https://uberduck.ai/ logo
Powered by
# machine-learning
  • u
    If someone were to make a macro that did the sentence marking part automatic, thank you SO much
  • h
    audacity has a silence detector but in my (one) experience it's not very useful, still gotta check through the entire audio to see where it messed up
  • h
    though i have heard of a notebook that automates clipping and transcription, but there's some setup involved
  • h
  • h
    [nsfw language because 4chan]
  • p
    That happened to me too.
  • j
    some decent stonks on Brock
  • m
    The Easter Bunny so far, I did CSTT already.
  • u
    Interesting..
  • m
    They're from Wow Wow Wubbzy, and surprisingly I would've made 10 wubbzy voices if him and Captain Sweet Tooth Tom were on the site.
  • t
    hey, anyone here know if training fastspeech2 is similar to training tacotron2?
  • h
    can someone test loading a dataset from google drive (any type, doesn't matter) because for me it only works on files i've loaded before and i wanna know if it's a global thing or a me thing
  • t
    Currently working on the TalkNet version of Gene Wilder.
  • t
    Here's how it turned out.
  • h
    hm maybe i should revisit the idea of training on vocals from ccmixter
  • h
    my excuse being that all the high-quality singer models are female (ponies) so let's get a male one
  • p
    So uh, I'm considering making a project out of training a model on the voices of a certain couple of youtubers who always come as a pair and who have a LOT of content, much of which is captioned
  • p
    as in like, actually manually captioned
  • p
    so uh. now i'm wondering a) is there a way to download a video from youtube with the captions as metadata and b) how...helpful would that be to tacotron?
  • p
    because like...yeah youtube captions are timestamped but they're only marked every few seconds or so, and they're not always PERFECTLY in time
  • p
    i guess just...how much manual phonetic marking is needed for a dataset? does that have to be done by hand, or does part of the neural net handle that?
  • p
    ...actually, the guys i'm thinking of resynthesizing are pretty damn good singers themselves, though they DO tend to do a lot of harmonies with each other, which...miiight be problematic
  • p
    i know neural nets can separate voice from instruments, but...have they been able to separate a voice from another voice?
  • p
    i know some versions of melodyne have "direct note access" but i've only seen that used mostly on instrument channels
  • h
    youtube-dl can download subtitles, and there's a tool to use those for clipping https://github.com/HexCodeFFF/aidatasetfromvideo
  • p
    ooooooo
  • h
    vocal remover has a thing to separate background vocals from lead ones but idk how well that would work if they're both lead
  • p
    oh wow
  • h
    you could potentially get the bits where they sing the same thing separately, overlay them, then finetune the nn to separate them
  • h
    though you'd need 11gb of vram for one
1...406407408...1068Latest