Anda di halaman 1dari 122

SPRINGER BRIEFS IN ELEC TRIC AL AND

COMPUTER ENGINEERING  SIGNAL PROCESSING

Chiong Ching Lai
Sven Erik Nordholm
Yee Hong Leung

A Study into
the Design
of Steerable
Microphone
Arrays

123
SpringerBriefs in Electrical and Computer
Engineering

Signal Processing

Series editors
Woon-Seng Gan, Singapore, Singapore
C.-C. Jay Kuo, Los Angeles, USA
Thomas Fang Zheng, Beijing, China
Mauro Barni, Siena, Italy
More information about this series at http://www.springer.com/series/11560
Chiong Ching Lai Sven Erik Nordholm

Yee Hong Leung

A Study into the Design


of Steerable Microphone
Arrays

123
Chiong Ching Lai Yee Hong Leung
Department of Electrical and Computer Department of Electrical and Computer
Engineering Engineering
Curtin University Curtin University
Perth, WA Perth, WA
Australia Australia

Sven Erik Nordholm


Department of Electrical and Computer
Engineering
Curtin University
Perth, WA
Australia

ISSN 2191-8112 ISSN 2191-8120 (electronic)


SpringerBriefs in Electrical and Computer Engineering
ISSN 2196-4076 ISSN 2196-4084 (electronic)
SpringerBriefs in Signal Processing
ISBN 978-981-10-1689-9 ISBN 978-981-10-1691-2 (eBook)
DOI 10.1007/978-981-10-1691-2

Library of Congress Control Number: 2016944176

© The Author(s) 2017


This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained
herein or for any errors or omissions that may have been made.

Printed on acid-free paper

This Springer imprint is published by Springer Nature


The registered company is Springer Science+Business Media Singapore Pte Ltd.
Preface

The book aims to provide discussions on the design of robust steerable broadband
beamformer, from modelling (signal source, acoustic environment and sensor array)
to designing beamformer weights to achieve desired response. The focus of this
book is on nearfield-only, farfield-only, mixed nearfield–farfield, fixed and steerable
robust broadband beamformer designs.
This book has been structured such that each subsequent chapter extends the
previous chapter to provide an additional feature. The technical discussion starts
from Chap. 2, which provides discussions on the signal source models, acoustic
environments and sensor arrays. Chapter 3 starts to discuss design formulations for
the fixed broadband beamformer, including the beamformer structure used. This
provides a good starting point as fixed broadband beamformer design is simple and
easy to understand without the complicated formulation to include beam steering
and robustness. Chapter 4 extends the design formulation in the previous chapter to
include beam steering capability. The formulation in Chap. 4 is later extended in
Chap. 5 to include robustness against practical mismatches and errors. The for-
mulations in this chapter encapsulate all the properties from the chapters before it.
Readers will be able to understand and possibly design robust steerable beam-
former after reading this book. Interested readers can further refer to the references
cited for detailed discussion on specific beamformer topics not covered by this
book.

Australia Chiong Ching Lai


October 2015 Sven Erik Nordholm
Yee Hong Leung

v
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Beamforming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Practical Applications. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Chapter Organisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Acoustic Environment, Source Models and Sensor
Arrays Theory. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Environment and Channel Modelling . . . . . . . . . . . . . . . . . . . . . 8
2.3 Source Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.1 Distributed and Point Source . . . . . . . . . . . . . . . . . . . . . 10
2.3.2 Nearfield and Farfield Source . . . . . . . . . . . . . . . . . . . . . 11
2.4 Sensor Arrays Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4.1 Spatial Aliasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4.2 Array Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5 Spiral Arm Array Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5.1 Ring Radii . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5.2 Twist Angle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3 Broadband Beamformer Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Beamformer Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2.1 Weight-and-Sum Beamformer . . . . . . . . . . . . . . . . . . . . . 28
3.2.2 Filter-and-Sum Beamformer . . . . . . . . . . . . . . . . . . . . . . 31
3.3 Design Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.1 Weighted LS Formulation . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.2 Weighted TLS Formulation . . . . . . . . . . . . . . . . . . . . . . 36

vii
viii Contents

3.4 Mixed Nearfield–Farfield Design Formulation . . . . . . . . . . . . . . . 38


3.5 Performance Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.6 Design Examples and Evaluation. . . . . . . . . . . . . . . . . . . . . . . . 40
3.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4 Steerable Broadband Beamformer Design. . . . . . . . . . . . . . . . . . . . 53
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2 Beamformer Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.3 Design Formulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.3.1 Weighted LS Formulation . . . . . . . . . . . . . . . . . . . . . . . 57
4.3.2 Weighted TLS Formulation . . . . . . . . . . . . . . . . . . . . . . 60
4.4 Mixed Nearfield–Farfield Design Formulation . . . . . . . . . . . . . . . 61
4.5 Steering Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.5.1 Steering Range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.5.2 Normalisation of Steering Function . . . . . . . . . . . . . . . . . 64
4.6 Performance Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.7 Design Examples and Evaluation. . . . . . . . . . . . . . . . . . . . . . . . 65
4.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5 Robust Formulation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.2 Conventional White Noise Gain Constraint . . . . . . . . . . . . . . . . . 79
5.3 Stochastic Error Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.3.1 Multiplicative Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.3.2 Additive Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.3.3 Multiplicative and Additive Error . . . . . . . . . . . . . . . . . . 85
5.4 Robust Formulation Using Stochastic Error Model. . . . . . . . . . . . 85
5.4.1 Weighted LS Formulation . . . . . . . . . . . . . . . . . . . . . . . 85
5.4.2 Weighted TLS Formulation . . . . . . . . . . . . . . . . . . . . . . 86
5.5 Performance Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.6 Design Examples and Evaluation. . . . . . . . . . . . . . . . . . . . . . . . 87
5.6.1 Design Specifications. . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.6.2 Array Gain and Sensitivity . . . . . . . . . . . . . . . . . . . . . . . 88
5.6.3 Perturbation in Sensor Characteristics . . . . . . . . . . . . . . . 89
5.6.4 Perturbation in Sensor Positions . . . . . . . . . . . . . . . . . . . 91
5.6.5 Perturbation Due to Local Scattering . . . . . . . . . . . . . . . . 95
5.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
Contents ix

6 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103


6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
6.3 Final Remarks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

Appendix A: Closed Form Integration for Fixed


Beamformer Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

Appendix B: Closed Form Integrations for Steerable


Beamformer Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Acronyms

DSP Digital signal processing


FIR Finite impulse response
FT Fourier transform
IFT Inverse Fourier transform
LS Least squares
LTI Linear time invariant
PDF Probability density function
SBBF Steerable broadband beamformer
TLS Total least squares
WNG White noise gain
w.r.t With respect to

xi
Chapter 1
Introduction

Abstract Beamforming is a signal processing technique that is used in conjunction


with an array of sensors to provide useful spatial filtering. Beamforming requires
processing the data collected over a spatial aperture in order to achieve spatial dis-
crimination of received signals. This spatial discrimination serves as an additional
degree of signal separation that can be used with other signal processing techniques
for performance improvement. As a result, beamforming has been applied in wide
variety of application fields such as communication, radio astronomy, biomedical,
imaging, geophysical exploration, navigation, radio detection and ranging.

Keywords Beamforming · Spatial filtering

1.1 Beamforming

Signals received by arrays of sensors located in space can be filtered either construc-
tively or destructively to achieve spatial selectivity. Such spatial filtering technique
is called beamforming, which aims to form beams towards a desired direction in
space in order to receive signal radiating from that direction while attenuating sig-
nals from other directions [1–3]. As such, beamforming is often applied to separate
signals that are overlapping in spectral domain but originate from different spatial
locations. Since beamforming operates on the spatial domain, it requires processing
the data collected over a spatial aperture. As such, beamformer are always used in
conjunction with an array of sensors distributed in space to provide a versatile form
of spatial filtering.
The major advantage of beamforming is that it provides spatial discrimination or
directivity of received signals, an additional degree of separation that can be used
together with other signal processing techniques. For example, signals originating
from different spatial locations but occupying same spectral band that otherwise
cannot be separated using temporal filtering can still be separated using beamforming.
In addition, the spatial discrimination of beamforming also allows optimised signal
processing to be selectively applied to the spatially separated signals. This divide-
and-conquer approach can be applied for demanding signal processing requirements
which are otherwise very complicated to solve.

© The Author(s) 2017 1


C.C. Lai et al., A Study into the Design of Steerable Microphone Arrays,
SpringerBriefs in Signal Processing, DOI 10.1007/978-981-10-1691-2_1
2 1 Introduction

One interesting aspect of beamforming is that its main beam can be made steerable
to cater for spatially moving signal source. The steerability in this context means that
the main beam can be steered electronically without any mechanical movement of the
sensor array. The steering is normally achieved by using a steering parameter without
the need to redesign the beamformer every time the steering angle changes. Hence,
a steerable beamformers can provide a dynamic response, as opposed to a fixed
beamformer, where its main beam can be steered on-the-fly after the beamformer is
deployed to the field. Steerable beamformers have been shown to be useful in various
applications such as wireless communications and audio communications [4–7].

1.2 Practical Applications

Beamforming has found its way into wide range of practical applications, ranging
from day-to-day applications to specialised fields. In day-to-day applications, beam-
forming using microphone arrays has become a common technology used in end-
user electronic devices. Modern audio conferencing, either through dedicated audio
conferencing devices or through personal computers, employs microphone array
beamforming for optimised speech enhancement, and noise and echo suppression.
This provides better speech quality and user experience for conference participants
than single microphone systems [8, 9]. For speech-activated commands in gaming
consoles, mobile phones and smart televisions, microphone array beamforming is
utilised to improve accuracy and robustness of the command triggering, including in
noisy environments [10]. In professional audio recordings, microphone arrays have
been used to provide high fidelity surround sound recordings and reproductions for
audio entertainment such as live orchestras, concerts and surround audio for movies
[11–13].
In biomedical field, beamforming is used in hearing aids, where the main beam is
normally formed towards the front, while a null is placed at the back of a patient. In
foetal heart monitoring system, an array of ultrasonic transducers is used to form a
beam localised towards the foetal heart. This can improve the accuracy of the moni-
toring system as noises picked up from other physiological sources such as maternal
aorta and movement of the foetus can be suppressed [14]. For cancer treatment,
non-invasive microwave beamforming is used for localised selective hyperthermia
treatment or heat-activated chemotherapeutic drug release [15]. In medical ultrasound
imaging, the imaged medium is insonified with focused beams. The backscattered
echoes are then beamformed to eliminate the contribution of signals backscattered by
other structures off the imaging beam. Recent development utilises more advanced
beamforming techniques to improve resolution, contrast and depth penetration with-
out sacrificing its lateral resolution [16, 17].
Geophysical imaging and exploration have also exploited the use of beamform-
ing technology. Seismic beamforming with beam steering has been used as high-
resolution tool for mapping earth’s subsurfaces, which can be used for mineral explo-
ration [18]. For mapping and monitoring earth surface, beamforming-based Synthetic
1.2 Practical Applications 3

Aperture Radar (SAR) technology is used to provide high-resolution, automated aer-


ial imaging of remote areas [19, 20].
Radio detection and ranging (RADAR) application, which is vital in air traffic
control and military, has also seen significant improvement due to beamforming
technology. Phased array radar uses the concept of beamforming to electronically
excite and radiate planar wavefront to desired directions from an array of elementary
antenna with omnidirectional characteristics. On the receiving path, beamforming is
again applied to coherently sum all the received signals. Unlike conventional radar
with constant mechanical movement, beamforming in phased array radar allows for
fast switching of the look direction, search and track multiple target using a single-
phased array radar [21]. Advanced beamforming techniques in radar application have
been an active research interest [22–25]. Likewise, beamforming also plays a key role
in underwater sound navigation and ranging (SONAR), which share similar concept
as RADAR applications [26–28].
Beamforming has also become an indispensable technology in radio astron-
omy. Small omnidirectional antennas distributed in different locations are connected
together to form a large radio telescope array such as the Square Kilometre Array
(SKA) and LOw-Frequency ARray (LOFAR). Beamforming is then applied in real
time to the signal received in order to form a beam towards a particular direction
in space to sweep or survey the interstellar space [29–31]. Beamforming in radio
astronomy has shown to improve angular resolution and signal-to-noise ratio which
is otherwise limited by the size of the antenna in a single-antenna radio telescopes.
It is evident that applications of beamforming technology are already very wide-
spread and the examples above are only a small part of it. It is not surprising to see
the interest for beamforming technology to expand and pioneer into new-found areas
of applications. With no loss of generality, in this book, we will focus on microphone
arrays where the emitted acoustic signals are generally broadband.

1.3 Chapter Organisation

Preliminary design decisions on the physical attributes of beamformers are discussed


in Chap. 2. These include the selection of array geometries, signal propagation mod-
els and source models. In general, such design decisions depend heavily on target
applications and the acoustics of the operating environment. There is no single global
solution that works for all applications and environments. Each model and structure
has its own merits and drawbacks. Accordingly, care must be exercised when making
these choices in order to avoid known limitations of certain models or structures and
to maximise the overall performance of beamformers.
Chapter 3 presents the design formulations of fixed beamformers in both weighted
least squares (LS) and weighted total least squares (TLS) sense. The design formu-
lations cover all three types of beamformer, namely nearfield-only, farfield-only
and mixed nearfield–farfield beamformers. Mixed nearfield–farfield beamformer is
operable for both nearfield and farfield sources. The trade-off for achieving such
4 1 Introduction

operability is a loss in performance relative to nearfield-only and farfield-only beam-


formers.
In Chap. 4, the fixed beamformer design formulations are extended to allow beam
steering. Beam steering is achieved by utilising polynomial or Farrow filter structure.
The structure allows the steering angle to be used directly as the beam steering
parameter for steering the main beam on-the-fly after beamformer weights have
been selected. Although the beam steering feature increases the design problem
size, appropriate selection of array geometry and the steering function can keep the
problem size manageable.
Chapter 5 further extends the beamformer design formulations to include robust-
ness towards practical errors and mismatches. Such errors and mismatches are mod-
elled as random variables where their stochastic properties are captured. This allows
the errors and mismatches to be linked to robustness, where the designs are then opti-
mised based on the mean performance. This stochastic model is incorporated into
beamformer design formulations in such a way that it still follows the conventional
design procedure.
Finally, conclusions are drawn in Chap. 6 which provides the summary of the
discussion. Future research directions are also discussed based on the materials pre-
sented in the previous chapters. These include analysis on different steering func-
tions, beamformer designs using different optimisation methods, incorporating track-
ing capability into steerable beamformers and optimisation of sensor placement for
given applications.

References

1. B.D. Van Veen, K.M. Buckley, Beamforming: a versatile approach to spatial filtering. IEEE
Signal Process. Mag. 5(2), 4–24 (1988)
2. D.H. Johnson, D.E. Dudgeon, Array Signal Processing—Concepts and techniques (Prentice
Hall, 1993)
3. S. Nordholm, H. Dam, C. Lai, E. Lehmann, Broadband Beamforming and Optimization, in
Academic Press Library in Signal Processing: Array and Statistical Signal Processing, ed. by
A.M. Zoubir, Vol. 3 (Massachusetts: Elsevier, 2014), pp. 553–598
4. C. Sun, A. Hirata, T. Ohira, N.C. Karmakar, Fast beamforming of electronically steerable
parasitic array radiator antennas: theory and experiment. IEEE Trans. Antennas Propag. 52(7),
1819–1832 (2004)
5. L.C. Parra, Steerable frequency-invariant beamforming for arbitrary arrays. J. Acoust. Soc.
Am. 119(6), 3839–3847 (2006)
6. C.C. Lai, S. Nordholm, Y.H. Leung, Design of robust steerable broadband beamformers with
spiral arrays and the farrow filter structure, in Proceedings of the International Workshop
Acoustics, Echo, Noise Control, Tel Aviv, Israel, 30 Aug–2 Sep 2010
7. C.C. Lai, S. Nordholm, Y.H. Leung, Design of steerable spherical broadband beamformers
with flexible sensor configurations. IEEE Trans. Audio, Speech, Lang. Process. 21(2), 427–
438 (2013)
8. M. Brandstein, D. Wards (eds.), Microphone arrays—Signal processing techniques and appli-
cations (Springer, Berlin, 2001)
9. J. Benesty, J. Chen, Y. Huang, Microphone Array Signal Processing, vol. 1 (Springer Science
& Business Media, 2008)
References 5

10. K. Chan, S. Low, S. Nordholm, K. Yiu, S. Ling, Speech recognition enhancement using beam-
forming and a genetic algorithm, in Proceedings of the International Conference on Network
and System Security, Oct 2009, pp. 510–515
11. A. Fukada, A challenge in multichannel music recording, in Audio Engineering Society Con-
ference: 19th International Conference: Surround Sound—Techniques, Technology, and Per-
ception, 2001
12. R. Glasgal, Ambiophonics. achieving physiological realism in music recording and reproduc-
tion, in Audio Engineering Society Convention, vol. 111, 2001
13. A. Farina, R. Glasgal, E. Armelloni, A. Torger, Ambiophonic principles for the recording and
reproduction of surround sound for music, in Audio Engineering Society Conference: 19th
International Conference: Surround Sound—Techniques, Technology, and Perception, 2001
14. R. Hoctor, K. Thomenius, Method and apparatus for non-invasive ultrasonic fetal heart rate
monitoring, Patent, Dec 2008, US Patent 7,470,232
15. E. Zastrow, S. Hagness, B. Van Veen, J. Medow, Time-multiplexed beamforming for noninva-
sive microwave hyperthermia treatment. IEEE Trans. Biomed. Eng. 58(6), 1574–1584 (2011)
16. F. Vignon, M. Burcher, Capon beamforming in medical ultrasound imaging with focused beams.
IEEE Trans. Ultrason. Ferroelectr. Freq. Control 55(3), 619–628 (2008)
17. J.F. Synnevag, A. Austeng, S. Holm, Benefits of minimum-variance beamforming in medical
ultrasound imaging. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 56(9), 1868–1879 (2009)
18. J. Guign, A. Stacey, C. Clements, S. Azad, A. Pant, A. Gogacz, W. Hunt, N. Pace, Acoustic
zoom high-resolution seismic beamforming for imaging specular and non-specular energy of
deep oil and gas bearing geological formations. J. Nat. Gas Sci. Eng. 21, 568–591 (2014)
19. A. Patyuchenko, C. Tienda, M. Younis, S. Bertl, P. Lopez-Dekker, G. Krieger, Digital beam-
forming sar interferometer based on a multi-beam reflectarray antenna, in Proceedings of the
European Conference on Synthetic Aperture Radar, June 2014, pp. 1–4
20. W.Q. Wang, Q. Peng, J. Cai, Digital beamforming for near-space wide-swath sar imaging, in
International Symposium Antennas, Propagation and EM Theory, Nov 2008, pp. 1270–1273
21. U. Nickel, Fundamentals of signal processing for phased array radar, in Research Institute for
High-Frequency Physics and Radar Techniques (FHR), 2006
22. Z. Li, R. Duraiswami, Flexible and optimal design of spherical microphone arrays for beam-
forming. IEEE Trans. Audio Speech Lang. Process. 15(2), 702–714 (2007)
23. D. Fuhrmann, G. San Antonio, Transmit beamforming for mimo radar systems using signal
cross-correlation. IEEE Trans. Aerosp. Electron. Syst. 44(1), 171–186 (2008)
24. A. Hassanien, S. Vorobyov, Transmit/receive beamforming for mimo radar with colocated
antennas, in IEEE Conference on Acoustics, Speech and Signal Processing, Apr 2009, pp.
2089–2092
25. S. Cao, Y. Zheng, R. Ewing, Transform sensing of phased array radar, in IEEE Radar Confer-
ence, May 2014, pp. 1143–1148
26. J. Thorner, Approaches to sonar beamforming, in IEEE Proceedings of the Southern Tier
Technical Conference, Apr 1990, pp. 69–78
27. S. Kogon, Robust adaptive beamforming for passive sonar using eigenvector/beam associa-
tion and excision, in Proceedings of the Sensor Array and Multichannel Signal Processing
Workshop, Aug 2002, pp. 33–37
28. T. Zhou, S. Li, H. Li, L. Yin, Application of aperture extrapolation beamformer in multibeam
bathymetric sonar, in IEEE Conference on Signal Processing, Oct 2010, pp. 2349–2352
29. A. Faulkner, P. Alexander, J. de Vaate, System design for ska capable aperture arrays, in
International Conference Electromagnetics in Advanced Applications (ICEAA), Sept 2012,
pp. 752–755
30. J.D. Mol, J.W. Romein, The LOFAR beam former: implementation and performance analysis,
in EuroPar’, vol. LNCS, vol. 6853, Part II, Bordeaux, France, Aug 2011, pp. 328–339
31. G. Hovey, T. Burgess, B. Carlson, Z. Ljusic, B. Veidt, H. Zhang, A broadband fpga digital beam-
former for the advanced focal array demonstrator (afad), in General Assembly and Scientific
Symposium, Aug 2011, pp. 1–4
Chapter 2
Acoustic Environment, Source Models
and Sensor Arrays Theory

Abstract In the simplest form, a signal propagating from one point to another
undergoes signal filtering from the propagation medium. Beamforming is a signal
processing method to undo this filtering, such that the desired signal is retained at
the receiver end, in addition to possibly suppression of unwanted signals. Hence,
the underlying acoustic environment of the target application needs to be studied
and modelled before a beamformer can be designed. This chapter discusses the fun-
damentals of acoustic environment modelling including different models for signal
sources, propagation mediums and sensor arrays.

Keywords Acoustic modelling · Microphone array · Array geometry

2.1 Introduction

A signal propagating from one spatial location to another, where it is then observed,
undergoes magnitude attenuation and time delay. This attenuation and time delay
can be considered as filtering by the propagation medium. In the simplest form,
a beamformer is usually used to undo or equalise this medium filtering such that
the original signal can be extracted, with possibly further suppression of unwanted
interferences and noise, as shown in Fig. 2.1. Of course, this concept can be easily
extended to design advanced beamformers such as beamformers with steerable main
beam and beamformers that works for both nearfield and farfield sources. Never-
theless, the design of beamformers consists of two parts: (a) modelling of acoustic
environment and (b) designing the actual beamformers. This chapter provides the
study for acoustic environment modelling, including the models for signal sources,
propagation mediums and array geometries. All these models have direct influences
in the design process of beamformer weights in the subsequent chapters.

© The Author(s) 2017 7


C.C. Lai et al., A Study into the Design of Steerable Microphone Arrays,
SpringerBriefs in Signal Processing, DOI 10.1007/978-981-10-1691-2_2
8 2 Acoustic Environment, Source Models and Sensor Arrays Theory

Interference source

Moving sound
source

Sensor
array

...

Background noise
Beamforming

Desired output

Interference source

Fig. 2.1 General beamformer system

2.2 Environment and Channel Modelling

When a signal propagates from point A to point B in an open space, it is actu-


ally travelling through a complex time-dependent medium governed by its physical
properties including temperature, viscosity, density, presence of foreign particles and
much more. In fact, between point A and B, there is no guarantee that those physical
quantities remain constant over time. The underlying mathematics becomes much
more complicated when both points A and B are in an enclosed space with foreign
objects, for example, an enclosed room with furnitures. In this case, the signal as
observed at Point B not only consists of the direct propagation path from point A, but
also the reflections from the wall of the enclosure as well as the scattering from the
foreign objects in the enclosure. If the line of sight between A and B is obstructed,
there will not be any direct path from A to B. From the signal processing point of
view, each signal propagation path from point A to B can be modelled as separate,
corresponding signal processing paths as shown in Fig. 2.2. The purpose of acoustic
environment and channel modelling is to identify and characterise these paths in
terms of transfer functions. Detailed acoustic environment and channel modelling
can be found in [1, 2].
2.2 Environment and Channel Modelling 9

(a)

th 2
ted pa B
Reflec

ath
ect p
Dir

Reflected path 1
Foreign
A
object

Room enclosure

(b)
Reflected path 1

Signal at
Direct path
point B
Signal at
point A

Reflected path 2

Fig. 2.2 Signal propagation through a medium. a Signal propagation in an enclosure. b Signal
propagation block diagram

For the purpose of this book, an isotropic, homogeneous, non-dispersive, time-


invariant propagation medium without any reflection nor reverberation is considered.
Hence, the signal propagating from point A to B undergoes a linear, time-invariant
(LTI) process as shown in Fig. 2.3. Signal as observed at any other points are also
described by the same system. Mathematically, for a signal source represented in
frequency domain as S0 (ω) and located at an arbitrary point r in space, the signal as
observed at any point rk is given by

X (rk , ω) = A (r, rk , ω) S0 (ω) . (2.1)


10 2 Acoustic Environment, Source Models and Sensor Arrays Theory

LTI system
Signal at Signal at
point A point B

Fig. 2.3 Signal propagation in an isotropic, homogeneous, non-dispersive, time-invariant medium


modelled as a LTI system

where A (·) is the transfer function describing the signal propagation from point r to
point rk in the modelled propagation medium, and ω = 2π f , with f the frequency
of the signal. In array signal processing, A (·) is also called array response. Note
that (2.1) is only possible if the propagation medium is a LTI system or slowly time
varying linear system. A more elaborate expression which includes solving the wave
equations for a general propagation medium can be found in [3].

2.3 Source Models

A source is a physical device which generates energy, in this case sound energy, to
be transmitted through a propagation medium. From signal processing point of view,
it is the origin of an excitation in a system. In beamforming, the signal source needs
to be appropriately modelled since its mathematical model forms part of the array
response.

2.3.1 Distributed and Point Source

For an arbitrarily shaped signal source occupying a volume in space as shown in


Fig. 2.4, the sound energy from the source will generate vibration on its surface. This
vibration transmits the sound energy from the source’s surface to the particles of the
propagation medium that are in contact with the source surface. This sound energy
or vibration will then propagate through the medium, which is normally indicated
by pressure waves before they reach a receiver or sensor. The amount of energy
transmitted from the source to the propagation medium depends on the intensity
of the vibration on the source’s surface and loss or dispersion of the medium. In
general, the intensity is not uniformly distributed throughout the source’s surface
and the mathematical model for such source requires solving for complex wave
equations [1, 3].
However, the mathematics can be simplified significantly if the physical size of
the source is reduced to infinitesimally small, such that it only occupies a point in
space. Due to its very small size, the sound intensity on its surface can be regarded
as constant and its shape as a sphere. This source is called a point source and it
has concentric, spherical wave fronts as shown in Fig. 2.5. Due to its much simpler
2.3 Source Models 11

Non-constant sound intensity


throughout source's surface
y

Body of source
x
Fig. 2.4 Cross-sectional view of distributed source model

Wave propagation with


constant intensity

Point source
x

Fig. 2.5 Cross-sectional view of point source model

mathematical model, point source model is widely used in array signal processing.
With the concept of point source, it is possible to model approximately a complex
source by sampling its surface and treat each spatial point as a point source.

2.3.2 Nearfield and Farfield Source

For a point source located in space, the wave fronts that radiate outwards from it
are spherical in shape. Hence, when signal observation is performed close to this
signal source, such curvature in these wave fronts that impinge onto the observation
12 2 Acoustic Environment, Source Models and Sensor Arrays Theory

sensors need to be accounted for. Mathematically, for a point source propagating


from a point r = (r, φ) in cylindrical coordinate system, and travelling through a
homogeneous, isotropic, non-dispersive and time-invariant medium (see Fig. 2.6),
the signal received at a sensor (or receiver) is phase delayed and attenuated by
its propagation response. This response, also known as array response or Green’s
function, is given by [4]

1  ω 
A (r, rk , ω) = exp −j rk − r (2.2)
4π rk − r c

where rk = (rk , φk ) is the position of the kth sensor, and · denotes Cartesian dis-
tance. The attenuation in (2.2) is due to the decay of signal amplitude as it propagates
outwards from its source. The constant 4π can be dropped for convenience since only
the relative gain and phase difference between the sensors are important.
Although (2.2) gives a generic frequency response from an arbitrarily located
signal source to an arbitrarily located sensor, the nonlinear Cartesian distance in the
equation may complicate beamformer designs. A simplified source model can be
obtained by considering the source to be at infinite distance away from the sensor
array, i.e. r → ∞. The reason for this is that when the source is far enough from
the sensor array, the wave front impinging on the array becomes planar (as opposed
to spherical), which can simplify the propagation model. However, in this farfield

P oint source

rk − r

rk  r
y Wavefront
propagation from
x
a nearfield source
Sensor

Arbitrary
sensor array

Fig. 2.6 Nearfield source model


2.3 Source Models 13

source model, a reference point is required and is normally taken as the origin of the
coordinate system or the centre of mass of the sensor array. The received signal at
each sensor is then modelled relative to this reference point. For the reference point
taken as the coordinate system’s origin, the response of the kth sensor is then given
by

r  ω 
A0 (r, rk , ω) = exp −j (rk − r − r) . (2.3)
rk − r c

Observe that as r → ∞,


r
lim =1 (2.4)
r→∞ rk− r

and

lim rk − r − r = rk  cos (φk − φ)


r→∞
= rk · r̂ (2.5)

where r̂ = rr
is the unit vector or normalised source position. The array response
for a farfield source is thus given by [4]

Afar (rk , ω, r) = lim A0 (rk , ω, r)


r→∞
 ω 
= exp −j rk · r̂ . (2.6)
c
For common array geometries such as uniform linear array and circular array (shown
in Fig. 2.7), their farfield array responses are, respectively,
 ω 
Alin (φ, k, ω) = exp −j kd cos (φ) (2.7)
 ωrc 
k
Acir (φ, k, ω) = exp −j cos (φ − φk ) . (2.8)
c
Although farfield source model gives a simpler expression for the array response
(c.f. (2.2) with (2.6)), the model is only valid for open space environment (e.g. free-
field) and not for enclosed environment, such as small rooms due to the requirement
r → ∞. However, studies have shown that in practice, r → ∞ is not a strict
requirement for (2.6) to be valid. For example, Eq. (2.6) may still be valid for medium-
sized office and indoor stadium, depending on the array size. One of the widely used
quantitative lower bounds as a practical criterion for the farfield source model to be
valid is [5–7]

2La2
r> (2.9)
λ
14 2 Acoustic Environment, Source Models and Sensor Arrays Theory

(a) Planar wavefront


from a farfield
source

φ
d

Reference sensor

Uniform linear array.

(b) Planar wavefront


from a farfield
source

y
φk
φx

Center of array as
reference point

Uniform circular array.

Fig. 2.7 Farfield source model for uniform a linear and b circular arrays

where r = r, La is the largest array aperture size and λ is the operating signal’s
wavelength. This criterion is based on the acceptable quadratic phase error and its
detailed discussion can be found in [8–11]. Due to the simplicity of farfield source
model, various nearfield beamforming solutions are derived from this model, such as
radial transformation [4], nearfield compensation [5], and radial reciprocity method
[12].
2.4 Sensor Arrays Theory 15

2.4 Sensor Arrays Theory

Sensors placed in space play the role of sampling received signal in space. This spatial
sampling is similar to temporal sampling in digital signal processing (DSP) systems.
Therefore, the sensors must be well distributed in space such that sufficient spatial
information can be captured from the spatial sampling of the received signal. At the
same time, the spacing of the sensors is required to be less than half the wavelengths
of the received signals to avoid spatial aliasing. Generally, spatial discrimination
capability depends on array aperture size, i.e. as aperture size increase, discrimination
improves. The absolute aperture size is not important, rather its size relative to the
wavelengths of the received signals is critical. However, the aperture size is normally
restricted due to practical reasons such as cost and design decision. There is no
single golden rule on the placement of sensors, and the placement choice is entirely
application specific.

2.4.1 Spatial Aliasing

In beamforming or spatial filtering, sensors placed in space play the role of spatially
sampling the received wave. Hence, similar to Nyquist criterion in temporal sampling,
the smallest distance d between adjacent sensors must be [13]
λ
d≤ (2.10)
2
in order to avoid spatial aliasing. The wavelength λ and frequency f of a signal are
related by

c = fλ (2.11)

where the constant c is the signal propagation speed (e.g. c = 343ms−1 for air in room
temperature and pressure). Criterion (2.10) is a necessary condition for narrowband
beamformers to avoid spatial aliasing. As an example, consider an endfire linear array
with 6 elements and its inter-element spacing of 4 cm. The highest signal frequency
that it can resolve before spatial aliasing occurs is fmax ≈ 4.3 kHz. The beampatterns
of a weight-and-sum beamformer designed for frequency f = 3 kHz and 7 kHz using
this array are shown in Fig. 2.8. The figure clearly demonstrates the occurrence of
spatial aliasing when (2.10) is violated. For broadband beamformers, the wavelength
λ is chosen to be the smallest signal wavelength, which corresponds to the highest
frequency component in a broadband signal. This selection guarantees no spatial
aliasing for all frequencies up to the chosen frequency [14].
16 2 Acoustic Environment, Source Models and Sensor Arrays Theory

(a) 0

−10

Magnitude response (dB)


−20

−30

−40

−50

−60

−70
−200 −150 −100 −50 0 50 100 150 200
Azimuth angle, φ (degree)
No spatial aliasing.

(b) 0

−10
Magnitude response (dB)

−20

−30

−40

−50

−60
−200 −150 −100 −50 0 50 100 150 200
Azimuth angle, φ (degree)
Spatial aliasing occurs.

Fig. 2.8 Endfire linear array with inter-element spacing of 4 cm for a without and b with spatial
aliasing

Fig. 2.9 Example of one-dimensional array


2.4 Sensor Arrays Theory 17

(a)

(b)

(c)

Fig. 2.10 Examples of two-dimensional arrays. a Circular array b Concentric circular array
c Rectangular planar array
18 2 Acoustic Environment, Source Models and Sensor Arrays Theory

(a)
z−axis

y−axis x−axis

(b)
z−axis

y−axis x−axis

Fig. 2.11 Examples of three-dimensional arrays. a Spherical surface array. b Spherical volume
array
2.4 Sensor Arrays Theory 19

2.4.2 Array Geometry

The placement of sensors in space to form a sensor array is dependent on design


requirements, which include the number of sensors, the size of the array relative to
the operating frequency and the performance of the array. The sensors can either
be placed arbitrarily or follow a known geometry shape. Regardless, the choice of
array geometry is important in beamformer designs as it plays a major role in the
performance of the beamformers. This is because different array geometries have
different advantages and limitations [15, 16]. For example, a uniform linear array
(see Fig. 2.9) has the best spatial resolution either at broadside or endfire, depending
on the target application, whereas a uniform circular array (see Fig. 2.10a) has a
uniform spatial resolution for the whole azimuth range.
In general, array geometries can be categorised into three main categories, namely
one-dimensional, two-dimensional and three-dimensional arrays. One-dimensional
arrays comprise of placing sensors in a line as shown in Fig. 2.9. Its variants include
uniformly or non-uniformly spaced array elements, and broadside or endfire config-
uration types. Two-dimensional arrays consist of placing sensors on a plane, which
can either fill up an enclosed area or along its perimeter. Common two-dimensional
arrays include planar, circular and multiring concentric circular arrays as shown
in Fig. 2.10. In the case of three-dimensional arrays, the sensors can be placed on
the surface of three-dimensional solids, or they can be placed on frames to fill up
the volume of three-dimensional solids, such as cylinder or sphere (see Fig. 2.11).
The choice of array patterns depends heavily on the target applications, and some
interesting array geometries specific to their applications can be found in [17–19].

2.5 Spiral Arm Array Geometry

The array geometry that is used extensively to illustrate beamformer design formu-
lations in the following chapters is a modified concentric circular array as shown
in Fig. 2.12. This array geometry can be called spiral arm array since its sen-
sors are extending spirally outwards from its centre. It consists of Kring concen-
tric rings, indexed by kring = 0, . . . , Kring − 1, with Ksen sensors, indexed by
ksen = 0, . . . , Ksen − 1, uniformly spaced along the circumference of each ring.
th
The kring ring is further twisted by an angle φkring . Hence, the total number of
sensors for such array is K = Kring Ksen , and each sensor can also be indexed by
k = (kring − 1)Ksen + ksen .
The positions (in cylindrical coordinate system) of the sensors are given by
 
2π ksen
rk = rkring , + φkring (2.12)
Ksen
20 2 Acoustic Environment, Source Models and Sensor Arrays Theory

Direction of
signal propagation

y
φ
r0 φkring
rKring −1 x

r1 Ksen

0th ring

1st ring

(Kring − 1)th ring

Fig. 2.12 Proposed spiral arm array geometry

where the centre of the array is taken as the origin of the coordinate system. The ring
radii rkring and the twist angle φkring are design parameters. Its array response (with the
centre of array taken as the reference point) to a farfield source impinging the array
from azimuth angle φ is given by
  
ωrkring 2π ksen
Afar (r, rk , ω) = exp −j cos φ − − φkring . (2.13)
c Ksen

This spiral arm array geometry possesses a few desirable characteristics that make
it an attractive candidate for broadband beamforming. Firstly, its multiring nature
allows each ring to compensate for separate frequency bands in a cooperative manner
to achieve larger bandwidth for broadband beamforming [20]. Besides, since it is a
two-dimensional array, it provides full 360◦ coverage of the azimuthal dimension,
without any ambiguity (as opposed to linear array).
Secondly, its circular symmetry property means that it has uniform resolution
throughout the entire azimuthal dimension [15]. This allows the beamformer to have
a response that is symmetric about its look direction. Moreover, the circular symmetry
property can be exploited in the design of steerable beamformer to provide full 360◦
beam steering (see Sect. 4.5).
Thirdly, each ring of the spiral arm array geometry has undergone a slight twist
(c.f. Fig. 2.10b). This rotation introduces irregularity and reduces the periodicity in
its geometry, thus providing irregular spatial sampling of the received signals, which
2.5 Spiral Arm Array Geometry 21

can help to suppress spatial aliasing [19, 21]. This property is useful in broadband
beamformer designs, especially with limited number of sensor, due to conflicting
array aperture size requirements, i.e. the spacing between sensors need to be small
enough to avoid spatial aliasing for high-frequency components but large enough to
maintain directivity for low-frequency components.

2.5.1 Ring Radii

One of the design parameters for the spiral arm array is its ring radii. From Nyquist
sampling theorem (2.10), the spacing between adjacent sensors must not be larger
than half the wavelength of the highest operating frequency in order to avoid spatial
aliasing. In contrast, the array aperture need to be sufficiently large to provide the
required spatial resolution for the low-frequency components. In order to satisfy these
contrasting requirements, the concept of narrowband signal processing is employed,
where each concentric ring from the proposed spiral arm array is designed to handle
a single-frequency component. Under this scheme, each ring radius is then selected
to satisfy the Nyquist criterion for its corresponding operating frequency given by
c
rkring ≤   (2.14)
π
4fkring sin Kring

0.8
Linear
Logarithmic
Reference
0.7

0.6

0.5
Ring radius, r (m)

0.4

0.3

0.2

0.1

0
0 500 1000 1500 2000 2500 3000 3500 4000
Frequency, f (Hz)

Fig. 2.13 Discretisation of ring radii


22 2 Acoustic Environment, Source Models and Sensor Arrays Theory

where fkring ∈ Ω is the maximum operating frequency for the kring


th
ring, and Ω is the
spectral range of interest. As an example, for Ω = [0.20, 3.8] kHz and Kring = 5,
one possible choice (following linear discretisation) is f0 = 3.8 kHz, f1 = 2.9 kHz,
f2 = 2.0 kHz, f3 = 1.1 Hz and f4 = 0.2 kHz.

Fig. 2.14 Beampatterns for fixed beamformer using a logarithmic and b linear discretisation of
ring radii
2.5 Spiral Arm Array Geometry 23

Table 2.1 Design parameters for fixed beamformer to illustrate different ring radii discretisation
Design parameters Value
Number of rings, Kring 5
Number of sensors per ring, Ksen 5
Ring twist angle, φkring 0◦
Sampling frequency, fS 8 kHz
Spectral range, Ω [0.2, 3.8] kHz
Spatial pass region, Φpb |φ| ≤ 15◦
Spatial stop region, Φsb |φ| ≥ 25◦
FIR filter length, N 64
Speed of propagating wave, c 343 m/s

However, with the finite number of rings covering a broadband signal, equation
(2.14) results in the discretisation of the broadband frequency range into Kring bands.
Judging from (2.14), which involves an inverse relationship between rkring and fkring ,
the logarithmic discretisation of fkring will outperform the linear discretisation. This
is because the uniform step size in linear discretisation does not provide sufficient
resolution at low frequencies where the value of the function (2.14) changes more
rapidly than at high frequencies. On the other hand, the logarithmic discretisation
with non-uniform step size fits nicely for (2.14), both at low and high frequencies.
This observation is shown in Fig. 2.13, where
c
rkring =   (2.15)
π
4fkring sin Kring

for fkring ∈ [0.2, 3.8] kHz (the reference) is discretised into Kring = 5 bands
using both linear and logarithmic discretisation schemes. To further highlight this
observation, the beampatterns for a fixed beamformer with linear and logarithmic
ring radii sampling are shown in Fig. 2.14. The beamformers are designed using
the weighted LS formulation in Sect. 3.3.1 with the parameters in Table 2.1, and
rkring = {0.0319, 0.0666, 0.1391, 0.2904, 0.6063} m for linear discretisation and
rkring = {0.0319, 0.0418, 0.0606, 0.1102, 0.6063} m for logarithmic discretisation.

2.5.2 Twist Angle

Unfortunately, the selection of the ring twist angle is not as straightforward as for
the ring radii. The amount of twist for each ring can be different and independent of
one another. However, if φkring is a multiple of K2π
ring
, then the spiral arm array will be
similar to the array in Fig. 2.10b.
24 2 Acoustic Environment, Source Models and Sensor Arrays Theory

(a) −7.2

−7.4
LS design error (dB)

−7.6

−7.8

−8

−8.2

−8.25 dB −12o 12o

−8.4
−40 −30 −20 −10 0 10 20 30 40
Rotation angle, φa (degree)

(b) −9

−9.5

−10
LS design error (dB)

−10.5

−11

−11.24 dB −14.4o 14.4o

−11.5
−40 −30 −20 −10 0 10 20 30 40
Rotation angle, φa (degree)

Fig. 2.15 Plot of weighted LS design errors versus φa for a Kring = 4 and b Kring = 5
2.5 Spiral Arm Array Geometry 25

In order to simplify the selection of φkring , each ring twist is restricted to be a


multiple of a scalar twist φa , i.e.

φkring = kring φa . (2.16)

Then, a simple line search algorithm can be used to find the optimum candidate for
φa , which is highly dependent on the overall beamformer design formulation and
specification. Figure 2.15 shows the cost (3.63) for the weighted LS farfield-only
steerable beamformer designs in Chap. 4 with φa ∈ [−36◦ , 36◦ ] for Kring = 4 and
5. Other design parameters are as given in Tables 4.1. From Fig. 2.15 and due to
the circular
 symmetry of the spiral arm array, the optimum

  Kring = 4 is
values for

φa = ± 12◦ + 180 Ksen
z
and for Kring = 5 is φa = ± 14.4◦ + 180 Ksen
z
, where z is a
non-negative integer.
Note that (2.16) is only one of many possible choices for φkring and results in the
proposed spiral arm array shown in Fig. 2.12. Other choices will result in different
variants of spiral arm array geometries.

2.6 Conclusions

In conclusion, before a beamformer can be designed, it is necessary to understand


the target application of the beamformer as well as the environment that it will be
operating in. This is because most of the designs of beamformers are formulated upon
a certain acoustic environment model, including types of signal source, propagation
medium and sensor array. Proper study and modelling of these environmental factors
are necessary to ensure designed beamformers to work. It is also essential to capture
and model these factors as close as possible to its practical counterparts so that the
designed beamformers, verified theoretically or through simulations, will continue
to work when deployed to its real-world application environment.

References

1. H. Kuttruff, Room Acoustics, 5th edn. (Spon Press, 2009)


2. M. Kleiner, J. Tichy, Acoustics of Small Rooms (CRC Press, 2014)
3. E.G. Williams, Fourier Acoustics: Sound Radiation and Nearfield Acoustical Holography (Aca-
demic Press, New York, 1999)
4. P.T.D. Abhayapala, Modal analysis and synthesis of broadband nearfield beamforming arrays,
Ph.D. dissertation, Telecommunications Engineering Group, The Australian National Univer-
sity (2008)
5. R.A. Kennedy, T.D. Abhayapala, D.B. Ward, Broadband nearfield beamforming using a radial
beampattern transformation. IEEE Trans. Signal Process. 46(8), 2147–2156 (1998)
6. Y.R. Zheng, R.A. Goubran, M. El-Tanany, Robust near-field adaptive beamforming with dis-
tance discrimination. IEEE Trans. Speech Audio Process. 12(5), 478–488 (2004)
26 2 Acoustic Environment, Source Models and Sensor Arrays Theory

7. H. Chen, S. Wee, Y. Zhu Liang, Optimal design of nearfield wideband beamformers robust
against errors in microphone array characteristics. IEEE Trans. Circuits Syst. I, Reg. Papers
54(9), 1950–1959 (2007)
8. P. Hacker, H. Schrank, Range distance requirements for measuring low and ultralow sidelobe
antenna patterns. IEEE Trans. Antennas Propag. 30(5), 956–966 (1982)
9. R. Hansen, Measurement distance effects on low sidelobe patterns. IEEE Trans. Antennas
Propag. 32(6), 591–594 (1984)
10. L. Ziomek, Three necessary conditions for the validity of the fresnel phase approximation for
the near-field beam pattern of an aperture. IEEE J. Ocean. Eng. 18(1), 73–75 (1993)
11. J.G. Ryan, Criterion for the minimum source distance at which plane-wave beamforming can
be applied. J. Acoust. Soc. Am. 104(1), 595–598 (1998)
12. R. Kennedy, D. Ward, T. Abhayapala, Nearfield beamforming using radial reciprocity. IEEE
Trans. Signal Process. 47(1), 33–40 (1999)
13. B.D. Van Veen, K.M. Buckley, Beamforming: a versatile approach to spatial filtering. IEEE
Signal Process. Mag. 5(2), 4–24 (1988)
14. J. Dmochowski, J. Benesty, S. Affes, On spatial aliasing in microphone arrays. IEEE Trans.
Signal Process. 57(4), 1383–1395 (2009)
15. A. Manikas, A. Alexiou, H. Karimi, Comparison of the ultimate direction-finding capabilities
of a number of planar array geometries, in Proceedings of the IEE Radar, Sonar and Navigation,
vol. 144, no. 6, Dec 1997, pp. 321–329
16. A. Sleiman, A. Manikas, The impact of sensor positioning on the array manifold. IEEE Trans.
Antennas Propag. 51(9), 2227–2237 (2003)
17. S.M. Jaeger, W.C. Horne, C.S. Allen, Effect of surface treatment on array microphone self-
noise, in Proceedings of the AIAA/CEAS Aeroacoustics Conference, Lahaina, HI, 12–14 June
2000
18. D. Wetzel, F. Liu, B. Rosenberg, L. Cattafesta, Acoustic characteristics of a circulation control
airfoil, in Proceedings of the AIAA/CEAS Aeroacoustics Conference, Miami, FL, 11–13 May
2009
19. J. Hald, J.J. Christensen, A class of optimal broadband phased array geometries designed for
easy construction, in International Congress and Exposition on Noise Control Engineering,
Dearborn, MI, 19–21 Aug 2002
20. S.C. Chan, H.H. Chen, Uniform concentric circular arrays with frequency-invariant
characteristics—theory, design, adaptive beamforming and DOA estimation. IEEE Trans. Sig-
nal Process. 55(1), 165–177 (2007)
21. D.H. Johnson, D.E. Dudgeon, Array Signal Processing—Concepts and Techniques (Prentice
Hall, 1993)
Chapter 3
Broadband Beamformer Design

Abstract A beamformer is a spatial filter used to achieve spatial selectivity. How-


ever, it can be combined with a temporal filter to achieve both spatial and tempo-
ral selectivity. Such combined filter is essentially a multidimensional filter which
is normally known as broadband beamformer. Broadband beamformers are useful
in applications involving acoustic signal, for examples, speech acquisition for per-
sonal computers, teleconferencing and built-in hands-free communication in vehi-
cles. Various design methods can be applied to design broadband beamformers. One
such method is an optimisation-based approach where different optimisation criteria
can be used to design beamformer weights for target applications. Typical beam-
former designs are based on either nearfield or farfield source model, resulting in
nearfield-only or farfield-only beamformers. However, it is possible to generalise the
design formulation to cover both nearfield and farfield cases, thus achieving mixed
nearfield–farfield beamformers, i.e. beamformers that work for both nearfield and
farfield sources simultaneously.

Keywords Broadband beamformer · Nearfield beamforming · Farfield beamforming

3.1 Introduction

Typically, a narrowband beamformer combines spatially sampled data from each


sensor to obtain an output data in a similar manner as a temporal filter combines
temporally sampled data. The difference between the two is that the beamformer
operates in spatial domain, whereas the temporal filter operates in temporal domain.
However, these two types of filter can be combined together to result in a multidi-
mensional filter which operates upon received signals in both spatial and temporal
domains. This combined filter is also known as broadband beamformer. One of the
major advantages of broadband beamformers is that they allow for both spatial and
temporal selectivity at the same time. This characteristic is desired in most applica-
tions involving acoustic signals such as speech acquisition for personal computers,
teleconferencing and built-in hands-free communication in vehicles [1].

© The Author(s) 2017 27


C.C. Lai et al., A Study into the Design of Steerable Microphone Arrays,
SpringerBriefs in Signal Processing, DOI 10.1007/978-981-10-1691-2_3
28 3 Broadband Beamformer Design

Known design methods for temporal filter design can be applied to design broad-
band beamformers [2, 3]. One such method is to minimise the error between the
beamformer response and a desired response. In this chapter, such approach will
be used in formulating the design of broadband beamformers. The design formula-
tions are generalised for both nearfield and farfield source models in order to achieve
mixed nearfield–farfield beamformer, i.e. beamformer that works for both nearfield
and farfield sources. These generalised formulations are much more flexible as the
same formulations can be used to design nearfield-only and farfield-only beamform-
ers, since both of them are special cases of the design formulations.

3.2 Beamformer Structure

In practice, a beamformer structure can take up any form and its choice is closely
related to its target application. However, two general beamformer structures are
the weight-and-sum and filter-and-sum beamformers which will be discussed in this
section.

3.2.1 Weight-and-Sum Beamformer

A typical narrowband beamformer, in its simplest form, possesses similar structure


to a finite impulse response (FIR) filter, except for its delay taps (c.f. Fig. 3.1a, b).
Note that the weights Wk at kth sensor are considered to be complex for generality.
Unlike FIR filter, the delays at each delay tap in a narrowband beamformer may not
be constant (the delays change with the direction of arrival of impinging wavefronts)
and non-uniform (the delay at one tap is not an integer multiple of the delay at other
taps), depending on the signal source model and array geometry used. These non-
uniform delays come from spatially distributed sensors used to capture signals over
a spatial aperture. Usually, these delays are closely related to array geometry and are
accounted for as part of the array response rather than in the beamformer response.
Taking away the delays from Fig. 3.1b results in Fig. 3.2, which is also known as
weight-and-sum beamformer structure.
In this structure, a complex weight is applied to the received signal at each sensor,
after which they are summed to produce a single output signal. These complex
weights scale the received signals such that they are constructively summed if they
come from a certain desired direction and destructively summed otherwise, thus
resulting in spatial selectivity [2]. A weight-and-sum beamformer is normally used
for narrowband beamforming, where the bandwidth of the signal is much smaller than
its centre frequency. In the frequency domain, the output Y (ω) of a weight-and-sum
beamformer is given by
3.2 Beamformer Structure 29

(a) x(n)
z −1 ··· z −1

w(0) w(1) w(N − 2) w(N − 1)


y(n)
···

(b)

x(n) 0th mic W (0)


τ (0)

1st mic W (1)


τ (1)

.. ..
. .

(K − 1)th
mic W (K − 1)
y(n)
τ (K − 1)

Delays from Delay-and-sum


array response beamformer
structure

Fig. 3.1 Similarity in FIR filter structure and weight-and-sum beamformer structure. a N-tap FIR
filter structure, b generic array response and delay-and-sum beamformer model arranged in similar
structure to FIR filter structure


K−1
Y (ω) = X (rk , ω) W (k) (3.1)
k=0

where X (rk , ω) is the received narrowband signal and W (k) is the complex weight
at kth sensor.
The narrowband beamformer structure in Fig. 3.2 operates only at one frequency
point. However, simple extension by means of Fourier transform (FT) and frequency-
dependent complex weights, or transfer function, Wk (ω) can be used to extend the
structure to operate over a broadband frequency range (see Fig. 3.3). In such struc-
ture, a broadband signal is decomposed into separate frequency components and
a weight-and-sum beamformer with frequency-dependent complex weights is then
30 3 Broadband Beamformer Design

0 th mic W (0)

1 st mic W (1)

.. ..
. .

(K − 1) th
mic W (K − 1)
y(n)

Fig. 3.2 Delay-and-sum beamformer structure

0th mic W (0, ω)


FT

1st mic W (1, ω)


FT

.. ..
. .

(K − 1)th
mic W (K − 1, ω)
y(n)
FT IFT

Fig. 3.3 Frequency domain broadband beamformer structure

applied to the corresponding components in order to achieve broadband beamform-


ing. This structure is also called frequency domain broadband beamformer since it
essentially decomposes the received signals into different frequency components,
performs signal processing in frequency domain, and finally reconstructs the signals
back to time domain signal. In practice, such processing can be performed using the
overlap-add or overlap-save methods. For this structure, its output signal is given by

y (t) = F −1 {Y (ω)} (3.2)


3.2 Beamformer Structure 31

where


K−1
Y (ω) = X (rk , ω) W (k, ω) , (3.3)
k=0

X (rk , ω) = F {x (rk , t)} is the FT of the received signal x (rk , t) at the kth sensor and
F {·} is the FT operation. By incorporating the array response, the narrowband signal
received at the kth sensor can be written as

X (rk , ω) = X (ω) A (r = rs , rk , ω) (3.4)

where X (ω) = F{x (t)} is the FT of the source signal x (t) and A (r = rs , rk , ω) is
the transfer function of the propagation medium between the source signal, located
at spatial position rs , and the kth sensor. Substituting (3.4) into (3.3) results in the
transfer function of the weight-and-sum beamformer, due to a source at the single
point rs ,

Y (ω) 
K−1
H (ω) = = W (k, ω) A (r = rs , rk , ω) . (3.5)
X (ω)
k=0

The beamformer response, due to a source located at any point r is thus given by


K−1
H (r, ω) = W (k, ω) A (r, rk , ω) . (3.6)
k=0

3.2.2 Filter-and-Sum Beamformer

Taking the inverse Fourier transform (IFT) of (3.6) results in its impulse response

h (r, t) = F −1 {H (r, ω)} (3.7)



K−1
= F −1 {W (k, ω) A (r, rk , ω)} (3.8)
k=0


K−1
= w (k, t)  a (r, rk , t) (3.9)
k=0

where w (k, t) and a (r, rk , t) are impulse responses of transfer function W (k, ω)
and propagation medium A (r, rk , ω), respectively, and  denotes convolution. The
output of such beamformer is given by
32 3 Broadband Beamformer Design

Fig. 3.4 Filter-and-sum 0th mic


beamformer structure
w (0, t)

1st mic
w (1, t)

.. ..
. .

(K − 1)th
mic
y(n)
w (K − 1, t)


K−1
y (t) = w (k, t)  x (rk , t) (3.10)
k=0
= h (r = rs , t)  x (t) (3.11)

where

x (rk , t) = x (t)  a (r = rs , rk , t) (3.12)

is the received signal at the kth sensor, with the source signal x (t) located at position
rs . Equation (3.10) results in the filter-and-sum beamformer structure as shown in
Fig. 3.4. Often, the filters w (k, t) is implemented by N-taps FIR filters shown in
Fig. 3.1a, i.e.

N−1
W (k, ω) = w (k, n) exp (−jωnTS ) (3.13)
n=0

where n is the time index and TS is the sampling period.


In general, a broadband beamformer is more complex and challenging to design
compared to a narrowband beamformer. This is due to trade-off between conflicting
requirements such as robustness, low-frequency response, spatial aliasing at high
frequency and array aperture size that need to be systematically addressed. The fre-
quency domain beamformer structure shown in Fig. 3.3 allows broadband frequency
components to be broken down into individual components upon which a set of
narrowband beamformers can be applied to those components. However, this is not
possible for filter-and-sum approach and thus, the complexity of broadband beam-
formers design is much greater.
3.3 Design Formulation 33

3.3 Design Formulation

In an optimised beamformer weight design, the beamformer’s response is optimised


against a certain desired response, i.e. the error between the beamformer’s actual and
desired response is minimised. Different method for calculating such error forms
different class of design formulation. Two classes of design formulations, namely
weighted LS and weighted TLS, are investigated.

3.3.1 Weighted LS Formulation

Define the error between the beamformer’s response H (r, ω) and a desired response
Hd (r, ω) to be

ξ (r, ω) = H (r, ω) − Hd (r, ω) . (3.14)

The weighted LS cost function for a given frequency ω is



JLS (ω) = V (r, ω) |ξ (r, ω)|2 dr (3.15)
R
 
where R dr = drdφ is a double integral over the spatial region of interest R
and V (r, ω) is a weighting function. Solving the minimisation problem of

min JLS (ω) ∀ω ∈ Ω (3.16)


W (k,ω)

over a frequency range of interest Ω results in an optimum transfer function


Wopt (k, ω) for the desired response Hd (r, ω). This transfer function can be used
to implement the weight-and-sum beamformer structure in Fig. 3.3. The optimum
impulse response for filter-and-sum beamformer structure in Fig. 3.4 is obtained as

hopt (k, t) = F −1 {Wopt (k, ω)}. (3.17)

Alternatively, the minimisation problem of (3.16) can be solved using matrix


calculus. Define
 H
w (ω) = W ∗ (0, ω) , . . . , W ∗ (K − 1, ω) (3.18)
a (r, ω) = [A (r, 0, ω) , . . . , A (r, K − 1, ω)] ,H
(3.19)
34 3 Broadband Beamformer Design

where ∗ denotes complex conjugate. The response from (3.6) can be written com-
pactly as

H (r, ω) = aH (r, ω) w (ω) . (3.20)

Substituting (3.20) into (3.14) results in the absolute error squared


 H  
|ξ (r, ω)|2 = aH (r, ω) w (ω) − Hd (r, ω) aH (r, ω) w (ω) − Hd (r, ω) (3.21)
= wH (ω) Q (r, ω) w (ω) − wH (ω) (Hd (r, ω) a (r, ω))
− (Hd (r, ω) a (r, ω))H w (ω) + |Hd (r, ω)|2 . (3.22)

The weighted LS cost function in (3.15) becomes

JLS (ω) = wH (ω) Q (ω) w (ω) − wH (ω) q (ω) − qH (ω) w (ω) + h (ω) (3.23)

where

Q (ω) = V (r, ω) a (r, ω) aH (r, ω) dr (3.24)
R

q (ω) = V (r, ω) Hd (r, ω) a (r, ω) dr (3.25)
R
h (ω) = V (r, ω) |Hd (r, ω)|2 dr. (3.26)
R

Hence, solving the minimisation problem of (3.16) is equivalent to differentiating


(3.23) with respect to (w.r.t.) wH (ω) and equating it to zero for all ω ∈ Ω, i.e.

dJLS (ω)
=0 (3.27)
dwH (ω)
Q (ω) w (ω) − q (ω) = 0 (3.28)
w (ω) = Q−1 (ω) q (ω) . (3.29)

So far, no restriction on the frequency ω has been imposed, and it is considered to


be continuous. In practice, the frequency ω and the frequency range Ω are normally
discretised into finite samples.
Likewise, obtaining the optimum impulse response hopt (k, t) by means of (3.17)
assumes continuous frequency ω. However, in practice, (3.17) is difficult to solve and
approximations may be sufficient. Suppose that the filter W (k, ω) is implemented
with a N-taps FIR filter, i.e. from (3.13),

W (k, ω) = eH (ω) wk (3.30)


3.3 Design Formulation 35

where wk is the real FIR filter weights attached to the kth sensor, and e (ω) is the
complex exponentials for FT, and they are given by

wk = [w (k, 0) , · · · , w (k, N − 1)]T (3.31)


 H
e (ω) = exp (−j(0)ωTS ) , exp (−jωTS ) , . . . , exp (−j (N − 1) ωTS ) . (3.32)

with the sampling period TS = f1S and fS is the sampling frequency. The FIR filter
weights can be obtained by solving (3.30) for all k. One way of solving (3.30) is by
minimising the mean squared error given by the cost function

H
Jw (k, ω) = e (ω) wk − Wopt (k, ω) 2 dω ∀k (3.33)
Ω

where Wopt (k, ω) is the optimum transfer function obtained by solving (3.29).
Another method of solving for the FIR filter weights is to embed (3.30) directly
into the design formulation. Define

aFIR (r, ω) = a (r, ω) ⊗ e (ω) (3.34)


 T
wFIR = wT0 , wT1 , . . . , wTK−1 , (3.35)

where ⊗ denotes Kronecker product. The beamformer’s response is thus given by

FIR (r, ω) wFIR .


HFIR (r, ω) = aH (3.36)

Replacing H (r, ω) in (3.14) with HFIR (r, ω) results in

ξFIR (r, ω) = HFIR (r, ω) − Hd (r, ω) (3.37)

and the new weighted LS cost function is


 
JLS,FIR = V (r, ω) |ξFIR (r, ω)|2 drdω (3.38)
Ω R

where the cost function JLS,FIR now includes an additional integration over the fre-
quency ω for minimising the weighted LS error during optimisation. Solving the
minimisation problem of

min JLS,FIR (3.39)


wFIR

results in the optimum FIR filter weights wopt,FIR for the desired response Hd (r, ω).
Likewise, the minimisation problem of (3.39) can be solved using matrix calculus.
Substituting (3.34) and (3.35) into (3.38) allows JLS,FIR to be written in matrix form

JLS,FIR = wTFIR QFIR wFIR − wTFIR qFIR − qH


FIR wFIR + hFIR (3.40)
36 3 Broadband Beamformer Design

where
 
QFIR = V (r, ω) aFIR (r, ω) aH
FIR (r, ω) drdω (3.41)
Ω R
 
qFIR = V (r, ω) Hd (r, ω) aFIR (r, ω) drdω (3.42)
Ω R
 
hFIR = V (r, ω) |Hd (r, ω)|2 drdω. (3.43)
Ω R

Differentiate (3.40) w.r.t. wTFIR and equates it to zero results in

dJLS,FIR
=0 (3.44)
dwTFIR

T

QFIR + QFIR wFIR − qFIR + q∗FIR = 0 (3.45)
−1
wFIR = {QFIR } {qFIR }. (3.46)


Note

that since
QFIR is a Hermitian matrix, QTFIR + QFIR = 2{QFIR } and
qFIR + q∗FIR = 2{qFIR }, where {·} denotes real part.
It is important to note the differences between (3.29) and (3.46). Equation (3.29)
solves for the optimum transfer function frequency by frequency, whereas (3.46)
solves for the optimum FIR filter weights by minimising the sum of weighted mean
squared error across frequency range of interest. In addition, the problem size of
(3.46) is much larger than (3.29) due to the Kronecker product.

3.3.2 Weighted TLS Formulation

The design formulations of (3.16) and (3.39) minimise the error between the beam-
former’s actual and desired response in the weighted LS sense. It is also possible
to design optimised beamformer weights by means of minimising different types of
error, such as TLS error defined as [4–6]

R V (r, ω) |ξ (r, ω)|2 dr
JTLS (ω) =  , (3.47)
R0 U (r, ω) |H (r, ω)|2 dr + 1

where the weighting function U (r, ω) can be different to V (r, ω) and integration
region R0 can be different to R, too. Solving the minimisation problem of

min JTLS (ω) ∀ω ∈ Ω (3.48)


W (k,ω)
3.3 Design Formulation 37

results in an optimum transfer function Wopt (k, ω) for the desired response Hd (r, ω)
in TLS sense. In theory, the optimum impulse response can be obtained by taking
the IFT of Wopt (ω).
Define

Q (ω) q (ω)
QTLS (ω) = (3.49)
qH (ω) h (ω)

Q0 (ω) 0
Q0,TLS (ω) = (3.50)
0 1

w (ω)
wTLS (ω) = (3.51)
−1

Q0 (ω) = U (r, ω) a (r, ω) aH (r, ω) dr, (3.52)
R0

the TLS cost function in (3.47) can be rewritten in matrix form as

wHTLS (ω) QTLS (ω) wTLS (ω)


JTLS (ω) = . (3.53)
wTLS (ω) Q0,TLS (ω) wTLS (ω)
H

Note that (3.53) is the Rayleigh–Ritz ratio whose minimum is given by the smallest
generalised eigenvalue of QTLS (ω) and Q0,TLS (ω). Hence, the minimisation prob-
lem (3.48) can be solved analytically and the solution vector wTLS (ω) is simply the
generalised eigenvector of QTLS (ω) and Q0,TLS (ω) that corresponds to their small-
est generalised eigenvalue [4, 6, 7]. The optimum beamformer transfer functions
wopt (ω) are extracted from wTLS (ω) after scaling its last element to −1. Since the
analytical solution of this design formulation is given by the eigenvector of its matri-
ces, this formulation is also called eigenfilter design method.
Again, the optimum impulse response hopt (k, t) can be obtained by taking the
IFT of wopt (ω). Alternatively, if N-taps FIR filter is used to implement the filter-
and-sum beamformer, the optimum FIR filter weights can be solved directly from
the optimisation problem in the similar way as in Sect. 3.3.1. Define a weighted TLS
cost function
 
V (r, ω) |ξFIR (r, ω)|2 drdω
JTLS,FIR =  Ω R . (3.54)
Ω R0 U (r, ω)
|HFIR (r, ω)|2 drdω + 1

Solving the minimisation problem of

min JTLS,FIR (3.55)


wFIR
38 3 Broadband Beamformer Design

results in the optimum FIR filter weights wopt,FIR . Using (3.34) and (3.35), JTLS,FIR
in (3.54) can be written as

wTTLS,FIR QTLS,FIR wTLS,FIR


JTLS,FIR = . (3.56)
wTTLS,FIR Q0,TLS,FIR wTLS,FIR

where

QFIR qFIR
QTLS,FIR = H (3.57)
qFIR hFIR

Q0,FIR 0
Q0,TLS,FIR = (3.58)
0 1

wFIR
wTLS,FIR = (3.59)
−1
 
Q0,FIR = U (r, ω) aFIR (r, ω) aH
FIR (r, ω) drdω. (3.60)
Ω R0

The optimum FIR filter weights are obtained by solving for the generalised eigen-
vector of QTLS,FIR and Q0,TLS,FIR that corresponds to their smallest generalised eigen-
value, and scaling the last element of vector wTLS,FIR to −1.
The main advantage of this weighted TLS design formulation is that no matrix
inversion is required to solve the design analytically [8] and the formulation is unbi-
ased [9], as opposed to the weighted LS design formulation. The weighted TLS
design formulation, which minimises the Rayleigh quotient, can be solved using
SVD which is numerically robust [9].

3.4 Mixed Nearfield–Farfield Design Formulation

The design formulations discussed in Sects. 3.3.1 and 3.3.2 can be readily extended to
design beamformers that work for both nearfield and farfield sources at the same time.
For such beamformer, which is also known as mixed nearfield–farfield beamformer,
its response is invariant to radial distance r. In order to achieve such design, the
desired beampattern needs to be independent of r, i.e.

Hd (r, φ, ω) = Hd (φ, ω) . (3.61)

From here onward, the vector r which gives the spatial position in polar coordinate
is separated into (r, φ) for clarity. As such, the integrations for r in both the LS and
the TLS cost functions are replaced by
  
·dr = ·drdφ, (3.62)
R R Φ
3.4 Mixed Nearfield–Farfield Design Formulation 39

where R is the range of interest for r, Φ is the range of interest for azimuth angle φ and
R = {r, φ : r ∈ R, φ ∈ Φ}. The range R needs to cover both nearfield and farfield
regions so that both the nearfield and the farfield design formulations are mixed into
a single design when designing mixed nearfield–farfield beamformers [6, 10]. In this
design, the array response (2.3) is used as it describes a generic normalised array
response, which covers both nearfield and farfield sources.
Note that the mixed nearfield–farfield beamformer design is achieved by simply
forcing the desired response Hd (r, φ, ω) to be independent of r as in (3.61). There
is no extra design formulation required to achieve such design. Previously discussed
design formulations (weighted LS in Sect. 3.3.1 and weighted TLS in Sect. 3.3.2)
can be readily applied to obtain the optimum transfer function or optimum FIR filter
weights for mixed nearfield–farfield beamformer designs.

3.5 Performance Metrics

There are a number of performance metrics that can be used to evaluate the per-
formance of a beamformer. Different metrics attempt to quantify different aspects
of a beamformer. Firstly, the performance error, defined as the error between the
actual beamformer response and the desired response, is used to analyse the behav-
iour of a beamformer at different design parameter values, both within and outside
its design specifications. This provides the overall picture on the performance of a
beamformer as well as conveying the situation where the beamformer breaks down.
The performance errors, defined as
 
ξLS (r) = |H (r, φ, ω) − Hd (r, φ, ω)|2 dφdω (3.63)
Ω Φ
 
Ω Φ
|H (r, φ, ω) − Hd (r, φ, ω)|2 dφdω
ξTLS (r) =   (3.64)
Ω Φ
|H (rd , φ, ω)|2 dφdω

are used for LS- and TLS-based beamformer, respectively. The parameter rd in (3.64)
refers to the value of r which the beamformer is designed for.
Secondly, the directivity, defined as [11, 12]

|H (r, φ, ω)|2
D (r, φ, ω) =  (3.65)
1
2π Φ
|H (r, φ0 , ω)|2 dφ0

is used to evaluate the beamformer gain against isotropic noise. This is because the
numerator represents the power of the signal arriving from φ and the denominator
represents the isotropic noise power at the array. Note that φ0 is an integration variable
and is used to distinguish from φ in the numerator.
40 3 Broadband Beamformer Design

Thirdly, the array gain measurement can be used to measure the signal-to-noise
ratio (SNR) improvement of a beamformer. It is defined as the ratio of SNR at the
output of beamformer to the SNR at an input sensor and is given by [12]

−1
Aw (ω) = wH (ω) w (ω) (3.66)
K−1 −1

= k e (ω) e (ω) wk
wH H
. (3.67)
k=0

3.6 Design Examples and Evaluation

A number of design examples are presented here to illustrate the design formulations
discussed. For all the design examples provided in this section, the array geometry
used is the spiral arm array discussed in Sect. 2.5. The sensor array and beamformer
design specifications used are as tabulated in Table 3.1. For each design formulation,
different design examples as given in Table 3.2 are provided to cover the nearfield-
only, farfield-only and mixed nearfield–farfield beamformer designs.
The nearfield–farfield boundaries for a broadband source with f ∈ Ω are derived
from (2.9) and (2.11). They are given by

2La2 min (f )
r1 = (3.68)
c

Table 3.1 Specifications for spiral arm array geometry and fixed beamformer design
Design parameters Value
Number of rings, Kring 4
Number of sensors per ring, Ksen 5
Ring radii, rkring 0.0319, 0.0852, 0.2272,
0.6063 m
Ring twist angle, φkring 12◦
Sampling frequency, fS 8 kHz
Spectral range, Ω [0.2, 3.8] kHz
Spatial pass region, Φpb |φ| ≤ 15◦
Spatial stop region, Φsb |φ| ≥ 25◦
FIR filter length, N 64
Speed of propagating wave, c 343 m/s
3.6 Design Examples and Evaluation 41

Table 3.2 Design examples for each design formulation


Design formulation Name Value of r Description
Weighted LS (Eq. 3.46) LSnear 1 Nearfield LS design
LSfar 100 Farfield LS design
LSmixed 1 ≤ r ≤ 100 Mixed nearfield–farfield LS design
Weighted TLS (Eq. 3.56) TLSnear 1 Nearfield TLS design
TLSfar 100 Farfield TLS design
TLSmixed 1 ≤ r ≤ 100 Mixed nearfield–farfield TLS
design

and

2La2 max (f )
r2 = . (3.69)
c

That is a broadband source, for all f ∈ Ω, is a nearfield source for r < r1 , a farfield
source for r > r2 and a mixture of both for r1 ≤ r ≤ r2 . The largest array aperture
size La is the diameter of the outermost ring, i.e. La = 1.2126 m from Table 3.1. This
results in r1 = 1.71 m and r2 = 32.58 m. Hence, r = 1 m is chosen for nearfield-only
designs and r = 100 m for farfield-only designs. For mixed nearfield–farfield designs,
1 ≤ r ≤ 100 m is chosen such that both nearfield and farfield regions are covered. In
these examples, the sensors are assumed to be omnidirectional microphones operat-
ing in air. For simplicity, the weighting functions are selected to be V (r, ω) = 1 and
U (r, ω) = 1. The desired beamformer response Hd (r, φ, ω), which is chosen to be
invariant to r, is defined as


exp −jωTS N−1 , φ ∈ Φpb , ω ∈ Ωpb
Hd (r, φ, ω) = 2 . (3.70)
0 , φ ∈ Φsb

The integral w.r.t. ω in the design formulations can be solved analytically (see Appen-
dix A), while the other integrals are approximated by uniformly spaced Riemann sum
with the number of points as specified in Table 3.3.

Table 3.3 Number of uniform discretisation points for numerical evaluation of integrals
Parameter Number of points
Azimuth angle range, Φ 360
Source radial distance, R 50
42 3 Broadband Beamformer Design

30

25

20
Performance error, ξLS(r) (dB)

15

LSnear
10 LS
far
LSmixed

−5

−10
0 1 2
10 10 10
Source radial distance, r (m)

Fig. 3.5 Performance error for LS-based designs

The plot of performance error (3.63) for LS-based designs is shown in Fig. 3.5.
The figure shows that the LSnear design only works at its designed distance of r = 1 m
and is inoperable for r
= 1 m. For the LSfar design, its performance error decreases as
the value of r increases from nearfield to farfield distance. When the value of r is suf-
ficiently large, around r ≥ 40 m, its performance error remains almost constant. This
means that the LSfar design will work for r ≥ 40 m even though the design specifica-
tion is only for r = 100 m. This is in agreement with (2.9) and the fact that as r → ∞,
the wavefronts that impinge on the sensor array approach planar wavefronts, which
is the concept behind farfield source model. Lastly, as expected, the LSmixed design
is operable for 1 ≤ r ≤ 100 m, which spans both nearfield and farfield distances, as
shown by its constant, low performance error. The mixed nearfield–farfield design
formulation essentially averages the beamformer performance in both nearfield and
farfield in order to achieve mixed nearfield–farfield capability. Consequently, its min-
imum performance error is higher compared to both the LSnear and LSfar designs.
The directivity plots for these LS-based designs, shown in Figs. 3.6, 3.7 and 3.8,
are also in line with the observations from Fig. 3.5. Note that the x-axes for the direc-
tivity plots are in logarithmic scale, and the directivities are evaluated with φ = 0.
3.6 Design Examples and Evaluation 43

15

3500

3000
10

Directivity index, DI (dB)


2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
0 1 2
10 10 10
Source radial distance, r (m)

Fig. 3.6 Directivity plot for the LSnear design

Figures 3.6 and 3.7 show that for the LSnear and LSfar designs, when the values of
r are outside their design specifications, their directivity values at low frequencies
decrease significantly. This implies that the designs only fail at low frequencies and
they continue to work at high frequencies. This observation is in agreement with the
array gain plot in Fig. 3.9, which shows positive array gain (i.e. SNR improvement)
at frequency above 1800 Hz but negative array gain at low frequencies. This negative
gain implies that any mismatch or error between the design model and practical envi-
ronment will be significantly amplified, thus degrading the SNR at low frequencies.
This is because at low frequencies, the designs behave as super-directive beamform-
ers which are sensitive to any mismatch, in this case, the mismatch in r between
the design specifications and the actual operating r. The LSmixed design has slightly
higher array gain than both LSnear and LSfar designs as it is designed to account for
such mismatch and, thus, works for both nearfield and farfield sources. However, its
array gain is still quite low and it is possible to further improve the array gain, as will
be discussed in Chap. 5.
For TLS-based designs, their performance errors are shown in Fig. 3.10 with
their directivity plots shown in Figs. 3.11, 3.12 and 3.13 and their array gains shown
in Fig. 3.14. In general, TLS-based designs achieve similar performance as their
corresponding LS counterparts. However, the performance errors for TLS-based
designs in Fig. 3.10 cannot be compared to that of LS-based designs in Fig. 3.5
44 3 Broadband Beamformer Design

15

3500

3000
10

Directivity index, DI (dB)


2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
0 1 2
10 10 10
Source radial distance, r (m)

Fig. 3.7 Directivity plot for the LSfar design

15

3500

3000
10
Directivity index, DI (dB)

2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
0 1 2
10 10 10
Source radial distance, r (m)

Fig. 3.8 Directivity plot for the LSmixed design


3.6 Design Examples and Evaluation 45

10

−10

−20
Array gain, A (dB)

−30 LSnear
w

LSfar
LSmixed
−40

−50

−60

−70

−80
0 500 1000 1500 2000 2500 3000 3500 4000
Frequency, f (Hz)

Fig. 3.9 Array gain plot for LS-based design

45
TLS
near
TLS
40 far
TLS mixed

35

30
(r) (dB)

25
TLS
Performance error, ξ

20

15

10

−5
0 1 2
10 10 10
Source radial distance, r (m)

Fig. 3.10 Performance error for TLS-based designs


46 3 Broadband Beamformer Design

15

3500

3000
10

Directivity index, DI (dB)


2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
0 1 2
10 10 10
Source radial distance, r (m)

Fig. 3.11 Directivity plot for the TLSnear design

15

3500

3000
10
Directivity index, DI (dB)

2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
0 1 2
10 10 10
Source radial distance, r (m)

Fig. 3.12 Directivity plot for the TLSfar design


3.6 Design Examples and Evaluation 47

15

3500

3000
10

Directivity index, DI (dB)


2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
0 1 2
10 10 10
Source radial distance, r (m)

Fig. 3.13 Directivity plot for the TLSmixed design

20

−20
Array gain, A (dB)

−40
w

TLS
near
TLSfar
TLS
mixed
−60

−80

−100

−120
0 500 1000 1500 2000 2500 3000 3500 4000
Frequency, f (Hz)

Fig. 3.14 Array gain plot for TLS-based design


48 3 Broadband Beamformer Design

Fig. 3.15 Beampatterns for TLSnear design evaluated for a r = 1 m and b r = 100 m
3.6 Design Examples and Evaluation 49

Fig. 3.16 Beampatterns for TLSmixed design evaluated at a r = 1 m and b r = 100 m


50 3 Broadband Beamformer Design

because their performance metrics are different (c.f. (3.64) with (3.63)). Due to
the scaling term in the denominator of (3.64), it appears that TLS-based designs
achieve lower performance error than their corresponding LS counterparts. This
scaling also gives the impression that there is no trade-off (increase in performance
error for achieving mixed nearfield–farfield capability) for the TLSmixed design as
compared to the TLSnear and TLSfar designs due to different scaling between the three
designs. Beampatterns for the TLSnear and TLSmixed designs, evaluated at r = 1 m
and r = 100 m are shown in Figs. 3.15 and 3.16 to further illustrate the observations
made.

3.7 Conclusions

As a summary, broadband beamformers can achieve both spatial and spectral selec-
tivity at the same time. This is because they are multidimensional filters acting on
both spatial and spectral domains. The design formulations for broadband beam-
formers presented in this chapter achieve such selectivity by minimising the error
between the beamformer response and a desired response. These design formula-
tions are generalised to cover both nearfield and farfield source models. This allows
three different types of beamformer, namely nearfield-only, farfield-only and mixed
nearfield–farfield beamformers to be designed from the same design formulations.
Design examples provided show beamformer characteristics of spatial selectivity,
frequency invariant response as well as operability for both nearfield and farfield
sources (for mixed nearfield–farfield designs) as expected from the design formula-
tions. When operating outside the design specifications, the beamformers start to fail
at the low frequencies. This is because at low frequencies the beamformers behave
as super-directive beamformers which are very sensitive to errors, i.e. the discrep-
ancy between the operating environment and the assumptions made in the design
formulations.

References

1. G.W. Elko, Microphone array systems for hands-free telecommunication. Speech Commun.
20(3–4), 229–240 (1996)
2. B.D. Van Veen, K.M. Buckley, Beamforming: a versatile approach to spatial filtering. IEEE
Signal Process. Mag. 5(2), 4–24 (1988)
3. H. Krim, M. Viberg, Two decades of array signal processing research: the parametric approach.
IEEE Signal Process. Mag. 13(4), 67–94 (1996)
4. S.C. Pei, C.C. Tseng, A new eigenfilter based on total least squares error criterion. IEEE Trans.
Circuits Syst. I Fundam. Theory Appl. 48(6), 699–709 (2001)
5. S.V. Huffel, J. Vandewalle, The total least squares problem: computational aspects and analysis.
Soc. Ind. Appl. Math. (1991)
6. S. Doclo, M. Moonen, Design of far-field and near-field broadband beamformers using eigen-
filters. Signal Process. 83(12), 2641–2673 (2003)
References 51

7. A.H. Sayed, Adaptive filters (Wiley, 2008)


8. S. Doclo, M. Moonen, Design of broadband beamformers robust against gain and phase errors
in the microphone array characteristics. IEEE Trans. Signal Process. 51(10), 2511–2526 (2003)
9. I. Markovsky, S. Van Huffel, Overview of total least-squares methods. Signal Process. 87(10),
2283–2302 (2007)
10. D.B. Ward, G.W. Elko, Mixed nearfield/farfield beamforming: a new technique for speech
acquisition in a reverberant environment, in Proceedings of IEEE workshhop on Applications
of Signal Processing to Audio and Acoustics, New Paltz, NY, 19–22 Oct 1997, pp. 4–7
11. J. Bitzer, K.U. Simmer, Superdirective microphone arrays, in Microphone Arrays, ch. 2, ed. by
M. Brandstein, D. Ward (Springer, Berlin, 2001), pp. 19–38
12. H.L.V. Trees, Optimum array processing (Wiley, 2004)
Chapter 4
Steerable Broadband Beamformer Design

Abstract The design formulation for fixed broadband beamformers can be extended
to steerable broadband beamformers using a polynomial filter structure to achieve
beam steering. The main advantage of a steerable broadband beamformer is that once
its coefficients are designed, its main beam can still be steered dynamically without
the need to redesign the beamformer weights. This feature is useful in applications
where a desired signal source does not always remain fixed at a single spatial location,
but is moving. These electronically steerable broadband beamformers allow the main
beam to be beamed and locked onto the same signal source even if it moves to another
spatial location.

Keywords Steerable beamformer · Polynomial filter structure · Farrow structure

4.1 Introduction

A fixed beamformer, though simple, has a fixed response. It is limited therefore to


applications where the location of the signal source is fixed. If the signal source
moves to a new position, then the fixed beamformer will need to be redesigned to
cater for such a change. In contrast, a steerable broadband beamformer (SBBF) is
able to steer its main beam on the fly, without the need to redesign the weights.
This steering capability offers dynamic beamforming which is extremely useful
in applications that involve moving sources. Some examples include audio–video
conferencing, hands-free communication systems, audio surveillance systems, and
human–machine interface systems where the human speaker (or signal source) is
likely to move around. In these cases, acoustic signal reception using fixed beam-
formers is not feasible and SBBFs provide a better fit since they can be steered readily
to track the moving source.
In the light of these demands, SBBFs have attracted much research attention. One
of the existing attempts to design SBBFs involves using a polynomial FIR filter or the
Farrow structure [1] to provide nonlinear mixing of the FIR filters to achieve beam
steering [2, 3]. Another approach of beam steering is proposed by Parra, where the
spatial–temporal dependency of a broadband beamformer is separated and steering is

© The Author(s) 2017 53


C.C. Lai et al., A Study into the Design of Steerable Microphone Arrays,
SpringerBriefs in Signal Processing, DOI 10.1007/978-981-10-1691-2_4
54 4 Steerable Broadband Beamformer Design

achieved by using the Wigner rotation matrix [4]. Other methods include designing a
modal beamformer, where the received signals are first decomposed into orthogonal
modes and then linearly combined to achieve a desired response [5, 6]. For this
modal beamformer, beam steering is achieved by means of modulating the modes.
In this chapter, the design formulations of SBBFs based on weighted LS and
weighted TLS are presented. These formulations are extensions from the design of
fixed broadband beamformer in Chap. 3 by replacing the filter at each sensor with
a polynomial filter structure. This structure extends the fixed beamformer with an
additional dimension for beam steering, parameterised by a single real coefficient.

4.2 Beamformer Structure

In the design of SBBFs, a beamformer structure having both broadband beamforming


capability and ease of beam steering is desired. One structure having such charac-
teristics is the beamformer with polynomial filter structure shown in Fig. 4.1 for
frequency-domain implementation using transfer functions and Fig. 4.2 for time-
domain implementation using filters. The polynomial filters provide an additional
degree of freedom which is necessary to achieve beam steering while maintaining
broadband signal processing capability. Beam steering is achieved by a steering func-
tion f (m, ψ) which maps the steering angle ψ to a real scalar to provide mixing of
filtered signals. Note that the structure in Fig. 4.2 is generalised from the polynomial
or Farrow filter structure in [2].
The response of such beamformer structure is given by


M−1 
K−1
H (r, ω, ψ) = f (m, ψ) A (r, rk , ω) W (k, m, ω) (4.1)
m=0 k=0
= d (r, ω, ψ) wSBBF (ω) ,
H
(4.2)

where M is the number of filters per sensor, W (k, m, ω) is the transfer function for
the mth filter at the kth sensor, and Eq. (4.2) is obtained by defining

d (r, ω, ψ) = a (r, ω) ⊗ f (ψ) (4.3)


 T
f (ψ) = f (0, ψ) , f (1, ψ) , · · · , f (M − 1, ψ) (4.4)
 H
wSBBF (ω) = wH SBBF (0, ω) , · · · , wSBBF (K − 1, ω)
H
(4.5)
 H
wSBBF (k, ω) = W ∗ (k, 0, ω) , · · · , W ∗ (k, M − 1, ω) , (4.6)

where a (r, ω) as defined in (3.19) is repeated below.

a (r, ω) = [A (r, 0, ω) , . . . , A (r, K − 1, ω)]H


4.2 Beamformer Structure 55

0th mic W (0, 0, ω) f (0, ψ)


FT

W (0, 1, ω) f (1, ψ)

.. ..
. .

W (0, M − 1, ω) f (M − 1, ψ)

1st mic W (1, 0, ω) f (0, ψ)


FT

W (1, 1, ω) f (1, ψ)

.. ..
. .

W (1, M − 1, ω) f (M − 1, ψ)

.. ..
. .
th W (K − 1, 0, ω) f (0, ψ)
(K − 1) mic
FT

W (K − 1, 1, ω) f (1, ψ)

.. ..
. .

W (K − 1,
M − 1, ω) f (M − 1, ψ)
y(n)
IFT

Fig. 4.1 Frequency-domain SBBF structure using the generalised Farrow structure
56 4 Steerable Broadband Beamformer Design

0th mic f (0, ψ)


w (0, 0, t)

f (1, ψ)
w (0, 1, t)

.. ..
. .

f (M − 1, ψ)
w (0, M − 1, t)

1st mic f (0, ψ)


w (1, 0, t)

f (1, ψ)
w (1, 1, t)

.. ..
. .

f (M − 1, ψ)
w (1, M − 1, t)

.. ..
. .
th f (0, ψ)
(K − 1) mic
w (K − 1, 0, t)

f (1, ψ)
w (K − 1, 1, t)

.. ..
. .

f (M − 1, ψ)
y(n)
w (K − 1, M − 1, t)

Fig. 4.2 Time-domain SBBF structure using the generalised Farrow structure
4.2 Beamformer Structure 57

The impulse response of such beamformer can be obtained by taking the IFT of
(4.1), i.e.

M−1 
K−1
h (r, t, ψ) = f (m, ψ) a (r, rk , t)  w (k, m, t) . (4.7)
m=0 k=0

4.3 Design Formulations

Similar to the design formulations of fixed broadband beamformers in Sect. 3.3,


the design of SBBFs can be posed as an optimisation problem to minimise the
error between the actual beamformer response H (r, ω, ψ) and a desired response
Hd (r, ω, ψ). This approach will be used to solve for the optimum transfer func-
tions and optimum FIR filter weights for a given desired beamformer response. The
obtained optimum transfer functions and optimum FIR filter weights can be used to
implement the SBBF structure in Figs. 4.1 and 4.2, respectively.

4.3.1 Weighted LS Formulation

Define the error function

ξ (r, ω, ψ) = H (r, ω, ψ) − Hd (r, ω, ψ) . (4.8)

The weighted LS cost function for a given frequency ω is


 
JLS (ω) = V (r, ω, ψ) |ξ (r, ω, ψ)|2 drdψ (4.9)
Ψ R(ψ)

where Ψ is the steering range of interest, V (r, ω, ψ) is a weighting function, and


the integration region R (ψ) is now dependent on the steering angle ψ.
Solving the minimisation problem

min JLS (ω) ∀ω ∈ Ω (4.10)


W (k,m,ω)

over the frequency range Ω results in an optimum transfer function Wopt (k, m, ω) for
the desired response Hd (r, ω, ψ). The optimum impulse response can be obtained
by taking the IFT of Wopt (k, m, ω).
The minimisation problem (4.10) can also be solved using matrix calculus. Using
(4.3) to (4.6), the cost function in (4.9) can be written as

JLS (ω) = wH
SBBF (ω) QSBBF (ω) wSBBF (ω) − wSBBF (ω) qSBBF (ω)
H

− qH
SBBF (ω) wSBBF (ω) + hSBBF (ω) (4.11)
58 4 Steerable Broadband Beamformer Design

where
 
QSBBF (ω) = V (r, ω, ψ) d (r, ω, ψ) dH (r, ω, ψ) drdψ (4.12)
Ψ R(ψ)
 
qSBBF (ω) = V (r, ω, ψ) Hd (r, ω, ψ) d (r, ω, ψ) drdψ (4.13)
Ψ R(ψ)
 
hSBBF (ω) = V (r, ω, ψ) |Hd (r, ω, ψ)|2 drdψ. (4.14)
Ψ R(ψ)

Solving the minimisation problem (4.10) is thus equivalent to differentiating


SBBF (ω) and equating it to zero for all ω ∈ Ω, i.e.
(4.11) w.r.t. complex wH

dJLS
=0 (4.15)
dwSBBF (ω)
H

QSBBF (ω) wSBBF (ω) − qSBBF (ω) = 0 (4.16)


wSBBF (ω) = Q−1
SBBF (ω) qSBBF (ω) . (4.17)

Note that (4.17) has the same form as (3.29) for fixed beamformer design, since
both of them follow the same design formulation process. The differences that dis-
tinguish the two formulations are in the matrix Q (ω) and the vector q (ω).
Now, suppose that N-taps FIR filter is used to implement the SBBF structure in
Fig. 4.2, i.e.


N−1
W (k, m, ω) = w (k, m, n) exp (−jnω) (4.18)
n=0
= e (ω) wSBBF,FIR (k, m)
H
(4.19)

where

wSBBF,FIR (k, m) = [w (k, m, 0) , · · · , w (k, m, N − 1)]T (4.20)

The FIR filter weights wSBBF,FIR (k, m), which are real, can be obtained by solving
(4.19) for all k and m. One way of doing this is to minimise the mean-squared error
given by the cost function

 H 
Jw (ω) = e (ω) wSBBF,FIR (k, m) − Wopt (k, m, ω)2 dω ∀m, k (4.21)
Ω

where Wopt (k, m, ω) is the optimum transfer function obtained by solving (4.17).
Alternatively, the optimum FIR filter weights wSBBF,FIR (k, m) can be obtained
directly from the design formulation. Define
4.3 Design Formulations 59

 T
wSBBF,FIR (k) = wTSBBF,FIR (k, 0) , · · · , wTSBBF,FIR (k, M − 1) (4.22)
 T
wSBBF,FIR = wTSBBF,FIR (0) , · · · , wTSBBF,FIR (K − 1) (4.23)
dFIR (r, ω, ψ) = d (r, ω, ψ) ⊗ e (ω) . (4.24)

Note that wSBBF,FIR is the FIR filter weights vector with the weights w (k, m, n),
for all k, m, and n, stacked into a long vector following the stacking pattern in
(4.20) and (4.22). This vector can be stacked differently, but requires the vector
dFIR (r, ω, ψ) in (4.24) to be reordered to match the new stacking order.
Using (4.23) and (4.24), the SBBF response is thus given by

FIR (r, ω, ψ) wSBBF,FIR ,


HFIR (r, ω, ψ) = dH (4.25)

Replacing H (r, ω, ψ) in (4.8) with HFIR (r, ω, ψ) results in

ξFIR (r, ω, ψ) = HFIR (r, ω, ψ) − Hd (r, ω, ψ) (4.26)

and the new weighted LS cost function is


  
JLS,FIR = V (r, ω, ψ) |ξFIR (r, ω, ψ)|2 drdωdψ (4.27)
Ψ Ω R(ψ)

where the cost function JLS,FIR now includes an additional integration over the fre-
quency ω for minimising the weighted LS error during optimisation. Solving the
minimisation problem
min JLS,FIR (4.28)
wSBBF,FIR

results in the optimum FIR filter weights wopt,SBBF,FIR for the desired response
Hd (r, ω, ψ). Likewise, the minimisation problem (4.28) can be solved using matrix
calculus. Using (4.23) and (4.24), the cost function (4.27) can be written in matrix
form as

JLS,FIR = wTSBBF,FIR QSBBF,FIR wSBBF,FIR − wTSBBF,FIR qSBBF,FIR


−qH
SBBF,FIR wSBBF,FIR + hSBBF,FIR (4.29)

where
  
QSBBF,FIR = V (r, ω, ψ) dFIR (r, ω, ψ) dFIR
H
(r, ω, ψ) drdωdψ (4.30)
Ψ Ω R(ψ)
  
qSBBF,FIR = V (r, ω, ψ) Hd (r, ω, ψ) dFIR (r, ω, ψ) drdωdψ (4.31)
Ψ Ω R(ψ)
  
hSBBF,FIR = V (r, ω, ψ) |Hd (r, ω, ψ)|2 drdωdψ. (4.32)
Ψ Ω R(ψ)
60 4 Steerable Broadband Beamformer Design

Differentiating (4.29) w.r.t. wTSBBF,FIR and equating it to zero result in

dJLS,FIR
=0 (4.33)
dwTSBBF,FIR
wSBBF,FIR = {QSBBF,FIR }−1 {qSBBF,FIR }. (4.34)

Note that (4.34) has the same form as (3.46), since the SBBF design method in this
section follows the same weighted LS design method for the fixed beamformer design
in Sect. 3.3.1. The difference is that the design formulation for SBBF is extended to
include beam steering parameter.

4.3.2 Weighted TLS Formulation

In terms of weighted TLS design formulation, define the weighted TLS cost function
as  
Ψ R(ψ) V (r, ω, ψ)
|ξ (r, ω, ψ)|2 drdψ
JTLS (ω) =   . (4.35)
Ψ R0 U (r, ω, ψ)
|H (r, ω, ψ)|2 drdψ + 1

Solving the minimisation problem

min JTLS (ω) ∀ω ∈ Ω (4.36)


W (k,m,ω)

results in an optimum transfer function Wopt (k, m, ω) for the desired response
Hd (r, ω) in TLS sense. Again, the cost function (4.35) can be written in matrix
form as
TLS,SBBF (ω) QTLS,SBBF (ω) wTLS,SBBF (ω)
wH
JTLS (ω) = . (4.37)
wHTLS,SBBF (ω) Q0,TLS (ω) wTLS,SBBF (ω)

where

QSBBF (ω) qSBBF (ω)
QTLS,SBBF (ω) = (4.38)
H
qSBBF (ω) hSBBF (ω)

Q0,SBBF (ω) 0
Q0,TLS,SBBF (ω) = (4.39)
0 1

wSBBF (ω)
wTLS,SBBF (ω) = (4.40)
−1
 
Q0,SBBF (ω) = U (r, ω, ψ) d (r, ω, ψ) dH (r, ω, ψ) drdψ (4.41)
Ψ R0
4.3 Design Formulations 61

The optimum transfer function is obtained by solving for the generalised eigen-
vector of QTLS,SBBF and Q0,TLS,SBBF that corresponds to their smallest generalised
eigenvalue and scaling the last element of vector wTLS,SBBF (ω) to −1.
For the filter implementation using N-taps FIR filter, the weighted TLS cost func-
tion becomes
  

Ψ Ω R(ψ) V (r, ω, ψ) FIR (r, ω, ψ)| drdωdψ
2
JTLS,FIR =    . (4.42)
Ψ Ω R0 U (r, ω, ψ)
|HFIR (r, ω, ψ)|2 drdωdψ + 1

The minimisation problem then becomes

min JTLS,FIR (4.43)


wSBBF,FIR

to solve for the optimum FIR filter weights wSBBF,FIR . In matrix form, JTLS,FIR is
given by
wTTLS,SBBF,FIR QTLS,SBBF,FIR wTLS,SBBF,FIR
JTLS,FIR = T (4.44)
wTLS,SBBF,FIR Q0,TLS,SBBF,FIR wTLS,SBBF,FIR

where


QSBBF,FIR qSBBF,FIR
QTLS,SBBF,FIR = H (4.45)
qSBBF,FIR hSBBF,FIR

Q0,SBBF,FIR 0
Q0,TLS,SBBF,FIR = (4.46)
0 1

wSBBF,FIR
wTLS,SBBF,FIR = (4.47)
−1
  
Q0,SBBF,FIR = U (r, ω, ψ) dFIR (r, ω, ψ) dH
FIR (r, ω, ψ) drdωdψ. (4.48)
Ψ Ω R0

Likewise, the optimum FIR filter coefficients are obtained by solving for the
generalised eigenvector of QTLS,SBBF,FIR and Q0,TLS,SBBF,FIR that correspond to their
smallest generalised eigenvalue and scaling the last element of vector wTLS,SBBF,TLS
to −1.

4.4 Mixed Nearfield–Farfield Design Formulation

The SBBF design formulations discussed can also be extended to mixed nearfield–
farfield SBBFs following the similar procedure in Sect. 3.4. For such beamformers,
their response, which is invariant to radial distance r, is given by

Hd (r, φ, ω, ψ) = Hd (φ, ω, ψ) , (4.49)


62 4 Steerable Broadband Beamformer Design

where the vector r that gives the spatial position in polar coordinate is separated into
(r, φ) for clarity.
Note that this mixed nearfield–farfield beamformer design is achieved by simply
forcing the desired response Hd (r, φ, ω, ψ) to be independent of r as in (4.49). For
this design, the integration range for r needs to cover both nearfield and farfield
distances. There is no extra design formulation required, and the formulations in
Sect. 4.3 can be readily applied to obtain the optimum transfer function or optimum
FIR filter weights for mixed nearfield–farfield beamformer designs.

4.5 Steering Function

The steering function f (m, ψ) introduces an extra degree of freedom necessary to


achieve beam steering. This function is entirely a design decision, but careful design
of f (m, ψ) allows for a reduced design problem size and yet achieves full 360◦ beam
steering.

4.5.1 Steering Range

For a SBBF in two-dimensional space, it is desirable for the main beam to be steerable
for the whole (360◦ ) azimuthal plane. This feature can be achieved without neces-
sarily designing the steering function f (m, ψ) to cover for the whole 360◦ steering
range. This can be done by exploiting the circular symmetry in the array geometry
used. Consider a circular symmetric sensor array as shown in Fig. 4.3, which can

Fig. 4.3 A circular


symmetry sensor array Sector 2 Sector 1

Sector 3 Sector 0

y
Sector 4 Sector 5
x
4.5 Steering Function 63

(a)

w7

w8 w2 w1 w6
Steering
w3 w0 range, Ψ

w9 w4 w5 w11

y
w10
x

(b) 80◦
Steering
range, Ψ
w6

w7 w1 w0 w11

w2 w5

w8 w3 w4 w10

y
w9
x

Fig. 4.4 Exploiting circular symmetry in sensor array by permuting beamformer weights to achieve
full 360◦ steering angle. a No weight permutation. Steering range is −30◦ ≤ ψ ≤ 30◦ , i.e. within
Sector 0. b Weights permuted to Sector 1. Steering range is now 30◦ < ψ ≤ 90◦ , i.e. steering angle
of ψ = 80◦ is achievable

be divided into 6 equiangular sectors. Suppose that a SBBF is designed using this
sensor array such that the steering range only covers ψ ∈ [−30◦ , 30◦ ]. The resulting
optimum beamformer weights associated with each sensor are illustrated in Fig. 4.4a.
Due to the circular symmetry of the array, steering the main beam to, for example, 80◦
is still possible with this design. This is achieved by first permuting the beamformer
64 4 Steerable Broadband Beamformer Design

weights as shown in Fig. 4.4b, and then, apply a steering of 20◦ using the steering
function f (m, ψ). The rule for designing f (m, ψ) by exploiting this property is that
given a L circular symmetric sensor array, i.e. an array that can be divided into L
equiangular
  the steering function f (m,
sectors, ψ) needs to be designed to cover

180◦ ◦
ψ ∈ − 180 L
, L
in order to achieve full 360 steering by means of beamformer
weights permutation.
Apart from circular symmetry in array geometry, the required steering range for ψ
also depends on the definition of the steering function f (m, ψ). For example, define

f (m, ψ) = ψ m , (4.50)

which results in the Farrow structure given in [2]. For ψ < 0,



|ψ|m , m = even
f (m, ψ) = (4.51)
− |ψ|m , m = odd.

Equation (4.51) means that if f (m, ψ) is as defined in (4.50), the steering range
for ψ needs to cover only the positive angle. For example, for a L circular symmetric
sensor array with f (m, ψ) as defined in (4.50), the design range for ψ only needs to

cover ψ ∈ 0, 180 L
, and yet full 360◦ steering can be achieved.

4.5.2 Normalisation of Steering Function

The steering function f (m, ψ) can also be viewed as mixing coefficients, parame-
terised by the steering angle ψ, for the filtered signal. Generally, f (m, ψ) is nor-
malised to avoid huge variation at different steering angles. One way is to normalise
f (m, ψ) such that 
f (m, ψ) = 1, ∀ψ. (4.52)
m

The normalisation scheme is entirely a design decision and can be dependent on


the choice of f (m, ψ). As an example, a normalised steering function fnorm (m, ψ)
having the same form as (4.50) can be defined as
m
ψ
fnorm (m, ψ) = , (4.53)
α

where α is the normalising constant.


4.6 Performance Metrics 65

4.6 Performance Metrics

Similar performance metrics as in Sect. 3.5 are adopted to evaluate the performance of
the SBBF design formulations. However, the metrics need to be modified to account
for beam steering. Hence, the performance errors for SBBF are now defined as
 
ξLS (r, ψ) = |H (r, φ, ω, ψ) − Hd (r, φ, ω, ψ)|2 dφdω, (4.54)
Ω Φ
 
Ω Φ
|H (r, φ, ω, ψ) − Hd (r, φ, ω, ψ)|2 dφdω
ξTLS (r, ψ) =   (4.55)
Ω Φ
|H (rd , φ, ω, ψ)|2 dφdω

and are used to evaluate LS- and TLS-based SBBF designs, respectively. The para-
meter rd in (4.55) refers to the value of r that the beamformers are designed for.
Likewise, the directivity for SBBF is defined as

|H (r, φ, ω, ψ)|2
D (r, φ, ω, ψ) =  . (4.56)
1
2π Φ
|H (r, φ0 , ω, ψ)|2 dφ0

In contrast to Sect. 3.5, the array gain averaged across designed steering range as
defined by

Aw (ω) = Aw (ω, ψ) dψ (4.57)
Ψ

where
M−1 −1
 
K−1
Aw (ω, ψ) = f (m, ψ) wH
SBBF,FIR
H
(k, m) e (ω) e (ω) wSBBF,FIR (k, m)
m=0 k=0
(4.58)

is used to evaluate and compare the SNR improvement for SBBFs. This averaged
metrics allow the array gains for multiple SBBF designs to be superimposed on a
single plot for comparison.

4.7 Design Examples and Evaluation

For all the design examples provided in this section, the same spiral arm array as
discussed in Sect. 2.5 is used with all required design specifications as tabulated in
Table 4.1, where appropriate 2π wrapping has been considered for spatial pass region
and stop region. With these design parameters, the same nearfield–farfield boundary
as in Sect. 3.6 applies.
66 4 Steerable Broadband Beamformer Design

Table 4.1 Specifications for spiral arm array geometry and SBBF design
Design parameters Value
Number of rings, Kring 4
Number of sensors per ring, Ksen 5
Ring radii, rkring 0.0319, 0.0852, 0.2272, 0.6063 m
Ring twist angle, φkring 12◦
Sampling frequency, fS 8 kHz
Steering range, Ψ −36◦ ≤ ψ ≤ 36◦
Spectral range, Ω [0.2, 3.8] kHz
BWφ
Spatial pass region, Φpb (ψ) |φ − ψ| ≤ 2
BWφ
Spatial stop region, Φsb (ψ) |φ − ψ| ≥ 2 + T Wφ
Spatial passband width, BWφ 30◦
Spatial transition width, T Wφ 10◦
FIR filter length, N 64
Speed of propagating wave, c 343 m/s

Table 4.2 Design examples for each design formulation


Design Name Value of r Description
formulation
Weighted LS LSnear 1 Nearfield LS design
Eq. (4.34)
LSfar 100 Farfield LS design
LSmixed 1 ≤ r ≤ 100 Mixed nearfield–farfield LS design
Weighted TLS TLSnear 1 Nearfield TLS design
Eq. (4.44)
TLSfar 100 Farfield TLS design
TLSmixed 1 ≤ r ≤ 100 Mixed nearfield–farfield TLS design

For each of the design formulations, the design examples as given in Table 4.2 are
provided. The weighting functions are selected to be V (r, ω) = 1 and U (r, ω) = 1.
The integrals w.r.t. ω and ψ are solved analytically (see Appendix B), while the other
integrals are approximated by uniformly spaced Riemann sum with the number of
points as specified in Table 4.3. The steering function is as defined in (4.53) with
α = 72◦ . The desired beamformer response Hd (r, φ, ω, ψ), which is chosen to be
invariant to r, is defined as
 
exp −jωTS N−1 , φ ∈ Φpb (ψ) , ω ∈ Ωpb
Hd (r, φ, ω, ψ) = 2 . (4.59)
0 , φ ∈ Φsb (ψ)
4.7 Design Examples and Evaluation 67

Table 4.3 Number of uniform discretisation points for numerical evaluation of integrals
Parameter Number of points
Azimuth angle range, Φ 360
Source radial distance, R 50

Figure 4.5 shows the performance error (4.54) for the LSnear design, with the
dashed lines indicating the boundaries of the steering range −36◦ ≤ ψ ≤ 36◦ . Note
that the y-axis is in logarithmic scale. The plot clearly shows that the beamformer
only works within its designed specifications, i.e. at r = 1 m, and its main beam can
be steered to any steering angle in Ψ . Outside these specifications, its performance
error increases significantly, implying that the beamformer is no longer operable.
The performance error for the LSfar design is shown in Fig. 4.6. Its performance
error decreases as the value of r increases from nearfield to farfield distance and
remains at about −10 dB for roughly r ≥ 40 m. This means that the design will work
for r ≥ 40 m even though the design specification is only for r = 100 m. This is
in agreement with (2.9) that provides the quantitative boundary between nearfield
and farfield distances. Lastly, for the LSmixed design shown in Fig. 4.7, it has low

30
2
10

25

20
Source radial distance, r (m)

Performance error, ξLS (dB)

15
1
10

10

0
0
10

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 4.5 Performance error for LSnear design


68 4 Steerable Broadband Beamformer Design

30
2
10

25

20
Source radial distance, r (m)

Performance error, ξLS (dB)


15
1
10

10

0
0
10

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 4.6 Performance error for LSfar design

performance error within the range 1 ≤ r ≤ 100 and ψ ∈ Ψ , showing that the LSmixed
design works for both nearfield and farfield sources and have beam steering capability
as designed.
The directivity (4.56), evaluated at r = 5 m for both LSnear and LSfar , is shown
in Figs. 4.8 and 4.9, respectively. The figures show that the steerable broadband
beamformers exhibit similar behaviour to their corresponding fixed broadband beam-
formers in Sect. 3.6; that is, when operating outside their design specifications, the
designs only fail at low frequencies. This is in agreement with the array gains shown
in Fig. 4.10, which shows huge negative gain (i.e. SNR degradation) at low frequen-
cies. The reason is that at low frequencies, the designs behave as super-directive
beamformers and are sensitive to the mismatch in r between the design specification
and the actual operating r. This mismatch is significantly amplified and thus causing
degradation in the SNR at low frequencies. The LSmixed beamformer is designed to
operate in both nearfield and farfield regions, possess consistent directivity as shown
in Fig. 4.11. All directivity plots are evaluated with φ = ψ.
For TLS-based designs, their performance errors (4.55) are shown in Figs. 4.12,
4.13, and 4.14, with their directivity plots shown in Figs. 4.15, 4.16, and 4.17 and
their array gains shown in Fig. 4.18. In general, TLS-based designs provide similar
observations to their LS counterparts.
4.7 Design Examples and Evaluation 69

30
2
10

25

20
Source radial distance, r (m)

Performance error, ξLS (dB)


15
1
10

10

0
0
10

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 4.7 Performance error for LSmixed design

15

3500

3000
10
Directivity index, DI (dB)

2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 4.8 Directivity plot for LSnear design evaluated at r = 5 m


70 4 Steerable Broadband Beamformer Design

15

3500

3000
10

Directivity index, DI (dB)


2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 4.9 Directivity plot for LSfar design evaluated at r = 5 m

10

−10

−20
Array gain, Aw (dB)

LS
near
−30 LS
far
LSmixed

−40

−50

−60

−70
0 500 1000 1500 2000 2500 3000 3500 4000
Frequency, f (Hz)

Fig. 4.10 Averaged array gain for LS-based designs


4.7 Design Examples and Evaluation 71

15

3500

3000
10

Directivity index, DI (dB)


2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 4.11 Directivity plot for LSmixed design evaluated at r = 5 m

30
2
10

25

20
Source radial distance, r (m)

Performance error, ξLS (dB)

15
1
10

10

0
0
10

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 4.12 Performance error for TLSnear design


72 4 Steerable Broadband Beamformer Design

30
2
10

25

20
Source radial distance, r (m)

Performance error, ξLS (dB)


15
1
10

10

0
0
10

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 4.13 Performance error for TLSfar design

30
2
10

25

20
Source radial distance, r (m)

Performance error, ξLS (dB)

15
1
10

10

0
0
10

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 4.14 Performance error for TLSmixed design


4.7 Design Examples and Evaluation 73

15

3500

3000
10

Directivity index, DI (dB)


2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 4.15 Directivity plot for TLSnear design evaluated at r = 5 m

15

3500

3000
10
Directivity index, DI (dB)

2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 4.16 Directivity plot for TLSfar design evaluated at r = 5 m


74 4 Steerable Broadband Beamformer Design

15

3500

3000
10

Directivity index, DI (dB)


2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 4.17 Directivity plot for TLSfar design evaluated at r = 5 m

20

−20
Array gain, A (dB)

−40
w

TLS
near
TLS
far
TLS
mixed
−60

−80

−100

−120
0 500 1000 1500 2000 2500 3000 3500 4000
Frequency, f (Hz)

Fig. 4.18 Averaged array gain for TLS-based designs


4.7 Design Examples and Evaluation 75

Fig. 4.19 Beampattern for TLSmixed design evaluated at r = 1 m and ψ = 100◦

Fig. 4.20 Beampattern for TLSmixed design evaluated at r = 100 m and ψ = −50◦
76 4 Steerable Broadband Beamformer Design

Although all the examples are designed only for −36◦ ≤ ψ ≤ 36◦ , full 360◦ beam
steering is still possible by exploiting the circular symmetry of the array as discussed
in Sect. 4.5.1. As an illustration, Figs. 4.19 and 4.20 show the beampatterns for the
TLSmixed design, evaluated at r = 1 m and ψ = 100◦ , and r = 100 m and ψ = −50◦ ,
respectively.

4.8 Conclusions

In this chapter, the design formulation of SBBFs, realised with the polynomial filter
structure, is provided. The major advantage of such structure is that the main beam of
the beamformer can be steered easily and directly with a single real parameter. These
design formulations are extended from the design formulations of non-steerable
beamformers in Chap. 3. The extension involves introducing an extra dimension of
freedom, i.e. the order of polynomial filter structure, to achieve beam steering, which
is an additional design dimension on top of the existing features of spatial selectivity
and frequency-invariant response. These features are validated through the design
examples provided.

References

1. C.W. Farrow, A continuously variable digital delay element, in IEEE International Symposium
on Circuits Systems (ISCAS), vol. 3, Espoo, Finland, 7–9 June 1988, pp. 2641–2645
2. M. Kajala, M. Hamalainen, Filter-and-sum beamformer with adjustable filter characteristics, in
Proceedings of the IEEE International Conference on Acoustics on Speech Signal Processing
(ICASSP), vol. 5, Salt Lake City, UT, 7–11 May 2001, pp. 2917–2920
3. C.C. Lai, S. Nordholm, Y.H. Leung, Design of robust steerable broadband beamformers with spi-
ral arrays and the farrow filter structure, in Proceedings of the International Workshop Acoustics,
Echo, Noise Control, Tel Aviv, Israel, 30 Aug–2 Sep 2010
4. L.C. Parra, Steerable frequency-invariant beamforming for arbitrary arrays. J. Acoust. Soc. Am.
119(6), 3839–3847 (2006)
5. J. Meyer, G. Elko, A highly scalable spherical microphone array based on an orthonormal decom-
position of the soundfield, in Proceedings of the IEEE International Conference on Acoustics
Speech Signal Processing (ICASSP), vol. 2, Orlando, FL, 13–17 May 2002, pp. 1781–1784
6. C.C. Lai, S. Nordholm, Y.H. Leung, Design of steerable spherical broadband beamformers
with flexible sensor configurations. IEEE Trans. Audio, Speech, Lang. Process. 21(2), 427–438
(2013)
Chapter 5
Robust Formulation

Abstract It is important to incorporate robustness into broadband beamformer


designs for them to work in the practical environment. This is because errors and
mismatches between the practical environment and theoretical model can be detri-
mental to the operation of beamformers. This chapter discusses a robust beamformer
design formulation by modelling practical errors and mismatches as random vari-
ables. The beamformers are then optimised based on the mean of these stochastic
models, resulting in robust beamformers.

Keywords Robust beamformer · Stochastic error models

5.1 Introduction

In practice, it is impossible to have a completely error-free model for designing a


beamformer. Hence, robustness to such errors is a major consideration in the design of
a practical beamformer. Beamformers, especially super-directive beamformers and
small array beamformers, are known to be very sensitive to slight errors and devia-
tion between the presumed and actual models [1–5]. Any violation of the underlying
assumptions can significantly degrade their performance. Causes for such violations
can be due to mismatches between the presumed and actual array element character-
istics [6], imperfect array calibration [7], error in the sensor positions [8], electronic
self-noise, medium inhomogeneity [9], nearfield–farfield mismatch [10], mutual cou-
pling between sensors [11], local scattering and source spreading [12–15], to name a
few. One such example is shown in Sects. 3.6 and 4.7 where the nearfield-only beam-
formers and farfield-only beamformers are evaluated for the value of r that is outside
their design specifications. The significance of these errors depends heavily on the
type of sensor array used as well as the area of application. For example, the effect
of mutual coupling between sensors is often negligible in acoustic beamforming but
not in radio antenna beamforming [16].
A major issue in acoustic broadband beamformers is that at low frequencies,
they behave like super-directive or small array beamformers. In these beamformers,
the element spacings are normally small relative to the operating signal wavelength

© The Author(s) 2017 77


C.C. Lai et al., A Study into the Design of Steerable Microphone Arrays,
SpringerBriefs in Signal Processing, DOI 10.1007/978-981-10-1691-2_5
78 5 Robust Formulation

[1–5]. As a result, the array aperture size is not sufficient to provide good signal
directivity and every array element essentially “sees” the same signal sample. In
order to achieve high directivity in such beamformers, the dynamic range of the
beamformer weights needs to be very large. Although these large weights can increase
the beamformer’s gain theoretically, which is desired, it causes the beamformers to
be extremely sensitive to errors and perturbations which exist in practice.
The most common method to introduce robustness to such errors is to include a
white noise gain (WNG) constraint in beamformer weights design. This is equivalent
to the diagonal loading method if the designs are expressed in matrix form [2–5, 17–
19]. Although perturbations and deviations from practical models can be translated
to WNG, there is no clear link between the two. Hence, it is difficult to select an
appropriate level of WNG for any given set of errors in practice. This is the main
limitation of the WNG method, though it does provide a quick and simple method
to achieve robustness.
The other method to achieve robustness is to include tolerance towards errors
in the ideal models to account for practical imperfections. Beamformers are then
designed by optimising an objective function, which includes the tolerance, based
on either their worst-case or mean performances [4, 20–22]. Mean performance is
defined as the beamformer performance is averaged using a probability density func-
tion that is used to model the error distribution. Both of these approaches have their
own advantages and drawbacks. Optimising for the worst-case performance ensures
the resulting beamformers can operate for all conditions, including the worst-case
scenario. However, such designs are too pessimistic in the sense that the worst-case
scenario may be too far from the mean scenario and may only occur infrequently.
On the other hand, optimising for the mean performance ensures the beamformers
can operate in the vicinity of the mean conditions. Hence, if there is a sudden occur-
rence which shifts the operating condition far away from the mean condition, the
beamformer may fail.
For the mean performance optimisation, it can be extended to include a stochastic
model to describe the error characteristics [6, 23]. This enables explicit quantifica-
tion of the parameters related to practical environments, sources, and array models
which are known probabilistically, thus allowing a more direct and meaningful quan-
tification of physical parameters and their desired tolerance. Besides, such stochastic
model is applicable in most cases, where the errors are random and only their sto-
chastic characteristics are known. In addition, the stochastic error model is more
appealing in the sense that the errors are weighted by their PDF, i.e. errors that occur
more frequently are weighted higher than those that occur less frequently. Hence, its
mean performance, where the errors are concentrated, is optimised to achieve opti-
mum performance. In order to provide sufficient robustness against the actual error
in practical applications, the variance of the modelled error should be greater than the
variance of the actual error (e.g. from manufacturer’s datasheet or measurements).
This chapter discusses a stochastic error model for designing robust broadband
beamformers, which is an extension to the model in [6, 23]. The discussion includes
formulations involving multiplicative errors, additive errors, and their combinations.
The multiplicative error model is useful to model errors that can be translated into
5.1 Introduction 79

errors in the array elements, such as their non-ideal characteristics, mismatches


between the array elements, errors in the sensor positions, and errors in the pre-
sumed source position. The additive error model is useful to model errors due to
source spreading and local scattering [24]. The discussion of robust formulation in
this chapter focuses on the SBBF design formulations in Chap. 4. However, it can
also be readily applied to the fixed beamformers in Chap. 3, by letting ψ = 0 and
M = 1.

5.2 Conventional White Noise Gain Constraint

It is known that the performance of beamformers will degrade in the presence of errors
and their robustness can be measured in terms of WNG [18, 25]. Thus, conventional
robust design formulation involves imposing the WNG constraint given by


K −1 M−1

|W (k, m, ω)|2 ≤ ρW
2
NG ∀ω (5.1)
k=0 m=0

or

wHS B B F w S B B F ≤ ρW
2
NG ∀ω (5.2)

to the minimisation problems (4.10) and (4.36) for frequency-domain beamformer


designs. The parameter ρW N G is an upper bound for WNG and is usually a design
decision. For time-domain beamformer designs using FIR filter implementations,
the constraint


K −1 M−1
 N −1
|w (k, m, n)|2 ≤ ρW
2
NG (5.3)
k=0 m=0 n=0

or

wTS B B F,F I R w S B B F,F I R ≤ ρW


2
NG (5.4)

is imposed to the minimisation problems (4.28) and (4.43). As an example, the


resulting robust SBBF for weighted LS design formulation from (4.28) then becomes

min JL S,F I R (5.5)


w S B B F,F I R

subject to wTS B B F,F I R w S B B F,F I R ≤ ρW


2
NG

Both the constraints in (5.1) and (5.3) essentially limit the magnitude of the transfer
function W (k, m, ω) and FIR filter weights w (k, m, n) such that their amplification
80 5 Robust Formulation

towards uncorrelated noise or mismatches is limited to ρW N G . There is no clear


relationship between real-world mismatches and ρW N G . Hence, it is difficult to select
an appropriate value of ρW N G for any given set of practical perturbations and errors.
This is the main limitation of the WNG method, though it does provide a quick and
simple method to achieve robustness.

5.3 Stochastic Error Model

In practice, it is difficult to have a complete knowledge of real-world perturbations or


mismatches to be included into any robust design formulations. However, the stochas-
tic characteristics of such perturbations or mismatches are attainable, e.g. through
numerous measurements or given by manufacturer’s datasheets. Such information
can be included into the design formulations to achieve robust designs.

5.3.1 Multiplicative Error

Sensor errors ε (r, rk , ω) such as gain and phase errors can often be modelled as
multiplicative errors [6, 23], i.e.

ε (r, rk , ω) = ρ (r, rk , ω) exp ( jγ (r, rk , ω)) (5.6)

where ρ (r, rk , ω) is the gain error factor and γ (r, rk , ω) is the phase error. The
perturbed array response can then be written as

 (r, rk , ω) = ε (r, rk , ω) A (r, rk , ω) (5.7)

or in vector form as

â (r, ω) = ε (r, ω)  a (r, ω) (5.8)

where

ε (r, ω) = [ε (r, 0, ω) , · · · , ε (r, K − 1, ω)]H (5.9)

and  represent element-wise or Hadamard multiplication. Note that the vector


ε (r, ω) follows the same stacking as a (r, ω) in (3.19). Replacing a (r, ω) in (4.3)
with (5.8) yields

d̂ (r, ω, ψ) = â (r, ω) ⊗ f (ψ)


= (ε (r, ω) ⊗ 1 M )  d (r, ω, ψ) (5.10)
5.3 Stochastic Error Model 81

where 1 is a column vector with all unity elements and its subscript denotes its length.
The critical equation for the beamformer design formulations in Chap. 4 stems from
the absolute error squared |ξ (r, ω, ψ)|2 which exists in both the weighted LS and TLS
formulations. Hence, in order to incorporate the perturbed array response d̂ (r, ω, ψ)
into those design formulations, the ideal, non-perturbed array response d (r, ω, ψ)
and

Q (r, ω, ψ) = d (r, ω, ψ) dH (r, ω, ψ) (5.11)

in |ξ (r, ω, ψ)|2 needs to be replaced with (5.10). This results in

Q̂ (r, ω, ψ) = d̂ (r, ω, ψ) d̂H (r, ω, ψ)


  
= ε (ω, r) ε H (r, ω) ⊗ 1 M 1TM  Q (r, ω, ψ) . (5.12)

Note that the sensor gain and phase errors can be considered as random variables
and it is the error vector ε (r, ω) that is of interest. Let

E (r, ω) = ε (r, ω) ε H (r, ω) (5.13)

and suppose we want to optimise for the mean performance by using the gain and
phase PDF as weighting functions for the weighted sum of cost functions for all
feasible sensors, i.e.

Ē (r, ω) = E {E (r, ω)}


 
= · · · E (r, ω) f E0 ,...,E K −1 (ε0 , . . . , ε K −1 )dε0 . . . dε K −1 (5.14)

and

ε̄ (r, ω) = E {ε (r, ω)}


 
= · · · ε (r, ω) f E0 ,...,E K −1 (ε0 , . . . , ε K −1 )dε0 . . . dε K −1 (5.15)

where f E0 ,...,E K −1 (ε0 , . . . , ε K −1 ) is the joint PDF for all the sensor’s errors. From now
on, the dependencies (r, ω) are dropped from ε for notational convenience (their
dependencies are understood from the context), and the kth element of a vector is
denoted by [·]k or simply by a subscript k. Assuming independence between errors
from different sensors, then

[ε̄ (r, ω)]∗k = εk f Ek (εk ) dεk (5.16)

where f Ek (εk ) is the PDF of the kth sensor’s error. The entry at the k1 th row and k2 th
column (for k1 = k2 ) in matrix Ē (r, ω) is given by
82 5 Robust Formulation
 
 ∗    
Ē (r, ω) k1 ,k2 = εk1 εk∗2 f Ek1 εk1 f Ek2 εk2 dεk1 dεk2
 
  ∗
 
= εk1 f Ek1 εk1 dεk1 εk2 f Ek2 εk2 dεk2 (5.17)

and for k1 = k2 (diagonal elements),



 

2  
Ē (r, ω) k1 ,k2 =
εk
f k εk dεk = σ 2 (5.18)
1 1 1 1 k1

where σk21 is the second moment of the gain random variable. Let
 
σ = diag σ02 , · · · , σ K2 −1 (5.19)

where diag (·) stacks its parameters into a diagonal matrix. The matrix Ē (r, ω) can
be written as
   
Ē (r, ω) = ε̄ (r, ω) ε̄ H (r, ω)  1 K 1TK − I K + σ (5.20)

where I K is K × K identity matrix. Hence, taking the expectation or mean of the


absolute error squared results in

E |ξ (r, ω, ψ)|2 = wHS B B F (ω) Q̄ (r, ω, ψ) w S B B F (ω)
 
− wHS B B F (ω) Hd (r, ω, ψ) d̄ (r, ω, ψ)
 H
− Hd (r, ω, ψ) d̄ (r, ω, ψ) w S B B F (ω)
+ |Hd (r, ω, ψ)|2 (5.21)

where

Q̄ (r, ω, ψ) = Ēmul (r, ω)  Q (r, ω, ψ) (5.22)


Ēmul (r, ω) = Ē (r, ω) ⊗ 1 M 1TM (5.23)
d̄ (r, ω, ψ) = ε̄ mul (r, ω)  d (r, ω, ψ) (5.24)
ε̄ mul (r, ω) = ε̄ (r, ω) ⊗ 1 M (5.25)

and the subscript mul denotes multiplicative error. If the gain and phase errors are
assumed to be independent, (5.16) can be simplified into
5.3 Stochastic Error Model 83
 
[ε̄ (r, ω)]∗k = ρk exp ( jγk ) f Pk (ρk ) f Γk (γk ) dρk dγk
 
= ρk f Pk (ρk ) dρk cos (γk ) f Γk (γk ) dγk

+ j sin (γk ) f Γk (γk ) dγk (5.26)

and therefore,  
ε̄ (r, ω) = ρ̄  γ̄ c + j γ̄ s (5.27)

where

 
ρ̄ k = ρk f Pk (ρk ) dρk (5.28)

 c
γ̄ k = cos (γk ) f Γk (γk ) dγk (5.29)

 s
γ̄ k = sin (γk ) f Γk (γk ) dγk (5.30)

with f Pk (ρk ) and f Γk (γk ) the PDFs of the gain and phase errors of the kth sensor.
The superscripts c and s in (5.27) are to distinguish between the cosine and sine
terms and show that they can be solved separately. In [6, 8, 23], it is shown that
stochastic error modelling with multiplicative errors is useful for modelling errors
such as mismatches between array elements, errors in sensor positions, and errors in
presumed source positions.
Applying the same formulation process, the following perturbed array response
and error models for design formulations using FIR implementation are obtained,

Q̄ F I R (r, ω, ψ) = Ē F I R,mul (r, ω)  Q F I R (r, ω, ψ) (5.31)


Ē F I R,mul (r, ω) = Ē (r, ω) ⊗ 1 M N 1TM N (5.32)
d̄ F I R (r, ω, ψ) = ε̄ F I R,mul (r, ω)  d F I R (r, ω, ψ) (5.33)
ε̄ F I R,mul (r, ω) = ε̄ (r, ω) ⊗ 1 M N (5.34)

where
Q F I R (r, ω, ψ) = d F I R (r, ω, ψ) dHF I R (r, ω, ψ) . (5.35)

5.3.2 Additive Error

Instead of multiplicative errors, suppose the sensor’s errors are additive due to, for
example, source spreading or local scattering [24], i.e.
84 5 Robust Formulation

â (r, ω) = ε (r, ω) + a (r, ω) . (5.36)

Then, following the same procedure discussed in Sect. 5.3.1, it can be derived that

Q̂ (r, ω, ψ) = d̂ (r, ω, ψ) d̂H (r, ω, ψ)


= (ε (r, ω) ⊗ f (ψ) + d (r, ω, ψ))
× (ε (r, ω) ⊗ f (ψ) + d (r, ω, ψ))H

= Q (r, ω, ψ) + ε (r, ω) ε H (r, ω) + ε (r, ω) aH (r, ω)

+ a (r, ω) ε H (r, ω) ⊗ G (ω, ψ) (5.37)

and

d̂ (r, ω, ψ) = ε (r, ω) ⊗ f (ψ) + d (r, ω, ψ)


= d (r, ω, ψ) + ε (r, ω) ⊗ g (ω, ψ) (5.38)

where G (ω, ψ) = g (ω, ψ) gH (ω, ψ) and g (ω, ψ) = f (ψ). Following the same
procedure to optimise for the mean performance as in Sect. 5.3.1 yields

Q̄ (r, ω, ψ) = Q (r, ω, ψ) + Ēadd (r, ω, ψ) (5.39)


d̄ (r, ω, ψ) = d (r, ω, ψ) + ε̄add (r, ω, ψ) (5.40)

where

Ēadd (r, ω, ψ) = Ē (r, ω) + ε̄ (r, ω) aH (r, ω)

+ a (r, ω) ε̄ H (r, ω) ⊗ G (ω, ψ) (5.41)
ε̄add (r, ω, ψ) = ε̄ (r, ω) ⊗ g (ω, ψ) (5.42)

and the subscript add denotes additive error. Likewise, for FIR implementation, the
following additive error model is obtained

Q̄ F I R (r, ω, ψ) = Q F I R (r, ω, ψ) + Ē F I R,add (r, ω, ψ) (5.43)


d̄ F I R (r, ω, ψ) = d F I R (r, ω, ψ) + ε̄ F I R,add (r, ω, ψ) (5.44)


Ē F I R,add (r, ω, ψ) = Ē (r, ω) + ε̄ (r, ω) aH (r, ω)

+ a (r, ω) ε̄H (r, ω) ⊗ G F I R (ω, ψ) (5.45)
ε̄ F I R,add (r, ω, ψ) = ε̄ (r, ω) ⊗ g F I R (ω, ψ) (5.46)

where G F I R (ω, ψ) = g F I R (ω, ψ) gHF I R (ω, ψ) and g F I R (ω, ψ) = f (ψ) ⊗ e(ω).


5.3 Stochastic Error Model 85

5.3.3 Multiplicative and Additive Error

The derivations in Sects. 5.3.1 and 5.3.2 can be combined to give a general error model
that covers both multiplicative and additive errors. This results in the perturbed array
response given by

â (r, ω) = ε mul (r, ω)  a (r, ω) + εadd (r, ω) . (5.47)

With this perturbed array response, it can be shown that

Q̄ (r, ω, ψ) = Ēmul (r, ω)  Q (r, ω, ψ) + Ēadd (r, ω, ψ) (5.48)


d̄ (r, ω, ψ) = ε̄mul (r, ω) d (r, ω, ψ) + ε̄add (r, ω, ψ) (5.49)

where Ēmul (r, ω), ε̄mul (r, ω), Ēadd (r, ω, ψ), and ε̄add (r, ω, ψ) are defined, respec-
tively, in (5.22), (5.24), (5.41), and (5.42). It should be noted that although the sub-
scripts mul and add distinguish the multiplicative and additive errors, their deriva-
tions are essentially based on the derivations of Ē (r, ω) in (5.14) and ε̄ (r, ω) in
(5.15). The model for FIR implementation has the same form and is given by

Q̄ F I R (r, ω, ψ) = Ē F I R,mul (r, ω)  Q F I R (r, ω, ψ) + Ē F I R,add (r, ω, ψ) (5.50)


d̄ F I R (r, ω, ψ) = ε̄ F I R,mul (r, ω) d F I R (r, ω, ψ) + ε̄ F I R,add (r, ω, ψ) . (5.51)

5.4 Robust Formulation Using Stochastic Error Model

Apart from providing a more meaningful quantification of practical models, this


stochastic approach merges the error modelling into the design formulation itself.
Hence, conventional weighted LS and weighted TLS design techniques, which are
formulated for non-robust designs, can also be used directly in the proposed robust
design methods.

5.4.1 Weighted LS Formulation

In order to incorporate the error model in Sect. 5.3 into the weighted LS beamformer
design formulation in (4.9), its objective function needs to be modified slightly. Let
the new objective function be the weighted sum of mean absolute error squared, i.e.
 

J¯L S (ω) = V (r, ω, ψ) E |ξ (r, ω, ψ)|2 drdψ
Ψ R
= wHS B B F (ω) Q̄ S B B F (ω) w S B B F (ω) − wHS B B F (ω) q̄ S B B F (ω)
− q̄HS B B F (ω) w S B B F (ω) + h S B B F (ω) (5.52)
86 5 Robust Formulation

where
 
Q̄ S B B F (ω) = V (r, ω, ψ) Q̄ (r, ω, ψ) drdψ (5.53)
Ψ R
q̄ S B B F (ω) = V (r, ω, ψ) Hd (r, ω, ψ) d̄ (r, ω, ψ) drdψ. (5.54)
Ψ R

The matrix Q̄ (r, ω, ψ) and vector d̄ (r, ω, ψ) are as defined in Sect. 5.3, depending
on the error model used, i.e. either as multiplicative error, or as additive error, or both.
The design of robust weighted LS SBBF can be achieved by minimising (5.52). Its
analytical solution is given by

w S B B F (ω) = Q̄−1
S B B F (ω) q̄ S B B F (ω) ∀ω ∈ Ω. (5.55)

Likewise, for FIR implementation, its new objective function is given by


  

J¯L S,F I R = V (r, ω, ψ) E |ξ F I R (r, ω, ψ)|2 drdωdψ
Ψ Ω R
= wTS B B F,F I R Q̄ S B B F,F I R w S B B F,F I R − wTS B B F,F I R q̄ S B B F,F I R
− q̄HS B B F,F I R w S B B F,F I R + h S B B F,F I R (5.56)

where
  
Q̄ S B B F,F I R = V (r, ω, ψ) Q̄ F I R (r, ω, ψ) drdωdψ (5.57)
Ψ Ω R
  
q̄ S B B F,F I R = V (r, ω, ψ) Hd (r, ω, ψ) d̄ (r, ω, ψ) drdωdψ. (5.58)
Ψ Ω R

Its analytical solution is given by

w S B B F,F I R = {Q̄ S B B F,F I R }−1 {q̄ S B B F,F I R }. (5.59)

5.4.2 Weighted TLS Formulation

As for the weighted TLS design formulation, define the new objective function as

Ψ R0 V (r, ω, ψ) E{|ξ (r, ω, ψ)| }drdψ
2
¯
JT L S (ω) = (5.60)
Ψ R0 U (r, ω, ψ) E{|H (r, ω, ψ)| }drdψ + 1
2

which gives
5.4 Robust Formulation Using Stochastic Error Model 87
 
Q̄0,S B B F (ω) = U (r, ω, ψ) Q̄ (r, ω, ψ) drdψ. (5.61)
Ψ R0

Then, the design of robust weighted TLS SBBF can be achieved by minimising
(5.60), which can be solved similarly to that described in Sect. 4.3.2 by substituting
Q S B B F (ω) with Q̄ S B B F (ω), Q0,S B B F (ω) with Q̄0,S B B F (ω), and q S B B F (ω) with
q̄ S B B F (ω)
Likewise, for FIR implementation, its objective function with the stochastic error
model is given by

Ψ Ω R0 V (r, ω, ψ) E{|ξ F I R (r, ω, ψ)|2 }drdψ
J¯T L S,F I R = (5.62)
Ψ R0 U (r, ω, ψ) E{|HF I R (r, ω, ψ)|2 }drdωdψ + 1

and
 
Q̄0,S B B F,F I R = U (r, ω, ψ) Q̄ F I R (r, ω, ψ) drdωdψ. (5.63)
Ψ R0

5.5 Performance Metrics

For evaluating and comparing between non-robust and robust beamformer designs,
the performance error (4.54) is used for both LS- and TLS-based beamformer designs.
The reason for removing the denominator term (c.f. (4.55)) when evaluating the TLS-
based beamformers is to remove the scaling effect due to different denominator values
between non-robust and robust TLS designs. In addition to the performance error,
the directivity in (4.56) and the array gain in (4.57) are also used. Note that the array
gain is also related to sensitivity or tolerance factor of a beamformer against errors
and perturbations, where such sensitivity is defined as [26]
1
Tse (ω) = . (5.64)
Aw (ω)

Equation (5.64) shows that as the array gain increases, sensitivity decreases, which
translates to better robustness against errors and perturbations.

5.6 Design Examples and Evaluation

5.6.1 Design Specifications

In order to illustrate the robustness achieved by using the stochastic error model,
a number of robust SBBF design examples are presented and compared with their
88 5 Robust Formulation

Table 5.1 Number of uniform discretisation points for numerical evaluation of integrals
Parameter Number of points
Frequency range, Ω 256
Steering range, Ψ 73
Azimuth angle range, Φ 181

Table 5.2 Design examples for each design formulation


Design formulation Name Value of r Description
Weighted LS (Eq. 5.59) LS f ar 100 Farfield LS design
Weighted TLS (Eq. 5.62) TLS f ar 100 Farfield TLS design

corresponding non-robust counterparts in Sect. 4.7. The same spiral array as described
in Sect. 2.5 and the design specifications as listed in Table 4.1 are used for designing
the robust beamformers.
The stochastic error model introduces additional complexity into the integrals in
the robust design formulation. As such, the integrals are difficult to solve analytically
and they are approximated by uniformly spaced Riemann sum with the number of
discretisation points as specified in Table 5.1. This numerical approach in approxi-
mating the integrals causes the design problem size to be large. As such, only robust
farfield SBBFs given by Table 5.2 are provided.
For the robust designs, both the multiplicative-only and additive-only error models
are used. The errors in all sensors are assumed to follow the same PDF model, which
is independent of both frequency ω and azimuth angle φ, with the gain and phase
error PDFs given by

N (1, 0.05) , ρk ≥ 0
f Pk (ρk ) = (5.65)
0 , otherwise
f Γk (γk ) = U (−0.05 rad, 0.05 rad) (5.66)

where N (μ, σ ) is the Gaussian PDF with mean μ and standard deviation σ and
U (a, b) is the uniform PDF with minimum value a and maximum value b. Note that
(5.65) is essentially a cropped Gaussian PDF. For comparison purposes, the same
PDF model of (5.65) and (5.66) is used in both multiplicative-only and additive-only
robust designs. In practice, the PDF model used should match the perturbation model
of the target applications.

5.6.2 Array Gain and Sensitivity

The array gain (as well as the sensitivity) of the designed beamformers is shown in
Fig. 5.1. The figure shows that the low array gain at low frequencies for both the
5.6 Design Examples and Evaluation 89

20

−20
Array gain, A w (dB)

TLS robust
TLS non−robust
−40
LS robust
LS non−robust

−60

−80

−100
0 500 1000 1500 2000 2500 3000 3500 4000
Frequency, f (Hz)

Fig. 5.1 Array gain for the designed beamformers

non-robust LS and TLS designs has been improved in their robust counterparts.
The array gain for both the robust LS and TLS designs is more consistent across
frequencies, suggesting consistent SNR improvement across frequencies. In terms
of sensitivity, the improved array gain in the robust designs suggests that both the
robust designs are less sensitive (i.e. more robust) towards mismatches, errors, and
perturbations.

5.6.3 Perturbation in Sensor Characteristics

In order to evaluate the robustness performance, the following perturbation model is


simulated and introduced into all sensors. The ideal sensor response is assumed to be
a bandpass filter with unity gain and linear phase shift within the spectral passband.
This response is then modelled with a 64-tap FIR filter, which will introduce a phase
delay into the sensor response. The ideal filter coefficients br (k, l) are then perturbed
with a uniform random variable as in

b̂r (k, l) ∼ br (k, l) + U (−0.1, 0.1) (5.67)

where b̂r (k, l) is the perturbed lth filter coefficient of the kth sensor. Figure 5.2 shows
the perturbed sensor responses, where each line corresponds to the response for each
sensor. Here, it is noted that the perturbation model (5.67) is pessimistic relative to
the actual sensor response from the calibration graph provided by the manufacturers.
90 5 Robust Formulation

(a) 5

−5

−10
Magnitude (dB)

−15

−20

−25

−30

−35
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Frequency, ω (xπ rad/s)

(b)
−20

−21

−22

−23
Bulk delay (number of samples)

−24

−25

−26

−27

−28

−29

−30
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Frequency, ω (xπ rad/s)

Fig. 5.2 a Magnitude and b phase delay for perturbed sensor responses
5.6 Design Examples and Evaluation 91

100
TLS robust
TLS non−robust
LS robust
LS non−robust
80

60
(dB)
LS
Performance error, ξ

40

20

Designed steering range, Ψ


−20
−80 −60 −40 −20 0 20 40 60 80
Steering angle, ψ (degree)

Fig. 5.3 Averaged performance error with perturbation in sensor characteristics for non-robust and
robust multiplicative-only designs

The reason for choosing this model is that it is simple, and if the design is robust
against such perturbations, then they will most likely be robust against the actual
perturbations and mismatches in real sensors.
The performance error for the design examples with this perturbation is shown
in Fig. 5.3, where each plot is obtained by averaging the performance error from
50 different realisations of the perturbation model in (5.67). It is clear from this
figure that robustness is achieved in the designs with the stochastic error model. The
trade-off for achieving this robustness is the increased performance error relative
to the ideal situation (without perturbation) as shown in Fig. 5.4. This trade-off is
typical in any robust design. A further highlight of the achieved robustness using
the stochastic error model is illustrated in Figs. 5.5 and 5.6, where the directivity,
with perturbation, for both non-robust and robust LS and TLS designs, are shown.
From these figures, the robust designs successfully maintain their directivity in the
presence of perturbations, unlike their non-robust counterparts.

5.6.4 Perturbation in Sensor Positions

The robustness achieved in the design examples is not limited to perturbation in the
sensor characteristics, but also to other perturbations such as in the sensor positions.
92 5 Robust Formulation

40
TLS robust
TLS non−robust
LS robust
LS non−robust
30

20
Performance error, ξLS (dB)

Designed steering range, Ψ

10

−10

−20
−80 −60 −40 −20 0 20 40 60 80
Steering angle, ψ (degree)

Fig. 5.4 Performance error without perturbation for non-robust and robust multiplicative-only
designs

Errors in sensor position cause variations in gain and phase delay of the signal arriving
at the sensor, which fit readily into the multiplicative error model.
Here, the same design examples are evaluated in the presence of errors in sensor
positions. The sensor positions are perturbed
 within a circular region (with the radius
given by the Gaussian distribution N 0, (0.001)2 ) around their nominal values,
and the perturbed positions (in x–y coordinate) are given by

     
r̂k ∼ xk + N 0, (0.001)2 cos (U (0, π )) , yk + N 0, (0.001)2 sin (U (0, π ))
(5.68)
where [xk , yk ] is the nominal position of the kth sensor in x–y coordinate. Figure 5.7
shows the performance error for the robust and non-robust designs in the presence of
perturbation in sensor positions. Each plot is obtained by averaging the performance
error from 50 different realisations of the perturbation model in (5.68). In the presence
of perturbation, the robust designs still achieve low performance error, suggesting
that they still work under the introduced perturbations in the sensor positions.
5.6 Design Examples and Evaluation 93

(a) 15

3500

3000
10

Directivity index, DI (dB)


2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

(b) 15

3500

3000
10 Directivity index, DI (dB)

2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 5.5 DIs with perturbation in sensor characteristics for non-robust a LS and b TLS designs
94 5 Robust Formulation

15
(a)
3500

3000
10

Directivity index, DI (dB)


2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

15
(b)
3500

3000
10
Directivity index, DI (dB)

2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 5.6 DIs with perturbation in sensor characteristics for robust multiplicative-only a LS and b
TLS designs
5.6 Design Examples and Evaluation 95

100
TLS robust
TLS non−robust
LS robust
LS non−robust
80

60
Performance error, ξLS (dB)

40

20

Designed steering range, Ψ


−20
−80 −60 −40 −20 0 20 40 60 80
Steering angle, ψ (degree)

Fig. 5.7 Averaged performance error with perturbation in sensor positions for non-robust and
robust multiplicative-only designs

5.6.5 Perturbation Due to Local Scattering

In order to evaluate the design examples against additive error model, errors due to
local scattering are considered [24]. In this perturbation model, additional propaga-
tion paths from signal source to the sensor array are present in addition to the direct
line-of-sight propagation path as shown in Fig. 5.8. The array element response with
such perturbation model is given by

1
L
 (r, k, ω) = A (r, k, ω) + ρi A (ri , k, ω) (5.69)
L i=1

where ri = (r, φ + φi ), L is the number of propagation paths, φi is the variation in


angle of arrival, and ρi is the variation in gain of the impinging signal due to local
scattering. Both the variations φi and ρi are taken to be
 π π
φi ∼ U − , (5.70)
9 9
ρi ∼ Rayleigh (0.01) (5.71)

where Rayleigh (σ ) is the Rayleigh PDF with scaling parameter σ .


96 5 Robust Formulation

Direct path
in ideal case
φ φi
Additional path due
to local scattering

Sensor array

Point source

Fig. 5.8 Model used for perturbation due to local scattering

100
TLS robust
TLS non−robust
LS robust
LS non−robust
80

60
Performance error, ξLS (dB)

40

20

Designed steering range, Ψ


−20
−80 −60 −40 −20 0 20 40 60 80
Steering angle, ψ (degree)

Fig. 5.9 Average performance error with local scattering perturbation for non-robust and robust
additive-only designs

Figure 5.9 shows the performance error for the design examples, where each plot
is obtained by averaging the performance error from 50 different realisations of the
perturbation model in (5.69) to (5.71). As expected, the robust additive-only designs
5.6 Design Examples and Evaluation 97

40
TLS robust
TLS non−robust
LS robust
LS non−robust
30

20
Performance error, ξLS (dB)

Designed steering range, Ψ

10

−10

−20
−80 −60 −40 −20 0 20 40 60 80
Steering angle, ψ (degree)

Fig. 5.10 Performance error without perturbation for non-robust and robust additive-only designs

have lower performance error as compared to their non-robust counterparts in the


presence of the simulated perturbation. Similarly, the trade-off for achieving this
robustness is the increased performance error in the absence of perturbation as shown
in Fig. 5.10. Figures 5.11 and 5.12 show the directivity with simulated local scattering
for the non-robust and robust additive-only designs, respectively. It is evident that
the directivity for the robust additive-only designs is maintained in the presence of
simulated local scattering.

5.7 Conclusions

In conclusion, stochastic error models offer an effective approach for modelling real-
world perturbations and errors into a robust beamformer design formulation. This is
because in this formulation, errors are modelled as random variables, which makes
sense since real-world perturbations can be considered as random. Hence, this error
model can capture the stochastic properties of the errors to be integrated into the
design model, where the errors are weighted by their rate of occurrence or PDFs.
98 5 Robust Formulation

15
(a)
3500

3000
10

Directivity index, DI (dB)


2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

15
(b)
3500

3000
10
Directivity index, DI (dB)

2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 5.11 Directivity with simulated local scattering for non-robust a LS and b TLS designs
5.7 Conclusions 99

15
(a)
3500

3000
10

Directivity index, DI (dB)


2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

15
(b)
3500

3000
10
Directivity index, DI (dB)

2500
Frequency, f (Hz)

2000 5

1500

0
1000

500

−5
−60 −40 −20 0 20 40 60
Steering angle, ψ (degree)

Fig. 5.12 Directivity with simulated local scattering for robust additive-only a LS and b TLS
designs
100 5 Robust Formulation

The design optimisation in this chapter provides a good way to achieve robustness
in the sense that the designs are optimised for the mean performance, where the
errors are more likely to happen. This provides a fair balance between robustness
and performance as it is not as pessimistic as the method of optimising for the
worst-case error. Furthermore, the method of optimising for the mean performance
effectively embeds the error model into the beamformer design models. Hence, the
robust design formulations can be extended, modified, and solved in similar ways as
their non-robust counterparts.
In addition, as most practical errors can be translated into errors occurring during
the sampling of signals by a sensor, they can be modelled as either multiplicative
complex error, or additive complex error, or both. This error modelling provides a
better connection between the real-world error and design model, unlike the WNG
method where such connection is vague. Hence, a more quantitative robustness spec-
ification is possible with this error model.

References

1. J. Bitzer, K.U. Simmer, in Superdirective Microphone Arrays, ed. by M. Brandstein, D. Ward.


Microphone Arrays. (Springer, Berlin, 2001), ch. 2, pp. 19–38
2. H. Cox, R. Zeskind, T. Kooij, Practical supergain. IEEE Trans. Acoust. Speech Signal Process.
34(3), 393–398 (1986). Jun
3. S. Yan, Y. Ma, Robust supergain beamforming for circular array via second-order cone pro-
gramming. Appl. Acoust. 66(9), 1018–1032 (2005)
4. S. Doclo, M. Moonen, Superdirective beamforming robust against microphone mismatch. IEEE
Trans. Audio Speech Lang. Process. 15(2), 617–631 (2007). Feb
5. E. Mabande, A. Schad, W. Kellermann, Design of robust superdirective beamformers as a
convex optimization problem, in Proceedings of IEEE International Conference on Acoustics
Speech Signal Processing (ICASSP), Taipei, Taiwan, 19–24 April 2009, pp. 77–80
6. S. Doclo, M. Moonen, Design of broadband beamformers robust against gain and phase errors
in the microphone array characteristics. IEEE Trans. Signal Process. 51(10), 2511–2526 (2003).
Oct
7. J. Sachar, H. Silverman, W. Patterson, Microphone position and gain calibration for a large-
aperture microphone array. IEEE Trans. Speech Audio Process. 13(1), 42–52 (2005). Jan
8. S. Doclo, M. Moonen, Design of broadband beamformers robust against microphone position
errors, in Proceedings of International Workshop Acoustics Echo, Noise Control, Kyoto, Japan,
8–11 Sep 2003, pp. 267–270
9. J. Ringelstein, A. Gershman, J. Bohme, Direction finding in random inhomogeneous media in
the presence of multiplicative noise. IEEE Signal Process. Lett. 7(10), 269–272 (2000). Oct
10. Y.J. Hong, C.C. Yeh, D.R. Ucci, The effect of a finite-distance signal source on a far-field
steering Applebaum array-two dimensional array case. IEEE Trans. Antennas Propag. 36(4),
468–475 (1988). Apr
11. R. Goossens, I. Bogaert, H. Rogier, Phase-mode processing for spherical antenna arrays with
a finite number of antenna elements and including mutual coupling. IEEE Trans. Antennas
Propag. 57(12), 3783–3790 (2009). Dec
12. J. Meyer, Microphone array for hearing aids taking into account the scattering of the head, in
IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, New Platz, NY,
21–24 Oct 2001, pp. 27–30
13. J. Goldberg, H. Messer, Inherent limitations in the localization of a coherently scattered source.
IEEE Trans. Signal Process. 46(12), 3441–3444 (1998). Dec
References 101

14. O. Besson, P. Stoica, Decoupled estimation of DOA and angular spread for a spatially distributed
source. IEEE Trans. Signal Process. 48(7), 1872–1882 (2000). Jul
15. D. Astely, B. Ottersten, The effects of local scattering on direction of arrival estimation with
MUSIC. IEEE Trans. Signal Process. 47(12), 3220–3234 (1999). Dec
16. W. Liu, D. McLernon, M. Ghogho, Design of frequency invariant beamformer without temporal
filtering. IEEE Trans. Signal Process. 57(2), 798–802 (2009). Feb
17. C.C. Lai, S. Nordholm, Y.H. Leung, Design of robust steerable broadband beamformers with
spiral arrays and the farrow filter structure, in Proceedings of International Workshop on
Acoustic Echo and Noise Control, Tel Aviv, Israel, 30 Aug–2 Sep 2010
18. E.N. Gilbert, S.P. Morgan, Optimum design of directive antenna arrays subject to random
deviations. Bell Syst. Technol. J. 34, 637–663 (1955). May
19. A. Elnashar, S. Elnoubi, H. El-Mikati, Further study on robust adaptive beamforming with
optimum diagonal loading. IEEE Trans. Antennas Propag. 54(12), 3647–3658 (2006). Dec
20. S.A. Vorobyov, A.B. Gershman, L. Zhi-Quan, Robust adaptive beamforming using worst-case
performance optimization: a solution to the signal mismatch problem. IEEE Trans. Signal.
Process. 51(2), 313–324 (2003). Feb
21. S. Vorobyov, H. Chen, A. Gershman, On the relationship between robust minimum variance
beamformers with probabilistic and worst-case distortionless response constraints. IEEE Trans.
Signal. Process. 56(11), 5719–5724 (2008). Nov
22. Z.L. Yu, W. Ser, M.H. Er, Z. Gu, Y. Li, Robust adaptive beamformers based on worst-case
optimization and constraints on magnitude response. IEEE Trans. Signal. Process. 57(7), 2615–
2628 (2009). Jul
23. C.C. Lai, S. Nordholm, Y.H. Leung, Design of robust steerable broadband beamformers incor-
porating microphone gain and phase error characteristics, in IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP), Prague, Czech Rep., 22–27 May 2011,
pp. 101–104
24. D. Astely, B. Ottersten, The effects of local scattering on direction of arrival estimation with
music. IEEE Trans. Signal Process. 47(12), 3220–3234 (1999)
25. S. Yan, H. Sun, U. Svensson, X. Ma, J. Hovem, Optimal modal beamforming for spherical
microphone arrays. IEEE Trans. Audio Speech Lang. Process. 19(2), 361–371 (2011). Feb
26. H.L.V. Trees, Optimum Array Processing (Wiley, 2004)
Chapter 6
Conclusions and Future Work

Abstract Beamforming is a signal processing technique which allows signals orig-


inating from non-overlapping spatial location to be selectively processed to achieve
desired signal enhancements. Beamformers’ design process is a two-step process,
i.e. (1) modelling of the underlying signal source, acoustic environment, and sensor
geometry of target applications and (2) careful selection of the beamformer weights
to achieve a desired response. A proper design allows properties of chosen sensor
geometry and beamformer structure to be exploited. Additionally, some degree of
robustness is normally designed into the beamformer in order to account for prac-
tical imperfection in real-world applications. Although the main focus of this book
is on robust SBBF design formulations, a number of interesting research directions
can be extended from this work, e.g. optimisation of array geometry, optimisation
of steering function, adaptive SBBFs, and tracking SBBFs, to name a few. With
beamforming technology pioneering into more and more practical applications and
advancement in low-cost DSP hardware, it is expected that beamforming will attract
a growing research interest.

Keywords Acoustic source tracking · Adaptive beamforming

6.1 Summary

The design of beamformer weights, both for fixed and for steerable beamformers,
requires the underlying signal source and acoustic environment of the target appli-
cation to be identified and modelled. In general, there is no single generic model that
works best for any application and the models are highly dependent on the target
applications. As such, utmost care is needed to provide these mathematical models
before designing the beamformer weights. The main aim of modelling the signal
source and acoustic environment is to capture the essential information on the prop-
erties of the signal source, propagation medium, and how the signal energy is being
transferred from a source to a receiver. In most cases, reasonable assumptions are
made to significantly simplify the underlying mathematical models.

© The Author(s) 2017 103


C.C. Lai et al., A Study into the Design of Steerable Microphone Arrays,
SpringerBriefs in Signal Processing, DOI 10.1007/978-981-10-1691-2_6
104 6 Conclusions and Future Work

Once the mathematical models are defined, a beamformer design process involves
selecting beamformer weights such that a certain desired response is achieved. This
selection can be done by optimising the weights such that the error between actual
and desired responses is minimised. This weight selection by means of optimisation
depends on how the error is calculated (e.g. LS or TLS) and different optimisa-
tion approaches (e.g. constrained or unconstrained optimisations and minimising the
maximum error), which results in beamformers with different performances. One
of the challenges in these approaches is that the design problem size can be large,
especially for SBBFs due to the additional steering dimension. In some cases, the
inherent properties of the design formulations can be exploited, such as (1) exploit-
ing the symmetry nature of the matrices in the design formulations and (2) solving
analytically two of the integrations in weighted TLS and LS design formulations.
In practice, the mathematical models used for a beamformer design do not always
capture mismatches, uncertainties, and errors in real-world environment. As such,
robustness in beamformers is necessary to ensure that the designed beamformers
will work when deployed into practical environments. One technique to achieve such
robustness is the stochastic approach where these practical imperfections are mod-
elled as random variables. The main benefit of using such a model is that the design
is optimised for the mean performance, which is more likely the operating condition
in practice compared to the pessimistic approach of optimising for the worst-case
scenario. This stochastic model is embedded directly into the mathematical model
of the sensor arrays, which means that the same non-robust design formulation can
be used to achieve robustness.

6.2 Future Work

A number of future directions that can be further pursued from the discussion of this
book are as follows:

1. Tracking beamformer
One of the interesting extensions of the work presented is to integrate a source
detection and tracking algorithm [1–3] together with SBBFs for automatic audio
reception with source-tracking capability [4]. Such integration is depicted in
Fig. 6.1 where the output of the source-tracking system, normally in terms of
estimated source location or direction, is used to steer the main beam towards the
direction of the signal source for audio reception. This automation releases the
necessity for a human operator in audio acquisition and recording applications
such as smart homes and robots.
2. Investigation on the steering function
Proper selection of the steering function f (m, ψ), as well as array geometry for
SBBFs, results in desirable characteristics that can be exploited in the design of
SBBFs (see Sect. 4.5). Such characteristics lead to a reduced design problem size
yet achieve full steering range. In the light of this, detailed investigation on the
6.2 Future Work 105

Fig. 6.1 Tracking Sensor array


beamformer system
···

.. Source
. tracking

Estimated
location

.. Steerable Output signal


.
beamformer

steering function can be performed to obtain a steering function that is matched to


the underlying physical properties of the array geometry and beamformer struc-
ture. This allows for a steering mechanism that does not impose any extra cost
on the design and implementation of the beamformer. One such approach is to
first decompose the received signal using some basis functions and then modu-
late these decomposed signals to achieve desired beam steering. For example, in
spherical arrays, beam steering can be achieved by modulating each individual
spherical harmonic component decomposed from the received signal [5–8].
3. Investigation on optimum sensor placement
Section 2.5 has shown that the selection of array geometry plays an important
role in the overall beamformer design. However, the main focus of this book is on
the beamformer weights design and not on the array geometry design. Therefore,
research on optimum array geometry can be extended from this work in order to
further improve the beamformer performances.
4. Constrained and min-max design formulation
The design formulations presented are based on unconstrained LS and TLS opti-
misations. Hence, there is no control on sidelobes level or the tolerance in the pass
region. Clearly, the design formulations presented can be extended to include con-
straints for sidelobes and tolerance in the pass band. In addition, different design
formulations such as constrained minimax optimisation can also be further inves-
tigated.
5. Adaptive SBBFs
The design of SBBFs presented is non-adaptive and data-independent. If the
characteristics of the received signals are known, this information can be used to
further improve the reception quality of the desired signals by means of adaptive
beamforming. Ultimately, this leads to adaptive SBBFs, where the beamformer
weights are updated progressively to approach optimality and to track a dynamic
source. However, one of the major problems in adaptive beamforming is the
106 6 Conclusions and Future Work

cancellation of target signal caused by reverberation or mismatches between the


presumed and the actual models [9]. This is an interesting and challenging research
topic.

6.3 Final Remarks

The current trend for sensor array processing is targeting low-cost and low-sensor-
count array. A lot of emphasis is put on signal processing to achieve required per-
formance with limited hardware. As such, beamforming is normally accompanied
with other single-channel or multichannel signal processing (depending on the target
applications) for further performance boost [10, 11]. However, with beamforming
and other signal processing algorithms maturing, manufacturers are looking at other
approach for performance boost. Since the cost for DSP computations continues to
drop, DSP cost is no longer a bottleneck to impede on beamforming technology.
This trend will spark more research interest in array signal processing including
beamforming.

References

1. E.A. Lehmann, Particle filtering methods for acoustic source localisation and tracking. Ph.D.
dissertation, Department of Telecommunications Engineering, The Australian National Uni-
versity, 2004
2. E.A. Lehmann, A.M. Johansson, Particle filter with integrated voice activity detection for
acoustic source tracking. EURASIP J. Adv. Signal Process. 2007(1), 28–38 (2007)
3. A.M. Johansson, Acoustic sound source localisation and tracking in indoor environments.
Ph.D. dissertation, School of Engineering, Blekinge Institute of Technology, 2008
4. S. Timofeev, A.R.S. Bahai, P. Varaiya, Adaptive acoustic beamformer with source tracking
capabilities. IEEE Trans. Signal Process. 56(7), 2812–2820 (2008)
5. J. Meyer, G. Elko, A highly scalable spherical microphone array based on an orthonormal
decomposition of the soundfield, in Proceedings of the IEEE International Conference on
Acoustics Speech Signal Processing (ICASSP), vol. 2, Orlando, FL, 13-17 May 2002, pp.
1781–1784
6. T.D. Abhayapala, D.B. Ward, Theory and design of high order sound field microphones using
spherical microphone array, in Proceedings of the IEEE International Conference on Acoustics
Speech Signal Processing (ICASSP), vol. 2, Orlando, FL, 13–17 May 2002, pp. 1949–1952
7. Z. Li, R. Duraiswami, Flexible and optimal design of spherical microphone arrays for beam-
forming. IEEE Trans. Audio Speech Lang. Process. 15(2), 702–714 (2007)
8. C.C. Lai, S. Nordholm, Y.H. Leung, Design of steerable spherical broadband beamformers
with flexible sensor configurations. IEEE Trans. Audio Speech Lang. Process. 21(2), 427–438
(2013)
9. O. Hoshuyama, A. Sugiyama, Robust adaptive baemforming, in Microphone Arrays, ed. by
M. Brandstein, D. Ward (Springer, Berlin, 2001), Chap. 5, pp. 87–110
10. M. Brandstein, D. Wards (eds.), Microphone Arrays—Signal Processing Techniques and Appli-
cations (Springer, Berlin, 2001)
11. J. Benesty, J. Chen, Y. Huang, Microphone Array Signal Processing, vol. 1 (Springer Science
& Business Media, 2008)
Appendix A
Closed Form Integration for Fixed
Beamformer Design


A.1 Solution for ·dω

The design formulations for fixed beamformers in Chap. 3 involve the integration
w.r.t. the operating frequency ω. For the FIR implementation design formulations,
a closed form solution for this integration is possible if: (1) the weighting functions
are independent of frequency, ω, and (2) the desired response is taken to be the ideal
brick wall response as given by (3.70).
Under these two assumptions, (3.41) can be simplified to
 
QFIR = V (r) aFIR (r, ω) aH
FIR (r, ω) dωdr (A.1)
R Ω

= V (r) AFIR (r) dr (A.2)
R

and (3.42) to
 
qFIR = V (r) Hd (r, ω) aFIR (r, ω) dωdr (A.3)
R Ω

= V (r) qFIR (r) dr (A.4)


R

where

AFIR (r) = aFIR (r, ω) aFIR
H
(r, ω) dω (A.5)
Ω

qFIR (r) = Hd (r, ω) aFIR (r, ω) dω. (A.6)
Ω

It is possible to obtain closed form solutions for both (A.5) and (A.6). Let

AFIR (r, ω) = aFIR (r, ω) aFIR


H
(r, ω) , (A.7)
© The Author(s) 2017 107
C.C. Lai et al., A Study into the Design of Steerable Microphone Arrays,
SpringerBriefs in Signal Processing, DOI 10.1007/978-981-10-1691-2
108 Appendix A: Closed Form Integration for Fixed Beamformer Design

the element of matrix AFIR (r, ω) at the lr th row and the lc th column is given by

r2
[AFIR (r, ω)]lr ,lc = ×
rk1 − rrk2 − r
  
fs  
exp jω rk1 − r − rk2 − r + (n1 − n2 ) (A.8)
c
= ρa (lr , lc ) exp (jωγa (lr , lc )) (A.9)

where

lr = k1 N + n1 (A.10)
lc = k2 N + n2 (A.11)
r2
ρa (lr , lc ) = (A.12)
rk − rrk2 − r
 1 
fs  
γa (lr , lc ) = rk1 − r − rk2 − r + (n1 − n2 ) (A.13)
c

Suppose that the integration region is taken as

Ω = Ωpb ∪ Ωsb (A.14)

where the spectral passband Ωpb and stopband Ωsb are given by

(1) (2)
Ωpb = {ω : −π ≤ ω ≤ π, ωpb ≤ |ω| ≤ ωpb } (A.15)
(1) (2)
Ωsb = {ω : −π ≤ ω ≤ π, |ω| ≤ ωsb , ωsb ≤ |ω| ≤ π } (A.16)

(1) (1) (2) (2)


such that 0 < ωsb < ωpb < ωpb < ωsb < π . Hence, for lr = lc ,

[AFIR (r)]lr ,lc = ρa (lr , lc ) exp (jωγa (lr , lc )) dω (A.17)
Ω

(1) (1)
= 2ρa (lr , lc ) ωsb sinc ωsb γa (lr , lc ) +



(2) (2) (1) (1)
ωpb sinc ωpb γa (lr , lc ) − ωpb sinc ωpb γa (lr , lc ) +


(2) (2)
π sinc (π γa (lr , lc )) − ωsb sinc ωsb γa (lr , lc ) . (A.18)

and for lr = lc ,



(1) (2) (1) (2)
[AFIR (r)]lr ,lc = 2ρa (lr , lc ) ωsb + ωpb − ωpb + π − ωsb . (A.19)

Note that the matrix AFIR (r) is Hermitian and any of its N × N submatrices
are toeplitz. These properties can be exploited to reduce computational load in
Appendix A: Closed Form Integration for Fixed Beamformer Design 109

populating this matrix and solving for beamformer weights. The Hermitian prop-
erty is as expected since the FIR weights are real.
For the integration of A.6, its analytical integration is given by



(2) (2) (1) (1)
qFIR (r) l = 2ρq (l) ωpb sinc ωpb γq (l) − ωpb sinc ωpb γq (l) (A.20)

where

l = kN + n (A.21)
r
ρq (l) = (A.22)
r − r
 k  
fs N −1
γq (l) = (rk − r − r) + n − . (A.23)
c 2
Appendix B
Closed Form Integrations for Steerable
Beamformer Design


B.1 Solution for ·dω

Similarly for the steerable beamformers in Chap. 4, a closed form solution for the
integration w.r.t. ω can be obtained if: (1) the weighting functions are independent of
frequency ω, and (2) the desired response is taken to be the ideal brick wall response
as given by (4.59). Under these assumptions, (4.30) can be simplified to
  
QSBBF,FIR = V (r, ψ) [aFIR (r, ω) ⊗ f (ψ)] [aFIR (r, ω) ⊗ f (ψ)]H drdωdψ
Ψ Ω R(ψ)
(B.1)
   

= V (r, ψ) aFIR (r, ω) aFIR
H
(r, ω) dω ⊗ f (ψ) f H (ψ) drdψ
Ψ R(ψ) Ω
(B.2)
 

= V (r, ψ) AFIR (r) ⊗ f (ψ) f H (ψ) drdψ (B.3)
Ψ R(ψ)

and (4.31) can be simplified to


  
qSBBF,FIR = V (r, ψ) Hd (r, ω, ψ) [aFIR (r, ω) ⊗ f (ψ)] drdωdψ
Ψ Ω R(ψ)
(B.4)
   
= V (r, ψ) Hd (r, ω, ψ) aFIR (r, ω) dω ⊗ f (ψ) drdψ
Ψ R(ψ) Ω
(B.5)
 
= V (r, ψ) qFIR (r, ψ) ⊗ f (ψ) drdψ (B.6)
Ψ R(ψ)

© The Author(s) 2017 111


C.C. Lai et al., A Study into the Design of Steerable Microphone Arrays,
SpringerBriefs in Signal Processing, DOI 10.1007/978-981-10-1691-2
112 Appendix B: Closed Form Integrations for Steerable Beamformer Design

where AFIR (r) is as given by (A.5) and



qFIR (r, ψ) = Hd (r, ω, ψ) aFIR (r, ω) dω. (B.7)
Ω

The closed form solution for AFIR (r) is given by (A.18) and (A.19), and from
(A.20) the closed form solution for (B.7) is given by

qFIR (r) , φ ∈ Φpb (ψ)
qFIR (r, ψ) = . (B.8)
0 , φ ∈ Φsb (ψ)


B.2 Solution for ·dψ

A closed form solution for the integration w.r.t. ψ in the design formulation of
steerable beamformer can also be obtained if the weighting function V (r,ψ) is
separable into V (r, ψ) = V1 (r) V2 (ψ). Specifically, the aim was to solve ·dψ
in (B.3) and (B.6) without solving any other integrations. This requires reordering
the integrals in those two equations. From here onwards, the notation R(ψ) ·dr is
 
explicitly written as R Φ(ψ) ·dφdr.
From the steerable beamformer specifications in Table 4.1,

Ψ = {ψ : ψ (1) ≤ ψ ≤ ψ (2) } (B.9)


Φ (ψ) = Φpb (ψ) ∪ Φsb (ψ) (B.10)
BWφ
Φpb (ψ) = {φ : |φ − ψ| ≤ } (B.11)
2
BWφ
Φsb (ψ) = {φ : |φ − ψ| ≥ + T Wφ } (B.12)
2

where ψ (1) ≤ ψ (2) . These regions are depicted graphically in Fig. B.1 for the case
ψ (2) − ψ (1) > BWφ and in Fig. B.2 for ψ (2) − ψ (1) ≤ BWφ , with the regions in green
for Φpb (ψ) and red for Φsb (ψ).
Due to the definition of Hd (r, ω, ψ) in (4.59), Eq. (B.8) is non-zero only at Φ =
Φpb (ψ). This region Φpb (ψ) is depicted by the green region in Figs. B.1 and B.2.
For the case of ψ (2) − ψ (1) > BWφ in Fig. B.1, the three parts of Φpb (ψ) are given,
respectively, by

BWφ BWφ
Part 1: Φ1 = {φ : ψ (1) − ≤ φ ≤ ψ (1) + },
2 2
BWφ
Ψ1 (φ) = {ψ : ψ (1) ≤ ψ ≤ φ + }; (B.13)
2
BWφ BWφ
Part 2: Φ2 = {φ : ψ (1) + < φ ≤ ψ (2) − },
2 2
Appendix B: Closed Form Integrations for Steerable Beamformer Design 113

Fig. B.1 Integration region bounded by Φ (ψ) and Ψ for ψ (2) − ψ (1) > BWφ

Fig. B.2 Integration region bounded by Φ (ψ) and Ψ for ψ (2) − ψ (1) ≤ BWφ
114 Appendix B: Closed Form Integrations for Steerable Beamformer Design

BWφ BWφ
Ψ2 (φ) = {ψ : φ − ≤ψ ≤φ+ }; (B.14)
2 2
BWφ BWφ
Part 3: Φ3 = {φ : ψ (2) − < φ ≤ ψ (2) + },
2 2
BWφ
Ψ3 (φ) = {ψ : φ − ≤ ψ ≤ ψ (2) }. (B.15)
2

Substituting (B.13)–(B.15) into (B.8) results in


    
qSBBF,FIR = qFIR (r, φ) ⊗ f (ψ) dψ dφdr (B.16)
R Φ Ψ (φ)
    
= qFIR (r, φ) ⊗ f (ψ) dψ dφ +
R Φ1 Ψ1 (φ)
   
qFIR (r, φ) ⊗ f (ψ) dψ dφ +
Φ2 Ψ2 (φ)
    
qFIR (r, φ) ⊗ f (ψ) dψ dφ dr. (B.17)
Φ3 Ψ3 (φ)

Now, it is possible to solve for



f (φ) = f (ψ) dψ (B.18)
Ψ (φ)

in (B.16). Suppose that the steering function f (m, ψ) is as defined in (4.53), the
element in the mth row of f (φ) is given by
 g2 (φ)  m
ψ
[f (φ)]m = dψ (B.19)
g1 (φ) α
(g2 (φ))m+1 − (g1 (φ))m+1
= . (B.20)
(m + 1) α m

The same integral, but for the case ψ (2) − ψ (1) ≤ BWφ , can be solved similarly.
The difference is that its integration region (see Fig. B.2) is given by

BWφ BWφ
Part 1: Φ1 = {φ : ψ (1) − ≤ φ ≤ ψ (2) − },
2 2
BWφ
Ψ1 (φ) = {ψ : ψ (1) ≤ ψ ≤ φ + }; (B.21)
2
BWφ BWφ
Part 2: Φ2 = {φ : ψ (2) − < φ ≤ ψ (1) + },
2 2
Ψ2 (φ) = {ψ : ψ (1) ≤ ψ ≤ ψ (2) }; (B.22)
Appendix B: Closed Form Integrations for Steerable Beamformer Design 115

BWφ BWφ
Part 3: Φ3 = {φ : ψ (1) + < φ ≤ ψ (2) + },
2 2
BWφ
Ψ3 (φ) = {ψ : φ − ≤ ψ ≤ ψ (2) }. (B.23)
2

Now that the analytical integration w.r.t. ψ for (B.8) is obtained, the next task
is to solve the same integral for (B.7), whose integration region Φ (ψ) covers both
spatial pass region and stop region (see (B.9)). For the case of ψ (2) − ψ (1) > BWφ
in Fig. B.1, the region Φpb (ψ) is as given by (B.13)– (B.15), and the region Φsb (ψ)
is given by
 
(1) BWφ
Part 4: Φ4 = {φ : −π ≤ φ ≤ ψ − + T Wφ },
2
Ψ4 (φ) = {ψ : ψ (1) ≤ ψ ≤ ψ (2) }; (B.24)
   
BWφ BWφ
Part 5: Φ5 = {φ : ψ (1) − + T Wφ < φ ≤ ψ (2) − + T Wφ },
2 2
 
BWφ
Ψ5 (φ) = {ψ : φ + + T Wφ ≤ ψ ≤ ψ (2) }; (B.25)
2
   
BWφ BWφ
Part 6: Φ6 = {φ : ψ (1) + + T Wφ ≤ φ ≤ ψ (2) + + T Wφ },
2 2
 
BW φ
Ψ6 (φ) = {ψ : ψ (1) ≤ ψ ≤ φ − + T Wφ }. (B.26)
2
 
BWφ
Part 7: Φ7 = {φ : ψ (2) + + T Wφ < φ ≤ π },
2
Ψ7 (φ) = {ψ : ψ (1) ≤ ψ ≤ ψ (2) }. (B.27)

Note that (B.24)–(B.27) are also true for the region Φsb (ψ) in Fig. B.2 (for the
case of ψ (2) − ψ (1) ≤ BWφ ). Substituting (B.13)–(B.15) and (B.24)–(B.27) into
(B.7) results in
    

QSBBF,FIR = AFIR (r, φ) ⊗ f (ψ) f H (ψ) dψ dφdr (B.28)
R Φ Ψ (φ)
    

= AFIR (r, φ) ⊗ f (ψ) f H (ψ) dψ dφ + . . . +
R Φ1 Ψ1 (φ)
    

AFIR (r, φ) ⊗ f (ψ) f H (ψ) dψ dφ dr (B.29)
Φ7 Ψ7 (φ)

The aim was to solve for




F (φ) = f (ψ) f H (ψ) dψ (B.30)
Ψ (φ)
116 Appendix B: Closed Form Integrations for Steerable Beamformer Design

in (B.28). The element in the m1 th row and m2 th column of F (φ) is given by


 g2 (φ)  m1 +m2
ψ
[F (φ)]m1 ,m2 = dψ (B.31)
g1 (φ) α
(g2 (φ))m1 +m2 +1 − (g1 (φ))m1 +m2 +1
= . (B.32)
(m1 + m2 + 1) α m1 +m2

Note that the matrix fliplr (F (φ)) is toeplitz and only contains 2M − 1 unique
entries, where fliplr (·) is the operation of flipping a matrix horizontally.