https://uberduck.ai/ logo
Powered by
# machine-learning
  • h
    https://mvsep.com/ with either 
    Ultimate Vocal Remover HQ
    or 
    MDX B
  • h
    in my experience the latter is worse at actually removing stuff but the output is hella crisp
  • h
    UVRv5 Demucs
    also exists but i haven't tried it
  • j
    Seems like UVR is the best for this
  • j
    Need some input on the reference audio
  • j
    Would this be good enough for Talknet?
  • g
    I think so
  • j
    I'm a bit worried with some of the string instrumentation still being somewhat audible in the background
  • j
    Which those seem impossible to completely remove since a lot of the string instruments also leak into the human vocal range.
  • j
    Specifically violins and cellos.
  • g
    Try converting the audio into 22,050 Hz
  • j
    In which step of the process?
  • g
    I mean converting the reference audio into 22kHz will lower the high frequencies somewhat
  • j
    Dimitri from Sly Cooper series test (prolly gonna train this for a bit and then a hifi-gan)
  • t
    Epoch 45 progress after restarting, looking much cleaner
  • t
    Is AITCH really the best model to use for this?
  • h
    ¯\_(ツ)_/¯
  • j
    Dimitri from Sly Cooper (probably finished and now with hifigan)
  • c
    https://github.com/bshall/soft-vc I was checking this, the results are amazing! does anyone know how hard is it to train a model for it?
  • m
    it doesnt seem they released any code
  • c
    Oh, I didn't noticed, that's sad
  • j
  • t
    My training on the Japanese models gives me zero useful results for English speaking. The audio files aren't the proper length.
  • t
    Japanese input text also seems to be awash, and the closest I can get is this
  • u
    Hmm How are you training again? With one dataset transcribed in english and one in japanese..?
  • t
    I was just riding on the 100 speaker Japanese dataset while mounting pretrained model AITCH
  • t
    No ARPAbet
  • u
    Ah that wouldn't really work for english then And it would sorta mess up japanese synthesis
  • t
    Hey dude. I'm sorry for the ping, but I wanted to let you know that yesterday I made two new Wiggles voice models: Captain Feathersword, and Murray Cook! Soon, I plan on doing Jeff Fatt and an updated Greg Page if both go well and sound decent.
  • t
    Are the models out?
1...950951952...1068Latest