#Articulation Issues

1 messages · Page 1 of 1 (latest)

hearty vortex
#

I am facing issues with articulation in the voice of my models. The tonality seems to be good, but there is inconsistency in articulation. It is correct in some areas, but lacks consistency.

Model (10min Dataset) trained on the new SuperPresets with a batch size of 16 and 300 epochs.

Here is a picture of my Graph:

received_7080384208721129.png
devout joltBOT
#

Ayo? @hearty vortex level 2 !!! lfg

long mango
hearty vortex
#

Update : Trained up to 500 same issue

twin meadow
hearty vortex
#

so you think it's the settings I have to mess with to fix this issue ?

#

cause my dataset can't get any cleaner I've followed every UVR / Rx / Adobe method @long mango guided me thru

twin meadow
hearty vortex
#

alright I will inform you tomorrow thanks for helping me

twin meadow
twin meadow
hearty vortex
#

I'll leave the thread as non resolved for now goodnight

twin meadow
#

alrighty. good night!

hearty vortex
#

@twin meadow ok so I haven't trained with the suggested parameters you told me but I have a question

twin meadow
hearty vortex
#

The dataset I'm training with the OvSuper40k (which is based in English) is in Greek Language

#

Do you think this is what causes articulation Issues?

twin meadow
hearty vortex
#

yeah funny thing is I have errors only in syphons

#

cause Greek language consists of 80% phonemes

twin meadow
#

It's possible that a Greek pretrain would be better in that case... hm

hearty vortex
#

Okay so how do I make one then?

twin meadow
hearty vortex
#

Does the server have any docs I can study

twin meadow
hearty vortex
#

about pre model training

twin meadow
#

you'd have to get like

#

50 hours of speech or something

hearty vortex
#

Where can I learn more about it

twin meadow
#

and like, literally, days

#

of training

twin meadow
#

maybe ask @slim pumice, they made the ov2 pretrains

jaunty finch
#

you can try asking simplecup himself

#

yeah 💀

twin meadow
#

yeah

hearty vortex
#

I'm willing to pay for this knowledge

jaunty finch
#

im sure he would help but i dont think it would help you that much in the long run

hearty vortex
#

will he see the messages tho I suppose he is busy

twin meadow
#

eh maybe

jaunty finch
#

nah he will just dm him

hearty vortex
#

Then what do you suggest

#

I want to fix the articulation in Greek

#

Parameters or Pretrain are the issue because my graphs are very decent

jaunty finch
#

are you sure your dataset is cleaned properly? and how big is it?

twin meadow
twin meadow
hearty vortex
#

yeah

twin meadow
#

And it's around 10 mins iirc

jaunty finch
#

yeah i saw, but it still could be bad lol

#

didnt see the 10min part sorry

#

then its either the parameters thing or you need to expand your dataset

#

10min isnt a lot

#

but then again you used the new pretrains

hearty vortex
#

So you think this will fix articulation?

#

yeah which are based in English

devout joltBOT
#

Ayo? @hearty vortex level 3 !!! lfg

long mango
hearty vortex
#

Alright I'll try Mj's params and let you guys know

jaunty finch
#

what i would do is expand the dataset to around 30min and try using the normal pretrain with the parameters fixed

#

and if you really want to learn more about the pretrains try asking simplcup

hearty vortex
#

Are the new pretrains non reliable?

long mango
#

That's to be expected though

jaunty finch
#

i didnt notice many improvements on my end and im still a bit sceptical about the smaller dataset/epochs part tbh

hearty vortex
#

Why tho I supposed having more voices would make them more flexible

jaunty finch
#

yeah in your dataset, not the pretrain

long mango
jaunty finch
#

thats why i suggested the stock one, just to be safe

#

nothing wrong with the new one

#

but still

hearty vortex
#

alright I really thank you for your time guys

#

I'll try everything and let you know

twin meadow
twin meadow
long mango
twin meadow
#

"I doesn't believe that exists yet"

long mango
slim pumice
hearty vortex