Distortion Only Became Audible at 104-112 dBA: What a Published Headphone Listening Study Actually Measured

For years, headphone reviews have printed total harmonic distortion (THD) figures down to fractions of a percent, implying that lower is always better and that listeners can tell the difference. A research write-up from RTINGS.com, authored by Pierre Lelievre and co-authored with Dr. Sean Olive, and drawing on measurement work published by Listen, Inc., set out to test that assumption directly with blind listening experiments. The RTINGS article carries an "Updated Mar 19, 2026" date; related Listen, Inc. posts appeared in April 2026 and on 4 May 2026. The findings, now sitting in the public record for several months, are worth walking through in detail because the actual numbers are more nuanced — and more heavily caveated by the authors themselves — than a headline might suggest.

How the detection thresholds were measured

The first experiment asked a simple question: how loud does distortion have to get before someone can actually hear it? Researchers recorded real headphones binaurally while they played music at levels from 85 dBA up to 110+ dBA SPL, in 3dB steps, inside an audiometric booth. They then stripped the frequency response back out of those recordings, isolating the nonlinear distortion component, and normalised every file to -16 LUFS so loudness itself wasn't a variable.

Three headphones under test — the TRUTHEAR x Crinacle ZERO: RED, the AKG K361, and the Audeze LCD-S20 — were played back through an Audeze LCD-X, chosen specifically because it measures lower distortion than the units being evaluated. Thirteen listeners took part, nine of whom had roughly 30 hours of formal distortion-discrimination training spread over three months using the Klippel Online Listening test. Detection was tested using an ABX method with an adaptive staircase (the Kaernbach method): step size started at 6dB and halved after the second reversal, with the threshold calculated as the mean of the last four reversals, corresponding to a 67% correct-detection rate (a d' of roughly 1.0). The musical stimuli were short clips — 11 seconds of Tracy Chapman's "Fast Car" and 9 seconds of Rebecca Pidgeon's "Spanish Harlem" — played back at a calibrated 83 dBA SPL.

The result: detection thresholds landed at 104-112 dBA SPL. For context, the authors give typical listening levels as 70-85 dBA SPL — meaning the distortion only became audible at levels well beyond how most people actually listen. It's worth noting the RTINGS piece contains this 104-112 dBA figure in its summary, results section, and conclusion, though one later passage in the same article states "98-112 dBA SPL." The 104-112 dBA range is the one repeated consistently throughout the piece.

The LCD-S20 showed the lowest thresholds of the three, requiring about 105 dB SPL on average before listeners could detect distortion. Notably, training level did not correlate with threshold — the extensively trained listeners weren't reliably more sensitive than anyone else. One participant, listener L12, stood out by averaging 92.5 dBA SPL across six tests, at least 10dB below every other listener, making them meaningfully more sensitive to distortion than the group as a whole.

Can listeners tell headphones apart once frequency response is removed?

The second experiment tackled a related but distinct question: at normal listening levels, with frequency response differences removed, can people tell headphones apart at all? This used a MUSHRA-like similarity test with a hidden "Clean" reference (the unprocessed original file) and no low-quality anchor to bias responses. Five headphones were compared — the ZERO: RED, Beyerdynamic DT 770 Pro, Sennheiser HD 650, AKG K361, and Lorelei X6 — with recordings made at 94 dBA SPL and playback again calibrated to 83 dBA SPL. Nine listeners, eight of them trained, ran five trials each, rating similarity to the reference on a 0-100 scale.

The hidden Clean reference scored highest, with a mean rating of 69.0, and was rated significantly higher than four of the five headphones (p < 0.05). But the picture is messier underneath that headline number: only one listener, L10, rated the hidden Clean reference as 100% similar across every trial — and that same listener also posted the highest standard deviation (34.8) on their other ratings, suggesting inconsistency rather than unusually sharp hearing. More strikingly, five of the nine listeners (56%) did not rate the Clean reference as most similar on average across their trials. The overall mean standard deviation across all ratings was 26.3, indicating considerable spread in how confidently people were judging similarity. The ZERO: RED scored lowest among the tested headphones, at 40.2.

A follow-up ABX test pushed further: two listeners dropped out after concluding they simply couldn't hear differences between the headphones, leaving seven who completed ten trials per pair. Of those seven, only one — listener L11 — achieved statistically significant discrimination, correctly identifying the ZERO: RED versus the HD 650 in nine of ten trials (p < 0.05). An ANOVA across the results found at least four of the five headphones statistically equivalent to each other, with all pairwise comparisons showing p > 0.83.

What the measurements showed, and what correlated with what

Alongside the listening tests, six headphones — the five above plus the LCD-S20 — were measured at 94 dB SPL using several distortion metrics: normalised and unnormalised THD via sine sweeps (all six measured below 1% THD through most of the audible range), intermodulation distortion (IMD) per IEC 60268-7 using two tones at 70Hz and 600Hz at a 4:1 ratio, difference frequency distortion (two tones 80Hz apart swept from 100Hz to 20kHz), and non-coherent distortion measured by comparing the processed listening-test files against the Clean reference. The normalised THD method used traces back to a technique first proposed by Steve Temme in 1993 at Listen, Inc.

The published IMD table (Ld2 dB / Ld3 dB / total IMD at 600Hz) reads: ZERO: RED -61.84 / -83.83 / 0.01%; DT 770 Pro -41.72 / -58.71 / 0.24%; HD 650 -56.52 / -61.61 / 0.05%; K361 -61.40 / -64.55 / 0.04%; LCD-S20 -66.73 / -54.09 / 0.03%; Lorelei X6 -52.09 / -53.50 / 0.13%.

On the correlation question, results diverged depending on which experiment was in view. Normalised THD correlated with the Experiment 1 detection thresholds at r ≈ 0.85 (p ≈ 0.03) — but the authors are explicit that this rests on only three headphones, which sharply limits the statistical power of that finding. Against the Experiment 2 similarity scores, none of THD, IMD, or DFD showed a significant correlation. Only non-coherent distortion, measured specifically on "Fast Car," correlated with what listeners reported.

The caveats the authors put on their own results

The write-up is unusually candid about its own limitations, and those caveats matter as much as the topline numbers. The ZERO: RED measured 0.24% non-coherent distortion on "Fast Car," well above the 0.03-0.09% range of the other headphones — but only 0.05% on "Spanish Harlem" at the same SPL, roughly a 14dB swing depending on the track. The authors say this may not reflect anything intrinsic to that IEM, and could instead be a recording artefact or an effect of the test fixture and insertion depth.

The LCD-S20 also came with a flag: it showed reduced bass compared with RTINGS' own published review of the same headphone, hinting at a possible seal issue during testing. Some threshold tests involving it were excluded, on the judgment that listeners might have been detecting noise rather than distortion, and it was left out of Experiment 2 entirely. The authors state that correcting the seal issue could only have pushed its threshold higher, never lower. Separately, they note that booth noise in the bass region was non-negligible and was accepted as a compromise in the test design.

Most broadly, the authors state plainly that these findings are specific to the headphones actually tested and cannot be generalised to every headphone on the market — nor, by extension, to loudspeakers or amplifiers. RTINGS notes it has updated its own review methodology, in what it calls Test Bench 2.1, to reflect distortion's reduced measured impact on perceived sound quality, though the site continues to measure and publish distortion figures.

The research also credits outside support: Steve Temme and the SoundCheck team at Listen, Inc. loaned the measurement software used, while Jeremie Voix and the Ecole de technologie superieure (ETS) provided the audiometric room in which the listening tests were conducted.

Taken as a whole, the study doesn't argue that distortion specifications are meaningless — the correlation between normalised THD and detection threshold, thin as the underlying dataset is, points toward some relationship. But it does suggest that the fractional-percent THD numbers printed in most reviews sit well beneath what ears can actually distinguish at normal listening levels, and that turning five different headphones into meaningfully different listening experiences, once frequency response is equalised out, is harder to demonstrate in a blind test than most gear discourse assumes.

This article was written by an AI system from Into The HiFi AI Desk, generated from the following sources, with no human editing pass before publication: RTINGS.com, "The Surprising Truth About Headphone Distortion: Most Of It Is Inaudible" (updated 19 March 2026), Listen, Inc., "Headphone Distortion Audibility" (April 2026), Listen, Inc., "Headphone Distortion Audibility: Subjective Perception and Psychoacoustic Measurements" (4 May 2026).