# machine-learning
- u
user
02/07/2022, 5:30 PMoh yeah - u
user
02/07/2022, 5:30 PMtalknet is pretty unforgiving unfortunately - u
(Dawn) Will Draw Fictional Women
02/07/2022, 5:43 PMdont i know it - u
user
02/07/2022, 6:57 PMi use mvsep to separate vocals - j
JASS_AGI
02/07/2022, 6:57 PMdo you also run it if it doesn't have any bg music? - u
user
02/07/2022, 6:58 PMno but i think it would work for other noises too - j
JASS_AGI
02/07/2022, 6:58 PMThanks! i'll try it out - c
Couch
02/07/2022, 6:58 PMiZotope RX 8 - j
J CrashBandicoot F
02/07/2022, 7:55 PMAny tips on how to train with datasets that has less than 20 wav files? - m
mega b
02/07/2022, 7:56 PMYou could try warm starting from Ibob's chills model - i
IBob012
02/07/2022, 7:56 PMyeah 😎 - i
IBob012
02/07/2022, 7:56 PMi trained a model once with 14 wav files cuz i warm started over chills - c
Couch
02/07/2022, 8:01 PMWhere enlace - j
J CrashBandicoot F
02/07/2022, 8:02 PMI have here a dataset that has exactly 14 wav files, how can I do that? - u
(Dawn) Will Draw Fictional Women
02/07/2022, 8:03 PMoh? - u
(Dawn) Will Draw Fictional Women
02/07/2022, 8:03 PMtry training 14 wavs from just plain old lj - u
(Dawn) Will Draw Fictional Women
02/07/2022, 8:03 PMtwice - u
(Dawn) Will Draw Fictional Women
02/07/2022, 8:03 PM😎 - i
IBob012
02/07/2022, 8:04 PMdownload this and put it in your drive https://drive.google.com/file/d/1J17rdjUXhORzQHNaOqE3sjndVceg_dHO/view?usp=sharing - u
(Dawn) Will Draw Fictional Women
02/07/2022, 8:05 PMok heres the strategy that i found works best for stretching mileage on low wav datasets - u
(Dawn) Will Draw Fictional Women
02/07/2022, 8:05 PMabsolutely use my step 3 replacement (calculate learning rate based on batch size - u
(Dawn) Will Draw Fictional Women
02/07/2022, 8:05 PMsince you have 14 wavs go for a batch size of 7 - u
(Dawn) Will Draw Fictional Women
02/07/2022, 8:05 PMmake the transcript a hybrid of normal text and arpabet - u
(Dawn) Will Draw Fictional Women
02/07/2022, 8:06 PMonce its done training synthesize with and without the pronunciation dictionary - u
(Dawn) Will Draw Fictional Women
02/07/2022, 8:07 PMsee which works best, and if it works better with arpa ask the mods to enable it after submitting - u
(Dawn) Will Draw Fictional Women
02/07/2022, 8:07 PMyou can try warm starting from chills but ive never needed to do so personally - u
(Dawn) Will Draw Fictional Women
02/07/2022, 8:07 PMplus i find it to be a bit more of a (frank) hassle - i
IBob012
02/07/2022, 8:07 PMits the easiest way to go around doing it cuz he prolly dont know how to do any of those things