https://uberduck.ai/ logo
Join Discord
Powered by
# machine-learning
  • h

    hecko

    12/10/2022, 1:41 PM
    i recommend going with batch size 2 or even 1 it'll take longer but turn out better
  • h

    hecko

    12/10/2022, 1:41 PM
    who told you that
  • h

    hecko

    12/10/2022, 1:42 PM
    30 wavs with 10 minutes total is 20 seconds per wav on average, which is way too long the standard is to have them between 2 and 12 seconds each
  • a

    Amizade | Pony's voice creator

    12/10/2022, 1:42 PM
    I tought that the batch size work varying with the amount of wavs
  • h

    hecko

    12/10/2022, 1:46 PM
    well it's like if you have fewer wav·s then you should let the ai train more slowly for best results and lower batch size does just that
  • h

    hecko

    12/10/2022, 1:47 PM
    though you should adjust the learning rate too, to something like 1e-4
  • a

    Amizade | Pony's voice creator

    12/10/2022, 1:48 PM
    I had put it like this in the learning rate
  • a

    Amizade | Pony's voice creator

    12/10/2022, 1:48 PM
    now in the min learning rate, I left it at 1e-5
  • r

    Reclezon

    12/10/2022, 1:54 PM
    For 30 wavs — I'd use 3. None are thrown out
  • a

    Amizade | Pony's voice creator

    12/10/2022, 1:55 PM
    okay, I'll try that. Upload 30 wavs and put 3 in the batch size
  • a

    Amizade | Pony's voice creator

    12/10/2022, 1:55 PM
    I hope it works
  • h

    HarmonyOmega

    12/10/2022, 2:14 PM
    Larry Lovage model I've been trying to do for the past few days
  • r

    Reclezon

    12/10/2022, 3:12 PM
    Can't really decide if the new model is actually better or if it's just of the HiFi-gan
  • r

    Reclezon

    12/10/2022, 3:14 PM
    Also annoying that each app renders audio noticably different
  • l

    Lexi (delulu posts on the daily)

    12/10/2022, 3:24 PM
    bro i continued sza's album photo it looks so cool
  • j

    Justin

    12/10/2022, 3:34 PM
    Is the ship almost about to sink or what?
  • a

    Amizade | Pony's voice creator

    12/10/2022, 3:36 PM
    Is there a site that transcribes a voice?
  • j

    Justin

    12/10/2022, 3:36 PM
    #841437191073955920
  • j

    Justin

    12/10/2022, 3:37 PM
    It's a colab nb tho
  • a

    Amizade | Pony's voice creator

    12/10/2022, 3:37 PM
    ok, thx
  • a

    Amizade | Pony's voice creator

    12/10/2022, 5:10 PM
    Copy code
    /content/wavs/1.wav Does not have 22050 KHz!
    /content/wavs/1.wav is not mono!
  • a

    Amizade | Pony's voice creator

    12/10/2022, 5:10 PM
    The wavsample is not changing to 22khz
  • g

    Gosmokeless28

    12/10/2022, 11:00 PM
    Why 400 epochs?
  • a

    Amizade | Pony's voice creator

    12/10/2022, 11:24 PM
    Because I tought the model will be training good
  • g

    Gosmokeless28

    12/10/2022, 11:31 PM
    Lol, no Haven't you heard of overfitting?
  • a

    Amizade | Pony's voice creator

    12/11/2022, 12:10 AM
    nope. I'm still learning to create voice.
  • g

    Gosmokeless28

    12/11/2022, 12:13 AM
    To put it simply, overfitting is the action of a model learning too much of its dataset in the course of it being trained, which ironically causes it to be unable to generate good results.
  • a

    Amizade | Pony's voice creator

    12/11/2022, 12:57 AM
  • a

    Amizade | Pony's voice creator

    12/11/2022, 2:12 PM
    Something is wrong...
  • a

    Amizade | Pony's voice creator

    12/11/2022, 2:18 PM
    is it normal when start training?
1...101810191020...1068Latest