https://uberduck.ai/ logo
Join Discord
Powered by
# talknet-support
  • x
    Could you explain to me what I'm doing wrong?
  • j
  • j
    Use this nb for training your TalkNET models
  • j
    also, make sure to run this cell and restart the runtime
  • x
    Thank you.
  • p
    can someone please help 💀💀
  • j
    dm your dataset
  • j
    lemme take a look
  • x
    Ohmygod it worked tysm-
  • x
    The actual model file isn't showing up in the Talknet folder on export, in my drive-
  • c
    No gpu
  • c
    change runtime to GPU and try again
  • c
    Training started at step 2500010 and it suggests deleting this folder which I am unable to do
  • c
    so do I have to keep having it train at 2500000
  • c
    alright
  • x
    For transcription, is it okay to put commas, or will that ruin it?
  • h
    Hello
  • c
    Is it OK to train the spectrogram generator for 1000+ epochs
  • h
    you're not gonna get that far
  • u
    pretty sure the notebook physically wont let you do that
  • h
    even if it does the notebook will disconnect far before 1000 epochs
  • j
    why do you wanna overtrain it?
  • p
    Overfitting is bad you know...
  • j
    200-250 should be enough
  • p
    Now that I'm training it, should it start from 25000 epoch by default?
  • p
    does anyone know how to make a dataset sound more natural with only like 150 clips or so?
  • p
    The only two options that come to mind are data augmentation and pre-training
  • u
    It really depends on the amount of data you have Ik i've gotta 200 clip models that have sounded bad because the data only added up to 2 minutes Try pretraining using crust or aitch on pipeline, and make sure to train a hifigan model after
  • u
    If it's still not enough data, you can make a multispeaker
  • p
    can you send the link to crust for me?
1...8910...74Latest