Localization is worst on the front–back axis
Aliases: front-back confusion · front-back reversal · HRTF ambiguity
What it is
A cue that was meant to sit due rear is often heard due front — the wearer turns the wrong way. In the horizontal plane, front–back is the worst axis of auditory localization. Interaural time and level differences can be identical for “a certain angle in front” and “the symmetric angle behind”; those points lie on the cone of confusion. Left and right are usually easy; front and back need pinna spectra, and a generic HRTF is least stable on that axis.
Why it happens
ITD and ILD are set by the interaural axis and are constant on a cone of rotation about that axis. On that cone, front versus back and up versus down are not unique solutions of the binaural difference. Breaking the cone takes direction-dependent filtering by the pinna at high frequencies: each ear’s geometry is individual, so the spectral fingerprint is personal. A product HRTF is an average ear; the front–back spectral difference is smeared, and due rear often collapses into due front — everyday sources more often lie in front, and perception uses that prior to break the tie.
A small head turn can break the cone: the direction in which ITD changes is opposite for front and back. If the headset does not re-spatialise on the new heading, or does so too late, that dynamic cue is unused and the front–back confusion is locked in on the first sample.
Studying it
Place sources uniformly in azimuth on the horizontal plane, including 0° and 180°. People point with the head or a controller at “where the sound is.” Look at the error distribution: left–right error is usually a small standard deviation; front–back error goes bimodal — a substantial fraction of 180° sources are pointed near 0°. Compare a generic HRTF, an individualised HRTF, and a condition that allows a small head turn before pointing.
Independent variables: azimuth, whether the HRTF is individualised, whether a head turn is allowed before localisation, whether the headphones cut high frequencies. Dependent variables: front–back reversal rate, pointing error, confidence.
Measuring only a few left–right points produces a false “spatial audio is accurate.” Front and back have to be planned conditions, not a surprise in the log. Open loudspeakers and closed headphones do not supply the same pinna cues; do not pool them.
Where it stops holding
An individualised HRTF, enough high frequency, and a small head turn during the sound can push reversals down; “worst” is not a physical forever. Low-frequency-dominated sounds (an engine rumble) have almost no pinna spectrum, so front and back are even less separable. Unilateral hearing loss wrecks ITD/ILD and the whole horizontal plane suffers, so front–back is no longer the only weak axis. Once vision gives a location (the loudspeaker is seen), audition is captured and front–back confusion will not show. Elevation is hard too, but a front–back reversal in a product costs a 180° turn, which is usually the larger price.
Applying it
- Do not treat “directly behind you” as a sentence spatial audio can say on its own. Pair a rear target with a brief visual anchor, or add the words “behind you.”
- Let the cue play while a small head turn is possible, and keep spatialisation following the head. Do not freeze a wrong-front on the first frame.
- Critical alerts should not distinguish two states only on the front–back axis (for example “car ahead / car behind” by image alone). Left–right, or a spoken word, is more stable.
- How to check: play the same cue several times at 0° and at 180°, eyes closed, pointing. Write down the fraction of 180° trials pointed into the front hemisphere. Under a generic HRTF treat that reversal rate as a known defect. If the product depends on “behind” as a standalone direction, either get the reversal rate down or stop using sound alone to split front from back.