https://uberduck.ai/ logo
Powered by
# machine-learning
  • u
    obviously it would sound perfect with what i assume is just directly reaching into the dataset
  • a
    here's another test case. Greet.wav is without reference audio. Greet2.wav is with ref audio where I try speaking. Anyone have any ideas on why it's so scratchy and how to fix it?
  • a
  • a
    i went and submitted it to uberduck already
  • a
    done with talknet, 600 epochs and 4000 steps for hifigan
  • u
    hmm
  • a
    tacotron seems to have better results, but you don't have control over the pitch and how the sentence forms. If you can combine tacotron results + talknet reference control, then that would be perfect
  • u
    try 6000 steps hifigan
  • a
    would 6000 do it? From 3000 to 4000, the numbers just seem to jump up and down with no improvements
  • u
    normal hifigan recommends 5000+ steps so
  • a
    alright, I'll try it tomorrow. it's getting late. Gonna sleep first
  • j
    So how exactly do I continue from a checkpoint? I still have the "begin training" step open from my last session, and it has the default pretrained model at 28%. Do I just run it again, or is there something else I'm supposed to do?
  • h
    as long as the model path and model name are the same you can just start training again, it'll load the model from your google drive and keep going from where it last saved
  • h
    you'll need to reupload the dataset though
  • k
    hecko. maybe you could answer a question?
  • k
    or maybe explain something?
  • k
    is there a way to install tacotron2 voice packages as a voice in Windows ten?
  • k
    for my program
  • k
    it uses any installed voice
  • h
    i'm not aware of a way to do so unfortunately
  • h
    though i'm sure someone who knows their stuff could make an app to do so
  • k
    do you know how cereproc s?
  • k
    s
  • k
    installs
  • k
    sorry
  • k
    I just use custom voices in Windows
  • k
    and wanted to use one from here
  • h
    someone would have to make an app that connects the windows text-to-speech system to the uberduck models, and we're short on programmers unfortunately
  • h
    even then you'd have to be able to download the model which is only possible with some of them
  • k
    I can't actually speak is why I ask
1...767768769...1068Latest