applio is a lot better than the other shit I was using before
#✨│ai-help
flint solar
applio is a lot better than the other shit I was using before
U prolly was using mangio fork
flint solar
im gonna try 12 batch size
How long is ur dataset
flint solar
about 90 minutes
8 would work fine
frail umbra
kk im gonna try it again now, Im pretty sure it was the audio sample rate or something
steep stirrup
30 min
simple ore
when you download audio from 15 year old answering machine with shitty dac
jovial maple
Hello guys, Im new here and Im just wondering about the best app or open source AI that I can use to clone some voices
simple ore
more like 20-year old
i remember I had one and could not even understand the messages people were leaving on it
fading yew
Hey guys i need help ... what is this effect called ??? Where can i generate similar kind of effect for my images ... Pls help TT
https://youtu.be/o1OsDWT_DUc?t=29
steel kite
Hina Mod doesn't seem to be working, it's not giving me the gradio link
Like it's not working at all
idk what this means
frosty ferry
nvm
steel kite
@potent badger @odd shale Can someone help me with this?
simple ore
idk what this means
it means the requirements have not been installed
steel kite
How do I fix it?
simple ore
i imagine this would happen if you just ran a clone of the app and skipped ths step that installs the required libraries
but since this colab is not in the list of workin colabs, I assume it is dead
steel kite
I tried hitting Run all and I even tried them in order.
There isn't other cloud based solutions that have mango-crepe. Only just crepe.
simple ore
mangio crepe is crepe with flexible hop length
serene horizon
RVC erroring in inference.
“AttributeError: 'NoneType' object has no attribute 'tobytes'”
Any ideas why this is happening?
simple ore
which does not fking work anyway
steel kite
How so?
simple ore
RVC erroring in inference.
“AttributeError: 'NoneType' object has no attribute ...
full error dump pls
serene horizon
full error dump pls
What does that include?
simple ore
the whole stack of errors so we can see what started it
I can only guess you're trying to infer a file that ends up as None object that gets sent to the pipeline and that is not good
serene horizon
the whole stack of errors so we can see what started it
pitch, pitchf = self.get_f0(
File "/Users/me/pinokio/api/rvc.pinokio.git/app/infer/modules/vc/pipeline.py", line 149, in get_f0
self.model_rmvpe = RMVPE(
File "/Users/me/pinokio/api/rvc.pinokio.git/app/infer/lib/rmvpe.py", line 563, in init
self.model = get_default_model()
File "/Users/me/pinokio/api/rvc.pinokio.git/app/infer/lib/rmvpe.py", line 544, in get_default_model
ckpt = torch.load(model_path, map_location="cpu")
File "/Users/me/pinokio/api/rvc.pinokio.git/app/env/lib/python3.10/site-packages/torch/serialization.py", line 986, in load
with _open_file_like(f, 'rb') as opened_file:
File "/Users/me/pinokio/api/rvc.pinokio.git/app/env/lib/python3.10/site-packages/torch/serialization.py", line 435, in _open_file_like
return _open_file(name_or_buffer, mode)
File "/Users/me/pinokio/api/rvc.pinokio.git/app/env/lib/python3.10/site-packages/torch/serialization.py", line 416, in init
super().init(open(name, mode))
FileNotFoundError: [Errno 2] No such file or directory: 'assets/rmvpe/rmvpe.pt'
serene horizon
the whole stack of errors so we can see what started it
Traceback (most recent call last):
File "/Users/me/pinokio/api/rvc.pinokio.git/app/env/lib/python3.10/site-packages/gradio/routes.py", line 437, in run_predict
output = await app.get_blocks().process_api(
File "/Users/me/pinokio/api/rvc.pinokio.git/app/env/lib/python3.10/site-packages/gradio/blocks.py", line 1349, in process_api
data = self.postprocess_data(fn_index, result["prediction"], state)
File "/Users/me/pinokio/api/rvc.pinokio.git/app/env/lib/python3.10/site-packages/gradio/blocks.py", line 1283, in postprocess_data
prediction_value = block.postprocess(prediction_value)
File "/Users/me/pinokio/api/rvc.pinokio.git/app/env/lib/python3.10/site-packages/gradio/components.py", line 2586, in postprocess
file_path = self.audio_to_temp_file(
File "/Users/me/pinokio/api/rvc.pinokio.git/app/env/lib/python3.10/site-packages/gradio/components.py", line 360, in audio_to_temp_file
temp_dir = Path(dir) / self.hash_bytes(data.tobytes())
AttributeError: 'NoneType' object has no attribute 'tobytes'
simple ore
FileNotFoundError: [Errno 2] No such file or directory: 'assets/rmvpe/rmvpe.pt'
you dont have the rmvpe model for some reason
using mangio(?) instead of applio on mac?
and pinokio of all things?
yeah, good luck with that
serene horizon
Ha! Well. I’m trying to run mainline RVC on Mac.
simple ore
okay, so whatever reason it failed to download the f0 extractor model
serene horizon
As I’m having issues with Applio sound quality on Mac.
serene horizon
okay, so whatever reason it failed to download the f0 extractor model
Only file in the rmvpe folder is rvmpe_inputs.pth
simple ore
you can grab it here and put where it needs to be https://huggingface.co/IAHispano/Applio/tree/main/Resources
serene horizon
you can grab it here and put where it needs to be https://huggingface.co/IAHispa...
Okay, I have done that. But still giving me an error message.
flint solar
Okay, I have done that. But still giving me an error message.
Send the error message
serene horizon
Send the error message
I don’t get an error message. It just stops at “2025-03-31 22:36:58 | INFO | infer.modules.vc.pipeline | Loading rmvpe model,assets/rmvpe/rmvpe.pt”
flint solar
I don’t get an error message. It just stops at “2025-03-31 22:36:58 | INFO | inf...
Did u install all the dependencies
serene horizon
Did u install all the dependencies
Not sure.
simple ore
as long as it does not throw 'module not found' it is fne
flint solar
As I’m having issues with Applio sound quality on Mac.
How so
serene horizon
How so
The file it produces is crackly and not of as good quality as running the same inference model in RVC Mangio on Windows.
edgy tangle
[W socket.cpp:697] [c10d] The client socket has failed to connect to [localhost]:28625 (errno: 99 - Cannot assign requested address).
simple ore
<:skull_sob:1159372531992645662>
i recommend using one big file as a source and let Applio cut it
because whatever you're doing is not working
knotty moth
show me the contents of dataset folder, there must be something wrong
knotty moth
screenshot the folder contents in imjoy file manager pls
there should be like /program_ml/assets/datasets/YourModel that should contain wav file(s)
edgy tangle
screenshot the folder contents in imjoy file manager pls
ok I died
let me do everything again
knotty moth
as said above, a single wav file is recommended for the applio slicing method
edgy tangle
yeah, I just simple slicing
using one big file
I think its a problem with kaggle or idk, because I do the same on local and works fine
I just notice that in kaggle doesnt create .spec.pt files
I think that its the problem
simple ore
spec pt are created on the fly as the files get loaded into the loader
in your case nothing got loaded
edgy tangle
oh
knotty moth
no that is in the processed model folder, I meant your dataset folder
simple ore
show the content of filelist.txt
edgy tangle
ok wait me a sec
knotty moth
but if like so, it should be fine
edgy tangle
no that is in the processed model folder, I meant your dataset folder
I just put this on dataset creator
simple ore
can you show kaggle ui for preprocess and extract features?
and log for extract features cell
edgy tangle
can you show kaggle ui for preprocess and extract features?
Im using crepe, but with rmvpe gives me the same error
when training
knotty moth
then check if the index file has already been created
edgy tangle
I can create index
knotty moth
I can create index
not "can" but the whether the index file has already been created
edgy tangle
knotty moth
no, in the model folder in file manager
edgy tangle
there it is
simple ore
okay, so extractr features runs, but the filelist is still empty?
knotty moth
there it is
well
knotty moth
there it is
then try start training
edgy tangle
okay, so extractr features runs, but the filelist is still empty?
Yeah still empty
knotty moth
is it in index training process or not?
edgy tangle
is it in index training process or not?
no
aaah sorry i dont understand
when I press start training gives me this
knotty moth
are u using applio main branch? (without -3.2.8)
edgy tangle
yeah
eh
wait me a sec
I removed it
but I think I removed something else with it
--branch 3.2.8-bugfix
y removed this
It is ok?
knotty moth
--branch 3.2.8-bugfix
it is the release branch, not the main one which is more recent
simple ore
repeat after me - you can not train with an empty filelist.txt
but it should not be empty
if the extracted file is full of files, it means the process ran
just did not save the list for some reason
vale radish
Can someone help whenever i open go-web.bat it says "D:\RVC 2025>runtime\python.exe infer-web.py --pycmd runtime\python.exe --port 7897
'runtime\python.exe' is not recognized as an internal or external command,
operable program or batch file.
D:\RVC 2025>pause
Press any key to continue . . .'
knotty moth
Can someone help whenever i open go-web.bat it says "D:\RVC 2025>runtime\python....
there is a space, go rename the folder to RVC2025
vale radish
there is a space, go rename the folder to `RVC2025`
Now what?
knotty moth
I dont know why it is happening
there shouldn't be empty filelist.txt in the main branch
edgy tangle
there shouldn't be empty filelist.txt in the main branch
Yeah, but I dont know why applio its keeping filelist empty
vale radish
https://tenor.com/view/just-do-it-shia-la-beouf-do-it-gif-4531935
still saying the same thing though
edgy tangle
knotty moth
still saying the same thing though
try doing manually
vale radish
try doing manually
How? Im new to this
knotty moth
still saying the same thing though
please show the error message again
after you rename the folder to D:\RVC2025
vale radish
D:\RVC2025>runtime\python.exe infer-web.py --pycmd runtime\python.exe --port 7897
'runtime\python.exe' is not recognized as an internal or external command,
operable program or batch file.
D:\RVC2025>pause
Press any key to continue . . .
knotty moth
D:\RVC2025>runtime\python.exe infer-web.py --pycmd runtime\python.exe --port 789...
if the python.exe doesn't exist, redownload the whole installation
vale radish
k
atomic sparrow
how should I generate the index from colab correctly?
knotty moth
how should I generate the index from colab correctly?
I don't get what you mean, please show the screenshot with error message
and make sure you have done previous steps without issues
atomic sparrow
hello, I don't mean any error, my question is what is the correct way to get the index when you are training a model for the first time.
pd: i cant send images
knotty moth
hello, I don't mean any error, my question is what is the correct way to get the...
please read the guide in this section https://docs.aihub.gg/rvc/local/applio/#3-train
atomic sparrow
thanks a lot!
whenever I create a model, and generate its respective index, I get an error when I use it in playground for example, in the applio web. Is it because I generated it after starting the training?
knotty moth
you didn't mean any error and then you mention the error? 
knotty moth
pd: i cant send images
!give-media-perms 1h @atomic sparrow
atomic sparrow
no, the doubt before was precisely because, from previous times, the index had given me an error XD
this is the error, any message, only this
hallow thistle
Generate index
It's download model
atomic sparrow
no, I am showing the section to add files, which is just below, the top is something else.
knotty moth
no, the doubt before was precisely because, from previous times, the index had g...
because you asked how to "generate index" I suppose you mean for training models
for existing models to add, it should have the pth file, and the index file within the zip
some older models might not have index file, and there's no way to generate it, except by the model maker with the model's preprocessed files
the inference can work without index file tho
atomic sparrow
the pth, I add it in the normal file and the index I have to put it only in a compressed zip?
ok I didn't know about this, thank very much
atomic sparrow
How long does it usually take for the colab to hold without finishing the execution with the training? or if there is any way to lengthen that time?
royal sierra
yo guys what voice changer for this is the best
knotty moth
How long does it usually take for the colab to hold without finishing the execut...
depend on time per epoch and total epochs (the cell can be stopped during the training session whenever needed)
knotty moth
yo guys what voice changer for this is the best
go to #🔍│help-w-okada then read the pinned guide there
simple ore
the pth, I add it in the normal file and the index I have to put it only in a co...
no, it just needs to have the index file name in a specific format
MyModel.pth and MyModel.index should work, I think
bleak nymph
im gonna do this but i'm gonna feel insane
like i'm literally talking to myself
it's easier in vc cause i'm actually talking to someone 😭
shadow sparrow
yo i cant find the pth file using Applio, how do i generate it or where do i find them?
flint solar
yo i cant find the pth file using Applio, how do i generate it or where do i fin...
Can u elaborate
shadow sparrow
i have absolutley zero pth files on logs
flint solar
i have absolutley zero pth files on logs
Are u training a model or just to infer
shadow sparrow
training
shadow sparrow
Are u training a model or just to infer
should it they be on the Applio backup folder
flint solar
should it they be on the Applio backup folder
Ss the colab file explorer
If ur using google colab
shadow sparrow
flint solar
They should be at program_ml/logs/modelName
shadow sparrow
flint solar
In ur case it’s the nett1-PossoMigliorare-1
shadow sparrow
which one is the right one?
flint solar
which one is the right one?
Use tensorboard to find out the best epoch
Lowest point on the g/total graph
shadow sparrow
because i got this error so i thought it would help me
shadow sparrow
It’s not accurate at all
what should i do?
flint solar
because i got this error so i thought it would help me
I don’t see an error
flint solar
damn so i just dont know how to use it lmao
Restart the training without the overtraining thingy
flint solar
how do i see the graph bro
Click refresh
flint solar
from scratch
Yea
shadow sparrow
ok but the tensorboard doesnt show me anything even when i reload it
flint solar
ok but the tensorboard doesnt show me anything even when i reload it
Are u sure
shadow sparrow
yeah im going to restart now and imma tell u if it works
flint solar
yeah im going to restart now and imma tell u if it works
Word
simple ore
yo i cant find the pth file using Applio, how do i generate it or where do i fin...
there was a bug that if you ran training with save every epoch = max epoch, the final model did not get saved
shadow sparrow
there was a bug that if you ran training with save every epoch = max epoch, the ...
so basically don't use overtraining detector?
simple ore
the overtraining detector is pretty much useless
but it was not the issue
the issue was that you did not save models every 10 or 20 epochs
anyway, I've fixed it in the latest code
shadow sparrow
so what should i do differently sorry bro but im a dumbass
simple ore
and dont bother with overtraining detector, use the tensorboard
shadow sparrow
what if i max out to 1k the total epoch is it bad?
tender kite
What's a recommended amount of training data for a voice model? 10 minutes?
analog obsidian
What's a recommended amount of training data for a voice model? 10 minutes?
depends in how expressive and diverse the dataset is
a 15 minute expressive dataset can beat a 30 minute monotone one
but also a 30 minute expressive dataset can beat a 15 minute expressive one
shadow sparrow
do i need to download this if im using the collab?
tender kite
depends in how expressive and diverse the dataset is
a 15 minute expressive dat...
Already. I got like an hour total of training expressive good quality audio, I just thought it'd be too much lol
analog obsidian
Already. I got like an hour total of training expressive good quality audio, I j...
1 hour is already enough to train models, it is not recommended to use more than that 🦈 🤙
shadow sparrow
1 hour is already enough to train models, it is not recommended to use more than...
bro help me out too
flint solar
do i need to download this if im using the collab?
U don’t need to download shi if u using colab
Everything is ran in the cloud
shadow sparrow
Everything is ran in the cloud
ok its working now bro, Big thanks fr
shadow sparrow
U don’t need to download shi if u using colab
y'all last thing is it possible to OT at epoch 38 with 20 mins of content?
guys i got this error again
silent stratus
guys i got this error again
idrk how to fix bc i just came accross this but try kaggle if its not working on colab
silent stratus
y'all last thing is it possible to OT at epoch 38 with 20 mins of content?
unlikeley but depends on ur batch size too
knotty moth
because i got this error so i thought it would help me
I'm not sure but I only use kaggle one with working tensorboard
obsidian marten
can someone teach me how to make an ai voice model?
knotty moth
can someone teach me how to make an ai voice model?
start from this guide https://docs.aihub.gg/essentials/how-to-make-voice-models/
rustic bobcat
what services do you guys use to implement the zip file
light meadow
guys what iss the best custom pretrained
knotty moth
guys what iss the best custom pretrained
the default pretrain if not sure, or this https://discord.com/channels/1159260121998827560/1339155300720054316
light meadow
😮
oh and
if i already split my dataset
into smaller files
how can i prevent applio do it again
is disabling cut_preprocess help
flint solar
hmm i forgot how many minute i need for dataset
10 minutes is a recommended minimum
crude flame
hmm i forgot how many minute i need for dataset
our docs have a good amount of useful info https://docs.aihub.gg/rvc/resources/dataset-isolation/
analog obsidian
15 minutes of audio, where the person uses their full vocal range to get decent results
increase that to 30-40 minutes to get even more natural results
knotty moth
into smaller files
better not do so, or join into a whole file then use the applio slicer
weak cipher
I made vosk stt api and don't know how to change voice like anime girl =))
I use hugface as api
light meadow
thx guys
light meadow
err,
errm, i forgot where can i find the output of the trains
so that i can download in case overtrained
i used rvc disconnected before, but this is my first time use applio
knotty moth
errm, i forgot where can i find the output of the trains
Applio/logs/yourmodel
knotty moth
so that i can download in case overtrained
and the tensorboard file (tfevents) is at the eval folder
knotty moth
wait till it saves some checkpoints
in Applio it should be in the same model logs folder
knotty moth
but i already do about 160 epochs
you didn't enable save_every_weights
knotty moth
you have to start over training (delete D_2333333.pth and G_2333333.pth files first)
and also eval folder
light meadow
;-; damm i have already trained like 3 hours.. are there any chance that i can continue train from the current status, but with save every weight ?
knotty moth
switch to another account, then load the preprocessed model files
again, since you missed the model checkpoints you should start over training
light meadow
its okay, i mean i just want to prevent OT in the future
it still ok for now
if now i press this button, can i export current model and continue training after that ?
knotty moth
its okay, i mean i just want to prevent OT in the future
the better way to tell if the model is overtraining or not is to test several checkpoints besides the tensorboard graph
light meadow
😮
light meadow
if now i press this button, can i export current model and continue training aft...
can i do this ._. ?
dammm i stopped and now need to train from 0 again
simple ore
dammm i stopped and now need to train from 0 again
there's resume training cell
that loads files from back-up back into the training folder
so you can resume from the last saved weights
knotty moth
there's resume training cell
he didn't check save_every_weights and he missed saved model checkpoints in the last training session, so I'd recommend starting over
simple ore
can always resume from d/g weights and do more epoch with the setting checked
or simply extract the model from g
brittle wing
-colab
half kayak
link rvc?
sweet oak
can rvc be used real time in voice chats like discord or gta?
knotty moth
can rvc be used real time in voice chats like discord or gta?
nope, only prerecorded ones
but if you want that, go to #🔍│help-w-okada and read the pinned guide there
sweet oak
but if you want that, go to <#1159290161683767298> and read the pinned guide the...
thanks
leaden belfry
10x!!!! Worked perfectly!
simple ore
10x!!!! Worked perfectly!
well, inference works, trainining wont
rich moss
hey I have two AI voice models (.pth and .index) that I trained separately, and I was wondering if there’s a way to merge them into a single model in Applio.
Is it possible to combine both .pth files and .index files to create a single voice?
tame mica
?
pth is the model
index is the accent
you cant mix them
if you have two pth files then yes
rich moss
if you have two pth files then yes
i want to mix the 2 pth files
flint solar
you cant mix index (or not yet(?)
U can use the index of ur choice
Between the 2 pths
rich moss
that is possible yeah
So, how can I merge two .pth files to create a single model? Can it be done within Applio?
analog obsidian
So, how can I merge two .pth files to create a single model? Can it be done with...
the two models must have been trained on the same sample rate
the first model is the main model
the second model is the secondary model
blend ratio is how much data of the main model do you want to keep, for example a blend ratio of 0.5 will retain only 50% of the information of the main model, the other 50% will come from the secondary model
trim frost
How do I upload my rvc model here?
sudden girder
anyone have a tutorial on how to create an ai song with these models?
nvm i got it
karmic oliveBOT
Not available yet
gilded charm
Does anyone know what's the best ai cover maker that I can run locally?
hot lagoon
simple ore
update gradio to 5.23.1
opal kestrel
what is the difference between crepe and crepe tiny im confused, i was told crepe is best for male voice but only crepe_tiny is avalible, how do i download crepe?
edgy fjord
Who have link for Collab training?
low shard
Who have link for Collab training?
First, what's your PC GPU?
fathom reef
Anyone else having trouble getting Ilaria to work?
Because every single time I use it with any model, I get an error instead.
I've uploaded my own voice models five times in a row now, testing them with different files to try and get it to work, yet nothing has changed.
edgy fjord
First, what's your PC GPU?
Phone
knotty moth
Because every single time I use it with any model, I get an error instead.
you didnt mention the error message, so I'd have to speculate:
- "GPU quota exceeded", try using smaller audio file, or wait for another hour, login your account
- audio file not readable, try re-exporting using audacity or use another software with "include metadata" disabled
- invalid/corrupted model pth/index file?
- 500 internal server error or random error in some internal code, you're cooked at this moment
fathom reef
Looks to be a 500 or a random error, because I tested the model with Applio and it worked just fine.
low shard
Anyone else having trouble getting Ilaria to work?
tell:
- your pc gpu
- what u want to do
- what link are you using
fathom reef
It's Ilaria, not local.
low shard
Train (make) RVC Models on cloud:
- Prepare the Dataset
- Setup RVC:
Choose a cloud way to use RVC,
- Google Colabs (4 hours of daily gpu for free, not much hours, but easy to use):
- Kaggles (a bit harder to use and needs phone number but gives 30 hours weekly of better gpus):
- Mainline (UI)
- Applio by Vidal (UI)
- Lightning.ai (Kinda hard, needs login, no issue with web uis or anything, but only free 15 credits monthly):
- Be sure to know about the tensorboard
Google Colab = Easier but risk of getting disconnected
Kaggle = Harder but way more gpu time
If you are looking for the easiest way and for free, is using https://weights.com/ which ofc uses RVC
RVC Inference (use models) on pre-recorded audio on Cloud
You can use either:
- Weights.com: Easiest Possible Ever Automatic
- Ilaria RVC Zero: Fastest free on cloud
- Applio UI Colab: RVC Fork with some extra features like TTS
- RVC AI Cover Maker UI: Automatically Separates the Vocals and Instrumentals, converts the voice and mixes them back
lament eagle
@low shard Yo something wrong with Kaggle applio I'm not sure what
low shard
It's Ilaria, not local.
yes, and I asked for your pc gpu, and what link are you using
i asked your pc gpu so i could tell you if it’s good enough to run it
and i asked your link to check if u are using the updated one
low shard
Ilaria is on the cloud, though.
I meant good enough to run it locally, since local doesn’t have cloud limits
low shard
What "updated one"?
-spaces
fathom reef
I'm not even using Colab.
I'm using the HuggingFace space, exactly as you mentioned.
lament eagle
<@911742715019001897> Yo something wrong with Kaggle applio I'm not sure what
When restarted
edgy fjord
"If you are going to train models, upload your dataset to your Google Drive storage" what is a dataset
How I get it
Is the audio file?
knotty moth
"If you are going to train models, upload your dataset to your Google Drive stor...
you should understand the term from the guide
https://docs.aihub.gg/essentials/how-to-make-voice-models/ https://docs.aihub.gg/rvc/resources/training/
edgy fjord
Thx
flint solar
@simple ore
Traceback (most recent call last):
File "/usr/local/lib/python3.10/dist-packages/gradio/queueing.py", line 625, in process_events
response = await route_utils.call_process_api(
File "/usr/local/lib/python3.10/dist-packages/gradio/route_utils.py", line 322, in call_process_api
output = await app.get_blocks().process_api(
File "/usr/local/lib/python3.10/dist-packages/gradio/blocks.py", line 2137, in process_api
result = await self.call_function(
File "/usr/local/lib/python3.10/dist-packages/gradio/blocks.py", line 1663, in call_function
prediction = await anyio.to_thread.run_sync( # type: ignore
File "/usr/local/lib/python3.10/dist-packages/anyio/to_thread.py", line 56, in run_sync
return await get_async_backend().run_sync_in_worker_thread(
File "/usr/local/lib/python3.10/dist-packages/anyio/_backends/_asyncio.py", line 2470, in run_sync_in_worker_thread
return await future
File "/usr/local/lib/python3.10/dist-packages/anyio/_backends/_asyncio.py", line 967, in run
result = context.run(func, *args)
File "/usr/local/lib/python3.10/dist-packages/gradio/utils.py", line 890, in wrapper
response = f(*args, **kwargs)
File "/content/program_ml/core.py", line 605, in run_model_blender_script
message, model_blended = model_blender(model_name, pth_path_1, pth_path_2, ratio)
ValueError: too many values to unpack (expected 2)
got this error when i tried to blend
low shard
I've uploaded my own voice models five times in a row now, testing them with dif...
I just tested it works
what’s your model download link and pc gpu? I’m asking your model download link so i can test it myself, and your gpu in case it’s good enough to use local tools instead of cloud which is limited
fathom reef
Sorry, that model is private.
low shard
<@911742715019001897> Yo something wrong with Kaggle applio I'm not sure what
there is another program running your ngrok token, maybe delete the notebook and retry all the guide
fathom reef
And for the last time, I'm not giving my PC GPU, since Ilaria runs on the cloud.
simple ore
<@155030383648440320>
Traceback (most recent call last):
File "/usr/local/lib...
models have to be the same sr/type
flint solar
models have to be the same sr/type
they are?
lament eagle
there is another program running your ngrok token, maybe delete the notebook and...
I exactly did that 😭
simple ore
they are?
there can be an issue if one model has sr 40000 and other older model has 40k
simple ore
When restarted
show the end of it
it looks like gradio/pydantic issue, a fresh clone should work
low shard
Sorry, that model is private.
It has a .index and .pth right?
you sure there aren’t other useless files?
i can’t help much with little info and without testing but alright
fathom reef
Of course it does.
lament eagle
it looks like gradio/pydantic issue, a fresh clone should work
I'll show when I got home
fathom reef
I know what I'm doing, and it just started randomly working all of a sudden.
I don't know what happened, but that was weird.
I'm sorry if I'm being rude, I'm in a really bad mood rn.
low shard
And for the last time, I'm not giving my PC GPU, since Ilaria runs on the cloud.
I just asked so you could run it locally if it’s good enough, there’s local tools that run on it and don’t have limitations, unlike, for example, huggingface zerogpu used by ilaria rvc which limits how long your file input can be, and google colab that has max 4 hours daily of gpu that can be even less. There are limitations because those run on a remote good pc gpu, which are expensive
Sayung the gpu name publically isn’t a unique unifier, there are thousands of pc with a gpu like yours, im saying this if you were worried about being tracked by your gpu name
Anyways, your choice
i hope I was clear enough, anyways I will respect your choice and understand
low shard
I know what I'm doing, and it just started randomly working all of a sudden.
does it not upload at all, or just not inference?
fathom reef
Fine. I don't think I'll ever be able to run anything at all locally, considering my GPU is an RX 580.
low shard
I'm sorry if I'm being rude, I'm in a really bad mood rn.
it’s fine, everyone has a bad day, sorry if i seemed rude too
fathom reef
It's resolved now, but there was a random error when using inference.
low shard
I exactly did that 😭
it worked?
low shard
Fine. I don't think I'll ever be able to run anything at all locally, considerin...
it could actually run it locally via applio on zluda iirc (which is a cuda emulator) on windows, except that it will be slower,
I asked you because we have seen users with high end rtx or 1k dollars gpu using google colab, because many think that they need 100k dollars pc to run it locally or think ai is only in sites
low shard
It's resolved now, but there was a random error when using inference.
ohh, that’s great to hear
I hope you have a great day then, for any other issues be free to ask there
I hope you understand I was just trying to be helpful, and sorry for any inconvenience,
About that random error, it might be zerogpu sometimes having issues, it’s a service that ilaria pays for her huggingface space, which basically uses shared high end gpus for multiple huggingface spaces
knotty moth
Fine. I don't think I'll ever be able to run anything at all locally, considerin...
tl;dr the zerogpu spaces are usually meant for "trying stuffs/demos". so as said above it's recommended to use applio zluda or run in cpu mode (slower but still feasible)
flint solar
there can be an issue if one model has sr 40000 and other older model has 40k
Ohh it’s prolly dat cause one is trained on applio and the other on mainline
There isn’t a way to make it work?
simple ore
delete these two lines in rvc/train/process/model_blender.py ``` if ckpt1["sr"] != ckpt2["sr"]:
return "The sample rates of the two models are not the same."
it may still fail with a very old model that has no settings saved
need to check the .pth values
brittle wing
-hf
simple ore
somebody forgot to connect the notebook to the gpu
serene horizon
I’m getting a lot of static /metallic/rattling sounds on my model. Any ideas why?
Ss sounds in particular.
Apart from that, it sounds good.
analog obsidian
I’m getting a lot of static /metallic/rattling sounds on my model. Any ideas why...
robotic esses happens due to either overtraining or the dataset not having enough sibilants
serene horizon
robotic esses happens due to either overtraining or the dataset not having enoug...
I did 350 epochs. 15 to 18 mins of clean audio.
blazing rivet
help its not working and i can upload a photo to show my problem
The system cannot find the path specified.
serene horizon
a bit like this @simple ore
hot lagoon
ot?
simple ore
click off the middle button, then fit the whole thing in the chart
odd sedge
how download RVC on MAC OS ?
simple ore
how download RVC on MAC OS ?
RVC as in realtime voice changer or as in retrieval-based voice conversion?
vale osprey
-colab
knotty moth
I’m getting a lot of static /metallic/rattling sounds on my model. Any ideas why...
suboptimal epoch, "overtraining", poor lossy source and/or muddy voice
serene horizon
suboptimal epoch, "overtraining", poor lossy source and/or muddy voice
Source is good and clear. I recorded it myself through a condenser mic.
As I said, I did 350 epochs with 15 to 18 mins of audio.
Too much epochs? Not enough audio clips?
knotty moth
Source is good and clear. I recorded it myself through a condenser mic.
As I sa...
try comparing it with lower epochs
marsh schooner
what speed do yall think my 3060 ti can run on without having a monitor plugged in or anything
only the voice changer
hallow thistle
only the voice changer
Mate, this isn't where you looking for W-Okada. For W-Okada, go to #🔍│help-w-okada. This #✨│ai-help is all about RVC programs and related tools.
serene horizon
try comparing it with lower epochs
How much lower?
wind fulcrum
im having trouble when dropping a index file in applio it just say error can someone please help me
knotty moth
How much lower?
not a good idea to ask anyone before trying to see
knotty moth
im having trouble when dropping a index file in applio it just say error can som...
error what? pls show it
wind fulcrum
error what? pls show it
just says error
knotty moth
just says error
error what?
wind fulcrum
but it was just that model i think
knotty moth
screenshot pls?
!give-media-perms 1h @wind fulcrum
don't forget the console window
wind fulcrum
error what?
im really new to this
serene horizon
not a good idea to ask anyone before trying to see
So there isn’t a recommended average of epochs to use? Because at the moment, I’m just guessing.
knotty moth
So there isn’t a recommended average of epochs to use? Because at the moment, I’...
the tensorboard graph can be the guidance but something unpredictable could happen
serene horizon
the tensorboard graph can be the guidance but something unpredictable could happ...
So I have to run again, and then look at the tensorboard?
formal wind
So I have to run again, and then look at the tensorboard?
Theres alot more to it but basically yes
serene horizon
Theres alot more to it but basically yes
Best amount of audio to use?
formal wind
Best amount of audio to use?
Like The mod said, use tensorboard. You can ask some mods how to 'Find the best epoch / detect for overtraining'
Which you mainly do with tensorboard and by ear
hallow thistle
im really new to this
This is the console window for Applio. It can be either CMD or Windows Terminal.
hallow thistle
So I have to run again, and then look at the tensorboard?
This is the actual Tensorboard. https://cdn.discordapp.com/attachments/1159290139609137264/1355196249476960519/image.png
vapid mantle
kaç yapmam gerekiyor?
mossy wyvern
what should i do my voice just disappears and sounds strange?
low shard
what should i do my voice just disappears and sounds strange?
Are you talking about realtime voice changing?
If so
Wrong channel
mossy wyvern
Are you talking about realtime voice changing?
yes
low shard
RVC = Retrieval-based-Voice-Conversion, the best Speech To Speech AI Models (on v2), Inferences (use models) pre-recorded audio (ai covers) and train (make) models
Wokada = uses RVC for realtime inference
mossy wyvern
and a voice delay of 5+ seconds
low shard
yes
Then elaborate your request in #🔍│help-w-okada
wary onyx
Are there websites like Weights.gg or anything cloud based stuff that i can do for free since my GPU is a bit poopookaka
Where you can train a model using the cloud?
I remember there being a command but i forgot it xd
patent trellisBOT
karmic oliveBOT
verbal oasis
Hi, I have an audio generated that sometimes is read fast and sometimes slow. Do you know if there's any tool, possibly AI-based, that can automatically recognize those parts and retime their speed? Perhaps by first establishing what a normal speed should sound like, and then making the fast and slow parts coherent with that standar
low shard
Are there websites like Weights.gg or anything cloud based stuff that i can do f...
there's google colab, kaggle and ligtning ai, those aren't as easy as weights.gg tho
all 3 got a free tier
btw weigths has a free tier too
if you're looking for a site that gives you remote good pc 24/7 for free without any limitations, nope that doesn't exist, gpus are expensive
rich moss
https://docs.applio.org/applio/getting-started/voice-blender
Hey! I tried blending two voice models in Applio, making sure both .pth files use the same sample rate. I set the first one as the main model and the second as the secondary
But I’m getting this error:
File "C:\aplio3\Applio-3.2.8-bugfix\core.py", line 605, in run_model_blender_script
message, model_blended = model_blender(model_name, pth_path_1, pth_path_2, ratio)
ValueError: too many values to unpack (expected 2)
analog obsidian
Hey! I tried blending two voice models in Applio, making sure both .pth files us...
the two models were trained in applio? or one was trained in applio and the other in mainline?
back then someone had the same problem due to one model being trained in applio and the other in mainline
rich moss
the two models were trained in applio? or one was trained in applio and the othe...
I'm not sure actually. I trained my model in Applio, but the other one I downloaded from weights.gg, so I don’t know which training method was used for it
analog obsidian
I'm not sure actually. I trained my model in Applio, but the other one I downloa...
if the model was trained using weights trainer then it's using mainline
simple ore
I'm not sure actually. I trained my model in Applio, but the other one I downloa...
use extras/model info
on both files
and see what they show
analog obsidian
oh yeah thats a good way to tell the trainer used for each model
rich moss
if the model was trained using weights trainer then it's using mainline
mainline is RVC1006Nvidia.7z ?
analog obsidian
mainline is RVC1006Nvidia.7z ?
yup
simple ore
not what I asked
in applio there's a screen that you put the model path in and it shows the info
analog obsidian
rich moss
do that for each model
analog obsidian
looks like the first one was trained using mainline
makes sense why u can't merge it with the applio one
simple ore
32k vs 32000
rich moss
looks like the first one was trained using mainline
So, do I need to retrain one of the voices with mainline? can I merge them using mainline ?
analog obsidian
So, do I need to retrain one of the voices with mainline? can I merge them using...
would be better to train both voices in applio because mainline havent got an update since almost 3 years now
but just in case, don't use ai outputs as dataset
get the actual real voice and train that instead
rich moss
would be better to train both voices in applio because mainline havent got an up...
Ty, but I can’t retrain the voice since it’s an AI model specific from weights.gg 😦
rich moss
but just in case, don't use ai outputs as dataset
What do you mean by that?
analog obsidian
What do you mean by that?
for example, training a dataset using audio made by AI will give mediocre results compared to a dataset with real human voice audio
she's the voice of the duckus model
train that instead of using the ai model's outputs
simple ore
option 1) fix model blender to remove the sample rate check
option 2) fix the model and change the sr value in the model
you need to do the opposite with '32k' model
analog obsidian
^ this also works btw
rich moss
option 1) fix model blender to remove the sample rate check
How can I do that
simple ore
edit this file
rich moss
edit this file
its work ty🫶
daring skiff
Hi. I saw a guy on youtube that did some fun, and really accurate ai voice trolling in overwatch with overwatch characters voices. And he said that the tech used is RVC/so-vits-svc-fork. I thought i wanted to try to do something similar. How?
hollow bolt
can someone give me the best examples of setups rvc. I want to see its full potential bc it seems to me there are way better voice swap AIs(not free) - even 5$ elevenlabs does a way better job and requires no setup
simple ore
Hi. I saw a guy on youtube that did some fun, and really accurate ai voice troll...
sovits is so old
pliant pewter
Any tts recommendation?
pliant pewter
and do you need voice cloning?
English. Not really, I'm gonna use voice model if available
I just want to create a voice tag that uses my fav character voices
simple ore
for Orpheus TTS, also needs LM Studio to run the model
nice expressive reads
less expressive, but very fast - kokoro tts
pliant pewter
Thank you
patent trellisBOT
karmic oliveBOT
marsh aspen
wha'ts this server about, guys selling voice ai models?
brittle wing
whats wrong with my rvc??
knotty moth
whats wrong with my rvc??
-gui
brittle wing
-gui
what do i use then bro
i found rvc to be better than weights for that specific model
gloomy gate
guys i using gtx 1650 super.. can you give me link the lastest version for this my gpu?
stable dagger
-rvc
AI HUB Docs
stable dagger
Anyone has tutorial for RTX 50 series (blackwell)? All of the RVC are not updated yet.
simple ore
Anyone has tutorial for RTX 50 series (blackwell)? All of the RVC are not update...
for mainline rvc runtime\python -m pip install torch==2.7.0.dev20250311 torchaudio==2.6.0.dev20250312 torchvision==0.22.0.dev20250312 --index-url https://download.pytorch.org/whl/nightly/cu128
but it wont work until you go around all modules and also fix torch.load calls and add 'weights_only=True' parameter
knotty moth
^@brittle wing
brittle wing
^<@456226577798135808>
thanks
pliant sail
what happend for this?
knotty moth
what happend for this?
try the applio one
pliant sail
try the applio one
alr
hallow thistle
guys i using gtx 1650 super.. can you give me link the lastest version for this ...
Try Applio. Not really suitable for voice model training with this GPU, but fine for doing AI cover. https://docs.applio.org/applio
hallow thistle
what do i use then bro
Try Applio instead. 
gloomy gate
Try Applio. Not really suitable for voice model training with this GPU, but fine...
I using windows not macOS
hallow thistle
I using windows not macOS
gloomy gate
tell me sir what voice changer device is suitable for my gpu gtx 1650 super
hallow thistle
tell me sir what voice changer device is suitable for my gpu gtx 1650 super
What do you mean? Are you looking for W-Okada the realtime voice changer and not RVC program?
!howtoask
knotty moth
tell me sir what voice changer device is suitable for my gpu gtx 1650 super
go to #🔍│help-w-okada and read the pinned guide there
gloomy gate
What do you mean? Are you looking for W-Okada the realtime voice changer and not...
Yes am looking for real time voice changer
knotty moth
Yes am looking for real time voice changer
#🔍│help-w-okada is for voice changer
hallow thistle
For W-Okada, go to #🔍│help-w-okada . But if you're looking for RVC the regular voice changer, stay here.
gloomy gate
Okay sir thank you
hallow thistle
No, W-Okada and RVC (Applio for example) are two different programs.
gloomy gate
So? w-okada for real time voice changer?
hallow thistle
gloomy gate
😭🙏🏻
hallow thistle
Next time, better write more context on what you're looking for instead of "I have a good GPU, but where can I download this thing?".
gloomy gate
instead i asked what voice changer is suitable for ("my gpu is gtx 1650 super")
Am a new ,so i dont know
hallow thistle
Still.
knotty moth
instead i asked what voice changer is suitable for ("my gpu is gtx 1650 super")
still not reading the guide in #🔍│help-w-okada ?
hallow thistle
shell sundial
which voice changer is good for real time changing with GPU 4070ti?
can u provide link download please thank u
knotty moth
which voice changer is good for real time changing with GPU 4070ti?
please go to #🔍│help-w-okada and read the pinned guide there
pliant sail
can someone help me?
i wanna backup the models and retraining but when i start training, is not continue model, is make new model... what should i do?
knotty moth
can someone help me?
delete the existing model folder before loading the backup
unless u mean to save to a backup
pliant sail
okay
oh sorry wrong translate, i mean i wanna continue the model btw i have already the backup but when starts training, is not load the backup and start make new model
knotty moth
you just need to do the "load backup", but not the preprocessing & feature
since the latter would create a new designated model folder
pliant sail
alr
errant acorn
i keep having this problem when opening the voice changer it keeps saying failed to load URL
knotty moth
i keep having this problem when opening the voice changer it keeps saying failed...
again please go to #🔍│help-w-okada and read the pinned guide there
knotty moth
and you another one 
#🔍│help-w-okada
clever birch
mbmb i didnt see channel
lament anvil
if there are 3 inputs and I have 3 hidden layers, will one neuron, for instance, take all 3 inputs but increase the weights of 2 inputs and not the third, while the second neuron focuses on increasing the weights of the first and third inputs and reduces the weight of the second, and so on? Is this correct?
serene horizon
I checked and I have 11.26 minutes of good audio. Can I train with that? @knotty moth
simple ore
if there are 3 inputs and I have 3 hidden layers, will one neuron, for instance,...
hidden layers are layers between the input layer and the output layer
lament anvil
hidden layers are layers between the input layer and the output layer
talking about a "perceptron" or a neuron in a neural network. If you have 3 inputs (x1, x2, x3), for example, one perceptron might focus on the first and third inputs (x1 and x3) and give them high weights (e.g., 0.9 and 0.8) while giving the second input (x2) a very small weight or zero (e.g., 0.1 or 0). Meanwhile, another perceptron might focus on the second and third inputs (x2 and x3), giving them high weights (e.g., 0.7 and 0.9) and reducing the weight of the first input (x1) to something close to zero? is that right?
simple ore
weights are usually learnable, you only control how many neurons on the input layer contribute to the calculated weight for the neuron on the next level
lament anvil
weights are usually learnable, you only control how many neurons on the input la...
so what i said is right?
simple ore
technically you can set the initial weights
lament anvil
technically you can set the initial weights
so the training ai means change the weights
simple ore
the training means the model adjusts the weights so you get the result you want from the input you provide
lament anvil
the training means the model adjusts the weights so you get the result you want ...
so what i said is right
?
simple ore
not exactly
idle osprey
how do i make models
serene horizon
Any idea why Applio on Kaggle keeps erroring on me? @simple ore
simple ore
Need to update gradio, I guess?
serene horizon
Need to update gradio, I guess?
"An error occurred launching Gradio: When localhost is not accessible, a shareable link must be created. Please set share=True or check your proxy settings to allow access to localhost."
simple ore
no idea
idle osprey
<https://docs.aihub.gg/essentials/how-to-make-voice-models/>
i dont understand
isnt there like a website to put my audio file and make it make the model
unreal lichen
!rvc websites
#websites
bro please some one help i want the link for training in colaab
tame mica
-kaggle
tame mica
?
unreal lichen
tnx tho
low shard
bro please some one help i want the link for training in colaab
Have I tried checking if ur PC is good enough first btw
low shard
isnt there like a website to put my audio file and make it make the model
AI training never worked like that, this is open source ai, not ChatGPT
The only site that does this is https://weights.gg/ which uses RVC with a simpler version that automatically detects overtraining
Btw RVC can also be used locally on your PC GPU if it's good enough, weights.gg is a cloud remote good PC service
tame mica
It's with 1 l
mb gng
torn seal
hello do we have to cut the sample vocal to 59 sec for better results while converting (infer) the sample vocals to the model anymore?
low shard
what happened to illyria collab
the ilaria rvc mainline google colab is broken forever #📰│dev-updates message and got replaced by the ZeroGPU huggingface space, Ilaria RVC Zero
simple ore
hello do we have to cut the sample vocal to 59 sec for better results while conv...
what is it based on?
serene horizon
@simple ore can you look at my tensor graphs and tell what you think?
weak needle
Hey 😊
I have a, probably really dumb, question.
I can hear myself talking when I have the program open, but it's not being "piped" to the actual output?
I assume I activated some kind of setting by accident, considering I haven't had this issue before.
Does anyone have a quick idea ? ):
stable dagger
Why does my TTS applio result has breathing sound? It's from TTS there's no breathing, it's not from real person audio
simple ore
Why does my TTS applio result has breathing sound? It's from TTS there's no brea...
example?
stable dagger
example?
I can't send a file in this channel, mind if i DM you to send the audio?.
Nevermind the advanced settings has filter radius to decrease respiration and it works just fine with other model.
simple ore
breathing may be coming from the model, usually there should be no such thing with tts
lament anvil
Sparse Connections make the input such that a group of inputs connects to a specific neuron in the hidden layer if, for example, you know a specific domain. But if you don’t know that specific domain and you make it fully connected, meaning you connect all the inputs to the entire hidden layer, will the fully connected network then focus and try to achieve something like Sparse Connections can someone say that im right or not?
viscid cairn
how do i change the batch someone send me a video please
hallow thistle
how do i change the batch someone send me a video please
!howtoask
hallow thistle
Which RVC program/link are you using?
hallow thistle
📥 **Download:** https://huggingface.co/Razer112/Public_Models/resolve/main/E-Wo...
That's not RVC program. That's a link to an RVC voice model.
viscid cairn
oh well it keeps making a weird noise well whatever ill just stop using it
hallow thistle
I'm asking you which RVC link you're using. Applio or W-Okada for example.
Crazy. Instead of continuing to solve the problem, you're just giving up at this point. 
knotty moth
Why does my TTS applio result has breathing sound? It's from TTS there's no brea...
ye more likely the model itself, perhaps trained using the old mainline/mangio with improper silence truncating and lack of denoising/noise gate, so it could hallucinate silences with some random breaths and noises
knotty moth
oh well it keeps making a weird noise well whatever ill just stop using it
you should denoise the input audio before inferring
simple ore
Sparse Connections make the input such that a group of inputs connects to a spec...
what are you trying to do? anyway, a fully connected network can set some weights to 0 during training, thus acting like a sparse network
serene horizon
ye more likely the model itself, perhaps trained using the old mainline/mangio w...
I'm using Mainline on Kaggle. Is there a more up to date version I should be using?
serene horizon
https://www.kaggle.com/code/deiant/applio/notebook
Thanks. I'll try that again.
But last time I tried it, it kept giving me errors.
serene horizon
https://www.kaggle.com/code/deiant/applio/notebook
Where do I put the dataset to train with?
"TypeError: argument of type 'bool' is not iterable
An error occurred launching Gradio: When localhost is not accessible, a shareable link must be created. Please set share=True or check your proxy settings to allow access to localhost."
I keep getting this error with Applio on Kaggle.
knotty moth
"TypeError: argument of type 'bool' is not iterable
An error occurred launching ...
mine is still rock solid, so try copying these codes to the Install and Start cells respectively (in the Start code, don't forget to set your ngrok token)
low shard
"TypeError: argument of type 'bool' is not iterable
An error occurred launching ...
u can also try to delete the kaggle and redo the tutorial
low shard
how do i change the batch someone send me a video please
all youtube tutorials are outdated, use ONLY written guides
serene horizon
mine is still rock solid, so try copying these codes to the Install and Start ce...
Thanks. That worked.
300 epochs about right for 12/13 mins of audio?
Also, is there a custom pretrained I should use, or will standard be okay?
knotty moth
Thanks. That worked.
300 epochs about right for 12/13 mins of audio?
Also, is...
unfortunately you can't rely much on experts for predicting the next stock prices
so you have to test the results and decide
serene horizon
so you have to test the results and decide
Got it. Okay, I have started training and turned on overtraining preventer.
hallow thistle
what tutorial link did you use?
Bro gave up before I even ask him about his GPU.
knotty moth
mine is still rock solid, so try copying these codes to the Install and Start ce...
@low shard are there anyone else having the similar issue on applio kaggle? if so, you could have my working script pinned here or in #📰│dev-updates
serene horizon
so you have to test the results and decide
Just checking I understand this correctly -
"New best epoch 300 with smoothed loss_g 21.384 and loss_d 3.517
Training has been successfully completed with 300 epoch, 6900 steps and 33.725 loss gen.
Lowest generator loss: 20.398 at epoch 276, step 6325
Saved model '/kaggle/working/program_ml/logs/my-project40kapp/my-project40kapp_300e_6900s.pth' (epoch 300 and step 6900)
Saved model '/kaggle/working/program_ml/logs/my-project40kapp/my-project40kapp_300e_6900s_best_epoch.pth' (epoch 300 and step 6900)
my-project40kapp | epoch=300 | step=6900 | time=13:15:31 | training_speed=0:00:35 | lowest_value=20.398 (epoch 276 and step 6325) | Number of epochs remaining for overtraining: g/total: 50 d/total: 99 | smoothed_loss_gen=21.384 | smoothed_loss_disc=3.517"
Meaning 276 was the best epoch before overtraining and so I should use "my-project40kapp_300e_6900s_best_epoch.pth?"
low shard
<@911742715019001897> are there anyone else having the similar issue on applio k...
I haven't seen anyone else having issues on applio kaggle
did you modify anything on the code?
it looked like to me that the user just did a copy of an old applio kaggle time ago
knotty moth
I haven't seen anyone else having issues on applio kaggle
did you modify anythi...
it was since the first 3.2.8 release was out and I did some tweaks
knotty moth
Just checking I understand this correctly -
*"New best epoch 300 with smoothed...
why still using overtraining detector?
serene horizon
why still using overtraining detector?
What do you mean?
I used the overtraining detector in order to avoid overtraining.
low shard
#google- colabe
what?
low shard
@fading vault @unreal lichen do -colab to check a list of google colabs in #🤖│bots ,and you will have sapphire's message
also, reminder that google colab is a cloud service (remote good pc) for people with a bad pc, and you should prob check your pc gpu first to see if it's good enough to run RVC or Wokada
low shard
btw is the colab working?
google colab is a cloud service, an entire platform where anyone can share notebooks that contain code to run ai programs
the site itself is up, but i don't know which google colab notebook you're talking about, you should send the link
there can be colabs that are old, some other not working, some other works
send the link here, also remember to check #📰│dev-updates
simple ore
I used the overtraining detector in order to avoid overtraining.
it does not work the way you expect it to work, it only check that g total does not go up for x epochs
use tensorboard and test the actual model
low shard
it was since the first 3.2.8 release was out and I did some tweaks
@simple ore could u check if the kaggle applio works
serene horizon
it does not work the way you expect it to work, it only check that g total does ...
Hasn’t applio selected the best epoch at the end?
simple ore
i dont use kaggle
simple ore
Hasn’t applio selected the best epoch at the end?
no
it selected "best" using the metric above, which does not work properly
simple ore
you check the tensorboard charts
serene horizon
you check the tensorboard charts
For?
knotty moth
inspect the graphs and also test several checkpoints
simple ore
so it looks like this
norm_g - stays below 1000, fm - flat or going down, mel - should always be going down, total g - converging at some value
if it looks like that, then you train until total g stabilizes
serene horizon
norm_g - stays below 1000, fm - flat or going down, mel - should always be going...
simple ore
use charts from the scalar tab, avg 50 ones
since your fm is weird, you may want to play with smaller or larger batch size
or a different pretrain like klm 4.9 for hifigan
serene horizon
since your fm is weird, you may want to play with smaller or larger batch size
I did 4 batch size.
serene horizon
or a different pretrain like klm 4.9 for hifigan
where do I get this?
serene horizon
https://discord.com/channels/1159260121998827560/1339155300720054316
Thank you.
So I should try again with KLM pretrain. And different batch size, higher or lower?
simple ore
how big is your dataset?
low shard
i dont use kaggle
could you update it with @knotty moth 's tweaks? i just checked myself and the kaggle notebook is acting weirdly
serene horizon
how big is your dataset?
13 minutes.
knotty moth
13 minutes.
not only that but also how diverse the dataset is (rather monotone or more expressive)
analog obsidian
expressive as the person using their whole vocal range
serene horizon
not only that but also how diverse the dataset is (rather monotone or more expre...
It’s of me singing. It covers all my ranges , high and low. For example, me singing Iris by the goo goo dolls is one of the files.
Also, other songs with good ranges.
knotty moth
It’s of me singing. It covers all my ranges , high and low. For example, me sing...
did you change learning rate or not?
serene horizon
did you change learning rate or not?
I set batch size to 4.
knotty moth
I set batch size to 4.
I assume you didnt modify the other configurations
serene horizon
I assume you didnt modify the other configurations
the only other setting I changed was to 300 epochs.
serene horizon
how stable the loss & gradient graphs tho?
Which graphs are these?
brave swallow
does this looks like good dataset to start work on it ??
serene horizon
So I'm going to try batch size of two with 250 epochs.
brave swallow
tops looks like they've been chewed by mp3
very interesting
this the the raw !
is the raw also looks chewed ?
if yes then I got huge problem
simple ore
you have noise at the bottom
brave swallow
you have noise at the bottom
yep under the 100 you mean right ?
but I didn't understand where you guys know the diff if its mp3 or compressed becuase both were .WAV
simple ore
need to use a basic highpass filter to cut the noise
the difference is mp3 eats low level values
brave swallow
I'll try to find examples for this
simple ore
low shard
could you update it with <@681186927151546397> 's tweaks? i just checked myself ...
cus I dunno what other Applio staffer to ask 
simple ore
you can see chunks eaten out of the sibilant
it is okay to use mp3 for inference, but for training it is no ideal
brave swallow
it is okay to use mp3 for inference, but for training it is no ideal
yeah , I'm into training
I think my raw data already compressed that's why
like here it has the same gaps you marked
knotty moth
yep under the 100 you mean right ?
just denoise it
serene horizon
@knotty moth @simple ore batch size 2/250 epochs/KLM pre-train
How is this?
small grail
anyway to make the local version of AICoverGen work on m1 mac?
brave swallow
<@681186927151546397> <@155030383648440320> batch size 2/250 epochs/KLM pre-trai...
batch size 2 ? you lack of VRAM or you aim to get realistic audio ?
simple ore
@serene horizon
knotty moth
<@681186927151546397> <@155030383648440320> batch size 2/250 epochs/KLM pre-trai...
bs 2 is worse
serene horizon
<@813202583903797258>
serene horizon
bs 2 is worse
so I go higher?
knotty moth
so I go higher?
minimum 4
simple ore
so you're using klm 4.9?
serene horizon
so you're using klm 4.9?
Yes.
simple ore
yeah, dont go less than 4 batch
knotty moth
does this looks better ?
not sure if sounds a bit muddy but not bad after all
brave swallow
not sure if sounds a bit muddy but not bad after all
idk how muddy sounds like so I can't really tell
😂
oblique glacier
pls help voice cuts
low shard
pls help voice cuts
Are you talking about realtime voice changer for calls/games?
If so, wrong channel
oblique glacier
oh yea sorry
low shard
RVC = Retrieval-based-Voice-Conversion, the best Speech To Speech AI Models (on v2), Inferences (use models) pre-recorded audio (ai covers) and train (make) models
Wokada = uses RVC for realtime inference
low shard
oh yea sorry
Use #🔍│help-w-okada and elaborate it here
low shard
yeah i missclicked mb
It's fine dw brother
long trail
serene horizon
@knotty moth @simple ore This with batch size 5, 250 epoch, and custom KLM pretain. Is this an improvement?
simple ore
give it a test
serene horizon
give it a test
with the 250 epoch file?
simple ore
yeah
serene horizon
yeah
I'm getting alot of static before and after each line of singing, but the actual singing sounds clean.
latent cypress
is it okay if i add audio from youtube that has been upscaled using AudioSR to the dataset?
analog obsidian
is it okay if i add audio from youtube that has been upscaled using AudioSR to t...
no
just use the original audio
latent cypress
just use the original audio
so, is it fine if my dataset is a mix of lossless and lossy audio?
analog obsidian
so, is it fine if my dataset is a mix of lossless and lossy audio?
no idea, i never tried that
cedar canopy
why was MPS for mac ever canned on applio? i downloaded the older version that still has mps support and set the threads to 6 and everything works flawlessly (even after the first inference)
im getting 30 second inferences on an m4 with 24gb of unified memory (usage peaked at about 22gb which might be a concern? started at 13gb)
i think i was chatting with you two about it last time @simple ore @low shard
sorry for the @, but in case you wanted to know for anyone else whos wondering how to run these in apple silicon or something
regarding overheating i cant say that happened to me either, the fans didnt kick up after 4 inferences in a row
simple ore
mps torch blows
and whatever the math libraries there are hang up
there were multiple attempt to fix it, but the best we could do was just set opm_num_threads to 1 to prevent it from hanging up
you can manually re-enable mps
cedar canopy
right okay i see whats going on:
Hanging after the first inference: Memory leak/lack of garbage collection. During an inference python reserves 7-13gb of ram which keeps getting bumped up by about 4-5gb per run, which would cause weak 8gb mac configurations to hang right after the first one. On my 24gb machine i get to 20 ish on the 4th inference
Overheating: People think they can do AI on a fanless laptop and then complain when it gets hot, works fine on a mac with a fan
theres no inference running rn and its staying at 20 with nothing going on after starting sub 10
hope this info helps anyone who comes across it 
simple ore
I did add torch.cuda.empty_cache() after each inference, but it did not help
something else is leaking
cedar canopy
manually unloading the voice does make it go down by a fair bit again
knotty moth
right okay i see whats going on:
**Hanging after the first inference**: Memory ...
afaik the mainline rvc inference doesn't spawn some child processes upon starting inference, so it would bump the memory usage after a long audio inference
unique rock
Does Applio Colab work without the cell that needs to be created or do I still need to insert it?
!uv pip install pydantic==2.10.6
teal coral
hey, i have a lot of time that i dont use it rvc. and i noticed that i cant no use it in colab, i shall download it in my pc using phyton for try models, not for training
?
grim bay
Should I just keep this the same?
low shard
Does Applio Colab work without the cell that needs to be created or do I still n...
Iirc it should be already fixed
low shard
hey, i have a lot of time that i dont use it rvc. and i noticed that i cant no u...
You can use RVC in colab, it's just many colab notebooks are having issues
Google colab is not a platform dedicated for cloud (remote good PC) RVC, it's used for machine learning
The issue is related to google updates and the rvc notebooks creators, not colab
First, what's your PC GPU
You should use cloud only if your PC is bad
teal coral
mi gpu is NVIDIA Geforce RTX 2050
grim bay
Should I just keep this the same?
.-.
teal coral
You can use RVC in colab, it's just many colab notebooks are having issues
Goog...
Can u send me a link of a rvc colab notebook that is working?
knotty moth
mi gpu is NVIDIA Geforce RTX 2050
still enough for inference
and better performance than the comparable GTX ones
low shard
mi gpu is NVIDIA Geforce RTX 2050
are you looking for training or inference?
grim bay
Since Unwa's big mel roformer beta 4 and Mel roformer karaoke doesn't work like expected, does anyone remember the old recommedation of models for vocal isolation & cleaning?
Those new models just doesn't right
knotty moth
Since Unwa's big mel roformer beta 4 and Mel roformer karaoke doesn't work like ...
not the older one, but try Gabox vocal Fv4
there is no better karaoke model yet, or consider BVE v2 in uvronline.app (formerly x-minus)
in mvsep.com there is also male/female (duet) separation
latent cypress
is copying the latest release file enough to update Codename’s fork, or do I still need to run run-install.bat?
grim bay
there is no better karaoke model yet, or consider BVE v2 in uvronline.app (forme...
Well, the problem of Mel roformer karaoke is it removed some important parts of the vocals
That's why I need another model
knotty moth
Well, the problem of Mel roformer karaoke is it removed some important parts of ...
what "important parts"? some choruses, keyboard synths or guitars?
grim bay
what "important parts"? some choruses, keyboard synths or guitars?
The lyrics to be exact
knotty moth
The lyrics to be exact
I'm sure the vocals may not be centered
which is supposed to be removed after all
since it is called backing vocals
grim bay
I'm sure the vocals may not be centered
So which model should I use? I cannot upload the vocal here since it might against the rule of uploading copyright
knotty moth
So which model should I use? I cannot upload the vocal here since it might again...
you don't
grim bay
you don't
I remembered there was an old model that I can use that used to appear in the aihub docs
But I cannot remember what it is, really
Wait, is it HP Karaoke?
Yeah, I think that
Yeah, that worked
warm salmon
Hello, where can I train a better ai voice models?
copper terrace
knotty moth
Hello, where can I train a better ai voice models?
read the docs https://docs.aihub.gg/
jaunty quarry
so i have an issue right my rvc used to work with a response time of 1 sec now it doesnt work at all but it says its responding in 2.6 sec but even if i use my mic it doesnt work and ive tried changing version but it didnt help
knotty moth
so i have an issue right my rvc used to work with a response time of 1 sec now i...
please go to #🔍│help-w-okada and read the pinned guide there
jaunty quarry
please go to <#1159290161683767298> and read the pinned guide there
alright thanks
karmic oliveBOT
patent trellisBOT
jaunty quarry
please go to <#1159290161683767298> and read the pinned guide there
yeah nothing there seemed to help its just like when i start it my voice barely works and even after i stop it it keeps going for about 30 sec to a min
knotty moth
yeah nothing there seemed to help its just like when i start it my voice barely ...
you need a capable gpu as mentioned in the system requirements
jaunty quarry
you need a capable gpu as mentioned in the system requirements
i have 4060 Ti it was working great a couple days ago and for cpu i have i5 13th gen 13400F
knotty moth
so i have an issue right my rvc used to work with a response time of 1 sec now i...
uh mb if it is rvc inference
if the ui seems stuck, try restarting the program
jaunty quarry
uh mb if it is rvc inference
if the ui seems stuck, try restarting the program
its not that its like if i try to start it right and i talk its really laggy no matter what settings i use and when i stop it it keep going for about 1 min or 30 sec and i cant do anything with the program while its still going on
knotty moth
its not that its like if i try to start it right and i talk its really laggy no ...
make sure you're not running any cpu intensive game/program
also close/debloat unnecessary background processes
jaunty quarry
also close/debloat unnecessary background processes
already tried that didnt really help anything i think it might be becuase of something with my pc becuase a couple diffrent apps have been weird aswell so i dont think its an app issue so im planning on factory reseting soon to test that aswell
buoyant frost
ill dm u
elfin sand
i dont know how to download voice to my pc for okada any helps?
3060ti - i5 12400f
from weights
low shard
i dont know how to download voice to my pc for okada any helps?
3060ti - i5 124...
RVC = Retrieval-based-Voice-Conversion, the best Speech To Speech AI Models (on v2), Inferences (use models) pre-recorded audio (ai covers) and train (make) models
Wokada = uses RVC for realtime inference
which one do you want? wokada?
because this is help-rvc, so if you want wokada, you should ask in #🔍│help-w-okada
elfin sand
oh mb
analog obsidian
RVC = Retrieval-based-Voice-Conversion, the best Speech To Speech AI Models (on ...
blazing solar
Why did it says no file exist with the name gradio
low shard
Why did it says no file exist with the name gradio
elaborate
- what's your pc gpu
- what you want to do
- what tutorial link are you using
- what's the specific issue
low shard
<:KEKW:788766717379215410>
real
blazing solar
Colab on applio
low shard
Colab on applio
elaborate to all my questions please
simple ore
Why did it says no file exist with the name gradio
requirements were not installed properly
i assume you're using the UI colab
latent cypress
is copying the latest release file enough to update Codename’s fork, or do I sti...
hi, does anyone know the answer?
simple ore
hi, does anyone know the answer?
what latest release file?
latent cypress
what latest release file?
this file
simple ore
it is old and not compatible with latest refinegan
but if you want to just update what you have, yes, you can download the source and unzip it over
it is very unlikely you need all the advanced stuff from it anyway
use latest applio main
latent cypress
so i don't need to run the run-install.bat right?
simple ore
no
in most cases if you already have an installation, you just need to overwrite the code
in rare cases an extra module may need to be installed with pip
latent cypress
okay, thanks for answering my question, kinda confused to update the fork, but now i got the answer
latent cypress
in rare cases an extra module may need to be installed with pip
if that happen to me, i'll just go straight to the applio main branch😅
indigo dagger
i cant get rvc gui no matter what i try
pastel oak
<:KEKW:788766717379215410>
we need to make this a command
pastel oak
i cant get rvc gui no matter what i try
rvc gui is outdated and old use something more up to date
whats your gpu and what are you trying to do
analog obsidian
in rare cases an extra module may need to be installed with pip
is it normal that kaggle uses less steps despite using the same batch size?
first pic: batch size 12 local
second pic: batch size 6 kaggle
simple ore
2 gpus?
analog obsidian
yes
simple ore
so you use batch x2
analog obsidian
yes
simple ore
thus less steps
analog obsidian
but i used batch size 6 on kaggle
simple ore
hm
analog obsidian
wouldn't that have the same steps as bs12 local?
simple ore
please check you did
analog obsidian
please check you did
so i've found that kaggle does this if you want to use both gpus
first pic: 2gpus
second pic: 1 gpu
it doesn't extract the features of the whole set
unless you only use 1 gpu
local
kaggle
im using the latest main in kaggle, not 3.2.8
local is also the latest main branch
both have checkpointing enabled (because i only have 8gb of vram in local)
im gonna try using only one gpu
also got this when i stopped the training in kaggle
simple ore
so you have 329 files in sliced_audios_16k on both local and kaggle?
analog obsidian
so you have 329 files in sliced_audios_16k on both local and kaggle?
yes
simple ore
hm.. could be just some loader shenanigans
analog obsidian
hm.. could be just some loader shenanigans
using only 1 gpu matches local steps
i hope this gets fixed 
🙏
analog obsidian
so it's only a visual bug?
simple ore
I guess? is the result of the training much different?
yeah, I think it is fine
you have 15 steps on one node and 15 steps on another note working in parallel, total is 30 steps
which is rounded/ceiled from 329 / 2 nodes / 6 batch
analog obsidian
thats good to know
analog obsidian
which is rounded/ceiled from 329 / 2 nodes / 6 batch
but this does screw up tensorboard logging?
its less accurate?
simple ore
only the main process logs the values
so technically it will be using every other step to calculate averages and skip even steps
not a big deal
analog obsidian
indigo dagger
rvc gui is outdated and old use something more up to date
whats your gpu and wh...
just text to speech using a model but the models here are not safetensor so i don't know any other program, and my gpu is a rtx 4060, with 8 gb of vram
indigo dagger
Sorry I ment to say I want to make AI covers, I have found songs I want to have a different character sing
pastel oak
Sorry I ment to say I want to make AI covers, I have found songs I want to have ...
Ai covers is called Inference on there
indigo dagger
K thanks I'll take a look
gloomy lynx
Man, this used to be as simple as just booting up AICoverGen
Now no matter what tool I set up, they just give one error or another and refuse to boot up
Doesn't matter if it's old or new, local or cloud, none of them work
Tried Codename, that's a whole mess in itself, would post a screenshot of the console but I don't have the perms
simple ore
Man, this used to be as simple as just booting up AICoverGen
if you're using something not up to date, there''s very high chance some libraries got updated and no longer compatible with 1-2 year old code
gloomy lynx
That's in reference to the fact things just worked in the past; this is all with the dependencies being downloaded fresh
grim bay
Is HP Karaoke still usable?
viscid moss
Is HP Karaoke still usable?
nah
grim bay
in worse quality tho
So which model works similar to that?
I don't wanna use any new karaoke model since it just meh
knotty moth
I don't wanna use any new karaoke model since it just meh
it's just your point of view, what kind of track where the older models work better than the new one?
remember, you don't need to send the audio, just tell me what kind of it is
if not sure, listen carefully whether the vocals are centered or panned, and see the spectrogram for pitch difference
btw I bet you haven't tried UVR BVE
grim bay
remember, you don't need to send the audio, just tell me what kind of it is
Can I maybe name the song?
grim bay
btw I bet you haven't tried UVR BVE
Well, to be honest, I don't know where to find that model
knotty moth
Can I maybe name the song?
perhaps
tender tusk
Can someone help me? Every time I download the voice changer, it doesn't open, and when I click on the start, the command prompt opens and closes instantly.
knotty moth
It's Goodbye Sengen covered by Ayame
not sure but im used to most 70's-00's songs that dont have difficult harmonies and adlibs
unfortunately there are no decent polyphonic separation models better than melodyne afaik
knotty moth
Can someone help me? Every time I download the voice changer, it doesn't open, a...
please go to #🔍│help-w-okada and read the pinned guide there
hallow thistle
Can someone help me? Every time I download the voice changer, it doesn't open, a...
For W-Okada, go to #🔍│help-w-okada . This channel #✨│ai-help is about RVC programs and related tools, not the realtime voice changer.
Make sure you don't ask anyone to ask you back and then going away with no more response. 
latent cypress
So which model works similar to that?
melroformer dereverb by anvuew is really good at removing backing vocals, even though the model is called dereverb
grim bay
melroformer dereverb by anvuew is really good at removing backing vocals, even t...
I tried, and it doesn't work for the specific song
knotty moth
melroformer dereverb by anvuew is really good at removing backing vocals, even t...
it actually extracts the centered part of lead vocals, though may not exactly work like center channel extractor
latent cypress
I tried, and it doesn't work for the specific song
yes, every model can work perfectly depending on the song
it's very difficult to extract lead and background vocals, because what we have now is an old model that hasn't been updated.
knotty moth
the difficult parts are usually harmonies, adlibs, and perhaps some overlapped "medley"
knotty moth
I tried, and it doesn't work for the specific song
^tell me if that song has those parts
at last, you can try melodyne (polyphonic detection mode) and do cover on it despite possibly damaged quality
unique rock
yes, every model can work perfectly depending on the song
is it the Mel Band DeRverb v2 ?
latent cypress
is it the Mel Band DeRverb v2 ?
yes