outer moth
Hello Model It's okay to say long sentences but is it a problem for the model if I say the words that I don't mean when I say short? How can I solve this? For example, if I say this, the model says gis.
There's no problem if I pronounce it clearly,
● but I just want it to copy its voice, not to presume what I says.
Does using more quality data solve the problem?
I speak in Korean. I teached Korean. But is it because the base model of which rvc is English? For example.. hubert or tokenizer?
fierce basinBOT
Ayo? @outer moth level 2 !!! 
thick jolt
Hello Model It's okay to say long sentences but is it a problem for the model if...
Probably because the dataset lacks the specific phoneme
You can also disable index (set feature index ratio to 0) and see if it that helps
outer moth
Probably because the dataset lacks the specific phoneme
I don't think it's an index problem. Maybe it's because the dataset is small. I used about 10~15 minutes of data.
thick jolt
10-15 mins would already be fine
Maybe it's because your dataset is entirely based in Korean and it doesn't have certain phonemes
You can add an English part for the dataset if that's the case
outer moth
10-15 mins would already be fine
I think same. But it seems like I'm not good enough to talk like it's not really rvc. I wonder if increasing the dataset for about an hour will solve pronunciation or word estimation problems.
outer moth
You can add an English part for the dataset if that's the case
Oh I didnt say Eng. For example I say "이제", But model say "이게". 이제 means now. 이게 means this
thick jolt
I think same. But it seems like I'm not good enough to talk like it's not really...
RVC works simply.
Give it data, it'll replicate that data. If you don't give it data it won't be able to properly give back that data.
So I guess it's just a phoneme issue on your dataset - maybe get 20 mins more and see what happens
outer moth
RVC works simply.
Give it data, it'll replicate that data. If you don't give it...
Ok I think same you. I will train more than 1h
thick jolt
Ok I think same you. I will train more than 1h
Nah 1 hour is excessive imo
Try 20-30 mins, will probably be the same as an 1hr model
outer moth
Try 20-30 mins, will probably be the same as an 1hr model
The people says not to spend a lot of time because it's hard on people? Is it because the model's performance gets worse? On the premise that the dataset is good.
thick jolt
The people says not to spend a lot of time because it's hard on people? Is it be...
The model doesn't get worse, but the improvements significantly decrease once you add more and more time
Like, 10 mins of dataset to 30 mins of datsset will be a pretty good bump, but to 1hr compared to 30 mins it's negligible
outer moth
Like, 10 mins of dataset to 30 mins of datsset will be a pretty good bump, but t...
I understand. Thank you