Stems from Finished Tracks – Splitting Tracks with Spleeter

I know a lot of folks have been looking for track splitting technology as a holy grail for a while, and figured it’d be worth dropping a post. I originally posted this over at MPC Forums so if you’re looking for an MPC head’s take on this tech… check it out over that way.

First word of warning… this is super nerdy, and requires a fair bit of knowledge to get working. Here’s the video I accidentally stumbled upon that turned me on to this tech

If you haven’t seen anything about it, definitely check out Spleeter

At a high level, you give Spleeter a track, and Spleeter uses some trained neural network technology to spit out a bass stem, a drum stem, a vocal stem, and an other stem. If you know what you’re doing with it, I believe it can also get you a piano stem, and presumably it can be trained for other things… I digress… I’m lazy as heck, and don’t mind using somebody else’s work. The stems that it kicks out are pretty good, they’re not great, and depending on the track your milage may vary. To my ears this sounds like spectral band processing and you hear artifacts that sound like this in the tracks. Honestly though, it is still pretty good.

Now off to the races. Specifically, the Azuki Max4Live Spleeter Azuki put together a video showing off what she did here:

Azuki put together a Docker container that has all of the bits that you need to run this by itself, and a Max4Live device that can work on the selected track. So once you’ve got it setup and working it’s as easy as loading the plugin, selecting a track, and then telling it to process.

Installation is fairly straightforward if you are fairly technical. Azuki’s site has a $1 band camp donation link, but you don’t need to pay for this. She has a link where you can get it for free (and all this tech is free) at [url][/url]. So donate if you end up enjoying the work, but it definitely is not necessary.

What you’ll need:

  • Ableton Live
  • Max4Live
  • Docker

Given these parameters, you’ll need to be running either a Windows or OS X environment. For me, I got this working on an OS X Catalina deployment. I might try this on Windows later… but we’ll see.

You’ll notice that in the instructions it says to give Docker 16GB of memory. For small clips you can probably be able to get away with 2GB, but I found as soon as I started trying to process anything in the 4+ minute timeframe I needed more memory. This is an area where if you want it to work fast more memory and more CPU will probably help.

My first attempt at using this technology for real was in the BB 278 For my entry, I took the track and split it up into the drum, bass, vocal, and other sections. That was the only processing that I did on the PC, everything else I did on the MPC including normalizing, chopping, etc. While each track individually had that spectral sound, I’d say when you mix the whole track back together it really fades away.

Splitting the tracks made it very easy to adjust volumes of different parts individually, from filtering, normalization, and I’m actually quite happy with the results… especially given that this is free software.

Here is the track I ended up with

Anyways… if you’re looking for some new tricks to keep up your sleeve definitely check this out. You can probably get it to work without Ableton Live, but what Azuki put together is super slick, especially given that this is free.

I’m probably not going to be able to support people too much who are looking to set this up, but for those curious, definitely reply to the thread and I’ll do what I can.