deMIX Pro is pretty damn good at vocal separations! Check this out

QuadraphonicQuad

Help Support QuadraphonicQuad:

This site may earn a commission from merchant affiliate links, including eBay, Amazon, and others.

JonUrban

Forum Curmudgeon
Staff member
Admin
Moderator
Since 2002/2003
Joined
Mar 2, 2002
Messages
17,741
Location
Connecticut
Holy Crap. You all have to listen to this. It's only 45 seconds or so, but the results are amazing, even to my old rotted ears.

This is a vocal separation from Fleetwood Mac's "Everywhere". It's the lead vocal in the stereo fronts, and the background vocals in the stereo rears. No other audio is present.

This is how good the new version of deMIX Pro is, which I just spent $399 on. Wow. This is pretty amazing.

If you play it from Mega it will probably play in stereo, so it's best to download it and play it in your DAW. It's a 6 CH WAV file

It was too big to add to the post, and since I can't change that anymore, I posted it on my Mega. Here's the link:


https://mega.nz/file/aJQH0SjC#4HfygPs_kwHi7PFcmH-VKQPYq84_GvFBH_H0YjaFsPo
deMux Fleetwood.jpg
 
Demucs 4 has just been released, and it has been in a dev branch for awhile. Some of these commercial programs are using demucs in the background (you can tell by examining the names of the files in the programs folder).

https://github.com/facebookresearch/demucs
They went from an SDR score of 7.1, with v3, to 9.0 with v4.

I haven't checked it out yet. But yeah new stuff all the time now with these AI/ML based solutions.

Ultimate Vocal Remover GIU is supposed to have demucs 4, as a choice by the end of the month. That should be a pretty accessible (free and self contained) approach if you want to run locally vs. in the cloud with sites like Online music/voice separator based on neural nets (free) or lalal.ai (paid).

Great stuff!

How did Dmix Pro separate the lead from the background vocals? I haven't seen anyone offering that capability to date. Combination of de-panning and vocal separation?
 
How did Dmix Pro separate the lead from the background vocals? I haven't seen anyone offering that capability to date. Combination of de-panning and vocal separation?

I have no idea! I had the program for about 45 minutes and right clicked the extracted vocal stem, then selected something like "Separate Main Vocal from Background" or something like that (I'm not at my big computer right now) and it did it. Believe me, I am clueless with this software and in about an hour I had a full 5.1 surround version of "Everywhere" that sounded great to me, with the caveat that I cannot hear like I used to
 
That main vs. vocal split is quite impressive, Jon. Thanks for sharing this information. I also have not found any service such as LALAL.AI or MVSEP that can break out background vocals.

I have started using Demucs4 on MVSEP. Despite the higher rating that Glenn mentioned, I find the quality of the sound of the vocals to be just slightly better than Demucs3. However, I haven't done enough to compare the degree of extraction between the two. I did find that vocals extracted with the new Phoenix algorithm on LALAL.AI sound significantly better than Demucs3. Specifically, the Phoenix stem has less of a "grainy" sound to it.

Unfortunately, Phoenix had a significant issue extracting the vocals on the McCartney song Another Day. It left out several seconds of the vocals and what was missing did not fully remain with the instrumental track. LALAL.AI tech support (which is quite responsive) couldn't do anything about it. See comparison below:

LALAL vs MVSEP [Another Day].jpg


(As Glenn can tell you, I have the uncanny ability to choose the worst possible song when first testing out software. LOL)
 
Listening to the stereo source of Everywhere I do think they must be using the position in the stereo field (panning) to separate the backing vocals from the lead.

Lead is in the center, backing vocals are on the sides.

So, what you would do is first split the stereo using an upmix tool, one of mine or penteo, or just centercut, etc. Using whatever knobs are there just to get good separation between the lead, vs. backing vocals, then split those channels from the upmix and send them (separately) to the vocal separator software.

e.g. C --> separate vocals = lead vocal stem (or if you prefer c + fronts --> separate vocals = lead vocal stem) and rears --> separate vocals = backing vocals stem.

Hope that's clear.

On Everywhere I bet you could get at least three vocal parts that way, first upmixing and then sending C, fronts, and rears, separately to be vocal separated from the music.

However, I would point out that if you just sent the original stereo --> separate vocal = all vocals stem --> upmix you'd end up in the same place. So the above method probably is only needed if you wanted to pan the different vocal parts differently than where a straight upmix would place them.
 
Listening to the stereo source of Everywhere I do think they must be using the position in the stereo field (panning) to separate the backing vocals from the lead.

Lead is in the center, backing vocals are on the sides.

So, what you would do is first split the stereo using an upmix tool, one of mine or penteo, or just centercut, etc. Using whatever knobs are there just to get good separation between the lead, vs. backing vocals, then split those channels from the upmix and send them (separately) to the vocal separator software.

e.g. C --> separate vocals = lead vocal stem (or if you prefer c + fronts --> separate vocals = lead vocal stem) and rears --> separate vocals = backing vocals stem.

Hope that's clear.

On Everywhere I bet you could get at least three vocal parts that way, first upmixing and then sending C, fronts, and rears, separately to be vocal separated from the music.

However, I would point out that if you just sent the original stereo --> separate vocal = all vocals stem --> upmix you'd end up in the same place. So the above method probably is only needed if you wanted to pan the different vocal parts differently than where a straight upmix would place them.
Interesting point. I've found it quite easy to separate backing from lead vocals when the lead is in the center and backing vocals are on the sides. In Audacity its just a matter of taking the vocal stem and using the Remove Center effect.

I'm wondering if deMix Pro can separate the lead and backing vocals if both are embedded in a mono track.
 
Unfortunately, Phoenix had a significant issue extracting the vocals on the McCartney song Another Day. It left out several seconds of the vocals and what was missing did not fully remain with the instrumental track. LALAL.AI tech support (which is quite responsive) couldn't do anything about it.

I did 'Another Day', from the HiRex flac file from the McCartney Archive series. There are artifacts, but the extraction revealed some parts I've never heard solo, especially Paul's harmony lines. Check it out. Stereo flac of the background vocals, less than a minute sample, taken from towards the end of the tune
 

Attachments

  • Sample.flac
    3.2 MB · Views: 0
UPDATE: I just tried this on a mono vocal track, and the program was able to pull the lead vocal out of the background vocal, leaving an isolated lead vocal track and an isolated background vocal track. I was sorta shocked that it worked. Very impressive. This really opens up a whole new dimension to upmixing!
 
I’m intrigued by this for learning parts of songs. I dabble in a number of different instruments (none played really well) and it’s nice to be able to have instruments separated to hear nuances or even parts that are a bit buried in the mix.

I’ve been trying MVSEP.com and have enjoyed the results. But I’m always interested in other tools that do a similar task.
 
UPDATE: I just tried this on a mono vocal track, and the program was able to pull the lead vocal out of the background vocal, leaving an isolated lead vocal track and an isolated background vocal track. I was sorta shocked that it worked. Very impressive. This really opens up a whole new dimension to upmixing!

I'm surprised as well. Is there anything in the documentation then mentions how they do that?
 
UPDATE: I just tried this on a mono vocal track, and the program was able to pull the lead vocal out of the background vocal, leaving an isolated lead vocal track and an isolated background vocal track. I was sorta shocked that it worked. Very impressive. This really opens up a whole new dimension to upmixing!
Yes, very encouraging Jon, thank you!

Helps with 2 of my main Penteo pet-peeves (in post up-mix) in separating some main vocals and background. Now if they can just figure out stuff like wide drum kit cymbals out of the rears better, that would also really help (any processes for that yet?)
 
Here's a link to a file result that John (Ar Surround) sent to me to test out. It's some Beatles mono snippets that I ran through the program to see if it would pull out the background vocals.

IT DID!!

NOTE: This is a stereo file with the lead vocal in the left and the backgrounds in the right

https://mega.nz/file/LARDSJia#GfNuzODyTYkP53YtmqQJpcu-3TWXnFB5QvlJjBLFPIE
teate.jpg
 
Check this out. It's a bit from the Beatles "I'll Follow the Sun"

It's just the vocals, and I did the vocal split on them to see if it would separate John from Paul on the harmonies. It did!!

But it didn't quite get it perfect, as it took John's vocals to be the lead, and Paul's to be the background, only when they both sing together.

It's quite remarkable. Listen to this file. It's stereo lead vocals in the front channels, stereo background vocals in the rear channels, empty center and LFE

Sun.jpg
 

Attachments

  • Sample.flac
    2.2 MB · Views: 0
Check this out. It's a bit from the Beatles "I'll Follow the Sun"

It's just the vocals, and I did the vocal split on them to see if it would separate John from Paul on the harmonies. It did!!

But it didn't quite get it perfect, as it took John's vocals to be the lead, and Paul's to be the background, only when they both sing together.

It's quite remarkable. Listen to this file. It's stereo lead vocals in the front channels, stereo background vocals in the rear channels, empty center and LFE

View attachment 85511
This is a pretty cool new toy you're getting to play with. And it's not even Christmas!
 
Wow Jon.....This sample sounds incredible.....Thanks for sharing....I may need to purchase this software to have some fun myself! :)

It's so easy to use. I bet there's a lot more to it than I have found so far. I thought that Beatles thing was pretty amazing. It would be very east to swap the audio from one channel to another to get John in the back and Paul in the front using an editor. Many possibilities with this thing
 
It's so easy to use. I bet there's a lot more to it than I have found so far. I thought that Beatles thing was pretty amazing. It would be very east to swap the audio from one channel to another to get John in the back and Paul in the front using an editor. Many possibilities with this thing
I bought DeMIX Pro yesterday and got it up and running this afternoon. I made a nice frontal spread of the vocals on The Beatles Chains. I put George's lead in the LF and RF which renders as a mono center. Then I took the extracted harmonies, stereotized them, and put them in the LF and RF. Finally, I took the entire vocal track and put it in the Center Channel to anchor the whole presentation and reduce the perception of artifacts. It sounds good through headphones. How it renders on the big system remains to be seen.

I don't know if putting the backing vocals in the rear speakers by themselves will work, but I think that having them in both the fronts and rears might mask any artifacts.

Nice new toy.

Screen Shot 2022-11-21 at 2.51.24 PM.jpg
 
Back
Top