https://uberduck.ai/ logo
Join Discord
Powered by
# tacotron-2-support
  • c
    840 epochs, that's too much
  • c
    you overfitted the model
  • a
    is there any way making it less janky with less epochs
  • a
    also thre is no refrance adio o use
  • a
    there is no more refrance audio to use*
  • a
    well i think it sounds janky because im too smart and name 2 wavs with the same name
  • a
    do i need to retrain?
  • f
    using the legacy tacotron 2 model and got this error, can someone help me figure this out?
  • f
    nvm issue fixed, i just switched to the FakeYou tacotron 2 notebook
  • g
    840?? No wonder it sounds janky. You overfit the model.
  • g
    Train the model for 200 epochs.
  • a
    mhmm okay thanks!
  • a
    RETAINING T I M E
  • f
    noo FakeYou model tester is giving me this message
  • f
    i already have it set to anyone can edit with the link
  • g
    I see the problem: You pasted the whole link instead of only the ID part of the URL.
  • a
    so i trained it for 200 but i think it needs to be 300
  • c
    No, it doesn't
  • c
    If you model didn't came out good with 200 epochs using only 6 minutes, your 6 minutes of audio isn't good
  • c
    Or you transcription can have typos
  • c
    Around 150 epochs you would already get a good model with 18 batch size
  • p
    Hello! It's been a while since I've typed here I missed silly talking robots But yeah I've encountered an interesting problem on the pipeline interference notebook while testing my Polish model
  • p
    I've recorded a fresh dataset earlier and the training seemed successful, but for some reason __the model rolls from accented letters back to Latin letters when rendering speech__, making the pronunciation wack af
  • p
    As an example 
    Częściowo śmierć się ułaskawiła
    gets pronounced as 
    Czesciowo smierc sie ulaskawila
  • p
    Do we know what might be causing that ? The model was fine and pronounced everything correctly when I listened to tensorboard samples
  • p
    Just now was able to sit on the PC to render the sample The input was the first sentence, but the output sounds like the second one (incorrect)
  • p
    This manner repeats with other inputs as well
  • p
    Again, input text: 
    Śmieszny ten błąd, który nie wiadomo skąd wynika.
    Gets pronounced as 
    Smieszny ten blad, ktory nie wiadomo skad wynika.
  • h
    uhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhh did you select the polish base
  • h
    -or sorry no if it's a synthesis problem then
1...909192...158Latest