# machine-learning
- u
{K EY1} (Kei)
03/24/2022, 9:15 PMAnd it's already past 200 epochs... - h
HolyArapaima
03/24/2022, 9:15 PMpain - m
matty1406
03/24/2022, 9:15 PMhow much minutes of data is that? musn't be a lot of if it's only one batch - h
HolyArapaima
03/24/2022, 9:16 PMit isn't a lot, I am pretty limited with what I have to work with - u
{K EY1} (Kei)
03/24/2022, 9:16 PMWhat voice is it? - h
HolyArapaima
03/24/2022, 9:16 PMI have gotten some okay results with talknet though - m
matty1406
03/24/2022, 9:16 PMhow many wavs? - h
HolyArapaima
03/24/2022, 9:17 PM48 lol - m
matty1406
03/24/2022, 9:17 PMeh, thats alright i guess, it's possible - u
{K EY1} (Kei)
03/24/2022, 9:17 PMOh jeez that's even less than my crappy iroha eng tacotron model - m
mega b
03/24/2022, 9:17 PMit might be in the dataset - m
mega b
03/24/2022, 9:17 PMcan you send transcript file? - h
HolyArapaima
03/24/2022, 9:20 PMignore the wav names been cutting out some of the more garbage samples - h
HolyArapaima
03/24/2022, 9:23 PMI might need to clean up the hifigan or something - h
HolyArapaima
03/24/2022, 9:26 PM^ that one is trained with talknet I thought I would try to train a tacotron since it's been a while since I last trained one, thought maybe I could use the hifigan or something just to push some of that engine noise out - u
{K EY1} (Kei)
03/24/2022, 9:59 PMHm, some of them aren't in english, so that could pose a problem However, i don't see a problem with the rest I'm surprised the graph looks like that - h
HolyArapaima
03/24/2022, 10:04 PMyeah, I tried to translate it to arabet as best I could - h
HolyArapaima
03/24/2022, 10:08 PMis there a hifigan tuner for talknet? - u
(Dawn) Will Draw Fictional Women
03/24/2022, 10:08 PMthats what the hifigan step is for - h
HolyArapaima
03/24/2022, 10:09 PMwhat do you mean? - u
(Dawn) Will Draw Fictional Women
03/24/2022, 10:09 PMthe hifigan step in talknet training is literally finetuning - h
HolyArapaima
03/24/2022, 10:10 PMso I just need to train it more maybe? change some values and stuff - u
(Dawn) Will Draw Fictional Women
03/24/2022, 10:10 PMyeah best bet is to train it longer - h
HolyArapaima
03/24/2022, 10:13 PMfor step 9 of talknet for the spectrogram generator I have been training it to about 600 how far should I push this lol - u
(Dawn) Will Draw Fictional Women
03/24/2022, 10:14 PMeither until the step stops working or the gpu privs get pulled - u
(Dawn) Will Draw Fictional Women
03/24/2022, 10:14 PMreturns might be diminishing tho - h
HolyArapaima
03/24/2022, 10:17 PMI have just never been able to get my tacotron to sound decent at least for what I have been doing. - h
HolyArapaima
03/24/2022, 10:18 PMtalknet I have just been getting a bit more engine noise though it's been a long time since I last used tacotron2