Ultrasonic Source Localization with a

Beamformer Implemented on an FPGA

Using a High-density Microphone Array
Master’s thesis in Embedded Electronic System Design

Daniel Antonsson
Mengxuan Li

Department of Computer Science and Engineering

C HALMERS U NIVERSITY OF T ECHNOLOGY
U NIVERSITY OF G OTHENBURG
Gothenburg, Sweden 2018
 Master’s thesis 2018

Ultrasonic Source Localization with a

Beamformer Implemented on an FPGA
Using a High-density Microphone Array

Daniel Antonsson
Mengxuan Li

Department of Computer Science and Engineering

Programme of Embedded Electronic System Design
Chalmers University of Technology
University of Gothenburg
Gothenburg, Sweden 2018
Ultrasonic Source Localization with a Beamformer Implemented on an FPGA Using
a High-density Microphone Array
DANIEL ANTONSSON
MENGXUAN LI

Supervisor: Lena Peterson, Department of Computer Science and Engineering

Supervisor: Steinar Thorvaldsson, Syntronic
Examiner: Per Larsson-Edefors, Department of Computer Science and Engineering

Master’s Thesis 2018

Department of Computer Science and Engineering
Programme of Embedded Electronic System Design
Chalmers University of Technology
University of Gothenburg
SE-412 96 Gothenburg
Telephone +46 31 772 1000

iv
Ultrasonic Source Localization with a Beamformer Implemented on an FPGA Using
a High-density Microphone Array
DANIEL ANTONSSON
MENGXUAN LI
Department of Computer Science and Engineering
Chalmers University of Technology

Abstract
A system has been designed which uses a high-density MEMS microphone array and
a beamforming algorithm implemented on an FPGA to perform real-time positioning
of ultrasonic sources. The array consists of 48 microphones which is a limit derived
from the latency of the system; if the latency were to increase further the system
could no longer be said to operate in real time.

The system calculates signal energy in front of the array and displays the energy on a
heatmap which updates at a frequency of 5.2 Hz. It can locate near-monochromatic
sources at 20 kHz and narrowband sources at 35 kHz. However, the ability of the
system to locate these sources was proven to be unreliable due to a phenomenon
called spatial aliasing which causes an ambiguous pattern in the heatmap. Very brief
sounds could not be located by the system and the system is also easily disturbed
by ambient noise, such as noise from conversations.

It was concluded that the first step of any potential future development should be
to completely revise the implementation of the beamforming algorithm in order to
increase the amount of microphones that can be used and decrease the latency of
the system.

Keywords: microphone array, ultrasonic, FPGA, beamforming, delay-and-sum

v
Acknowledgements
We would like to express our sincere and heartfelt thanks to our supervisors Steinar
Thorvaldsson and Lena Peterson for providing excellent guidance during this thesis
work.

We would also like to thank all the people working at Syntronic, especially Bhav-
ishya Goel and Erik Ikdal, for providing us with resources which made our research
possible and also for their helpful suggestions, valuable advice and detailed com-
ments.

Daniel Antonsson & Mengxuan Li, Gothenburg, June 2018

vii
 Contents

Abbreviations 1

1 Introduction 2
1.1 Microphone arrays and beamforming . . . . . . . . . . . . . . . . . . 2
1.2 Field-Programmable Gate Array . . . . . . . . . . . . . . . . . . . . . 3
1.3 Research statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Problem description . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.5 Limitations to mitigate risks . . . . . . . . . . . . . . . . . . . . . . . 5
1.6 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2 Basic Delay-and-Sum Beamforming Theory 7

2.1 Beamforming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Spatial Aliasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3 System Design Process and Testing Methodology 12

3.1 Hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.1 Sensor Boards . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.2 Breakout Board . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.3 Ultrasonic Transducer . . . . . . . . . . . . . . . . . . . . . . 14
3.2 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2.1 Audio Processing . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.2.2 Delay-and-sum beamforming . . . . . . . . . . . . . . . . . . . 15
3.2.3 Directional Energy . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.4 Finite state machine . . . . . . . . . . . . . . . . . . . . . . . 18
3.2.5 Digilent Parallel Transfer Interface . . . . . . . . . . . . . . . 20
3.2.6 Reading the Transferred Data and Calculating the Moving
Average . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2.7 Heat Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Design Decisions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.4 Testing Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4 Results 29
4.1 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Implementation results . . . . . . . . . . . . . . . . . . . . . . . . . . 29

5 Discussion and Possible Improvements 38

5.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

ix
Contents

5.1.1 General performance . . . . . . . . . . . . . . . . . . . . . . . 38

5.1.2 Brief sounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.1.3 FPGA Resource Utilization . . . . . . . . . . . . . . . . . . . 39
5.1.4 Mirror images in the vertical array . . . . . . . . . . . . . . . 41
5.1.5 Noisy environment . . . . . . . . . . . . . . . . . . . . . . . . 42
5.2 Potential improvements . . . . . . . . . . . . . . . . . . . . . . . . . . 42

6 Conclusion 44

x
 Abbreviations

ADC Analog-to-Digital Conversion

API Application Programming Interface
BRAM Block Random-Access Memory
BUFG Buffer Global
CIC Cascaded Integrator-Comb
DAS Delay and Sum
DOA Direction of Arrival
DPTI Digilent Parallel Transfer Interface
DSP Digital Signal Processor
DTFT Discrete-Time Fourier Transforms
FF Flip-Flop
FIR Finite Impulse Response
FPGA Field-Programmable Gate Array
FSM Finite State Machine
IO Input/Output
IP Internet Protocol
LUT Lookup Table
LUTRAM Lookup Table Random-Access Memory
MEMS Microelectromechanical Systems
MMCM Mixed/Mode Clock Manager
PC Personal Computer
PCM Pulse-Coded Modulation
PDM Pulse-Density Modulation
RAM Random-Access Memory
ROM Read-Only Memory
SMD Surface Mounted Devices
SNR Signal-to-Noise Ratio
SRL Shift Register Lookup Table
UART Universal Asynchronous Receiver/Transmitter

1
 1
Introduction

This is a project report that documents the work carried out in a master thesis
project at Chalmers University of Technology in collaboration with the consult-
ing company Syntronic. This chapter describes the background, purpose, problem
description and limitation of the thesis work.

1.1 Microphone arrays and beamforming

In many fields, such as sonar, radar, wireless communications and seismology, there
has been considerable research effort into sensor arrays [1], [2]. This is a group of
sensors that is arranged in some geometrical pattern with the aim to enhance signals
and separate them from noise [2]. For acoustic signals they are called microphone
arrays.

Some applications in which microphone arrays have been found useful are separating
sound sources to improve performance in hearing aids, security monitoring, studying
wildlife, detecting air leaks and detecting small unmanned aerial vehicles [3]–[7].
Most of these applications depend on the array’s ability to localize sound, which is
the focus of this report.

Source localization is done by means of a technique called beamforming. A beam-

former can be described as a spatial filter which enhances signals coming from a
certain direction, the look direction, and attenuates signals originating from any
other direction [8]. Conforming to this analogy: filters have passbands and stop-
bands and similarly beamformers have main lobes and side lobes. The main lobe
is the high amplitude portion of the beamformer and the side lobes are the attenu-
ated portions. The usefulness of this filtering property lies in steering the beam to
various locations and determine which direction corresponds to the highest signal
intensity - in our case: the loudest sound. How the beam is steered will be explained
in Chapter 2. The ability of microphone arrays to accurately identify the correct
position depends on the following factors [8]–[12]:

• Number of microphones
• Spacing between microphones
• Weighting of individual microphone gains

2
 1. Introduction

• The size of the array

A higher number of microphones makes the beam more narrow, which means that
sound can be spatially filtered more precisely. Additionally, the signal-to-noise ratio
(SNR) of the array increases as the number of microphones increases [13]. Less
spacing between microphones prevents spatial aliasing [14], a phenomenon that is
explained in Chapter 2, and it also increases the array’s SNR [13]. More spacing
between microphones, on the other hand, makes the beam narrower. The weighting
of individual sensors, which may be considered as gain factors, affects the level and
shape of the side lobes. Finally, the accuracy increases as the size of the array is
increased [9].

In a nutshell: the array’s ability to locate sound sources improves as the number
of microphones is increased, the size of the array is increased and the density of
microphones is increased. But the ability to decrease spacing between microphones
is obviously limited by the size of the microphones and increasing the amount of
microphones increases the amount of data to be processed.

The problem with size is somewhat mitigated by using microelectromechanical sys-

tems (MEMS) microphones; this technology has led to the development of small
microphones with very high performance. However, processing large amounts of
data from many sensors is likely not feasible on a regular computer processor, es-
pecially if the system is to operate in real time. Additionally, using a microphone
array with a large amount of microphones means there will be a need to handle a
large amount of input/output (I/O). Managing these problems can be made easier
by implementing such a system on a Field-Programmable Gate Array (FPGA).

1.2 Field-Programmable Gate Array

An FPGA is a device which is comprised of programmable logic blocks and pro-

grammable routing and I/O blocks. They can be programmed with hardware de-
scription languages to achieve a vast variety of functions [15]. The most fundamen-
tal building blocks of FPGAs are registers, multiplexers (mux) and lookup tables
(LUTs). Registers, also called flip-flops, are used to store one bit of data; muxes
are used to select one of several input signals and forward the selected input to the
output; LUTs can be seen as a multiplexer with fixed input values. FPGAs have
programmable routing, which means wiring between logic blocks inside the FPGA
can be arranged by the designer. FPGAs can have a high I/O count and each I/O
pin can be configured as an input, output or bidirectional I/O [16].

Implementing the system on an FPGA opens up the possibility of an application

specific architecture. For instance, if the system turns out to be too slow, the ar-
chitecture can be designed to carry out all operations in parallel to the extent that
resources permit it. Conversely, if the system has too high hardware requirements,
it can be designed to carry out operations in series. Additionally, the MEMS mi-

3
1. Introduction

crophones (from this point on referred to as sensors) have built-in analog-to-digital

conversion (ADC) [17], which make them convenient to use with an FPGA.

The company Xilinx, the inventor of the FPGA [18], provides a software suite called
Vivado for synthesis and analysis of designs made with hardware descriptive lan-
guage (HDL). Besides many other features, this tool can also be used to program
the FPGA with the synthesized design [19]. In this project Vivado has been used
for all HDL design.

1.3 Research statement

The goal of this master thesis is to design a system that uses beamforming for real-
time positioning of ultrasonic sources, using an FPGA platform and data that are
collected by a high-density MEMS microphone array. As a metric of how successful
the project is, the accuracy with which the system can localize a static source will be
used. This, by extension, means that the system is to be designed to accommodate as
many sensors in the array as reasonable within the FPGA resource constraints, since
increasing the number of sensors improves the accuracy of the array. To maximize
the chance for success in this project the system is to use an algorithm that is as
simple as possible without critically compromising its accuracy.

1.4 Problem description

The section presents a more detailed analysis of what the goals asserted in Section 1.3
entail. Firstly, some considerations of processing ultrasound are explored, which is
followed by the introduction of a simple algorithm and finally the ramifications of
requiring operations to occur in real time are explored.

This project focuses on the problem of source localization for ultrasound, i.e. sound
that is beyond the range of what is audible for humans. Focusing on ultrasound is
interesting for many applications: air leaks in pressurized air systems, for instance,
produce ultrasonic signals [6] and might be difficult to detect at lower frequencies;
the sonar-like abilities of bats use ultrasound to prevent the sound from scatter-
ing [20] and thus only microphone arrays with the ability to process ultrasound can
localize this sound. One consequence of working with ultrasound is that processing
it inherently requires a higher sample rate compared to lower frequency sound which
in turn puts higher requirements on the system hardware. Another problem, which
is detailed in Chapter 2, is that a phenomenon called spatial aliasing negatively
affects performance at higher frequencies.

As mentioned in Section 1.3 a decision to use the simplest algorithm was made. The
simplest and most straightforward algorithm is the delay-and-sum (DAS) beamform-
ing algorithm, which also poses the most demanding hardware requirements [21].

4
 1. Introduction

This means that upon scaling up the amount of microphones to be used in a sys-
tem using the DAS algorithm, one can expect resource consumption to become
prohibitively large and that some measures to lower resource consumption would
have to be made. One such potential measure could be to use ring buffers to avoid
unnecessary storage of calculated delay values [22].

Also mentioned in Section 1.3 is that one of the goals of the project was for the system
to operate in real time. In this context real time means that the system must be
able to respond quickly enough to an acoustic event, which may last less than a
second. For instance, an acoustic camera - a device for visualizing sound sources -
may be considered to operate in real time at 10 image frames per second [22]. The
requirement of having a system operating in real time puts the following restrictions
on the system:

• For every look direction a number of audio samples must be collected to de-
termine the energy of the sound. The amount of audio samples collected for
each angle must be kept to a minimum.

• The size of the array cannot be too large, since the bigger the array is the
larger delays are required to perform the DAS algorithm.

• The number of locations the beam is steered to must be limited, since every
new look direction will require some time for processing.

1.5 Limitations to mitigate risks

A significant risk in a project like this is potentially setting too ambitious goals,
which is why we try to limit ourselves as much as possible. We therefore limit
our focus on algorithms and only do comparison between algorithms using high-
level simulation tools, specifically MATLAB. As previously mentioned we gravitate
strongly towards an algorithm that is thought to maximize our chance of a successful
FPGA implementation.

Projects like this may often fail due to uncontrollable factors, such as not having
access to hardware that is a prerequisite to the system. Prior to the onset of this
project the design of custom hardware is already finished which mitigated this risk.
In retrospect this consideration is likely crucial as a design error in the hardware is
in fact found and could be corrected in the nick of time.

Many research projects focus on broadband beamformers since the goal is often to
process speech and there is a desire to have consistent beamforming over a span of
frequencies. In this project we spend no time considering broadband beamformers.
This can be justified by the fact that we have no interest in the quality of sound but
rather only of the sound source position and thus we conclude that a narrowband
beamformer (which are simpler to design [8]) would suffice.

5
1. Introduction

1.6 Thesis outline

Chapter 2 reviews the basic theory relevant to the thesis work. It contains two
sections that introduce beamforming and spatial aliasing.

Chapter 3 explains the design steps taken to achieve the goals stated in Section 1.3.
It is divided into three parts: system architecture, design decisions and testing
methodology.

Chapter 4 presents the results of our work. This section shows both results from
simulations and tests conducted on actual hardware.

Chapter 5 discusses the significance of the results presented in Chapter 4. Sugges-

tions for future improvements is also included in this section.

Chapter 6 lists the conclusions which are responding to Section 1.3.

6
 2
Basic Delay-and-Sum
Beamforming Theory

The purpose of this chapter is to base this project in a mathematically stable foun-
dation. In the following sections, the theory behind the implementation in the thesis
project is described briefly.

2.1 Beamforming

Originally developed in the fields of radar and sonar, beamforming is a long re-
searched technique with a multitude of applications. Two interesting techniques are
the ability to detect the direction of arrival (DOA) and enhancing a desired signal.
A beamformer can be thought of as a spatial filter which grants a microphone array
the ability to “listen” in a single direction, cancelling out sound coming from other
directions.

The process of spatially filtering a signal can in fact be conceptualized as two sub-
processes. The first consists of delaying signals from each microphone in a way so
that the sum of the signals aligns in time when coming from a certain direction and
thus making signals from that direction constructively interfere while signals coming
from other direction destructively interfere. This process is illustrated in Fig. 2.1.
The second process works by applying weight coefficients to the signals and then
add them together. Synchronizing the signals steers the direction of the beam while
applying weights controls the beamwidth of the main lobe and the attenuation of the
side lobes [8]. In this project we have focused only on the synchronization process.

One beamforming technique is the DAS beamformer. It is the least complex algo-
rithm to implement but is the most resource intensive [21]. This algorithm works
by delaying the signal at each sensor. Given a linear array with uniform distance
between each sensor the output of this algorithm from the direction of α is described
as

1 NX
−1
ida
g(a, t) = s(t − ), (2.1)
N i=0 c

7
2. Basic Delay-and-Sum Beamforming Theory

Source
Wavefront

Aligned
t2 Delay Signals
Stage

t1 t2

t1

0
Mircophone

Figure 2.1: Block diagram of spatial filter.

where N is the number of sensors, s is the received signal, d is the distance between
sensors, a = cos(α), ω = 2πf is the angular frequency and c is the wave propagation
speed. If we represent s(t) by its continuous Fourier transform S(ω)

1 Z∞
s(t) = S(ω)ejωt dω
2π −∞

then

1 Z∞ ωa
g(a, t) = S(ω)W ( )ejωt dω,
2π −∞ c

where

1 NX
−1
W (α) = e−j2πf n cos(α)d/c . (2.2)
N n=0

W (α) is called the array pattern [23] and it is a plot of the array’s response to a
signal. In order to draw the array pattern we rewrite (2.2) as a function of ψ and
θ, where ψ is a variable ranging 0 o ≥ ψ ≤ 180 o and θ is the look direction of the
beam:

1 NX
−1
W (ψ, θ) = e−j2πnf (cos ψ−cos θ)d/c , (2.3)
N n=0

the absolute of which can be shown to be

 
sin N πf d(cos ψ − cos θ)/c
|W (ψ, θ)| =   (2.4)
N sin πf d(cos ψ − cos θ)/c

as provided by [8].

8
 2. Basic Delay-and-Sum Beamforming Theory

The array pattern can then be plotted and the resulting figure can be seen in
Fig. 2.2a. This gives us good insight into how the beamformer will spatially fil-
ter signals; it shows the width of the main lobe and the attenuation factor of the
side lobes.
90o 90o

120o 60o 120o 60o

150o 30o 150o 30o

0 [dB] 0 [dB]
-10 -10
-20 -20
-30 -30
-40 -40
180o 0 o 180o 0o

(a) (b)

Figure 2.2: (a) shows an example beam pattern of an array with eight microphones,
f = 5 kHz, d = 4 cm, θ = 90 o . The polar plot shows (2.4) plotted as a function of
ψ. (b) shows and example of spatial aliasing of beam pattern of an array with eight
microphones, f = 15 kHz, d = 4 cm and ψ = 90 o .

2.2 Spatial Aliasing

The Nyquist sampling theorem tells us that signals must be bandlimited to half
the sampling rate to prevent aliasing from occurring. This phenomenon extends to
signals that are functions of space coordinates and is then called spatial aliasing. The
components created by spatial aliasing are called grating lobes [23]. If the spatial
equivalent of the Nyquist criterion is not satisfied grating lobes will be produced,
with the consequence that the array pattern will be ambiguous.

Fig. 2.2b shows an example of spatial aliasing occurring. Besides the main lobe at
ψ = 90 o there are two grating lobes at ψ = 55 o and ψ = 125 o . The relationship
of the beamformer response versus the incident angle and frequency can be seen in
Fig. 2.3, which displays a 3D plot and a heat map of the relationship. It is apparent
that spatial aliasing becomes more prevalent for higher frequency signals.

It turns out, however, that spatial aliasing may not be as problematic as it can first
appear. An intuition to this is that humans can localize sounds quite well even
though an average spacing of 20 cm between the ears would suggest that spatial
aliasing would occur for any sound above 850 Hz [24]. The patterns displayed in
Figs. 2.2 and 2.3 were both generated using a monochrome source signal. The
undesirable effects of spatial aliasing are somewhat mitigated when wideband signals
are used [24]. This effect is illustrated in Fig. 2.4, where it can be seen that the
two grating lobes shrink in amplitude as more (monochromatic) signals are added
together and used as the source.

9
2. Basic Delay-and-Sum Beamforming Theory

(a)

(b)

Figure 2.3: These plots show the relationship between frequency, incident angle
and the beamformer response with eight sensors and d = 11.43 mm in the form of a
3D plot in (a) and a heat map in (b).

10
 2. Basic Delay-and-Sum Beamforming Theory

One signal Two signals

90o 90o

120o 60o 120o 60o

150o 30o 150o 30o

0 [dB] 0 [dB]
-10 -10
-20 -20
-30 -30
-40 -40
180o 0 o 180o 0o

(a) (b)
Four signals Ten signals
90o 90o

120o 60o 120o 60o

150o 30o 150o 30o

0 [dB] 0 [dB]
-10 -10
-20 -20
-30 -30
-40 -40
180o 0 o 180o 0o

(c) (d)

Figure 2.4: The figure depicts beamforming patterns of a system with eight sensors,
d = 11.43 mm θ = 90 o and various input. In (a) f = 40 kHz, in (b) the signal has
two components of f = 20 kHz and f = 80 kHz, in (c) the signal consists of four
components ranging 20 kHz ≤ f ≤ 80 kHz and in (d) the signal consists of ten
components ranging 20 kHz ≤ f ≤ 80 kHz.

11
 3
System Design Process and
Testing Methodology

This chapter is meant to provide an overview of our system and document the system
design process. Firstly the available hardware is presented in order to clearly show
where some of the system’s parameters originate from. Then the system architecture
is explored in Section 3.2 which is divided into several subsections: audio processing,
DAS beamforming, acquisition of directional energy, finite state machine (FSM),
Digilent parallel transfer interface (DPTI), moving average in MATLAB and heat
map. Section 3.3 contains explanations for any major design decision. Finally, our
testing methodology is detailed.

3.1 Hardware

Some design parameters are derived directly from the available hardware. For that
reason this section will present the available hardware.

3.1.1 Sensor Boards

The sensor boards (see Fig. 3.1) are custom designed by Syntronic and each houses
24 MEMS microphones [25] spaced 11.43 mm apart. This means that in this project
we are constrained to this distance and our ability to vary the geometry of the array
is limited to the configuration of the boards relative each other. In total we have
six sensor boards available, which amounts to a total of 144 sensors.

3.1.2 Breakout Board

The breakout board (see Fig. 3.2) is custom designed by Syntronic to interface
between the Genesys 2 Kintex 7 FPGA [26] development board and the sensor
boards.

12
 3. System Design Process and Testing Methodology

(a)

(b)

Figure 3.1: The front (a) and the back (b) of the sensor boards.

(a) (b)

Figure 3.2: The front (a) and the back (b) of the breakout board. (a) shows the
SMD connectors on the front of the board that will connect the sensor boards to the
breakout board. At the time of taking this picture only two connectors are soldered
on the board. (b) shows a large SMD connector on the back which fits the FPGA
Mezzanine Card connector on the Genesys 2 Kintex 7 FPGA development board.

13
3. System Design Process and Testing Methodology

3.1.3 Ultrasonic Transducer

Figure 3.3: The ultrasonic transducer.

The ultrasonic transducer (see Fig. 3.3) is made of a piezoelectric material which
vibrates when an alternating voltage is applied to it. Its datasheet claims it has a
center frequency of 40 kHz [27].

3.2 System Architecture

The system architecture overview is depicted in Fig. 3.4. It contains the FSM and
the beamforming system. The beamforming system includes audio processing, DAS
beamforming, acquisition of directional energy and interface.
Cosine Wave

Halfband
FIR Low-
CIC Filter Low-pass Delay
Microphone 1 pass Filter
Filter

Adder Multiplier Accumulator

Cosine Wave

Halfband
FIR Low-
CIC Filter Low-pass Delay
Microphone N pass Filter
Filter

Microphone PDM to PCM Conversion Delay and Sum Directional Accumulated Energy Interface

Input Output
FSM System

Figure 3.4: System overview.

3.2.1 Audio Processing

Conversion from pulse-density modulation (PDM) to pulse-coded modulation (PCM)

is done on every individual data stream. An illustration of each component in this

14
 3. System Design Process and Testing Methodology

conversion is shown in Fig. 3.5; as can be seen each one of these conversion blocks
consists of several digital filters and a multiplier. The first filter is a cascaded
integrator-comb (CIC) filter which is used to low pass filter the PDM stream in
order to convert it into a PCM format; the PDM stream is essentially averaged.
Then half band filters are used as compensation filters since the passband of CIC
filters is not sufficiently flat [28]. The purpose of the finite impulse response (FIR)
filter is to remove any high frequency noise left over from the conversion. Finally,
the multiplier is used as a mixer to move the signal down to a lower frequency for
making handling the signal easier. All of these components are implemented using
Xilinx IPs.

The PDM clock is 4.608 MHz which is within the ultrasonic operation range accord-
ing to the datasheet of the digital MEMS microphones [25] and a number evenly
divisible by 192 kHz (a common sampling frequency).

Cosine Wave PCM

PDM
Halfband Low-
CIC Filter FIR Low-
pass Filter
16 pass Filter
Microphone @ 4.608 MHz @ 384 kHz 2 @ 192 kHz

Sound source

Figure 3.5: PDM to PCM conversion architecture.

3.2.2 Delay-and-sum beamforming

Delays are implemented with a Xilinx IP called RAM-based shift register. This
IP has the option of variable length delays, which fits perfectly in this project since
every scanning angle needs a different set of delays. The desired delay td is converted
into a number of clock cycles Nclk = ceil (td fclk ), where fclk is the frequency of the
clock, and this number is driven to the input address of the shift register. A value for
Nclk must be provided for each stream of input data and each scanning angle. These
values are pre-calculated in MATLAB and saved into text files that are initialized as
ROM upon compiling the system. A set of delays can then be chosen by pointing to
an address in the ROMs containing the pre-calculated delay values which are then
driven to the shift registers. Delay values for 128 different angles between 30 o and
150 o are pre-calculated in MATLAB.

Assuming speed of sound is 343 m/s, the largest possible delay using the shift register
IP is 1024 clock cycles. With the PDM clock being 4.608 MHz this yields a maximum
delay of td = 1024/4.608 MHz ≈ 222 µs. Assuming room temperature, and thus the
speed of sound c = 343 m/s, the maximum needed delay for 24 sensors spaced
d = 11.43 mm apart is

(N − 1)d 23 · 11.43 mm
tmax = cos(θmax/min ) = cos(30o ) ≈ 664 µs, (3.1)
c 343 m/s

15
3. System Design Process and Testing Methodology

for a system with maximum scanning angles of 30 o and 150 o . This means that
is necessary to either use three shift register IPs in a row or divide the clock that
drives the shift register IPs by three. The current system uses the former solution,
for reasons explained in Section 3.3.

The adding together of data streams is done with an adder IP which is controlled by
enable signals from the FIR filters (which must traverse the shift registers). The mux
in Fig. 3.6 selects the enable signal which is delayed the most to ensure that data
from all the sensors are being added together and that there is no period of adding
together data from just a few sensors. As the data need one clock cycle to be added
together the enable signals are delayed one clock cycle for each stage of adders by
being passed through a register, as shown in Fig. 3.6. This figure illustrates a nested
loop which is defined by two constants, width and depth, that the designer decides.
Arranging the adders in a nested loop like this allows for convenient changing of
the number of sensors to be used. By generating an adder every time i mod 2j+1
returns true for 0 < i < width and 0 < j < depth where depth = log 2(width) the
formation shown in Fig. 3.6 is generated. Changing the number of sensors to, for
instance, 16 would automatically add another layer of adders which decreases the
amount of manual coding that has to be done.

Enable signals

Data
Depth
0 1
Width

PDM to RAM-
0 PCM based SR FIFO

PDM to RAM-
1 PCM based SR

PDM to RAM-
2 PCM based SR

PDM to RAM-
3 PCM based SR

Figure 3.6: Four-microphone architecture. Clock signals are not shown to prevent
cluttering. SR stands for shift register.

As earlier mentioned, delay values are pre-calculated. This is done using a custom
designed MATLAB function which accepts as argument the number of sensors and
the desired angles for scanning. For purposes of user feedback this function also
creates an illustration of the wavefront incident on the microphone array, which can
be seen in Fig. 3.7. This function calculates the delay values and converts them
into a number of clock cycles, Nclk , which are written into plain text files that are
subsequently used for initializing ROMs in the system.

16
 3. System Design Process and Testing Methodology

0.025 0.025
Microphones Microphones
0.02 Wavefront 0.02 Wavefront

0.015 0.015

0.01 0.01

0.005 0.005

0 0
0 28.9 57.7 86.6 [ s] 0 16.5 33.1 49.6 [ s]

-0.005 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 -0.005 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04

(a) (b)

Figure 3.7: This figure illustrates how the delay values are calculated for a system
using four sensors. The distance between microphones is hard coded to 11.43 mm.
The value for the speed of sound used is c = 343 m/s.

3.2.3 Directional Energy

We would like to analyze the energy of the audio rather than just the raw audio.
Therefore we need to do some processing in order to identify the energy of our
summed audio data. The energy of a signal is calculated as

Z +∞
Ec = |xc (t)|2 dt,
−∞

where xc (t) is a continuous signal [29]. For discrete signals we have

+∞
|x[n]|2 .
X
Ed = (3.2)
−∞

The physical implementation of this is straightforward and is depicted in Fig. 3.8.

The Xilinx IP multiplier is used to square the output of the DAS beamformer. The
accumulator adds together a number of samples to acquire the signal energy.

Summation from
delay-and-sum Accumulated
beamforming Energy
Multiplier Accumulator

Figure 3.8: Acquiring directional energy as per (3.2).

17
3. System Design Process and Testing Methodology

3.2.4 Finite state machine

In the previous section we have created a beamformer with adjustable delays so that
it can be steered to different angles. To get the directional energy we have to adjust
the delays to point the beam in the desired angle, sample some data, and store it.
Then we must do this for every angle we want to examine and then see which angle
provides the largest energy, which will be the candidate for the direction of arrival
(DOA). In order to do this we need a FSM, whose design is illustrated in Fig. 3.9.

idle

sta
mo
0

rt =
rt =

de

1
=0
sta

sta
1
=1
rt =

rt =
sta

de

0
mo

mode = 0

static beam run system

mode = 1

wait for shift

registers to
clear out

Figure 3.9: State diagram for the FSM. The input signals start and mode are set
by the user via switches on the FPGA development board.

The idle state is for doing nothing while the start button is set to zero. The run
system state is for gathering data samples and the wait state is for waiting until
data from all sensors have propagated out of the shift registers. The delay values are
provided to the shift registers by a ROM in which pre-calculated delay values are
stored. Every time the system returns to the run system state a counter ticks up
and a new set of data is read from the ROM. This setup is illustrated in Fig. 3.10.
To get a better sense of the timing of the FSM Fig. 3.11 can be studied; any “Valid
data” will occur during the run system state and the waiting periods in between
occur during the wait state. Note that in Fig. 3.11 the length of the data vectors
is still the same after applying the delay. This is because the shift registers use
the SRL16/SRL32 mode of the slice LUTs which cannot be reset. Only the output
registers of the IP block can be reset [30]. Because of this we have chosen to let
any data still left in the shift registers propagate out before starting the delaying of
a new set of data. In the actual system there are 128 sets of delays corresponding
to a scan between 30 o and 150 o , which means that the counter in Fig. 3.11 for the
actual system needs to reach 127 before the entire sweep is finished. At that point

18
 3. System Design Process and Testing Methodology

the FSM asserts a flag to indicate the finished sweep, which is helpful for verifying
the frequency of the system.
SR
Counter ROM
A
D Q

A
D Q

A
D Q

A
D Q
Data

Figure 3.10: Illustration of how the delay lines are controlled. The ports on the
shift registers are the address input A the data input D and the data output Q.

Start

Raw 1

Raw 2

Raw 3

Raw 4

Counter 0 1 2 0

Wr. FIFO

Delayed 1

Delayed 2

Delayed 3

Delayed 4
Valid data #0 Valid data #1 Valid data #2

Figure 3.11: Timing diagram of the delay lines for a simplified version of the
system which steers the beam to only three different angles. The signals Raw 1-4
represent raw data streams from the sensors. The signals Delayed 1-4 represent the
data streams after they have passed through the shift registers. The signal Counter
selects the set of delay values associated with Valid data #0, #1 or #2 and routes
them to the address port of the shift registers. The signal Wr.FIFO tells the DPTI
that data is ready for transfer (this data is a single bit vector which is attained after
squaring and summing a whole data stream). In input A whole data streams are
allowed to propagate out of the shift registers before starting a new set of delays.

The purpose for the static beam state is to fix the beamformer at 90 o to enable
verification of the system. In this state the system changes from transmitting an
directional energy to the computer and instead transmits the output of the DAS
beamformer, i.e. audio. This has enabled, for instance, the comparison between the
system with and without the last FIR filter stage (implementing the system without
the last FIR filter resulted in a large amount of noise and clipping and thus this
filter was kept). This has also allowed for downloading samples of audio which has
been useful for analyzing the spectrum of sound sources (see Fig. 4.4).

19
3. System Design Process and Testing Methodology

The total amount of time it takes for the system to finish a complete sweep can
be calculated using Fig. 3.12 as a guideline: if τmax (i) = Max delay, τdata = Valid
data and τhold (i) = Hold time then

N
X
τtot = τdata + τmax (i) + τhold (i),
i=1

where τtot is the total delay, and N is the number of microphones. In the current
system three shift register IPs are being used which means that the time required
to empty them is divided by a factor of three, hence

N
X τhold (i)
τtot = τdata + τmax (i) + . (3.3)
i=1 3

For the current system with N = 24 and delay values calculated for angles between
30 o and 150 o τtot = 96.2 ms, meaning that the frequency of the system is fsys. =
τtot
−1
≈ 10.4 Hz.

Valid data

Delayed 1

Delayed 2

Delayed 3

Delayed 4
Max delay Hold time

Figure 3.12: This figure depicts how much time the sampling of data from a
single angle takes. Max delay represents the time it takes for sound to propagate
across the whole array. Valid data represents the time it takes to sample data. Hold
time represents the time it takes to let the shift registers clear out, at which point
scanning of the next angle can begin.

Finally, the FSM also controls a flag that controls the multiplexing of the two board
of microphones, the reason for which will be explained in Section 3.2.7.

3.2.5 Digilent Parallel Transfer Interface

For the FPGA used in this project an interfacing mechanism called DPTI is avail-
able which facilitates high-bandwidth communication between a host computer and
the FPGA board. It contains two interfaces, of which one is the application pro-
gramming interface (API) for the host, and another is the hardware interface on
the FPGA. The DPTI subsystem uses an 8-bit bidirectional data bus and several
control signals to achieve an asynchronous or synchronous parallel interface [31].

20
 3. System Design Process and Testing Methodology

The timing diagram of the synchronous read mode that is used the system is shown
in Fig. 3.13. In this mode, the transfer rate is 60 MHz and the maximum theoretical
bandwidth is 480 Mbps [32].

Figure 3.13: DPTI timing graph [32].

The synchronous read operation is started when PROG_RXFN is driven low by

host. Then the PROG_OEN is driven low by the peripheral system to turn around
the bus drivers before the PROG_RDN going low to acknowledge the data. After
PROG_OEN goes low, the first byte of data is on the PROG_D bus. The data
can be released from host by peripheral system when PROG_RDN is keeping low
or the wait state is inserted in the PROG_RDN. If more data need to be read, the
clock will be changed following PROG_RDN sampled low. When all data have been
consumed, the PROG_RXFN will be driven high by the host. After PROG_RXFN
is set to high, any data which appear on the data bus should be ignored [32].

The frame of input signal wr_di is shown in Table 3.1. Of this signal 24 bits are
dedicated to the directional energy data, one bit is dedicated for the Board Flag
and 7 bits are dedicated for the angle index.
Steered Response Power Board Flag Index
31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

Table 3.1: The frame of wr_di.

The protocol is initiated through the API in the host computer and terminates
when the user-specified number of bytes has been transferred. Currently the system
scans 128 angles which thus results in 128 directional energy values, each of which
is a 32-bit word. This means that transferring 128 · 32 bits = 4096 bits or 512 bytes
will ensure a complete set of directional energy data, but it is not known at which
point in the scan the transfer started. This is why the seven least significant bits
(log2 (128) = 7) of wr_di are used for providing the information about which angle
the directional energy values correspond to. In the same fashion it is also not known
whether the transferred data is coming from the horizontal or the vertical board,
which is why another bit, Board Flag, is dedicated to provide this information.

21
3. System Design Process and Testing Methodology

Thus the eight least significant bit of the 32-bit directional energy values are being
discarded.

3.2.6 Reading the Transferred Data and Calculating the

Moving Average

The API used to control the DPTI is written in C++, but we wanted to use MAT-
LAB for processing the data in order to get access to a multitude of MATLAB
functions. The currently implemented solution is that the DPTI API continuously
writes the directional energy data to a bit file and MATLAB continuously reads
from the same file as long as there is a complete set of data from a sweep. Upon
activating the MATLAB script it uses file position indicators to make sure the most
recent data is being read from the file.

Once ported into MATLAB each complete set of directional energy data is averaged
with a few of the previous sets to reduce transient behavior. Averaging just two
sets showed significantly reduced flickering in the subsequent heatmap image. When
averaging more than five sets of data, however, the image started to become sluggish;
it would take approximately a second for it to react to changes in sound source
locations. The MATLAB script has two modes: one which displays the two DOAs
and one which displays a heatmap. If the mode for two DOAs is selected then after
averaging, data from the two sensor boards are separated using the Board Flag and
displayed in polar plots. If the mode for displaying a heatmap is selected, the data
from the two sensor boards are separated and then combined in the fashion shown
in Section 3.2.7. The combined directional energy data is mapped onto a circle as in
Fig. 3.16b and plotted using the MATLAB function surf. This function allows for
displaying both a heatmap and a 3D plot, both of which are displayed in chapter 4.

Finally, the MATLAB function max is used for picking out the largest value of
the combined directional energy data, which is the candidate for the source. The
corresponding angles are then displayed on the axes of the heatmap.

3.2.7 Heat Map

In order to expand the system from calculating a DOA to calculating a position two
boards arranged perpendicularly to each other are needed, as per the arrangement
in Fig. 3.14a. Sensors from two array boards are multiplexed so that they take turn
in going through the processes described in Sections 3.2.1, 3.2.2, 3.2.3 and 3.2.5.
This expansion can be done with virtually no increased resource consumption, al-
beit with the trade-off that the number of locations per seconds is halved which,
according to (3.3), yields a frequency of fsys = 5.2 Hz. The boards are arranged in
the geometrical constellation shown in Fig. 3.14a.

Since the DPTI (as it is currently used) starts at arbitrary points in the system’s

22
 3. System Design Process and Testing Methodology

Board Flag
23:0
24#
Mics

Array B 23:0
Beamforming
System
23:0
24#
Mics

Array A

(b)

(a)

Figure 3.14: (a) shows the geometrical configuration of the arrays and (b) shows
the multiplexing of the two arrays shown in (a).

execution it is necessary to dedicate a bit of the transferred data to indicate whether

the data are associated with the horizontal or the vertical board, as earlier explained
in Section 3.2.5 and as can be seen in Fig. 3.1. With this information it is now
possible to “demux” the data from the two arrays which results in two vectors of
directional energy data, which correspond to θ and φ of the spherical coordinate
system. This means that the positions being scanned are in a “cone” with its vertex
at the sensor array’s center as seen in Fig. 3.15a.

To go from the aforementioned two 1-dimensional vectors to a 2-dimensional matrix

the outer product of the vectors is calculated:

C = b ⊗ a = bT × a, (3.4)

where a and b are 1 × 128 vectors representing directional energy data from sensor
array A and B (see Fig. 3.14) and C is a 128 × 128 matrix which represents the
2-dimensional directional energy. A way to conceptualize how this works is to put
some idealized data in (3.4) and see what happens. Say, for the sake of simplicity,
that five angles per sensor array are to be scanned and that there is a single sound
source somewhere in front of the arrays:

h i h i
a = 1 1 10 1 1 , b = 10 1 1 1 1

and

23
3. System Design Process and Testing Methodology

(a)

(b) (c)

Figure 3.15: This figure illustrates the region in front of the array which is being
scanned. In (a) the array and its scanning “cone” can be seen at an angle. In (b)
the array and the cone can be seen from the side and in (c) the array and the cone
can be seen from above. Note that this is just an illustration and in actuality the
angles θ and φ are much larger.

24
 3. System Design Process and Testing Methodology

   
10 10 10 100 10 10
 1 h 1 1 10 1 1 
   
  i  
 1  1 1 10 1 1 =  1
C= 1 10 1 1  . (3.5)
 
1 1 1 10 1 1 
  

1 1 1 10 1 1

The horizontal vector, a, has a large directional energy value in its middlemost
element, which represents a sound source at θ = 90 o ; the vertical vector, b, has
a large directional energy value in its leftmost element, which represents a sound
source at the array’s smallest angle (φ = 30 o for this system). When these matrices
are multiplied as in (3.4) it results in a combined set of directional energy values as
in (3.5). Finally, the C array can be used to visualize the combined beamformers.
One such visualization can be seen in Fig. 3.16, where the size of the dots corresponds
to the directional energy values. As seen in Fig. 3.15a the scanning area is in
actuality a circular disk, so to truthfully visualize the 2-dimensional directional
energy values from array C must be mapped onto a disk as seen illustrated in
Fig. 3.16b.

(a) (b)

Figure 3.16: This figure illustrates the results of combining the two perpendicular
beamformer. In (a) the directional energy values are straightforwardly mapped on a
square. In (b) they are mapped onto a circular disk, which represents a cross section
of the combined beamformers.

Before ending this section a comment should be made on the seemingly arbitrary
geometrical configuration of the sensor boards illustrated in Fig. 3.14a. Conceivably,
it would be more desirable to have an entirely symmetrical configuration, but it is not
possible to create a symmetrical configuration with two sensor boards perpendicular
to each other without blocking some sensors of one of the boards.

25
3. System Design Process and Testing Methodology

3.3 Design Decisions

This section is dedicated to explaining the various major design decisions which were
not obvious choices.

Choosing the DAS algorithm

As asserted in this project’s research statement we want to limit the complexity of

this project in order to maximize our chances of successfully implementing a working
prototype. This was the main factor in our decision to base the system on the DAS
beamformer.

Limiting the scanning angles between θ = 30 o and θ = 150 o

As mentioned in Section 3.2.2 the system only scans angles between θ = 30 o and
θ = 150 o . There are two reasons for designing the system this way. The first reason
is that the beamformer’s main lobe becomes less and less sharp as it approaches
0 o or 180 o . This can be observed in the simulation results in Section 4.1 and in
particular Fig. 4.2. This would have as a consequence that the system would not be
able to distinguish between angles close to θ = 0 o and θ = 180 o . The second reason
is that the maximum delay needed for the DAS algorithm increases depending on
how close to θ = 0 o or θ = 180 o the beamformer is steered. This can be seen in (3.1)
which, assuming θ is the only variable, will obviously be maximized for θ = 0 o or
θ = 180 o . A larger maximum delay means that the system frequency will be lower,
which provides further incentive to limit the scanning angles.

Sending additional information besides directional energy over the DPTI

As explained in Section 3.2.5 the eight least significant bits of the directional energy
values are being sacrificed to transfer one bit for indicating which sensor board is
being used and which angle each directional energy value belongs to. The reason
for this decision is simply because it was simple to implement. A more elegantly
designed system would, for instance, have the DPTI send a start signal to the FSM
which would then mean that the first transferred word would be known to correspond
to whichever angle and sensor board that the FSM starts at, which would allow for
using all 32 bits for the directional energy values. This was not done, however, as
the development time of the DPTI became too time consuming.

The choice to use DPTI

DPTI provides a higher bandwidth transfer of data compared to, e.g., UART. But
the system transfers 4096 bits every 92.6 ms as seen in (3.3) which means that its
transfer rate is only 4096 Bytes/92.6 ms ≈ 44.2 kbps, which can be comfortably han-
dled by the UART protocol [26]. The reason for choosing the higher-rate protocol,
DPTI, is to allow us to download and analyze data that have undergone only parts
of the algorithm or potentially even raw audio from the microphones. Both of these
actions require the transfer of much larger amounts of data and to keep options open
the decision was made in favor of DPTI.

26
 3. System Design Process and Testing Methodology

The number of scanning angles

The system’s ability to differentiate between scanning angles depends on the shortest
amount of time it can delay the data streams. With some trivial trigonometry it
can be shown that

cτmin
∆θ = arcsin( ), (3.6)
d

where c is the speed of sound, d is the distance between sensors and τmin is the
minimum delay. As mentioned in Section 3.2.2 there are three shift registers in
a row which means that τmin = 3 · fclk−1
= 3 · (4.608 MHz)−1 ≈ 651 ns. This with
d = 11.42 mm and c = 343 m/s (3.6) yields ∆θ ≈ 1.12 o . This means that in the
angles between θ = 30 o and θ = 150 o it should be possible to distinguish between
(150 − 30)/1.12 ≈ 107 angles. 107 values require seven bits to be represented in
the wr_di - bus, so the decision was made to use all the 128 values that can be
represented by seven bits. The strongest incentive for this decision was to increase
the resolution of the scan (every angle can be thought of as a pixel in an image).
Using slightly more than 107 values can be rationalized by remembering that only
adjacent sensors may have identical delay values and sensors spaced further apart
may not, so the DAS algorithm should by no means fail because of this. With 128
angles in the interval 30 o ≥ θ ≤ 150 o there will be 120 o /128 = 0.9375 o between
each angle. This would fundamentally be the smallest possible angle difference for
the system to distinguish, but as indicated in (3.6) the system may be unable to
distinguish between adjacent discrete angles.

Three shift register IPs in a row

As mentioned in Section 3.2.2 the current system uses three shift register IPs in a
row to achieve the desired delay. The most obvious alternative to doing this would
be to divide the clock by a factor of three or more and use only one shift register
IP, which was how the system was originally designed. There are two disadvantages
with this approach. The first disadvantage is that clock domain crossing requires
a measure to prevent metastability [33] in data, such as using FIFOs. The second
disadvantage is that clocking the shift registers with a slower clock will mean that
the minimum delay, τmin , will increase. The main incentive for using three shift
register IPs in a row was to achieve a lower τmin and thus increase the resolution of
the directional energy sweep. Unfortunately, the design is incomplete in the sense
that currently the delayed data must pass through all three shift registers, which
causes τmin to be three clock cycles. The reason this is not trivial to correct is
because the MATLAB script that calculates the delay values was designed for one
shift register IP and would have to be fundamentally redesigned for this problem
to be solved. Additionally, the FSM would have to be modified to control the data
streams so that they are able to pass through only one shift register IP and bypass
the other two.

27
3. System Design Process and Testing Methodology

Multiplexing of whole system

The obvious benefit of this is that the system could be expanded upon to relay two
DOAs instead of just one and that it could use twice the amount of microphones,
virtually without consuming any additional resources. The obvious drawback to
this is that the frequency of the system is halved, as mentioned in Section 3.2.7.
Another major advantage with the decision to do this, however, was the simplicity
of implementing it.

3.4 Testing Methodology

The testing methodology is based on the sound sources we have available, which
is any sound source that can be produced by regular consumer speakers, i.e. any-
thing below 22 kHz, and a piezoelectric ultrasonic transducer which has a center
frequency at 40 kHz. A square wave of 40 kHz, the center frequency of according to
its datasheet [27], is applied to the pins of the ultrasonic transducer. This produces
an ultrasonic sound which can be used for testing the system’s ability to localize
such sounds. The Static Beam mode will be used to analyze the spectrum of these
sound sources to ensure that the simulations and the system tests are performed
with similar sound sources.

The first category of tests will be to have a static source in front of the system using
one board with 24 sensors and thus only detecting the horizontal DOA. The second
category of tests will all be done on the system that uses two boards with a total of
48 sensors. The results of all these tests will be presented in a heatmap. The tests
will include:

• Testing with a single static source at 20 kHz

• Testing multiple static sources at 20 kHz

• Testing with a single static source at ultrasonic frequency

• Testing the system’s performance while varying the distance to the arrays

• Testing how well the system can detect very brief sounds

• Testing how well the system can detect sounds in a noisy environment

• Testing how sensitive the system is to small movements of the sound source

28
 4
Results

In this chapter, the project results are presented along with some observations.
First the simulation results are reviewed which shows what to expect out of the
implemented system and, subsequently, the implementation results are reviewed.

4.1 Simulation results

Fig. 4.1 shows the results of plotting (2.3) in MATLAB while to some extent at-
tempting to replicate the sound source we have available. The frequency spectrum
of the two different sound sources used in the tests can be seen in Fig. 4.4. The
sub-ultrasonic sound source, seen in Fig. 4.4b appears well-behaved and can thus
be simulated using a monochromatic source. The ultrasonic sound source seen in
Fig. 4.4a appears to have several harmonics and we have attempted to replicate
this spectrum in simulations. Specifically, the five peaks with the most power were
selected and a signal with power equally distributed on these five frequencies was
created for the simulation.

As predicted the sharpness of the beamformer is increased as the number of sensors

is increased. It should be noted that the grating lobes have a slightly decreased
amplitude, which is predicted in Section 2.2.

Another simulation was performed to investigate what happens at angles close to

θ = 180 o (and thus also θ = 0 o since the patterns are symmetrical). The results can
be seen in Fig. 4.2 and it should be noted that the width of the main lobe increases
significantly as the beamformer is steered towards θ = 180 o .

4.2 Implementation results

Currently the largest number of sensors we have been able to use is 48 sensors,
which is all the sensors on two sensor boards. The implementation result report from
Vivado indicates that we will not be able to add many more sensors. An overview of
the resource consumption can be seen in Table 4.3, which shows a large consumption

29
4. Results

90o 90o
o o o
120 60 120 60o

150o 30o 150o 30o

0 [dB] 0 [dB]
-10 -10
-20 -20
-30 -30
-40 -40
180o 0 o 180o 0o

(a) (b)
90o 90o
o o o
120 60 120 60o

150o 30o 150o 30o

0 [dB] 0 [dB]
-10 -10
-20 -20
-30 -30
-40 -40
180o 0 o 180o 0o

(c) (d)

Figure 4.1: Simulation of our system with parameters decided by our hardware;
f = 40 kHz, d = 11.43 mm and θ = 90 o . The number of sensors is 8 in (a) and
24 in (b). In (c) and (d) the system with 24 sensors, with θ = 120 o and θ =
150 o respectively, are plotted. The sound source is modeled after the ultrasonic
transducer.

90o 90o
o o o
120 60 120 60o

150o 30o 150o 30o

0 [dB] 0 [dB]
-10 -10
-20 -20
-30 -30
-40 -40
180o 0 o 180o 0o

(a) (b)

Figure 4.2: Simulation of our system with 24 sensors, f = 34 kHz, d = 11.43 mm

and θ = 150 o in (a) and θ = 175 o in (b).

30
 4. Results

of FPGAs LUTRAM. As mentioned already in Section 3.2.4 the system relays one
set of directional energy data at a frequency of fsys = 10.4 Hz when one sensor board
of 24 sensors is used. When two boards, with a total of 48 sensors, are used and
the two directional energy vectors are combined in MATLAB the resulting updating
frequency of the heatmap image is, as mentioned in Section 3.2.7, f = 5.2 Hz.

LUT 28%

LUTRAM 71%

FF 7%

BRAM 4%

DSP 54%

IO 14%

BUFG 22%

MMCM 20%

0 10 20 30 40 50 60 70 80 90 100

Figure 4.3: This figure displays the FPGA resource utilization of the system
using 48 sensors.

In Fig. 4.4a the frequency spectrum of the ultrasonic transducer collected using the
Static Beam mode. This is the only sound source we have had access to which reli-
ably emits sound of which the majority resides in the ultrasonic range. A lower fre-
quency sound generated by a tone generator from a cell phone is shown in Fig. 4.4b.
This result is what facilitated more accurate simulation results in Section 4.1.

In Fig. 4.6 the results from testing the system with 24 sensors with the 35 kHz
sound source can be seen. The patterns are very similar to those of Fig. 4.1, which
tells us the DAS algorithm has been implemented correctly. Note that the beamform
pattern is now limited to 30 o −150 o , the reason for which is explained in Section 3.3.

Upon multiplexing the whole system and separating it in MATLAB as described in

Section 3.2.7 two directional energy vectors are attained which can be seen plotted in
Fig. 4.7. It was at this point possible to verify in real time that the θ plot responded
to horizontal movements and the φ plot responded to vertical movement and work
on attaining the outer product of the two vectors could begin.

The results of the first test with a heatmap can be seen in Fig. 4.8. It is apparent

31
4. Results

Amplitude spectrum of summed sound

40

20

-20
Sound intensity [dB]

-40

-60

-80

-100

-120

-140

-160
0 1 2 3 4 5 6 7 8 9 10
Frequency [Hz] 104

(a) of summed sound

Amplitude spectrum
40

20

-20
Sound intensity [dB]

-40

-60

-80

-100

-120

-140

-160
0 1 2 3 4 5 6 7 8 9 10
Frequency [Hz] 104

(b)

Figure 4.4: This figure displays the analysis result of the spectrum of the two
sources. Subfigure (a) shows the ultrasonic source at f = 35 kHz and (b) shows a
sine wave at f = 20 kHz generated by a cell phone.

32
 4. Results

90o 90o
o o o
120 60 120 60o

150o 30o 150o 30o

0 [dB] 0 [dB]
-10 -10
-20 -20
-30 -30
-40 -40
180o 0 o 180o 0o

(a) (b)
90o

120o 60o

150o 30o

0 [dB]
-10
-20
-30
-40
180o 0o

(c)

Figure 4.5: This figure displays the test result of the implemented system with 24
sensors spaced d = 11.43 mm apart using a sound source with f = 20 kHz placed at
(a) θ = 90 o , (b) θ = 120 o and (c) θ = 150 o .

that the source position is more clearly discernible in a linear scale which is why all
subsequent test results will be presented this way.

Testing with the ultrasonic source provided an ambiguous pattern in the heatmap,
as can be seen in Fig. 4.9, but the highest directional energy appears in the correct
place and both angles displayed are correct as well.

The purpose of the next test is to find out how the distance between the source and
the arrays affects the system’s performance. It was carried out by placing a sound
source of 20 kHz and various distances right in front of the arrays. The results can
be seen in Fig. 4.10 where it becomes apparent that the system’s ability to locate
the source deteriorates drastically somewhere between 0.5 m - 1.0 m. The strongest
candidate appears at θ = 90 o and φ = 30 o . Additionally at 4.0 m there appears
to be mirror images near θ = 90 o and φ = 90 o . In general, mirror images seem
occasionally to appear in the vertical directional energy vector, but never in the
horizontal directional energy vector.

A great deal of testing with transient sounds such as claps, finger snaps and rattling
with keys have been tried with the array. All such testing has strongly indicated that
the system is incapable of locating such sounds, i.e. sounds that last for fractions
of a second. Ostensibly, sounds had to last for roughly one second for there to be a
distinguished peak of energy in the heatmap.

33
4. Results

90o 90o
o o o
120 60 120 60o

150o 30o 150o 30o

0 [dB] 0 [dB]
-10 -10
-20 -20
-30 -30
-40 -40
180o 0 o 180o 0o

(a) (b)
90o

120o 60o

150o 30o

0 [dB]
-10
-20
-30
-40
180o 0o

(c)

Figure 4.6: This figure displays the test result of the implemented system with 24
sensors spaced d = 11.43 mm apart using a sound source with f = 35 kHz placed at
(a) θ = 90 o , (b) θ = 120 o and (c) θ = 150 o .

0o
0 [dB] 30o

-10

90o -20 60o

o o
120 60
-30

-40
150o 30o
90o

0 [dB]
-10
-20
-30
-40
180o 0o
120o

(a)
150o
o
180

(b)

Figure 4.7: This figure displays the results of the implementation of two perpendic-
ular arrays with 48 sensors by plotting θ and φ separately, (a) shows θ and (b) shows
φ. The sound source is of the characteristics shown in Fig. 4.4b and is arbitrarily
placed in the room.

34
 4. Results

(a) (b)

(c) (d)

Figure 4.8: This plot shows the result of taking the outer product of the two
directional energy vectors. Subfigures (a) and (b) show the plot on a logarithmic
scale and (c) and (d) show the plot on a linear scale.

35
4. Results

(a) (b)

Figure 4.9: This image shows the results of testing source localization with the
ultrasonic transducer which is of the characteristics shown in Fig. 4.4a. The sound
source were placed at (a) θ = 90 o φ = 90 o and (b) θ = 102 o φ = 90 o .

Testing with steady sources in noisy environments showed that the sound source is
easily drowned out in sounds such as a normal conversation. Even sounds from our
footsteps as we performed the test could occasionally have a noticeable effect on
the system. Especially sounds at low frequency, specifically any sound below 4 kHz
would notably result in chaotic patterns in the heatmap. It should be noted that
since our system gathers 50 samples per angle at 192 kHz it only captures audio for
50 · (192 kHz)−1 ≈ 260 µs and can thus only capture a full wavelength, λ, of the
sound if λ ≤ c · 260 µs = 89.18 mm, where c = 343 m/s. This means that if the
frequency is below c/λ ≈ 3846 Hz the system will capture less than one wavelength
of the sound.

Finally, as mentioned in Section 3.4 tests were carried out to estimate the system’s
sensitivity to small movements of the sound source, i.e. to estimate the system’s
accuracy. The most interesting question to answer was the one raised in Section 3.3
regarding whether the system is actually able to distinguish between all the 128
discrete angles. Because the problems with spatial aliasing persisted and were more
prevalent with the ultrasonic source these accuracy tests were performed with the
lower-frequency source of 20 kHz. The testing method was to simply place the source
in front of the array, note the angles calculated by MATLABs max function and then
move the source in the horizontal direction until the displayed value for θ changed.
Unfortunately, the results of this test must be said to be inconclusive because the
calculated angles generally tended to fluctuate between two values.

36
 4. Results

(a) (b)

(c) (d)

Figure 4.10: This figure shows the results of testing the system with a sound
source at different distances from the array which is of the characteristics shown in
Fig. 4.4b and at θ = 90 o and φ = 90 o . The distance is (a) 0.5 m, (b) 1.0 m, (c)
2.0 m and (d) 4.0 m.

Figure 4.11: Two sound sources at 20 kHz.

37
 5
Discussion and Possible
Improvements

The chapter analyzes the results and observations presented in Chapter 4. First the
reasons for the results are discussed and this is followed by suggestions on possible
improvements that can be made to the system.

5.1 Discussion

A system has been designed that employs two one-dimensional arrays to create a
heatmap representing the directional energy in two dimensions.

5.1.1 General performance

As is clear from the results it is possible to perform localization on the source we

are using in our experiments, but this localization is not very reliable. According
to the simulation results shown in Fig. 2.4 spatial aliasing should become less of
a problem when power is more evenly distributed over a broad range of frequency,
which we were partly successful in verifying. The ultrasonic sound source we had
access to emitted sound in a rather narrow band so the difference in amplitude
between the main lobe and the grating lobes was just barely distinguishable. Sources
for ultrasound that span over broader bands of frequency must be employed to verify
the system’s capabilities, but it is far from trivial to produce such a source as any
equipment that produces sound at ultrasonic frequencies is not common. Research
into how to produce such a source should be carried out. It would also be interesting
to use ultrasound that has the spectral characteristics of animals that vocalize in
the ultrasonic range.

38
 5. Discussion and Possible Improvements

5.1.2 Brief sounds

Currently, the system cannot pick up sounds that last a very brief time, such as
clapping and snapping sounds. A possible reason for this is the low system frequency,
fsys . As mentioned in Section 3.2.4 it takes 96 ms for one full sweep to be finished.
This means that a sound which lasts a shorter time than this could occur when the
beamformer is pointing in another direction and a few milliseconds later, when the
beamformer is pointing in the direction which the sound originated from, the sound
could be gone. This would mean that the sound at its highest intensity would not
be picked up by the main lobe of the beamformer, but would nevertheless be strong
enough to overpower all other directional energy values. This would then result in
a strong directional energy in the wrong location on the heatmap.

One potential way for eliminating this problem would be to completely revise the
system so that it instead works by storing one audio sample set in a buffer for
each sensor and performing all processing on the same sets of sampled audio. This
idea is illustrated in Figs. 5.1a and 5.1b where four sampled sets are being cropped
out at different times. In Figs. 5.1c and 5.1d the resulting signals from adding the
cropped signals show that there is a combination of time-shifted data that results in
a maximum correlation which suggests that this method could be equivalent to the
DAS algorithm, if not mathematically then at least practically. The implications
of this would be that, while the audio data would have to be stored at 4.608 MHz,
the algorithm could be carried out at the maximum tolerable rate of the FPGA -
450 MHz [26]. Theoretically, this also means that all combinations of time-shifted
data, i.e. data from all angles, could be processed in parallel which would drastically
reduce latency. In the current implementation there was a lot of focus on the RAM-
based shift register IP which would not be used in such an implementation as just
described. The approach of storing audio would require a RAM that allows for
accessing any and all portions of the stored audio; this approach would essentially
require redesigning the entire system.

5.1.3 FPGA Resource Utilization

As seen in Fig. 4.3 the the LUTRAM resources come closest to being exhausted.
The resources that are a secondary candidate of being exhausted are the DSP slices.

The high consumption of LUTRAM resources is a result of the way the delay lines
are implemented; there is a need for large delays to allow for large arrays and
there is a need for small delays to allow for finer scanning resolution. These two
requirements combined results in a prohibitively large consumption of LUTRAMs.
The most straightforward remedy to this would be to just use one shift register
IP. This is a highly recommended course of action as the solution of using three
cascaded shift register IPs did not work entirely as intended; the main incentive
was to have a smaller minimum delay while still being able to produce the required
maximum delay (which is derived from the size of the physical size of the array),

39
5. Discussion and Possible Improvements

4 4

3.5 3.5

3 3

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0

-0.5 -0.5
-20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20

(a) (b)
2 2

1.5 1.5

1 1

0.5 0.5

0 0

-0.5 -0.5
0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140

(c) (d)

Figure 5.1: This figure illustrates the idea of performing all processing on the same
sets of audio data using an example with four audio streams. The blue waveforms in
(a) and (b) represent sampled audio coming from some direction (and thus arriving
at the sensors at different times) and the dashed boxes represent the act of selecting a
portion of the data for further processing. In (a) the four audio streams are cropped
with no time shift and if the data from these windows are added together it results
in the waveform seen in (c). In (b) the four audio streams are cropped with a time
shift causing the waveforms to synchronize and if the data from these windows are
added together it results in the waveform seen in (d).

40
 5. Discussion and Possible Improvements

but the minimum delay possible with this solution turned out to be one clock cycle
per shift register IP, i.e. three clock cycles. Currently the only benefit from this
solution is that the need for clock-domain crossing was eliminated.

Another, more ambitious, approach for reducing the consumption of LUTRAMs

would be to take the approach of storing audio in buffers and perform all processing
on these same sets of audio samples, as described in Section 5.1.2. This would
eliminate the need for the shift register IPs altogether and instead require, if there
is sufficient space for it, only one block RAM for each audio stream. The FPGA
used in this project has 445 available block RAM blocks of 36 kbit [34] which at
192 kHz facilitates (192 kHz · 16b)−1 · 36 kB ≈ 11.72 ms of stored audio in each block
RAM. Sound travelling at 343 m/s will have travelled about four metres in 11.72 ms,
which would suggest that a block RAM of 36 kb is more than sufficient to store all
the data needed for producing the delays required for a very large array.

To bring down the high consumption of DSP slices the multichannel capabilities of
the CIC filter and FIR filter IPs should be exploited. As seen in Fig. 5.2 several
data streams would be multiplexed and run through the filters. It is possible to run
the filters at a much higher clock rate than 4.608 MHz, which means that one single
chain of filters could handle several streams simultaneously without having to slow
down the rate of the data.

Cosine Wave

CDC

Microphone
Halfband Low- FIR Low-
MUX CIC Filter DeMUX Cosine Wave
pass Filter pass Filter

CDC
Microphone

Figure 5.2: This image illustrates how the chain of filters that perform the
PDM-to-PCM conversion could be serialized.

5.1.4 Mirror images in the vertical array

It can be seen from some of the results that there is often an mirror image present in
the vertical directional energy component of the heatmap. This is especially obvious
in Fig. 4.9a and Fig. 4.10d. A mirror image in the horizontal, however, has never
been observed during testing. A possible reason for this could be that, as can be
seen in Fig. 5.3a, the center of the horizontal array is aligned with the vertical board,
but the center of the vertical array is offset from the horizontal board. If this is the
case then the problem should disappear when the system is expanded to use four
sensor boards, because at this point an entirely symmetrical configuration, as the
one suggested in Fig. 5.3b, could be constructed.

41
5. Discussion and Possible Improvements

5.1.5 Noisy environment

It is very clear from testing that ambient sound, such as people conversing, adversely
inhibits the systems ability to distinguish sound sources. A great improvement to
the system could be to combat such effects by high-pass filtering the sound. This
filtering stage would have to be put somewhere before the integration stage in the
FPGA; it is not possible to do it on the PC end.

Center of vertical array

Center of vertical array

Center of horizontal array Center of horizontal array

(a) (b)

Figure 5.3: This figure displays (a) the geometrical configuration of the current
array and (b) a symmetrical configuration for a system using four sensor boards.

5.2 Potential improvements

There are some potential improvements that could be made to the system. This
section contains a list of recommended actions arranged in descending order of ur-
gency.

Investigating how to pick up short-duration sounds

The systems inability to locate brief sounds is perhaps the most unfavorable problem
with the current system; many applications may require the array to be capable of
picking up sounds of very short duration, such as footsteps. At the moment the only
hypothesis we have as to why the system cannot locate brief sounds relates to the
latency of the system, so our recommendation would be to investigate this.

Implementing a high-pass filter

As mentioned in Section 5.1.5 the system does not perform well in noisy environment.
Arguably the most important improvements is to implement a high-pass filter to
remove low frequency components, which is likely to drastically improve the system’s
performance. For instance, FIR high-pass filters could be implemented for each data

42
 5. Discussion and Possible Improvements

stream as seen in Fig. 5.4a and thus consume 24 DSP slices since one FIR high-pass
filter consumes one DSP slice which can be known from FIR IP in Vivado. It is
also conceivable that a single high-pass filter could be implemented after the DAS
algorithm as seen in Fig. 5.4b and thus consume only a single DSP slice.

Data Stream High-pass Data Stream

Filter
Data Stream High-pass Data Stream
Filter
Data Stream High-pass Data Stream High-pass
Filter Filter
24 Data
24 Data Streams Streams

Data Stream High-pass Data Stream

Filter

(a) (b)

Figure 5.4: This figure displays (a) implementing high-pass filter in parallel and
(b) implementing high-pass filter in serial.

Investigating ways to circumvent the fundamental delay

Some literature suggest systems that use much larger arrays compared to the one
used in this project. This would mean the time for the sound to propagate from
the first microphone to the last would be too large to be able to achieve real-time
localization with the approach we are using. This indicates to us that there are
methods to improve the system in terms of latency and this would have to be
investigated.

Refining the transfer interface between FPGA and PC

Currently the transfer of information between the host computer and the FPGA is
one-sided; information can only be sent from the FPGA to the host computer. As
explained in Section 3.2.5 the angle index and Board Flag is sent along with the
directional energy data, which is a non-ideal solution. A natural refinement to the
system would be to allow for the host computer to send a signal to the FSM so that
it is known which set of directional energy data will be transferred first.

Serializing the PDM to PCM conversion

One way to save FPGA resources is possible to serialize the PDM to PCM conversion.
The filters could potentially be run at a much higher frequency and several audio
data streams could be multiplexed to save hardware resources (see Fig. 5.2).

Using ring buffers

To avoid unnecessary storage of calculated delay values the possibility of using ring
buffers should be investigated.

43
 6
Conclusion

The goal of this master thesis was to design a system that uses a massive MEMS
microphone array and beamforming to perform real-time positioning of ultrasonic
sources. Another key objective was to maximize the amount of microphones in the
array, since this improves the accuracy of which the array can localize sound sources,
and using an FPGA was hypothesized to be a suitable tool for this purpose.

A system that can perform sound source localization has been designed and it can,
arguably, be said to operate in real time at a rate of 5.2 locations per second.
The localization is visualized in a heatmap which displays energy originating from
different directions.

In regards to the goal of maximizing the number of microphones: the final version
of the system uses 48 microphones. The reason it was not feasible to increase this
number was because of the system’s large latency, which originates from the way
the system implements the DAS algorithm, the speed of sound and the physical size
of the array. If the latency were to become even higher the system could no longer
be said to operate in real time. Thus it was this latency, and not a shortage of
FPGA resources, that limited the number of microphones, leading to the conclusion
that the first step of any future development should be to completely revise how
the system implements the DAS algorithm as discussed in Section 5.1.2. If this
revised version of the system could be implemented and the new limiting factor
instead becomes high consumption of FPGA resources there are ways to reduce the
resource consumption as discussed in Section 5.1.3.

Any conclusions regarding the accuracy of the system were, as stated in Section 4.2,
difficult to make. What can be said for certain is that the fundamentally smallest
angle difference the system can distinguish is 0.9375 o but it is possible that this
angle is slightly larger in practice, as discussed in Section 3.3. Decreasing this angle,
and thus increasing the resolution of the heatmap, would lead to a larger latency.
Again, it must be concluded that the way the system implements the DAS algorithm
should be revised.

44
 Bibliography

[1] L. Dai, Intelligent Macromolecules for Smart Devices. Springer London, Jan.
2004, pp. 451–452.
[2] H. Krim and M. Viberg, “Two decades of array signal processing research:
The parametric approach”, IEEE Signal Processing Magazine, vol. 13, no. 4,
pp. 67–94, Jul. 1996.
[3] S. Li and M. Stanaćević, “Source separation in noisy and reverberant environ-
ment using miniature microphone array”, in 2014 48th Asilomar Conference
on Signals, Systems and Computers, Nov. 2014, pp. 446–449.
[4] A. R. Abu-El-Quran, R. A. Goubran, and A. D. C. Chan, “Security monitor-
ing using microphone arrays and audio classification”, IEEE Transactions on
Instrumentation and Measurement, vol. 55, no. 4, pp. 1025–1032, Aug. 2006.
[5] D. T. Blumstein, D. J. Mennill, et al., “Acoustic monitoring in terrestrial envi-
ronments using microphone arrays: Applications, technological considerations
and prospectus”, Journal of Applied Ecology, vol. 48, no. 3, pp. 758–767, Jun.
2011.
[6] J. Steckel and H. Peremans, “Ultrasound-based air leak detection using a
random microphone array and sparse representations”, in IEEE SENSORS
2014 Proceedings, Nov. 2014, pp. 1026–1029.
[7] E. E. Case, A. M. Zelnio, and B. D. Rigling, “Low-cost acoustic array for
small UAV detection and tracking”, in 2008 IEEE National Aerospace and
Electronics Conference, Jul. 2008, pp. 110–113.
[8] J. Benesty, J. Chen, and Y. Huang, Microphone Array Signal Processing.
Springer-Verlag Berlin Heidelberg, 2008, pp. 40–46, 192–193.
[9] M. Fallahi, M. Blau, et al., “Optimizing the microphone array size for a virtual
artificial head”, The international Symposium on Auditory and Audiological
Research (ISAAR), vol. 6, pp. 359–366, Jan. 2018.
[10] M. Brandstein and D. Ward, Signal Processing Techniques and Applications.
Springer-Verlag Berlin Heidelberg, 2001, pp. 158–164, 181–215.
[11] R. Etienne-Cummings and M. Clapp, “Architecture for source localization
with a linear ultrasonic array”, in ISCAS 2001. The 2001 IEEE International
Symposium on Circuits and Systems (Cat. No.01CH37196), vol. 3, May 2001,
181–184 vol. 2.
[12] J. Grythe, Beamforming algorithms - beamformers, https : / / wpstatic .
idium . no / web2 . norsonic . com / 2016 / 10 / TN - beamformers . pdf, Oslo,
Norway.

45
Bibliography

[13] J. Tiete, F. Domínguez, et al., “Soundcompass: A distributed MEMS mi-

crophone array-based sensor for sound source localization”, Sensors (Basel,
Switzerland), vol. 14, no. 2, pp. 1918–1949, Jan. 2014.
[14] D. Petersen and C. Howard, “Simulation and design of a microphone array for
beamforming on a moving acoustic source”, Acoustics 2013 - Victor Harbor,
Nov. 2013.
[15] U. Farooq, Z. Marrakchi, and H. Mehrez, Tree-based Heterogeneous FPGA
Architectures, Application Specific Exploration and Optimization. Springer-
Verlag New York, 2012, pp. 7–8.
[16] B. Goel, Introduction to FPGA, 2017.
[17] D. Todorović, I. Saln, et al., “Implementation and application of FPGA plat-
form with digital MEMS microphone array”, in 4th International Conference
on Electrical, Electronics and Computing Engineering, Kladovo, Serbia, Jun.
2017, AKI2.2.1–6.
[18] Xilinx. (May 2018). Company Overview, [Online]. Available: https://www.
xilinx.com/about/company-overview.html (visited on 06/19/2018).
[19] ——, (May 2018). Vivado Design Suite - HLx Editions, [Online]. Available:
https://www.xilinx.com/products/design-tools/vivado.html (visited
on 06/19/2018).
[20] D. S. Jacobs and A. Bastian, Predator–Prey Interactions: Co-evolution between
Bats and Their Prey. Springer International Publishing, 2016.
[21] R. Mucci, “A comparison of efficient beamforming algorithms”, IEEE Trans-
actions on Acoustics, Speech, and Signal Processing, vol. 32, no. 3, pp. 548–
558, Jun. 1984.
[22] B. Zimmermann and C. Studer, “FPGA-based real-time acoustic camera pro-
totype”, in Proceedings of 2010 IEEE International Symposium on Circuits
and Systems, May 2010, pp. 1419–1419.
[23] D. E. Dudgeon, “Fundamentals of digital array processing”, Proceedings of the
IEEE, vol. 65, no. 6, pp. 898–904, Jun. 1977.
[24] J. Dmochowski, J. Benesty, and S. Affes, “On spatial aliasing in microphone
arrays”, IEEE Transactions on Signal Processing, vol. 57, no. 4, pp. 1383–
1395, Apr. 2009.
[25] Digital zero-height sisonic microphone with multi-mode and ultrasonic support,
Knowles Electronics, May 2014.
[26] Genesys 2 FPGA board reference manual, Digilent, Aug. 2017.
[27] Ultrasonic transducer, Kobitone Audio Company, May 2006.
[28] R. Lyons, “Understanding cascaded integrator-comb filters”, Besser Associates,
Tech. Rep., Mar. 2005.
[29] A. V. Oppenheim, A. S. Willsky, and S. H. Nawab, Signals & systems, English,
2. ed. Upper Saddle River, N.J: Prentice Hall, 1997.
[30] RAM-based shift register v12.0, XILINX, Nov. 2015.
[31] Digilent adept parallel transfer interface (DPTI) programmer’s reference man-
ual, DIGILENT, Jun. 2015.
[32] Digilent parallel transfer interface (DPTI), DIGILENT, Jun. 2016.
[33] R. Ginosar, “Metastability and synchronizers: A tutorial”, IEEE Design Test
of Computers, vol. 28, no. 5, pp. 23–35, Sep. 2011.

46
 Bibliography

[34] 7 series FPGAs memory resources, XILINX, Sep. 2016.

47