Code & Data

Statistical analysis code, full paired-comparison results, and pointers to the system implementations used in the study.

Analysis code

analysis/grouped_experiment_analysis.py is the script used to run the paired t-tests (with FDR-BH correction) and generate the grouped figures for Experiments 1–3.

python3 analysis/grouped_experiment_analysis.py

It expects a listener-level scores table with the same schema as listening_scores_dataframe.csv below (columns: source_file, participant_id, row_in_sheet, question_no, question_name, score, model_name).

Statistical results (full pairwise tables)

The paper's Table 2 reports only the comparisons discussed in the narrative. The CSVs below contain every paired comparison within each experiment: pooled means, paired t-statistics, raw and FDR-BH corrected q-values, paired Cohen's d_z, and the number of valid paired observations.

File	Experiment
exp1_intra_group_quality_paired_ttests.csv	Exp 1 — intra-group mixing quality (RQ1)
exp2a_grouping_compensation_paired_ttests.csv	Exp 2a — grouping-error compensation (RQ2)
exp2b_loudness_compensation_paired_ttests.csv	Exp 2b — loudness-error compensation (RQ2)
exp3_full_mix_ablation_paired_ttests.csv	Exp 3 — full-mix ablation (RQ3)
listening_scores_dataframe.csv	Anonymized listener-level scores used for all analyses (participant IDs replaced with L01, L02, …)
q10_q13_consistency.csv	Intra-rater reliability on hidden repeated trials (Pearson / Spearman / Kendall, MAE, RMSE)

Figures

Intra-group mixing quality box plot — Figure 2 — Intra-group mixing quality (Experiment 1).

Grouping and loudness error compensation box plot — Figure 3 — Compensation of grouping and loudness errors (Experiment 2).

Full-mix ablation box plot — Figure 4 — Full-mix quality, two-stage vs. single-stage (Experiment 3).

System implementations

This repository hosts the analysis code and audio examples. The mixing systems compared in the study live in their own repositories:

Automatic-level-balance-for-multitrack-music
Multi-track level-balancing tool supporting AIFF / FLAC / MP3 / OGG / WAV input, producing a final mixdown through several level-balance strategies.

ELL — intra-group balancing method
tracks-auto-grouping-using-CLAP
CLAP-based automatic track-grouping exploration, related to functional grouping of multitrack sessions.

Grouping exploration
Diff-MST
Fork of sai-soum/Diff-MST — multitrack mixing style transfer via a differentiable mixing console, used as one of the full-mix / intra-group / inter-group models evaluated in the paper.

Diff-MST model
cambridge-mt_scrapper
Fork of sai-soum/cambridge-mt_scrapper — downloads mixes and metadata from the Cambridge Multitrack forum, used to source the evaluated excerpts.

Data sourcing
AES2025-materials
Related hybrid signal-processing + neural style-transfer approach targeting scalable automatic mixing for large, un-predefined track counts.

Related prior work

MEGAMI (Moliner et al., 2025) is used via its released checkpoint and is not reproduced in this repository; see the paper's bibliography for the original source.