Improving Speech Privacy in Personal Sound Zones

Abstract

This paper proposes two methods for providing speech privacy between spatial zones in anechoic and reverberant environments. The methods are based on masking the content leaked between regions. The masking is optimised to maximise the speech intelligibility contrast (SIC) between the zones. The first method uses a uniform masker signal that is combined with desired multizone loudspeaker signals and requires acoustic contrast between zones. The second method computes a space-time domain masker signal in parallel with the loudspeaker signals so that the combination of the two emphasises the spectral masking in the targeted quiet zone. Simulations show that it is possible to achieve a significant SIC in anechoic environments whilst maintaining speech quality in the bright zone.

Full text

IMPROVING SPEECH PRIVACY IN PERSONAL SOUND ZONES
Jacob Donley⋆, Christian Ritz⋆and W. Bastiaan Kleijn†
⋆School of Electrical, Computer and Telecommunications Engineering, University of Wollongong, Australia
†School of Engineering and Computer Science, Victoria University of Wellington, New Zealand
ABSTRACT
This paper proposes two methods for providing speech privacy be-
tween spatial zones in anechoic and reverberant environments. The
methods are based on masking the content leaked between regions.
The masking is optimised to maximise the speech intelligibility con-
trast (SIC) between the zones. The first method uses a uniform
masker signal that is combined with desired multizone loudspeaker
signals and requires acoustic contrast between zones. The second
method computes a space-time domain masker signal in parallel with
the loudspeaker signals so that the combination of the two empha-
sises the spectral masking in the targeted quiet zone. Simulations
show that it is possible to achieve a significant SIC in anechoic envi-
ronments whilst maintaining speech quality in the bright zone.
Index Terms—multizone soundfield reproduction, personal
sound zones, speech privacy, speech intelligibility
1. INTRODUCTION
Using an array of loudspeakers, multizone soundfield reproduction
[1] aims to provide listeners in a target zone with their own indi-
vidual soundfield that does not interfere with other zones within the
reproduction region. In some cases, it is desirable to create zones of
quiet, where audio from neighbouring zones is suppressed or can-
celled [1, 2, 3]. The multizone approach can be used for applications
such as the creation of personal sound zones [4] in multi-participant
teleconferencing, restaurants/caf´
es, entertainment/cinema, vehicle
cabins and public announcement locations where the reproduction
can be optimised to provide private quiet zones.
In order to keep the sounds zones personal it is necessary to min-
imise the interzone audio interference (leakage) to maximise the in-
dividual experience. The existence of leakage means that the re-
production of speech in a particular zone may be intelligible in other
zones, deviating from the desired personal sound zones. Some of the
earlier methods treat the leakage with hard constraints and attempt to
completely remove it [1, 2]. This results in zones that are mostly free
of the interference but 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 multizone occlusion problem [1, 4, 5] and has been
dealt with in various ways such as the control of planarity [6], or-
thogonal basis planewaves [3] and alleviated zone constraints [3, 7].
Reproduction in reverberant rooms has also been accomplished with
enhanced acoustic contrast using sparse methods [8].
More recent work has focused on alleviating the constraint so
that the amount of leakage is controlled by a weighting function [3,
7]. Allowing the sound to leak into other zones can improve the
practicality of the system but decreases the individuality of zones.
Existing methods focus on single frequency soundfields, al-
though there has been work attempting to create multizone sound-
fields for wideband speech [9]. More recently, work has been done
[10] to extend a method [3] to the reproduction of weighted wide-
band speech soundfields by using the spatial weighting function.
This is shown in [11] to allow each zone’s acoustic content to be
controlled by dynamic space-time-frequency weighting.
To maintain speech privacy amongst the zones it is necessary to
keep the leaked speech unintelligible [12]. If the leaked speech is
at a level below the threshold of hearing then it may be expected to
start becoming inaudible and/or masked. To reproduce clear speech
in a weighted multizone soundfield at a level of 60 dBA in a zone,
known as the ‘bright’ zone, the level of leaked speech in the quiet
zone could be reduced to around 30 dBA to 35 dBA [8, 11] which
is still well above the threshold of hearing (≈0 dBA).
In this work it is shown for the first time, as far as the authors are
aware, a difference, or contrast, in intelligibility across the personal
sound zones which corresponds to private sound zones. Contribu-
tions are made by evaluating the objective intelligibility of repro-
duced speech and providing methods of control for increased pri-
vacy between zones as a baseline study. A method is provided and
evaluated for increasing privacy in multizone speech soundfields in
anechoic and reverberant environments by using noise to mask the
leaked spectrum into the target quiet zone so that it becomes unin-
telligible. A third contribution is the description and analysis of an
enhanced method for increasing privacy and at the same time im-
proving perceived quality in reproductions, analysed using objective
(instrumental) measures. This is achieved by performing a weighted
multizone reproduction on the noise masker so that it has more in-
fluence in the target quiet zone and less in the target bright zone.
This paper begins with an explanation of the weighted multizone
speech soundfield method used in this work in Section 2. Noise
masking and its relation to speech intelligibility and speech privacy
are explained in Section 3. Results of the noise masking methods
and conclusions are given in Section 4 and Section 5, respectively.
2. WEIGHTED MULTIZONE SPEECH SOUNDFIELDS
The following section provides an overview of the weighted orthog-
onal basis expansion synthesis [3] and the cylindrical harmonic ex-
pansion reproduction [2] used in this work to reproduce speech in
one zone and suppress it in another. This initial step creates a wide-
band controllable contrast in the level between zones which is then
used to reduce leakage between zones.
A multizone soundfield reproduction is depicted in Fig. 1. The
circular reproduction region, D, of radius R, contains three sub re-
gions called the bright, quiet and unattended zone, denoted by Db,
Dqand D∩(Db∪Dq)′, respectively. The radius of Dband Dqis
© 2016 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.

Unattended Zone
Bright Zone
Quiet Zone
rz
rz
rπ
Fig. 1. A weighted multizone soundfield reproduction layout is
shown. The shading depicts the desired bright zone soundfield par-
tially directed towards the quiet zone causing the occlusion problem.
rand their centres are located on a circle of radius rzconcentric
with D. The angle of the desired planewave in Dbis θand is repro-
duced by loudspeakers positioned on an arc of angle φL, radius Rl,
concentric with Dand with the first loudspeaker at angle φ.
Any arbitrary soundfield, including the reproduction of planewave
speech, can be described by an infinite set of planewaves arriving
from all angles [13]. In the orthogonal basis expansion approach to
multizone soundfield reproduction [3] it is shown that a soundfield
function, S(x, k), that fulfils the wave equation, where x∈Dis
an arbitrary spatial sampling point and kis the wavenumber of the
soundfield, can be described with an additional weighting function,
w(x). This weighting function provides relative importance to the
reproduction in different zones and the weighted soundfield function
used throughout this work can be written as
S(x, k) = X
j
Pj(k)Fj(x, k),(1)
where the coefficients for the orthogonal wavefields, Fj(x, k), for a
given weighting function are Pj(k)and j∈ {1,...,N}where N
is the number of basis planewaves [3].
The complex loudspeaker weights used to reproduce the sound-
field in the time-frequency domain are defined as [14]
˜
Ql(k) =
M
X
m=−M
2eimφl∆φsPjPj(k)ime−imφp
iπH(1)
m(kRl),(2)
where M=⌈kR⌉is the truncation length [3], i=√−1,Rand Rl
are from Fig. 1, φp= (j−1)∆φare the wavefield angles, ∆φ=
2π/N,φlis the angle of the lth loudspeaker from the horizontal
axis and ∆φsis the angular spacing of the loudspeakers. Here, Pj
is chosen to minimise the difference between the desired soundfield
and the actual soundfield [3]. In this work frequency, f=kc/2π
[13] and c= 343 m s−1is the speed of sound.
In order to reproduce planewave speech soundfields ˜
Ql(k)must
be applied to the speech in the time-frequency domain and inverse
transformed back to the time-domain to obtain the set of loudspeaker
signals. This can be done by means of a Gabor transform or any
unitary time-frequency transformation as
˜qal (n) = 1
2K
K−1
X
m=0
˜
Ql(m∆k)˜
Ya(m∆k)eiπmn/K ,(3)
where ˜
Ya(k)is the discrete Fourier transform of the ath overlap-
ping windowed frame of the input speech signal, y(n). Each loud-
speaker signal, ql(n), is reconstructed by performing overlap-add
reconstruction with the synthesis window. This results in the loud-
speaker signals, which will reproduce the multiple zones.
The observed signals, p(x, n), can be found at any arbitrary
point in the soundfield by convolving each of the loudspeaker sig-
nals with the transfer function, H(x,xl, k), and summing, as
p(x, n) = 1
2KX
l
K−1
X
m=0
Ql(m∆k)H(x,xl, m∆k)eiπmn/K ,(4)
where xlis the position of the lth loudspeaker and Ql(k)is the time-
frequency transform of ql(n). The soundfield can now be evaluated
at any given point in the reproduction region for different input sig-
nals and the resulting pressure, p(x, n), can be observed in the bright
zone and quiet zone. From this it is possible to analyse the speech
intelligibility in each zone as presented in the following section.
3. PRIVATE SOUND ZONES
This section discusses the relationship between speech privacy and
intelligibility and how they are affected in a multizone soundfield
reproduction scenario. The use of the Speech Intelligibility Contrast
(SIC) is proposed for improving the privacy in personal sound zones.
3.1. Speech Privacy and Intelligibility Contrast
A measure is required to optimally design and evaluate the per-
formance of a method to control privacy in the multizone sound-
field reproduction. The relationship between speech intelligibility
and privacy is highly correlated. Two measurement standards cur-
rently published for assessing speech privacy in closed and open plan
spaces are ASTM E2638 [15] and ASTM E1130 [16], respectively.
These standards are based on two different measures, which are the
Speech Privacy Class (SPC) and the Articulation Index (AI). Both
are highly correlated to speech intelligibility and the SPC has been
shown to be a better measure for higher privacy situations [12] mak-
ing it reasonable to maximise a measure of intelligibility contrast
between zones to obtain privacy.
It has been shown that objective intelligibility measures are
highly correlated with subjective measures and are based on analysing
spectral band powers. High mutual information between the clean
speech (talker), y(n), and the degraded speech (listener), p(x, n)
from (4), is attained at high signal to noise ratio (SNR) [17], hence
indicating that reducing the SNR, for example by adding noise,
reduces intelligibility. In this work the intelligibility for two signals
x1(n)and x2(n)is denoted as I(x1;x2). The particular measure
Mcan be the mutual information, such as that provided by the
Short-Time Objective Intelligibility (STOI) [18] or Speech Trans-
mission Index (STI) [19]. The intelligibility of the pressure signal at
a spatial point xand the signal y(n)is then IM(p(x,·); y).
In this work the SIC is defined as
SICM=1
kDbkZDbIMdx−1
kDqkZDqIMdx,(5)
where kDbkand kDqkare the sizes of Dband Dq, respectively, and
the domain is restricted such that IMfor any x∈Dbis greater than
or equal to IMfor any x∈Dq. The following two subsections pro-
vide two methods to maximise SICM.

3.2. Improving Multizone Privacy
To maximise the SIC, IMmust be zero at all points in Dqwhilst
maintaining maximum IMat all points in Db. Ideally, the mean
SNR of p(x, n)over Dbshould be maintained as high as possible,
so to increase SICMthe mean SNR of p(x, n)over Dqshould be
reduced. To maximise the SIC noise is added to ql(n)under the con-
straint that the mean amplitude of p(x, n)over Dqis less than that
of p(x, n)over Db. This then becomes a constrained optimisation
dependent on the reproduced signals in the bright and quiet zones as
max
G∈RSICM,(6)
where the noise levels, GdB, of ql(n)are optimised.
To increase the SIC a time-domain noise mask, u(n), is added
to each loudspeaker signal, ql(n), which is derived from its time-
frequency domain representation from (3). Noise is added at differ-
ent gain, GdB, relative to the maximum amplitude among Lloud-
speaker signals, A= max({ql(n) : l= 1, ..., L}). The noise mask
is added as
q′
l(n) = ql(n) + u(n)A10
GdB
20 dB ,(7)
where the new loudspeaker signals are q′
l(n). In this work u(n)is
chosen to be uniform white noise with no directivity and this method
is referred to as the ‘Flat Mask’ due to its spatial and spectral unifor-
mity.
Then by transforming q′
l(n)for use in (4), SICMis obtained
from (4) and (5). Now SICMcan be optimised with (6) using GdB
in (7). However, this method does not control u(n)in the spatial
domain and so the mean IMover Dbis also reduced even though
the SIC is maintained.
3.3. Improving Multizone Privacy and Quality
Ideally a private personal sound zone system would have a maximum
SIC whilst maintaining high perceptual quality in the bright zone.
Adding u(n)to ql(n)adds error to p(x, n)for all xwhich as a side-
effect reduces the quality of p(x, n)for any x∈Dband a trade-off
between target quality and privacy becomes necessary. Following
a similar notation to IM, the quality of p(x, n),x∈Dbdegraded
from y(n)is any speech quality assessment model of measure, ´
M,
denoted by B´
M(p(x,·); y)∈ {0, ..., 1}, scaled to match that of IM.
Now a new optimisation can be defined as
max
GdB∈RSICM+λ
kDbkZDbB´
Mdx,(8)
where the noise levels, GdB, are defined below, λis a weighting pa-
rameter for the importance of quality in the optimisation and IM≥
B´
Mfor x∈Db. This optimisation also requires minimum mean
SNR of p(x, n)over Dqand maximum mean SNR of p(x, n)over
Dbachieved here by applying zone weighting to u(n).
To simplify the optimisation of (8) in this work, constraints are
applied to the multizone reproduction of u(n), which is a planewave
field in Dqand quiet in Db. The constraints are θ= 0°, so that the
masker source is collocated with the leakage, and a new weighting
function, ¯w(x), is constrained to an importance in Dqof unity, 104in
Dband 0.05 in the unattended zone. The remainder of the multizone
reproduction is the same as used to generate ql(n)for y(n).
The goal is to find another set of loudspeaker signals that would
reproduce u(n)in Dqto control the mean SNR of p(x, n)over Dq,
therefore solving (8). To do this, u(n)is transformed to the time-
frequency domain as ˜
Ua(k)and used as the input signal in (3). New
loudspeaker weights, ˜
Q′l(k), are derived from (2). Then, from (3),
the loudspeaker signals, ˆql(n), are reconstructed and these become
the new noise mask signals as
q′′
l(n) = ql(n) + ˆql(n)A10
GdB
20 dB ,(9)
where the new loudspeaker signals are q′′
l(n)with noise levels, GdB.
In this work this method is referred to as the ‘Zone Weighted Mask’
due to the masker signal being dependent on the multizone scenario.
Then by transforming q′′
l(n)for use in (4), SICMis obtained
from (4) and (5). Now SICMcan be optimised with (8) using GdB
in (9). The optimisation problem can now be analysed by measuring
IMfor x∈Db∩Dq,B´
Mfor x∈Dband for various GdB.
4. RESULTS
This section presents objective intelligibility results for the bright
and quiet zones in anechoic and reverberant reproduction environ-
ments and discusses the SIC and quality trade-off.
4.1. Multizone Reproduction Evaluation
The layout of Fig. 1 is evaluated, where r= 0.3 m,rz= 0.6 m,R=
1 m and Rl= 1.5 m. The value of θ={0°,15°,90°}for the angle
of the desired planewave virtual source in the bright zone. These
angles are chosen to represent multizone occlusion scenarios. Input
speech signals sampled at 16 kHz are transformed to the frequency
domain using an FFT and 64 ms windows with 50% overlapping.
The loudspeaker signals, q′
l(n)and q′′
l(n), are generated using the
methods outlined in section 2 and 3. The reproduction is performed
for L= 295 and φL= 2πwhich, for the cases in this work, is free
of aliasing problems below 8 kHz [2, 3].
The zone weights are constant and are chosen so that the bright
zone weight is unity, the unattended zone weight is 0.05 the re-
production importance of the bright zone following [3, 10] and the
weight of the quiet zone is set to 104. Frequency dependent zone
weighting and signal filtering may give further improvements. The
noise masking methods, ‘Flat Mask’ and ‘Zone Weighted Mask’, are
applied with GdB ranging from −40 dB to 20 dB in (7) and (9).
Speech files for the evaluation were taken from the TIMIT cor-
pus [20]. Twenty files were randomly selected such that the selection
was constrained to have a male to female speaker ratio of 50 : 50.
Three reverberant rooms and one anechoic are evaluated. The
rooms walls have an absorption coefficient of 0.3and are 4 m ×
9 m ×3 m,8 m×10 m ×3 m and 9 m ×14 m ×3 m, sizes that were
selected to match a small office, medium office and restaurant/caf´
e,
respectively. The multizone setup is placed in the centre of the rooms
and recordings are analysed from both zones where 32 receivers are
positioned randomly in each zone. Room reflections are simulated
using the image method [21] with approximately 446 ×103,206 ×
103and 149 ×103images for each of the respective rooms, at 0.5 s
in length and sampling frequency of 16 kHz.
The reproductions are analysed using the STOI, STI and Percep-
tual Evaluation of Speech Quality (PESQ) [22] measures to evaluate
the performance with SICSTOI and SICSTI in anechoic and rever-
berant environments, respectively. The STOI measure is designed
for the prediction of time-frequency weighted noisy speech like the
simulated recordings in this work. The STI measure is currently

Noise Mask (GdB )
-40 -20 0 20
θ = 90°
PESQ (MOS)
1
1.7
2.4
3.2
3.9
4.6
BZ STOI
QZ STOI
BZ PESQ
STOI (%WC)
0
20
40
60
80
100
θ = 0°
Flat Mask
θ = 0°
Zone Weighted Mask
STOI (%WC)
0
20
40
60
80
100
θ = 15° θ = 15°
Noise Mask (GdB )
-40 -20 0 20
STOI (%WC)
0
20
40
60
80
100
θ = 90°
PESQ (MOS)
1
1.7
2.4
3.2
3.9
4.6
PESQ (MOS)
1
1.7
2.4
3.2
3.9
4.6
Fig. 2. Mean STOI and PESQ are shown for the anechoic environ-
ment and 95% confidence intervals are indicated. BZ and QZ are
the bright and quiet zone, respectively. Black dashed lines indicate
optimum GdB and λ= 1.
the only choice for a reverberant objective intelligibility measure. A
good objective measure for speech quality is the PESQ measure.
The STOI and PESQ are measured in this work with the clean,
y(n), and degraded, p(x, n), speech for each file and receiver com-
bination. The STI is measured for each receiver using the systems
impulse response found with a logarithmic sine sweep. The intelli-
gibility and quality results are then averaged over each zone like that
of (5) and (8). This results in three object measures, two weighting
methods, four rooms, 13 levels of added noise, 20 speech files and
64 receiver positions totalling ≈332,800 data points.
4.2. Intelligibility Contrast from Noise-Based Sound Masking
Fig. 2 shows that by using the ‘Flat Mask’ method to obtain privacy
between zones it is possible to obtain upwards of 85% SICSTOI
but this is only possible within a small range of GdB (−25 dB to
−20 dB). The range remains the same size as the angle is increased
but the GdB which is required to maintain SICSTOI is increased to
approximately −15 dB. In each case of the ‘Flat Mask’ method the
signal in the bright zone is of poor quality as shown by the corre-
sponding PESQ curve (which is undesirable).
It can be seen that θhas a small impact on the range of increased
SICSTOI and it is possible to maintain above 80% SICSTOI for dif-
ferent angles. The effect of θis only minor due to the large zone
weighting used in the reproduction process. Fig. 2 shows that with
a small change in angle, 15°, SICSTOI remains the same and the
PESQ curve starts to rise.
The effect of the spatially weighted noise maskers can be clearly
seen in Fig. 2 where the use of a ‘Zone Weighted Mask’ improves
SICSTOI across all scenarios. The maximum improvement occurs
when GdB is between −5 dB and 20 dB and provides a SICSTOI
of greater than 95% for every scenario. Even when the occlusion
problem is present it is still possible to obtain privacy with greater
STI
0
0.2
0.4
0.6
0.8
1
θ = 0°
Flat Mask
θ = 0°
Zone Weighted Mask
STI
0
0.2
0.4
0.6
0.8
1
θ = 15° θ = 15°
Noise Mask (GdB )
-40 -20 0 20
STI
0
0.2
0.4
0.6
0.8
1
θ = 90°
Noise Mask (GdB )
-40 -20 0 20
θ = 90°
PESQ (MOS)
1
1.7
2.4
3.2
3.9
4.6
PESQ (MOS)
1
1.7
2.4
3.2
3.9
4.6
PESQ (MOS)
1
1.7
2.4
3.2
3.9
4.6
BZ STI
QZ STI
BZ PESQ
Room 1
Room 2
Room 3
Fig. 3. Mean STI and PESQ are shown for the small office, medium
office and restaurant/caf´
e labelled as Room 1, Room 2 and Room 3,
respectively. BZ and QZ are the bright and quiet zones, respectively.
Vertical black lines indicate optimum GdB and λ= 1.
than 95% SICSTOI when GdB is between 0 dB and 15 dB.
Another benefit of using the ‘Zone Weight Mask’ is that the qual-
ity of the bright zone reproduction is increased within the region
where SICSTOI is significantly large. With a SICSTOI of greater
than 70% it is also possible to obtain a PESQ of greater than 3.4 re-
ducing to 3.2 and 2.8 for a SICSTOI of 80% and 90%, respectively.
This shows the trade-off between reproduction quality and zone pri-
vacy which is controlled using λand may depend on the application
of the private multizone system.
With the multizone reproduction in different reverberant rooms
it can be seen in Fig. 3 that a contrast in intelligibility is still possible
without room equalisation. The quality is reduced most likely due
to uncontrolled early reflections inhibiting the bright zone, however,
the SICSTI still remains high at various GdB albeit reduced from
an ideal anechoic environment. The maximum SICSTI is 40% and
occurs with the ‘Zone Weighted Mask’ for Room 3.
5. CONCLUSIONS
This paper has investigated speech privacy between bright and quiet
zones in multizone reproduction scenarios. Methods have been pro-
posed and evaluated for increasing the speech intelligibility contrast
(SIC) in anechoic and reverberant environments showing that added
noise can be used to mask the leaked spectrum to provide a sig-
nificant SIC of higher than 95%. It has also been shown that it is
possible to maintain quality in the bright zone with a PESQ MOS of
3.2 whilst providing a SIC above 80% by using space-time domain
masker signals and that speech privacy can be achieved in reverber-
ant rooms using the methods outlined in this paper. Future work will
look into further improvement of the quality and privacy in reverber-
ant environments as well as a reduction in the number of required
loudspeakers.

6. REFERENCES
[1] M. Poletti, “An investigation of 2-D multizone surround sound
systems,” in Proc. 125th Conv. Audio Eng. Soc. 2008, Audio
Eng. Soc.
[2] Y. J. Wu and T. D Abhayapala, “Spatial multizone soundfield
reproduction: Theory and design,” IEEE Trans. Audio, Speech,
Lang. Process., vol. 19, pp. 1711–1720, 2011.
[3] W. Jin, W. B. Kleijn, and D. Virette, “Multizone soundfield
reproduction using orthogonal basis expansion,” in Int. Conf.
on Acoust., Speech and Signal Process. (ICASSP). 2013, pp.
311–315, IEEE.
[4] T. Betlehem, W. Zhang, M. Poletti, and T. D. Abhayapala,
“Personal Sound Zones: Delivering interface-free audio to
multiple listeners,” IEEE Signal Process. Mag., vol. 32, pp.
81–91, 2015.
[5] T. Betlehem and P. D. Teal, “A constrained optimization
approach for multi-zone surround sound,” in Int. Conf. on
Acoust., Speech and Signal Process. (ICASSP). 2011, pp. 437–
440, IEEE.
[6] P. Coleman, P. Jackson, M. Olik, and J. A. Pedersen, “Personal
audio with a planar bright zone,” J. Acoust. Soc. of Am., vol.
136, pp. 1725–1735, 2014.
[7] H. Chen, T. D. Abhayapala, and W. Zhang, “Enhanced sound
field reproduction within prioritized control region,” in INTER-
NOISE and NOISE-CON Congr. and Conf. Proc. 2014, vol.
249, pp. 4055–4064, Inst. of Noise Control Eng.
[8] W. Jin and W. B. Kleijn, “Theory and design of multizone
soundfield reproduction using sparse methods,” IEEE Trans.
Audio, Speech, Lang. Process., vol. 23, pp. 2343–2355, 2015.
[9] N. Radmanesh and I. S. Burnett, “Generation of isolated wide-
band sound fields using a combined two-stage lasso-ls algo-
rithm,” IEEE Trans. Audio, Speech, Lang. Process., vol. 21,
pp. 378–387, 2013.
[10] J. Donley and C. Ritz, “An efficient approach to dynamically
weighted multizone wideband reproduction of speech sound-
fields,” in China Summit & Int. Conf. Signal and Inform. Pro-
cess. (ChinaSIP). 2015, pp. 60–64, IEEE.
[11] J. Donley and C. Ritz, “Multizone reproduction of speech
soundfields: A perceptually weighted approach,” in Asia-
Pacific Signal & Inform. Process. Assoc. Annu. Summit and
Conf. (APSIPA ASC). 2015, pp. 342–345, IEEE.
[12] B. N. Gover and J. S. Bradley, “ASTM metrics for rating
speech privacy of closed rooms and open plan spaces,” Cana-
dian Acoust., vol. 39, pp. 50–51, 2011.
[13] E. G. Williams, Fourier Acoustics: Sound Radiation and
Nearfield Acoustical Holography, Academic Press, 1999.
[14] Y. J. Wu and T. D. Abhayapala, “Theory and design of
soundfield reproduction using continuous loudspeaker con-
cept,” IEEE Trans. Audio, Speech, Lang. Process., vol. 17,
pp. 107–116, 2009.
[15] Standard test method for objective measurement of the speech
privacy provided by a closed room, ASTM Int. E2638-10,
2010.
[16] Standard test method for objective measurement of speech pri-
vacy in open plan spaces using articulation index, ASTM Int.
E1130-08, 2008.
[17] W. B. Kleijn, J. B. Crespo, R. C. Hendriks, P. Petkov, B. Sauert,
and P. Vary, “Optimizing speech intelligibility in a noisy envi-
ronment: A unified view,” IEEE Signal Process. Mag., vol. 32,
pp. 43–54, 2015.
[18] C. H. Taal, R. C. Hendriks, R. Heusdens, and J. Jensen,
“An algorithm for intelligibility prediction of timefrequency
weighted noisy speech,” IEEE Trans. Audio, Speech, Lang.
Process., vol. 19, pp. 2125–2136, 2011.
[19] Sound system equipment-Part 16: Objective rating of speech
intelligibility by speech transmission index, IEC 60268-16,
2003.
[20] J. Garofolo, L. Lamel, W. Fisher, J. Fiscus, D. Pallett,
N. Dahlgren, and V. Zue, “TIMIT acoustic-phonetic contin-
uous speech corpus,” Linguistic Data Consortium, 1993.
[21] J. B. Allen and D. A. Berkley, “Image method for efficiently
simulating small-room acoustics,” J. Acoust. Soc. of Am., vol.
65, pp. 943–950, 1979.
[22] A. W. Rix, J. G. Beerends, M. P. Hollier, and A. P. Hek-
stra, “Perceptual evaluation of speech quality (PESQ)-a new
method for speech quality assessment of telephone networks
and codecs,” in Int. Conf. on Acoust., Speech and Signal Pro-
cess. (ICASSP). 2001, pp. 749–752, IEEE.