Académique Documents
Professionnel Documents
Culture Documents
ENHANCEMENT
ABSTRACT
This paper proposes the application of the Wiener filter in an adaptive manner in
speech enhancement. The proposed adaptive Wiener filter depends on the adaptation of the
filter transfer function from sample to sample based on the speech signal statistics(mean
and variance). The adaptive Wiener filter is implemented in time domain rather than in
frequency domain to accommodate for the varying nature of the speech signal. The
proposed method is compared to the traditional Wiener filter and spectral subtraction
methods and the results reveal its superiority.
1 INTRODUCTION
Speech enhancement is one of the most is removed first. Decomposition of the vector space
important topics in speech signal processing. of the noisy signal is performed by applying an
Several techniques have been proposed for this eigenvalue or singular value decomposition or by
purpose like the spectral subtraction approach, the applying the Karhunen-Loeve transform (KLT)[8].
signal subspace approach, adaptive noise canceling Mi. et. al. have proposed the signal / noise KLT
and the iterative Wiener filter[1-5] . The based approach for colored noise removal[9]. The
performances of these techniques depend on idea of this approach is that noisy speech frames
quality and intelligibility of the processed speech are classified into speech-dominated frames and
signal. The improvement of the speech signal-to- noise-dominated frames. In the speech-dominated
noise ratio (SNR) is the target of most techniques. frames, the signal KLT matrix is used and in the
noise-dominated frames, the noise KLT matrix is
Spectral subtraction is the earliest method for used.
enhancing speech degraded by additive noise[1]. In this paper, we present a new technique to
This technique estimates the spectrum of the clean improve the signal-to-noise ratio in the enhanced
(noise-free) signal by the subtraction of the speech signal by using an adaptive implementation
estimated noise magnitude spectrum from the noisy of the Wiener filter. This implementation is
signal magnitude spectrum while keeping the phase performed in time domain to accommodate for the
spectrum of the noisy signal. The drawback of this varying nature of the signal.
technique is the residual noise.
The paper is organized as follows: in section
Another technique is a signal subspace II, a review of the spectral subtraction technique is
approach [3]. It is used for enhancing a speech presented. In section III, the traditional Wiener
signal degraded by uncorrelated additive noise or filter in frequency domain is revisited. Section IV,
colored noise [6,7]. The idea of this algorithm is proposes the adaptive Wiener filtering approach for
based on the fact that the vector space of the noisy speech enhancement. In section V, a comparative
signal can be decomposed into a signal plus noise study between the proposed adaptive Wiener filter,
subspace and an orthogonal noise subspace. the Wiener filter in frequency domain and the
Processing is performed on the vectors in the signal spectral subtraction approach is presented.
plus noise subspace only, while the noise subspace
σs
2
A priori knowledge sˆ(n) = mx + ( x(n) - mx ) ∗ δ ( n)
σs + σv
2 2
Space-
σs
2
Degraded speech variant Enhanced
x(n) speech
= mx + ( x(n) − mx )
σs + σv
2 2
h(n)
signal sˆ( n) (14)
If it is assumed that mx and σs are updated at
each sample, we can say:
Measure of
σs (n) ( x(n) − mx(n))
2
approximated by:
performed.
σˆs (n) = ⎨
2 60
O u tp u t P S N R (d B )
⎩0, otherwise
50
(17.a)
Where 40
30
n+ M
1
σˆx (n) =
2
(2 M + 1)
∑ ( x(k ) − mˆ (n))
k =n−M
x
2
20
Spectral Subtraction
(17.b)
10 Wiener Filter
By this proposed method, we guarantee that Adaptive Wiener Filter
the filter transfer function is adapted from sample
to sample based on the speech signal statistics. 0
-10 -5 0 5 10 15 20 25 30 35
Input SNR (dB)
5 EXPERIMENTAL RESULTS
For evaluation purposes, we use different Figure 2: PSNR results for white noise
speech signals like the handel, laughter and gong case at-10 dB to +35 dB SNR levels for Handle signal
signals. White Gaussian noise is added to each
speech signal with different SNRs. The different
speech enhancement algorithms such as the
spectral subtraction method, the Weiner filter in
frequency domain and the proposed adaptive
Wiener filter are carried out on the noisy speech
signals. The peak signal to noise ratio (PSNR)
results for each enhancement algorithm are
compared.
1 1
50
A m p lit u d e
A m p lit u d e
0 0
40 -1 -1
O u tp u t P S N R (d B )
30 1 1
A m p lit u d e
A m p lit u d e
0 0
20
-1 -1
0 2000 4000 6000 8000 0 2000 4000 6000 8000
Spectral Subtraction (c) (d)
10 Wiener Filter 1
A m p lit u d e
Adaptive Wiener Filter 0
0
-10 -5 0 5 10 15 20 25 30 35 -1
0 2000 4000 6000 8000
Input SNR (dB) (e) Time(msec)
70
Amplitude (dB)
Amplitude (dB)
0 0
60 -20 -20
-40
O u tp u t P S N R (d B )
50 -40
0 1000 2000 3000 4000 0 1000 2000 3000 4000
(a) (b)
40
Amplitude (dB)
Amplitude (dB)
0 0
30 -20
-20
20 Spectral Subtraction -40 -40
0 1000 2000 3000 4000 0 1000 2000 3000 4000
Wiener Filter (c) (d)
10
Adaptive Wiener Filter
Amplitude (dB)
0
0
-10 -5 0 5 10 15 20 25 30 35 -20
Input SNR (dB) -40
0 1000 2000 3000 4000
Figure 4: PSNR results for white noise case at -10 dB (e) Freq.(Hz)
to +35 dB SNR levels for Gong signal
The results of the different enhancement Figure 6:The spectrum of the Handel sig. in Fig.(5) (a)
original sig. (b) noisy sig. (c) spectral subtraction. (d)
algorithms for the handle signal with SNRs of 5,
Wiener filtering. (e) adaptive Wiener filtering.
10,15 and 20 dB in the both time and frequency
domain are given in Figs. (5) to (12). These results
A m p lit u d e
1 1
A m p lit u d e
A m p lit u d e
1 1
0 0
0 0
-1 -1
0 2000 4000 6000 8000 0 2000 4000 6000 8000
-1 -1
(a) (b) 0 2000 4000 6000 8000 0 2000 4000 6000 8000
A m p lit u d e
A m p lit u d e
A m p lit u d e
1 1
0 0
0 0
-1 -1
0 2000 4000 6000 8000 0 2000 4000 6000 8000 -1 -1
(c) (d) 0 2000 4000 6000 8000 0 2000 4000 6000 8000
A m p lit u d e
(c) (d)
A m p lit u d e
1
1
0 0
-1 -1
0 2000 4000 6000 8000 0 2000 4000 6000 8000
(e) Time (msec) (e) Time(msec)
0 0 0 0
-20 -20 -20
-20
-40 -40
-40 -40
0 1000 2000 3000 4000 0 1000 2000 3000 4000 0 1000 2000 3000 4000 0 1000 2000 3000 4000
(a) (b) (a) (b)
0 0 0 0
-20 -20 -20 -20
-40 -40
0 1000 2000 3000 4000 -40 -40
0 1000 2000 3000 4000 0 1000 2000 3000 4000 0 1000 2000 3000 4000
(c)
Amplitude (dB)
(d) (c)
Amplitude (dB)
(d)
0
0
-20
-20
-40
-40
0 1000 2000 3000 4000 0 1000 2000 3000 4000
(e)Freq. (Hz) (e)Freq. (Hz)
Figure 8: The spectrum of the Handel sig. in Fig.(7) Figure 10: The spectrum of the Handel sig. in Fig.(9)
(a) original sig. (b) noisy sig. (c) spectral subtraction. (d) (a) original sig. (b) noisy sig. (c) spectral subtraction. (d)
Wiener filtering. (e) adaptive Wiener filtering. Wiener filtering. (e) adaptive Wiener filtering.
A m p lit u d e
A m p lit u d e
An adaptive Wiener filter approach for
0 0 speech enhancement is proposed in this papaper.
This approach depends on the adaptation of the
-1 -1
0 2000 4000 6000 8000 0 2000 4000 6000 8000 filter transfer function from sample to sample
(a) (b) based on the speech signal statistics(mean and
variance). This results indicates that the proposed
1 1 approach provides the best SNR improvement
A m p lit u d e
A m p lit u d e
0 0 among the spectral subtraction approach and the
traditional Wiener filter approach in frequency
-1 -1 domain. The results also indicate that the proposed
0 2000 4000 6000 8000 0 2000 4000 6000 8000 approach can treat musical noise better than the
(c) (d) spectral subtraction approach and it can avoid the
1 drawbacks of Wiener filter in frequency domain .
A m p lit u d e
0 REFERENCES
0
[3] Y. Ephriam and H. L. Van Trees: A signal
0
subspace approach for speech enhancement, in
-20 -20 Proc. International Conference on Acoustic,
-40 Speech and Signal Processing, vol. II, Detroit,
-40
0 1000 2000 3000 4000 0 1000 2000 3000 4000 MI, U.S.A., pp. 355-358, May (1993).
(a) (b) [4] Simon Haykin: Adaptive Filter Theory,
Prentice-Hall, ISBN 0-13-322760-X, (1996).
Amplitude (dB)
Amplitude (dB)
ABSTRACT.
This work is a study of the effects of frequency selectivity on multi-carrier wideband signals in three different
environments; indoors, outdoor to indoor and outdoors. The investigation was made using measurements carried
out with a sounder with a 300 MHz bandwidth. The main part of this work is related to evaluate the contribution
of several parameters; frequency selective fading, coherence bandwidth and delay spread on the frequency
selectivity of the channel. A description of the sounder parameters and the sounded environments are given. The
300 MHz bandwidth is divided in segments of 60 kHz to perform the evaluation of frequency selective fading.
Sub channels of 20 MHz for OFDM systems and 5 MHz for WCDMA were evaluated. Figures are provided for
a number of bands, parameters and locations in the three environments. It is also shown the variation of the
signal level due to frequency selective fading. The practical assumptions about the coherence bandwidth and
delay spread are reviewed and a comparison is made with actual measurements. Statistical analysis was
performed over some of the results.
Keywords. Coherence bandwidth, frequency correlation, frequency selective fading and multi-carrier
modulation
.
characteristics of the wireless communications
I. INTRODUCTION. channel. The dispersive channel characteristics
arise from the different propagation paths, i.e.
To simulate and evaluate the performance of a multipath, between the receiver and the transmitter.
wireless mobile system a good channel model is This dispersion could be measured, if we could
needed. Mobile communication systems are using measure the channel impulse response (CIR). As a
larger bandwidths and higher frequencies and these general rule the effects of ISI on the transmission
characteristics impose new challenges on channel errors is negligible if the delay spread is
estimation. The channel models that have been significantly shorter than the duration of the
developed for the mobile systems in use may not be transmitted symbol. Due to the expected increase in
applicable anymore. To validate that the old demand of higher data rates, wideband multi-
models can be used for future systems or to design carrier systems such as; OFDM and WCDMA are
new models, it is necessary to answer the question expected to be technologies of choice [1], [12] and
about how the same parameters performs at higher [14]. This is because these two technologies can
bandwidths? Also, we have to be able to measure provide both; high data rates and an acceptable
and validate some parameters and compare them to level of quality of service. However, these systems
well known practical assumptions. Measurements need first to address better the problem regarding
for analysis of the fading statistics at common channel prediction or estimation, because this
frequencies have been performed before, but they condition is the main boundary for higher data
have been performed at small bandwidths, it is rates. The study of correlation of the mobile radio
necessary to update the models with higher channel in frequency and time domains has helped
bandwidths. to understand the problem of channel estimation.
One of this work objectives is to evaluate
As the data rate (the bandwidth) increases the frequency selective fading (FSF) in several
communication limitations come from the Inter environments. This work begins with the results of
Symbol Interference (ISI) due to the dispersive
∫ E {h(τ ) }e
∞
actual results and a comparison is provided. 2 − j 2π∆fτ
R H ( ∆f ) = dτ
The rest of this document is structured as follow; in −∞
part II the theoretical foundations of the channel (5)
impulse response frequency selective fading and
coherence bandwidth are reviewed. Also in this
part, the characteristics of the three environments where
{
E h (τ )
2
}
is the average Power Delay
sounded are described. In part III, the frequency Profile PDP of the channel. So, under the above
selective fading evaluation and analysis are conditions, RH is the Fourier transform of the
presented. Plots of the dependency of fading deep average PDP.
and frequency separation of two specific points in
the response are studied. At part IV, data about the 2.2 Coherence bandwidth.
relationship between delay spread and coherence The multipath effect of the channel, the arrival of
bandwidth are provided. At the end in part V, different signals in different time delays causes the
conclusions and future work are mentioned. statistical properties of two signals of different
frequencies to become independent if the frequency
II. MATHEMATICAL BACKGROUND separation is large enough. The maximum
frequency separation for which the signals are still
2.1 The wideband channel model. strongly correlated is called coherence bandwidth
The radio propagation channel is normally (Bc). Besides to contribute to the understanding of
represented in terms of a time-varying linear filter, the channel, the coherence bandwidth is useful in
with complex low-pass impulse response, h(t, τ). Its evaluating the performance and limitations of
time-varying low-pass transfer function is [4] [6] different modulations and diversity models.
[8] [10]:
∞ The coherence bandwidth of a fading channel is
∫ h(t;τ )e τ dτ
− j 2πf
H (t , f ) = probed by sending two sinusoids, separated in
−∞ frequency by ∆f = f1- f2 Hz, through the channel.
(1) The coherence bandwidth is defined as ∆f, over
Where τ represents delay, using (1) the frequency which the cross correlation coefficient between r1
correlation function for the channel can be written and r2 is greater than a preset threshold, say, η0=
as: 0.9. Namely:
(8) 1 1
BC =0.9 = (13) BC =0.5 = (14)
50σ rms 5σ rms
r1r 2 ⎡ r + r ⎤ ⎛ r1r 2 λ ⎞ 2 2
= exp⎢− I ⎜
2 ⎥ 0⎜
1
⎟
2 ⎟
2
1 1
µ (1 − λ )
2 2
⎣ 2 µ (1 − λ ) ⎦ ⎝ µ 1 − λ ⎠ BC =0.9 = (15) BC = (16)
8σ mean 2πσ rms
Figure 7. RMS delay spread for indoors Table 1, shows the comparisons of Bc for the
three environments with the different versions of
Figure 11 shows the frequency correlation for the expressions 13 -16 and measured results. This
outdoors environment. Figure 12, shows the Bc at table shows that the values of the expressions are
frequency correlation of 0.9. In this case the Bc can always lower than the measured results, which
not be compared to the Bc for the other two induce to conclude that the expressions were
environments, since in this environment a lower underestimated, at least in these environments.
bandwidth is evaluated, 120 MHz instead of 300 Moreover, it is possible to conclude that these
MHz. Despite this difference and observing expressions were deduced with not enough
figures 11 and 12, Bc is not significantly lower measured results. Also, table 1 show that the
even when we have higher distances and higher relationship between delay spread and coherence
delay spread. In outdoors the Bc is not bigger than 2 bandwidth, not necessarily is a single constant.
MHz in average. In the other hand, the RMS delay
spread is 1.5 µS in average.
References.
lienfa@cc.kyu.edu.tw,chencc@mail.stut.edu.tw
ABSTRACT
The query service for the location of an object is called Location Based Services
(LBSs), and Reverse Nearest Neighbor (RNN) queries are one of them. RNN queries
have diversified applications, such as decision support system, market decision,
query of database document, and biological information. Studies of RNN in the past,
however, focused on inquirers in immobile status without consideration of
continuous demand for RNN queries in moving conditions. In the environment of
wireless network, users often remain in moving conditions, and sending a query
command while moving is a natural behavior. Availability of such service therefore
becomes very important; we refer to this type of issue as Continuous Reverse
Nearest Neighbor (CRNN) queries. Because an inquirer’s location changes
according to time, RNN queries will return different results according to different
locations. For a CRNN query, executing RNN search for every point of time during a
continuous query period will require a tremendously large price to pay. In this work,
an efficient algorithm is designed to provide precise results of a CRNN query in just
one execution. In addition, a large amount of experiments were conducted to verify
the above-mentioned method, of which results of the experiments showed significant
enhancement in efficiency.
2.1.1R-Tree
What CRNN query pursues is a period of Figure 6. Finding segment point of CRNN search
continuous time; the moving distance of query
objects is very short among some adjacent points of Figure 7 illustrates the time segmentation
time for the query, thus possibly resulting in the same process described above. For object a, b, and c, their
RNN result. That is, the entire period of continuous respective NNs are identified first: NN(a)=b,
query is divided into several segments, and the RNN NN(b)=a, and NN(c)=b. Next, use each object as the
results in each segment are the same. If these points center of a circle, and the distance to its respective
of time share the same RNN result, then it is not NN as the radius to make circles of a, b, and c. Then,
necessary to execute RNN search for each of the intersections of the circles and qlines, Ps, P1, P2, P3,
points of time; one-time calculation is enough. P4, and Pe , are sorted according to time, and every
Therefore, CRNN query does not require executing two intersection points define a time segment. The
RNN search for all points of time. Instead, points of entire CRNN query is cut into five time segments, [Ps,
time that share the same RNN result are grouped into P1] , [ P1, P2] , [P2,P3] , [P3,P4] , and [P4,Pe].
time segments, and one-time RNN search is executed Every segment has a unique RNN query result.
for each of the segments. RNN of query q is a
collection of the objects of which the NN is query q.
If the distance, or N, is realized in advance, then
these objects are the RNN for query q when the
distances from query q to the objects are shorter than
the distances from the objects to their respective NN.
Y axis
Y axis
Y axis
0 0
0 0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
X axis X axis
X axis
In addition to the object distribution described CRNN query is executing RNN algorithm for every
above, the influences that the amount of query time point of time which is continuous, and it is
(qline) and the number of objects may impose on impossible to calculate the required count of
efficiency are also considered. Three data sets of execution. Therefore, the CRNN query time must be
Uniform, Gaussian, and Zipf are considered in object segmented before the total execution time required
distribution. The amount of query time (qline) for CRNN query may be calculated. The more the
changes from query length 1 to query length 10. The time is segmented, the more executions of RNN are
number of objects changes from 1K to 10K. required. If a period of time is segmented into m
Parameters and settings used in the experiment are segments, then time complexity will be O(m×n3), and
listed in Table 1. if time is not adequately segmented, then the RNN
Table 1: Parameter settings of experiment result may be erroneous. These make it an inefficient
CRNN search algorithm, and it will not be compared
Parameter Description Settings in this experiment. Efficiency of two methods is
distribution Data distribution Uniform, compared in this experiment: one uses Rdnn-tree as
Gaussian, Zipf the index, and the other uses no index. To evaluate
interval Time interval of 1, 2, 5, 8, 10 these two methods, comparison of the time required
Query for one CRNN search execution can be used, and this
object-no Number of Data 1, 10. 30, 50 , comparison is referred to as total cost in this Study.
Objects 100(k)
5.4 Performance Results and Discussion
5.2 Compared Algorithms and Performance
Metrics Based on the changes of metrics (distribution,
interval, and object-no), different types of
The most intuitive method for finding RNN is experiments have been conducted. Results are
looking for the NN of every object. If the number of summarized by object-no and query interval in the
query objects is N, then time Complexity is O(n2). next section.
Next, determine which objects’ NNs are query points.
If the NNs are the query points, then the objects will 5.4.1 The effect of object-no parameter
be the RNNs for the query points. The required time
complexity for the RNN algorithm is O(n3). First, the fixed query interval is set at 5. The
However, the most intuitive method for finding influence imposed on efficiency by object-no
)
)
103
103 103
2 2 2
10 10 10
10 10 10
1 1 1
1K 10K 30K 50K 100K 1K 10K 30K 50K 100K 1K 10K 30K 50K 100K
object-no object-no object-no
Figure 11: Influences on different types of data distribution from changing object-no
5.4.2 The effect of query interval parameter distance of MinD(qline , Rect) decreases, causing
pruning efficiency to reduce. On the contrary, when
This section focuses on the influence from the the interval increases, the number of time
length of query interval on each method under segmentation by CRNN query increases.
different types of object distribution. Results of the Consequently, the number of RNN searches for every
experiment are shown in Figure 12. Generally segment increases, and total cost of CRNN query
speaking, when the query interval is lengthened, the increases as well.
105 105 10
5
104 104 10
4
)
)
3
10 103 103 CRNN without index
CRNN with index
102 10
2
10
2
10 10 10
1 1 1
1 2 5 8 10 1 2 5 8 10 1 2 5 8 10
Query Interval Query Interval Query Interval
6. Conclusions and Future Works Study also prove the efficiency of the proposed
method. As wireless communication and mobile
An efficient CRNN search algorithm is proposed device technology become mature, more and more
in this Study. Such algorithm requires only one users access information from wireless information
execution to find out RNN results from all continuous systems through mobile devices. To process requests
RNN searches. The diversified experiments in this from more and more mobile users, data dissemination
vulnerability to frequency offset errors between the where the Xm’s are the base band symbols on each
transmitted and received signals, which may be caused by sub-carrier. The digital-to-analog (D/A) converter then creates
Doppler shift in the channel or by the difference between the an analog time-domain signal which is transmitted through the
transmitter and receiver local oscillator frequencies [4]. In channel.
such situations, the orthogonality of the carriers is no longer At the receiver, the signal is converted back to a
maintained, which results in Intercarrier Interference (ICI). ICI discrete N point sequence y(n), corresponding to each sub-
results from the other sub-channels in the same data block of carrier. This discrete signal is demodulated using an N-point
the same user. ICI problem would become more complicated Fast Fourier Transform (FFT) operation at the receiver.
when the multipath fading is present [5]. If ICI is not properly
compensated it results in power leakage among the S/P IFFT P/S D/A
subcarriers, thus degrading the system performance.
Channel
given by
(7)
Sin(π (l + ε − k )) (3) Subsequently, the ICI coefficients for this received signal
Sin(l − k ) = exp( jπ (1 − 1 / N )(l + ε − k ))
NSin(π (l + ε − k ) / N ) becomes
S '' (l-k) = – S (l-k-1) + 2S (l-k) – S (l-k+1) (8)
3.1.1 ICI Canceling Modulation When compared to the two previous ICI coefficients
The ICI self-cancellation scheme requires that the |S (l-k)| for the standard OFDM system and |S'(l-k)| for the ICI
transmitted signals be constrained such that X (1) = - X (0), X canceling modulation, |S''(l-k)| has the smallest ICI
(3) = - X (2) …X (N-1) = - X (N-2).The received signal on coefficients, for the majority of l-k values, followed by
subcarriers k and k + 1 to be written as |S' (l-k)| and |S (l-k)|. This is shown in Figure 2 for N = 64 and
N −2 ε = 0.5. The combined modulation and demodulation method
Y ' (k ) = ∑ X (l )[S (l − k ) − S (l + 1 − k )] + n
l = 0 , 2 , 4 , 6 ,..
k (4) is called the ICI self-cancellation scheme. The reduction of the
ICI signal levels in the ICI self-cancellation scheme leads to a
N −2 higher CIR. The theoretical CIR is given by
Y ' (k + 1) = ∑ X (l )[S (l − k − 1) − S (l − k )] + n k +1
− S (−1) + 2 S (0) − S (1)
2
l = 0 , 2 , 4 , 6 ,..
CIR = 2
(9)
(5) N −1
where nk and nk+1 is the noise added to it.
And the ICI coefficient S ' (l-k) is denoted as
∑ − S (l − 1) + 2S (l ) − S (l + 1)
l = 2 , 4 , 6 ,..
S '(l-k) = S (l-k) – S (l+1-k) (6)
As mentioned previously, the redundancy in this
scheme reduces the bandwidth efficiency by half. There is a
tradeoff between bandwidth and power tradeoff in the ICI self-
cancellation scheme.
3.2 Maximum Likelihood Estimation
The second method for frequency offset correction
in OFDM systems was suggested by Moose in [12]. In this
approach, the frequency offset is first statistically estimated
using a maximum likelihood algorithm and then cancelled at
the receiver. This technique involves the replication of an
OFDM symbol before transmission and comparison of the
∑ Im Y (k )Y * (k )
⎪ ⎪
⎪ ⎪
⎪
2 1
⎪ where y(n) denotes the received preamble symbols
∧ 1
tan − 1
⎪ ⎪
distorted in the channel, w(n) the AWGN, and x(n) the IFFT
ε=
⎪ ⎪
⎪ k =− K ⎪
⎨ ⎬ (13) of the preambles X(k) that are transmitted, which are known at
2π ⎪ K ⎪
∑ Re Y (k )Y * (k )
⎪ ⎪
⎪
⎪
⎪
⎪
the receiver. Assume there are Np preambles preceding the
⎪ 2 1 ⎪
⎪⎩ k =− K ⎪⎭ data symbols in each frame are used as a training sequence
This maximum likelihood estimate is a conditionally and the variance σ2 of the AWGN w(n) is stationary.
unbiased estimate of the frequency offset and was computed
using the received data. Once the frequency offset is known, 4.2.2 Offset Correction Scheme
the ICI distortion in the data symbols is reduced by The ICI distortion in the data symbols x(n) that
multiplying the received symbols with a complex conjugate of follow the training sequence can then be mitigated by
the frequency shift and applying the FFT, multiplying the received data symbols y(n) with a complex
X (n) = FFT {y (n) e-j2π nε / N} (14) conjugate of the estimated frequency offset and applying FFT,
i.e.
3.3 Extended Kalman Filtering xˆ(n) = FFT{ y(n) e -j 2 π n ε(n) / N} (25)
A state space model of the discrete Kalman filter is As the estimation of the frequency offset by the EKF
defined as scheme is pretty efficient and accurate, it is expected that the
z(n) = a(n) d(n) + v(n) (15) performance will be mainly influenced by the variation of the
In this model, the observation z(n) has a linear AWGN.
relationship with the desired value d(n). By using the discrete
Kalman filter, d(n) can be recursively estimated based on the 4.3 Algorithm
observation of z(n) and the updated estimation in each 1. Initialize the estimate εˆ(0) and corresponding state
recursion is optimum in the minimum mean square sense. error P(0)
The received symbols in OFDM System are 2. Compute the H(n), the derivative of y(n) with respect to
y(n) = x(n) ej 2 π n ε(n) / N + w(n) (16) ε(n) at εˆ(n-1) the estimate obtained in the previous
where y(n) the received symbol and x(n) is the FFT iteration.
of transmitted symbol. It is obvious that the observation y(n) is 3. Compute the time-varying Kalman gain K(n) using the
in a nonlinear relationship with the desired value ε(n), i.e error variance p (n-1), H(n), and σ2
y(n) = f(ε(n)) + w(n) (17) 4. Compute the estimate yˆ(n) using x(n) and εˆ(n-1) i.e.
where f(ε(n)) = x(n) ej 2 π n ε(n) / N (18) based on the observations up to time n-1, compute the
In order to estimate ε(n) efficiently in computation, error between the true observation y(n) and yˆ(n)
we build an approximate linear relationship using the first- 5. Update the estimate εˆ(n) by adding the K(n)-weighted
order Taylor’s expansion: error between the observation y(n) and yˆ(n) to the
y(n)≈f(εˆ(n-1))+f'(εˆ(n-1))[ε(n)-εˆ(n-1)]+w(n) (19) previous estimate εˆ(n-1)
6. Compute the state error P(n) with the Kalman gain K(n),
where εˆ(n-1) is the estimate of ε(n-1). H(n), and the previous error P(n-1).
To Define 7. If n is less than Np, increment n by 1 and go to step 2;
z(n) = y(n) – f(εˆ(n-1) (20)
d(n) = ε(n) - εˆ(n-1) (21) otherwise stop.
and the following relationship It is observed that the actual errors of the estimation εˆ(n) from
z(n) = f'(ε(n-1)) d(n) + w(n) (22) the ideal value ε(n) are computed in each step and are used for
adjustment of estimation in the next step.
5. CONCLUSION
It is observed from the figures that Extended
Kalman filter method indicates that for very small frequency
offset, it does not perform very well, as it hardly improves
BER. However, for high frequency offset the Kalman filter
does perform extremely well. Important advantage of EKF
method is that it does not reduce bandwidth efficiency as in
self cancellation method because the frequency offset can be
estimated from the preamble of the data sequence in each
OFDM frame.
Self cancellation does not require very complex
hardware or software for implementation. However, it is not
Figure 3: BER performance with ICI bandwidth efficient as there is a redundancy of 2 for each
Cancellation for ε=0.25 carrier. The ML method also introduces the same level of
Figure 3 shows that for small frequency offset redundancy but provides better BER performance, since it
values, ML and SC methods have a similar performance. accurately estimates the frequency offset. EKF
However, ML method has a lower bit error rate for increasing implementation is more complex than the ML method but
values of Eb/No. provides better BER performance.
Further work can be done by extending the concept
of self-ICI cancellation and by performing simulations to
investigate the performance of these ICI cancellation schemes
in multipath fading channels.
6. REFERENCEs
[1] Ramjee Prasad, “OFDM for wireless communication
system”,Artech House,2004.
[2]S.Weinstein and P.Ebert, ‘Data transmission by
frequency-division multiplexing using the discrete Fourier
transform,’ IEEE Trans. Commun.,vol.19, pp. 628-634, Oct.
1971.
[3] L.J. Cimini, “Analysis and Simulation of a Digital Mobile
Channel Using Orthogonal Frequency Division Multiplexing”,
Figure 4: BER performance with ICI IEEE Transactions on Communication. no.7 July 1985.
Cancellation for ε=0.5 [4] Russell, M.; Stuber, G.L.; “Interchannel interference
Figure 4 illustrates that for frequency offset value of analysis of OFDM in a mobile environment”, Vehicular
0.5, BER increases for both the methods but ML method Technology Conference, 1995 IEEE 45th, vol. 2, pp. 820 –
maintains a lower bit error rate than SC.EKF is better than SC 824,.Jul. 1995
method. [5] X.Cai, G.B.Giannakis,”Bounding performance and
In figure 5, for frequency offset value of 0.75, self- suppressing intercarrier interference in wireless mobile
cancellation method has a BER similar to standard OFDM OFDM”, IEEE Transaction on communications, vol.51, pp.
system since the self-cancellation technique does not 2047-2056, no.12, Dec.2003.
completely cancel the ICI from adjacent sub-carriers and
ABSTRACT
AAAL
HLR
AU
AAAH
FA
HA
UE
(E/UPE) ME/UPE)
Figure 2 : The proposed mobility management architecture for next generation all-IP-based wireless systems
3.1 Authentication and Authorization and its HSS. IEEE 802.1x uses a special frame
format known as Extensible Authentication Protocol
The working principle of this architecture is as (EAP) over LAN (EAPOL) for transportation of
follows. When a mobile user requests service from a authentication messages between a UE and an access
FN and the FN determines that it has no SLA with point (AP). EAP [9] over RADIUS [10] or Diameter
user's home service subscriber (HSS), it forwards the [11] is used for the transportation of authentication
request to aGW to authenticate the user. Then, aGW messages between other entities. When the UE
talks to user's HSS and mediates between FN and roams into a FN, the authentication and MIP
HSS for authentication message exchanges. Once the registration are carried out as described below. Here,
user is authenticated, aGW also creates security EAP-SIM [12] is used to illustrate the authentication
associations/keys required between different network process. Note that any other authentication schemes,
entities. Finally the HSS and FN will be mutually e.g. EAP-AKA [13], EAP-SKE [14], EAP-TLS [15]
authenticated, and will have session keys for secured etc. can also be used. Figure 3 shows the location
data transfer. registration procedure.
The authentication and Mobile IP
registration processes are integrated in the proposed
architecture using the procedures defined in [7].
IEEE 802.1x port access control standard [8] is used
for end-to-end mutual authentication between a UE
7. Selection of Intersystem
Mobility Anchor GW
1. The UE discovers new access system and Request to the Home AAAH server (AAAH).
performs access system and network selection. Once the AAAH receives the MIP Registration
2. The UE sends an attach request, MIP Request containing the SIM Key Request
Registration Request including Mobile-AAA extension, first it verifies the Mobile-AAA
Authentication extension (as defined in [16]) to authentication extension. If the authentication is
the aGW. The UE also includes a SIM Key successful, it contacts the home authentication
Request extension [19] and a Network Access center (AuC) of the UE and obtains n number of
Identifier (NAI) [18], e.g. UE@relam, in its MIP triplets (RAND, SRES, Kc), where RAND
Registration Request. The SIM Key Request denotes a random number, SRES denotes the
extension contains a random number response and Kc is the key used for encryption.
(NONCE_UE) picked up by the UE, which is Then it forwards a copy of these triplets to aGW.
used for new authentication key generation as When aGW receives n triplets it derives a
discussed later in this section. UE_AAAH key (KUE_AAAH) and calculates
3. When the aGW receives the MIP Registration message authentication code (MAC) for the
Request and finds the Mobile-AAA RANDs (MAC_RAND) using [19]
Authentication extension, it learns that the UE is
a roaming user. Based on the NAI in the MIP KUE_AAAH = h(n * Kc│NONCE_UE) and
Registration Request, the aGW recognizes that
the operator does not have direct SLA with the MAC_RAND = PRF(KUE_AAAH, α) (1)
UE's HN and forwards the MIP Registration
6 4 REFERENCES
4
NIA
[1] ETSI TS 129 120 V3.0.0, “Universal mobile
eSAE telecommunications systems (UMTS); mobile
2
0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 application part (MAP) specification for
Probability of intersystem roaming
gateway location register (GLR)”, 3GPP/ETSI
2000, 2000-2003.
Figure 4 : Total cost of location registration [2] I.F. Akyildiz, W. Wang, “A new signaling
protocol for intersystem roaming in next
generation wireless systems”, IEEE Journal on
Selected Area in Communications, vol.19, no.
10, Oct. 2001, pp. 2040-2052.
[3] I.F. Akyildiz, W. Wang, “A novel distributed
dynamic location management scheme for
minimizing signaling costs in mobile IP”, IEEE
Transactions on Mobile Computing, vol. 1, No 3,
July 2002, pp. 163-175.
[4] N. Shenoy, “A framework for seamless roaming
across heterogeneous next generation wireless
Abstract - Printed microstrip antennas and simple rectangular patch fed at the center of
arrays are known to have limitations in terms radiating wall. A microstrip patch antenna is a
of bandwidth and efficiency, all imposed by the radiating patch on one side of a dielectric
very presence of the dielectric substrate. The substrate, which has a ground plane on the
paper deals with the design of a probe fed and underside. The EM waves fringe off the top
edge fed rectangular microstrip patch antenna patch into the substrate, reflecting off the
with the basic parameters W,h,L,ε r,fo to ground plane and radiates out into the air.
achieve better bandwidth and directivity with Radiation occurs mostly due to the fringing
efficient radiation pattern and Gain. The field between the patch and ground. The
analytical results for various possible radiation efficiency of the patch antenna
dimensions and different dielectric values were depends largely on the substrate permittivity
calculated for achieving bandwidth and (εr) of the dielectric[2]. The basic geometry of
directivity without any structural complexities the microstrip patch is shown in fig (1)
.The analytical results were tested by
simulating with basic design software
PCAAD,MSTRIP40. To obtain an optimum
value for the design parameters of the
microstrip antenna Support Vector Machines
(SVM), Generalised Regularisation Neural
Network (GRNN) and Back Propagation
Network (BPN) were implemented to train the
network to attain optimized values to yield
wide bandwidth and better directivity with
high Gain. The application of artificial neural
network ensures an optimum design
methodology for microstrip antenna design
which is revealed when comparing the results
with analytical methods and the results of the Figure 1.Microstrip Patch Antenna Geometry
simulation softwares.
1. Introduction
Microstrip patch antennas have been
attractive due to their conformal properties.
Mathematical modeling of the basic microstrip
radiator was initially carried out by the Ideally, a thick dielectric is preferred
application of transmission-line analogies to for broadband purposes. Small values of width
W of patch result in low antenna efficiencies
x0=1
x1 wk,0
wk,1
x2 wk,2 vk (k) yk
.
.
.
xn wk,n
Figure 2. Mathematical
definition of a neural network
X1
wk, bk
1
wk,
X2 2 Σ
.
UbiCC Journal - Volume 3 . 45
.
Fig 3. Signal flow diagram of a
neuron
Input
S
S
fo D
S
L
W S
BW
ε
r
D
P S- Summation Unit
D- Division Unit
fo
Input
With no data points between H1 and H2, and LD = Σi αi - Σi Σj αi αj yi yj Φ(xi) · Φ(xj) -(22)
distance between H1 and H2 maximized.
Some training point will lie on the hyper plane as Φ(xi ) · Φ(xj ) = K(xi , xj )
H1 and H2, they are called support vector Where K is the kernel function. This is
machines because they define the separating equivalent as the dot product in high
plane and the other training points can be dimension is equal to kernel function in input
removed or moved provided they don’t cross space. The common kernel function used is
the planes H1 and H2. The distance between Gaussian kernel,
hyper plane H1 and H2 is 2/ || w||. To
maximize the distance between the two data K (xi , xj) = e - || xi – xj || 2 / σ2 -(23)
Figure 15. Output of the Optimized output of the Fig 14. Plot to show the radiation pattern
rectangular patch antenna using MSTRIP using PCAAD
1.Dipak K.Neog, Shyam S.Pattnaik, C.Panda, Swapna Devi, Bonomali Khuntia, and Malaya Dutta, “Design of a Wideband Microstrip
Antenna and the use of Artificial Neural Networks in Parameter Calculation”, IEEE Antennas and Propagation Magazine, Vol.47,
No.3, June 2005,pp.60-65.
2. Inder J.Bahl, Prakash Bhartla and Stanislaw S. Stuchly, “ Design of Microstrip Antennas Covered with a Dielectric Layer”, IEEE
Transactions on Antennas and Propagation, Vol. AP-30, No. 2, MARCH 1982, pp. 314-318.
3. Kin-Lu Wong and Yi-Fang Lin, “Small broadband rectangular microstrip antenna with chip-resistor loading”, ELECTRONICS
LETTERS, 1 l September 1997,Vol. 33 No. 79,pp.1593, 1594.
4. S.Lebbar, Z.Guennoun, M.Drissi, and F.Riouch, “A Compact and Broadband Antenna Design Using a Geometrical- Methodology-
Based Artificial Neural Network”, IEEE Antennas and Propagation Magazine, Vol.48, No.2, April 2006,pp.146-154.
5. C. L. Mak, K. M. Luk, Senior Member, IEEE, K. F. Lee, Fellow, IEEE, and Y. L. Chow, “Experimental Study of a Microstrip Patch
Antenna with an L-Shaped Probe,” IEEE Transactions on Antennas and Propagation, VOL. 48, NO. 5, MAY 2000,pp.777-783.
6. R.K.Mishra and Patnaik, “Designing Rectangular Patch Antenna Using the Neurospectral Method”, IEEE Transactions on Antennas
and Propagation,AP-51,8 August 2003,pp.1914-1921.
7. S.S.Pattnaik, D.C.Panda and S.Devi, “Input Impedance of Rectangular Microstrip Patch Antenna Using Artificial Neural Networks”,
Microwave and Optical Technology Letters,32,5,5 March 2002,pp.381-383.
8. S.S.Pattnaik, D.C.Panda and S.Devi, “Radiation Resistance of Coax-Fed Rectangular Microstrip Patch Antenna Using Artificial
Neural Networks”, Microwave and Optical Technology Letters, 34,1,5 July 2002,pp.51-53.
9. D.M.Pozar, “Microstrip Patch Antennas,” in L.C.Godara (ed), Handbook of Antennas in Wireless Communications, New York, CRC
Press, 2001,Chapter 6.
10.Ye Bin Hu Gu Yu , “The analyze and improve TCP performance using a DSR route protocol based on signal strength”, IEEE
Wireless Communications, Networking and Mobile Computing, pp. 846 – 849, 2005.
11. Dongkyun Kim , Hanseok Bae, Jeomki Song, “Analysis of the interaction between TCP variants and routing protocols in
MANETs”, IEEE Parallel Processing, ICPP 2005 Workshops, pp 380-386, 2005.
12. Prabakaran, M. Mahasenan, A. , “Analysis and enhancement of TCP performance over an IEEE 802.11 multi-hop wireless network:
single session case”, IEEE International Conference on Personal Wireless Communications, pp-29- 33, 2005
13. Caihong Kai Yuzhong Chen Nenghai Yu, “An Improvement Scheme Applied to TCP Protocol in Mobile Ad Hoc Networks”, IEEE
International Conference on Mobile Technology, Applications and Systems, pp.1-6, 2005
ABSTRACT
The performance of transmission spaced selection diversity (SD) placed at base
station (BS) in DS-CDMA system remains insufficiently clear. This performance
will be evaluated by considering the effect of space distance between antennas
and the maximum Doppler frequency (fd) on bit error rate (BER) performance under
optimum conditions which are not clarified until now. Moreover, analysis of this
system is presented under the effect of Rayleigh fading.
Ref. [3] introduces a transformation matrix T to DS-CDMA system with three antennas at the
BS and one antenna at the MS is assumed. Fig. 1
transform the correlated received signals r1(t) and
shows propagation model at the BS. Table 1 shows
r2(t) into two new uncorrelated signals r3(t) and simulation parameters.
r4(t) therefore, Incident waves
⎡ r3 (t ) ⎤ ⎡ r1 (t ) ⎤
⎢r (t )⎥ = T ⎢r (t )⎥ (6)
⎣4 ⎦ ⎣2 ⎦
φ
⎡ 2 2⎤ θ
⎢ ⎥ #3
Where T = ⎢ 2 2 ⎥ #2 #1
(7)
⎢− 2 2⎥ d/λ
⎢⎣ 2 2 ⎥⎦
Figure1: Linear array and propagation model at
The two new received signals can be expressed as, BS.
BER
channels were disturbed by AWGN. 1.0E-02
M=1
The Performance of the diversity system
depends on correlation between antenna elements.
1.0E-03
The correlation is determined by antenna elements 80 100 120 140 160 180 200
spacing, angle spread of incident waves φ and
fd (Hz)
direction of arrival θ [5]. Thus, we have to optimize
these values to get better BER performance.
Figure 4: Maximum Doppler frequencies (fd ) vs.
BER for Eb/N0=10 dB
4 COMPUTER SIMULATION RESULTS
1.0E+00 rho=0
Fig. 2 shows the effect of arrival angle, θ, of the rho=0.1
signal on BER performance at Eb/N0=10dB. From 1.0E-01 rho=0.5
this figure, it can be concluded that changing the rho=0.9
value of θ gives slightly small effect. Therefore, we 1.0E-02
rho=1.0
BER
use in our simulation the value 300 of θ. 1.0E-03
1.0E-04
1.0E-02
1.0E-05
0 5 10 15 20
BER
SNR dB
1.0E-05
1.0E-02 0 2 4 6 8
d/λ
1.0E-03
BER
M=3/0.5λ
1.0E-04
M=2/5.25λ
6 REFERENCES
1.0E-05 M=3/5.25λ
1.0E-02
1.0E-03
1.0E-04
0 2 4 6 8 10
Eb/N0 dB
5 CONCLUSIONS
Yuri Boreisha
Minnesota State University Moorhead, USA
Boreisha@mnstate.edu
Oksana Myronovych
North Dakota State University, USA
Oksana.Myronovych@ndsu.nodak.edu
ABSTRACT
Problem solving and learning processes conducted on the basis of contemporary Web-
based DSS provide for development and enhancement of knowledge management
systems. Knowledge objects form the foundation of the conceptual approach to the
knowledge management based on the contemporary Internet technologies and
knowledge accumulated in DSS.
Alternative
Standards solutions
(Desired state) Problem
(DSS)
solver
(Manager)
Information Constraints
(Current state)
Solution
The dialog manager is also in charge for the and especially a time-series of internal company data
information visualization. Finally, access to the and sometimes external data. Relational databases
Internet, networks, and other computer-based systems accessed by query and retrieval tools provide an
permits the DSS to tie into other powerful systems, elementary level of functionality. Data warehouse
including the TPS or function-specific subsystems. systems that allow the manipulation of data by
There are many kinds of DSS. The first generic computerized tools tailored to a specific task and
type of DSS is a Data-Driven DSS. These systems setting or by more general tools and operations
include file drawer and management reporting provided additional functionality. Data-Driven DSS
systems, data warehousing and analysis systems, with Online Analytical Processing (OLAP) provide
Executive Information Systems and Spatial DSS. the highest level of functionality and decision support
Data-Driven DSS emphasize access to and that is linked to analysis of large collections of
manipulation of large databases of structured data historical data.
Internal External
Data Data
DSS
Database/
Data Warehouse
User Interface
(Dialog Manager)
Users
Report
Writing Mathematical
Models Groupware
Software
Solutions Outputs
Periodic Outputs from Outputs and from
and mathematical from explanations OLAP
special models groupware
reports
Internal External
Data Data
DSS
Database/
Data Warehouse
Web Services
provide access to
DSS Software System
Ajax-Enabled Applications
implement
Dialog Manager
Internet Users
The service-oriented architecture (SOA) provides consisting of different software components working
the theoretical model for all Web services. The model together. Consuming Web services is based on open
behind Web services is a loosely coupled architecture, standards managed by broad consortia (e.g., World
DSS
Database/
Data Warehouse
Web Service
Ajax-Enabled Application
Web-enabled DSS provide for expandable built up over the years. This organizational knowledge
collections of the knowledge objects that constitute the can be captured and stored using case-based reasoning
knowledge repository of the corresponding KMS. From (CBR). In CBR description of the past experiences of
this point of view the knowledge objects can be human specialists, represented as cases, are stored in a
considered as a knowledge representation technique. database for the later retrieval when the user encounters
a new case with similar parameters. The system
4 PROBLEM SOLVING AND LEARNING searches for stored cases with problem characteristic
similar to the new one, finds the closest fit, and applies
AI distinguishes two general kinds of learning. The the solution of the old case to the new case. Successful
first kind is based on coupling new information to solutions are tagged to the new case and both are stored
previously acquired knowledge. Typical examples together with the other cases in the knowledge base.
include learning by analyzing differences, by managing Unsuccessful solutions are also appended to the case
multiple models, by explaining experience, and by database along with explanations as why the solutions
correcting mistakes. The second kind is based on did not work.
digging useful regularity out of data; a practice often Problem-based learning (PBL) is (along with active
refers as data mining. Typical examples include learning and cooperative/collaborative learning) one of
learning by recording cases, by building identification the most important developments in contemporary
trees, by training neural nets, by training perceptrons, higher education. PBL is based on the assumption that
by training approximation nets, and by simulation human beings evolved as individuals who are motivated
evolution (e.g. genetic algorithms). to solve problems, and that problem solvers will seek
Expert systems primarily capture the tacit and learn whatever knowledge is needed for successful
knowledge of individual experts, but organizations also problem solving. PBL is a typical example of an
have collective knowledge and expertise that they have
System searches
Repository of Repository of
knowledge objects for knowledge objects
the suitable ones (based on a
Web-enabled DSS)
System stores
System finds the problem description
the closest fit and and the knowledge
provides access to object in the repository
knowledge objects
5 CONCLUSIONS 6 REFERENCES
Knowledge is a complex phenomenon, and there are [1] V. Supyuenyong, N. Islam: Knowledge
many aspects to the process of managing knowledge. Management Architecture: Building Blocks and
Knowledge-based core competencies of firms are key Their Relationships, Technology Management for
organizational assets. Knowing how to do things the Global Future, Vol. 3, pp. 1210-1219 (2006).
effectively and efficiently in ways that other [2] K.C. Laudon, J.P. Laudon: Management
organizations cannot duplicate is a primary source of Information Systems. Managing the Digital Farm,
profit and competitive advantage that cannot be Prentice Hall, pp. 428-508 (2006).
purchased easily by competitors in the marketplace. [3] R. McLeod, G. Schell: Management Information
This paper discusses Web-enabled DSS, related Systems, 10th Edition, Prentice Hall, pp. 250-274
knowledge repositories, and KMS that facilitate the (2006).
problem solving and learning. The knowledge objects [4] P.J. Deitel, H.M. Deitel: Internet and World Wide
approach to the knowledge representation allows Web. How to Program, 4th Edition, Prentice Hall,
considering contemporary DSS as integrated parts of pp. 50-117 (2008).
the corresponding KMS. [5] T. Berners-Lee, et al: A Framework for Web
Science, Foundations and Trends in Web Science,
Vol. 1, No 1, pp. 1-130 (2006).
Antipolis, France The important reason for being able to simulate so-
† Chalmers University, Goteborg, Sweden cial encounters is that it allows us to study the effect
0
10 0 1
10 10
k
Figure 3: Social network as extracted from movement
Figure 2: The degree of connectivity k of a node plot- data from one day of simulation. Black and white cir-
ted against the frequency of nodes with degree k on a cles indicate researchers and developers, respectively.
log-log scale. Each point represents data points from
one numerical simulations over a 7 period.
ple from different teams are loosely connected. Re-
sults vary across different runs but the a two-clusters
they do not consider physical interactions and face- structure was already present. Similar results were
to-face communications that are at the basis of hu- obtained by analyzing the tracking data from the
man behaviors. In this study, we focused on this last real-life sensor network (see Fig. 2 in Ref. [2]).
feature, we estimate social relationships from pat- We simulated one week of activity and measured
terns of collocation in the workplace. This approach the properties of the resulting social network. Fig. 2
will be integrated with data collected from electronic shows the degree of connectivity k versus the fre-
communications in future studies, to better specify quency of nodes with degree k for one simulation.
the structure of the network and to investigate the In general, the observed distributions do no follow a
(possible) different topologies of electronic and phys- power law (straight line in the log-log plots). This is
ical social networks. probably due to the limited sampling size: there are
We inferred the structure of the social network in few agents and a short duration of the simulation.
the office by simulating the movement of a group of In fact, due to the small size of the environment,
people for long periods and considering a simple prox- this frequency distribution converges to a delta af-
imity rule: two individuals share a link if they spend ter 6 − 8 weeks of simulations, at this time every
enough time in the vicinity of one another. In addi- agent is directly connected to everybody else. Fur-
tion, we added to the system some context specific ther investigations and more experimental data are
rules, e.g., we excluded the entrance hall. This sim- clearly required to fully characterize the topology of
ple rule can lead to a number of false positives, e.g., this network, and to assess whether the structure of
two individuals may share the same location with- the social network in a real world physical space dif-
out interacting. However, we expect that in the long fers from those measured with email or chat log files,
run and with a large number of users it provides a where spatial extension and physical constraints are
gross estimation of global structure of the network of not taken in account.
interactions and of its evolution in time. Extending the period of simulation to 4 weeks, we
Fig. 3 illustrates the social network amongst two observed the network becomes fully connected after
departments (Research and Development) after one 9 working days (on average), even if the clusters cor-
day of simulation; it shows, for example, that peo- responding to the different teams are still present at
2
[3] K. Chan and J. Liebowitz. The synergy of so-
cial network analysis and knowledge mapping:
Hops a case study. International Journal of Manage-
1.5 ment and Decision Making, 7(1):19–35, 2006.
[4] S. Counts and J. Geraci. Incorporating physi-
1 cal co-presence at events into digital social net-
working. In CHI ’05: CHI ’05 extended abstracts
on Human factors in computing systems, pages
0.5
5 10 15 20 25 30 1308–1311, New York, NY, USA, 2005. ACM
Days
Press.
Figure 4: Triangles indicate the average number of [5] R. Cross and A. Parker. The Hidden Power of
hops to connect one worker with all the workforce Social Networks: Understanding how Work Re-
(average shortest path). In the x-axis the number ally Gets Done in Organizations. Harvard Busi-
of days are reported, starting from day 8. Previous ness School Press, 2004.
days are not shown because the network is not fully [6] H. Ebel, L.-I. Mielsch, and S. Bornholdt. Scale-
connected. Vertical bars indicate standard deviations free topology of e-mail networks. Phys. Rev. E,
taken over 50 simulations. Square dots indicate the 66(3):035103, Sep 2002.
same quantity considering random encounters only.
Standard deviations for random encounters are not [7] R. Guimera, B. Uzzi, J. Spiro, and L. A. N. Ama-
shown for clarity. ral. Team Assembly Mechanisms Determine Col-
laboration Network Structure and Team Perfor-
mance. Science, 308(5722):697–702, 2005.
tions and their effects on collective behaviors.
[8] Q. Jones and S. A. Grandhi. P3 systems:
Putting the place back into social networks.
Acknowledgments IEEE Internet Computing, 9(5):38–46, 2005.
[9] H. Kautz, B. Selman, and M. Shah. Referral
Frederick Schlereth contributed to this study dur- web: combining social networks and collabo-
ing his internship at Accenture Technology Labs in rative filtering. Commun. ACM, 40(3):63–65,
Sophia Antipolis. We thank Valery Petrushin and 1997.
Gang Wei for providing tracking data obtained from
Accenture Technology Labs in Chicago. [10] D. Kirsh. Distributed Cognition, Coordination
and Environment Design. Proceedings of the Eu-
ropean conference on Cognitive Science, pages
References 1–11, 1999.
[11] S. Milgram. The small world problem. Psychol-
[1] T. Allen. Architecture and Communication ogy Today, 2(1):60–67, 1967.
Among Product Development Engineers. MIT
Press, Cambridge, MA, 1997. [12] A. Omicini, A. Ricci, M. Viroli, C. Castel-
franchi, and L. Tummolini. Coordination Arti-
[2] M. Bezzi and R. Groenevelt. Towards under- facts: Environment-Based Coordination for In-
standing and modeling office daily life. In 2nd telligent Agents. Proceedings of the Third In-
International Workshop on Exploiting Context ternational Joint Conference on Autonomous
ABSTRACT
In all types of data communication systems, errors may occur. Therefore error
control is necessary for reliable data communication. Error control involves both
error detection and error correction. Previously error detection can be done by
Cyclic Redundancy Check (CRC) codes and error correction can be performed by
retransmitting the corrupted data block popularly known as Automatic Repeat
Request (ARQ). But CRC codes can only detect errors after the entire block of
data has been received and processed. In this work we use a new and “continuous”
technique for error detection namely, Continuous Error Detection (CED). The
“continuous” nature of error detection comes from using arithmetic coding. This
CED technique improves the overall performance of communication systems
because it can detect errors while the data block is being processed. We focus only
on ARQ based transmission systems. We will show have the proposed CED
technique can improve the throughput of ARQ systems by up to 15%.
Keywords: Cyclic redundancy check codes, arithmetic coding, automatic repeat request.
ABSTRACT
This unified research describes spoken Urdu numbers investigative analysis from ‘siffar’ (zero) to
‘nau’ (nine) for making a concrete foundation in recognition of Urdu language. Sound samples
from multiple speakers were utilized to extract different features using Fourier descriptor and
Neural networks. Initial processing of data, i.e. normalizing and time-slicing was done using a
combination of Simulink in MATLAB. Afterwards, the MATLAB tool box commands were used for
calculation of Fourier descriptions and correlations. The correlation allowed comparison of the
same words spoken by the same and different speakers.
The analysis presented in this paper laid a foundation step in exploring Urdu language in
developing an Urdu speech recognition system. In this paper the speech recognition feed-forward
neural network models in Matlab were developed. The models and algorithm exhibited high
training and testing accuracies. Our major goal work involves in the future use of TI
TMS320C6000 DSK series or linear predictive coding. Such a system can be potentially utilized in
implementation of a voice-driven help setup in different systems. Such as multi media, voice
controlled tele-customer services.
Keywords: Spoken number, Fourier, Correlation, Feature extraction, Feed-forward neural networks, Learning rate
Correlation exists between objects, phenomena, or Thus X [ k ω ] has real and imaginary components
signals and occurs in such a way that it cannot be by in general, so that for the kth harmonic
chance alone. Unconsciously, the correlation is used
in everyday life. When one looks at a person, car or X ( k ) = R ( k ) + j I( k )
house, one’s brain tries to match the incoming image X (k ) = [R 2
(k ) + I 2 (k ) ]
1/ 2
(2.1)
with hundreds (or thousands) of images that are
3. FEEDFORWARD VS RECURRENT
NETWORKS |FFT| 2 yout1
Magnitude Signal To
Neural networks have proven to be a power tool for FFT Workspace1
4.4 Windowing
The MATLAB algorithm for the two dimensional Fig.5 Correlation of the spoken Urdu number
FFT routine is as follows [9]: dau (two)
600
500
400
300
200
100
0
0 50 100 150 200 250 300
300
SPEAKER: s1 w0
800 200
700
100
600
0
0 50 100 150 200 250 300
500
100 SPEAKER: s1 w4
400
0 350
0 50 100 150 200 250 300
300
150
100
SPEAKER: s1 w1 50
500
0
0 50 100 150 200 250 300
450
400
Fig.13 FFT magnitude spectrum for spoken
350
Urdu number char (four)
300
250
200
SPEAKER: s1 w5
500
150
450
100 400
50 350
300
0
0 50 100 150 200 250 300
250
200
50
0
0 50 100 150 200 250 300
300
number aik (one). 200
100
0
0 2000 4000 6000 8000 10000 12000
SPEAKER ONE: s1 w0
1800
1600
Fig.18 Magnitude spectrum of the correlation of the
1400
spoken Urdu number spoken chaar(four) by speaker-
1
1200
1000
800
SPEAKER ONE: s1 w5
1200
600
400 1000
200
800
0
0 2000 4000 6000 8000 10000 12000
600
0
0 2000 4000 6000 8000 10000 12000
SPEAKER TWO: s2 w0
600
Fig.19 Magnitude spectrum of the correlation of the
500 spoken Urdu number spoken paanch (five) by speaker-
1
400
300
SPEAKERS 15 WORD 1 surface
200
100 0.8
0.6
0
0 2000 4000 6000 8000 10000 12000 0.4
0.2
SPEAKER THREE:s3 w0
1800 Fig.20 The surface plot of the correlation of the
1600 spoken Urdu numbers spoken aik (one) by speaker-15
1400
1200
1000
4.8 Creating a Network
800
600
400
To create a feedforward neural network (MLP)
200
object in an interactive way to use the command
0
nntool to open the Neural Network Manager. We can
0 2000 4000 6000 8000 10000 12000
then import or define data and train our network in
Fig.17 Magnitude spectrum of the spoken Urdu number the GUI or export our network to the command line
siffar (zero) by speaker-3 for training [5].
% Read data from .wav files and calculate Once the network is trained, it can be used to detect
magnitudes from FFTs the correct output for the given input vector. The
% 15 speakers have spoken the same word following functions take the input pattern as an array
pronounced as "siffar" and use the function to retrieve the output from the
s1_w0 = wavread('s1_w0.wav'); network.
s1_w0_norm = s1_w0/max(s1_w0); % Creating new array 'alldat (:,:)' for NN use ...
s1_w0_fft0 = abs(fft2(s1_w0_norm));
min_arr_size= min(min_arr_size, alldat(1,1) = 1;
size(s1_w0_norm)); alldat(1,2) = 1;
………….. alldat(1,3:258) = s1_w1_fft0(1:256);
Before using the Neural Networks in the demo 4.16 Setting and Loading Test Data
program to detect and classify correct word structure
as well as able to find out the alternate word for the The demo program is capable of loading the required
incorrectly recognized word (by the Speech input for the neural network from the wave file. The
Recognition engine), we need to train the Word wave file that contains the input sentences to be
Network with the training data. The train network detected by the neural network, however should be in
uses the object to train the word frequency and a specific format so that the program can correctly
network 6000 and 1500 times respectively. The load and feed the input data to the neural network
following code snippet shows how the word network and generate the correct output.
gets trained.
load 'neural_network.mat'
cutoff = 64;
for i=1:150; % The data of 15 speakers spoken in Urdu from
for j=3:cutoff+2; siffar (zero) to nau (nine).
ptrans(i,j-2) = log(alldat(i,j)); % Minimum array size has been kept 1e10
end
end min_arr_size = 1e10;
0
10
s3_w0_fft(1:min_arr_size,1)=s3_w0_fft0(1:min_arr_
size, 1); -1
s4_w0_fft(1:min_arr_size,1)=s4_w0_fft0(1:min_arr_ 10
size, 1);
-2
10
figure(200),plot(t,'g'); 8 Epochs
0
10
5. ANALYSIS & RESULTS
-1
10
2
Performance is 0.00987841, Goal is 0.01
10
2
Performance is 0.00979687, Goal is 0.01
1
10 10
Training-Blue Goal-Black
1 0
10 10
Training-Blue Goal-Black
0 -1
10 10
-1 -2
10 10
-2 -3
10 10
0 500 1000 1500 2000
2459 Epochs
-3
10
0 10 20 30 40 50 60
Fig.26 Training on 15 speakers
65 Epochs
Layers (64,35, 35,1)
Fig.23 Training on number 7 with 12 speakers Learning rate 0.01
Layers (64,35, 35,1) Goal 0.01
Learning rate 0.01
Goal 0.01 2
10
Performance is 0.00994686, Goal is 0.01
2
Performance is 0.00912022, Goal is 0.01
1
10 10
Training-Blue Goal-Black
1 0
10 10
Training-Blue Goal-Black
0 -1
10 10
-1 -2
10 10
-2 -3
10 10
0 500 1000 1500 2000
2285 Epochs
-3
10
0 1 2 3 4 5 6 7 8 9 10 11
Fig.27 Training on 12 speakers
11 Epochs
Layers (64,35, 35,1)
Fig.24 Training on number 8 with 12 speakers Learning rate 0.01
Layers (64,35, 35,1) Goal 0.01
Learning rate 0.01
Goal 0.01 2
10
Performance is 0.00955099, Goal is 0.01
2
Performance is 0.00967774, Goal is 0.01
1
10 10
Training-Blue Goal-Black
1 0
10 10
Training-Blue Goal-Black
0 -1
10 10
-1 -2
10 10
-2 -3
10 10
0 500 1000 1500
1569 Epochs
-3
10
0 2 4 6 8 10 12
Fig.28 Training on 10 speakers
12 Epochs
Layers (64,35, 35,1)
Fig.25 Training on number 9 with 12 speakers Learning rate 0.01
Layers (64,35, 35,1) Goal 0.01
10
0
12 64, 30, 30, 1 72.59%
13 64, 35, 35, 1 94%
10
-1
Table 1. Training of Neural Network with
different number of speakers
-2
10
10
0 5 5 2 100
Table 2. Training of Neural Network with
10
-1 different number of speakers showing accuracy
rate.
-2
10
Relationship between network layers and training
10
-3 data can be formulated from the effect of learning
0 200 400 600
1243 Epochs
800 1000 1200
rate (α) on training is shown in Fig 21-30. In
experiments, a large value of α took lesser time to
Fig.30 Training on 5 speakers
train, although convergence to the steady state value
Layers (64,35, 35,1)
was noisy. It shows that there exists a trade off
Learning rate 0.01
between layers of the network (network parameter
Goal 0.01
setting), learning rate (α), and the epoch for which
the algorithm was used.
We created and tested the networks in different
configurations, especially the hidden layer size. The
following table shows the learning accuracy with 6. CONCLUSION
some of the networks. A maximum accuracy of 94%
was achieved with double hidden layers network (64, In this paper, we presented recognition analysis of Urdu
35, 35 1). Therefore double-hidden layers should numbers siffar to nau (one to nine), which is totally an
suffice for application as shown in table 1. unique idea. The data was acquired in moderate noisy
environment by word utterances of 15 different
speakers. FFT algorithm was used in MATLAB to
S No Learning
analyze the data. As expected, we found high
Neuron count accuracy
correlation among frequency contents of the same
1 64, 10, 10 81.48% word, when spoken by many different speakers.
2 64, 20, 10 81.48%
3 64, 30, 10 72.59% We have investigated creation of neural network
4 64, 40, 10 83.70% models for automatically recognizing individual Urdu
5 64, 60, 10 62.96% numbers to be specific. The feedforward neural
6 64, 80, 10 45.93% network was trained for different learning rates;
combined and individually for different goals and
[3] T. Parsons. Voice and Speech Processing. [16] S K Hasnain, Nighat Jamil, “Implementation
McGraw-Hill College Div., Inc, 1986. of Digital Signal Processing real time Concepts
Using Code Composer Studio 3.1, TI DSK
[4] S K Hasnain, Pervez Akhter, “Digital Signal TMS 320C6713 and DSP Simulink Blocksets,”
Processing, Theory and Worked Examples,” IC-4 conference, Pakistan Navy Engineering
2007. College, Karachi, Nov. 2007
[5] MATLAB User’s Guide. Mathworks Inc., 2006. [17] M M El Choubassi, H E El Khoury, C E Jabra
Alagha, J A Skaf, M A AL Alaoui, “Arabic
[6] DSP Blockset (For use with Simulink) User’s Speech Recognition Using Recurrent Neural
Guide Mathworks Inc., 2007. Networks,” Symp. Signal Processing & Info.
Tech., 2003, ISSPIT 2003, Dec. 2003, pp. 543-
[7] Samuel D Stearns, Ruth A David, “Signal 547.
Processing Algorithms in MATLAB,” Prentice
Hall, 1996. [18] M A Al-Alaoui, R Mouci, M M Mansour, R
Ferzli, “A Cloning Approach to Classifier
[8] “TMS320C6713 DSK User’s Guide,” Texas Training,” IEEE Trans. Systems, Man and
Instruments Inc., 2005. Cybernetics – Part A: Systems and Humans, vol.
32, no. 6, pp. 746-752, 2002.
[9] G. C. M. Fant. Acoustic Theory of Speech
Production. Mouton, Gravenhage, 1960.