Put a sound somewhere around the head. Each ear hears a source through its own filter, shaped by the head, the shoulders and the folds of the outer ear, and those two filters are what tell the brain where a sound is. This effect convolves the input with the measured pair for the position asked for, so on headphones the sound arrives from that angle and height rather than from a point between the ears.
Azimuth, elevation and width
Azimuth runs from -180 through 0 at the front to 180, so 90 is hard right and either end is directly behind. Elevation runs from 40 degrees below ear level to straight overhead. A stereo input is placed as two sources, its left channel at the azimuth minus half the width and its right at the azimuth plus half, so a stereo file keeps its own spread while the whole image moves; at width 0 both land in one place. A mono input is one source.
Distance
Level falls as one over distance from a reference of 1 metre, and a low-pass whose corner falls with distance stands in for the air. There is no room: for one, chain the convolution reverb after this step.
The measured set
The responses are the MIT KEMAR compact set, measured in 1994: 368 positions on the right side of the head at elevations from -40 to 90 degrees, 128 taps each at 44.1 kHz, with the loudspeaker equalised out and the interaural delay kept. Positions between the measured ones are blended from the four nearest; a source on the left uses the mirrored position with the ears swapped. On a file at another sample rate the responses are resampled once.
Measured responses carry the resonance of the dummy head’s own ear canal, which headphones are already compensated for, so the set is diffuse-field equalised by default: the average response over all positions is taken out of each one with zero phase, leaving the interaural delays untouched. The switch turns that off for the raw set.
Moving a source
Changing a position glides the filter taps rather than switching them, so a slider drag or an automation sweep in the audio workflow does not click. The step declares no latency: the direct sound sits a couple of samples into every response, which is the delay a head imposes.
Where it stops
A measured head is one head. Front and back are the cues that depend most on the individual ear, so a source placed behind reads as in front for some listeners, and elevation is weaker than azimuth for everyone. In the audio workflow, where the measured set is not loaded, a spherical-head model stands in: left and right are convincing, height only roughly.