Validation

DEBUSSY's parameter set was validated on a 60-track benchmark spanning the stimulus families used in the BELL-001 protocol (breath / ambient / musical excerpts). The benchmark exists to check two things:

1. Numerical stability — every parameter is finite and reproducible across repeated runs and across the 44.1 kHz / 48 kHz sample rates commonly found in stimulus libraries (mosqito resamples internally to 48 kHz; see the resample notes).
2. Discrimination — the parameters separate acoustically different material in the directions predicted by the psychoacoustic literature (e.g. higher roughness and sharpness for harsher excerpts).

The 60-track benchmark

Three categories, all standardised to 44.1 kHz mono 16-bit PCM WAV before analysis:

Category	Material	n
A — "relaxation"	DEAM static-annotation corpus, low-arousal subset	10
B — "frequently listened"	DEAM mid-to-high arousal + genre-stratified FMA-medium subset	30
C — clinical	BELL-001 SleepThera breath-paced biofeedback stimuli	20

The pipeline produced a complete set of measured values for 58 of 60 tracks. The two exceptions are ~10 s breath clips with no autocorrelation peak above the voicing gate, so HNR is undefined for them rather than wrong.

Two commercial reference lists are excluded pending separate licensing; adding them would bring group A to n=30. Tracking issue: #9.

What the benchmark shows

The A-vs-B contrast is exploratory and underpowered. Three parameters reach uncorrected p < 0.05, but none survive Benjamini–Hochberg correction across the twelve tests (smallest q = .19), and group A is small.

Do not read the A-vs-B effect sizes as findings

At n=10 versus n=30 the δ values are unstable in sign, not merely in magnitude. Redrawn at 150 per group within the same arousal bands — so the only thing that changed is n — eleven of twelve parameters separate, and four reverse sign: LAeq, spectral centroid, sharpness and spectral flatness all move from positive to strongly negative. A second draw at the arousal extremes and an independent corpus (FMA) agree on the direction.

Simulated power explains the discrepancy rather than excusing it: at the effect sizes actually present, the published design had 11–30 % power, so its null was uninformative and its signs were close to coin flips.

Run python validation/sensitivity_power.py --power for the full table. The 60-track benchmark itself is unchanged; this is an additional analysis, not a correction to the released data.

Category C separates far more sharply, with very large effect sizes (|δ| ≥ 0.7) against both music categories on seven of the twelve measured quantities — LAeq, dynamic range, attack-time median, roughness, sharpness, spectral slope β and spectral flatness. Crest factor is close behind (δ = +0.81 against A, +0.68 against B). This is the acoustic fingerprint expected of quiet, dynamically wide, sharp-onset breath recordings, and it is detected across all four family modules.

The point of the benchmark is not a substantive claim about music or breath stimuli — that belongs to the companion review. It is evidence that every parameter returns a sensible, bounded, finite-variance distribution on acoustically very different material.

Reference ranges

The per-parameter reference values used by the tier check are tabulated in reference_ranges.md. They follow Zwicker & Fastl (2007) for sharpness and roughness and the companion review for the level, spectral, and temporal descriptors. They are reference values the reviewed literature reports, not validated cut-offs — and only Tier 1 is graded pass/fail.

Reproducing the benchmark

The benchmark tracks are analysed exactly as any short file — in full, byte for byte — so the published values are reproducible from the public API:

from debussy import analyze_audio, write_csv

results = [analyze_audio(p) for p in benchmark_paths]
write_csv(results, "benchmark.csv")

See examples/batch_report.py for a folder-level driver. The manifest, the raw per-track parameter matrix and the scripts that regenerate both are released in validation/; paper/paper.md carries the summary figure.

Sample-rate handling

Level and spectral descriptors are computed at the file's native rate. The psychoacoustic descriptors (roughness, sharpness) are computed by mosqito, which requires 48 kHz and resamples other rates internally, printing a one-line notice. Pass suppress_warnings=True to silence it in batch runs. The reported sample_rate is always the file's native rate.