# machine-learning
- h
hecko
03/20/2022, 11:58 PMyou probably mass-exported from audacity which adds the label names as metadata it's not an error just a warning, nothing to worry about - w
Waltman13 (semi-active)
03/21/2022, 2:56 AMTried training Tay Zonday with an arpabet-converted transcript. While he can speak arpabet now, he still strokes.. starting to think that the dataset or default settings I'm using are shit - u
(Dawn) Will Draw Fictional Women
03/21/2022, 2:57 AMi train all my models with a modified step 3 - u
(Dawn) Will Draw Fictional Women
03/21/2022, 2:57 AMand my suzanne vega singing model works pretty good - u
(Dawn) Will Draw Fictional Women
03/21/2022, 2:57 AMwell actually i didnt test it with the stress test - u
(Dawn) Will Draw Fictional Women
03/21/2022, 2:58 AMand when i was making a video with it it stroked out twice… - u
(Dawn) Will Draw Fictional Women
03/21/2022, 2:58 AMforget i said anything - g
Gosmokeless28
03/21/2022, 3:15 AMTry training the model for 200 epochs with the batch size of 14. - g
gremlin.pk
03/21/2022, 12:04 PM@User to create dataset..... can some clips with only 1 second ? or < 1 second ? like a word speak: "hello" my dataset have some clips with <= 1 second - h
hecko
03/21/2022, 12:13 PMit's not ideal but it'll work though if you can merge them without making it sound unnatural i highly recommend doing that - p
prism
03/21/2022, 2:36 PMSo I was wondering how the new ARPAbet taco Tron 2 training book works and if I need to rewrite my transcript into ARPAbet - h
hecko
03/21/2022, 3:20 PMsame as the usual one, and it does that for you - p
prism
03/21/2022, 3:21 PM@hecko I had some non standard English words is there anyway to input them in ARPAbet - g
Gosmokeless28
03/21/2022, 6:50 PMJust use the notebook without having your dataset be altered. - p
prism
03/21/2022, 8:16 PM@Gosmokeless28 is the format. 1.wav|{L AY K} {TH IH S}. And does this also work with the talk net trainer - g
Gosmokeless28
03/21/2022, 8:17 PMwavs/1.wav|A sentence. - p
prism
03/21/2022, 8:19 PMOh then how do you make it know it's in ARPAbet - u
{K EY1} (Kei)
03/21/2022, 8:19 PMThe arpabet for "this" is {dh ih s} btw - p
prism
03/21/2022, 8:23 PMOkay sorry I just been making a bank of the Pokemon announcer from battle revolution and it's been having a hard time with figuring out pokemon names and thought it would be best just to do it manually - h
Hugh G. Rection
03/21/2022, 10:43 PMI tried training the TalkNet version of Rico (the funni yes rico KABOOM penguin), and... WTF?? It works? - a
Alphabet
03/22/2022, 11:23 PMalright, so I got my hands on the attenborough dataset. 5000 wav files and about 7+ hours of audio - a
Alphabet
03/22/2022, 11:24 PMran it through talknet. 200 = 200 epochs and 2000 steps. 500 = 500 epochs and 4000 steps. For 200 epochs, val loss reached .362 and for 500 epochs, val loss reached .359 - a
Alphabet
03/22/2022, 11:26 PMoverral, it sounds amazing. After all this testing with 5+ datasets, there's very little reason to train past 200 epochs. You're not going to detect too much difference after 200 epochs for talknet - g
Gosmokeless28
03/22/2022, 11:26 PMWhich TalkNet model trainer notebook did you use? - g
Gosmokeless28
03/22/2022, 11:26 PMJust curious - a
Alphabet
03/22/2022, 11:27 PMthe default one with 1e3/3e6 learning. https://colab.research.google.com/drive/1Nb8TWjUBJIVg7QtIazMl64PAY4-QznzI?usp=sharing - a
Alphabet
03/22/2022, 11:28 PMthe one at the top of the list. Technically, the backup is exactly the same as the default but has different learning rates i think - g
Gosmokeless28
03/22/2022, 11:29 PMYeah. Do you think those alternative learning rates do a better job than the default ones, in your opinion? - a
Alphabet
03/22/2022, 11:30 PMhonestly? makes very little difference