no u gotta select it
#Guide for Applio - Help me
pallid breach
ok so first choose spin here
enable custom pretrain
and use dat
done
now u training spin
important shit to remember
in the inference tab u also need to select spin
otherwise ur model kinda
dies
😭
scarlet rapids
keep in mind that models trained with spin would also require spin embedder in inference, otherwise the voice would speak gibberish
scarlet rapids
important shit to remember
in the inference tab u also need to select spin
^
the inference option should also have like that
pallid breach
now click that
and choose the finger spinner
also in case u wanna use your model in realtime, this doesn't work in w-okada
but works in another voice changer named vonovox
no i mean using the model as a voice changer
vonovox is literally a better wokada just in case
scarlet rapids
check out the vonovox guide here https://docs.aihub.gg/rvc-voice-changer/local/vonovox/
pallid breach
shit can run fine while playing games without having to decrease graphics
and
it doesnt open in ur browser
give it a try
spin is there
click the link sybau (😭 ) sent
same shit as applio
pallid breach
download source (zip)
run setup.bat
let it install its shit
run start.bat
way better
trust
dev is known here
yea sure
vb audio cable is shit why u using it?
thats
actually quite rare
lmao
bro is lucky asf
pray vb audio cable works there anyway lmao
no one has ever tried it
but it should
lets say vb audio cable is maybe a tiny bit buggy
scarlet rapids
scarlet rapids
at least the UI system (tkinter) isn't bloated since it doesn't use chromium framework
pallid breach
block size is chunk ms
server mode wokada
use wasapi
to use spin
u always used client mode in w-okada?
scarlet rapids
the audio device type, WASAPI has less latency than MME like used in wokada client mode
pallid breach
audio backend is the thing that sends audio between apps
client mode was a silly idea wok had at the last time
it may be the reason why his voice changer app runs extremely bad
block size = chunk
scarlet rapids
u always used client mode in w-okada?
haven't tried vonovox actually, but does it also have ASIO?
pallid breach
lower block = less delay, higher gpu usage
higher = u get the idea
pallid breach
haven't tried vonovox actually, but does it also have ASIO?
yes
play with the block size value until u find one that suits best for u
but dev is now working on that actually
bad for realtime
yea like u can try it after u finish ur training
if u set your index value to 0 wokada wont use it btw
index is bad because it can potentially add more delay and runs poorly on cpu
also messes up pronunciation
like super bad
index idea was because models trained with super small datasets usually don't sound like the intended target, rvc-boss made the index files so they can sound more closer and true to the original source
bigger datasets dont have that problem
g and d uses 1gb of space together iirc
epochs are 54 mb
you're not saving every g and d, right? lmao
scarlet rapids
bad for realtime
I think the index option should be provided anyway, some ppl with capable spec may find it useful for better accent
pallid breach
are they named g_23333?
pallid breach
I think the index option should be provided anyway, some ppl with capable spec m...
ok nice so you're not saving every g and d
u cant, your model needs the g and d files in order to continue training
scarlet rapids
at least hide it with warning like that
pallid breach
you're not saving every g and d
dont worry
"save only latest"
be sure thats enabled
never disable it
for original rvc training? (og pretrain, cvec) nothing
for spin, select spin where i told u, and use custom pretrained
ignore the rest of stuff
yeah
deletes everything and starts the training from 0
useful if u wanna train the same dataset but with a different batch size
btw for vonovox u need this
ah
im restarted
lmao
weird
no
applio doesnt install that
nor vonovox
also hear ur epochs
machine learning has this thing named overtraining that like the name implies, if u train a model too much, it may not be able to do predictions anymore
in rvc the only way to know if your model is overtrained is by hearing it
overtrained models sounds very robotic and unnatural
yea trust, this is pretty ez
kaggle uses a gpu from 2016 lmao
i mean its free
ai shit is heavy what else i can say
my 5 hour dataset log folder is 15gb
pallid breach
my 5 hour dataset log folder is 15gb
but it's totally worth it, the more u add, the less robotic the output 
the model also knows how to reproduce volume changes more naturally, and knows which sample has to use depending on the context (pretty cool)
for realism i'd say 1-2 hours minimum of data is ok (if u want the best possible results obviously you need more than that, up to max 24 hours maybe), better results if the dataset is super expressive, in my case my dataset was super monotone
pallid breach
did u ran start.bat as admin?
either you ran setup.bat as admin or start.bat as admin
dont run any bat as admin
just double click them
pallid breach
just in case reinstall the whole thing
this time be sure you dont run none of the two bat file as admin
pallid breach
i mean u can just use audacity or something
im gonna tell the dev this suggestion
btw vonovox has a disc server too
dude who made the app it's there (dr87)
scroll down and find this
click where it says discord server
trim plume
Just set the output to your headphones