Grouping Rules
Definitions of the three grouping strategies used in the two-stage analysis framework (paper Section 2.2).
A grouping function partitions the input tracks X = {x₁, …, xₙ} into functional groups before intra-group processing. We compare three grouping strategies, used as the upstream condition in Experiment 2a and as the two-stage input in Experiment 3.
1. Four-group baseline
A widely used grouping scheme in prior work and multitrack datasets (e.g. MedleyDB), consisting of four coarse categories:
| Group | Contents |
|---|---|
| Bass | Bass instruments. |
| Drums | All percussive / drum-kit elements. |
| Vocal | All vocal tracks. |
| Other | Everything else (harmonic and melodic instruments, pads, effects). |
2. Seven-group mixing-oriented scheme
Designed specifically for group-based processing in automatic mixing. It starts from the 4-group organization and introduces functional distinctions relevant to mixing, while keeping the overall design as simple as possible.
| Group | Definition |
|---|---|
| Low-Frequency | The Bass category, kept as a dedicated group. |
| Low–Mid Percussion | Kick drums, snare drums, toms, and related low/mid-frequency percussive components (split from the original Drums category). |
| High Percussion | Hi-hats, cymbals, overheads, and other high-frequency percussive components (split from the original Drums category). |
| Lead | Lead vocals, solo instruments, principal melodies, and other perceptually dominant foreground tracks (reassigned from Vocal / Other). |
| Accompaniment | Backing and harmony vocals, plus instruments providing harmonic support, rhythmic accompaniment, or arrangement-level support (reassigned from Vocal / Other). |
| Pad | Sustained background layers, spatial textures, and tracks with weak attacks or limited rhythmic definition (reassigned from Other). |
| Other (residual) | Small effects, transitional sounds, and exceptional tracks that do not satisfy the definitions of the preceding groups. |
Relationship to the 4-group baseline
| 4-group category | Corresponding 7-group assignment |
|---|---|
| Bass | Low-Frequency |
| Drums | Low–Mid Percussion; High Percussion |
| Vocal | Lead; Accompaniment |
| Other | Lead; Accompaniment; Pad; Other |
Vocal and Other tracks from the 4-group scheme are reassigned to the 7-group scheme according to their functional role in the mix, not their source category — e.g. a backing instrument originally in “Other” may become part of Accompaniment or Pad, and a solo instrument may become Lead.
3. Instrument-based grouping (control condition)
A label-driven grouping derived directly from track/instrument metadata, used as a control condition against the mixing-oriented 7-group scheme (this is the grouping strategy released with MEGAMI). Unlike the 7-group scheme, tracks are grouped by instrument name rather than by functional role in the mix, and typically produces more than seven groups per session.
Where each scheme is used
| Experiment | Grouping conditions compared |
|---|---|
| Experiment 2a — grouping-error compensation | Instrument-based, 4-group, 7-group (fixed intra-group processing / inter-group model) |
| Experiment 3 — full-mix ablation | 7-group + ELL as the two-stage input, vs. the single-stage full-mix baseline |