Open HRTF Dataset - Artificial Head
Overview
The dataset contains a high precision measurement of the ITA Artificial Head [1]. It is provided with 5 degree resolution in azimuth and elevation.
The available file formats are SOFA and OpenDAFF.
Measurement Setup
The HRTF is acquired according to the free field definition by Blauert [2]. One loudspeaker is moved along a circular path in elevation using a swivel arm.
The artificial head is positioned in the center of this circle and can be rotated around its own axis using a turntable turning the azimuth angle.The radius of the speaker to the center of both rotations is 2 meters.
The measurement takes place in a hemi-anechoic chamber with dimensions ( 11m x 5,97m x 4.5m ). The measurement center position is located two meters above the ground.
The reference measurement is taken in the center of the sampled sphere without the head. The microphone used is a 1/2" microphone from Brüel & Kjaer (Type 4192). The used loudspeaker is a Omnes Audio BB 2.01 housed a custom enclosure to reduce influences on the directivity. The frequency response of the loudspeaker can be seen below!
For all measurements, logarithmic sine sweeps with an FFT-degree of 14 and a sampling rate of 44.1 kHz covering a frequency range from 20 Hz to 20 kHz are used. The resulting measurements have an SNR of approximately 80 dB.
As the floor of the measurement chamber gives a large reflection, not a whole sphere can be measured. Instead, the upper hemisphere and the lower hemisphere are measured separately with and overlapping area of 30 degrees, and the artificial head is flipped upside down for the lower hemisphere. As the switch from one hemisphere to the other cannot be performed without small positional deviation, data from the upper hemisphere measurement is used up until an elevation of 100 degree. For all lower elevations, the second hemisphere measurement is used. This avoids switching between hemisphere datasets near the horizontal plane.
Postprocessing
As per the definition of a free field HRTF, the ear transfer functions are referenced to the centered reference measurement. Additionally, time window and time crop functions are used to clean the measurement of reflections and noise.
In a first step all directions are processed with the same window. The starting slope is set between 0.0020s and 0.0040s and the trailing slope between 0.010s and 0.012s. For more information look into the ita_time_window.
After the time window, the division by the reference measurement is done and time windowed with the same windows. This division needs to be regularized and is performed between 100 Hz and 20 kHz. For more information, look into the ita_divide_spk function.
In a final step, the low frequency information needs to be added. As the used loudspeaker can not cover frequencies below 200 Hz, and the measurement room is not
anachoic below that frequency, no usable information is present.
It is assumed that HRTFs are not affecting low frequencies, as the head is very small compared to the wavelength. Instead, theoretical HRTFs will converge to 0 dB towards low frequencies. To account for this known behavior, low frequency data is interpolated. For more information, look into the ita_HRTF_postProcessing_smoothLowFreq function.
Finally, a cyclic time shift by a fixed number of 128 samples is applied to avoid anti-causal filters for all directions and ears.
Resulting Dataset
Two slices of horizontal and median plane of the final data is plotted below.
Helpful Software
All measurements and post-processing has been done with the open source ita-toolbox. Postprocessing has been done with toolbox version.
Download HRTF Dataset (coming soon)
[1] Schmitz, A. (1995). Ein neues digitales Kunstkopfmeßsystem. Acta Acustica United with Acustica, 81, 416–420
[2] Blauert, J. (1997). Spatial hearing: The psychophysics of human sound localization. Handbuch der Audiotechnik (2nd ed.). MIT Press Cambridge.