oxygine-sound/dependencies/openal/hrtf.txt

HRTF Support
============

Starting with OpenAL Soft 1.14, HRTFs can be used to enable enhanced
spatialization for both 3D (mono) and multi-channel sources, when used with
headphones/stereo output. This can be enabled using the 'hrtf' config option.

For multi-channel sources this creates a virtual speaker effect, making it
sound as if speakers provide a discrete position for each channel around the
listener. For mono sources this provides much more versatility in the perceived
placement of sounds, making it seem as though they are coming from all around,
including above and below the listener, instead of just to the front, back, and
sides.

The built-in data set is based on the KEMAR HRTF diffuse data provided by MIT,
which can be found at <http://sound.media.mit.edu/resources/KEMAR.html>. It's
only available when using 44100hz playback.


External HRTF Data Sets
=======================

OpenAL Soft also provides an option to use user-specified data sets, in
addition to or in place of the built-in set. This allows users to provide their
own data sets, which could be better suited for their heads, or to work with
stereo speakers instead of headphones, or to support more playback sample
rates, for example.

The file format for the data sets is specified below. It uses little-endian
byte order. Certain data fields are restricted to specific values (these
restriction may be lifted in future versions of the lib).

==
ALchar   magic[8] = "MinPHR00";
ALuint   sampleRate;

ALushort hrirCount; /* Required value: 828 */
ALushort hrirSize;  /* Required value: 32 */
ALubyte  evCount;   /* Required value: 19 */

ALushort evOffset[evCount]; /* Required values:
  { 0, 1, 13, 37, 73, 118, 174, 234, 306, 378, 450, 522, 594, 654, 710, 755,
    791, 815, 827 } */

ALshort coefficients[hrirCount][hrirSize];
ALubyte delays[hrirCount]; /* Element values must not exceed 127 */
==

The data is described as thus:

The file first starts with the 8-byte marker, "MinPHR00", to identify it as an
HRTF data set. This is followed by an unsigned 32-bit integer, specifying the
sample rate the data set is designed for (OpenAL Soft will not use it if the
output device's playback rate doesn't match).

Afterward, an unsigned 16-bit integer specifies the total number of HRIR sets
(each HRIR set is a collection of impulse responses forming the coefficients
for a convolution filter). The next unsigned 16-bit integer specifies how many
samples are in each HRIR set (the number of coefficients in the filter). The
following unsigned 8-bit integer specifies the number of elevations used by the
data set. The elevations start at the bottom, and increment upwards.

Following this is an array of unsigned 16-bit integers, one for each elevation
which specifies the index offset to the start of the HRIR sets for each given
elevation (the number of HRIR sets at each elevation is infered by the offset
to the next elevation, or by the total count for the last elevation).

The actual coefficients follow. Each coefficient is a signed 16-bit sample,
with each HRIR set being a consecutive number of samples. For each elevation,
the HRIR sets first start with a neutral "in-front" set (that is, one that is
applied equally to the left and right outputs). After this, the sets follow a
clockwise pattern, constructing a full circle for the left ear only. The right
ear uses the same sets but in reverse (ie, left = angle, right = 360-angle).

After the coefficients is an array of unsigned 8-bit delay values, one for each
HRIR set. This is the delay, in samples, after recieving an input sample before
before it's added in to the convolution filter that the corresponding HRIR set
operates on and gets heard.


Note that the HRTF data is expected to be minimum-phase reconstructed. The
time delays are handled by OpenAL Soft according to the specified delay[]
values, and afterward the samples are fed into the convolution filter using the
corresponding coefficients. This allows for less processing by using a shorter
convolution filter, as it skips the first coefficients that do little more than
cause a timed delay, as well as the tailing coefficients that are used to
equalize the length of all the sets and contribute nothing.

For reference, the built-in data set uses a 32-sample convolution filter while
even the smallest data set provided by MIT used a 128-sample filter (a 4x
reduction by applying minimum-phase reconstruction). Theoretically, one could
further reduce the minimum-phase version down to a 16-sample convolution filter
with little quality loss.