Multiband SMR, ICC and ICLD Dataset of Grammy-Nominated Tracks (1995–2026)

Abstract

Description This repository contains the derived audio features, statistical results, robustness analyses, and Python source code accompanying: Long-Term Changes in Multiband Stereo Characteristics of Grammy-Nominated Popular Music Recordings (1995–2026) The study analyzes 536 Grammy-nominated recordings from 1995–2026 using: Side-to-Mid Ratio (SMR) Zero-lag inter-channel correlation coefficient (ICC) Inter-Channel Level Difference (ICLD) as an auxiliary descriptor The primary descriptors were calculated for broadband audio and ten frequency bands ranging from 20 to 21,950 Hz. Supplementary descriptors (lag-tolerant correlation, out-of-phase energy fraction, and within-recording variability) were calculated for the same bands. The corpus consists of two sampling periods: 1995–2020 (484 tracks), based on the annual Grammy Nominees compilation CDs, and 2021–2026 (52 tracks), based on Record of the Year nominees following discontinuation of the compilation series. The original audio recordings are not included because they remain protected by copyright. Naming conventions Year is the Grammy award year; Track is the source file name. ICC is the normalized zero-lag inter-channel cross-correlation (not the intraclass correlation). Legacy file and column names use "Phase" for this quantity. Band keys: 20_50, …, 6400_12800, 12800_21950 (also written 12800_Nyq in the extraction scripts; both denote 12,800–21,950 Hz), and Full for the unfiltered broadband signal. Column names in the released CSV files were harmonized after export. LagGain = small-lag correlation gain (G_small), LateAbs = late-lag absolute cross-correlation (C_late), Range90 = 5th–95th percentile range. LagGain is not stored in track_lag.csv; it equals IACC1abs − |CC0|. Repository Contents Track-level data and annual summaries grammy_ms_analysis_result.csv — track-level SMR grammy_phase_band_analysis_result.csv — track-level zero-lag ICC grammy_icld_band_analysis_result.csv — track-level ICLD and absolute ICLD grammy_ms_yearly_summary.csv, grammy_phase_band_yearly_summary.csv, grammy_icld_band_yearly_summary.csv — annual means and track counts annual_summary_with_variance.csv — annual means, SDs, and variance estimates track_lag.csv — lag-tolerant inter-channel correlation descriptors track_pan.csv, track_pan_hist_long.csv — time–frequency-bin panning descriptors, including the out-of-phase energy fraction track_timevar.csv — within-recording variability from 1-s frames track_third_octave.csv — 1/3-octave SMR, ICC, and ICLD (exploratory) Statistical results wls_regression_summary_corrected.csv — SMR annual WLS results wls_phase_band_regression_summary_corrected.csv — ICC annual WLS results grammy_icld_wls_regression_summary.csv, grammy_abs_icld_wls_regression_summary.csv — ICLD and absolute ICLD WLS results wls_results_all_periods.csv — WLS results for the supplementary descriptor families (1995–2026, 1995–2020, 2021–2026) Frequency-dependence and robustness analyses frequency_dependence_summary.csv — track-clustered and year-clustered omnibus Wald tests year_clustered_omnibus_results.csv — year-clustered Wald, small-cluster F-adjusted, and wild cluster bootstrap omnibus tests frequency_band_slopes.csv, temporal_dependence_sensitivity.csv, track_clustered_model_coefficients.csv — band-specific slopes and track- versus year-clustered inference regression_specification_sensitivity.csv — WLS, inverse-variance WLS, and OLS–HC3 comparison robustness_period_comparison.csv, robustness_period_comparison_holm.csv — sampling-period comparison robustness_leave_one_year_out.csv, robustness_recent_track_influence.csv — leave-one-year-out and single-recording influence robustness_weighting_sensitivity.csv — WLS versus unweighted OLS (exploratory) within_between_year_variability.csv — between-year and within-year variance decomposition Exploratory stereo-balance analysis (not reported in the manuscript) Short-time L/R energy balance based on 1-s windows with 0.5-s hops: grammy_stereo_balance_track_analysis.csv grammy_stereo_balance_yearly_summary.csv grammy_stereo_balance_wls_results.csv Python scripts stereo_smr_analysis.py — SMR extraction and annual WLS stereo_phase_analysis.py — zero-lag ICC extraction and annual WLS ICLD.py — ICLD and absolute ICLD Three_added_analyses.py — Year × Band interaction (track- and year-clustered) and weighting sensitivity year_clustered_omnibus.py — year-clustered omnibus test, F adjustment, wild cluster bootstrap (seed 20260101) 2021–2026 Robustness Analysis.py — period comparison, leave-one-year-out, recent-recording influence Within-Year vs Between-Year Variability.py — variance decomposition Supplementary_Stereo_Descriptors_Analysis.py — lag-tolerant correlation, panning/out-of-phase, 1/3-octave, within-recording variability Short-Time Panning Balance Proxy.py — short-time L/R balance (exploratory) Audio and output paths are set at the top of each script and must be edited before use. Figure-generation scripts are not included. Analysis Overview The pipeline includes multiband filtering (second-order Butterworth, zero-phase filtfilt, band edges ≈ −6 dB), SMR, ICC, and ICLD calculation, annual WLS regression (weights = annual track count) with Holm–Bonferroni correction, a track-level Year × Frequency-band interaction model with track- and year-clustered covariance, small-cluster and bootstrap omnibus tests, and robustness analyses. Welch-based and frame-level analyses use the effective power response |H|⁴ of the zero-phase filters. The strongest and most robust changes were in the 400–1600 Hz region, where SMR increased and zero-lag ICC decreased over time. The 400–800 and 800–1600 Hz focal bands were selected after inspecting band-wise results; the study is therefore exploratory. Supplementary descriptor families were corrected for multiplicity only within each family, and not all computed descriptors are reported in the manuscript. Comparisons between early and late years are confounded with the 2020/2021 change in sampling frame. Environment Verified with Python 3.13.14, NumPy 2.3.2, SciPy 1.16.1, statsmodels 0.14.5, SoundFile 0.13.1, pandas 2.3.2, and matplotlib 3.10.5 (macOS). Audio extraction was run in earlier sessions; the statistical scripts were re-run in this environment. Copyright Original FLAC recordings are not redistributed. Only derived numerical data, statistical outputs, and source code are provided. Citation Please cite both the associated publication and this Zenodo repository. Publication: Not published yet.