Dynamic Range ABX — Compression

A blind dynamic range test compares uncompressed audio passages against versions with restricted dynamic range under double-blind conditions. Listeners evaluate randomized presentations to determine the threshold where peak limiting and dynamic compression become perceptible, distinguishing subtle mastering control from audible reduction in musical crest factor and transient punch. The open question this drill settles, for you personally, is where “transparent” limiting stops being transparent.

How to use it

1. Start on the subtle rung. Signal A is an open-dynamics reference at crest factor 14 dB; Signal B is the same material at 12 dB. If that is too easy, work down to the 9 dB and 6 dB pairings.
2. Learn A and B openly. Toggle the references until you can hear the difference on the chosen rung, then commit to the blind section.
3. Answer X trials. X is randomly one of the two; the engine tallies trials, correct count, streak, running p-value, and confidence.
4. Keep the volume knob still. Both signals are level-matched, so touching the volume mid-session replaces dynamics as the cue and corrupts the score.
5. Log the rung where you went from chance to significant. That boundary, not a headline number, is your usable result.

Crest factor: the arithmetic of dynamics

Crest factor is the peak-to-RMS ratio expressed in decibels: CF = 20 × log10(peak / RMS). It is a pure amplitude statistic, and for standard shapes it is exact: a square wave sits at 0 dB because its peak equals its RMS, a sine wave is 20 × log10(√2) = 3.01 dB, a triangle 20 × log10(√3) = 4.77 dB. Real program material is far more peaky, because transients live briefly above the average level.

Signal	Peak/RMS	Crest factor
Square wave	1.00	0 dB
Sine wave	1.41	3.01 dB
Triangle wave	1.73	4.77 dB
Percussive acoustic program	5–10	14–20 dB
Brickwalled pop master	2–2.5	6–8 dB

The loudness-war mechanic follows directly from these numbers. A limiter squashes peaks while the makeup gain raises the floor, so a master taken from 14 dB of crest factor down to 6 dB, with peaks pinned at the same ceiling, carries 8 dB more average energy: roughly twice the perceived loudness against an open master on any system without loudness normalization. Program loudness is standardized by ITU-R BS.1770 (current revision 5, approved 2023), which defines the K-weighted, gated measurement reported in LUFS. The normalization generation that followed quietly defanged the tactic: Spotify normalizes playback to −14 LUFS integrated, YouTube and several other services target the same neighborhood, and Apple Music sits near −16, so a crushed −8 LUFS master is simply turned down 6 dB and arrives with less punch than the open mix that was never pushed. What remains audible is the damage, not the advantage, and that is exactly what this drill isolates.

Mechanically, a digital limiter applies gain reduction over milliseconds. Lookahead peeks a few ms into the future so the reduction ramps in before the peak lands; release then rides the signal back up, which is where the “ambient swell” and pumping cues come from: the noise floor and reverb tails of the passage rise while transients are clamped, then sink as gain recovers. Pushed past its knee, the limiter’s transfer curve also folds in low-level harmonic distortion, coupling dynamics processing to the harmonic spectrum the THD Test page measures separately.

Why level-matching and forced choice are the whole method

The loudest-sounding clip wins almost every unblinded preference test, regardless of which is better reproduced, and that bias is strong enough to invalidate casual comparisons. This drill defuses it in two stages. First, Signal A and Signal B are integrated-loudness matched before presentation: the processing under test is crest-factor reduction at equal measured loudness, so loudness itself is no longer available as a cue, and any gain-difference tell is engineered away. Second, the comparison is forced-choice ABX with a fresh random assignment each trial and no option to abstain, so guessing behaves like the coin flips it is.

Scoring then asks the only honest statistical question: how often would chance produce your score? The engine’s running p-value is the upper-tail binomial probability of getting at least your number correct at 50% per-trial odds, with verdicts significant at p ≤ 0.05, needs more trials while the evidence is soft, and chance when accumulated misses make detection untenable. The three rungs (12 vs 14, 9 vs 14, 6 vs 14 dB) function as a difficulty staircase: your detection boundary between rungs is your personal audibility threshold for limiting, measured rather than asserted.

Accuracy, limits and troubleshooting

• Material matters. Dense, sustained program hides a limiter’s gain moves better than sparse percussive content, where the release ride is exposed. A clean sweep on one rung with drums does not generalize to pads on another.
• Your output chain can process without telling you. Phone and laptop “volume balancing,” TV dynamic-range compression, and Bluetooth codec quirks all compress; run the drill on a wired path with effects off or you may be testing the OS, not the master.
• Fatigue skews late-session results downward. Loudness discrimination degrades over minutes of level-comparison work; take short runs across days.
• A chance verdict is informative, not embarrassing. Modern lookahead limiters genuinely shave 1–2 dB nearly transparently; finding 12-vs-14 indistinguishable is the expected outcome for many trained listeners.

Related tools

• Loudness Meter — measure LUFS and true peak of any material in your browser.
• Dynamic Range Test — the measurement-style sibling: sweep system crest factor and headroom.
• Blind ABX hub — the full threshold series sharing this scoring engine.
• Decibel Meter — sanity-check your playback level before comparing rungs.