly
  • h
    including things that are hard to catch with a screen reader
  • s
    @hecko what should i press first to train something?
  • s
    because I don't know what to press to start making the Model
  • h
    all the play buttons in order
  • h
    except the first one, i forgot to remove it
  • s
    And wher can i inport my dataset?
  • h
    in step 2 you can paste a discord file link
  • s
    Ok
  • e
    https://twitter.com/niconicoHaruqa/status/1481972139037106181 anyone knows how to make Tacotron2 learn punctuation, exclamation, and question mark pronunciation? I think the original prebuild model for the LJSpeech dataset on uberduck doesn't include those, so I'm curious to see or hear the opinions of some people in here
  • e
    Ignore their laughing phoneme tho lol
  • h
    tacotron only looks at the text it's currently reading so it's hard to make a question mark affect a whole sentence the folks over at mekatron sometimes made separate models trained on just the question mark sentences, maybe that'll help
  • e
    Interesting, that's what one of my friends assumes too, how would you combine several tacotron models to then use all of these in conjunction?
  • e
    I'll have to check out mekatron tho
  • h
    mekatron is unfortunately dead, supposedly due for migration to uberduck but it's been months now
  • h
    i think talqu doesn't use separate models though, for one the words are short enough that maybe the question mark is close enough to affect the words
  • e
    I'd ask the Dev personally, but I can't write japanese unfortunately, so I can't really just hit them up lol
1...725726727...1068Latest