https://uberduck.ai/ logo
Powered by
# machine-learning
  • c
    I think you need English and Japanese data
  • c
    If you want to do Japanese to English
  • t
    The GitHub they sent gave me a Japanese dataset in the ten provided languages None of which were English though
  • t
    It's all a matter of understanding how to apply all this
  • d
    Noob question here
  • d
    Is there a difference between training a model straight up with 400 wav files
  • d
    And training the model once with 200 and then another 200 after?
  • c
    sounds like it would be too much work i dont know what it would achieve to do 200 at a time
  • d
    U mean training it two separate times?
  • d
    Just want to test what the model sounds like at 200, 400, 600 etc
  • c
    its better to just train the entire dataset rather than parts of it
  • c
    but you can split the dataset for validation and training
  • c
    (validation is basically the model telling you how good it is at replicating the source material (lower loss==better))
  • d
    So i have a model trained with 200 now
  • d
    Should i add the 200 to the folder?
  • d
    Or only train with the new ones
  • c
    if the rest of the 200 is consistent with the rest of the dataset go for it
  • c
    as for dataset swapping during training mega told me to just train a new model for it instead of resuming with the new dataset
  • d
    What do you mean by consistent?
  • d
    Say theres a few words my model tends to sound kind of robotic
  • u
    same speaker same tone of voice similar speech pattern
  • c
    the wavs should sound more or less the same in tone (you can see this by inspecting the spectrogram of each clip)
  • d
    Would it be a dumb idea to train with say 5 files that contains that word?
  • c
    not at all
  • u
    definitely not
  • d
    Alrighty then
  • d
    Thanks a lot
  • c
    @User you know your guide from github or whatever it is right. you should make a page for terminology
  • a
    i still forgor about local training guide
  • c
    ‼️
1...711712713...1068Latest