Cooper

The Electric Network Frequency - An Automated Approach

THE ELECTRIC NETWORK FREQUENCY (ENF) AS AN AID TO
AUTHENTICATING FORENSIC DIGITAL AUDIO RECORDINGS – AN
AUTOMATED APPROACH
ALAN J. COOPER
Metropolitan Police Service, London, UK
alan.j.cooper@met.police.uk

A recent forensic technique developed to establish the authenticity of recorded digital audio evidence is the Electric
Network Frequency (ENF) Criterion. This paper confirms the applicability of the ENF criterion for use in mainland
UK and introduces an automated approach to matching ENF estimates taken from a questioned recording to a database
of ENF values. The signal processing procedures described have been used successfully by the Metropolitan Police
Forensic Audio Laboratory in London to extract and match ENF data from evidential recordings.

INTRODUCTION
The term ‘forensic audio’ refers to audio material that
may provide evidence in legal proceedings. The party
introducing the recorded audio evidence must show that
it has not been altered since the time of its production
and this is achieved using standardised evidencehandling procedures and chain-of-custody records.
When these safeguards fail, the reliability of the
evidence may depend upon a technical analysis to prove
the originality and integrity of the recording.
The field of audio recording authenticity examination is
a complex forensic science forming an important
function of a forensic audio laboratory. An authentic
recording is defined by the AES as:
A recording made simultaneously with the
acoustic events it purports to have recorded,
and in a manner fully and completely
consistent with the methods of recording
claimed by the party who produced the
recording; a recording free from
unexplained artefacts, alterations, additions,
deletions, or edits [1].
Methods for the authenticity examination of analogue
recordings are well established [2-4] and recently
introduced magnetic feature visualisation techniques
may be applied to both analogue and digital recordings
stored on magnetic media [5,6]. However, the signal
conditioning principles associated with digital
recordings are completely different to that of analogue
and additional techniques for assessing the integrity of a
suspect recording are required [7-9]. One such
technique is the ‘Electric Network Frequency (ENF)
Criterion’ proposed by Grigoras [10-12].
ENF relates to the frequency of a networked electricity
supply transmission system. The ENF criterion is based
on estimating the frequency of power line related

signals that may have been recorded alongside the
wanted audio data as a result of the digital recorder
being connected directly to the network supply, or in the
case of battery powered recorders, being in the
proximity of electro magnetic fields emanating from the
network supply.
The ENF is not constant but changes by small amounts
in a random fashion over a period of time, further, the
same frequency value is experienced throughout the
network [10]. Thus, extracted ENF data from a
recording may be compared to a database of ENF
information, allowing the date and time of the recording
to be ascertained. Additionally, it may be used to
establish if and where the recording has been edited.
Therefore, the ENF criterion provides a powerful
method to assess the evidential integrity of audio data
that has been recorded onto audio, video, computer and
telecommunications equipment [10-13].
This paper confirms the validity of the ENF criterion for
use in mainland UK and describes a method to extract
ENF data from evidential recordings that allows
automatic searching and matching of the data to an ENF
database.
1
PRINCIPLES OF THE ENF CRITERION
Previous studies on the ENF criterion have been limited
to transmission systems for countries in continental
Europe who form a large single network controlled by
the “Union for the Co-ordination of Transmission of
Electricity” (UCTE) [14].
The electric power transmission system in England and
Wales, which is not part of UCTE, is run by the
National Grid Company (NGC) [15] and connects
power stations and substations in a high voltage network
and distribution system known as ‘The National Grid’.
The NGC also operate electricity interconnection

AES 33rd International Conference, Denver, CO, USA, 2008 June 5–7

1

Cooper
systems linking the transmission network in England
and Wales to the transmission systems in Scotland and
direct current inter-connectors with France and Northern
Ireland [16]. A map of the high voltage network
distribution in the UK (275 kV or above) is shown in fig
1 [17]. The majority of power introduced into a
network comes from turbines which drive alternating
current generators. The turbine’s speed of rotation
determines the ENF and standards adopted by countries
worldwide are based on either 50 Hz or 60 Hz
transmission systems; in the UK and Europe the ENF is
50Hz while in North America it is 60 Hz. An electrical
distribution network, or grid, is organised and powergenerating facilities distributed in a way that allows the
grid operators to cope with wide changes in the
dynamics of supply and demand. Within a network, the
generating systems operate in synchronicity, and the
ENF will remain constant if the sum of all loads and
losses equals the total generation of the network [18].
When there is not enough power available to meet the
demand on the grid, the generators all slow down
together and the ENF falls, conversely, if the demand
for power drops the generators speed up and the ENF
increases. If the average rate of demand on the system
differs from the average rate of supply in any given
period, the network operators will be presented with a
change in frequency for which it must immediately
compensate by shedding load for under-frequency or
shedding generation for over-frequency [18]. System
frequency will therefore vary around the 50 or 60 Hz
target and the network operators have statutory
obligations to maintain the frequency within certain
limits. For the NGC this is +/- 0.5 Hz, however, it is
normally kept within more stringent 'operational limits'
which are set at +/- 0.2 Hz [19].

The Electric Network Frequency - An Automated Approach
The Metropolitan Police Forensic Audio Laboratory has
been collating a database of ENF estimates for over four
years. The database located in London has been used to
corroborate the date and time of both test and evidential
recordings made in places located around England
Wales and Scotland. A typical histogram produced
from one month of ENF data is shown in fig 2, the
distribution is Gaussian having a mean of 50 Hz and a
standard deviation of 0.06 Hz.

Mean=50 Hz
STD=0.06 Hz

Frequency

Fig 2: Histogram of ENF data taken from the archive
for November 2006. A Gaussian distribution has been
overlaid for reference.

ENF components may find their way onto a recording
due to poor power supply regulation, earth loops
between recording equipment or more likely via
inductive coupling of ENF currents into high gain
recorder circuitry as the result of electromagnetic fields
emanating from recorder power supply components
such as transformers. In the same way, battery powered
recorders may have ENF signals induced from nearby
mains-operated equipment.
The recorded ENF signal may contain harmonics with
one or more of the harmonics having a higher power
than the fundamental. These harmonics may also be
used for analysis; however, frequencies above the third
harmonic are unlikely to be useful due to masking
caused by lower frequency acoustic signals [20].
In summary, the synchronicity of the generators produce
a uniformity of ENF across any geographic part of the
grid, and over a period of time the dynamic behaviour
of the supply and demand provides a unique ENF
deviation pattern. The combination of these two factors
makes the ENF a powerful forensic tool when an audio
recording has captured the ENF as a by-product of the
recording process.

Fig 1: High voltage power line distribution across
England, Wales and Scotland [17].

AES 33rd International Conference, Denver, CO, USA, 2008 June 5–7

2

Cooper
2

SIGNAL PROCESSING OF ENF SIGNALS
AN OVERVIEW
To date there have been three main methods used to
extract ENF data from an evidential recording:
‘time/frequency domain analysis’ based on the
spectrogram, ‘frequency domain analysis’ based on
selecting the maximum magnitude of a series of power
spectrums calculated from consecutive time segments of
data, and ‘time domain analysis’ based on zero crossing
measurements of a band-pass filtered signal, [10-12].
For all methods, the results reported have relied on a
visual comparison of the extracted data to a database of
ENF information. Visual searching may therefore
require many hours of work on the part of the examiner,
making an automated search process attractive.
The methods described in this paper are based on a
frequency domain approach. Producing a practical
method for automatically searching and matching the
ENF requires the resulting data from a Fast Fourier
Transform (FFT) to be reduced so as to minimise the
computational overheads of the search routine. The
obvious method is to use an algorithm that stores only
the peak value of the frequency estimate taken over a
well-defined bandwidth.
Two related problems are encountered when extracting
ENF data from evidential recordings. The first relates
to the precision of the measurements on the ENF
signals. This is limited not only by practical
considerations, but additionally by the time-bandwidth
product in Fourier Transformation theory known as the
‘uncertainty principle’, which states that you can not
obtain arbitrarily high resolution in both the time and
frequency domains simultaneously, making low
frequency signals that vary with time very difficult to
estimate [21]. The second being that in general the
recorded ENF signal energy is usually small, producing
frequency estimates that are susceptible to error due to
noise. A number of signal processing techniques have
been proposed to help overcome the limitations of the
Fourier transform uncertainty principle, including those
based on parametric frequency estimators [21], zero
padding/interpolation schemes [22] and signal
derivatives [23].
The paper is split into three parts. The first part
introduces a method for ENF extraction for
database/archiving purposes that allows a commercially
available real time FFT analyser and peak frequency
data logger to be used. Suitable time and frequency
resolution are obtained by using a technique that trades
time for bandwidth. The second part describes a
method for ENF extraction from evidential recordings
based on an overlapping Short Time Fourier Transform
(STFT) combined with a peak interpolation scheme
[22], allowing good time and frequency resolution for
minimal computational overheads. The third part
describes the automated matching process of the

The Electric Network Frequency - An Automated Approach
extracted data from the evidential recording to the
archive of database ENF values, and is based on a
simple Mean Squared Error (MSE) search. Both the
extraction and matching algorithms have been
developed using MathWorks MATLAB.
3

PRODUCING AN ARCHIVE OF ENF
ESTIMATES
The archiving process to be described produces ENF
estimates every 1.4 seconds at a spectral resolution of
0.0009 Hz and along with associated time for each
estimate, stores them to a simple text file which has the
start date of the file contained in a header. Each file
will contain approximately one month of ENF data,
allowing the automated searches to be carried out using
easily manageable data sets.
Transformation to the frequency domain is achieved
using a standard FFT having inherent time and
frequency resolution trade-offs. In practice, this means
that suitable frequency resolution requires a very long
length of signal. A method has been used that provides
acceptable time and frequency resolution by increasing
the analysis bandwidth for a directly proportional
reduction in the analysis time window. This is achieved
by making the original ENF sine-wave signal nonlinear, producing higher order harmonic components.
The analysis is carried out using one of these ENF
harmonics, where compared to the fundamental, its
frequency will have a directly proportional increase in
bandwidth, and for a given frequency resolution will
require a directly proportional reduction in FFT size and
therefore a directly proportional reduction in the amount
of input data required. Thus allowing proportionally
more ENF estimates to be made over a period of time.
The non-linearisation process is achieved using a simple
pulse generator locked to a full wave rectified signal
taken from the output of a step-down transformer
connected directly to the electrical network. The output
of the locked pulse generator is then fed to a high
quality 16 bit PC soundcard. The impulse response of
the soundcard will be convolved with the incoming ENF
locked pulses which in the frequency domain produces a
spectrum consisting of a set of harmonically related
frequencies under an approximate sin x/x envelope. The
full-wave rectification produces even order harmonics
only. It is expected that the ENF frequency will
normally be within the range 49.5 Hz to 50.5 Hz [19].
The harmonic of the fundamental 50 Hz signal chosen
for analysis is the 100th, equating to a centre frequency
of 5 kHz and at this frequency the bandwidth of interest
will be between: 49.5 × 100 = 4950 Hz and
50.5 × 100 = 5050 Hz, a one hundred-fold increase.
After each Fourier transform the peak frequency is
selected from within the band of interest, building a
vector of peak frequency estimates representing the
changing pattern of the ENF. The resulting frequency

AES 33rd International Conference, Denver, CO, USA, 2008 June 5–7

3

Cooper

The Electric Network Frequency - An Automated Approach

estimates are then simply scaled back down to the
baseband, achieving the same frequency resolution as
would normally be obtained for 100 times the data
length, thus allowing 100 times more ENF estimates for
a given time period. The FFT, frequency windowing,
peak picking and storage operations are all carried out
in real time using Sound Technology Inc, SpectraLab
software, running under Microsoft Windows XP. The
overall process is described by fig 3.
In practice the ENF signal to noise ratio (SNR) ranges
from between 50 dB to 70 dB [18] and using the nonlinear technique described, the 100th harmonic still
produces a signal to noise circa 45 dB. Therefore, the
proposed technique produces negligible estimation error
due to noise.

240v rms
50 Hz

Step down
transformer
and rectifier

ENF locked
pulse
generator

leads to an approximately maximum likelihood
estimator for moderate signal to noise ratios.
4.1 Short time Fourier transform (STFT)
The STFT may be derived by splitting the original data
sequence x[ n ], 0 ≤ n ≤ N − 1 , into J overlapping
segments of length M samples:

xm[n] = x[mL + n], 0 ≤ n ≤ M − 1

(1)

Where xm is the mth frame of the input signal and L is
the number of samples advanced between each
consecutive frame, known as the ‘hop size’. Each frame
is then multiplied by a length M spectral analysis
weighting window w producing:

Sound
Card

Log
power
spectrum

Window
100th
harmonic
+/- 50 Hz

PC

Peak
detection
and storage
including
date & time
information

Fig 3: Process for estimating and archiving ENF data.
4

EXTRACTION OF ENF DATA FROM
RECORDINGS
The problem of extracting the ENF data from an
evidential recording may be defined as: ‘track and
estimate at regular time intervals a single sinusoidal
component having a finite SNR, a relatively narrow
bandwidth and a slow rate of change of frequency’.
From network operational practices as described by the
NGC [19], the total bandwidth for the ENF may be
defined over the range 49.5 Hz to 50.5 Hz. The SNR is
determined by the induced level of ENF into the
recording system and the relative levels of electronic
and acoustic noise found over the ENF bandwidth of the
recording.
The process must provide adequate frequency resolution
and sampling interval between ENF estimates that are
compatible with the archive database. This will allow a
simple and efficient matching process to be used.
The method chosen to track the dynamic behaviour of
the ENF is based on the Short Time Fourier Transform
(STFT) [24]. Peak magnitude estimation is achieved
using a quadratic interpolation and mild zero padding
scheme as described by Abe & Smith [22]. The process

xm[n] = xm[n] â‹… w[n]

(2)

The data from the windowed frame is then extended by
zeros using a factor of b to produce a zero-padded
windowed frame x ′m[ n] . Converting each frame to the
frequency domain using a length P FFT produces the
STFT at frame m:

Xm[k ] =

− j 2π kn
1 P −1
P
′

x
n
e
m
[
]
∑
P n=0

(3)

Where k is the kth frequency bin. The hop size L is set
to match the sampling time interval of the archiving
database (1.4 seconds) and the FFT transform size P is
variable and dependent on a user-defined parameter D
expressed in terms of multiples of hop size L:

P = L â‹… fs â‹… b â‹… D

(4)

Where fs is the recorded data sampling rate. The STFT
segmentation process is described by fig 4. Each frame
is analysed to find the prominent local maximum or
peak in the magnitude spectrum corresponding to the

AES 33rd International Conference, Denver, CO, USA, 2008 June 5–7

4

Cooper

The Electric Network Frequency - An Automated Approach

ENF. The ENF peak value in any frame is unlikely to
coincide with the exact frequency position of an FFT
transform point. One advantage of zero padding in the
time domain is that it increases the FFT size, making
each FFT bin bandwidth fs/p proportionally narrower.
This produces a more densely sampled spectrum,
providing accurate interpolation in the frequency
domain. However, to gain reasonable accuracy, the
zero padding factor has to be very large, resulting in a
very high FFT size. This has obvious implications for
processing efficiency, where time scales are increased
according to P â‹… log( P ) .

From eq 3, the three frequency samples expressed in dB
from each frame as defined in a) and b) are given by:

α = 20 log10 Xm(kβ −1 )

(5)

β = 20 log10 Xm(kβ )

(6)

λ = 20 log10 Xm(kβ +1 )

(7)

Total data length
DxL

Frame 1
Frame 2

DxL

L

Frame 3

DxL

L

Frame 4

DxL

L

Frame 5

DxL

L
L

Frame 6

DxL

L: hop size (sampling interval)
D: Sample interval multiplication factor
D x L: Frame length

DxL

L

Frame J

Fig 4: Segmentation and overlap scheme used for the STFT.

4.2 Quadratic interpolation
In order to overcome the computational limitations of
high zero-padding factors, a quadratic interpolation
scheme has been used in conjunction with a low zero
padding factor as described by Abe & Smith [22]. The
procedure is straightforward: compute the log power
spectrum of each STFT frame using a zero padding
factor of 4, and then apply quadratic interpolation
(QIFFT) to each frame as follows:
a)

b)
c)
d)

Select the FFT bin β having maximum
magnitude over the spectral bandwidth of
interest (coarse estimate).
Select the adjacent FFT bins
β − 1 and β + 1 either side of the peak.
Fit a second order (quadratic) model to the
3 values of data.
The estimated peak value (ν ) of the
QIFFT is the peak value of the quadratic
model.

Solving for the peak location ν of the quadratic model
using α , β and λ [25]:

1
α −λ
2 α − 2β + λ

ν= ⋅

(8)

Estimation error bias inherent in quadratic interpolation
is the difference between the true peak value and the
peak value of the fitted quadratic model. The bias is
reduced to acceptable levels by the application of the
mild zero padding factor as described [22].
4.3 The overall extraction process
To improve the efficiency of the processing the initial
sampled audio data is decimated to 300Hz. This is
followed by a band-pass filter set to the ENF frequency
region of interest (49.5 Hz to 50.5 Hz). The time
domain signal is then split into J overlapping frames
and each frame processed as previously described. The
overall extraction process is shown in fig 5.

AES 33rd International Conference, Denver, CO, USA, 2008 June 5–7

5

Cooper

The Electric Network Frequency - An Automated Approach
Input data
Decimation
to
300 Hz

Bandpass filter
49.5 Hz - 50.5 Hz
Split data into J
overlapping frames

Weighting
window
Frame 1

Weighting
window
Frame 2

Weighting
window
Frame J
Segments
3 to J-1

Zero pad
data by a
factor 4

Zero pad
data by a
factor 4

Zero pad
data by a
factor 4

FFT frame 1

FFT frame 2

FFT frame J

Quadratic
interpolation
of peak value
Frame 1

Quadratic
interpolation
of peak value
Frame 2

Quadratic
interpolation
of peak value
Frame J

Vector of ENF estimates:

f1 , f 2 , "" f J

Fig 5: Overall ENF signal processing procedure.
4.4 Noise robustness
This section discusses the practical problem of
estimating the peak value of a single slowly varying
sinusoid from a finite number of noisy discrete time
observations. The application of the band-pass filter
provides frequency selectivity, confining the data to the
ENF region of interest only. The band-pass filter,
which is applied before the signal is segmented, will
also prevent spectral leakage components produced
from signals outside the region of interest masking the
ENF. This allows a rectangular weighting window to be
deployed, leading to greater noise immunity, due to it
having the narrowest main-lobe in the frequency
domain of all weighting windows [26].
Stochastic noise is also a major problem for the
extraction process, as the level of induced ENF is often
very small. For decreasing SNR’s peak estimation
errors increase and there is usually a point at which the
error rises very rapidly [27]. The FFT can be very

effective in picking out periodic components of a signal,
even when it is affected by relatively high noise levels.
A P point FFT may be considered as P/2 contiguous
band-pass filters, with the bandwidth of each filter being
dependant on the sample rate and the number of points
used in the FFT (fs/p). Considering the noise to be
white, a SNR improvement can be achieved by
spreading the noise over P filters producing a noise
power reduction Ï• of:

⎛1⎞
⎟
⎝PâŽ

ϕ = 10 log10 ⎜

(9)

Thus, within the FFT frequency bin containing a
sinusoid, the SNR will be improved by 3dB for every
doubling of FFT size. However, doubling the FFT size
requires a doubling of data size and therefore a trade-off
between SNR and time resolution between frames.

AES 33rd International Conference, Denver, CO, USA, 2008 June 5–7

6

Cooper

The Electric Network Frequency - An Automated Approach

For typical evidential recordings, the SNR’s for ENF
estimation needs a large frame size requiring many
seconds of data for each frame. In the overlapping
frame system previously described (fig 4), the sampling
interval multiplication factor, D, is used to increase the
transform/frame size to exploit the noise suppression
properties of the FFT, allowing a trade-off between
noise performance and ENF envelope resolution.
An example of noise performance verses frame length is
shown, where a noisy ENF signal has been extracted
from a 6-minute long section of recording. Fig 6 shows
an extraction using a frame length of 1 sample interval
(D=1, 1.4 seconds), fig 7 shows an extraction of the
same data using a frame length of 5 sample intervals
(D=5, 7 seconds) and fig 8 shows the same data
extracted using a frame length of 15 sample intervals
(D=15, 21 seconds). It can be seen that increasing the
frame length reduces the effects of the noise.

ENF samples

Fig 6: Frame length set to 1 sample interval.

ENF samples

Fig 8: Frame length set to 15 sample intervals.

5
AUTOMATED MATCHING OF ENF DATA
An algorithm is required that searches for the extracted
ENF pattern in the archived ENF data. The process
involves overlaying two length N vectors a and t and
computing a metric that determines the degree to which
the two vectors match. The metric used for this
matching process is the mean squared error (MSE).
The extracted data is overlaid at the start of the archive
file and the MSE between the extracted data and the part
of the archive that has been overlaid is calculated, the
result is then stored. The extracted data is advanced one
sample and the process is repeated until the extracted
data has slid across the entire archive file. A vector of
‘error’ values is therefore formed and the minimum
value is tagged as the best match. This error value
directly indicates the start position of the match in the
archive and therefore the date and time. The error ε is
given by the logarithm of the MSE:

⎛1
⎝N

ε = log ⎜

ENF samples

Fig 7: Frame length set to 5 sample intervals.

i= N

∑ (a − t )
i =0

i

i

2

⎞
⎟
âŽ

(10)

Where ai is the i th element of the overlapped archive
file and ti is the i th element of the extracted file, N is
the number of samples or elements in the extracted file.
The logarithm provides better visual discrimination
when the errors are plotted graphically and produces
less skew in the overall error distribution aiding
statistical analysis.
An example showing the results of an automated
matching process using the techniques described is
shown in fig 9. The extracted data is from a 70 minute
recording produced in Glasgow Scotland UK and the
archive produced in London England UK, a distance of
420 miles (676 km), the archived data has been offset
by 0.1 Hz to aid visual comparison.

AES 33rd International Conference, Denver, CO, USA, 2008 June 5–7

7

Cooper

The Electric Network Frequency - An Automated Approach

Archive (London)

Extract (Scotland)

Fig 9: Automated match found using the
techniques described. The extracted waveform
has been offset by 0.1 Hz to aid visual
comparison.
Statistically, the strength of the match may be
ascertained by examining the minimum MSE value in
relation to all the other MSE values obtained during the
search of the archive. It is found that for relatively long
recordings > 30 minutes, the error distribution may be
close to Gaussian. From the Scotland/London example,
fig 10 shows the standardised errors and fig 11 shows a
histogram of the standardised errors with a Gaussian
distribution overlaid for comparison purposes. It can be
seen that the minimum error is >6 standard errors below
the mean, indicating that the match has almost certainly
not resulted by chance.

Fig 11: Histogram of errors with a Gaussian overlay.
As the recording length diminishes the ENF pattern
becomes less complex. The differentiation between
lower error values therefore decreases and the
probability of a match occurring by chance increases.
However, if the extracted ENF signal is of good quality,
reliable automated matches can still be obtained with
relatively short recording lengths. As an example of
this, a two minute ENF extract was matched
6
from 2 × 10 comparisons taken over a 36 day archive
file and the results are shown in fig’s 12, 13 and 14.
Over this short recording, very good visual correlation
can be seen between the archive and extracted ENF
values (fig 12). A high discrimination between the
lowest error value and its nearest neighbours are shown
in fig 13. The error distribution is shown to be skewed
(fig 14) making statistical inference more difficult.
However, transforming the results to Gaussian may be
possible.

Archive
Extract
point of minimum error

Fig 10: Standardised errors showing the point of match
being >6 standard errors below the mean.
Fig 12: Automated match found using a 2 minute ENF
extract. The extracted waveform has been offset by
0.01 Hz to aid visual comparison.

AES 33rd International Conference, Denver, CO, USA, 2008 June 5–7

8

Cooper

The Electric Network Frequency - An Automated Approach
ACKNOWLEDGEMENTS
The author would like to thank Dr Catalin Grigoras of
the National Institute of Forensic Expertise in Bucharest
Rumania for introducing him to the ENF criterion and
for the useful discussions and exchange of ideas over
the past several years. Thanks also to Robin How of the
Metropolitan Police Audio Laboratory for helpful
suggestions and feedback relating to draft versions of
this paper.
point of minimum error

Fig 13: Even for a 2 minute section the minimum
error value of the match is still well below its
nearest neighbours.

Fig 14: The overlaid Guassian distribution
highlights the skew in the error histogram.

6
CONCLUSIONS
This paper further demonstrates the ENF criterion as a
powerful methodology to authenticate digital audio
recordings and validates its use in mainland UK by
establishing ENF correlation over large geographical
distances. Relatively simple processing procedures
have been instigated allowing reliable automated date
and time matching of extracted ENF data to an archived
ENF file, even for a recording as low as two minutes in
duration. Obviously, the reliability of a match
diminishes with decreasing ENF to noise ratio and is the
single biggest limitation of the ENF criterion. It is
therefore anticipated that future research will target the
development of robust and efficient ENF extraction
algorithms.

REFERENCES
[1]
AES27:1996, Recommended Practice for
Forensic Purposes – Managing Recorded
Audio Materials Intended for Examination,
Audio Engineering Society (1996).
http://aes.org/publications/standards

[2]

B. E. Koenig, “Authentication of Forensic
Audio Recordings”, Journal of the Audio
Engineering Society, vol. 38, no. 1/2, pp. 3-33,
January/February (1990 Jan/Feb.).

[3]

D. J. Dean, The relevance of replay transients
in the forensic examination of analogue tape
recordings, Police Scientific Development
Branch, Home Office, Science and Technology
Group, publication 16/91 (1991).

[4]

J. P. French, “Authenticity examinations of
audio recordings”, Proceedings of the Institute
of Acoustics, Reproduced Sound 9, vol. 15, no.
7 (1993).

[5]

D. Boss, S. Gfroerer, N. Neoustroev, “A New
Tool for the Visualisation of Magnetic Features
on Audiotapes”, The International Journal of
Speech Language and the Law, Forensic
Linguistics, vol. 10, no. 2, pp. 255-276 (2003).

[6]

J. Bouten, M. Van Rijsbergen, and S. Donkers,
“Derivation of a transfer function for imaging
polarimetry used in magneto-optical
investigations of audio tapes in authenticity
investigations”, Journal of the Audio
Engineering Society, vol. 55, no. 4, pp. 257265 (2007 Apr.).

[7]

A. J. Cooper, “The Significance of the Serial
Copying Management System (SCMS) in the
Forensic Analysis of Digital Audio
Recordings”, The International Journal of
Speech Language and the Law, vol. 12, no. 1,
pp. 49-62 (2005).

AES 33rd International Conference, Denver, CO, USA, 2008 June 5–7

9

Cooper

The Electric Network Frequency - An Automated Approach

[8]

A.J. Cooper, “Detection of Copies of Digital
Audio Recordings for Forensic Purposes”, PhD
thesis, Open University, Milton Keynes, UK,
(2006). http://www.robinhow.com/ac

[19]

Electricity-Real Time Operational Data,
National Grid Company.
http://www.nationalgrid.com/uk/electricity/data
/Realtime/

[9]

E. B. Brixen, “Techniques for the
authentication of digital audio Recordings”,
Audio Engineering Society, 122nd Convention,
Vienna, Austria (2007).

[20]

[10]

C. Grigoras, “Digital Audio Recording
Analysis: The Electric Network Frequency
(ENF) Criterion”, The International Journal of
Speech Language and the Law, vol. 12, no. 1,
pp. 63-76 (2005).

E. B. Brixen, “Further Investigations into the
ENF Criterion for Forensic Authentication”,
123rd Audio Engineering Society Convention,
New York, USA, (2007 Oct.), Convention
paper 7275.

[21]

A. Czyzewski, A. Ciarkowski, A. Kaczmarek,
J. Kotus, M. Kulesza and P. Maziewski, “DSP
Techniques for Determining “Wow”
Distortion”, Journal of the Audio Engineering
Society, vol. 55, no. 4, pp. 266-284 (2007
Apr.).

[22]

M. Abe and J. O. Smith III, “Design Criteria
for Simple Sinusoidal Parameter Estimation
based on Quadratic Interpolation of FFT
Magnitude Peaks”, 117th Audio Engineering
Society Convention, San Francisco, CA, USA,
(2004 Oct.), Convention paper 6256.

[23]

M. Desainte-Catherine and S. Marchand, “High
Precision Fourier Analysis of Sounds Using
Signal Derivatives”, Journal of the Audio
Engineering Society, vol. 48, no. 7/8, pp. 654667 (2000 Jul/Aug.).

[24]

J. B. Allen, “Short Term Spectral Analysis,
Synthesis, and Modification by Discrete
Fourier Transform”, IEEE Transactions on
Acoustics Speech and Signal Processing, vol.
25, no. 3, pp. 235-238, (1978 Jun.).

[25]

J. O. Smith and X. Serra, “PARSHL: An
Analysis/Synthesis Program for Non-Harmonic
Sounds Based on a Sinusoidal Representation”,
Proceedings of the International Computer
Music Conference, Tokyo, (1987).

[11]

[12]

[13]

C. Grigoras, “Applications of ENF Criterion in
Forensic Audio, Video, Computer and
Telecommunication Analysis”, Forensic
Science international, vol. 167, pp. 136-143
(2007).
C. Grigoras, “Application of ENF Analysis
Method in Authentication of Digital Audio and
Video Recordings”, 123rd Audio Engineering
Society Convention, New York, USA, (2007
Oct.), Convention paper 1273.
M. Kajstura, A. Trawinska, J. Hebenstreit,
“Application of the Electrical Network
Frequency (ENF) Criterion, A Case of a Digital
Recording”, Forensic Science international,
vol. 155, pp. 165-171 (2005).

[14]

Union for the Co-ordination of Transmission of
Electricity, Homepage. http://www.ucte.org

[15]

National Grid Company, Homepage.
http://www.nationalgrid.com

[16]

UK Electricity Networks, Postnote,
Parliamentary Office of Science and
Technology, no. 163 (2001 Oct.).
http://www.parliament.uk/post/pn163.pdf

[26]

Electricity in the UK, Postnote, Parliamentary
Office of Science and Technology, no. 280
(2007 Feb.).
http://www.parliament.uk/post/pn280.pdf

F. J. Harris, “On the Use of Windows for
Harmonic Analysis with the Discrete Fourier
Transform”, Proceedings of the IEEE, vol. 66,
no. 1, pp. 51-83, (1978 Jan.).

[27]

D. C. Rife and R. R. Boorstyn, “Single Tone
Parameter Estimation from Discrete-Time
Observations”, IEEE Transactions on
Information Theory, vol. 20, no. 5, pp. 591598, (1974 Sep.).

[17]

[18]

T. S. Sidhu, “Accurate Measurement of Power
System Frequency Using a Digital Signal
Processing Technique”, IEEE Transactions on
Instrumentation and Measurement, vol. 48, no.
1, pp. 75-81 (1999 Feb.).

AES 33rd International Conference, Denver, CO, USA, 2008 June 5–7

10