Computers and Electrical Engineering 38 (2012) 15791594

Contents lists available at SciVerse ScienceDirect

Computers and Electrical Engineering

journal homepage: www.elsevier.com/locate/compeleceng

Performance analysis of under-modelling stereophonic acoustic echo

cancellation by adaptive ltering LMS algorithm q
Mohamed Djendi , Aouda Bounif
University Saad Dahleb of Blida, Signal Processing and Image Laboratory (LATSI), Route de Soumaa, B.P. 270, Blida 09000, Algeria

a r t i c l e i n f o a b s t r a c t

Article history: This paper addresses the eld of stereophonic acoustic echo cancellation (SAEC) with
Received 10 February 2012 adaptive ltering algorithms. In SAEC applications, using the least mean square (LMS) algo-
Received in revised form 19 June 2012 rithm, it is usually assumed that the lengths of the adaptive lters are equal to that of the
Accepted 19 June 2012
unidentied system responses. Although, in many realistic situations, under-modelled
Available online 21 July 2012
lengths adaptive lters, whose lengths are less than that of the unidentied systems
(under-modelled systems), are employed, and analysis results for the exact modelled
stereophonic LMS algorithm are not automatically appropriate to the under-modeled
lengths. In this paper, we present a statistical analysis of the under-modeled stereophonic
LMS algorithm. Exact expressions and deterministic recursive equations to the mean coef-
cients behavior of the adaptive LMS lters are derived to completely characterize and
assess the performances (transient and steady-state) of the under-modeling stereophonic
LMS algorithm. The expected theoretical behaviour is compared with Monte Carlo simula-
tions and practical experimental results, showing a very good agreement.
2012 Elsevier Ltd. All rights reserved.

1. Introduction

Acoustic echo cancellers (AECs) are necessary in applications such as mobile phones, hands-free telephony, speaker-
phones and for communication systems (i.e. audio and video conferencing) in order to reduce or fully cancel the echoes phe-
nomenon which make worse the quality of connections and communications. Theoretically, stereophonic acoustic echo
cancellation (SAEC) can be viewed as a simple generalisation of the usual single-channel acoustic echo cancellation principle
to the two channel case [13]. The purpose of echo canceller is to identify the receiving room echo paths and subtract an
estimated replica of the echo, thereby achieving cancellation. An adaptive lter is used to identify the echo paths. The output
of the adaptive lter, which is an estimate of the echo signal, can be used to reduce undesirable echoes [47].
The SAEC systems allow to have far better sound quality and sound localisation than what has been provided before. The
improvements in quality are brought by increasing the signal bandwidth and also by adding more audio channels to the sys-
tem. This last fact spurred the need for multi-channel acoustic echo cancellers [8].
The two-channel SAEC application is most attractive since only complexity issues differ for the more general multi-chan-
nel case. A basic scheme for SAEC is sketched in Fig. 1, where we illustrate the concept with a transmission room on the left
and a receiving room on the right. The transmission room is sometimes referred to as the far-end and the receiving room as
the near-end. As depicted in Fig. 1, the echo is due to acoustic coupling between the loud-speakers and microphones in the
receiving room. In this scheme, the acoustic echo paths h1 and h2 in the local room are modelled by adaptive FIR lters h ~ v 1 n
~ v 2 ^
and h n, from which their added outputs produces an estimate y of the true echo y. Indeed, the physical impulse responses

q
Reviews processed and approved for publication by Editor-in-Chief Dr. Manu Malek.
Corresponding author.
E-mail addresses: m_djendi@yahoo.fr, djendi.mohamed@gmail.com (M. Djendi).

0045-7906/$ - see front matter 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.compeleceng.2012.06.008
1580 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594

Fig. 1. Schematic diagram of a stereophonic echo canceller. Two adaptive ltering algorithms are used between Room 1 and Room 2 where: Room 1 is the
transmitting room and Room 2 is the receiving room.

h1 and h2 are of innite length; nevertheless it is assumed that the lters hv1 and hv2 are sufciently long, in the sense that
the tails of h1 and h2 not modelled by h~ v 1 n and h
~ v 2 n have low energy and thus can be neglected. Speaking in the sequel of
true impulse responses means that we only consider the rst parts of h1 and h2 which contain most of the energy, and
which are assumed to be of the same size L as the model lters h ~ v 1 n and h
~ v 2 n.
In SAEC for teleconferencing, we have a fundamental problem of the possibility to identify the true impulse responses of
the acoustic echo paths. This problem arises from the correlation between the two signals picked up in the remote room in
this request. SAEC is fundamentally different from traditional mono echo cancellation. A SAEC, straightforwardly imple-
mented, not only would have to track changing echo paths in the receiving room but also in the transmission room. Thus,
a generalisation of the mono AEC in the stereo case does not result in satisfactory performance. The problems of SAEC were
rst described in [3] and later on in [4]. The fundamental problem is that the two channels may carry linearly related signals
which in turn may make the normal equations, to be solved by the adaptive algorithm, singular. This implies that there is no
unique solution to the equation but an innite number of solutions and it can be shown that all solutions (but the physically
true one) depend on the transmission room. As a result, intensive studies have been made of how to handle this properly [8].
Generalisation of the solution to the normal equations in a more practical sense was addressed in references [4,5,8]. It was
explained that in practice, the problem is not actually singular but extremely ill-conditioned due to the fact that the length of
the adaptive lter is shorter than the echo paths of the transmission room. Furthermore, in practice, the transmission room is
not completely stationary, i.e. smooth continuous changes exist, which slightly improves the situation by making the prob-
lem somewhat less ill-conditioned [9,10]. A complete theory of non-uniqueness and characterisation of the SAEC solution
was presented in [11,12]. It is shown that the only solution to the non-uniqueness problem is to reduce the correlation be-
tween the stereo signals and an efcient low complexity method for this purpose was also given in [1113].
In [14], the authors present a combination of mono and stereo echo cancellation which has the benet of lower complex-
ity than a pure stereo solution. Currently, attention has been focused on the investigation of other methods that decrease the
cross-correlation between the channels in order to get well-behaved estimates of the echo paths [15]. The main problem is
how to reduce the correlation sufciently without affecting stereo perception and sound quality. Early examples of SAEC
implementations can be found in [1618]. These proposed solutions were presented before the theory and limitations of
SAEC were fully understood, and were mainly based on the use of a single adaptive lter for each return channel. Recently,
several methods and technique are proposed to solve this problem in the time domain as in [1923], and also in the fre-
quency domain as proposed in [2426]. The performance of the SAEC is strictly affected by the choice of algorithm more than
in the monophonic case. This is easily recognised since the performance of most adaptive algorithms depends on the con-
dition number of the input signal covariance matrix.
In SAEC applications, the condition number is very high, and algorithms such as the LMS or the NLMS that do not take the
coherence between the input signals into account, converge very slowly to the theoretical solution. It is consequently very
interesting to study multi-channel adaptive ltering algorithms. A framework for multi-channel adaptive ltering can be
found in references [18,27].
In this paper, we focus our interest on the case where the length of the adaptive LMS lters, employed with the SAEC
application in the receiving room, are less than the length of the real lters. We propose a new study of this case, in terms
of the mean behaviour convergence of the coefcient error vectors, and show theoretical results that are very close to the
performed ones.
 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594 1581

This paper is organised as follows: Section 2 explains the SAEC problem and describes the fundamental differences be-
tween mono and stereo acoustic echo cancellation. In Section 3, we present the performance analysis of the under-modelled
stereophonic LMS algorithm in terms of the mean solutions of the coefcients error vector and of the steady-state values.
Finally, in Section 4, we give simulation results obtained with: (1) highly correlated Gaussian inputs, (2) white Gaussian in-
puts and (3) weakly correlated Gaussian inputs.
The notations that we have used in this paper are fairly standard. Boldface symbols are used for vectors and matrices. We
also have the following notations:

L ~ v 1 n and h
the length of the adaptive lters h ~ v 2 n,
L L
M the length of the impulse responses in the transmitting room GvM1 and GvM2 ,
N v1 v2
the length of the impulse responses in the receiving room hN and hN ,
E denotes mathematical expectation,
n denotes discrete time index,
(.)T denotes transpose operator,

2. The stereophonic acoustic echo cancellation SAEC problem

In this section, we show and recall from [11] that the solution of the normal equation of SAEC applications is not as evi-
dent as in the single-channel case [2831]. Indeed, since the two input signals are obtained by ltering from a common
source, a problem of non-uniqueness is expected [2]. In the following discussion, we consider the following lengths: (i)
(M) is the length of the impulse responses in the transmission room, (ii) (L) is the length of the modelling lters, and nally
(iii) (N) represents the length of the impulse responses in the receiving room [11]. In our study, we suppose that the distant
room system is stationary, linear and time invariant therefore, we have the following relation:
 T  T
v1
XM n GvM2 XM
v2
n GMv1
1
 T  T
where the two vectors GM g 1v 1 ; g v2 1 ;    :; g vM1 and GvM2 g v1 2 ; g v2 2 ;    ; g vM2 are the impulse responses of the sources-to-
v1

microphones acoustic paths in the remote0 room as indicated in Fig. 1, and the two following vectors XvM1 n and XvM2 n:
v1
XM n x1 n; x1 n  1;    ; x1 n  M 1T
v2
XM n x2 n; x2 n  1;    ; x2 n  M 1T
represents the vectors of signal samples of the microphones outputs in the same room (transmission room the following, we
suppose the recursive least square (RLS) cost function (see Fig. 1 for notations) as follows:
X
n
Jn wnp en2 2
p1

where
 T  T
en yn  h~ v 1 n Xv 1 n  h~ v 2 n Xv 2 n 3
L L L L

is the error signal at time n between the microphone output

 T  T
v1 v2
yn hN n XNv 1 n  hN n XNv 2 n 4

is the stereophonic acoustic echo estimated and to be cancelled by the two-channel adaptive lters h ~ v 1 n and h
~ v 2 n, and w
L L
(0 < w 6 1) is an exponential forgetting factor. The vectors included in (3) and (4) are dened as follows:
h iT h i
v1 v1 v1 v1 v2 v2 v2 v2 T
hN h1 ; h2 ;  ; hN and hN h1 ; h2 ;    ; hN

XNv 1 n x1 n; x1 n  1;    ; x1 n  N 1T and XNv 2 n x2 n; x2 n  1;    ; x2 n  N 1T

h iT h iT
~ v1 h
h ~v 1 ; h
~v 1 ;    ; h
~v 1 and h
~ v2 h~v 2 ; h
~v 2 ;    ; h
~v 2
L 1 2 L L 1 2 L

XL n x1 n; x1 n  1;    ; x1 n  L 1 and XLv 2 n x2 n; x2 n  1;    :; x2 n  L 1T

v1 T

In the following, we will address the problem of SAEC by the minimisation of the cost function of (2). This mathematic
minimisation leads to the following solution [5]:
" #
~ v 1 n
h L
RL n rL n 5
~ v 2 n
h L
1582 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594

where RL(n) is the correlation matrix which is given by the following expression:
" #
Xn
Xv 1 n  v 1 T  T   R v 1 v 1 RX v 1 X v 2

v2 X X
RL n wnp Lv 2 XL n XL n 6
p1 XL n RX v 2 X v 1 RX v 2 X v 2

We note that the parameter rL(n) represents the correlation vector between the input signals and the output signal in the
local room and given by the following:
" #
Xn
Xv 1 n
rL n wnp yp Lv 2 7
p1 XL n

is an estimate of the cross-correlation vector between the input and output signals.
Our aim is to obtain the optimum lters from (5). Now, consider the vector:
h iT
v2 v1
U GM  GM 8

It can be readily veried by using (1) that RL(n)U = 0, which means that the matrix RL(n) is not invertible. Therefore, there
is no unique solution to the problem of minimising (2), and the adaptive algorithm drives to any of the possible solutions,
which can be very different from the optimal predictable solution h ~ v 1 n hv 1 and h
~ v 2 n hv 2 , where these vectors are
L 1 L 2
dened as follows:
h i h i
v1 v1 v1 v1 T v2 v2 v2 v2 T
hL h1 ; h2 ;    ; hL and hL h1 ; h2 ;    ; hL

Now, we can discuss the possibilities that this matrix can or not to be invertible. However, in practical situation, there are
-at least- two reasons that make this matrix invertible:

(i) In real situation, the two vector signals XvL 1 n and XvL 2 n at the outputs of the distant room and at the input of the
local room contain noise components that are uncorrelated and
(ii) the two adaptive lters h ~ v 1 n and h
~ v 2 n that model the impulse responses of the local room are of nite FIR length, so
L L
the size of the two input vectors XvL 1 n and XvL 2 n is much smaller that the length of the vectors GvM1 and GvM2 , and thus
the relation (1) is not satised. For this reason, the matrix RL(n) becomes invertible (but is ill-conditioned because the
two input signals are strongly correlated) and the true solution h ~ v 1 n hv 1 and h
~ v 2 n hv 2 can be found
L L L L
accordingly.

In order to study the possibilities of inverting the matrix RL (n) according to the length M of the impulse response on the
transmitting room and also according to the length L of the adaptive lters, we have selected two obvious cases that can be
discussed [2] as follows:
 T  T T
(i) In the rst case, we assume that L P MWe consider the vector U GvM2 0    0;  GvM1 0    0 containing

2x(L  M) zeros coefcients. In this case, we can easily verify using (1) that R(n)U = 02L1, as a result R(n) is not
invertible
(ii) In the second case, we assume that L < MWe recall that this is the practical case, since the real two-channel lters GvM1
and GvM2 are actually of innite length. However, relation (1) can be rewritten as follows:
 T  T
XvL 1 n GLv 2 q1 n  L XLv 2 n GLv 1 q2 n  L 9

where the two scalars q1(n  L)q2(n  L) are given by:

X
M
q1 n  L x1 n  ig iv 2
iL1

X
M
q2 n  L x2 n  ig iv 1
iL1

From (1), we know that XvM1 n and XvM2 n are linearly related, but from (9) we can see that the same is not true for XvL 1 n and
XvL 2 n [except if q1(n  L) = q2(n  L) which happens only when GvM2 and GvM2 have at least M  L common zerosan event that
rarely occurs in practice]. Hence, in principle, the covariance matrix is full-rank, but is very ill-conditioned because q1(n  L)
and q2(n  L) are in general very small. Thus, for the practical case when L < M, there is a unique solution to the normal equa-
tion, although the covariance matrix is very ill-conditioned.
In the next Section 3, we will present a new study of the under-modelling stereophonic acoustic echo cancellation (SAEC)
situation of the receiving room impulse responses by the use of the stereophonic LMS algorithm. To our knowledge, this pro-
posed study is new and highlights the stereophonic LMS algorithm behaviour in the decient case where the two-channel
 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594 1583

adaptive lters lengths are under-modelled. In this analysis, we propose new theoretical model and formulas that describe
and predict perfectly the mean behaviour of the adaptive LMS lters of the SAEC applications. The following proposed the-
oretical analysis allows to perfectly predict the behaviour of the decient LMS algorithm used in SAEC congurations in the
transient and steady state phases. The accuracy of our proposed theoretical analysis and model are conrmed through
Monte-Carlo simulations results which will be nely described and presented in Section 4.

3. New performance analysis of the under-modelling stereophonic LMS algorithms

In this section, we present the new proposed performance analysis of the under-modelling SAEC case with the LMS algo-
rithm. Therefore, in this proposed analysis, we suppose that impulse responses in the receiving room, between the two loud-
speakers and the microphone, can be rewritten as follows:
v1     T
hN h1 h 1 10
v2    T
hN h 2 h2 11
where

h1and h  represents the new modelled and under-modelled adaptive lter of the rst channel (see Fig. 1), and h ; h

1 2 2
represent the modelled and under-modelled adaptive lter of the second channel and are given as follows:
h i h iT
 v1 v1 v1 T   hv 1 ; hv 1 ;    ; hv 1 and
h1; h1 ; h1 ;    ; hL ; h 1 L1 1 N
h i h iT
 v2 v2 v2 T   hv 2 ; hv 2 ;    ; hv 2
h2; h1 ; h1 ;    ; hL ; h 2; L1 1 N

The input signals for the under-modelled SAEC application of the LMS algorithm for the two-channel are given:
h iT
XNv 1 n Xv 1 nXv 1 n 12
h iT
XNv 2 n Xv 2 nXv 2 n 13
v1
X n x1 n  1; x1 n  2;    ; x1 n  L ; T
Xv 1 n x1 n  L  1; x1 n  M  2;    ; x1 n  NT
Xv 2 n x2 n  1; x2 n  2;    ; x2 n  LT ; Xv 2 n x2 n  L  1; x2 n  M  2;    ; x2 n  NT
In the following, we propose to dene new errors ltering Z1(n) and Z2(n) for the two channels. These error vectors are
computed between modelled lters coefcients and impulse responses coefcients in the receiving room and given as
follows:
~ v 1 n  h n
Z1 n h 14
L 1
~ v 2 n  h n
Z2 n h 15
L 2

According to Fig. 1, the wanted output signal (the microphone output in the receiving room) is:
 T  T
v1 v2
yn XNv 1 n hN XNv 2 n hN 16

The output signal estimated by adaptive lters is given by:

 T  T
~n XLv 1 n h
y ~ v 1 n Xv 2 n h
L L
~ v 2 n
L 17
h iT h iT
~ v 1 n and h
Where the vectors h ~ v 2 n are given by h
~ v1 h~v 1 ; h
~v 1 ;    ; h
~v 1 and h
~ v2 h~v 2 ; h
~v 2 ;    ; h
~v 2 .
L L L 1 2 L L 1 2 L

~n is:
The error signal at time n between the microphone output in the receiving room y(n) and its estimate y
~n
en yn  y 18
From (16) and (17), the error in (18) for the two channels can be expressed as:
 T   v1
e1 n Xv 1 n h T
1  X n Z 1 n n1 n 19
   Xv 2 nT Z2 n n n
e2 n Xv 2 nT h 20
2 2

Where n1(n) and n2(n) are stationary, zero-mean, and independent noise sequences that are uncorrelated with any other
signal. The combination of these last two equations gives the nal ltering error (see Fig. 1.):
en e1 n e2 n 21
Or equivalently, we get the new expression of the global ltering errors:

en Xv 1 nT h    Xv 1 nT Z1 n  Xv 2 nT Z2 n nn
  Xv 2 nT h 22
1 2

where n(n) = n1(n) + n2(n).

1584 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594

3.1. New update formulas of adaptive stereophonic LMS lters

In the classical SAEC application showed by Fig. 1, we have used two adaptive lters based LMS algorithms h~ v 1 n and
L
~ v 2 n for the two-channel according to Fig. 1. These two adaptive LMS algorithms increment their coefcients according
h L
to the following recursion:

h ~ v 1 n l enXv 1 n
~ v 1 n 1 h 23
L L 1

h ~ v 2 n l enXv 2 n
~ v 2 n 1 h 24
L L 2

with l1, l2 are the step sizes of the adaptive lters and e(n) is the lter error. We recall that Xv1(n) Xv2(n) are the two
input signals of the two adaptive ltering algorithms h~ v 1 n and h
~ v 2 n, respectively. In the next section, we will provide
L L
a theoretical understanding for the under-modelling stereophonic LMS error vectors evolutions in times and its mean
convergence.

3.2. Error vectors update of under-modelling stereophonic adaptive LMS coefcients

In this section, we present the proposed formulas update for the stereophonic adaptive LMS algorithm coefcient in the
under-modelled situation. For that reason, we dene the error vectors between modelled lters coefcients and impulse re-
sponses coefcients in the receiving room at time (n + 1) as follows:
~ v 1 n 1  h n 1
Z1 n 1 h 25
L 1
~ v 2 n 1  h n 1
Z2 n 1 h 26
L 2
 
since the real coefcients lters h1 n 1 and h2 n 1 have constant values, i.e.
  
h1 n 1 h1 n h1 27
  
h2 n 1 h2 n h2 28
and by the introduction of (25) and (27) in (23), (26) and 28 in (24), we get the following new recursive formulas for the
stereophonic LMS coefcient error vectors Z1(n + 1) and Z2(n + 1), respectively
~ v 1 n l enXv 1 n  h
Z1 n 1 h L 1 1
  T  h i h i
v1 v1   Xv 1 n l Xv 2 nT h
  Xv 1 n  l Xv 1 n
I  l1 X n X n Z1 n l1 Xv 1 nT h 1 1 2 1

 Xv 2 nT Z2 n l1 nnXv 1 n 29
~ v 2 n l enX n  v2 
Z2 n 1 h L 2 h2
h i h i
I  l2 Xv 2 nXv 2 nT Z2 n l2 Xv 2 nT h
  Xv 2 n l Xv 1 nT h
2 2
  Xv 2 n  l Xv 2 n
1 2

 Xv 1 nT Z1 n l2 nnXv 2 n 30
As it is explained before, we will present, in the next Section 3.3, the exact derivation of the new and proposed determin-
istic recursive formulas for the mean weight of the stereophonic adaptive LMS coefcients and also its steady state in the
permanent regime.

3.3. Mean convergence of under-modelling stereophonic adaptive LMS coefcients

~ v 1 n and h
The mean behaviour of the under-modelled LMS coefcient error vectors of the two adaptive lters h ~ v 2 n can
L L
now be determined by taking the expected value of both sides of (29) and (30), and using the independence assumption to
yield the following relations:
nh i o n h i o
EfZ1 n 1g EfI  l1 Xv 1 nXv 1 nT Z1 ng l1 E Xv 1 nT h
  Xv 1 n E l Xv 2 nT h
1 1
  Xv 1 n
2

 l1 EfXv 1 nXv 2 nT Z2 ng 31

nh i o nh i o
EfZ2 n 1g EfI  l2 Xv 2 nXv 2 nT Z2 ng l2 E Xv 2 nT h
  Xv 2 n l E Xv 1 nT h
2 2
  Xv 2 n
1

 l2 EfXv 2 nXv 1 nT Z1 ng 32

or equivalently
EfZ1 n 1g I  l1 R1 nEfZ1 ng l1 b1 n C1 n  l1 R12 nEfZ2 ng 33
EfZ2 n 1g I  l2 R2 nEfZ2 ng l2 b2 n C2 n  l2 R21 nEfZ1 ng 34
 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594 1585

where the matrix R1, R2, R12 and R21 are dened by the following expressions:

(i) R1 = E{Xv1(n)(Xv1(n))T} and R2 = E{Xv2(n)(Xv2(n))T} are the autocorrelation matrix of the two sources signals in the
transmitting room , respectively,
(ii) R12 = E{Xv1(n)(Xv2 (n))T} and R21 = E{Xv2(n)(Xv1 (n))T} are the crosscorrelation matrix of the two sources signals in the
receiving room, respectively, and the vectors b1 , b2 , C1 and C2 are respectively dened by the following expressions:

b1 b11 ; b12 ;    ; b1L T ; b2 b21 ; b22 ;    ; b2L T ;

T
C1 c11 ; c12 ;    ; c1L  ; C2 c21 ; c22 ;    ; c2L T

whose ith components are given, respectively

N 
X  X
N
v1 v1
b1i hj Efx1 n  j 1x1 n  i 1g hj r x1x1 j  i 35
jL1 jL1

N 
X  N 
X 
v2 v2
b2i hj Efx2 n  j 1x2 n  i 1g hj r x2x2 j  i 36
jL1 jL1

N 
X  X
N
v2 v2
c1i hj Efx2 n  j 1x1 n  i 1g hj rx2x1 j  i 37
jL1 jL1

N 
X  N 
X 
v1 v1
c2i hj Efx1 n  j 1x2 n  i 1g hj r x1x2 j  i 38
jL1 jL1

According to these last expression, we can say that when L = N, the components b1i = b2i = c1i = c2i = 0. Recall that
k1 max 6 tr(R1) and k2 max 6 tr(R2) , where k1 max and k2 max are the maximum eigenvalues of R1 and R2, respectively, and
tr(.) is the trace operator; then, from (33) and (34), the under-modelled length stereophonic LMS algorithms are convergent
in the mean if the step sizes satises the following conditions
2 2
0 6 l1 6 and 0 6 l2 6 39
k1 max k2 max
According to this last relation, we can easily conclude that the stability conditions for the decient stereophonic adaptive
LMS algorithms are similar to that of the sufcient exact adaptive LMS algorithm for L = N case.

3.4. Steady state formulas of under-modelling stereophonic adaptive LMS coefcients

In the following, we will present the new formulas that describe theoretically the steady state of the under-modelling
stereophonic adaptive LMS coefcients. Therefore, the steady state solutions of relations (33) and (34) are given, respec-
tively, by
EfZ1 1g I  l1 R1 nEfZ1 1g l1 b1 n C1 n  l1 R12 nEfZ2 1g 40
EfZ2 1g I  l2 R2 nEfZ2 1g l2 b2 n C2 n  l2 R21 nEfZ1 1g 41
and the nal steady state solution values of (40) and (41) are, respectively, given by
h i1 h i
EfZ1 1g R1 n R12 nR1
2 nR 21 n b1 n C1 n  R12 nR1
2 nb2 n C2 n 42
h i1 h i
EfZ2 1g R2 n R21 nR1
1 nR 12 n b2 n C2 n  R21 nR1
1 nb1 n C1 n 43

From the two formulas of (14) and (42), then from (15) and (43), we obtain the steady-state mean coefcient vectors of
~ v 1 n and h
the stereophonic LMS adaptive lters h ~ v 2 n respectively, as follows
L L
n o
E h~ v 1 1 h w n 44
L 1 1
n o
E h~ 1 h w n
v 2 
45
L 2 2

Where the two vectors w1(n) and w2(n) are respectively expressed by the following relations:
h i1 h i
w1 n R1 n R12 nR1
2 nR 21 n b1 n C1 n  R12 nR1
2 nb2 n C2 n 46
h i1 h i
w2 n R2 n R21 nR1
1 nR 12 n b2 n C2 n  R21 nR1
1 nb1 n C1 n 47
1586 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594

However, the two Eqs. (44) and (45) indicate that correlated input signals Xv1(n) and Xv2(n) result in some nonzero stea-
dy-state coefcients biases quantied by w1(n) (see (46)) and w2(n)(see (47)), respectively. These two biases depend on the
following points:

(i) The degree of the under-modelled SAEC under-modelled lters inuenced by L relative to N.
(ii) The coefcients values h   and h
 .
1 2
(iii) The input signal statistics.

In the case when the input signals Xv1(n) and Xv2(n) are white, we get b1(n) = b2(n) = C1(n) = C2(n) = 0NL1, hence the two
parameter vectors w1(n) and w2(n), given by (46) and (47) respectively, are zero and
n o
E h~ v 1 1 
h1 48
L
n oWhite
E h~ v 2 1 
h2 49
L
White

We notice that Eqs. (48) and (49) prove the well-known fact that for uncorrelated input signals Xv1 (n) and Xv2(n), the
under-modelled stereophonic LMS adaptive lters h ~ v 1 n and h
~ v 2 n converge in the mean to the rst coefcients of un-
L L
 
known system impulse responses h1 and h2 , respectively. It should be mentioned that the bias phenomenon in the stereo-
phonic under-modelled length LMS algorithm when the input is correlated appears also in the hierarchical LMS algorithm
[32], which was shown to converge to a biased solution for correlated input signals and, under certain circumstances, for
white inputs [3335]. All these theoretical expressions will be approved and validated in the next section by simulations.

4. Experimental results

Simulation examples are presented here in order to verify the proposed mean theoretical results derived in this paper.
Experiments are performed in a system identication setup. A zero-mean white Gaussian noise is added to the desired signal
y(n). The adaptive lter coefcients were initialized with zeros. In this paper, we have used Monte Carlo simulation principle
to validate theoretically and experimentally the decient stereophonic LMS algorithms behaviour by our proposed analysis.
Therefore, the Monte-Carlo simulations results of our theoretical analysis and that of the experimentation are obtained by
mean averaging over 1000 independent runs of both the theoretical simulated formulas and that of the experimentation.
We have used Monte Carlo method to carry out three kinds of simulations when the observations and the inputs are: (i)
highly correlated Gaussian; (ii) Gaussian white; and (iii) weakly correlated Gaussian.

4.1. Highly correlated Gaussian input

In this experiment, the unknown system is a 20-coefcient nite impulse response (FIR) lter, i.e. N = 20, with,
v1
hN 0:01;0:02; 0:04; 0:08;0:15;0:45; 0:15; 0:6; 0:6;0:45;0:3; 0:15; 0:08;0:04;0:02; 0:01;0:5;0:3; 0:2;0:2T
v2
hN 0:02;0:03;0:45;0:81;0:61; 0:30;0:72;0:81;0:35; 0:47; 0:7;0:45;0:31; 0:25; 0:4;0:32;
 0:21; 0:18; 0:15; 0:001T ;
v1 v2
This two lters hN and hN are shown by the following gures Figs. 2 and 3, respectively. In the chosen model for sim-
ulation, both the unknown systems and the adaptive FIR lters are excited with a highly correlated signal generated by the
following recurrences:

xv 1 n 0:9xv 1 n  1 g1 n 50
xv 2 n 0:9xv 2 n  1 g2 n 51

where g1(n) and g2(n) are a zero-mean uncorrelated Gaussian sequences with unity variances. We select
L = 15, 2 (n) = E(f2) = 0.00015, and use small step sizes l1 = l2 = 0.0004 and large ones l1 = l2 = 0.001. The large step-sizes
values are selected close to the maximum one that ensures algorithm convergence. We note that this correlated signals input
will be used 1000 times to test the SAEC system of Fig. 1.
However, the obtained results for the theory and for the simulation with this input (observations) will be mean averaged
at the end of the ltering process to satisfy the Monte Carlo simulation principle. We note also that in all the carried out
simulations, we have selected the lengths of the SAEC adaptive LMS lters (L) according to two different situations where
the under-modelization is slightly decient (i.e. in the case of N = 20 and L = 15), or is seriously under-modelled and decient
(i.e. in the case where N = 20 and L = 10).
In Figs. 4 and 5, we have shown the mean behaviour of the sixth and eighth coefcient errors obtained from Monte Carlo
simulations and determined from the theoretical expressions in (33) and (34), respectively. These two gures highlight the
performances of the proposed analysis in the situation when the SAEC adaptive LMS lters are slightly decient and under-
modelled. The obtained mean averaged results conrm the accuracy of our analyses. We have also observed that the con-
 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594 1587

0.8
V1
hN
0.6

0.4

0.2
Magnitude

-0.2

-0.4

-0.6

-0.8
0 2 4 6 8 10 12 14 16 18 20
Samples
v1
Fig. 2. The impulse responses hN in the receiving room.

1
hV2
N
0.8

0.6

0.4

0.2
Magnitude

-0.2

-0.4

-0.6

-0.8

-1
0 2 4 6 8 10 12 14 16 18 20
Samples
v2
Fig. 3. The impulse responses hN in the receiving room.

vergence speed characteristic of the two adaptive lters is inversely proportional to the step size values, i.e. when the step
size values of l1 and l2 are chosen small, the convergence speeds of the coefcients error vectors toward zero (the sixth and
the eighth ones) are slow and, in contrary when the step size values are large (close to the maximum ones that ensures algo-
1588 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594

Fig. 4. Comparison of theoretical and simulation curves of the mean behaviour of the sixth and eighth coefcient errors of the under-modelled length LMS
adaptive h~ v 1 n algorithm for correlated Gaussian input data, and N = 20; L = 15; input SNR = 40dB. Results obtained through 1000 independent runs of a
L
Monte-Carlo simulation.

Fig. 5. Comparison of theoretical and simulation curves of the mean behaviour of the sixth and eighth coefcient errors of the under-modelled length LMS
adaptive h~ v 2 n algorithm for correlated Gaussian input data, and N = 20; L = 15; input SNR = 40dB. Results obtained through 1000 independent runs of a
L
Monte-Carlo simulation.
 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594 1589

rithm convergence), the convergence speed toward zeros is fast. Finally, we can say that these Mont Carlo simulation results
conrm the accuracy of our proposed analysis when the SAEC adaptive LMS lters are slightly under-modelled.
In Figs. 6 and 7, we illustrate the mean behaviour of the sixth and eighth coefcient errors obtained from Monte Carlo
simulations and determined from the theoretical expressions in (33) (for the adaptive lter h ~ v 1 n and (34) (for the adaptive
L
lter h~ v 2 n, respectively.
L
The only difference between this experiment and the previous one is the under-modelled length of the adaptive lter that
is taken equal to L = 10 (i.e. in the case of seriously under-modelled condition). We have carry out this experience with the
same values of the adaptive step-sizes as in the previous experiment. This experiment also validates our analysis even when
the under-determined length of the adaptive lters (L = 10) is very small according to the exact length modelisation. We note
that the convergence speed of the two coefcients, in this case, is slow according to the rst experiment when the length of
the adaptive lters L is close to that of the real ones (N). We observe that in this case when then adaptive lters are seriously
under-modelled, the convergence speed of the SAEC adaptive LMS lters is degraded but this does not affect the performance
and the accuracy of our proposed analysis. This Monte Carlo simulation results proved another time the validity of the anal-
ysis, as the theoretical results were shown to accurately predict the actual behaviour of the under-modelled length stereo-
phonic LMS algorithm, both during the transient and steady state phases.

4.2. White Gaussian input

In this experiment, we have used a similar model to that previously described in the highly correlated input case, ex-
cept that both the unknown system and the adaptive lter are excited by zero-mean white Gaussian signal inputs of unity
variance. We have selected for the two adaptive lters h ~ v 1 n and h
~ v 2 n, a large step sizes close to the maximum step size
L L
value (l1 = l2 = 0.005), and a small ones (l1 = l2 = 0.0005). We note that this experiment is carried out for the situation
where the SAEC adaptive LMS lters h ~ v 1 n and h
~ v 2 n lengths are slightly under-modelled (L = 15, N = 20). We recall here
L L
that the white Gaussian signals input will be used 1000 times to test the SAEC system of Fig. 1 and also to allow com-
parison between the proposed analysis with the practice one. Nevertheless, the simulation results follow the Monte Carlo
principle and are mean averaged at the end of the ltering process. The obtained results for the two SAEC adaptive lters
~ v 1 n and h
h ~ v 2 n, are reported on Figs. 8 and 9, respectively. On these two gures, we have plotted the Monte Carlo mean
L L
behaviour of the sixth and eighth coefcient errors obtained from simulations and from the theoretical expressions of rela-
tions (33) and (34) ,respectively. From these two gures, we have noted a good agreement between theory and simulation
values.

Fig. 6. Comparison of theoretical and simulation curves of the mean behaviour of the sixth and eighth coefcient errors of the under-modelled length LMS
adaptive h~ v 1 n algorithm for correlated Gaussian input data, and N = 20; L = 10; input SNR = 40dB. Results obtained through 1000 independent runs of a
L
Monte-Carlo simulation.
1590 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594

Fig. 7. Comparison of theoretical and simulation curves of the mean behaviour of the sixth and eighth coefcient errors of the under-modelled length LMS
adaptive h~ v 2 n algorithm for correlated Gaussian input data, and N = 20; L = 10 ; input SNR = 40dB. Results obtained through 1000 independent runs of a
L
Monte-Carlo simulation.

Fig. 8. Comparison of theoretical and simulation curves of the mean behaviour of the sixth and eighth coefcient errors of the under-modelled length LMS
adaptive h~ v 1 n algorithm for white Gaussian input data, and N = 20; L = 15; input SNR = 40dB. Results obtained through 1000 independent runs of a Monte-
L
Carlo simulation.
 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594 1591

Fig. 9. Comparison of theoretical and simulation curves of the mean behaviour of the sixth and eighth coefcient errors of the under-modelled length LMS
adaptive h~ v 1 n algorithm for white Gaussian input data, and N = 20; L = 15; input SNR = 40dB. Results obtained through 1000 independent runs of a Monte-
L
Carlo simulation.

We have also noted that the convergence speed to the truth coefcients for the two adaptive lters, in the case of white
Gaussian inputs, is much faster than the case of correlated Gaussian input data. Clearly, theoretical convergence using small
or large step sizes exhibits a perfect agreement with the simulation result and allows an accurate prediction of the algorithm
behaviour during the transient stages of adaptation.

4.3. Weakly correlated Gaussian input

This experiment can be classied between the both experiments cited above. For that reason, we have used weakly cor-
related Gaussian inputs xv1(n) and xv1(n) that are generated as follows:

xv 1 n 0:1xv 1 n  1 g1 n 52
xv 2 n 0:1xv 2 n  1 g2 n 53
These two input signals are convolved with the impulse response of an actual room truncated to 512 points (see Fig. 10),
i.e. N = 512.
~ v 1 n and h
The under-modelled length of the adaptive lters h ~ v 2 nis chosen equal to L = L = 320 (this case is considered
L L 1 2
as a strongly under-modelled conditions), and the variance of the additive noise at the input is cho-
sen 2 (n) = E(f2) =;0.000015. We have selected a large step sizes l1 = l2 = 0.001 close to the maximum values, an average
ones l1 = l2 = 0.0005 and small ones l1 = l2 = 0.0001. The large step-sizes values are close to the maximum ones that en-
sures SAEC LMS algorithms convergences. In this experiment, we have used the System Mismatch criterion expressed in
dB to compare the theory and simulations obtained results. We dened this criterion for the two adaptive lters h ~ v 1 n
L
and h ~ v 2 n as follows:
L
!
vi ~ v i nk2
khL n  h L
System Mismatch 10 log 10 vi ; i 2 f1; 2g 54
khL nk2
For the theory, the System Mismatch criterion is dened by
!
kEzi k2
System Mismatch 10 log 10 vi ; i 2 f1; 2g 55
khL nk2
1592 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594

4
x 10
4

2
Magnitude

-1

-2

-3
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Samples

Fig. 10. Impulse response of an actual room of 5000 points.

0
Theory
Simulation

-10

1=0.0001
-20
System Mismatch (dB)

-30

-40
1=0.0005

-50 1=0.001

0 0.5 1 1.5 2.5 3

Samples 4
x 10
Fig. 11. Comparison of theoretical and simulation curves of the System Mismatch behaviour of the under-modeled length LMS adaptive h~ v1 n algorithm
L
for white Gaussian input data, and N = 512; L = 320; input SNR=60dB. Results obtained through 1000 independent runs of a Monte-Carlo simulation.
 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594 1593

0
Theory
Simulation

-10
2=0.0001
System Mismatch (dB)

-20

-30

2=0.0005
-40

2=0.001
-50

0 0.5 1 1.5 2 2.5 3

Samples 4
x 10
Fig. 12. Comparison of theoretical and simulation curves of the System Mismatch behaviour of the under-modeled length LMS adaptive h~ v2 n algorithm
L
for white Gaussian input data, and N = 512; L = 320; input SNR=60dB. Results obtained through 1000 independent runs of a Monte-Carlo simulation.

We recall here that the obtained simulation results on this experiment are mean averaged and obtained through the
Monte Carlo method as explain in the two previous Sections. The obtained simulation results for the System Mismatch evo-
lution of the two adaptive lters h ~ v 1 n and h
~ v 2 n, computed by (54), are reported on Figs. 11 and 12, respectively.
L L
We have plotted along these two gures their corresponding System Mismatch of (55), obtained by the theory through
Eqs. (33) and (34), respectively. We have clearly noted that, for long adaptive lters, there is a perfect agreement between the
theory and the simulations results for a small step-sizes (in the case where l1 = l2 = 0.0001 for the two gures), whereas for
a large step size (in the other two cases of l1 and l2 values), the theoretical curve is slightly less accurate in predicting the
two adaptive lters h ~ v 1 n and h
~ v 2 n behaviours during the transient stages of adaptation. These results and the previous
L L
ones have conrmed and validated the good properties of our proposed analysis of the SAEC by decient two-channel LMS
algorithm in different situation whether the problem of under-modelization is slightly or strongly decient.

5. Conclusion

In this paper, we have derived an analytical model for predicting and analysing the stochastic behaviour of the stereo-
phonic least mean square LMS algorithm, when the adaptive lters are under-modelled. With our study, and using the
independence assumption, we have provided exact theoretical expressions and recursive equations for the ltering error
and for the mean coefcient error vectors during the transient and steady state phases. The analytical expressions that
we have proposed in this paper illustrate that the SAEC with under-modelled LMS algorithms and with highly/weakly cor-
related inputs have a different convergence behaviour that is signicantly different from that of the stereophonic exact
modelisation of the LMS adaptive lters. The obtained Monte Carlo simulation results for slightly and seriously under-mod-
elization of the SAEC applications by adaptive LMS lters have conrmed the accuracy of our performance analysis through
several performance criteria as the mean behaviour of the adaptive lters coefcients or with their system mismatch val-
ues. The obtained Monte Carlo results for these different criteria and for different lengths of the under-determined stereo-
phonic adaptive LMS algorithms are in excellent agreement with the provided theoretical predictions. Finally, we believe
that the proposed analysis has tremendous potential in solving many of the problems faced by the researchers in SAEC sys-
tems regarding lack of this interesting application. As a future work, we intend to develop a complete analysis in the MSE
sense and also in the case of presence of diffuse noise components which is inherent in practice and constantly present
with adaptive SAEC systems.
1594 M. Djendi, A. Bounif / Computers and Electrical Engineering 38 (2012) 15791594

Acknowledgments

The authors would like to thank the anonymous reviewers and the handling Editor for the useful comments that they
provided and for their overall objective recommendations which have largely improved the paper.

References

[1] Sondhi MM. An adaptive echo canceller. The Bell Sys Tech J 1967;XLVI:497510.
[2] Sondhi MM, Morgan DR, Hall JL. Stereophonic acoustic echo cancellation, an overview of the fundamental problem. IEEE Signal Process Lett
1995;2:14851.
[3] Sondhi MM, Morgan DR. Acoustic echo cancellation for stereophonic teleconferencing. In: Proceedings of IEEE ASSP workshop on applications of signal
processing to audio and acoustics; 1991. p. 1412.
[4] Benesty J, Amand F, Gilloire A, Grenier Y. Adaptive ltering algorithms for stereophonic acoustic echo cancellation. Proc IEEE ICASSP 1995:3099102.
[5] Gansler T, Benesty J. Stereophonic acoustic echo cancellation and two-channel adaptive ltering: an overview. Int J Adapt Contr Signal Process
2000;14(September):56586.
[6] Djendi M. An efcient stabilized fast Newton adaptive ltering algorithm for stereophonic acoustic echo cancellation SAEC. Comput Electr Eng
2012;38:93852.
[7] Djendi M, Guessoum A. A new fast Newton-type adaptive ltering algorithm for stereophonic acoustic echo cancellation (SAEC). Int J Adapt Contr
Signal Process 2010;35(June):43544.
[8] Benesty J, Paleologu C, Gnsler T, Ciochina S. A perspective on stereophonic acoustic echo cancellation. Springer-Verlag; 2011. ISBN 978-3-642-22573-
4.
[9] Shimauchi S, Makino S. Stereo projection echo canceller with true echo path estimation. Proc IEEE ICASSP 1996:305962.
[10] Makino S, Strauss K, Shimauchi S, Haneda Y, Nakagawa A. Subband stereo echo canceller using the projection algorithm with fast convergence to the
true echo path. Proc IEEE ICASSP 1997:299302.
[11] Benesty J, Morgan DR, Sondhi MM. A better understanding and an improved solution to the problems of stereophonic acoustic echo cancellation. IEEE
Trans Speech Audio Process 1998:15665.
[12] Benesty J, Morgan DR, Sondhi MM. A better understanding and an improved solution to the problems of stereophonic acoustic echo cancellation. Proc
IEEE ICASSP 1997:3036.
[13] Emura S, Haneda Y, Kataoka A, Makino S. Stereo echo cancellation algorithm using adaptive update on the basis of enhanced input-signal vector. Signal
Process 2006;86(7):115767.
[14] Benesty J, Morgan DR, Sondhi MM. A hybrid mono/stereo acoustic echo canceler. IEEE Trans Speech Audio Process 1998:6; 468475.
[15] Gansler T, Eneroth P. Inuence of audio coding on stereophonic acoustic echo cancellation. Proc IEEE ICASSP 1998:364952.
[16] Hirano A, Sugiyama A. A new multi-channel echo canceller with a single adaptive lter per channel. Proc IEEE ICASSP 1992:19225.
[17] Minami S. A stereophonic echo canceller using single adaptive lter. Proc IEEE ICASSP 1995:302730.
[18] Gilloire A, Turbin V. Using auditory properties to improve the behavior of stereophonic acoustic echo cancellers. Proc IEEE ICASSP 1998:36814.
[19] Romoli L, Cecchi S, Peretti P, Piazza F. A mixed decorrelation approach for stereo acoustic echo cancellation based on the estimation of the fundamental
frequency. IEEE Trans Audio Speech Lang Process 2012;20(2):6908.
[20] Romoli L, Cecchi S, Palestini L, Peretti P, Piazza F. A novel approach to channel decorrelation for stereo acoustic echo cancellation based on missing
fundamental theory. In: Proceedings IEEE ICASSP, Dallas, TX, USA; 2010. p. 32932.
[21] Romoli L, Cecchi S, Peretti P, Piazza F. Improved approach to stereophonic channel decorrelation based on missing fundamental theory. In: Proceedings
workshop on acoustic echo and noise control, Tel Aviv, Israel; 2010. p. 14.
[22] Cho N, Cho J, Lee J, Kim Y. Stereophonic acoustic echo cancellation using spatial deceorrelation. In: 2011 IEEE workshop on applications of signal
processing to audio and acoustics, New Paltz, New York, USA; October 1619, 2011. p. 936.
[23] Cecchi S, Romoli L, Peretti P, Piazza F. A combined psychoacoustic approach for stereo acoustic echo cancellation. IEEE Trans Audio Speech Lang Process
2011;19(6):15309.
[24] Romoli L, Squartini S, Piazza F. A variable step-size frequency-domain adaptive ltering algorithm for stereophonic acoustic echo cancellation. In:
Proceedings 18th EUSIPCO, Aalborg, Denmark; 2010. p. 2630.
[25] Yang F, Wu M, Yang J. Stereophonic acoustic echo suppression based on Wiener lter in the short-time Fourier transform domain. IEEE Signal Process
Lett 2012;19(4):22730.
[26] Wu S, Qiu X, Wu M. Stereo acoustic echo cancellation employing frequency-domain preprocessing and adaptive lter. IEEE Trans Audio Speech Lang
Process ASLP 2011;19(3):61423.
[27] Benesty J, Duhamel P, Grenier Y. Multi-channel adaptive ltering applied to multi-channel acoustic echo cancellation. Proc EUSIPCO 1996:14058.
[28] Mayyas K. Performance analysis of the decient length LMS adaptive algorithm. IEEE Trans Signal Process 2005;53(8):272734.
[29] Yousef NR, Sayed AH. A unied approach to the steady-state and tracking analyzes of adaptive lters. IEEE Trans Signal Process 2001;49(2):31424.
[30] Gu Y, Tang K, Cui H, Du W. Convergence analysis of a decient-length LMS lter and optimal-length sequence to model exponential decay impulse
response. IEEE Signal Process Lett 2003;10(1):47.
[31] Homer J, Bitmead RR, Mareels I. Quantifying the effects of dimension on the convergence rate of the LMS adaptive FIR estimator. IEEE Trans Signal
Process 1998;46(10):26115.
[32] Woo TK. Fast hierarchical least mean square algorithm. IEEE Signal Process Lett 2001;8(11):28991.
[33] Stoica P, Agrawal M, Ahgren P. On the hierarchical least-squares algorithm. IEEE Commun Lett 2002;6(4):1535.
[34] Macleod MD. Performance of the hierarchical LMS algorithm. IEEE Signal Process Lett 2002;9(12):4367.
[35] Nascimento VH. Analysis of the hierarchical LMS algorithm. IEEE Signal Process Lett 2003;10(3):7881.

Mohamed Djendi received his rst Ph.D. degree in Electronics-Signal and Telecommunications from the ENSP School of Algiers, Algeria, in 2006. He
received his second Ph.D degree in Signal Processing and Telecommunications from the University of Rennes, France, in 2010. From 2000 to present, he has
been a full Professor and researcher at Blida University and in LATSI research Laboratory. Currently, (September 2011October 2012), he holds a Post-
doctoral position at University of Rennes,IRISA/ENSSAT, engaging in research on Signal Processing and Communication Systems.

Aouda Bounif received her Engineer state and Master degrees in Signal Processing and Telecommunications from the University of Blida, Algeria in 1988
and 2007 respectively. She is now working as a permanent teacher at the College school of Chelf, Algeria and is an assistant researcher at the research
Laboratory (LATSI), Blida University, Algeria, where she is working on her Ph.D degree since September 2009. Her researches are on stereophonic acoustic
echo cancellation, speech enhancement, speech coding and ICA.