#Newbie on this and i need help (ive already read the doc located in guides) (japanese voice)
1 messages · Page 1 of 1 (latest)
Check https://docs.aihub.gg/rvc/resources/dataset-isolation/ & https://docs.aihub.gg/rvc/resources/training
Last update: Dec 24, 2024
Last update: Dec 24, 2024
Also 10 hours are a bit too much lol
U need to be sure they are good quality
One message removed from a suspended account.
One message removed from a suspended account.
you could ask codename if you can have his Kurisu model
One message removed from a suspended account.
It took him months to make his Kurisu model 😭
One message removed from a suspended account.
One message removed from a suspended account.
you don't have to train 10 hours
iirc codename used 10 mins
for the kurisu model
One message removed from a suspended account.
One message removed from a suspended account.
rvc can't correctly clone mouth clicks
he meant 10 hrs of the entire source with multiple speakers
oh
i think rvc models can whisper only if the pretrain was trained using whispers
i believe klm do have whisper support
One message removed from a suspended account.
how much the length of each speaker in the entire source?
One message removed from a suspended account.
better remove mouth clicks
hold on
no, he meant 10 hours of kurisu talking
she has tons of voice lines
it's usually 24 mins each anime episode and 12 episodes in a season
One message removed from a suspended account.
you used the vn?
the only person who knows how to get the perfect kurisu model is codename tho
One message removed from a suspended account.
xD
One message removed from a suspended account.
he thought you wanted to do a rvc model using multiple speakers
One message removed from a suspended account.
for training whispering use this pretrain https://discord.com/channels/1159260121998827560/1339155300720054316
the original doesn't have whispering
remember to remove mouth clicks from the dataset
One message removed from a suspended account.
entrenar diferentes hablantes, diferentes personas
🦈
One message removed from a suspended account.
I think you should use one of the games as the source
One message removed from a suspended account.
One message removed from a suspended account.
i have no idea how codename trained his kurisu model
but i remember that is bad to use different sources for a dataset
One message removed from a suspended account.
One message removed from a suspended account.
yes, only include one person in the dataset
and also remove every mouth clicking sound
One message removed from a suspended account.
One message removed from a suspended account.
#1298338943410110474 message
I had a model from a VN (and it got anime adaptation), it is 2 hrs total of around 2k voicelines that are easy to group according to the character (other characters only have less than 50 min length). it was trained using og pretrain & mainline, so I could improve it in future but not rn.
One message removed from a suspended account.
no idea, you seem to overcomplicate it
One message removed from a suspended account.
One message removed from a suspended account.
Dang even when you offer to pay he said no
I mean it makes sense considering all of the effort that went into it
weeks cleaning the set + months training it... yeah tbh i would also not share a model like that
Mhm
One message removed from a suspended account.
One message removed from a suspended account.
One message removed from a suspended account.
yup
you'd better remove mouth clicks in the dataset audio, perhaps using izotope RX mouth de-click
though it may also depend on the pretrain whether it also has mouth clicks or not
One message removed from a suspended account.
One message removed from a suspended account.
not the dataset, but spectrogram of some part is fine
One message removed from a suspended account.