#Every sample/character i train in the end the outcome is always the same
1 messages · Page 1 of 1 (latest)
are you putting the dataset in the right spot?
it has to be the path of the folder where the dataset is
i created a folder inside of the rvc stuff and put my dataset in the folder i created and made sure to enter the right path of my folder
Ayo? @summer loom level 1 !!! 
I dont recommend doing that. Try putting the models you wanna make seperate from the Mangio folder
also what is the file types that you are using for your dataset
ive had a similar issue and try to not put anything i work on inside the folder
alr , will do, i use .wav files for the samples
now that i'm trying to use it on a different path outside of mangio it gives me this error
tried to use different paths , everywhere but still no success , it only let me train if i have my samples inside of mangio
full message of the error
I see the problem
The last file at the end of the path is a log
Don’t use logs
That’s used for the index which you have none yet
how do i not use logs , i'm kinda new to this
The logs and weights folder should only have something in it after you’ve trained a model
You should copy the path of the folder with your dataset in it
that's what i'm doing
Ayo? @summer loom level 2 !!! 
Are you able to join a vc so I can show you
i can but i cant listen or talk
Are you able to stream your screen
yes
thats ok
better if you can talk in here
so i will explain you where my file and everything is
open up your mangio rvc folder
ok
ok
what now
you saw nothing
show me the folder of your dataset
now copy the path of the folder
put it on v2
paste the path in the the enter the path of the training folder
process data
i guess so
once it says end process you can go to the next step
feature extraction
now i press feature extraction right
yes
wait for it to load
while that loads, open up task manager
performance
gpu 0
scroll down a bit
ok
you are going to train on a batchsize of 4
save frequency of 20 or 25
your choice
i normally do 50 epoch because my sample is over 30 seconds
it will take alot with 400 epochs
my graphic card not the best
itll take a bout 15 - 20 minutes
once it is done, you can always check the different evolutions of it
well ty for the help atleast i fixed that error
but idk if you wanna wait here
it will take some time
if it sounds bad at 400, you can listen to it on model inference with 300 and so on
i thought that if the sample was short i should use low amount?
because there isn't much to train
the main things that determine a good model is epochs, quality of the dataset, and how much you have
yup but if i use to much epoch it can sound robotic too yea?
best to go for a long time and see the best out of the many epochs you trained on
the first is always the longest
the reason i'm not using headphones is because i was heading to sleep xD
i understand
kinda weird how its taking longer now
its gonna take a bit
before i would do it and it would start training immediatly
maybe because something was wrong
maybe
now all i gotta do is pray that the voice doesn't sound the same
fr
like the audio i sent before
is the audio you used cleaned up and without noise
there is some minor noise/echo
Ayo? @summer loom level 3 !!! 
then it might not turn out good tbh
yea , i was just testing
to see if it really worked
and if it did then i would redo it again but cleaner
the thing is if the model turns out nothing like the voice you want it to, then it is probably the pytorch version you have
Ayo? @vocal dagger level 32 !!! 
it's a processing thing for python
hmm i see
just keep that prompt open
alr
any errors you see will be in there
training takes time
well i gotta go sleep
this will take alot of time
and i dont wanna make y ou wait
and i also wanna sleep xD
when i wake up and if it finished i will ping you , if you dont mind
plz do
and then i'l also test if the voice sounds the same
sounds good
alr man thanks alot for the help
ofc
i'l be heading to bed rn
gn
Gm for you and Gn for me xD
@vocal dagger ok it finished
i tried out the voice model
it turned out the same voice
its the same voice as the previous audio i sent
idk what to do on how to fix it now
@vocal dagger
i did try that before opening this help forum but i will try fixes rn
when you open fixes, double click on LOCAL_CREPE_FIX.bat
see if that gives you any errors
@vocal dagger
it also downloaded a bunch of stuff the first time i opened
or updated
i cant do what it tells me to do for some reason
do you have python downloaded
if not, download the latest version
this is a new problem so i am trying to see what can be the problem
yup i do
3.10
3.10
alr
then put this in your console
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
all of it
do i select the last option , (last ime i had to because it wasn't detecting python or some kind of error but when i did it it fixed)
do i disable this
Ayo? @summer loom level 4 !!! 
or can i keep it
disable it
done
ok
yea it dint do anything now
which one
for the pip one, wait for it to get to the root command
oh it started doing some stuff
alr
ok it only shows me to do this
uh
i did like it said but dint go so well again
@vocal dagger
do the whole line
ah alr
alr
yea this is pretty annoying
run the command in a fresh command prompts instead of fixes
alr
tbh mangio rvc is pretty old, try the newest ver
it worked
good
i thought i had the newest one?
its not
newest is 1008 for mainline
or 1228 patch
or applio 3.0.5
where can i get the newest version ?
hes using mangio fork
there's other versions of mangio rvc?
does this version have rmvpe
yes
yes
I still use it
good
i make my models with it
then no need to update
mb bro i only use mainline💀
idk what u mean by that
it's a different program for training models
like the main rvc released on github
is it normal for these 2 messages showing up everytime i did the fix , yes it showed before too
the one i'm using i downloaded it from github too
yeah it's normal
i think...
anything in red bad
yeah every rvc here is open source
alr
everything good then
let it close out then
should i try opening and do the voice then?
it should do so automatically
yup it did
i will do it at 50 epoch , may sound kinda eh but i just wanna see if the voice will change and stop giving me the same voice
Gtx 1650
aight
keep it at low batch size
i'l warn you about how it goes
coz only 4gb vram
yeah
if i have 4 what is a good amount to use ?
Ayo? @summer loom level 5 !!! 
the amount of VRAM you have
keep it at 4
alr
@vocal dagger it dint work 😭
the voice is the same
broooooooo
what the hell is this curse
it keeps giving me that damn voice its always the same
I might have to get some more minds on this problem
like it doesn't matter what sample i use its always this voice
i can use audio samples of vtubers like gura , or youtubers , girls , everything
it will ALWAYS sound like that voice
it doesn't let me do other voice
whack
this is very weird , before i used the google one where i could train , i think it was google lab or something , it worked just fine but then they couldn't keep that up so it got discontinued i think so i changed to this one and now it gives me this problem
cuz it is run on your computer
yea makes sence
but i did everything right
this is not supposed to be happening
this is so weird
ur obviously just not training the model here
this is all based off of the pretrains
give me a quick run down of what u do exactly to start training
I walked him through the model making process in mangio and it still turns out like this
let me see
first and foremost
show me ur settings on the last step
@vocal dagger they seem to have ghosted me
theyll be back
when
Grub this is single-handedly the most like
I know nothing on helping that you can do
Why are you advising them to install 3.12, when everything should be 3.11 which most these projects call for
And on top of that, the zip already has python and torch in it
Ignoring how you also told them to install torch, globally, instead of in a containerized environment
Like bruh
All they gotta do is unzip and run
If that doesn’t work, it’s either on a different drive from the C drive
In a one drive folder, or the extraction failed
Let me go to pc and ill reread
i even looked through the files that he had
this right here, is, he is skipping the first step, or, the first step is failing or putting it somewhere else
sorry for the late response but you mean settings on what?
yeah can u give me a run down of how u would go on about training a model with screenshots
are you selecting a voice?
yes
ok so heres whats gonna happen, ill walk you through step by step
first i put the name of the voice and out this settings
then i precess the data with the right path
just press feature extraction
so, can you show me this logs folder
alr
could be processing the same thing because it has the same name
what is in 0_gt_wavs
alright, so preprocess is working
so, when you are inferring
are you selecting the voice?
yes
for curiosity, what audio format is the dataset in
i even have a separate application that is solely to use ai voices and its still the same voice on it
.wav
ok
can you send the model and part of the dataset?
just the model.pt
and some example of like, what it should sound like
the one in weights?
like a 5 second clip
yeah
you can upload it somewhere and link me it in dms if you dont want to share it
you want me to send it?
alr
sending it
i will give you the sample
how its supposed to sound like
ok 1s
and its not only with this sample
i also tried other samples but it turns out the same voice
and im going to assume that is correct
nope that is the same voice but higher pitch
so this is part of the dataset?
yes the sample i used to train
by dataset you mean sample right
or audio
yeah
so, my main recommendation is, i would try to get a little more audio, and try to keep it consistent
cause there are parts in that where i can hear like the beeping, where he is close to the mic, far from the mic
inconsistencies with rvc typically cause it to like, do the stupid "make a model sound like robotic hiss" issue
yea ik what you mean but i tried with other samples and it turned how the same
is this a youtuber or
its my friend
if you want to make a model of ur friend i recommend just like, having them record their mic directly in audacity
If its ok by you , what if you try to use that sample and train it on your pc? just to see if it turns out the same for you , if you dont mind
1s
alr
like
at the top you can find it
batch size per Gpu?
i used 4
ok
50 epoch
i also used 400 epoch before
but the one i sent you is 50
so better do it with 50
yeah its not gon be long
damn your gpu is way better than mine then
not really
how is it that fast xD
like, its better, but, its not a god tier one
for me 50 epoch take like maybe 30min or something
because training speed is dependent on dataset and the gpu speed
so like
1 epoch of 60 minutes of data, should be roughly the same speed as like, 60 epochs of 1 minute
oh damn
asking because i'm curious
i can also train in an pc without gpu?
i forgot if you can or cannot
can but will be slow
i see
gpus are used because its a bunch of tiny calculations
because i can just try do it on my second pc
nut would take some tim , my second pc isn't that good
if that works for you i will shocked lol
1650's may not seem super fast, and are not really
so, how long was the dataset for yours that you trained
that was the whole dataset?
yes
weird cause i did it and it sounds diff from the model u gave me, like
it sounds like i made a 50 epoch, and you made higher
ill run it for a bit longer and see if same
i used from the mangio that i used to train
but i also tried with other application i have solely to use ai vocals
Ayo? @summer loom level 6 !!! 
this is what i get
using this, with the 50 epoch
i can see a resemblance
it worked
pretty sure
yea pretty sure that one is right
i think primarily that it is working, but there seems to be an issue with how you are inferring with it i think
my personal recommendation, is using regular rvc https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/releases/tag/updated1006v2
完整包 Complete package
For Nvidia GPU users:
https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/RVC1006Nvidia.7z
For AMD/Intel GPU users:
https://huggingface.co/lj1995/VoiceConversionWeb...
i dont see how it is working i used other samples not from the same voice to see if it would lead to the same result and it did end up being the same result
i was only using this one because it was the only one i knew about
yeah that is the most up to date simple one
is this one also used locally?
the UI is ugly, but, that is old reliable
mangio hasn't been updated since like, august of last year
i see
also
could you possible ssend me the pth file of the voice you made just now using the sample i gave you?
i wanna mess around with it
because i think it worked
also does it have an tutorial on how to use/install this? because i dont wanna screw around and mess up stuff lol
you download the zip if you have nvidia or amd/intel (tells you which one)
you unzip it, onto your c drive
then you just run the go-web.bat
like this?
dont move it out of the folder, you can create a shortcut, then move the shortcut
but go-web has to be in the folder
alr
as for guides, we got plenty in #1159513888199540817
😭
i extracted it to my C drive and then press go web
and it gave me this message
are you missing the runtime folder as shown here?
because the zip you download, should be like 5 GB
and you unzip the entire thing, and run
tensorboard
ignore the audios thing, and tensorboard
no i dint run
yup
so what does the folder look like
i linked you to releases
dint download the one from my graphics card
yeah source code is just, only the source code
it does not come with all the dependencies in it
yea i was dumb
i just use the one i linked you
oh i see
well as long it fixes my problem i will be happy 😭
i'm extracting it
65%
well what i was implying earlier is, the model will sound off if you dont clean the dataset
and make sure the voice is consistent
yup i'm aware i was only using that sample more of a test to see if it actually worked
will do thanks for the tips
but why is wrong to have pauses in the sample?
pauses aint bad, what im saying is the voice isn't consistent
theres parts where he talks close, then far like across the room
theres the beep in the background in the beginning
stuff like that is what causes it to mess up
ah yes i see
and did you run it through UVR
the reason for that is because we were playing lethal company XD
ye
yup but idk good settings for it so i did my best XD
if ur gonna do that, id recommend just, find something to record one audio program
and record only discord
and i could tell because you dont got denoise on in uvr
i see , that's a good idea
actually no , because we were talking inside of the game xD
if you open a spectrogram of the sample, you will see there is a big streak in the high frequencies, and that is typical /w UVR
yeah im saying like
u want to sound like him, but in game
if u make a regular model, and talk in game, its gonna add that sound
oh i see
if you make a model of him in lethal company, right, its gonna be impossible to make it consistently the same
because if he turns, or you turn
the audio its distorted by the positional audio
in wich option can i find that
i'm a beginner when it comes to this stuff xD
under the adv settings in UVR
ik it exists in one of the menus, i just dont really use UVR that much
may this be the one?
i can press to press data ? @grizzled drift
i just wanna make sure
i would click v1, then v2 again
and it will give you a 32k option for sample rate
i recommend using 32k for most people unless your source audio strictly has frequencies higher than 16 kHz
consistently at least
no use
its buggy af iirc
idk if they ever fixed it
its better to just, do each step, make sure it doesnt error out, then let it train
vs hope one click does all the steps in order and doesnt mess up, but, last time i used it, it worked like, 20% of the time
alr ty
i'm training it rn and then i will ping you when its done if you dont mind
also could you send me the pth file of the voice you made?
from the sample i gave you
i alr deleted it, sorry g
ik its ur bud, i just deleted it all cause, me no need archival of some persons friend lol
@grizzled drift IT WORKED
IT FRICKING WORKJED
FINALLY
THANK YOU
Thank you for everyone's help
U r right usually it has to do with some missing file. It would be best if the person ditched mangio fork for now as it's outdated
But hey good stuff

im not trying to be elitiest, that is not goal, just
ik i got super offensive and emotional, but
i hate when people say "go install python"
and even moreso when they say "go install pytorch globally"
I am not even bringig that up heh
I'm just saying
U did a good job explaining and resolving
Yup i was looking to fix the problem but if installing other program is the solution i'm also happy with that i had no idea there was other way of doing ai voices so thank you for that ! 🙂
Ayo? @summer loom level 7 !!! 
