Stems Explained: How AI Stem Separation Works for Producers

What Are Stems?

Stems are individual audio tracks isolated from a mixed recording. A full song can be separated into stems like:

• Vocals — lead and backing vocals
• Drums — kick, snare, hi-hat, cymbals, percussion
• Bass — bass guitar, synth bass, upright bass
• Guitar — electric, acoustic, rhythm, lead
• Piano/Keys — piano, organ, synth pads
• Other — everything else (horns, strings, effects)

Stems let producers isolate the exact elements they want to sample — grab just the drum break, just the bass line, or just the vocal hook.

How AI Stem Separation Works

Modern AI models can separate a mixed audio file into individual stems without needing the original multitrack session. The most widely used model is Demucs, developed by Meta/Facebook Research.

Here's how it works at a high level:

1. The mixed audio is converted to a spectrogram (visual representation of frequencies over time)
2. A deep neural network, trained on thousands of songs with known stems, learns to identify which frequencies belong to which instrument
3. The model generates separate spectrograms for each source
4. Each spectrogram is converted back to audio — producing isolated stems

Demucs uses a hybrid architecture (both time-domain and frequency-domain) for higher quality separation. The latest version (htdemucs_6s) separates into 6 sources: vocals, drums, bass, guitar, piano, and other.

What Selekt Audio Does Differently

Most stem separation tools give you 2 or 4 stems. Selekt Audio runs a multi-model pipeline:

• Demucs 6-source — vocals, drums, bass, guitar, piano, other
• DrumSep — further splits the drum stem into kick, snare, hi-hat, toms, ride, crash
• Wind/Woodwind isolation — extracts saxophone, clarinet, flute, trumpet, oboe from the "other" stem

Every sample in the library has pre-separated stems available for streaming and download — no processing wait time.

How Producers Use Stems

• Sampling — Isolate a drum break, vocal phrase, or bass line from a vintage recording and build a new beat around it
• Remixing — Rearrange or reimagine a track using its individual elements
• Layering — Combine stems from different sources to create unique textures
• Practice — Remove vocals to create karaoke tracks, or isolate an instrument to learn a part
• Sound design — Process individual stems through effects, pitch-shifting, or time-stretching

Quality Considerations

AI separation is not perfect — you may hear artifacts like bleed (traces of other instruments) or slight audio degradation. Quality depends on:

• The source recording quality
• How densely mixed the original track is
• The separation model and its configuration
• The type of instrument being isolated

Drums and vocals typically separate cleanly. Instruments with overlapping frequency ranges (guitar and piano, for example) can be harder to isolate.

Try It Yourself

Selekt Audio's Stem Separator tool lets you upload any audio file and get AI-separated stems — free, no signup required. Or browse the full library where every track already has pre-separated stems ready to stream and download.