# tacotron-2-support
- j
Justin
11/21/2022, 4:22 AM - t
TheWu-Cepticon
11/21/2022, 4:23 AMthank you👍 - j
Justin
11/21/2022, 4:23 AMnp - a
AhmadGT
11/21/2022, 9:09 AMOkay Thanks! - f
Floofy
11/21/2022, 7:33 PMFollowing the tutorial ^^ still curious on which would be easier to use, legacy or pipeline? - a
AhmadGT
11/21/2022, 8:22 PMi don't find either of them hard - a
AhmadGT
11/21/2022, 8:23 PMjust follow the instructions and you will be good to go - a
AhmadGT
11/21/2022, 8:23 PMalso use the pipeline its better - f
Floofy
11/21/2022, 8:34 PMtysm! - d
Deleted User
11/21/2022, 11:51 PMif someone sends me a model link (the long tacotron ID) , can I download it and save it? - g
Gosmokeless28
11/21/2022, 11:55 PMYeah - d
Deleted User
11/21/2022, 11:56 PMhow do I do that - g
Gosmokeless28
11/21/2022, 11:56 PMDo you have a Google Drive link to a Tacotron 2 model? - d
Deleted User
11/21/2022, 11:56 PMyeah - g
Gosmokeless28
11/21/2022, 11:57 PMPress the ⬇️ download button in the top right corner - r
realyp19
11/21/2022, 11:58 PMok cool, so now even if the person who sent it deletes their model it'll still be useable? - d
Deleted User
11/22/2022, 2:30 AMthank you - g
Gosmokeless28
11/22/2022, 2:30 AMNo problem - t
tylerdurdenceketi
11/22/2022, 6:54 PMhello what is the difference between arpabet 1.0 vs 0.0 i am training a non english model - r
Reclezon
11/22/2022, 7:02 PMArpabet is just a text format for American English pronunciation. Other languages don't really need it it also converts the regular tezt (like this message) into arpz - t
tylerdurdenceketi
11/22/2022, 7:40 PMthanks do you have any advice for my training? i have created a dataset from mozilla common voice restored wav files with voice fixer used preprocess_audio.py loss is staying at 0.30 what can i do? - t
tylerdurdenceketi
11/22/2022, 7:40 PMi am using ipa transcription btw - r
Reclezon
11/22/2022, 8:00 PMI haven't made any models that use IPA so far, but the loss is normal, if it doesn't change after however many epochs you feel, start with just checking the settings you're training with, or if you're using a large dataset, make sure the transcript is right, although larger datasets do take longer to train - r
Reclezon
11/22/2022, 8:22 PMThe reason * say check the transcript is for any misspellings that could cause a model to become confused, I've had that happen before. - t
tylerdurdenceketi
11/22/2022, 8:40 PMI have 10k wav files in the dataset I can't check that shit i think I have tried batch size of 8, 32, 64 and 128 Is it better to have smaller batch size for better generalization or larger batch for faster training? - m
matty1406
11/22/2022, 8:42 PMbatch size should be small, and it's better to make sure it's a number that can be divided by the amount of wavs to make sure it goes through them all... unless of course the amount is a prime number... - t
tylerdurdenceketi
11/22/2022, 8:44 PMThanks Before i turned off sample inference and validation, sample inference graph was good. Diagonal and yellow pixels in it Then after one epoch it leaned left I have always paid attention to sample inference graph. Is it wrong? - r
Reclezon
11/22/2022, 11:06 PMIt's fine that if one epoch looks like a mess, but as long as the average like a yellow diagonal line then the model will be good - r
Reclezon
11/22/2022, 11:07 PMSole inference graph shouldn't be wrong, but also shouldn't be the only graph you look at, either. - r
Reclezon
11/22/2022, 11:10 PMSome models just do better or worse with some sentences