# tacotron-2-support
- t
tylerdurdenceketi
11/26/2022, 9:20 AMAdd a new code block and type this: !pip install --upgrade gdown - l
lil beby
11/26/2022, 10:03 AMGuys, how to get in? - h
hecko
11/26/2022, 12:01 PMdid you run the previous steps - h
hecko
11/26/2022, 12:02 PMare you using a school computer or school google account? if so then you can't - h
hecko
11/26/2022, 12:03 PMyou'd have to use your own device with your own personal account - h
hecko
11/26/2022, 12:03 PM(or take it up with the computer person in your school but i doubt it'll work) - u
9 x x
11/26/2022, 5:03 PMyeah make sure you are not on a work account - u
9 x x
11/26/2022, 5:03 PMalso - u
9 x x
11/26/2022, 5:03 PM#841437191073955920 - u
9 x x
11/26/2022, 5:04 PMmake sure the notebook is there - u
9 x x
11/26/2022, 5:04 PMand when you make a copy you are not switching accounts - s
Sonic2022_mario
11/26/2022, 6:57 PMI Just Uploaded Sonic (Roger Craig Smith, Frontiers) And Tested It And Sounds Like This - d
Daft
11/26/2022, 11:12 PMhow come my model is pronouncing certain words wrong - l
Lexi (delulu posts on the daily)
11/27/2022, 12:42 AMbecause dataset could be silly or u could use arpabet - m
mynameisNegan
11/27/2022, 1:51 AM@hecko What's the best epoch for AITCH using Uberduck Tacotron 2 Pipeline? I have over 547 wav files total of 32:19 mins. - h
hecko
11/27/2022, 1:52 AMthere's no hard number for things like that - h
hecko
11/27/2022, 1:52 AMjust train it for some time and test a few of the checkpoints in the synthesis notebook - m
mynameisNegan
11/27/2022, 1:52 AMOkay. - t
tylerdurdenceketi
11/27/2022, 11:31 AMI have preprocessed files with preprocess_audio.py. Graphs were good. But speeches cut off early. So i ran preprocess files with preprocess_audio.py again with 300ms padding. But mel spectogram shows a black region at the end. I have read something about adding noise to audio tracks in github issues section of the tacotron2 repo. What should I do? - u
(Dawn) Will Draw Fictional Women
11/27/2022, 11:32 AM>But speeches cut off early. - u
(Dawn) Will Draw Fictional Women
11/27/2022, 11:32 AMyou overtrained - u
(Dawn) Will Draw Fictional Women
11/27/2022, 11:32 AMeven then a few tries would probably yield the full prompt - t
tylerdurdenceketi
11/27/2022, 11:33 AMWell loss ratio says otherwise. - t
tylerdurdenceketi
11/27/2022, 11:33 AMIt was 0.30 - u
(Dawn) Will Draw Fictional Women
11/27/2022, 11:33 AMif you get it TOO low it could be overtrained - u
(Dawn) Will Draw Fictional Women
11/27/2022, 11:34 AMif you would like tips on a third attempt i suggest multiple sentences in a single wav file - u
(Dawn) Will Draw Fictional Women
11/27/2022, 11:34 AMmake sure they dont exceed 12 seconds tho - t
tylerdurdenceketi
11/27/2022, 11:35 AMWell I have 10000 wav files. - u
(Dawn) Will Draw Fictional Women
11/27/2022, 11:35 AM@hecko has a wav merger cell that should alleviate all the labor - t
tylerdurdenceketi
11/27/2022, 11:36 AMI set batch size 32 for 20 epoch. It stalled at 0.30 Then I set it to 128 Then it triggered lr_decay Learning rate got smaller and loss started to decrease