only train button
#✨│ai-help
covert anchor
rare gobletBOT
Ayo? @covert anchor level 3 !!! 
knotty moth
sorry no idea cuz I'm not using the latest applio
covert anchor
oh :c
knotty moth
btw check the model name, should be only alphanumeric without spaces as safe bet
covert anchor
well, dw, you tried and i apreciate that
this is the name
it's a champion from league, on spanish
the name of the dataset, "mezcla" means mix
and that's all
probably im just gonna wait untill tomorrow and get help of a friend that uses the latest version of applio
anyways, thx!
covert anchor
slice the audio. You cant train with one 9 min file unsliced.
what, why?
simple ore
because that's how it works. Either you slice the file youself in using 3-5s slices, or you let Applio slice it in preprocess
flint solar
does anybody knows how to use kaggle?
there is a guide
boreal sluice
https://rentry.co/RVC-Mainline-Kaggle
yhanks so muchh
rare gobletBOT
Ayo? @boreal sluice level 1 !!!
dusty bone
whenever i use any voice it rlly just cracks alot lol
is there a fix or is my mic the problem
knotty moth
whenever i use any voice it rlly just cracks alot lol
try not to infer doubled vocals
odd shale
Crash alot
I would suggest Deiteris' w-okada fork
odd shale
Tried voice.ai
It's kinda better than the OG version.
And if you ask, nope, we can't give support nor troubleshooting for voice.ai issues
odd shale
whenever i use any voice it rlly just cracks alot lol
The source of the problem can be either 2 of these things: Or the model you're using wasn't properly made/trained or it simply doesn't fit your voice.
knotty moth
Crash alot
odd shale
Ty
You're welcome.
Tiny fact: The models made with tiny/moderated-size dataset prolly won't perform properly on W-Okada nor Voice.ai
Tho there can be tiny/lucky exceptions
gritty zinc
Make sense
rare gobletBOT
Ayo? @gritty zinc level 1 !!!
gritty zinc
So this One work on everything?
odd shale
So this One work on everything?
Yep, it should.
Deiteris fork is kinda more optimized than others.
Also it's matter of playing around with settings and reading the guide.
odd shale
hm
Read the docs i gave you above.
I'm not sure which version of W-Okada you're using
gritty zinc
oh yh ty
pure pecan
I have f5-tts installed, how do I put weighted voices on it? it only seems to let me upload audio samples, and not .zip of weighted
dusty bone
The source of the problem can be either 2 of these things: Or the model you're u...
thank u! Im guessing it must be voice
i do have a deeper voice
soo ill find smth that fits more
low shard
I have f5-tts installed, how do I put weighted voices on it? it only seems to le...
F5 tts is only 0shot
pure pecan
F5 tts is only 0shot
Sorry, what is 0shot?
low shard
Sorry, what is 0shot?
0shot = no actual training needed, just an audio file to work, inferior in terms of quality to few-shots
few-shots = training needed, an example is GPT-SoVITS (TTS) & RVC (STS)
You can't upload RVC models to GPT-SoVITS nor F5 TTS
pure pecan
huh, could have sworn I installed a version of f5-tts that could do training based on the github tutorial, oh well, thanks fo the information
low shard
RVC models can be used only in programs that use RVC, such as W-okada
low shard
huh, could have sworn I installed a version of f5-tts that could do training bas...
The training is related to the model who does 0shot, not actual single model training
gritty zinc
hm
low shard
You can train the big model that's being used for 0shot, not train every voice you want to a model in F5 tts
gritty zinc
You can train the big model that's being used for 0shot, not train every voice y...
can i smh link it to disc?
low shard
can i smh link it to disc?
You mean as a realtime voice changer for calls? you could technically do that for any TTS
But it's not really realtime
You'd have to type the words and let the audio play so
gritty zinc
eh beter than nothing
low shard
eh beter than nothing
There's actually Wokada, which uses RVC models in realtime
So it's Speech To Speech, rather than Text To Speech
What's your pc gpu?
gritty zinc
this is too complicated for my dump ass i just want act in vc for online class
thought it would be good idea
prob no
low shard
this is too complicated for my dump ass i just want act in vc for online class
It's AI
Open Source AI
low shard
prob no
There's guides
Leo gave you the wokada guide above
which has everything you need to know
gritty zinc
yes buttt
rare gobletBOT
Ayo? @gritty zinc level 2 !!!
gritty zinc
it kinda NOT what im looking for excatly
low shard
it kinda NOT what im looking for excatly
what are u looking for exactly
low shard
smt like the text speech thing
Yea you can install any TTS in the guide and use it for calls
Really depends tho
If you want generic voices and easy: edge tts
If you want custom voices and easy: F5 TTS or Fish Speech
If you want custom voices and best quality (but more complex): gpt-sovits
the process for using each of those in calls is kind of the same
Also, depends if your pc gpu is good enough tho
gritty zinc
alr tyy
true ravine
pls help
rare gobletBOT
Ayo? @true ravine level 1 !!!
pure pecan
ok I downloaded RVC off of pinokio, what tab and where do I put the weighted file?
rare gobletBOT
Ayo? @pure pecan level 1 !!!
pure pecan
wait, is RVC only for singing, I thought it was a text to speech
timid valve
pure pecan
well I founded a folder called weights, and put my model in there, but when im in the rvc ui, I don't see any option to select it.
pure pecan
nvm figrued it out, had to take the pth file out
just wish i could use tts instead of voice...
I want to script something and have it read it all out
blah i just keep getting errors trying to convert voice, oh well, thanks anyway
rare gobletBOT
Ayo? @pure pecan level 2 !!!
pure pecan
sounds awful with my voice though
brittle wing
what am i doing wrong
What is that
pure pecan
What is that
rvc
brittle wing
Colab or local
pure pecan
local
crude flame
How to hide app?
windows button + tab then create a new desktop and use that
pure pecan
maybe the model was just bad, tried another one and it's really good (1000 epocs hatsune miku)
seems to not work with recordings that are longer than a minute
and rmvpe doesnt work at all
yeah i think im gonna need some 1 on 1 help in call or something i just dont get what im doing
neon hemlock
hirari
pure pecan
yeah?
blazing crane
Need help
No errors, just the RVC bugs out and stops whenever i try to train
INFO:torch.nn.parallel.distributed:Reducer buckets have been rebuilt in this iteration
rare gobletBOT
Ayo? @blazing crane level 1 !!!
blazing crane
Using Ov2 pretrain at 40k
brittle wing
Last time, which one is the best alternative for denoising?
brittle wing
-colab
brittle wing
Last time, which one is the best alternative for denoising?
@crude flame
crude flame
<@673327878288703519>
2
brittle wing
i thought it was several images
Which one to be exact
It's a collage...
Yh
Just asking
crude flame
for colab melband it is the second one for mvsep you can use either standard or aggressive for denoise by aufr33 and xminus you can use either of the melband
brittle wing
for colab melband it is the second one for mvsep you can use either standard or ...
What's the best option among all MelBand Roformers
Cause results are different everywhere
crude flame
the first two are the same (mvsep and colab) and idk about xminus
brittle wing
the first two are the same (mvsep and colab) and idk about xminus
They're the same?
crude flame
just use the one that sounds best
brittle wing
Which one is it
brittle wing
the first two are the same (mvsep and colab) and idk about xminus
Number 1 or two in the Colab
crude flame
2
brittle wing
I don't wanna waste time or GPU
crude flame
yes
brittle wing
With high Chunk size and overlap 16?
crude flame
thats fine
covert anchor
dont uncheck the box in preprocess
tsm it worked ❤️
blazing crane
could i get some help please?
marsh schooner
what do u think about this?
simple ore
1hr+ - 8
<1hr - 4
10 hr+ - 16
something like that
more data makes a batch more uniform
simple ore
so if its less than 1 hour 4 right?
it really depends on the variety of the dataset, if there's a different content using larger batch kinda levels all outliers
with smaller batch those outliers have more noticeable effeect
marsh schooner
it really depends on the variety of the dataset, if there's a different content ...
what if there was like a 2-8 hour data set i would still use 8 right?
simple ore
sure you can try batch 8 for 30m+
marsh schooner
sure you can try batch 8 for 30m+
do they affect quality at all?
or is it just the time
simple ore
it does affect the training result
the model is trying to find optimal parameters to generate a voice, with different batches it may overshoot or undershoot the target or circle around local minima
marsh schooner
oh dang should i try to reatrain then bc yesterday i made like a 22 min data set on 8 batch size and it doesnt sound like the voice i was making but it sounds realistic
covert anchor
there's any way i can improve the speed on the training?
simple ore
there's any way i can improve the speed on the training?
what's your gpu and dataset size?
covert anchor
i should be training at 2 or 3 secs on every epoch, or at least i think i should
covert anchor
what's your gpu and dataset size?
rtx 4060
marsh schooner
https://pytorch-optimizers.readthedocs.io/en/latest/visualization/
no clue on wwhat these pics r
simple ore
it was training 2-3sec/epoch because you had an empty set lol
covert anchor
it was training 2-3sec/epoch because you had an empty set lol
empty set?
simple ore
with 2 mute files and discarded 9 min audio
simple ore
no clue on wwhat these pics r
it is an example how the model looks for the optimal solution (the peak in the middle) doing small incremental steps towards it
covert anchor
with 2 mute files and discarded 9 min audio
discarded? im not understanding, idk if it was for me
simple ore
discarded? im not understanding, idk if it was for me
when you tried to use unsliced 9 min file, it was not used at all
but now that you've sliced it the training actually uses it and for 9 min file and 4060 52/s epoch is a good speed
marsh schooner
it is an example how the model looks for the optimal solution (the peak in the m...
this better?
covert anchor
oh
rare gobletBOT
Ayo? @marsh schooner level 4 !!!
marsh schooner
the goal is to get it as far down as possible right?
the further down means more realism most likely
and is it fine going backwards like this?
modern surge
-rvc
AI HUB Docs
modern surge
-colab
simple ore
and is it fine going backwards like this?
that's restarting from some previous saved set of weights
simple ore
the further down means more realism most likely
g total is a summof feature match (most important), mel loss (medium), and kl
mel loss is how close the spectrogram of the generated audio is close to the original
less is better
fm is more complex measure, generally it should go down or stay somewhat stable
marsh schooner
g total is a summof feature match (most important), mel loss (medium), and kl
so if i trained something and its too high from my liking is there something i can change in my training settings to get it to sound closer to the original
sharp rampart
hey does rvc work in ableton or any daw for real time or no
like can you talk through rvc and get the outputs /inputs into your daw like fl studio, ableton, etc in realtime with no delayt
or is that simply not possible
marsh schooner
like can you talk through rvc and get the outputs /inputs into your daw like fl ...
do u mean rvc okada aka the voice changere
sharp rampart
this
rare gobletBOT
Ayo? @sharp rampart level 1 !!!
sharp rampart
or do i have the wrong thing to do it
marsh schooner
that i have no clue i use the voice changer i thought that was just text to speech though
marsh schooner
oh thats probably what thats for then
theres a voice changer for real time called okada but it impossible not to have delay but u could still have full convos with ppl using it and u dont need a crazy gpu
i run mine using my 3060 ti
sharp rampart
i need it to have no delay so i can make music with it
thats not possible yet is it?
marsh schooner
i have no clue i dont do music ai
marsh schooner
so if i trained something and its too high from my liking is there something i c...
bc im trying to get my voice to sound closer to the original is there a setting that helps that or no?
unique rock
For how many minutes of dataset can I use this batch size without the colab disconnecting and for how many times so that the model does not sound bad?
marsh schooner
isnt this slow for a 4090 22-25 min dataset
steady geyser
Where do i get a virtual mic
simple ore
isnt this slow for a 4090 22-25 min dataset
33s per 25 min set, about right?
chilly ridge
Hello everyone! I would need the advice of someone who knows about RVC please, I can't find what I'm looking for on the forums or it's old (or I don't know how to search). I'm looking to train an rvc model, the problem is that I only have 2min of audio in wav files of about 2s. So I've tried 50, 100, 300 Epoch, changing the batch size... I still have a "robotic" voice and all the Ss or CHs are "metallic". Is it possible to train my model with all these wav files, or is it better to increase the size of my dataset? All my audio files are studio quality, without reverb ect ect... Thank you!
I run it locally on a quadro RTX 4000 Ada 8Go
analog obsidian
Hello everyone! I would need the advice of someone who knows about RVC please, I...
increase the size of the dataset that fixes the robotic S, ch
it does not really fix them but makes them appear way less
kindred swallow
"frequent errors occured" anyone know how to fix
chilly ridge
increase the size of the dataset that fixes the robotic S, ch
Do you think it's better to concatenate all my files in one 2min wav file ?
analog obsidian
Do you think it's better to concatenate all my files in one 2min wav file ?
never let rvc to slice full wav files for you, is bad
ideally every sample should be 3s-5s long
use audacity audio labelling
chilly ridge
All my files is around 2s long, it's better to increase manually the duration ?
I can make a python script for that, 3s of full audio, without silence
3s average
analog obsidian
I can make a python script for that, 3s of full audio, without silence
is not going to help because your dataset is just 2 minutes of audio
chilly ridge
Okay thx
analog obsidian
10 minutes is bare minimum for okay perfomance
models hit the "realistic" tone at around 40 minutes
chilly ridge
I have a another model with 8min of audio also in 2s wave file, but it's a really "specific" voice, it's really glitched when i use it. (It's Pat from Mickey in french), even if i try to sound like the original voice to help rvc it's the same. But if i set the pitch to -12 it's super clean, but, tooooooo deep
chilly ridge
10 minutes is bare minimum for okay perfomance
models hit the "realistic" tone a...
I'll try 10min audio
oak inlet
yii
analog obsidian
I have a another model with 8min of audio also in 2s wave file, but it's a reall...
use batch size 4
train for not longer than 200 epochs
chilly ridge
use batch size 4
train for not longer than 200 epochs
And increase the dataset size ?
rare gobletBOT
Ayo? @chilly ridge level 2 !!!
analog obsidian
And increase the dataset size ?
yes also when a voice does not match the original is due to:
undertraining (most common)
dataset has too much variety, for example a person using two "voices" at the same time
oak inlet
yes also when a voice does not match the original is due to:
undertraining (most...
does uvr5 also give both
vocal and audio
analog obsidian
does uvr5 also give both
for separating vocals? yea, use mvsep bs roformer
or mel roformer kim
oak inlet
for separating vocals? yea, use mvsep bs roformer
alr
oak inlet
or mel roformer kim
which do u prefer personally
rare gobletBOT
Ayo? @oak inlet level 1 !!!
chilly ridge
yes also when a voice does not match the original is due to:
undertraining (most...
Audio came from kingdom hearts game, so, i have file with just him yelling. Does i need to use these kind of files ?
analog obsidian
which do u prefer personally
i prefer mel roformer, bs roformer is known to have muddy instrumentals and vocals
oak inlet
oh
analog obsidian
Audio came from kingdom hearts game, so, i have file with just him yelling. Does...
avoid yelling and laughing audios
you can keep a few laughs
but be sure to not add too much of them
might confuse ai believing thats how the person sounds
oak inlet
how do i add ffprobe and ffmpeg to root. Im kinda restarted
chilly ridge
might confuse ai believing thats how the person sounds
Thx for all these advice ! I'll try theme all
analog obsidian
how do i add ffprobe and ffmpeg to root. Im kinda restarted
are u on linux? sorry idk
analog obsidian
if ur new to rvc check this https://docs.ai-hub.wtf/
oak inlet
ok
simple ore
Hello everyone! I would need the advice of someone who knows about RVC please, I...
sibilants are made out of noise "columns" that is shifted in frequency
so RVC needs about 5-10k attempts to make a proper 'S' or 'Ch' or similar sounds out of pure noise. That is why undertrained model produces metallic S.
chilly ridge
sibilants are made out of noise "columns" that is shifted in frequency
Can you explain please ?
Ho i see
analog obsidian
Can you explain please ?
too small dataset causes them to appear often due to lack of data
simple ore
with a too small dataset you need to train for like 2000+ epochs
but chances are that while it may fix S sounds it may ruin voiced parts (those wavy lines)
chilly ridge
I'll use audio file from all kingdom hearts game instead of 1 to increase the dataset size
I think i can reach 8-10min
chilly ridge
with a too small dataset you need to train for like 2000+ epochs
That's a lot, even with my gpu
simple ore
the example aboive took 5000 loops
chilly ridge
The image ?
simple ore
analog obsidian
I think i can reach 8-10min
it will not fully remove them but they're still going to appear moderately
he's trying to explain you that the more steps you train, they appear less
xD
simple ore
not, they dont appear less, they take a proper shape
simple ore
the noise column shifts into the right frequency range
so instead of metallic z it is a proper hissing s
analog obsidian
not, they dont appear less, they take a proper shape
well this is a better way to say it lol
analog obsidian
Ok so, with 10min dataset. 3 batch size and 2000epoch ? 🙃
do 4 and no more than 200 epochs
analog obsidian
simple ore
it would depend on how many times there's S in the dataset and if you're lucky enough for the training to hit that S slice of audio
knotty moth
10 hr+ - 16
I dont think even 10h+ dataset from 3k voice recordings for a specific person's voice is necessary like this lol https://www.techspot.com/news/105764-panasonic-resurrects-long-dead-founder-ai-share-management.html
simple ore
anyway, with a pretrain the requirements are not that high.. i've been testing it from scratch
simple ore
I dont think even 10h+ dataset from 3k voice recordings for a specific person's ...
with a good dataset 10-20 epochs with a pretrain get you a recognizable voice of a specific person
nowhere perfect, but like 50% of the work is done there
after that it is just small touches here and there that slowly shape up the voice
but anyway, rvc is not a real voice clone, there are very important characteristics it can not reproduce such as peronal inter-phoneme microdelays and mannerisms
knotty moth
i should be training at 2 or 3 secs on every epoch, or at least i think i should
~30-40 sec per epoch for 10 min dataset & batch size 8 should be normal
dull ledge
@analog obsidian Hi Lyery, are you there?
analog obsidian
<@775545133448953856> Hi Lyery, are you there?
dull ledge
You have a really cool Goth Mommy model, unfortunately the link is not working, did you delete it? If not, mind to share it with me if its still public? Thank you 😄
The link is this one: https://voice-models.com/model/1ucea3z45g5
analog obsidian
You have a really cool Goth Mommy model, unfortunately the link is not working, ...
i no longer have it sry, lost it when i upgraded my ssd
dull ledge
Was the best one I heard 🤭 . No worries, thank you
azure marshBOT
i got an error
oak inlet
File "C:\Users\vedant\Desktop\Retrieval-based-Voice-Conversion-WebUI-main\infer\modules\vc\modules.py", line 172, in vc_single
self.hubert_model = load_hubert(self.config)
File "C:\Users\vedant\Desktop\Retrieval-based-Voice-Conversion-WebUI-main\infer\modules\vc\utils.py", line 23, in load_hubert
models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
File "C:\Users\vedant\AppData\Local\Programs\Python\Python310\lib\site-packages\fairseq\checkpoint_utils.py", line 423, in load_model_ensemble_and_task
raise IOError("Model file not found: {}".format(filename))
OSError: Model file not found: assets/hubert/hubert_base.pt
!howtoask
patent trellis
azure marshBOT
magic badge
wet fable
hi there, i'd like to ask whether RVC is better than weights.gg
RVC needs lots of time to generate, while weights.gg not
idk if there's differences. Thanks!
crude flame
hi there, i'd like to ask whether RVC is better than weights.gg
RVC needs lots o...
weights uses rvc
wet fable
I'm sorry so it's exactly the same thing?
knotty moth
hi there, i'd like to ask whether RVC is better than weights.gg
RVC needs lots o...
if u have a 5090, it might be faster
crude flame
I'm sorry so it's exactly the same thing?
yup
wet fable
Ok thanks!
low shard
hi there, i'd like to ask whether RVC is better than weights.gg
RVC needs lots o...
The difference is the GPU
I don't remember if it was an AI specialized GPU like A100 or an rtx 4090 tho
prisma kite
pomocy wyskoczyło mi No module named 'gradio'
i nie wiem jak to naprawić
help, I got No module named 'gradio'
and I don't know how to fix it****
simple ore
help, I got No module named 'gradio'
and I don't know how to fix it****
if you got the source, you need to install all requirements
but better just get a compiled version
with all the required packages included
gentle hollow
Guys plsss tell me wat Ai vocal remover y’all useee
young perch
Guys, what should I do if I experience a large delay in my voice changer? When I test it in the program, delay is about 3-4 seconds, but when I join the game, it shows around 30 000 ms. Please Help(((
low shard
when i opened voice models channel theres literally 0 models
refresh discord, be sure your connection is stable
might be a discord moment
sage wind
refresh discord, be sure your connection is stable
bro 100mb per sec and also cable how is that can be unstable 😭😭😭
low shard
bro 100mb per sec and also cable how is that can be unstable 😭😭😭
see if there's an 'X' next to the search thing, could be that your last research got stuck
sage wind
see if there's an 'X' next to the search thing, could be that your last research...
o ty
sage wind
did that solve it?
yeah
rare gobletBOT
Ayo? @sage wind level 1 !!!
sage wind
i can even send screenshot but i dont have permission
low shard
i can even send screenshot but i dont have permission
yea you need to be level 1 by chatting to send images in help channels
low shard
yw
chilly ridge
use batch size 4
train for not longer than 200 epochs
I increased the dataset to 10min with every file duration between 5 and 8s (idk why but audacity don't want to work with less duration), it's actually in training for 200 epochs
20s/epoch
dawn bronze
@low shard How can I convert vocals from huggingface.co to A Mp3
analog obsidian
I increased the dataset to 10min with every file duration between 5 and 8s (idk ...
8s is too long, but it doesn't really matter because rvc is going to slice those long samples
(and audacity already removed long silences by itself so the chances of the rvc slider adding long silences are very low)
low shard
<@911742715019001897> How can I convert vocals from huggingface.co to A Mp3
I'm guessing you're talking about the vocals output you get from Ilaria RVC Zero?
Why would you want to convert them from .wav to mp3? mp3 is lossy and lower quality than wav
wav is way better, it's lossless
There's no reason to convert them,
If you really want to do that, you can just google any wav to mp3 converting site like this
But It's really not suggested
knotty moth
<@911742715019001897> How can I convert vocals from huggingface.co to A Mp3
use audacity (with ffmpeg) or any other audio editor, or the online sites like suggested above
true ravine
windows button + tab then create a new desktop and use that
thanks men
azure marshBOT
chilly ridge
8s is too long, but it doesn't really matter because rvc is going to slice those...
It's better but sometimes it's robotic again
but it's really really better
analog obsidian
It's better but sometimes it's robotic again
yes its normal, usually you stop getting robotic sibilances at the 30 minute mark
anything below that use a batch size of 4
and pray for getting them less often
chilly ridge
okay, and if i try more epoch ? like 400 ? it will just getting worst ?
analog obsidian
okay, and if i try more epoch ? like 400 ? it will just getting worst ?
not worse but the improvement is very marginal and you have the risk of overfitting
not audible difference + unnecesary risks
chilly ridge
ok i'll try to increase the dataset again
do you know how i can extract automatically a voice from a certain caracter from a movie file ?
it look really long to do it manually
analog obsidian
do you know how i can extract automatically a voice from a certain caracter from...
that is called speaker diarization, currently the only one that exists is named pyannote and is extremely ass
chilly ridge
i tryed with python and speechbrain but it's not convincing
rare gobletBOT
Ayo? @chilly ridge level 3 !!!
analog obsidian
sadly is better to separate speakers manually
low shard
that is called speaker diarization, currently the only one that exists is named ...
nah not that ass
chilly ridge
that is called speaker diarization, currently the only one that exists is named ...
i tryed that too
chilly ridge
it juste identifyed the whole movie like it's the character i'm looking for
low shard
was very ass for me
idk why it wasn't for me
analog obsidian
it juste identifyed the whole movie like it's the character i'm looking for
yea its trash dont use it
low shard
I haven't played with it since a while tho https://github.com/sanctuary-osai/Pyannote-Speaker-Diarization-3.1
chilly ridge
idk why it wasn't for me
if it work for you, so explained how to use it
analog obsidian
I haven't played with it since a while tho https://github.com/sanctuary-osai/Pya...
better not having the risks of ai getting confused at voices
they do have a paid version i havent tried
probably he tried that
low shard
probably he tried that
How could I?
analog obsidian
How could I?
idk i was guessing
analog obsidian
me 2
low shard
if it work for you, so explained how to use it
the parameters are explained in the ui of the project i sent
chilly ridge
the parameters are explained in the ui of the project i sent
thx
low shard
better not having the risks of ai getting confused at voices
That's true tho, better double check it
but maybe could help, im saying it just for that
imagine having to watch a whole movie to make a model 😭
chilly ridge
yea its trash dont use it
diarization does't worked
any idea how i can export all the movie scene with that character ?
clearelly i don't like the idea to do it mannualy
analog obsidian
any idea how i can export all the movie scene with that character ?
do it manually
oak inlet
File "C:\Users\vedant\Desktop\Retrieval-based-Voice-Conversion-WebUI-main\infer\...
anyone know fix
chilly ridge
yes its normal, usually you stop getting robotic sibilances at the 30 minute mar...
i have 1h of dataset, do i also use 4 for batch size and 200Epoch ?
glacial pollen
so, where is it
analog obsidian
i have 1h of dataset, do i also use 4 for batch size and 200Epoch ?
hard to answer this since it depends on the variety of the dataset
glacial pollen
ah
I mean
It should be pretty simple
🤔
you lack a model ( a component of rvc, that is )
oak inlet
oh
chilly ridge
hard to answer this since it depends on the variety of the dataset
same as always, i have 1h of audio with only 3s-4s audio file
analog obsidian
same as always, i have 1h of audio with only 3s-4s audio file
for simplicity sake use batch 8
glacial pollen
Dl all you lack from there
chilly ridge
thx
oak inlet
<@1311544405361954909>
THAT WAS IT
oak inlet
I DIDNT KNOW WHAT TO DO ON THAT TY
glacial pollen
( + additionally, the root of the issue
but typically the last line tells you what is the main deal
oak inlet
( + additionally, the root of the issue
ye
glacial pollen
Alr, glad I could help. best of luck man
glacial pollen
yes
oak inlet
oh ty
glacial pollen
rmvpe is for f0 extraction
hubert is for features
waxen kelp
hello how do i send pictures for help?
oak inlet
rmvpe is for f0 extraction
hubert is for features
ohh
glacial pollen
hello how do i send pictures for help?
First you must get to some level, I think 5, 7 or 10 to have an ability to ( in case you aren't able to rn
You level up by being active on chat
oak inlet
its levek 2
glacial pollen
oh, then 2 then. Seems like the threshold got lowered
oak inlet
i can do it rn
waxen kelp
First you must get to some level, I think 5, 7 or 10 to have an ability to ( in ...
oh so i just yap?
oak inlet
oh so i just yap?
yues
glacial pollen
I mean yea kinda, but you can just go #🤖│bots I suppose
or somethin'
to not clog the main chats
Anyway, I gotta go back to work now
oak inlet
yo i got another error now there is no trace back it procceses for a few sec then it says error
2024-12-01 11:13:13 | INFO | fairseq.tasks.hubert_pretraining | current directory is C:\Users\vedant\Desktop\Retrieval-based-Voice-Conversion-WebUI-main
2024-12-01 11:13:13 | INFO | fairseq.tasks.hubert_pretraining | HubertPretrainingTask Config {'_name': 'hubert_pretraining', 'data': 'metadata', 'fine_tuning': False, 'labels': ['km'], 'label_dir': 'label', 'label_rate': 50.0, 'sample_rate': 16000, 'normalize': False, 'enable_padding': False, 'max_keep_size': None, 'max_sample_size': 250000, 'min_sample_size': 32000, 'single_target': False, 'random_crop': True, 'pad_audio': False}
2024-12-01 11:13:13 | INFO | fairseq.models.hubert.hubert | HubertModel Config: {'_name': 'hubert', 'label_rate': 50.0, 'extractor_mode': default, 'encoder_layers': 12, 'encoder_embed_dim': 768, 'encoder_ffn_embed_dim': 3072, 'encoder_attention_heads': 12, 'activation_fn': gelu, 'layer_type': transformer, 'dropout': 0.1, 'attention_dropout': 0.1, 'activation_dropout': 0.0, 'encoder_layerdrop': 0.05, 'dropout_input': 0.1, 'dropout_features': 0.1, 'final_dim': 256, 'untie_final_proj': True, 'layer_norm_first': False, 'conv_feature_layers': '[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2', 'conv_bias': False, 'logit_temp': 0.1, 'target_glu': False, 'feature_grad_mult': 0.1, 'mask_length': 10, 'mask_prob': 0.8, 'mask_selection': static, 'mask_other': 0.0, 'no_mask_overlap': False, 'mask_min_space': 1, 'mask_channel_length': 10, 'mask_channel_prob': 0.0, 'mask_channel_selection': static, 'mask_channel_other': 0.0, 'no_mask_channel_overlap': False, 'mask_channel_min_space': 1, 'conv_pos': 128, 'conv_pos_groups': 16, 'latent_temp': [2.0, 0.5, 0.999995], 'skip_masked': False, 'skip_nomask': False, 'checkpoint_activations': False, 'required_seq_len_multiple': 2, 'depthwise_conv_kernel_size': 31, 'attn_type': '', 'pos_enc_type': 'abs', 'fp16': False}
2024-12-01 11:13:14 | INFO | infer.modules.vc.pipeline | Loading rmvpe model,assets/rmvpe/rmvpe.pt
flint solar
2024-12-01 11:13:13 | INFO | fairseq.tasks.hubert_pretraining | current director...
this isn't an error
oak inlet
oh
flint solar
its info
flint solar
in the webui
its normal
flint solar
yo i got another error now there is no trace back it procceses for a few sec the...
nothing seems wrong here
waxen kelp
hi! i try to use RVC but in the end it says this.does anyone know whats the problem and how to fix it
oak inlet
nothing seems wrong here
should i rec it
flint solar
hi! i try to use RVC but in the end it says this.does anyone know whats the prob...
this gui is extremely outdated
distant turtle
-colab
waxen kelp
this is new
oh snap is there a tuto on how to download it?
seems a lot of stuff
im not aware of much github
oak inlet
oh snap is there a tuto on how to download it?
nah there isnt
wait
make sure u have pip 24.0
3.10
rc1
waxen kelp
make sure u have pip 24.0
tf is that?
oak inlet
tf is that?
u need it to install the requirements
rare gobletBOT
Ayo? @oak inlet level 3 !!!
waxen kelp
nah its aight i was just tryna make a silly cover
oak inlet
oh
waxen kelp
thank you very much helping tho!
rare gobletBOT
Ayo? @waxen kelp level 2 !!!
waxen kelp
what's ur pc gpu?
its GX 840
low shard
its GX 840
you sure it's that?
You can check your pc gpu via:
ctrl+shift+esc (task manager) -> Performance tab -> GPU
and here u should find the gpu memory
I have never heard of that gpu + can't find it online
gloomy lynx
I just realized my version of rvc gui is outdated too damn
rare gobletBOT
Ayo? @gloomy lynx level 1 !!!
low shard
I just realized my version of rvc gui is outdated too damn
yea nah u should never use rvc gui
what's ur pc gpu
gloomy lynx
nvidia geforce rtx 3060 ti
low shard
nvidia geforce rtx 3060 ti
low shard
what do you thinks better?
applio imo
gloomy lynx
Alr lemme try it
chilly ridge
for simplicity sake use batch 8
10min25/Epoch is normal ??
analog obsidian
10min25/Epoch is normal ??
what, 10 minutes per epoch or 25 seconds?
analog obsidian
thats not normal, it should be 1-2-3 minute per epoch if your dataset has 1 hour of data
how much vram do you have
chilly ridge
I reach 48min without silence in audio
chilly ridge
how much vram do you have
8Go Vram, 32Go ram and a big xenon processor
I have an IA gpu normally
I dont' know if it's change something
analog obsidian
oh that explains
chilly ridge
Quadro ADA RTX 4000 8Go
analog obsidian
Quadro ADA RTX 4000 8Go
this is what are you using rn?
analog obsidian
idk sorry i have no idea, i don't use these types of gpu
chilly ridge
With the 10min dataset and 4batch it was 7s/epoch
I checked the "charge dataset in gpu"
analog obsidian
I checked the "charge dataset in gpu"
oohhh
this explains
caching dataset in gpu in huge datasets cause massive vram usage
you're using system RAM
this is why became so slow
system is using fallback memory
disable that on big datasets
chilly ridge
Ok so i need to uncheck this option ?
analog obsidian
yes
chilly ridge
caching dataset in gpu in huge datasets cause massive vram usage
Now it's 2min/Epoch
rare gobletBOT
Ayo? @chilly ridge level 4 !!!
chilly ridge
Thx 🙏
chilly ridge
I'll credit you if i finish my project one day
oak hearth
so uhwhere do i download
rare gobletBOT
Ayo? @oak hearth level 2 !!!
oak hearth
my thingy flatlined, it TECHnically is going down and it isnt going up do i just download the latest .pth saved?
or do i let it go until it goes up?
simple ore
click
knotty moth
I checked the "charge dataset in gpu"
you need an A100 to do that
there's no point of using that option tho
knotty moth
Quadro ADA RTX 4000 8Go
I dont recall there's RTX ada with only 8 GB (you could get a cheaper normal rtx one with 12/16 gb vram lmao)
oak hearth
so uh is there any way to resume training without a g and d file
chilly ridge
I dont recall there's RTX ada with only 8 GB (you could get a cheaper normal rtx...
I get it for free, so i don't really care 🙃
chilly ridge
I dont recall there's RTX ada with only 8 GB (you could get a cheaper normal rtx...
I just checked and you'r right, i have a quadro rtx 4000 8Go, the 4000 ada 20Go is in my boss's workstation
Mb
knotty moth
I get it for free, so i don't really care 🙃
you can resell it, profit! 
chilly ridge
Haha, no, i prefer keep it
pastel fiber
can someone help
azure marshBOT
can someone help
chilly ridge
you can resell it, profit! <:goofy:1159397241199534152>
I have a 3060 8Go in my laptop. With same dataset and setting, wich one you think have less min/epoch ?
timber hamlet
What’s the
pastel fiber
why does it say the input is the vb output, it also says output is vb input
rare gobletBOT
Ayo? @timber hamlet level 1 !!!
meager comet
how much faster is an a100 compared to colab t4
knotty moth
I just checked and you'r right, i have a quadro rtx 4000 8Go, the 4000 ada 20Go ...
there's only RTX 4000 ada 20Gogs, your boss must have modded yours to 8Gogs (if not physically, perhaps only bios tweaking). performance wise, it's roughly between 4060 Ti (including 16Gogs) and 4070 (12 Gogs).
chilly ridge
No no
I said i don't have an ada
Just a quadro rtx 4000
My boss's gpu is an rtx 4000 ada
knotty moth
Quadro ADA RTX 4000 8Go
bruh you did say this
chilly ridge
I just checked and you'r right, i have a quadro rtx 4000 8Go, the 4000 ada 20Go ...
Here
knotty moth
Here
the name seems a bit ambiguous, perhaps the old Turing one?
https://www.techpowerup.com/gpu-specs/quadro-rtx-4000.c3336
chilly ridge
Haha yeah, it's this one
I thought I had the same computer as my boss
But it look like he keep the big gpu for him
knotty moth
Haha yeah, it's this one
in comparison, there's RTX 2070 with the same gogs and similar performance
chilly ridge
I see, so, my laptop with the 3060 8Go could be faster/epoch ?
knotty moth
I see, so, my laptop with the 3060 8Go could be faster/epoch ?
also RTX laptops are usually easier to overheat than desktop ones
chilly ridge
I put my laptop upside down in front of the AC 🤫
chilly ridge
I don't understand the graph
knotty moth
I don't understand the graph
skill issue
it literally shows the same performance as 2060
chilly ridge
Okay, not bad for a free gpu
shut goblet
how the fuck do i fix this
simple ore
how the fuck do i fix this
probbaly something like vc redist is missing
or cuda tools
shut goblet
probbaly something like vc redist is missing
well when i was opening the rvc file
it showed up a error
is there a proper way to open the file?
simple ore
is failing to load because some dependency is missing
simple ore
is there a proper way to open the file?
there's a way to find the missing depencency
https://github.com/lucasg/Dependencies/releases/tag/v1.11.1 - download x64 release, unzip, run the dependenciesgui.exe
from that open the dll shown on the screenshot, it will list what it needs
but my guess is either vc++ redist or CUDA toolkit
frank owl
Check out this creation I made on Weights.gg! https://www.weights.gg/shared/cm43ggokn1qqwogar8nn0oibe?inviteCode=4619f
wet fable
Are there any accessible applications that could have the same effect as RVC?
blazing solar
-cOLAB
rare gobletBOT
Ayo? @blazing solar level 1 !!!
low shard
Are there any accessible applications that could have the same effect as RVC?
wdym same effects as RVC
what’s ur pc gpu and what are u looking for?
wet fable
wdym same effects as RVC
Can do the same thing as RVC I mean
low shard
Can do the same thing as RVC I mean
You can just use RVC
RVC is the best speech to speech program
low shard
what’s ur pc gpu and what are u looking for?
Reply to this
knotty moth
Are there any accessible applications that could have the same effect as RVC?
do you mean any frontend applications using RVC? I'd recommend any of following:
-rvc
low shard
tf is a gradio
gradio is one of the most used framework for making webuis in AI projects
so a requirement for eddy’s UVR UI
by checking its github, https://github.com/Eddycrack864/UVR5-UI
you seem to be on windows, did you run UVR5-UI-installer.bat Without administrator?
shut goblet
you seem to be on windows, did you run UVR5-UI-installer.bat Without administrat...
Yes
low shard
Yes
Are you sure it’s in the C drive and there’s no special characters in whatever folder it is?
it doesn’t seem to be on the C drive
marsh schooner
how much hours do yall think is too much hours on a data set to be used on a voice changer
like if i wanted to avoid it being overtrained what would i do
for the maximum
frosty coral
why does rvc keep crashing whenever i click start audio conversion? (not responding)
brittle wing
-hf
flint solar
why does rvc keep crashing whenever i click start audio conversion? (not respond...
What are u using to run rvc
frosty coral
as in gpu?
marsh schooner
What are u using to run rvc
do u know that maximum dataset length a voice can have without being overtrained
flint solar
do u know that maximum dataset length a voice can have without being overtrained
There is no constant length u can’t predict anything
But if I were to choose
I’d go for 7-10 min dataset of clean and diverse data
marsh schooner
minutes???
flint solar
minutes???
Yes
flint solar
i was ab to go for like a 2-3 hours
No dats bad
flint solar
be sure to NOT follow yt tutd
Someone need to put out a new one 😭
low shard
Someone need to put out a new one 😭
I can’t do it for local rvc
flint solar
Ik what he using
low shard
It’s the gui wit harvest and pm
smh
marsh schooner
No dats bad
how is that bad
flint solar
how is that bad
Dats what will most likely happen
knotty moth
Dats what will most likely happen
it seems he resumed training with different batch size from before
flint solar
it seems he resumed training with different batch size from before
It was fucked before they even stopped training
knotty moth
he also kept overtraining, perhaps till more than 1k epochs 💀
vale raptor
hi I have problem, I tried to import voice model to gui but got this error:
size mismatch for enc_p.emb_phone.weight: copying a param with shape torch.Size([192, 768]) from checkpoint, the shape in current model is torch.Size([192, 256])
how can I fix it?
I'm sorry for my mistakes, english is not my native language
flint solar
hi I have problem, I tried to import voice model to gui but got this error:
size...
What are u using to run rvc
locally or colab
vale raptor
What are u using to run rvc
rvc gui from there https://github.com/Tiger14n/RVC-GUI?tab=readme-ov-file
marsh schooner
it seems he resumed training with different batch size from before
so 2-3 hours would be fine?
flint solar
rvc gui from there https://github.com/Tiger14n/RVC-GUI?tab=readme-ov-file
This is an architecture mismatch error
The gui ur using is very out of date
It’s only compatible with rvc v1 models
flint solar
so 2-3 hours would be fine?
1 hour is my maximum length
vale raptor
This is an architecture mismatch error
oh what have I use instead?
flint solar
oh what have I use instead?
-colab
flint solar
oh what have I use instead?
Use applio it’s the first link
vale raptor
Use applio it’s the first link
thanks I'll try
flint solar
thanks I'll try
Let me know if something goes wrong
low shard
-colab
it’s better to ask whats the user pc gpu first than giving colab
in case he has a good pc to run it on
flint solar
in case he has a good pc to run it on
True
low shard
oh what have I use instead?
what’s ur pc gpu?
low shard
True
yea bc i seen people using colab while they got an rtx 3060, just because the helper gave them a colab
The user should always be asked what hardware they got so the helper can give them the tools to use
flint solar
Could you explain how to download it?
Download what
amber fjord
Suggestions for <@1177104427363090445>
he
vale raptor
what’s ur pc gpu?
rtx 2060s
low shard
U can run rvc locally
see
that’s why every staff should always ask for the hardware first
low shard
rtx 2060s
your pc is good enough to run it locally (use it on ur pc), colab is just a cloud service (run it on remote good pc, for people who got bad pc)
low shard
Could you explain how to download it?
elaborate whats ur pc gpu and what u want to do
amber fjord
rtx 4060
rare gobletBOT
Ayo? @amber fjord level 1 !!!
amber fjord
and what do u want to do
trolling my friends
low shard
trolling my friends
trolling can mean anything
im guessing you want realtime voice changer for calls?
amber fjord
yeah
low shard
Yea its the best u always be specific when asking for help
low shard
?
1st link, its the wokada (program for using rvc , speech to speech, models in realtime for calls) fork (modified version, this one has better performance)
the 1st link is a guide to install it (which u will use the nvidia way as you got an rtx)
By reading the guide, you have everything you need to know, if you run into any issues, ask in #🔍│help-w-okada as this isn’t the correct help channel for that
amber fjord
thank
low shard
yw
vale raptor
Let me know if something goes wrong
I'm sorry for silly questions but how can I add voice model
rare gobletBOT
Ayo? @vale raptor level 2 !!!
flint solar
I'm sorry for silly questions but how can I add voice model
Download tab then paste the huggingface model link
vale raptor
I see
vale raptor
Download tab then paste the huggingface model link
damn i did it I finally did it it's victory
thank you very much
simple ore
It was fucked before they even stopped training
since the whole thing is in 35.3-36.4 range, the growing value is not a big deal. but the value itself is high, so yeah, likely just bad dataset
too much variety in the samples, i guess?
flint solar
too much variety in the samples, i guess?
Too much of something is always bad not just in rvc 
flint solar
since the whole thing is in 35.3-36.4 range, the growing value is not a big deal...
He trained it for way too long I think dats why the value is high
simple ore
it started with 35 and ended with 36.4?...so training duration played little difference
I have an opposite problem
amber fjord
what is it?
simple ore
I wish people stopped placing projects into 'special' folders
simple ore
what is it?
move the vc folder to C:\vc
in short, programs like voice changer are not designed with the best Windows SDK guidelines. VC executed by you, without admin permissions, is trying to write stuff into semi-protected Program Files folder, which Windows does not allow, because the software should utilize a user folder instead
flint solar
I never really understood what the fm graph is good for
opal verge
hey im using google RVC colab added index didnt crated ); everthing eles goes well what shuold i do?
low shard
hey im using google RVC colab added index didnt crated ); everthing eles goes w...
elaborate more, which colab and whats the error when training the index?
opal verge
rvc V2
simple ore
I never really understood what the fm graph is good for
there's a comparison between original slice of audio and a generated slice of audio, the distriminator (the teacher in GAN algorithm) compares them together by looking thru several filters and calculates a mean difference
since it does not compare the exact values, it is possible that one part of the generated audio gets better while the other gets worse, but the average difference stays about the same
amber fjord
maybe I don't understand something, but no matter where I put it, the program doesn't work for me
flint solar
since it does not compare the exact values, it is possible that one part of the ...
How do I know if the model is improving or not thru feature matching
simple ore
How do I know if the model is improving or not thru feature matching
generally the difference between real and fake audio, and thus fm value, should be going down
flint solar
generally the difference between real and fake audio, and thus fm value, should...
My fm be going down sometimes
Actually most of the time
simple ore
here I had a model try to produce a single sample
you can hear the difference and the FM chart reflects that
flint solar
My fm be going down sometimes
I meant up** 
flint solar
you can hear the difference and the FM chart reflects that
What are the factors that influence fm
simple ore
yeah, best case I see something like
low shard
👉 https://discord.com/channels/1159260121998827560/1159290161683767298
real, so many people still use this channel for wokada
simple ore
cuz someone needs to rename the channel 🙂
help-rvc (not VC!)
help-w-okada (the VC!)
flint solar
yeah, best case I see something like
Yeah dats what I be getting too
I thought it was bad
simple ore
I think the value going up is expected as the model figures out the parameters to use, but it should stabilize and settle at some value or around it without much deviation
like that 'cement' example above
the discriminator used in RVC is not stellar, so in most cases the model just finds some local minima for parameters and settles around that
and it does not get out of that hole no matter how much more you train it
there's a chance that fm going up is just that hump on the chart after which the value would go down once the model learns to reproduce the audio better, but I've yet too train for that long to see it
flint solar
the discriminator used in RVC is not stellar, so in most cases the model just fi...
I guess improving the discriminator is the next move
simple ore
that would require a whole new set of pretrains
I've tested a new loss function that does not require much change, seems to be doing better
flint solar
I've tested a new loss function that does not require much change, seems to be d...
When is it releasing
fallen grotto
does anyone have any colab i can use to make ai covers
rare gobletBOT
Ayo? @fallen grotto level 2 !!!
flint solar
does anyone have any colab i can use to make ai covers
-colab
flint solar
Use applio
fallen grotto
thanks ho
low shard
help-rvc (not VC!)
😭
azure marshBOT
Not available yet
grim bay
Any order of model to use to extract vocals from a song to make AI cover?
I'm both using MVSEP and UVR
simple ore
When is it releasing
i have a branch created for applio 3.2.7
rare bear
Sorry if this is the wrong place to ask this question but I'm brand new to AI voices and what can be done with them.
I was wondering if the w-okada voice changer is the only 'app' available for real-time changing of voices on a Mac or are there any other apps I could try RVC files out on a Mac with? Thanks.
knotty moth
Sorry if this is the wrong place to ask this question but I'm brand new to AI vo...
read the pinned guide there: #🔍│help-w-okada
rare bear
read the pinned guide there: <#1159290161683767298>
So basically w-okada is the only way to use RVC voices in real-time on a Mac then? There aren't any other apps that can be used instead?
flint solar
i have a branch created for applio 3.2.7
Can u link it
simple ore
knotty moth
So basically w-okada is the only way to use RVC voices in real-time on a Mac the...
nope, not even voice.ai
rare bear
nope, not even voice.ai
Okay thanks. Seems rather strange but it is what it is I guess.
rare gobletBOT
Ayo? @rare bear level 1 !!!
low shard
So basically w-okada is the only way to use RVC voices in real-time on a Mac the...
W-okada is the best
rare bear
W-okada is the best <:lfg:1159355870119993496>
Well from what I've been told so far it's the only one so would be the best wouldn't it? 😉
low shard
lol
flint solar
nope, not even voice.ai
voice ai so ass bro omg
low shard
voice ai so ass bro omg
🙏
inner cloak
When I try to start Applio (colab), I am getting this:
An error occurred connecting to Discord: Could not find Discord installed and running on this machine.
Traceback (most recent call last):
File "/content/program_ml/app.py", line 90, in <module>
inference_tab()
File "/content/program_ml/tabs/inference/inference.py", line 418, in inference_tab
choices=get_speakers_id(model_file.value),
File "/content/program_ml/tabs/inference/inference.py", line 325, in get_speakers_id
model_data = torch.load(model, map_location="cpu")
File "/usr/local/lib/python3.10/dist-packages/torch/serialization.py", line 1004, in load
with _open_zipfile_reader(opened_file) as opened_zipfile:
File "/usr/local/lib/python3.10/dist-packages/torch/serialization.py", line 456, in init
super().init(torch._C.PyTorchFileReader(name_or_buffer))
OSError: [Errno 22] Invalid argument
It just stays like this with no link to gradio. Anyone else?
It was working last night, so I am guessing it needs to be fixed. Edit: nvm
rare gobletBOT
Ayo? @shut goblet level 3 !!!
simple ore
if I had to guess, some translation is messed up
spice wing
Where can i find Voice models for AllTalk Ai?
pth files dont work there
shut goblet
if I had to guess, some translation is messed up
i dont understand what this means
unique rock
Hey, I need to remove the breaths from my dataset, for example: the breaths when a singer is about to sing du verse and takes a breath or releases it when he finishes singing.
low shard
Where can i find Voice models for AllTalk Ai?
pth files dont work there
There aren't here
80% of the models are RVC (Best few shots Speech To Speech)
20% GPT-SoVITS (Best few shots Text To Speech)
wild yoke
Why is .wav not recognized as a format?
simple ore
i dont understand what this means <:skullsob:1159372531992645662>
i18n is a package that is used to translate UI to different languages
simple ore
Why is .wav not recognized as a format?
not an actual .wav maybe?
viscid moss
how do i fix this
Did you solve it?
I thought that bug was fixed. Anyway, how to fix that is described here
viscid moss
also <@274566299349155851>
Thanks for lmk Nick
oak hearth
What are pretrains? If I want to clone a voice, should I use a pretrain or spend more time making it myself? What will be more quality?
rare gobletBOT
Ayo? @oak hearth level 3 !!!
oak hearth
Are pretrains just faster and better and should i use them to make the most lifelike realistic clone?
analog obsidian
What are pretrains? If I want to clone a voice, should I use a pretrain or spend...
pretrains are pre-made models trained on days worth of audio (in rvc case)
making them from scratch is a very hard process even for people that have the knowledge
always use a pretrain model when training on rvc
their purpose is to have a baseline during the finetuning process (when you are training a model)
They can affect the final result quality since audio upscaling is involved
If you train without a pretrain your model is going to sound like shit because rvc has no prior knowledge of sounds
analog obsidian
Are pretrains just faster and better and should i use them to make the most life...
always use the original pretrain as the custom made pretrains dont reconstruct your dataset frequencies as good as the original
for a realistic model train an expressive non monotone dataset of 30 minutes and above
the more data, the better and realistic
be sure your dataset has variety, for example avoid monotone dialogue or audios repeating the same sentence/words etc
coral frigate
Can someone give me advice? I’m training a model with 15 hours of data and I don’t want to use a pretrain on it. Which one should I select? Custom makes me add my own pretrain I think but I’m not sure.
rare gobletBOT
Ayo? @coral frigate level 5 !!!
analog obsidian
Can someone give me advice? I’m training a model with 15 hours of data and I don...
15 hours is too short for a pretrain
aim for the amount of time as vctk dataset
crude flame
Can someone give me advice? I’m training a model with 15 hours of data and I don...
are you making a from scratch pretrain or a finetune
coral frigate
I’m not aiming to make a pretrain. I just want to make a voice model with my dataset but I don’t want to use any of the pretrains applio offers
analog obsidian
I’m not aiming to make a pretrain. I just want to make a voice model with my dat...
what i said still applies
coral frigate
If that makes sense
analog obsidian
rvc will not be able to give you a good result
use a pretrain instead
when you remove a pretrain you remove the knowledge of sounds
coral frigate
Im still new to making models but everytime I’ve used a pretrain, the voice has sounded strange and very off. Is there a way I can fix that then if I have to use a pretrain?
crude flame
I’m not aiming to make a pretrain. I just want to make a voice model with my dat...
de-select this
analog obsidian
Im still new to making models but everytime I’ve used a pretrain, the voice has ...
you've used a pretrain with this exact dataset?
coral frigate
Yes with 150 epoch. It sounds nothing like the dataset and just sounds really off. It has been the case whenever I use a pretrain. In this case I used the contentvec pretrain.
analog obsidian
Yes with 150 epoch. It sounds nothing like the dataset and just sounds really of...
contentvec is not a pretrained model
is an embedder
crude flame
Yes with 150 epoch. It sounds nothing like the dataset and just sounds really of...
contentvec is for feature extraction
analog obsidian
for your huge dataset of 15 hours you can try batch 16, and it should take 1 or 2 days to give good results with a pretrain
coral frigate
contentvec is for feature extraction
Ow I assumed that they were all just pretrains.
coral frigate
for your huge dataset of 15 hours you can try batch 16, and it should take 1 or ...
Can you suggest which pretrain I should use from the list for an English model?
analog obsidian
Can you suggest which pretrain I should use from the list for an English model?
original pretrain
coral frigate
Sorry if this is a dumb question but How do I use an original pretrain?
crude flame
Sorry if this is a dumb question but How do I use an original pretrain?
dont touch this
analog obsidian
Sorry if this is a dumb question but How do I use an original pretrain?
are you on applio? simply don't change anything related to pretrains, it'll use the original
yea that
leave it like that
coral frigate
Ow ok I just keep that unchecked and I’m good to go?
coral frigate
are you on applio? simply don't change anything related to pretrains, it'll use ...
And yes I’m on applio
analog obsidian
Ow ok I just keep that unchecked and I’m good to go?
yep
coral frigate
Ok thank you so much guys
coral frigate
Can someone give me advice? I’m training a model with 15 hours of data and I don...
So would it matter which one of these I select if I’m unchecking the box anyway?
crude flame
So would it matter which one of these I select if I’m unchecking the box anyway?
leave it on contentvec
coral frigate
Ok thank you
knotty moth
I’m not aiming to make a pretrain. I just want to make a voice model with my dat...
you dont need shitton hours of a single speaker dataset from 3k voice recordings like this https://www.techspot.com/news/105764-panasonic-resurrects-long-dead-founder-ai-share-management.html
30m-2h can already produce good results
coral frigate
you dont need shitton hours of a single speaker dataset from 3k voice recordings...
I was just told when I started that the bigger the dataset the t he better
knotty moth
I was just told when I started that the bigger the dataset the t he better
do you even have enough time and effort to clean the massive dataset lmao
remember: quality > quantity
analog obsidian
the more the better but i agree 15 hours is too much 😭
also you have to be sure your dataset has no possible sounds that could cause problems later
coral frigate
Yeah sadly I did have to listen through all 15 hours like a podcast
analog obsidian
i mean technically a 15 hour dataset is better than a 30 min or an 1 hour one
but rvc already becomes realistic at the 30 minute mark
technically speaking is not bad
but is too much effort
coral frigate
Idk I’m dumb and still figuring rvc out. I’ve actually been trying to train this for 3 weeks now but applio hasn’t been able to train it. Just keep getting an error
analog obsidian
Idk I’m dumb and still figuring rvc out. I’ve actually been trying to train this...
for realistic purposes i recommend using a dataset of 1 hour
max 2 hours
higher than that is not bad, but not worth the effort
covert axle
-colab
coral frigate
max 2 hours
higher than that is not bad, but not worth the effort
Would’ve been nice to know before going through hours and hours of data
analog obsidian
Would’ve been nice to know before going through hours and hours of data
yea... we tell people the more the better because most believe a model is going to sound realistic at 5 minutes of data
knotty moth
Yeah sadly I did have to listen through all 15 hours like a podcast
surely the cleaning effort could take more than double the audio duration itself, unless you hire a few freelancers (that still waste of money)
coral frigate
yea... we tell people the more the better because most believe a model is going ...
Is it normal for feature extraction to take 10-15 minutes? I’m trying to troubleshoot why I still can’t train the model
analog obsidian
Is it normal for feature extraction to take 10-15 minutes? I’m trying to trouble...
yes because your dataset is 15 hours
@knotty moth is rmvpe on applio using gpu now?
or they're still using rmvpe cpu
if applio is still using rmvpe cpu, is going to be super slow
if applio is using rmvpe gpu it should be done in 30 mins approx
analog obsidian
surely the cleaning effort could take more than double the audio duration itself...
what i can say about this is at least his 15 hour model will be able to do much more than a 1 or 2 hour one
so the effort will at least be worth a little bit
analog obsidian
what i can say about this is at least his 15 hour model will be able to do much ...
but still, rvc is really limited because hifigan and bad code
coral frigate
ive been getting this error in the cmd for a month and im probably doing something dumb thats failing the extraction. have any suggestions on how else i can trouble shoot it?
analog obsidian
ive been getting this error in the cmd for a month and im probably doing somethi...
reduce cpu cores to 4
knotty moth
<@681186927151546397> is rmvpe on applio using gpu now?
not sure but it still has some gpu usage
coral frigate
ill try that
analog obsidian
not sure but it still has some gpu usage
hmm he got oom so i suppose is using rmvpe gpu
coral frigate
reduce cpu cores to 4
do i use 4 for both preprocess and extract?
analog obsidian
do i use 4 for both preprocess and extract?
yep, it doesnt really a speed difference, or at least i havent noticed it
maybe it has one but is very marginal
knotty moth
hmm he got oom so i suppose is using rmvpe gpu
I only recall oom happens when inferring a big whole audio depending on ram
analog obsidian
I only recall oom happens when inferring a big whole audio depending on ram
oom also happens when you're using too much cpu cores during feature extraction
staring at crepe
knotty moth
oom also happens when you're using too much cpu cores during feature extraction
not sure, it simply hangs the system for a while and then crashes
analog obsidian
not sure, it simply hangs the system for a while and then crashes
it gave me a bsod once 😭
crashed that bad
coral frigate
reduce cpu cores to 4
It’s in the middle of extracting all the audio files but as its going in seeing that it still gives the same error line for each one
analog obsidian
It’s in the middle of extracting all the audio files but as its going in seeing ...
uh.... are you sure you reduced the cpu cores during feature extraction?
whats going on is that your system is getting out of vram during the feature extraction
coral frigate
Yh it’s at 4 now for both
Could it be because I have 33 files in the dataset folder maybe?
analog obsidian
Could it be because I have 33 files in the dataset folder maybe?
i dont think so, first process is the preprocessing, this slices your audios since hifigan does not work in long audios
after that feature extraction, extracts the features of the sliced audios
so is doing it on 3s samples
it goes out of vram when is doing a lot of them
at the same time
coral frigate
This has been an issue for a while now. Idk what else I can do to fix the issue
analog obsidian
This has been an issue for a while now. Idk what else I can do to fix the issue
last resort is to try using 2 cpu cores instead of 4
if that doesn't work im lost sorry, have never faced an issue like that on applio before
might be related to applio only
coral frigate
last resort is to try using 2 cpu cores instead of 4
if that doesn't work im los...
and nothing is wrong in this section?
analog obsidian
and nothing is wrong in this section?
nope, looks fine to me
coral frigate
Well then I’m doomed
analog obsidian
Well then I’m doomed
cpu cores to 2 also shouldn't affect the speed quality
since it looks applio is finally using rmvpe gpu
like u could try that
cpu cores only affect feature speed when you're using an cpu based extractor
coral frigate
I’ll try that but I don’t think it’ll change much seeing that my cpu should’ve been able to handle 4
Maybe redownloading applio will do something
analog obsidian
I’ll try that but I don’t think it’ll change much seeing that my cpu should’ve b...
your cpu is not being used here, your gpu is
the gpu is doing the extractor
hence why it gets out of vram
in older versions of applio rmvpe was actually cpu based
but is very slow compared to rmvpe gpu
coral frigate
Even then my gpu should’ve been able to handle 4 without vram issues. I’ll try redownloading applio and hope that fixes something
analog obsidian
yeah exactly it should be able to do it, i find weird its getting out of vram
i hope redownloading fixes the issue
coral frigate
Thanks. I hope so
rare gobletBOT
Ayo? @coral frigate level 6 !!!
coral frigate
I’m out of options after that
glacial pollen
Could it be because I have 33 files in the dataset folder maybe?
Limit each file's length to 10 mins ( maybe 15 )
use 4 or 8 threads and switch to rmvpe ( non gpu variant )
btw, would be helpful if you provided info on your amount of ram, ( and vram if you actually did use rmvpe gpu variant )
Also, total length of the set and length per file ( can be avg )
analog obsidian
btw, would be helpful if you provided info on your amount of ram, ( and vram if ...
is applio using rmvpe gpu?
is weird he's getting oom with a 4090
glacial pollen
is applio using rmvpe gpu?
not sure as I don't use applio
perhaps it's a 2 in 1 ?
in any case, 4090 shouldn't have any ooms like that
analog obsidian
Limit each file's length to 10 mins ( maybe 15 )
its a 15 hour dataset
he is using rvc slicer to slice those files
glacial pollen
only reasonable way out of it is.. the extraction's done on cpu and / or per-file length
oh
That's an overkill
what's the length per file? ( as I assume it's not big sample but split
coral frigate
use 4 or 8 threads and switch to rmvpe ( non gpu variant )
How Do I switch to the none gpu variant? I have 32gb ram and the file lengths vary from 20 mins to the longest being 50 minutes
glacial pollen
How Do I switch to the none gpu variant? I have 32gb ram and the file lengths va...
ye that's the thing
above 20 or maybe 25 ( but I'd stick to 20 per file ) mins, it can cause issues
analog obsidian
oh but even if he lets the slicer to do it?
analog obsidian
yea this is new to me too xD
glacial pollen
oh yea
that can surely be the case cause, 1 set where extraction is loading 15 hours of data vs sequential 1 by 1 sid processing ( even 40 or 100 hours )
is different
analog obsidian
interesting, rvc always giving surprises
glacial pollen
or so I suspect at least
there's no other way out of it
other than applio being jammed ( wouldn't be surprised at this point
analog obsidian
other than applio being jammed ( wouldn't be surprised at this point
i believe this might be the problem actually lmao
glacial pollen
lemme prepare something
might have a 'temp' solution
@coral frigate Amount of threads you have?
I'll assume you can afford to use 8
coral frigate
How do I check? But I definitely would
glacial pollen
that's alright then, gimme a sec
knotty moth
is weird he's getting oom with a 4090
have we known his cpu? if it's intel 13/14th, could be its degradation issue
coral frigate
Ryzen 9 9950x
knotty moth
Ryzen 9 9950x
check your cpu & gpu stability, shouldnt be overvolted
coral frigate
now the question is, are you willing to spend a lil bit of time on that?
Sure on what?
glacial pollen
@coral frigate Cause basically, you'd have try 2 things
- Try to feature extract on single 30-50 min file
- If that failed, you'd have to use another fork or just, mainline for the sake of extraction ( and then move stuff back to applio ) (( that'd confirm whether applio is the issue or just the situation itself isn't favored by rvc in general ))
- If that failed too.. rip, in that case it'd mean rvc's not optimized for single speaker 15h at once extraction
coral frigate
Ok how do I move the feature extract back into applio if that’s the issue ?