unkempt steeple
first model that is not a vtuber/anime character related yippee
trained on 1 minute of data from the yapdollar tiktok/short. (mel roformer + resemble enhance)
pitch extraction: rmvpe
steps: ???k (don't know, trained this when i was away from home)
batch size: 2
pretrain: klm 4.1 / 32k
huggingface: https://huggingface.co/sxndypz/rvc-v2-models/resolve/main/yapdollar.zip
sonic mauveBOT
sonic mauveBOT
unkempt steeple
unkempt steeple
slow internet
finite torrent
appeared before samples
finite torrent
slow internet
its ok
finally a model from sxndy that isnt anime/vtuber hehe
tho why klm 4.1? u should have used og pretrain :p
unkempt steeple
klm 4.1 has better range imo
finite torrent
klm 4.1 has better range imo
but no distorted harmonics/distorted high ends/low end noise?
unkempt steeple
klm 4.1 has less severe harmonics compared to the 4.2 one
finite torrent
tho i still prefer og pretrain, keeps the accent of the character and has no issues imo
unkempt steeple
i mean i will choose og sometimes, but i choose klm for higher range
finite torrent
high range of what?
unkempt steeple
higher voice range
since on the original pretrain, whenever the inference hits the higher voice range it just breaks
i have also trained the original fiveish voice
will post that as well
radiant plank
Is it possible to get a model without a single instance of like
Whats it called
Idk artifacting or shit
unkempt steeple
Idk artifacting or shit
don't think so. hifigan would still cause some artifacts, even if the dataset is the highest quality.
finite torrent
don't think so. hifigan would still cause some artifacts, even if the dataset is...
if hifigan dosent have limits we would get way better models
unkempt steeple
exactly
radiant plank
Hifigan jumped my family
unkempt steeple
btw surt how's my yapdollar model?
radiant plank
Better than mines
finite torrent
@unkempt steeple razer confirmed me that there are distorted harmonics on klm 4.1. even tho seoul says that there arent distorted harmonics, according to razer it has a distorted harmonic on the right side of the spectogram if im not wrong.
unkempt steeple
i mean yeah it has some distorted harmonics
finite torrent
idk if there are other distorted harmonics than this but yea, i personally prefer og pretrain to prevent issues like this.
unkempt steeple
but they're not as severe as the klm 4.2 one
maybe it's because of the dataset i used for training
finite torrent
its the pretrains fault, it couldnt be the dataset ig
finite torrent
like this one?
ye
native swift
XIAO HONG SHUUUUUUUU 🔥 🔥 🔥 🔥 🔥 🗣️ 🗣️ 🗣️ 🗣️ 🗣️
unkempt steeple
next model is going to be... you guessed it. a vtuber.
and it's an another idolen vtuber
uneven gyro
and it's an another idolen vtuber
finite torrent
XIAO HONG SHUUUUUUUUUUU
uneven gyro
XIAO HONG SHUUUUUUUUUUU
native finchBOT
finite osprey