It's how DJing with stems should work. Just pull up a track—any track—and instantly solo or mute the vocals, instrumentals, drums or bass. With the latest machine learning-powered version of their software, Serato may have finally made this a reality for DJs.

The Evolution of Source Separation: From EQ to AI

Even with just an analogue mixer, there have always been tricks to try to isolate vocals and instrumentals, such as using the filters and EQ. Or if you play the instrumental and vocal version of a track over one another with the phase inverted.

In 2015, Native Instruments' STEMS format tried a different approach, asking producers and labels to upload four independent stereo stems as part of the digital distribution process. Despite being labeled "open," though, support has been largely limited to TRAKTOR and a handful of tracks. The fidelity and flexibility are unmatched, but the musical choices are extremely limited.

Enter AI. Machine learning-based source separation works by "training" a model on a large data set of examples. That data model can then very quickly analyze stereo tracks and separate audio material into separate content, with categories like vocals, bass, drums and instrumentals. The draw: you can use any tracks you like.

Various free and paid tools offer AI-powered source separation in Web services and production hosts and plug-ins. Algoriddim was the first major DJ vendor to add real-time mixing of stems, with Neural Mixing in Djay Pro AI. Now, a similar feature comes to industry heavyweight Serato.

Stems for DJing, Ready for Primetime

Serato Stems is now available in both Serato DJ Pro 3.0 and Lite 3.0. The big difference with Serato isn't being first; it’s that the results work and sound far better than the alternatives.

The workflow is simple. You can opt to have Serato analyze all tracks for Stems separation or on-demand when you launch a track. The easier approach is to use a Stems Prepare Crate to batch-process only the files you need so they're ready for you once you start mixing, just as you'd do with tempo and key analysis. Once analyzed, you can break down tracks into two parts (acapella and instrumental) or four (vocal, melody, bass and drums).

Then the magic starts. Suddenly you can solo a vocal out of any track, at any point you like. You can try remixing a track, dropping out the "melody" section so there's a more spare production with just bass and drums—or any other combination. You can create wild mash-ups, instant a capellas, alternative mixes and blends, all on the fly.

What makes this more satisfying is how much better the audio quality is than the competition. I tested a range of music compared to other alternatives, including djay Pro AI, Deezer's Spleeter and browser-based LALAL.AI. The others often struggled to get consistent results, even with fairly sparsely mixed dance music with clear vocals. They either removed too much information, leaving stems with unpleasant artifacts, or too little, leaving in mashed-up bits of digital background sounds. On drums, that could make for unpleasant artificial sounds that removed punch; on vocals, the results might sound unintelligible or distorted.

Serato Stems, by contrast, was nearly impossible to trip up. Clear mixes with a lot of dynamic range still work best, and there were still some audible artifacts, especially depending on source material. But Serato's results were good enough that they usually sounded usable in a mix. Across genres, even with material that lacked clear drums or melodies, with vocals in other languages (including Arabic and Cantonese), Serato returned usable results. I got a workable a capella out of some Circassian ritual singing. In fact, the only vocals that didn't work were either completely buried in a mix, or vocoded. (The latter was more correctly identified as "melodic" anyway.)

You can't mix the level of individual stems–there's only mute for each, for now. But Serato has included vocal effects, with echoes and brake, which also help to cover any abrupt transitions. That adds another fun dimension for stems, as you can apply those effects to just a part of the mix, with more musicality and precision than filtering/EQ alone would deliver.

Hardware Integration

As usual with Serato, you'll get the most out of the experience if you have Serato-ready hardware. There's no separate add-on license for Stems support—any Serato Pro or Lite license includes Stems for free. But without hardware connected, you'll only see a capella and instrumental options on each deck. Once you do connect any hardware with pads, you'll enable all four stems plus effects for each. (If you just want to use a mouse, a Serato Play license will also work.)

A range of hardware maps pads to stem mute and effects by default, including the new entry-level Pioneer DDJ-FLX4. But any Serato-compatible DJ controller with pads will work—I tried a Roland DJ-202M and DJ-707M, and was able to make my own pad mapping using MIDI learn.

With dedicated pads, the fun really begins. Drop a bunch of tracks in the Stems Prepare Crate, then start jamming, trying out unusual combinations and experimenting. It'll surely be an instant hint for streamers on sites like Twitch. There's no way to export the results yet, though you can always record your session if there's a mash-up you want to save or take with you on a USB stick. A beta release of Serato's production tool, Serato Studio, also incorporates the same technology, so this could soon shake up remixing and production, too.

Our Take

Serato Stems is a revelation. The technology was already promising, but this is the first time it's genuinely usable, with results that at least some of the time would be usable in a club, and almost all of the time are fun to play with. It's also refreshing to see some AI applications that are pure fun and not frightening or dystopian. Here, machine learning enables something that simply wasn't possible before, separating parts of a track with results that sound closer to how you perceive them, and less like an analogue or digital process. You can make horrifying mash-ups, but also discover happy accidents. Really, the only danger of this technology is that you may lose some hours into the night messing around with your music library. It's gone beyond gimmick: this is a genuinely new way of mixing.