Differential Analysis of Acoustical Smartphone Recording Capabilities - a Contribution towards Smartphone-modulated Perception of Tinnitus [original]

Faculty of

Engineering,

Computer Science

and Psychology

Institute of Databases

and Information Sys-

tems (DBIS)

Differential Analysis of Acoustical Smart-

phone Recording Capabilities - a Contribu-

tion towards Smartphone-modulated Per-

ception of Tinnitus

Bachelor Thesis at Ulm University

Presented by:

Johannes Vedder

johannes.v[email protected]

691733

Examiner:

Prof. Dr. Manfred Reichert

Advisor:

Robin Kraft, M.Sc.

2020

Last updated: October 13, 2020

Typesetting: PDF-L

EX 2ε

Abstract

Loud noise is a common risk factor for physical and mental health in our industrial-

ized world, which can trigger different sorts of health issues like permanent hearing

loss and tinnitus. To mitigate noise-induced problems in daily life, smartphones can

be used as an easy way to observe noise levels. As recording quality differs de-

pending on smartphone models and calibration techniques, standardized methods

are needed to acquire comparable results. To examine such possibilities in more

detail, several acoustical experiments were performed regarding the recording ca-

pabilities of in-build smartphone microphones compared to an external microphone

to figure out optimal smartphone recording conditions as this further increases mea-

surement accuracy. Additionally, various different calibration approaches differing

in effort and accuracy are evaluated. Results show that smartphones are capable

of measuring sound pressure levels accurately with only small deviations of about

±3 dB(A). Moreover, smartphone microphones are heavily frequency dependent,

which is why an approach was presented to normalize for these variations. Gath-

ered calibration data was further brought in conjunction with sound perception data

of tinnitus probands, to show an application in health issues. The presented meth-

ods provide a straightforward approach to measure sound levels with a smartphone

and compare them to other device conditions, opening the use of smartphones in

the modulation of sound perception in tinnitus and other conditions.

iii

Contents

Abstract iii

1 Introduction 2

1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 Purpose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.3 Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.4 Structure of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Fundamentals 6

2.1 Hearing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.2 Auditory Perception . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.3 Tinnitus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.4 Measuring Sound Levels . . . . . . . . . . . . . . . . . . . . . . . . 10

2.4.1 Decibel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.4.2 Sound Pressure Level . . . . . . . . . . . . . . . . . . . . . . 11

2.4.3 A-weighting . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.4.4 Sound Level Meter . . . . . . . . . . . . . . . . . . . . . . . 13

2.5 Smartphone Sound Processing . . . . . . . . . . . . . . . . . . . . . 15

2.5.1 Microphone . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.5.2 Hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.5.3 Operating System . . . . . . . . . . . . . . . . . . . . . . . . 17

2.6 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3 Concept 19

4 Experimental Design 22

4.1 Calibration Techniques . . . . . . . . . . . . . . . . . . . . . . . . . 22

4.1.1 Basic Calibration . . . . . . . . . . . . . . . . . . . . . . . . 23

Contents

4.1.2 Frequency-based Calibration . . . . . . . . . . . . . . . . . . 27

4.1.3 Internal Calibration . . . . . . . . . . . . . . . . . . . . . . . 30

4.2 Acoustical Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4.2.1 Noise Samples . . . . . . . . . . . . . . . . . . . . . . . . . 32

4.2.2 Recording Environment . . . . . . . . . . . . . . . . . . . . . 36

4.2.3 Technical Requirements . . . . . . . . . . . . . . . . . . . . . 37

4.2.4 Monodirectional Behavior . . . . . . . . . . . . . . . . . . . . 40

4.2.5 Omnidirectional Behavior . . . . . . . . . . . . . . . . . . . . 41

5 Results 43

5.1 Calibration Techniques . . . . . . . . . . . . . . . . . . . . . . . . . 44

5.1.1 Basic Calibration . . . . . . . . . . . . . . . . . . . . . . . . 45

5.1.2 Frequency-based Calibration . . . . . . . . . . . . . . . . . . 49

5.1.3 Analysis of already gathered data . . . . . . . . . . . . . . . 55

5.2 Acoustical Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

5.2.1 Monodirectional Behavior . . . . . . . . . . . . . . . . . . . . 60

5.2.2 Omnidirectional Behavior . . . . . . . . . . . . . . . . . . . . 66

6 Discussion 68

7 Conclusion 79

7.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

7.2 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

A Source Codes 82

Bibliography 103

Acronyms

ADC Analog-to-digital Converter.

AGC Automatic Gain Control.

API Application Programming Interface.

dB Decibel.

Hz Hertz.

NIHL Noise-induced Hearing Loss.

OS Operating System.

Pa Pascal.

PCM Pulse Code Modulation.

SLM Sound Level Meter.

SPL Sound Pressure Level.

STFT Short-time Fourier Transform.

TFFT Time-based fast Fourier Transformation.

ToH Threshold of Hearing.

1 Introduction

Loud noise in daily life can often cause reduced well-being, stress, sleep and con-

centration problems or in extreme situations even hearing problems, like hearing

loss or tinnitus. To quantify the amount of noise people are exposed during the day,

an easy and effortless way is needed.

Since measuring the loudness of noise is not an easy task, this can normally only

be fulfilled by appropriate hardware, which is capable of measuring sound levels

e.g. a Sound Level Meter (SLM). To measure sound levels properly, the device

needs to be calibrated beforehand, which can only be achieved by using another

very precise reference device. All in all, this is a very cumbersome task. Also,

sound measuring devices are not widely used, since they are expensive, large,

heavy and only usable for the task of measuring. Besides that, paying attention to

high noise levels, e.g. at a concert or a nightclub is not that common and socially

accepted. Nevertheless, a frequent exposure to high noise levels has permanent

consequences on hearing. To bring a change to this fact, the high prevalence of

smartphones can help, since they are equipped with a microphone and a permanent

companion. Therefore, adding the possibility to measure noise with a smartphone

could increase the awareness for high sound levels and help avoiding them. As a

result, this could mitigate the exposure time to high sound levels, thus decreasing

the risk for permanent hearing problems and provide a positive impact on general

health. Moreover, people with hearing disabilities like tinnitus have the possibility

to get acoustical information about their surroundings easily, which might help to

better understand their disease.

However, a lot of different smartphones with a varying equipment of hardware exist,

and at least for the largest operating mobile systems Android by Google and iOS by

Apple, it is not possible to measure the current sound level with a smartphone out of

the box. Rather than that, they all differentiate in the way of how they record sound

1 Introduction

levels since not only the build-in hardware is different, but also can other software

lead to different amplitude levels. Additionally, neither the manufacturers of the

operating system nor the smartphone producers, provide a calculation to convert

the measured amplitudes to a comparable unit. Thus, the question arises, if there

is a way to measure the current loudness with a smartphone in a comparable unit,

so that measurements can be compared independently of the used device.

To improve the current situation, three calibration approaches are compiled in the

following work, which address the problem of the heterogeneous hardware. Two

of the approaches are furthermore evaluated and assessed in terms of accuracy,

feasibility and convenience.

Additionally, to further review the measuring capabilities of smartphone microphones

under different conditions like distance and direction, the acoustical recording capa-

bilities of smartphones as well as the usage of an external microphone, are tested

and compared. This shall give information about the accuracy of sound measuring

results in certain situations and how reliable they are.

1.1 Motivation

To better analyze the impact of noise in daily life, it is necessary to measure infor-

mation about surrounding noise multiple times a day. This can help to get a better

understanding of the loudness for noise-induced mental health problems and hear-

ing disabilities, such as hearing loss and tinnitus.

Since the analysis of surrounding noise shall be available to as many people as

possible, the process of collecting noise levels should rather be uncomplicated to

provide an easy use. Therefore, it is not favorable to use any expensive or com-

plicated external hardware. Instead, using a smartphone-based application that

makes it easy to collect and save noise data for anyone seems to be a better ap-

proach.

Additionally, by offering the possibility to measure noise levels with a smartphone,

this opens an uncomplicated way to get quick information about the sound level

around oneself to assess the current noise situation. Observing high sound levels

in the surrounding with the smartphone, could probably convince a person to leave

1 Introduction

an area that could possibly lead to hearing damage. Thus, an application could start

to raise and increase social awareness for noise-induced health problems since its

easy availability.

1.2 Purpose

Gathering a lot of acoustical data for a longer period can not only increase personal

well-being and reduce the proneness for noise-induced disabilities, but also ben-

efit scientific medical research. Therefore, this work shall aid the research for the

TrackYourTinnitus1project, by providing a solution to gather acoustical data with a

crowd-based technique to assist the analysis of reasons for tinnitus.

Using the already well working TrackYourTinnitus application, which analyzes tin-

nitus fluctuation via a survey, it is possible to extend the built-in loudness measur-

ing function. For further analysis, calibration of measurements taken from various

smartphones allows a better understanding of how tinnitus is perceived in different

noise environments. All in all, this can support further tinnitus research.

1.3 Approach

The following work is based on the already implemented noise measuring function

of the TrackYourTinnitus application. Although, the function to record amplitudes

is already working, the retrieved data was not satisfactory, due to the hardware

differences and no standardized recording methods of the various smartphones.

Therefore, several calibration approaches were developed to adjust the measured

amplitudes to a common scale. For this purpose, already working approaches were

found and verified, however further options which might give the chance for more

precise results have been explored in addition to that.

Also, the idea to conduct further experiments regarding acoustical behavior of smart-

phones has come up, since the quality of the measuring results might change under

different conditions. This includes measuring amplitudes in decreasing distances

1https://www.trackyourtinnitus.org/

1 Introduction

of the sound source and analyzing the fluctuations and changes. Therefore, this

should help gaining knowledge about the fact whether the recorded values of smart-

phones do accurately reflect reality since decrease of sound level follows specific

physical laws. Additionally, the impact of measuring the sound level from different

directions has been considered as significant, to examine the possible omnidirec-

tional recording capabilities of smartphone microphones in more detail.

1.4 Structure of the Thesis

The thesis is structured in seven chapters. The following chapter deals with the

audiological fundamentals about hearing and the technical expertise of sound mea-

suring. The third chapter describes the general concept. Afterwards, in the fourth

chapter the experiments which are examined in the course of this thesis are de-

scribed. The fifth chapter presents the results of the experiments which are then

discussed in the sixth chapter. Finally, a conclusion with an outlook is given in the

last chapter.

2 Fundamentals

The following chapter explains the audiological fundamentals of hearing and the

perception of sound in the brain. Moreover, it is described how sound measuring

works in general as well as how it is performed with smartphones. Lastly, additional

reference to related studies is given.

2.1 Hearing

The human ear gives us the ability to sense sounds perceived via vibrations, which

are released by an audio source. The outer ear collects the waves in the air, which

then go through the ear canal until they reach the eardrum, which vibrates. Three

bones, malleus, incus and stapes then transfer the sound vibrations further to the

snail-shaped cochlea. As the cochlea is filled with fluid, the vibrations cause ripples,

which are sensed at the basilar membrane. As the cochlea is not shaped evenly,

sound waves of high frequency are sensed at the beginning part, which is narrow

and stiff. Lower frequencies are sensed at the more flexible and wider top of the

cochlea (apex). At the basilar membrane, the organ of Corti picks up the waves

by using about 20.000 small hair cells, which convert the sound waves into nerve

impulses. The electrical signal then travels via the auditory nerve to the cerebral

cortex of the brain, thus enabling us to hear [31]. Figure 2.1 shows this.

Being exposed to noise over a longer period may decrease the ability to hear. This

is referenced to as Noise-induced Hearing Loss (NIHL). NIHL can be caused by

damage to the hair cells in the cochlea or by damage to the synapses of the auditory

nerve [1]. High sound levels cause destruction to hair cells and degenerated hair

cells will not get replaced and therefore die, resulting in a permanent deterioration to

hearing. NIHL can be temporary by exposure to sounds of more than 80 dB(A), e.g.

2 Fundamentals

Incus

Malleus

Semicircular

Canals

Vestibular

Nerve

Cochlear

Nerve

Eustachian Tube

Tympanic

Membrane

External

Auditory Canal

Stapes

(attached to

oval window)

Cochlea

Round

Window

Auricle

Tympanic

Cavity

16 kHz

6 kHz

0.5 kHz

Figure 2.1: Overview of the auditory system [7]

by industrial noise, jet engine, visiting a nightclub. Usually hearing then recovers

within 24 hours. But reoccurring exposure and exposure over a longer time to high-

intensity sounds (> 120 dB(A)) will cause the hearing loss to be persistent [26].

2.2 Auditory Perception

Hearing a sound and understanding what it means does not mean the same thing.

While the pure task of hearing is performed solely by our ears, the actual process of

recognizing the heard sound and interpreting it is a complex task for our brain and

is called auditory perception [5].

The auditory perception can be divided into different process steps [8].

• Detection

A sound wave is within the human hearing range and can be heard.

• Discrimination

The ability to differentiate sounds from each other correctly.

2 Fundamentals

• Identification

A sound needs to be identified and assigned to where it comes from and what

the sound source is.

• Comprehension

The information of the sound needs to be understood, so that the perceived

message can be used.

Further studies, involving research and analysis of the functionality of the auditory

system and the auditory cortex during the last years, have shown that the path to

perceiving auditory information is quite complex.

Following the explanation of the auditory system in section 2.1, the acoustic stim-

uli which have been transferred to the brain, are divided into perceptual features.

This enables us to split a mixture of auditory sources into distinct auditory objects.

Therefore, we can differentiate the incoming sound stream and assign each sound

to its corresponding source [5]. The brain’s auditory cortex makes this possible by

analyzing the perceptual components of sound, such as pitch ("height" of the tone),

timbre ("tone color") and loudness [29]. By having defined the auditory object, it

can be further converted into an abstract level of perception, which then makes it

possible to include information from the brain to combine them with the perceived

object. Then, this abstract representation can be interpreted to make use of the

information [5].

However, the sound we hear might not be heard by other humans the same way

since the level of perception is not the same for everyone. Different intensity lev-

els and noise frequencies might have a stressful or more unwanted effect to some

humans more than others. Therefore, also the definition between sound and noise

can be of highly subjective nature [10]. Since the exact parameters that lead to

noise cannot be clearly defined, a research branch called psychoacoustics deals

with the “relationship between physical stimuli and the induced perception“ [42].

The most important parameters of psychoacoustics are loudness (critical ranges

of human hearing) and sharpness (proportion of high-frequency spectral compo-

nents to total loudness) as well as roughness, fluctuations and strength (time and

level differences in the sound signal, which modifies both amplitude and frequency)

[10]. There are several ways to measure each parameter, e.g. by calculating the

psychoacoustical perceived loudness with a method by E. Zwicker. Considering

2 Fundamentals

psychoacoustical parameters, it is evidently that there is a huge gap between the

way sound is measured and how it is perceived by the individual, as other factors

such as psychological experience, expectation and attitude have to be included into

a measurement which concerns human perception [10]. Therefore, it is not enough

to only include A-weighted levels when measuring noise since no adequate answer

can be given by that [10].

2.3 Tinnitus

Tinnitus, derived from the Latin word tinnire (to ring), describes the perception of

sound in absence of an external sound source [16]. Persons affected can sense

a lot of different sounds ranging from a typical ringing noise to hissing, whistling,

sounds analogous to animals like cicadas or crickets or complex sounds resem-

bling human voices or music [12] [3]. Data from the National Health and Nutrition

Examination Survey (NHANES) in the U.S. has shown that more than 50 million US

adults with about 15 % between 60 and 69 years old reported having experienced

any form of tinnitus [37].

Tinnitus can be classified into objective and subjective tinnitus. The former one

is the more seldom occurring form of tinnitus, with only 1% of all tinnitus reported

cases in the U.S. affected [2]. This is the case if another person can measure it.

The more often occurring form is subjective tinnitus, which is only perceived by

the person affected [3]. This makes it very hard for objective measurements, as

there are only certain ways to examine the severity of the case and ways for further

investigation and comparison rely on subjective impressions.

The main cause for tinnitus is hearing loss, which is often caused by reduced

cochlear nerve activity. Beside otological causes, there are numerous risk factors,

which include exposure to loud environments, head or neck injuries, acoustic shock

traumas and high blood pressure. Since there is no general rule for what causes

tinnitus initially, people with limited hearing capabilities can develop it as well as

people with normal hearing. Generally, the likelihood for tinnitus increases with age

up to 70 years [3].

Symptoms related with tinnitus are sleeping difficulties, emotional stress, physical

exhaustion and difficulties staying in too noisy or very quiet environments. Severe

2 Fundamentals

degrees of tinnitus can lead to a large impairment to health including working dis-

ability, anxiety and depression [12].

Possible treatments for tinnitus can either reduce the intensity of the tinnitus by

means of pharmacotherapy or try to weaken the disturbance of subjective percep-

tion with cognitive, sound or habituation therapy [12]. Also hearing aids proved to

be a supporting factor when tinnitus comes in combination with hearing loss. Not

only do hearing aids offer the possibility to increase the audibility of speech, they

can also support their wearer by amplifying only specific non-related tinnitus fre-

quencies which helps in decreasing the perception of these frequencies and aiming

towards better recognition of speech as well as taking away the stress associated

to them [12]. Other approaches using hearing aids include masking the tinnitus

by using a sound generator, which usually creates white noise or pleasant ambi-

ent sounds based on the specific type of tinnitus to reduce its loudness and thus

reducing the awareness [32].

2.4 Measuring Sound Levels

Converting sound intensity and sound pressure into a numerical unit can be a quite

confusing task since the logarithmic scale differs to typical measurement scales.

However, due to the way sound is sensed by the human ear, the logarithmic scale

better depicts changes in loudness. A short introduction on how sound levels are

measured and a connection to the human hearing are given in the following chapter.

2.4.1 Decibel

Sound travels in waves and consists of frequency and amplitude. The measurement

unit of sound is Decibel (dB) with a logarithmic scale. That is the case because

of the way our hearing works. The logarithmic scale starts from 0 and increases

in multiples of 10 dB. An increase of 10 dB will sound twice as loud to our ears.

However, an increase of 3 dB will already double the sound intensity. This can be

explained by looking into what loudness and sound intensity actually means. While

sound power is measured by the energy per united time which is emitted by a sound

source, the sound intensity is the sound power per area. The intensity measured

2 Fundamentals

in dB is defined as dB(I) := 10 ∗log10(I

I0)[27]. Loudness on the other hand is

subjective and depends on the ability of the listener’s hearing and characteristics of

the sound like duration and frequency. For example, sounds with the same intensity

but different frequencies do not necessarily need to have the same loudness. Figure

2.2 shows the coherence between intensity and loudness to our hearing. The unit

phone describes the perceived loudness of pure frequencies. The perception of

lower frequencies increases with rising intensity, as higher frequencies around 3

kHz can be perceived with even lower intensity [27].

Figure 2.2: Loudness is different for the perception of our hearing depending on the

frequency and the intensity [27].

2.4.2 Sound Pressure Level

To effectively use the decibel scale for measuring loudness or intensity, it is neces-

sary to use a reference unit, as decibel itself is a relative scale and does not have

a magnitude. To make use of the decibel unit regarding measuring sound levels,

the loudness can be measured in dB(SPL). The unit Sound Pressure Level (SPL)

is dependent on the atmospheric pressure and is measured in Pascal (Pa). It is

referenced as 20 µPa, which was assumed to be the Threshold of Hearing (ToH).

2 Fundamentals

This makes 1 Pa correspond to an SPL of 94 dB. An increase of 6 dB will double

the sound pressure [36]. A comparison between the sound level of different sounds

can be seen in figure 2.3. In the following work the term sound pressure level is

abbreviated to sound level.

Figure 2.3: Comparison of different decibel values and their effects on hearing [39].

2.4.3 A-weighting

The way of measuring sound can differ according to its use. In the case of measur-

ing for hearing-based applications, it is often measured using the A-weighted sound

pressure level dB(A).

This scale focuses more on the way humans hear. As seen in Figure 2.2 frequen-

cies are not perceived equally by the ear, as humans are only able to hear frequen-

cies between 16 Hertz (Hz) and 16 kHz with the highest sensitivity between 2 and

4 kHz. Due to the minor sensitivity for low frequencies, a consistent scale would

only provide limited meaning to noise measurements, as non-perceivable, lower

2 Fundamentals

frequencies would be given the same significance as higher frequencies which are

perceived more. Weighting helps to directly show the effects of loudness concern-

ing hearing damage [30]. The curve of the A-weighting can be seen in figure 2.4.

A-weighting comes in use for sound pressure levels below 55 dB. For higher levels,

B and C-weighting are chosen.

Figure 2.4: The A-weighting filter curve [36].

However, studies have shown, that the role of lower frequencies is much more im-

portant to physical and psychological factors than it is depicted with the A-weighting

sound pressure level. Although high-frequency hearing loss is much more common,

the effect of a vast amount of health factors for lower frequencies has been proven

in recent studies, making the general use of the A-weighted sound pressure level

questionable [30].

2.4.4 Sound Level Meter

Measuring the sound pressure with a decent degree of precision is achieved by us-

ing a SLM e.g. the one in figure 2.5. Basic devices have a range of +30 dB to +130

dB and can measure frequencies from 20 Hz to 8 kHz. Usable devices are low-cost

and can be used for simple measurements, however, they have quite a high amount

of tolerance range, which must be taken into account if an accurate measurement is

2 Fundamentals

needed. Dependent on the standardization class, devices have different tolerance

limits for different reference frequencies. In the most-used frequencies between 20

Hz and 10 kHz lower Class 2 devices have an error-rate between ±1.5 and 5.6 dB

[15]. For basic sound measurements though the difference between a Class 2 and

a Class 1 SLM is inconsiderable.

Figure 2.5: A sound level meter with a pop filter

The functionality of a SLM is achieved by using a microphone, amplifier, weighting

network, rectifier, and a display. In the first step, the signal of the sound wave is

converted into an electrical signal, which is then increased by using a preamplifier.

Afterwards, a weighting network alters the signal by filtering its frequencies to adapt

to a defined weighting standard. Most SLMs use the widespread A-weighting, which

is discussed in chapter 2.4.3. Following this, the signal is amplified once more

and converted from alternating current to direct current. Finally, the signal can be

2 Fundamentals

displayed [24]. This process is visualized in figure 2.6.

Microphone

Preampliﬁer

Weighting network

Ampliﬁer

Rectiﬁer Display

Figure 2.6: The functionality of a sound level meter, based on Kumar (2018)

2.5 Smartphone Sound Processing

As briefly touched in the introduction chapter, collecting acoustic data with a smart-

phone poses an ambitious challenge to its executor. To understand the difficulties

one has to overcome, it is important to get a brief knowledge about how the actual

functionality behind recording sound with a smartphone works and what stages a

signal has to pass until it can be measured by an application. Faber (2017) has

made some important research about the sound measurement possibilities and

limitations of smartphones, which are summarized briefly in the following sections

[9].

An acoustic signal can be recorded with either the build-in or an external micro-

phone, which can be connected via the audio jack or wireless. Following this, the

analog signal is transmitted through an Analog-to-digital Converter (ADC), which

creates a numerical representation out of the measured signal. Lastly, it can then

be delivered by the Operating System (OS) to requesting applications [9]. The pro-

cess is visualized in figure 2.7.

All mentioned steps will manipulate the incoming sound signal until it reaches the

application which measures it. It is important to state, that a smartphone was initially

not designed and built to provide proper sound measuring tools, as it is mainly intent

to be used as a communication device. Nevertheless, due to its vast amount of

functionality and tools, it can be used for other applications as well, like loudness

measuring. Most of the tools, which are build-in to modify the incoming signal are

designed to enhance the sound signal to understand someone better while making

a call for example. Since there was no need to get a clear amplitude signal, most

systems did not provide a method to disable these modifiers in the past or even do

2 Fundamentals

Microphone

(either internal or

external)

Acoustical Signal

(pascals, Pa)

Analog to Digital

Converter (ADC)

Analog signal Operating System

Digital representation

Digital Signal

(numeric)

Electrical Signal

(volts, V)

Figure 2.7: The process of converting an acoustical signal into a measurement

value that can be used by a smartphone application, based on Faber

(2017)

not provide it today. Moreover, the build-in hardware like the microphone does have

limited quality as it is designed to fit into the phone which only has a small outlet,

thus reducing its capability to record clear signals. This explains, why it is often not

that easy to get a good measurement accuracy.

2.5.1 Microphone

In the first step the microphone is a very import factor, when it comes to deviations

of precision.

In most cases, the build-in microphone is used for measurement purposes. If there

are multiple microphones available, normally only one, which provides the best om-

nidirectional features is selected. Given the position the microphone is located, it

can only record sound properly in a monodirectional way, even though its omnidi-

rectional design [9]. This changes the recorded signal strength depending on where

the signal originates, due to the acoustic shadow of the smartphone.

Better results have been achieved by using an external microphone. Possible op-

tions to connect can either be wireless (headset) microphones via Bluetooth. Other

non-wireless microphones are connected directly wired to the audio jack or data

port of the phone. Usually, wired measurements using the audio jack provide a

higher quality than wireless [9].

2 Fundamentals

2.5.2 Hardware

The sheer amount of different smartphone models are an important factor to con-

sider when obtaining sound measurements. Different build-in hardware parts such

as microphone or ADC, which are involved in signal processing, will manipulate

the quality of the outcome. Furthermore, smartphone manufacturers do not rely on

the same parts for all their smartphones, so the quality can differ for each smart-

phone model, too. Also, systems like Automatic Gain Control (AGC) can completely

corrupt the signal, as they are normally used to keep the outgoing signal at a con-

stant level independent of the incoming signal’s amplitude to enhance the quality of

phone calls [9].

2.5.3 Operating System

Just as the hardware does, the OS, which runs on the system, plays an important

role when it comes to the quality of the signal a phone can provide. As the OS

has direct control to the underlying hardware, it can often control to a certain part

the way the hardware operates. Whether this helps to improve the outgoing signal

does not only rely on the interface the hardware offers, but also on the Application

Programming Interface (API) the OS provides [9]. If one of both is not available, the

application on the top of the chain does not have any possibility to get an accurate

signal.

For example, Apple’s operating system (iOS) did constantly provide a high-pass

filter to all microphone inputs, due to their limitation to only partly record frequencies

below 200 Hz. Only iOS 6 in 2012 has added a function to disable this behavior and

the in-build AGC via the system’s API which now makes reliable measurements

possible [9].

2.6 Related Work

As of the current state of writing, performing acoustical measurements is a niche

topic since smartphones initially are not built for this. However, there is an increased

need to measure sound levels of noise. Since it is necessary to measure sound

2 Fundamentals

levels as reliable as possible, several smartphone applications are available, which

try to achieve a high degree of accuracy. There are a lot of studies which analyze the

performance of these applications like Kardous and Shaw (2014) or Faber (2017)

[17] [9]. Despite the increasing number of applications, only some of them provide

decent results. Since for most of the applications with good results the complete

functionality is unclear and most of them are subject to change, additional ways

are needed to make sound measuring possible for the average user. To further

investigate the acoustical behavior of smartphones in regard to measure sound in

changing environments as for directional or omnidirectional recordings, Hawley and

McClain (2018) provided valuable findings [13].

Moreover, this work provides a contribution to smartphone-based mobile crowd-

sensing studies which aim at gathering information about environmental data on a

large scale. Additional work in this field has been done by Kraft et al. (2020) who

presented an approach on how geospatial data could be helpful for concerned per-

sons and tinnitus patients to track noise levels in their surrounding area [22]. For

this purpose, an architecture was developed to process large amounts of data ef-

fectively and give the user a visual overview of loudness data that was gathered by

other users [21].

3 Concept

The work of this thesis builds up on the already well-established Track Your Tinnitus

application, which can be used to easily track tinnitus perception by using a smart-

phone. At the moment this involves observing the progress of the user’s tinnitus

measured by questionnaires throughout the day. This enables the affected persons

to measure fluctuations and working out daily habits and coherences, that have a

positive or negative impact on the perception of tinnitus [20].

To further extend this well-functioning system, improved noise measuring with a

smartphone will be introduced. Noise poses an important factor for broader tinnitus

evaluation, which will enable the system to gather more precise information about

the actual situation, hence using the noise measurements in combination with user

given data for extended evaluation. Recorded measurements e.g. environment

noise include data such as loudness and frequencies, which can give direct feed-

back to the user, thus supporting him to gain more knowledge of his specific tinnitus.

Consequently, including sound data for tinnitus studies allows analyzing the precise

characteristics and effects tinnitus has on the health of the affected person.

For uncomplicated sound measuring purposes the smartphone microphone should

be the primary way for measurements, as the aim for further studies is to have a

lot of participants who can join the study without the need for additional hardware.

However, due to better omnidirectional performance, an external microphone will be

tested and compared against the qualities of the internal microphone. This should

compensate for the different types of smartphones, the various analysis procedures

and the position of the smartphone while recording.

Recording with a smartphone seems like a trivial thing, since every conventional

smartphone is equipped with a microphone and a lot of apps make heavily use of

this technology to send voice messages or record videos with audio. Therefore, the

technological part is already established. However, measuring sound levels with

3 Concept

this equipment is a completely new application area for which neither the smart-

phone nor the hardware was designed for. Hence, a workaround by developing an

adequate applicable recording method and building an application around it, that al-

lows to take precise measurements is something completely new. Measuring noise

with a smartphone, which would enable almost everyone who has a smartphone

to do so, is only a niche sector at the moment and a standardized process to cope

with the hardware differences of the smartphones has yet to be developed.

Altogether, the following work will examine the characteristics of different smart-

phone types, and the use of internal and external microphones for sound measuring

purposes by using different frequencies and changing volume levels. Additionally,

the results of the measurements will aid to create a method, that ensures com-

parability between the recordings and to achieve necessary modulations to adjust

the measurement abilities of various smartphone types. This includes setting up

a standardized concept to guarantee objectivity on one hand, and simultaneously

working close to reality to reach a high degree of validity.

One way to achieve this for instance, is the positioning of the smartphone’s micro-

phone, as it is mostly located at the bottom of the phone. Most of the time, the

microphone is facing the user, which results being in the acoustic shadow without

acquiring a direct source of the signal, but an altered and mitigated version as seen

in figure 3.1. As a consequence, this does not only influence loudness, but also

implies recording a mutated frequency spectrum, which could both interfere with a

reliable result. Nonetheless, it has yet to be proven how much influence the limited

omnidirectional features of the internal microphone have.

Other important factors are effects of hardware and software differences of smart-

phones as already discussed in section 2.5. It has to be shown, if and to what extent

different microphones and varying processing hardware account for deviating data

outcomes. To create a whole spectrum of multiple factors, different frequencies

have to be tested together with multiple volume levels to find similarities and dis-

tinctions.

3 Concept

Figure 3.1: Most of the sound waves are reflected and mitigated before they are

recorded by the smartphone.

4 Experimental Design

Measuring amplitudes with smartphones poses a challenge to a lot of different

tasks. The first issue that needs to be handled are the various amount of smart-

phone models, which measure amplitudes on a separate scale. That is the case

why a calibration of all models is necessary. This calibration topic will be evalu-

ated with different approaches, as there are multiple possible methods to execute

a calibration. As part of the methods will be tested and evaluated in the next chap-

ter, the following text should give an introduction on the ideas, the functionality and

use-cases of the method. Additionally, experiments are being conducted, that an-

alyze the acoustical modalities of the involved smartphone microphones. This will

be done to give an overview of different aspects, that need to be considered when

amplitude measurements are performed by a smartphone. This includes acoustical

behavior tests, which are performed to evaluate how results vary for different situa-

tions. Since the sound environment varies during daily life, some frequencies may

be recorded differently than others. This is why varying sound samples are being

part of the evaluation and will also be compared against each other.

4.1 Calibration Techniques

As introduced earlier, smartphone models have different characteristics when it

comes to recording audio, which has an effect on the measured amplitudes. How-

ever, this mostly does not have an impact on the comprehension of the recorded

data (e.g. listening to a conversation that was recorded with a low amplitude can

be compensated by increasing the volume). On the contrary, when amplitudes are

measured this fact does play a huge role, as these amplitudes cannot be converted

directly to a standardized scale. Hence, a need for calibration is obviously neces-

sary.

4 Experimental Design

A calibration of a smartphone recording can be achieved in many possible ways. To

decide which calibration method should be chosen, it is necessary to assess the ob-

jectives of the project. Particularly, the required accuracy of the measurement and

the effort willing to apply in order to receive the desired degree of accuracy, should

be considered accurately. Evaluation of several ways for calibration has shown that

gaining a higher degree of accuracy is not possible without the use of external hard-

ware such as a calibrated SLM or a tone generator, which can either measure the

amplitude on a normalized scale or play a tone with a standardized volume. How-

ever, as this greatly increases the difficulty for measurements that only require a

rough estimation of the measured sound level, other means of calibration are re-

quired, that reduce the necessary work until a measurement can be performed.

Especially when crowd-sourcing shall be performed and many people with different

technical skills are put into place, it is inevitable to keep the process of calibration as

uncomplicated as possible (or even perform it on behalf of the user as seen later)

to eliminate any high deviation errors in calibration since this would decrease the

accuracy of the whole data set. Consequently, having diversified opportunities for

calibration depending on purpose of use has its justification.

In the following section two calibration methods will be covered, that have been

classified applicable for general use. For every method there will be a way on

how the calibration can be performed in detail. This includes not only a detailed

description of the functionality, but also associated assumptions, that will be verified

to evaluate the further use of the respective method. Additionally, a theoretical

method which could serve as a convenient method for future calibration uses will

be presented.

4.1.1 Basic Calibration

To perform a straightforward sound level measurement with a smartphone, no ex-

ternal hardware is necessary. Nevertheless, getting an estimate of the sound level

requires performing a calibration of the microphone. To keep this as simple as pos-

sible, Dr. M. Ziegler outlined an approach which involves an easy-to-use method

for a simple calibration in combination with the application "Spaichinger Schall-

analysator" 1. The calibration can either be performed by tearing apart sheets of

1https://spaichinger-schallpegelmesser.de/schallanalysator.html

4 Experimental Design

copy paper, or with a tone which has a consistently equal volume that must be

verified beforehand. Since the former method is usable without any additional hard-

ware, the main focus will be put on it. The following summary is taken from the

instruction document of the application [40].

Figure 4.1: The application "Spaichinger Schallanalysator" in measurement mode,

displaying sound pressure level, base frequency and corresponding

tone [40]

A screenshot of the user interface of the application is given in figure 4.1. The cur-

rent volume in dB(A) is given as well as the base frequency with its corresponding

tone. Additionally, the app offers more features like measuring the sound intensity

in W/m2, the effective sound pressure in Pa and much more.

A calibration with the paper tearing method consists of two steps. First, the smart-

phone or an external microphone is placed approximately 90 cm (which equals

three length of paper sheets) away of the person who will perform the calibration. It

is important to state, that the smartphone should be positioned at the same height

as the paper that is getting teared apart and that the direction of the smartphone

microphone is aimed at the person. The calibration can then be started by letting

the smartphone measure the amount of background noise so that a difference cal-

culation can be performed afterwards. Also, a sound level correction value can be

4 Experimental Design

given depending on where the calibration was performed. The second step of the

calibration process includes tearing apart 10 DIN-A4 sheets of white copy paper

(80 g/m2) one after another. The application measures the sound level of every rip

and calculates a standard deviation, which is about 2 dB afterwards. The smart-

phone is then ready to be used for basic sound level measurements. Figure 4.2

shows the setup that was established for the experiment.

Figure 4.2: Execution of the calibration by tearing apart paper

The method works by the fact, that the noise, which is created by tearing apart

paper at a specific distance away from the microphone, has the same loudness for

every performed paper tearing. Therefore, the recorded amplitude will serve as a

comparison value to calculate amplitude values to dB(A), since the noise of tearing

apart paper in dB(A) is known to the application. This enables the application to

calculate a compensation value for later measurements, based on the loudness of

the performed calibration.

Anyway, depending on where the calibration has been performed, a sound level

correction value can be applied at the time of calibration. This is due to the fact,

that the sound which is recorded by the microphone consists of primary sound from

the sound source, as well as sound that has been reflected by walls and furniture.

As diffused sound superimposes the direct sound of the loudspeaker, the correction

value needs to be increased for small rooms. An approximate value for different type

4 Experimental Design

of rooms would need to be gathered in further experiments.

Moreover, if an external microphone is used, the possibility is given to manually

configure the frequency response of the microphone. However, as this is already

covered in the next section 4.1.2, no further research will be done in this direction

with the application.

Another way to calibrate the microphone with the help of the application is offered

by using a tone generator, which creates a 1000 Hz tone between 75 and 95 dB

or by using an external sound source and a reference SLM, that can confirm the

correct decibel value. Since this is also part of the next chapter 4.1.2, it will not be

elaborated further at this place.

The paper tearing technique was then reconstructed and examined for its accuracy.

To do so, different frequencies were outputted on a Bluetooth loudspeaker and the

sound level in dB(A) was measured by the smartphone using the already mentioned

"Spaichinger Schallanalysator" application. Beyond that, an SLM was used next to

the smartphone to validate the measurements. Thus, both results can be analyzed

and compared. This approach is shown in figure 4.3.

Figure 4.3: Measuring the sound level of different frequencies after calibrating the

smartphone with the paper tearing technique

4 Experimental Design

This experiment will check the following hypotheses:

1. Calibration of the microphone will lead to reliable measurements

2. Different frequencies will not exceed a certain deviation level

3. The deviation level varies for different frequencies, so each frequency is

recorded with various sound levels

4.1.2 Frequency-based Calibration

Every sound consists of several frequencies, which are all measured with varying

intensity by a microphone. This fact is referred to as the frequency response of the

microphone. Each microphone has a defined frequency operating range, which de-

termines what frequencies can be recorded by it. However, this does not imply that

all frequencies in this range are recorded equally. This is why a frequency response

curve is necessary to specifically compare the capabilities of microphones in detail.

The frequency response is specified by testing the entire frequency spectrum of

the microphone and then measuring the amplitude of each tested frequency which

yields a microphone-specific frequency response curve. An example for a curve is

given in figure 4.4.

Figure 4.4: A frequency response curve of a microphone. The output varies de-

pending on the frequency. [6]

4 Experimental Design

Since smartphone models are not equipped with the same type of microphone,

sound level measuring with a smartphone microphone will undoubtedly face the

same issues, meaning they will certainly not only differ in the frequency ranges,

which can be sensed, but will also have differences in their frequency responses.

Aside from the deviations of the frequency responses, comparing loudness with

smartphones is additionally made complicated by the specific characteristics of the

smartphone microphone and the way the sound is processed by the system, which

is further discussed in chapter 2.5. This leads to the fact that measured ampli-

tude values cannot be converted to standardized decibel values without previous

calibration.

Nevertheless, by creating a frequency response curve, the deviations of every fre-

quency as well as the overall amplitude shift caused by the microphone and the

system characteristics can be represented together. Moreover, by comparing fre-

quency response curves of different smartphones with each other, an adjustment

can be made for each frequency. This makes it possible to measure how much

smartphone microphones deviate from each other independently of the hard- and

software that is being used.

A calibration is performed by calculating a specific difference value per frequency.

In other words, a specific frequency curve cfor smartphone Ashall be referenced to

as cA. This curve shall be compared to cB, the frequency curve of smartphone B.

Each frequency curve consists of multiple frequencies nthat shall be tested. For the

following experiment, nis defined as n∈ {500,1000,2000,4000,5000,7000,8000,

10000}. After the creation of the whole device curve, each frequency will have a

corresponding measured amplitude.

cA(n) : Amplitude of smartphone A for frequency n

Advantages of this method should be a good precision since the measurement

deviation of all frequencies are taken into account for calibration.

However, other than for the basic calibration in 4.1.1, additional hardware is nec-

essary to perform this type of calibration approach. Firstly, to play the different fre-

quencies it is necessary to have a loudspeaker which can play all the frequencies

that have been defined. Secondly, a calibrated reference device is necessary.

4 Experimental Design

Figure 4.5: Performing a frequency calibration by the help of a sound level meter

This includes either a SLM as shown in figure 4.5 to measure the sound level for

which the calibration should take place, or alternatively by using an already cali-

brated smartphone. This works by finding out the sound level with the help of the

already calibrated smartphone before the start of the calibration.

In both cases a smartphone or another device needs to be paired to the loudspeaker

to play all the defined frequencies, but the smartphone that should be calibrated can

also take over this role. The same process must be carried out for each smartphone,

that shall be calibrated.

Luckily, once a frequency response curve c(n)has been created for a smartphone

model, it can be shared with other smartphones of the same model as they are

supposed to have the same acoustics. In this case a cloud database could be

established to store the model specific c(n)values. For every new non-calibrated

smartphone the model identifier could be compared with the database and if avail-

able, automatically download the calibration values. Therefore, only one calibration

for each smartphone model would be required.

The idea is based on the company Eardial2, which already uses a similar approach

2https://eardial.com/calibration/

4 Experimental Design

to calibrate smartphones, however, without the focus on different frequencies.

Beyond testing for multiple frequencies, the sound level plays an important role. It

is necessary to calibrate all smartphones with the same defined sound level, so

the sound level of a starting frequency should be constant and verified before the

start of a calibration attempt. Therefore, an initial volume calibration at about 80

dB(A) for 1000 Hz must be made initially by using a SLM or an already calibrated

smartphone.

However, the extent of the sound level regarding the measured amplitude for dif-

ferent smartphones is unknown. To further analyze the effects of loudness on the

measurement accuracy, an Android application has been developed to test the fre-

quency responses of microphones at different volume levels. For this purpose, the

smartphone must be paired to a Bluetooth loudspeaker, which is placed at a 20 cm

distance away from the smartphone microphone. Then, a specially defined list of

frequencies is played one after another for 2 seconds each with a 1 second pause

afterwards to measure from silence again. Furthermore, to calculate the deviation

level for each frequency, several volume levels are analyzed, too. Therefore, the fre-

quency output starts at the lowest selectable volume level of the smartphone and

is then gradually increased until 15, which is the highest sound level for an Android

smartphone. During the output of the frequencies, the smartphone measures the

amplitude of the frequency that is played. The process of outputting and recording

the amplitude has been developed to run fully automatically. This should provide

knowledge about the impact of loudness on the overall amplitude shift.

The following hypotheses will be examined in combination with this experiment:

1. Varying frequencies have an impact on the recorded amplitude

2. Different smartphone models do not reflect varying frequencies evenly in am-

plitude.

3. Smartphones do not represent rising sound levels evenly in amplitude.

4.1.3 Internal Calibration

Another calibration approach is presented by performing a self-calibration with a

smartphone. However, as this method deals with the internal sound processing of

4 Experimental Design

the OS, it is unknown if it can practically be used at the time of writing, therefore it

is only described theoretically. Nevertheless, there might be a chance to develop a

calibration method out of this idea later since it should be very precise and would

not require additional user effort.

To get a practical use out of the previously described methods, a lot of

time-consuming work is necessary to be done by the user. Beyond that, addi-

tional items are necessary like the paper sheets for the method presented in 4.1.1,

a SLM or a calibrated smartphone shown in 4.1.2. This needs to be done to obtain

a reference volume where all executions of the experiments can refer to. However,

for a self-calibration this is not necessary, as it is only performed internally by the

smartphone itself.

To perform a calibration without relying on a monitored volume, the application

needs to have access to the digital sound input interface. This enables the ap-

plication to pass testing frequencies with a fixed volume directly to the digital input

processing of the system without using the loudspeaker and microphone. Previous

calibration methods always relied on outputting sound by using a loudspeaker or

creating it in other ways and then recording it via the microphone. This always in-

volves an analog to digital conversion, which is prone to measurement errors and

involves the use of an ADC, that alters the input unpredictably.

By the method of passing digital sound directly to a digital interface, all previously

necessary digital to analog conversion and a re-conversion into a digital signal

would thus be skipped. The schematic for this approach can be seen in figure 4.6.

Any sound leaving the operating system for calibration purposes would directly be

diverted to the digital input line, to avoid any conversation interferences. This pure

digital to digital interface would allow identifying differences in signal conversion and

internal processing in a more precise way and correct them.

4 Experimental Design

Microphone

(either internal or

external)

Analog to Digital

Converter (ADC)

Analog

Operating System

Digital

Digital to Analog

Converter (DAC)

Loudspeaker

Example for a conventional

calibration approach

Acoustical

Internal calibration

Reference sound from external source

(does not have to be a smartphone)

Figure 4.6: Schematic of a calibration approach with a pure digital-to-digital inter-

face

4.2 Acoustical Behavior

To achieve knowledge about the behavior of smartphone microphones, a variety

of experiments will be performed. An overview of each experiment is given in the

following section.

4.2.1 Noise Samples

To test the overall microphone recording performance, four different audio samples

were selected varying in frequency, tone and base. This is done to figure out if

and to what extend microphones record volumes for various frequencies differently.

Analyzing the audio samples is performed by utilizing the Time-based fast Fourier

Transformation (TFFT), which transforms the input signal of a time domain into a

frequency domain. This allows comparing the distinctions of the characteristics of

the audio samples. All samples are uncompressed and lossless WAV files, which

are shortened to a duration of about half a minute. All samples are free to use

and taken from the website soundeffects+ 3. Every audio sample will be shortly

evaluated below. The figures with the time-based fast fourier transformation of the

3https://www.soundeffectsplus.com/

4 Experimental Design

samples were made with the software WavePad4, the figures with the audio spec-

trum of the samples are created with the help of the software Audacity5.

1. Violin

The first sample is 30 seconds long and consists of a violin only, playing a

part of the Minuet from String Quintet in E Major, Op. 11, No. 5 by Luigi

Boccherini. As seen in figure 4.7a the sample consists of a tone pattern with

mostly small, high intensity tones ranging from 0 to 5000 Hz. Figure 4.7b

shows large fluctuations of the loudness.

(a) Time-based fast fourier transformation (TFFT) of a violin playing

(b) Audio spectrum of a violin in the relative dB unit

Figure 4.7: Characteristics of the violin sample

2. Street Traffic

The second sample also has a duration of 30 seconds and includes mostly

of street noise caused by traffic. Additionally, some chatter here and there

can be noticed. The TFFT graph in figure 4.8a shows a loud and stacked

frequency image for smaller frequencies up to 1000 Hz, which represent the

motor sounds. Higher frequencies in the range from 1000 to 4000 Hz are

4https://www.nch.com.au/wavepad/fft.html

5https://www.audacityteam.org/

4 Experimental Design

quieter and less dense. Analyzing the loudness of the traffic noise in figure

4.8b shows a changing volume, however this change is performed steadily

and no fluctuations in a short time can be observed.

(a) TFFT of street traffic

(b) Audio spectrum of street traffic in the relative unit

Figure 4.8: Characteristics of the traffic sample

3. Airliner Takeoff

The third sample is 26 seconds long and consists of a fixed microphone

recording heavy sound of an aircraft engine spooling up to takeoff thrust with

decreasing sound level as the aircraft accelerates away from the microphone.

The transformation in figure 4.9a shows a base-heavy and low frequency pat-

tern, which is measured continuously throughout the recording. Two increas-

ing lines can be seen, one starting from 1800 Hz increases up to 2900 Hz and

the other higher one starts at 3000 Hz and increases to 6000 Hz. Increas-

ing frequencies up to 10 kHz can be measured at about 7 seconds until 13

seconds, as the aircraft passes by the microphone. Afterwards the distance

to the aircraft increases, hence higher frequencies are declining. The loud-

ness of the airliner can be seen in figure 4.9b, it shows a gradually increasing

loudness until the turbines are fully ready, then some regularly occurring small

fluctuations, until the aircraft takes off. Thereafter, the sound decreases rather

fast.

4 Experimental Design

(a) TFFT of a large airliner taking off

(b) Audio spectrum of a large airliner

Figure 4.9: Characteristics of the airliner sample

4. Helicopter

The fourth and last sample is a 30 seconds long noise caused by a helicopter.

As seen in figure 4.10a, due to the rotation of the rotor blades, it creates a

loud and bass-heavy noise between 0 and 2000 Hz. Additionally, a lot of

higher and more quiet frequencies up to 9000 Hz can be measured, which

is created by the oscillating wind. Other than the first samples, this sample

does hardly have any variations, so it provides a good comparability source

for some experiments later. The audio spectrum of the helicopter is visualized

in figure 4.10b. Here, a generally even loudness can be observed with no

visible fluctuations.

4 Experimental Design

(a) TFFT of a helicopter

(b) Audio spectrum of a helicopter in the relative

Figure 4.10: Characteristics of the helicopter sample

4.2.2 Recording Environment

For the recording of the samples a specific experimental set-up was created. This

included setting up a sound system in stereo mode with two side and one center

speaker to achieve a wider angle of sound source. The center speaker was placed

in the middle directly in front of the smartphone. The two side speakers are placed

40 cm to the side and turned 45 degrees inwards. To eliminate possible post pro-

cessing of the sound system as much as possible, any audio enhancing modes and

bass amplifier were disabled. The samples were transferred from a laptop over an

AUX IN cable.

To conduct experiments from different distances, markers were installed on the floor.

Every meter, stripes of different color were installed. Due to limited room space

about 5 meters were covered.

During the experiments, the ambient noise was minimized as much as possible to

about 30 dB(A).

4 Experimental Design

4.2.3 Technical Requirements

For the process of measuring sound levels with a smartphone, a custom android

application was created.

One feature of the application was a selection of the recording mode the application

should use for the upcoming measurement. Although, Android only mentions the

MediaRecorder

class on their official documentation for recording audio6, there is

in fact one more possibility to record audio, which includes the

AudioRecord

class.

The differences are rather complex.

MediaRecorder

only offers a straightforward

solution for developers who just want to record audio and not process it further

in any way. Therefore, data is directly written into a file without any possibilities

to intervene during the process of recording. Nevertheless, in terms of measuring

sound levels,

MediaRecorder

does include the

getMaxAmplitude()

method, which

allows to receive the maximum amplitude recorded since its last call. After some

tests, it seems that the return value is converted to signed 16-bit integer values (0

to 32767), which correspond to the sound pressure recorded at the microphone.

However, the Android documentation does not explicitly state this7.

On the other hand,

AudioRecord

offers a greater number of choices, as it provides

the raw sound stream in byte format, which needs to be compressed manually.

Even though this seems rather obstructive since it makes the whole process of just

retrieving the amplitude more complicated, it is actually a huge benefit, as sound

compressing can be accomplished by the app directly. Hence, there is no need of

being dependent on the unknown underlying algorithm of the

MediaRecorder

. Ad-

ditionally, small experiments on different devices have shown, that retrieved sound

levels are much higher, when using

AudioRecord

, which leads to a higher accuracy

when sensing small level changes. A comparison between both modes is shown in

figure 4.11.

However, both classes support the deactivation of any possible activated AGC that

usually normalizes any incoming signal to the same amplitude, which would make

any sound level measurements pointless.

6https://developer.android.com/guide/topics/media/index.html

7https://developer.android.com/reference/android/media/MediaRecorder

4 Experimental Design

Figure 4.11: Different amplitude levels can be measured, depending on the selected

recorder mode.

Before the start of a recording, different parameter need to be set. This includes the

audio sampling rate, audio encoding bit rate and the audio channel.

The

MediaRecorder

class also includes setting an output format and the audio en-

coder, which should be used since the recorded data is directly encoded and saved

into a file. On the contrary,

AudioRecord

allows setting an audio format, which is

set to a linear Pulse Code Modulation (PCM). The application automatically selects

a representation of a 16-bit signed integer, or if unavailable, it falls back to 8 bit.

Same applies for the sample rate, which has a default value of 48000 and will fall

back to smaller values, if higher ones are not available. The number of channels is

set to mono on default, to increase comparability, as only some smartphones are

capable of stereo recordings.

Furthermore, the application measures the current sound level in both modes every

50 milliseconds for 40 seconds and logs the amplitude including a timestamp in a

CSV file. The uncompressed audio data is also recorded in both modes and saved

if further analysis needs to be performed. Screenshots of the application can be

seen in figure 4.12.

4 Experimental Design

(a) The start screen with a toggle switch

to start a direct recording with Au-

dioRecord or the rather unsophisticated

MediaRecorder.

(b) The application records the amplitudes

Figure 4.12: Screenshots of the self-developed application to measure amplitudes

4 Experimental Design

4.2.4 Monodirectional Behavior

In the same way loudness does, distance is having an impact on the measured

sound level, as sound waves are traveling longer until they are recorded by a mi-

crophone. To find out what extent distance has on the measured sound level an

experiment will be conducted where the distance from the acoustic source to the

microphone will be increased consistently.

The monodirectional behavior experiment will be performed in a way to achieve a

high degree of realism to serve as a reference statement for future studies. There-

fore, three speakers are positioned in one line in front of the person taking the

measurements to get a distributed sound source over a range of about 1 meter.

This is due to the fact, that typical noise is mostly not released from a selected point

but from a broader range. Additionally, it is assumed that the source of noise orig-

inates at or below of the ears of the person and that the smartphone is generally

placed lower than the head. As a general rule, it shall be defined, that the most

frequently taken posture for recording measurements is a person standing upright,

with a smartphone in the (right) hand at the height of the thorax region and with the

smartphone microphone facing the body.This posture with its common position of

the smartphone was selected because it can be found very frequently in everyday

life of a lot of persons. Therefore, having this height of the smartphone might help

to receive more comparable results for the analysis of data of future experiments

and studies. Since this effect was prioritized, no preference was given to measure

the exact loudness coming out of the speaker. In conclusion, this posture can serve

as a general standard for future measurement research since it has been defined

to best resemble the conditions where future measurements should be taken.

However, due to the individual sound characteristics of the narrow and tall room

where the experiment was conducted and due to the fact, that no sound absorber

or other precautionary measures were installed to mitigate sound reflection, the

outcome will result in a lot of inevitable deviations. Therefore, the results cannot be

compared to similar experiments taken in a sound laboratory, since the experiment

was not conducted in perfect conditions. Nevertheless, it may serve as a general

orientation for future measurements, since the results should still provide significant

data.

By conducting the experiment, the following hypotheses shall be validated:

4 Experimental Design

1. The measured sound level will decrease with increasing distance.

2. Decrementation of measured sound levels will occur proportionally with in-

creasing distance. Only minor measurement deviations might occur.

4.2.5 Omnidirectional Behavior

The validity of sound level measurement likely depends heavily on the direction the

sound is coming from. This is assumed due to the fact, that smartphones usually

only have one primary microphone, which is positioned at one side of the device.

Although, often smartphone manufacturers claim their microphone to be omnidirec-

tional, this cannot be achieved reliably because of the construction type of smart-

phones. This fact has already been investigated in other studies like Faber (2017),

therefore the focus shall be put on the posture in which the smartphone is held.

This can make a difference, due to coverage of the microphone by parts of the body

like the hand or the upper body, when facing away from the sound source.

To investigate this issue further, the following experiment was conducted. One per-

son stands in front of a speaker with a smartphone in his hand. Then he rotates

around his own axis to let the sound hit the smartphone from different sides. As the

sound is coming from the front most of the time, the smartphone itself deflects part

of it and a lot is reflected by the body of the person into the microphone. However,

when the person is turning to the side by 90 degrees, sound will pass by the mi-

crophone unchanged, as there is only little reflection, but it will not directly hit the

smartphone. Turning by 90 degrees further to 180 degrees, the person is stand-

ing between microphone and sound source, so amplitude values will most likely be

much lower than before. The last 90 degree for a total of 270 degrees is similar

to the 90-degree turn, nevertheless as the smartphone is held in the right hand,

sound waves could slightly be blocked by the right arm, although it is unknown if

this deviation is measurable with the given experimental setup.

Besides testing the results of the internal smartphone microphone, the same exper-

iment shall be conducted with an external microphone to compare the results. It can

be assumed, that even a cheap, external microphone with omnidirectional capabili-

ties, is performing better by delivering much more precise results, than the build-in

microphone of the smartphone, since it is less likely, that the external microphone

4 Experimental Design

gets covered up, when attached correctly.

In summary, the following hypotheses can be made:

1. A smartphone microphone reflects direction differences in the sound level it

measures depending on a predefined posture.

2. Sound coming from another direction then the smartphone microphone points

to, will be measured with a lower sound level.

3. An external microphone provides steady results, independent on the direction

of sound.

4. According to 1., 2. and 3. an external microphone will deliver more precise

results than the internal microphone with variable direction of sound.

5 Results

During the process of collecting data for the experiments, various smartphones run-

ning the Android operating system were used. To represent a selection of varying

hardware, smartphones of the three companies Samsung, Huawei and Oneplus

were used. The utilized smartphones are listed below:

• Huawei Honor 6X, running Android Version 7.0

• OnePlus 7 Pro, running Android Version 10.0.7

• Samsung Galaxy S4, running Android Version 6.0.2

• Samsung Galaxy S7, running Android Version 8.0

• Samsung Galaxy S8, running Android Version 9.0

Additionally, an external condenser microphone with an omnidirectional polar pat-

tern is used, which has a frequency range of 20 Hz to 16 kHz and a sensitivity of

-30 dB±2dB.

For a lot of experiments during this study, a SLM of the type Tenmars TM 103 was

used which has a frequency range of 31.5 Hz to 8 kHz, a measurement range of 30

to 130 dB(A) and an accuracy of ±1.5dB.

The analysis of gathered data was performed by using python with the mathematical

and scientific libraries NumPy and SciPy. To process the gathered measurements,

the database format SQLite was used. Other diagrams were created with Microsoft

Excel with the CSV file format.

During the analysis of the results an interesting observation was made. The results

often showed a significant amount of double measurement values consecutively,

indicating a transmission smaller than measured since the polling rate of the app

was set to 50 ms. However, these values have not distorted the comparison of

5 Results

multiple data sets, as not the amount of polling values was decisive for the length of

analyzed data but the length of the audio sample. Anyhow, all consecutive double

values have been filtered out.

The following chapter gives an overview of the results that were gathered while

performing different acoustical smartphone experiments. The first section deals

with the results of the two experiments that were performed to achieve a reasonable

degree of precision among different smartphone models. The section afterwards

investigates the acoustical behavior of smartphones, especially how reliable the

measurement values are they provide. Therefore, data is gathered under different

conditions to obtain a wide range of information, which will allow analyzing gathered

data in the future more efficiently. Finally, the third section deals with the evaluation

of the already collected data of the Track Your Tinnitus application.

5.1 Calibration Techniques

The acoustical calibration of smartphones is a vital task for a possible model-wide

comparison. Since each smartphone manufacturer or even smartphone model

measures amplitudes at a varying specific value range, it is hard to calculate a

comparable loudness value, e.g. in dB purely based on the given amplitude values.

As an example, the violin sample seen in figure 5.1 shows, that the Samsung smart-

phone records amplitude values mostly in the range of 6000 and 8000, sometimes

tearing out to 5000 and 9000. On the other hand, values of the Huawei are generally

lower at around 3000 to 6500. Unlike the previous models, the Oneplus shows lower

values around 3000 to 4000 almost continuously. In addition to that, the fluctuations

in amplitude on the Oneplus are very low, which could be due to the smartphone

characteristics in general or it could be a sign for a built-in, non-deactivatable AGC.

Another possible reason for low indicated fluctuations might be a low rate of trans-

mission of the current amplitude, either by the ADC to the system or by the system

to its API. Nevertheless, it is shown, that a calibration is necessary before taking

any sound level measurements.

5 Results

Figure 5.1: Comparison of smartphone models recording at varying amplitude

ranges

The following section presents the results of the calibration experiments.

5.1.1 Basic Calibration

The following experiment was performed by using the application "Spaichinger Schal-

lanalysator" for Android with version 2.2 (last update: 05.05.2020)1.

The experiment is split into two parts. Firstly, conducting the calibration approach

of the application, which includes tearing apart 10 sheets of copy paper. Secondly,

measuring sound levels with the smartphone and a SLM next to it. For this pur-

pose, multiple frequencies (100 Hz, 500 Hz, 1000 Hz, 2000 Hz, 4000 Hz, 5000 Hz,

7000 Hz, 8000 Hz, 10000 Hz) were played at a consistent volume with a Bluetooth

1https://spaichinger-schallpegelmesser.de/schallanalysator.html

5 Results

speaker. This makes it possible to find out how much the sound levels of the smart-

phone deviate from the ones measured with the SLM and to find out if the deviation

changes for different frequencies. The sound level for each frequency was consec-

utively measured 12 times simultaneously on the Oneplus smartphone and on the

SLM.

The calibration part of the application is performed with the help of an interactive

process. Initially, the process starts by measuring the background noise to calculate

a threshold value for later measurements. Then, the actual calibration process can

be initiated by the user, by pressing a button and starting to tear apart one sheet

of paper. Meanwhile, the application measures the amplitude of the tearing and

allocates it a decibel value. This process is repeated for 10 times. After that, a

standard deviation is calculated and outputted. For this measurement, a deviation

value of 2 dB(A) has been given by the application.

106

113

120

127

134

141

148

155

162

169

176

183

190

197

204

211

218

225

232

239

246

253

260

267

274

281

288

295

302

309

316

323

330

337

344

351

358

365

dB(A)

Measurement interval (about 180ms each)

Figure 5.2: Measuring sound levels while tearing apart five sheets of paper to cal-

culate the deviations of the peaks

To prove how the standard deviation of the tearing process can be computed, five

sheets of paper were teared apart, and the noise was measured by using the men-

5 Results

tioned application in an already calibrated phase. The audio spectrum of the pro-

cess can be seen in figure 5.2. A standard deviation was then calculated with the

maximum values of each tearing approach. By doing this, the standard deviation

results in 1.7 dB(A), which is close to the 2 dB(A), that were calculated by the ap-

plication while performing the calibration with ten sheets. Additionally, the average

of all maximum values is calculated, which corresponds to 74.8 dB(A).

100

500

1000

2000

4000

5000

7000

8000

10000

dB(A)

Frequency

Smartphone

SLM

Figure 5.3: Measured sound levels of different frequencies on a smartphone and on

the SLM after having performed a smartphone calibration

The second part of the experiment gives insights on the deviations of the sound

levels measured by the smartphone application and a SLM. To do this, the sound

levels of each frequency were mean averaged for the smartphone and SLM each.

Then, both results were compared, which is shown in figure 5.3. Here, the frequen-

cies in Hz are depicted, as well as the recorded sound level in dB(A). Both devices

show a high increasing sound level between 100 Hz to 500 Hz and then only a

lower increase until 4000 Hz. As it was expected since both SLM and smartphone

measure the sound level in dB(A) this does approximately match the A-weighting

filter curve. Although, the A-weighting peaks at around 3000 Hz and further shows

a slow decrease. Calculating a two-paired, homoscedastic t-test of smartphone

and SLM values (without 100 Hz), provides a result of p= 0.68. Nevertheless, it

5 Results

can clearly be said, that the measurements of the smartphone are reliable, which

proves the first hypothesis.

-9

-8

-7

-6

-5

-4

-3

-2

-1

2000

4000

6000

8000

10000

12000

Smartphone devia�ons from SLM in

dB(A)

Frequency

Smartphone devia�on

Maximum devia�on of 3 dB(A)

Maximum devia�on of 2 dB(A)

Figure 5.4: Average deviation of sound levels measured with a smartphone in re-

gard to the SLM at different frequencies (without 100 Hz).

To further evaluate the deviations of the smartphone, the values were analyzed in

detail. The deviations of the smartphone in regard to the SLM for each measured

frequency can be seen in figure 5.4. Although, the deviations of the smartphone

seem to be very high with about -14 dB(A) at 100 Hz, the error decreases dras-

tically at 2000 Hz in the range of -8 dB(A) to -9 dB(A) and then further improves

for 4000 Hz to around 5 dB(A). After that, the error rate decreases even further to

a range of -4 dB(A) to -2 dB(A) for the highest frequencies measured. The rea-

son for these differences could be the result of different frequency response rates

of the microphone for each frequency, which will be discussed later. Additionally,

some minor measurement errors shall be considered. Nonetheless, it is neces-

sary to say, that no sound level correction value has been added to the measured

smartphone values, which would be needed to settle the effects of the room, where

the measurement has been taken place. However, by adding a correction value of

5.5 dB(A), the measurement error reduces to ±3 dB(A) for frequencies between

500 and 10000 Hz. An even further reduction to a maximum error of ±2 dB(A)

5 Results

is also ensured for frequencies between 500 and 8000 Hz. Therefore, the second

and the third hypothesis, that different frequencies are recorded with different sound

pressure levels and do not exceed a certain deviation level, have also be proven.

Interestingly, different frequencies do not generally show large fluctuations in devi-

ation, which shows that different frequencies only have a limited impact on the ac-

curacy of the measured sound level for the tested smartphone microphone. Thus,

applying only a single correction value for all frequencies is feasible. Nevertheless,

to achieve even better results with smaller deviations the frequency response values

of the microphone should be taken into account if known, to limit the compensation

of the frequency difference. Another possibility, if the exact frequency response

value is unknown, is to use the frequency-based calibration method, which is ex-

plained in the next section 5.1.2.

5.1.2 Frequency-based Calibration

To include the frequency response of the smartphone microphone in the calibra-

tion, a more sophisticated approach is necessary. For this purpose, a set of fre-

quencies was outputted with a loudspeaker 20 cm away from a smartphone mi-

crophone with an initial calibration value of 80 dB(A) at 1000 Hz. This defined

volume level was not changed during the experiment. For each smartphone, every

frequency was measured three times and average values were calculated. Calcu-

lating a two-paired, homoscedastic t-test of the results, provides significant values

with α= 0.05 and p < α for p= 0.016, p = 0.005 and p= 0.004 for the Galaxy S4

and p= 0.049, p = 0.0and p= 0.0for the Galaxy S7. The Oneplus results were

not significant for the second try with p= 0.0, p = 0.787 and p= 0.0.

The results of each measurement are compared in figure 5.5. Both the Oneplus

7 Pro and the Galaxy S7 have consistently higher amplitudes, which are at least

twice as large for 5000 Hz and seventeen times as large for 10 kHz than the Galaxy

S4. This shows the differences that the smartphones have due to their specific

characteristics. Additionally, a distinct frequency response for each smartphone can

clearly be seen as well as the results for alternating each frequency. This shows

that a single calibration value for every smartphone is not sufficient in every case

for a general well-performing calibration. Therefore, the first hypothesis that varying

frequencies have an impact on the recorded amplitude is correct.

5 Results

2000

4000

6000

8000

10000

12000

500

1000

2000

4000

5000

7000

8000

10000

Amplitude

Frequency

OnePlus 7 Pro

Galaxy S4

Galaxy S7

Figure 5.5: Measured amplitudes for the frequency calibration with indicated stan-

dard deviations.

The next step was to generate a dependency between amplitude and the mea-

sured sound level. Since the sound level of the A-weighting is not constant for all

frequencies, a difference value must be added or subtracted to the 80 dB(A) at 1000

Hz. This is further explained in section 2.4.3 with figure 2.4. However, measuring

the sound level in dB(A) with a SLM, the recorded values are not the same due to

measurement inaccuracies like direction, position or reflection. The maximum devi-

ations were 4.06 % for 20 kHz, which corresponds to +3.3 dB(A) or 4.26 % for 10

kHz, with -3.3 dB(A). For all other frequencies, the deviations were lower or even

zero. The full list of deviations is shown in table 5.1.

Frequency in Hz 500 1000 2000 4000 5000 7000 8000 10000

A-weighting in dB(A) 76,8 80 81,2 81 80,5 79,48 78,9 77,5

SLM in dB(A) 74,4 80 84,5 81 81,8 79,6 77,8 74,2

Relative deviations in dB(A) -2,4 0 3,3 0 1,3 0,12 -1,1 -3,3

Absolute deviations in % -3,12 0,00 4,06 0,00 1,61 0,15 -1,39 -4,26

Table 5.1: Deviations of the SLM to the A-weighting

5 Results

All in all, there are two options to calibrate the amplitude values with: the opti-

mum A-weighting decibel values or the recorded experimental ones which include

the measurement deviations. For the following calculations, the reference curve

was created with the decibel values, which were taken with the sound level meter

under experimental conditions. This decision was made, because the specific cir-

cumstances under which the data with the sound level meter was gathered are the

same as when the amplitudes were recorded with the smartphone. To display the

maximum deviations to the optimal A-weighting curve, the standard deviation is in-

dicated in the plot, however as already discussed, it is very minor and not detectable

in figure 5.6.

0,00

500,00

1000,00

1500,00

2000,00

2500,00

3000,00

3500,00

100

200

300

400

500

600

700

500

1000

2000

4000

5000

7000

8000

10000

Calibra�on factor

Frequency

OnePlus 7 Pro

Galaxy S7

Galaxy S4

Figure 5.6: The calibration factors for the Galaxy S4 (right axis) are much higher

and show a greater deviation rate than the ones of the Oneplus and the

Galaxy S7 (left axis).

To achieve a general comparison for each frequency across all devices, the dB(A)

value measured with the SLM is divided by the measured amplitude of the smart-

phone, which gives a frequency- and device-specific calibration factor. Due to norm

purposes, it is multiplied by 10000. This factor has been chosen to adapt to the

current amplitude values of the smartphones. Nonetheless, should the amplitude

values rise for smartphones in the next generations, this factor must be scaled fur-

ther up or the calibration values will be rather small or result in a high deviation. The

current decibel deviation of the calibration factor varies due to the amplitude range

5 Results

of the smartphone and is larger for higher amplitudes. For the Oneplus 7 Pro at

8000 Hz, the highest amplitude was measured in this experiment, which leads to a

deviation of 1 calibration factor of 0.9 dB(A). All calibration factors can be seen in

figure 5.6. The calibration factors of the Oneplus and the Galaxy S7 are in a similar

range, due to their similar operating amplitudes and are denoted on the primary

vertical axis. Since the Galaxy S4 records amplitudes with a much lower value, a

larger calibration factor is needed, shown on the secondary vertical axis. Another

fact is that the graphs of the Oneplus and the Galaxy S7 are very steady and both

show low fluctuations, however, the fluctuations in amplitudes are much higher for

the Galaxy S4. For example, the calibration factor is six times as high for the Galaxy

S4 between 5000 Hz and 8000 Hz due to the high drop in amplitude. For the One-

plus and the Galaxy S7, the deviation is not greater than three times between 500

Hz. and 10000 Hz. This indicates that the deviations between the microphone char-

acteristics of the Galaxy S4 in regard to the A-weighting are much greater than with

the other smartphones. Since different smartphones do not record frequencies with

the same amplitude, the second statement is proven.

Additionally, the question is raised by how much the accuracy changes if multiple

sound levels are included into the calibration. With the use of the self-developed

application, many frequencies have been outputted on a Bluetooth speaker with

the use of the complete volume spectrum of the smartphone. The results of the

self-calibration can be seen in figure 5.7.

5 Results

(a) Samsung Galaxy S6

(b) Samsung Galaxy S8

Figure 5.7: Measuring volumes of different frequencies on two different Samsung

Galaxy models with external speaker.

5 Results

The graph of figure 5.7a shows multiple frequency lines, which were outputted and

measured at different volumes with a Galaxy S6. Generally, frequencies between

8 Hz and 128 Hz have overlapping amplitudes and form a steady increasing line.

Interestingly, the same applies for 1024 Hz. Higher frequencies with 256 Hz, 512

Hz and 2048 Hz have continuously much steeper rising curves, due to microphone

characteristics. Frequencies of 4096 Hz, 8192 Hz and 14000 Hz have not been

generated and outputted correctly by the tone generator used with the application,

however they are not relevant for this example, which is why they have been left out.

The average background amplitude level was 249 for both measurements.

The same experiment has been conducted with the Galaxy S8 which is another

smartphone of the same manufacturer, however in a later version. This was done

to see how large the deviations in model variants are. In figure 5.7b, a relatively

similar image to the first tested version can be seen. However, the former bundled

frequency ranges are now stretched further apart and more fluctuating for higher

selected sound levels. In contrast, for lower sound levels the measured frequencies

are dense. Moreover, the higher, correctly outputted frequencies, are not measured

at higher levels, thus a huge change in microphone characteristics can be detected.

Looking at the general changes for increasing volumes, the overall rise is far lower

compared to the curves, than what is shown in the curves of the former model.

This fact can obviously be identified, when calculating a moving average curve out

of all frequencies as seen in figure 5.8.

This has been done to efficiently compare the average amplitude level with the other

smartphone variant. Both curves show a similar behavior for the first five volume

levels with deviations below 1000, then the curves split apart and increase with

different rates until volume level 12 with a deviation of about 9000. While the mea-

sured amplitude levels of the Galaxy S6 rise much higher, the Galaxy S8 does not

increase that fast, thus the distance to the S6 is getting bigger. The Galaxy S6 stops

its steep rise at a volume level of 13 and only increases marginally thereafter. For

the Galaxy S8, the curve peaks at level 14 and stays constant after that. Deviations

are between amplitudes of 8000 and 9000. However, as discussed below, the stan-

dard deviation of the mean is much higher at the end for the Galaxy S8 and more

visible at the middle levels for the Galaxy S6. This shows, that smartphones do not

represent rising sound levels evenly in amplitude, which is the answer to the third

statement.

5 Results

Figure 5.8: Moving average of amplitudes for frequencies 8 hz to 2048 hz of two

smartphone variants

5.1.3 Analysis of already gathered data

Before the work for this thesis started, a large data set has already been gath-

ered with the TrackYourTinnitus application for Android and iOS. The functionality

included a set of defined questions regarding the current state and extent of the

tinnitus perception at the time of answering [20]. For a more detailed analysis,

the sound level data should be combined with information about the current noise

situation. However, since the involved smartphones have not been calibrated be-

forehand, it is necessary to find a calibration approach afterwards. The gathered

data include the amplitude values, a user id, hard- and software information about

the device and the current time.

To determine the actual loudness data for a specific device it is possible to directly

convert the amplitude data into sound pressure values. However, this requires to

have information at least about the characteristics of the microphone, as already

discussed in section 4.1, since a more detailed approximation of the sound level is

only possible by having knowledge about the frequencies of the sound sample, as

shown in section 4.1.2.

5 Results

Nevertheless, to make a simplified statement about the approximate sound pres-

sure value for a given amplitude, an average amplitude calculated out of different

sound scenes was measured to include a wide range of frequencies. Therefore, the

violin and the traffic sound sample, which were already used in section 4.2, were

played by a Bluetooth loudspeaker at an average sound level of 80 dB(A) and 60

dB(A) at a distance of about 20 cm. The sound levels have been verified by an

SLM, the same way as it is shown in figure 4.5. At the same time, the amplitudes

were measured by the Galaxy S7 smartphone. The measured amplitudes of the

complete sample duration were averaged for each sample and a whole average

value was calculated out of that. The results are shown in table 5.2.

dB(A) 80 60

Amplitudes for the violin sample 3691 409

Amplitudes for the traffic sample 4282 401

Average amplitudes for both samples 3987 405

Table 5.2: Averaged amplitudes for the violin and traffic sound samples played at

two different sound pressure levels.

The first average amplitude of 3987 shall be defined as a0and was used for the

calibration at 80 dB(A). The second average amplitude of 405 shall be defined as

a1and was used later for validation purposes at 60 dB(A). Before calculating the

calibration factor, b= 100 amplitude values were subtracted from the amplitude at

80 dB(A), due to static background noise. Thereafter, the resulting amplitude is

divided by sound pressure of the corresponding sound level of 80 dB(A), which cor-

responds to the calibration value c. This is done by converting the sound pressure

level Lp, which is measured in dB(SPL) to the sound pressure p, which is measured

in Pascal, as introduced in section 2.4. Additionally, p0is defined as the reference

sound pressure given as p0= 20 µPa = 2 ×10−5. The formula is defined as:

˜p=p0∗10

[36]

For the calibration sound level of Lp= 80 dB(A), the sound pressure corresponds

to ˜p= 0.2Pa. This results in the following reference value, which serves as a

device specific calibration.

5 Results

c=a0−b

˜p=3987 −100

0.2= 19435

For the next step, an amplitude value for which the sound level is calculated is

divided by the calibration value. For this example, the amplitude of 405 is taken,

which corresponds to the sound pressure of this amplitude.

˜p=a1

c=405

19435 ≈0.0208 Pa

Lastly, this sound pressure can then be converted back to its corresponding sound

pressure level Lp[36]. For the measured amplitude value of 405 with its sound

pressure of about 0.0208 Pa, the sound pressure level is calculated as:

Lp= 20 ∗log10

˜p

= 20 ∗log10

0.02084

0.00002 ≈60.36 dB(SPL)

This corresponds roughly to the average volume of our sound samples which was

60 dB(A). This shows, that a conversion of amplitudes to a sound pressure level is

indeed possible. The complete source code is given in listing A.1.

This knowledge can be transferred to the data samples for the Galaxy S7 of the

TrackYourTinnitus dataset. For the Galaxy S7 140 data samples exist, which were

measured by one user. Calculating the amplitudes to their corresponding dB(SPL)

values by using the already explained method, creates the histogram in figure 5.9.

This shows, that all samples might be measured between 48 dB and 93 dB, which

appear to be reasonable values.

Since this approach is only feasible if loudness information about the smartphone is

known, another approach shall be presented if no calibration was done beforehand.

Nevertheless, this requires that the data set is large enough and the amplitude

range is known.

However, it is necessary to consider, that the sound levels have been calculated

by neglecting their corresponding frequencies, which is an important factor as seen

in section 5.1.2. Additionally, no defined measurement posture was defined, which

could interfere with the results, as seen in section 5.2.2. Nevertheless, the method

serves as a fairly accurate value to evaluate part of the already gathered data under

5 Results

22 20

dB(SPL)

[ 48, 53] ( 53, 58] ( 58, 63] ( 63, 68] ( 68, 73] ( 73, 78] ( 78, 83] ( 83, 88] ( 88, 93]

Number of samples

Figure 5.9: A histogram of the converted sound level values for the Galaxy S7 data

samples

the condition, that an amplitude value for a defined sound level is known.

Looking at the TrackYourTinnitus data set again, shows that the amount of individ-

ual data points is 78767, however, since this thesis only deals with Android smart-

phones, data set from Apple devices are omitted. The largest amount of measure-

ments (1596) was taken with an LG Optimus smartphone with an amplitude range

between 0 and 29050, which roughly corresponds to the default amplitude range on

Android. Other Android devices which show a high occurrence in the data set are

the Samsung Galaxy S III mini (1030) with a range of 89 to 31157 and the Samsung

Galaxy Note 3 (916) with a range of 0 to 4633. Since the maximum amplitude of the

Note 3 is much lower, compared to the other ones, this indicates, that the amplitude

range is different for varying devices, even from the same manufacturer.

Doing a further analysis with a selected number of devices which supposedly have

the same amplitude range of about 0 to 32767 (LG Optimus, Samsung Galaxy S III

mini, Samsung Galaxy S4), provides a sample of 33710 data samples by 41 indi-

vidual users. Falsified values with a sound level of 0 were filtered out. The analysis

can be seen in a histogram in figure 5.10. By dividing the amplitude scale into sev-

eral regions this indicates that with 30% about third of the samples were taken in an

amplitude range between 15 and 1515. The next group with amplitudes up to 3015

takes up 21% of all samples. Then amplitudes to a value of 4515, which is about a

third of the whole amplitude range, correspond to 11% of all samples. It can clearly

be seen that most of the amplitude data is taken in rather quiet surroundings like at

home for example in a calm room. This fact can help to further analyze the sound

5 Results

1012

698

385

300 236

153 133 97 78 58 38 44 29 19 17 18 21 15 7 7 5

Amplitude ranges

[15, 1515]

(1515, 3015]

(3015, 4515]

(4515, 6015]

(6015, 7515]

(7515, 9015]

(9015, 10515]

(10515, 12015]

(12015, 13515]

(13515, 15015]

(15015, 16515]

(16515, 18015]

(18015, 19515]

(19515, 21015]

(21015, 22515]

(22515, 24015]

(24015, 25515]

(25515, 27015]

(27015, 28515]

(28515, 30015]

(30015, 31515]

Number of data samples

200

400

600

800

1000

1200

Figure 5.10: A histogram with the number of data samples for multiple Android

smartphones of the same amplitude range.

level for example by setting a constant volume level for the most common amplitude

values.

For a very quiet room the sound level could be about 20 dB(A). Since the lowest

measured amplitude is 15, this will serve as the corresponding calibration ampli-

tude. Calculating the sound pressure levels for every data point as shown before

creates the histogram in figure 5.11.

Here, it can easily be seen, that most of the data samples could be taken at around

60 dB(SPL) to 70 dB(SPL), which corresponds to the sound level of a conversation

or street traffic. Since these are both frequently occurring occasions, the calibration

values seem to be well selected. Nevertheless, once again these values do only

represent an approximation of the actual sound level and should therefore not be

taken at face value.

5 Results

179

757

1133

1012

220

dB(SPL)

≤ 40 (40, 50] (50, 60] (60, 70] (70, 80] (80, 90]

Number of data samples

200

400

600

800

1000

1200

Figure 5.11: A histogram with the number of data samples and their corresponding

sound level values for multiple Android smartphones of the same am-

plitude range.

5.2 Acoustical Behavior

Investigation of the acoustical behavior of smartphones, was performed by using

a stereo sound system and the already in section 4.2.1 described samples. All

experimental data was taken on a consistent volume level. The number of tested

smartphones was n= 3 with different manufacturers to cover a diverse spectrum

of hardware. Furthermore, the experiments were repeated with an external micro-

phone, connected via an AUX-IN cable instead of the in-built smartphone micro-

phone.

5.2.1 Monodirectional Behavior

To find out if and to what extent distance has an influence on the amplitude and

how the quality varies for increasing distance, further experiments have been con-

5 Results

ducted. As sound only emerges from one direction and directly hits the smartphone,

the experiment measures the monodirectional behavior of the smartphone. The

experiment is conducted by playing different sound samples on a stereo system

and recording the emitted sound waves on different smartphones at three distinct

distances from the main speaker (1 meter, 2 meters, 3 meters). After that, the

experiment was repeated by using an external microphone as the choice of input

source.

During the experiment the smartphone was held in the hand at a height of 1.5 meter

in a posture explained in subsection 4.2.4.

00 10 20

2000

4000

6000

8000

00 10 20

Figure 5.12: Measuring amplitudes with increasing distance with the aircraft takeoff

sample on the Galaxy S7 (left) and the Oneplus (right)

The results show that the measured sound level indeed decreases with further dis-

tance with every device. The measured amplitudes were consistently lower on all

devices with consistent behavior for the aircraft takeoff sample. This can be seen in

figure 5.12 for the Galaxy S7 and the Oneplus smartphones. For both smartphones,

the measurement provides similar results when comparing both smartphones in

terms of their ability to reflect distance correctly in the amplitude values. Noticeably,

the Galaxy shows more deviant values for different distances, consequently it might

have a better distance measurement than the Oneplus, which values lie more to-

gether. Another interesting point is that the course of the polynomial trend line is

not the same for all distances. For the first distance on both phones more excessive

5 Results

movements can be observed, thus the sound level characteristics of the sample are

described better. For increasing distances, the amplitude movements decrease.

00 10 20

1000

2000

3000

4000

5000

00 10 20

Figure 5.13: Measuring amplitudes with increasing distance with the helicopter

sound sample on the Galaxy S7 (left) and the Oneplus (right)

A similar result can be seen with the helicopter sample on both phones in figure

5.13. Again, the Galaxy S7 shows much more split apart values for the first distance,

while the amplitude values for the Oneplus seem to have spread more equally.

Nonetheless, all smartphones show the ability to reflect distance in their amplitude

values for both samples.

However, the ability to distinguish distance well is not consistent with all samples.

The violin sample does not give any coherent clues about the difference of 2 meters

and 3 meters distance as seen in figure 5.14. While a distance of 1 meter is clearly

identifiable with the use of a polynomial trend line, for an increasing distance this

is not possible anymore. The reason for this might be the high frequencies of the

violin which might be more prone to a higher rate of room reflection. Similar results

are measured on all smartphones and even the use of an external microphone did

not show any improvement in this regard.

5 Results

00 10 20 30

2000

4000

6000

8000

Figure 5.14: No statements about the distance of 2 m and 3 m can be made for the

violin sample across all devices

Repeating the whole experiment with the use of an external microphone, rather

leads to similar results for the traffic sample with the Galaxy S7 smartphone. A

comparison for the internal and external microphone can be seen in figure 5.15.

This similar behavior was also noticed on other devices. Analysis of the raw ampli-

tude data of the devices however shows, that the 1 meter and 2 meters measure-

ment values are split more apart when using an external microphone. This creates

a noticeable larger margin between the amplitudes of the first distance and the ones

thereafter, which can also be confirmed consistently on all devices. Although, this

behavior is not evenly distributed on all smartphones, as intensity varies according

to which smartphone is used. For the Samsung smartphone this effect can slightly

be noticed at around the 25 second mark, as the curves do not overlap that much.

Nevertheless, for the monodirectional measurement, the differences can only be

measured slightly.

5 Results

Figure 5.15: Comparison of amplitudes with increasing distance by using a Sam-

sung Smartphone (left) in combination with an external microphone

(right)

To find out, whether the obtained results of the monodirectional behavior can be

valued, it is necessary to prove them in some way. Following the scientific Inverse-

square law, the sound pressure decreases proportionally with distance by 1/r. Since

this rule should apply to the executed measurements as well, it should be possible

to find out if the law can be transferred to smartphone recordings. Proving this would

allow the comparison of two or more values received from a sound source playing

sound with a steady loudness. Furthermore, this technique would allow to roughly

determine the distance from the sound source to the smartphone. To evaluate if

the executed experiment does have any correlation to the Inverse-square law, three

distinct amplitude values from each measured distance (1 meter, 2 meters, 3 me-

ters) have been compared to the proportional decrease of the Inverse-square law.

However, since only three measurement points exist, just an approximate guess

with three data points can be made.

For analysis purposes, the measured amplitude value of each distance is plotted

together with the curve of the Inverse-square law 1/r, where r is defined as the dis-

tance in meters. This attenuation factor is then applied by multiplying the measured

amplitude at 1 meter to it, which allows to adapt the formula to the measurement.

The results can be seen in figure 5.16. A clear decrease in amplitude for increas-

5 Results

1 2 3

Distance in [m]

1500

2000

2500

3000

3500

4000

Amplitude

1/r

Measured sound level

Figure 5.16: Comparison between proportional signal reduction and measured val-

ues

ing distance can be analyzed, as it was observed before. However, signal levels

do not decrease proportional but proceed more linear. The reason for this might

be the acoustical characteristics of the room, where the measurement was taken.

Due to sound reflection of the walls, the amplitude degradation is far less than as-

sumed. If the experiment was replicated with less reflection, the derivation of the

measured levels with regard to the expected values should be far less. Additionally,

it is unknown, if the amplitude of the smartphone directly corresponds to the sound

pressure, since it is an arbitrary value of the Android OS. Hence, it might not entirely

describe the sound pressure at the smartphone microphone.

Even though this experiment was only performed at three distances and the acous-

tical characteristics of the room might have modified the measurement values, the

results still indicate a clear reduction of amplitude with accumulated distance. Com-

paring the results with the curve of the Inverse-square law, they are in a proximate

range, although in spite of the given limitations a highly certain prove cannot be

made.

5 Results

5.2.2 Omnidirectional Behavior

To figure out the impact of the direction the sound is coming from, the sound level

was measured at different directions from the sound source. As sound is not always

hitting the smartphone from straight ahead, it must be evaluated how other sound

directions influence the measured amplitude of the smartphone.

1000

2000

3000

4000

5000

6000

Figure 5.17: Measuring the amplitude with the internal microphone of the Oneplus

(left) and an external microphone (right) at different positions to the

sound source. The use of an external microphone leads to lower devi-

ations.

The omnidirectional experiment has been performed with the traffic sample. As fig-

ure 5.17 shows, variations in loudness can clearly be seen for different positions.

In the first orange marked range, the sound is coming from the front (0 degrees).

After that, the person holding the smartphone in his right hand turns to the right (90

degrees). Interestingly, an increasing amplitude can be observed. Then, the person

is turning right again and stands with his back to the sound source (180 degrees).

The amplitude declines a lot in consequence of this and less amplitude spikes are

measured. As a last position, the person turns right again (270 degrees), which in-

creases the amplitude again, which can be seen in the yellow range. Nevertheless,

a slight less average amplitude value than in the green range can be observed, al-

though the person stands sideways to the sound source during both ranges. This

might be the case due to the shielding of the right arm, which blocks the sound

5 Results

waves, that would directly hit the microphone, therefore the amplitude is lower in

the yellow range than in the green. On the other hand, the green range shows a

large increase, which could be the case of the reflection of the right hand, which

reflects the sound waves into the microphone, thus the spike in amplitude is seen.

Repeating the experiment with an external microphone does show some differences

as expected. The use of a true omnidirectional microphone does indeed mitigate the

position differences since the amplitude does not show a lot of excessive changes

as the internal microphone does. Obviously, the amplitude drops significantly, when

standing behind the sound source (blue range), since the body blocks a lot of the

incoming sound waves. However, this effect is much less, when compared to the

internal microphone. Amplitude changes between the other three positions are only

slightly visible and could also be part of the sample. Hence, it is clearly deducible,

that the use of an external microphone does lead to more accurate results, when

measuring sound from different positions.

6 Discussion

The performed experiments have shown that smartphones are indeed capable to

measure noise in daily life. Therefore, they can also be used as an easy tool to be

aware of the surrounding noise level in real time to lower the risk of health issues

like stress, hearing loss or tinnitus caused by high noise levels. Moreover, this type

of smartphone application can also be used to mitigate the effects of tinnitus for

affected people. This can be done for example by identifying unpleasant frequen-

cies or uncomfortable noise levels (too quiet or too loud) and issuing a warning if

a threshold is exceeded. In summary, this research about noise measuring with a

smartphone represents only the beginning of a larger field of application, which can

be adopted quite easily, after the basic issues are solved.

Basic Calibration Approaches

Several calibration approaches have been introduced to convert smartphone am-

plitude values to the comparable dB(A) unit. Two methods, the basic and the fre-

quency calibration, have been presented. Although, both methods have only been

tested with Android smartphones in this thesis, the theory behind it works for all

smartphones the same if their operating system provides an amplitude value that

corresponds to the surrounding loudness. This has been proven for iPhone smart-

phones as well by M. Ziegler [40]. Beyond that, the calibration can be reproduced

by anyone, which shows that the external validation can be ensured.

The results of the basic calibration show that it only takes about 5 minutes and

some sheets of paper to gain an accuracy of a maximum of 3 dB(A). However, a

compensation value of 5.5 dB(A) has to be taken into account to compensate for

any room impacts. Since this correction value has only been created by using the

help of a SLM, results may change if an estimation of this value must be assumed.

6 Discussion

This topic is further discussed in the section The Effects of Reflection. Nonetheless,

this approach has proven to work as an uncomplicated method to provide decent

results.

The measured results are comparable with the results of other testing measure-

ments by M. Ziegler (Sept. 2019), which have been gained by using this calibration

method. First, the standard deviation of the calibration is said to be about 2 dB(A).

For the conducted measurement, these 2 dB(A) have exactly been measured dur-

ing this study. Additionally, the accuracy of the measurement is supposed to be

about 2 to 3 dB(A), which has also been shown here with a maximum deviation of

3 dB(A). The compensation value was chosen a bit higher with 5.5 dB(A) than the

values with 0 to 3.5 dB(A) in the comparison measurement of Ziegler (2019). How-

ever, since the compensation value is dependent on the characteristics of room and

microphone, it is hard to compare [41].

Nonetheless, only one compensation value is added for the basic calibration, that

corresponds to all frequencies and sound levels. This shows a limitation of the

basic calibration. Since the process of tearing a paper covers a specific range of

frequencies and only has a defined loudness range, it does not take the specific

frequency response and the response for different sound levels of the microphone

into account. While a difference in frequency response was suspected before, its

existence has already been verified with the measurements of the basic approach.

Despite that, most of these limitations shall be solved by using the frequency-based

calibration, which aims at providing a solution for such issues.

Frequency-dependent Calibration

The frequency calibration focuses on two things. Firstly, to compensate any devia-

tions resulting from a different frequency response of the microphone and secondly,

provide compensation for deviations in different sound levels.

Additionally, the frequency calibration emphasizes the importance of performing an

individual calibration for different frequencies, since the amplitude curve of each

smartphone clearly shows an own behavior. By including an individual calibration

factor, the frequency differences can indeed be compensated. Nonetheless, it is

necessary to apply further research to bring the concept to application. Especially,

6 Discussion

when measured sound consists of multiple frequencies, a single calibration factor is

not available, due to the unique characteristics of each frequency to the recording.

Hence, it is necessary to determine a new calibration factor out of all recorded

frequencies. One solution could be to use the calibration factor value of the lowest

measured frequency of the sample in question, which is referred as the fundamental

frequency. This frequency has the feature of being the lowest and the loudest

perceived frequency of the human ear [4]. However, it must be shown, that the

calibration factor identified, results in a reliable sound level measurement.

Another approach to accomplish a valid calibration factor is by performing a time-

frequency analysis like the Short-time Fourier Transform (STFT). This is used for

continuous signal streams with changing frequencies over time. To determine the

corresponding frequencies, the time signal is split into segments, which can then

be analyzed further by dividing them into a frequency spectrogram [35]. A graphical

representation of the result is given in figure 6.1. Information about specific fre-

quency components of the sound could then be applied to calculate an appropriate

calibration factor using the already collected calibration data.

Figure 6.1: A 3D example for the short-time Fourier transform, which includes fre-

quency (x-axis), colored intensity (y-axis) and time (z-axis) [38].

6 Discussion

Additional Factor: Smartphone Age

One more issue that is worth mentioning, is that the age of a smartphone seems

to have a negative effect on the accuracy. The result of the frequency calibra-

tion shows a larger scale of amplitude for the newer Oneplus 7 Pro and Galaxy

S7 phones, which indicates a higher accuracy, due to a larger scale. Additionally,

looking at the calibration factors for each frequency of the involved smartphones,

shows that the frequency fluctuations are much higher for the Galaxy S4, whereas

the Oneplus and the Galaxy S7 show an almost constant trend. Hence, the extent

of necessary calibration is much higher for the Galaxy S4 since it does not depict

the A-weighting accurately. This finding is comparable to the results of Murphy and

King (2016), who observed that accuracy increases with newer phones. The rea-

son behind this might be more advanced microphone technology or a deterioration

for smartphone microphones with age [28]. Similarities between the Galaxy S4 and

the Galaxy S7, which are smartphones by the same manufacturer could not be

identified, however due to the age difference this is not surprising.

The Impact of Sound Levels for Calibration

The results of the impact of loudness for the calibration do not give a clear outcome.

Resulting from the measurements, a different increasing rate of amplitudes for vary-

ing sound levels was shown. This indicates that the amplitude does not increase

gradually. However, this was only shown by increasing the internal Android volume

level and testing for pure frequencies. It is necessary to state, that the Android

volume level generally cannot be used to calibrate a smartphone since it was not

proven that the output sound level gradually increases evenly with increasing vol-

ume levels. Furthermore, different loudspeakers could possibly output the volume

levels at different sound levels, too. Thus, a better procedure for future applications

would be to measure the different sound levels in dB(A), as it would increase com-

parability. Additionally, several loudness values in dB(A) could be defined to use

for future frequency measurements. Nevertheless, this includes additional hard-

ware like a SLM or an already calibrated smartphone. However, looking at the re-

sults of the monodirectional experiment shows that by decreasing the distance and

thus loudness indeed not every sample has shown an equally decreasing amplitude

(apart from possible impacts of reflection).

6 Discussion

Besides that, it is important to notice, that for this experiment and the behavior

measurements, no amplitude to dB(A) conversion was performed to avoid falsifying

the measurement results and showing the raw amplitude values. This was done

to prevent calculation errors tamper with the validity of the results. Therefore, no

conclusions from microphone pressure to dB(A) should be made.

To say if an inclusion of a loudness compensation value in future calibrations can be

done, further knowledge about the smartphone behavior for decreasing loudness is

needed. However, information about loudness cannot directly be retrieved since this

is part of the calibration result and the actual sound level is unknown. But retrieving

additional information about the type of noise and its direction could be achieved

by detailed sound analysis with the use of machine learning. Thus, the recorded

sound could be classified, which on one hand might be beneficial to the calibration

and on the other hand could be an advantage in tinnitus research, since the specific

type of noise which correlates to the perception of tinnitus at that moment could be

identified. Hence, this might serve as a future approach to improve calibration and

well-being as well.

Comparison to Related Studies

For a comprehensive comparison of the results, other studies are included, which

deal with noise levels measured by a smartphone. Kardous and Shaw 2014 have

tested several sound measuring applications for iPhone devices without perform-

ing any type of calibration and 3 of 10 applications had deviations below 2 dB(A)

[17]. A large experiment has been performed by Murphy and King (2016), who con-

ducted 1472 tests on 100 phones. Interestingly, only one application for Android

but four applications for iOS have accurately measured deviations below 2 dB(A).

Additionally, it was stated, that Android applications do generally provide “less reli-

able“ results, than iOS applications, which might be the case due to more diverse

hardware [28].

All in all, the relatively straightforward and uncomplicated calibration approaches

evaluated in the course of this study give promising results, which can be improved

even more by some modifications regarding the calibration. Nonetheless, the re-

sults show that a comparable level to professional applications can be achieved,

6 Discussion

although it is necessary to have in mind that the measurements for this work have

been gathered under limited conditions.

Additional Factors: Distance to the Microphone and its Surroundings

Results of the data for the acoustical behavior give different results depending on

the samples, which have been tested. A general capability to reflect even small

distance changes of one meter in sound levels was proven. Despite an even de-

crease for the traffic sample for all 3 meters, other samples have shown problems

to correctly reflect changes between 2 and 3 meters. Possible reasons for this are

again room reflection parameters. Interestingly, the impact does change depending

on the type of frequency, which is explained in further detail in section 6. Thus, it is

no surprise, that for higher frequency samples like the violin tones, a much higher

degree of reflection is measured.

The results of the omnidirectional experiment show, that the direction of the micro-

phone to the sound source does indeed make a difference. All in all, this experiment

provides knowledge if and to what extent different directions of sound sources have

an influence on the amplitudes measured. Obtained results can possibly help to

design future experiments which include the originating direction of sound into the

calculation of sound level, as sound can be sensed with a higher intensity by the

person when it is coming from behind, than it is measured with the smartphone

in front. Therefore, a compensation value might be added to account for direction

parameters. Having in mind, that the smartphone is not exposed directly to the

surrounding sound every time, the same applies, if the smartphone is covert up,

e.g. in a pocket. Under these conditions, it is harder to make any reliable assump-

tions about the results of the internal microphone, as low sound levels might not be

recognized correctly anymore.

Additional Facotrs: Use of external Microphones

The performance of the external microphone is varying. While for the Samsung

smartphone the use of an external microphone provides similar results as to the

internal microphone, the results showed a much lower amplitude for the Oneplus

phone.

6 Discussion

The reason for this lies in the custom characteristics of the external microphone

completely differing from the one of the internal microphone of the smartphone.

Here the question arises, whether it is possible to achieve the same acoustical be-

havior by using the same external microphone on different smartphones. However,

when comparing two different smartphones using the same external microphone,

the results still differ, as the sound processing modalities of the smartphones are

not the same, as it is also discussed in 2.5. For the external microphone in the

monodirectional experiment the amplitudes are diverging (about 500 higher with

the Samsung, and 2000 lower with the Oneplus). Nevertheless, for another exter-

nal microphone, this is different again. If that was the case, a re-calibration would

have to be performed every time, when an external microphone is used in combi-

nation with a new smartphone.

As a general consequence, values of the distance measurements increased at least

between the distances of 1 meter and 2 meters for all devices by the use of an

external microphone, which certainly improves the measurement quality. With 2

meters and 3 meters, the improvement can still be measured with the Oneplus

phone in a mitigated form, the Samsung phone seems to be more prone to sound

reflection since the positive effect of the external microphone can no longer be

detected. Nonetheless, the positive effect of the external microphone can be seen,

that is why the use of an external microphone is mostly better when compared to

the internal microphone. However, considering the used external microphone was a

low-price model, better results might be even achieved with a higher quality model.

These findings are close to the results of Faber (2017), who compared two inexpen-

sive external microphones and concluded that “even a very inexpensive microphone

[...] can be used for reasonably accurate measurements with a smartphone“ [9].

Additionally, a follow-up study of Kardous, Chucri and Shaw (2016) with external

microphones has shown an enhancement of the mean deviation from 2 dB(A) with

the internal microphone to 1 dB(A) with an external microphone, stating that exter-

nal microphones “greatly improve the overall accuracy and precision of smartphone

sound measurements“ [18].

6 Discussion

A Guidance for future Measurements

Following the outcome of the results, the general recommendation to retrieve loud-

ness data as accurate as possible, is to point the smartphone microphone in the

direction of the sound to avoid sound reflection as much as possible. For the same

reason, a lot of space should be around the microphone, which is the case with

wide rooms being more favorable than small rooms. Additionally, nothing should

block the path between the sound source and the microphone, so the sound level

can be recorded directly in an unmitigated form. This includes removing any smart-

phone bumpers, cases or other form of protection parts, which could interfere with

the measurement. Moreover, the height of the smartphone should be the same as

the sound source to achieve an accurate recording. For an unsophisticated mea-

surement, holding the smartphone at the general posture is sufficient. However,

depending on the aims of the measurement, a defined height for holding the smart-

phone could be predefined, e.g. at face level to examine the consequences for

hearing.

Since all measurements in this study have only been conducted under limited con-

ditions in a living room, they need to be proven under some real life conditions,

which includes a lot more external influences that might have an enormous impact

on the accuracy. This includes for example increased background noise like wind

or walking sounds, reduced sound reflection, changing distances and directions, a

covert smartphone in a pocket or accidentally covering the microphone for a short

time, which can lead to high amplitude spikes. Therefore, methods to identify and

exclude such errors need to be developed to take correct measurements during

large surveys. Some additional factors that are needed to consider when perform-

ing complex measurements will be elaborated in the next section.

The Effects of Reflection

Every calibration attempt and behavior experiment was made in a room with no

sound absorbing materials, so it is possible that the results heavily depend on the

room characteristics of the place where the experiment was conducted. This in-

cludes sound waves, which are either absorbed or reflected according to the mate-

rial they hit. Lower frequencies have a longer wavelength and spread very much, so

6 Discussion

they are only reflected by larger objects, while higher frequencies are more direc-

tional and can even bounce off from smaller objects and thus are more likely to add

to the measurement [33]. Thus, the measurement values deviate from exact results,

which would have been gained under optimal conditions. A good example for this

is demonstrated in figure 6.2, which is taken from a monodirectional measurement,

which has been taken during an early experimental stage. Here, the impacts of

reflection can be seen very clearly, since the measurement results at 5 meters dis-

tance have been recorded with a higher amplitude than the ones at 3 meters. The

reason for this abnormality is the wall of the room, since due to the diffused sound

reflection of the wall, the recorded amplitude is much higher. Thus, even though the

sound at 3 meter distance has a much smaller distance to the sound source, it is

measured at a lower amplitude, because of the impact, that reflection has on the 5

meter measurements.

500

1000

1500

2000

2500

3000

3500

4000

115

134

153

172

191

210

229

248

267

286

305

324

343

362

381

400

419

438

457

476

495

514

533

Amplitude

Time in 50 ms intervals

Figure 6.2: The impact of sound reflection is the reason why the sound is recorded

at a much higher amplitude than it has.

To counteract this effect in future calibration approaches, it is necessary to choose

a balancing value, which need to be further evaluated. This value can be estimated

based on experience or calculated as a better alternative. For this purpose, new

methods need to be developed to calculate an approximate balancing value de-

pending on the size, equipment, floor and wall materials of the room. Additional

help could be given by having the possibility to locate the current position of the

user to change the balancing value, depending on the surrounding conditions. For

6 Discussion

example, choose a lower compensation value for reflection when outdoors and a

higher one when in a small room indoors. These approaches can even include real

time balancing values, which might help to reduce the impact of reflection on the

result.

Considering Accuracy and Use Case

In any case one can assume, that for more precise results also more complex mea-

surement procedures are needed. This can either include more advanced software

and experience taken in experiments beforehand, the help of additional hardware

or even both. Although, the aim of gathering loudness data with a smartphone is

to achieve as much as precise results as possible, this does not mean to make

this task too hard to be accomplished by the average user. Thus, a careful consid-

eration between a high accuracy and a straightforward calibration and measuring

procedure needs to be found. There are several reasons for this: Firstly, a too

advanced approach leaves higher chances for the user to make errors during cali-

bration or measurements, which could result in small deviations of measurements

in the best case, or even falsify the complete result in the worst case. Secondly,

the user could be unable to carry out the calibration due to missing hardware or

additional needed equipment. Another reason for this might be incomprehensible

or too complex instructions, which deter the user to even start the calibration. Con-

sequently, the selected approach should be as easy as possible to perform.

Additionally, if microphone characteristics of a specific smartphone are known or

have been examined, it is possible to share these over the internet to let other users

benefit from it. Therefore, only one calibration per smartphone would be necessary

to perform since calibration data is the same for identical smartphone models. This

would reduce the difficulty for sound level measurements even further, if a large

database is established, since the average user would only need to deal with the

calibration, if no one else has already performed it. Hence, this would make it

possible to perform a calibration on behalf of the user, which is especially useful for

large crowd-based data collections.

6 Discussion

Noise Measuring in Context of Tinnitus Perception

When measuring noise to investigate the course of tinnitus perception, it is im-

portant to consider the sensation of the individual person, since subjective tinnitus

cannot be measured objectively [11]. Studies have already shown that the percep-

tion and annoyance of tinnitus fluctuates during the day, however, this varies from

patient to patient [14]. Hence, in order to find a way to reduce the distress, fur-

ther studies to measure the individual fluctuations in the perception of tinnitus are

needed, which can be achieved by the use of questionnaires, as they are provided

as part of the TrackYourTinnitus application [34]. The advantage of this is to further

analyze the health condition of patients by gathering relevant medical data in their

daily life, which is also referred to as Ecological Momentary Assessment (EMA). By

combining the collected data of a large number of different participants, this helps

to gain addition insights [19].

Moreover, including data about noise can additionally provide important information,

since exposure to noise can likely cause emotional stress, which in turn has nega-

tive effects on tinnitus [25] [23]. Therefore, specific types of noises that can cause

stress to the individual person need to be identified. This may include but might

not be limited to high frequency sounds, high intensity sound levels, specific noise

occurrences, which include a certain type of frequencies that cause annoyance or

a combination of all. Thus, it is necessary to examine probable causes further by

measuring and analyzing frequencies and noise levels.

The presented frequency calibration approach, in combination with a standardized

measuring method, is well suited for this type of investigation since it does not only

provide the ability to measure the sound level, but it also assures that the measure-

ment is accurately done for different types of frequency responses of smartphone

microphones. The acquired noise data can then further be used by performing a

correlation with subjective data of the patient’s questionnaire about the current tin-

nitus condition to obtain a widespread result. This method allows to do a much

more differential analysis and utilization of the collected data since the relationship

between noise and tinnitus perception can be investigated much easier and in more

detail. As a result, the frequency method serves as a valuable approach for mea-

suring noise levels for a general use, as well as to determine and modulate the

tinnitus perception with a smartphone.

7 Conclusion

Following the results of the study it can be concluded, that the complex procedure

of measuring sound levels with a smartphone is a task that can be accomplished.

With the acoustical behavior experiments it has been proven, that smartphones are

indeed capable of measuring sound correctly in relation to its distance. However,

all results should be considered with caution, due to the impact of sound reflection

and other factors.

All in all, the experiments and data do not only make it possible to convert amplitude

values into sound pressure levels, but additionally lead to a better understanding in

the way of how smartphones record sound and how changing factors like distance

or position have an influence on the recorded amplitude. This enables further appli-

cations in the field of noise measuring for personal and health usage, as measuring

sound levels (in a basic fashion) does not require more additional equipment than

a smartphone and a calibration procedure. Therefore, anyone has the possibility

to perform sound level measurements in daily life, which can provide approximate

answers to basic questions, like: Could the current sound level be harmful for my

hearing? Or: How long can I expose myself to this sound level until health risk might

occur?

An additional use case is given by collecting and analyzing information about fre-

quency and loudness of tinnitus patients in their daily life, which might result in

further knowledge about the impact of tinnitus especially in its subjective occur-

rence. This can be achieved by putting the gathered data into context with informa-

tion about the perception of the affected person to provide further improvements in

dealing with tinnitus.

A summary of the findings and an outlook to future opportunities is given below.

7 Conclusion

7.1 Summary

This work deals with different topics, which include the analysis of smartphone am-

plitude data, collecting and developing own approaches to calibrate a smartphone

microphone and performing different experiments regarding the acoustical behavior

of an internal as well as an external microphone.

Results from the basic calibration have shown promising results with regard to gen-

eral application, with a good accuracy to start with and a small effort for the calibra-

tion. Additionally, to compensate for the microphone response and to deal with dif-

ferences in loudness, the frequency-based calibration has been introduced. While

it provides a higher degree of accuracy, it requires much more effort to be accom-

plished. However, it has been shown, that it is worth to further develop the basic

calibration technique to include reference values into the calibration that account for

changing conditions and therefore achieve more precise results in these cases.

The results of the calibration have consequently led to the analysis of sound data

samples of tinnitus probands since the data could be determined by a subsequent

calibration approach.

Moreover, additional information when measuring amplitudes at different distances

and positions were gathered. This has demonstrated that smartphones are able to

show distance differences in amplitude, however, due to measurement limitations,

the accuracy could not be proven. Another important finding is that the position

to the sound source is important as well as the general posture as the time of

measuring since this is highly influential for the result. Therefore, standardization

suggestions for future measurements have been made and tested with multiple

smartphones to increase the reliability.

Together with the achieved knowledge in the course of this study, a way to efficiently

gather sound level data has been presented, which does not only benefit the aver-

age user, but could be very beneficial for tinnitus patients to monitor their health

issues as well as for other conditions.

7 Conclusion

7.2 Outlook

The results achieved in this work may serve as a starting point to measure loudness

with the help of a smartphone. As discussed before, a calibration is necessary

beforehand and can be carried out in different ways, which all pose different difficulty

and accuracy levels. Additionally, the number of overall calibrations to be performed

can be reduced drastically, if they do not need to be performed by every user, e.g.

by sharing them over the internet with regard to the type of smartphone used. This

method would need to be developed and tested.

Nevertheless, all calibration measurements have not been in use outside up to now.

Thus, more effort will be necessary to achieve good results also in daily life mea-

surements. Although the presented calibration methods are state of the art, it will be

necessary to improve and adapt them to changing indoor and outdoor conditions.

Beside the introduced and currently working methods, with changing technology

and more knowledge on this topic, there is a lot more potential to develop new ap-

proaches in the future, as the presented internal calibration shows. Hopefully, this

will give everyone in the future the possibility to measure sound levels with a smart-

phone. Moreover, it might lead to a general awareness to sound as sound levels

and sound types are factors for well-being and illness. The further use of the results

in tinnitus evaluation and treatment shows a first application in this direction.

A Source Codes

Listing A.1: Conversion of a measured amplitude value to the sound pressure level

in dB(SPL), by using a predefined amplitude and its sound pressure

level as a calibration value.

package de.uulm.dbis;

// Idea based on https://stackoverflow.com/a/14870458/3964954

// conversion of smartphone amplitudes to dB(SPL) after a

calibration was performed

,→

public class Amp2Dec {

// reference sound pressure

private final double REFERENCE = 0.00002;

// ------- defined calibration values ----------------

private final int BG_AMP = 50;

// measured amplitude of

background noise

,→

private final int X= 4000 - BG_AMP;

// measured amplitude

during calibration

,→

private final double Y= 80.0;

// corresponding dB(SPL)

// -------------------------------------------

// calculated dB(SPL)

private double db;

public Amp2Dec(int amp) {

db =amplitude_to_decibel(amp);

A Source Codes

}

private double amplitude_to_decibel(int amp) {

// calibration value, given that amplitude value X

corresponds to Y dB(SPL)

,→

double calibration_value =X/spl_to_pressure(Y);

// pressure in pascal, given that amplitude behaves

relative to the pressure

,→

double pressure =amp /calibration_value;

// dB(SPL), using the formula for the sound pressure level

return 20 * Math.log10(pressure /REFERENCE);

}

//sound pressure in pascal for a given sound level

private double spl_to_pressure(double spl) {

return Math.round(1000000000 * REFERENCE *Math.pow(10, spl

/ 20)) / 1000000000.0;

,→

}

public double getDb() {

return db;

}

A Source Codes

Listing A.2: Measure amplitude levels with an Android smartphone aiming for the

highest possible quality.

package de.uulm.dbis;

// Imports are shortened due to their large amount

public class AudioAnalyzer {

// Adapted from

// https://gist.github.com/pardom-zz/9644274 audioRecorder

implementation

,→

// https://stackoverflow.com/a/21124025 audioRecorder to wav

implementation

,→

private static int RECORDER_BPP;

// bits per sample

private static int RECORDER_CHANNELS_INT;

// 1=MONO, 2=STEREO

private final int SECONDS_TO_RUN = 40;

Thread recordingThread =null;

private boolean USE_DIRECT_REC;

// true: audioRecord false:

mediaRecorder

,→

private ScrollView scroller;

private File filename;

private TextView myText;

private Context context;

private String agc_status ="";

private String recorder_settings ="";

private File recording_file;

private File recording_file_dir;

private MediaRecorder mediaRecorder =null;

private AudioRecord audioRecorder =null;

private int mSampleRate;

private short mAudioFormat;

private short mChannelConfig;

A Source Codes

private short[] mBuffer;

private int mBufferSize =AudioRecord.ERROR_BAD_VALUE;

private int mLocks = 0;

private ArrayList<Integer>amplitude_list =new ArrayList<>();

private File recording_file_dir_temp;

public AudioAnalyzer(TextView con, ScrollView scroller, Context

context, boolean dir) {

,→

this.myText =con;

this.scroller =scroller;

this.context =context;

this.USE_DIRECT_REC =dir;

analyze();

}

public void analyze() {

Handler handler =new Handler();

Runnable runnable =() -> {

Date date =new Date();

SimpleDateFormat formatter =new

SimpleDateFormat("dd-MM-yyyy_HH-mm-ss-SSS");

,→

SimpleDateFormat formatter2 =new

SimpleDateFormat("HH-mm-ss-SSS");

,→

File path =new File(Environment.

getExternalStorageDirectory().getAbsolutePath() +

"/AudioAnalyzer");

,→

if (!(path.isDirectory())) {

if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.O)

{

,→

try {

Files.createDirectory(Paths.get(path.

getAbsolutePath()));

,→

}catch (IOException e) {

e.printStackTrace();

}

A Source Codes

}else {

path.mkdir();

}

System.out.println("AudioAnalyzer Directory

created");

,→

}else {

System.out.println("AudioAnalyzer Directory is not

created");

,→

}

filename =new File(path, "audioanalyzer_" +

formatter.format(date) +".csv");

,→

recording_file =new File(path, "recording_" +

formatter.format(date) +".aac");

,→

recording_file_dir_temp =new File(path, "recording_" +

formatter.format(date) +"_temp.raw");

,→

recording_file_dir =new File(path, "recording_" +

formatter.format(date) +".wav");

,→

initRecorder();

try {

FileWriter writer =new FileWriter(filename);

writer.append("DEBUG,");

writer.append(" RECORDING MODE: " +(USE_DIRECT_REC

?"DIRECT (AUDIO REC) " :"MEDIA REC "));

,→

writer.append(" HARDWARE INFOS: " +"ManuFacturer

:" +Build.MANUFACTURER +"Version OS: " +

Build.VERSION.BASE_OS +

,→

"Model: " +Build.MODEL +

"Product: " +Build.PRODUCT +" "

);

writer.append(" " +agc_status +" " +

recorder_settings +" ");

,→

writer.append("\n");

// 20 * seconds

for (int i= 0; i < 20 * SECONDS_TO_RUN; i++) {

int amp =getAmp();

A Source Codes

String currdate_log =formatter.format(new

Date());

,→

String currdate_disp =formatter2.format(new

Date());

,→

String str =currdate_log +"," +amp +"\n";

writer.append(str);

System.out.println("MAX AMP: " +amp);

handler.post(() -> {

myText.append(currdate_disp +", " +amp +

"\n");

,→

scroller.fullScroll(ScrollView.FOCUS_DOWN);

});

try {

Thread.sleep(50);

}catch (InterruptedException e) {

e.printStackTrace();

}

writer.flush();

writer.close();

}catch (IOException e) {

100

e.printStackTrace();

101

}

102

stopRecorder();

103

};

104

new Thread(runnable).start();

105

}

106

107

private void stopRecorder() {

108

if (USE_DIRECT_REC)

109

stopAudioRecord();

110

else {

111

mediaRecorder.stop();

112

mediaRecorder.reset();

113

mediaRecorder.release();

A Source Codes

114

}

115

}

116

117

private void initRecorder() {

118

if (USE_DIRECT_REC) {

119

initializeAudioRecord();

120

startAudioRecord();

121

}else

122

initializeMediaRecord();

123

}

124

125

private int getAmp() {

126

if (USE_DIRECT_REC)

127

return getRawAmplitude();

128

else

129

return mediaRecorder.getMaxAmplitude();

130

}

131

132

public void initializeMediaRecord() {

133

mediaRecorder =new MediaRecorder();

134

// disable any agc

135

AudioManager audioManager =(AudioManager)

context.getSystemService(Context.AUDIO_SERVICE);

,→

136

if (audioManager.getProperty(AudioManager.

PROPERTY_SUPPORT_AUDIO_SOURCE_UNPROCESSED)!= null)

{

,→

137

mediaRecorder.setAudioSource(MediaRecorder.

AudioSource.UNPROCESSED);

,→

138

agc_status += " AGC: UNPROCESSED ";

139

}else {

140

mediaRecorder.setAudioSource(MediaRecorder.

AudioSource.VOICE_RECOGNITION);

,→

141

agc_status += " AGC: VOICE RECOGNITION ";

142

}

143

// testing

A Source Codes

144

int sampr = 96000;

145

int encbitr = 384000;

146

int auch = 1;

147

// or 44100?

148

mediaRecorder.setAudioSamplingRate(sampr);

149

// or 96000?

150

mediaRecorder.setAudioEncodingBitRate(encbitr);

151

// or AMR_NB?

152

mediaRecorder.setOutputFormat(MediaRecorder.OutputFormat.

MPEG_4);

,→

153

// or AAC or AMR_NB?

154

mediaRecorder.setAudioEncoder(MediaRecorder.AudioEncoder.

HE_AAC);

,→

155

// could also be "/dev/null"

156

mediaRecorder.setOutputFile(String.

valueOf(recording_file));

,→

157

// mono or stereo?

158

mediaRecorder.setAudioChannels(auch);

159

recorder_settings += "AUDIO RECORDER

SAMPLERATE/ENCODINGBITRATE/CHANNELCONFINT/AUDIOFORMAT:

,→

160

+sampr +"/" +encbitr +"/" +auch +"/" +

"output: MPEG_4 / encoder: HE_AAC";

,→

161

try {

162

mediaRecorder.prepare();

163

}catch (IOException e) {

164

e.printStackTrace();

165

}

166

mediaRecorder.start();

167

}

168

169

public void initializeAudioRecord() {

170

if (mSampleRate > 0 && mAudioFormat > 0 && mChannelConfig >

0) {

,→

A Source Codes

171

audioRecorder =new

AudioRecord(MediaRecorder.AudioSource.MIC,

mSampleRate, mChannelConfig, mAudioFormat,

mBufferSize);

,→

172

return;

173

}

174

175

// Find best/compatible AudioRecord

176

for (int sampleRate : new int[]{48000,44100,32000,16000,

8000}) {

,→

177

for (short audioFormat : new

short[]{AudioFormat.ENCODING_PCM_16BIT,

AudioFormat.ENCODING_PCM_8BIT}) {

,→

178

for (short channelConfig : new

short[]{AudioFormat.CHANNEL_IN_MONO,

AudioFormat.CHANNEL_IN_STEREO}) {

,→

179

// Try to initialize

180

try {

181

mBufferSize =AudioRecord.

getMinBufferSize(sampleRate,

channelConfig, audioFormat);

,→

182

if (mBufferSize < 0) {

183

continue;

184

}

185

mBuffer =new short[mBufferSize];

186

audioRecorder =new

AudioRecord(MediaRecorder.AudioSource.

MIC, sampleRate, channelConfig,

audioFormat, mBufferSize);

,→

187

188

if (audioRecorder.getState() ==

AudioRecord.STATE_INITIALIZED) {

,→

189

mSampleRate =sampleRate;

190

mAudioFormat =audioFormat;

191

mChannelConfig =channelConfig;

A Source Codes

192

if (channelConfig ==

AudioFormat.CHANNEL_IN_MONO) {

,→

193

RECORDER_CHANNELS_INT = 1;

194

}else {

195

RECORDER_CHANNELS_INT = 2;

196

}

197

if (audioFormat ==

AudioFormat.ENCODING_PCM_16BIT) {

,→

198

RECORDER_BPP = 16;

199

}else {

200

RECORDER_BPP = 8;

201

}

202

System.out.println("AUDIO RECORDER

INITIALIZED");

,→

203

String rec_set ="AUDIO RECORDER

SAMPLERATE/CHANNELCONFINT/

AUDIOFORMAT/mBUFFERSIZE:

,→

204

+sampleRate +"/" +

RECORDER_CHANNELS_INT +"/"

+RECORDER_BPP +"/" +

mBufferSize;

,→

205

System.out.println(rec_set);

206

recorder_settings += rec_set;

207

return;

208

}

209

System.out.println("AUDIO RECORDER ERROR");

210

audioRecorder.release();

211

audioRecorder =null;

212

throw new IllegalStateException("Could not

init audioRecorder");

,→

213

}catch (Exception ignored) {

214

}

215

}

216

}

A Source Codes

217

}

218

}

219

220

public synchronized void startAudioRecord() {

221

// AGC DISABLE

222

if (AutomaticGainControl.isAvailable()) {

223

AutomaticGainControl agc =AutomaticGainControl.create(

224

audioRecorder.getAudioSessionId()

225

);

226

agc_status += " AGC STATUS BEFORE: " +agc.getEnabled()

+" ";

,→

227

agc.setEnabled(false);

228

agc_status += " AGC STATUS AFTER: " +agc.getEnabled()

+" ";

,→

229

}else {

230

agc_status =" NO AGC AVAILABLE ";

231

}

232

if (audioRecorder == null || audioRecorder.getState() !=

AudioRecord.STATE_INITIALIZED) {

,→

233

throw new IllegalStateException("startRecording()

called on an uninitialized AudioRecord.");

,→

234

}

235

if (mLocks == 0) {

236

audioRecorder.startRecording();

237

recordingThread =new

Thread(this::writeAudioDataToFile, "AudioRecorder

Thread");

,→

238

recordingThread.start();

239

}else {

240

throw new IllegalStateException("Could not start

audioRecorder");

,→

241

}

242

243

mLocks++;

244

}

A Source Codes

245

246

// Write the output audio in byte

247

private void writeAudioDataToFile() {

248

String filename =getTempFilename();

249

FileOutputStream os =null;

250

try {

251

os =new FileOutputStream(filename);

252

}catch (FileNotFoundException e) {

253

e.printStackTrace();

254

}

255

256

while (mLocks != 0) {

257

// gets the voice output from microphone to byte format

258

int bufferReadSize =audioRecorder.read(mBuffer, 0,

mBufferSize);

,→

259

calcRawAmplitude(mBuffer, bufferReadSize);

260

try {

261

// writes the data to file from buffer

262

// stores the voice buffer

263

// short[] shorts = new short[bytes.length/2];

264

// to turn bytes to shorts as either big endian or

little

,→

265

// endian.

266

// ByteBuffer.wrap(bytes).order(ByteOrder.

LITTLE_ENDIAN).asShortBuffer().get(shorts);

,→

267

// to turn shorts back to bytes.

268

byte[] bytes2 =new byte[mBuffer.length * 2];

269

ByteBuffer.wrap(bytes2).order(ByteOrder.

LITTLE_ENDIAN)

,→

270

.asShortBuffer().put(mBuffer);

271

os.write(bytes2);

272

// ServerInteractor.SendAudio(buffer);

273

}catch (IOException e) {

274

e.printStackTrace();

275

}

A Source Codes

276

}

277

278

try {

279

os.close();

280

}catch (IOException e) {

281

e.printStackTrace();

282

}

283

}

284

285

public synchronized void stopAudioRecord() {

286

mLocks--;

287

if (mLocks == 0) {

288

if (audioRecorder != null) {

289

audioRecorder.stop();

290

audioRecorder.release();

291

audioRecorder =null;

292

recordingThread =null;

293

// copy the recorded file to original copy & delete

the recorded copy

,→

294

copyWaveFile(getTempFilename(), getFilename());

295

deleteTempFile();

296

}

297

}

298

}

299

300

private String getFilename() {

301

return recording_file_dir.getAbsolutePath();

302

}

303

304

private void deleteTempFile() {

305

File file =new File(getTempFilename());

306

file.delete();

307

}

308

309

// stores the file into the SDCARD

A Source Codes

310

private String getTempFilename() {

311

return recording_file_dir_temp.getAbsolutePath();

312

}

313

314

private void copyWaveFile(String inFilename, String

outFilename) {

,→

315

FileInputStream in;

316

FileOutputStream out;

317

long totalAudioLen;

318

long totalDataLen;

319

int channels =RECORDER_CHANNELS_INT;

320

long byteRate =RECORDER_BPP *mSampleRate *channels / 8;

321

322

try {

323

in =new FileInputStream(inFilename);

324

out =new FileOutputStream(outFilename);

325

totalAudioLen =in.getChannel().size();

326

totalDataLen =totalAudioLen + 36;

327

328

WriteWaveFileHeader(out, totalAudioLen, totalDataLen,

329

mSampleRate, channels, byteRate);

330

byte[] bytes2 =new byte[mBuffer.length * 2];

331

ByteBuffer.wrap(bytes2).order(ByteOrder.LITTLE_ENDIAN)

332

.asShortBuffer().put(mBuffer);

333

while (in.read(bytes2) != -1) {

334

out.write(bytes2);

335

}

336

337

in.close();

338

out.close();

339

}catch (IOException e) {

340

e.printStackTrace();

341

}

342

}

343

A Source Codes

344

345

private void WriteWaveFileHeader(FileOutputStream out, long

totalAudioLen,

,→

346

long totalDataLen, long

longSampleRate, int

channels, long byteRate)

,→

347

throws IOException {

348

byte[] header =new byte[4088];

349

350

header[0] =

;

// RIFF/WAVE header

351

header[1] =

;

352

header[2] =

;

353

header[3] =

;

354

header[4] = (byte) (totalDataLen & 0xff);

355

header[5] = (byte) ((totalDataLen >> 8)& 0xff);

356

header[6] = (byte) ((totalDataLen >> 16)& 0xff);

357

header[7] = (byte) ((totalDataLen >> 24)& 0xff);

358

header[8] =

;

359

header[9] =

;

360

header[10] =

;

361

header[11] =

;

362

header[12] =

;

fmt

chunk

363

header[13] =

;

364

header[14] =

;

365

header[15] =

' '

;

366

header[16] = 16;

// 4 bytes: size of

fmt

chunk

367

header[17] = 0;

368

header[18] = 0;

369

header[19] = 0;

370

header[20] = 1;

// format = 1

371

header[21] = 0;

372

header[22] = (byte) channels;

373

header[23] = 0;

374

header[24] = (byte) (longSampleRate & 0xff);

375

header[25] = (byte) ((longSampleRate >> 8)& 0xff);

A Source Codes

376

header[26] = (byte) ((longSampleRate >> 16)& 0xff);

377

header[27] = (byte) ((longSampleRate >> 24)& 0xff);

378

header[28] = (byte) (byteRate & 0xff);

379

header[29] = (byte) ((byteRate >> 8)& 0xff);

380

header[30] = (byte) ((byteRate >> 16)& 0xff);

381

header[31] = (byte) ((byteRate >> 24)& 0xff);

382

header[32] = (byte) (channels *RECORDER_BPP / 8);

// block

align

,→

383

header[33] = 0;

384

header[34] = (byte) RECORDER_BPP;

// bits per sample

385

header[35] = 0;

386

header[36] =

;

387

header[37] =

;

388

header[38] =

;

389

header[39] =

;

390

header[40] = (byte) (totalAudioLen & 0xff);

391

header[41] = (byte) ((totalAudioLen >> 8)& 0xff);

392

header[42] = (byte) ((totalAudioLen >> 16)& 0xff);

393

header[43] = (byte) ((totalAudioLen >> 24)& 0xff);

394

395

out.write(header, 0,4088);

396

}

397

398

private void calcRawAmplitude(short[] mBuffer, int

bufferReadSize) {

,→

399

if (bufferReadSize < 0) {

400

amplitude_list.add(0);

401

return;

402

}

403

int sum = 0;

404

for (int i= 0; i <bufferReadSize; i++) {

405

sum += Math.abs(mBuffer[i]);

406

}

407

if (bufferReadSize > 0) {

408

amplitude_list.add(sum /bufferReadSize);

A Source Codes

409

return;

410

}

411

amplitude_list.add(0);

412

}

413

414

private int getRawAmplitude() {

415

if (amplitude_list.size() == 0) {

416

return 0;

417

}else

418

return amplitude_list.get(amplitude_list.size() - 1);

419

}

420

}

List of Figures

2.1 Overview of the auditory system [7] . . . . . . . . . . . . . . . . . . 7

2.2 Loudness is different for the perception of our hearing depending on

the frequency and the intensity [27]. . . . . . . . . . . . . . . . . . . 11

2.3 Comparison of different decibel values and their effects on hearing

[39]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.4 The A-weighting filter curve [36]. . . . . . . . . . . . . . . . . . . . . 13

2.5 A sound level meter with a pop filter . . . . . . . . . . . . . . . . . . 14

2.6 The functionality of a sound level meter, based on Kumar (2018) . . . 15

2.7 The process of converting an acoustical signal into a measurement

value that can be used by a smartphone application, based on Faber

(2017) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.1 Most of the sound waves are reflected and mitigated before they are

recorded by the smartphone. . . . . . . . . . . . . . . . . . . . . . . 21

4.1 The application "Spaichinger Schallanalysator" in measurement mode,

displaying sound pressure level, base frequency and corresponding

tone [40] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

4.2 Execution of the calibration by tearing apart paper . . . . . . . . . . 25

4.3 Measuring the sound level of different frequencies after calibrating

the smartphone with the paper tearing technique . . . . . . . . . . . 26

4.4 A frequency response curve of a microphone. The output varies de-

pending on the frequency. [6] . . . . . . . . . . . . . . . . . . . . . . 27

4.5 Performing a frequency calibration by the help of a sound level meter 29

4.6 Schematic of a calibration approach with a pure digital-to-digital in-

terface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4.7 Characteristics of the violin sample . . . . . . . . . . . . . . . . . . . 33

4.8 Characteristics of the traffic sample . . . . . . . . . . . . . . . . . . 34

List of Figures

4.9 Characteristics of the airliner sample . . . . . . . . . . . . . . . . . . 35

4.10 Characteristics of the helicopter sample . . . . . . . . . . . . . . . . 36

4.11 Different amplitude levels can be measured, depending on the se-

lected recorder mode. . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4.12 Screenshots of the self-developed application to measure amplitudes 39

5.1 Comparison of smartphone models recording at varying amplitude

ranges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

5.2 Measuring sound levels while tearing apart five sheets of paper to

calculate the deviations of the peaks . . . . . . . . . . . . . . . . . . 46

5.3 Measured sound levels of different frequencies on a smartphone and

on the SLM after having performed a smartphone calibration . . . . . 47

5.4 Average deviation of sound levels measured with a smartphone in

regard to the SLM at different frequencies (without 100 Hz). . . . . . 48

5.5 Measured amplitudes for the frequency calibration with indicated stan-

dard deviations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

5.6 The calibration factors for the Galaxy S4 (right axis) are much higher

and show a greater deviation rate than the ones of the Oneplus and

the Galaxy S7 (left axis). . . . . . . . . . . . . . . . . . . . . . . . . 51

5.7 Measuring volumes of different frequencies on two different Samsung

Galaxy models with external speaker. . . . . . . . . . . . . . . . . . 53

5.8 Moving average of amplitudes for frequencies 8 hz to 2048 hz of two

smartphone variants . . . . . . . . . . . . . . . . . . . . . . . . . . 55

5.9 A histogram of the converted sound level values for the Galaxy S7

data samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

5.10 A histogram with the number of data samples for multiple Android

smartphones of the same amplitude range. . . . . . . . . . . . . . . 59

5.11 A histogram with the number of data samples and their correspond-

ing sound level values for multiple Android smartphones of the same

amplitude range. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

5.12 Measuring amplitudes with increasing distance with the aircraft take-

off sample on the Galaxy S7 (left) and the Oneplus (right) . . . . . . 61

5.13 Measuring amplitudes with increasing distance with the helicopter

sound sample on the Galaxy S7 (left) and the Oneplus (right) . . . . 62

100

List of Figures

5.14 No statements about the distance of 2 m and 3 m can be made for

the violin sample across all devices . . . . . . . . . . . . . . . . . . 63

5.15 Comparison of amplitudes with increasing distance by using a Sam-

sung Smartphone (left) in combination with an external microphone

(right) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5.16 Comparison between proportional signal reduction and measured

values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

5.17 Measuring the amplitude with the internal microphone of the Oneplus

(left) and an external microphone (right) at different positions to the

sound source. The use of an external microphone leads to lower

deviations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

6.1 A 3D example for the short-time Fourier transform, which includes

frequency (x-axis), colored intensity (y-axis) and time (z-axis) [38]. . . 70

6.2 The impact of sound reflection is the reason why the sound is recorded

at a much higher amplitude than it has. . . . . . . . . . . . . . . . . 76

101

List of Listings

A.1 Conversion of a measured amplitude value to the sound pressure

level in dB(SPL), by using a predefined amplitude and its sound pres-

sure level as a calibration value. . . . . . . . . . . . . . . . . . . . . 82

A.2 Measure amplitude levels with an Android smartphone aiming for the

highest possible quality. . . . . . . . . . . . . . . . . . . . . . . . . . 84

102

Bibliography

[1] Peter Alberti. “Noise induced hearing loss”. In: BMJ (Clinical research ed.)

304 (Mar. 1992), p. 522. DOI:

10.1136/bmj.304.6826.522

[2] American Tinnitus Association. Understanding the Facts. 2020. URL:

https:

//www.ata.org/understanding-facts

[3] David Baguley, Don McFerran, and Deborah Hall. “Tinnitus”. In: The Lancet

382.9904 (2013), pp. 1600–1607. ISSN: 0140-6736. DOI:

10.1016/S0140-

6736(13)60142-7

.URL:

https://doi.org/10.1016/S0140-6736(13)

60142-7

[4] B. Benward and M.N. Saker. Music in Theory and Practice. Music in Theory

and Practice Vol. 1. McGraw-Hill, 2003. ISBN: 9780072942620.

[5] Jennifer Bizley and Yale Cohen. “The what, where and how of auditory-object

perception”. In: Nature reviews. Neuroscience 14 (Oct. 2013), pp. 693–707.

DOI:

10.1038/nrn3565

[6] Hristo Bojinov et al. “Mobile Device Identification via Sensor Fingerprinting”.

In: CoRR abs/1408.1416 (2014). arXiv:

1408.1416

.URL:

http://arxiv.

org/abs/1408.1416

[7] Wikimedia Commons. Anatomy of Human Ear with Cochlear Frequency Map-

ping. 2009. URL:

https://commons.wikimedia.org/wiki/File:Anatomy_

of_Human_Ear_with_Cochlear_Frequency_Mapping.svg

[8] Norman Erber. Auditory Training. Jan. 1982. ISBN: 0-88200-149-3.

[9] Benjamin M. Faber. “Acoustical Measurements with Smartphones : Possibili-

ties and Limitations”. In: 2017, pp. 10–17.

[10] Klaus Genuit. “Status of psychoacoustics in noise analysis”. In: Journal of

the Acoustical Society of America 137 (Apr. 2015), pp. 2291–2291. DOI:

10.

1121/1.4920362

103

Bibliography

[11] Haúla Faruk Haider et al. “Pathophysiology of Subjective Tinnitus: Triggers

and Maintenance”. In: Frontiers in Neuroscience 12 (2018), p. 866. ISSN:

1662-453X. DOI:

10.3389/fnins.2018.00866

.URL:

https://www.fron

tiersin.org/article/10.3389/fnins.2018.00866

[12] Byung Han et al. “Tinnitus: Characteristics, Causes, Mechanisms, and Treat-

ments”. In: Journal of clinical neurology (Seoul, Korea) 5 (Apr. 2009), pp. 11–

9. DOI:

10.3988/jcn.2009.5.1.11

[13] Scott H. Hawley and Robert E. McClain. “Visualizing Sound Directivity via

Smartphone Sensors”. In: The Physics Teacher 56.2 (2018), pp. 72–74. DOI:

10.1119/1.5021430

. eprint:

https://doi.org/10.1119/1.5021430

.URL:

https://doi.org/10.1119/1.5021430

[14] James Henry et al. “Pilot Study to Evaluate Ecological Momentary Assess-

ment of Tinnitus”. In: Ear and hearing 33 (Sept. 2011), pp. 179–290. DOI:

10.1097/AUD.0b013e31822f6740

[15] IEC. IEC 61672-1:2013 Electroacoustics - Sound level meters - Part 1: Spec-

ifications. 2013.

[16] Pawel Jastreboff. “Phantom Auditory Perception (Tinnitus): Mechanisms of

Generation and Perception”. In: Neuroscience research 8 (Sept. 1990), pp. 221–

54. DOI:

10.1016/0168-0102(90)90031-9

[17] Chucri A. Kardous and Peter B. Shaw. “Evaluation of smartphone sound mea-

surement applications”. In: The Journal of the Acoustical Society of Amer-

ica 135.4 (2014), EL186–EL192. DOI:

10.1121/1.4865269

. eprint:

https:

//doi.org/10.1121/1.4865269

.URL:

https://doi.org/10.1121/1.

4865269

[18] Chucri Kardous, Peter Shaw, and William Murphy. “Evaluation of smartphone

sound measurement applications using external microphones - A follow-up

study”. In: vol. 139. May 2016, pp. 2036–2036. DOI:

10.1121/1.4950017

[19] Robin Kraft et al. “Combining Mobile Crowdsensing and Ecological Momen-

tary Assessments in the Healthcare Domain”. In: Frontiers in Neuroscience

14 (2020), p. 164. ISSN: 1662-453X. DOI:

10.3389/fnins.2020.00164

.URL:

https://www.frontiersin.org/article/10.3389/fnins.2020.00164

104

Bibliography

[20] Robin Kraft et al. “Comprehensive insights into the TrackYourTinnitus database”.

In: Procedia Computer Science 175 (2020). The 17th International Confer-

ence on Mobile Systems and Pervasive Computing (MobiSPC),The 15th In-

ternational Conference on Future Networks and Communications (FNC),The

10th International Conference on Sustainable Energy Information Technol-

ogy, pp. 28 –35. ISSN: 1877-0509. DOI:

https://doi.org/10.1016/j.

procs.2020.07.008

.URL:

http://www.sciencedirect.com/science/

article/pii/S1877050920316872

[21] Robin Kraft et al. “Design and Implementation of a Scalable Crowdsensing

Platform for Geospatial Data of Tinnitus Patients”. In: 2019 IEEE 32nd Inter-

national Symposium on Computer-Based Medical Systems (CBMS). 2019,

pp. 294–299.

[22] Robin Kraft et al. “Efficient Processing of Geospatial mHealth Data Using

a Scalable Crowdsensing Platform”. In: Sensors 20.12 (2020). ISSN: 1424-

8220. URL:

https://www.mdpi.com/1424-8220/20/12/3456

[23] Kari Suzanne Kraus and Barbara Canlon. “Neuronal connectivity and inter-

actions between the auditory and limbic systems. Effects of noise and tin-

nitus”. In: Hearing Research 288.1 (2012). Annual Reviews 2012, pp. 34 –

46. ISSN: 0378-5955. DOI:

https://doi.org/10.1016/j.heares.2012.

02.009

.URL:

http://www.sciencedirect.com/science/article/pii/

S0378595512000457

[24] Dilip Kumar and Deepak Kumar. “Chapter 13 - Abatement of Noise Pollution”.

In: Sustainable Management of Coal Preparation. Ed. by Dilip Kumar and

Deepak Kumar. Woodhead Publishing, 2018, pp. 279 –291. ISBN: 978-0-12-

812632-5. DOI:

https://doi.org/10.1016/B978-0-12-812632-5.

00013-6

.URL:

http://www.sciencedirect.com/science/article/pii/

B9780128126325000136

[25] Berthold Langguth. “A review of tinnitus symptoms beyond ‘ringing in the

ears’: a call to action”. In: Current Medical Research and Opinion 27.8 (2011).

PMID: 21699365, pp. 1635–1643. DOI:

10.1185/03007995.2011.595781

eprint:

https://doi.org/10.1185/03007995.2011.595781

.URL:

https:

//doi.org/10.1185/03007995.2011.595781

105

Bibliography

[26] Mark Lutman. “What Is the Risk of Noise-Induced Hearing Loss at 80, 85,

90 dB(A) and Above?” In: Occupational medicine (Oxford, England) 50 (June

2000), pp. 274–5. DOI:

10.1093/occmed/50.4.274

[27] Meinard Müller. Fundamentals of Music Processing. Springer Verlag, 2015.

ISBN: 978-3-319-21944-8.

[28] Enda Murphy and Eoin A. King. “Testing the accuracy of smartphones and

sound level meter applications for measuring environmental noise”. In: Ap-

plied Acoustics 106 (2016), pp. 16 –22. ISSN: 0003-682X. DOI:

https://

doi.org /10.1016/j.apacoust. 2015.12.012

.URL:

http:/ /www.

sciencedirect.com/science/article/pii/S0003682X15003667

[29] Casey O’Callaghan. “Auditory Perception”. In: The Stanford Encyclopedia of

Philosophy. Ed. by Edward N. Zalta. Summer 2020. Metaphysics Research

Lab, Stanford University, 2020.

[30] R. Pierre and D. Maguire. “The Impact of A-weighting Sound Pressure Level

Measurements during the Evaluation of Noise Exposure”. In: NOISE-CON.

2004.

[31] Christopher J. Plack. The Sense of Hearing - Second Edition. London: Psy-

chology Press, 2013. ISBN: 978-1-134-71668-5.

[32] Kelly Reavis, Janice Chang, and Fan-Gang Zeng. “Patterned sound therapy

for the treatment of tinnitus”. In: The Hearing Journal 63 (Nov. 2010), pp. 21–

22,24. DOI:

10.1097/01.HJ.0000390817.79500.ed

[33] Jens Rindel. “Book Review: Room Acoustics, sixth edition, by Heinrich Kut-

truff”. In: Aug. 2017. DOI:

10.13140/RG.2.2.28652.46724

[34] Winfried Schlee et al. “Measuring the Moment-to-Moment Variability of Tin-

nitus: The TrackYourTinnitus Smart Phone App”. In: Frontiers in Aging Neu-

roscience 8 (2016), p. 294. ISSN: 1663-4365. DOI:

10.3389/fnagi.2016.

00294

.URL:

https://www.frontiersin.org/article/10.3389/fnagi.

2016.00294

[35] Ervin Sejdi´

c, Igor Djurovi´

c, and Jin Jiang. “Time–frequency feature repre-

sentation using energy concentration: An overview of recent advances”. In:

Digital Signal Processing 19.1 (2009), pp. 153 –183. ISSN: 1051-2004. DOI:

https://doi.org/10.1016/j.dsp.2007.12.004

.URL:

http://www.

sciencedirect.com/science/article/pii/S105120040800002X

106

Bibliography

[36] Eberhard Sengpiel. Sengpiel Audio Berlin.URL:

http://www.sengpielaudi

o.com/

[37] Josef Shargorodsky, Gary Curhan, and Wildon Farwell. “Prevalence and Char-

acteristics of Tinnitus among US Adults”. In: The American journal of medicine

123 (Aug. 2010), pp. 711–8. DOI:

10.1016/j.amjmed.2010.02.015

[38] Ethan Weil. Short time fourier transform. 2006. URL:

https://upload.wiki

media.org/wikipedia/commons/9/9b/Short_time_fourier_transform.

PNG

[39] Xueqing Zhang. “Noise Prediction of a NACA 0015 Airfoil with Gurney Flap: A

study based on Time-Resolved Particle Image Velocimetry”. PhD thesis. Jan.

2016.

[40] Dr. Markus Ziegler. Bedienungsanleitung Schallanalyator. Available at

https

://spaichinger-schallpegelmesser.de/Anleitung_Schallanalysator

.pdf

, version 2.2. Sept. 2019.

[41] Dr. Markus Ziegler. Spaichinger Schallpegelmesser. Available at

https://

spaichinger-schallpegelmesser.de/Anleitung_Spaichinger_Schallp

egelmesser.pdf

, version 4.3. Dec. 2019.

[42] E. Zwicker. Psychoakustik. Hochschultext. Springer Berlin Heidelberg, 2013.

ISBN: 9783642685101. URL:

https://books.google.de/books?id=xXOiBg

AAQBAJ

107

Name: Johannes Vedder Matriculation Number: 691733

Declaration

I hereby confirm that I created this work on my own and that I have not used any

other materials than the ones referenced to in this thesis.

Ulm, .............................................................................

Johannes Vedder

13.10.2020