Skip to content

How stem separation works

Jam uses an AI separation model to split a recording into six parts: vocals, drums, bass, guitar, piano and other. The result is an estimate, so a little of one part can leak into another.

Last reviewed

The six parts

PartWhat it holds
VocalsLead and backing voices.
DrumsKit and most percussion.
BassBass guitar, upright bass, synth bass.
GuitarAcoustic and electric guitars.
PianoPiano and similar keyboards.
OtherEverything else, such as strings, horns, synths and effects.

The model was trained to recognise what each instrument sounds like. It rebuilds each part from the finished mix; it doesn’t have the original multitrack recording.

Why some sound leaks between parts

Instruments that play the same notes in the same range share frequencies, and the model has to guess which part each sound belongs to. You may hear:

  • a faint trace of the vocal in the guitar part (often called bleed);
  • a watery or “phasey” sound on a sustained note;
  • an instrument the model doesn’t know, such as an organ or a sitar, split between two parts or placed in Other.

What gives the best results

  • Clean source audio. Lossless or high-bitrate files separate better than low-quality or re-recorded audio.
  • Studio recordings. Live recordings with crowd noise or room reverb are harder.
  • Clear arrangements. Dense, heavily processed mixes leave more overlap.

What you can do about it

You can’t re-run separation with different settings, but the mixer helps: lower a part rather than muting it, or use a mixer preset that highlights one instrument (the mixer). See also separated parts sound wrong.

Was this article helpful?