# machine-learning
- u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:45 PMrobotic sound is either caused by noise learned in the dataset or simply the hifigan being mean - u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:46 PMbad pronunciation is a fickle thing, everyone seems to have conflicting solutions for it - u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:46 PMsome say larger batch, others say smaller - u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:46 PMme personally i train with a step 3 replacement with autocalculates learning rate based on batch size - u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:47 PMi only occasionally get pronunciation weirdness with some models - u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:47 PMarpabet completely alleviates this in most cases - u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:47 PMlong sentences causing the model to break down is either caused by bad alignment or low val loss - u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:48 PMif it still happens with the model at 0.05 val loss then its probably bad alignment - u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:48 PMmost that can be done there is a revision of the transcript and dataset to make sure everything's completely accurate - k
Komori-uta Kitsune
03/13/2022, 6:48 PMsent dm - m
mega b
03/13/2022, 6:49 PMwhen you open the exe, does it show all the options n such? - m
mega b
03/13/2022, 6:50 PMsomething like this - j
JamesVincentWalsh
03/13/2022, 6:55 PMSo what I'm getting so far is to try arpabet- sounds like there's nothing to lose there-and revise the data. I do have a lot of sentences broken up into pieces- would it be better to put some of them back together so it has more complete sentences to learn from? Not sure there's anything else I can do to fix the data, unless adding in more would be worth trying - u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:56 PMyeah, you can just append the arpa transcript alongside the ljspeech one for a hybrid, those work fine and you wont have to throw away the existing one - u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:56 PMas for the split sentences, i dont really know - u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:57 PMdont think it would make too much of a difference outside of maybe inflictions - u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:58 PMand i would try isolating the voice from clothe rustling with something like audo.ai - u
(Dawn) Will Draw Fictional Women
03/13/2022, 6:58 PMif its too intense it might be interfering with alignment - j
JamesVincentWalsh
03/13/2022, 7:01 PMGotcha. Thanks so much! I think I have a plan now. - k
Komori-uta Kitsune
03/13/2022, 7:03 PMsorry mega - k
Komori-uta Kitsune
03/13/2022, 7:03 PMwas in dm - m
mega b
03/13/2022, 7:03 PMokay could you download this waveglow model? - m
mega b
03/13/2022, 7:03 PM - k
Komori-uta Kitsune
03/13/2022, 7:04 PMgetting it - k
Komori-uta Kitsune
03/13/2022, 7:04 PMsec - m
mega b
03/13/2022, 7:05 PMand also download this: https://github.com/johnpaulbin/tacotron2/releases/download/Main/tacotron2_statedict.pt - m
mega b
03/13/2022, 7:06 PMit is the basic tacotron2 model - k
Komori-uta Kitsune
03/13/2022, 7:10 PMok - m
mega b
03/13/2022, 7:12 PMokay now rename the waveglow model to have
at the end of the file name.pt