Music Production
Stem Separation: How It Works and Where It Fails
A mixed song is a single waveform. The individual parts were added together and can't be pulled apart exactly. Separation models estimate what each part probably sounded like. This guide explains that estimate, using the open-source four-stem model that runs in ANYANO's free separation tools.
Who this is for: For producers, DJs and remixers who separate stems from finished songs.
By ANYANO Editorial · Published · 8 min read
Why un-mixing is hard
When a bass and a kick play at the same time and the same low frequencies, their energy combines. Nothing in the mixed file records which part of that energy came from which instrument. A separation model has to guess, based on what it learned from thousands of songs where the parts were known.
How the model works
ANYANO's separation tools use HT-Demucs, an open-source model that predicts four stems: vocals, drums, bass and "other". It looks at the audio both as a waveform and as a spectrogram, which helps it handle sharp transients and sustained tones.
Long songs are processed in overlapping segments at 44.1 kHz, and the overlaps are blended back together so no seams are audible. In the free tools, all of this runs inside your browser tab after the model downloads once.
Where separation fails
| Artifact | Cause | Typical example |
|---|---|---|
| Bleed | Parts share frequencies | Hi-hat hiss in the vocal stem |
| Watery or phasey sound | Energy split between stems | Thin, swirling vocal on sustained notes |
| Lost reverb tails | Ambience assigned to the wrong stem | Dry vocal with reverb left in "other" |
| Merged instruments | Only four output stems | Guitar, piano and synths together in "other" |
| Backing vocals left behind | Harmonies blend with pads | Faint harmonies in the instrumental |
What separates well
Dense rock mixes, live recordings with room sound and heavily distorted material separate less cleanly.
- Modern studio productions with a clear lead vocal.
- Sparse arrangements where parts occupy different frequency ranges.
- Lossless or high-bitrate sources. MP3 compression artifacts get separated too.
Getting cleaner stems
- Start from the highest-quality file you have.
- Use the higher-quality setting when your device can handle it.
- Listen to each stem solo before using it. Artifacts hide in the full mix.
- Clean up with EQ and gating: high-pass the vocal and gate the drums.
- Credit and clear rights. Separating a song doesn't change who owns it.
Key takeaways
- Separation estimates parts. It can't recover the original multitracks.
- Four stems means everything that isn't vocals, drums or bass ends up in "other".
- Source quality matters. Artifacts in the file become artifacts in the stems.
Limitations
- Results vary by song. No separation model is artifact-free on every mix.
- Using a separated vocal or instrumental publicly still requires the rights holder's permission.