Christoph Pörschmann, Johannes M. Arend
Analysis and visualization of dynamic human voice
directivity
Open Access via institutional repository of Technische Universität Berlin
Document type
Conference paper | Published version
(i. e. publisher-created published version, that has been (peer-) reviewed and copyedited; also known as:
Version of Record (VOR), Final Published Version)
This version is available at
https://doi.org/10.14279/depositonce-15551
Citation details
Pörschmann, C. & Arend, J. M. (2022): Analysis and visualization of dynamic human voice directivity. In:
AIA-DAGA 2022 : proceedings of the International Conference on Acoustics. Berlin: Deutsche Gesellschaft für
Akustik e.V. pp. 1444-1447.
Terms of use
This work is protected by copyright and/or related rights. You are free to use this work in any way permitted by
the copyright and related rights legislation that applies to your usage. For other uses, you must obtain
permission from the rights-holder(s).
Analysis and Visualization of Dynamic Human Voice Directivity
Christoph P¨orschmann1, Johannes M. Arend1,2
1TH K¨oln, Institute of Communications Engineering, Cologne, Germany
2TU Berlin, Audio Communication Group, Berlin, Germany
Email: christoph.poerschmann@th-koeln.de
Introduction
In many everyday situations, we experience the influence
of the human voice directivity. We perceive loudness and
timbre differently when a speaker faces us or turns away
from us. Often, we use voice directivity intuitively, for ex-
ample when facing a person in a meeting or a casual con-
versation. Such effects of human voice directivity have
long been a topic of research. Early studies were carried
out more than 200 years ago analyzing the directional
radiation of speech in general [1, 2, 3]. In 1929 a first ap-
proach determining directivity patterns for several vowels
and fricatives in the horizontal plane has been presented
by Trendelenburg [4]. In 1939 Dunn and Farnsworth
[5] determined spherical directivity patterns for a spoken
sentence at different distances in third-octave bands from
63 Hz up to 12 kHz. Since then, voice directivity has been
subject to many studies, either applying human speakers
or dummy heads with integrated mouth simulators.
A specific characteristic of the human voice that cannot
be analyzed with a dummy head is its dynamic directiv-
ity, i.e., time-variant changes that occur when speaking
or singing. To adequately determine these features, the
sound radiation needs to be captured for an appropri-
ately large number of directions. Katz and D’Alessandro
[6] analyzed the voice directivity in the horizontal plane
for sustained vowels articulated by a professional opera
singer. Even though the study showed no systematic
differences between the different vowels, it gave first in-
sights into the vowel-dependencies. Kocon and Mon-
son [7] examined articulation-dependent effects of voice
directivity for different vowels in fluent speech. They
observed an effect of the vowel on the directivity pat-
tern and determined the strongest directivity for an [a].
Monson et al. [8] analyzed phoneme-dependencies in the
horizontal plane for measurements of the directivity in
angular steps of 15◦. They showed that the directivity
varies strongly for different articulations, e.g., between
voiceless fricatives with a more directive sound radiation
of an [s] than of an [f]. However, this study did not
show significant differences between speech and singing,
and only minor dependencies on the articulation level.
In contrast, Chu and Warnock [9] observed significant
differences depending on the articulation level. Postma
and Katz [10] as well as Postma et al. [11] analyzed the
influence of dynamic voice directivity for auralizations.
Their results indicate that auralizations involving dy-
namic voice directivity are perceived as more plausible
and exhibit a wider apparent source width than aural-
izations with static voice directivity or omnidirectional
sources. In contrast, a recent study by Ehret et al. [12]
concluded that dynamic directivity is perceptually indis-
tinguishable from a static voice directivity.
In general, voice directivity measurements can either be
performed sequentially for an arbitrary number of direc-
tions or simultaneously using a surrounding microphone
array. In the case of sequential measurements, the spec-
trum of the radiated sound is typically analyzed averaged
over time. Thus, time-variances influencing the speaker’s
directivity can hardly be resolved and are in most evalua-
tions not considered. For determining the spherical direc-
tivity patterns, it is advantageous to apply surrounding
microphone arrays (SMAs) [13, 14, 15, 16]. Generally,
the setup is restricted to a limited number of sampling
points. Thus, the spatial resolution of the directivity sets
is sparse. Consequently, methods are required for spatial
upsampling of the (sparsely) measured datasets by ap-
propriate interpolation between the measured directions.
Up to now, only a limited number of studies on spherical
voice directivity have been published (e.g., [17, 9, 15]),
but none of them analyzed the time-variant effects of
fluent speech. This pilot study aims to investigate to
what extent high-density directivity sets can be deter-
mined from fluent speech measured on a sparse sam-
pling grid with an SMA. In this context, we presented
the SUpDEq (Spatial Upsampling by Directional Equal-
ization) method [18], which originally was designed for
spatial upsampling of head-related transfer functions. In
P¨orschmann and Arend [19, 20], we evaluated the method
for a dummy head with mouth simulator and showed that
reasonable dense directivity sets can be obtained from
sparse measurements. Our studies revealed that mea-
surements in a surrounding microphone array with 32 mi-
crophones are sufficient to generate a decent full-spherical
dense directivity set, with an error averaged over the en-
tire sphere below 4 dB for frequencies up to 8 kHz. In
[21, 22], we applied the SUpDEq method to human voice
directivity and determined full-spherical directivity pat-
terns of five vowels and three fricatives for 13 persons
using a surrounding spherical microphone array with 32
microphones. The results showed significant differences
between the two groups of phonemes, and between some
of the phonemes of each group. Furthermore, in this con-
text we studied the influence of hand postures on voice
directivity and showed that cupping the hands around
the mouth or holding a hand in front of the mouth have
stronger effects than differences between the phonemes
[23]. In this paper, we present a pilot study that investi-
gates dynamic voice directivity of fluent speech and thus
resolves in which way voice directivity changes over time.
DAGA 2022 Stuttgart
1444
Figure 1: Human speaker inside the SMA during the
measurements.
Method and Materials
Measurements
All measurements of this pilot study were performed in
the anechoic chamber of TH K¨oln, sized 4.5 m ×11.7 m
×2.3 m (W×D×H) and with a lower cut-offfrequency
of about 200 Hz. The directivity patterns were measured
with an SMA having a basic shape of a pentakis dodeca-
hedron with 32 Rode NT5 cardioid microphones located
at the vertices of this shape on a constant radius of 1 m.
This sampling scheme allows resolving the directivity up
to a spatial order of N= 4 [13]. In the present study,
an additional Rode NT5 microphone was positioned at
the front serving as a reference and for spectral equal-
ization in postprocessing. Four RME Octamic II devices
served as preamplifiers and AD / DA converters for the
32 microphones of the SMA. All signals of the SMA were
processed with two RME Fireface UFX audio interfaces.
For a more detailed description of the SMA setup, please
refer to Arend et al. [14, 16]. In the present study, one
of these audio interfaces was also used as a preamplifier
and AD / DA converter for the reference microphone.
As test material, we recorded excerpts from Antoine de
Saint-Exupery’s book ”The little prince” [24], spoken by
two German speakers in English and German. To ob-
tain a visual representation of the time-variant changes
of the mouth, we filmed the speakers using a Sony PXW-
FS7 camera at a frame rate of 180 fps. The video can be
superimposed with the determined directivity patterns
or used to create slow-motion sequences of mouth move-
ments. Fig. 1 shows the SMA with a person inside.
Postprocessing
The postprocessing is only briefly summarized here, as
it was carried out similarly as presented in P¨orschmann
and Arend [20, 22]. To eliminate the influence of re-
flections and of room modes of the anechoic chamber,
which in our case become prominent below 200 Hz, a low-
frequency extension was applied substituting the origi-
nal low-frequency component in the frequency domain by
an adequately matched one of an analytic low-frequency
model. Furthermore, in the postprocessing, the inaccura-
cies in positioning the subjects in the center of the micro-
01234567
Time [s]
0
2
4
6
8
10
Directivity Index [dB]
Once... old, magni�cent...
I...
[ʌ][əʊ][aɪ] [æ]
1 kHz 2 kHz 4 kHz 8 kHz
Figure 2: Directivity index (DI) over time for one sen-
tence for octave bands of 1 kHz, 2 kHz, 4 kHz, and 8 kHz.
Four selected phonemes are marked on the time axis.
phone array were compensated. As small deviations of
some centimeters already lead to strong impairments in
the spatial upsampling process [25], we applied a method
for distance error compensation that reduces the impair-
ments of distance errors of the measured directivity pat-
terns [20]. Then the compensated multichannel record-
ings were partitioned into frames of 67 ms (3200 samples
at 48 kHz sampling rate) and Hann-windowed (50 % over-
lap). Further processing was done separately for each au-
dio frame. The sparse datasets were spatially upsampled
to a dense grid with 2702 sampling points on a Lebedev
sampling scheme applying the SUpDEq method1, which
we described in detail in P¨orschmann et al. [18] and evalu-
ated for upsampling voice directivity in P¨orschmann and
Arend [20]. However, in contrast to the processing used
in the studies mentioned above, we applied the postpro-
cessing and the upsampling not to transfer functions but
directly to the multichannel audio frames. As output of
the processing chain, a high-density dataset was stored
for each audio frame, which can be used for the further
analysis of dynamic voice directivity.
Results
The results presented here are based on one selected sen-
tence of the recordings, which was the first sentence of
[24]: ”Once when I was six years old I saw a magnifi-
cent picture in a book, called True Stories from Nature,
about the primeval forest.” In a first step, we calculated
the directivity index (DI) for each of the frames:
DI(f, n) = 10 lg 4π|p(φ0,θ0, f, n)|2
2π
�
0
π/2
�
−π/2
|p(φ,θ, f, n)|2cosθdθdφ
,(1)
with φthe azimuth, θthe elevation, φ0,θ0the frontal
direction, fthe frequency, and nthe index of the frame.
1A Matlab-based implementation of the SUpDEq method can
be accessed on https://github.com/AudioGroupCologne/SUpDEq
DAGA 2022 Stuttgart
1445
0°
30°
60°
90°
120°
150°
180°
210°
240°
270°
300°
330°
-18
-12
-6
0 dB
[
�
]
0°
30°
60°
90°
120°
150°
180°
210°
240°
270°
300°
330°
-18
-12
-6
0 dB
[
��
]
0°
30°
60°
90°
120°
150°
180°
210°
240°
270°
300°
330°
-18
-12
-6
0 dB
[a
�
]
0°
30°
60°
90°
120°
150°
180°
210°
240°
270°
300°
330°
-18
-12
-6
0 dB
[æ]
1 kHz 2 kHz 4 kHz 8 kHz
Figure 3: Directivity in the horizontal plane for an [
�
], an [
��
], an [a
�
], and an [æ] in octave bands of 1 kHz, 2 kHz,
4 kHz, and 8 kHz normalized to a maximal value of 0 dB.
90°
60°
30°
Front
-30°
-60°
-90°
-60°
-30°
Back
30°
60°
-18
-12
-6
0 dB
[
�
]
90°
60°
30°
Front
-30°
-60°
-90°
-60°
-30°
Back
30°
60°
-18
-12
-6
0 dB
[
��
]
90°
60°
30°
Front
-30°
-60°
-90°
-60°
-30°
Back
30°
60°
-18
-12
-6
0 dB
[a
�
]
90°
60°
30°
Front
-30°
-60°
-90°
-60°
-30°
Back
30°
60°
-18
-12
-6
0 dB
[æ]
1 kHz 2 kHz 4 kHz 8 kHz
Figure 4: Directivity in the vertical plane for an [
�
], an [
��
], an [a
�
], and an [æ] in octave bands of 1 kHz, 2 kHz, 4 kHz,
and 8 kHz normalized to a maximal value of 0 dB.
Fig. 2 shows the DI over time in octave bands. Since no
stable values were obtained for speech pauses and sec-
tions with low energy, we refrained from plotting frames
with an energy of 14 dB or more below the average energy
of the entire sentence. Generally, the DI increases with
frequency, which is in line with other studies [15, 22] and
can easily be explained by frequency-dependent diffrac-
tion effects of the head. Furthermore, we found varia-
tions in DI over time caused by phoneme-dependencies.
These variations occurred in a similar way as in our pre-
vious studies [21, 22], in which we analyzed directivity
patterns of a variety of separately articulated phonemes.
For our test sentence, we observed the highest DI for an
[
�
] at the beginning of the sentence.
Then we determined directivity patterns for selected
phonemes within the sentence. Fig. 3 and 4 show the
directivity patterns in the horizontal and vertical plane,
for four different time frames in which an [
�
], an [
��
], an
[a
�
], and an [æ] were spoken. The respective time frames
are also marked in Fig. 2. It can be observed that sig-
nificant differences between the phonemes occur in the
horizontal plane. While for an [
�
] and an [a
�
], the direc-
tivity patterns in the 4 kHz and 8 kHz octave band are
very similar, they are stronger directed to the frontal di-
rection for an [
��
] at 8 kHz. Furthermore, the plots show
slight asymmetries, e.g., for the [
��
] or the [æ]. In the
vertical plane, the frequency-dependent differences seem
to be smaller than in the horizontal plane, at least in
the frontal hemisphere. In the vertical plane, both, the
variations between the phonemes as well as between the
different frequency bands, tend to be reduced. However,
in future work, these findings need to be examined in
more detail based on larger parts of the test material.
Conclusion
In this pilot study, we analyzed dynamic voice directiv-
ity of fluent speech. We presented and demonstrated a
method to determine high-density directivity sets from
sparse measurements carried out with an SMA with 32
microphones. The results reveal how the dynamic direc-
tivity changes in fluent speech over time and how this
affects the DI. Furthermore, superimposing the directiv-
ity plots with a filmed sequence of the subject allows
studying the dynamics of the lip movements as well as of
the form and the size of the mouth opening. The results
presented in this pilot study are only based on one single
sentence. Next, we plan to systematically segment and
evaluate larger recordings. This will allow for a more de-
tailed analysis of variations and fluctuations of the direc-
tivity. For this, processing tools for speech segmentation
need to be applied and appropriately adapted.
The results of this study are of high relevance for different
purposes. First, it is of general interest when studying
voice production to analyze the dynamic voice directivity
DAGA 2022 Stuttgart
1446
in detail for fluent speech. Second, methods and datasets
are required for applications in the field of virtual reality,
augmented reality, or room acoustic simulation, to inte-
grate adequate voice radiation patterns into the process
of sound-field calculation. The question of whether and
to what extent time-variant aspects have to be consid-
ered has a strong influence on the system design and the
required methods for obtaining the directivity datasets.
Finally, when reproducing one’s own voice in a virtual
acoustic environment [26, 27, 28, 29], its dynamic direc-
tivity could be perceptible for the speaker himself.
Acknowledgements
We thank Raphael Gillioz and Kai Altwicker for their
work on the measurements and the segmentation of the
raw recordings. The research has been carried out in the
research project NarDasS, funded by the Federal Min-
istry of Education and Research in Germany, support
code: BMBF 03FH014IX5-NarDasS.
References
[1] Saunders, G., Treatise on Theaters, I. and J. Taylor, London,
1790.
[2] Wyatt, B., Observation on the Design for the Theatre Royal,
Drury Lane, J. Taylor, London, 1813.
[3] Henry, J., “Annual Report of the Board of Regents of the
Smithsonian Institution,” Technical report, A. G. F. Nichol-
son, Washington, DC, 1857.
[4] Trendelenburg, F., “Beitrag zur Frage der Stimm-
richtwirkung,” Zeitschrift f¨ur techn. Physik, 11, pp. 558–563,
1929.
[5] Dunn, H. K. and Farnsworth, D. W., “Exploration of pressure
field around the human head during speech,” The Journal of
the Acoustical Society of America, 10, pp. 184–199, 1939, doi:
https://doi.org/10.1121/1.1915975.
[6] Katz, B. and D’Alessandro, C., “Directivity measurements of
the singing voice,” in Proceedings of the 19th International
Congress on Acoustics, 2007.
[7] Kocon, P. and Monson, B. B., “Horizontal directivity pat-
terns differ between vowels extracted from running speech,”
The Journal of the Acoustical Society of America, 144(1), pp.
EL7–EL12, 2018, doi:10.1121/1.5044508.
[8] Monson, B. B., Hunter, E. J., and Story, B. H., “Horizontal
directivity of low- and high-frequency energy in speech and
singing,” The Journal of the Acoustical Society of America,
132(1), pp. 433–441, 2012, doi:10.1121/1.4725963.
[9] Chu, W. T. and Warnock, A. C. C., “Detailed Directivity of
Sound Fields Around Human Talkers,” Technical report, 2002,
doi:10.4224/20378930.
[10] Postma, B. N. J. and Katz, B. F., “Dynamic voice directivity
in room acoustic auralizations,” in Proceedings of the 42th
DAGA, pp. 352–355, 2016.
[11] Postma, B. N. J., Demontis, H., and Katz, B. F. G., “Sub-
jective Evaluation of Dynamic Voice Directivity for Auraliza-
tions,” Acta Acustica united with Acustica, 103(2), pp. 181–
184, 2017.
[12] Ehret, J., Stienen, J., Brozdowski, C., B¨onsch, A., Mittelberg,
I., Vorl¨ander, M., and Kuhlen, T. W., “Evaluating the Influ-
ence of Phoneme-Dependent Dynamic Speaker Directivity of
Embodied Conversational Agents’ Speech,” in Proceedings of
the 20th ACM International Conference on Intelligent Vir-
tual Agents, pp. 1–8, ACM, New York, NY, USA, 2020, doi:
10.1145/3383652.3423863.
[13] Pollow, M., Directivity Patterns for Room Acoustical Mea-
surements and Simulations, Logos Verlag Berlin, 2015.
[14] Arend, J. M., Stade, P., and P¨orschmann, C., “Binaural re-
production of self-generated sound in virtual acoustic envi-
ronments,” in Proceedings of the 173rd Meeting of the Acous-
tical Society of America, volume 30, pp. 1–13, 2017, doi:
10.1121/2.0000574.
[15] Brandner, M., Frank, M., and Rudrich, D., “DirPat -
Database and Viewer of 2D/3D Directivity Patterns of Sound
Sources and Receivers,” in Proceedings of the 144th AES Con-
vention, e-Brief 425, 1, pp. 1–5, 2018.
[16] Arend, J. M., L¨ubeck, T., and P¨orschmann, C., “A Reactive
Virtual Acoustic Environment for Interactive Immersive Au-
dio,” in Proceedings of the AES Conference on Immersive and
Interactive Audio, 2019.
[17] Kob, M. and Jers, H., “Directivity measurement of a singer,”
The Journal of the Acoustical Society of America, 105, p.
1003, 1999, doi:10.1121/1.425813.
[18] P¨orschmann, C., Arend, J. M., and Brinkmann, F., “Direc-
tional Equalization of Sparse Head-Related Transfer Function
Sets for Spatial Upsampling,” IEEE/ACM Transactions on
Audio, Speech, and Language Processing, 27(6), pp. 1060 –
1071, 2019, doi:10.1109/TASLP.2019.2908057.
[19] P¨orschmann, C. and Arend, J. M., “A Method for Spatial
Upsampling of Directivity Patterns of Human Speakers by Di-
rectional Equalization,” in Proceedings of the 45th DAGA, pp.
1458 – 1461, 2019.
[20] P¨orschmann, C. and Arend, J. M., “A Method for Spatial
Upsampling of Voice Directivity by Directional Equalization,”
Journal of the Audio Engineering Society, 68(9), pp. 649–663,
2020, doi:10.17743/jaes.2020.0033.
[21] P¨orschmann, C. and Arend, J. M., “Analyzing the Directiv-
ity Patterns of Human Speakers,” in Proceedings of the 46th
DAGA, pp. 1141 – 1144, 2020.
[22] P¨orschmann, C. and Arend, J. M., “Investigating phoneme-
dependencies of spherical voice directivity patterns,” The
Journal of the Acoustical Society of America, 149(6), pp. 4553
– 4564, 2021, doi:10.1121/10.0005401.
[23] P¨orschmann, C. and Arend, J. M., “Effects of hand postures
on voice directivity,” JASA Express Letters, 2(3), p. 035203,
2022, doi:10.1121/10.0009748.
[24] de Saint-Exupery, A., The Little Prince, Reynal & Hitchcock,
1943.
[25] P¨orschmann, C. and Arend, J. M., “How positioning inaccu-
racies influence the spatial upsampling of sparse head-related
transfer function sets,” in Proceedings of the International
Conference on Spatial Audio - ICSA 2019, pp. 1–8, 2019.
[26] P¨orschmann, C., “One’s own voice in auditory virtual environ-
ments,” Acta Acustica united with Acustica, 87(3), pp. 378–
388, 2001.
[27] P¨orschmann, C. and Pellegrini, R. S., “3-D Audio in Mobile
Communication Devices: Effects of Self-Created and Exter-
nal Sounds on Presence in Auditory Virtual Environments,”
JVRB - Journal of Virtual Reality and Broadcasting, 7(11),
pp. 3–11, 2010.
[28] Neidhardt, A., “Detection of a nearby wall in a virtual echolo-
cation scenario based on measured and simulated OBRIRs,”
in Proceedings of the AES Conference on Spatial Reproduc-
tion, 2018.
[29] Frank, M. and Brandner, M., “Perceptual Evaluation of
Spatial Resolution in Directivity Patterns 2: coincident
source/listener positions,” in Proceedings of the International
Conference on Spatial Audio - ICSA 2019, September, pp.
1–5, 2019.
DAGA 2022 Stuttgart
1447
