T de-ess is
#✨│ai-help
peak falcon
cracked
crude flame
perhaps
low shard
but you can get RX "for free"
🏴☠️
peak falcon
I have a de esser from FabFilter but I have no idea how to use it properly
It's called Pro-DS
proven hill
please no cracked stuff pay full price.
crude flame
Everything cracked
you mean you payed full price right? because you are an outstanding citizen
crude flame
ight, thank you for your honesty
bold yarrow
Everything cracked
someones about to get banned
peak falcon
For using cracked software?
peak falcon
I don't know, I didn't read them
I just joined for the Ai models because I tried to make a producer tag
bold yarrow
oh
low shard
isnt that in the rules
well sharing the link of it causes us troubles
bold yarrow
well i embarrassed myself
low shard
but just saying "google how to be a pirate" won't get u banned
bold yarrow
could piss off your isp
well i just downloaded de-esser
but i can't find it anywhere
did i download a virus from ruislip or something
ill show them
peak falcon
what DAW are you using?
bold yarrow
daw?
peak falcon
how do you mix the vocals?
tame mica
get a free daw like reaper
audacity is not ideal for audio mixing
or you can yk "buy" other daws
proven hill
reaper best
vapid mantle
@hot ledge hocam kaç Target Sample Rate olmalı ?
hot ledge
vapid mantle
tm hocam eyvallah
crystal jetty
Hello everyone, can someone help me? I'm generally 0 in these matters(
low shard
can sm1 send me the voice changer file for windows i cant fibd it
that's wokada, tell me ur pc gpu in #🔍│help-w-okada
low shard
Hello everyone, can someone help me? I'm generally 0 in these matters(
in what?
latent cypress
do you guys prefer klm 5.0 mini or klm 4.3 x2?
jaunty shale
is there a way to convert .safetensors file to .pth file so I can use it in applio?
simple ore
safetensors of what? gptsovits?
jaunty shale
I used okada to make a merged voice
it created a safetensor file in the model_dir folder
i run it on browser (it works better than window one for me)
static oar
so i just used Google Collab to make a voice model from audio clips right, works good with the voice changer but i was wondering what i can use to be able to apply that voice model to audio. any ideas?
rough star
Does that mean that I have to create a new user on my PC?
lime otter
-colab
simple ore
Does that mean that I have to create a new user on my PC?
it mean literally do not run it using Administrator user
you can run it as a normal user that has local admin priveleges
red gale
does anyone of you know any uhh,good free voice changers?
rough star
you can run it as a normal user that has local admin priveleges
Okay so double click wouldn't run it as admin right?
simple ore
Okay so double click wouldn't run it as admin right?
just download a compiled version, my dude
knotty moth
<@687355869884973059> hocam kaç Target Sample Rate olmalı ?
the result will be always bad even if you upscale it to 32k
32k'ya yükseltseniz bile sonuç her zaman kötü olacaktır
latent kettle
reinstall it
What
proven hill
reinstall it
reinstall what?
red gale
<#1159290161683767298>
Figured it out already,no worries
jaunty shale
is there a way to convert .safetensors file to .pth file so I can use it in appl...
is there way to convert it so I don't have to make a dataset for the merged model?
bold yarrow
hey guys
is there a website where i can accurately predict how many epochs my model would need without becoming either undertrained or overtrained
proven hill
is there a website where i can accurately predict how many epochs my model would...
tensorboard
proven hill
is there way to convert it so I don't have to make a dataset for the merged mode...
why would you do that, no its not possible
jaunty shale
why would you do that, no its not possible
wanted to create a merged voice and use it to convert audio voices into that voice-
proven hill
you can simply merge two voices
jaunty shale
I can't.
I tried in applio it doesn't work anymore.
The merged voice I have has more than 2 voices combined
proven hill
are the files youre merging the same sample rate?
jaunty shale
Yep
proven hill
can you show me the error?
jaunty shale
will do
craggy brook
Is there any other way to access the 15.ai site? Or is there a site as good as this site with better sound?
proven hill
Is there any other way to access the 15.ai site? Or is there a site as good as t...
what do you need?
craggy brook
voicing for a character I want to use. realistically, with emotions and without distorting the tone of voice
proven hill
voicing for a character I want to use. realistically, with emotions and without ...
try weights.gg
jaunty shale
ok now it worked-
..somehow
but I already have premade merged voice. Is there like a long enough audio to convert in real time?
proven hill
but I already have premade merged voice. Is there like a long enough audio to co...
wdym
jaunty shale
wdym
So you know how you can use voice in real time? Not only I can use my voice but also audio to convert into in real time pretty much. It works very well and I did it multiple times, but I just need some audio that is long enough to train the merged audio I did in okada.
(I use soundpad to make it work)
proven hill
So you know how you can use voice in real time? Not only I can use my voice but ...
-rt
proven hill
first link
proven hill
how can i use my gpu to rvc?
what do you need to do?
whats your gpu?
jaunty shale
then what you need exactly?
I only need some audio to play through my soundboard to convert it to that merged voice and record it.
Some audio that is long enough, so I can make a dataset.
knotty moth
they are already doing that
nvm I scrolled too much
proven hill
I only need some audio to play through my soundboard to convert it to that merge...
i dont understand this at all honestly
solid arch
or it is set that way
proven hill
im using rx 570
cooked badly
solid arch
idk abt anything but its working
proven hill
so no file input?
knotty moth
how can i use my gpu to rvc?
if u mean the inference, use applio with zluda. I suppose it will be okay with at least 4 gb vram
heady gorge
Do I really need winzip for the voice changer
proven hill
you have to unpack it somehow.
but i suggest 7zip
solid arch
best way to like stop cutting the audio out?
knotty moth
Do I really need winzip for the voice changer
7zip or winrar is preferrable
proven hill
best way to like stop cutting the audio out?
realtime?
solid arch
proven hill
7zip or winrar is preferrable
hun, 7zip>>>
proven hill
set gpu to your gpu
heady gorge
Why do voice changers have to cost?
proven hill
set f0 do rmvpe
solid arch
set gpu to your gpu
aint working tho only cpu
proven hill
Why do voice changers have to cost?
the one we have its free
solid arch
oh
heady gorge
Yeah, but I need winzip
proven hill
Yeah, but I need winzip
download 7zip
heady gorge
Is that free
knotty moth
hun, 7zip>>>
both work at least
proven hill
yes
solid arch
i download it after
proven hill
both work at least
true
knotty moth
Is that free
7zip is open source
solid arch
this
jaunty shale
i dont understand this at all honestly
..okay let me explain it differently.
I use Voicemeeter Banana to make this happen. B2 is input device that is used to convert anything that can be played in it. B1 is something that will be played through the actual microphone. When I play audio through soundboard, I can play it in the B2 to convert it into the different voice in real time. (since I cannot convert it in applio). With that, I can make a dataset.
heady gorge
7zip is open source
Can you send me the link
proven hill
Can you send me the link
google is free
solid arch
I have okada already.
ure telling me that u want some audios to be converted into a voice which is to okada so okada would transmit that and turn into a different voice and will play it thru to ur mic
a audio came from ur soundboard?
soundboard > okada > mic
jaunty shale
ure telling me that u want some audios to be converted into a voice which is to ...
correct. So I can make a dataset.
I really don't wanna spend more hours to make it in applio again.
heady gorge
Are you sure this is safe
solid arch
sends any kind of audio thru my mic
proven hill
..okay let me explain it differently.
I use Voicemeeter Banana to make this hap...
you wanna make a dataset out of converted audios?
proven hill
Are you sure this is safe
yes and if you dont trust check internet
jaunty shale
you wanna make a dataset out of converted audios?
yes. I did it before in applio and it workedd out just fine.
proven hill
yes. I did it before in applio and it workedd out just fine.
why would you make a dataset of an already existing model
solid arch
imma extract the file
knotty moth
I cannot use safetensor file in applio
use the original model file before you added it into the voice changer
jaunty shale
merged voice makes only a safetensor file.
knotty moth
merged voice makes only a safetensor file.
only if you do it within the voice changer
proven hill
What's the link to the voice changer
-rt
proven hill
first link
proven hill
only if you do it within the voice changer
that explains it
yea you need to merge in mainline (discontinued), applio (suggested) or ilaria rvc mainline (discontinued)
jaunty shale
only if you do it within the voice changer
in okada, yes.
heady gorge
I'm not good with PC stuff But I can try
proven hill
I'm not good with PC stuff But I can try
just follow the guide, its easy :)
solid arch
chat why im one my browser 😭
jaunty shale
I'll just figure it out. I only need a long audio from youtube.
proven hill
please follow the guide
proven hill
I'll just figure it out. I only need a long audio from youtube.
again follow what i said
low shard
chat why im one my browser 😭
That's wokada, use #🔍│help-w-okada
jaunty shale
again follow what i said
?
heady gorge
Okay. Can I have the link to the voice changer
proven hill
-rt
proven hill
first link
jaunty shale
dont merge with the voice changer
oh right, will keep in mind for now, like I said I never knew it makes safetensor file in the first place.
proven hill
oh right, will keep in mind for now, like I said I never knew it makes safetenso...
merged files in applio will be in pth, exactly what you need
heady gorge
I don't see the Voice changer in 7zip
low shard
I don't see the Voice changer in 7zip
Realtime voice changer for calls? Tell me ur PC GPU in #🔍│help-w-okada
heady gorge
it keeps on changing
craggy brook
try weights.gg
How to make characters voices with emotions such as angry, confused, sad etc.
proven hill
How to make characters voices with emotions such as angry, confused, sad etc.
you cant
assuming youre using tts
craggy brook
assuming youre using tts
what ?
craggy brook
yes
winged crane
yo mods can i get permission to share my screen?
proven hill
yo mods can i get permission to share my screen?
for?
winged crane
i wannashre my screen when im playing dragonball sparkeling zero with a friend
if thats ok?
share
winged crane
you can join and see aswell if you want to
knotty moth
you can join and see aswell if you want to
ask mods for streaming permission, and be sure to avoid showing any inappropriate stuffs
hallow thistle
proven hill
winged crane
@quick jungle can i get permission to share my screen
quick jungleBOT
<@1158132286169022555> can i get permission to share my screen
Watchu want? :3
proven hill
its a bot
flint geyser
yo do you want me to still fix it
low shard
yo do you want me to still fix it
I mean if u could fix it it would be good, but it's left to rot since months
honest junco
Sorry I tried couples times but my RVC keep showing Frequent errors occur. Please check if the model of the framework being targeted is loaded.
And my colab are showing my server is not an accepted origin. (further occurrences of this error will be logged with level INFO)
I tried to search in github but I still can't find anyway to fix it
(Fun story I used 3 hours to do pip install pip==24.0
proven hill
(Fun story I used 3 hours to do pip install pip==24.0
download a precompiled version
honest junco
download a precompiled version
What that mean (Sorry I'm a script nub
proven hill
What that mean (Sorry I'm a script nub
it comes with everything preinstalled basically
proven hill
Oh you mean the rar
ye
proven hill
why?
proven hill
what colab are you using?
honest junco
google colab
honest junco
Alr
honest junco
give me the link
dmed u
honest junco
alr mb
proven hill
-ngrok
honest junco
yeah
proven hill
damn
proven hill
ohhh its a voice changer
honest junco
yeah
honest junco
Cuz when I use my gpu
the res of it
are 20k ms
basically it take 20s to tranfer my voice to u know
proven hill
what gpu do you have?
honest junco
proven hill
are you using the forked version?
honest junco
forked version?
proven hill
the new version
honest junco
Yes
proven hill
better support for amd cards
honest junco
1.5.3.18a
proven hill
first link
honest junco
alr tysm
low shard
alr tysm
also use #🔍│help-w-okada
proven hill
np
low shard
https://tenor.com/view/boo-gif-19787475173016375
skill issue
proven hill
skill issue
smh
proven hill
Fun fact I paid the google colab
why….
honest junco
I spent 56.5 for using that
proven hill
lmao
low shard
I though google were p2w 💀
u can use it for free 4 hours daily not granted, or use ur own gpu
honest junco
u can use it for free 4 hours daily not granted, or use ur own gpu
I even used the limit of ngrok
I need to use my nd google acount to log in to use it 
low shard
I even used the limit of ngrok
u can use horizon
it's another tunnel
anyways, ur gpu should be good enough
honest junco
honest junco
anyways, ur gpu should be good enough
More than enough ig
low shard
old colab too, https://colab.research.google.com/github/hinabl/voice-changer-colab/blob/master/Hina_Modified_Realtime_Voice_Changer_on_Colab.ipynb is the updated one
low shard
More than enough ig
yeah, u should follow the wokada deiteris fork guide
proven hill
let him download the fork
honest junco
yeah, u should follow the wokada deiteris fork guide
Idid
low shard
ye i was just saying
honest junco
seting up tho
sudden tree
where can you find models like the melband roformer karoake model by viper?
do they upload them I am too nervous to download some random ckpt on huggingface lmfao
proven hill
where can you find models like the melband roformer karoake model by viper?
@viscid moss you got this
sudden tree
he already sent me the mega but idk where he obtained file from
I wanna know the source
viscid moss
where can you find models like the melband roformer karoake model by viper?
The ViperX models are here:
https://github.com/TRvlvr/model_repo/releases/tag/all_public_uvr_models
And most of the UVR5 models
proven hill
he already sent me the mega but idk where he obtained file from
*she, btw
viscid moss
do they upload them I am too nervous to download some random ckpt on huggingface...
well u can check the model hash, that way you make sure it's the same original just re-uploaded
proven hill
also huggingface is safe :)
viscid moss
ye
sudden tree
thanks a lot guys
well doesnt huggingface literally virus check each file regardless
just wondering why isnt the viperx karaoke and stuff included in uvr 5 model download set
viscid moss
just wondering why isnt the viperx karaoke and stuff included in uvr 5 model dow...
anjok is working on that
sudden tree
oh thats sick
viscid moss
Will be available soon ig
sudden tree
nice
thank you
i wonder why everyone recs the mvsep when the queue is ungodly lmfao
not worth
valid spruce
What sample rate should I use?
proven hill
What sample rate should I use?
32k
proven hill
no problem!
sudden tree
everyone always says you dont need to cut audio yourself, but I realized when I train with my own clips, my Crepe models even beat the RVMPE models! I think the auto clipping of applio causes the ai to be confused. Has anyone else experienced this>?
analog obsidian
everyone always says you dont need to cut audio yourself, but I realized when I ...
wdym by train with my own clips
like, slicing the dataset yourself and disabling rvc's splitting?
analog obsidian
exactly
are you using the script found in the docs? you're supposed to slice the whole dataset in chunks of 3 seconds with an overlap of 0.3
and crepe vs rmvpe the difference is subtle, crepe models are softer while rmvpe are more harsh
sudden tree
i just basically use a audio ceiling to prevent white noise
and then split the audio into like 5 min chunks
rmvpe has always sounded better to me tho
analog obsidian
rmvpe has always sounded better to me tho
yes bc is more robust and new, crepe its literally from 2018
sudden tree
i wish there was just a vid of someone training with the doc settings
analog obsidian
rmvpe was made in 2023
sudden tree
most vids just do what I do and throw the audio in the training
I did not use the settings in the doc lmfao
analog obsidian
and then split the audio into like 5 min chunks
this is bad, hifigan can't read files over 5 secs
you're not slicing it yourself, rvc is slicing the dataset for u
sudden tree
no, but the pretrain split it for me automatically in applio
Yes i use the applio pre cut setting
analog obsidian
so you didn't disabled rvc splitting
sudden tree
and the process
no I didnt
but still my old crepe model sounds better when i split it by hand
analog obsidian
every training is different + batch size matters
sudden tree
true I just run like 8 batch size even though i have 12 vram bc i train on 32fp
analog obsidian
u can actually get different results using the same exact parameters
sudden tree
fp32 was big mistake activating maybe?
analog obsidian
fp32 was big mistake activating maybe?
enabling fp32 is a W move
fp16 is too unstable
sudden tree
or maybe i should deactivate the process audio preset in applio?
analog obsidian
or maybe i should deactivate the process audio preset in applio?
nope leave it enabled
sudden tree
also should I make the input audio loud or just leave it as is?
some of the audio i train is raw vocal dataset and is quite
analog obsidian
i can tell you the """right""""(not really) way to preprocess a dataset
sudden tree
sure
analog obsidian
so open audacity, open your dataset (if your dataset are multiple audio files, merge them into one audio before doing this), select the whole audio, find truncate silence and use these settings:
damn i forgot
before doing that, convert the dataset to mono
analog obsidian
bc rvc cant read stereo files
sudden tree
oh shit so could that have ruined my training?
analog obsidian
oh shit so could that have ruined my training?
no bc applio converts it to mono anyways
sudden tree
ah lmfao
analog obsidian
but since you're doing this method, you should convert it to mono
sudden tree
isnt truncate silence same as doing noise gate in fl
analog obsidian
isnt truncate silence same as doing noise gate in fl
no
this literally removes silences
and leaves only the speech audio
analog obsidian
so like this
analog obsidian
its probably training on the silences?
yup
after you have your truncate silence dataset do the next step
use these settings and you should be fine
only use simple mode if you truncated the silence
never use it for datasets that have silence
sudden tree
ok thank you a lot for this
is there a full guide so i can train
My only question is why truncate instead of just using the auto setting?
analog obsidian
is there a full guide so i can train
analog obsidian
My only question is why truncate instead of just using the auto setting?
automatic mode leaves more silences, is not technically bad but you require more epochs to train
sudden tree
ok thanks where do you find these links
analog obsidian
-docs
AI HUB Docs
analog obsidian
^
azure marshBOT
what is batch size can you help me
sudden tree
is 8 batch size good for 13 min vocal data
analog obsidian
is 8 batch size good for 13 min vocal data
yes
sudden tree
what about 12 or 4
analog obsidian
4 is unsafe but works in some cases where the dataset is very monotone and repetitive
12 works in some cases as well
where 8 fails
sudden tree
do you just use the built in pretrain?
analog obsidian
yes, original pretrain
sudden tree
ok just wondering if going to 12 batch size would help it
analog obsidian
ok just wondering if going to 12 batch size would help it
8 is safer
sudden tree
hopefully truncate silence will remove all random noises in a studio sesh
analog obsidian
as long they're below -42,5 db
sudden tree
also do you personally use the melband roformer karaoke models to isolate leads?
analog obsidian
also do you personally use the melband roformer karaoke models to isolate leads?
yuh, only good model we have for vocals that have harmonies
sudden tree
lmfao, audacity doesnt take m4as jeez
analog obsidian
😭
glacial pollen
lmfao, audacity doesnt take m4as jeez
you can use ffmpeg
a must have for everyone working with audio
analog obsidian
my beloved
glacial pollen
well no, m4a is not a codec
analog obsidian
bro convert it to wav
glacial pollen
it is a container
analog obsidian
lmao
sudden tree
oh i thought m4a was codec
sudden tree
nice for a while i thought youtube had opus but i dont see it anymore to download
glacial pollen
it's kinda mp4 counterpart
just the difference is, it doesn't contain " video layer "
glacial pollen
essentially, it's " MPEG-4 Audio Layer "
sudden tree
lmfao now i gotta learn cmd ffmpeg fuck!
glacial pollen
Ye, Opus is in fact a very good codec
for lossy stuff
direct successor of vorbis ( ogg )
analog obsidian
.\ffmpeg -i audio.m4a audio.wav
glacial pollen
^ ye, will unwrap the container and get it to wave
Tho as Lyery said, if you're working on audio from youtube, use yt-dlp.exe ( a cli tool )
it'll fetch the audio from yt servers in best possible available quality ( mostly opus and rarely aac ) And then convert that using ffmpeg
(( that's my exact workflow for ' yt sourced audio ' ))
sudden tree
damn never new m4a was container does it just hold mp3 atp?
glacial pollen
I believe aac
sudden tree
it'll fetch the audio from yt servers in best possible available quality ( mostl...
i use ytdlp but cmd version and i can only see m4as and mp4s i just grab best m4a tbh
glacial pollen
use -x arg
sudden tree
damn wtf converting it to wav made filesize 10x
analog obsidian
.\yt-dlp.exe -x url
glacial pollen
else, aac or m4a ( still aac I believe. )
sudden tree
yeah -x usually grabs the video tho ngl
glacial pollen
well no
analog obsidian
no
sudden tree
oh sh
if wav is container, how does it make file size increase if bitrate stays exact same
glacial pollen
The thing is
wave pcm is not using any compression
so effectively, whatever would be ( which is not as file comes from lossy compression )
gets ' 0 filled '
that's a thing that has to be done, no other way.
All the missing data is just filled in
sudden tree
ok so i am a noob
glacial pollen
So yea, whatever you have or get from yt-dlp -> wave
that wave after editing / processing -> 32 bit float 44.1khz
sudden tree
and i just did a zoom out on the audacity and the audio is completely peaked out now
tf
oh i zoomed on the db lmfao hahaha
sudden tree
that wave after editing / processing -> 32 bit float 44.1khz
32 bit float wave?
glacial pollen
yes
it's the bit depth
32 bit float is the " target end " for files that rvc processes anyways
sudden tree
and if you are using different songs etc we can import multiple files into the training at different db normalizations?
glacial pollen
- you avoid potential issues during editing
sudden tree
or should each seperate session be normalized to equal level
glacial pollen
Well, the whole dynamics aspect of rvc is a lil skewed up anyways
Biggest issue is, if you mangle with dynamics on your own ( be it rms, peak norm or compression )
it can screw up model's ability to express itself well at high volumes. it'll cause tearing
so at best... if you have to, tame the peaks and maybe add a tiny bit of compression
sudden tree
export audio as mono wav or stereo?
glacial pollen
wave
sudden tree
i forgot to make the audio mono lmfao
glacial pollen
just copy one channel into a blank file
( aka, do not use any " merge channels " algos or such )
sudden tree
?
analog obsidian
export audio as mono wav or stereo?
mono wav
glacial pollen
you wanna delete 1 channel from the audio
either L or R
and just save it as mono wave
analog obsidian
i just downloaded audacity so i dont know how to do that
tracks > mix > convert stereo to mono
glacial pollen
Alternatively, copy / highlight just 1 channel of your choice and paste it over
Cause like, depending on what audacity does
sudden tree
i see
glacial pollen
if it fuses the channels / centers em, it's pretty bad
that's a " merged mono " and not true mono
sudden tree
lyery is just trying to get me to fuse you are just saying delete one to prevent the wrongful merge and distortion
sudden tree
how to delete one track?
analog obsidian
he knows better than me
glacial pollen
Because rather than raw mono, you get a fuse of channels, ish ( as long audacity does that which I am not 100% sure of
sudden tree
yeah how do you do it
glacial pollen
as it does kind of algo magic and averaging of phase and such
glacial pollen
hold on
sudden tree
also clipping every 3 sec with a .3 sec overlap seems like a disaster intuitevely to me idk why
glacial pollen
You'll have uhhh
sudden tree
if you are training on only 3 sec you are guaranteed to get clipping on harmonization it seems like that would mess up the fluency
analog obsidian
if you are training on only 3 sec you are guaranteed to get clipping on harmoniz...
no
glacial pollen
Then select the other one ( which ever you want but I personally use RX and do params measurements on both channels to pick the better one ) and delete it
leaving only just 1 channel ( and so, you have your file mono in the end
analog obsidian
sudden tree
wait couldnt you just pan 100% stereo lmfao
glacial pollen
Cause well ye, you have 2 channels visible
analog obsidian
Then select the other one ( which ever you want but I personally use RX and do p...
im interested in this 🥹
glacial pollen
I mean sure, panning
but it's just 1 click
then x one the other track
Done
That simple
sudden tree
ah nice i figured it out thanks'
glacial pollen
Nice
sudden tree
and you can train multiple files or do you need to merge them into 1 audio file
glacial pollen
no difference really
But best is imo to just use 1 track, 1 file. And do processing on 1 file ( to keep the uniformness
sudden tree
yeah
analog obsidian
i wonder why
uuuh i can't remember the explanation lmao 😭 but dw this method is fine
glacial pollen
also clipping every 3 sec with a .3 sec overlap seems like a disaster intuitevel...
well
the only reason overlap exists is to avoid the discontinuity in " context "
naturally, if you can afford to split it all on your own, properly, you can bypass it
But that's the best we have if it's automation
( I tried various methods to better it, sadly didn't work out well / significantly. Such as envelope or better rms methods )
Tho ye, dw about it. As lyery said, it's alright
sudden tree
i see
thank you
the last question i have is regarding the normalization
like one of my sessions is higher peaks and normalization
just wondering if it will mess up training
should i lower the gain on the loud one
like i just ran a normalization of -10db to match to look of each
glacial pollen
like i just ran a normalization of -10db to match to look of each
tbf, normalization won't help here much either way ( not as much as compression which in the same time can break stuff but whatev. )
sudden tree
i see so dont worry about it
glacial pollen
But then, it all should go well if your audio comes from " same source "
glacial pollen
in that case, you can more or less match the " overal " volume levels
per clip / track
sudden tree
some are way loiuder than others so i am just normalizing until the clips look the same
glacial pollen
doesn't have to be perfect but it'll help
you can just do each track at -3 dB norm
that's because rvc does normalize each anyways ( each cut 3 sec segment )
glacial pollen
but ye, getting em to similar levels is a nice thing to do regardless
sudden tree
is there a way to collapse all tracks in audacity to one continuous singular
glacial pollen
Yes
That's actually the only reason I keep audacity ( and use it just for that lol )
glacial pollen
Polish
analog obsidian
doesn't have to be perfect but it'll help
do i need to use this before exporting one of the stereo tracks as mono?
glacial pollen
Nicee haha
glacial pollen
do i need to use this before exporting one of the stereo tracks as mono?
Oh, well no
the way I do it is copy 1 channel
and paste it into new blank file
length will auto match
glacial pollen
pick one that seems better btw, the channel
analog obsidian
glacial pollen
for instance, one that has better sdr levels or dc or peaks, you know the deal
sudden tree
do you recommend alternativing sides like left-right-left for clips or only using left-left-left
glacial pollen
tracks > align tracks > end to end
After that you export it
and import again
Then you have one continuous file
sudden tree
i see thanks
glacial pollen
do you recommend alternativing sides like left-right-left for clips or only usin...
Not really
as in, doesn't matter tbh
best channel per track
but if it's the same source..
for instance, 1 anime but different episodes
I always want to believe the recording session and so on was set more or less similar
so in that case I do pick the same channel throughout my project ( just in case ✨ ) (( Unless one is explicitly bad or worse
sudden tree
damn the align track thing isnt working ugh!
glacial pollen
how so?
sudden tree
nevermind i figured it out
glacial pollen
a
glacial pollen
no
sudden tree
aight
glacial pollen
At least I never do that
guess you can try one time
but I don't see any point in that personally
sudden tree
shit i just figured out something
glacial pollen
I guess the biggest clue whether you should try that or not is seeing how phase behaves
if channels differ significantly in that aspect, perhaps you could try
sudden tree
if you export as mono and you have left and right tracks it just mutes the right one
hahahaha
ill just pan left on the rights
glacial pollen
oh then problem's solved, if there's no mixing or centering algo involved
sudden tree
haha
glacial pollen
then you good to go
sudden tree
thanks for all the help brother
means a lot
and then what settings for applio you recommend for the preprocess?
manual splitting setttings?
glacial pollen
Given the most propable case for you uhhh, go for default
and do include preprocessing
it's the normalization + butterworth filtering ( 0-57hz iirc )
sudden tree
i see i was just asking bc the laf dude was saying 3sec with .3 sec overlap
glacial pollen
ye, that's the default
glacial pollen
automatic + 3 / 0.3
Simple can work too
but if you truncate stuff, that is
Silence truncation
sudden tree
i see i guess since we already truncated so simple makes sense
i will do
thank you for the help
glacial pollen
It really depends
for instance I used to work with bs 12 / 14 and 16 for most of my above 10 or 13 min sets
yet sometimes that works like crap and 7,8, 9 are safer
As always I recommend bs range finding
train the model at: 4, 8, 12, 16 ( each for 400-500 epochs )
sudden tree
oh wow, I didnt know people did that
glacial pollen
if you're aiming for " perfectionist " model
glacial pollen
well no, people do not do that
but I just recommend that workflow if you're a perfectionist like me lol
sudden tree
yeah i am lmfao
glacial pollen
( tho in reality, both learning rate and batch size should be picked individually )
Oh ye, in that case def go for that
and from there see on graphs + do some inference testing on various epochs
sudden tree
yeah i dont even know how to modify learning rate in applio
glacial pollen
n see which one does the well
from there, you can finetune it even further as in, do -/+ 1 batch from the base batch size ( one that performed the best )
glacial pollen
what's the frequency response of your files?
sudden tree
not sure haha i am super noob
glacial pollen
also how do you change LR in appluo
nah, that thing don't touch was just giving examples
sudden tree
lmfao
glacial pollen
not sure haha i am super noob
You can try spek software
sudden tree
i usually just put 48k because its highest level
glacial pollen
pretty basic but will do
sudden tree
is that bad
glacial pollen
model's sr should be more or less aligned with your files
with some minor exceptions
for example, a deviation of 1-2khz shouldn't hurt or 3
sudden tree
can i use spectrograph on audacity
glacial pollen
For instance, if I have somewhat imperfect audio ( can be compression ) that's ranging anywhere from 41 to 43khz or even 44
sudden tree
its peaking at 19k
glacial pollen
I'll use 48k model ( because those extra 2,3 or 4 khz does mean clarity and fidelity, esp in respiration
glacial pollen
In that case 40khz model ye
sudden tree
i wonder why
glacial pollen
yt should never be used for 48k
glacial pollen
because that's nyquist range
sudden tree
i see haha
sudden tree
damn you have to become a audio expert for this
glacial pollen
for them you do *2 the sr and that's your true sr
sudden tree
so using 48khz was ruining my models possibly?
glacial pollen
Quite possible yes
glacial pollen
because the models are trained on specific frequency ranges ( pretrained models
glacial pollen
they are accustomed to working within a giving frequency spectrum ye
Yup
but it's not a 'hardcoded rule'
sudden tree
damn thanks so i gotta download the legit studio rips to be able to go to the 48khz range
glacial pollen
For instance
sudden tree
or find raw vocals with really good mics
glacial pollen
My Kurisu ( best model I ever made )
was 38-42 ( variable ) sr
yet trained on 48k
Yet she sounds nice, right
sudden tree
yeah
glacial pollen
So there's no strict strict rule, yet it's highly advisable to stick to what I mentioned yup
sudden tree
yeah i wonder if the mismatch causes audio ripping or the glitching noises
glacial pollen
not quite
it primarily affects the model's potential / generalization or generally adapting to your voice ( finetuning potential
sudden tree
damn cant find the custom cutting in applio
sudden tree
it primarily affects the model's potential / generalization or generally adaptin...
that makes sense actually haha
where is audio cutting setting located in training
cant seem to find
glacial pollen
show ss
sudden tree
newest
glacial pollen
hmmm
sudden tree
i saw it a while ago it disappeared for somereason
sudden tree
im just gonna reboot rq
analog obsidian
Ah thats the latest compiled, yes its outdated
sudden tree
so fucking weird i cant find the simple cutting @analog obsidian
analog obsidian
use latest main branch repo
glacial pollen
ah, if it's compiled and not from repo
sudden tree
im using 3.2.8
glacial pollen
then outdated
ye but precompiled / zip packages aren't updated in-line with repo atm
sudden tree
can i just not use simple lmfao
glacial pollen
download the repo and use 3.2.8's env folder
and if that doesn't work, delete the borrowed 3.2.8's env folder and redownload all ( using install-applio .bat file )
Lyery will help you hopefully as I have to get back to my work
knotty moth
im using 3.2.8
without the bugfix is broken
analog obsidian
without the bugfix is broken
tldr of this convo:
we teach him the truncate silence method of slicing
he can't do it because he's using the latest compiled version, which is outdated
just do this and decompress it in your applio folder
don't use run-install.bat
no need to reinstall anything
run applio
knotty moth
or use codename's fork
sudden tree
i mean @analog obsidian i can still use my current version and just use default splitting?
analog obsidian
i mean <@775545133448953856> i can still use my current version and just use def...
no
sudden tree
it seems to have worked
what
fuck i am at epoch 100 alr
why not codename said i could
analog obsidian
if u want to use default splitting then don't use truncate silence
analog obsidian
it shouldnt matter
well the idea of this method is that the chunks are meant to be consistent, rvc old slicing is not consistent so yeah
sudden tree
i see
analog obsidian
does not affect quality
analog obsidian
just means your model is gonna take more epochs
sudden tree
i see but all the 16k splits are legit 3 seconds anyways
analog obsidian
the truncate silence method helps rvc to learn the dataset faster
analog obsidian
there is not an audible difference between this method and the casual old method of automatic slicing anyways
only thing that changes is how fast rvc learns the dataset
sudden tree
i see it just prevents those like 1 second clipped audios?
so the epochs for same audio is lower?
analog obsidian
quality degredation ?
rvc kinda ignores those samples
sudden tree
oh i didnt know that
lmfao
i wonder what would happen if you set the time for each to 10 seconds
analog obsidian
not really ignoring but it separates them from the rest
sudden tree
i used to use like 7 sec samples
analog obsidian
so the model learns the dataset even slower
since it has to learn two things at the same time
instead of 1
analog obsidian
every 3 sec chunk get paired and every 1 sec chunk gets paired
and rvc learns them individually
smth like that
analog obsidian
why dont we just use 7 sec samples or 10 sec
iirc hifigan only accepts 5 sec max
i might be wrong with this tho no idea
analog obsidian
dont worry it will not kill your quality
the model will learn the dataset a bit slower
but thats really it
you can continue using the old slicing method if you wish
sudden tree
ngl you said truncating wouldnt help but my model seems to be improving way more conistently this time
just looking at the loss values in cmd
analog obsidian
yea because like i told you, it learns it faster
sudden tree
i see
analog obsidian
so you notice it sounds good because its learning faster
sudden tree
just removing the silences = less dead space = faster training
makes sense
you are really just maximizing the roi
analog obsidian
u still need silence for training
sudden tree
just not a lot
analog obsidian
thats why the setting set it to kept a bit of it
yuh
- rvc injects 2 silences in your dataset
this is bc you have to teach the model to understand what silence is
at least that was noobies told me
glacial pollen
2 of them is really enough for typical dataset
sudden tree
haha do you think rvmpe is the ultimate development of this technology?
i wonder if rvc can even improve atp
analog obsidian
haha do you think rvmpe is the ultimate development of this technology?
very robust and good
sudden tree
the problem is realtime translation
it still sounds blocky on my end with large chunk size
analog obsidian
well realtime perfomance heavily depends in the dataset
singing datasets are bad for speech
sudden tree
yessir
analog obsidian
while speech datasets are okayish-mid for singing (it depends)
sudden tree
we need to be able to develop some agi ig to make the tech flawless
glacial pollen
while speech datasets are okayish-mid for singing (it depends)
^ There
a colorful and vibrant in emotions and pitch set can sing well
sudden tree
yeah i try speech on juice wrld model and it works since he raps haha
sudden tree
a colorful and vibrant in emotions and pitch set can sing well
lmfao reminds me of alex jones set
sudden tree
what epoch level do yall tend to set the models at
like 300 for 15 mins is peak usually?
analog obsidian
sadly old graphs are not accurate enough to show you which epoch to choose since they only tell you the latest value in that specific epoch
sudden tree
have any of yall looked into onnx model conversion for realtime
apparently you can offload to cpu?
analog obsidian
at least when i tested onnx it degraded my model quality a bit
and also on nvidia is slow af
sudden tree
sadly old graphs are not accurate enough to show you which epoch to choose since...
i just use g/loss ngl
total g loss
then max the smoothing
analog obsidian
yea g/total (from 3.2.8) is outdated by now
analog obsidian
in simple words, it only tells you the latest value of that specific epoch
this means you might have a better value in another epoch and you'll never know
so if ur lowest g/total was 29
u might actually have another low one hidden
the new graphs fixed this
sudden tree
what cant you see on the graph every epoch?
i never had that issue
i can see all of them on the graph
analog obsidian
no like you see that if you hover your mouse in a random point it tells you something like "value: 32,5"
well that value is one of multiple values in each epoch
so 32,5 in the epoch 100 (for example) is just its latest value
sudden tree
damn just no updates to the package installer huh
sudden tree
so just zip the github repo and unzip in the same file and replace all folders?
sudden tree
rip to all my premade model logs lmfao
analog obsidian
F
simple ore
there's some cleanup in progress, so your stuff may break
simple ore
like need to update filelist.txt and replace "mute/v2_" with "mute/"
analog obsidian
oo
simple ore
and need to add soxr using env/python -m pip install soxr
sudden tree
shit cant i just pip install entire project haha
simple ore
waste of time
sudden tree
does anyone else have the issue when converting vocals?
its like the converted file is longer than input causing the vocals to be off beat
analog obsidian
its like the converted file is longer than input causing the vocals to be off be...
enable split audio option in applio's inference
sudden tree
wait is g/loss/total reliable
apparently the smoothed one starts consistently increasing at 14k steps
but the higher steps still sounds vastly better
like 20k even sounds amazing
it really sounds more real
analog obsidian
wait is g/loss/total reliable
g/total is the average of mel, fm and kl
analog obsidian
is mel better?
mel is the clarity of your model, this metric always improve the longer you train so this is why you feel it sounds more real
sudden tree
hmmmm didnt know that
analog obsidian
but your g/total graph stopped improving the moment it started rising
sudden tree
what would you do w this
analog obsidian
what would you do w this
i choose generalization over everything (g/total) so i always use the lowest point in g/total before overtraining
sudden tree
is generalization just the ability to be applied in any scenario?
analog obsidian
is generalization just the ability to be applied in any scenario?
generalization is the ability to the model of generating new audio
overtrained epochs have distorted frequencies and other bad stuff
sudden tree
because it seriously sounds much much more accurate to juice wrld at even 30k steps
even though it started rising at 14k
analog obsidian
yuh because mel is still improving
so the spectogram is clearer
which gives the feel the model is more realistic
sudden tree
i see, why does everyone use generalization when clarity is much more important for realism
knotty moth
mel is more likely to keep going down even more than 1k epochs
sudden tree
so generalization = more flexibility
analog obsidian
i see, why does everyone use generalization when clarity is much more important ...
because a model that can generalize well sounds good no matter the audio u give to it
sudden tree
i see that makes sense
analog obsidian
but anyways like i said before the old graphs only logs the last step of the epoch, we can't tell if your model is actually overtrained or not since the graph is innacurate
sudden tree
so at .999 smooth the graph start increasing at 240 epochs but the lowest recorded loss was at 270
analog obsidian
the new ones are like this
sudden tree
should i use 280 epoch or 240 as the base
knotty moth
wait is g/loss/total reliable
the applio main branch and codename's fork use more accurate loss values which are average of each epoch instead of the epoch's last step (since the mainline rvc versions)
analog obsidian
the new ones are like this
and i can confirm every epoch in the rising zone sounds like shit
sudden tree
the applio main branch and codename's fork use more accurate loss values which a...
so if i update will i be able to see the averages or will need to retrain?
analog obsidian
i believe your model is not overtrained and the g/total is just fluctuating
but we will never know
the log is innacurate
sudden tree
oh really?
sudden tree
wow so i should just take whichever sounds best then since its inaccurate
knotty moth
so if i update will i be able to see the averages or will need to retrain?
you cant change the logged values unless you start over training
sudden tree
makes sense
analog obsidian
i remember back then i used to choose an epoch based in the mel graph
i felt it was a bit more reliable than g/total
sudden tree
yeah it is honestly sounding like 30k steps is the best
500 epochs sounds a bit overtrained ngl
idk tho
it would theoretically make sense to choose the mel graph if converting rap vocals to rap vocals imo
analog obsidian
a rap model will always be good at inferencing rap songs regardless
its literally whats made for
sudden tree
true haha
yeah i am getting robotic noises at 30k steps
thats overtraining correct?
analog obsidian
yeah
robotic sounds happen when the model is overtrained
as long the model is not robotic, its fine
sudden tree
imma just download a conversion and compare
haha
ill update because the new g total is accurate right?
analog obsidian
overtraining is pretty easy to spot, literally if it sounds robotic, its overtrained
analog obsidian
ill update because the new g total is accurate right?
yeah new g/total is accurate
every new graph is reliable now
u can trust them
astral jungle
-train