The words behind the listening.

The process of estimating individual sound sources from a recording in which they are mixed together.

Why it matters. It is the central task behind isolating a voice, drums, bass, guitar, or piano from a finished song.

An audio track containing one part or a group of related parts, such as vocals, drums, or everything classed as other.

Why it matters. A stem name describes the requested grouping. It does not by itself say whether the audio came from a studio session or a model.

The combined recording given to a separation system, usually the finished stereo mix that a listener would play normally.

Why it matters. The model has to estimate hidden parts from this combined signal, where instruments often overlap.

A known source track or grouping supplied by the original multitrack session and used as a listening or evaluation reference.

Why it matters. It can describe the target, but it cannot demonstrate the quality of a model because it was not created by separation.

An audio estimate produced by a specific model build from a specific mixture and set of options.

Why it matters. A valid demonstration must identify the build and preserve the output so another person can inspect the same evidence.

Sound from an unwanted part that remains audible in the estimated stem.

Why it matters. Leakage can distract a musician or obscure a line even when the target part is otherwise clear.

A sound introduced or altered by processing, such as warbling, a metallic edge, a missing attack, or an unstable ambience.

Why it matters. Artefacts may affect naturalness and practical usefulness without being captured fully by a single numerical score.

A comparison in which listeners do not know which model or method produced each candidate while they make their judgement.

Why it matters. Hiding the identity reduces brand and expectation bias, but the prompt, ordering, and participant count still need to be reported.

Adjusting candidate clips to a comparable perceived or measured loudness before listening.

Why it matters. A louder clip can seem clearer or better even when the underlying separation is not better, so gain changes must be controlled and disclosed.