What Are Stems in Music? A Beginner's Guide to Stem Separation

Tracks, stems and mixes: the vocabulary

When a song is recorded, each instrument and microphone usually gets its own track. A modern pop song might have 60 or more: separate kick, snare, overhead and room mics for the drums; double-tracked guitars; a dozen vocal layers; synths; effects.

A stem is a group of related tracks mixed down into a single stereo file. A typical set might be:

• Drums: every drum and percussion track
• Bass: bass guitar, synth bass, sub
• Vocals: lead and backing vocals with their effects
• Other / music: guitars, keys, strings, pads, everything else

The mix is all stems summed together, which is the stereo song you hear on streaming services. A key property of stems is that if you play them all at their original levels, they add back up to the mix.

Where stems are used

• Stem mastering: some mastering engineers prefer a handful of stems over one stereo file, so they can adjust the vocal or bass balance during mastering.
• Film and TV: music, dialogue and effects are delivered as separate stems so they can be re-balanced for different versions and languages.
• Remix releases: artists share official stems for remix contests.
• Live performance: bands play along with stems of parts they can't perform live.
• DJing: modern DJ software can mute or isolate stems on the fly.

How AI separates stems from a finished song

Official stems are rarely available. Instead, music source separation models estimate them from the finished mix. These neural networks are trained on datasets of songs for which the real stems exist. MUSDB18, a widely used research dataset of 150 songs, is split into exactly the four stems listed above. During training, the model hears the mix, tries to predict each stem, and is corrected against the real ones, millions of times over.

The model used on this site, HTDemucs v4 from Meta AI Research, processes audio in two ways at once: as a raw waveform (good at preserving sharp transients like drum hits) and as a spectrogram (good at separating pitched instruments). A transformer layer connects the two, letting the network use several seconds of musical context. More recent "band-split" RoFormer models push quality further, especially for vocals, but are considerably larger.

What AI stems are good, and not so good, at

Small amounts of bleed (traces of one instrument in another stem) are normal. In a mix they're usually inaudible; when soloed, you may hear them.

Things you can do with stems

1. Practice: mute the bass stem and play the bass part yourself, or loop the drum stem for timing practice.
2. Transcribe: solo the bass or "other" stem to hear notes buried in the mix.
3. Make karaoke tracks: combine everything except vocals, or simply use the Vocal Remover.
4. Remix: import stems into any DAW and build a new arrangement around the acapella.
5. Rebalance: turn the vocal down for a sing-along version, or the drums up for a workout mix.

To try it, open the Stem Splitter, drop in a song, and download the four stems as WAV for editing or MP3 for listening.

Stems and copyright

Separating a recording doesn't create new rights. The stems remain part of the original copyrighted recording and composition. Private practice and study are one thing; releasing remixes, samples or covers built from separated stems requires permission from the rights holders, unless an artist has explicitly released stems under a licence that allows it.

Frequently asked questions