Olive's loudspeaker preference model (US 8,311,232 B2)
Sean Olive, assigned to Harman International. Filed 2005-03-02 (priority 2004-03-02), granted 2012-11-13. The first source in this spoke that connects a measurement to a listener, which is the gap coverage edge 8 names and the line iec-60268-5-replacement keeps drawing: IEC 60268-21 states outright that it "does not assess the perception and cognitive evaluation of the reproduced sound." This document is an attempt to bridge exactly that, and its verdict is that anechoic measurement predicts preference well enough to be worth patenting.
What it measures
Seven anechoic curves, at 1/20-octave smoothing and 2 Hz resolution: on-axis, the listening window (a spatial average over ±30° horizontal and ±10° vertical), early reflections, predicted in-room response, sound power, and the two directivity indices (early-reflections and sound power). This is the curve set now published under the name spinorama, and the reason a modern review publishes six curves instead of one frequency-response trace.
Six statistics are computed over those curves and fed to an ordinary multiple linear regression
(Y = b₀ + b₁X₁ + … + bₙXₙ):
- NBD, narrow-band deviation — mean absolute deviation inside each ½-octave band, 100 Hz–12 kHz
- AAD, absolute average deviation from the level in a 200–400 Hz reference band
- SM, smoothness — the Pearson r² of a regression line through the curve, so higher is smoother
- SL, slope, against a target slope
- LFX, low-frequency extension — log₁₀ of the −6 dB point below 300 Hz
- LFQ, low-frequency quality, from amplitude deviations in the bass
What it claims
On the first set of 13 loudspeakers, four variables account for essentially all the variance (r² = 0.99, adjusted r = 0.96, RMS error 0.26 preference units, F(4,8) = 137.34, p < 0.0001). Generalised across 70 loudspeakers drawn from 18 separate listening tests, the correlation falls to r = 0.70 with a residual error of 0.8 preference units — still highly significant, and a useful reminder of how much of that first r² was the small sample. In the 70-speaker model the weight sits on four terms: NBD on-axis 31.5%, low-frequency extension 30.5%, NBD predicted-in-room 20.5%, smoothness predicted-in-room 17.5%.
A likely source of the "r = 0.86" figure (added 2026-10-11). Some secondary sites quote r = 0.86 for the 70-speaker model. This page keeps 0.70. Olive's own review article, The Perception and Measurement of Headphone Sound Quality (Acoustics Today, 2022), gives 0.86 as the correlation of his later around-ear/on-ear headphone model (and 0.91 for the in-ear one). That makes a mix-up between the two models a plausible explanation. It is not confirmed: the secondary quotes were not traced back to their own source.
The negative result is the sharper one. Consumer Union's model, based largely on 1/3-octave sound power, showed no correlation at all with listener preference across 13 speakers reviewed in Consumer Reports (r = 0.05, p = 0.81). Sound power alone does not predict perceived quality; the on-axis and in-room behaviour, and the bass extension, do most of the work.
Why this is graded T3
The load-bearing claim is not the method, it is how well the method works — and the correlations
were produced by the model's own author at a company that sells loudspeakers, in a document examined
for novelty rather than for statistics. Per ../QUALITY.md ("tier by the claim, not the URL"), a
maker's own effectiveness figure is T3 whatever the venue's standing, the same reading applied to
the ACM Queue case in ../programming-languages-wiki. The curve definitions and the variable
formulas are artifact fact and would be T1 on their own.
The peer-reviewed record does exist and this spoke cannot reach it: Olive's AES 116th Convention papers 6113 and 6190 (2004), Parts I and II, sit in the AES e-library behind the same paywall that already blocked ts-parameter-manufacturing-variation's figures. That is now the second time the audio-engineering-society's library has been the barrier between this wiki and a number it needs.
What it changes here
thiele-small-parameters and the alignment maths predict what a box will do; nothing here said which of the resulting curves a person would rather listen to. The answer this document gives is narrower than the spinorama's popularity suggests: flatness on-axis and in the predicted in-room curve, plus bass extension, with directivity mattering through the in-room curve rather than in its own right. And at r = 0.70 across 70 speakers, roughly half the variance in preference is still unexplained by any measurement in the set.