Incorporating the Biometric Voice Technology into the E-Government
Systems to Enhance the User Verification

KHALID T. AL-SARAYREH AND RAFA E. AL-QUTAISH
Faculty of Information Technology
Applied Science University
Amman 11931
JORDAN
khalid_sar@yahoo.com, rafa@rafa-elayyan.net

MOHAMMED D. AL-MAJALI
College of Educational Sciences
Mu’tah University
Karak 61710
JORDAN
mdmajali@mutah.edu.jo


Abstract:- Many courtiers around the world have started their e-government programs. E-government portals
will be increasingly used by the citizens of many countries to access a set of services. Currently, the use of the
e-government portals arises many challenges; one of these challenges is the security issues. E-government
portals security is a very important characteristic in which it should be taken into account. In this paper, we
have incorporated the biometric voice technology into the e-government portals in order to increase the security
and enhance the user verification. In this way, the security should be increased since the user needs to use his
voice along with his password. Therefore, no any unauthorized person can access the e-government portal even
if he/she knows the required password.


Keywords:- Security Systems, E-Government Portals, Biometric Voice, Speaker Verification, Authentication.

1 Introduction
Security approaches can be broken down into two
approaches: passive (authentication) and active
(identification). Passive approaches are like a shield
in that they protect against a clear and present
danger such as a hacker attempting to access a
computer system, while active approaches are more
like prevention via a preemptive strike as in
arresting terrorists before they plant a bomb.
All cards, keys, and username/password
combinations have a common flaw: anybody can
use them. Credit cards can also be easily
counterfeited, and even the most sophisticated card
can be lost, stolen, or maliciously taken away. The
use of PINs and passwords somehow improves the
situation, but the fundamental problem with PINs is
that they identify a card but not its user. Obtaining
both the card and the PIN might be more difficult
than obtaining the card alone, but is quite feasible,
particularly if the owner of the card is forced to
cooperate. Thus, cards, PINs and passwords can
hardly provide highly secure solutions. The same
flaw applies to the username/password combination:
the password really identifies the username, not the
actual user! While all of the traditional approaches
have their strengths, they also have corresponding
weaknesses.
Whereas the requirement for physical access
security has existed since time immemorial,
computer threats and problems gained prominence
during the 1990s due to the explosive growth of
Internet, e-commerce and other computer
technologies. Table 1 summarizes the computer
threats as perceived in 1992 and 2002. The Table
shows that the number and severity of threats to
networked computers have caused into question
traditional approaches to security and demand a new
response from the IT to deal with the E-world of
tomorrow.
From Table 1, we noted that typing a text
password to open a system is not enough to identify
the current user because:
1. Authorized persons give their own passwords to
unauthorized ones.
WSEAS TRANSACTIONS on COMPUTERS
Manuscript received Nov. 1, 2007; revised May 3, 2008
Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 435 Issue 5, Volume 7, May 2008

2. Clever intruders may guise the password of this
system if the given time is enough.
3. Smart programs with intelligence methods that
can find the password in short time are available.
4. This password is used by a group of people.
5. Tracking and identifying the unauthorized
persons who tried to inter the system.

Table 1: The Computer Threats as Perceived in 1992 and 2002 [1]
Most severe threats in 1992 Most severe threats in 2002

- Natural Hazards
- Inadequate Control over
Media
- Weak and Ineffective
Controls Hacking
- Access to System by
Competitors
- Hacking

- Viruses
- System penetration:
Hacking/Espionage
- Fictitious people/ Perpetrators
- Denial of Service
- Insider abuse of net access
- Unauthorized access by Insiders
- Natural Hazards
- Human Error

- Infringement of IP rights
- Spoofing
- Implied trust Exploitation
- Active Wiretap
- Sabotage
- Telecom Eavesdropping
- Repudiation
- Credit Card Fraud

Speaker verification is verifying a user’s identity
by his voice, which is assumed to be unique for that
person. Even a very secure cryptology system has
the chance of being cracked; a unique human voice
is very attractive to be used in systems where
security should be provided. In this paper, a text-
dependent speaker verification system is
implemented by using various feature extraction and
feature comparison methods.
The feature extraction methods are mel-
cepstrum, Linear Prediction (LP) coefficients and
fundamental frequency and magnitude. The feature
comparison methods are Vector Quantization (VQ)
and Dynamic Time Warping (DTW).
Many courtiers around the world just have
started their e-government programs. E-government
portals will be increasingly used by the citizens of
many countries to access a set of services.
Currently, there are many challenges of the use of
the e-government portals; one of these challenges is
the security issues. E-government portals security
is a very important characteristic in which it should
be taken into account. In this paper, we have
incorporated the biometric voice technology into the
e-government portals in order to increase the
security and enhance the user verification. In this
way, the security should be increased since the user
needs to use his voice along with his password.
Therefore, no any unauthorized person can access
the e-government portal even if he gets a password.
The rest of this paper is organized as follows:
Section 2 introduces an overview of the previous
work in this topic, Section 3 presents an overview of
the proposed system, Section 4 describes the
methods which we used in our system, Section 5
gives the details of implementation of the speaker
verification, Section 6 illustrates some testing
examples with their results, Section 7 discusses the
results in the previous section. Finally, Section 8
concludes the paper and gives some trends on future
works.


2 Literature Review
Voice is a combination of physiological and
behavioral biometrics. The features of an
individual’s voice are based on the shape and size of
the appendages (e.g., vocal tracts, mouth, nasal
cavities, and lips) that are used in the synthesis of
the sound. These physiological characteristics of
human speech are invariant for an individual, but
the behavioral part of the speech of a person
changes over time due to age, medical conditions
(such as a common cold), and emotional state, etc.
Voice is also not very distinctive and may not be
appropriate for large-scale identification.
Based on the spoken text, there are two types of
the biometric voice recognition systems, that is, text
dependent and text independent. A text dependent
voice recognition system is based on the utterance
of a fixed predetermined phrase. A text independent
voice recognition system recognizes the speaker
independent of what he/she speaks, as we used in
our proposed system.
A text-independent system is more difficult to
design than a text-dependent system but offers more
protection against fraud. The problem with the voice
recognition is that speech features are sensitive to a
number of factors such as background noise,
medical conditions, etc [2, 3]. We solved this
problem by adding different voices in different
situations (for the same person) to learn the system
WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 436 Issue 5, Volume 7, May 2008

on the different possibilities of variant voices for the
same person.
Furthermore, for the text dependent, Monrose et
al. [4] have introduced an algorithm to generate a
cryptographic key from a user’s utterance of a
password. In addition, Jian and Ross [5] have stated
that the NIST (National Institute of Standards and
Technology) in 2002 implemented a text dependent
biometric voice system and get 10-20% of false
reject rate and 2-5% false accept rate.


3 The Proposed System: A General
Overview
A speaker verification system consists of two parts:
feature extraction and template based comparison.
In the first part, the system records the user’s speech
and applies short-term analysis to find the
parameters that distinguish the user. These
parameters are then recorded as the reference
template for any future comparison [7]. The next
time, when the user needs to be verified, he speaks
the same utterance. Short term analysis is again
performed, and the parameters are extracted once
more. In order to give access to the user, the new
extracted parameters are compared with the
previously recorded (reference) parameters. If the
similarity between the two parameter sets is above a
given threshold, the user is verified. Otherwise,
access is denied. The block diagram of the system is
shown in Figure 1.



Figure 1: Block diagram of the speaker verification

4. Methods Used in the System
4.1 End Point Detection
The first technique that applies to each speech
waveform is end detection. Due to the nature of the
recording process, speech waveforms have blank
parts in the beginning and in the end. These parts
are purely noise [6], and they do not contain any
information. As a result, the speech parameters of
these parts are independent of the Speaker and
identical in terms of statistics. Because of this fact,
these parts increase the correlation between different
users. This will cause the false acceptance ratio to
increase. Moreover, it is easier to overlap identical
parts of speech when the actual speech data starts at
time t0. Since the duration of the speech is reduced,
by using this method, the computational load is also
reduced. Endpoint detection algorithms generally
use the combination of zero-crossing rate and
energy OR the combination of zero-crossing rate
and average magnitude of a given utterance. Steps
of the end-point detection algorithm are listed as
follows:
• Removing the DC part (mean) of the speech
signal, which is considered to be an important
step because the zero-crossing rate of the signal
is calculated and it plays a role in determining
where the unvoiced sections of speech exist. If
the DC offset is not removed, we will be unable
to find the zero-crossing rate of noise in order
to eliminate it from our signal.
• Framing speech signal.
• Computing the total zero-crossings number of
each frame such that:

∑
=
−−=
N
m
mkxmkxkZ
1
))1(sgn())(sgn()( (1)
WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 437 Issue 5, Volume 7, May 2008

Where Z(k) is the average magnitude of the kth
frame and sgn(x) is the signum function that
finds the sign of x.
• Compute noise statistics (mean & standard
deviation) and the maximum of average
Magnitude and zero-crossings of the first five
frames assuming that the first five frames are
inconsiderable parts.
• Find end-points according to the average
magnitudes assuming that the average
magnitude of the spoken parts is greater than a
threshold TA (sum of the mean and standard
deviation in our paper) determined by the mean
and the standard deviation of the average
magnitude of the noisy part. End-points should
be in pair such that one of them is the
beginning of a specific spoken part and the
other one is the end of that part.
• Correct end-points according to the zero-
crossings assuming that the zero crossings of
spoken parts are smaller than a threshold TZ
(sum of the mean and standard deviation in our
paper) determined by the mean and the
standard deviation of the average magnitude of
the noise part. End point detection is applied to
all of our speech signals before they are
processed.
Figures 2 and 3 below show the plot of a two-
speech signal with and without end-detection. In
addition, Figure 4 shows the block diagram of the
end point detection algorithm.



Figure 2: Speech signal before end detection



Figure 3: Speech signal after end detection
WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 438 Issue 5, Volume 7, May 2008

Figure 4: block diagram of end point detection algorithm

4.2 Feature Extraction
In order to compare different users, we need to
extract the parameters from the speech signal. These
parameters help us to distinguish one user from the
others. However, the speech signal must first be
divided into frames. The parameters are calculated
separately for each frame. For the framing part, we
choose to use a hamming window with a window
size of 30msec and skip a rate (frame overlap) of
10msec. After each feature extraction algorithm, we
get a matrix as a result. The columns of these
matrices are the feature vectors for each frame [8].

4.2.1 Mel-Spectrum
Mel-spectrum coefficients are known to model the
human ear well, and it is mentioned in the literature
few times that the performance of the mel-cepstrum
coefficients is better than the performance of other
parameters. Mel-cepstrum coefficient extraction is
based on the fact that the sensitivity of the human
ear varies with frequency. Our ear is more sensitive
to low frequency sounds than to high frequency
ones. The computation of the extraction of these
parameters will be explained next.
1. Apply framing to the end detected speech signal.
2. Then the 512-point fast Fourier transform of
each frame is applied and the algorithm of the
magnitude response is taken. Due to the
symmetry property of the FFT, only the first 256
samples (Si) are processed.
3. The frequency range (0-fs/2 Hz) is divided into
uniformly spaced triangular overlapping
windows (K windows in total). Fig. 5 illustrates
the Mel spectrum scale.



Figure 5: Mel spectrum scale

4. Then the Mel scale is converted back to the
linear scale by the following equation:
f =700× (10 1/2595) (2)
In this case, the resulting mel-wrapped frequency
filter bank triangles become linearly spaced in low
frequencies and exponentially spaced at high
frequencies as shown in Figure 6.
WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 439 Issue 5, Volume 7, May 2008

e
Figure 6: Linear Space

• Each triangular window (wk) is multiplied with
the corresponding parts of the FFT, and the
resultant window becomes (Sik *wk). Then for
each window, a single number is found
according to the following equation:
∑ ×= kWikSikm (3)
• After applying the above procedure to K
window, a vector Pi of length K is obtained
such that [ ]ikmimimiP . . . 2 1= (4)
• Finally, the Inverse Discrete Cosine Transform
(IDCT) of Pi is calculated, and the first n many
coefficients are taken. The obtained mel-
cepstrum coefficient for the frame will be:

` [ ] IDCTinciciciC == . . . 2 1 (5)
When this procedure is applied to I frames, the
mel-cepstrum matrix is found as:

⎥⎥
⎥⎥
⎦
⎤
⎢⎢
⎢⎢
⎣
⎡
=
IncAncnc
MOMM
IcAcc
IcAcc
C
21
22212
12111
(6)

In our simulations, we have used 30 filter banks
(K) and 12 Cepstrum coefficients (n).
The block diagram of the feature extraction part
is given in Figure 7.


Figure 7: Block diagram of the feature extraction

4.2.2 LP-Coefficients
LPC based feature extraction is the most widely
used method by developers of speech recognition
systems [9]. The main reason is that speech
production can be modeled completely by using LP
analysis. Besides, LPC based feature extraction can
WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 440 Issue 5, Volume 7, May 2008

also be used in speaker recognition systems, where
the main purpose is to extract the vocal tract
parameters from a given utterance [10]. In speech
synthesis [11], linear prediction coefficients are the
coefficients of the FIR filter representing a vocal
tract transfer function. Therefore, linear prediction
coefficients are suitable to use as a feature set in
speaker verification systems.
The general idea of LP is to determine the
current sample by a linear combination of p
previous samples where the linear combination
weights are the linear prediction coefficients [12].
Therefore, the LP polynomial can be written as:
∑
=
−=
∧ p
k
knxkanx
1
)()()( (7)
Where x is the input speech frame. In our algorithm,
LP coefficients must be calculated for each frame.
The autocorrelation method is used to extract LP
coefficients.
The algorithm of the autocorrelation method is as
follows:
• Find (p+1) autocorrelation matrix elements
R(k)’s such that:

∑−−
=
=+=
kN
m
kmxmxkR
1
0
p . . . 1, 0,k ),()()( (8)

Where N is equal to the frame size in samples
• Compute p autocorrelation coefficients α(k)’s
as follows:

⎥⎥
⎥⎥
⎦
⎤
⎢⎢
⎢⎢
⎣
⎡−
⎥⎥
⎥⎥
⎦
⎤
⎢⎢
⎢⎢
⎣
⎡
−−
−
−
=
⎥⎥
⎥⎥
⎦
⎤
⎢⎢
⎢⎢
⎣
⎡
⇒
⎥⎥
⎥⎥
⎦
⎤
⎢⎢
⎢⎢
⎣
⎡
=
⎥⎥
⎥⎥
⎦
⎤
⎢⎢
⎢⎢
⎣
⎡
⎥⎥
⎥⎥
⎦
⎤
⎢⎢
⎢⎢
⎣
⎡
−−
−
−
)(
...
)1(
)0(1
)0(...)2()1(
............
)2(...)1()1(
)1(...)1()0(
)(
...
)2(
)1(
)(
...
)2(
)1(
)(
...
)2(
)1(
)0(...)2()1(
............
)2(...)1()1(
)1(...)1()0(
pR
R
R
RpRpR
pRRR
pRRR
pa
a
a
pR
R
R
pa
a
a
RpRpR
pRRR
pRRR
(9)

• Repeat Steps 1 and 2 for each frame of the
utterance.
To see the effect of the linear prediction order in
our system, we have used different values for “p”.
One of the values has been chosen according to the
following formula:
2
1000
+= sfP (10)
The sampling frequency is 16kHz, The block
diagram of the LPC calculation is given in Figure 8.



Figure 8: Block diagram of the LPC calculation

4.3 Fundamental Frequency (Pitch) and
Magnitude
4.3.1 Short-Time Magnitude
Using an endpoint detection algorithm, the speech is
selected from the two input waveforms, and then
their short-time magnitudes are determined. The
short-time magnitude characterizes the envelope of
a speech signal by lowpass, filtering it with a
rectangular window [13]. The magnitude function
follows these steps:
• The bounds of the signal are determined, and
each end is zero-padded.
• The signal is convolved with a rectangular
window. As the window is swept across the
signal, the magnitude contained within the signal
is summed and plotted at the midpoint of the
window's location. This provides the speech with
the cover needed for the speech to be uttered
correctly.
One magnitude plot is discrete time warped onto
the other, see Figure 9. The dot product of the two
waveforms is computed, and this number is divided
by the product of the signals' norms. This
calculation results in a percentage of how similar
one signal is to another.

WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 441 Issue 5, Volume 7, May 2008

Figure 9: short time magnitude of the signal

4.3.2 Short-Time Frequency
A simple model of the human vocal tract is a
cylinder with a flap at one end. When air is forced
through the tract, the vibration of the flap is
periodic. The inverse of this period is known as the
fundamental frequency, or pitch. This frequency,
combined with the shape of the vocal tract, produces
the tone that distinguishes your voice. Variations in
people's vocal tracts result in different fundamentals
even when the same word is said. Therefore, pitch is
another characteristic of speech that can be matched
[14].
Since harmonic peaks occur at integer multiples
of the pitch frequency, we can compare peak
frequencies at each time t to locate the fundamental
to extract pitch from signals, making use of a
harmonic-peak-based method.. The implementation
finds the three highest-magnitude peaks for each
time, then thw implementation computes the
differences between them. Since the peaks should be
found at multiples of the fundamental, now their
differences should represent multiples as well. Thus,
the differences should be integer multiples of one
another. Using the differences, the implementation
derives the closest voice for the fundamental
frequency.
The pitch frequency is computed at each time,
and this gives the pitch track of the signal. A major
advantage of this method is that it is very noise-
resistive. Even as the noise increases, the peak
frequencies should still be detectable above the
noise. It is also easily implemented in MATLAB
[15].
The first step is to find the signal's spectrogram.
The spectrogram parameters decided here are a
window length of 512 points and a sampling rate of
10000 Hz. assuming that the fundamental frequency
(pitch) of any person's voice will be at or below
1200 Hz, so when finding the three largest peaks,
only consider sub-1200 Hz frequencies, cutting out
the rest of the spectrogram. Before that, use the
whole spectrogram to find the signal's energy [16].
The signal's energy at each time is very
important as it shows the voiced and unvoiced areas
of the signal, with voiced areas having the higher
energies. Since using our pitch track to compare the
pitch between signals, be certain that the
comparison held only for the voiced portions, and
the areas where the pitch will be distinct between
two different people. A plot of energy versus time
can actually be used to window the pitch track so
that only the voiced portions are taken.
To find the energy versus time window, take the
absolute magnitude of the spectrogram and then
square it. According to Parseval's Theorem, adding
up the squares of the frequencies at each time gives
us the energy of the signal there Plotting this versus
time gives us our window.
Once this is done, cut the spectrogram and move
on to finding the three largest peaks at each time. A
frequency is designated as a "peak" if the
frequencies directly above and below it have smaller
magnitudes than it does. If a frequency is a peak,
then its magnitude is compared to the three
magnitude values stored in the "peak matrix" (a
matrix of magnitudes and locations for the three
highest peaks which start out as zeros at each time).
If it is greater than the minimum matrix value, then
its magnitude and location replace the magnitude
and location of the matrix's smallest peak.
WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 442 Issue 5, Volume 7, May 2008

Figure 10: pitch track for entire signal (include background noise)

The matrix of peak values and locations at each
time is then fed through the fundamental frequency
algorithm, and we have our uncut pitch track
(above) [17]. At this point, go back to the energy
versus time plot and use it to find the energy
threshold of the noise and unvoiced areas that cut
out of the pitch track. This is done by finding the
mean and standard deviation of the very beginning
of the signal (assumed to be noise as the person
never begins speaking until at least half a second
into the recording due to mental processing time)
and using these to develop the threshold. Then, the
pitch track is windowed with the energy signal, and
everything below the threshold is cut out (below).
This gives a pitch track of the voiced portions of the
signal. It is now ready for comparison with another
signal.



Figure 11: Energy in signal derived from spectrogram
WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 443 Issue 5, Volume 7, May 2008

Figure 12: pitch track with unvoiced regions reduced to Zero

Pitch track comparison takes two signals and
finds each of their pitch tracks. It then maps the
pitch tracks onto one another using dynamic time
warping. After mapping, take the dot product of the
two tracks and equally divide it by the norms of the
tracks to find the percent that they match. This is
done twice, mapping the first signal onto the second
and then vice versa, and then the highest dot product
is taken as the matching correlation.

4.4 Feature Comparison
After the feature extraction step, the similarity
between the parameters derived from the spoken
utterance [18] and the reference parameters need to
be computed. The three most commonly
encountered algorithms in the literature are:
Dynamic Time Warping (DTW), Hidden Markov
Modeling [19] (HMM) and Vector Quantization
(VQ). In our paper, we use DTW for comparing the
parameter matrices extracted in the previous section.

4.5 Dynamic Time Warping
One of the difficulties in speech recognition is that
although different recordings of the same words
may include more or less the same sounds in the
same order, the precise timing - the durations of
each sub word within the word - will not match. As
a result, efforts to recognize words by matching
them to templates will give inaccurate results if
there is no temporal alignment [20].
Although Hidden Markov Models (HMM) have
largely superseded it, early speech recognizers used
a dynamic-programming technique called Dynamic
Time Warping (DTW) to accommodate differences
in timing between sample words and templates. The
basic principle is to allow a range of 'steps' in the
space of (time frames in samples, time frames in
templates) and to find the path through that space
that maximizes the local match between the aligned
time frames, subject to the constraints implicit in the
allowable steps. The total `similarity cost' found by
this algorithm is a good indication of how well the
sample and template match, which can be used to
choose the best-matching template.
Because of the emotional instability of the user,
some utterances of the user might differ from each
other. The reference recording might sound
“project”, and the compared recording might sound
like “prooject” depending on the mood of the user.
Obviously, a simple linear squeezing of this longer
password will not match the key signal because the
user slowed down the first syllable while he kept a
normal speed for the "ject" syllable. (i.e., can
compare "Prrrooo" to "Pro" and "ject" to "ject").
DTW is used to align the two different samples
of the same utterance in the time domain.
Furthermore, the algorithm is given as follows:
• Each column of the parameter matrix is
considered as a point in the vector space.
• A Local Distance Matrix (LDM) is defined
such that each element of the matrix is a
distance from one point (parameter matrix
column) to another. As (wrong use of as, unless
you mean as it's here, it's there, so it should be
without and) the distance metric one is used for
the LP coefficients, another distance is used for
the MFCC and spectral coefficients. If the
numbers of columns in the reference and the
new parameter matrices are M and N
respectively, the LDM becomes an MxN
matrix.
• An accumulated distance matrix is computed
such that each element of the matrix contains
the corresponding LDM matrix element plus
the smallest neighboring accumulated distance.
WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 444 Issue 5, Volume 7, May 2008

• To calculate the ADM, we start by equating
ADM(1,1) to LDM(1,1)
• The other elements of the ADM are calculated
as follows:
ADM (m,n) = LDM(m,n) + min { ADM(m,n-1),
ADM(m-1,n-1), ADM(m-1,n)}
• When the lower right corner of the matrix is
reached, this value shows the Minimum global
distance between the two-parameter matrices.
If two identical matrices are fed as the input to
this algorithm, the minimum global distance
becomes zero. The output of the DTW algorithm is
the minimum global distance value. Small values
mean similar parameter matrices, and high values
indicate that the two parameter matrices are highly
unlikely.

4.6 Decision Making and its Function
There are usually 3 approaches to construct the
decision rules:
• Geometric.
• Topological.
• Probabilistic rules.
The most popular speech verification models are
based on probabilistic rules [21].
In classical speech verification systems, when no
prior information is given on the cost of the
different kinds of errors, the Bayes Decision rule is
applied by selecting the value of Δi in (5) that
minimizes the Half Total Error Rate:

HTER = 1/2(%FA +%FR) (11)
where %FA is the rate of false acceptances, and
%FR is the rate of false rejects. Note that this cost
function changes the relative weight of client and
impostor accesses in order to give them equal
weight, instead of the one induced by the training
data.
If the probabilities are perfectly estimated, then
the Bayes Decision is the optimal decision.
Unfortunately, this is usually not the case. In that
case, the Bayes Decision might not be the optimal
solution, and we should thus explore other forms of
decision rules. In this thesis, we will discuss two
sorts of decision rules, which are based either on
linear functions or on more complex functions such
as Support Vector Machines.


5. Implementation of the Speaker
Verification
5.1 Step 1: Trained the System and Build a
Dataset
When a speaker attempts to verify himself with this
system, his or her incoming signal is compared to
that of a "key" [22]. This key should be a signal that
produces a high correlation for both magnitude and
pitch data when the authorized user utters the
password, but not in cases where:
• the user says the wrong word (the password is
forgotten).
• an intruder says either the password or a wrong
word.

5.2 Step 2: Verification
Once a key has been established, an authorization
attempt breaks down into the following steps:
1. The person (for example, person 1) utters the
password using the microphone.



Figure 13: Person 1 signal (sig1)
WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 445 Issue 5, Volume 7, May 2008

2. The short-time frequency of the signal is found
and recorded just as the pitch track is.
PITCH takes a spectogram-MxN matrix
returned from spectrogram function and a
corresponding sample - Scalar holding sample
rate - to create a pitch track and a plot of the
fundamental frequency versus time. It then
eliminates false pitch estimates by zeroing the
areas of low energy because no one is talking or
because a fricative voice occurs, see Figures 14,
15 and 16.



Figure 14: Spectrogram of the key signal



Figure 15: Spectrogram of the new signal



Track1 Track2

Figure 16: Track signals

To prepare for a Dynamic Time Warp (DTW) [23]:
1- Cut the trailing and leading zeros of track1, see
Figure 17.a.
2- Cut the trailing and leading zeros of track2, see
Figure 17.b.
3- Perform the Dynamic Time Warping by mapping
track2 onto track1, and calculating the dot
product, then map track1 onto track2 and
calculate the dot product, and finally, map track2
WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 446 Issue 5, Volume 7, May 2008

onto track1 and calculate the dot product (see
Figure 18).
Where:
track2dtw=track2cut (path2on1)
dotWDTW1 = dot(track1cut,track2dtw)
/(norm(track1cut)*norm(track2dtw));
4- Map track1 onto track2 and calculate the dot
product, as in Figure 19.



a. Track1cut b. Track2cut

Figure 17: Track signals after cutting



Path2on1 Track2dwt


Figure 18: Path2on1 and Track2dtw



Path1on2 Track1dwt

Figure 19: Path1on2 and Track1dtw

WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 447 Issue 5, Volume 7, May 2008

Where:
track1dtw = track1cut (path1on2);
dotWDTW2=
dot(track2cut,track1dtw)/(norm(track2cut)*norm
(track1dtw));
Take Larger Dot Product as Result
corr = pitchmaster (sig, locksig)
ans = 0.9996 so that pitch=0.9996
5. The short-time magnitude of the signal is found
and recorded, as is the magnitude. Then, find out
where to cut the two signals and find the average
magnitude for the two cutting signals, as in
Figure 20.
6- Perform Dynamic Time Warping (DTW) linking
one signal to another in order to let the peaks
match up. The time warp is performed both ways
(matching sig1 to sig2 and vice versa).
The output will be the best match of the two
signals.
1- Let sig1 be the key and sig2 be time-warped to
sig1 (see Figure 21.a).
2- Let sig2 be the key and sig1 be time-warped to
sig2 (see Figure 21.b).



mag1 mag2

Figure 20: Magnitude for the two cutting signals



a. sig1cut b. sig2cut

Figure 21: Sign1 and Sign2 cuts

WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 448 Issue 5, Volume 7, May 2008

Then for each signal (DTW), the following
calculations are performed (see Figure 22):
num = dot(mag1,mag2(path1));
den = (norm(mag1))*(norm(mag2(path1)));
final1 = num/den;
num = dot(mag1(path2),mag2);
den = (norm(mag1(path2)))*(norm(mag2));
final2 = num/den;
Take the larger result which represents the mag
value
>> final1
final1 = 0.9916
>> final2
final2 = 0.9952
So that mag=0.9916
7- These numbers are compared to the thresholds
(0.95). If both the magnitude and pitch
correlations are above this threshold, the speaker
has been verified.



path1 path2

Figure 22: Path1 and Path2

6. Experimentation and Results
To develop such a key, the system is trained for the
recognition of the speaker. In this instance, the
speaker first chooses a password, which is acquired
five separate times as shown in Figure 23. The pitch
and magnitude information are recorded for each.
The signal that matches the other four signals best in
both cases is chosen as the key as shown in Figure
24.
Furthermore, the system will return lower
thresholds for matching to magnitude vs. time and
pitch vs. time. Thresholds below 0.90 are not
returned, and to eliminate the possibility of an
extremely high threshold like 0.99, an upper bound
is placed on the thresholds of 0.95
In MATLAB, the function makelock.m was
written to determine the key signal from five
possible signals. The call is: [Lock signal] =
makelock (sig1, sig2, sig3, sig4, sig5)
In this instance, Khalid records his chosen
password "project" five times and saves them as
sig1, sig2, etc. When the call is made, the results are
assigned to an array lock (which holds the time
signal, which is a large array of points) and two
scalars with the lower threshold bounds.
>> [Lock, pitchthreshold, magthreshold]
=makelock (sig1, sig2, sig3, sig4,sig5);
>> pitchthreshold
pitchthreshold=0.9500
>> magthreshold
magthreshold=0.9500
>> >> sum (sig1-locksig)
ans =-2.4974
>> sum (sig2-locksig)
ans = -2.4199
>> sum (sig3-locksig)
ans = 0
>> sum (sig4-locksig)
ans = -2.8141
>> sum (sig5-locksig)
ans = -1.8363
Based on the above, sig3 has been chosen as the
key.
WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 449 Issue 5, Volume 7, May 2008

Figure 23: Five Signals of the Same Person



Figure 24: Key Lock Signal

6.1 Examples
>> speak now
Key and results: Pitch 0.99955,
Magnitude 0.98929,
*****MATCH****
>> speak now
Key and results: Pitch 0.99957,
Magnitude 0.98537,
*****MATCH****
>> speak now
Key and results: Pitch 0.94346,
Magnitude 0.97949,
***** NO MATCH *****
>> speak now
Key and results: Pitch 0.95588,
Magnitude 0.95045,
*****MATCH****
>> speak now
Key and results: Pitch 0.99937,
Magnitude 0.98412,
*****MATCH****
>> speak now
Key and results: Pitch 0.95457,
Magnitude 0.82971,
*****NO MATCH *****

6.2 Results
To test the system, the two members of the group
(Person 1, Person 2) were set as keys during the
runtime of the system. The basic results are shown
in Table 2 and Table 3.

Table 2: Comparing Ratio

Person 1 Person 2
Person 1 67% 17%
Person 2 19% 74.5%

WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 450 Issue 5, Volume 7, May 2008

Table 3: Result table

Person 1 Person 2
Person 1 and Person 2
(with Password) 67% 74.5%
Intruder (with
Password) 21% 18.5%


7. Discussion
Through the previous section, we have provided a
set of examples to test the systems. In Table 2, we
gave two passwords for two persons; as a result we
noted that person 1 has accessed the system
successfully with ratio 67% using the given
password, while person 2 has accessed the system
successfully with ratio 74.5%, also with the given
password, using 0.95 thresholds. Subsequently, we
reduced the threshold to 0.90 and asked person 1
and person 2 to access the system, as a result, person
1 has accessed the system successfully with ratio
17%, while person 2 has accessed successfully with
ratio 19%
We gave an intruder the two passwords, and then
he has successfully accessed the system with ratios
21% and 18.5% using person 1 and person 2
passwords, respectively. Table 3 represents these
results.
In order to accomplish the results in Tables 2 and
3, person 1, person 2 and the intruder have tried to
access the system 100 times using the given
passwords. Furthermore, person 1 with person 2
password has tried to access the system 100 times,
and the same has been done for person 2 with
person 1 password. In addition, an intruder has been
given the passwords of person 1 and person 2, then
he has tried to access the system 200 times using the
passwords of person 1 and person 2.


8. Conclusion and Future Work
Many courtiers around the world have just started
their e-government programs. E-government portals
will be increasingly used by the citizens of many
countries to access a set of services. Currently, there
are many challenges of the use of the e-government
portals; one of these challenges is the security
issues. E-government portals security is a very
important characteristic in which it should be taken
into account. In this paper, we have incorporated
the biometric voice technology into the e-
government portals in order to increase the security
and enhance the user verification. In this way, the
security should be increased since the user needs to
use his voice along with his password. Therefore, no
any unauthorized person can access the e-
government portal even if he gets a password.
The proposed system was designed to be
installed on the server of the e-government to
connect clients (especially G2G) using the voice
instead of or beside the traditional username and
password. The reason behind that is that most of the
financial transactions are done via the portal of e-
Government. Using this system, it would not be
possible for a person to deny his or her access to the
portal as it will authenticate the person and make
sure it is the person allowed to access the data. This
system was not designed for Business to
Government (B2G) and Consumer to Government
(C2G) as they both have unlimited number of
people, which is not the case in the Government to
Government model.
Another application that makes use of all the
calculations and theorems mentioned in this thesis is
the handicapped chair, which was implemented and
scored a great success.
The training should be done over a period of
time to take into account differences in background
noise, the speaker's health, microphones, and other
various factors. Furthermore, an instantaneous key
could be skewed because the user will attempt to
sound similar at each training session, which will
actually produce dissimilarities.
A composite key could help alleviate this
problem by taking the five signals that are used in
training and producing an "average" of them in
terms of magnitude and fundamental frequency.
Furthermore, the results could improve over time if
the system took successful attempts and added them
to the average key. In this case, the recognition
could get better over time without extra training
sessions.
The current lower thresholds for matching
determined by makelock.m are set rather high, in a
range between 0.90 and 0.95. The reason why the
number is set high is to provide the utmost security
for your password in the worst case scenario, which
is when it is known to the intruder. As seen in the
results, intruders without knowledge of the
password were never successful in accessing the
system. Therefore, it is possible to set the thresholds
lower so that the owner's acceptance rates increase
while an intruder's acceptance rate remains nominal.
Finally, a speaker's voice has many other
characteristics that make it identifiable to the human
ear. Computationally, a system can also try to draw
WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 451 Issue 5, Volume 7, May 2008

out such unique differences and match those to users
as well. For example, formants are a function of a
person's vocal tract, so tracking such data could
improve further results. As mentioned earlier, using
Linear Predictive Coding (LPC) and spectrum
techniques may produce more accurate results that
increase the rates of acceptance and denial.
This system can be developed in the future
through connecting it to other software programs or
through the authentication process, which is done
through the Iris or the fingerprint. The voice will be
widely used in the coming software programs. The
system could also be installed on special chips as an
embedded system that would work on its own.


References :
[1] Jain, A. K.; Ross, A. and Prabhakar, S., “An
introduction to biometric recognition”, IEEE
Transactions on Circuits and Systems for Video
Technology, Vol. 14, No. 1, 2004, pp. 4-20.
[2] Monrose, F.; Reiter, M. K.; Li, Q. and Wetzel,
S., “Cryptographic Key Generation from
Voice”, in Proceedings of the IEEE Symposium
on Security and Privacy - S&P’01, Oakland,
CA, USA, May 14-16, 2001, pp. 202-213.
[3] Bagge N. and Donnica C., Text Independent
Speaker Recognition, online: http://www-
dsp.rice.edu/courses/elec301/Projects01/speak
er_id/index.html, accessed on March 16, 2008.
[4] Jian A. K. and Ross, A., “Multibiometric
Systems”, Communications of the ACM, Vol.
47, No. 1, 2004, pp. 34-40.
[5] O’Shaughnessy, D., Speech Communications:
Human and Machine, the Institute of Electrical
and Electronics Engineers - IEEE, Inc., New
York, 2000.
[6] Guo, S. and Liddel, H, “Support Vector
Regression and Classification Based Multi-
view Detection and Recognition”, in
Proceedings of the IEEE International
Conference on Automatic Face & Geusture
Recognition, 2000, pp. 300-305.
[7] Peacocks, R. and Graf, D., An Introduction to
Speech and Speaker Recognition, IEEE
Computer Magazine, Vol. 23, No. 8, 1990, pp.
26-33.
[8] Davis, S. B. and Mermelstein P., "Comparison
of parametric representations for Monosyllabic
word recognition in continuously spoken
sentences", IEEE Transactions on Signal
Processing, Vol. 28, No. 4, 1980, pp. 357-366.
[9] Gasem, M., Vector Quantization, online:
http://www.geocities.com/mohamedqasem/vect
orquantization/vq.html, accessed on March,
2008.
[10] Xafopoulus, A., Speaker Verification: an
Overview, Technical Report, Artificial
Intelligence & Information Analysis laboratory,
Informatics Department, Aristotle Univ. of
Thessaloniki, Thessaloniki, Greece, 2001.
[11] Childers, D. G., Speech Processing and
Synthesis Toolboxes, John Wiley & Sons Inc.,
New York, 2000.
[12] Gold, B. and Morgan, N., Speech and Audio
Signal Processing: Processing and Perception
of Speech and Music, John Wiley & Sons Inc.,
New York, 2000.
[13] L. Rabiner and B.-H. Juang, Fundamentals of
Speech Recognition, Prentice-Hall, Englewood
Cliffs, N.J., 1993.
[14] Minh, N., An Automatic Speaker Recognition
System, online: http://lcavwww.epfl.ch/
~minhdo/asr_project/asr_project.html,
accessed on March 16, 2008.
[15] L.R Rabiner and R.W. Schafer, Digital
Processing of Speech Signals, Prentice-Hall,
Englewood Cliffs, N.J., 1978.
[16] Matlab VoiveBox, online: http://www.ee.ic.ac.
uk/hp/staff/dmb/voicebox/voicebox.html,
accessed on March 16, 2008.
[17] Yu, K.; Mason, J. and Oglesby, J., “Speaker
Recognition Using Hidden Markov Models,
Dynamic Time Warping and Vector
Quantization”, in IEE proceedings on Vision,
Image and Signal Processing, Vol. 142, No. 5,
1995, pp. 313-318.
[18] Deller, J. R.; Proakis, J. G and Hansen J. H.,
Discrete-Time Processing of Speech Signals,
MacMillan, New York 1993.
[19] Arslan, L., Digital Speech Processing, EE 578
Lecture Notes, Boðaziçi University, Turkey,
spring 2001.
[20] Rafael, G. and Richard, W., Digital Image
Processing, 2nd edition, Prentice-hall, 2002.
[21] Fant, G., Acoustic Theory of Speech
Production and Some of its Implications,
Journal of Speech and Hearing Research, Vol.
4, pp. 303-320, 1961.
[22] Minh, N., An Automatic Speaker Recognition
System, Audio Visual Communications
Laboratory, Swiss Federal Institute of
Technology, Lausanne, Switzerland, February
2003.
[23] Ross, A. and Jain, A., Biometrics: Speaker
Verification, online: http://biometrics.cse.msu.
edu/speaker.html, accessed on March 16, 2008.

WSEAS TRANSACTIONS on COMPUTERS Khalid T. Al-Sarayrh, Rafa E. Al-Qutaish
and Mohammed D. Al-Majali
ISSN: 1109-2750 452 Issue 5, Volume 7, May 2008