for a singing dataset is it better to leave the breaths in or cut them out

hm maybe annotate the breaths as hyphens

-no wait that won't work with talknet because the training duration extraction wouldn't know about that

use comas

for me pause between sentences == breathing

hmm maybe

i just don't know whether i want to burden the network with having to learn how to breathe

hold on here's my main concern

do i - try to clip out the vocals in between the breaths - have it as one big thing and also bo back and re-clip the previous 7 songs to have breaths in them - ignore this part entirely and hope there's enough data elsewhere (current)

is this overtrained?

not overtrained but probnably something with your dataset

you should have cut out some samples where he makes a funny noise (see the first clip)

Left it in for the intro

was just asking because i was training for a while, and concerned about this

it doesn't really have trouble with speaking other things now though

i think i'll stop training it now since it already matches the files almost perfectly if you put in the same words

before i break it haha

what's with this jarring high frequency noise at the end btw? happens a fair bit

i think its the clipping that the ai tries to replicate

is this using a fine tuned hifi gan

no idea if anomalys mic is just Bad and clips sometimes

Also yes, the model will learn how to create noise which is a big no no

i think it clipped maybe once or twice

in the entire dataset

oh well

doesn't seem to clip much on the other notebook

the one that's not hifi

i think it was just a glitch of some kind, this time it almost doesn't clip, i just reloaded the notebook

thanks everyone for the help!

Dylan Dalmatian trained so far