German pretrained model
# tacotron-2-supportu
{K EY1} (Kei)
09/29/2022, 5:30 PMGerman pretrained model
{K EY1} (Kei)
09/29/2022, 5:30 PM@hecko
So pikachu told me to ask you about this because you're knowledgeable
But i started training a single speaker german pretrained model on ipa, and its synthesis output has just been static no matter what notebook or pretrained model i train it on
I can send an example of the static if needed, and the dataset if needed
I accordingly modified the cleaners and symbol set, and this did not happen with the ipa french base model i trained with the same method
h
hecko
09/29/2022, 5:30 PMhmhm
hecko
09/29/2022, 5:30 PMi haven't touched ipa ever so if it's related to that then idk
hecko
09/29/2022, 5:31 PMcould you send a copy of the notebook you used with all the settings in it
u
{K EY1} (Kei)
09/29/2022, 5:31 PMYeah i'll do that
{K EY1} (Kei)
09/29/2022, 5:31 PMWhen i get home
{K EY1} (Kei)
09/29/2022, 5:31 PMI tried my notebook, as well as the notebook in #841437191073955920
{K EY1} (Kei)
09/29/2022, 5:32 PMLegacy btw
h
hecko
09/29/2022, 5:33 PMoh legacy huh
hecko
09/29/2022, 5:33 PMdefinitely haven't tried ipa on legacy
hecko
09/29/2022, 5:33 PMoh and also send me a few of the wav·s
u
{K EY1} (Kei)
09/29/2022, 5:34 PMI'll send the whole thing if you want, or just 3 ish?
{K EY1} (Kei)
09/29/2022, 5:35 PMAlso i'm going to try training with a different dataset to see if the same thing happens
h
hecko
09/29/2022, 5:37 PMyeah just a few, i don't wanna burn through my data cap
hecko
09/29/2022, 5:37 PMjust wanna make sure they're formatted right because it's really easy to forget one of the settings
u
{K EY1} (Kei)
09/29/2022, 6:11 PMYeah i checked but maybe i'm forgetting something weird
{K EY1} (Kei)
10/07/2022, 5:57 PMsorry for the delay
h
hecko
10/07/2022, 5:58 PMhmm all seems fine
u
{K EY1} (Kei)
10/07/2022, 5:58 PMyeah thats what i thought
h
hecko
10/07/2022, 5:58 PMthen it's probably an issue with the transcripts
u
{K EY1} (Kei)
10/07/2022, 5:58 PMit outputs static like this
h
hecko
10/07/2022, 5:58 PMor the settings
u
{K EY1} (Kei)
10/07/2022, 5:59 PMthis is the training transcript
{K EY1} (Kei)
10/07/2022, 5:59 PMi'll make a copy of the notebook with the settings i used
{K EY1} (Kei)
10/07/2022, 6:01 PM{K EY1} (Kei)
10/07/2022, 6:03 PMah my message about the symbol set got deleted as spam
{K EY1} (Kei)
10/07/2022, 6:04 PMhere we go
this was under 'letters'
{K EY1} (Kei)
10/07/2022, 6:06 PMcleaners were set to a new definition called "return_text"
Copy code
def return_text(text):
return texth
hecko
10/07/2022, 6:18 PMnot sure if this is the issue but it's weird that the min_learning_rate is higher than learning_rate
hecko
10/07/2022, 6:21 PM-oh maybe
hparams.ignore_layershecko
10/07/2022, 6:21 PMnot sure what to set it to since the legacy layout might be different
hecko
10/07/2022, 6:21 PMin pipeline it's
["embedding.weight"]u
{K EY1} (Kei)
10/07/2022, 6:26 PMembedding.weight was being ignored i think
I'll redo training and make sure of that
{K EY1} (Kei)
10/07/2022, 6:26 PMWait it is?
{K EY1} (Kei)
10/07/2022, 6:27 PMOh
I put it in backwards i think
{K EY1} (Kei)
10/07/2022, 6:28 PMI think i meant for the values to be switched around