Curated listening examples for the three experiments reported in the paper.
All excerpts are drawn from the Cambridge Multitrack
Library. Three dense pop, rock, and metal excerpts (labelled April V1,
April V2, and It’s below, matching the
internal session names) are used throughout. Audio has been level-normalized and re-encoded as AAC
(~160 kbps) for web playback; this does not affect the relative comparisons.
Experiment 3 — Full-Mix Ablation (RQ3)
Single-stage full-mix models compared against their two-stage counterparts
(7-group grouping + ELL intra-group processing, same unchanged inter-group model). A human-mixed
reference is included as a perceptual anchor.
April V1
Condition
Single-stage
Two-stage
Human reference
MEGAMI
Diff-MST
April V2
Condition
Single-stage
Two-stage
Human reference
MEGAMI
Diff-MST
It's
Condition
Single-stage
Two-stage
Human reference
MEGAMI
Diff-MST
Experiment 2a — Compensation of Grouping Errors (RQ2)
Same intra-group processing and inter-group model; only the upstream
grouping strategy changes: the mixing-oriented 7-group scheme, the
4-group baseline, or instrument-based grouping (label-driven control,
typically more than 7 groups). See the grouping rules for definitions.
April V1
Grouping condition
MEGAMI
Diff-MST
7-group
4-group
Instrument-based
April V2
Grouping condition
MEGAMI
Diff-MST
7-group
4-group
Instrument-based
It's
Grouping condition
MEGAMI
Diff-MST
7-group
4-group
Instrument-based
Experiment 2b — Compensation of Loudness Errors (RQ2)
Fixed 7-group processing; only the intra-group loudness
relationship changes between with-balance and no-balance
conditions.
April V1
Balance condition
MEGAMI
Diff-MST
With balance
No balance
April V2
Balance condition
MEGAMI
Diff-MST
With balance
No balance
It's
Balance condition
MEGAMI
Diff-MST
With balance
No balance
Experiment 1 — Intra-group Mixing Quality (RQ1)
Can models trained for full mixing transfer to intra-group mixing?
One representative excerpt (April Drums) mixed by four methods:
ELL (rule-based equal local loudness), NoMix (unprocessed control),
MEGAMI, and Diff-MST.
April Drums
Condition
Audio
ELL (balanced)
NoMix (no balance)
MEGAMI
Diff-MST
Full Audio Set
These curated examples are one representative excerpt per condition per experiment. The complete set
of session exports (all excerpts, all repeated model samples) is not included here to keep the site
light; see Code & Data for the analysis scripts and full statistical results
behind every comparison.