AI-powered stem separation — the ability to isolate vocals, drums, bass, and other instruments from a mixed audio file — has improved dramatically in the past two years. What used to produce ghostly, artefact-laden results now generates surprisingly clean separations.

I’ve tested the current crop of tools on three different tracks (a dense rock mix, a clean pop production, and an electronic track) to see how they stack up for practical production use.

The Tools Tested

Demucs (Meta/Facebook Research): Open-source, free, runs locally on your computer. The academic research model that many commercial tools are built on.

LALAL.AI: Commercial web-based service with a freemium model. Offers stem separation plus noise removal and other audio processing.

iZotope RX 11: The industry-standard audio repair suite, which now includes stem separation alongside its other tools. Not cheap at around $400 AUD for the standard version.

Moises.ai: Mobile and web-based service focused on musicians who want to isolate parts for practice or remixing.

BandLab’s Splitter: Free, web-based, integrated into BandLab’s platform.

Vocal Isolation

This is the most common use case, and it’s where all the tools perform best.

Demucs produced the cleanest vocal isolation across all three test tracks. The rock track, which had heavily layered guitars and dense arrangements, still yielded a vocal stem with minimal bleed. The algorithm has clearly been trained on a massive dataset, and it shows.

iZotope RX 11 was a close second. The vocal isolation was slightly less clean than Demucs on the rock track, but the built-in spectral editing tools mean you can manually clean up any remaining artefacts. For professional work, this combination of AI separation and manual editing is hard to beat.

LALAL.AI performed well on the pop track and electronic track but struggled more with the dense rock mix. Vocals were clear but with audible “breathing” artefacts in quieter passages.

Moises.ai was solid for casual use. The separations are good enough for practice tracks (where you want to remove or isolate vocals for karaoke or rehearsal purposes) but not quite clean enough for professional remixing.

BandLab’s Splitter was the weakest performer, which is expected given it’s completely free. Usable for basic purposes but noticeably artefact-heavy.

Drum Isolation

This is harder than vocal isolation, and the results are more variable.

Demucs again led the pack. Drum stems were clean with good transient preservation — the kick and snare attacks were intact, which is critical for rhythmic accuracy.

iZotope RX 11 was competitive. The drum isolation was slightly less clean than Demucs, but the precision tools available in RX for post-separation editing give it an edge in professional workflows.

The commercial web-based tools (LALAL.AI, Moises) produced usable but imperfect drum stems. Cymbal bleed and ghost notes from other instruments were present in most results.

Bass Isolation

Bass remains the most challenging instrument to isolate because low frequencies are often shared between bass, kick drum, and other instruments.

None of the tools produced a truly clean bass stem from the rock track. Demucs came closest, but there was noticeable kick drum bleed.

The pop track (with a cleaner, more separated mix to begin with) yielded better results across all tools. If your source material is well-produced pop or electronic music, bass isolation is workable. For dense live recordings, expect compromises.

Practical Applications

Remixing: If you’re creating a remix and need stems, AI separation has reached a point where it’s genuinely useful. The vocal and drum separations from Demucs or iZotope are clean enough for professional release in many cases.

Sampling: For producers who sample records, stem separation opens up creative possibilities that weren’t available before. Isolating a vocal phrase, a drum break, or a bass line from a mixed track is now practical rather than theoretical.

Live performance: DJs and live performers use stem separation to create acapellas, instrumental versions, and custom edits. Moises.ai is particularly popular in this space because of its mobile accessibility.

Music education and practice: Isolating or removing specific instruments from recordings allows musicians to practice along with songs. This is probably the most widespread use case.

The Legal Question

It’s worth noting that stem separation doesn’t change copyright law. Isolating vocals from a copyrighted recording doesn’t give you the right to use them in your own work without permission. If you’re using separated stems in commercial releases, make sure you’ve cleared the rights.

For practice, education, and personal use, stem separation is generally fine. For public release, get your samples cleared.

The Verdict

For professional production work, Demucs (free, local processing) and iZotope RX 11 (paid, comprehensive workflow) are the clear leaders. For casual use, Moises.ai and LALAL.AI are convenient and affordable.

The technology has crossed a threshold where it’s genuinely useful in real production scenarios. It’s not perfect — dense mixes still produce artefacts, and bass isolation remains challenging — but for vocal and drum separation, the results are remarkable.

The AI consultants in Sydney working with audio technology companies say we’re still in the early stages of what’s possible with AI audio processing. The next generation of tools will likely handle bass and other difficult separations much better. For now, what we have is already more than good enough to be useful.