or well not all all it's better to have 800 calm samples than 800 calm + 100 angery

but as many roughly-samey clips as you have the patience to get

What is this supposed to mean? [WARNING] wavs/47x.wav in filelist while expecting .npy .

yeah im doing peko rn and ive already gotten and transcribed over 100 samples

i might try to get more but i dont have much more patience -_-

100 wont be that good, it will work but it would be better if you got more

i used the vocal remover and wow, its actually pretty good

there's actually an official vocal mate

hey there - any recommends on good hifiGAN characters? i had a pretty crappy dataset of about 80 samples and i seem to be getting mostly gibberish. using Rarity(fine tuned). just wondering if there's big differences in behavior.

im gonna try to get near 200 samples then ill try

you might wanna uncheck the

dictionary

checkbox if you have it checked, that converts the text into phonemes and we don't want that because uberduck tacotron models are trained directly on text

you skipped the step where it turned .wav audio into .npy numbers you have to play all the steps in sequence (except maybe the google drive ones)

yeah it was unchecked. made no difference. i'm getting Reached Max Decoder Steps a lot. is that an indication of bad data?

definitely something bad

could you send me the google drive code so i can check how bad it is

sure its: 1-B8_4fZJfUXCPwOJ5wNQC5Vktrwr3beH

Well, I did every step, it even creates npy

ok yeah wow it's very bad

186 voice samples danganronpa 2 spoilers // ||i mostly avoided the sparkling justice lines because they were mostly of her shouting||

what does your data look like

like is it made of short clips because that'd be bad

define short - how long should each clip be range wise?

a few seconds

a sentence or two

but based on this it seems that that's not the issue

i've heard that with less than 100 samples one should limit oneself to ~100 epochs so it doesn't overfit

well the majority are around 2-3 secs i think. there's at least one sentence. i'm trying for galatea's voice from JLU and found the actress did a sitcom. so most of those clips are decent. but the galatea ones were really difficult and had bg music almost always to filter out.

okay now that might be a possibility - i did 500 epoch but stopped it short of 250 when it hit 0.14 or so

ooooof yeah that's too much

allright - so need to retrain for 80 files - what epoch would be good, and will that bring the error range down quicker? because at 80 it was still pretty high when i had 500 epoch.