https://uberduck.ai/ logo
Powered by
# machine-learning
  • c
    (mel spectrograms, not attention/alignment graphs)
  • b
    On colab? How are you getting all the graphs
  • c
    im doing this locally
  • b
    Ah I had a feeling you were
  • c
    i just plotted the datasets npys to graphs so its easier to identify which mel spectrograms were farting
  • b
    Ah
  • c
    i have a feeling attention learning is greatly being impacted by those 3 faulty melspecs
  • c
    because i still havent gotten proper aligment image
  • b
    Hmm
  • c
    also these are the parameters im using 
    wile-e_bs4_lr1.00e-04_cmu
    im using 4 because 32 / 4 = 8 and my gpu can handle it
  • c
    looks like somebody had the same issue that i have right now i think i found the problem if you save the wav as "32bit float" mysterious things will happen and the stft will shit
  • h
    I am trying to update my model with 20 voice clips. I'm thinking of updating it with twice as much clips as before (40 clips). Do you guys think 40 clips are enough?
  • b
    i usually use at least 100 but 40 is definitely an improvement
  • h
    Ok thanks
  • h
    Maybe I'll use a little more than that
  • j
    Input
  • j
    Output
  • h
    Can any of you fix my voice clips? I'm having issues with MEL spectograms again
  • h
    I'm trying to update my model
  • j
    Some of the clips have background noise in them
  • j
    Clean them and it should work fine
  • h
    Sorry but last time someone did that for me
  • j
    That was me
  • j
    But I didn't clean the dataset haha
  • h
    What did you do
  • j
    I fixed the dataset for you
  • h
    how did you fix the dataset?
  • j
    It was straight up videos
  • j
    Not in the correct format and all that stuff
  • j
    Go through #841422801965416538 channel
1...677678679...1068Latest