Content uploaded by Jacob Donley
All content in this area was uploaded by Jacob Donley on Sep 03, 2015
Content may be subject to copyright.
Multizone Reproduction of Speech Soundﬁelds:
A Perceptually Weighted Approach
Jacob Donley and Christian Ritz
School of Electrical, Computer and Telecommunications Engineering, University of Wollongong
Wollongong, NSW, Australia, 2522
E-mail: email@example.com, firstname.lastname@example.org
Abstract—In this paper a method for the reproduction of
multizone speech soundﬁelds using perceptual weighting crite-
ria is proposed. Psychoacoustic models are used to derive a
space-time-frequency weighting function to control leakage of
perceptually unimportant energy from the bright zone into the
quiet zone. This is combined with a method for regulating the
number of basis planewaves used in the reproduction to allow for
an efﬁcient implementation using a codebook of predetermined
weights based on desired soundﬁeld energy in the zones. The
approach is capable of improving the mean squared error for
reproduced speech in the bright zone by -10.5 decibels. Results
also show that the approach leads to a signiﬁcant reduction in
the spatial error within the bright zone whilst requiring 65%
less loudspeaker signal power for the case where the soundﬁeld
in this zone is in line with, and hence partially directed to, the
Spatial audio reproduction gives listeners a full experience
of the acoustic environment, including the sound source, and
has been further extended to multizone soundﬁeld repro-
duction, which provides audio in spatially separated regions
from a single set of loudspeakers, originally proposed in
. They may also be used for suppressing, or cancelling,
audio outside a targeted listening zone . The multizone
approach has many applications such as the creation of
personal sound zones in multi-participant teleconferencing,
entertainment/cinema and vehicle cabins where personal sound
zones are optimised to provide one, or many, listener(s) with
individual acoustic material .
In order to keep the sounds zones personal it is necessary
to minimise the interzone interference to maximise the indi-
vidualistic experience. Some of the earlier methods treat the
interference with hard constraints and attempt to completely
remove it , . This results in zones that are mostly
free of the interference, however, this is difﬁcult to achieve
in situations where a desired soundﬁeld in the bright zone
is obscured by or directed to another zone, as the system
requires reproduction signals many times the amplitude of
what is reproduced within any zone. This is known as the
occlusion problem , ,  and has been dealt with in
various ways such as the control of planarity , orthogonal
basis planewaves  and alleviated zone constraints , .
Requiring large signals in relation to the reproduced zones
means the system is inefﬁciently directing its energy for the
multizone reproduction, with most sound energy present in
unattended regions. This may be undesirable at times where
listeners commute between sound zones and could put unnec-
essary strain on loudspeaker drivers. More recent work has
focused on alleviating the constraint such that the interference
(or leakage) is allowed into other zones, though, the amount
can be controlled with a weighting function , . Allowing
the sound to leak into other zones can improve the practicality
of the system but decrease the individuality of zones.
Existing methods focus on single frequency soundﬁelds,
although there has been work attempting to create multizone
soundﬁelds for wideband speech . More recently, work has
been done  to extend a method  to the reproduction
of weighted wideband speech soundﬁelds whilst efﬁciently
maintaining the weighting function in the spatial, time and
frequency domain. This allows for dynamic weighting of the
zones as well as individual frequency components in time thus
allowing each zone’s acoustic content to be controlled.
The control of acoustic components to enhance the percep-
tion of a signal has been researched thoroughly for applications
such as compression . The relationship between the quality
in the bright zone and interference in other zones has been
subjectively tested , however, the occlusion problem is not
directly addressed and the planarity control does not consider
human perception. Hence, perceptual models are employed in
this work in order to enhance the experience in personal sound
zones, especially where the occlusion problem is present.
Leaked sound energy is treated as unwanted noise in other
zones and controlled such that it is perceptually less noticeable
as indicated by established psychoacoustic models.
We begin with an explanation of the weighted multizone
soundﬁeld method used in this work in Section II. Psychoa-
coustic models are introduced in Section III as well as the
need for regulating the number of weighted basis planewaves.
Results of the perceptual weighting and conclusions are given
in Section IV and Section V, respectively.
II. WEIGHTED MU LTIZ ON E WIDEBAND SOU NDFIELDS
The multizone soundﬁeld reproduction layout considered in
this work is shown in Fig. 1 and contains a reproduction
region, D, which has a radius R. The reproduction region
consists of three zones called the bright, quiet and unattended
zones which are represented by Db,Dqand D∩(Db∪Dq)0,
respectively. The centres of Dband Dqhave a distance of rz
from the centre of Dand each of these zones has a radius of
© 2015 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or
future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for
resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
Fig. 1: A weighted multizone soundﬁeld reproduction layout is shown. The
ﬁgure depicts a situation where the desired soundﬁeld in the bright zone is
partially directed towards the quiet zone causing the occlusion problem.
r. Loudspeakers are positioned with a distance of Rlfrom the
centre of Don an arc of angle φLwhich starts at angle φand
reproduces planewave speech soundﬁelds in Dbwith an angle
In the method of weighting multizone soundﬁelds , a
spatial weighting ﬁlter as a function of space, w(x), is used
to control the reproduction of sound within each of the
zones. Subsequent work  extended this approach to al-
low for space-time-frequency dependent weighting functions,
w(x, n, k), which allows for weighting functions to be adapted
based on the signal characteristics of the target soundﬁeld. We
denote wb,wqand wuas the weighting functions for xb∈Db,
xq∈Dqand xu∈D∩(Db∪Dq)0, respectively. The reproduced
soundﬁeld pressure at any point in the reproduction region
is deﬁned as the sum of space-time-frequency dependant
weighted soundﬁeld values ,
ˆpw(x, n) =
w(x, n, k)(1)
w(x, n, k)=fSd(x, n, k), w(x, n, k )is a repro-
duced soundﬁeld, which is derived as a function of a desired
soundﬁeld, Sd(x, n, k), and a weighting function, w(x, n, k)
using the approaches outlined in , . Here, xis a given
position, nis a given time and kis a given frequency.
w(x, n, k)is summed for Kdifferent sinusoidal components.
In this work k= 2πf /c and c= 343 m s−1.
In (1) w(x, n, k)allows independently weighting soundﬁeld
components in space and time. It is then possible to deﬁne
the reproduced space-time-frequency domain signal for a
particular input as ,
Yw(x, n, k) =
w(x, n, k)
Yw(x, n, k)is the time-frequency signal at an arbitrary
location, x, in the reproduction region, D,Y(n, k)is obtained
from the short-time Fourier transform of the windowed frame
of input y(n)and |·| denotes the absolute value. Using overlap-
add reconstruction we can obtain the time-domain signal at
any point in Dwhere a different weighting function can be
used for each space-time-frequency. The weighting function
can now be used to control the leaked content into the quiet
zone in the space-time-frequency domain.
III. PSYCHOACOUSTIC WEIGHTING MODELS
The capability of controlling the energy leakage between
zones then allows the weighting function to become dependent
on the signal being reproduced. For instance, the leaked audio
spectrum may be controlled, altered, suppressed or designed
to be masked by another spectrum. From this, psychoacoustic
modelling can be applied to the weighting function in order to
reduce the perceptual affect of the leakage in the quiet zone.
A. The Hearing Threshold
The beneﬁt of using zone weighting is that the hard con-
straint of zero energy is alleviated and sound energy may be
allowed to leak into the quiet zone. However, this then means
the quiet zone is no longer completely quiet.
Due to the human threshold of hearing in quiet, a quiet
zone could be redeﬁned so that the sound pressure level is
imperceptible. This would then make a weighted multizone
system practical (from a relieved constraint) and remain quiet
(perceptually). The threshold in quiet has been well established
with frequency dependent functions that provide a good ap-
proximation , .
Using the new space-time-frequency domain weighting it
is possible to apply the threshold in quiet approximation to
(2) where w(x, n, k)is chosen so that the output in the quiet
Yw(xq, n, k), is as close to the threshold in quiet as
possible. Then, using the codebook method , w(x, n, k)
can be chosen to minimise the difference,
Yw(xq, n, k)−A(xq, n, k)(3)
where A(xq, n, k)is a space-time-frequency dependent func-
tion describing the perceptual criteria. In this work Sound
Pressure Level (SPL) in dB is relative to the threshold of
hearing pr= 20 µPa.
B. Spreading Functions to Reduce Multizone Error
Analysing the weighted multizone reproductions in 
reveals that larger weighting increases the error in the bright
zone whilst suppressing the quiet zone. The quality of the
spatial reproduction in the bright zone is less erroneous when
the weighting is eased for the quiet zone allowing more energy
to leak. If the quiet zone is controlled to have minimal energy
leaked into it there becomes an erroneous bright zone.
The spatial errors shown in Fig. 2 are calculated from :
b(n, k) = RDb
Sd(x, n, k)−Sa
w(x, n, k)
RDb|Sd(x, n, k)|2dx
where b(n, k)is the spatial error in the bright zone.
60 Bright Zone SPL
Threshold in Quiet
Bright Zone Spatial Error
60 Quiet Zone SPL
Fig. 2: Multizone soundﬁeld reproduction with perceptual weighting in the quiet zone. The desired bright zone signal is an equal loudness curve at 30 phon
 and a 2kHz masker signal at 30 dB SPL is present in the quiet zone. The red and green dashed lines show the worst and best case scenarios,
respectively. The bright zone error is calculated using (4). The “Leaked SPL” shows the result after controlling the interzone interference with wq.
0Quiet Zone SPL Limits in Codebook
N = 80
Fig. 3: Shows the maximum (solid line) and minimum (dash-dot line) levels
in a codebook that the weighting function provides for wq= 10−2→104.
The number of basis planewaves, N, used to generate the codebooks are a
constant number, 80, (red lines) and a regulated number (blue lines).
The work in  shows that for k= 2 kHz the spatial error
is greater than −5dB when the quiet zone is occluded by the
bright zone and has a large weight (equivalent to wq= 10),
however, the spatial error is less than −20 dB when the weight
is alleviated (equivalent to wq= 0.1).
In Fig. 2 it is shown that using a spreading function to
mask apparent sounds, in the “target” quiet zone, can reduce
the error of the reproduction in the bright zone. This is because
we can safely allow the sound energy at particular frequencies
to leak into the quiet zone with no perceptual affect. If
the “target” quiet zone contained many different frequency
components then it is possible that the bright zone energy
could be completely leaked into the quiet zone unperceivably
and thus reduce the error in the bright zone to a minimum.
C. Maintaining Broad Control for Wideband Speech
Inaccurate reproductions can be caused by spatial aliasing
and poorly conditioned matrices which cause the control range
of the weighting function per frequency to become reduced
and less accurate as can be seen in Fig. 3. In order to maintain
accurate estimation of the affect of different zone weights
the number of basis planewaves needed to reproduce the
soundﬁelds can be regulated .
Fig. 3 shows that with a regulated Nthe difference in level
that a given wqcan provide in the quiet zone is improved for
low and high frequencies and the response is smoother for low
frequencies. N= 80 is well balanced between spatial aliasing
and ill-conditioning for 2kHz , .
IV. RES ULT S
A. Multizone Reproduction Evaluation Setup
The multizone soundﬁeld layout of Fig. 1 is evaluated,
where r= 0.3m, rz= 0.6m, R= 1 m, Rl= 1.5m,
θ= sin−1(r/2rz)≈14.5° and π≈3.141 59 rad. The value
of θis chosen such that an evanescent planewave with instant
decay would interfere with half the quiet zone , .
This choice results in a slight occlusion problem where the
range of weighting control is larger than for no occlusion
and full occlusion. Signals sampled at 16 kHz are converted to
the time-frequency domain using a Hamming window (50%
overlap) and Fast Fourier transform (FFT) of length 1024.
For the evaluation we use the efﬁcient method of codebooks
described in  to store pre-determined weighted soundﬁeld
values to be used for a given setup or wideband reproduction.
The codebooks are constructed for a reproduction where
L= 65 and φL= 2πwhich for this particular setup is
free of signiﬁcant aliasing problems in the quiet zone below
The codebooks are built with spatial pressure samples for
all x∈Db∩Dqwith each soundﬁeld zone approximated from
2724 samples. The zone weights are chosen as wb= 1 and
wu= 0.05 following ,  and the variable weight is wq.
Then, using (3), wqis chosen to match the quiet zone to a
given level, A(xq, n, k). In this work we choose A(xq, n, k)
to be the threshold in quiet using the ISO226 standard 
with additional masking curves using the ISO/IEC MPEG
Psychoacoustic Model 2 spreading function .
Speech ﬁles for the evaluation are taken from the TIMIT
corpus  where 20 ﬁles are chosen randomly. The male to
female speaker ratio of these ﬁles is 50 : 50. Evaluations using
these speech ﬁles are shown with 95% conﬁdence intervals.
B. Reduced Bright Zone Error from Psychoacoustic Masking
The error induced from the multizone reproduction of the
speech soundﬁelds is evaluated using the Mean Squared Error
(MSE) of the reproduced speech where the reference signal
for the MSE is the original speech signal. To obtain an
approximation of the reproduced speech the mean of the
-65 MSE of Weighted Reproduced Speech
Fig. 4: Shows the MSE of reproduced speech ﬁles in the bright zone for
different uniform weighting functions (wq).
Fig. 5: Difference between Sd(xb, n, k)and Sa
w(xb, n, k)for f= 2 kHz.
Aand Bshow the magnitude difference and Cand Dshow the phase
difference. Aand Care for wq= 10−2and Band Dare for wq= 104.
simulated spatial pressure samples obtained with the approach
of section III are used across Dband Dq.
Upon analysing the MSE of different reproduced speech
ﬁles it becomes apparent that the majority of error measured
in the bright zone from the reproduction is in the spatial
domain. The sampling theory used to obtain the reproduced
speech means that spatial information is neglected, however,
(4) evaluates the spatial error and is similar to the measure
of planarity . This then means that the application of
perceptual criteria primarily reduces the spatial error of the
The maximum improvement in MSE of the bright zone
reproduced speech is −10.5dB, from −69.8dB for wq= 104
to −80.3dB for wq= 10−2, and can be seen in Fig. 4.
Even though there is a difference of −10.5dB, the MSE in
the reproduced speech is minimal. However, the maximum
improvement in spatial error for the bright zone, b, averaged
for all frequencies is −24.0dB, from −7.4dB for wq= 104
to −31.5dB for wq= 10−2, and can be seen in Fig. 2.
Also shown in Fig. 2 is that a 2kHz masker signal in
the quite zone can allow the spatial error in the bright zone
to be reduced. This reduction in spatial error is depicted
in Fig. 5 where the perceptual weighting uses wq= 10−2
instead of wq= 104which gives a smaller difference between
the desired soundﬁeld and reproduced soundﬁeld. In Fig.
5 the magnitude difference is calculated from |Sd|−|Sa
and the phase difference from arg(Sd/Sa
w). The equivalent
improvement in band required loudspeaker power due to the
perceptual weighting is −28 dB and 65% less, respectively.
In this paper we have proposed a method for perceptually
weighting multizone speech soundﬁelds which can improve
error in bright zones, especially when the occlusion problem is
apparent. We have shown the need for regulating the number of
basis planewaves used for the reproduction. Perceptual weight-
ing is shown to improve the MSE for reproduced speech in
the bright zone from −69.8dB to −80.3dB and signiﬁcantly
reduce the spatial error on average from −7.4dB to −31.5dB
whilst requiring less power. Future work includes testing
methods for maximising the speech intelligibility difference
and privacy between zones in multizone speech soundﬁelds.
 M. Poletti, “An Investigation of 2-D Multizone Surround Sound Sys-
tems,” in Audio Engineering Society Convention 125. Audio Engineer-
ing Society, Oct. 2008.
 W. Jin and W. Kleijn, “Multizone soundﬁeld reproduction in rever-
berant rooms using compressed sensing techniques,” in 2014 IEEE
International Conference on Acoustics, Speech and Signal Processing
(ICASSP), May 2014, pp. 4728–4732.
 T. Betlehem, W. Zhang, M. Poletti, T. D. Abhayapala, and others,
“Personal Sound Zones: Delivering interface-free audio to multiple
listeners,” Signal Processing Magazine, IEEE, vol. 32, no. 2, pp. 81–91,
 Y. J. Wu and T. D. Abhayapala, “Spatial multizone soundﬁeld repro-
duction: Theory and design,” IEEE Transactions on Audio, Speech, and
Language Processing, vol. 19, no. 6, pp. 1711–1720, 2011.
 T. Betlehem and P. D. Teal, “A constrained optimization approach for
multi-zone surround sound,” in 2011 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2011, pp.
 P. Coleman, P. J. Jackson, M. Olik, and J. A. Pedersen, “Personal audio
with a planar bright zone,” vol. 136, no. 4, pp. 1725–1735.
 W. Jin, W. B. Kleijn, and D. Virette, “Multizone soundﬁeld reproduction
using orthogonal basis expansion,” in International Conference on
Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2013, pp.
 H. Chen, T. D. Abhayapala, and W. Zhang, “Enhanced sound ﬁeld
reproduction within prioritized control region,” in INTER-NOISE and
NOISE-CON Congress and Conference Proceedings, vol. 249. Institute
of Noise Control Engineering, 2014, pp. 4055–4064.
 N. Radmanesh and I. S. Burnett, “Generation of isolated wideband
sound ﬁelds using a combined two-stage lasso-ls algorithm,” IEEE
Transactions on Audio, Speech, and Language Processing, vol. 21, no. 2,
pp. 378–387, 2013.
 J. Donley and C. Ritz, “An efﬁcient approach to dynamically weighted
multizone wideband reproduction of speech soundﬁelds,” in China Sum-
mit & International Conference on Signal and Information Processing
(ChinaSIP). IEEE, 2015, pp. 60–64.
 M. Bosi and R. E. Goldberg, Introduction to digital audio coding and
standards. Springer, 2003.
 K. Baykaner, P. Coleman, R. Mason, P. J. Jackson, J. Francombe,
M. Olik, and S. Bech, “The relationship between target quality and
interference in sound zone,” vol. 63, no. 1, pp. 78–89.
 B. ISO, “226: 2003:AcousticsNormal equalloudness-level contours,”
International Organization for Standardization, 2003.
 J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, and D. S. Pallett,
“DARPA TIMIT acoustic-phonetic continous speech corpus CD-ROM.
NIST speech disc 1-1.1,” vol. 93, p. 27403.