AN IMPROVED SOLUTION TO THE FREQUENCY-INVARIANT BEAMFORMING WITH

CONCENTRIC CIRCULAR MICROPHONE ARRAYS

Xudong Zhao1 , Gongping Huang2 , Jingdong Chen1 , and Jacob Benesty3

1
CIAIC, Northwestern Polytechnical University, Xi’an, Shaanxi 710072, China
2
Technion Israel Institute of Technology, Technion City, Haifa 32000, Israel
3
INRS-EMT, University of Quebec, Montreal, QC H5A 1K6, Canada

ABSTRACT is the angular position of the mth element in the pth ring, and the
Frequency-invariant beamforming with circular microphone arrays superscript T is the transpose operator.
(CMAs) has drawn a significant amount of attention for its steering Assume that the desired signal comes from the direction θs .
flexibility and high directivity. However, frequency-invariant beam- Then, our objective is to design a frequency-invariant beampattern,
forming with CMAs often suffers from the so-called null problem, which is as close as possible to a desired directivity pattern with its
which is caused by the zeros of the Bessel functions; then, concen- main beam pointing in the direction θs . For that, a complex weight
∗
tric CMAs (CCMAs) are used to deal with this problem. While Hp,m (ω) is applied at the output of the mth microphone on the pth
frequency-invariant beamforming with CCMAs can mitigate the null ring, where the superscript ∗ is the complex-conjugate operator, and
problem, the beampattern is still suffering from distortion due to s- then the weighted outputs are summed together to form the beam-
patial aliasing at high frequencies. In this paper, we find that the forming output. Putting together all the complex weights in a vector
spatial aliasing problem is caused by higher-order circular harmon- form, we have
ics. To deal with this problem, we take the aliasing harmonics into [ ]T
h(ω) = hT1 (ω) hT2 (ω) · · · hTP (ω) , (4)
account and approximate the beampattern with a higher truncation [ ]T
order of the Jacobi-Anger expansion than required. Then, the beam- hp (ω) = Hp,1 (ω) Hp,2 (ω) · · · Hp,Mp (ω) . (5)
forming filter is determined by minimizing the errors between the In this paper, we will use the three commonly used measures
desired directivity pattern and the approximated one. Simulation re- to evaluate the performance of the designed beamformer, i.e., the
sults show that the developed method can mitigate the distortion of beampattern, the white noise gain (WNG), and the directivity factor
the beampattern caused by spatial aliasing. (DF) [1].
Index Terms— Microphone arrays, concentric circular arrays, The beampattern describes the sensitivity of the beamformer to
frequency-invariant beamforming, spatial aliasing. plane waves arriving at the array from different directions; it is de-
fined as
1. SIGNAL MODEL, PROBLEM FORMULATION, AND B [h (ω) , θ] = hH (ω) d (ω, θ)
PERFORMANCE MEASURES
∑
P ∑
Mp
∗
Consider a UCCMA with P rings, where the pth ring, of radius = Hp,m (ω) eȷϖp cos(θ−ψp,m ) , (6)
p=1 m=1
rp , consists of Mp omnidirectional microphones. The total number
∑P
of microphones of the entire UCCMA is M = p=1 Mp . Sup-
where the superscript H is the conjugate-transpose operator.
pose that we want to steer the beamformer in the direction θ, where The WNG reflects the robustness of the beamformer; it is de-
θ ∈ [0, 2π). In this scenario, the steering vector, of length M , cor- fined as [1]
responding to a plane wave of a farfield source is defined as  H 
h (ω) d (ω, θs )2
[ ]T W [h (ω)] = . (7)
d (ω, θ) = dT1 (ω, θ) dT2 (ω, θ) · · · dTP (ω, θ) , (1) hH (ω) h (ω)
where The DF is defined as the ratio of the beampattern gain at the
[ look direction to the average beampattern gain at other directions. It
dp (ω, θ) = ȷϖp cos(θ−ψp,1 ) ȷϖp cos(θ−ψp,2 )
e e is equal to the signal-to-noise ratio (SNR) gain of the beamformer in
]T the spherically isotropic noise field [1, 2], i.e.,
··· eȷϖp cos(θ−ψp,Mp ) (2)  H 
h (ω) d (ω, θs )2
D [h (ω)] = H , (8)
is a vector of length Mp , ȷ is the imaginary unit with ȷ2 = −1, h (ω) Γd (ω) h (ω)
ϖp = ωrp /c, c is the speed of sound in the air, i.e., c = 340 m/s,
where the elements of the M × M matrix Γd (ω) are given by
ω = 2πf is the angular frequency, f > 0 is the temporal frequency, ( )
ωδij
2π(m − 1) [Γd (ω)]ij = sinc , (9)
ψp,m = , m = 1, 2, . . . , Mp (3) c
Mp
with δij = ∥ri − rj ∥2 being the distance between microphone i and
This work was supported in part by the National Key Research and De- j, ri , rj ∈ {r1,1 , r1,2 , . . . , rp,Mp , . . . , rP,MP }, and
velopment Program of China under Grant No. 2018AAA0102200 and in
part by the Key Program of National Science Foundation of China (NSFC)
[ ]T
rp,m = rp cos ψp,m sin ψp,m (10)
Grant No. 61831019 and the NSFC and Israel Science Foundation (ISF) joint
research program under Grant No. 61761146001. being the coordinate of the mth microphone in the pth ring.

978-1-5090-6631-5/20/$31.00 ©2020 IEEE 556 ICASSP 2020

 2. BEAMPATTERN DESIGN WITH THE JACOBI-ANGER 3. THE SPATIAL ALIASING PROBLEM
EXPANSION Substitute (22) into (6), we can write the beampattern as
According to the DMA study in the literature [1, 2], an N th-order B [hMN (ω) , θ]
frequency-independent directivity pattern with its main beam point-
ing in the direction θs , is defined as [3] ∑
N
c2N,n (θs )
= ∑P ψ H (ω) d (ω, θ)
Mp Jn2 (ϖp ) n
∑
N n=−N p=1
B (b2N , θ − θs ) = b2N,n eȷn(θ−θs ) ∑
N ∑P
c2N,n (θs )
n=−N = ∑P (ȷn )∗ Jn (ϖp ) eH
n,p dp (ω, θ)
2
p=1 Mp Jn (ϖp )
= [Υ (θs ) b2N ]T pe (θ) , (11) n=−N p=1

∑
N ∑
P
where b2N,n , n = −N, . . . , 0, . . . , N are real coefficients that de- = u∗n,p (ω) εn,p (ω, θ) , (23)
termine the shape of the desired directivity pattern, n=−N p=1

[ ]T
b2N = b2N,−N · · · b2N,0 · · · b2N,N , (12) where
[ −ȷN θ ]T
pe (θ) = e ··· 1 ··· e ȷN θ
(13) c∗2N,n (θs )
un,p (ω) = ∑P ȷn Jn (ϖp ) (24)
2
p=1 Mp Jn (ϖp )
are vectors of length 2N + 1, and
( ) is the weight on the pth ring for the order n,
Υ (θs ) = diag eȷN θs , . . . , 1, . . . , e−ȷN θs (14)
∑
Mp

is a (2N + 1) × (2N + 1) diagonal matrix. εn,p (ω, θ) = eH

n,p dp (ω, θ) = eȷnψp,m eȷϖp cos(θ−ψp,m )
With the Jacobi-Anger expansion [4], we have the following re- m=1
lation: ∑
Mp
∑
∞
′ ′
∑
∞ = eȷnψp,m ȷn Jn′ (ϖp ) eȷn (θ−ψp,m )
ȷϖp cos(θ−ψp,m ) n −ȷnψp,m ȷnθ
e = ȷ Jn (ϖp )e e , (15) m=1 n′ =−∞
n=−∞
∑
∞
′ ′ ∑
Mp
′
= eȷn θ ȷn Jn′ (ϖp ) eȷ(n−n )ψp,m (25)
where Jn (ϖp ) is the Bessel function of the first kind [4]. Substitut-
n′ =−∞ m=1
ing (15) into (6) gives
is the summation of circular harmonics, and
∑
P ∑
Mp
∗
∑
∞
−ȷnψp,m ȷnθ
B [h (ω) , θ] = Hp,m (ω) n
ȷ Jn (ϖp )e e ∑
Mp
∑
Mp
′ ȷ(n−n′ ) M
2π (m−1)
p=1 m=1 n=−∞ eȷ(n−n )ψp,m = e p

∑
N m=1 m=1
{
≈ eȷnθ hH (ω) ψ n (ω) , (16) Mp , n − n′ = qMp
n=−N = , (26)
0, others
where
where q = 0, ±1, ±2, . . . are integers.
[ ]T
ψ n (ω) = ȷn Jn (ϖ1 ) eTn,1 · · · Jn (ϖP ) eTn,P , (17) Suppose that we limit the order of the Jacobi-Anger expansion
[ −ȷnψ ] to ±N ′ with N ′ > N , and for circular harmonics whose orders
T
en,p = e p,1
e−ȷnψp,2 · · · e−ȷnψp,Mp . (18) are larger than N ′ or smaller than −N ′ , the magnitudes of their
coefficients are small enough to be neglected. Then, we have
Now, forcing B [h (ω) , θ] to be equal to B (b2N , θ − θs ), we obtain
′
∑
N ∑
Mp
Ψ (ω) h (ω) = Υ∗ (θs ) b2N , (19) εn,p (ω, θ) =
′ ′
eȷn θ ȷn Jn′ (ϖp )
′
eȷ(n−n )ψp,m
n′ =−N ′ m=1
where [
[ ]H = Mp eȷnθ ȷn Jn (ϖp ) + eȷN1 θ ȷN1 JN1 (ϖp ) + · · ·
Ψ (ω) = ψ −N (ω) ··· ψ 0 (ω) ··· ψ N (ω) (20) ]
+ eȷNLn θ ȷNLn JNLn (ϖp ) , (27)
is a (2N + 1) × M matrix. Assume that Mp > (2N + 1), then
ΨH (ω) is a full-column rank matrix. The minimum-norm solution where Nl = n − ql Mp , l = 1, 2, . . . , Ln , are the orders of aliasing
of (19) gives harmonics, and ql are nonzero integers with (n − N ′ ) /Mp 6 q1 <
[ ]−1 q2 < · · · < qLn 6 (n + N ′ ) /Mp . For example, if we choose
hMN (ω) = ΨH (ω) Ψ (ω) ΨH (ω) Υ∗ (θs ) b2N , (21) N = 2, Mp = 6, and N ′ = 8, the orders of aliasing harmonics in
the pth ring are described in Table 1.
which can also be written as
∑
N
c∗2N,n (θs ) Table 1. The Orders of Aliasing Harmonics.
hMN (ω) = ∑P 2
ψ n (ω) , (22) n −2 −1 0 1 2
p=1 Mp Jn (ϖp )
n=−N
Nl 4, −8 5, −7 6, −6 7, −5 8, −4
where c2N,n (θs ) = e−ȷnθs b2N,n .

557
 8.0 8.0 0dB
4. BEAMPATTERN DESIGN WITH THE SUPPRESSION
OF ALIASING HARMONICS 6.0 6.0
-10dB

-20dB
Taking those aliasing harmonics into account, the beampattern 4.0 4.0
B [h (ω) , θ] can be approximated as -30dB
2.0 2.0
′
-40dB
∑
N
B [h (ω) , θ] ≈ e ȷnθ H
h (ω) ψ n (ω) . (28) 0 0 -50dB
-180 -90 0 90 180 -180 -90 0 90 180
n=−N ′
8.0 8.0 0dB
Now, the objective of frequency-invariant beamforming is to find
an optimal filter, which makes B [h (ω) , θ] as close as possible to 6.0 6.0
-10dB
B (b2N ′ , θ − θs ) and suppresses the effect of aliasing harmonics. -20dB
This can be expressed as the following optimization problem: 4.0 4.0
-30dB
∑ 2
N
H 2.0 2.0
-40dB
min h (ω) ψ n (ω) − c2N,n (θs ) +
h(ω) n=−N 2
0 0 -50dB
N′
∑ 2 ∑−1
−N 2 -180 -90 0 90 180 -180 -90 0 90 180
H H
h (ω) ψ n (ω) + h (ω) ψ n (ω) .
n=N +1 2 n=−N ′ 2
8.0 8.0 0dB
(29)
-10dB
6.0 6.0
Solving (29), we get -20dB
4.0 4.0
−1 -30dB
hopt (ω) = Φ (ω) g (ω) , (30)
2.0 2.0
-40dB
where
0 0 -50dB
′ -180 -90 0 90 180 -180 -90 0 90 180
∑
N
Φ (ω) = ψ n (ω) ψ H
n
(ω) (31)
n=−N ′
Fig. 1. Performance of the developed and conventional beam-
is an M × M matrix and
formers with a UCCMA for the design of the second-order su-
∑
N percardioid: (a.1) beampattern B [hMN (ω) , θ], (a.2) beampattern
g (ω) = c∗2N,n (θs ) ψ n (ω) (32) BD [hMN (ω)[ , θ], (a.3) beampattern
] BA [hMN (ω) ,[θ], (b.1) beam-
]
pattern B hopt,µ (ω) , θ , (b.2) beampattern BD hopt,µ (ω) , θ ,
n=−N
[ ]
and (b.3) beampattern BA hopt,µ (ω) , θ .
is a vector of length M . In order to make this approach more robust,
we can add a small (and positive) regularization parameter, µ, i.e.,

hopt,µ (ω) = [Φ (ω) + µIM ]−1 g (ω) , (33) is a matrix of size P × P with
where IM is the M × M identity matrix. [ ]T
For the particular case of M1 = M2 = · · · = MP = M , the jn (ω) = ȷn Jn (ϖ1 ) Jn (ϖ2 ) ··· Jn (ϖP ) , (37)
analytical expression of (33) can be written as

∑
N the definition of en is the same as in (18) where we drop the param-
hopt,µ (ω) = wn (ω) ⊗ en , (34) eter p, and ⊗ is the Kronecker product.
n=−N Substituting (34) into (28), we get the equation at the bottom of
this page. It is seen that the beampattern
[ can be] expressed as two
where parts: the desired beampattern BD hopt,µ (ω) , θ , which approach-
c∗2N,n (θs ) −1 es to
[ the desired directivity
] pattern, and the aliasing beampattern
wn (ω) =
M
Dn (ω) jn (ω) (35) BA hopt,µ (ω) , θ , which reflects the effect of aliasing harmonics.
Similarly, we can get BD [hMN (ω) , θ] and BA [hMN (ω) , θ] for the
is a vector of length P , minimum-norm filter, where we need to substitute the vector wn (ω)
in (38) with un (ω) and
n (ω) + jN1 (ω) jN1 (ω) + · · ·
Dn (ω) = jn (ω) jH H

+ jNLn (ω) jH
µ [ ]T
NLn (ω) + IP (36) un (ω) = un,1 (ω) un,2 (ω) ··· un,P (ω) . (39)
M

[ ] ∑
N ∑
N [ ]
B hopt,µ (ω) , θ = M eȷnθ wnH (ω) jn (ω) + M eȷN1 θ wnH (ω) jN1 (ω) + · · · + eȷNLn θ wnH (ω) jNLn (ω) . (38)
n=−N n=−N
| {z } | {z }
BD [hopt,µ (ω),θ ] BA [hopt,µ (ω),θ ]

558
 and the minimum-norm beamforming filters in decibel scale, i.e.,
-20
20 log10 |wn,p (ω) | and 20 log10 |un,p (ω) |, for different n and p.
Note that wn,p (ω) is the pth element of wn (ω) according to (35).
-40 It is seen that the coefficients of the minimum-norm beamformer are
larger in magnitude at the outer rings and low frequencies to achieve
the maximum WNG while the coefficients of proposed beamformer
-60
are larger in magnitude at the inner rings and high frequencies to
avoid spatial aliasing. This explains the underlying reason that how
-80 the proposed beamformer avoid spatial aliasing, i.e., it emphasizes
0 2.0 4.0 6.0 8.0 more on the contribution of the innermost rings to avoid the spatial
0 aliasing since the interelement spacing of the inner rings is much s-
maller than that of the outer rings and aliasing is less likely to happen
-20
for inner rings than for outer rings.

-40 6. CONCLUSIONS
Frequency-invariant beamforming with CCMAs may suffer from
-60 the spatial aliasing problem, particularly at high frequencies when
the size of the array is large. This paper studied the problem and
-80
we showed through theoretically analysis that the spatial aliasing
0 2.0 4.0 6.0 8.0 of frequency-invariant beamforming with CCMAs is caused by
higher-order circular harmonics. With the presented analysis, one
can easily determine the harmonics that cause aliasing. Then, using
0
the Jacobi-Anger expansion, we developed a beamforming method
-20
that can make the designed beampattern resembles the desired di-
rectivity pattern while suppressing the effect of aliasing harmonics
-40
to form frequency-invariant beampatterns in a large frequency range
of interest. Simulations were performed to compare the developed
-60 method with the existing beamforming methods with UCCMA and
results justified the effectiveness of the developed method.
-80
0 2.0 4.0 6.0 8.0
7. RELATION TO PRIOR WORK
Frequency-invariant beamforming with circular microphone arrays
(CMAs) are widely used to pick up audio and speech signals of
Fig. 2. The magnitudes of the weights of the designed and conven- interest from noisy observations [5–12]. Among several frequency-
tional beamforming filters for different orders and rings: (a) n = 0, invariant beamforming methods developed in the literature, the dif-
(b) n = 1, and (c) n = 2. ferential beamforming one with CMAs has attracted much research
interest for its high directivity, frequency-invariant beampatterns,
and steering flexibility [3, 11]. However, differential beamformers
5. SIMULATIONS with CMAs may suffer from white noise amplification problem, and
In this section, we study the performance of the developed frequency- null problems in the white noise gain (WNG) and in the directivity
invariant beamformer through simulations. We consider a UCCMA factor (DF). To deal with those problems, the approach with concen-
consisting of 3 rings with M1 = M2 = M3 = 6, r1 = 1.5 cm, tric CMAs (CCMAs) was developed [13–15], which does not only
r2 = 3 cm, r3 = 5 cm. The desired directivity pattern is chosen as improve the WNG but also mitigate the null problems in WNG and
the second-order supercardioid whose coefficients are DF.
[ ]T However, it is well known that microphone array systems often
b2N = 0.1032 0.242 0.3095 0.242 0.1032 , suffer from the spatial aliasing problem, especially when the size of
the array is large. This problem can seriously distort the beampat-
◦ ′
the desired direction is set to θs = 0 , and we choose N = 8 and tern at high frequencies; therefore, many efforts have been made to
µ = 10−13 . deal with it [12,16–20]. In this paper, we analyze the spatial aliasing
We compare the performance of the designed beamformers: problem of the frequency-invariant beamformer with uniform CC-
hMN (ω) and hopt,µ (ω). The results are plotted in Fig. 1. Fig- MAs (UCCMAs) and show that the aliasing is caused by higher-
ure[ 1 (a.1) and ](b.1) plot the beampatterns B [hMN (ω) , θ] and order circular harmonics. With our analysis, one can easily deter-
B hopt,µ (ω) , θ . Figure 1 (a.2) and (b.2) plot the desired beam- mine the different orders of harmonics that cause aliasing. Follow-
[ ]
patterns BD [hMN (ω) , θ] and BD hopt,µ (ω) , θ respectively. It ing the principle in [15], we then develop a method to improve the
[ ]
is seen that B hopt,µ (ω) , θ does not change much in the stud- performance of the beamformers. Same as the technique in [15], the
ied frequency range and is close to the desired directivity pat- beamforming filter with the developed method is found by minimiz-
tern. By comparison, B [hMN (ω) , θ] is considerably distorted ing the error between the desired directivity pattern and the beam-
at high frequencies due to the spatial aliasing problem. Figure 1 pattern approximated with the Jacobi-Anger expansion. The major
(a.3)[ and (b.3) plot] the aliasing beampatterns BA [hMN (ω) , θ] and difference is that we approximate the beampattern with the Jacobi-
BA hopt,µ (ω) , θ , respectively. It is seen clearly that the devel- Anger expansion of higher truncation orders and take into accoun-
oped beamformer hopt,µ (ω) can suppress almost all the aliasing t the aliasing harmonics during beamforming. The developed ap-
harmonics. proach can mitigate the distortion of the beampattern caused by the
Figure 2 plots the magnitude of the weights of the proposed spatial aliasing, which is corroborated by simulation results.

559
 8. REFERENCES [18] D. L. Alon and B. Rafaely, “Beamforming with optimal alias-
ing cancellation in spherical microphone arrays,” IEEE/ACM
[1] J. Benesty and J. Chen, Study and Design of Differential Mi- Trans. Audio, Speech, Lang. Process., vol. 24, no. 1, pp. 196–
crophone Arrays. Berlin: Springer-Verlag, 2012. 210, 2016.
[2] G. W. Elko and J. Meyer, “Microphone arrays,” in Springer [19] D. L. Alon and B. Rafaely, “Spatial aliasing-cancellation for
Handbook of Speech Processing, J. Benesty, M. M. Sondhi, circular microphone arrays,” in Proc. 4th Joint Workshop
and Y. Huang, Eds. Berlin, Germany: Springer-Verlag, 2008, Hands-Free Speech Communication and Microphone Arrays
ch. 50, pp. 1021–1041. (HSCMA), 2014, pp. 137–141.
[3] G. Huang, J. Benesty, and J. Chen, “On the design of [20] D. L. Alon and B. Rafaely, “Spatial decomposition by spher-
frequency-invariant beampatterns with uniform circular micro- ical array processing,” in Parametric Time-Frequency Domain
phone arrays,” IEEE/ACM Trans. Audio, Speech, Lang. Pro- Spatial Audio, V. Pulkki, S. Delikaris-Manias, and A. Politis,
cess., vol. 25, no. 5, pp. 1140–1153, 2017. Eds. Hoboken, NJ, USA: Wiley, 2017.
[4] M. Abramowitz and I. A. Stegun, Handbook of Mathematical
Functions: with Formulas, Graphs, and Mathematical Tables.
New York: Dover, 1970.
[5] J. Meyer, “Beamforming for a circular microphone array
mounted on spherically shaped objects,” J. Acoust. Soc. Am.,
vol. 109, pp. 185–193, Jan. 2001.
[6] S. Yan, Y. Ma, and C. Hou, “Optimal array pattern synthesis
for broadband arrays,” J. Acoust. Soc. Am., vol. 122, no. 5, pp.
2686–2696, 2007.
[7] H. Teutsch and W. Kellermann, “Acoustic source detection and
localization based on wavefield decomposition using circular
microphone arrays,” J. Acoust. Soc. Am., vol. 120, pp. 2724–
2736, Nov. 2006.
[8] S. Chan and H. Chen, “Uniform concentric circular arrays with
frequency-invariant characteristics: theory, design, adaptive
beamforming and doa estimation,” IEEE Trans. Signal Pro-
cess., vol. 55, no. 1, pp. 165–177, 2007.
[9] J. Meyer and G. Elko, “Spherical harmonic modal beamform-
ing for an augmented circular microphone array,” in Proc.
IEEE ICASSP, 2008, pp. 5280–5283.
[10] S. Yan, “Optimal design of modal beamformers for circular
arrays,” J. Acoust. Soc. Am., vol. 138, no. 4, pp. 2140–2151,
2015.
[11] J. Benesty, J. Chen, and I. Cohen, Design of Circular Differen-
tial Microphone Arrays. Berlin, Germany: Springer-Verlag,
2015.
[12] B. Rafaely, B. Weiss, and E. Bachmat, “Spatial aliasing in
spherical microphone arrays,” IEEE Trans. Signal Process.,
vol. 55, no. 3, pp. 1003–1010, 2007.
[13] G. Huang, J. Benesty, and J. Chen, “Design of robust con-
centric circular differential microphone arrays,” J. Acoust. Soc.
Am., vol. 141, no. 5, pp. 3236–3249, 2017.
[14] G. Huang, J. Chen, and J. Benesty, “On the design of robust s-
teerable frequency-invariant beampatterns with concentric cir-
cular microphone arrays,” in Proc. IEEE ICASSP, 2018, pp.
506–510.
[15] G. Huang, J. Chen, and J. Benesty, “Insights into frequency-
invariant beamforming with concentric circular microphone
arrays,” IEEE/ACM Trans. Audio, Speech, Lang. Process.,
vol. 26, no. 12, pp. 2305–2318, 2018.
[16] Z. Li and R. Duraiswami, “Flexible and optimal design of
spherical microphone arrays for beamforming,” IEEE Trans.
Audio, Speech, Lang. Process., vol. 15, no. 2, pp. 702–714,
2007.
[17] J. Meyer and G. W. Elko, “Handling spatial aliasing in spher-
ical array applications,” in Proc. Hands-Free Speech Commu-
nication and Microphone Arrays (HSCMA), 2008, pp. 1–4.

560