AES_108_preprint

____________________________

* Corresponding author: dbegault@mail.arc.nasa.gov

Direct comparison of the impact of head tracking, reverberation,

and individualized head-related transfer functions on the spatial

perception of a virtual speech source

Audio Engineering Society 108

Conference, 19-22 February 2000, Paris

Durand R. Begault / NASA Ames Research Center *

Elizabeth M. Wenzel / NASA Ames Research Center

Alexandra S. Lee / San José State University

Mark R. Anderson / Raytheon STX Corporation

ABSTRACT

A study was performed using headphone-delivered virtual speech stimuli,
rendered via HRTF-based acoustic auralization software and hardware, and
blocked-meatus HRTF measurements. The independent variables were
chosen to evaluate commonly held assumptions in the literature regarding
improved localization: inclusion of head tracking, individualized HRTFs, and
early and diffuse reflections. Significant differences were found for azimuth
and elevation error; reversal rates; and externalization.

0. INTRODUCTION

A review of the literature related to perception of headphone-delivered 3-D sound

suggests that optimal localization performance and perceived realism result from
inclusion of factors in virtual acoustic  synthesis that emulate our everyday spatial
hearing experiences. These include (1) head-tracked virtual stimuli  [1-3], (2)
synthesis of a realistic diffuse field via auralization techniques  [4-6], and (3) the use
of individualized (as opposed to non-individualized) head-related transfer functions
(HRTFs) ([7-9]). However, these  conclusions have never been tested within a single
study that directly compared the perceptual effects of these factors to determine
their comparative advantages. The motivation of the current study was therefore to
essentially pit these factors against one another in a contest, using speech stimuli.

In reviewing the large number of studies that have evaluated virtual localization

accuracy, different types of stimuli and experimental paradigms are used, making
comparison difficult. In particular, noise stimuli probably bias performance

Begault, D. R., et al., “Direct comparison……”

differentially when compared to the use of more familiar everyday stimuli, such as
speech. For example, we cannot use any sort of real-world, cognitive reference
regarding the distance of a noise burst. Noise stimuli are particularly inapplicable to
reverberation studies, since the diffuse field of a sound decay bears similarity in the
time domain to frequency-weighted noise, resulting in “noise convolved with
decaying noise”. Such a stimulus would lack meaningful sensory cues from the
reverberant field.

Brief (3 s) segments of speech stimuli were used in the current experiment both

because reverberation was evaluated, and because of its obvious relevance to 3-D
audio applications such as teleconferencing. The trade-off in using speech for
evaluating localization accuracy is that the average speech spectrum does not
contain a significant level of acoustical power at those frequencies where the HRTF
yields elevation cues that can be used by a listener. For that reason, only eye-level
elevations were evaluated in the current study, as in a previous speech localization
study we had conducted [10].

Individualized HRTFs have been cited in the literature as improving localization

accuracy, increasing externalization, and reducing reversal errors [2, 9, 11]. Møller
et al. conducted a source-identification experiment using speech stimuli that
suggested that non-individualized HRTFs resulted in an increased number of
reversals, but had no effect on externalization; however, all of the conditions in that
study were made under reverberant conditions and without head-tracking [9].

The literature also indicates that reverberation, and even only a few “early

reflections” (or even attenuated delays) are sufficient to produce image
externalization [5, 12, 13]. In the current study, the experimental variables included
anechoic stimuli, stimuli with HRTF-filtered early reflections (0-80 ms), and stimuli
with “full auralization” reverberation (the same early reflections plus additional
simulation of the late reverberant sound field from 80 ms - 2.2 s).

2. SUBJECTS

Nine paid participants (4 female 5 male, age range 18-40, hearing ability 15 dB

HL) inexperienced with virtual acoustic experiments participated. They were not told
the purpose of the experiment.

Prior to the experiment, subjects were instructed how to make azimuth, elevation,

distance and realism judgments on the interactive, self-paced software interface
described below. For distance judgments, they were instructed to pay attention to
whether or not the sound seemed inside or outside of the head, and that the
maximum distance judgment possible on the graphic corresponded to “more than 4
inches outside the edge of the head.” Instructions were given on the computer
screen before each experimental trial that utilized head tracking requesting that the
subject move their head; they were not trained to move their head in any particular
manner. Subjects were encouraged to stop when fatigued, and were given breaks
between experimental blocks at their request. Subjects required three to four days to
complete the entire experiment

Begault, D. R., et al., “Direct comparison……”

3. EXPERIMENTAL DESIGN

The experimental design shown in Figure 1 was developed to evaluate all

combinations of independent variables. “Anechoic”, “early reflection” and “full
auralization” refers to the level of diffuse field simulation used. “Individualized”
HRTFs were measured for each subject, while “generic” HRTFs were derived from
the Institut für Technische Akustik Kunstkopf from the University of Aachen (supplied
as part of the room modeling software). ”Tracking on-off” refers to whether or not the
direct sound HRTFs and (when applicable) the early reflection HRTFs were updated
in real time in response to head movement.

Each subject was run under each of the treatment combinations shown in the first

three rows of Figure 1, resulting in a total of 12 conditions. Only six azimuth positions
were used, all at eye level elevation: left and right 45 degrees; left and right 135
degrees; and 0 and 180 degrees. These positions represent pairs that lie on two
different “cones of confusion” and a pair directly on the median plane  [14]. The
elements of each pair are similar in that the interaural  time delay (ITD) is the same
for left 45 and 135 degrees, 0 and 180 degrees, and right 45 and 135 degrees. All
positions simulated were at eye level (0 degrees elevation). Each azimuth position
was evaluated 5 times in each block, resulting in 30 trials per block. The order of
blocks, azimuth positions, and speech stimuli were randomized, while the particular
combination of experimental treatments held constant for each block. Each trial took
approximately 10 seconds to complete.

4. STIMULI

4.1 Experimental hardware and software

Stimuli were monitored by subjects over headphones (Sennheiser HD-430) in a

sound proof booth, at a level of approximately 60 dBA. This level corresponds to a
normal speech level at a distance of 1 meter. A head-tracking device (Polhemus
FastTrak) was attached at all times to the subjects’ headphone. The head tracker
was interfaced via a tcp/ip socket connection with the simulation software and
hardware (Lake DSP “Headscape” and “Vrack” software, CP-4 hardware) updating
the stimuli in response to head movement at an update interval of 33 Hz (30 ms).
The end-to-end latency throughout the entire hardware system averaged 45.3 ms
(SD = 13.1 ms), as measured via a swing-arm apparatus (described in [15, 16]).

For simulating different azimuths, the virtual simulation turned the receiver about

their central axis to face different directions, rather than moving the sound source.
The hardware did not re-synthesize the stimuli in response to subject  tilt or roll, only
to yaw (azimuth).  Nevertheless, the simulated reverberation always included a full 3-
D simulation.

Custom software drove the experimental trials and data collection, and generated

playback of speech recordings from a digital sampler (randomized 3 s anechoic

Begault, D. R., et al., “Direct comparison……”

female and male voice segments from the Archimedes test CD [17]). The software
also coordinated the Lake Vrack program, the sound generator, the head tracking
device, and the stimuli generation.

Subjects indicated their responses using an interactive graphic shown in Figure

2. The details of the interface were described previously in [3]

Subjects were

required first to indicate the azimuth and the distance on the left panel, and then the
elevation in the right panel of the display, and then finally to adjust the slider at the
bottom to indicate “perceived realism”. Subjects were forced to make all judgments
before they could proceed to the next trial.

The distance of the outer circle from the center of the head was twice the radius

from the center to the edge of the head in order to prevent biasing the
externalization judgments. For example, a very large head relative to the available
area on the display would probably yield more inside-the-head localization
judgments. The “perceived realism” was indicated on a continuous scale, but
indicated the words “bad”, “poor”, fair”, “good” and “excellent” at locations
corresponding to a 0 – 4 rating scale. No special instructions were given on how to
interpret ”realism” or what specifically to listen for.

4.2 HRTF measurement

Before starting the experiment, subjects had their HRTFs measured using a

modified Crystal River Engineering “Snapshot” system. This system is based a
blocked-meatus measurement technique (see, e.g., [18]) and has been used in our
previous experiments. The HRTF is measured iteratively via a single speaker, with
the subject moving on a rotating chair for each measurement; the speaker is moved
vertically along a pole to obtain a set of azimuth measurements at a specific
elevation. Room reflections are eliminated from the measurement, and
compensation for subject movement is made in the calculation of interaural time
delays.

A software script (Matlab from The MathWorks, Inc.) guided the experimenter in

measuring the HRTF map at 30 degree azimuth increments, at 5 elevations:  -36,  -
18, 0, 18, 36 and 54 degrees relative to eye level. A head-tracking device was
monitored by the experimenter to position the subject for each measurement. Golay
code pairs are used to obtain the raw impulse response; subsequently, a diffuse field
equalization that compensates for the headphone, microphone and loudspeaker
transfer functions is applied.  The measurements are eventually processed into a
HRTF “map”, consisting of an  array of 256 point, minimum-phase impulse responses
at a sample rate of 44.1 kHz. Further details will appear in a later version of this
paper.

4.3 Reverberation simulation

A room prediction-auralization software package (CATT-Acoustic) was used to

generate both an individualized and a generic binaural version of an existing multi-

Begault, D. R., et al., “Direct comparison……”

purpose performance space. This space has a volume of about 1000 m

(8.9 m

width, 14.3 m length, 7.9 m height) and is essentially a rectangular box with gypsum
board mounted on stud walls, a cement ceiling, and a wooden parquet floor with
seating. Velour curtains are on the two opposite long sides of the hall. The modeled
source and receiver were positioned asymmetrically in the room 5.3 m apart, 1 m off
the center line of the room and with the source 4.7 m off of the back wall. The source
directivity was specified in octave bands based on a model of the human voice. A
total of 10,088 rays were calculated using a hybrid cone tracing technique (see [19,
20]).

The acoustical characteristics of the modeled room in the experiment were

initially compared to measurements made in the real room to verify the match
between significant acoustical parameters such as octave-band reverberation times
and early reflection arrivals (see [21]). The agreement between the modeled and real
room was within 0.1 s in each octave band from 125 Hz- 4 kHz, with a mid-band RT
of 0.9 s. However, the existing space was considered overly dry (unreverberant) for
purposes of the current experiment; the reverberation for the full auralization was
meant to be obvious to the non-expert subjects. Consequently, the absorptive
materials in the virtual room were adjusted to achieve a mid-band reverberation time
of 1.5 s, partly by removing seats in the room model. The resulting reverberation
times were as follows:

Octave band center frequency (Hz)

125

250 500

Reverberation time T-30 (s)

2.2

1.9

1.5

1.4

The early reflections out to 80 ms contained no noticeable echoes, and were
uniformly dense from all directions after about 12 ms.

A customized Matlab script took the complete measured HRTF map made by

Snapshot and  produced an interpolated, up-sampled (44.1  – 48 kHz) version for
implementation in the room modeling program. The room modeling program then
generated 256 binaural impulse responses of the first 80 ms of the room response,
representing 1.5 degree increments of listener motion relative to the source. In
addition, two binaural impulse responses were generated representing the diffuse
field from 80 ms  – 2 s. The early reflection binaural impulse responses were updated
in real time for both the anechoic and early reflection conditions by the Lake
hardware/software for real-time, low-latency convolution. The anechoic condition
was achieved by using only the part of the impulse response representing the direct
arrival. The “full auralization” condition included the late diffuse reverberation
response (80 ms-2.2 s), which did not vary with head motion.

5. RESULTS

A 2 * 2 * 3 (head tracking on/off * generic/individual HRTF type * reverberation

treatment) univariate repeated measures analysis of variance (ANOVA) was
conducted for each of the five dependent measures: azimuth error, elevation error,

Begault, D. R., et al., “Direct comparison……”

front-back reversal rate, externalization, and subjective rating of sound realism. An
alpha level of .05 was used for all statistical tests, unless otherwise indicated. The
Geisser-Greenhouse correction was used to adjust for assumed violations of
sphericity when testing repeated measures effects involving more than two levels.

5.1 Azimuth error

Azimuth error was defined as the unsigned deviation, in degrees, of each

azimuth judgment from the target azimuth location, corrected for frontal plane and
median plane reversals. A significant main effect was found for reverberation type,
even after reducing the degrees of freedom by applying the Geisser-Greenhouse
correction: F(1.38, 11.05) = 5.19, p = .035; see Figure 3, top. The main effects for
head tracking and HRTF type were non-significant.

A post-hoc test (Scheffe) found no significant difference between the early

reflection and full auralization conditions. An additional post-hoc test of the combined
reverberant (early reflection and full auralization) conditions and  the anechoic
condition was also non-significant,  F(1, 8) = 6.08,  p  > .05, perhaps due to the
conservative nature of Scheffe test. Despite this lack of significance in the post-hoc
test, an estimation of partial  ù

indicated that the presence of reverberation in the

stimuli, excluding all other treatment effects, had a large effect on azimuth error
(partial ù

= .22)

Specifically, the reverberant conditions resulted in less azimuth

error (mean = 17.32, SD = 5.82) than anechoic conditions (mean = 23.19, SD =
9.07).

Although non-significant, the effect of head tracking on azimuth error, excluding

all other treatments, was moderate, having an estimated partial ù

of .14. Head-

tracked stimuli resulted in a smaller azimuth error (mean =17.64, SD =7.96)
compared to non-head--tracked stimuli (mean = 20.91, SD = 6.83).

The analysis also revealed a significant two-way interaction between head

tracking and HRTF for azimuth error,  F(1, 8) = 20.07,  p = .002; see Figure 3
(bottom). Subsequent post-hoc tests of the simple effects of head tracking in the
individualized and generic HRTF conditions (Bonferroni,  alpha = .0167) indicated no
significant effect of head tracking for either HRTF type, perhaps due to the
conservative nature of the Bonferroni test. Although non-significant, an estimated
partial  ù

of .30 indicates that the effect of head tracking in non-individualized HRTF

conditions was large, independent of all other treatments. For non-individualized
HRTF conditions, head-tracked stimuli resulted in smaller azimuth error (mean =
16.70, SD = 7.71) than no head tracking (mean = 21.74, SD = 7.82).

Partial ù

is a measure of relative treatment magnitude; it reflects the proportion of total variability

explained by an experimental treatment, independent of all other treatments [22]. In general, ù

values on the order of .01, .06, and .15 indicate small, medium and large treatment effects,
respectively.

Begault, D. R., et al., “Direct comparison……”

5.2 Elevation error

Elevation error was defined as the unsigned deviation, in degrees, of an

elevation judgment from an eye-level target (“0 degrees elevation”). A significant
main effect was found for reverberation, F(1.20, 9.57) = 5.15, p = .043, applying the
Geisser-Greenhouse correction; see Figure 4. No significant main effects or
interactions were found for head tracking or HRTF type.

In contrast to its facilitating effect in reducing azimuth error, the presence of

reverberation in the stimuli increased the magnitude of elevation errors (mean  =
28.67,  SD  = 17.76), compared to anechoic conditions (mean = 17.64, SD = 14.60).
A post-hoc comparison (Bonferroni,  alpha =  .0167) revealed no significant difference
in elevation error between early and full auralization conditions. In addition, a
Scheffe test found no significant difference between the combined reverberant
conditions and the anechoic conditions. However, an estimated partial  ù

of .21

indicates that the effect of reverberation on elevation error is rather large compared
to the anechoic condition, with reverberation increasing elevation error.

5.3 Reversal rates

The ANOVA indicated that head tracking significantly reduced reversals (front-

back and back-front confusion rates) compared to stimuli without head tracking,
F(1,8) = 31.14, p= .001; see Figure 5. The overall mean reversal rate for head-
tracked conditions was 28% (SD = 25%) compared to the 59% (SD = 12%) reversal
rate for non-head-tracked conditions. No other treatments or interactions yielded
significant effects on reversal rates. (A separate analysis of front-back and back-front
confusions has not been made at this time, but will be presented in a future paper).

5.4 Externalization

A sound is externalized if its location is judged to be outside of the head. For this

analysis, the cut-off point for treating a judgment as “externalized” was set to > 5
inches instead of > 4 inches in order to yield a conservative estimate that eliminated
judgments perceived at the edge of the head (“verged cranial”; see [23]). Note that
the unit of an “inch” is relative to the graphic shown in Figure 2, and does not
represent a literal judgment in inches by the listener. The edge of the head is set at 4
inches, and the maximum distance judgment possible by the subject was 8 inches.

A significant main effect of reverberation was found for the proportion of

externalized distance judgments, F(1.43, 11.43) = 13.43, p = .002 (Geisser-
Greenhouse correction); see Figure 6. Subjects externalized judgments at a mean
rate of 79% (SD = 23%) under the combined reverberant conditions, compared to
40% (SD = 29%) under the anechoic condition. No other treatments or interactions

Begault, D. R., et al., “Direct comparison……”

yielded significant effects for externalized stimuli. A post-hoc test (Bonferroni, alpha
= .0167) found no significant difference between the two reverberant conditions, but
the reverberation conditions combined (Scheffe test) resulted in significantly more
externalized judgments than the anechoic conditions, F(1,8) = 18.10, p < .05.

5.5 Perceived realism

Listeners were asked to rate the perceived realism of each stimulus on a

continuous scale that was subsequently encoded from 0 (least realistic) to 4 (most
realistic). No significant main effects or interactions were found. Realism ratings for
each of the 12 conditions, averaged over all nine participants, varied only from 2.42
and 2.97, with an overall mean realism rating of 2.71 (SD = .55) on the 0-4 scale.
This lack of ability suggests that the participants did not differentiate among
conditions based on perceived realism, or that they did not have a common
understanding of what “realism” meant. It can be assumed, then, that all 12
conditions were equally, moderately, “real” to the participants.

6. DISCUSSION

It is possible to evaluate these results in light of system requirements for

improved virtual acoustic simulation. Overall, these results would seem to indicate
that stimuli which include reverberation will yield lower azimuth errors and higher
externalization rates (here, by a ratio of about 2:1), but at the sacrifice of elevation
accuracy. The inclusion of head tracking will significantly reduce reversal rates, also
by a ratio of about 2:1, but does not improve localization accuracy or externalization.
Except for the interaction of head tracking and HRTF for azimuth error, there is no
clear advantage to including individualized HRTFs for improving localization
accuracy, externalization or reversal rates, within a virtual acoustic display of
speech.

The fact that individualized HRTFs did not significantly increase azimuth

accuracy is not surprising since most of the spectral energy of speech is in those
frequencies where ITD cues are more significant than spectral cues. Møller et al.
also found that individualized HRTFs gave no advantage for localization accuracy for
speech [9]. Further analyses will be conducted to examine the correlation between
the magnitude of the ITD difference between individual and generic HRTFs, versus
individual localization performance. One might expect a strong correlation between
localization error and the ITD difference magnitude.

It has been reported previously that reversals are mitigated by individualized

HRTFs  [7-9, 24, 25]. For instance, for noise stimuli in a non-dynamic simulation, an
almost 3:1 increase in reversals (11% to 31%) was reported when using non-
individualized HRTFs  [7, 25]. This is not the case here for speech stimuli. For non-
dynamic stimuli, the mean reversal rate found here (59%) was much higher than a
previous study using speech stimuli (37%)  [10]. This is probably because the current
study had one-third of the total stimuli on the median plane, where differences in

Begault, D. R., et al., “Direct comparison……”

spectral cues are minimal between front and back. Only one-twelfth of the stimuli in
the previous study were on the median plane.

Previous studies have reported on the problem of inside-the-head localization

and the externalization advantage of using reverberation [5, 13, 26]. The current
study indicates that only the presence of 80 ms of early reflections, as opposed to a
full auralization out to 1.5 s, is sufficient. The early reflections cause a lowering of the
interaural cross-correlation of the binaural signal over time as the speech is voiced,
relative to the anechoic stimuli [27]. It is likely that this increased differentiation of the
binaural signal over time is responsible for the externalization effect, as opposed to
the cognitive recognition of a room.

It was surprising that the presence of reverberation caused the accuracy of

azimuth estimation to improve by approximately 5 degrees. Typically, reverberation
introduces a smearing effect in the form of image broadening that would make the
loci of the localized image less precise. On the other hand, non-externalized
responses are not actually “localized” in the normal sense; it is possible that sounds
heard within the head are less precisely localized, and that localization resolution
improves beyond a certain distance. There could also have been a response bias; it
may have been easier on the graphic response screen to represent a perceived
azimuth more accurately when the distance judgment had increased.

Elevation biases were observed previously for virtual speech stimuli using non-

individualized HRTFs  [10, 28]. Particularly for sound sources at 0 degrees azimuth
and elevation, virtual acoustic stimuli  (as well as dummy head recordings) are
frequently perceived as located upwards, within or at the edge of the head, when
auditioned through headphones. We have not yet fully analyzed the elevation data
presented here in detail, but we can informally note that the bias is upwards as was
previously reported [10, 28]. There is no explanation for this phenomenon at this
point.

Head-tracking did not significantly increase externalization rates, nor did it yield

more accurate judgments of azimuth. Head-tracking primarily served to eliminate
reversals, a phenomenon explained by the differential integration of interaural cues
over time as the cone of confusion is resolved by head motion  [27]. These findings
contrast previous results indicating that  head movements enhance source
externalization  [13, 29, 30]. In the post-hoc analysis of the data presented here,
there was a strong effect for head tracking found for azimuth error, with head-tracked
stimuli resulting in about a 3-degree improvement. A future version of this paper will
examine head movement data to determine any correlations between accuracy and
the characteristics of head movement used by each listener. The fact that the stimuli
were only three seconds long could have been a contributing factor.

The perceived realism could have been affected by many factors. No instructions

were given, perhaps causing subjects to utilize different criteria that all resulted in a
relatively “neutral” judgment. It was surprising that full auralization with head tracking
was not judged as significantly more realistic than anechoic conditions. It may simply
be that realism is difficult to associate with a three-second segment of speech in a
laboratory simulation.

Begault, D. R., et al., “Direct comparison……”

In a future version of this paper, a more extensive analysis of the data will be

presented that will include reference to individual differences amongst the subjects,
an analysis of head tracking, and an exploration of the data for specific azimuths.
This latter issue is important since the proportion of median-plane stimuli in the
present study was large. This should be taken into account when comparing our
localization and reversal findings to previous studies.

ACKNOWLEDGEMENTS

Many thanks are due to Jonathan Abel and Kevin Jordan for their assistance in

the completion of this work. This paper was partially supported by NASA-San Jose
State University co-operative research grant NCC 2-1095.

REFERENCES

[1] Wightman, F.L. and Kistler, D.J., "The importance of head movements for
localizing virtual auditory display objects", in Proceedings of the 1994 International
Conference on Auditory Displays, Santa Fe, N.M. (1995).

[2] Wightman, F.L. and Kistler, D.J., "Factors affecting the relative salience of sound
localization cues", in Binaural and Spatial Hearing in Real and Virtual
Environments, R. Gilkey and T. Anderson, Eds., (Lawrence Earlbaum, Mahwah,
N.J., 1997).

[3] Wenzel, E.M., "Effect of increasing system latency on localization of virtual
sounds", in Audio Engineering Society 16th International Conference on Spatial
Sound Reproduction, (1999).

[4] Lehnert, H. and Blauert, J., "Principals of Binaural Room Simulation", Applied
Acoustics, vol. 36, pp. 259-91 (1992).

[5] Begault, D.R., "Perceptual effects of synthetic reverberation on three-
dimensional audio systems", Journal of the Audio Engineering Society, vol. 40, pp.
895-904 (1992).

[6] Kleiner, M., Dalenbäck, B.-I., and Svensson, P., "Auralization-an overview",
Journal of the Audio Engineering Society, vol. 41, pp. 861-875 (1993).

[7] Wenzel, E.M., Arruda, M., Kistler, D.J. and Wightman, F.L., "Localization using
non-individualized head-related transfer functions", Journal of the Acoustical
Society of America, vol. 94, pp. 111-123 (1993).

[8] Wightman, F.L., Kistler, D. J., and Perkins, M. E., "A new approach to the study
of human sound localization", in Directional Hearing, W. Yost and G. Gourevitch,
Eds., (Springer Verlag, New York, 1987).

[9] Møller, H., Sørensen, M. F., Jensen, C. B., and Hammershøi, D., "Binaural
technique: do we need individual recordings?", Journal of the Audio Engineering
Society, vol. 44, pp. 451-469 (1996).

Begault, D. R., et al., “Direct comparison……”

[10] Begault, D.R. and Wenzel, E. M., "Headphone Localization of Speech", Human
Factors, vol. 35, pp. 361-376 (1993).

[11] Wenzel, E.M. and Foster, S.H., "Perceptual consequences of interpolating
head-related transfer functions during spatial synthesis", in Proceedings of the
ASSP (IEEE) Workshop on Applications of Signal Processing to Audio and
Acoustics, October 17-20, New Paltz, NY, (1993).

[12] Laws, P., "Entfernungshören und das Problem der Im-Kopf-Lokalisiertheit von
Hörereignissen [Auditory distance perception and the problem of "in-head
localization" of sound images], Acustica, 29, pp. 243-259 (NASA Technical
Translation TT-20833) (1973) .

[13] Durlach, N. I., Rigopulos, A., Pang, X. D., Woods, W. S., Kulkkarni, A., Colburn,
H. S., and Wenzel, E., "On the externalization of auditory images", Presence:
Teleoperators and Virtual Environments, vol. 1, pp. (1992).

[14] Mills, W., "Auditory localization", in Foundations of Modern Auditory Theory,
J.V. Tobias, Ed. (Academic Press, New York, 1972).

[15] Wenzel, E.M., "The impact of system latency on dynamic performance in virtual
acoustic environments", in Proceedings of the 15th International Congress on
Acoustics and 135th Meeting of the Acoustical Society of America, Seattle, WA,
(1998).

[16] Adelstein, B.D., Johnston, E.R., and Ellis, S., "Dynamic response of
electromagnetic spatial displacement trackers", Presence: Teleoperators and
Virtual Environments, vol. 5, pp. 302-318 (1996).

[17] Bang and Olufsen, Music for Archimedes, compact disc B&0 101. 1992.

[18] Møller, H., Sørensen, M. F., Hammershøi, D., and Jensen, C. B., "Head-related
transfer functions of human subjects", Journal of the Audio Engineering Society,
vol. 43, pp. 300-321 (1995).

[19] Dalenbäck, B.-I., "Verification of prediction based on randomized tail-corrected
cone-tracing and array modeling", in Proceedings of the 137th meeting of the
Acoustical Society of America, 2nd Convention of the European Acoustics
Association (Forum Acusticum 99) and the 25th German DAGA Conference, Berlin
(1999).

[20] Dalenbäck, B.-I., A new model for room acoustic predicition and absorption, .
Dissertation, Chalmers University of Technology (1995).

[21] Begault, D.R. and Abel, J.S., "Studying room acoustics using a monopole-
dipole microphone array", in 16th International Congress on Acoustics and 135th
Meeting of the Acoustical Society of America, Seattle, WA, (1998).

[22] Keppel, G., Design and Analysis: a Researcher's Handbook, 3

ed. (Prentice-

Hall, Upper Saddle River, NJ, 1991).

[23] Plenge, G., "On the difference between localization and lateralization", Journal
of the Acoustical Society of America, vol. 56, pp. 944-951 (1974).

[24] Weinrich, S., "The problem of front-back localization in binaural hearing", in
Tenth Danavox Symposium on Binaural Effects in Normal and Impaired Hearing,
Denmark, (1982)..

Begault, D. R., et al., “Direct comparison……”

[25] Wightman, F.L. and Kistler, D.J., "Headphone simulation of free-field listening.
II: Psychophysical validation", Journal of the Acoustical Society of America, vol. 85,
pp. 868-878 (1989).

[26] Toole, F.E., "In-head localization of acoustic images", Journal of the Acoustical
Society of America, vol. 48, pp. 943-949 (1970).

[27] Blauert, J., Spatial hearing: The psychophysics of human sound localization
(MIT Press, Cambridge, 1983).

[28] Begault, D.R., "Perceptual similarity of measured and synthetic HRTF filtered
speech stimuli", Journal of the Acoustical Society of America, vol. 92, pp. 2334
(1992).

[29] Wenzel, E.M., "What perception implies about implementation of interactive
virtual acoustic environments", in Audio Engineering Society 101st Convention, Los
Angeles, preprint 4353 (1996).

[30] Loomis, J.M., Hebert, C., and Cincinelli, J.G., "Active localization of virtual
sounds", Journal of the Acoustical Society of America, vol. 88, pp. 757-1764 (1990).

Begault, D. R., et al., “Direct comparison……”

REVERB

TYPE

anechoic

early reflections

full auralization

HRTF

USED

individual generic

individual

generic

Individual

generic

HEAD

TRACKING

off

externalization

•

reversals

•

Azimuth

•

Elevation

•

Realism

•

FIGURE 1. Experimental conditions by block (top three rows); dependent variables

shown in 1

column in italics.

FIGURE 2. Response screen graphic. Subjects indicate azimuth and distance on

left view; elevation on right view; perceived realism on slider.

Begault, D. R., et al., “Direct comparison……”

anechoic

early reflections

full auralization

reverberation treatment

Unsigned azimuth error (degrees)

FIGURE 3.Azimuth error: significant main effect (top) and interaction (below). Mean

values and standard error bars shown.

Individual HRTF

Unsigned azimuth error (degrees)

Static

Tracked

Generic HRTF

Begault, D. R., et al., “Direct comparison……”

anechoic

early reflections

full auralization

Reverberation treatment

Unsigned elevation error (degrees)

referenced to eye level

FIGURE 4. Elevation: Main effect for reverberation. Mean values and standard error

bars shown.

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

head-tracked

static

Front-back & back-front reversed stimuli

(percent total)

FIGURE 5. Reversals: main effect for head tracking. Mean values and standard

error bars shown.

Begault, D. R., et al., “Direct comparison……”

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

anechoic

early reflections

full auralization

Reverberation treatment

Percentage externalized judgments

FIGURE 6. Externalized judgments: main effect for reverberation. Mean values and

standard error bars shown.