Steerable microphone arrays.

© All Rights Reserved

4 tayangan

Steerable microphone arrays.

© All Rights Reserved

- MITRES_6_007S11_hw05
- Time Delay Doa
- 5G_ebook.pdf
- Military Radars
- border security using wireless integrated network sensor by vinay kumar IPEC
- Article VCB2010 FMarin
- Optimum Use of Weather Radar Airbus Safety First Nr22
- 05.08100403
- Enabling Technologies for Ultra Wideband Phased Array - Gunnarsson
- TWR-850 - Installation Manual.pdf
- FAR3000 Operator's Manual
- report.docx
- Advances in Laboratory Geophysics Using Bender Elements (2006) - Thesis (382)
- RadarLosses(7)2011
- Bel Project Report
- 3 Manual Radar Plotting
- main4
- Solutions Final Exam 2016
- Awn Part03
- 10.1.1.11

Anda di halaman 1dari 122

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

•

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

Department of Electrical and Computer

Engineering

Curtin University

Perth, WA

Australia

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

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.

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.

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

3.5 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

6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

6.3 Final Remarks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

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

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

Acronyms

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

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.

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.

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].

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

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.

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

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.

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.

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

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

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

10 2 Acoustic Environment, Source Models and Sensor Arrays Theory

LTI system

Signal at Signal at

point A point B

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].

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.

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

throughout source's surface

y

Body of source

x

Fig. 2.4 Cross-sectional view of distributed source model

constant intensity

Point source

x

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.

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

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

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

r

lim =1 (2.4)

r→∞ rk− r

and

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]

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

from a farfield

source

φ

d

Reference sensor

from a farfield

source

y

φk

φx

Center of array as

reference point

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

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.

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

−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

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

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].

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

2π

r1 Ksen

0th ring

1st ring

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.

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)

22 2 Acoustic Environment, Source Models and Sensor Arrays Theory

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.

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

−8.4

−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

−11.5

−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

multiple of a scalar twist φa , i.e.

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

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

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

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.

3.1 Introduction

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].

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.

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.

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

y(n)

···

(b)

τ (0)

τ (1)

.. ..

. .

(K − 1)th

mic W (K − 1)

y(n)

τ (K − 1)

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)

FT

FT

.. ..

. .

(K − 1)th

mic W (K − 1, ω)

y(n)

FT IFT

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

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

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

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

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

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

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

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

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.

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

Hd (r, ω) to be

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

W (k,ω)

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

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

|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)

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

(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)

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),

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

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

T

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

HFIR (r, ω) = aH (3.36)

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

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

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

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.

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

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

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

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

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].

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.

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.

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

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

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)

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

2500

Frequency, f (Hz)

2000 5

1500

0

1000

500

−5

0 1 2

10 10 10

Source radial distance, r (m)

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

2500

Frequency, f (Hz)

2000 5

1500

0

1000

500

−5

0 1 2

10 10 10

Source radial distance, r (m)

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)

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)

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)

46 3 Broadband Beamformer Design

15

3500

3000

10

2500

Frequency, f (Hz)

2000 5

1500

0

1000

500

−5

0 1 2

10 10 10

Source radial distance, r (m)

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)

3.6 Design Examples and Evaluation 47

15

3500

3000

10

2500

Frequency, f (Hz)

2000 5

1500

0

1000

500

−5

0 1 2

10 10 10

Source radial distance, r (m)

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)

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

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

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.

4.1 Introduction

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

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.

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

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)

4.2 Beamformer Structure 55

FT

W (0, 1, ω) f (1, ψ)

.. ..

. .

W (0, M − 1, ω) f (M − 1, 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

w (0, 0, t)

f (1, ψ)

w (0, 1, t)

.. ..

. .

f (M − 1, ψ)

w (0, M − 1, t)

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

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.

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

Ψ R(ψ)

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

Solving the minimisation problem

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)

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

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

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

HFIR (r, ω, ψ) = dH (4.25)

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

−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

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.

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

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

|ξ

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

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.

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

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.

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.

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

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)

|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.

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 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

4.6 Performance Metrics 65

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.

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

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)

15

1

10

10

0

0

10

−5

−60 −40 −20 0 20 40 60

Steering angle, ψ (degree)

68 4 Steerable Broadband Beamformer Design

30

2

10

25

20

Source radial distance, r (m)

15

1

10

10

0

0

10

−5

−60 −40 −20 0 20 40 60

Steering angle, ψ (degree)

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)

15

1

10

10

0

0

10

−5

−60 −40 −20 0 20 40 60

Steering angle, ψ (degree)

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)

70 4 Steerable Broadband Beamformer Design

15

3500

3000

10

2500

Frequency, f (Hz)

2000 5

1500

0

1000

500

−5

−60 −40 −20 0 20 40 60

Steering angle, ψ (degree)

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)

4.7 Design Examples and Evaluation 71

15

3500

3000

10

2500

Frequency, f (Hz)

2000 5

1500

0

1000

500

−5

−60 −40 −20 0 20 40 60

Steering angle, ψ (degree)

30

2

10

25

20

Source radial distance, r (m)

15

1

10

10

0

0

10

−5

−60 −40 −20 0 20 40 60

Steering angle, ψ (degree)

72 4 Steerable Broadband Beamformer Design

30

2

10

25

20

Source radial distance, r (m)

15

1

10

10

0

0

10

−5

−60 −40 −20 0 20 40 60

Steering angle, ψ (degree)

30

2

10

25

20

Source radial distance, r (m)

15

1

10

10

0

0

10

−5

−60 −40 −20 0 20 40 60

Steering angle, ψ (degree)

4.7 Design Examples and Evaluation 73

15

3500

3000

10

2500

Frequency, f (Hz)

2000 5

1500

0

1000

500

−5

−60 −40 −20 0 20 40 60

Steering angle, ψ (degree)

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)

74 4 Steerable Broadband Beamformer Design

15

3500

3000

10

2500

Frequency, f (Hz)

2000 5

1500

0

1000

500

−5

−60 −40 −20 0 20 40 60

Steering angle, ψ (degree)

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)

4.7 Design Examples and Evaluation 75

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

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.

5.1 Introduction

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

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

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.

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)

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

2

NG (5.4)

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

w S B B F,F I R

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

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.

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.

Sensor errors ε (r, rk , ω) such as gain and phase errors can often be modelled as

multiplicative errors [6, 23], i.e.

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

or in vector form as

where

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

= (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

= 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

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.

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

and

= · · · 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)

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)

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

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,

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)

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

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

= (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

= 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

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

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)

5.3 Stochastic Error Model 85

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

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

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)

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.

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)

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

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

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.

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

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.

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)

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.

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

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

−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.

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)

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

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

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

−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

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

Â (r, k, ω) = A (r, k, ω) + ρi A (ri , k, ω) (5.69)

L i=1

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)

96 5 Robust Formulation

Direct path

in ideal case

φ φi

Additional path due

to local scattering

Sensor array

Point source

100

TLS robust

TLS non−robust

LS robust

LS non−robust

80

60

Performance error, ξLS (dB)

40

20

−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)

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

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

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

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

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

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.

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.

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.

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

beamformer system

···

.. Source

. tracking

Estimated

location

.

beamformer

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

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

topic.

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

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

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

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)

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)

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)

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

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,

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

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)

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)

F (φ) = f (ψ) f H (ψ) dψ (B.30)

Ψ (φ)

116 Appendix B: Closed Form Integrations for Steerable Beamformer Design

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.

- MITRES_6_007S11_hw05Diunggah olehLenin Kumar
- Time Delay DoaDiunggah olehSanchit Gupta
- 5G_ebook.pdfDiunggah olehRekha Nair
- Military RadarsDiunggah olehAnurag Reddy
- border security using wireless integrated network sensor by vinay kumar IPECDiunggah oleh2687
- Article VCB2010 FMarinDiunggah olehcorsaiolo
- Optimum Use of Weather Radar Airbus Safety First Nr22Diunggah olehJustin Ong
- 05.08100403Diunggah olehNguyễn Hồng Dương
- Enabling Technologies for Ultra Wideband Phased Array - GunnarssonDiunggah olehvs11
- TWR-850 - Installation Manual.pdfDiunggah olehGee
- FAR3000 Operator's ManualDiunggah olehZaw Khaing Win
- report.docxDiunggah olehvishal agnal
- Advances in Laboratory Geophysics Using Bender Elements (2006) - Thesis (382)Diunggah olehJulio Humberto Díaz Rondán
- RadarLosses(7)2011Diunggah olehRohit Sharma
- Bel Project ReportDiunggah olehMasoom Sharma
- 3 Manual Radar PlottingDiunggah olehosama
- main4Diunggah olehBaşak Esin Köktürk Güzel
- Solutions Final Exam 2016Diunggah olehayadman
- Awn Part03Diunggah olehabdulaziz saif ali mansoor
- 10.1.1.11Diunggah olehmahdi114
- The Review of Energy Efficient NetworkDiunggah olehEditor IJTSRD
- Shih Physical 01 PaperDiunggah olehAnjum Preet Singh
- Free Field Vibrations During the PassageDiunggah olehFajar Sugih Rizqillah
- 9. Receivers_2013.pdfDiunggah olehHemal Vyas
- lecture5_2Diunggah olehM Kashif Junaid
- Ground Penetrating Radar for the Detection of Water Leakage in Buried PipeDiunggah olehAnonymous WNWk1h2
- applsci-07-01144-v2Diunggah olehMihai Amariei
- soundASR.pdfDiunggah oleh陳韋佳
- VibrationsDiunggah olehdustinshay
- Lecture Part7Diunggah olehAnonymous mXicTi8hB

- The Healing Power of MindDiunggah olehS Sharath Chandra
- The Ascent of Man - Jacob BronowskiDiunggah olehMarko Verhere
- Addictions Native Americans, NFB and Substance Abuse TheoryDiunggah olehTom Neser
- Magic of RadionicsDiunggah olehgalaxy5111
- Critical Thinking Mini GuideDiunggah olehJoe Lau
- The Debate of King MilindaDiunggah olehDzoJo
- Gary E. Schwartz - The Afterlife ExperimentsDiunggah olehcesargnomo
- Thinking From a to ZDiunggah olehS Sharath Chandra
- Dialogues Between Buddhism and Science Attention,Memory and MindDiunggah olehS Sharath Chandra
- Buddhadasa Buddhas Doctrine of AnattaDiunggah olehanon7535
- Ancient MesopotamiaDiunggah olehS Sharath Chandra
- Netti Ppakaranam by Bhikku NanamoliDiunggah olehS Sharath Chandra
- The Hundredth Monkey Phenomenon and Mass MesmerismDiunggah olehS Sharath Chandra

- Fluid MechanicsDiunggah olehflokaai
- UPMREB FORM 4(a)2012 Format of the Minutes of the Meeting_0Diunggah olehkarl_poor
- JCSOS-REV-ReadingGuide-chapter5 wordDiunggah olehcollinmanderson
- Manual Detector de FluorescenciaDiunggah olehPablo C. Vicencio
- NMPED 2017 School Grades, August 2017Diunggah olehAlbuquerque Journal
- 2012 Legislative Report on Remedial EducationDiunggah olehEdNews
- Great Expectations.docxDiunggah olehlovleshruby
- Program EnDiunggah olehAl'smbrreketekketeck Dgreketeckketeck
- Rule 87Diunggah olehbryanmarquez22
- FIFA-Manager-14.pdfDiunggah olehAnonymous C4BikQxRji
- Oracle Unified MethodDiunggah olehJyothiM
- Case31MerckDiunggah olehGaLb Aboha
- As Level Physics 2011 Smak Gs Kbi MeasurementDiunggah olehJoshuaUntung
- Melaleuca CatalogueDiunggah olehNatasha Lim
- GNLD's Omega 3 - help for age related mental decline.Diunggah olehNishit Kotak
- Matlab Summary and TutorialDiunggah olehtraisimdo
- tcsDiunggah olehphani98
- Ban Duras TheoryDiunggah olehtmilic6
- 299303777 57 Atrium Management Corp vs CA DigestDiunggah olehkeithnavalta
- Aries Lagan Transit of Saturn on 26 Jan 2017 in SaggiDiunggah olehVikrant Kaushal
- Chapt 12Diunggah olehKamauWafulaWanyama
- Light and Life 2014 HKDSE High FlyersDiunggah olehUcb Bioe
- Crypto Algorithms.pdfDiunggah olehTaha Khan
- The Wilcoxon Sign Rank TestDiunggah olehApam Benjamin
- exams_knockout_test06.pdfDiunggah olehDiego González Sancho
- falmaterolDiunggah olehnely
- EMS 2013Diunggah olehJessica Pool
- DOs and DONTs in FranceDiunggah olehMustika Cahya P
- Dec_2008_low resDiunggah olehsachinsaini
- 1524705382017_Kasus+gout+dan+hiperuricemiaDiunggah olehreffada