# machine-learning
- d
Draconiator
05/26/2022, 9:52 PMI'd make one of Elliot Rodger so people can make memes with it, but that guy is too disgusting to even record audio from. - r
Reclezon
05/26/2022, 11:51 PMOops... - r
Reclezon
05/26/2022, 11:52 PMtbh the first few epoch are always the most fun for gettung stuff like this. - h
hecko
05/27/2022, 12:02 AMoh are you training a new base model - r
Reclezon
05/27/2022, 12:03 AMUnfortunatly, no. Still redoing Imogen Heap. Felt like messing with the Learning rate and set it to 5e-2 - h
hecko
05/27/2022, 12:03 AMah - r
Reclezon
05/27/2022, 12:04 AMI could train from scratch, but Idk how to do that - u
(Dawn) Will Draw Fictional Women
05/27/2022, 12:05 AMif you mean without a base model - u
(Dawn) Will Draw Fictional Women
05/27/2022, 12:05 AMdont - r
Reclezon
05/27/2022, 12:05 AMConsidering there's a lot of voices of Uberduck that are masculine. I could train a male one for people to use if that'd make it easier since the default is LJSpeech - u
(Dawn) Will Draw Fictional Women
05/27/2022, 12:06 AMit works fine enough - r
Reclezon
05/27/2022, 12:06 AMtrue - r
Reclezon
05/27/2022, 12:06 AMI'm guessing it'd just take too long? - r
Reclezon
05/27/2022, 12:06 AMor is it something else - u
(Dawn) Will Draw Fictional Women
05/27/2022, 12:07 AMyeah - u
(Dawn) Will Draw Fictional Women
05/27/2022, 12:08 AMitd take quite the amount of time - u
(Dawn) Will Draw Fictional Women
05/27/2022, 12:08 AMand not really worth it - u
{K EY1} (Kei)
05/27/2022, 12:26 AMI mean you could train a normal male vb off of ljspeech and then use that as a warm start model - u
(Dawn) Will Draw Fictional Women
05/27/2022, 12:30 AM^ - u
(Dawn) Will Draw Fictional Women
05/27/2022, 12:30 AMchills is good for that - h
hecko
05/27/2022, 12:32 AMthe arpabet notebook uses chills yeah - r
Reclezon
05/27/2022, 1:03 AMoh huh - r
Reclezon
05/27/2022, 1:05 AMalso minor, but am wanting to transcribe this as romaji but idk the accuracy of the text. What I have so far is
. Is it at least somewhat similar to the audio?私 は イメチェン です 初めまして - u
{K EY1} (Kei)
05/27/2022, 1:46 AMThey said
Keep in mind that there are no spaces in japanese. Also, that's not romaji so i'm not sure why you said romaji Romaji would be "Watashi wa Imojen desu. Hajimemashite."私はイモジェンです。はじめまして。 - r
Reclezon
05/27/2022, 2:06 AMThat's the output of the STT I was using formatted to 1 line, I doubt the #841437191073955920 would accept it that way without modification anyway. - u
{K EY1} (Kei)
05/27/2022, 2:07 AMWhat? It would accept it in phonetic input if you used basic_cleaners or convert_to_ascii Ofc it wouldn't accept it with kanji You could do it with kana but that would take like 484884 hours of data - r
Reclezon
05/27/2022, 2:13 AMEh? That seems quite specific - u
{K EY1} (Kei)
05/27/2022, 2:24 AMOh it was a hyperbole; i just spammed my keyboard - r
Reclezon
05/27/2022, 2:25 AMoh - j
jens
05/27/2022, 9:11 AMI assume this wont affect tacotron 2 fakes but this basically killed any face swapping colab that trains for long periods of time (DFL)