On the design of differential loudspeaker

arrays with broadside radiation patterns

Cite as: JASA Express Lett. 1, 084804 (2021); https://doi.org/10.1121/10.0005760
Submitted: 11 May 2021 • Accepted: 12 July 2021 • Published Online: 31 August 2021

Jiale Wang, Wen Zhang, Chao Pan, et al.

ARTICLES YOU MAY BE INTERESTED IN

Machine learning in acoustics: Theory and applications

The Journal of the Acoustical Society of America 146, 3590 (2019); https://
doi.org/10.1121/1.5133944

Spheroidal ambisonics: Spatial audio in spheroidal bases

JASA Express Letters 1, 084803 (2021); https://doi.org/10.1121/10.0005942

Use of the filtered-x least-mean-squares algorithm to adapt personal sound zones in a car
cabin
The Journal of the Acoustical Society of America 150, 1779 (2021); https://
doi.org/10.1121/10.0005875

JASA Express Lett. 1, 084804 (2021); https://doi.org/10.1121/10.0005760 1, 084804

© 2021 Author(s).
 ARTICLE asa.scitation.org/journal/jel

On the design of differential loudspeaker arrays

with broadside radiation patterns
Jiale Wang,1 Wen Zhang,1,a) Chao Pan,1 Jingdong Chen,1 and Jacob Benesty2
1
Center of Intelligent Acoustics and Immersive Communications, School of Marine Science and Technology,
Northwestern Polytechnical University, Xi’an, Shaanxi 710072, China
2 
Institut National de la Recherche Scientifique, Energie, Materiaux et Telecommunications, University of Quebec,
Montreal, Quebec, Canada
wang.jl@mail.nwpu.edu.cn, wen.zhang@nwpu.edu.cn, cpan@nwpu.edu.cn, jingdongchen@ieee.org, Jacob.Benesty@inrs.ca

Abstract: Loudspeaker arrays with high directivity are desirable in many acoustic and sound applications to direct sounds
into a desired region. One way of designing such arrays is through the differential operator to maximize the directivity factor.
However, this method generally works for linear arrays with endfire steering direction and its usage to generate a broadside
radiation pattern is restricted to the second-order with three loudspeakers. This paper presents a general approach to the
design of differential linear loudspeaker arrays with broadside radiation patterns of any-order. Three methods are presented
to find the beamforming filter with design examples provided. V C 2021 Author(s). All article content, except where otherwise noted, is

licensed under a Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

[Editor: David C Swanson] https://doi.org/10.1121/10.0005760
Received: 11 May 2021 Accepted: 12 July 2021 Published Online: 31 August 2021

1. Introduction
Directional sound radiation,1–3 which is achieved by an array of loudspeakers, has a wide range of applications such as
acoustic contrast control in multi-zone reproduction (or personal sound zones),4 room reverberation reduction,5 and effi-
ciency improvement in spatial active noise control.6 This problem can be regarded as a reciprocal problem of microphone
array beamforming.7,8 The resulting arrays can be broadly divided into two categories, i.e., additive arrays, which are based
on the sum operator, and differential arrays, which are based on the differential operator.8
An additive array generates a radiation pattern (or beampattern) using the constructive interference of the sound
fields generated by each element of the loudspeaker array. It has advantages of high radiation efficiency and high robust-
ness against system mismatches. However, due to the diffraction limit,9 the low-frequency directivity of the additive array
is limited by its aperture size with respect to the wavelength and, as a result, a large-sized array is needed to achieve direc-
tional sound radiation at low frequencies.
In contrast, a differential array of small size can be used to generate a narrow beampattern. The differential
configuration can be applied on each element of the array, i.e., using two or more closely spaced transducers to form a
directional source/sensor as the array element,10 or on the entire array, which has a small aperture size to achieve super-
directivity.11,12 It has the advantage of high array directivity, but has low radiation efficiency and is sensitive to system
mismatches,11,12 known as the low white-noise-gain (WNG) problem.9 Therefore, recent efforts have been focusing on
designing robust differential arrays by formulating the design problem as one of optimization with additional constraints
to improve WNG.4,13 Another issue with differential arrays is that it is difficult to generate broadside patterns with a linear
array (i.e., the direction of the mainlobe is perpendicular to the line that connects all the loudspeakers as shown in Fig. 1).
So far, only the case of using three elements to generate a second-order differential broadside pattern has been investi-
gated,8,9 where the differential and delay-and-sum patterns are combined to improve its efficiency.
This paper is devoted to the design of broadside patterns with linear differential loudspeaker arrays, which was
shown to be possible with linear microphone arrays through numerical results of steering the mainlobe direction into the
broadside direction.14 In this work, we first derive a general formula of the broadside differential radiation pattern for lin-
ear loudspeaker arrays, where we assume that the number of loudspeakers is odd ( 3) so that the broadside differential
pattern is always symmetric and has a gain of one at the broadside direction and an even number of nulls. Then, the
design problem is formulated as one of optimization. Three kinds of solutions are subsequently deduced, which include
the method with equality constraints of the broadside direction and distinct nulls, the method with maximum WNG to
improve radiation efficiency, and a trade-off approach. The effectiveness of the proposed techniques is demonstrated

a)
Author to whom correspondence should be addressed. ORCID: 0000-0002-0752-6123.

JASA Express Lett. 1 (8), 084804 (2021) C Author(s) 2021.

V 1, 084804-1
 ARTICLE asa.scitation.org/journal/jel

Fig. 1. Illustration of a differential loudspeaker array with a fourth-order broadside differential pattern.

through simulations and compared to a reference method of combining the differential and delay-and-sum (DAS) patterns
to improve radiation efficiency.

2. Broadside differential radiation patterns

2.1 Problem formulation
As depicted in Fig. 1, an M-element uniform linear loudspeaker array lies on the x axis, whose center is located at the ori-
gin, with an interelement spacing equal to r. In this work, small-spacing loudspeaker arrays with differential beampatterns
are considered for low-frequency sound radiation and the size of each loudspeaker and the spacing between neighboring
loudspeakers are assumed to be much smaller than the wavelength. Therefore, the loudspeaker array is compact and the
loudspeakers are modelled as omnidirectional point sources. In this case, we focus on the sound pressure generated by the
array at a far-field location of distance r and angle h written as follows:

eıkr X
M0
pðk; r; hÞ¢ w ðkÞeıkxm sin h ; (1)
4pr m¼M0 m

where ı is the imaginary unit, wm ðkÞ denotes the weight of the mth loudspeaker at xm ¢mr with m ¼ M0 ; M0
þ1; …; M0  1; M0 and M0 ¢ðM  1Þ=2, the superscript ðÞ stands for the complex-conjugate operator, the angle h is
defined with respect to the positive y axis, k¢2pf =c is the wave number, f denotes the frequency, and c is the speed of
sound in air (around 340 m/s).
Putting Eq. (1) into a vector form gives
eıkr H
pðk; r; hÞ ¼ w ðkÞdðk; hÞ; (2)
4pr
where the superscript ðÞH is the conjugate-transpose operator,
 T
wðkÞ¢ wM0 ðkÞ    wM0 ðkÞ ; (3)
sin h T
dðk; hÞ¢½eıkxM0 sin h    eıkxM0 ; (4)
T
and the superscript ðÞ is the transpose operator.
With the signal model given in Eq. (2), the objective of this work is to design the loudspeaker weights (or driv-
ing signals) in wðkÞ in order to achieve a highly directive or superdirective radiation in the broadside direction, i.e.,
h ¼ 0 . We focus on the design of the low-frequency region using a small-spacing array with the differential excitation
function adopted to overcome the diffraction limit.9 To achieve this objective, we make the following assumptions: (1) the
spacing between neighboring loudspeakers in the array is much smaller than the smallest wavelength in the frequency
band of interest, (2) an odd number of loudspeakers (i.e., M ¼ 2N þ 1 with N being a positive integer number) is used to
design a broadside pattern, which is symmetric with respect to the y axis and has a gain of one at h ¼ 0 and 2 N nulls,
and (3) the weights are symmetric, i.e., wm ðkÞ ¼ wm ðkÞ, in order to generate the symmetric broadside
eıkr X M0
pðk; r; hÞ ¼ w ðkÞeıkmr sin h
4pr m¼M0 m
" #
eıkr X
M0
 
¼ w ðkÞ þ 2 wm ðkÞ cos ðkmr sin hÞ
4pr 0 m¼1

¼ pðk; r; hÞ:

JASA Express Lett. 1 (8), 084804 (2021) 1, 084804-2

 ARTICLE asa.scitation.org/journal/jel

Before ending this subsection, let us define two functions, which will be used in the following for beamforming.
The first one is the array excitation function, which is defined as
X
M0
qðk; xÞ¢ wm ðkÞdðx  xm Þ: (5)
m¼M0

It represents the driving signals of the loudspeaker array in a continuous form as a function of the wavenumber and spa-
tial position. The second function is the far-field broadside beampattern, which is defined as
 1
eıkr
bðk; hÞ¢pðk; r; hÞ 
4pr
ð1
ıkx sin h
¼ qðk; xÞe dx: (6)
1

2.2 Second-order radiation pattern

In this subsection, we review the formula of the second-order differential pattern, where three point sources are used to
form the pattern. The array excitation function can be expressed as
pffiffiffiffiffi
ð2Þ dðx þ rÞ  2 cos n1 dðxÞ þ dðx  rÞ
q ðk; xÞ ¼ ; (7)
n1
which is a function of the speaker positions and weights, where the superscript ðÞð2Þ denotes the second order and
n1 ¢ðkrb1 Þ2 , with b1 being a tuning parameter for the broadside beampattern. The excitation function in Eq. (7) corre-
sponds to the case where the three loudspeakers are at r; 0; r; with weight wðkÞ ¼ ð1=n1 Þ  ½1  2 cos ðkrb1 Þ 1T .
Then, the radiation pattern in the far-field15 can be approximated as8
2
bð2Þ ðk; hÞ ¼ ½cos ðkr sin hÞ  cos ðkrb1 Þ
n1
sin2 h
1 2 : (8)
b1
ð1Þ
The nulls of this pattern are decided by the parameter b1, i.e., hnull ¼ 6arcsinðb1 Þ, with 0 < b1 1.

2.3 General higher-order radiation patterns

Now, we derive a general formula of the 2Nth-order broadside differential pattern based on a recurrence relation of the
array excitation function. That is, a 2Nth-order excitation function can be written in term of combining three 2ðN  1Þ
th-order excitation functions, or a spatial convolution of the 2ðN  1Þ th-order and the second-order excitation function,
i.e.,

q2ðN1Þ ðk; x þ rÞ  2 cos ðkrbN Þq2ðN1Þ ðk; xÞ þ q2ðN1Þ ðk; x  rÞ

qð2NÞ ðk; xÞ ¼
nN
¼ q2ðN1Þ ðk; xÞ qð2Þ ðk; x; bN Þ
¼ qð2Þ ðk; x; b1 Þ qð2Þ ðk; x; b2 Þ  qð2Þ ðk; x; bN Þ; (9)
where denotes the spatial convolution operator. Equation (9) shows that a 2Nth-order excitation function can also be
represented as the N-fold spatial convolution of the second-order excitation function with tuning parameters
fbn ; n ¼ 1; 2; …; Ng. It follows then that the radiation pattern can be written as
Y
N
bð2NÞ ðk; hÞ ¼ bð2Þ ðk; h; bn Þ
n¼1
!
Y
N
sin2 h
1 2 : (10)
n¼1 bn

Note that the deduction of Eq. (10) is based on the fact that the far-field radiation pattern can be regarded as the wave-
number spectrum of the array excitation function (or its spatial Fourier transform with the variable k0 ¢k sin h).9,16 Then,
the radiation pattern of the N-fold convolution of the second-order excitation function is equivalent to the product of N
second-order radiation patterns.

JASA Express Lett. 1 (8), 084804 (2021) 1, 084804-3

 ARTICLE asa.scitation.org/journal/jel

Equation (10) shows that a 2Nth-order differential pattern is uniquely determined by its 2 N nulls located at
ð2NÞ
hnull ¼ 6arcsinðb1 Þ; 6arcsinðb2 Þ; …; 6arcsinðbN Þ, where 0 < b1 b2    bN 1. Note that at least 2N þ 1 loud-
speakers are required to generate such a pattern.
The general formula in Eq. (10) can also be represented in a sum form, i.e.,
X
N
bð2NÞ ðk; hÞ ¼ 1 þ aN;n sin2n h; (11)
n¼1
where aN;n ; n ¼ 1; 2; …; N are also functions of fbn ; n ¼ 1; 2; …; Ng. The directivity of the higher-order differential pat-
tern is represented by sin2N h as in Eq. (11). The higher the order, the narrower is the mainlobe.

3. Design of loudspeaker array weights

The objective of designing loudspeaker array weights is to derive wðkÞ in Eq. (2) so that its radiation pattern would be as
close as possible to an ideal differential pattern in Eq. (11). It is shown from Sec. 2 that an ideal 2Nth-order differential
pattern has a one at the broadside direction h ¼ 0 and 2 N nulls. In this section, we study the case where 2 N nulls are
distinct. We formulate the design problem as one of optimization as in Ref. 13 to derive the corresponding solutions.

3.1 Design with equality constraints

We first study the design of the array weights with M ¼ 2N þ 1 speakers using equality constraints (EC), i.e., the main-
lobe direction of 0 and 2 N distinct nulls in the angle range of ½90 ; 0 Þ [ ð0 ; 90 . Note that the broadside radiation
pattern is symmetric. Therefore, the nulls are symmetric with respect to the mainlobe direction. So, only N distinct nulls
hN;n ; n ¼ 1; …; N, located in the range of ð0 ; 90  need to be identified.
We express these fundamental constraints in a matrix form as
Aðk; hÞwðkÞ ¼ i; (12)
where
Dðk; hÞ
Aðk; hÞ¢ ; (13)
C
 T
h¢ 0 hN;1    hN;N ; (14)
 H
Dðk; hÞ¢ dðk; 0 Þ dðk; hN;1 Þ    dðk; hN;N Þ ; (15)
h
^
C¢ I 0  I ; (16)
T
i¢½1 0    0 ; (17)
I is the identity matrix of size M0  M0 , and I is the flip of the matrix I. The matrix Dðk; hÞ, of size ðN þ 1Þ  M, is the main-
lobe and null direction constraints. The matrix C, of size M0  M, is used to maintain the symmetry of the array weights.
When the number of constraints is equal to the number of weights, Aðk; hÞ is a square matrix. Given that hN;n are
6 0 ; 8n; Aðk; hÞ is of full rank. It follows then that the array weights can be determined uniquely as
distinct nulls, and hN;n ¼
w EC ðkÞ ¼ A1 ðk; hÞi: (18)

3.2 WNG improvement

The differential loudspeaker array with a highly directive radiation pattern generally suffers from the intrinsic low-
radiation efficiency problem as a large number of evanescent sound waves cannot propagate to far-field.9 The radiation
efficiency is represented by WNG,8 which is expressed as
jbð2NÞ ðk; 0 Þj2
WNGðkÞ ¼ ð 1
jqð2NÞ ðk; xÞj2 dx
1
1
¼ : (19)
wH ðkÞwðkÞ
Note that WNG is often expressed in the decibel scale. The higher the value of WNG, the more efficient is the
radiation.
A common approach to improving WNG is by using M > 2N þ 1 speakers to design a 2Nth-order broadside
differential pattern.17 According to Eq. (19), we formulate the following optimization problem to maximize WNG, i.e.,
min wH ðkÞwðkÞ s: t:Aðk; hÞwðkÞ ¼ i: (20)
wðkÞ

JASA Express Lett. 1 (8), 084804 (2021) 1, 084804-4

 ARTICLE asa.scitation.org/journal/jel

Note that in this case, we have N þ 1 beampattern constraints and M0 constraints to maintain the symmetry of the array
weights.
The solution to Eq. (20) is known as the minimum norm (MN) solution,13,17
 1
wMN ðkÞ ¼ AH ðk; hÞ Aðk; hÞAH ðk; hÞ i; (21)
which gives the maximum WNG. However, as will be shown in Sec. 4, this design may lead to frequency-variant radiation
patterns.
To further improve the stability of the design, more constraints can be included, in addition to the N þ 1 funda-
mental beampattern constraints. Assume that we have L more constraints at (hL;l , bl), where 0 < hL;1 <    < hL;L 90
and hL;l is different from hN;n , with bl being the corresponding value of the desired gain at hL;l . Then, with N þ L þ 1
beampattern constraints and M0 constraints to maintain the symmetry of the array weights, a linear system is constructed
as follows:13
~ h; hÞwðkÞ ¼ ~i;
Aðk; (22)
where
" #
Aðk; hÞ
~ h; hÞ ¢
Aðk; ; (23)
Dðk; hÞ
 H
Dðk; hÞ ¢ dðk; hL;1 Þ    dðk; hL;L Þ ; (24)
T
h ¢½hL;1    hL;L  ; (25)
T
b ¢½b1    bL  ; (26)
and

~i¢ i : (27)
b
The array uses M  2ðN þ LÞ þ 1 loudspeakers and the problem is also formulated as one of optimization to
maximize WNG, i.e.,

Fig. 2. The desired (theoretical) and designed radiation patterns at frequency of 500 Hz: (a), (b), and (c) are, respectively, the desired second-
order, fourth-order, and sixth-order radiation patterns with the maximum directivity index, and (d), (e), and (f) are the patterns designed
with the EC method.

JASA Express Lett. 1 (8), 084804 (2021) 1, 084804-5

 ARTICLE asa.scitation.org/journal/jel

Fig. 3. Performance of the studied methods to design a fourth-order broadside differential pattern: (a) WNG and (b) DI.

min w H ðkÞwðkÞ ~ h; hÞwðkÞ ¼ ~i;

s: t: Aðk; (28)
wðkÞ

whose solution is
h i1
~ H ðk; h; hÞ Aðk;
wMNA ðkÞ ¼ A ~ H ðk; h; hÞ ~i;
~ h; hÞA (29)

which is called the minimum-norm solution with additional constraints (MNA). This gives another WNG improved solu-
tion but using additional constraints, which are different from the null constraints. The resulting radiation pattern will bet-
ter match the desired differential pattern than the MN method.

4. Simulations
In this section, we perform simulations to validate the proposed methods and compare the results with the existing
approach that combines the differential and DAS patterns to improve the WNG.8 The loudspeakers are assumed to be
omnidirectional point sources and the interelement spacing is r ¼ 5 cm.
We first study the design of SLA using the method presented in Sec. 3.1 with 2N þ 1 equality constraints, i.e., a
gain of 1 at the broadside direction and 2 N distinct nulls. The desired radiation patterns includes the second-order,
fourth-order, and sixth-order broadside differential patterns with the maximum directivity index (DI).8 With these pat-
terns, the maximum achievable acoustic power is concentrated in the preferred broadside direction given a fixed total radi-
ation power.3 The designed patterns are plotted in Fig. 2. As seen, the radiation patterns designed by the proposed method
are almost the same as the desired patterns.
Next, we compare the three proposed methods, i.e., the EC, MN, and MNA solutions, to design a fourth-order
broadside differential pattern with the maximum DI, using 21 omnidirectional loudspeakers. For EC, the central 5 loud-
speakers are used. For MNA, an additional constraint is set at the angle of 16 , which is within the 3 dB (or half-power)
bandwidth of the mainlobe. The performance measures are WNG and DI. We also plot the results by the method that
combines the differential and DAS patterns (abbreviated as DD in Fig. 3)8 to improve the WNG for comparison. As
shown in Fig. 3(a), the reference method (DD) and the proposed MN and MNA solutions can significantly improve the
WNG, i.e., the radiation efficiency, compared with the standard EC solution. The MN method gives the maximum WNG.
Figure 3(b) plots the DI of different methods as a function of frequency. While the EC method has an almost constant DI
over frequency, the DI of the MN and DD methods increases with frequency, i.e., the generated radiation pattern is vari-
ant with frequency. The MNA solution can approximately maintain the same DI as EC and also achieve a higher WNG.
Figure 4 plots the radiation patterns with the MN and MNA methods in a wider frequency range. It can be seen
that the mainlobe of the MN solution becomes narrower as frequency increases, while the radiation pattern of the MNA
solution keeps almost frequency invariant. In summary, the MN solution achieves the maximum WNG, i.e., the maximum
radiation efficiency, while the MNA solution can maintain a frequency-invariant radiation pattern with slightly reduced
WNG.

Fig. 4. Two dimensional plots of a fourth-order broadside differential pattern designed by (a) MN and (b) MNA solutions.

JASA Express Lett. 1 (8), 084804 (2021) 1, 084804-6

 ARTICLE asa.scitation.org/journal/jel

5. Conclusion
In this paper, we derived the general formula of higher-order broadside differential patterns for linear loudspeaker array.
We then proposed three methods for the design of the loudspeaker array weights. The EC method can successfully design
the desired radiation pattern, but it has low radiation efficiency. The MN solution gives the maximum WNG but the
resulting radiation pattern changes with frequency. In comparison, the MNA solution (at a price of reduced directivity)
maintains a frequency-invariant radiation pattern with high radiation efficiency.
Acknowledgments
This work was partially supported by the National Natural Science Foundation of China (NSFC) under Grants Nos. 61671380
and 61831019.
References and links
1
D. G. Meyer, “Digital control of loudspeaker array directivity,” J. Audio Eng. Soc. 32(10), 747–754 (1984).
2
D. Kim, K. Kim, S. Wang, S. Q. Lee, and M. J. Crocker, “Maximization of the directivity ratio with the desired audible gain level for
broadband design of near field loudspeaker arrays,” J. Sound Vib. 330(23), 5517–5529 (2011).
3
M. M. Boone, W.-H. Cho, and J.-G. Ih, “Design of a highly directional endfire loudspeaker array,” J. Audio Eng. Soc. 57(5), 309–325
(2009).
4
J.-W. Choi, Y. Kim, S. Ko, and J.-H. Kim, “Super-directive loudspeaker array for the generation of personal sound zone,” in
Proceedings of the 125th Convention of the Audio Engineering Society, San Francisco, CA (2008).
5
M. A. Poletti, T. Betlehem, and T. Abhayapala, “Higher-order loudspeakers and active compensation for improved 2D sound field
reproduction in rooms,” J. Audio Eng. Soc. 63(1–2), 31–45 (2015).
6
N. Murata, J. Zhang, Y. Maeno, and Y. Mitsufuji, “Global and local mode-domain adaptive algorithms for spatial active noise control
using higher-order sources,” in Proceedings of IEEE ICASSP, Brighton, UK (2019), pp. 526–530.
7
M. R. Bai and Y.-H. Hsieh, “Point focusing using loudspeaker arrays from the perspective of optimal beamforming,” J. Acoust. Soc.
Am. 137(6), 3393–3410 (2015).
8
Y.-H. Kim and J.-W. Choi, Sound Visualization and Manipulation (Wiley, Singapore, 2013), Chap. 5, pp. 219–281.
9
J.-W. Choi, Y. Kim, S. Ko, and, and J. Kim, “A differential approach for the implementation of superdirective loudspeaker array,” in
Proceedings of the 128th Convention of the Audio Engineering Society, London, UK (2010).
10
H. F. Olson, “Gradient loudspeakers,” J. Audio Eng. Soc. 21(2), 86–93 (1973).
11
E. Mabande and W. Kellermann, “Towards superdirective beamforming with loudspeaker arrays,” in Proceedings of the 19th
International Congress on Acoustics, Madrid, Spain (2007).
12
M. F. Sim on Galvez, S. J. Elliott, and J. Cheer, “A superdirective array of phase shift sources,” J. Acoust. Soc. Am. 132(2), 746–756
(2012).
13
J. Chen, J. Benesty, and C. Pan, “On the design and implementation of linear differential microphone arrays,” J. Acoust. Soc. Am.
136(6), 3097–3113 (2014).
14
J. Jin, G. Huang, X. Wang, J. Chen, J. Benesty, and I. Cohen, “Steering study of linear differential microphone arrays,” IEEE Trans.
Audio Speech Lang. Proc. 29, 158–170 (2021).
15
The approximation is based on the fact that kDx 1 and the cos term can be expanded using a Taylor series, that is, cos x
1  x2 =2.
16
E. G. Williams, Fourier Acoustics: Sound Radiation and Nearfield Acoustical Holography (Elsevier, Amsterdam, 1999), Chap. 2, pp.
15–83.
17
J. Benesty and J. Chen, Study and Design of Differential Microphone Arrays (Springer Science & Business Media, New York, 2012), pp.
1–282.

JASA Express Lett. 1 (8), 084804 (2021) 1, 084804-7