# machine-learning
- p
postmates!!
03/19/2022, 1:29 AMwhere? - u
{K EY1} (Kei)
03/19/2022, 1:43 AM#841437191073955920 - u
{K EY1} (Kei)
03/19/2022, 1:43 AMWait - u
{K EY1} (Kei)
03/19/2022, 1:43 AMMaybe not - g
Gosmokeless28
03/19/2022, 1:44 AMOne moment. - g
Gosmokeless28
03/19/2022, 1:45 AMThere. - a
Alphabet
03/19/2022, 7:04 AMso i'm guessing the "ARPAbet enabler version of Uberduck's TT2 model trainer notebook:" will convert the transcription to arpabet first before training? - g
Gosmokeless28
03/19/2022, 7:16 AMYes - m
matty1406
03/19/2022, 2:23 PMtheres something seriously wrong going on lol - m
matty1406
03/19/2022, 2:23 PMepoch 5 and its already below 0.09, and the graph is very bad lol - s
Sweetie Belle
03/19/2022, 2:24 PMhorizontal almost transparent purple lines - m
matty1406
03/19/2022, 2:38 PMah, i think it was because it wasn't signed 16 bit PCM - s
Sweetie Belle
03/19/2022, 2:40 PMit can be. All audios should strongly have 22050 Hz of frequency, 16 bit PCM and 1 channel mono - m
matty1406
03/19/2022, 2:41 PMyeah i know, i dunno why it wasn't defaulting to it though, but its working now - h
hecko
03/19/2022, 2:52 PM,,what the hell uhh send your list.txt maybe - m
matty1406
03/19/2022, 2:58 PMdon't worry, its been fixed, just a small problem with the bit rate - m
matty1406
03/19/2022, 3:40 PMbeautiful stonks!!! - u
{K EY1} (Kei)
03/19/2022, 4:07 PMMy graphs do that when the audio is shit It happened when i was training iroha eng talk which only had single words lol But i see with you it was an encoding error - a
Alphabet
03/19/2022, 7:22 PMSo more testing of talknet. talk 200 = train 200 epochs and 2000 step hifigan. talk 500 = epoch 500 and step 7000. talk 1100 = epoch 1100 and step 6000 with different learning rates (1e-4 instead of 1e-3, min 1e-7 instead of 1e-6). This is with about 580 wavs and 1 to 1.5 hours worth of audio - a
Alphabet
03/19/2022, 7:22 PM6 wav files attached so hopefully everyone can see it - a
Alphabet
03/19/2022, 7:23 PMmy conclusion. Training more than 200 epochs is worth it. There is very little risk of overtraining if you have enough wav files. For smaller dataset, risk for overtraining increases - a
Alphabet
03/19/2022, 7:23 PMat around 400 epochs, there was no more gain in training since it kept saying "no result in top 3" and nothing was saved - a
Alphabet
03/19/2022, 7:24 PMand there's no point in changing learning rate. changing learning rate to try to squeeze in more learning seems to made things worse. However, this is only testing with one dataset though - a
Alphabet
03/19/2022, 7:26 PMtraining eventually bottomed out at around .14. It was .15 at 200 epochs and eventually at .142 by epoch 400. Oh, and this is using Aerith from FF7 remake wav files - a
Alphabet
03/19/2022, 7:30 PMcompared to tacotron though, Aerith from tacotron sounds better. This is with only 500 epochs and final .12 training value for tacotron+arpabet. The only problem is that you can't force AI to talk in a certain way so tacotron has less usage. - a
Alphabet
03/19/2022, 7:31 PMit's a shame you can't combine tacotron results with talknet manipulation somehow - h
hecko
03/19/2022, 8:53 PMidk about you but i (mostly) prefer the 200 smile clip to the others, and the bitch clips all sound identical - h
hecko
03/19/2022, 8:53 PMthere was a thing being adapted called mellotron that would allow that, unfortunately it didn't work out - a
a cat came into my house
03/20/2022, 11:39 PM"/content/tacotron2/utils.py:14: WavFileWarning: Chunk (non-data) not understood, skipping it." - a
a cat came into my house
03/20/2022, 11:39 PMdid i do something wrong or...