Marsland Press
Journal of American Science 2009:5(2) 1-12
1
Abstract:
The lighting is available for the highways to avoid accidents and to make the driving safe and easy, but
turning the lights on all the nights will consume a lot of energy which it might be used in another important
issues. This paper aims at using the sound recognition techniques in order to turn the lights on only when
there are cars on the highway and only for some period of time. In more details, Linear Predictive Coding
(LPC) method and feature extraction will be used to apply the sound recognition. Furthermore, the Vector
Quantization (VQ) will be used to map the sounds into groups in order to compare the tested sounds.
[Journal of American Science 2009:5(2) 1-12] ( ISSN: 1545-1003)
Key word: Linear Predictive Analysis; Sound Recognition; Speaker Verification; Electricity Consumption
1. Introduction
Conserving Energy is one of the most
important issues in many countries since they have
limited resources of fuel to depend on, and they
may be import all their need of energy from other
countries. Therefore many conferences have been
held urge the people to conduct the consumption of
energy. This paper will introduce a system to
control the lighting of lamps in highways. The
system will turn the lights on only if there is a car
in the highway for a pre-defined period of time, and
will keep the lights off for any other sound.
Conserving energy of highways lights system
could be used to reduce the power invoice by
controlling the lights of lamps in the highways and
will save a lot of energy. The algorithms that define
the Conserving Energy of Street Lights system use
the Database which consists of 250 sounds of cars
and a lot of sounds from other domains.
2. An Overview of the Related Techniques
2.1 Voice Recognition
Voice recognition consists of two major tasks,
that is, Feature Extraction and Pattern Recognition.
Feature extraction attempts to discover
characteristics of the sound signal, while pattern
recognition refers to the matching of features in
such a way as to determine, within probabilistic
limits, whether two sets of features are from the
same or different domain [Rabiner and Juang,
1993]. In general, speaker recognition can be
subdivided into speaker identification, and speaker
verification. Speaker verification will be used in
this paper to recognize the sound of cars.
2.2 Linear Predictive Coding (LPC)
Linear predictive coding (LPC) is defined as a
digital method for encoding an analogue signal in
which a particular value is predicted by a linear
Using the Sound Recognition Techniques to Reduce the Electricity
Consumption in Highways
1 Khalid T. Al-Sarayreh, 2 Rafa E. Al-Qutaish, 3 Basil M. Al-Kasasbeh
1 School of Higher Technology (ÉTS), University of Québec, Montréal, Québec H3C 1K3, Canada
2 Faculty of IT, Alzaytoonah University of Jordan, Airport Street, PO Box: 130, Amman 11733, Jordan
3 Faculty of IT, Applied Science University, PO Box: 926296, Amman 11931, Jordan

Using the Sound Recognition Techniques Khalid T. Al-Sarayreh et al.
2
function of the past values of the signal. It was first
proposed as a method for encoding human speech
by the United States Department of Defence (DoD)
in federal standard, published in 1984. The LPC
model is based on a mathematical approximation of
the vocal tract. The most important aspect of LPC
is the linear predictive filter which allows
determining the current sample by a linear
combination of P previous samples. Where, the
linear combination weights are the linear prediction
coefficient.
The LPC based feature extraction is the most
widely used method by developers of speech
recognition. The main reason is that speech
production can be modelled completely by using
linear predictive analysis, beside, LPC based
feature extraction can also be used in speaker
recognition system where the main purpose is to
extract the vocal tract [Nelson and Gailly, 1995].
2.3 Vector Quantization (VQ)
The quantization is the process of representing
a large set of values with a much smaller set
[Sayood, 2005]. Whereas, the Vector Quantization
(VQ) is the process of taking a large set of feature
vectors, and producing a smaller set of feature
vectors, that represent the centroids of the
distribution, i.e. points spaced so as to minimize the
average distance to every other point.
However, optimization of the system is
achieved by using vector quantization in order to
compress and subsequently reduce the variability
among the feature vectors derived from the frames.
In vector quantization, a reproduction vector in a
pre-designed set of K vectors approximates each
feature vector of the input signal. The feature
vector space is divided into K regions, and all
subsequent feature vectors are classified into one of
the corresponding codebook-elements (i.e. the
centroids of the K regions), according to the least
distance criterion (Euclidian distance) [Kinnunen
and Frunti, 2001].
2.4 Digital Signal Processing (DSP)
The Digital Signal Processing (DSP) is the
study of signals in a digital representation and the
processing methods of these signals [Huo and Gan,
2004]. The DSP and analogue signal processing
are subfields of signal processing. Furthermore, the
DSP includes subfields such as audio signal
processing, control engineering, digital image
processing, and speech processing. RADAR Signal
processing, and communications signal processing
are two other important subfields of DSP [Lyons,
1996].
2.5 Frequency Domain
The signals are converted from time or space
domain to the frequency domain usually through
the Fourier transform. The Fourier transform
converts the signal information to a magnitude and
phase component of each frequency. Often the
Fourier transform is converted to the power
spectrum, which is the magnitude of each
frequency component squared. This is one of the
features that we have depended on in our analysis
[Fukunaga, 1990].
2.6 Time Domain
The most common processing approach in the
time or space domain is enhancement of the input
signal through a method called filtering. Filtering
generally consists of some transformation of a
number of surrounding samples around the current
sample of the input or output signal. There are
various ways to characterize filters [Smith, 2001].
Most filters can be described in Z-domain (a
superset of the frequency domain) by their transfer

Marsland Press
Journal of American Science 2009:5(2) 1-12
3
functions. A filter may also be described as a
difference equation, a collection of zeroes and
poles or, if it is an FIR filter, an impulse response
or step response. The output of an FIR filter to any
given input may be calculated by convolving the
input signal with the impulse response. Filters can
also be represented by block diagrams which can
then be used to derive a sample processing
algorithm to implement the filter using hardware
instructions [Garg, 1998].
3. The Methodology
In this paper, first we have collect many
samples for cars sound from many areas. Then the
feature extraction was applied on the sound. The
sound was passed through a high-pass filter to
eliminate the noise. The extraction of the LPC
coefficient, the magnitude of the signal, and the
pitch of the signal were made. These features were
normalized and clustered into codebooks using
vector quantization and the Linde-Buzo-Gray
(LBG) algorithm for clustering which based on the
k-mean algorithm. Finally a comparison with the
template database that we have built before was
made.
3.1 Database
The database which was used in this system
was built from the recorded sounds which we
record from different places, also from sounds for
rain, thunder, and plane which we have brought
them from internet, also from different human
sounds. For the cars, rain, and plane groups, vector
quantization method is used for clustering based on
LBG algorithm and k-mean algorithm, and the
Euclidian distance for matching. Statistical
analyses were used for the human group, since the
sounds of the human are very different and can’t be
bounded. Statistical analyses were based on the
power spectrum of the sound then the mean and
slandered deviation was taken to make the
comparison.
3.2 Collecting Samples
We have collected about 250 sample of car’s
sound from different places beside the highways.
These samples were taken after the mid night to
assure that we have taken the pure sound of the car
with the least possible noise. A microphone
connected to a laptop was used; it was at a high
place to assure to collect all the sound, since the
proposed hardware should be beside the light of the
highway, which is about 5 to 50 meters above the
cars. We have used a program called Sound Forge
for recording the sounds. Most of the sounds were
recorded at a sample frequency of 44Khz to make
sure that the sound has a high quality, and all the
component of the sound will be shown when
converting the sound to the frequency domain.
3.3 Feature Extraction
In order to recognize the sound of the car
among other sounds we need to extract the
parameters from the sound signal, these parameters
help us to distinguish the sounds domain from
others (car, plane, weather, and human sounds).
Feature extraction consists of choosing those
features which are most effective for preserving
class separately [Fukunaga, 1990]. The main
features that we have chosen which most
effectively describe the sounds are LPC analysis,
magnitude of the signal, and pitch of the signal.
3.4 Pitch Extraction
The harmonic-peak-based method has been
used to extract pitch from the wave sound. Since
harmonic peaks occur at integer multiples of the
pitch frequency, then we compared peak

Using the Sound Recognition Techniques Khalid T. Al-Sarayreh et al.
4
frequencies at each time (t) to locate the
fundamental frequency in order to find the highest
three magnitude peaks for each frame. Therefore,
the differences between them computed. Since the
peaks should be found at multiples of the
fundamental, we know that their differences should
represent multiples as well. Thus, the differences
should be integer multiples of one another. Using
the differences, we can derive our estimate for the
fundamental frequency.
The peak vector consists of the largest three
peaks in each frame. This forms a track of the pitch
for the signal [Ayuso-Rubio and Lopez-Soler,
1995]. First we have found the spectrogram of the
signal; spectrogram computes the windowed
discrete-time Fourier transform of a signal using a
sliding window.
The spectrogram is the magnitude of this
function which shows the areas where the energy is
mostly appear, after that we have take the largest
three peaks in each frame. A major advantage to
this method is its very noise-resistive. Even as
noise increases, the peak frequencies should still be
detectable above the noise.
3.5 Feature Comparison
After the feature extraction, the similarity
between the parameters derived from the collected
sound and the reference parameters need to be
computed. The three most commonly encountered
algorithms in the literature are Dynamic Time
Warping (DTW), Hidden Markov Modelling
(HMM) and Vector Quantization (VQ). In this
paper, we use the VQ to compare the parameter
matrices.
3.6 Decision Function
There are usually three approaches to
construct the decision rules [Gonzales and Woods,
2002], that is; Geometric, Topological, or
Probabilistic rules.
If the probabilities are perfectly estimated,
then the Bayes Decision theory 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 paper, we will
discuss two types of decision rules, which are
based either on linear functions or on more
complex functions such as Support Vector
Machines (SVM).
4. Theoretical Implementation
4.1 LPC Analysis
LPC based feature extraction is the most
widely used method by developers of speech
recognition. The main reason is that speech
production can be modelled completely by using
linear predictive analysis, beside, LPC based
feature extraction can also be used in speaker
recognition system where the main purpose is to
extract the vocal tract parameters from a given
sound, in speech synthesis, linear prediction
coefficient are the coefficient of the FIR filter
representing a vocal tract transfer function,
therefore linear prediction coefficient are suitable
to use as a feature set in speaker verification system.
The general idea of LPC is to determine the current
sample by a linear combination of P previous
samples where the linear combination weights are
the linear prediction coefficient. Since LPC is one
of the most powerful speech analysis techniques for
extracting good quality features and hence
encoding the speech. The LPC coefficients (ai) is
the coefficients of the all pass transfer function H(z)
modelling the vocal tract, and the order of the LPC
(P) is also the order of H(z), which has been
defined to be 10 in this paper.

Marsland Press
Journal of American Science 2009:5(2) 1-12
5
Linear predictive coding (LPC) offers a
powerful and simple method to exactly provide this
type of information. Basically, the LPC algorithm
produces a vector of coefficients that represent a
smooth spectral envelope of the DFT magnitude of
a temporal input signal. These coefficients are
found by modelling each temporal sample as a
linear combination of the previous P samples.
To be noted that the order of the LPC which
used in this paper is 10. The LPC filter is given by:
10
10...
2
2
1
11
1
)(


ZaaZaZa
ZH
This is equivalent to saying that the
input-output relationship of the filter is given by the
linear difference equation:
¦


10
1
)()()(
i
insiansnu
Where u(n) is the innovation of the signal, s(n) is
the original signal, H(Z) is LPC filter, and ai are
the coefficient of the filter.
Another important equation that is used to
predicate the next output from previous samples is:
¦¦

#

˜
10
1
][
10
1
][][][ˆ
k
knska
k
knskanuGns
Where dž[n] (the prediction for the next output value)
is a function of the current input and previous
outputs, G is the gain.
The optimal values of the filter coefficients
are gotten by minimizing the Mean Square Error
(MSE) of the estimate, that is:
¸
¹
·¨
©
§
¹¸
·
©¨
§o ][2min][ˆ][][ neEnsnsne
Where E[n] is the mean square error.
A popular method to get a Minimum Mean
Square Error (MMSE) is called the autocorrelation
method, where the minimum is found by applying
the principle of orthogonality. To find the LPC
parameters, the Toeplitz autocorrelation matrix is
used:
»
»
»
»
»
»
»
»
»
»
»
»
»
»
¼
º
«
«
«
«
«
«
«
«
«
«
«
«
«
«
¬
ª











»
»
»
»
»
»
»
»
»
»
»
»
»
»
¼
º
«
«
«
«
«
«
«
«
«
«
«
«
«
«
¬
ª
»
»
»
»
»
»
»
»
»
»
»
»
»
»
¼
º
«
«
«
«
«
«
«
«
«
«
«
«
«
«
¬
ª
)10(
)9(
)8(
)7(
)6(
)5(
)4(
)3(
)2(
)1(
)0()1()2()3()4()5()6()7()8()9(
)1()0()1()2()3()4()5()6()7()8(
)2()1()0()1()2()3()4()5()6()7(
)3()2()1()0()1()2()3()4()5()6(
)4()3()2()1()0()1()2()3()4()5(
)5()4()3()2()1()0()1()2()3()4(
)6()5()4()3()2()1()0()1()2()3(
)7()6()5()4()3()2()1()0()1()2(
)8()7()6()5()4()3()2()1()0()1(
)9()8()7()6()5()4()3()2()1()0(
10
9
8
7
6
5
4
3
2
1
R
R
R
R
R
R
R
R
R
R
a
a
a
a
a
a
a
a
a
a
RRRRRRRRRR
RRRRRRRRRR
RRRRRRRRRR
RRRRRRRRRR
RRRRRRRRRR
RRRRRRRRRR
RRRRRRRRRR
RRRRRRRRRR
RRRRRRRRRR
RRRRRRRRRR
Where:
¦



k
n
knsnskR
159
0
)()()(
and R(k) is the autocorrelation of the signal.
The above matrix equation could be solved
using the Gaussian elimination method. Any matrix
inversion method or The Levinson-Durbin
recursion (described below). To compute this
vector, the recursive Levinson-Durbin Algorithm
(LDR) was used.
4.2 Pre-emphasis
In general, the digitized speech waveform has
a high dynamic range and suffers from additive
noise. In order to reduce this range pre-emphasis is
applied. By pre-emphasis [Robiner and Juang,
1993], we imply the application of a high pass filter,
which is usually a first-order FIR of the form:
H (z) =1í az-1, 9 ” a ”1.0
The pre-emphasis is implemented as a
fixed-coefficient filter or as an adaptive one, where
the coefficient ai is adjusted with time according to
the autocorrelation values of the speech. The
pre-emphasize has the effect of spectral flattening
which renders the signal less susceptible to finite
precision effects (such as overflow and underflow)
in any subsequent processing of the signal. The
selected value for a in our work was 0.9375. Fig.1
and Fig.2 below represent the process of LPC
analysis.

Marsland Press
Journal of American Science 2009:5(2) 1-12
7
4.4 The k-means algorithm
The K-means is one of the simplest
unsupervised learning algorithms that solve the
well known clustering problem. The procedure
follows a simple and easy way to classify a given
data set through a certain number of clusters
(assume k clusters) fixed a priori. The main idea is
to define k centroids, one for each cluster. These
centroids should be placed in a cunning way
because of different location causes different result.
So, the better choice is to place them as much as
possible far away from each other.
The next step is to take each point belonging
to a given data set and associate it to the nearest
centroid. When no point is pending, the first step is
completed and an early group is done. At this point
we need to re-calculate k new centroids as bar
centers of the clusters resulting from the previous
step. After we have these k new centroids, a new
binding has to be done between the same data set
points and the nearest new centroid. A loop has
been generated. As a result of this loop we may
notice that the k centroids change their location
step by step until no more changes are done. In
other words centroids do not move any more
[MacQueen, 1997].
Finally, this algorithm aims at minimizing an
objective function, in this case a squared error
function. The objective function:
¦¦


k
j
n
i
jC
j
ixj
1 1
)2()(
Where
)2()(
jC
j
ix  is a chosen distance
measure between a data point
( )j
ix and the
cluster centre is an indicator of the distance of the n
data points from their respective cluster canters
[Zha et al, 2001]. The algorithm is composed of the
following steps:
1. Place K points into the space represented by
the objects that are being clustered. These
points represent initial group centroids.
2. Assign each object to the group that has the
closest centroid.
3. When all objects have been assigned,
recalculate the positions of the K centroids.
Repeat Steps 2 and 3 until the centroids no
longer move. This produces a separation of the
objects into groups from which the metric to be
minimized can be calculated [Moore, 2007].
4.5 Distance Measure
After quantizing a sound into its codebook, we
need a way to measure the similarity/dissimilarity
between and two sound domains. It is common in
the field to use a simple Euclidean distance
measure, and so this is what we have used.
4.5.1 Euclidian distance
Euclidean metric is the distance between two
points that one would measure with a ruler, The
Euclidean distance between two points P=[p1 p2 p3
pn]
T and Q=[q1 q2 q3 qn]
T and, in Euclidean
n-space, is defined as:
¦


n
i
iqip
1
2)(
4.5.2 Distance Weighting Coefficients
We follow the algorithm proposed above to
weight distances so as to increase the likelihood of
choosing the true sound domain. This algorithm,
basically, reflects the greater importance of unique
codewords as opposed to similar ones. This is a
very important portion of our system. The training
system architecture we created consists of two
main parts. The first part consists of processing

Using the Sound Recognition Techniques Khalid T. Al-Sarayreh et al.
8
each sound input voice sample to condense and
summaries the characteristics of the sound features.
The second part involves pulling each sound’s data
together into a single, easily manipulated, three
dimensional matrix.
5. Practical Implementation
First we have high-pass filtered the signal
because the more important information for speech
processing lies in the higher frequencies according
to Kinnunen and Franti [2004]. Then we split the
signal into frames, each about 30ms long. By
breaking the signal into frames, we approximate
these distinct sounds in our analysis. For each
frame we calculate the LPC coefficient. We also
calculated the Magnitude and the pitch of the sound.
These coefficients characterize each sound domain.
The next step is to map these data. This is called
vector quantization (VQ) and is accomplished by a
clustering algorithm.
However, the clustering algorithm takes a
number of random vectors and condenses the
vectors that are nearest to it, iterating until the least
mean error between the vectors is reached. We
clustered the data into vectors, and each of these
vectors is called a codeword. This set of vectors, or
codewords is created for each sound. The
codewords for a given sound are then stored
together in a codebook for that sound domain. Each
speaker’s codebook (sound group) is then stored
together in a master codebook which is compared
to the test sample during the testing phase to
determine the sound domain.
Suppose there is a region of space where
codeword vectors from several different sounds
were laid. If a test vector also falls in this region,
the codewords do not help determine the identity of
the sound domain because the errors between the
test vector and the various codewords will be
roughly equal.
Kinnunen and Franti [2004] present an
algorithm for discriminating between code vectors
during the testing phase to help solve this problem.
Their idea is to give a higher precedence to code
vectors which give a more definite idea of a
domain’s identity by weighting them. We used the
algorithm they presented and computed weighting
coefficients for my codebook data during the
training phase. In parts (a) and (b) of Fig.3, we
present the features of the sounds for the different
domains before and after clustering [MacQueen,
1997].

(a) (b)
Fig.3 (a) the pitch feature for all sounds in all domains before clustering, (b) the pitch feature for all sounds in all
domains after clustering

Marsland Press
Journal of American Science 2009:5(2) 1-12
9
The figures above depict code vectors before
and after clustering, and each shape represents a
different group (car, plane, weather). You can see
that the distribution of the vectors before and after
clustering is different. Weighting the vectors does
make a difference.
6. Testing, Results, and Discussion
6.1 Testing
For all of the following tables, denote
"recognizing the sound as a car" to be T, "not
recognizing the sound as a car" to be F, and we
assume that the acceptable interval is between
0.607 and 1.0622. However, we have conducted
two types of testing as the following:
1. Using the first method in testing (Feature
Extraction based on Statistical analysis), the
percentage of recognizing the cars sounds as a
car is 92.5% and the percentage of
recognizing the plane and rain sounds as a car
is 6%; see Tables 1 and 2 below for more
details.
Table 1: The performance of the first method in recognition for cars samples
Number of Sample Cars Recognized Value of the mean calculation for the sound
1 T 0.70
5 F 0.45
9 F 0.59
10 T 0.77
13 T 0.70
20 T 0.69
35 T 0.80
49 T 0.88
57 T 0.90
62 T 0.96
77 T 0.92
99 T 0.85
Table 2: The performance of the first method in recognition for planes samples
Number of sample Planes and rains Recognized Value of the mean calculation for the sound
1 T 0.650
5 F 0.005
9 F 0.020
10 F 0.100
13 F 0.217
15 F 0.168
19 F 0.020
22 F 0.180
25 F 0.054
29 T 0.600
33 F 0.322

Using the Sound Recognition Techniques Khalid T. Al-Sarayreh et al.
10
2. Second Method in Testing (Vector
Quantization); see Tables 3 and 4.
Table 3: The performance of the second method in
recognition for cars samples
Number of Sample Cars Recognized
1 T
5 T
9 T
10 T
13 T
20 T
35 T
49 T
57 T
62 T
77 T
99 T
Table 4: The performance of the second method in
recognition for planes and rains samples
Number of sample
Planes and rains
Recognized
1 F
5 F
9 F
10 F
13 F
15 F
19 F
22 F
25 F
29 F
33 F
6.2 Results
The voice recognition system that we have
built identifies the three main sound domains which
are planes, cars, and weather. The recognition for
these domains was 100% according to the vector
quantization method but the vector quantization
method divided all the sound into three mainly
regions so any new sound will be approximated
into one of these region so to get more accuracy
and to avoid any similar sound to the car sound we
have developed a new code called feature
extraction based on statistical analysis to discard
any sound that is similar to the car sound but really
not car sound this method with the vector
quantization method get accuracy 100% and we
assure to make energy conserving with probability
reached 100% since the lights will turn only for
cars .the feature extraction based on statistical
analysis method gain accuracy of 92.5% when
work alone.
6.3 Discussion
The idea of the feature selection based on
statistical hypothesis testing is to set two classes
let's say x1 and x2 and each class has its own shape,
distribution, mean, and standard deviation then for
a specific sample we will try to investigate whether
the values which it takes differ significantly or
belongs to these sets. For the feature extraction, we
have take the power spectral density for each sound
and the mean of all the samples and their standard
deviation, then we have make dot multiplication
between the mean (power spectral density), of all
the samples and the new test sound and upon this
number(the result of multiplication) we have decide
if this test sound lie in the pre-determined classes
or not and the range of comparison was (ȝ - 2*ı ,
ȝ + 2*ı) so the probability that the data will lie
within this range is 95% according to statistical
analysis and rules. However, the results of this
method were as the following three choices:
1. When executing the code as it is (with

Marsland Press
Journal of American Science 2009:5(2) 1-12
11
acceptable interval [av-2*std, av+2*std]), the
results were as the following:
- Average = 0.8346,
- STD = 0.1138
- The probability of data that will lie here =
95%.
- The correct recognition of car sound =
92.5%.
- Recognizing sound of planes as car= 4%
- Recognizing sound of rain as car= 2%
- Recognizing sound of animals as car= 0%
2. When executing the code and taking the
acceptable interval to be [av-sd , av+sd], the
results were as the following:
- Average = 0.8346
- Std = 0.1138
- The probability of data that will lie here =
67%.
- The correct recognition of car sound= 87%
- Recognizing sound of planes as car= 0%
- Recognizing sound of rain as car= 0%
- Recognizing sound of animals as car =0%
3. When executing the code and taking the
FFT(sound, 11000) and summing the area
under the curve from 0-5500 and interval to
be [av-2*sd, av+2*sd] the results were as the
following:
- Average = 0.8324
- Std = 0.1155
- Acceptable interval [av-2*std, av+2*std].
- The correct recognition of car sound =91%
- Recognizing sound of planes as car= 0%
- Recognizing sound of rain as car= 8%
- Recognizing sound of animals as car= 0%.
Based on the results of this paper, we can note
that the choice number one is the best
7. Conclusion
The lighting is available for the highways to
avoid accidents and to make the driving more
safety and more easy, but turning the lights on all
the nights will consume a lot of energy which it
might be used in another important issues. This
paper presented a methodology of using the sound
recognition techniques in order to turn the lights on
only when there are cars on the highway and only
for some period of time. In more details, Linear
Predictive Coding (LPC) method and feature
extraction have been used to apply the sound
recognition. Furthermore, the Vector Quantization
(VQ) has been used to map the sounds into groups
in order to compare the tested sounds.
However, this paper made the following
contributions:
1. Designing a new system for conserving
energy based on the voice recognition of the
car sound.
2. This system is the first application of this type
that concern the street lights.
3. This paper also demonstrates that the
weighted Euclidian distance with the LBG
algorithm was very helpful and achieved high
accuracy.
This paper shows that the feature of the sound
that we have extracted is very valuable and really
can distinguish between different sounds, so it is
differentiate sounds from other or which we can
call that it makes speaker identification with high
accuracy.
References
Ayuso-Rubio, A. J., Lopez-Soler, J. M., 1995,
Speech Recognition and Coding: New
Advances and Trends, Springer, Berlin,
Germany.
Fukunaga, K., 1990, Introduction to Statistical
Pattern Recognition, 2nd edition, Academic
Press, London, UK.
Goldberg, R., Riek, L., 2000, A Practical Handbook

Using the Sound Recognition Techniques Khalid T. Al-Sarayreh et al.
12
of Speech Coders, CRC Press, Boca Raton, FL,
USA.
Gonzales, R., Woods, R., 2002, Digital image
processing, 2nd edition, Prentice-Hall.
Garg, H. K., 1998, Digital Signal Processing
Algorithms: Number Theory, Convolution, Fast
Fourier Transforms, and Applications, CRC
Press, Boca Raton, FL, USA.
Kinnunen, T., Frunti, P., 2001, Speaker
Discriminative Weighting Method for
VQ-Based Speaker Identification, In
proceedings of the. 3rd International
Conference on Audio-and Video-Based
Biometric Person Authentication (AVBPA’01),
Halmstad, Sweden, pp. 150-156.
Klevans, R., Rodman, R., 1997, Voice Recognition,
Artech House Inc., Northfield, MN, USA.
Kuo, S. M., Gan, W.-S., 2004, Digital Signal
Processors: Architectures, Implementations, and
Applications, Prentice Hall, Englewood Cliffs,
New Jersey, USA.
Lyons, R. G., 1996, Understanding Digital Signal
Processing, Pearson Education, London, UK.
MacQueen, J., 1967, Some Methods for
Classification and Analysis of Multivariate
Observations, In proceedings of the 5th
Berkeley Symposium on Mathematical
statistics and probability, Berkeley, California,
University of California Press, pp. 281-297.
Moore, A., 2007, K-means and Hierarchical
Clustering, Tutorial Slides, School of Computer
Science, Carnegie Mellon University, Online:
http://web.cecs.pdx.edu/~york/cs510w04/kmea
ns09.pdf, Accessed on Dec. 3, 2007.
Nelson, M., Gailly, J.-L., 1995, The Data
Compression Book, M&T Books.
Rabiner, L. R. and Juang, H., 1993, Fundamentals
of Speech Recognition, Prentice Hall,
Englewood Cliffs, New Jersey, USA.
Sayood, K., 2005, Introduction to Data
Compression, Morgan Kaufmann, San
Francisco, CA, USA.
Smith, D., 2001, Digital Signal Processing
Technology: Essentials of the Communications
Revolution, American Radio Relay League,
Newington, CT, USA.
Zha, H., Ding, C., Gu, M., He, X., Simon, H. D.,
2001, Spectral Relaxation for K-means
Clustering, Advances in Neural Information
Processing Systems, Vol. 14, pp. 1057-1064.