#My own voice for singing (I have 20 mins clean audio)

1 messages · Page 1 of 1 (latest)

civic quiver
#

Hey guys, I need someone to help me create a model using my voice recorded. On a professional mic for a singing model.
I have a pc that can run the training - but I need help with what khz to choose (e.g, 32k, 48k etc), what method (hifigan, refinegan etc), what epoch to use, and whatever else. I’ve made my own model already but I’d like it to be the best quality possible.
My audio set is or easy cleaned and denoised

limber notch
#

is it 20 min non stop? and was it already litle noise before the denoise (since denoise itself is damaging)

civic quiver
#

Left some small pauses and breaths here and there

#

I don’t know if muffled is the right word

#

It sounds very clear

#

But it’s not as clear as real singing

#

I’m not talking about mispronounced consonants etc

#

But I just want it to be as near to as real singing

#

In terms of crispness

limber notch
#

So you've made the model alrady?

#

or is the probelm you dont know how to train at all

civic quiver
#

E.g. if there’s a certain setting I should use, certain pretrain I should be using,
Certain rvc/svc architecture

limber notch
#

I mean the standard is still cvec hifigan

#

which is default

#

and one of the legacy core pretrains should be good

#

and thats pretty much the best you can go unless you find someone who's willing to give a personal pretrain (they won't)

civic quiver
#

Still slightly unsure as some people say 32k better, some say 48k

limber notch
#

32k is better

#

Cuz all the architectures or something are made for 32k

#

the lower the better i heard

#

but 40k is just bad cuz wth is 40k