https://uberduck.ai/ logo
Powered by
# machine-learning
  • u
    Sure thing I guess!
  • a
    anyone ever tried to upload like 800mb of wavs to collab notebook? it keeps timing out
  • h
    try uploading them to google drive instead, and then loading them with some code like 
    !cp /content/drive/MyDrive/sans-undertale-dataset/wavs /content/tacotron2/wavs
  • a
    lol, i was thinking of changing the google collab code, but that's a lot easier to do
  • a
    good idea there
  • h
    my personal version of the notebook has a downloader that can get datasets from e.g. discord https://colab.research.google.com/drive/1xnFT5gXY5CIdsyG0YZbCulXdJksJR4ei
  • h
    just so i can resume training with one click (plus google drive login) without having to reupload the dataset
  • j
    im so confused on how to use the pipeline
  • j
    where the finished model go
  • j
    nvm
  • m
  • h
    Yeah, art is a human’s job
  • h
    ??
  • h
    yeah i think i've read something about that in the context of text unsupervised output has no human involved so can't be copyrighted, but if you do some curation then you can claim copyright over the collection (and if control then over the piece)
  • m
    I'm now working on the improved Smash Announcer with 3 more minutes of data than before! I'm gonna try it with normal pretrained, then try it with a reverb pretrained voice, and then try with lower learning rate...
  • m
    I'd like to avoid using lower learning rate cause the model doesn't pronounce things great if I do so...
  • m
    it's doing something....
  • m
    warm starting now since it didn't work... seems to be doing better:
  • m
    this is only epoch 12 while last time was epoch 34
  • m
    yo! this is good!
  • m
    it only lasted one epoch tho 😦
  • u
  • m
    getting better! this could be the new model!
  • m
    ok its done! it speaks english properly, seems to be good with long sentences (most of the time) and pronounces better!
  • m
    i'm trying to compare the difference in other parts other than pronouncing things...
  • u
  • u
    fuck dude
  • u
    makes me wanna try the fighting layer announcer again
  • m
    it's better and not better at the same time is my final judgement! they both seem to be better at different things... new one is better at longer sentences and pronunciation and old one is better at speaking without stuttering or stroking unless it is too long of a sentence...
  • m
    i'm just gonna publish it, and hope people like it...
1...749750751...1068Latest