# machine-learning
- u
(Dawn) Will Draw Fictional Women
02/15/2022, 7:53 PMuh - u
(Dawn) Will Draw Fictional Women
02/15/2022, 7:53 PMare we talking about the same thing here - m
mega b
02/15/2022, 7:54 PMthe model shouldn't be affected if it misses an arpabet conversation, since it is already massively full of data - u
(Dawn) Will Draw Fictional Women
02/15/2022, 7:54 PMso - u
(Dawn) Will Draw Fictional Women
02/15/2022, 7:54 PMit just doesnt bother converting it? - m
mega b
02/15/2022, 7:54 PMyea - u
(Dawn) Will Draw Fictional Women
02/15/2022, 7:54 PMoh - u
(Dawn) Will Draw Fictional Women
02/15/2022, 7:54 PMwell would it still be beneficial to convert them manually? - m
mega b
02/15/2022, 7:55 PMnot exactly, since we can already improve on an existing dictionary - u
(Dawn) Will Draw Fictional Women
02/15/2022, 7:55 PMweird! - u
(Dawn) Will Draw Fictional Women
02/15/2022, 7:56 PMthis whole thing is weird i feel weird - m
matty1406
02/15/2022, 7:58 PMdo you think arpabet will help pronouncing words for models with low learning rate? since most of the time models trained with a low learning rate has worse pronunciation than normal learning rate. - u
(Dawn) Will Draw Fictional Women
02/15/2022, 7:58 PMreally?? - m
matty1406
02/15/2022, 7:58 PMfrom tests, i believe so... - u
(Dawn) Will Draw Fictional Women
02/15/2022, 7:58 PMall this time buster's been telling me the only reason he does lower learning rate is because it IMPROVES pronunciation - m
matty1406
02/15/2022, 7:59 PMlearning rate helps with making it actually talk and function... - m
matty1406
02/15/2022, 7:59 PMthats why my models sound actually decent with little data - u
(Dawn) Will Draw Fictional Women
02/15/2022, 7:59 PMgaaah why is everything turning so weeeird!! - s
samkoelle
02/15/2022, 8:03 PMthe arpabet pipeline uses G2P which guesses the arpabet if its not in the dictionary - u
(Dawn) Will Draw Fictional Women
02/15/2022, 8:03 PMah i see - u
(Dawn) Will Draw Fictional Women
02/15/2022, 8:03 PMlike the script i was using - u
(Dawn) Will Draw Fictional Women
02/15/2022, 8:04 PMbut is it more beneficial to have a 100% accurate full arpa transcript - u
(Dawn) Will Draw Fictional Women
02/15/2022, 8:04 PMor to have a partially converted transcript with all foreign words still in lj - s
samkoelle
02/15/2022, 8:05 PMwe only ever read the pipeline without converting at all prior to calling train, so im not really sure. i just wouldnt convert if using the pipeline - s
samkoelle
02/15/2022, 8:06 PMwe did make an app that allowed us to update the dictionary to include words that werent in there before - s
samkoelle
02/15/2022, 8:06 PMwhich theoretically we could release (although we'd have to figure out how to host it) - s
samkoelle
02/15/2022, 8:06 PMif there was interest - u
(Dawn) Will Draw Fictional Women
02/15/2022, 8:06 PMincheresting - u
(Dawn) Will Draw Fictional Women
02/15/2022, 8:06 PMstill a little lost tho - s
samkoelle
02/15/2022, 8:08 PMhow so?