https://uberduck.ai/ logo
Powered by
# machine-learning
  • y
    you are awesome
  • h
    same tone yeah, unless you mean going to falsetto at a certain pitch which might be fine (i'm not sure because i think it ignores the octave of the pitch)
  • h
    yeah absolutely
  • p
    For voices trained on little data the validation list should it be the best samples or should the best samples be in the normal training data or in both
  • p
    My experience with making singing voices it seems to ignore it. Tried to make one with belting on the high notes but it just randomly belts.
  • h
    Has anynone tried or managed to make tacotron training on TPU?
  • h
    I know there is not working waveglow training on TPU
  • p
    Can Talknet take more than 22khz in training datasets? I remember hearing that it can, and it does synthesize like 32khz clips, but in the training notebook it says: only 22khz, my current model could benefit from higher frequency clips.
  • c
    You can change the rate in the code
  • p
    @User Would it make training better or worse?
  • p
    Like longer or shorter training time
  • h
    you'd have to train the base network from scratch i think??
  • h
    which would take a week or more
  • p
    oh damn, guess I won't be training it then
  • p
    How does it output 32khz if it is trained on 22khz?
  • c
    hifigan can upsample (I Think)
  • p
    oh, that explains it
  • c
    Probably that
  • c
    You can try using Tacotron 2 MMI
  • s
    My model still stutters even though I trained it on normal settings and it has 85 WAV files.
  • s
    What to do?
  • u
    give the dataset s'more wav files of him talking
  • u
    get it?
  • s
    Will changing some settings in the colab also fix the problem?
  • s
    As in batch size and stuff?
  • u
    i believe so
  • t
    So I've been doing this new multilingual voice cloning method
  • t
    And at the training step I got this error
  • t
    This is me attempting to run it with only one of the datasets though rather than all 10 languages (because all I need is one)
  • t
    So I don't quite know what to remedy if it's possible to work with just Japanese
1...714715716...1068Latest