Sound Chambers Have a Data Problem: 845 Measured Ears

Every headphone target curve you have ever read about is applied to ears nobody in particular owns — an average built on dummy heads, which are themselves approximations of an average that has never actually been measured. That gap gets treated as an unavoidable cost of doing acoustics. It isn't a cost problem. It's a sample problem, and the entire publicly usable pool of properly measured human head-related transfer functions is 845 people.

Your Ear Shape Is Half the Measurement

A sound is a pressure wave, and before it reaches the eardrum it has been shaped by the ridges of the outer ear, the bowl of the pinna, the side of the skull, the shoulder, even hair. Each surface reflects, absorbs and diffracts different bands differently depending on where the sound came from. What arrives at the eardrum is a spectral fingerprint, and the brain has learned to read fingerprints as directions. Fill the pinna ridges with putty and subjects lose the ability to say whether a sound came from above or below. Leave the putty in for a few days and most adapt, because the brain rebuilds the map.

That fingerprint, measured for one direction, is the head-related transfer function. A full set covers the sphere around your head. Collapse the sphere into a single curve — weighting each direction by the patch of sphere it represents, because a regular azimuth and elevation grid oversamples the poles — and you get the diffuse-field version: the aggregate colouration your ears apply to a field arriving from everywhere at once. That curve underpins headphone targets, including the Harman target, which came out of preference testing on speakers in a good room. Personal variation concentrates where pinna geometry does the most work, roughly 4 to 16 kHz.

The Dummy Heads Average Something Unmeasured

Fixtures like the Brüel & Kjær 5128 and KEMAR exist because measuring every buyer's ears is impossible. They are precision artificial heads with measurement mics sitting at the canal opening, built to pinna-shape standards and chosen for consistency between labs. What they are not is your ears. The 5128's outer ear is known to overstate treble between 5 and 10 kHz against the average human described by ISO 11904, and there is no measured diffuse-field HRTF for the pinna the Harman fixture uses — so no reference exists for what that ear does acoustically.

Worse, the population those fixtures guess at is thin on the ground. Add up the four datasets that constitute essentially all the properly measured human HRTFs available to researchers — HUTUBS at TU Berlin with 96 subjects, ARI in Vienna with 263, SONICOM at Imperial College with 408, and Meta's Sound Sphere 2 with 78 — and you get 845 people across 941 directions. That is the pool. It skews heavily toward male European academics. No amount of fixture precision repairs a corpus that small; it only makes the fixture's error more repeatable.

Ten Sweeps in a Field, Then a Network Fills the Sphere

The measurement chain is almost embarrassingly simple: a DIY in-ear microphone of the PIRATE type sits at the canal entrance, a consumer Bluetooth speaker holds a position, and REW plays a sweep while recording what comes back. Do that at ten positions: straight ahead, front-left, left, back-left, behind, back-right, right and front-right, plus one raised front-left and one overhead. That samples the regions carrying the most listener identity between 1 and 10 kHz.

The expensive part of a real rig is the room. An anechoic chamber kills reflections with deep wedge lining and runs into six figures to build and to maintain. Measuring outdoors does that job for free: with the ground five or six feet below the ears, the first reflection arrives roughly 10 milliseconds late and about 17 dB down, far enough away in time to gate out. The speaker's own colouration is measured once, stored as a calibration file and subtracted from every capture, so only the listener's contribution survives. The diffuse-field figure is magnitude-only by definition — which bands get lifted and by how much, not their arrival times — so phase a Bluetooth link can't supply is irrelevant here.

The rest is interpolation. Each capture is resampled to a 600-point grid, smoothed to a sixth of an octave and normalized, and the 199 bins between 1 and 10 kHz from all ten positions are concatenated into a 1,990-value vector standing in for this listener's acoustic identity. Two linear layers with GELU compress that to 192 values. Target directions go through Fourier positional encoding, borrowed from neural radiance fields, into 96 values. Three decoder layers emit a 600-point curve per direction, and five independently trained networks average their answers in decibels. Below 300 Hz nobody gets a personalized answer — the population median from 96 HUTUBS subjects is blended in from 300 to 500 Hz.

Average error on listeners the model never saw came out around 0.93 dB, across every direction and every frequency. The threshold for noticing a frequency response difference sits around 1 to 2 dB. That is the whole claim: ten sweeps and twenty minutes of a person's afternoon, outdoors, no chamber, landing inside the range where the difference stops being reliably audible.

What I Think

Treat the 0.93 dB as promising, not settled. The validation subjects come from the same four datasets used for training, no perceptual listening test was run, and the output stops at 10 kHz — precisely the band where individual variation is strongest and predictions are least trustworthy. A model accurate in the range it was built for, and silent above it, is a tool with a job description, not a chamber replacement. Anyone quoting it as proof the problem is solved skips the part where nobody has put real humans in front of it and asked what they hear.

But the thing worth stealing here is the economics, not the architecture. Acoustic measurement has been stuck in a loop where the honest answer to "how do real ears hear this" is "we cannot afford to find out", so the industry measures a plaster head and calls it an average. If ten outdoor sweeps produce a usable per-person diffuse-field curve, the constraint on better targets stops being lab time and becomes recruitment — 8,000 people instead of 845, from populations that aren't mostly European men. That is a far better problem to have, and it decides whether a target curve means anything. The dummy heads aren't going anywhere; they just shouldn't be the last word.