#Daitaku Helios (Singing v2) (Umamusume) (RVC v2) - 300 Epochs

1 messages · Page 1 of 1 (latest)

subtle spindle
#

-# i added more singing data to the dataset so she wouldn't strain in higher notes

this character's featured in umamusume (both the anime and game)

trained on 20 minutes of of her singing data extracted from the game
(cleaned using anvuew's de-reverb, uvr de-echo (in some parts), resampled to 40k)

songs used:

  • umapyoi legend

  • make debut!

  • instictive speed

  • girls' legend u

  • run for our dream!

  • gaze on me!

  • overrunner

  • ms. victoria

  • dramatic journey

  • tracen ondo

  • u.m.a. new world!!

  • uma summer

CV: Aya Yamane

pitch extraction: rmvpe
steps: 23.1k
batch size: 6
pretrain: klm 4 hfg / 40k

hf: https://huggingface.co/sxndypz/rvc-v2-models/resolve/main/daitakuhelios_sing_v2.zip

mtcrabbit-daitaku-helios.gif
subtle spindle
#

i'm actually surprised by how it turned out

winged delta
#

thanks for giving the step amount for your models, it's quite interesting to see how the sweetspot for these are 20k~

subtle spindle
#

you're welcome

#

earlier i went for 7 batch size 17k-19k ish steps

#

and now i just hear if there are less sibilant artifacts present in samples than the prev checkpoint

#

and also i don't train models further than 3-4 hours because i don't wanna risk overtraining it

#

ik there are models out here with 50k - 100k steps and they sound fine but i'm playing it safe

winged delta
subtle spindle
#

thanks

#

this model turned out pretty good for its pretrain it used

#

there are some problems with the pretrain like how some of the models i've trained with it where the dataset contained "long-held notes" didn't properly held long notes imo

#

if you know what i'm saying

wise furnace
winged delta
subtle spindle
wise furnace
subtle spindle
#

oh icic

#

can't wait till legacy core v2.9 fully releases so it could overtake klm in terms of range

#

i also can't wait till config.json gets fixed for 40k and 48k

subtle spindle
winged delta
#

1.5 also has that problem but i dont think is that bad (?

#

but its just a side effect of using singing datasets

subtle spindle
#

interesting 🤔

winged delta
# subtle spindle interesting 🤔

mixing speech and singing just makes that problem worse for pretrains (i tried that before), tho for finetuning i think it's fine, not sure

#

and finetuning those kinds of pretrains always leads to some weird problems with vibratos

subtle spindle
#

ohhh so i have to either choose a pretrain trained on speech or a pretrain trained on singing

winged delta
#

ye