Daniel

 Sanz  Ausín  

Experimental  
Investigation  Of  
Loudspeaker  Power  
Requirements  
 
Master’s  thesis,  January  2015  
Experimental  Investigation  Of  Loudspeaker  Power  Requirements  
 

Author:  
Daniel  Sanz  Ausín  
 
Supervisors:  
Michael  A.  E.  Andersen  
Henrik  Schneider  

DTU  Elektro  
Technical  University  of  Denmark  
2800  Kgs.  Lyngby  
Denmark  
 
[email protected]    

 
 
 
 
 
 
 
 
Project  period:   25/08/2014-­‐25/01/2015  
 
ECTS:   30  
 
Class:   Public  
   
Edition:   1.  edition  
 
 
 Abstract
The components of a sound system, whether it is a small system for
an electronic device or a high power one, need to be designed to meet
certain power requirements. For this, manufacturers often design these
components using test-signals, which have an unchanging amplitude.
These are very different from real audio signals, because, in audio,
the amplitude is continuously changing over the time, and high power
peaks can be found, which have a small duration in comparison with
the continuous low power values.
This project follows the work done by a group of PhD students and
professors of the Technical University of Denmark, where the power
requirements of a loudspeaker were investigated when it is reproduc-
ing different audio files. For this, over 400 songs were used, and the
simulations were done with the mathematical linear model of the loud-
speaker, which will also be explained in this document. The results
of the simulations showed, as expected, that loudspeakers have a high
power consumption for short amounts of time, and a low continuous
power consumption. The goal of the investigation is to avoid over-
sizing and unnecessary costs when designing all the units of a sound
system, such as the power supply unit, the audio amplifier and the
loudspeaker driver.
The main goal of this project is to validate the functionality of the
mathematical model of the loudspeaker, to make sure that the simu-
lations on this model give the same results as measurements in a real
loudspeaker. For this, a circuit has been designed and built, which
allows to send an audio signal to a real loudspeaker system and mea-
sure the power consumption of the loudspeaker driver. Measurements
taken on this circuit have been compared to simulations using the lin-
ear and non-linear models, which are also explained in the document.
The results show, indeed, that the models simulate the behaviour of
the real loudspeaker with high precision, so the results from the sim-
ulations are absolutely valid. Moreover, the difference between the
measurements on the linear and the non-linear models is very small,
which confirms that the linear model is enough for the simulations,
which is more simple than the non-linear one.
The validity of the mathematical model has been confirmed making
measurements with various audio files belonging to different musical
styles. The power requirements give very similar results both for the
real loudspeaker and the simulation models.
Finally, a database of over 100 loudspeaker drivers of the same size
has been taken, and simulations have been done on each one of them:
all the loudspeakers must have the same sound pressure level response,
so a filter is applied to each one to make them all sound the same.

i
Then, simulations have been run and the power requirements have
been analysed. The results show how the range of power consumption
can be quite wide for the loudspeakers, even if they are similar and
are forced to have the same sound response. This explains why design
for the worst case needs to be done.

ii
Contents
1 Introduction 1
1.1 Thesis Objective . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Organisation of the Thesis . . . . . . . . . . . . . . . . . . . . 2

2 Loudspeaker Modelling 4
2.1 Loudspeaker Parts And Operation . . . . . . . . . . . . . . . . 4
2.2 Domains and Conversions . . . . . . . . . . . . . . . . . . . . 5
2.2.1 Electrical Domain . . . . . . . . . . . . . . . . . . . . . 5
2.2.2 Mechanical Domain . . . . . . . . . . . . . . . . . . . . 5
2.2.3 Acoustical Domain . . . . . . . . . . . . . . . . . . . . 6
2.3 Electro-Mechanical Conversion and Final Linear Model . . . . 7
2.4 Non-Liniarities . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5 Final Non-Linear Model . . . . . . . . . . . . . . . . . . . . . 11

3 Loudspeaker Power Measurement Circuit 13

3.1 Audio Power, RMS and Peak Power . . . . . . . . . . . . . . . 13
3.2 Sampling Theorem and Aliasing . . . . . . . . . . . . . . . . . 14
3.3 Power Measurement Circuit: Main Components . . . . . . . . 16
3.3.1 The Loudspeaker: Monacor SPH-170TC . . . . . . . . 18
3.3.2 The Amplifier: Texas Instruments TPA3116D2 . . . . 18
3.3.3 The 10mΩ Current Measurement Resistor . . . . . . . 19
3.3.4 The Instrumentation Amplifier: Texas Instruments INA2128 19
3.3.5 Operational Amplifiers for Filter design: MC33079 . . 19
3.4 Testing Anti-Aliasing Filters . . . . . . . . . . . . . . . . . . . 21
3.5 ADC & Software: Matlab & National Instruments USB-6356 . 24
3.5.1 Matlab . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.5.2 NI USB-6356 . . . . . . . . . . . . . . . . . . . . . . . 24
3.6 Audio Amplifier Model . . . . . . . . . . . . . . . . . . . . . . 25

4 Loudspeaker System Measurements, Simulations & Compar-

ison 29
4.1 Preparing the Audio File . . . . . . . . . . . . . . . . . . . . . 29
4.2 Power Measurements at two Sound Pressure Levels . . . . . . 30
4.2.1 Circuit Measurements . . . . . . . . . . . . . . . . . . 30
4.2.2 Simulations: Linear and Non-Linear Loudspeaker Models 30
4.3 Comparison & Conclusions . . . . . . . . . . . . . . . . . . . . 31

iii
5 Power Window Sweep 35
5.1 Power Window Sweep Algorithm . . . . . . . . . . . . . . . . 35
5.2 Comparing Results & Conclusions . . . . . . . . . . . . . . . . 37

6 Loudness Normalisation & Measurements with Different Songs 41

6.1 Loudness Normalisation . . . . . . . . . . . . . . . . . . . . . 41
6.2 Simulations with Different Songs . . . . . . . . . . . . . . . . 42
6.2.1 Preparation of the Songs . . . . . . . . . . . . . . . . . 43
6.2.2 Measurements . . . . . . . . . . . . . . . . . . . . . . . 43
6.2.3 Power Window Sweep . . . . . . . . . . . . . . . . . . 44

7 Statistical Analysis of Loudspeaker Power Requirements 49

7.1 Thiele-Small Parameters Database . . . . . . . . . . . . . . . 49
7.2 Target Sound Pressure Level & Filter design . . . . . . . . . . 50
7.3 Simulation & Power Window Sweep . . . . . . . . . . . . . . . 51
7.4 Comparison & Conclusions . . . . . . . . . . . . . . . . . . . . 54

8 Conclusion 57

9 Future Work 59

Appendices c

A Matlab Files c

B Other Files e

iv
List of Figures
2.1 Loudspeaker Driver. Source:[5] . . . . . . . . . . . . . . . . . . 4
2.2 Electrical model of the voice coil. Source:[3] . . . . . . . . . . 5
2.3 Mechanical model of the loudspeaker. Source:[3] . . . . . . . . 6
2.4 Electro-Mechanical equivalent circuit of the loudspeaker. Source:[3] 7
2.5 Linear model of the loudspeaker in Simulink . . . . . . . . . . 8
2.6 Linear model of the loudspeaker used in LTspice . . . . . . . . 9
2.7 force factor (Bl) vs diaphragm displacement (x). Source:[3] . . 10
2.8 voice coil inductance (Le) vs diaphragm displacement (x). Source:[3] 10
2.9 mechanical suspension compliance (Cm) vs diaphragm dis-
placement (x). Source:[3] . . . . . . . . . . . . . . . . . . . . . 11
2.10 Non-linear model of the loudspeaker in Simulink . . . . . . . . 12
3.1 Typical waveform of an audio file, showing high peak power
and low rms values. Source:[7] . . . . . . . . . . . . . . . . . . 14
3.2 Correctly and incorrectly sampled signals. Source:[8] . . . . . 14
3.3 Aliasing: frequencies being mirrored due to wrong sampling.
Source:[8] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.4 Typical frequency response of a first order low-pass filter . . . 16
3.5 Block diagram of the whole power measurement circuit . . . . 17
3.6 Current measurement resistor in series with the loudspeaker . 17
3.7 Picture of the power measurement circuit . . . . . . . . . . . . 18
3.8 Sallen-Key configuration for a 2nd order low-pass filter . . . . 20
3.9 Frequency response of the 8th order low-pass filter, using the
Sallen-Key configuration . . . . . . . . . . . . . . . . . . . . . 21
3.10 FFT of the unfiltered signal in the output of the DAC . . . . . 22
3.11 FFT of the low-pass filtered signal in the output of the DAC . 22
3.12 FFT of the unfiltered signal in the output of the instrumenta-
tion amplifier (current channel) . . . . . . . . . . . . . . . . . 23
3.13 FFT of the low-pass filtered signal in the output of the instru-
mentation amplifier (current channel) . . . . . . . . . . . . . . 23
3.14 Gain of the audio amplifier over the frequency, depending on
the connected load . . . . . . . . . . . . . . . . . . . . . . . . 26
3.15 Matlab figure showing the constant gain of the amplifier for a
high level signal: Resistive Load = Red , Loudspeaker Load
= Blue. (for more precision, only the peaks of the sinusoidal
waveforms are shown) . . . . . . . . . . . . . . . . . . . . . . 27
3.16 Matlab figure showing the constant gain of the amplifier for a
lower level signal: Resistive Load = Red , Loudspeaker Load
= Blue. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.1 Band-pass filter used for the audio file on Simulink . . . . . . 29

v
4.2 Comparison of the voltage between the real measurements
(green) and the simulation on the linear model (blue) and
non-linear model (red) for low level sound pressure level. The
red line cannot be seen because it is exactly the same as the
blue one. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3 Comparison of the current between the real measurements
(green) and the simulation on the linear model (blue) and
non-linear model (red) for low level sound pressure level . . . . 32
4.4 Comparison of the voltage between the real measurements
(green) and the simulation on the linear model (blue) and non-
linear model (red) for high level sound pressure level. The red
line cannot be seen because it is exactly the same as the blue
one. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.5 Comparison of the current between the real measurements
(green) and the simulation on the linear model (blue) and
non-linear model (red) for high level sound pressure level . . . 33
4.6 Comparison of the power curves of the real measurements
(green) and the simulation on the linear model (blue) and
non-linear model (red) . . . . . . . . . . . . . . . . . . . . . . 34
5.1 Window sweep over the power curve of the loudspeaker with
different window sizes. Source:[1] . . . . . . . . . . . . . . . . 36
5.2 Maximum rms values as a function of the window size. Source:[1] 37
5.3 Power window sweep algorithm applied to high level measure-
ments. Real measurement (green), simulation on the linear
model (blue) and non-linear model (red). . . . . . . . . . . . . 38
5.4 Power window sweep algorithm applied to low level measure-
ments. Real measurement (green), simulation on the linear
model (blue) and non-linear model (red). . . . . . . . . . . . . 38
5.5 Difference between the real measurements and each one of the
simulation model for high level (difference with linear model
= green, difference with non-linear model = red) . . . . . . . . 40
5.6 Difference between the real measurements and each one of the
simulation model for low level (difference with linear model =
green, difference with non-linear model = red) . . . . . . . . . 40
6.1 Frequency response of the two filters that compose the K-filter.
Source:[1] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.2 Short period of song no. 1 showing the comparison of the
power consumption (for high level). Real measurement (green),
simulation on the linear model (blue) and non-linear model (red). 44

vi
 6.3 Short period of song no. 4 showing the comparison of the
power consumption (for high level). Real measurement (green),
simulation on the linear model (blue) and non-linear model (red). 45
6.4 Song Number 1: Max RMS Power VS Window Size HIGH
LEVEL (Real Measurement = green , Simulation on Linear
Model = blue , Simulation on Non-Linear Model = red) . . . . 45
6.5 Song Number 4: Max RMS Power VS Window Size HIGH
LEVEL (Real Measurement = green , Simulation on Linear
Model = blue , Simulation on Non-Linear Model = red) . . . . 46
6.6 Power window sweep comparison for all the songs on high level.
Curves correspond to real measurements on the circuit (1-red,
2-blue, 3-green, 4-yellow, 5-magenta, 6-black). . . . . . . . . . 47
6.7 Power window sweep comparison for all the songs on low level.
Curves correspond to real measurements on the circuit (1-red,
2-blue, 3-green, 4-yellow, 5-magenta, 6-black). . . . . . . . . . 48
7.1 Bode diagram showing the SPL of each loudspeaker over the
frequency, theoretically measured at 1m distance with a 2.83V
input voltage . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
7.2 Bode diagram of the target SPL transfer function . . . . . . . 51
7.3 Bode diagram of the filter designed for each loudspeaker . . . 51
7.4 Simulink model used, where the linear model is combined with
the filter of each loudspeaker . . . . . . . . . . . . . . . . . . . 52
7.5 Power consumption of each loudspeaker for the same SPL.
Only a short period of the audio curve is shown . . . . . . . . 53
7.6 Result of power window sweep algorithm applied to 128 loud-
speakers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
7.7 histfit function used to display the distribution of the peak
apparent power values . . . . . . . . . . . . . . . . . . . . . . 55
7.8 histfit function used to display the distribution of the rms
apparent power values . . . . . . . . . . . . . . . . . . . . . . 55
7.9 Ratio of the peak apparent power to the rms apparent power
for the 128 loudspeakers . . . . . . . . . . . . . . . . . . . . . 56

List of Tables
1 Data from the measurement of the loudspeaker system when
playing an audio file . . . . . . . . . . . . . . . . . . . . . . . 30
2 Peak and rms power values and their ratio, the crest factor,
for each song . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

vii
1 Introduction
Nowadays loudspeakers can be found everywhere, from small electronic de-
vices such as MP3 players and mobile phones, to professional level sound
systems. The range of power requirements of these loudspeakers can also be
very extensive, and the components of the system, such as the power supply,
the cables, the amplifier, the filters and the loudspeakers need to be designed
to meet these power specifications. Before the design of all these parts, it is
important to perform investigation of different loudspeakers and configura-
tions when they are playing audio files, to get knowledge about the power
requirements.
The project presented here follows the work done by a group of PhD stu-
dents and professors of the Technical University of Denmark: Requirements
Specification for Amplifiers and Power Supplies in Active Loudspeakers [1].
Here, a mathematical model for the loudspeaker is described, which can
be used to simulate the behaviour of the loudspeaker in order to perform
measurements. This paper also introduces a method to get the power re-
quirements of the loudspeaker more clearly, when playing an audio file.
For this project, a circuit to measure the power of a real loudspeaker has
been designed and built in order to check that the model of the loudspeaker
introduced in the previous paragraph gives similar results to the real loud-
speaker. After that, different measurements have been performed and the
results of the real loudspeaker and the models have been compared.

1.1 Thesis Objective

The main motivation of the project is to avoid oversizing and unnecessary
cost in the elements of a sound system, such as the power supply, the amplifier
and the loudspeaker driver. For this, a lot of research needs to be done, and
this project focuses in part of this investigation.
The paper mentioned before [1] explains how loudspeakers have a high
power consumption for very short time periods and low consumption for
longer periods (high power peaks and low continuous values). However, man-
ufacturers usually test the power requirements of their loudspeakers using
test-signals that have a constant power, which is not the case of the common
audio files, so their loudspeakers are oversized. It is believed that this over-
sizing can be avoided and costs can be reduced if it is known for how long
these peaks of power last and how often they are repeated, because, this way,
a sound system that meets the power requirements of real audio files can be
designed.
In [1], a research of the power requirements of over 400 audio files is

1
explained, and the mathematical model of the loudspeaker is used to make
simulations. The main goal of this project is to compare results of measure-
ments in mathematical models with real measurements, to check that the
models are valid. Two models are used: the linear one and the non-linear
one, which is more complex. It will be analysed if the linear model is valid to
simulate the behaviour of the real loudspeaker or if, instead, the non-linear
model is needed.
Finally, an statistical analysis of power requirements of multiple similar
loudspeakers is performed, in order to know what is the power consumption
range of all of them. The goal of this is to be able to design the power
supply unit and the rest of the parts of the sound system so that they meet
the power requirements of all the loudspeakers, so design for the worst case
needs to be done. The design of these components is not part of this project,
but would be a future step.

1.2 Organisation of the Thesis

The chapters of this document are explained in this section:
In Chapter 2, the dynamic loudspeaker is explained. First, its physical
parts are explained, and after that the electrical, mechanical and acoustical
domains are analysed. All the domains are included in the mathematical
linear model of the loudspeaker. Finally, the non-linearities are introduced
to the model.
Chapter 3 introduces first the concept of audio power. After that, the
Sampling Theorem and aliasing are explained. It is also in this chapter where
the circuit to make the loudspeaker power measurements is explained: the
components and software tools used are introduced, and the design of the
anti-aliasing filters is described. Finally, the linearity of the audio amplifier
over the frequency is measured.
In Chapter 4 the first measurements are done in the circuit described
in the previous chapter. These measurements are afterwards compared to
results obtained with the linear and non-linear models of the loudspeaker.
Chapter 5 explains first the power window sweep algorithm used in this
project to compare the results of the power measured with different meth-
ods. First, the algorithm is explained, and then it is applied to the power
consumption curves measured in Chapter 4.
In Chapter 6, loudness normalisation is explained. Then, it is applied
to six different songs to make them sound equally loud. Finally, the same
measurements and comparisons explained in Chapters 5 and 6 are made for
these six songs.

2
 Chapter 7 is a little bit different than the rest. Here, a database of more
than 100 similar loudspeakers is taken and a target sound pressure level
response is designed, to make all of them sound equally loud. For this, one
filter is designed for each one of them, that will make all the loudspeakers have
the same sound response. Finally, the same measurements (only simulations)
and comparisons as in Chapters 5 and 6 are applied to all the loudspeakers,
and their power requirements are compared.
At the end, the final conclusions of the project are stated, and the future
work that follows is explained.
The appendices explain the files that are related to this project. However,
these files are not needed for the understanding of this document.

3
2 Loudspeaker Modelling
2.1 Loudspeaker Parts And Operation
[2] The loudspeaker is an electroacoustic transducer than converts an electri-
cal signal to an acoustic signal. The main parts of the dynamic loudspeaker
are the voice coil, the pole piece, the magnet, the diaphragm or cone, the
spider, the dust cap, the outer suspension, the top plate and the basket.

Figure 2.1: Loudspeaker Driver. Source:[5]

When the loudspeaker receives an electrical signal, the current flows in

the voice coil and an electro-magnetic force is induced in the permanent
magnetic field in the air gap. This force causes the displacement of the voice
coil, which is then transferred to the diaphragm and this one radiates sound.
The loudspeaker can be split into three domains:

• Electrical domain

• Mechanical domain

• Acoustical domain

Each one of these domains and the conversion between them is explained
in the next section.

4
2.2 Domains and Conversions
2.2.1 Electrical Domain
[2] The analogous electrical circuit of the loudspeaker is formed with the
impedance of the voice coil, which is modelled as a resistor Re in series with
an inductor Le . u(t) is the input voltage signal, which introduces a current
in the voice coil, i(t). Blv(t) is the force induced to the mechanical domain
in the voice coil, which makes the diaphragm move with a velocity v(t). Bl
is the force factor where B is the magnetic flux density in the air gap and l
is the effective length of the voice coil.
The parameters mentioned before, as well as some others, are called
Thiele-Small Parameters, and are usually found in the data sheets of the
loudspeakers. These electro-mechanical parameters define the performance
of a loudspeaker driver.

Figure 2.2: Electrical model of the voice coil. Source:[3]

Applying Kirchoff ’s voltage law to the previous circuit, the transfer func-
tion of the electrical circuit is shown in equation (2.1).

u (t) = Re i(t) + jωLe i(t) + Blv(t) (2.1)

where ω = 2πf and f is the frequency of the input signal.
This project is focused in Woofer loudspeakers, thus, in the low frequency
range. The previous electrical model is sufficient for this frequency range,
but, if higher frequencies were being analysed, the effect of Eddy currents
would have to be taken into account and a resistor Re0 would need to be
added to the model in parallel with Le .

2.2.2 Mechanical Domain

The mechanical circuit of the loudspeaker includes: Cm ; the suspension com-
pliance, Mm ; the mass of the diaphragm, and Rm ; the mechanical damping

5
factor. Bli(t) is the force induced from the electrical domain into the me-
chanical domain, which causes the diaphragm to move with a velocity v(t)
due to the input current i(t). Finally, ZA is the representation of the acoustic
impedance in the mechanical domain. When the loudspeaker is not mounted
in any kind of box (free air configuration), the term ZA can be neglected.
When representing the mechanical domain in an analogous electrical cir-
cuit, Rm acts as a resistor, Mm as an inductor and cm as a capacitor. The
force applied by the current running through the voice coil is represented as a
voltage source, and the velocity it generates in the diaphragm is represented
as current flowing through the circuit. This way, the analogous electrical
circuit can be represented as in figure 2.3.

Figure 2.3: Mechanical model of the loudspeaker. Source:[3]

Again, applying Kirchoff ’s voltage law, the transfer function is as in

equation (2.2).
1
Bli(t) = Rm v(t) + jωMm v(t) + v(t) + ZA (2.2)
jωCm
As explained before, the term ZA can be neglected when the loudspeaker
is in free air configuration. In this project, all the measurements and sim-
ulations have been done using this configuration, so from now on, the term
ZA will be neglected.

2.2.3 Acoustical Domain

[6] The most important acoustic parameters of a loudspeaker system are
those that are determined by the enclosure of the speaker. This is why, when
working in the free air configuration, no model is needed for the Acoustical
Domain.
The acoustic pressure level at a distance r from the loudspeaker is defined
as in equation (2.3).

6
 ρ
p(r) = sUd (s) (2.3)
2πr
Ud (s) is the displaced air volume, and is calculated as Ud (s) = ud (s)SD ,
where ud (s) is the velocity of the diaphragm calculated in the mechanical
domain, and SD is the effective surface area of the loudspeaker driver, usually
provided in the data sheet. ρ is the density of air.
The previous equation would give as a result the acoustic pressure in a
pressure unit, such as Pa. However, when working with sound, the pressure
level is often expressed in dB, due to the fact that this is more ’realistic’, in
the sense that it adjusts better to the human hearing. The sound pressure
level in dB (SPL) is calculated as shown in equation (2.4).
 
p(r)
SP L(dB) = 20log (2.4)
p0
where p0 is the reference sound pressure level. Usually p0 = 20µP a is
taken, as this value corresponds to the human hearing threshold (lowest
pressure value that a human can hear).

2.3 Electro-Mechanical Conversion and Final Linear

Model
Integrating the two circuits explained before into a single one, the loudspeaker
can be modelled as it is shown in figure 2.4.

Figure 2.4: Electro-Mechanical equivalent circuit of the loudspeaker.

Source:[3]

The force factor (Bl) is the parameter for the conversion of energy from
the electrical domain to the mechanical domain: the current flowing through
the voice coil induces a velocity in the diaphragm in the mechanical domain.
Equations (2.1) and (2.2) can also be written in the Laplace domain as
represented in equations (2.5) and (2.6).

7
 u(s) = (Re + Le s) i(t) + Blv(t) (2.5)
 
1
Bli(s) = Rm + Mm s + v(t) (2.6)
Cm s
In this project, all the simulations have been done using the software
Matlab. To integrate the previous model in Matlab, the third party software
Simulink has been used inside Matlab. Following equations (2.5) and (2.6),
the Simulink linear model of the loudspeaker used for this project is shown
in figure 2.5.

Figure 2.5: Linear model of the loudspeaker in Simulink

[4] However, sometimes it is convenient to have all the parameters of

the linear model as electrical components. For example, when using a com-
puter software implementing a SPICE simulator of electronic circuits (in this
project LTspice has been used). The parameters of the mechanical domain
can be transferred to the electrical domain, and the result would be equation
(2.7) and figure 2.6.
!
1
u(s) = Re + Le s + 1 1 i(s) (2.7)
Rel
+ Lel s
+ C el s
Rel , Lel and Cel are the electrical equivalents of Rm , Cm and Mm respec-
tively, and can be calculated as explained in equations (2.8) to (2.10).

(Bl)2
Rel = (2.8)
Rm

Lel = (Bl)2 Cm (2.9)

8
 Mm
Cel = (2.10)
(Bl)2

Figure 2.6: Linear model of the loudspeaker used in LTspice

2.4 Non-Liniarities
[2] The previously described linear model of the loudspeaker is valid for small
sound pressure levels, where the diaphragm displacement is short. However,
when the sound level is higher and the displacement of the diaphragm is
longer, non-linearities appear in the model: several loudspeaker parameters
are dependent on the displacement, so their value will change with the po-
sition of the diaphragm. The most important non-linear parameters are the
force factor Bl, the electrical inductance Le and the mechanical compliance
Cm .
In the case of Bl, when the diaphragm moves long distances (some mm) in
and out from its center position, the l parameter (effective length of the voice
coil) changes, so Bl changes depending on the position x(t) of the diaphragm.
In the case of Le , the inductance value changes with the position of the voice
coil as it moves in and out. Finally the mechanical compliance values change
when the voice coil is driven far away from its center position.
Figures 2.7 to 2.9 show the non-linearities of this parameters that depend
on the displacement.

9
 Figure 2.7: force factor (Bl) vs diaphragm displacement (x). Source:[3]

Figure 2.8: voice coil inductance (Le) vs diaphragm displacement (x).

Source:[3]

10
Figure 2.9: mechanical suspension compliance (Cm) vs diaphragm displace-
ment (x). Source:[3]

2.5 Final Non-Linear Model

For the non-linear model, the non-linearities of Bl, Le and Cm have to be
introduced in the model explained before. In this project, the Klippel analysis
software has been used to determine the coefficients for the equations of
these three parameters as a function of the position of the piston. The
resulting non-linear model is shown in figure 2.10. The main difference with
the linear model is that Bl, Cm and Le are now calculated as polynomials of
the position, using the coefficients of the Klippel analyser.

11
Figure 2.10: Non-linear model of the loudspeaker in Simulink

12
3 Loudspeaker Power Measurement Circuit
3.1 Audio Power, RMS and Peak Power
[7] Audio power is the electrical power transferred from the audio amplifier
to the loudspeaker. Although it is usually measured in Watts, when dealing
with AC signals, that unit corresponds to the active power. In this project
the apparent power is measured, so the unit used is VAR. The electrical power
delivered to the loudspeaker determines the sound pressure level generated
by this one.
The power of a loudspeaker can be calculated if the voltage in the termi-
nals of the loudspeaker and the current passing through it are known. The
average power of an audio signal played in a loudspeaker can be calculated
as in equation (3.1).

1 T
Z
Pavg = v(t)i(t) dt (3.1)
T 0
In this project, both the measurements and the simulations have been
done using discrete signals, with a sampling frequency of 44.1kHz, which
is the typical sampling frequency used in audio. In most cases, instead of
calculating the average power in the way shown above, the instantaneous
power has been calculated, which corresponds to the apparent power of each
sample, as in equation (3.2).

Pi = Vi Ii (3.2)
With the power values of each sample, the rms Power and the Peak
Power can be calculated. For n values {P1 , P2 , ..., Pn }, these parameters are
calculated as in equations (3.3) and (3.4).
v
u n
u1 X
Prms = t P2 (3.3)
n i=1 i

Ppeak = max{P1 , P2 , ..., Pn } (3.4)

These two parameters are very important when dealing with audio sig-
nals, because these signals have a wide amplitude range (large variation of
rms values over time). This means that the rms value and the peak value
will be very different depending on the amount of samples taken. Varying
the window size (amount of samples taken) and the position of the window
through the audio file can give extremely different results, due to the fact

13
that audio signals typically have a very high crest factor, which is calculated
as the ratio of the peak power of the audio signal to the rms value of it.

Figure 3.1: Typical waveform of an audio file, showing high peak power and
low rms values. Source:[7]

3.2 Sampling Theorem and Aliasing

[8] The Nyqvist-Shannon Sampling Theorem could be considered as a ’limi-
tation’ when working with continuous signals in the digital domain, which is
what happens in digital audio. The sampling theorem states that the sam-
pling frequency must be, at least, the double of the highest frequency of the
signal that is being sampled.

fs ≥ 2fmax (3.5)
Figure 3.2 shows the different resulting waveforms depending on whether
the sampling theorem has been accomplished or not.

Figure 3.2: Correctly and incorrectly sampled signals. Source:[8]

14
 The sampling theorem is also applied to audio signals: the human hearing
frequency range goes from 20Hz to 20kHz, which is the bandwidth of the
audio signals. The sampling frequency must be, then, over 40kHz. This is
why audio signals are often sampled at 44.1kHz or above.
[9] Aliasing is the effect that causes a signal to become indistinguish-
able when it is sampled without the sampling theorem being accomplished.
Aliasing can also happen if the sampling theorem is accomplished: sampling
generates low-frequency aliases, but their amplitude levels are very low, so
they do not mean a problem. But if the signal has frequency components
higher than fs /2 (sampling theorem not accomplished), these frequencies will
be mirrored back into the bandwidth of the sampling frequency. In this case,
there is wrong information about the measured signal.

Figure 3.3: Aliasing: frequencies being mirrored due to wrong sampling.

Source:[8]

Aliasing is avoided using low-pass filters. The signal is filtered with a

cutoff frequency (fc ) below fs /2, and this way higher frequency components
are eliminated. However, analog filters are not ideal and they still let some
signal pass through even if its frequency is higher than fc , but the amplitude
is reduced. The amount of reduction applied to higher frequencies is deter-
mined by the order of the filter. Each increment of the order has a roll-off
of 20dB/decade. So a higher order filter will be more effective to reduce
frequency components over fs .
In this project, the sampling frequency used was 44.1kHz, so the sampling
theorem was accomplished. However, anti-aliasing low-pass filters were used,
and there are two main reasons for this:
The first one is that the signals may contain frequency components over
the audio bandwidth (20Hz to 20kHz), such as noise and other sources.
These components might be mirrored back to the signal frequency spectrum
due to sampling, so they need to be cut off.
The second reason is something that can be understood more easily if the
next sections of this paper are read. Basically, there are two sampling steps

15
 Figure 3.4: Typical frequency response of a first order low-pass filter

in this project: the first one, when the signal is converted from digital to
the analog domain, to send it to the amplifier and then to the loudspeaker.
In this first step, there is no problem with aliasing, as the signal’s original
bandwidth remains unaffected. However, images higher than the original
bandwidth will appear due to sampling. The second stage of sampling is
when the current and voltage signals of the loudspeaker are sent back to the
computer. These two signals have the mirrored frequency components due
to the first stage of sampling, and if no filter is used, these components will
be mirrored in the original signal’s frequency bandwidth (audio frequency
range), affecting it wrongly.

3.3 Power Measurement Circuit: Main Components

In this section the circuit built for this project will be explained shortly, and
the main components and their functionality will be explained.
As explained before, a circuit has been built to measure the power of the
loudspeaker when it is playing a certain audio file. For this, the signal is sent
from the computer using a DAC and is first low-pass filtered, before it goes in
the audio amplifier. After that, the signal goes into the loudspeaker, which is
in series with a current-measuring resistor. A dual instrumentation amplifier
is used to measure the voltage in the terminals of the loudspeaker and the
current passing through it. The outputs of the instrumentation amplifier are
low-pass filtered again and finally stored in the computer via an ADC.

16
 The signal flow of this circuit is shown in figure 3.5. Figure 3.6 shows how
a small resistor is placed in series with the loudspeaker (which in this case is
simplified as an 8Ω resistor) to measure the current. The voltage measured
in the terminals of this resisistor is proportional to the current through it
(Ohm’s Law ). Finally, figure 3.7 shows a picture of the measurement circuit,
showing its most important parts.

Figure 3.5: Block diagram of the whole power measurement circuit

Figure 3.6: Current measurement resistor in series with the loudspeaker

17
 Figure 3.7: Picture of the power measurement circuit

3.3.1 The Loudspeaker: Monacor SPH-170TC

The loudspeaker chosen for this project is the SPH-170 TC, by Monacor. It
is a dual 6.500 Woofer (only one channel has been used) with a power handling
of 40Wrms and 60Wmax . It has an 8Ω impedance. Its resonant frequency is
at 33Hz and its frequency bandwidth goes up to 5000Hz.

3.3.2 The Amplifier: Texas Instruments TPA3116D2

The TPA3116D2, by Texas Instruments, is a Class-D stereo amplifier. Its
maximum output power is 50W , and it needs a 24V single power supply. It
is important to know that this is a full-bridge amplifier, which means that
its output is floating. This fact affects other components.

18
3.3.3 The 10mΩ Current Measurement Resistor
A 10mΩ current measuring resistor, placed in series with the loudspeaker, has
been used to measure the current going through it. This is done measuring
the voltage at the terminals of the resistor, which is directly proportional to
the current.

3.3.4 The Instrumentation Amplifier: Texas Instruments INA2128

The INA2128 dual instrumentation amplifier has been used to get the voltage
and current signals of the loudspeaker and send them to the computer. Each
one of the channels is designed with three operational amplifiers, and has
individual gain control using only one external resistor. The gain can be set
from 1 to 1000. This instrumentation amplifier is supplied at ±12V , and
voltage regulators have been used to avoid any unexpected change in the
supply voltage.
The reason of using this component is because both the voltage and cur-
rent signals are floating, and their amplitude is too high for the ADC. So the
functionality of the instrumentation amplifier is to make the signals ground-
referenced and have the proper amplitude to be sent to the ADC.
The gain of each channel is set by the following formula: G = 1 + 50kRG
. To
set this value, it must be known that the input voltage range for the ADC
is ±10V , so the outputs of the instrumentation amplifier must be between
these limits. For the current channel, the voltage in the terminals of the
10mΩ resistor is always smaller than 50mV , so it needs a high gain. A
resistor of 270Ω is used, which, according to the formula before, gives a
gain of around 186. For the voltage channel, however, the signal should be
reduced, because it can go up to 20V in the output of the amplifier. For this,
a voltage divider is placed before the instrumentation amplifier, and then the
signal is amplified again. The gain is set to be 4.125 using a resistor of 16kΩ.

3.3.5 Operational Amplifiers for Filter design: MC33079

As explained before, when converting audio signals from the digital to the
analog domain and vice versa, aliasing problems might appear due to im-
proper sampling. This is exactly the case of this project, and this is why
anti-aliasing filters need to be included.
In the power measurement circuit, aliasing can appear due to two elements
mainly: the first one is the DAC/ADC: the sampling frequency for this one
is 44.1kHz (typical for audio signals). The second one is the audio amplifier,
which has a switching frequency of 400kHz.

19
 To avoid this, two low-pass filters are used: one at the input of the am-
plifier and one (per channel) at the output of the instrumentation amplifier.
These two filters use the Sallen-Key topology for low pass filters [10]. The
software Filter Wizard, provided by Analog Devices [11] was also used as an
initial help for the design. Figure 3.8 shows the typical Sallen-Key configu-
ration 2nd order low-pass filter.

Figure 3.8: Sallen-Key configuration for a 2nd order low-pass filter

The main parameters of this filter are the cut-off frequency fc and the
quality factor Q. They are calculated as explained in equations (3.6) and
(3.7).
1
fc = √ (3.6)
2π R1 R2 C1 C2
√
R1 R2 C1 C2
Q= (3.7)
C2 (R1 + R2 )
For this project, the components chosen are:

• R1 = 7.5kΩ

• R1 = 130kΩ

• C1 = 680pF

• C2 = 56pF

This gives fc = 26.12kHz and Q = 0.791. fc is chosen to be around

25kHz to reduce the images due to the sampling frequency of the ADC
at 44.1kHz and the switching frequency of the audio amplifier at 400kHz,
but bearing in mind that all the audio signal, below 20kHz, must remain
unaffected.

20
 The operational amplifier chosen for the impementation of the filters is the
MC33079. It is a low noise quad operational amplifier with a high bandwidth,
up to 15M Hz, which makes it ideal for this application.
In the input of the amplifier, a 4th order filter is needed: to reach a 4th
order filter with the Sallen-Key topology, two stages as the one shown before
are needed, connected in series. For this, two channels of the MC33079 quad
operational amplifier were used. At the output of each one of the channels
of the instrumentation amplifier, another 4th order low-pass filter is needed,
so all the four channels of the operational amplifier are used (two per output
of the instrumentation amplifier).
As a result, the images due to the sampling frequency are reduced with a
8th order low-pass filter (two 4th order in series), and the images due to the
switching frequency of the amplifier are reduced with a 4th order filter. The
resulting frequency response of an 8th order filter is shown in figure 3.9.

Figure 3.9: Frequency response of the 8th order low-pass filter, using the
Sallen-Key configuration

3.4 Testing Anti-Aliasing Filters

After including the filters in the circuit, measurements are taken to check
if they are efficient to avoid aliasing. For this, the LeCroy Wavesurfer 10
oscilloscope is used. This oscilloscope has the option to do the Fourier Fre-
quency Transform (FFT) of an input signal, and it is used to see if there

21
are images near the two frequencies mentioned before due to aliasing. The
measurements confirmed that aliasing is happening (before the filters), as it
can be seen in the next figures, even if the amplitude of the mirrored signals
is very small.
The figure 3.10 shows the FFT of the output signal from the DAC. This
signal has low voltage because it is passed through the amplifier yet. In the
figure, images at the right side of 44.1kHz can be seen, which are a mirror
of the frequencies in the left, but with a decreased amplitude.

Figure 3.10: FFT of the unfiltered signal in the output of the DAC

After the 4th order low-pass filter applied at the output of the DAC, it can
be seen in figure 3.11 how the frequencies around 44.1kHz have disappeared.
This is very important because, even if the images have a small amplitude,
they could produce sampling when the signal is amplified and sent back to
the computer via the ADC, sampled at 44.1kHz.

Figure 3.11: FFT of the low-pass filtered signal in the output of the DAC

22
 The same explanation is applied for the 400kHz switching frequency of
the audio amplifier: figure 3.12 shows the images around 400kHz. The signal
corresponds to the output signal of the instrumentation amplifier channel
measuring the current through the loudspeaker.

Figure 3.12: FFT of the unfiltered signal in the output of the instrumentation
amplifier (current channel)

After the filter, it can be seen again in figure 3.13 how no images appear
around 400kHz so aliasing is avoided when the signal is sent to the ADC.

Figure 3.13: FFT of the low-pass filtered signal in the output of the instru-
mentation amplifier (current channel)

After seeing the previous images, the importance of the first filter is ob-
vious. However, it is not clear if the second filter is absolutely necessary due
to the fact that the images around 400kHz have a very small amplitude, and
after the filter there is still some very small amplitude noise which cannot be
eliminated by the filter.

23
3.5 ADC & Software: Matlab & National Instruments
USB-6356
The main system used in this project to send, receive and store audio signals
to and from the power measurement system is composed by the USB-6356
data acquisition toolbox, by National Instruments, and the software Matlab.

3.5.1 Matlab
Matlab is a numerical computing environment and high level programming
language. It is developed by Mathworks. It allows plotting of functions and
data, implementation of algorithms, creation of user interfaces, and interfac-
ing with other programs and software devices.
In this project Matlab is used with the previously mentioned data ac-
quisition device called USB-6356. Matlab includes some functions that let
it communicate with devices of National Instruments. A review of these
functions is done in the next section.
The other use of Matlab in this project is the realisation of simulations
using the loudspeaker model explained in the second chapter of this docu-
ment. For the simulations, the software Simulink is used, which is one of the
Matlab’s 3rd party software tools. Simulink is a data flow graphical program-
ming language tool. It includes a graphical block diagraming tool, which is
used here to implement the different loudspeaker models.

3.5.2 NI USB-6356
The USB-6356 by National Instruments is the data acquisition device used
at this project. It can be used for several purposes, but in this project it
is mainly used as an Digital-to-Analog and Analog-to-Digital Converter and
data storing element, always in communication with Matlab. Some of the
most important parameters of the device are:

• 8 Analog Inputs (floating or ground referenced, DC coupling)

• 2 Analog Outputs (ground referenced, DC coupling)

• Maximum Input/Output range: ±10V

• Input and output resolution: 16bits

• Maximum sampling frequency: 1.25M Hz

24
 The next few lines show some basic functions to use the National Instru-
ments devices on Matlab. In this simple example, a sinusoidal signal is sent
to the firs output of the device and is then sent back and stored in Matlab
via the first input.

%Initial example: 300 Hz sinusoidal output

%1st output is directly connected to 1st input

%Dectect and list the devices connected to the computer

d = daq.getDevices;
%Establish connection to the National Instruments device
s = daq.createSession('ni');

%Initialize first input and output. Voltage measurement

s.addAnalogInputChannel('Dev1',0,'Voltage');
s.addAnalogOutputChannel('Dev1',0,'Voltage');
%Set the sampling frequency for both input and output channels
fs = 44.1e3;
s.Rate = fs;
%Set the maximum voltage range
s.Channels(1).Range = [-10 10];

%create sinusoidal signal

t = 0:1/fs:10-1/fs; %Run for 10 seconds
signal = sin(2*pi*300*t);

%Output signal. The conversion starts and runs for the specified
%time
s.queueOutputData(signal');
%Get data from device input and store it in a Matlab variable
data = s.startForeground;

3.6 Audio Amplifier Model

The power measurement circuit has two main components: the audio am-
plifier and the loudspeaker. When sending an audio signal to the circuit,
this one is amplified and played in the loudspeaker. For the simulations, the
signal sent to the loudspeaker must be the same as in the real measurements,
so a model of the amplifier needs to be included in the simulation.
It is very important to check if the TPA3116D2 audio amplifier has a lin-
ear gain over the frequency. For this, a high precision audio analyser is used:
this device sends a frequency-swept signal to the amplifier and determines
if the output voltage of it is constant in amplitude over the frequency, or
instead, if the gain is different depending on the frequency. Three different

25
measurements have been done: first an 8Ω purely resistive load is connected
in the output of the amplifier. Secondly the loudspeaker is connected, and,
lastly, no load is connected. The results are shown in the next figure 3.14.

Figure 3.14: Gain of the audio amplifier over the frequency, depending on
the connected load

The figure shows that the gain of the amplifier is almost constant for an
purely resistive 8Ω load in the audio frequency range (20Hz to 20kHz). How-
ever, when the loudspeaker is connected, the results show a small resonance
peak around 60Hz. Around 500Hz the gain is the same as in the resistive
load, but at 1kHz it begins to fall, and finally raises exponentially when
passed 10kHz. The results without a load show a higher gain in general, but
this case is not so important for this project.
Although the amplifier is not linear, only the range from around 200Hz to
2kHz is important for this project, because this is more or less the frequencies
that a woofer loudspeaker will reproduce. In the mentioned range, the gain
is almost linear, even if it drops when it passes 1kHz. This means that in
the simulations the amplifier can be modelled using only a linear gain block
of 25.7dB.
To check this, measurements have been done in the power measurement
circuit: first a frequency-swept sinusoidal signal, from 20Hz to 4kHz is sent
to the loudspeaker and the output voltage of the amplifier is stored. Then
the same is done with a 8Ω resistor load, instead of the amplifier. This is
done for both a high level voltage signal and a lower level voltage signal. The
results of the comparison between the output voltage of each measurement
show that they are very similar and that the gain is very constant in the

26
 (a) Low frequency range (around 250Hz)

(b) Higher frequency range (around 1.7kHz)

Figure 3.15: Matlab figure showing the constant gain of the amplifier for a
high level signal: Resistive Load = Red , Loudspeaker Load = Blue. (for
more precision, only the peaks of the sinusoidal waveforms are shown)

interesting frequency range. All of this confirms that, in the simulations, the
amplifier can be modelled simply with a gain block. The difference errors in
both cases are under 1%.

27
 (a) Low frequency range (around 250Hz)

(b) Higher frequency range (around 1.7kHz)

Figure 3.16: Matlab figure showing the constant gain of the amplifier for a
lower level signal: Resistive Load = Red , Loudspeaker Load = Blue.

28
4 Loudspeaker System Measurements, Sim-
ulations & Comparison
After the power measurement circuit has been built and tested, measure-
ments are taken sending audio signals to the loudspeaker and measuring the
power consumption of the driver. The results of these measurements are then
compared to results of simulations using the linear and non-linear model of
the loudspeaker.

4.1 Preparing the Audio File

[12] Before playing any audio file in the loudspeaker, it needs to be taken
into account that it is a woofer. In a two-way loudspeaker system, the woofer
would be the part of the system reproducing the lower frequencies, and the
tweeter driver would reproduce the higher ones. In a three-way loudspeaker,
there would be a subwoofer, a woofer and a tweeter. This means that each
one of these drivers does not have to reproduce the entire audio frequency
range: instead, crossover filters are used, which divide the signal into the
ways of the sound system and filter them for the proper frequency range for
each driver. For example, as explained before, the audio signal of a woofer
would be band-pass filtered between 200Hz and 2kHz, more or less.
For this project, the same is applied: the audio file is band-pass filtered
before sending it to the loudspeaker system or the simulations. For this,
Simulink plug-in is used on Matlab: the wav audio file is loaded on a vari-
able and then sent to Simulink, where it is band-pass filtered and sent back
to the workspace, to be finally stored in a new wav file. The figure 4.1 shows
the Simulink diagram used. The commercial song used for most of the mea-
surements is ’Give Life Back To Music’, by ’Daft Punk ’, and it is chosen due
to being a quite ’standard’ song nowadays, that could easily be heard on the
media such as the radio, television or music websites.

Figure 4.1: Band-pass filter used for the audio file on Simulink

29
4.2 Power Measurements at two Sound Pressure Lev-
els
After the audio file has been filtered, the measurements are done. For all the
measurements (real and simulations) two sound pressure levels are used: the
low level corresponds to background listening level, where the listener can
easily have a normal conversation while listening to the background music,
for example. The high level corresponds to when the listener wants to focus
only on the music and wants to listen to it at a considerable volume, but not
enough to be painful. The voltage sent to the amplifier in the lower level
measurements is 10 times smaller (in linear scale) than the one in the higher
level measurements. The gain of the audio amplifier is set at the maximum
for all the simulations.

4.2.1 Circuit Measurements

First of all, real measurements are taken on the circuit: the filtered audio file
is cut and done only 20 seconds long, for time saving reasons. Then it is sent
to the power measurement circuit and the data from the instrumentation
amplifier measuring the current and the voltage through the loudspeaker is
sent back to the computer. After doing the necessary unit conversions, the
current and voltage are stored into variables.
Table 1 shows some important parameters of the current and voltage
signals stored for the two levels. As discussed before, it can be seen how the
peak values are much higher than the rms values (high crest factor ).

Level RMS Voltage Peak Voltage RMS Cur- Peak Current

(V) (V) rent (A) (A)

Low 0.1895 1.0664 0.0271 0.1461

High 1.8905 10.5369 0.2660 1.4240

Table 1: Data from the measurement of the loudspeaker system when playing
an audio file

4.2.2 Simulations: Linear and Non-Linear Loudspeaker Models

Simulations have done with the same 20 seconds of the audio file using the
linear and non-linear models of the loudspeaker in Simulink and the current
and voltage through the loudspeaker model are extracted, to be compared
with the results of the real measurements. The peak values of current and

30
voltage are very similar to those of the real measurements, in both low and
high sound pressure levels. The comparison between the results and the
precision of them is discussed in the next section.

4.3 Comparison & Conclusions

There are several different ways to compare the results form the measure-
ments in the circuit and in the simulations with different models. In figures
4.2 to 4.5, the voltage and current curves of the three different methods of
measurement (real, linear model and non-linear model) are compared di-
rectly. In order to see the small differences with higher detail, only a small
area is shown for these figures: from t = 6.954s to t = 6.964s.

Figure 4.2: Comparison of the voltage between the real measurements (green)
and the simulation on the linear model (blue) and non-linear model (red) for
low level sound pressure level. The red line cannot be seen because it is
exactly the same as the blue one.

31
Figure 4.3: Comparison of the current between the real measurements (green)
and the simulation on the linear model (blue) and non-linear model (red) for
low level sound pressure level

Figure 4.4: Comparison of the voltage between the real measurements (green)
and the simulation on the linear model (blue) and non-linear model (red) for
high level sound pressure level. The red line cannot be seen because it is
exactly the same as the blue one.

32
Figure 4.5: Comparison of the current between the real measurements (green)
and the simulation on the linear model (blue) and non-linear model (red) for
high level sound pressure level

As it can be seen in the figures, the results from the different types of
measurements are very satisfactory: the curves are very similar, both the
voltage and the current. They both follow the same waveform, although
they have slightly different values when peaks are reached. In general, the
current measurements seem to be less precise: this makes sense, because more
components have been used to measure the current through the loudspeaker
than those to the voltage. All the components used in the measurements
have small tolerances. However, all these tolerances have to be summed up:
this means that it is possible that the small differences appreciated in the
figures are introduced by the lack of precision of the components, rather than
by the linear and non-linear models.
If the voltage and current waveforms are multiplied sample by sample,
the apparent power can be displayed and compared. Figure 4.6 shows the
apparent power curve compared for the 3 methods of measurement. It can
be appreciated how the differences between the real and simulated measure-
ments increase. This can be seen most clearly in the peak values, where the
peaks of the real measurement can sometimes be slightly higher than those
of the simulations. This errors appear because, already in the voltage and
current curves, there were some tiny errors, so when these two curves are
multiplied, the errors raise with a power of two.
It is difficult to explain exactly why these small errors appear mostly on

33
Figure 4.6: Comparison of the power curves of the real measurements (green)
and the simulation on the linear model (blue) and non-linear model (red)

the peaks. They could also appear due to a lack of precision of the Thiele-
Small parameters, because their values can change when they are heated up
due to high power. The best example of this is the change of the resistance
value, Re , which has not been included in the model. However, it can be seen
clearly how the three curves follow the same waveform continuously, which
ensures that both the linear and the non-linear model have a high precision
and can simulate the behaviour of a real loudspeaker with high fidelity.
Even if some conclusions can be taken when comparing the apparent
power curves, an algorithm is introduced in the next section, which is more
suitable for comparing the different power curves and taking conclusions
about loudspeaker power requirements.

34
5 Power Window Sweep
In chapter 4 some measurements have been done on the power measurement
circuit and then compared to results of simulations with the linear and non-
linear model of the loudspeaker. For this, the resulting current and voltage
curves have been compared directly, and then multiplied to compare the
power curve. Even if this can give some useful information about the power,
it is difficult to see the differences between the three methods of measure-
ment, so an algorithm must be implemented so as to show the differences
of the power requirement measurement more clearly in the three cases. The
algorithm will show to be very useful to see how, when the loudspeaker is
playing an audio file, commonly the power peaks are very high in comparison
with the rms value of the power.
What it is explained in the next section is included in the content of a
Convention Paper for the Audio Engineering Society (AES) called ’Require-
ments Specification for Amplifiers and Power Supplies in Active Loudspeak-
ers’ [1], written by a group of professors and PhD students of the Technical
University of Denmark.

5.1 Power Window Sweep Algorithm

[1] Let’s assume that the voltage and current signals of the loudspeaker are
currently properly stored. As the signals have been digitalised with a certain
sampling frequency, they are stored in a discrete way: each value corresponds
to a sample. The instantaneous power can, then, be calculated multiplying
the current and voltage values that correspond to the same sample. After
doing this, the values of power of every sample are stored and the rms values
can be calculated. In the case of n values {P1 , P2 , ..., Pn }:
v
u n
u1 X
Prms = t Pi2 (5.1)
n i=1

The way this algorithm is implemented is by changing the number of

samples taken to calculate the rms value of the power, which from now on
will be called the window size. Assuming that n is the window size and N is
the total length or the audio file (the total number of samples), all the possible
windows are analysed for a given window size: this is done calculating first
the rms power of a window that starts on the 1st sample and finishes on the
sample N − n. Then, the window starts with the 2nd sample, then with the
3rd and so on, until the last window starts on the sample N − n and finishes

35
 (a) Small window size (b) Larger window size

(c) Window size equal to the length of

the audio file

Figure 5.1: Window sweep over the power curve of the loudspeaker with
different window sizes. Source:[1]

on the sample N . The maximum value of all the rms values taken will be
stored as the maximum rms value for a window size n.
The final goal of this algorithm is to get the maximum rms value for
several different window sizes and compare them: first, a window size of only
one sample is taken. In this particular case, the rms value of the window
will be the exactly the value of the sample. The maximum value will also be
the highest power peak.
The window size is then increased and, thus, the rms values of the win-
dows will begin to be lower and lower. The last case is when the window size
is equal to the length of the audio file, and there will be only one window,
whose value is the rms value of the song.
The importance of the algorithm resides in comparing the maximum rms
values of different window sizes and plotting them in the same graph. This
will show something similar to what is shown in figure 5.2, where the max-
imum rms power is high for short window sizes, but goes down very fast
as the window size is increased. From a particular window size and on, the

36
maximum values will be almost the same.

Figure 5.2: Maximum rms values as a function of the window size. Source:[1]

This algorithm gives important information about the power consumption

of a loudspeaker when reproducing an audio file: it shows how the value of the
power is very high for very short amounts of time. However, when a longer
time window is taken, the rms value of the power decreases exponentially,
which means that the power peaks are very short and high compared to the
power curve’s rms value.

5.2 Comparing Results & Conclusions

The algorithm explained in 5.1 has been implemented on Matlab and used
to analyse the measurements taken on the real loudspeaker system and the
simulations with linear and non-linear models. To save time, the window
size has been increased logarithmically, and not linearly. This way there is
more information on smaller window sizes, where the maximum rms power
decreases exponentially. For these measurements, only 5 seconds of the power
curves have been analysed, using 1000 different window sizes.
The figures 5.3 to 5.6 show the results of applying the power window
sweep algorithm to both low and high sound pressure levels.

37
Figure 5.3: Power window sweep algorithm applied to high level measure-
ments. Real measurement (green), simulation on the linear model (blue) and
non-linear model (red).

Figure 5.4: Power window sweep algorithm applied to low level measure-
ments. Real measurement (green), simulation on the linear model (blue) and
non-linear model (red).

38
 The figures show satisfactory results, that confirm that the linear and
non-linear models of the loudspeaker are trustable models for the behaviour
of the real loudspeaker driver. Both figures show how the curve corresponding
to the non-linear model has higher peak values than the linear model. This
is due to the fact that non-linearities affect most on the highest power peaks,
when the position of the piston of the loudspeaker moves a considerable
distance from its center position, and this makes some of the Thiele-Small
parameters of the loudspeaker change. For larger window sizes, the curves
of the two models are very similar.
Both for low and high sound pressure levels it can be seen how the power
curve of the real measurements is very similar to the simulations for short
window sizes (It should be closer to the one of the non-linear model than
to the one of the linear model, but maybe this difference is due to errors of
precision of the components in the circuit). As the window size increases, the
real measurement shows to have higher maximum rms power values than the
simulations, but this difference is very small. This can be seen most clearly
when the window size is 102 samples. Finally, for very long window sizes,
real measurements and simulations give very similar results. All of this can
be seen more clearly in figures 5.5 and 5.6, where the difference of the power
curve of the real measurement and each simulation is directly compared. It
can be appreciated how the difference remains very low for both cases, again,
confirming the fidelity of both models.

39
Figure 5.5: Difference between the real measurements and each one of the
simulation model for high level (difference with linear model = green, differ-
ence with non-linear model = red)

Figure 5.6: Difference between the real measurements and each one of the
simulation model for low level (difference with linear model = green, differ-
ence with non-linear model = red)

40
6 Loudness Normalisation & Measurements
with Different Songs
In the previous section it was demonstrated how the linear and non-linear
models can be used to simulate the behaviour of the loudspeaker driver.
However, all the measurements before have been done using the same audio
file. The aim of this project is not to compare the power requirements of
different musical styles, but it is still important to check that the model
is working for different audio files, belonging to various musical styles and
with changing waveform characteristics. This is why the same simulations
as before have been run with six audio files of different musical styles.
However, the audio files cannot be directly compared due to the fact that
the music pieces belong to different styles: they are from different times and
recorded in different ways. This means that the loudness perceived when we
listen to them is not the same, and thus, they need to be loudness normalised.

6.1 Loudness Normalisation

[1] [13] [14] Nowadays sound is important in almost every communication
media: from songs on conventional loudspeakers to the television, the radio,
the mobile phones, electronic devices, and so on. However, sometimes the
audio files that all these devices reproduce have different loudness levels, and
the listener finds himself adjusting the sound volume continuously. To avoid
this, loudness normalisation needs to be applied to the audio files, so that
they are all perceived with identical loudness level by the listener.
With digital audio, all the songs are peak normalised: their peak value
is below 0dB (digitally, not in SPL), but this does not mean that all the
digital audio files have the same loudness level, because for example their
rms values can be very different.
To perform loudness normalisation, the loudness level is first measured
with a K-weighted filter curve, which is based on how humans perceive the
loudness level of sound. This filter consists of two filters: the first one simu-
lates the acoustic effect of the human head as a rigid sphere, while the second
one, a high-pass filter, refers to the low frequency b-curve used to simulate
the frequency sensitivity of the human hearing. The K-weighting method is
an open standard defined by The International Telecommunication Union:
ITU BS.1770. This filter is shown in figure 6.1.
After the K-weighting filter, the loudness is measured using three sim-
ilar terms: LKFS, LUFS and LU. LKFS is an abbreviation of Loudness
K-weighted Full Scale and one unit of LKFS is equal to 1dB. LUFS stands

41
Figure 6.1: Frequency response of the two filters that compose the K-filter.
Source:[1]

for Loudness Units Full Scale: despite having a different name, it is identical
to LKFS, but it is used in different standards. Finally, LU stands for Loud-
ness Unit and is a relative loudness measurement unit (the other two were
absolute measures).
The audio file, measured in LKFS, is split into 400ms blocks with an
overlapping of 75%. Gating filters are used to remove audio content that
should not be taken into account, such as quiet parts or very loud exceptional
noises. Two threshold values are used for this, an absolute one and a relative
one. In this project the absolute value used is −70dB and the relative one is
−10dB (compared to the average loudness of the song).
Finally, this algorithm gives a value of the perceived loudness in dB for
the audio file analysed. In order for two or more audio files to be loudness
normalised, this loudness value in dB needs to be the same for all of them, so
the lowest one is taken as a reference, and a ’reduction gain’ is be applied to
all the others to make them match the reference loudness value (if the refer-
ence value taken is higher than the one corresponding to the audio file with
the lowest loudness, when applying a gain, saturation would be introduced,
because the highest peak of the lowest one would be over 0dB).
In this project the algorithm explained above has been implemented us-
ing the loudness itu function in Matlab. This function follows the ITU-R
BS.1770-2 specification.

6.2 Simulations with Different Songs

The algorithm explained above has been applied to six songs of different
musical styles. Then the steps explained in the previous chapters have been

42
applied and the power requirements of the loudspeaker driver reproducing
the six songs has been analysed. Simulations have been compared to real
measurents.

6.2.1 Preparation of the Songs

Six songs have been chosen for this part of the project. The style of these
audio pieces ranges from classical music to dubstep, which means that they
are from very different seasons and their loudness values will be very different.
The songs chosen are:

• 1 - Waking On A Pretty Daze - Kurt Vile

• 2 - Wilkie - Roman Flgel
• 3 - Creature Fear - Bon Iver
• 4 - Up Down Up Down - Koreless
• 5 - Bangarang feat. Sirah - Skrillex
• 6 - Requiem for soloists, chorus and orchestra, K. 626: Requiem aeter-
nam - Helmuth Rilling , Mozart

These songs have been loudness normalised as explained in section 6.1,

using the loudness itu function on Matlab. All of them have been matched to
the loudness level of the 6th song, which showed to be the one with the lowest
loudness level. After that, all the songs have been filtered to be suitable for
a 6.500 woofer driver, as explained in section 4.1.

6.2.2 Measurements
The steps followed to perform the measurements in the circuit and the sim-
ulations of the loudspeaker playing these audio files are the same as the ones
explained in section 4.2. In this case, the voltage and current values of each
sample taken from the real simulations, the linear and the non-linear models
of the loudspeaker have been directly multiplied, and the comparison of the
apparent power consumption of the loudspeaker for two of the songs is shown
in figures 6.2 and 6.3. The reason for showing only the results of songs 1 and
4 is because they are the most ’extreme’ cases, with the highest and lowest
peak values. The measurements have been done again in 2 different sound
pressure levels, but the results shown here correspond only to the high level.
Even if they belong to very different musical styles, it can be clearly
appreciated how the waveforms of the real measurements and the simulations

43
still match almost perfectly. Again, some small differences can be found in the
highest peaks, but the models behave almost exactly as the real loudspeaker.

Figure 6.2: Short period of song no. 1 showing the comparison of the power
consumption (for high level). Real measurement (green), simulation on the
linear model (blue) and non-linear model (red).

6.2.3 Power Window Sweep

The same power window sweep algorithm as the one explained in chapter 5
is applied here. In this case, only 4 seconds of each song are taken, for time
saving reasons, and the amount of window sizes taken for each audio piece is
1000. Figures 6.4 and 6.5 show the result of applying the algorithm to songs
1 and 4 (again, the most extreme cases). However, it needs to be taken into
account that only a small piece of the whole audio file is taken, so the peak
values do not correspond to the maximum peaks of the whole files, but only
of the parts taken.

44
Figure 6.3: Short period of song no. 4 showing the comparison of the power
consumption (for high level). Real measurement (green), simulation on the
linear model (blue) and non-linear model (red).

Figure 6.4: Song Number 1: Max RMS Power VS Window Size HIGH
LEVEL (Real Measurement = green , Simulation on Linear Model = blue ,
Simulation on Non-Linear Model = red)

It can be seen how the curves of the real measurements and the simula-

45
Figure 6.5: Song Number 4: Max RMS Power VS Window Size HIGH
LEVEL (Real Measurement = green , Simulation on Linear Model = blue ,
Simulation on Non-Linear Model = red)

tions match again, as expected. In this case, the real measurement shows
higher power consumption than the simulations, but with a very small differ-
ence. Comparing the results of the two simulation models, the non-linear one
shows higher peaks due to the displacement of the loudspeaker diaphragm.
Finally, figures 6.6 and 6.7 show the comparison of the algorithm applied
to all the songs, each one for one sound pressure level. The range of power
consumption is very different for each one of the songs. This can be due
to two reasons: first one, as loudness normalisation has been applied to all
the songs, it is possible that the maximum peaks of some of them have been
reduced. However, in this case, the rms values of all the songs should be
quite similar because they all have the same theoretical loudness, but this
is not true due to the second reason: only a small part of the audio file has
been considered (4 seconds), and it is unknown if this part corresponds to a
high or low level part of each song (a low level part could be the introduction
of the song, while a high level part would be the chorus, for example).
In order to display the data of each song more clearly, table2 shows the
rms power, peak power, and the ratio of the peak to the rms, which is the
crest factor, for each one of the songs. In this case, 30 seconds of the whole
audio piece are analysed, and not only 4 seconds, to give a much better idea
of these parameters in longer time. The large variation of the crest factor is

46
due to the different musical styles.

Figure 6.6: Power window sweep comparison for all the songs on high level.
Curves correspond to real measurements on the circuit (1-red, 2-blue, 3-green,
4-yellow, 5-magenta, 6-black).

47
Figure 6.7: Power window sweep comparison for all the songs on low level.
Curves correspond to real measurements on the circuit (1-red, 2-blue, 3-green,
4-yellow, 5-magenta, 6-black).

Song RMS Power Peak Power Crest Factor

(VAR) (VAR)
1 0.3968 6.0026 15.1279
2 0.3167 7.3907 23.3357
3 1.1519 18.7844 16.3080
4 0.1296 2.0323 15.6812
5 0.2732 3.6056 13.1965
6 0.2980 5.0153 16.8322

Table 2: Peak and rms power values and their ratio, the crest factor, for each
song

48
7 Statistical Analysis of Loudspeaker Power
Requirements
This section describes a work done in this project that differs a little bit
from the rest of the sections. For the statistical analysis of loudspeaker
power requirements, 128 loudspeakers with similar characteristics have been
taken, and their power consumption is compared.
This is done defining a target sound pressure level that every loudspeaker
should have. Then, the power consumption of each one of them is anal-
ysed. This gives an idea on the range of power consumption values of similar
loudspeakers, and their efficiency compared with the rest.

7.1 Thiele-Small Parameters Database

First of all, the Thiele-Small parameters of many loudspeakers are needed.
However, in order to make reasonable comparisons between them, they need
to be quite similar. To achieve this, only woofer drivers of 6.500 diameter have
been analysed. The Thiele-Small parameters have been taken from [15], and
an Excel file with 128 loudspeakers of 6.500 diameter has been created. The
parameters that are needed are:

• Sd : piston area of the diaphragm (cm2 )

• Re , Le : electrical resistance and inductance of the voice coil (Ω, mH)

• Mm , Cm , Rm : mechanical mass, suspension compliance and damping

factor of the diaphragm (g, N/m, N ∗ s/m)

• Bl : force factor (T ∗ m)

All these parameters are loaded on Matlab in order to define one trans-
fer function for each loudspeaker: the input to this transfer function is the
voltage applied to the loudspeaker, and the output is the sound pressure
level. So, this transfer function defines the behaviour and efficiency of the
loudspeaker driver: it determines what the sound pressure level is going to
be for a certain input voltage signal depending on its frequency and ampli-
tude. This can be much easier understood looking at figure 7.1, where the
sound pressure level of each loudspeaker is displayed for the audio frequency
spectrum in a bode diagram, from 20Hz up to 20kHz. The SPL is measured
theoretically at 1m from the loudspeaker, for a 2.83Vrms input.
Looking at figure 7.1 it can be seen how the frequency response and the
efficiency of each loudspeaker can change several dBs in comparison to others.

49
Figure 7.1: Bode diagram showing the SPL of each loudspeaker over the
frequency, theoretically measured at 1m distance with a 2.83V input voltage

This is why a ’target SPL curve’ response has been implemented, and a filter
has been designed for each loudspeaker in order to achieve this same ’target
SPL’ for all of them. This is explained in detail in the next section.

7.2 Target Sound Pressure Level & Filter design

To design the target sound pressure level curve, it has been taken into account
that all the loudspeakers being analysed are 6.500 woofer drivers. The typical
frequency range that this kind of loudspeaker reproduces is below 2kHz.
The target SPL curve has been designed as two filters in series: a 4th order
low-pass filter with a cut-off frequency of 1.5kHz and a 2nd order high-pass
filter with a cut-off at 60Hz. A gain of 90dB has also been applied, because
this is a typical SPL value for loudspeakers at 1m distance with 2.83Vrms
signal. Figure 7.2 shows the target SPL frequency response curve.
For each of the loudspeaker analysed before, one filter is designed: placing
this filter before the loudspeaker would make it have the same SPL response
over frequency as the target. Depending on the loudspeaker, this filter will
have a different curve, which can either apply or remove some gain for a
certain frequency. In figure 7.3, the filter designed for each loudspeaker can
be appreciated. It can be seen how in most of the cases the gain applied is
around 0, which means that the SPL of that loudspeaker and the target SPL
were quite similar before applying the filter.

50
 Figure 7.2: Bode diagram of the target SPL transfer function

Figure 7.3: Bode diagram of the filter designed for each loudspeaker

7.3 Simulation & Power Window Sweep

In order to get an interesting comparisons on how each loudspeaker behaves
and what efficiency it has, a simulation has been run using Simulink. The
same filtered song that was used in other simulations in chapters 4 and 5 has

51
been amplified and cut to 10 seconds This audio piece has been simulated
in each loudspeaker using their linear model. To get all the loudspeakers
to have the same theoretical sound pressure level, the filter designed in the
section before has been applied to the audio signal before sending it to the
loudspeaker. Figure 7.4 shows the Simulink file, where the filter is placed
just before the loudspeaker. Each loudspeaker has its own numerator and
denominator coefficients for the filter, calculated as explained in the previous
section.

Figure 7.4: Simulink model used, where the linear model is combined with
the filter of each loudspeaker

As it can be seen on figure 7.4, the current is extracted from each sim-
ulation. This current is then multiplied with the input voltage to the loud-
speaker in order to compare the apparent power of each loudspeaker: figure
7.5 shows a short period of the power curve of all the loudspeakers. It can
be appreciated how the power consumption of each one of them has very
different values, even if the sound pressure level will be the same thanks to
the filter.
In figure 7.5 it can be difficult to distinguish the different ranges of power
consumption for each loudspeaker. This is why the power window sweep
algorithm explained in Chapter 5 has been also applied to each one of these
curves of power. In this case, only 5 seconds of the audio power curve have
been analysed in the algorithm, and the number of different window sizes
taken is 500, in order for the simulation not to take too long (the algorithm
needs to be applied once for each loudspeaker, with a total of 128 times).
Figure 7.6 shows the result of the algorithm applied to all the loudspeak-
ers. It can be seen how from 104 samples window and on, the maximum rms
power remains constant. The figure shows how there is a quite wide range
of power consumption for all the loudspeakers: In the one with the highest

52
Figure 7.5: Power consumption of each loudspeaker for the same SPL. Only
a short period of the audio curve is shown

consumption, the peak power (that corresponds to the maximum rms for a
window size of one sample) goes over 12 VAR, while in others it is around
only 2 VAR. To compare the results with higher precessions, other functions
of Matlab have been used in the next section.

53
Figure 7.6: Result of power window sweep algorithm applied to 128 loud-
speakers

7.4 Comparison & Conclusions

Although it might not be easy to compare the power consumption of all the
128 loudspeakers from the figures in the previous section, some important
conclusions can be extracted. For this, the statistical function histfit of
Matlab has been used. This function has firstly been used to compare all
the peak apparent power values of the loudspeakers, which are equal to the
maximum rms power values with a window size of only one sample. Figure
7.7 shows how the mean value of the peak power consumption is around 6
VAR, but the range is very wide: from the ones with the highest efficiency
that have around 2 VAR peak power, to the less efficient ones, with a peak
power around 12 VAR.
As the window size is increased, the maximum rms value goes down
as expected, with the same form for all the loudspeakers, because all the
simulations are done using the same audio file. Figure 7.8 shows the results
of the maximum rms apparent power when the window size is 220500 samples
long (5 seconds of the audio file), which is the highest window size taken.
This value is very similar to the rms value of the whole audio file, and it has
a mean value around 0.3 VAR.
Finally, figure 7.9 shows the ratio of the peak apparent power to the rms
apparent power (crest factor) for each one of the 128 loudspeakers analysed.

54
Figure 7.7: histfit function used to display the distribution of the peak ap-
parent power values

Figure 7.8: histfit function used to display the distribution of the rms ap-
parent power values

It can be seen how, even if they have different power consumption ranges, the
ratios remain constant: all of them are between 16.5 and 18.5. This confirms
that the power consumption is proportional to the input audio file for all the
loudspeakers, whatever the power consumption is.

55
Figure 7.9: Ratio of the peak apparent power to the rms apparent power for
the 128 loudspeakers

56
8 Conclusion
The main objective of this thesis was to make measurements on a real loud-
speaker and compare the results obtained with those of simulations using
the linear and non-linear models of the loudspeaker, to investigate about
their power requirements. Different measurements have been done: first, the
power measurement circuit was built and tested, and measurements taken on
this circuit were compared to simulations, using the same audio file. After
that, six different audio files, belonging to various musical styles, were used
to perform the same comparisons. Finally, a database of over a hundred
loudspeakers was used to create models for each one of them and compare
their power requirements using simulations.
The first part of the project shows clearly how the linear and non-linear
models of the loudspeaker behave in a very similar way to the real loudspeaker
that is being modelled. The results achieved playing the same audio file
are very satisfactory, as it can be seen how the power curves belonging to
measurements on the real circuit, the linear model and the non-linear model
follow each other very closely, with some small differences when peaks are
reached. These small errors could also be due to the precision errors of the
components used in the circuit. The power window sweep algorithm has also
been applied to these measurements, and it can be clearly appreciated how
audio files usually have very short peaks, which are much larger than the rms
power of the whole audio piece (high crest factor ).
The comparison between the two simulation models shows how similar
they are: the only difference between the curves of power obtained with the
two models is that the non-linear model is slightly more precise when high
power peaks are reached, due to the fact that it takes into account how some
of the Thiele-Small parameters change when the diaphragm of the driver
has a high displacement. However, these are minor errors, and it has to
be taken into account that the non-linear coefficients need to be included
in this model, which are usually not provided in the loudspeaker datasheets
(they need to be obtained experimentally using the Klippel analyser or a
similar tool). This is why the linear model is clearly enough to simulate the
behaviour of the loudspeaker, and there is no need for a non-linear model,
unless very high precision is needed.
The results obtained from the measurements using six songs from different
musical styles show that the linear and non-linear models of the loudspeaker
behave almost exactly as the real loudspeaker, it does not matter what audio
file used is. Even if loudness normalisation has been applied to all the files to
make them sound equally loud, it can be seen how the crest factor is always
very high, although it changes for the different musical styles.

57
 Finally, the statistical analysis of loudspeaker power requirements shows
that the range of power consumption for the same sound pressure level re-
sponse can be very wide, even if the loudspeakers have similar characteristics,
because they have different efficiencies. This means that, when designing a
power supply for a certain kind of loudspeaker, it has to be taken into ac-
count that the power consumption may vary widely, so the worst case needs
to be taken into account, which would be the loudspeaker with the highest
power consumption, or the lowest efficiency.

58
9 Future Work
In this project an extensive investigation of loudspeaker power requirements
has been performed and important conclusions have been extracted. How-
ever, this does not mean that it is finished here: there is a great amount of
work that can be done following this project, from measurements of power
requirements on different loudspeakers and different configurations, to the
design of power supplies, amplifiers and loudspeakers that take into account
all the results of this work.
First of all, all the measurements and simulations in this project have been
done for woofer loudspeakers. The linear and non-linear models are only valid
for the low-frequency range. However, it would be very interesting to do a
similar investigation for tweeter loudspeaker drivers, using models that are
valid for the high frequency range. This way, the power requirements for the
entire audio spectrum would be known.
The measurements have been done using the free air configuration, where
the loudspeaker does not have any enclosure. Although this gives important
information, the power requirements might change when the loudspeaker is
inside an enclosure, such as in the closed box or ported box configurations. In
these cases it is important to take into account the acoustical domain of the
loudspeaker when creating its model. It would be interesting to investigate
and compare the power requirements of loudspeakers in these configurations,
and check that the models are also valid for them.
The final goal after so much investigation would be to design loudspeaker
systems with power supplies, amplifiers and loudspeaker drivers that take into
account the results obtained. Right now, loudspeaker system manufacturers
test their equipment using test-signals such as sinusoidal waveforms, so they
design for constant power consumption. But these signals are very different to
real audio signals, as in audio, typically, power requirements are very different
over time, because the crest factor is very high. This means that loudspeaker
systems that require a high maximum power but a low continuous power
should be designed, which would reduce the cost.

59
References
[1] Henrik Schneider, Lasse C. Jensen, Lars P. Petersen, Arnold Knott,
Michael A. E. Andersen, Requirements Specification for Amplifiers and
Power Supplies in Active Loudspeakers. Audio Engineering Society Con-
vention Paper. http://www.aes.org/e-lib/browse.cfm?elib=17536

[2] W. Marshall Leach Jr., Introduction To Electroacoustics & Audio Ampli-

fier Design. Kendall/Hunt Publishing, 3rd edition.

[3] Martin Bruun Werner, Measurement of Nonlinear Loudspeaker Parame-

ters. Technical University of Denmark.

[4] Knud Thorborg, Andrew D. Unruh and Christopher J. Struck, An Im-

proved Electrical Equivalent Circuit Model for Dynamic Moving Coil
Transducers. Audio Engineering Society Convention Paper. http://www.
aes.org/e-lib/browse.cfm?elib=14048

[5] Ruben Bjerregaard Berthold, Anders Normann Madsen, MATLAB Based

Measurement System for Loudspeaker Measurements. Technical Univer-
sity of Denmark.

[6] Rosalfonso Bortoni, Sidnei Noceti Filho, Rui Seara, Comparative Analy-
sis of Moving-Coil Loudspeakers Driven by Voltage and Current Sources.
Audio Engineering Society: Convention Paper 5910. http://www.aes.
org/e-lib/browse.cfm?elib=12454

[7] Wikipedia, Audio Power. http://en.wikipedia.org/wiki/Audio_

power.

[8] Wikipedia, Nyquist-Shannon Sampling Theorem. http://en.

wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampling_theorem

[9] Walt Hester, Oversampling Interpolating DACs. MT-017 Tutorial, ANA-

LOG DEVICES.

[10] Texas Instruments, Analysis of the Sallen-Key Architecture. Applica-

tion Report. http://www.vyssotski.ch/BasicsOfInstrumentation/
AnalysisOfTheSallen-KeyArchitecture.pdf

[11] Analog Devices, Filter Wizard. http://www.analog.com/

designtools/en/filterwizard/

a
[12] Ebay Guides, Woofer vs. Subwoofer: Is There Re-
ally a Difference? http://www.ebay.com/gds/
Woofer-vs-Subwoofer-Is-There-Really-a-Difference-/
10000000177630901/g.html

[13] TC Electronic, Loudness Explained. http://www.tcelectronic.com/

loudness/loudness-explained/

[14] Joseph Timoney, Thomas Lysaght, Marc Schoenwiesner, Implementing

Loudness Models in Matlab. Dept. of Elec. Engineering, DIT, Dublin,
Ireland.

[15] Loudspeaker Thiele-Small parameters database. http://

petoindominique.fr/

b
Appendices
A Matlab Files
In this appendix, the Matlab files related to each section of this document
are listed. These files are not necessary to understand this document, but
the list is provided in case the reader wants to see the work done with more
detail, or use some of the Matlab functions used during this project. The
name of each file starts with the section it belongs to, inside each chapter.
• Chapter 2 - Loudspeaker Modelling
– section3 Loudspeaker Linear Model.slx : linear model of
the loudspeaker in Simulink
– section5 Loudspeaker Non Linear Model.slx : non-linear
model of the loudspeaker in Simulink
• Chapter 3 - Loudspeaker Power Measurement Circuit
– section5 Sin Output NIdevice.m : simple example of the func-
tions of the NI USB-6356 data acquisition device.
– section6 Sinusoidals Loudspeaker Load.m : sinusoidal wave-
forms on the loudspeaker to check the linearity of the audio am-
plifier.
– section6 Sinusoidals Resistive Load.m : sinusoidal wave-
forms on a resistive load to check the linearity of the audio ampli-
fier.
– section6 Comparison.m : comparison of the linearity on the
resistive load and the loudspeaker.
• Chapter 4 - Loudspeaker System Measurements, Simulations
& Comparison
– section1 Applying filter to song.m : filter applied to the au-
dio file to make it suitable for a woofer loudspeaker.
– section2 Measurements Circuit 2 Levels.m : measurements
of current and voltage of the loudspeaker circuit playing the fil-
tered audio file.
– section3 Simulations Comparison.m : same measurements
on the linear and non-linear models using Simulink, and compar-
ison of results.

c
• Chapter 5 - Power Window Sweep

– section2 Applying Window.m : power window sweep algo-

rithm applied to the power curves obtained in the measurements
of chapter 4.

• Chapter 6 - Loudness Normalisation & Measurements with

Different Songs

– section2 1 Loudness Normalization Songs.m : loudness nor-

malisation of the six songs, using the loudness itu function.
– section2 2 Applying Filter Songs.m : filter applied to the
loudness normalised songs.
– section2 3 Circuit Measurements.m : measurements taken
on the loudspeaker circuit for the six songs.
– section2 4 Simulations And Comparison.m : measurements
with the linear and non-linear model for the six songs, and com-
parison of all the results.
– section2 5 Applying Window.m : power window sweep algo-
rithm applied to the power consumption curves of all the songs.
– section2 6 Normal Distribution.m : statistical analysis of
the peak and rms power values for each song.

• Chapter 7 - Statistical Analysis of Loudspeaker Power Re-

quirements

– section1 Target Filter.m : reading of all the Thiele-Small pa-

rameters from the loudspeakers database Excel file, design of tar-
get sound pressure level curve, and design for the filter for each
loudspeaker.
– section3 Simulations.m : simulations playing the filtered song
on each one of the loudspeaker models of the database, with their
filter.
– section3 Applying WIndow.m : power window sweep algo-
rithm applied to the power consumption curves of all the loud-
speakers.
– section4 Normal Distribution.m : statistical analysis of the
power characteristics of each loudspeaker.

d
B Other Files
In this appendix, other files related to some chapters of this document can
be found.

• Chapter 3 - Loudspeaker Power Measurement Circuit

– Datasheets folder : datasheets of all the components used in

the power measurement circuit
– Whole Circuit - LTSpice.pdf : schematic of the power mea-
surement circuit on LTSpice
– LoudspeakerLTSPICE.pdf : complex model of the loudspeaker
on LTSpice

• Chapter 7 - Statistical Analysis of Loudspeaker Power Re-

quirements

– 6.5 Woofer Data.xslx : database with the Thiele-Small pa-

rameters of the 128 loudspeakers analysed (the ones used are on
sheet 3)

e
www.elektro.dtu.dk  
Department  of  Electrical  Engineering  
 
Technical  University  of  Denmark  
Ørsteds  Plads  
Building  348  
DK-­‐2800  Kgs.  Lyngby  
Denmark  
Tel:   (+45)  45  25  38  00  
Fax:   (+45)  45  93  16  34  
Email:  [email protected]