https://uberduck.ai/ logo
Powered by
# machine-learning
  • u
    got the text cleaners part
  • u
    i give up
  • m
    does your pipeline notebook expect the wavs to already be converted to 
    .npy
    files? currently getting this when i try to start training: 
    FileNotFoundError: [Errno 2] No such file or directory: 'speakers/0000_SeanConnery_v01/wavs/083.npy'
  • h
    pipeline doesn't convert to .npy ever
  • h
    go into your list.txt and replace .npy with .wav
  • m
    ah ok
  • m
    if anyone else every runs into this:
    Copy code
    transcript_file = open("/content/project/transcription.txt", "r")
transcript_lines = transcript_file.readlines()
for index, line in enumerate(transcript_lines):
  line_splits = line.split('.npy')
  newline = line_splits[0] + '.wav' + line_splits[1]
  transcript_lines[index] = newline
  
transcript_file = open("/content/project/transcription.txt", "w")
transcript_file.writelines(transcript_lines)
  • h
    that feels overcomplicated
  • h
    i'd replace lines 3 through 6 with just 
    transcript_lines = transcript_lines.replace('.npy', '.wav')
  • m
    I think this needs to be done before the dataset is brought into the notebook as 
    /content/project/transcription_val.txt
    gets created with 
    .npy
    extensions after training is run. (even if the above is run on 
    /content/project/transcription.txt
    , before training)
  • m
    AttributeError: 'list' object has no attribute 'replace'
  • h
    -oh right because you converted into a list
  • h
    you can replace 
    readlines
    and 
    writelines
    with 
    read
    and 
    write
  • m
    Or just go to browserling.com and substitute text
  • m
    @hecko, now running into this, any thoughts?
  • h
    ,,honestly i have no idea what this could be
  • u
    or just use find and replace
  • j
  • j
  • m
    Hey guys. could someone explain me more about what is pipeline, TorchMoji and GSTs? I searched for information but didn't find anything useful. I'm new to the subject, I'm learning as much as I can about the tacotron.
  • d
    For training a talknet model should I use google colab or my own rig I have a gtx1660 and a i5-9400f
  • t
    Proto (not the one with a mustache) from the 2016 ad on TortoiseTTS
  • t
    Here's another sample:
  • t
    And a third one.
  • t
    Original audio file:
  • i
    actually very accurate
  • i
    how many minutes of source audio did you use?
  • t
    I used 1 WAV file, 8 seconds long.
  • i
    Incredible
  • h
    pipeline is short for the uberduck pipeline, meaning our own reimplementation of tacotron and eventually other models torchmoji is a network trained to predict emoji from text, 15.ai uses it for emotions and pipeline tries to do so too (though personally i haven't noticed much of an effect even with big models) gst (global style tokens) is i guess a general term for inputs that influence emotion, the original by google used an embedding from audio into a few dozen sliders
1...854855856...1068Latest