# machine-learning
- l
LexiB
09/07/2021, 2:56 AMUh oh - t
The Wiggles and High School D×D
09/07/2021, 3:01 AMPardon? Sorry, I didn't really understand. Can I use my own metadata file or can I just download one from another model to use for the synthesis process? - t
The Wiggles and High School D×D
09/07/2021, 3:02 AMWould it help if I linked the wiki despite it not telling me anything on this? - t
The Wiggles and High School D×D
09/07/2021, 3:02 AMhttps://github.com/vlomme/Multi-Tacotron-Voice-Cloning/wiki/Training-(and-for-other-languages - u
user
09/07/2021, 3:03 AMThis part is what you copy - j
JacobLenstar
09/07/2021, 4:15 AMbefore I make new talknet voices, I'm gonna train all the ones I've made that sound crap to not... sound crap - t
The Wiggles and High School D×D
09/07/2021, 11:51 AMA warning before I post the link (whenever that will be): I may have an overly low amount of audio I used for the encoder and synthesiser so let me know if I need to add more - t
The Wiggles and High School D×D
09/07/2021, 11:52 AMWon't push it past 15GB for anyone's sake, though; I think it's 10GB right now - j
JacobLenstar
09/07/2021, 1:01 PMI let sheen train all night and he still sounds kinda crap :( - j
JacobLenstar
09/07/2021, 1:01 PMwell - j
JacobLenstar
09/07/2021, 1:02 PMall night might be an exaggeration since my computer went into sleep mode about two hours after I went to sleep lol - j
JacobLenstar
09/07/2021, 1:02 PMhe made it to 2910 epoches and 2910 steps - m
MoarMosspato
09/07/2021, 4:34 PMNO NOT THE K80 - m
MoarMosspato
09/07/2021, 4:40 PMI keep factory reseting and I'm still getting K80s - h
haru0l
09/07/2021, 4:43 PMyou can still use it i believe - h
haru0l
09/07/2021, 4:43 PMjust lower the batch size and wait longer - j
JacobLenstar
09/07/2021, 5:04 PMI retrained Sheen Estevez and Jimbo Kern on talknet, but they still don’t sound great - j
JacobLenstar
09/07/2021, 5:05 PMshould I be training for longer than 2000 steps? - j
JacobLenstar
09/07/2021, 5:05 PMSandy sounded great and I trained her for 4000+ - h
haru0l
09/07/2021, 5:39 PMwhy not dkjhsdk - h
haru0l
09/07/2021, 5:40 PMi remembered megamind sounding bad at 2000 ish and sounded good when it disconnected me at 9900 steps - h
haru0l
09/07/2021, 5:40 PMit overfitted though 💀 - h
haru0l
09/07/2021, 6:08 PMinteresting, my cupcakke model uses the "calmer" speaking sentences if i use really short sentences but uses the "singing" ones if it's long - c
CerealYe
09/07/2021, 6:45 PMI found something interesting. https://www.sony.com/en/SonyInfo/technology/stories/AI_Sound_Separation/ - j
JacobLenstar
09/07/2021, 7:36 PMI’m getting the out of memory error a lot from talknet - b
BIA
09/07/2021, 7:54 PMI wonder if its possible to load models into ram instead of vram - h
hecko
09/07/2021, 8:02 PMfor synthesis yes, in fact the new synthesis notebook in #841437191073955920 uses cpu only for training that would slow things down extremely, not recommended - b
BIA
09/07/2021, 8:03 PMI might be stupid, but I im more thinking still using CUDA/GPU for rendering but loading the stuff into ram vs vram - h
hecko
09/07/2021, 8:04 PMa - h
hecko
09/07/2021, 8:04 PMyeah it'd still be too slow probably