I am impressed with UVR5 no Spectralayers or other cant match in quality and speed of export... I mean, come on, it extracts vocals with effects like reverb and delay... Insane, and program is free.. I was skeptical about it but, now.... Man...

 

Hi Jarredou, I use the paid version of your Colab MVSep-MDX23. It's by far the best stem isolation I've ever heard and I speak for all my friends that we thank you for creating this miracle! Unfortunately I've been unable to successfully use the new 2.3 version that was announced. I'm a music researcher but I don't understand how this all works. I have sort of a "dummy's approach" where I got to this address (https://colab.research.google.com/g...23-Colab_v2/blob/main/MVSep-MDX23-Colab.ipynb) and then I lower the chunk size and specify the folders on gdrive. It usually works but when I tried the 2.3 version, it showed only vocal and instrumental. I'm mostly interesting in rhythm section parts - your MDX23 stems for drums and bass are just amazingly clean and beautiful. Anyway, I panicked when I tried the new version and couldn't get back to being able to get drum stems. I'm not sure how I got back to normal but I did so I'm okay again, but I'm still curious about this new version and any other tips that might help me get better (or faster) results, or just to understand how all this magic works. Have you done any tutorials? If you did some YouTube videos where you explain about this I'm sure thousands of people like me would love to watch them. Anyway, thanks in advance for any help and instruction you can provide! For example, it might help to know what the chunk size and overlap settings do and how to get the best results. Also I don't understand why there's "instrum" and "instrum2". They seem to do exactly the same thing. Someone said that instrum is everything minus the vocals and the instrum2 is drums+bass+other, but that always turns out to be the same thing!

 

My fork runs greatly with free colab T4 GPUs, you don't need to pay to use it.

To keep up to date version, use the link provided here: https://github.com/jarredou/MVSEP-MDX23-Colab_v2/
There's no more chunk_size in settings as the chunking code was fully reworked (better quality, faster processsing and using less memory, joint work with Anjok from UVR)

For the 4-stems separation you have to uncheck "vocals_instru_only", which is enabled by default.



Instrum/instrum2 were part of original script by mvsep and are created as you were told. Instrum2 can have less vocal bleed then instrum, but most of the time, differences are not audible.

I haven't done any tutorial/videos, but I've answered lots of questions in Audio Separation discord server. Default settings are already optimized for good balance between speed and quality.

I've added some hints in latest version to explain quickly the settings.

 

Last edited by a moderator: Jan 17, 2024

I finally got around to installing UVR and learning it as I needed to separate things so that's the perfect opportunity to figure things out. Thank you for this thread, even though I didn't go through all of the algorithms the end results very impressive using the ensemble mode!

So, yeah, I don't know if it's the best, but it's so far the most configurable and the best I've heard.

 

so, I did some tests, and it seems old nVidia GTX1080Ti with old i7-5820K is around 14% faster than Apple Mac Studio M2 Max with 64GB memory

 

x-minus. best of best.

 

Hi guys!
A couple questions from noob.

1. I have tested stem separation algos and demucs4 ft sounds best for me. Even betyer than Ensemble of demucs and mdx. Should I search for the better models for UVR5? I have heard about BS-Reformer and mdx23. Does it exist in the form of ready to use model for UVR?
2. I want to train my own model for de-reverb and clean audio. Is there step by step guide? I am not familiar with python. I have trained models for RVC2.
3. Base models in UVR5 for denoise and deverb is the best? Or better ready to use models existing?

Thank you in advance!

 

If you haven't tried Jarredou's Colab MVSep-MDX23 version, you have yourself to blame. It's superfast and the result is incredible.

 

@itaudio . Could you please do a small tutorial , what to download and how to install Jarredou´s Colab MVSep-MDX23 into UVR ver. 5.6.0 ?

I only found one " MDX23C-InstVoc HQ " in the downloadcenter of UVR.
...dont think thats the right one.

Just got started with this steam-seperation-thingie , and got a hang of how UVR works.
Did not find any manual and -on how to add external models into UVR.

Already had a faboulus result ( for my expactations) with
Kimvocal 2 + MDX-Net Inst hq + htdemucs_ft in Ensemble mode.

--------
Edit: oh , now I see it´s a paid model. sry for bother.

 

Last edited: Feb 25, 2024

Where does one pay for and download this model?

 

Kimvocal 2 is the cleanest by far and the best to grab vocals for training in Applio V3.08

I use Kimvocal 2 then once ripped take the vocal track into Spectral layers 10 to clean and remove reverb and unwanted multilayers from a UV5 Kimvocal 2 seperated vocal
as you can delete on the fly sections within spectral layers 10 as you dont want over processed chorus or unison vocal for training vocals in Applio v 3.08
To be honest to train a vocal in Applio 3.08 you only need 30 to 45 seconds of single layered verse segment from a lead vocal from any song to get a decent result with Nvidia Cuda supported GPU with minimum 8 gig of ram

I use AMD Ryzen 5800 with 32gig
Nvidia GTX1080 - 8 gig

clean as possible 30 to 45 second lead vocal segment exported to wav from inside Spectral layers 10

setting for Applio v3.08

load the 30 to 45 second lead vocal segment and process the features section within Applio 3.08
set to Rvmpe
hop lenght dont know is this is needed but I set the hop length to 32 default is 128
set 100 epochs
once loaded into Applio 3.08 and the training is underway it takes approx 10 minutes to complete the training
Good hunting

 

Last edited: Mar 6, 2024

It's not a paid thing, and not a model you can't add to UVR, it's a python notebook to use on google colab made with public models (including that MDX23C-InstVoc HQ one) + some extra features that UVR doesn't provide (VitLarge23 model / (multiband) weighted ensembles / big shifts trick).

It's freely accessible here https://github.com/jarredou/MVSEP-MDX23-Colab_v2/

(you can also use inference.py as python CLI, locally)

 

Last edited by a moderator: Mar 12, 2024

Jarredou I have tried your model it wants access to my emails and the stuff I have on Google drive?
I can never get these things to work, python programs or Colab
I downloaded
MVSep-MDX23 Colab Fork v2.3 with the GUI and it will not install
and will not work, some syntax continuation error
When I open a command prompt a window flashes for a second and then nothing
where am I going wrong ? I can usually figure out my way through this stuff ,but I am lost with this
Thanks for any help

 

It's still not a model and it can't access your emails. The GUI is the python notebook to use on google colab, else you can use inference.py as python CLI, locally.

 

Jarredou Thanks for getting back to me

so I open Python 3.12 , at the prompt

>>> C:\MVSep-MDX23\inference.py
File "<stdin>", line 1
C:\MVSep-MDX23\inference.py
^
SyntaxError: unexpected character after line continuation character


Command Prompt

C:\Users\LISA>C:\MVSep-MDX23\inference.py
GPU use: 0
Traceback (most recent call last):
File "C:\MVSep-MDX23\inference.py", line 12, in <module>
import numpy as np
ModuleNotFoundError: No module named 'numpy'

I am not a programmer ,so if there is some secret language I am supposed to understand or know, I do not

and how do I get your program to run in colab any direction any help appreciated

Thanks
>>>

 

The colab notebook isnt data-mining your emails or trying to spy on you or hack you.
You need to grant the notebook access to your google Drive, so that when the Colab is finished rendering the stems, it has a place to export/save the rendered stems to.
It's not a security/safety liability.
You also dont need to be a hacker, it's pretty straightforward. Just drop the files you want Stemmed into a folder (tell it which folder on your google drive), and likewise output the rendered stems to either the default or custom folder path you want.
EZ Peazy

 

THANK You Dyslexicon

for your help ,when you cannot understand the code or what it is doing ,you need to be careful

 

well tried everything python ,,colab none of it works ,,starting to think something else is going on,
none of it works
/content/MVSEP-MDX23-Colab_v2
/bin/bash: -c: line 1: unexpected EOF while looking for matching ``'
/bin/bash: -c: line 2: syntax error: unexpected end of file ,,,,,,,,,,, none of it works

Tried again

[Errno 2] No such file or directory: '/content/MVSEP-MDX23-Colab_v2'
/content
python3: can't open file '/content/inference.py': [Errno 2] No such file or directory

I know this should not be this difficult ,programmers think differently ,do I need to upload the entire python file to my google drive ?

 

Last edited: Mar 13, 2024

Open https://colab.research.google.com/g...23-Colab_v2/blob/v2.3/MVSep-MDX23-Colab.ipynb

Run first install cell, allow access to drive. It's only used if you read audio and/or saved the separated results in gdrive. nothing more. the code is public and was reviewed/tweaked multiple times by multiple people. It does only audio source separation.

Once install is done, it takes around 1m30s, set your input, a direct link to an audio file or to a folder containing audio files (and it will process all the files in it). Set your output folder (better to have it on gdrive as colab virtual hard drive space is deleted when you disconnect, you would loose your separated files + download speed is better with gdrive).

Tweak the settings if you want, default ones are already well balanced between quality and processing time.

Run the cell, get your files and stop ranting like this with false allegations of being a data thief just because you can't run a python script, please. Or just use UVR.

 

Last edited by a moderator: Mar 13, 2024