Citation
A forensic investigation of the electrical properties of digital audio recording

Material Information

Title:
A forensic investigation of the electrical properties of digital audio recording
Creator:
Leroi, Jack ( author )
Place of Publication:
Denver, CO
Publisher:
University of Colorado Denver
Publication Date:
Language:
English
Physical Description:
1 electronic file (71 pages). : ;

Subjects

Subjects / Keywords:
Signal processing -- Digital techniques ( lcsh )
Sound -- Recording and reproducing -- Digital techniques ( lcsh )
Sound -- Recording and reproducing -- Equipment and supplies ( lcsh )
Genre:
bibliography ( marcgt )
theses ( marcgt )
non-fiction ( marcgt )

Notes

Review:
In media forensics, the devices; e.g. computers, smart phones, still/video cameras, audio recorders, and software; e.g. video, audio, and graphics editors, file and disk utilities, mathematical computation applications, are, for the most part, black boxes. The design specifications are usually proprietary and the operating specifications may be incomplete, inaccurate, or unavailable. This makes it difficult to validate the technology, but using it without validation could discredit a practitioner's findings or testimony. The alternative is to test the device or program to determine relevant characteristics of its performance. An important and common device in media forensics is the portable digital audio recorder used to record surveillance and interviews. This type can also be used to record the alternating current (AC) waveform from the mains power. While small variations in the AC frequency (ENF) can be forensically important, distortion in the recording can affect its value in adjudication or investigation. A method is presented to evaluate aspects of a recorder's operation that can cause distortion. Specifically, the method measures the noise generated by the recorder's electronics in its input and amplifier circuits. The method includes a procedure to isolate the recorder from environmental sources of noise. The method analyzes the broadband noise floor produced by the range of recording conditions and recorder settings. It also analyzes the noise amplitude for the harmonics for the mains frequency.
Thesis:
Thesis (M.S.)--University of Colorado Denver. Recording arts and media forensics
Bibliography:
Includes bibliographic references.
System Details:
System requirements.
General Note:
College of Arts and Media
Statement of Responsibility:
by Jack Leroi.

Record Information

Source Institution:
University of Colorado Denver
Holding Location:
|Auraria Library
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
900788248 ( OCLC )
ocn900788248

Downloads

This item has the following downloads:


Full Text
A FORENSIC INVESTIGATION OF THE ELECTRICAL PROPERTIES OF
DIGITAL AUDIO RECORDING
by
JACK LEROI
B.S., Metropolitan State College, 2006
A thesis submitted to the
Faculty of the Graduate School of the
University of Colorado in partial fulfillment
of the requirements for the degree of
Master of Science
Recording Arts Media Forensics
2014


2014
JACK LEROI
ALL RIGHTS RESERVED


This thesis for the Master of Science degree by
Jack LeRoi
has been approved for the
Recording Arts Program
by
Catalin Grigoras, Chair
Jeff Smith
Lome Bregitzer
Date: May 2, 2014


LeRoi, Jack (M.S., Recording Arts Media Forensics)
A Forensic Investigation of the Electrical Properties of Digital Audio Recording
Thesis directed by Associate Professor Catalin Grigoras.
ABSTRACT
In media forensics, the devices; e.g. computers, smart phones, still/video
cameras, audio recorders, and software; e.g. video, audio, and graphics editors,
file and disk utilities, mathematical computation applications, are, for the most
part, black boxes. The design specifications are usually proprietary and the
operating specifications may be incomplete, inaccurate, or unavailable. This
makes it difficult to validate the technology, but using it without validation could
discredit a practitioners findings or testimony. The alternative is to test the
device or program to determine relevant characteristics of its performance.
An important and common device in media forensics is the portable digital
audio recorder used to record surveillance and interviews. This type can also be
used to record the alternating current (AC) waveform from the mains power.
While small variations in the AC frequency (ENF) can be forensically important,
distortion in the recording can affect its value in adjudication or investigation. A
method is presented to evaluate aspects of a recorders operation that can cause
distortion. Specifically, the method measures the noise generated by the
recorders electronics in its input and amplifier circuits. The method includes a
procedure to isolate the recorder from environmental sources of noise. The
method analyzes the broadband noise floor produced by the range of recording


conditions and recorder settings. It also analyzes the noise amplitude for the
harmonics for the mains frequency.
The form and content of this abstract are approved. I recommend its
publication.
Approved: Catalin Grigoras
IV


TABLE OF CONTENTS
CHAPTER
I. INTRODUCTION...................................1
II. RATIONAL.......................................6
III. PROCEDURES.....................................8
IV. RESULTS.......................................17
V. ANALYSIS......................................26
VI. CONCLUSION....................................30
REFERENCES............................................32
APPENDIX
A. ADDITIONAL NOISE LEVEL GRAPHS.................34
B. ADDITIONAL HARMONIC LEVEL GRAPHS..............40
C. EQUATIONS.....................................65
v


CHAPTER I
INTRODUCTION
In a popular anecdote1, the 3rd century BCE Greek scientist and
mathematician Archimedes of Syracuse was asked by King Hieron II2 to
investigate whether the smith who had made the Kings crown had substituted
silver for some of the gold he was given. Archimedes was not allowed to melt
the crown to compare its volume to that of the same weight in pure gold. He had
to find a way, unknown at the time, to determine the volume of an irregularly
shaped object. The solution came to him in his bath, as he observed the amount
of water that his body displaced. Whether or not Archimedes then ran naked
through the streets shouting Eureka, Archimedes principle explains the
relationships among weight, density, and volume and is an important scientific
discovery3.
However, while Archimedes discovery is remarkable, in a larger sense, it is
not exceptional. The observation and analysis of the physical world is a primary
pursuit of Homo sapiens. From the earliest times, our survival, as creatures
without fang or claw, depended on comprehending and manipulating nature.
This ability of early humans to know and use the available materials, e.g. wood,
stone, animal tissues, to understand weather and the plants and animals in their
environment, was indispensable in the growth of culture and civilization.
1


Given these characteristics, and a human fossil record of 250,000 years4,
modern science is a curiously late development. One can only speculate over
the factors that prevented its earlier acceptance, but, paradoxically, the human
need for explanation is a likely candidate. The naivete of ancient peoples about
causality left gaps in their phenomenology. In the absence of natural reasons for
the world and its phenomena, they filled the gaps with their imagination. In their
imagining, the world was animated by deities, spirits, and other intentional, but
invisible beings.
Scientific methods depend on objective assessments of observable
phenomena to produce natural explanations of those phenomena. Proof may be
offered for natural explanations and such proof may either be accepted or
rejected. Supernatural explanations are beyond proof and must either be
believed or denied. The two types of explanations are incompatible. It cannot be
that lightning is both a plasma discharge of differences in electrical potential and
a thunderbolt thrown by Zeus from Mount Olympus. There cannot even be an
intermediate explanation.
Supernatural explanations are appealing and emotionally satisfying in a way
that natural explanations often are not. They are simple and inclusive and
certain while natural explanations are complex, partial, provisional. They meet
the human need for quick answers and risk avoidance that were inherited from
our prey ancestors.
Science and technology are today so embedded in the physical and social
culture of global civilization, that they supply the explanations and provisions that
2


formerly came from religion. They are so pervasive, it is impossible to
experience the world the way it was before they existed, but, absent an
alternative, rejecting supernatural explanations was likely very difficult. However,
once agriculture and writing developed, people had the time and tools to
investigate other possible reasons for the visible world. This fostered the ability
to reason and laid the foundation for what later became the scientific method.
The curiosity that drives scientific exploration is an inherent human
characteristic but the ways of thought it requires are not natural. Even those who
have no supernatural beliefs are not therefore automatically scientific. The
hardwired tendency, even need, to make quick decisions leaves us prone to
logical fallacies and other mental errors. These errors are often embedded in
heuristic methods like common sense. While common sense is often useful,
the innate errors produce biases of various types. It is important to stress that
these fallacies are not moral or intellectual failures but natural consequences of
our neurology5. With this understanding, it is possible to mitigate the harmful
effects of such fallacies.
It is instructive to consider the effect this tendency has had on forensic
science. For many years, courts, law enforcement organizations, and even
forensic practitioners themselves accepted the premise that every fingerprint was
unique. This assumption was based on the experience of fingerprint examiners.
Since identical fingerprints were never found, this uniqueness proposition was
not disproved; therefore it was assumed that the lack of disproof amounted to
3


proof. (This is the Argumentum ad Ignorantiam6; i.e., whatever is not false
must be true, whatever is not true must be false.)
Yet there is no known mechanism that would prevent prints from different
fingers, either from the same or different persons, from being the same. As a
result, it is impossible to use fingerprint evidence for positive identification7.
When Brandon Mayfield was arrested in 2004 for the Madrid bombings
based on fingermarks that closely matched Ouhnane Daoud, it seemed that the
uniqueness proposition was disproved. In fact, law enforcement had made an
inaccurate identification. However, the issues raised by this case fit with the
current movement in forensics towards scientific methods and away from
heuristics. This trend was accelerated with the publication of the NAS report in
2009 that directly addresses the challenges in media forensics.
The developments in electronic technology in the last 100 years have
produced a complex, bewildering, and ever changing technical environment.
Since the invention of digital electronics, the worldwide saturation of millions of
miniature recording devices has produced an overwhelming flood of image,
video, and audio files. These have strained the ability of forensic practitioners to
respond adequately to the demand. Given the limits on time, money, and
personnel, it is tempting for them to accept categorically that their tools are
adequate to the task. However, as with the uniqueness proposition, such
acceptance is unwise, because, as with most end users, they typically have little
or no knowledge about the inner workings of the technology on which they rely.
4


In particular, the operating characteristics of the hardware and software
used in media forensics can be difficult to determine. Manufacturers
specifications may be missing or incomplete and the accuracy of the available
information is not guaranteed. Failing to allow for this deficiency could discredit a
practitioners findings or testimony.
The better approach is to rigorously test technology to determine actual
operation. These tests may cover the overall performance of a device or
program, or may be focused on a specific factor or set of factors related to a
particular application of the technology. Ideally, the results will show that the
technology performs adequately for the intended task. If not, it must be rejected
for that particular purpose, though it may yet be useful for others.
Guidelines for evaluating tools, techniques, and procedures are found in the
SWGDE Recommended Guidelines for Validation Testing, published by the
Scientific Working Group on Digital Evidence.
5


CHAPTER II
RATIONAL
One of the applications of technology that is significant in media forensics is
the capture and analysis of the fluctuations in the frequency of the electrical grid.
For North America, the standard frequency is 60 Hz but this fluctuates as the
aggregate load on the grid varies. As the load increases, the frequency sags
until more power is added to the grid. As the load decreases, the frequency rises
until power is removed from the grid. Since the load is constantly changing, the
frequency is also constantly changing, though the power producers are required
to hold the nominal frequency to 60 Hz 0.02 Hz8. (The actual frequency may
vary up to 0.05 Hz before the system is considered over limit9.) Because the
variations are non-cyclical (non-deterministic), the actual frequencies that occur
between any two moments cannot be predicted or modeled. Thus, the sequence
of frequencies between those two moments can provide a timestamp that
indicates when that sequence of frequencies occurred.
The variations in the Electric Network Frequency (ENF) can be acquired
unintentionally during the recording of audio signals10. This occurs in a number
of ways. For one, when an electromagnetic field is generated by the electric
current flowing in the grid, the field can induce an ENF signal into the recorder's
microphones or into its audio amplifier circuits. This signal is combined with the
intended audio signal. For another, the recorder can receive an ENF signal
6


through an external (mains) power supply. If the 60 Hz AC from the mains is not
adequately filtered from the DC power supply, an ENF signal can be induced in
the recorder's audio amplifier circuits. Again, this signal is combined with the
intended audio signal.
The variations in the Electric Network Frequency (ENF) can also be
acquired deliberately. In order to use a recording of an unknown sequence of
ENF variations as a timestamp, there must be a known sequence of ENF
variations to compare it to. To this end, databases of the ENF are acquired and
maintained for each electrical grid11. The house voltage from the mains is
converted to audio levels and applied to the audio inputs of the recording device
used for the database.
The utility of this method depends on the accuracy and precision with which
the ENF is recorded. The discrepancy between the original signal and a
recording of it are classified as acquisition errors. These errors are generated at
different stages in the circuitry of the recording device; the way and extent to
which these errors affect the recorded ENF signal determine whether it can be
used to validate the time of a recording. This paper describes some of the
factors of digital recording that can lead to these acquisition errors.
7


CHAPTER III
PROCEDURES
The initial investigation of a recording device requires the evaluation of its
noise characteristics. This evaluation should measure the random fluctuations in
electrical potential produced in the input and processing circuits of the device.
The factors contributing to the noise figure are internal sources of noise as well
as noise that arises from the environment in which a recording is made.
Internal sources include thermal noise, shot noise, and flicker noise. These
are generated by the recorder's electronic components themselves12.
Environmental sources include atmospheric13 and industrial noise14 These
are generated by lightning and by discharges of electrical energy by machinery
and electrical devices. The sudden change in electrical potential generates
electro-magnetic fields which induce noise in the recorder's electronic
components.
Fluctuations in electrical potential can also be induced in the device's
components by other environmental phenomena. The most significant
phenomena are the electro-magnetic fields produced by current flowing in the
electric power network. The fluctuations in these fields are not random. The
variations in strength are synchronous with the amplitude and frequency of the
sine wave driving the electrical grid. The fluctuations may be recorded as hum at
the electric network frequency, 60 hertz for North America.
8


In a digital recorder, there is also noise added to the signal by the Analog-
to-Digital conversion. This noise is caused by quantization error in the ADC and
jitter in the sampling clock.
To properly evaluate a recorder's noise figure, it is necessary to eliminate all
influence by environmental sources. This may be accomplished in one of two
ways. Either the recorder can be located in an area where these fields are
absent or the recorder can be shielded from whatever fields are present. It is
possible to eliminate industrial noise and hum by locating the recorder away from
all electric power system components, machinery, and electrical devices, but this
will not eliminate atmospheric noise. However, the use of proper shielding will
reduce all ambient fields to insignificant levels.
The procedures for this paper were developed using the facilities of the
National Center for Media Forensics. The equipment included a Stanford
Research Systems SR770 FFT spectrum analyzer, a Ramsey Electronics
STE3000FAV faraday shielded enclosure, and an Olympus DM-520 battery
powered portable digital audio recorder. The SR770 contains a low distortion
sine wave generator.
Since the STE3000FAV is designed to block cell phone signals in the 20
megahertz to 8 gigahertz range, it was necessary to validate the enclosures
performance in the audio range, particularly at the electric network frequency of
60 hertz. The complication in performing this validation was the lack of a
calibrated test instrument. As a substitute, a portable recorder was used to
receive any electro-magnetic energy in the audio range that leaked through the
9


shielding. However, this was bootstrapping because the very object of these
procedures was to evaluate the recorder. The workaround was finally
accomplished by adding an initial step. In this first step, the recorder was used in
an environment that was free of electro-magnetic fields. (Note that it was
impossible to accomplish the entire procedure in this environment because of the
need for mains power for the SR770.) The field free recordings were analyzed
for periodic phenomena. Due to the absence of external electro-magnetic fields,
any phenomena must have been generated by the recorder itself. These results
were then used to calibrate the recordings made with the DM-520 placed in the
STE3000FAV in the lab.
Recordings were made with the DM-520 in the STE3000FAV, with the lid
closed and latched. This provided the maximum electro-magnetic isolation.
However, in this configuration there is no way to apply an external signal to the
DM-520. Leaving the lid unlatched and slightly open allowed connecting a thin
audio cable between the SR770 signal generator and the DM-520, but this spoils
the integrity of the enclosure. It was necessary to determine whether recordings
made with the configuration showed increased levels of environmental noise. To
this end, recordings were made with an audio cable connected to the signal
generator output of the SR770. It was run underneath the lid of the
STE3000FAV and the audio plug at the end of the cable was placed near the
input jack of the DM-520, but not connected to it.
The DM-520 has 16 recording levels, 3 input gain levels, and 2 sampling
rates, resulting in 96 unique settings. There were also 2 recording environments,
10


field and lab. The field tests had 1 input state, shorted. The lab tests had 2 input
states, shorted and open. The lab test also had 2 recording conditions, shielded
and partially shielded. Testing all of these combinations would require 480
recordings; however, evaluating the DM-520s noise performance only required
testing a representative sample of these settings. The recording levels were
reduced to half by selecting every other one, resulting in 240 combinations for
testing. The reduction halved the work effort required to collect and process the
recordings. Practically, the results for the missing combinations can be
extrapolated from these 240 combinations.
The input circuits of an audio recorder amplify the electrical signals from a
microphone or external audio source. These circuits produce internal noise from
the activity of the electrons that flow through them. This activity can be
influenced by the impedance of the microphone or external audio source. By
testing the DM520 with the input shorted and with the input open, the impedance
applied to the input circuits is either 0 ohms or infinite ohms. These two states
provide the bounds for the impedance of anything connected to the recorder.
The output impedance of the SR770 signal generator is specified as less than 5
ohms, so the self generated noise of the DM520s input circuits with the SR770
connected is near the results obtained with a shorted input.
The device to encode a continuously variable voltage (analog) to a series of
fixed binary values (digital) is called an Analog to Digital convertor. Such a
convertor is at the heart of every digital audio recorder. To complete a
conversion, the changing analog voltage must be fixed at a static level long
11


enough for it to be encoded into a digital value. This is done by a sample and
hold circuit that freezes the voltage until the conversion is complete. The time for
this conversion must be sufficiently short to fit within the time window between
samples. The time to charge the sample and hold circuit must also fit in the
window before the conversion time.
In a digital audio recorder, noise is also generated in the Analog to Digital
Convertor (ADC). One source of this ADC noise is non-linearity in its transition
levels. Figure 3.1 shows the relationship between an analog input and the digital
output for a 3 bit ADC.
Analog Voltage Input
Figure 3.1
As the voltage increases or decreases, the code changes when the voltage
crosses the transition point between one code and the one adjacent. If the
voltage difference between transition points varies, then the response of the ADC
is non-linear. This non-linearity produces noise in the digital output.
12


It is possible to measure the non-linearity of an ADC by applying a
waveform of known probability density to its voltage input and analyzing the code
output15. The voltage range of the input signal should cover the transition points
of interest. During the design and testing of ADC components, the voltage range
of the signal covers the entire input range of the ADC. For the purposes of this
investigation, the transition points of special interest are around the 0 voltage
point. Any non-linearity in this area can affect the accuracy and precision of zero
crossing measurements of the ENF. This is of particular importance in using the
recordings in ENF databases.
Three common waveforms used for testing are the saw tooth, triangle, and
sine. The saw tooth and triangle waveforms are the easiest to compute from
since they are theoretically linear, that is they have the same change in voltage
per unit time. Thus they have an equal probability for each transition point.
Flowever, it is difficult to generate these as truly linear and the equipment to
generate them is uncommon. Flowever, high quality sine wave generators are
readily available and, while the probability density computations are more
complicated than those for linear waveforms, this is the standard approach.
For an ideal ADC, the distribution of digital code values matches the
distribution of voltage values in the input waveform. Any non-linearity shows as
differences in the distribution. Flowever, using probability to measure non-
linearity requires the acquisition of large data sets. This is because the infinite
voltage values in the input have to be captured in the finite number of code
output values. (These code values are also sometimes called buckets.) In
13


addition, any random noise in the input will be reduced, as large data sets
average the input variations.
Field Noise Recording Procedure
Settings.
Recorder: Olympus DM-520
Serial Number: 100104915
Bit Depth: 16
Sampling Rate: 44.1 kHz, 48 kHz
Input: External Microphone
Input State: Shorted
Microphone Sensitivity: Low/Medium/High
Recording Levels: 2, 4, 6, 8, 10, 12, 14, 16
Step 1.
Insert the shorted plug into the external microphone jack.
Step 2.
Record samples approximately 36 seconds long.
@ 2 sampling rates
@ 3 microphone sensitivities
@ 8 recording levels
@ 1 input state
48 Samples Total
Step 3.
Trim the sample files to 30 seconds.
14


Save the trimmed samples to wave files.
Step 4.
Plot the waveform for each wave file created in Step 3.
Save the graph for each waveform plot.
Step 5.
Calculate the long term spectrum (Fourier transform) for each wave file
created in Step 3.
Plot each long term spectrum.
Save the graph for each long term spectrum plot.
Lab Noise Recording Procedure
Settings.
Recorder: Olympus DM-520
Bit Depth: 16
Sampling Rate: 44.1 kHz, 48 kHz
Input: External Microphone
Input State: Shorted, Open
Microphone Sensitivity: Low/Medium/High
Recording Levels: 2, 4, 6, 8, 10, 12, 14, 16
Step 1.
Insert the shorted/open plug into the external microphone jack.
Step 2.
Record samples approximately 36 seconds long.
15


@ 2 sampling rates
@ 3 microphone sensitivities
@ 8 recording levels
@ 2 input states
@ 2 recording conditions
192 Samples Total
Step 3.
Trim the sample files to 30 seconds.
Save the trimmed samples to wave files.
Step 4.
Plot the waveform for each wave file created in Step 3.
Save the graph for each waveform plot.
Step 5.
Calculate the long term spectrum (Fourier transform) for each wave file
created in Step 3.
Plot each long term spectrum.
Save the graph for each long term spectrum plot.
16


CHAPTER IV
RESULTS
The use of digital audio recording is common in forensics. Recordings are
created during surveillance and interviews, and are used in investigations and as
evidence in legal proceedings. The noise measurements made for this paper
provide the basis for choosing recording settings that have the least noise and
highest fidelity.
The first selection to consider is the sampling rate. For the DM-520, the
choice is between 44.1 kHz and 48 kHz. In the 240 sample files, there are 120
pairs of files whose settings differ only by their sampling rate. Figure 4.1 displays
the comparison between the 44.1 kHz and 48 kHz sample files for the field
recordings set for medium microphone sensitivity and recording level 8. The
input was shorted.
The 4 plots display the graphs of the Fourier Transforms for the 2 sample
files. The 44.1 kHz file is on the left; the 48 kHz file is on the right. The upper
plots are the left channels; the lower plots are the right channels. The frequency
range of the X axis is from 0 Hz (not including the DC component) to 20 kHz,
inclusive. The bin width, or resolution, is 0.033333333333333 Hz. The decibel
range of the Y axis is from -180 to -70. 0 DB is referenced to a full scale output
of the 16 bit ADC, which is 215 -1. The green line on each plot is the mean value
of the channel samples.
17


Left Channel
30 Second Noise Sample
44.1kHz Sampling Rate
Left Channel
30 Second Noise Sample
48kHz Sampling Rate
"Channel Values
Channel Mean
Frequency(Hz)
Right Channel
30 Second Noise Sample
44.1kHz Sampling Rate
x 10
- Channel Values
Channel Mean
Frequency(Hz)
Right Channel
30 Second Noise Sample
48kHz Sampling Rate
x 10
1 1.5
Frequency(Hz)
x 10
Frequency(Hz)
x 10
Figure 4.1
DM-520 30 Second Field Noise Test
Shorted Input, 16 Bit
Medium Microphone Sensitivity, Recording Level 8
Comparing the mean values for the corresponding channels in Figure 4.1
supplies a measure of the relative amounts of noise in the channels. This is
justified by the similar contours of the FFT plots. For these two recordings, the
difference for both channels is the same at 0.2 decibels. Comparing the mean
values for all the corresponding channels will show any relationship between
noise and sample rate.
18


Figure 4.2 displays the graphs of the differences between the mean values
for all the corresponding channels in the 240 sample files.
Left Channels
3§ Second Mails Samite
0 0 1 0 2 0 3
Decibel Differences for Channel Means
Figure 4.2
DM-520 30 Second Noise Tests
Corresponding Channel Comparisons
19


The histogram shows the distribution of the differences between the sample
means for the left and right channels. The greatest difference between any 2
corresponding channels is 0.3 decibels. (The values for the sample means are
rounded to 1 decimal place.)
Figure 4.3 displays the graph for the sample means for all 240 sample files.
For this graph, the mean values for the 2 channels in each sample file are
averaged. This is justified because the maximum difference for the channels for
any sample file is less than 0.21 decibels. Table 4.1 shows the key for the
recording parameters for the files whose means are displayed in Figure 4.3.
Table 4.1
DM-520 30 Second Noise Tests
Recording Conditions
Zone Sampling Rate Recording Location Shielding Input
1 Left 44.1 kHz Field None Shorted
1 Right 48 kHz Field None Shorted
2 Left 44.1 kHz Lab Full Shorted
2 Right 48 kHz Lab Full Shorted
3 Left 44.1 kHz Lab Full Open
3 Right 48 kHz Lab Full Open
4 Left 44.1 kHz Lab Partial Shorted
4 Right 48 kHz Lab Partial Shorted
5 Left 44.1 kHz Lab Partial Open
5 Right 48 kHz Lab Partial Open
20


Figure 4.3
DM-520 30 Second Noise Tests
Channel Averaged
Mean Noise Amplitude
Channel Averaged
Mean Noise A n "lit tide


The graph is divided into 3 areas according to the microphone sensitivity
setting when the recordings were made. Each area is divided into 5 zones.
Each zone number has the same set of recording conditions as shown in Table
4.1. The value on the left side of each zone is for a 44.1 kHz sampling rate. The
value on the right side of each zone is for a 48 kHz sampling rate.
It is also useful to look at the results aggregated by recording level. Figure
4.4 displays the relative noise figures for the 24 field recordings made at the 48
kHz sampling rate. From this graph, noise figures can be estimated for the
recording levels that were not tested. The complete set of 10 graphs is included
in Appendix A.
Left Channel
Figure 4.4
DM520 30 Second Field Noise Tests
Shorted Input, 48k Sampling Rate, 16 Bit
22


Besides addressing a recorders broad band noise performance, the
methodology being advocated here should also account for noise in the narrow
band around the mains frequency (ENF) and its upper harmonics. For this
evaluation, the ENF is 60 Flz. The first 7 upper harmonics are also included,
from 120 to 480 Hz.
Figure 4.5 displays the 48 kHz sample file for the field recording with
settings for medium microphone sensitivity and recording level 8. The input was
shorted. The 4 plots display the graphs of the Fourier Transforms for the sample
file. The upper plots are the left channel; the lower plots are the right channel.
The full range FFT is on the left. The frequency range of the X axis is from 0 Hz
(not including the DC component) to 20 kHz, inclusive. The bin width, or
resolution, is 0.033333333333333 Hz. The decibel range of the Y axis is from -
180 to -70. 0 DB is referenced to a full scale output of the 16 bit ADC, which is
215 -1. The amplitudes of the 60 Hz component and its upper harmonics are on
the right. The green line on each plot is the mean value of the channel samples.
23


Figure 4.5
DM-520 30 Second Field Noise Test
Shorted Input, 48k Sampling Rate, 16 Bit
Medium Microphone Sensitivity, Recording Level 8
Figure 4.6 displays the aggregate for the 10 sample files recorded with
medium microphone sensitivity at recording level 8. The complete set of 24
graphs is included in Appendix B.
24


eft Channels
Figure 4.6
DM520 30 Second Lab Noise Tests
60 Hz Harmonic Levels
Medium Microphone Sensitivity, Recording Level 8
25


CHAPTER V
ANALYSIS
In Figure 4.2, the histogram shows that the greatest difference between the
sample means for any corresponding right or left channels in the 204 sample files
is 0.3 decibels. The code that produced the graph measures only the absolute
difference between the sample means. It does not track whether the mean value
is greater for the channel in the 44.1 kHz or the 48 kHz file. This information is
lost when the results are aggregated. From this graph, it appears that there is no
significant difference in internal noise generated by the 2 different sampling rates.
However, a pattern does emerge when the sample means are graphed
sequentially. In Figure 4.3, the left points in each zone are the sample means for
the 44.1 kHz file for each recording level. The corresponding right points are the
sample means for the 48 kHz file for the recording level. It is clear from the
graph that the 48 kHz sampling rate produces a slight but consistent difference
from the 44.1 kHz rate. This appears to be a computational artifact due to the
difference in the bandwidths for the 2 sampling rates. (The equations are given
in Appendix C.)
The 48 kHz sampling rate has an upper frequency limit (Nyquist
frequency16) of 24 kHz. This is 1.95 kHz higher than the upper frequency limit of
the 44.1 kHz sampling rate. (The graphs in Figure 4.1 are by convention
26


restricted to the audio range of 20 Hz to 20 kHz so the additional bandwidth for
the 48 kHz sample file is not visible.)
Assuming that the upper frequency limit of the anti-aliasing is raised a
corresponding amount, and assuming that the noise generated in the recorder's
input has power in that additional range, there should be a wider noise spectrum
for the 48 kHz sample rate. If the long term spectrum for both sampling rates
were flat, the sample means should be equal. However, the slope of the LTS for
both sampling rates declines at the higher frequencies.
Including the higher frequencies in the mean likely produces a lower value
for the wider spectrum of the 48 kHz files. (The Further calculations would show
whether restricting averaging to the audio range would eliminate this consistent
difference. In either case, the difference is minor and is unlikely to have much
significance.
While the choice of sampling rate has little apparent influence on the
amount of noise in the consequent recording, other settings and conditions have
a much greater effect. These effects can be deduced from the graphs in Figures
4.3 and 4.4. It should be noted that the most striking aspect of Figure 4.3 is the
consistency in the profiles of the plot lines.
As might be expected, increasing the sensitivity (gain) of the microphone
preamp increases the amount of noise in the recording. The amount of noise
also increases with higher recording levels. In addition, it is discernible that for
each recording level in each microphone sensitivity area, there are two
correlations among the file types. The noise figures for sample files made with
27


the same sampling rate and shorted inputs are the same, as are the noise figures
for those with open inputs. It is clear that the noise figures for the files with open
inputs are significantly higher than for those with shorted inputs. The greatest
difference, however, is only 5 dB. Overall, the broadband noise performance of
the DM-520 is excellent considering that the greatest mean value is
approximately -178 dB below full range.
An issue not addressed by the previous analyses is the possible presence
of tones at frequencies that interfere with the accidental or deliberate acquisition
of the ENF. The following analysis speaks to this issue.
In North America, the ENF is 60 Hz, 0.02 Flz. When the ENF is recorded,
any periodic noise at this frequency or its upper harmonics will interfere with
measuring the actual ENF. Figure 4.5 displays the 48 kHz sample file for the
field recording set for medium microphone sensitivity and recording level 8. The
input was shorted.
The 60 Hz harmonic profiles in Figure 4.5 are within the noise band for both
channels. There is no evidence of periodic noise at any of the frequencies.
The same analysis was performed for the rest of the 240 sample files. The
results were aggregated by microphone sensitivity and recording level. Figure
4.6 displays the results for medium microphone sensitivity and recording level 8.
As with the sample means for the wide band spectra shown in Figure 4.3, the
harmonic content is greater for the files with open inputs. The values for the
harmonic frequencies are generally higher than the channel mean. This is partly
a statistical variation due to the random nature of the noise and partly a result of
28


the low frequency emphasis in the analog to digital conversion process. Since
these samples are relatively short at 30 seconds, using a longer sample should
reduce the variation from sample file to sample file at the same harmonic
frequency. It should also reduce the variation among different harmonic
frequencies in the same file.
29


CHAPTER VI
CONCLUSION
This document has presented a methodology for evaluating some aspects
of the performance of a single unit of one type of portable digital recorder. The
results obtained here would be useful in judging whether the performance of this
unit is adequate to allow a valid measurement of the fluctuations in the ENF
recorded by it. Yet, the results do not cover all the aspects of the recorder's
performance that might introduce errors during the acquisition of the ENF. The
methodology as developed so far concentrates on the performance of the audio
input circuits. The next step is to measure the performance of the ADC in the
unit, in particular its linearity, using a waveform of known probability density.
This analysis is usually done during the design of an ADC where other sources of
noise can be minimized. In this case, it is an open question whether the noise
from any non-linearity can be separated from that from other sources. Further
testing is necessary to determine whether it can be done successfully.
It is important to note that while the methodology may be valid for other
models of recorder, these results and analyses are not. They do not even apply
to other units of the same model. In order to be valid for any particular device,
they must be done on that unit. To do otherwise is not only foolish; it is at least
intellectually dishonest.
30


It is also important to consider that the most important part of this or any
other methodology is not the methodology itself. Technology changes;
understanding and knowledge increase; methodologies must develop to stay
relevant. The most important part is the skepticism that drives it. Answers and
conclusions are always partial and conditional. Continually questioning what is
known and what is believed is the antidote for false assumptions and lax
attitudes. By not blindly accepting the status quo and looking at things from more
than one perspective, it is possible to find hidden connections and reveal
problems that would otherwise go undetected. This effort is crucial to sustain
and enhance the usefulness and validity of forensic science in all its disciplines.
31


REFERENCES
1. Heath, T.L. The Works of Archimedes. Cambridge: The University Press,
1897. xix. Print.
2. Hieron II. Encyclopaedia Britannica. Encyclopaedia Britannica, Inc.,
2014. Web. 22 Apr. 2014.
3. Heath, T.L. The Works of Archimedes. Cambridge: The University Press,
1897. 257. Print.
4. Human Evolution Dating. What does it mean to be human?
Smithsonian Institution, 2014. Web. 22 Apr. 2014.
5. Dror, Itiel E., and David Charlton. Why Experts Make Errors. Journal of
Forensic Identification. 56.4 (2006): 612. Print.
6. Argumentum ad Ignorantiam. Philosophy103: Introduction to Logic.
Lander University, 2004-9. Web. 29 Apr. 2014
7. The Fingerprint Inquiry. The Fingerprint Inquiry Report. Edinburgh,
Scotland. 2011. 740. Print.
8. North American Electric Reliability Corporation. Standard BAL-004-0
Time Error Correction. 1 Apr. 2005. 1, Print.
9. NERC Balancing Authority Controls Standard Drafting Team. Time Error
Correction. 2. Print.
10. Grigoras, C. Applications of ENF Analysis in Forensic Authentication of
Digital and Video Recordings. Journal of Audio Engineering Society. 57.9
(2009): 643-661. Print.
11. Grigoras, C., Jeff Smith, and Chris Jenkins. Advances in ENF Database
Configuration for Forensic Authentication of Digital Media. Proceedings of the
131st Convention of the Audio Engineering Society. Print.
12. Van DerZiel, A. Noise in Solid-State Devices and Lasers. Proceedings
of the IEEE. 58.8 (1970). Print.
32


13. Furutsu, K., and T. Ishida. On the Theory of Amplitude Distribution of
Impulsive Random Noise. Journal of Applied Physics. 32.7 (1961): 1206-1221.
Print.
14. Bell, L. H., and Douglas H. Bell. Industrial Noise Control: Fundamentals
and Applications. CRC Press, 1993. Print.
15. Doernberg, J., Hae-Seung Lee, and David A. Hodges. Full-Speed
Testing of A/D Converters. IEEE Journal of Solid-State Circuits. SC-19.6 (1984):
820-827. Print.
16. Pohlmann, Ken C. Principles of Digital Audio. New York: McGraw-Hill,
2005. 25. Print.
33


APPENDIX A
ADDITIONAL NOISE LEVEL GRAPHS
34


Left Lhannsl
Figure A.1
DM520 30 Second Field Noise Test
Shorted Input, 44.1k Sampling Rate, 16 Bit
Figure A.2
DM520 30 Second Field Noise Test
Shorted Input, 48k Sampling Rate, 16 Bit
35


Figure A.3
DM520 30 Second Lab Noise Test, Fully Shielded
Shorted Input, 44.1k Sampling Rate, 16 Bit
Figure A.4
DM520 30 Second Lab Noise Test, Fully Shielded
Shorted Input, 48k Sampling Rate, 16 Bit
36


Lift Channel
V

i'
*5
Figure A.5
DM520 30 Second Lab Noise Test, Fully Shielded
Open Input, 44.1k Sampling Rate, 16 Bit
Left Channel
Ri_ ei
Figure A.6
DM520 30 Second Lab Noise Test, Fully Shielded
Open Input, 48k Sampling Rate, 16 Bit
37


Figure A.7
DM520 30 Second Lab Noise Test, Partially Shielded
Shorted Input, 44.1k Sampling Rate, 16 Bit
Figure A.8
DM520 30 Second Lab Noise Test, Partially Shielded
Shorted Input, 48k Sampling Rate, 16 Bit
38


Left Channel
£
__i
j<
o >?
^ .2
2) "C'
o
Vi
5: <
Figure A.9
DM520 30 Second Lab Noise Test, Partially Shielded
Open Input, 44.1k Sampling Rate, 16 Bit
Left Channel
Ki.;hi Channel
Figure A.10
DM520 30 Second Lab Noise Test, Partially Shielded
Open Input, 48k Sampling Rate, 16 Bit
39


APPENDIX B
ADDITIONAL HARMONIC LEVEL GRAPHS
40


Left Channels
Figure B.1
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
Low Microphone Sensitivity, Recording Level 2
41


60 Hz Harmonic Levels
Medium Microphone Sensitivity, Recording Level 2
To m
E Tfi
£ ,0
CL i./:i
42


Figure B.3
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
High Microphone Sensitivity, Recording Level 2
43
U hio!:s f Lat 44 1 ' Lj"4i< Lat44 1-' La;4:!-' La:44 1; L;;
ihoittc Int-jt i-icr'ed !n:us aliened irp-ut 0:;-n hi-rut Gaer in:-.it iicrod Iniiit Shonec. In.;jt l pen in rut 0:en hpit
Sh elded Shielded Sh elded Shielded P a 1! a I v Pailiall:artaih' Fartial


Left Channels
Figure B.4
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
Low Microphone Sensitivity, Recording Level 4
44


60 Hz Harmonic Levels
Medium Microphone Sensitivity, Recording Level 4
To m
E Tfi
£ ,0
CL i./:i
45


60 Hz Harmonic Levels
High Microphone Sensitivity, Recording Level 4
To m
E Tfi
£ ,0
CL i./:i
46


60 Hz Harmonic Levels
Low Microphone Sensitivity, Recording Level 6
To m
E Tfi
£ ,0
CL i./:i
47


60 Hz Harmonic Levels
Medium Microphone Sensitivity, Recording Level 6
To m
E Tfi
£ ,0
CL i./:i
48


60 Hz Harmonic Levels
High Microphone Sensitivity, Recording Level 6
To m
E Tfi
£ ,0
CL i./:i
49


Left Channels
Figure B.10
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
Low Microphone Sensitivity, Recording Level 8
50


Left Channels
Figure B.11
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
Medium Microphone Sensitivity, Recording Level 8
51


Left Channels
:::> o
c-i ( i.:-j Cr-
Figure B.12
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
High Microphone Sensitivity, Recording Level 8
52


Left Channels
Figure B.13
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
Low Microphone Sensitivity, Recording Level 10
53


Left Channels
Figure B.14
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
Medium Microphone Sensitivity, Recording Level 10
54


Left Channels
Figure B.15
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
High Microphone Sensitivity, Recording Level 10
55


Left Channels
Figure B.16
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
Low Microphone Sensitivity, Recording Level 12
56


Left Channels
Figure B.17
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
Medium Microphone Sensitivity, Recording Level 12
57


Left Channels
Figure B.18
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
High Microphone Sensitivity, Recording Level 12
58


Left Channels
Figure B.19
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
Low Microphone Sensitivity, Recording Level 14
59


Left Channels
:::> o
c-i ( i.:-j Cr-
Figure B.20
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
Medium Microphone Sensitivity, Recording Level 14
60


Left Channels
Figure B.21
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
High Microphone Sensitivity, Recording Level 14
61


Left Channels
Figure B.22
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
Low Microphone Sensitivity, Recording Level 16
62


Left Channels
Figure B.23
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
Medium Microphone Sensitivity, Recording Level 16
63


Left Channels
* +
*
:::> o
c-i ( i.:-j Cr-
Figure B.24
DM520 30 Second Lab Noise Test
60 Hz Harmonic Levels
High Microphone Sensitivity, Recording Level 16
64


APPENDIX C
EQUATIONS
For
Where
Then
And
M: Mean Amplitude
S: Sampling Rate
A: Bin Amplitude
n: Bin Number
n ^ oo as S ^ oo
s
2
lim n = 2
S co S
---1
2
M =
S
2
S ^
n = 2 r
S-1
65


Full Text

PAGE 1

A FORENSIC INVESTIGATION OF THE ELECTRICAL PROPERTI ES OF DIGITAL AUDIO RECORDING by JACK LEROI B.S., Metropolitan State College, 2006 A thesis submitted to the Faculty of the Graduate School of the University of Colorado in partial fulfillment of the requirements for the degree of Master of Science Recording Arts Media Forensics 2014

PAGE 2

2014 JACK LEROI ALL RIGHTS RESERVED

PAGE 3

ii This thesis for the Master of Science degree by Jack LeRoi has been approved for the Recording Arts Program by Catalin Grigoras, Chair Jeff Smith Lorne Bregitzer Date: May 2, 2014

PAGE 4

iii LeRoi, Jack (M.S., Recording Arts Media Forensics ) A Forensic Investigation of the Electrical Properti es of Digital Audio Recording Thesis directed by Associate Professor Catalin Grig oras. ABSTRACT In media forensics, the devices; e.g. computers, sm art phones, still/video cameras, audio recorders, and software; e.g. video, audio, and graphics editors, file and disk utilities, mathematical computation a pplications, are, for the most part, black boxes. The design specifications are u sually proprietary and the operating specifications may be incomplete, inaccur ate, or unavailable. This makes it difficult to validate the technology, but using it without validation could discredit a practitionerÂ’s findings or testimony. The alternative is to test the device or program to determine relevant characteris tics of its performance. An important and common device in media forensics i s the portable digital audio recorder used to record surveillance and inte rviews. This type can also be used to record the alternating current (AC) wavefor m from the mains power. While small variations in the AC frequency (ENF) ca n be forensically important, distortion in the recording can affect its value in adjudication or investigation. A method is presented to evaluate aspects of a record erÂ’s operation that can cause distortion. Specifically, the method measures the noise generated by the recorderÂ’s electronics in its input and amplifier c ircuits. The method includes a procedure to isolate the recorder from environmenta l sources of noise. The method analyzes the broadband noise floor produced by the range of recording

PAGE 5

ivconditions and recorder settings. It also analyzes the noise amplitude for the harmonics for the mains frequency. The form and content of this abstract are approved I recommend its publication. Approved: Catalin Grigoras

PAGE 6

v TABLE OF CONTENTS CHAPTER I. INTRODUCTION ................................... ............................................... 1 II. RATIONAL ...................................... ................................................... .. 6 III. PROCEDURES ................................... ................................................. 8 IV. RESULTS ....................................... ................................................... 17 V. ANALYSIS ....................................... ................................................... 26 VI. CONCLUSION .................................... ............................................... 30 REFERENCES ........................................ ................................................... ........ 32 APPENDIX A. ADDITIONAL NOISE LEVEL GRAPHS ................. ............................ 34 B. ADDITIONAL HARMONIC LEVEL GRAPHS ............... ...................... 40 C. EQUATIONS ...................................... ................................................ 65

PAGE 7

1 CHAPTER I INTRODUCTION In a popular anecdote1, the 3rd century BCE Greek scientist and mathematician Archimedes of Syracuse was asked by K ing Hieron II2 to investigate whether the smith who had made the King ’s crown had substituted silver for some of the gold he was given. Archimed es was not allowed to melt the crown to compare its volume to that of the same weight in pure gold. He had to find a way, unknown at the time, to determine th e volume of an irregularly shaped object. The solution came to him in his bat h, as he observed the amount of water that his body displaced. Whether or not A rchimedes then ran naked through the streets shouting “Eureka”, Archimedes’ principle explains the relationships among weight, density, and volume and is an important scientific discovery3. However, while Archimedes’ discovery is remarkable, in a larger sense, it is not exceptional. The observation and analysis of t he physical world is a primary pursuit of Homo sapiens From the earliest times, our survival, as creatu res without fang or claw, depended on comprehending and manipulating nature. This ability of early humans to know and use the av ailable materials, e.g. wood, stone, animal tissues, to understand weather and th e plants and animals in their environment, was indispensable in the growth of cul ture and civilization.

PAGE 8

2Given these characteristics, and a human fossil rec ord of 250,000 years4, modern science is a curiously late development. On e can only speculate over the factors that prevented its earlier acceptance, but, paradoxically, the human need for explanation is a likely candidate. The na ivet of ancient peoples about causality left gaps in their phenomenology. In the absence of natural reasons for the world and its phenomena, they filled the gaps w ith their imagination. In their imagining, the world was animated by deities, spiri ts, and other intentional, but invisible beings. Scientific methods depend on objective assessments of observable phenomena to produce natural explanations of those phenomena. Proof may be offered for natural explanations and such proof may either be accepted or rejected. Supernatural explanations are beyond pro of and must either be believed or denied. The two types of explanations are incompatible. It cannot be that lightning is both a plasma discharge of differ ences in electrical potential and a thunderbolt thrown by Zeus from Mount Olympus. T here cannot even be an intermediate explanation. Supernatural explanations are appealing and emotion ally satisfying in a way that natural explanations often are not. They are simple and inclusive and certain while natural explanations are complex, par tial, provisional. They meet the human need for quick answers and risk avoidance that were inherited from our prey ancestors. Science and technology are today so embedded in the physical and social culture of global civilization, that they supply th e explanations and provisions that

PAGE 9

3formerly came from religion. They are so pervasive it is impossible to experience the world the way it was before they exi sted, but, absent an alternative, rejecting supernatural explanations wa s likely very difficult. However, once agriculture and writing developed, people had the time and tools to investigate other possible reasons for the visible world. This fostered the ability to reason and laid the foundation for what later be came the scientific method. The curiosity that drives scientific exploration is an inherent human characteristic but the ways of thought it requires are not natural. Even those who have no supernatural beliefs are not therefore auto matically scientific. The hardwired tendency, even need, to make quick decisi ons leaves us prone to logical fallacies and other mental errors. These e rrors are often embedded in heuristic methods like “common sense”. While commo n sense is often useful, the innate errors produce biases of various types. It is important to stress that these fallacies are not moral or intellectual failu res but natural consequences of our neurology5. With this understanding, it is possible to mitig ate the harmful effects of such fallacies. It is instructive to consider the effect this tende ncy has had on forensic science. For many years, courts, law enforcement o rganizations, and even forensic practitioners themselves accepted the prem ise that every fingerprint was unique. This assumption was based on the experienc e of fingerprint examiners. Since identical fingerprints were never found, this uniqueness proposition was not disproved; therefore it was assumed that the la ck of disproof amounted to

PAGE 10

4proof. (This is the “Argumentum ad Ignorantiam6”; i.e., “whatever is not false must be true, whatever is not true must be false.”) Yet there is no known mechanism that would prevent prints from different fingers, either from the same or different persons, from being the same. As a result, it is impossible to use fingerprint evidenc e for positive identification7. When Brandon Mayfield was arrested in 2004 for the Madrid bombings based on fingermarks that closely matched Ouhnane D aoud, it seemed that the uniqueness proposition was disproved. In fact, law enforcement had made an inaccurate identification. However, the issues rai sed by this case fit with the current movement in forensics towards scientific me thods and away from heuristics. This trend was accelerated with the pu blication of the NAS report in 2009 that directly addresses the challenges in medi a forensics. The developments in electronic technology in the la st 100 years have produced a complex, bewildering, and ever changing technical environment. Since the invention of digital electronics, the wor ldwide saturation of millions of miniature recording devices has produced an overwhe lming flood of image, video, and audio files. These have strained the ab ility of forensic practitioners to respond adequately to the demand. Given the limits on time, money, and personnel, it is tempting for them to accept catego rically that their tools are adequate to the task. However, as with the uniquen ess proposition, such acceptance is unwise, because, as with most end use rs, they typically have little or no knowledge about the inner workings of the tec hnology on which they rely.

PAGE 11

5In particular, the operating characteristics of the hardware and software used in media forensics can be difficult to determi ne. Manufacturer’s specifications may be missing or incomplete and the accuracy of the available information is not guaranteed. Failing to allow fo r this deficiency could discredit a practitioner’s findings or testimony. The better approach is to rigorously test technolog y to determine actual operation. These tests may cover the overall perfo rmance of a device or program, or may be focused on a specific factor or set of factors related to a particular application of the technology. Ideally, the results will show that the technology performs adequately for the intended tas k. If not, it must be rejected for that particular purpose, though it may yet be u seful for others. Guidelines for evaluating tools, techniques, and pr ocedures are found in the “SWGDE Recommended Guidelines for Validation Testin g”, published by the Scientific Working Group on Digital Evidence.

PAGE 12

6 CHAPTER II RATIONAL One of the applications of technology that is signi ficant in media forensics is the capture and analysis of the fluctuations in the frequency of the electrical grid. For North America, the standard frequency is 60 Hz but this fluctuates as the aggregate load on the grid varies. As the load inc reases, the frequency sags until more power is added to the grid. As the load decreases, the frequency rises until power is removed from the grid. Since the lo ad is constantly changing, the frequency is also constantly changing, though the p ower producers are required to hold the nominal frequency to 60 Hz 0.02 Hz8. (The actual frequency may vary up to 0.05 Hz before the system is considered over limit9.) Because the variations are non-cyclical (non-deterministic), th e actual frequencies that occur between any two moments cannot be predicted or mode led. Thus, the sequence of frequencies between those two moments can provid e a timestamp that indicates when that sequence of frequencies occurre d. The variations in the Electric Network Frequency (E NF) can be acquired unintentionally during the recording of audio signa ls10. This occurs in a number of ways. For one, when an electromagnetic field is generated by the electric current flowing in the grid, the field can induce a n ENF signal into the recorder's microphones or into its audio amplifier circuits. This signal is combined with the intended audio signal. For another, the recorder c an receive an ENF signal

PAGE 13

7 through an external (mains) power supply. If the 6 0 Hz AC from the mains is not adequately filtered from the DC power supply, an EN F signal can be induced in the recorder's audio amplifier circuits. Again, th is signal is combined with the intended audio signal. The variations in the Electric Network Frequency (E NF) can also be acquired deliberately. In order to use a recording of an unknown sequence of ENF variations as a timestamp, there must be a know n sequence of ENF variations to compare it to. To this end, database s of the ENF are acquired and maintained for each electrical grid11. The house voltage from the mains is converted to audio levels and applied to the audio inputs of the recording device used for the database. The utility of this method depends on the accuracy and precision with which the ENF is recorded. The discrepancy between the o riginal signal and a recording of it are classified as acquisition error s. These errors are generated at different stages in the circuitry of the recording device; the way and extent to which these errors affect the recorded ENF signal d etermine whether it can be used to validate the time of a recording. This pap er describes some of the factors of digital recording that can lead to these acquisition errors.

PAGE 14

8 CHAPTER III PROCEDURES The initial investigation of a recording device req uires the evaluation of its noise characteristics. This evaluation should meas ure the random fluctuations in electrical potential produced in the input and proc essing circuits of the device. The factors contributing to the noise figure are in ternal sources of noise as well as noise that arises from the environment in which a recording is made. Internal sources include thermal noise, shot noise, and flicker noise. These are generated by the recorder's electronic componen ts themselves12. Environmental sources include atmospheric13 and industrial noise14. These are generated by lightning and by discharges of ele ctrical energy by machinery and electrical devices. The sudden change in elect rical potential generates electro-magnetic fields which induce noise in the r ecorder's electronic components. Fluctuations in electrical potential can also be in duced in the device's components by other environmental phenomena. The m ost significant phenomena are the electro-magnetic fields produced by current flowing in the electric power network. The fluctuations in these fields are not random. The variations in strength are synchronous with the amp litude and frequency of the sine wave driving the electrical grid. The fluctua tions may be recorded as hum at the electric network frequency, 60 hertz for North America.

PAGE 15

9In a digital recorder, there is also noise added to the signal by the Analogto-Digital conversion. This noise is caused by qua ntization error in the ADC and jitter in the sampling clock. To properly evaluate a recorder's noise figure, it is necessary to eliminate all influence by environmental sources. This may be ac complished in one of two ways. Either the recorder can be located in an are a where these fields are absent or the recorder can be shielded from whateve r fields are present. It is possible to eliminate industrial noise and hum by l ocating the recorder away from all electric power system components, machinery, an d electrical devices, but this will not eliminate atmospheric noise. However, the use of proper shielding will reduce all ambient fields to insignificant levels. The procedures for this paper were developed using the facilities of the National Center for Media Forensics. The equipment included a Stanford Research Systems SR770 FFT spectrum analyzer, a Ram sey Electronics STE3000FAV faraday shielded enclosure, and an Olymp us DM-520 battery powered portable digital audio recorder. The SR770 contains a low distortion sine wave generator. Since the STE3000FAV is designed to block cell phon e signals in the 20 megahertz to 8 gigahertz range, it was necessary to validate the enclosureÂ’s performance in the audio range, particularly at the electric network frequency of 60 hertz. The complication in performing this vali dation was the lack of a calibrated test instrument. As a substitute, a por table recorder was used to receive any electro-magnetic energy in the audio ra nge that leaked through the

PAGE 16

10 shielding. However, this was bootstrapping because the very object of these procedures was to evaluate the recorder. The worka round was finally accomplished by adding an initial step. In this fi rst step, the recorder was used in an environment that was free of electro-magnetic fi elds. (Note that it was impossible to accomplish the entire procedure in th is environment because of the need for mains power for the SR770.) The field fre e recordings were analyzed for periodic phenomena. Due to the absence of exte rnal electro-magnetic fields, any phenomena must have been generated by the recor der itself. These results were then used to calibrate the recordings made wit h the DM-520 placed in the STE3000FAV in the lab. Recordings were made with the DM-520 in the STE3000 FAV, with the lid closed and latched. This provided the maximum elec tro-magnetic isolation. However, in this configuration there is no way to a pply an external signal to the DM-520. Leaving the lid unlatched and slightly ope n allowed connecting a thin audio cable between the SR770 signal generator and the DM-520, but this spoils the integrity of the enclosure. It was necessary t o determine whether recordings made with the configuration showed increased levels of environmental noise. To this end, recordings were made with an audio cable connected to the signal generator output of the SR770. It was run undernea th the lid of the STE3000FAV and the audio plug at the end of the cab le was placed near the input jack of the DM-520, but not connected to it. The DM-520 has 16 recording levels, 3 input gain le vels, and 2 sampling rates, resulting in 96 unique settings. There were also 2 recording environments,

PAGE 17

11 field and lab. The field tests had 1 input state, shorted. The lab tests had 2 input states, shorted and open. The lab test also had 2 recording conditions, shielded and partially shielded. Testing all of these combi nations would require 480 recordings; however, evaluating the DM-520Â’s noise performance only required testing a representative sample of these settings. The recording levels were reduced to half by selecting every other one, resul ting in 240 combinations for testing. The reduction halved the work effort requ ired to collect and process the recordings. Practically, the results for the missi ng combinations can be extrapolated from these 240 combinations. The input circuits of an audio recorder amplify the electrical signals from a microphone or external audio source. These circuit s produce internal noise from the activity of the electrons that flow through the m. This activity can be influenced by the impedance of the microphone or ex ternal audio source. By testing the DM520 with the input shorted and with t he input open, the impedance applied to the input circuits is either 0 ohms or i nfinite ohms. These two states provide the bounds for the impedance of anything co nnected to the recorder. The output impedance of the SR770 signal generator is specified as less than 5 ohms, so the self generated noise of the DM520Â’s in put circuits with the SR770 connected is near the results obtained with a short ed input. The device to encode a continuously variable voltag e (analog) to a series of fixed binary values (digital) is called an Analog t o Digital convertor. Such a convertor is at the heart of every digital audio re corder. To complete a conversion, the changing analog voltage must be fix ed at a static level long

PAGE 18

12enough for it to be encoded into a digital value. This is done by a sample and hold circuit that freezes the voltage until the con version is complete. The time for this conversion must be sufficiently short to fit w ithin the time window between samples. The time to charge the sample and hold ci rcuit must also fit in the window before the conversion time. In a digital audio recorder, noise is also generate d in the Analog to Digital Convertor (ADC). One source of this ADC noise is n on-linearity in its transition levels. Figure 3.1 shows the relationship between an analog input and the digital output for a 3 bit ADC. Figure 3.1 As the voltage increases or decreases, the code cha nges when the voltage crosses the transition point between one code and t he one adjacent. If the voltage difference between transition points varies then the response of the ADC is non-linear. This non-linearity produces noise i n the digital output. 111 110 101 100 011 010 001 000 Code Center Code Width (Voltage Span per LSB) nrr r 0 1/8 1/4 3/8 1/2 5/8 3/4 7/8 F S

PAGE 19

13 It is possible to measure the non-linearity of an A DC by applying a waveform of known probability density to its voltag e input and analyzing the code output15. The voltage range of the input signal should cov er the transition points of interest. During the design and testing of ADC components, the voltage range of the signal covers the entire input range of the ADC. For the purposes of this investigation, the transition points of special int erest are around the 0 voltage point. Any non-linearity in this area can affect t he accuracy and precision of zero crossing measurements of the ENF. This is of parti cular importance in using the recordings in ENF databases. Three common waveforms used for testing are the saw tooth, triangle, and sine. The saw tooth and triangle waveforms are the easiest to compute from since they are theoretically linear, that is they h ave the same change in voltage per unit time. Thus they have an equal probability for each transition point. However, it is difficult to generate these as truly linear and the equipment to generate them is uncommon. However, high quality s ine wave generators are readily available and, while the probability densit y computations are more complicated than those for linear waveforms, this i s the standard approach. For an ideal ADC, the distribution of digital code values matches the distribution of voltage values in the input wavefor m. Any non-linearity shows as differences in the distribution. However, using pr obability to measure nonlinearity requires the acquisition of large data se ts. This is because the infinite voltage values in the input have to be captured in the finite number of code output values. (These code values are also sometim es called “buckets”.) In

PAGE 20

14 addition, any random noise in the input will be red uced, as large data sets average the input variations. Field Noise Recording Procedure Settings. Recorder: Olympus DM-520 Serial Number: 100104915 Bit Depth: 16 Sampling Rate: 44.1 kHz, 48 kHz Input: External Microphone Input State: Shorted Microphone Sensitivity: Low/Medium/High Recording Levels: 2, 4, 6, 8, 10, 12, 14, 16 Step 1. Insert the shorted plug into the external microphon e jack. Step 2. Record samples approximately 36 seconds long. @ 2 sampling rates @ 3 microphone sensitivities @ 8 recording levels @ 1 input state 48 Samples Total Step 3. Trim the sample files to 30 seconds.

PAGE 21

15 Save the trimmed samples to wave files. Step 4. Plot the waveform for each wave file created in Ste p 3. Save the graph for each waveform plot. Step 5. Calculate the long term spectrum (Fourier transform ) for each wave file created in Step 3. Plot each long term spectrum. Save the graph for each long term spectrum plot. Lab Noise Recording Procedure Settings. Recorder: Olympus DM-520 Bit Depth: 16 Sampling Rate: 44.1 kHz, 48 kHz Input: External Microphone Input State: Shorted, Open Microphone Sensitivity: Low/Medium/High Recording Levels: 2, 4, 6, 8, 10, 12, 14, 16 Step 1. Insert the shorted/open plug into the external micr ophone jack. Step 2. Record samples approximately 36 seconds long.

PAGE 22

16 @ 2 sampling rates @ 3 microphone sensitivities @ 8 recording levels @ 2 input states @ 2 recording conditions 192 Samples Total Step 3. Trim the sample files to 30 seconds. Save the trimmed samples to wave files. Step 4. Plot the waveform for each wave file created in Ste p 3. Save the graph for each waveform plot. Step 5. Calculate the long term spectrum (Fourier transform ) for each wave file created in Step 3. Plot each long term spectrum. Save the graph for each long term spectrum plot.

PAGE 23

17 CHAPTER IV RESULTS The use of digital audio recording is common in for ensics. Recordings are created during surveillance and interviews, and are used in investigations and as evidence in legal proceedings. The noise measureme nts made for this paper provide the basis for choosing recording settings t hat have the least noise and highest fidelity. The first selection to consider is the sampling rat e. For the DM-520, the choice is between 44.1 kHz and 48 kHz. In the 240 sample files, there are 120 pairs of files whose settings differ only by their sampling rate. Figure 4.1 displays the comparison between the 44.1 kHz and 48 kHz samp le files for the field recordings set for medium microphone sensitivity an d recording level 8. The input was shorted. The 4 plots display the graphs of the Fourier Trans forms for the 2 sample files. The 44.1 kHz file is on the left; the 48 kH z file is on the right. The upper plots are the left channels; the lower plots are th e right channels. The frequency range of the X axis is from 0 Hz (not including the DC component) to 20 kHz, inclusive. The bin width, or resolution, is 0.0333 33333333333 Hz. The decibel range of the Y axis is from -180 to -70. 0 DB is r eferenced to a full scale output of the 16 bit ADC, which is 215 -1. The green line on each plot is the mean value of the channel samples.

PAGE 24

18 Figure 4.1 DM-520 – 30 Second Field Noise Test Shorted Input, 16 Bit Medium Microphone Sensitivity, Recording Level 8 Comparing the mean values for the corresponding cha nnels in Figure 4.1 supplies a measure of the relative amounts of noise in the channels. This is justified by the similar contours of the FFT plots. For these two recordings, the difference for both channels is the same at 0.2 dec ibels. Comparing the mean values for all the corresponding channels will show any relationship between noise and sample rate.

PAGE 25

19Figure 4.2 displays the graphs of the differences b etween the mean values for all the corresponding channels in the 240 sampl e files. Figure 4.2 DM-520 – 30 Second Noise Tests Corresponding Channel Comparisons

PAGE 26

20The histogram shows the distribution of the differe nces between the sample means for the left and right channels. The greates t difference between any 2 corresponding channels is 0.3 decibels. (The value s for the sample means are rounded to 1 decimal place.) Figure 4.3 displays the graph for the sample means for all 240 sample files. For this graph, the mean values for the 2 channels in each sample file are averaged. This is justified because the maximum di fference for the channels for any sample file is less than 0.21 decibels. Table 4.1 shows the key for the recording parameters for the files whose means are displayed in Figure 4.3. Table 4.1 DM-520 – 30 Second Noise Tests Recording Conditions Zone Sampling Rate Recording Location Shielding Input 1 Left 44.1 kHz Field None Shorted 1 Right 48 kHz Field None Shorted 2 Left 44.1 kHz Lab Full Shorted 2 Right 48 kHz Lab Full Shorted 3 Left 44.1 kHz Lab Full Open 3 Right 48 kHz Lab Full Open 4 Left 44.1 kHz Lab Partial Shorted 4 Right 48 kHz Lab Partial Shorted 5 Left 44.1 kHz Lab Partial Open 5 Right 48 kHz Lab Partial Open

PAGE 27

21Figure 4.3 DM-520 – 30 Second Noise Tests Channel Averaged Mean Noise Amplitude

PAGE 28

22The graph is divided into 3 areas according to the microphone sensitivity setting when the recordings were made. Each area i s divided into 5 zones. Each zone number has the same set of recording cond itions as shown in Table 4.1. The value on the left side of each zone is fo r a 44.1 kHz sampling rate. The value on the right side of each zone is for a 48 kH z sampling rate. It is also useful to look at the results aggregated by recording level. Figure 4.4 displays the relative noise figures for the 24 field recordings made at the 48 kHz sampling rate. From this graph, noise figures can be estimated for the recording levels that were not tested. The complet e set of 10 graphs is included in Appendix A. Figure 4.4 DM520 – 30 Second Field Noise Tests Shorted Input, 48k Sampling Rate, 16 Bit

PAGE 29

23Besides addressing a recorderÂ’s broad band noise pe rformance, the methodology being advocated here should also accoun t for noise in the narrow band around the mains frequency (ENF) and its upper harmonics. For this evaluation, the ENF is 60 Hz. The first 7 upper ha rmonics are also included, from 120 to 480 Hz. Figure 4.5 displays the 48 kHz sample file for the field recording with settings for medium microphone sensitivity and reco rding level 8. The input was shorted. The 4 plots display the graphs of the Fou rier Transforms for the sample file. The upper plots are the left channel; the lo wer plots are the right channel. The full range FFT is on the left. The frequency r ange of the X axis is from 0 Hz (not including the DC component) to 20 kHz, inclusi ve. The bin width, or resolution, is 0.033333333333333 Hz. The decibel r ange of the Y axis is from 180 to -70. 0 DB is referenced to a full scale out put of the 16 bit ADC, which is 215 -1. The amplitudes of the 60 Hz component and its upper harmonics are on the right. The green line on each plot is the mean value of the channel samples.

PAGE 30

24 Figure 4.5 DM-520 – 30 Second Field Noise Test Shorted Input, 48k Sampling Rate, 16 Bit Medium Microphone Sensitivity, Recording Level 8 Figure 4.6 displays the aggregate for the 10 sample files recorded with medium microphone sensitivity at recording level 8. The complete set of 24 graphs is included in Appendix B.

PAGE 31

DM520 Medium 25Figure 4.6 DM520 – 30 Second Lab Noise Tests 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 8 rophone Sensitivity, Recording Level 8

PAGE 32

26 CHAPTER V ANALYSIS In Figure 4.2, the histogram shows that the greates t difference between the sample means for any corresponding right or left ch annels in the 204 sample files is 0.3 decibels. The code that produced the graph measures only the absolute difference between the sample means. It does not t rack whether the mean value is greater for the channel in the 44.1 kHz or the 4 8 kHz file. This information is lost when the results are aggregated. From this gr aph, it appears that there is no significant difference in internal noise generated by the 2 different sampling rates. However, a pattern does emerge when the sample mean s are graphed sequentially. In Figure 4.3, the left points in ea ch zone are the sample means for the 44.1 kHz file for each recording level. The co rresponding right points are the sample means for the 48 kHz file for the recording level. It is clear from the graph that the 48 kHz sampling rate produces a slig ht but consistent difference from the 44.1 kHz rate. This appears to be a compu tational artifact due to the difference in the bandwidths for the 2 sampling rat es. (The equations are given in Appendix C.) The 48 kHz sampling rate has an upper frequency lim it (Nyquist frequency16) of 24 kHz. This is 1.95 kHz higher than the uppe r frequency limit of the 44.1 kHz sampling rate. (The graphs in Figure 4.1 are by convention

PAGE 33

27restricted to the audio range of 20 Hz to 20 kHz so the additional bandwidth for the 48 kHz sample file is not visible.) Assuming that the upper frequency limit of the anti -aliasing is raised a corresponding amount, and assuming that the noise g enerated in the recorder's input has power in that additional range, there sho uld be a wider noise spectrum for the 48 kHz sample rate. If the long term spect rum for both sampling rates were flat, the sample means should be equal. Howev er, the slope of the LTS for both sampling rates declines at the higher frequenc ies. Including the higher frequencies in the mean likely produces a lower value for the wider spectrum of the 48 kHz files. (The F urther calculations would show whether restricting averaging to the audio range wo uld eliminate this consistent difference. In either case, the difference is mino r and is unlikely to have much significance. While the choice of sampling rate has little appare nt influence on the amount of noise in the consequent recording, other settings and conditions have a much greater effect. These effects can be deduce d from the graphs in Figures 4.3 and 4.4. It should be noted that the most stri king aspect of Figure 4.3 is the consistency in the profiles of the plot lines. As might be expected, increasing the sensitivity (g ain) of the microphone preamp increases the amount of noise in the recordi ng. The amount of noise also increases with higher recording levels. In ad dition, it is discernible that for each recording level in each microphone sensitivity area, there are two correlations among the file types. The noise figur es for sample files made with

PAGE 34

28the same sampling rate and shorted inputs are the s ame, as are the noise figures for those with open inputs. It is clear that the n oise figures for the files with open inputs are significantly higher than for those with shorted inputs. The greatest difference, however, is only 5 dB. Overall, the br oadband noise performance of the DM-520 is excellent considering that the greate st mean value is approximately -178 dB below full range. An issue not addressed by the previous analyses is the possible presence of tones at frequencies that interfere with the acc idental or deliberate acquisition of the ENF. The following analysis speaks to this issue. In North America, the ENF is 60 Hz, 0.02 Hz. When the ENF is recorded, any periodic noise at this frequency or its upper h armonics will interfere with measuring the actual ENF. Figure 4.5 displays the 48 kHz sample file for the field recording set for medium microphone sensitivi ty and recording level 8. The input was shorted. The 60 Hz harmonic profiles in Figure 4.5 are withi n the noise band for both channels. There is no evidence of periodic noise a t any of the frequencies. The same analysis was performed for the rest of the 240 sample files. The results were aggregated by microphone sensitivity a nd recording level. Figure 4.6 displays the results for medium microphone sens itivity and recording level 8. As with the sample means for the wide band spectra shown in Figure 4.3, the harmonic content is greater for the files with open inputs. The values for the harmonic frequencies are generally higher than the channel mean. This is partly a statistical variation due to the random nature of the noise and partly a result of

PAGE 35

29the low frequency emphasis in the analog to digital conversion process. Since these samples are relatively short at 30 seconds, u sing a longer sample should reduce the variation from sample file to sample fil e at the same harmonic frequency. It should also reduce the variation amo ng different harmonic frequencies in the same file.

PAGE 36

30 CHAPTER VI CONCLUSION This document has presented a methodology for evalu ating some aspects of the performance of a single unit of one type of portable digital recorder. The results obtained here would be useful in judging wh ether the performance of this unit is adequate to allow a valid measurement of th e fluctuations in the ENF recorded by it. Yet, the results do not cover all the aspects of the recorder's performance that might introduce errors during the acquisition of the ENF. The methodology as developed so far concentrates on the performance of the audio input circuits. The next step is to measure the pe rformance of the ADC in the unit, in particular its linearity, using a waveform of known probability density. This analysis is usually done during the design of an ADC where other sources of noise can be minimized. In this case, it is an ope n question whether the noise from any non-linearity can be separated from that f rom other sources. Further testing is necessary to determine whether it can be done successfully. It is important to note that while the methodology may be valid for other models of recorder, these results and analyses are not. They do not even apply to other units of the same model. In order to be v alid for any particular device, they must be done on that unit. To do otherwise is not only foolish; it is at least intellectually dishonest.

PAGE 37

31It is also important to consider that the most impo rtant part of this or any other methodology is not the methodology itself. T echnology changes; understanding and knowledge increase; methodologies must develop to stay relevant. The most important part is the skepticis m that drives it. Answers and conclusions are always partial and conditional. Co ntinually questioning what is known and what is believed is the antidote for fals e assumptions and lax attitudes. By not blindly accepting the status quo and looking at things from more than one perspective, it is possible to find hidden connections and reveal problems that would otherwise go undetected. This effort is crucial to sustain and enhance the usefulness and validity of forensic science in all its disciplines.

PAGE 38

32 REFERENCES 1. Heath, T.L. The Works of Archimedes. Cambridge: The University Press, 1897. xix. Print. 2. “Hieron II.” Encyclopaedia Britannica. Encyclopaedia Britannica, Inc., 2014. Web. 22 Apr. 2014. 3. Heath, T.L. The Works of Archimedes. Cambridge: The University Press, 1897. 257. Print. 4. “Human Evolution >> Dating.” What does it mean to be human? Smithsonian Institution, 2014. Web. 22 Apr. 2014. 5. Dror, Itiel E., and David Charlton. “Why Expert s Make Errors.” Journal of Forensic Identification. 56.4 (2006): 612. Print. 6. “Argumentum ad Ignorantiam.” Philosophy103: Introduction to Logic. Lander University, 2004-9. Web. 29 Apr. 2014 7. The Fingerprint Inquiry. The Fingerprint Inquiry Report. Edinburgh, Scotland. 2011. 740. Print. 8. North American Electric Reliability Corporation Standard BAL-004-0 Time Error Correction. 1 Apr. 2005. 1, Print. 9. NERC Balancing Authority Controls Standard Draf ting Team. Time Error Correction. 2. Print. 10. Grigoras, C. “Applications of ENF Analysis in Forensic Authentication of Digital and Video Recordings.” Journal of Audio Engineering Society 57.9 (2009): 643-661. Print. 11. Grigoras, C., Jeff Smith, and Chris Jenkins. “ Advances in ENF Database Configuration for Forensic Authentication of Digita l Media.” Proceedings of the 131st Convention of the Audio Engineering Society Print. 12. Van Der Ziel, A. “Noise in Solid-State Devices and Lasers.” Proceedings of the IEEE 58.8 (1970). Print.

PAGE 39

3313. Furutsu, K., and T. Ishida. “On the Theory of Amplitude Distribution of Impulsive Random Noise.” Journal of Applied Physics 32.7 (1961): 1206-1221. Print. 14. Bell, L. H., and Douglas H. Bell. Industrial Noise Control: Fundamentals and Applications CRC Press, 1993. Print. 15. Doernberg, J., Hae-Seung Lee, and David A. Hod ges. “Full-Speed Testing of A/D Converters.” IEEE Journal of Solid-State Circuits SC-19.6 (1984): 820-827. Print. 16. Pohlmann, Ken C. Principles of Digital Audio. New York: McGraw-Hill, 2005. 25. Print.

PAGE 40

34 APPENDIX A ADDITIONAL NOISE LEVEL GRAPHS

PAGE 41

35 Figure A.1 DM520 – 30 Second Field Noise Test Shorted Input, 44.1k Sampling Rate, 16 Bit Figure A.2 DM520 – 30 Second Field Noise Test Shorted Input, 48k Sampling Rate, 16 Bit

PAGE 42

36 Figure A.3 DM520 – 30 Second Lab Noise Test, Fully Shielded Shorted Input, 44.1k Sampling Rate, 16 Bit Figure A.4 DM520 – 30 Second Lab Noise Test, Fully Shielded Shorted Input, 48k Sampling Rate, 16 Bit

PAGE 43

37 Figure A.5 DM520 – 30 Second Lab Noise Test, Fully Shielded Open Input, 44.1k Sampling Rate, 16 Bit Figure A.6 DM520 – 30 Second Lab Noise Test, Fully Shielded Open Input, 48k Sampling Rate, 16 Bit

PAGE 44

38 Figure A.7 DM520 – 30 Second Lab Noise Test, Partially Shielde d Shorted Input, 44.1k Sampling Rate, 16 Bit Figure A.8 DM520 – 30 Second Lab Noise Test, Partially Shielde d Shorted Input, 48k Sampling Rate, 16 Bit

PAGE 45

39 Figure A.9 DM520 – 30 Second Lab Noise Test, Partially Shielde d Open Input, 44.1k Sampling Rate, 16 Bit Figure A.10 DM520 – 30 Second Lab Noise Test, Partially Shielde d Open Input, 48k Sampling Rate, 16 Bit

PAGE 46

40 APPENDIX B ADDITIONAL HARMONIC LEVEL GRAPHS

PAGE 47

DM520 Low Mic 41Figure B.1 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Low Mic rophone Sensitivity, Recording Level 2 Sensitivity, Recording Level 2

PAGE 48

DM520 Medium 42Figure B.2 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Microphone Sensitivity, Recording Level 2 Sensitivity, Recording Level 2

PAGE 49

DM520 High Mic 43Figure B.3 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 2 Sensitivity, Recording Level 2

PAGE 50

DM520 Low Mic 44Figure B.4 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 4 4

PAGE 51

DM520 Medium 45Figure B.5 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 4 rophone Sensitivity, Recording Level 4

PAGE 52

DM520 High Mic 46Figure B.6 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 4 Recording Level 4

PAGE 53

DM520 Low Mic 47Figure B.7 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 6 rophone Sensitivity, Recording Level 6

PAGE 54

DM520 Medium 48Figure B.8 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 6 rophone Sensitivity, Recording Level 6

PAGE 55

DM520 High Mic 49Figure B.9 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 6 rophone Sensitivity, Recording Level 6

PAGE 56

DM520 Low Mic 50Figure B.10 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 8 rophone Sensitivity, Recording Level 8

PAGE 57

DM520 Medium 51Figure B.11 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 8 rophone Sensitivity, Recording Level 8

PAGE 58

DM520 High Mic 52Figure B.12 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 8 rophone Sensitivity, Recording Level 8

PAGE 59

DM520 Low Mic 53Figure B.13 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 10 rophone Sensitivity, Recording Level 10

PAGE 60

DM520 Medium Mic 54Figure B.14 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 10 rophone Sensitivity, Recording Level 10

PAGE 61

DM520 High Mic 55Figure B.15 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 10 rophone Sensitivity, Recording Level 10

PAGE 62

DM520 Low Mic 56Figure B.16 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 12 rophone Sensitivity, Recording Level 12

PAGE 63

DM520 Medium Mic 57Figure B.17 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 12 rophone Sensitivity, Recording Level 12

PAGE 64

DM520 High Mic 58Figure B.18 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 12 rophone Sensitivity, Recording Level 12

PAGE 65

DM520 Low Mic 59Figure B.19 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 14 rophone Sensitivity, Recording Level 14

PAGE 66

DM520 Medium Mic 60Figure B.20 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 14 rophone Sensitivity, Recording Level 14

PAGE 67

DM520 High Mic 61Figure B.21 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 14 rophone Sensitivity, Recording Level 14

PAGE 68

DM520 Low Mic 62Figure B.22 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 16 Recording Level 16

PAGE 69

DM520 Medium Mic 63Figure B.23 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 16 rophone Sensitivity, Recording Level 16

PAGE 70

DM520 High Mic 64Figure B.24 DM520 – 30 Second Lab Noise Test 60 Hz Harmonic Levels Mic rophone Sensitivity, Recording Level 16 rophone Sensitivity, Recording Level 16

PAGE 71

65 APPENDIX C EQUATIONS For M : Mean Amplitude S : Sampling Rate A : Bin Amplitude n : Bin Number Where S n as Then 0 1 2 2 2 lim S S n n SA And 1 2 2 2 S S n n MA