Stem Separation: The Quality Loss Nobody Talks About
AI stem separation is a miracle for remixes and sampling. But the artifacts it introduces are invisible to meters and obvious to ears. Here's what happens under the hood.
Stem Separation: The Quality Loss Nobody Talks About
Services like LALAL.AI, Moises, and the new integrated separators in DJ software have made stem extraction a commodity. Drop a track in, get vocals, drums, bass, and other out in 30 seconds. The quality has improved dramatically from the 2017 era of watery artifacts.
But here's the engineering reality: stem separation is not compression. It doesn't lose information in the entropy sense — it loses information in the neural sense. The model reconstructs what it thinks should be there. And it leaves artifacts that sound like "bad reverb" to untrained ears.
The three artifact types
Spectral leakage. When the model separates a vocal, it can't perfectly detach the vocal from the snare drum that plays at the same time. Some snare energy bleeds into the vocal stem. This sounds like a quiet reverb tail or a subtle shimmer on the vocal. It's not actually reverb — it's leakage.
Transient smearing. Drum separation is the worst culprit. The attack of a kick drum occupies the same frequency range as the sub-bass of a synth. Separating them requires the model to decide where the kick ends and the sub begins. The result is often a kick with a softer attack than the original and a sub-bass that's missing its low-end body.
Phase incoherence. Each separated stem is phase-independent because the model processes each source through a different neural pathway. When you sum all the stems back, the phase relationships don't reconstruct the original stereo image. This is why AI-separated stems sound "wide" in a way that doesn't fold properly to mono.
When it matters and when it doesn't
For sampling — pulling a vocal phrase from a track to use in a new context — the artifacts are largely irrelevant. You're going to process the sample anyway. Any reverb, EQ, or time-stretching you apply will mask the separation artifacts.
For remixing where you're rearranging stems, the transients and phase coherence matter because the stems interact with new elements. A separated bass track might fight with your synth sub in ways you'd never experience from the original multitrack.
For DJ stems — Serato's active use case — the quality is more than sufficient. The artifacts are masked by room acoustics, crowd noise, and the fact that you're mixing stems in real time without critical listening.
What the separation companies aren't telling you
Every separation model is trained on a dataset. The quality reflects the dataset's density in that genre. Pop, electronic, and hip-hop separate well because there's more training data. Classical, jazz, and live recordings separate poorly because the instruments overlap more in both frequency and time.
The marketing implies universal quality. The reality is genre-dependent.
One Thing to Try This Week
Take a full mix and run it through your preferred stem separation tool. Export the full set of separated stems, then reimport them into your DAW and sum them back to one track using only level matching (no EQ or compression). Null test the summed stem mix against the original full mix. A good separation should null to -25 dB or better. If you're above -15 dB, the artifacts are significant enough to affect a remix.