# machine-learning
- e
Emoto
03/25/2022, 5:10 PMOr is it multispeaker in a sense that it doesn't blur voices yet consistently - e
Emoto
03/25/2022, 5:11 PMAka this - h
hecko
03/25/2022, 5:48 PMscruffygamer is training a 53-speaker model and pikachu is looking for someone to train a 161-speaker one, both in part for that purpose - e
Emoto
03/25/2022, 5:54 PMBut they're using all different corpora aren't they? - h
hecko
03/25/2022, 6:07 PMyeah they're assembled from whatever character datasets they could find - h
hecko
03/25/2022, 6:08 PMor well i think scruffy just used his own - h
hecko
03/25/2022, 6:08 PMpikachu got some from #835647732453605376, some from me, some from ppp - s
SpamZow (WoomyPearl)
03/25/2022, 6:21 PMNow training Erin Esurance. - s
SpamZow (WoomyPearl)
03/25/2022, 6:21 PM40 wavs this time around on default settings. - s
SpamZow (WoomyPearl)
03/25/2022, 6:22 PMSome of them have her speak long sentences but all of them are under 15 seconds. - s
SpamZow (WoomyPearl)
03/25/2022, 6:22 PMI’ll be leaving my PC on while I go out for a while. - s
SpamZow (WoomyPearl)
03/25/2022, 6:22 PMGotta touch some grass lol - s
samkoelle
03/25/2022, 10:15 PMwhat architecture was this? - h
hecko
03/25/2022, 10:24 PMprobably talknet because the intonation is the same as in previous clips - h
HolyArapaima
03/25/2022, 11:01 PMdoes anyone have any good recommendations for minimum and maximum learning rates? I lowered my learning rate to 2e-4 and 1e-5 and I got some good improvement on my small data set. - c
chicanobatman
03/25/2022, 11:55 PMyeah i hear a difference, i agree it just depends on the size of the dataset. i wish there was a way to like plug in the number of wavs you're using to calculate how many epochs you should train for without overtraining - s
SpamZow (WoomyPearl)
03/26/2022, 1:29 AMErin Esurance model test - s
SpamZow (WoomyPearl)
03/26/2022, 1:30 AMShe can’t pronounce some phonemes. - s
SpamZow (WoomyPearl)
03/26/2022, 1:30 AMI’m currently training to the 250th epoch - s
SpamZow (WoomyPearl)
03/26/2022, 1:30 AMWeird since some of my other models had better results with less WAV files - s
SpamZow (WoomyPearl)
03/26/2022, 1:31 AMAnd with less lines to train with - s
SpamZow (WoomyPearl)
03/26/2022, 1:38 AMEpoch 219 seems promising - u
(Dawn) Will Draw Fictional Women
03/26/2022, 1:39 AM>there’s a giant fucking hole in the graph - s
Sweetie Belle
03/26/2022, 1:40 AMbroken chain - m
mega b
03/26/2022, 1:48 AMI never thought I needed glasses before - h
HolyArapaima
03/26/2022, 1:55 AMsome people hear lee deep, lee dee, lee veem I need some opinions - h
HolyArapaima
03/26/2022, 2:23 AMalso if I want to go lower do I decrease the number after the e? I don't understand how the math works. something with the batch size? - g
Gosmokeless28
03/26/2022, 2:42 AMlee dee - h
HolyArapaima
03/26/2022, 2:43 AMI am also a believer in lee dee