ACTA ACUSTICA UNITED WITH ACUSTICA

Vol. 99 (2013) 836 – 844

DOI 10.3813/AAA.918662

Measurement of a Full 3D Set of HRTFs for In-Ear

and Hearing Aid Microphones on a Head and
Torso Simulator (HATS)

Chris Oreinos, Jörg M. Buchholz

National Acoustic Laboratories, 16 University Avenue, Macquarie University, Australia
Department of Linguistics, Faculty of Human Sciences, Macquarie University, Australia
The HEARing Cooperative Research Centre, Australia. chris.oreinos@nal.gov.au

Summary
The accurate reproduction of acoustic real-world environments is becoming of increasing importance in hearing
device research and development. It is thereby often required to accurately predict the sound pressure at the
microphones of a hearing device in a simulated or recorded acoustic environment. For that reason, an extensive
set of head-related transfer functions (HRTFs) was measured in free-field with a pair of behind-the-ear (BTE)
hearing aids placed on a Head and Torso Simulator (HATS). Transfer functions to the in-ear HATS microphones
were also measured. A spherical head model was applied to extend the useable frequency range towards low
frequencies. Special care was given to preserve the phase properties of the measurements so that the HRTFs could
be widely used in phase-sensitive technical applications, including the evaluation of spatial signal processing
algorithms (i.e. directional microphones, beamformers) in hearing devices and the evaluation of sound field
resynthesis methods. The extended HRTF set can also be used for research in psychoacoustics. It is available for
download at: http://www.nal.gov.au/download/HATS_BTE_hrirDatabase.zip.
PACS no. 43.66.Pn, 43.66.Qp, 43.20.Fn, 43.66.Ts

1. Introduction back methods, where listener-specific effects are not of

primary interest. In particular in the latter application, the
Head-related transfer functions (HRTFs) contain all the measurement of a full 3D HRTF data set is required. By
spectral, temporal, and spatial information that is avail- applying such a 3D HRTF data set, it is then possible to
able in a given acoustic environment. The HRTFs in their transform the loudspeaker output signals of basically any
free-field form [1] are calculated as the ratio of the Fourier multi-channel playback system (e.g., using higher-order
transform of the sound pressure at a point in the ear canal Ambisonics [3, 4, 5]) into ear signals. Although a large
of a subject to the sound pressure that would have been amount of 3D HRTF data sets are publically available (e.g.
measured at the centre of the head, with the subject not [6, 7, 8, 9, 10, 11], very limited HRTF data that addi-
being present. They contain the diffraction effect of the tionally contain responses to the microphones of hearing
head and torso as well as the resonances and scattering ef- aids (HA) fitted to a listener’s (or HATS’) ears are avail-
fects of the pinna. Given that HRTFs are determined by the able. Such HTRF sets are essential for the research and
detailed size and shape of the head, torso, and in particu- development of hearing devices, the evaluation of hearing
lar ears, they are highly individual. This is demonstrated, aid algorithms, and the verification of multi-channel loud-
for example, in localization experiments, where listeners’ speaker systems for recreating realistic acoustic scenes
performance decreases significantly when non-individual (e.g., [12]) aimed at hearing aid testing.
HRTFs are used, in particular along the cones of confu- When multi-microphone signal processing (or enhance-
sion [2]. However, non-individual HRTFs, as for exam- ment) techniques are considered in hearing devices, the
ple measured on a standardized head and torso simulator HRTFs need to provide accurate phase and amplitude in-
(HATS), are often used in applications where it is unprac- formation. Particularly when considering delay-and-sub-
tical to measure individual HRTFs, such as virtual audi- tract directional HA processing [13], even very small
tory spaces and computer games. Non-individual HRTFs phase and amplitude errors between the two HA micro-
can also be considered sufficient when analysing and veri- phones are critical at low frequencies due to the differ-
fying the general effect of multi-channel loudspeaker play- entiation and subsequent equalization of the high-pass
6 dB/octave roll-off. Such accuracy was difficult to achieve
Received 27 November 2012, here at low frequencies, by the measurements alone, for
accepted 1 August 2013. two reasons: (a) because a small measurement loudspea-

836 © S. Hirzel Verlag · EAA

Oreinos, Buchholz: Measurement of HRTFs ACTA ACUSTICA UNITED WITH ACUSTICA
Vol. 99 (2013)

ker, with poor low-frequency response, had to be used due

to weight restrictions of moving equipment and (b) be-
cause the used anechoic chamber was not rated anechoic
down to the required frequencies. To remedy that, the mea-
sured HRTFs were extended towards low frequencies by
applying a spherical head model.
Although spherical head models have been previously
used for this purpose [14], the applied methods for the
combination of the model with the measurements have
been evaluated mainly perceptually [15]. The signal pro-
cessing in hearing devices is very different from the au-
ditory signal processing, and in particular for the case
of multi-microphone signal processing techniques, it is
much more sensitive to small amplitude and phase vari-
ations. Hence, different (or modified) techniques are re-
quired for combining measured and modelled HRTFs. The
low-frequency extension methods developed and verified
in this study are important for any HRTF measurement
where the available anechoic chamber has a limited fre-
quency bandwidth or when only semi-anechoic or even
reverberant chambers are available and impulse response
truncation methods need to be applied to remove the room
artefacts from the measured HRTFs.
The present study comprises
• The measurement of an extensive 6-channel HRTF data
set for 1784 source locations covering the full 3D space
using a Brüel & Kjær type 4128C HATS with two Figure 1. The acoustic mannequin used for the measurements:
behind-the-ear (BTE) hearing aids, featuring two mi- (a) left ear of the HATS showing the hearing aid satellite with
crophones each, fitted above the left and right pinna. its two microphones; (b) the HATS standing upside down on a
• The application of a spherical head model to extend the tripod and held in place by a purpose-built plastic fitting.
usable frequency range of the measured HRTFs towards
low frequencies.
• A method to combine measured and modelled HRTFs P.58 recommendation for HATS telephonometry. It comes
that is adequate for multi-microphone hearing aid signal pre-fitted with 4158C/4159C type occluded ear simulators,
processing techniques. The applicability of this method which were connected to a pair of 2610 Brüel & Kjær
is verified by considering directivity plots of an exam- measuring amplifiers. For these HRTFs measurements, a
ple hyper-cardioid microphone that is realised by com- pair of –intentionally– slightly asymmetric pinnae (B&K
bining the signals from the two microphones of either models DZ-9763 and DZ-9764) were used. Two behind-
BTE hearing aid. the-ear (BTE) hearing aids were placed and secured, with
sticky tape, above the HATS’ left and right pinnae (Fig-
ure 1a). The hearing aids were provided by Phonak and
2. Methods featured cables connected to their microphones and in-
built receiver (not used for this set of measurements). The
2.1. Measurement setup signals picked up by the microphones were sent via bal-
HRTFs were measured in the anechoic chamber of the anced cables to a purpose-built preamplifier designed by
Auditory Neuroscience Laboratory, Department of Phys- Phonak. It must be noted that these hearing aid “satel-
iology, University of Sydney. The chamber is 3.5 m long, lites” did not perform any signal processing. Finally, the
4.6 m wide and 2.4 m high and is rated as anechoic above six channels (two in-ear plus two BTE microphones for
300 Hz. It contains a motor controlled semicircular hoop each side) were fed to an RME Fireface 400 soundcard
with a radius of 1.2 m, which supports an Audience A3 connected to a PC running MATLAB. Transfer functions
loudspeaker. The hoop is fully controlled by computer were measured using a logarithmic sweep of 1 s duration,
software with an angular precision of less than 0.1◦ and sampled at 44.1 kHz with 24-bit resolution.
is calibrated before the start of each set of measurements. An important issue that had to be resolved was the fact
The measurements of the HRTFs were performed on the that the lowest elevation angle where the hoop could be
4128C head and torso simulator (HATS) manufactured by positioned at was δ = −55◦ . In order to work around this
Brüel & Kjær. The HATS complies with the ANSI stan- technical restriction the HATS had to be mounted upside
dard for manikins for simulated in-situ airborne acoustic down supported by its head, facing the front as in the up-
measurements (ANSI S3.36-1985 [R2006]) and the ITU-T right placement, for some of the measurements. For that

837
ACTA ACUSTICA UNITED WITH ACUSTICA Oreinos, Buchholz: Measurement of HRTFs
Vol. 99 (2013)

Table I. The coordinates of the 1784 points on the sphere where free-space configuration. Instead, the loudspeaker magni-
the HRTFs and BTE transfer functions were measured. Resolu- tude response was equalized, with a 512-tap FIR filter,
tion of elevation range: 5◦ . Number of measurements: N using a measurement provided by our collaborators, and
the magnitude responses of the BTE microphones were
Elevation range Azimuth resolution N
equalized, with a 64-tap FIR filter, after being measured
◦
[−40 , −40 ]◦
5 ◦
1224 inside a hearing aid test box. The frequency response of
(−60◦ , −40◦ ) ∪ (40◦ , 60◦ ) 8◦ 270 the HATS ear simulators (4158C and 4159C) were already
(−70◦ , −60◦ ] ∪ [60◦ , 70◦ ) 10◦ 144 calibrated up to 20 kHz, on purchase of the product, so
(−80◦ , −70◦ ] ∪ [70◦ , 80◦ ) 15◦ 96 no further correction was necessary. The gain necessary
(−90◦ , −80◦ ] ∪ [80◦ , 90◦ ) 30◦ 48 to compensate for the in-ear microphones/amplifiers sen-
−90◦ , 90◦ 360◦ 2
sitivity was estimated by comparing the measured (0◦ , 0◦ )
HRTF to the calibration chart provided by B&K, for the
specific HATS that was used, as well as by comparing
purpose, it had to be placed on a custom-built mount that the measured HRTFs to the ITU-T P.58 recommendation
was shaped around the top of the head using a thermo- (“Head and torso simulator for telephonometry”) values
plastic material (polymorph pellets). Figure 1b shows the for HRTFs at 0◦ , 90◦ , 180◦ , 270◦ on the horizontal plane.
setup used for the upside down mounting of the HATS. For The sensitivity of the four BTE microphones/amplifiers
those measurements, the source (loudspeaker) coordinates was compensated for by minimizing the RMS error be-
(azimuth θ and elevation δ) were referenced to the upside tween their individual magnitude responses, averaged over
down HATS according to the transformations: δ = −δ and all source directions, and the corresponding averaged in-
θ = −θ, so that the role of the ears remained consistent. ear microphone responses in the frequency range of 300–
The use of asymmetric pinnae for the left and right ears of 600 Hz.
the HATS dictated this choice. If both pinnae were exactly
the same, another option would have been to transform the
coordinates according to δ = −δ and θ = θ, while ex- 2.2. Spherical head model realisation
changing the roles of left and right side microphones (in- In this work, a spherical head model was applied to extend
ear and BTE). the useable frequency range of the HRTFs, i.e. to extrapo-
In both measurement conditions (standing upright and late information to lower frequencies, while preserving the
upside down) the accurate alignment of the HATS was fa- sensitive phase information of the HA microphones. This
cilitated by using two laser beam pointers mounted on the model approximates the human head by a rigid sphere and
hoop at an arc distance of π/2. Correct alignment in all as such, it does not take into account the acoustic reso-
three axes could be attained by targeting the laser beams nances and scattering by the pinnae, at high frequencies,
to the ear canals and the tip of the nose of the HATS. That or the shadowing effects of the torso. Although it is an ide-
resulted in effectively aligning the centre of the interaural alised model, it has been suggested that it can be used to
axis with the centre of the hoop, which is the centre of the significantly improve the low frequency accuracy of mea-
coordinate system against which the measurements were sured HRTFs [15, 16, 17]. Using the Fourier-Bessel series
referenced. An observation that will prove useful in sec- expansion of the wave equation solution [18] for the in-
tion 2.2 is that the centre of the interaural axis does not cident (interior problem) and scattered (exterior problem)
coincide with the centre of the HATS’ head. sound field created by a single point source positioned at
In order to further verify the correct alignment of the (rs , θs , δs ) and imposing the boundary condition of zero
setup, all loudspeaker locations with elevation −55◦ ≤ δ ≤ total radial velocity on the rigid sphere of radius r = a, the
−40◦ were measured with the HATS standing both up- pressure on its surface can be computed as [19]
right and upside down. The corresponding transfer func-
tions were compared during the measurement process and p(r = a, θ, δ) = (1)
were found to match very well (absolute magnitude er- ∞
h−m (krs )
i−1
ror between the upright and upside down HRTF measure- (2m + 1)Pm (cos γ),
ments, averaged across all available overlapping locations, m=0 (ka)2 h−m (ka) h−0 (krs )
−55◦ ≤ δ ≤ −40◦ and 0◦ ≤ θ < 360◦ , < 1 dB up to at
least 4–5 kHz and < 2 dB up to at least 10 kHz), confirm- where h−m is the mth-order spherical Hankel function of the
ing the centred placement of the HATS in both conditions. second kind, h−m its derivative and γ is the angle between
Table I summarizes all 1784 measurement positions. the vector of the point source and that of the observation
HRTFs are typically normalized by the free-field re- point. Pm is the mth-degree Legendre polynomial defined
sponse measured with the same loudspeaker and micro- as
phones as used in the HRTF measurements, but with the 1 dm 2
microphones placed in the origin of the applied coordinate Pm (x) = (x − 1)m . (2)
2m m! dxm
system (here the centre of the hoop) and the head (HATS)
removed [1]. Unfortunately, due to time restrictions on the Equation (1) assumes waves normalized so that they have
use of the facility, the loudspeaker response couldn’t be unit amplitude and zero phase at the origin, having an im-
measured with the in-ear and the BTE microphones in a plied e+iωt time dependence.

838
Oreinos, Buchholz: Measurement of HRTFs ACTA ACUSTICA UNITED WITH ACUSTICA
Vol. 99 (2013)

Table II. Spherical head model parameters. centre of the hoop and [xic , 0, zic ] is the position vector
of the centre of the hoop (centre of the interaural axis)
Parameter Value referenced to the centre of the HATS head. This change of
Head radius 0.1 m coordinate systems is necessary so that, for example, the
Ear canal entrance locations L: (100◦ , −11◦ ) angle γ between the vector of the left ear (100◦ , −11◦ ), as
R: (−100◦ , −11◦ ) referenced to the HATS’ centre system, and the leftmost
Front BTE microphone locations L: (104◦ , 10◦ ) source point, whose coordinates expressed in the system
R: (−104◦ , 10◦ ) centred at the midpoint of the interaural axis are (90◦ , 0◦ ),
Rear BTE microphone locations L: (109◦ , 10◦ ) is correctly calculated to be zero.
R: (−109◦ , 10◦ )
“Point” source distance 1.2 m One-sided pressure spectra were calculated with the
spherical head model for each of the 6 microphone lo-
cations (see Table II) at 4097 (nfft = 8192) equidistant
frequency points between 0 Hz and fs /2. The correspond-
The spherical head model response was computed using ing real-valued head-related impulse responses (HRIRs)
the parameters shown in Table II. Given the non-spherical were then derived after first “mirroring” the resulting spec-
nature of the HATS’ head as well as the existence of tra to negative frequencies by applying the complex con-
the torso, which was disregarded by the applied spherical jugate transformation: H (−iω) = H ∗ (iω) and then apply-
head model, the parameters of TABLE II were adjusted ing an inverse (discrete) Fourier transform (nfft = 8192).
in a more heuristic, rather than rigorous, way. The head- Note that the nfft was chosen long enough to avoid time-
related values were chosen by comparison to their loca- aliasing after the successive time shifting and filtering of
tions on the HATS while iteratively minimizing the RMS the impulses.
error of the modelled and measured ILDs and ITDs. The
sound source distance was taken to be exactly the radius
of the hoop where the speaker was mounted. The result- 2.3. Combination of the head model with the mea-
ing mean absolute errors, averaged across all 1784 mea- sured data
dB
surement points, were |ILDdB meas − ILDmod | < 0.9 dB and A primary question that needed to be addressed was how
|ITDmeas − ITDmod | < 46 µs, where the ILDs were esti- to combine the measured HRTFs with the transfer func-
mated at the region 300–600 Hz. A formal optimisation tions derived from the spherical head model at low fre-
was impractical due to the computational complex nature quencies. Algazi et al. [14] have proposed cross-fading
of the spherical head model. the magnitudes of the model and the data (linearly in
It should be mentioned that alternative approaches exist dB) while either keeping the entire phase response of the
to the spherical head model, e.g. the snow-man model de- model, or re-inserting the estimated time of arrival of the
scribed by Algazi et al. [14], which are expected to provide data impulses to the minimum-phase inverse-DFT of the
a more accurate representation of the low frequency re- combined (cross-faded) magnitude responses. The latter
sponse of the HRTFs. However, the spherical head model method, called minimum-phase reconstruction [15], has
already resulted in very small mean absolute ITD and ILD been perceptually verified for sound localization. How-
errors and thus, such complex models were not further ever, the measured transfer functions of this HRTF set
considered here. will be used as inputs to hearing aids whose processing
In order to maximize the computational efficiency, is very different from human auditory processing. In par-
equation (1) was computed using a set of recursive for- ticular, multi-microphone signal enhancement techniques
mulas as described by Duda et al. [16] instead of being di- (e.g., directional microphones and adaptive beamformers)
rectly computed via the analytic expression. Given that the are highly sensitive to very small phase and amplitude
origin of the coordinate system used in the above spherical variations. Hence, the HRTFs must preserve as accurately
head model is defined in the centre of the spherical head as possible phase and amplitude relationships, especially
and thus is different from the one used in the HRTF mea- between the individual BTE hearing aid microphone pairs.
surements (see section 2.1), the source positions (radius In this regard the method proposed by Algazi et al. [14] did
rs , azimuth θs , elevation δs ) to be used for the spherical not provide satisfactory results and needed to be modified.
head model computation had to be derived from the source The combination method pursued in this study consists
positions of the HRTF data set according to of time aligning the model’s impulse responses to the cor-
 y  responding measured impulse responses and then com-
θs = arctan , bining them using “crossover” type low-pass and high-
x + xic pass filters with cut-off frequencies in the regions of 400–
 z + zic 
δs = arctan , (3) 500 Hz. First, the time of arrival (TOA) for all 1784 mea-
(x + xic )2 + y 2 sured and modelled HRIRs was estimated separately for

rs = (x + xic )2 + y 2 + (z + zic )2 , the in-ear and BTE microphones according to the method
proposed by Nam et al. [20] as the time instant of the max-
where [x, y, z] = [rs cos θs cos δs , rs sin θs cos δs , rs sin δs ] imum peak of the absolute of the cross-correlation func-
is the position vector of a given source, referenced to the tion between a given HRIR and its minimum-phase ver-

839
ACTA ACUSTICA UNITED WITH ACUSTICA Oreinos, Buchholz: Measurement of HRTFs
Vol. 99 (2013)

sion,

 
  ITU−T P.58 recom. values
 
τ̂ = argmax  h[n − τ]hmp [n] . (4) 20 Left in-ear mic.
τ  n  Right in-ear mic.
15

Magnitude (dB)
Since the BTE hearing aid microphones were only spaced
apart by dmic = 0.009 m, their maximum TOA difference 10
across all source locations was about 26 µs. Considering
5
that the applied sampling frequency was fs = 44.1 kHz
and thus, the resolution of the applied TOA estimation 0
method was only T = 1/fs ≈ 22.7 µs, the HRIRs had
to be up-sampled (a factor of 100 was chosen) before the −5
2 4
10 103 10
cross-correlation operation of equation (4) was performed. Frequency (Hz)
An alignment delay was then calculated from the esti-
mated TOA of the modelled HRIRs, τ̂mod , and the corre- Figure 2. Processed (0◦ , 0◦ ) HRTF magnitude response, extended
sponding measured HRIRs, τ̂meas , given by to low frequencies using the spherical head model, compared
against the values of the ITU-T P.58 recommendation.
[τ̂mod − τ̂meas ]
Δ= . (5)
100
otherwise be introduced by adding the low-pass and high-
To preserve even the smallest phase differences between pass HRTF components. The filtering was done in the fre-
each BTE microphone pair, the average alignment delay quency domain, as mentioned above, and an impulse re-
for each microphone pair was always applied to both mi- sponse was computed taking the inverse DFT (as in sec-
crophones. In this way any (small) TOA estimation er- tion 2.2) of the combined modelled/measured conjugate-
ror only affected the absolute timing of the BTE micro- mirrored transfer function.
phone pair signals, but the inter-microphone phase dif- The final impulse responses of the combined HRTFs
ferences provided by the spherical head model were pre- were then truncated to 256 samples using a one-sided
served. Since the combination of the measured and mod- Tukey window with a tapered-to-constant-section-ratio of
elled HRTFs was performed in the frequency domain, the 1/5. Special care was taken in the truncation process to
derived alignment delays Δ were transformed into com- maintain the very short pre-ringing that was introduced by
plex gain factors, the linear phase crossover filtering process. Finally, the im-
pulse responses were saved separately to MATLAB files
gΔ(if ) = e−i2πf Δ/fs , (6) (.mat extension), with every file having the impulse re-
sponses of all six microphones relevant to a given source
which were then multiplied with the corresponding mod- position. The final HRTF data set is publically available at
elled HRTFs. No further amplitude adjustments were re- http://www.nal.gov.au/download/HATS_BTE_hrirDatabase.zip.
quired than those already described in section 2.1.
Inspired by the design of loudspeaker crossovers, 256-
tap long linear-phase FIR filters were used to realise the 3. Results
crossover between measured and time-aligned modelled
HRTFs that resembled the magnitude spectrum of a com- Figure 2 shows the derived (0◦ , 0◦ ) HRTF response (mea-
plementary pair of asymmetric Linkwitz-Riley filters. An surement combined with the spherical head model), along
8th order low-pass filter with a cut-off frequency of 500 Hz with the corresponding ITU-T P.58 recommendation’s
and a 4th order high-pass filter with a cut-off frequency of range of values, highlighting the compliance with this
400 Hz were chosen. The 4th order HP filter slope com- standard. The HRTFs at 0◦ , 90◦ , 180◦ , 270◦ on the hor-
bined with the roll-off of the measurements effectively re- izontal plane were also compared to the relevant ITU-T
sulting in a higher order HP response, almost complimen- P.58 values (not shown here) and were similarly found to
tary to the 8th order LP filter. The different slopes and comply.
cut-off frequencies of the filters was chosen to: (1) ap- The magnitude spectra of the measured, modelled, and
ply a small gain-peak of 1–2 dB around 300–500 Hz and combined HRTFs are plotted in Figure 3 for an example
in turn to provide a smoother transition between measured source direction of (60◦ , −30◦ ). Responses for the left ear
and modelled HRTFs and (2) ensure good agreement of are shown in the left figures and for the right ear in the
the combined HRTF responses with data from the ITU- right figures. The top figures refer to the front microphone
T P.58 recommendation and the HATS’ own calibration of the BTE hearing aids and the bottom figures refer to the
chart. This filter design approach in the frequency domain in-ear microphones. In the top panels of Figure 3 it can be
controlled the steepness of the slopes of the cross-over fil- seen that the overall trend of the modelled HRTFs matches
ters and thereby limited the temporal ringing of the filters with the measured HRTF of the BTE microphone. In con-
in the time domain. Choosing a linear-phase filter design trast, the measured in-ear responses deviate significantly
minimized any phase (or comb-filter) effects that could from the model at higher frequencies. This deviation is

840
Oreinos, Buchholz: Measurement of HRTFs ACTA ACUSTICA UNITED WITH ACUSTICA
Vol. 99 (2013)

20 20 measurement
model
10 10 combined data
0 0
−10 −10
Magnitude (dB)

−20 −20
2 3 4 2 3 4
10 10 10 10 10 10

20 20
10 10
0 0
−10 −10
−20 −20
2 3 4 2 3 4
10 10 10 10 10 10
Frequency (Hz)

Figure 3. Spectrum magnitude of HRTF measurements (dashed

curve), spherical head model (dash-dotted curve) and final com-
bined and windowed responses (thick solid curve) relevant to the (a)
source direction . The left and right panels show left and right
ear microphone responses, respectively. The top curves refer to
the front BTE microphones and the bottom curves to the in-ear
microphones.

mainly due to the ear canal resonances, the most promi-

nent of which is at around 2.5 kHz, as well as the effect
of the pinnae. At low frequencies the application of the
spherical head model increases the usable bandwidth and
removes the low frequency dips and peaks, which are due
to the measurement chamber (which is not anechoic below
300 Hz) room modes.
When comparing the details contained in the modelled
and measured spectra, in particular at frequencies above
about 1 kHz, clear differences can be observed. These dif-
ferences can be even more pronounced at other source po-
sitions, especially at the contralateral side where more ex-
(b)
tensive rippling occurs. However, these discrepancies are
expected because the shape of the head of the HATS is
not strictly spherical and the presence of the torso and pin- Figure 4. Contour plots of the magnitude of HRTFs to the: (a)
nae has not been considered in the spherical head model. in-ear and (b) hearing aid (front) microphone of HATS’ left side.
The abscissa of the plots depicts the source direction on the hor-
Since the model is only applied to extend the usable fre-
izontal plane, with azimuth θ ∈ [−180◦ , 180◦ ].
quency range below about 300 Hz, the mentioned discrep-
ancies are not significant and can be neglected.
Figure 4 shows the HRTF magnitudes of the left in-ear
and front HA microphones for all available source posi- plotted in Figure 5 for sources located on the horizontal
tions on the horizontal plane. The contour plots of the plane. ILDs were calculated as the magnitude-difference
in-ear responses can be qualitatively compared to similar between the left and right ear’s HRTFs. A number of ob-
plots found in literature [11, 7, 17, 21]. However, a quanti- servations can be made from these plots. First, due to the
tative one-to-one comparison is not meaningful unless one different pinnae used on the measured HATS the ILD for
is comparing measurements from the same mannequin, fit- θ = 0◦ is slightly different from zero at high frequencies.
ted with the same pinnae and ear canal and simulator and A second observation is that for positive azimuths close to
driven by a similar measuring loudspeaker at the same dis- 90◦ (source on-axis with left ear) a positive ILD, which
tance. Comparing Figure 4a and 4b, it is evident that the takes substantial values at high frequencies, is observed,
ear-canal resonances at ∼2.5 kHz and ∼5 kHz are elimi- while an almost symmetrical picture up to about 5 kHz is
nated in the BTE responses. Additionally, some of the high seen for negative azimuths (above ∼5 kHz the used pin-
frequency fine structure appears smoothed and attenuated nae cause deviations to the symmetric nature of the ILDs).
due to the absence of pinna reflections. The theoretical bright spots at ±90◦ are also confirmed
As a next step to validate the combined HRTFs, the on that plot (i.e., a source at 90◦ creates a bright spot at
Interaural Level Differences (ILDs) versus frequency are 180◦ around the perimeter of the sphere, that is at −90◦ ,

841
ACTA ACUSTICA UNITED WITH ACUSTICA Oreinos, Buchholz: Measurement of HRTFs
Vol. 99 (2013)

where the contralateral ear lies). As mentioned above for

the HRTF magnitude case, it is not straightforward to com-
pare the ILD contours from another data set, unless it is en-
sured that all geometrical, structural and acoustical prop-
erties of the setup are the same. The ILDs measured at the
BTE-microphones (not shown here) look very similar to
the in-ear ILDs of Figure 5 at low frequencies, but above
about 5 kHz are significantly smaller due to absence of the
pinna and ear canal effects.
The broadband ITD of the measured HRTFs as well
as the modelled HRTFs before and after TOA alignment
(see section 2.3) are shown in Figure 6 for the in-ear
microphones as a function of source azimuth. The ITDs
were calculated as the TOA difference between the left
and right ear microphones using equation (4). The ITDs
for the measurement (solid line) as well as the TOA-
aligned model (dashed line) match very well, confirming
the accuracy of the TOA-alignment procedure described
in section 2.3. Moreover, these ITDs exhibit the typical
ITD shape measured in humans as seen for example in Figure 5. Interaural level differences between left and right in-ear
[20] or [22]. The ITDs for the modelled HRTFs before microphones (HL − HR ), versus frequency, relevant to positions
TOA-alignment (dashed-dotted line) show a slightly less lying on the horizontal plane with azimuth θ ∈ [−180◦ , 180◦ ].
“peaky” behaviour than the ITDs for the measured HRTFs,
but otherwise provide a very good fit and thus confirm the
applicability of the parameters given in Table II for the
calculation of the spherical head model responses. As ex- measurements
600
pected, the ITDs estimated from the HA microphones (not model
400 aligned model
shown here) look almost the same as those of Figure 6.
In order to verify the accurate representation of phase 200
ITD (us)

(or timing) of the combined HRTFs, in particular between 0

the front and back microphones of the BTE hearing aids, −200
3D directivity plots of a hyper-cardioid and an omnidirec-
−400
tional microphone were realised and shown in Figure 7
for two example frequencies, 300 Hz and 1600 Hz. The −600
radial distance of each directivity plot surface node rep- −800
resents the magnitude of the directional microphone out- −180 −120 −60 0 60 120 180
Azimuth (deg)
put response (in dB) due to a point-source located at that
node’s azimuth and elevation at a distance of r = 1.2 m
(distance between loudspeaker and HATS’ head centre). Figure 6. Interaural time differences between left and right in-ear
microphones relevant to positions lying on the horizontal plane
The hyper-cardioid output was generated by applying de-
with azimuth θ ∈ [−180◦ , 180◦ ]. The ITD estimates of the mea-
lay and subtraction beamforming [13] to the HRTFs of the
surements and the model, before and after time alignment, are
front and back microphones of the left BTE hearing aid. plotted.
Such an approach for creating directional microphones is
highly sensitive to phase and amplitude errors, in partic-
ular at low frequencies. If the phase difference between to the first order ideal free-field patterns. In order to illus-
the microphones is slightly disrupted (as was the case in trate the head shadow effect on the microphone direction-
the measured HRTFs at low frequencies), the directivity ality, the directivity patterns of an omnidirectional BTE
plots get severely distorted. The bottom polar plots of Fig- microphone on the head (i.e. only considering the front
ure 7 show the hyper-cardioid directivity attained at low microphone) are shown in the top figures of Figure 7 at
frequencies (300 Hz) to be smooth and resembling that at 300 Hz and 1600 Hz. They are very similar to the direc-
higher frequencies (1600 Hz). Thus the low frequency re- tivity patterns shown by Kates [13], further demonstrating
gion where the model dominates the combined response the validity of the measured HRTFs.
produces a directional result that resembles that at higher Lastly, Figure 8 shows the SNR, averaged across all six
frequencies where the measured data dominate. microphone channels (which gave very similar values) and
Note that the presented directivity patterns differ from all measurement positions (solid line). A noise estimate
the well-known ideal hyper-cardioid polar plot [13] due to was formed using the last 256 samples of the 2048-sample
the existence of the head. The head shadowing effectively long unprocessed measurement data. A signal estimate (in
creates a higher-order directional behaviour, as compared fact signal + noise) was formed using samples (100:999)

842
Oreinos, Buchholz: Measurement of HRTFs ACTA ACUSTICA UNITED WITH ACUSTICA
Vol. 99 (2013)

(a)

Figure 8. The frequency-dependent mean SNR of all microphone

channels averaged across all positions (solid curve) along with
the SNR of the measurement with the lowest broadband SNR
(dash-dotted curve).

but describes the effective “SNR” after averaging and

de-convolving the recorded sine sweeps. The frequency-
dependent SNR of the measurement with the worst broad-
band SNR, occurring at the position (30◦ , −85◦ ), is also
(b)
plotted as a worst-case scenario. It can be seen that the
SNR at low frequencies (below 300–400 Hz) is signif-
icantly lower than the values at mid and high frequen-
cies. Especially if the 6 dB/octave roll-off of a delay-and-
subtract directional HA microphone is considered, the out-
put SNR becomes negative at low frequencies. This con-
firms again the necessity of applying the spherical head
model for a “phase-preserving” extrapolation of the mag-
nitude to low frequencies.

4. Summary and conclusions

(c)
The procedure for the measurement of an extensive set of
HRTFs from 1748 directions covering the whole sphere
to the 2 in-ear and 4 BTE hearing aid microphones on
a HATS was presented. A spherical head model was ap-
plied to extend the usable frequency range of the mea-
sured HRTFs towards low frequencies. A method was also
proposed to combine the measurements with the model
responses while maintaining the very sensitive amplitude
and phase information carried by the HRTFs, which is par-
ticularly critical between the front and back hearing aid
microphones. The applicability of this approach as well as
(d)
the validity of the entire HRTF data set, especially when
used for hearing aid applications, was confirmed by ana-
lyzing magnitude spectra of the combined HRTFs, ITDs,
Figure 7. Directivity patterns of the left BTE microphone pair ILDs and directivity plots of hyper-cardioid and omni-
(worn on the HATS) for an omnidirectional configuration (i.e., directional microphones.
using only the signals of the front BTE microphone) at (a)
300 Hz, (b) 1600 Hz and a hyper-cardioid configuration at (c) The developed HRTF data set covering the full sphere
300 Hz, (d) 1600 Hz. The 2D shadow plots illustrate the directiv- at a high resolution can be used in a wide range of ap-
ity plots of the 3D patterns on the horizontal plane. plications ranging from recreating virtual auditory spaces
to research and development with hearing devices, eval-
uation of hearing aid algorithms, including psychoacous-
from the measurement data. The estimated SNR does not tic aspects, and verification of multi-channel loudspeaker
strictly depict the true microphone – amplifier – A/D SNR systems for recreating realistic acoustic scenes. The ex-

843
ACTA ACUSTICA UNITED WITH ACUSTICA Oreinos, Buchholz: Measurement of HRTFs
Vol. 99 (2013)

istence of measurements from both in-ear and BTE mi- function estimate. Proc. International Conference on Audi-
crophones permits realistic simulations of aided scenarios tory Display, 2003, 119–122.
where sound reaches the hearing aids of the wearer but at [9] T. Nakado, T. Nishino, K. Takeda: Head-related transfer
the same time leaks through to the eardrum (this “acous- function measurement in sagittal and frontal coordinates.
Acoust. Sci. & Tech. 29 (2008) 335–337.
tic path” will be affected by the fitting of the hearing aid,
which may include ear-molds with different vent sizes). [10] H. Kayser, S. D. Ewert, J. Anemüller, T. Rohdenburg, V.
Hohmann, B. Kollmeier: Database of multichannel in-ear
and behind-the-ear head-related and binaural room impulse
Acknowledgement responses. EURASIP, J. on Advances Signal Processing
(2009).
The authors acknowledge the financial support of the
HEARing CRC, established and supported under the Co- [11] B. P. Bovbjerg, F. Christensen, P. Minnaar, X. Chen: Mea-
suring the head-related transfer functions of an artificial
operative Research Centres Program – an initiative of the head with a high directional resolution. AES 109th Con-
Australian Government. In addition, we would like to vention, Los Angeles, USA, 2000.
thank Johahn Leung, Heather C. Kelly, and Prof. Simon [12] M. Pauli, S. Favrot, J. M. Buchholz: Improving hearing
Carlile for letting us use the Auditory Neuroscience Lab- aids through listening tests in a virtual sound environment.
oratory, Department of Physiology, University of Sydney, Hearing Journal 63 (2010) 40–44.
as well as two anonymous reviewers for their helpful sug- [13] J. M. Kates: Digital hearing aids. Plural Publishing, San
gestions. Diego, CA, 2008.
[14] V. R. Algazi, R. O. Duda, D. M. Thompson: The use of
References head-and-torso models for improved spatial sound synthe-
sis. AES 113rd Convention, Los Angeles, USA, 2002.
[1] J. Blauert: Spatial hearing: The psychophysics of human [15] D. J. Kistler, F. L. Wightman: A model of head-related
sound localization. MIT, Cambridge, MA, 1983. transfer functions based on principal components analysis
[2] E. M. Wenzel, M. Arruda, D. J. Kistler, F. L. Wightman: and minimum-phase reconstruction. J. Acoust. Soc. Am.
Localization using nonindividualized head-related transfer 91 (1992) 1637–1647.
functions. J. Acoust. Soc. Am. 94 (1993) 111–123. [16] R. O. Duda, W. L. Martens: Range dependence of the re-
[3] S. Moreau, J. Daniel, S. Bertet: 3D sound field recording sponse of a spherical head model. J. Acoust. Soc. Am. 104
with higher order ambisonics. Objective measurements and (1998) 3048–3058.
validation of spherical microphone. AES 120th Conven- [17] D. S. Brungart, W. M. Rabinowitz: Auditory localization of
tion, Paris, France, 2006. nearby sources. Head-related transfer functions. J. Acoust.
[4] D. B. Ward, T. D. Abhayapala: Reproduction of a plane- Soc. Am. 106 (1999) 1465–1479.
wave sound field using an array of loudspeakers. IEEE [18] E. G. Williams: Fourier acoustics. Academic, San Diego,
Trans. Speech Audio Process. 9 (2001) 697–707. CA, 1999.
[5] N. Epain, P. Guillon, A. Kan, R. Kosobrodov, D. Sun, [19] J. Daniel: Spatial sound encoding including near field ef-
C. Jin, A. van Schaik: Objective evaluation of a three- fect: Introducing distance coding filters and a viable, new
dimensional sound field reproduction system. Proceedings ambisonics format. AES 23rd International Conference,
of 20th International Conference on Acoustics, ICA, Syd- Copenhagen, Denmark, 2003.
ney, Australia, 2010. [20] J. Nam, J. S. Abel, J. O. Smith: A method for estimating
[6] W. G. Gardner, K. D. Martin: HRTF measurements of a interaural time difference for binaural synthesis. AES 125th
KEMAR. J. Acoust. Soc. Am. 97 (1995) 3907–3908. Convention, San Francisco, USA, 2008.
[7] V. R. Algazi, R. O. Duda, D. M. Thompson, C. Avendano: [21] R. Duraiswami, D. N. Zotkin, N. A. Gumerov: Interpola-
The CIPIC HRTF database. Proc. IEEE Workshop on Ap- tion and range extrapolation of HRTFs. Proceedings of the
plications on Signal Processing to Audio and Acoustics, ICASSP ’04, 2004.
2001, 99–102. [22] S. Busson, R. Nicol, B. F. G. Katz: Subjective investi-
[8] E. Grassi, J. Tulsi, S. Shamma: Measurement of head- gations of the interaural time difference in the horizontal
related transfer functions based on the empirical transfer plane. AES 188th Convention, Barcelona, Spain, 2005.

844