https://uberduck.ai/ logo
Powered by
# machine-learning
  • b
    That sounds like its the spectrogram generator’s fault. How many epochs did you train it for? I usually let it go to 1000-1500
  • m
    You mean the hifi gan or actual tacotron model? Tacotron I did the average 250
  • b
    Oh you’re using hifigan on tacotron? I assumed you were using TalkNet, sorry
  • b
    Did the tacotron model sound less metallic than the hifigan?
  • m
    The reason I tried finetuning hifi gan was too remove the metallic because someone said it could have been the vocoder, but training hifigan hasn’t changed anything
  • m
    I don’t know if I just have done something wrong or if its actually the model itself
  • b
    Hm
  • b
    what did your melspecs look like when training in tacotron
  • c
    Do you guys know how can I make hifigan output mp3 instead of wav?
  • c
    I tried converting here in the code but I got left with a noisy audio
  • b
    you can have it auto convert using ffmpeg
  • g
    i think my model is doing a good job, it has turned itself into a nature sound generator
  • g
    now if only i knew why...
  • c
    what arch is that
  • g
  • g
    its super broken
  • g
    ah apparently there was some incorrect indentation
  • g
    python moment
  • u
    I keep getting an error whenever i put the .txt file into filelists on Tacotron
  • h
    please post a screenshot of the error
  • b
    oh cool, I'm currently creating a notebook for a pix2pix-based voice style transfer network
  • b
  • b
    (its in japanese tho lol)
  • b
    the git repo is english tho
  • b
  • g
    oh thats pretty cool, how is that going?
  • b
    I started on adding the git repo to the model
  • s
    no
  • g
    apparently the true cause of the exploding loss is this stupid horrible line of code that produces a negative learning rate after a while https://github.com/ebadawy/voice_conversion/blob/master/src/models.py#L27 always love me some gradient ascent
  • b
    bruh
1...665666667...1068Latest