Rate the model
To rate this model, either use /rate or click the buttons below.
#-
cinder ospreyBOT
cinder ospreyBOT
winged delta
didnt collapse with that small batch size?
stark sequoia
didnt collapse with that small batch size?
no
winged delta
no
wow
36 mins of data, batch 4 and no collapse
tho idk if it generalized better than batch size 8
stark sequoia
36 mins of data, batch 4 and no collapse
i trained two 1 hour models with batch size 4
stark sequoia
tho idk if it generalized better than batch size 8
yes
winged delta
it sounds pretty good to have bg noise tho
winged delta
it has bg noise?
the inference audios, not sure about the dataset but probably the dataset may have slight bg noise
tho lyery is like experimenting on "non-cleaned" audios i guess
hasty arch
tho lyery is like experimenting on "non-cleaned" audios i guess
I wonder if they make the model have less mess ups
winged delta
I wonder if they make the model have less mess ups
using mel denoise aggressive may help a bit
tho it isnt that messed up, for me its good even tho it has bg noise, tongue clicks and mic noise (?)
stark sequoia
it has bg noise?
both dataset and inference audios has background noise
stark sequoia
I wonder if they make the model have less mess ups
allows the model to handle background noise better
stark sequoia
using mel denoise aggressive may help a bit
i used spectral denoise
winged delta
i used spectral denoise
btw mel denoise aggressive removes more things besides white noise
aufr33 confirms it
it still may help a bit i guess?
stark sequoia
it still may help a bit i guess?
i delete sound effects manually
i don't use any ai denoisers in my models
stark sequoia
I wonder if they make the model have less mess ups
if you want to leave background noise in your dataset make sure is barely audible
coarse trout
if you want to leave background noise in your dataset make sure is barely audibl...
Bro I need to do this so bad because my models are so sensitive to noise that it's hard to make samples for them. Also at what volume should you listen to your dataset to see if the noise is barely audible?
stark sequoia
Bro I need to do this so bad because my models are so sensitive to noise that it...
like this
coarse trout
So if the noise is under -70 db it's fine?
stark sequoia
after normalization the volume is going to increase to this value
stark sequoia
So if the noise is under -70 db it's fine?
yes
this is after rvc normalization
don't go over -63db
coarse trout
Would it be a good idea to add artificial white noise or does it have to be real noise
Artificial meaning using the signal generator and adding white noise everywhere and real noise meaning the noise that's in the dataset before cleaning
hasty arch
wouldn't adding noise damage it
coarse trout
wouldn't adding noise damage it
If it's quiet nah
stark sequoia
Would it be a good idea to add artificial white noise or does it have to be real...
has to be real noise
hasty arch
ai can still detect it tho :p
stark sequoia
wouldn't adding noise damage it
no, rvc can clone noise
coarse trout
ai can still detect it tho :p
Rvc expects noise
hasty arch
Rvc expects noise
oh
coarse trout
Thanks to the most noisy pretrain
hasty arch
no, rvc can clone noise
which is bad apparently since models used to get taken down ruthlessly for it
stark sequoia
which is bad apparently since models used to get taken down ruthlessly for it
loud noise is bad for models
but yes i also agree models with slight background noise shouldn't be taken down
as they can inference noise better
(like this model)
also too much denoising damages the dataset
hasty arch
but yes i also agree models with slight background noise shouldn't be taken down
I have been conditioned to remove noise from datasets no matter how small because of that bot that takes your models down for the slightest flaw
hasty arch
also too much denoising damages the dataset
how so?
stark sequoia
I have been conditioned to remove noise from datasets no matter how small becaus...
also i didn't de-ess this dataset
hasty arch
also i didn't de-ess this dataset
I don't do that either
stark sequoia
so both de-essing and denoising is not mandatory for models, thats just ai hub nonsense
cinder ospreyBOT
so both de-essing and denoising is not mandatory for models, thats just ai hub n...
stark sequoia
how so?
damages the dynamics of the model, which is the most important part of the dataset
hasty arch
damages the dynamics of the model, which is the most important part of the datas...
dynamics?
stark sequoia
dynamics?
makes the audio sound a bit muffled
thats bad for rvc, makes gradients go crazy
hasty arch
makes the audio sound a bit muffled
stark sequoia
and the end result is the model having volume issues
hasty arch
and the end result is the model having volume issues
what kind?
stark sequoia
what kind?
the model randomly sounds very low then very high
hasty arch
the model randomly sounds very low then very high
I've experienced that with pitch sometimes but never volume
stark sequoia
I've experienced that with pitch sometimes but never volume
yes bc that problem only happens when your dataset is full of damaged dynamics
so as long you denoise good enough, is going to be fine
hasty arch
ah
drowsy lotus
if i were u i would train that model on 800 epochs or 1000 with batch size 12
winged delta
so both de-essing and denoising is not mandatory for models, thats just ai hub n...
tho sometimes harsh esses may be bad if u dont de ess
winged delta
if i were u i would train that model on 800 epochs or 1000 with batch size 12
good for ya :p
drowsy lotus
good for ya :p
ye
stark sequoia
tho sometimes harsh esses may be bad if u dont de ess
this dataset had harsh esses
winged delta
this dataset had harsh esses
wow
drowsy lotus
https://tenor.com/view/cow-femur-breaker-gif-22814322
rn im in a bad mood and bad day :p
stark sequoia
thought that esses sound bad, tho iirc bad esses may be caused by overfitting
bad sibilance happens when the dataset is too small or not diverse enough
winged delta
bad sibilance happens when the dataset is too small or not diverse enough
oh, so it's not because of overfitting
stark sequoia
oh, so it's not because of overfitting
it is overfitting, small datasets overfits extremely fast
unstable training also causes bad sibilance
hasty arch
so as long you denoise good enough, is going to be fine
I just use the denoise on MVSEP by aufr33
it's pretty good since it basically removed the horrible mic noise from the hl1 scientists
winged delta
I just use the denoise on MVSEP by aufr33
the aggressive one?
hasty arch
the aggressive one?
ye
hasty arch
it was 65 epochs, not 63
my bad lmao
I do that all the time, only because I lazy and just copy my last post since most use same epoch count
I forget to change stuff sometimes bc excitement to post it
stark sequoia