Steerable frequency-invariant beamforming for arbitrary arrays

Lucas C. Parraa兲
Biomedical Engineering Department, City College of New York, New York 10031
共Received 9 September 2005; revised 27 March 2006; accepted 28 March 2006兲
Frequency-invariant beamforming aims to parameterize array filter coefficients such that the
spectral and spatial response profiles of the array can be adjusted independently. Solutions to this
problem have been presented for specific sensor configurations often requiring a larger number of
sensors. However, in practical applications, the number and location of sensors are often restricted.
This paper proposes to find an optimal linear basis transformation that decouples the frequency
response from the spatial response. A least-squares optimal basis transform can be computed
numerically for arbitrary sensor configurations, for which typically no exact analytical solutions are
available. This transform can be further combined with a spherical harmonics basis resulting in
readily steerable broadband beams. This solution to broadband beamforming effectively decouples
the array geometry from the steering geometry. Furthermore, for frequency-invariant beams, this
approach results in a significant reduction in the number of beam-design parameters. Here, the
method is demonstrated for an optimal design of far-field response for an irregular linear array with
as few as three sensors. © 2006 Acoustical Society of America. 关DOI: 10.1121/1.2197606兴
PACS number共s兲: 43.60.Fg 关EJS兴 Pages: 3839–3847

I. INTRODUCTION ment. This paper presents a numerical approach to construct

an optimal frequency-invariant response for an arbitrarily
An array of spatially distributed sensors can be made chosen array configuration.
selective in space and frequency by filtering and summing Following previous work,8 the resulting frequency-
the output of multiple sensors. For a fixed geometry, the invariant response is made steerable by combining it with the
spectro-spatial response profile is determined by the filter spherical harmonics decomposition of the beam pattern.2,9
coefficients. Changing a given coefficient will typically af- The required coefficients for rotating spherical harmonics are
fect both the frequency as well as the spatial response profile. given by the Wigner rotation matrix.10 A numerical approach
This spectro-spatial coupling complicates broadband filter related to the present method has been considered in the
design as well as adaptive beamforming algorithms. The goal context of near-field design but has not been fully
of this work is to find a parameterization of the filter coeffi- developed.3 A related technique has been presented for arbi-
cients that decouples the spatial selectivity from the fre- trarily placed sensors on a sphere but without addressing
quency selectivity for arbitrary array configurations. Once frequency invariance.11
decoupled, a frequency-invariant response is obtained by The paper is organized as follows: Secs. II and III intro-
choosing the same coefficients for multiple frequencies. This duce the notation and define the goal of frequency-invariant
simplifies broadband beamforming as the frequency response beamforming. Sections IV and V present existing solutions
and the spatial profile can be adjusted independently. based on analytic expansions of plane waves and the corre-
Broadband and frequency-invariant beamforming has sponding inversion formulas. The numerical least-squares
been addressed by Ward et al.,1 covering far-field problems, approach proposed in this work is presented in Sec. VI.
Beam steering is discussed in Sec. VII giving some special
and Abhayapala et al.2 and Kennedy et al.3 for near-field
considerations to the linear array. Section VIII discusses sen-
problems. Their approach is based on the spatial Fourier
sor noise and the resulting regularization of the least-squares
transform of a continuous aperture. In practical implementa-
solution. Section IX explains how the proposed basis trans-
tions, the aperture needs to be sampled with a discrete num-
formation can be used to make existing adaptive beamform-
ber of sensors. This leads to specific array configurations
ing algorithms frequency invariant. Finally, examples are
typically with a large number of sensors on a linear or rect-
presented using data-independent beam design for a linear
angular lattice.2,4,5 More recent work presents analytic inver-
array. The paper closes with a discussion on how the pro-
sion methods for linear,6 cylindrical,7 and spherical arrays.8 posed method can be applied to other array configurations,
Although these array configurations may be optimal in terms directional sensors, and near-field beam design. To better fol-
of frequency invariance, reduced aliasing, or spatial resolu- low the main argument, the reader may skip Secs. IV, V,
tion, they may not be practical in some applications. In par- VII A, and VIII in a first reading.
ticular, speech acquisition with embedded microphones re-
quires broadband arrays often with a very small number of II. ARRAY RESPONSE
microphones 共two to five兲 in a constrained spatial arrange-
This section defines some terminology commonly used
with sensor arrays and beam forming. Denote the signal
a兲
Electronic mail: [email protected] sampled by the nth sensor at time t as xn共t兲. A beamformer

J. Acoust. Soc. Am. 119 共6兲, June 2006 0001-4966/2006/119共6兲/3839/9/$22.50 © 2006 Acoustical Society of America 3839
convolves the N sensor signals with the corresponding filter L

coefficients cn共t兲 and sums the result to generate an array cn共k兲 = 兺 bnl共k兲c̃l共k兲, 共8兲
output signal: l=1

N such that the basis transform bnl共k兲 converts the array re-
y共t兲 = 兺 cn共t兲 ⴱ xn共t兲. 共1兲 sponse into a frequency-invariant array response,
n=1 N

The convolution is denoted here with the symbol ⴱ. 兺 gn共k,⍀兲bnl共k兲 = g̃l共⍀兲.

n=1
共9兲
For far-field design, it is customary to discuss the effect
of this processing by considering the signals produced by a The basis transform replaces the N sensors indexed by n
planar wave impinging on the sensors. Assume a steady-state by a new set of L virtual sensors indexed by l. These virtual
plane wave with radial frequency ␻ traveling in direction sensors are now frequency invariant. This basis transform, in
⍀ = 共␽ , ␸兲, where spherical angles ␽ and ␸ represent eleva- turn, factorizes the filter-array response, which is seen by
tion and azimuth, respectively. Orientation and frequency combining Eqs. 共7兲–共9兲:
can also be specified in Cartesian coordinates as a three- L
vector k pointing in direction ⍀ with k = 储k储 characterizing f共k,⍀兲 = 兺 c̃l共k兲g̃l共⍀兲. 共10兲
the wavenumber k = ␻ / c. Because of the simple linear rela- l=1
tionship, the wavenumber k will also be referred to as “fre-
quency” in this paper. The plane wave elicits, at location rn, By comparing Eq. 共10兲 with Eq. 共7兲, it becomes clear
a pressure signal:9 that modifying the new parameters c̃l共k兲 will affect the spa-
tial response profile f共k , ⍀兲 only at frequency k. In fact, the
xn共t兲 = ei共k·rn−␻t兲 , 共2兲 spatial response profile of the filter array is fully determined
by coefficients c̃l共k兲. In particular, a frequency-invariant
with i2 = −1. To obtain the angle-dependent frequency re-
beamformer is obtained by choosing the same coefficients
sponse, consider the temporal Fourier transform F of this
for all frequencies, c̃l共k兲 = c̃l.
pressure signal:
The challenge of uncoupling the spatial from the spec-
xn共␯兲 = F兵xn共t兲其 = 2␲␦共␯ − ␻兲eik·rn . 共3兲 tral response lies in finding a basis transform bnl共k兲 that sat-
isfies Eq. 共9兲—even if only approximately.
This is the response of the sensors to the plane wave which
defines the sensor response gn共k兲: IV. ARRAY RESPONSE FOR VARIOUS GEOMETRIES
gn共k兲 = e ik·rn
. 共4兲 To understand existing analytic solutions to this prob-
The ␦-function expresses the monochromatic nature of the lem, it will be useful to express the sensor response 关Eq. 共4兲兴
plane wave, and is typically omitted to facilitate a more com- in terms of the arrival direction ⍀. The plane wave response
pact notation.9 Similarly, the Fourier transform of the result- of a sensor located at rn = 共rn⬘ , ⍀n⬘兲 can be expanded10,9 as
ing response of the filter and sum: ⬁ l
gn共k,⍀兲 = 4␲ 兺 i jl共krn⬘兲 兺 Y m
l 共⍀兲Y l 共⍀n⬘兲,
l m*
共11兲
y共␯兲 = F兵y共t兲其 = 2␲␦共␯ − ␻兲f共k兲, 共5兲 l=0 m=−l

leads to the definition of the filter-array response f共k兲: where jl共kr兲 are spherical Bessel functions of the first kind,
N and Y ml 共⍀兲 are spherical harmonics. The response of a
f共k兲 = 兺 cn共k兲gn共k兲. 共6兲 spherical array is given by Eq. 共11兲 with rn⬘ = r⬘. For a hori-
n=1 zontally placed circular array with ␽n⬘ = ␲ / 2, and restricting
to the horizon as arrival direction ␽ = ␲ / 2, Eq. 共11兲 becomes7
⬁

III. FREQUENCY-INVARIANT BEAMFORMING

gn共k, ␸兲 = 兺 ilJl共kr⬘兲eil共␸−␸⬘兲 ,
l=−⬁
n 共12兲

Now consider Eq. 共6兲 with k rewritten in terms of the where Jl共kr兲 are Bessel functions of the first kind. In a ver-
arrival direction ⍀ and frequency k, tically aligned linear array with ␽n⬘ = 0, Eq. 共11兲 becomes
N
independent of ␸, and is therefore symmetric about the array
axis:6
f共k,⍀兲 = 兺 cn共k兲gn共k,⍀兲. 共7兲
n=1 gn共k,⍀兲 = e−ikrn⬘ cos ␽ . 共13兲
This equation can be seen as a parameterization of the
filter-array response for each frequency k with coefficients
V. ANALYTIC INVERSION APPROACHES
cn共k兲, and basis functions gn共k , ⍀兲. Modifying coefficients
cn共k兲 will affect the frequency and spatial response simulta- Let us now consider existing analytic solutions to the
neously because gn共k , ⍀兲 depends on both the frequency k factorization problem 共9兲. A general theory for continuous
and arrival direction ⍀. The goal of frequency-invariant apertures has been proposed1 with corresponding approxima-
beamforming is to find a new parameterization for cn共k兲: tions for discrete arrays.1,5 Some specific array configura-

3840 J. Acoust. Soc. Am., Vol. 119, No. 6, June 2006 Lucas C. Parra: Steerable frequency-invariant beamforming
tions make use of elegant analytic inversion formulas, as in basis transform is computed once for each frequency, k, and
the case of spherical,8 hemispherical,12 circular,13,7 linear,6 remains unaltered by subsequent beam design.
and rectangular4 arrays. Equations 共11兲–共13兲 were presented The accuracy of the proposed method depends on how
to give an indication of possible approaches. To give an idea well G can be inverted, that is, how well can G̃ be repre-
of these methods, consider the case of a circular array. For a sented by B : ␧ = 储GB − G̃储2 = 储GG†G̃ − G̃储2. This is the mini-
circular array, one may choose: mum attainable square error in satisfying the factorization
bnl共k兲 = N−1i−lJ−1
l 共kr兲e
il␸n
. 共14兲 condition 共9兲 for a specific set of angles ⍀q. Any other basis
transform is suboptimal 共in the least-squares sense兲.
Inserting Eqs. 共14兲 and 共12兲 into Eq. 共9兲 will indeed give Uniform sampling of angles ⍀ gives equal weight to the
a frequency-invariant function of space, g̃l共␸兲 = e−il␸, assum- optimality criteria ␧. Nonuniform sampling will place a big-
ing that the following orthogonality condition holds for any ger weight on the areas in angular space that are sampled
l , m: more densely. To avoid large errors for intermediate angles
N that have not been sampled, one should choose band-limited
1
兺 ei共m−l兲␸n = ␦ml .
N n=1
共15兲 basis functions G̃. Spherical harmonics represent such a ba-
sis and will be presented next.
Unfortunately, this is only correct for 兩l兩 ⬍ N, and only if
the N sensors are placed on an equidistant lattice along the
circular array. As a result, Eq. 共9兲 is only approximately cor- VII. BEAM STEERING
rect. For circular, as well as spherical, arrays the approxima-
tion can only be improved with a larger number of sensors, Ideally, the new spatial basis, g̃l共⍀兲, should be easily
while the lattice must be carefully arranged to match the steerable. This means that the overall orientation of a beam
analytic inversion formulas.13,8 The same is true for linear can be rotated without changing its spatial profile, i.e., with-
and rectangular arrays.1,5 In addition, for those configura- out changing coefficient c̃l共k兲. This implies that after rotation
tions, the angular response profile at any given frequency of the frame of reference, the basis can be expressed in terms
only partially determines the required spatial Fourier basis of the same basis. The basis of spherical harmonics satisfies
coefficients.6 Arbitrary Fourier coefficients have to be chosen this property;
outside of the determined range, further compromising the
l
accuracy of the approximations.
Ym ⬘
l 共⍀⬘兲 = 兺 Dmm
l

m=−l
⬘
共␣, ␤, ␥兲Y m
l 共⍀兲. 共18兲
VI. LEAST-SQUARES SOLUTION
The goal of this work is to directly minimize the ap- ⍀ and ⍀⬘ are the spherical angles before and after the rota-
proximation error resulting from the restricted number of tion of the frame of reference. The rotation can be specified
sensors, and to overcome the restrictions on sensor locations by the Euler angles ␣ , ␤ , ␥, where ␥ is an initial spin about
imposed by analytic inversions. The proposed solution is to the original z axis, ␤ changes its elevation, and ␣ is a sub-
invert Eq. 共9兲 numerically. To this end, discretize the arrival sequent change in longitude. Explicit expressions for coeffi-
directions with angles ⍀q, q = 1 , . . . , Q, and write Eq. 共9兲 in
l
cients Dmm ⬘
共␣ , ␤ , ␥兲 were first given by Wigner10,14 and are
matrix notation variably referred to as Wigner D-functions or Wigner rota-
tion matrix. Efficient algorithms are available to compute
G共k兲B共k兲 = G̃, 共16兲 these coefficients from a conventional 3 ⫻ 3 rotation matrix
given in Cartesian coordinates.15
where 关G共k兲兴qn = gn共k , ⍀q兲, 关B共k兲兴nl = bnl共k兲, and 关G̃兴ql An additional advantage of the spherical harmonics is
= g̃l共⍀q兲. The new desired spatial basis vectors can be de- their uniform resolution.16 Equation 共18兲 states that any ro-
fined in the columns of G̃共k兲, while G共k兲 is determined by tated version of the spherical harmonics of order L can be
the array configuration. For each frequency, k, Eq. 共16兲 represented exactly by harmonics of, at most, order L. This
specifies LQ conditions with LN unknowns. Typically, only implies that band-limited beam patterns, which do not oscil-
a few sensors are available; yet one would like to parameter- late faster than ␲ / L, can be represented exactly with har-
ize the response for many different arrival directions. There- monics of at most order L.17,18
fore, with Q ⬎ N, the problem is overdetermined. The least- Thanks to identity 共18兲 the spherical harmonics are
squares solution to this problem 共i.e., the B that will readily steerable and therefore are a natural choice for the
reproduce G with the smallest square error兲 is computed with new basis:8
the pseudo-inverse, G† = 共GHG兲−1GH,
g̃lm共⍀兲 = Y m
l 共⍀兲. 共19兲
B共k兲 = G†共k兲G̃. 共17兲
This requires that G is of full rank—a condition that has The virtual sensors now require double index lm instead of
to be verified in practice. Note that the discretization of the just l. With the matrix notation adopted for Eq. 共16兲, a pair
arrival directions is only used to compute the basis transform lm indexes a column of matrix G̃. Arranging the Wigner
bnl共k兲. When applying the new basis, the arrival direction can rotation coefficients as matrix 关D共␣ , ␤ , ␥兲兴lm,l⬘,m⬘
take on any arbitrary values. For a given array geometry, the l
= Dmm ⬘
共 ␣ , ␤ , ␥ 兲, the basis is then rotated with

J. Acoust. Soc. Am., Vol. 119, No. 6, June 2006 Lucas C. Parra: Steerable frequency-invariant beamforming 3841
 FIG. 1. Signal flow diagram.

L
G̃⬘ = G̃D共␣, ␤, ␥兲, 共20兲
Pl⬘共cos ␽⬘兲 ⬇ 兺 Dll⬘共␤兲Pl共cos ␽兲, 共23兲
and the basis transform that includes an arbitrary rotation is l=0
given by
B⬘共k, ␣, ␤, ␥兲 = B共k兲D共␣, ␤, ␥兲.
To summarize the overall design structure, let us com-
共21兲
Dll⬘共␤兲 =
2l + 1
2
冕 0
␲
Pl⬘共cos共␽ − ␤兲兲Pl共cos ␽兲 dcos ␽ ,

bine the previous equations in a compact form. The response 共24兲

of the array is now given by
f共k,⍀兲 = g共k,⍀兲B共k兲D共␣, ␤, ␥兲c̃共k兲. 共22兲 which defines a basis rotation or shift analogous to Eq. 共18兲.
The approximation results from the truncation of the sum at
The corresponding signal-flow diagram is shown in Fig. 1.
L ⬍ ⬁. An optimal choice of basis would minimize this ap-
Vector g共k , ⍀兲 is the response of the N sensors to a plane proximation error, implying that a shift of the basis can be
wave. B共k兲 uncouples the spatial response from the spectral expressed accurately within the same order 共l 艋 L兲. There-
response. D共␣ , ␤ , ␥兲 steers the spatial response into an arbi- fore, in an ideal basis the shift matrix, 关D共␤兲兴ll⬘ = Dll⬘共␤兲, will
trary direction; and c̃共k兲 defines its spatial profile for each have triangular structure such as in the case of spherical
frequency separately. Choosing the same coefficients c̃共k兲 harmonics.16 Figure 2 shows that the shift matrix D共␤兲 for
for all frequencies yields an array response that is frequency the Legendre basis is, in fact, approximately triangular re-
invariant. sulting in small truncation errors. Therefore, for the linear
array we suggest to use,
This structure uncouples the steering geometry from that
of the array geometry. In previous approaches, the ability to
g̃l共␽兲 = Pl共cos ␽兲. 共25兲
steer seemed inevitably linked to the choice of array archi-
tecture and the corresponding analytic inversion formulas.
Using the appropriate definitions for matrix D共␤兲 based
The problem of steering the array has also been uncoupled
on Eq. 共24兲, one can write the corresponding equations for
from that of choosing its spatial profile, and finally, the de-
this approximate shift of the beam pattern as in Eqs. 共20兲 and
sign of the frequency response has been separated from that
共21兲.
of the spatial response.
In principle, there should be a closed-form solution to
integral 共24兲. However, in practice, it may be more efficient
to evaluate the integral once 共perhaps numerically兲 for each
A. Steering in a linear array desired shift angle ␤, and store the L2 coefficients for later
For the special case of a linear array, the situation is use. To evaluate the Legendre coefficients of a general func-
complicated by the lack of full control over the two- tion, f共z兲, numerically one may discretize the angle with Q
dimensional beam pattern. In a vertical linear array, one can samples, zq = cos ␽q, and convert the integral into a discrete
specify the response profile only in elevation ␽. Rotating an sum. This can be expressed efficiently in matrix-vector nota-
axis-symmetric shape in elevation by ␤ violates the axial tion by defining matrix 关P兴ql = Pl共zq兲, and vector 关f兴q = f共zq兲.
symmetry around the z axis. Therefore, one cannot rotate a The Legendre coefficient for function f共z兲 arranged as vector
beam pattern by angle ␤ and at the same time preserve sym- 关c̃兴l = c̃l are computed with
metry in ␸. As a result, there is no basis in ␽ that is isomor-
phic with respect to a shift. However, one can aim to find a c̃ = P†f, 共26兲
transformation that approximately preserves a spatial profile
defined in ␽ 苸 关0 , ␲兴. One may choose, for instance, the Leg- where the pseudo-inverse P† implements the integration sum
endre basis and expand its shifted version in the same basis and proper quadrature weights for any sampling of ␽.19
function set; Transformation 共26兲 is inverted in approximation with

3842 J. Acoust. Soc. Am., Vol. 119, No. 6, June 2006 Lucas C. Parra: Steerable frequency-invariant beamforming
 FIG. 2. Legendre basis functions and their rotations.
Top two panels show Legendre basis functions as inten-
sity images up to order L = 20. Left and right panels
represent Pl共␽兲 and Pl共␽ − ␤兲. This example is for a
rotation angle ␤ = 30°. Bottom left panel shows the ap-
proximation error in Eq. 共23兲 due to truncation at L
= 20. Bottom right panel shows the corresponding ma-
trix Dll⬘ as defined in Eq. 共24兲. All four panels use the
same intensity gray scale.

f ⬇ Pc̃. 共27兲 xn共k兲 = gn共k,⍀兲 + wn共k兲. 共29兲

As with any pseudo-inverse, the approximation is exact if With the appropriate definition of matrix 关W共k兲兴qn = wn共k兲,
dim共f兲 = dim共c̃兲, while the square error is minimized for Eq. 共16兲 now becomes
dim共f兲 ⬎ dim共c̃兲. With this notation, the shift matrix D共␤兲 共G共k兲 + W共k兲兲B共k兲 = G̃. 共30兲
in Eq. 共24兲 is,
Note that the signal of the steady-state plane wave G共k兲 is
D共␤兲 = P†P⬘共␤兲, 共28兲 deterministic, while the noise W共k兲 is a random variable. To
find the optimal basis, one now has to minimize the expected
where matrix P⬘ captures the L Legendre polynomials evalu- value over the random noise:
ated at Q cosines shifted by ␤: 关P⬘共␤兲兴ql⬘ = Pl⬘共cos共␽q − ␤兲兲.
E关储共G共k兲 + W共k兲兲B共k兲 − G̃储2兴. 共31兲
For noise that is uncorrelated with the signal, this evaluates
VIII. EFFECTIVE BEAM CONTROL, MEASUREMENT 共omitting dependency on k兲 to the following trace:
NOISE, AND REGULARIZATION
Tr共BH共GHG + E关WHW兴兲B − BHGHG̃ − G̃HGB兲. 共32兲
When computing the basis B with Eq. 共17兲, it was re- This criterion is minimized by
quired that matrix G be of full rank. However, in practice,
the matrix GH共k兲G共k兲 is ill conditioned for the lowest fre- B = 共GHG + ⌺兲−1GHG̃, 共33兲
quencies and, therefore, cannot be accurately inverted when
computing the pseudo-inverse G†共k兲. This is to be expected which, as usual, is obtained by setting the derivative with
for very low frequencies with a large wavelength since a respect to BH equal to zero and solving for B. If the noise is
limited aperture prevents effective spatial resolution. Simi- spatially uncorrelated and homogeneous in space, then the
larly, for high frequencies, the finite spacing of sensors gen- N ⫻ N matrix ⌺共k兲 = E关W共k兲HW共k兲兴 is diagonal with powers
erates aliasing side lobes, once again leading to a noninvert- ␴2n共k兲 on the diagonal. This result is the conventional regu-
ible GH共k兲G共k兲 共recall the goal of using only a small number larization of the pseudo-inverse. The examples presented in
of sensors兲. This limitation is inherent to any beamforming Sec. XI will use this regularization.
design and is typically resolved by restricting the frequency
band of operation or carefully choosing sensor spacing and IX. APPLICATION TO ADAPTIVE BEAMFORMING
array aperture. In the present numerical inversion approach, DESIGN
the instability leads to rather large gains which may arbi-
trarily magnify sensor noise. This section shows how this The time domain output y共t兲 of the filter array in re-
problem can be resolved by considering the effect of mea- sponse to sensor readings xn共t兲 is given by the convolution
surement noise. and sum:
Assume zero-mean wide-sense stationary additive sen- N L
sor noise wn共t兲 with power spectrum ␴2共k兲. The sensor signal y共t兲 = 兺 cn共t兲 ⴱ xn共t兲 = 兺 c̃l共t兲 ⴱ x̃l共t兲. 共34兲
in response to a plane wave is now 共in the frequency domain兲 n=1 l=1

J. Acoust. Soc. Am., Vol. 119, No. 6, June 2006 Lucas C. Parra: Steerable frequency-invariant beamforming 3843
 The second equality here results from Eq. 共8兲 and the reproduces the desired response f. Equation 共38兲 does not
following definition of the frequency-invariant virtual sensor include rotation. Instead, rotation is introduced after de-
readings: termining the optimal coefficients for a desired response.
N With N sensors, one can satisfy, at most, N conditions on
x̃l共t兲 = 兺 bnl共t兲 ⴱ xn共t兲. 共35兲 the response. A larger number of conditions can only be
n=1 satisfied approximately. The coefficients that reproduce f
with the least-squares error are given by
This suggests that adaptive algorithms, that are driven
by the sensor observations, can be applied to the signals of c̃ = 共G共␪兲B兲†f. 共39兲
the newly defined frequency-invariant virtual sensors. In-
Without going into detail, one should note that it is also
stead of optimizing filter parameters, cn共k兲, based on sensor
useful to regularize this inverse, as in Eq. 共33兲.
readings, xn共t兲, the adaptive algorithm now adapts param- The Legendre basis uses L basis coefficients and can be
eters c̃l共k兲 based on virtual sensor readings, x̃l共t兲. All known rotated without having to recompute coefficients. One can
adaptive beamforming algorithms, such as generalized side- compare this with the coefficient obtained in a naive
lobe canceling, blind source separation, and others are, there-
frequency-invariant basis with 关G̃兴qm = ␦qm in Eq. 共17兲. This
fore, immediately applicable without further modification.
basis requires Q coefficients, which is typically significantly
To find an optimal frequency-invariant response, the
larger than L, and the coefficients have to be recomputed
adaptive algorithm will now optimize the parameters c̃l. If
when the response is to be rotated. For this “naive” basis, Eq.
the algorithm is based on a gradient of some cost function,
共39兲, simplifies to
J共兵c̃l共k兲其兲, with a set of frequency dependent parameters,
兵c̃l共k兲其, the frequency-invariant gradient is then simply the c̃ = 共G共␪兲G†兲†f. 共40兲
original gradient summed over all frequencies
In Eqs. 共38兲–共40兲, the dependence on frequency was
⳵J ⳵J
=兺
omitted for simplicity. The optimal parameters for each fre-
. 共36兲
⳵c̃l k ⳵c̃l共k兲 quency are computed with Eq. 共39兲 or 共40兲. For a frequency
invariant basis, one should use the same coefficients for all
Note that this has the potential to significantly reduce the frequency bands. One option is to use the coefficients com-
number of free parameters as the same coefficients are used puted with Eq. 共40兲 averaged across frequencies, c̃
for all frequencies. For most adaptive algorithms, this will = 1 / T兺kc̃共k兲. Though suboptimal, this approach is not only
result in significant improvements in convergence speed as simpler but in practice shows also better error behavior com-
well as estimation accuracy. This advantage is in addition to pared to the “optimal” solution, which would require com-
the potential advantage of a frequency-invariant response. bining all frequencies prior to computing the pseudo-inverses
in Eq. 共39兲 or 共40兲. Simulations show that this averaging of
X. APPLICATION TO IRREGULAR LINEAR ARRAY coefficients across frequencies results in more evenly distrib-
WITH LEAST-SQUARES BEAM DESIGN uted deviations from the desired solutions as compared to the
globally optimal solution.
The proposed method was implemented for a linear ar- Note that even after regularization, effective beam de-
ray of omnidirectional sensors and Legendre polynomials as sign is not possible at frequencies for which GHG is not
the beamforming basis, i.e., Eqs. 共13兲 and 共25兲. Notice that invertible. One measure for the instability of the inverse 共or
Eq. 共13兲 does not require equidistant sensor placement. The rank deficiency of G兲 is the condition number. One can use
least-squares solution given by Eq. 共17兲 is used to compute the condition number as a criterion to exclude frequency
the basis transform. Choosing Legendre polynomials, as in bands when computing the average c̃. The examples de-
Eq. 共25兲, as the frequency-invariant virtual array response in scribed below assume acoustic sensors with a sound propa-
Eq. 共17兲 means that gation speed of 342 ms−1. For an aperture of 10 cm, one
finds a useful frequency range 共with condition number
G̃ = P. 共37兲
⬍150 dB兲 of at least 100– 5000 Hz.
The effectiveness of the resulting basis is demonstrated
by estimating parameters c̃ that optimally reproduce a de-
XI. EXAMPLES
sired beam pattern. To this end, one could use a variety of
beam design methods.20 A simple data-independent method Figure 3 shows the results obtained for a linear array of
is the least-squares beamformer, which will be used here. omnidirectional sensors with irregular spacing and an aper-
Assume the prescribed response is specified as a vector f ture of 10 cm. Arbitrary spacing of a small number of sen-
with coefficients f d, each of which represents the desired sors 共N = 3 and N = 5兲 was used to highlight the advantage of
response for angles ␪d, d = 1 , . . . , D. The response of the ar- the present technique as compared to existing analytic meth-
ray at those angles can be written in matrix notation as ods, which typically require a larger number of sensors in a
regular arrangement.
f = G共␪兲Bc̃, 共38兲
The figure shows that frequency invariance is reasonably
where the coefficients of matrix G共␪兲 are given by 关G兴dn well maintained up to the Nyquist frequency of 5000 Hz
= gn共␪d兲 specifying the response of the nth sensor for the despite using only one set of coefficients 共instead of separate
desired angle ␪d. The goal is to find the coefficient c̃ that coefficients for each of the T / 2 + 1 frequency bins兲. Time

3844 J. Acoust. Soc. Am., Vol. 119, No. 6, June 2006 Lucas C. Parra: Steerable frequency-invariant beamforming
FIG. 3. 共Color online兲 Frequency-invariant design of null beam 共left兲 and main beam 共right兲. Top panels show desired responses. Each following row
compares results for different design parameters. First and third columns show time-domain filter coefficients cn共t兲. Second and third columns show magnitude
response as function of frequency k and angle ␽. Gray-scale represents response in dB. First row is obtained with a Legendre basis with five sensors assuming
−30 dB noise. The same parameters are used for the following rows except that: Second row assumes zero noise; Third row uses naive basis; Fourth row uses
only three microphones; Fifth row shifts the Legendre basis by 30° degrees.

domain coefficients were computed with T = 128, which re- In these examples, the Legendre basis included orders
sults in a reduction of the number of free parameters by a up to L = 20, and the basis transform was computed with Eq.
factor of 65. 共33兲 using Q = 200 equidistant samples ␽n. Compare this to

J. Acoust. Soc. Am., Vol. 119, No. 6, June 2006 Lucas C. Parra: Steerable frequency-invariant beamforming 3845
the third row showing the result with the naive basis. There planar array, analogous considerations to those of the linear
is no significant difference to the Legendre basis 共shown in array may be necessary to cope with the symmetry across the
the first row兲 despite the reduction of the number of param- array plane.
eters by a factor of Q / L = 10. The Legendre basis therefore The presentation in this paper was restricted to omnidi-
makes the array steerable and reduces the number of param- rectional sensors. However, the method applies equally well
eters without compromising accuracy. In practice, the order to sensors with directional response. If the sensor response is
of the Legendre basis should be chosen depending on the specified by r共k , ⍀兲, then g共k , ⍀兲 has to be replaced by
desired angular bandwidth. r共k , ⍀兲g共k , ⍀兲 everywhere.
White noise with a power of −30 dB was assumed, ex- Finally, the notation here considers only the far-field re-
cept in the second row where no regularization was used, i.e., sponse. One can generalize the argument to the near field by
the basis was computed with Eq. 共17兲. However, note that replacing plane waves of orientation ⍀ with point sources
when the condition number exceeded 150 dB, those fre- located at position r. In that case, the array response to a
quency bands were excluded regardless of the noise assump- planar wave, gn共k , ⍀兲, is to be replaced by the array response
tions. Notice the significant scale increase in the filter coef- to a spherical wave,9 gn共k , r兲 = eik·共r−rn兲 / 储共r − rn兲储, and r has
ficients indicating excessive low-frequency gain. Therefore, to be discretized over the desired range of point source po-
noise-based regularization should always be used in practice. sitions when computing B. The expansion for the spherical
In these examples, N = 5 sensors are located at rn = 0, 2, rather than a planar wave is given by Eq. 共11兲 whereby the
5, 7, and 10 cm. Compare this with the fourth row that has spherical Bessel function jl共krn⬘兲 is replaced by
only N = 3 sensors located at rn = 0, 5, and 10 cm. This com- jl共krn⬘兲 / jl共kr兲.9 The rationale leading to a steerable beam de-
parison indicates that the deviation from frequency invari- sign remains fully applicable.
ance is primarily due to the limited number of sensors. Not
surprisingly, as the number of sensors increases, frequency
ACKNOWLEDGMENTS
invariance is improved.
Finally, the last row demonstrates the effect of shifting This work was initially motivated by the work of Meyer
the Legendre basis by an angle of ␤ = 30° 共as in Fig. 2兲 and Elko on frequency-invariant spherical arrays,8 and the
leading to a corresponding shift of the beam pattern. need to develop adaptive frequency-invariant beams for
Simulations with different sensor locations 共not shown兲 acoustic source separation.22 I would also like to thank
indicate that the performance of the proposed method does Walter Kellermann for useful comments, Chris Alvino for his
not depend significantly on the specific arrangement of sen- careful review of multiple versions of this manuscript, and
sors. This is expected as the proposed method is guaranteed the anonymous reviewers of an earlier conference paper23
to make the best use of a fixed and known sensor configura- who highlighted the need to address the issue of noise and
tion. However, as with conventional beam-design methods, stability.
some sensor arrangements are better suited to minimize
1
aliasing and increase resolution. The present work did not D. B. Ward, R. A. Kennedy, and R. C. Williamson, “Theory and design of
aim to determine such optimal sensor locations. See, for in- broadband sensor arrays with frequency invariant far-field beam patterns,”
J. Acoust. Soc. Am. 97, 1023–1034 共1995兲.
stance, Ref. 21 for a modern technique to optimize locations. 2
T. Abhayapala, R. Kennedy, and R. Williamson, “Nearfield broadband
Similarly, the sensitivity to errors in sensor position is array design using a radially invariant modal expansion,” J. Acoust. Soc.
comparable to conventional beam design. Simulations on the Am. 107, 392–403 共2000兲.
3
examples above 共not shown兲 indicate only minor effects for R. Kennedy, T. Abhayapala, and D. Ward, “Broadband near-field beam-
forming using a radial beampattern transformation,” IEEE Trans. Signal
location errors of about 1 mm 共⬇5% of the microphone Process. 46, 2147–2156 共1998兲.
spacing兲 but a significant deterioration for larger position er- 4
W. Liu and S. Weiss, “A new class of broadband arrays with frequency
rors. As with conventional beam design, it is preferable to invariant beam patterns,” in Proceedings of the International Conference
use adaptive rather than data-independent methods as they on Acoustics, Speech, and Signal Processing 共IEEE, New York, 2004兲,
Vol. 2, pp. 185–188.
can adapt to position errors. Section IX outlined how to 5
D. B. Ward, R. A. Kennedy, and R. C. Williamson, “FIR filter design for
implement adaptive design methods using the frequency- frequency invariant beamformers,” IEEE Signal Process. Lett. 3, 69–71
invariant basis. 共1996兲.
6
T. Sekiguchi and Y. Karasawa, “Wideband beamspace adaptive array uti-
lizing FIR fan filters for multibeam forming,” IEEE Trans. Signal Process.
XII. CONCLUSION 48, 277–284 共2000兲.
7
H. Teutsch and W. Kellermann, “EB-ESPRIT: 2D Localization of multiple
The previous section demonstrated the proposed method wideband acoustic sources using eigenbeams,” in Proceedings of the In-
ternational Conference on Acoustics, Speech, and Signal Processing
on a linear array. The implementation for a volumetric array 共IEEE, New York, 2005兲, Vol. 3, pp. 89–92.
is straightforward using definition 共4兲 or expansion 共11兲 for 8
J. Meyer and G. Elko, “A highly scalable spherical microphone array
based on an orthonormal decomposition of the soundfield,” in Proceedings
G, and using spherical harmonics 共19兲 to define matrix G̃
of the International Conference on Acoustics, Speech, and Signal Process-
analogous to matrix P. The linear array was chosen here ing 共IEEE, New York, 2002兲, Vol. 2, pp. 1781–1784.
9
because the resulting frequency-invariant beam patterns are E. G. Williams, Fourier Acoustics 共Academic, New York, 1999兲.
10
easier to visualize on paper, and because steering required A. Edmonds, Angular Momentum in Quantum Mechanics 共Princeton Uni-
versity Press, Princeton, N.J., 1957兲.
special consideration as a result of the restricted linear ge- 11
Z. Li and R. Duraiswami, “A robust and self-reconfigurable design of
ometry. No such complication should arise for a volumetric spherical microphone array for multi-resolution beamforming,” in Pro-
array since the Wigner rotations in Eq. 共18兲 are exact. For a ceedings of the International Conference on Acoustics, Speech, and Signal

3846 J. Acoust. Soc. Am., Vol. 119, No. 6, June 2006 Lucas C. Parra: Steerable frequency-invariant beamforming
 Processing 共IEEE, New York, 2005兲, Vol. 4, pp. 1137–1140. of the Workshop on Applications of Signal Processing to Audio and Acous-
12
Z. Li and R. Duraiswami, “Hemispherical microphone arrays for sound tics 共IEEE, New York, 2005兲, pp. 150–153.
capture and beamforming,” in Proceedings of the Workshop on Applica- 18
P. N. Swarztrauber, “The spectral approximation of discrete scalar and
tions of Signal Processing to Audio and Acoustics 共IEEE, New York, vector functions on the sphere,” SIAM 共Soc. Ind. Appl. Math.兲 J. Numer.
2005兲, pp. 106–109. Anal. 16, 934–949 共1979兲.
13
S. Chan and C. Pun, “On the design of digital broadband beamformer for 19
P. N. Swarztrauber and W. F. Spotz, “Spherical harmonics projectors,”
uniform circular array with frequency invariant characteristics,” in Pro- Math. Comput. 73, 753–760 共2003兲.
ceedings of the International Symposium on Circuits and Systems 共IEEE, 20
B. Van Veen and K. Buckley, “Beamforming techniques for spatial filter-
New York, 2002兲, Vol. 1, pp. 693–696.
14 ing,” Digital Signal Processing Handbook 共CRC Press, New York, 1997兲,
E. Wigner, Gruppentheorie und ihre Anwendungen auf die Quanten-
pp. 61.1–61.20.
mechanik der Atomspektren (Group theory and its applications to the 21
S. Blank and M. Hutt, “On the empirical optimization of antenna arrays,”
quantum mechanics of atomic spectra) 共Friedr. Vieweg, and Sohn, Berlin,
1931兲. IEEE Antennas Propag. Mag. 47, 58–67 共2005兲.
22
15
C. H. Choi, J. Ivanic, M. S. Gordon, and K. Ruedenberg, “Rapid and W. Liu and D. Mandic, “Semi-blind source separation for convolutive
stable determination of rotation matrices between spherical harmonics by mixtures based on frequency-invariant transformation,” in Proceedings of
direct recursion,” J. Chem. Phys. 111, 8825–8831 共1999兲. the International Conference on Acoustics, Speech, and Signal Processing
16
R. Jakob-Chien and B. K. Alpert, “A fast spherical filter with uniform 共IEEE, New York, 2005兲.
23
resolution,” J. Comput. Phys. 136, 213–230 共1997兲. L. C. Parra, “Least squares frequency invariant beamforming,” in Pro-
17
R. Duraiswami, Z. Li, D. Zotkin, E. Grassi, and N. Gumerov, “Plane-wave ceedings of the Workshop on Applications of Signal Processing to Acous-
decomposition analysis for spherical microphone arrays,” in Proceedings tics and Audio 共IEEE, New York, 2005兲.

J. Acoust. Soc. Am., Vol. 119, No. 6, June 2006 Lucas C. Parra: Steerable frequency-invariant beamforming 3847