See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224338405

A Variable Step-Size Affine Projection Algorithm Designed for Acoustic Echo

Cancellation

Article in IEEE Transactions on Audio Speech and Language Processing · December 2008
DOI: 10.1109/TASL.2008.2002980 · Source: IEEE Xplore

CITATIONS READS
155 769

3 authors:

Constantin Paleologu Jacob Benesty

Polytechnic University of Bucharest Institut National de la Recherche Scientifique
220 PUBLICATIONS 2,652 CITATIONS 759 PUBLICATIONS 22,551 CITATIONS

SEE PROFILE SEE PROFILE

Silviu Ciochina
Polytechnic University of Bucharest
207 PUBLICATIONS 2,840 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Advanced adaptive algorithms View project

New convolutional coding schemes with reduced complexity operating over higher order Galois fields for channel error correction View project

All content following this page was uploaded by Silviu Ciochina on 31 May 2014.

The user has requested enhancement of the downloaded file.

1466 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 16, NO. 8, NOVEMBER 2008
A Variable Step-Size Affine Projection Algorithm
Designed for Acoustic Echo Cancellation
Constantin Paleologu, Member, IEEE, Jacob Benesty, Senior Member, IEEE, and Silviu Ciochină, Member, IEEE
Abstract—The adaptive algorithms used for acoustic echo can-
cellation (AEC) have to provide 1) high convergence rates and good
tracking capabilities, since the acoustic environments imply very
long and time-variant echo paths, and 2) low misadjustment and
robustness against background noise variations and double-talk.
In this context, the affine projection algorithm (APA) and different
versions of it are very attractive choices for AEC. However, an APA
with a constant step-size parameter has to compromise between
the performance criteria 1) and 2). Therefore, a variable step-size
APA (VSS-APA) represents a more reliable solution. In this paper,
we propose a VSS-APA derived in the context of AEC. Most of the
APAs aim to cancel (i.e., projection order) previous a posteriori
errors at every step of the algorithm. The proposed VSS-APA aims
to recover the near-end signal within the error signal of the adap-
tive filter. Consequently, it is robust against near-end signal varia-
tions (including double-talk). This algorithm does not require any a
priori information about the acoustic environment, so that it is easy
to control in practice. The simulation results indicate the good per-
formance of the proposed algorithm as compared to other mem-
bers of the APA family.
Index Terms—Acoustic echo cancellation (AEC), adaptive filters,
affine projection algorithm (APA), variable step-size affine projec-
tion algorithm (VSS-APA).
I. INTRODUCTION
COUSTIC echo cancellation (AEC) provides one of the
A best solutions to the control of acoustic echoes generated
by hands-free audio terminals [1]–[4]. In this type of applica-
tion, an adaptive filter identifies the acoustic echo path between
the terminal’s loudspeaker and microphone, i.e., the room im-
pulse response. The filter output, which provides an electronic
replica of the acoustic echo, is subtracted from the microphone
signal to cancel the echo. Nevertheless, there are several specific
and challenging problems associated with AEC applications.
First, the echo path is extremely long (on the order of hun-
dreds of milliseconds) and it may rapidly change at any time
during the connection. The excessive length of the acoustic echo
path in time is mainly due to the slow speed of sound through
air; moreover, multiple reflections of walls and objects in the
room increase this length. In addition, the impulse response of
the room is not static overtime, since it varies with the ambient
Manuscript received December 18, 2007; revised June 30, 2008. This work
was supported by the UEFISCSU under Grants PN-II no. 65/01.10.2007 and no.
331/01.10.2007. The associate editor coordinating the review of this manuscript
and approving it for publication was Prof. Sen M. Kuo.
C. Paleologu and S. Ciochină are with the Telecommunications Department,
University Politehnica of Bucharest, Bucharest 060042, Romania (e-mail:
pale@comm.pub.ro; silviu@comm.pub.ro).
J. Benesty is with the Universite du Quebec, INRS-EMT, Montreal, QC H5A
1K6, Canada (e-mail: benesty@emt.inrs.ca).
Digital Object Identifier 10.1109/TASL.2008.2002980
1558-7916/$25.00 © 2008 IEEE
Authorized licensed use limited to: Inst Natl de la Recherche Scientific EMT. Downloaded on October 21, 2008 at 13:45 from IEEE Xplore. Restrictions apply.
temperature, pressure, and humidity; also, movement of objects
and human bodies can rapidly modify the acoustic impulse re-
sponse. As a consequence of these aspects related to the acoustic
echo path characteristics, the adaptive filter works most likely
in an under-modeling situation, i.e., its length is smaller than
the length of the acoustic impulse response. Hence, the residual
echo caused by the part of the system that can not be mod-
eled acts like an additional noise and disturbs the overall per-
formance. Second, the background noise that corrupts the mi-
crophone signal can be strong and highly nonstationary.
Besides these specific problems associated with the acoustic
environment, there are some classical issues that have to be ad-
dressed in the general framework of echo cancellation. The first
one concerns the nonstationary character of the speech signal,
since it is well known that the performance of an adaptive filter
depends on the properties of the input signal. In addition, a
speech signal is highly correlated. Therefore, this type of signal
represents a challenge for any adaptive filter. Another major as-
pect that has to be considered in echo cancellation concerns
the behavior during double-talk, i.e., the talkers on both sides
speak simultaneously. In this case, besides the echo plus back-
ground noise, the microphone of the hands-free terminal cap-
tures a speech signal that acts like a large level of uncorrelated
disturbance to the adaptive filter, and it may cause its divergence.
For this reason, the echo canceller is usually equipped with a
double-talk detector (DTD) [2], in order to control the behavior
of the adaptive filter during these periods.
Each of the previously addressed problems implies some
special requirements for the adaptive algorithms used for AEC.
Summarizing, the “ideal” algorithm should have a high con-
vergence rate and good tracking capabilities (in order to deal
with the high length and time-varying nature of the acoustic
impulse response) but achieving low misadjustment. These
issues should be obtained despite the nonstationary character of
the input signal (i.e., speech). Also, the algorithm should be ro-
bust against the microphone signal variations (i.e., background
noise variations and double-talk) and in an under-modeling
case. Finally, its computational complexity should be moderate,
providing both efficient and low-cost real-time implementation.
Even if the adaptive filters literature contains a lot of very
interesting and useful algorithms [5]–[7], there is no adaptive
algorithm that satisfies all the previous requirements.
In this context, the affine projection algorithm (APA) (orig-
inally proposed in [8]) and some of its versions, e.g., [9]–[11],
were found to be very attractive choices for AEC applications.
The main advantage of this family of algorithms over the well-
known normalized least-mean-square (NLMS) algorithm con-
sists of a superior convergence rate, especially for speech signals.
PALEOLOGU et al.: VARIABLE STEP-SIZE AFFINE PROJECTION ALGORITHM 1467

Similar to the NLMS algorithm, the performance of the clas-

sical APA (in terms of convergence rate, misadjustment, and
stability) is governed by the step-size parameter. This param-
eter has to be chosen based on a compromise between fast con-
vergence rate and good tracking capabilities on the one hand,
and low misadjustment on the other hand. In order to meet this
conflicting requirement, a number of variable step-size NLMS
(VSS-NLMS) algorithms and variable step-size APAs (VSS-
APAs) were developed [12]–[16] (and reference therein). Nev-
ertheless, most of the existing VSS-APAs require the tuning of
some parameters which are not a priori available or have to be
estimated (e.g., background noise power). For real-world AEC Fig. 1. AEC configuration.
applications, it is highly desirable to use nonparametric algo-
rithms, in the sense that no information about the acoustic envi-
ronment is required. in AEC, where more complex DTDs are required. From the
Even if VSS-APAs could provide a reliable solution for AEC, complexity point of view, this last aspect represents also an
taking into account the previously discussed performance cri- advantage for a real-world AEC application.
teria, there is still a major issue that has to be addressed, i.e., The rest of the paper is organized as follows. Section II starts
the behavior during double-talk. As it was mentioned in a pre- with a presentation of the AEC configuration, together with the
vious paragraph, the overall performance of the adaptive filter notation; then, the classical APA is introduced, followed by the
could be seriously affected in this case, up to divergence. Ac- framework and derivation of the proposed VSS-APA. The sim-
cordingly, the standard procedure is to use a DTD in order to ulation results are presented in Section III, comparing the pro-
slow down or completely halt the adaptation process during posed algorithm with the classical APA and different other al-
double-talk periods. A lot of very interesting DTD algorithms gorithms from the same family; several scenarios are consid-
have been proposed [17]–[23]. Maybe the simplest one is the ered, e.g., echo path changes, background noise variations, and
well-known Geigel DTD [17], which provides an efficient and double-talk situation. Finally, in Section IV the main results of
low-complexity solution especially for network echo cancella- this work are discussed and the conclusions are drawn.
tion. Other more complex algorithms, e.g., based on coherence II. VSS-APA FOR AEC
and cross-correlation methods [18], [19] have proven to give
A general AEC configuration is depicted in Fig. 1. The goal
better results in AEC context. Nevertheless, there is some in-
of this scheme is to identify an unknown system (i.e., acoustic
herent delay in the decision of a DTD; during this small period,
echo path) using an adaptive filter. Both systems have finite-
a few undetected large amplitude samples can perturb the echo
impulse responses, defined by the real-valued vectors
path estimate considerably. Consequently, it is highly desirable
and ,
to improve the robustness of the adaptive algorithm in order to
where superscript denotes transposition and is the time
handle a certain amount of double-talk without diverging. This
index; is the length of the echo path, while is the length of
is the motivation behind the development of the so-called robust
the adaptive filter. The signal is the far-end speech which
algorithms. A solution of this kind, based on the theory of ro-
goes through the acoustic impulse response , resulting the
bust statistics [24], was proposed in [25] and applied to the APA.
echo signal, . This signal is picked up by the microphone
Other approaches try to minimize or even to annihilate the role
together with the near-end signal , resulting the micro-
of the DTD, e.g., using a postfilter to suppress the residual echo
phone signal . The near-end signal can contain both the
that remains after cancellation [26], or using an adaptive cross-
background noise, , and the near-end speech, . The
spectral technique instead of an adaptive algorithm [27], [28].
output of the adaptive filter provides a replica of the echo,
In this paper, we design a VSS-APA for AEC. The derivation
which will be subtracted from the microphone signal. The
of the variable step-size takes into account (in an explicit
DTD block controls the algorithm behavior during double-talk;
manner) the fact that the microphone signal of the hands-free
nevertheless, the proposed algorithm will be derived without
terminal contains the background noise or/and a speech se-
involving the DTD decision.
quence; these signals should be recovered in the error signal
The following relations define the classical APA [8]:
of the adaptive filter. In addition, the under-modeling case
is also considered. The resulting formula depends only on (1)
signals that are available within the AEC application, e.g., the
(2)
output signal of the adaptive filter and the microphone signal
(both in terms of power estimates). Thus, there is no need where is the desired
for a priori information about the acoustic environment, so signal vector of length , with denoting the projection order.
that the algorithm is nonparametric from this point of view. The matrix is the
Moreover, the proposed VSS-APA is robust to near-end signal input signal matrix, where
variations (i.e., background noise variations or double-talk). Its (with ) are the input
behavior during double-talk can be further improved by using signal vectors. The constant denotes the step-size parameter
a very simple Geigel DTD, which is not a very common choice of the algorithm.

Authorized licensed use limited to: Inst Natl de la Recherche Scientific EMT. Downloaded on October 21, 2008 at 13:45 from IEEE Xplore. Restrictions apply.
1468 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 16, NO. 8, NOVEMBER 2008

Let us rewrite (2) in a different form By solving the quadratic (9), two solutions are obtained, i.e.,

(3)
(10)
where
Following the analysis from [29], which states that a value of the
(4)
step-size between 0 and 1 is preferable over the one between
is a diagonal matrix. It is obvious that (2) is obtained when 1 and 2 (even if both solutions are stable, but the former has
. less steady-state mean square error with the same convergence
Using the adaptive filter coefficients at time , the a posteriori speed), it is reasonable to choose
error vector can be defined as
(11)
(5)

It can be noticed that the vector from (1) plays the role From a practical point of view, (11) has to be evaluated in terms
of the a priori error vector. Replacing (3) in (5) and taking (1) of power estimates as
into account, it results that
(12)
(6)

where denotes a identity matrix. In consistence with The variable in the denominator can be computed in a recursive
the basic idea of the APA, it can be imposed to cancel a poste- manner, i.e.,
riori errors, i.e., , where denotes a column
vector with all its elements equal to zeros. Assuming that (13)
, it results from (6) that . This corre- where is a weighting factor chosen as , with
sponds to the classical APA update (2), with the step-size . ; the initial value is .
In the absence of the near-end signal, i.e., , the scheme The estimation of is not straightforward in real-
from Fig. 1 is reduced to an ideal “system identification” con- world applications like AEC. In this case, is the near-end
figuration. In this case, the value of the step-size makes signal (i.e., background noise or/and near-end speech) which is
sense, because it leads to the best performance [8]. combined together with the acoustic echo, resulting the micro-
Nevertheless, it can be noticed that the AEC scheme from phone signal (the only signal that is practically available). Sev-
Fig. 1 can be interpreted as a combination between two classes eral scenarios could be considered, as follows.
of adaptive system configurations (according to adaptive filter 1) Single-Talk Scenario: In the single-talk case, the near-end
theory [5]–[7]). First, it represents a “system identification” con- signal consists only of the background noise, . Its power
figuration, because the goal is to identify an unknown system could be estimated during silences (and it can be assumed con-
(i.e., the acoustic echo path) with its output corrupted by an ap- stant), so that (12) becomes
parently “undesired” signal (i.e., the near-end signal). However,
it also can be viewed as an “interference cancelling” configu- (14)
ration, aiming to recover an “useful” signal (i.e., the near-end
signal) corrupted by an undesired perturbation (i.e., the acoustic
For a value of the projection order , the nonparametric
echo); consequently, the “useful” signal should be recovered in
VSS-NLMS (NPVSS-NLMS) algorithm proposed in [14] is ob-
the error signal of the adaptive filter. Therefore, since the exis-
tained. For , a VSS-APA can be derived, by computing
tence of the near-end signal cannot be omitted in AEC, a more
(14) for , then using a step-size matrix like
reasonable condition is , where the column vector
in (4), and updating the filter coefficients according to (3). Nev-
represents the
ertheless, the background noise can be time-variant, so that the
near-end signal vector of length . Taking (6) into account, it
power of the background noise should be periodically estimated.
results that
Moreover, when the background noise changes between two
(7) consecutive estimations or during the near-end speech, its new
power estimate will not be available immediately; consequently,
where the variables and denote the th until the next estimation period of the background noise, the al-
elements of the vectors and , with . gorithm behavior will be disturbed.
The goal is to find an expression for the step-size parameter 2) Double-Talk Scenario: In the double-talk case, the near-
such that end signal consists of both the background noise and
the near-end speech ; so that . It is
(8) very difficult to obtain an accurate estimate for the power of
where denotes mathematical expectation. Squaring (7) this combined signal, taking into account especially the nonsta-
and taking the expectations results in tionary character of the speech signal. Therefore, (12) is futile
and the presence of DTD is a must, in order to control the adap-
(9) tation process during these periods.

Authorized licensed use limited to: Inst Natl de la Recherche Scientific EMT. Downloaded on October 21, 2008 at 13:45 from IEEE Xplore. Restrictions apply.
PALEOLOGU et al.: VARIABLE STEP-SIZE AFFINE PROJECTION ALGORITHM 1469

3) Under-Modeling Scenario: In both previous cases, the The first term from the right-hand side of (21) represents the part
adaptive filter can work in an under-modeling situation of the acoustic echo that can be modeled by the adaptive filter.
, so that an under-modeling noise (i.e., the residual echo It can be written as
caused by the part of the system that cannot be modeled) ap-
pears. It can be interpreted as an additional noise that corrupts (22)
the near-end signal. Since it is unavailable in practice, the power
where the vector contains the first
of the under-modeling noise cannot be estimated in a direct
coefficients of the echo path vector . The second term from
manner, and consequently, its contribution to the near-end signal
the right-hand side of (21) is the under-modeling noise. This
power cannot be evaluated.
residual echo (which cannot be modeled by the adaptive filter)
In order to adapt (12) for a real-world AEC application, let
can be expressed as
us consider the previous cases in a more unified framework. To
begin, let us assume that , the microphone signal at time (23)
index can be expressed as
where
(15) and (i.e., the last coef-
ficients of the echo path vector ). The term from (23) acts like
where . Squaring (15) and taking the expec-
an additional noise for the adaptive process, so that (7) should
tation of both sides (assuming that and are uncorre-
be rewritten as
lated), it results that , so
that

(16) (24)

Supposing that the adaptive filter has converged to a certain de- with . Following the same procedure, (11)
gree, it can be considered that becomes

(17) (25)
where . Consequently
since the near-end signal and the under-modeling noise can be
(18) considered uncorrelated.
Unfortunately, expression (25) is useless in a real-world AEC
or in terms of power estimates [using a procedure similar to (13)] application since it depends on some sequences that are unavail-
able, i.e., the near-end signal and the under-modeling noise. In
(19)
order to solve this issue, the desired signal can be rewritten as
For case 1), when only the background noise is present, i.e., (26)
, an estimate of its power is obtained using the
right-hand term in (19). This expression holds even if the level Next, let us assume that and are uncorrelated. This
of the background noise changes, so that there is no need for holds for a white signal. When the input signal is speech, it is
the estimation of this parameter during silences. For the case 2), difficult to analytically state this assumption. Nevertheless, we
when the near-end speech is present (assuming that it is uncor- can extend it based on the fact that in AEC scenario,
related with the background noise), the near-end signal power and that for usual cases the correlation function has a decreasing
estimate can be expressed as ; the last trend with the time lag. Moreover, in general the first part of the
parameter denotes the power estimate of the near-end speech. acoustic impulse response is more significant as compared
Accordingly, the right-hand term in (19) provides a power esti- to the tail . Squaring, then taking the expectations of both
mate of the near-end signal. Most importantly, this term depends sides of (26), it results in
only on the signals that are available within the AEC applica-
tion, i.e., the microphone signal and the output of the adap-
tive filter . Based on these findings, (12) can be rewritten as (27)

Also, let us assume that the adaptive filter coefficients have con-
(20) verged to a certain degree, so that

(28)
for . As compared to (12), the previous rela-
tion is more suitable in practice. Consequently
Next, let us consider the under-modeling case 3), when
. In this situation, the echo signal at time index can be de- (29)
composed as
Rewriting (25) in terms of the power estimates and taking (29)
(21) into account, the expression of the step-size parameter is iden-

Authorized licensed use limited to: Inst Natl de la Recherche Scientific EMT. Downloaded on October 21, 2008 at 13:45 from IEEE Xplore. Restrictions apply.
1470 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 16, NO. 8, NOVEMBER 2008

TABLE I
VSS-APA ALGORITHM

tical to (20). Consequently, in all the analyzed cases, the vari-

able step-size parameters can be computed in a unified manner
as in (20).
The adaptive filter coefficients should be updated using (3),
with the step-sizes computed according to (20). In practice, (3)
has to be rewritten as

(30)
where is a positive scalar known as the regularization factor.
The reason behind this regularization process is to prevent the
problems associated with the inverse of the matrix ,
which could become ill-conditioned especially when highly cor-
related inputs (e.g., speech) are involved. An insightful analysis
Fig. 2. Block diagram of the proposed VSS-APA. The blocks with gray back-
about this factor, in the framework of APA, can be found in ground contain the specific operations of the proposed algorithm.
[15]. Considering the context of a “system identification” con-
figuration, the value of the regularization factor depends on the
level of the noise that corrupts the output of the system that has deviations from the previous theoretical conditions, so that we
to be identified. A lower signal-to-noise ratio requires a higher will take the absolute values in (20). Hence, the final step-sizes
value of the regularization factor. In the AEC context, different formula is written as
types of “noise” corrupt the output of the echo path, e.g., the
background noise or/and the near-end speech; in addition, the
under-modeling noise (if it is present) increases the overall level (31)
of “noise.” Also, the value of the regularization factor depends
on the value of the projection order of the algorithm. As the
value of the projection order of the APA becomes larger, the for . Summarizing, the proposed VSS-APA is
condition number of the matrix also grows; con- listed in Table I. A block diagram of this algorithm is presented
sequently, a higher value of is required. in Fig. 2, where the blocks with gray background indicate the
Finally, some practical issues have to be addressed. First, a specific operations of the proposed algorithm. As compared to
very small positive number should be added to the denom- the classical APA, the additional computational amount of the
inator in (20) to avoid division by zero. Second, under our VSS-APA consists of multiplication operations, divi-
assumptions, we have and sions, additions, and square-root operations. Taking into
. Nevertheless, account the fact that the value of the projection order in AEC ap-
the power estimates of these parameters could lead to some plications is usually smaller than 10 (e.g., common values could

Authorized licensed use limited to: Inst Natl de la Recherche Scientific EMT. Downloaded on October 21, 2008 at 13:45 from IEEE Xplore. Restrictions apply.
PALEOLOGU et al.: VARIABLE STEP-SIZE AFFINE PROJECTION ALGORITHM
be , and 8), and the length of the adaptive filter is large
(e.g., hundred of coefficients), it can be concluded that the com-
putational complexity of the proposed VSS-APA is moderate
and comparable with the classical APA or to any fast versions
of this algorithm.
For a value of the projection order , the VSS-NLMS
algorithm proposed in [16] (in the under-modeling context)
is obtained. This algorithm was found to be robust against
near-end signal variations (especially double-talk), but has a
slower convergence rate and tracking capability, as compared
to the NLMS algorithm [30]. The main motivation behind the
development of the proposed VSS-APA is to improve these
performance criteria.
Since it is based on the assumption that the adaptive filter
coefficients have converged to a certain degree, the proposed
VSS-APA could also experience a slower initial convergence
rate and a slower tracking capability as compared to the APA,
because (17) or (28) are biased in these situations. Concerning
the initial convergence rate, we could start the proposed algo-
rithm using a regular APA in the first iterations, with
Also, in order to deal with echo path changes, the proposed al-
gorithm could be equipped with an echo path changes detector
[21], [31]. Nevertheless, in our simulations none of the previous
scenarios are considered. The experimental results will prove
that the performance degradation is not very significant (espe-
cially when the value of the projection order is increased).
It is interesting to notice that the step-size of the proposed
VSS-APA does not depend explicitly on the near-end signal,
even if it was developed taking into account its presence; con-
sequently, a robust behavior under near-end signal variations
(e.g., background noise variations and double-talk) is expected.
Moreover, since only the parameters available from the adaptive
filter are required and there is no need for a priori information
about the acoustic environment, the proposed algorithm is easy
to control in practice.
III. SIMULATION RESULTS
It would be very interesting to compare the proposed
VSS-APA with most of the VSS algorithms from the APA
family and also with most of the double-talk robust algorithms.
However, it is beyond the scope of this paper. Consequently,
we choose to limit the framework only to the APA family by
comparing the proposed solution with the regular APA and also
with two other versions of it, as follows.
For the single-talk scenarios, the variable regularized APA
(VR-APA) recently proposed in [15] is included in our compar-
isons. The motivation behind this choice is twofold. First, due
to its nature, this algorithm can be also considered a VSS-APA.
Second, the experimental results presented in [15] show that it
outperforms other VSS-APAs. The VR-APA update is
Its variable regularization factor is given by
Authorized licensed use limited to: Inst Natl de la Recherche Scientific EMT. Downloaded on October 21, 2008 at 13:45 from IEEE Xplore. Restrictions apply.
1471
where is a positive design parameter. The power of the back-
ground noise and the power of the input signal have to be
known within the algorithm, while the term is evaluated
as
(34)
where denotes the norm and is a weighting parameter
as in (13).
For the scenarios where there are near-end signal variations
(i.e., background noise variations or double-talk), the robust
proportionate APA (R-PAPA) proposed in [25] is considered
for comparison. Developed in the framework of robust statistics,
this algorithm was initially applied for network echo cancella-
tion, but this idea was found to be efficient even for less sparse
echo path like in AEC [32]. The R-PAPA is not included in the
other single-talk scenarios since it is outperformed by the reg-
ular APA in these cases [25].
The simulations were performed in an AEC context, as shown
in Fig. 1. The length of the adaptive filter is set to 512 coef-
. ficients. The measured impulse response of the acoustic echo
path is plotted in Fig. 3(a) (the sampling rate is 8 kHz); its entire
length has 1024 coefficients. This length is truncated to the first
512 coefficients [before the dotted line in Fig. 3(a)] for a first
set of experiments performed in an exact modeling case. Then,
the entire length of the acoustic impulse response is used for
a second set of experiments performed in the under-modeling
case. The far-end signal is either an AR(1) process gen-
erated by filtering a white Gaussian noise through a first-order
system , or a speech sequence [Fig. 3(b)]. For
the double-talk scenarios, the near-end speech is plotted
in Fig. 3(c). An independent white Gaussian noise signal
is added to the echo signal , with 20-dB signal-to-noise
ratio (SNR) for most of the experiments. It is assumed that the
power of the background noise is known for the VR-APA. The
weighting factor (for VR-APA and VSS-APA) is computed
using [14]. The value of the parameter in the denomi-
nator of (31) is .
• Single-talk scenario
The first set of simulations is performed in an exact mod-
eling scenario , starting with the single-talk
case. The performance is evaluated in terms of the normalized
misalignment (in dB), defined as .
First, the parameters for APA and VR-APA have to be set, in
order to fulfill some performance criteria. It is known that the
overall behavior of the APA depends on its step-size parameter
. In Figs. 4(a)–(c), the misalignment curves for the APA with
different values of the step-size are shown, as compared to the
proposed VSS-APA. The input signal is the AR (1) process for
Fig. 4(a) and (b); the speech sequence is used in Fig. 4(c). The
SNR is equal to 20 dB for Figs. 4(a) and (c); SNR dB
is used in Fig. 4(b). The value of the projection order is
and the regularization factor is for all the algorithms.
Since the requirements are for both high convergence rate and
(32) low misalignment, a compromise choice has to be made in the
case of the APA. Even if the value leads to the fastest con-
vergence mode, the “middle” value of the step-size [ in
Fig. 4(a) and (c), or in Fig. 4(b)] seems to offer a more
(33)
proper solution, taking into account the previous performance
1472 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 16, NO. 8, NOVEMBER 2008
Fig. 3. (a) Measured room acoustic impulse response. Its length is truncated to
the first 512 coefficients (before the dotted line) for the set of experiments per-
formed in an exact modeling case (L = N = 512). The entire length is used for
the set of experiments performed in the under-modeling case (L = 512; N =
1024). (b) Far-end speech signal used in the experiments. (c) Near-end speech
signal used in the experiments performed in the double-talk case.
criteria. In this case, the convergence rate is slightly reduced as
compared to the situation when , but the final misalign-
ment is significantly lower. The proposed algorithm has an ini-
tial convergence rate similar to the APA with the “middle” value
of the step-size (it should be noted that the assumption (17) is
not yet fulfilled in the first part of the adaptive process), but it
achieves a significant lower misalignment, which is close to the
one obtained by the APA with the smallest step-size [
in Fig. 4(a), in Fig. 4(b), or in Fig. 4(c)].
In the case of the VR-APA, its behavior depends on the design
parameter . In Fig. 5(a)–(c), the misalignment curves for the
VR-APA with different values of are shown, as compared to
the proposed VSS-APA. The other conditions are the same as
in Fig. 4. It can be noticed that the overall performance of the
Authorized licensed use limited to: Inst Natl de la Recherche Scientific EMT. Downloaded on October 21, 2008 at 13:45 from IEEE Xplore. Restrictions apply.
Fig. 4. Misalignment of the algorithms in the single-talk case and exact mod-
eling scenario, L = N = 512;p = 2;  = 50 . (a) APA with three different
step-sizes ( = 1;  = 0:2, and  = 0:08), and VSS-APA; input signal is an
AR(1) process, SNR = 20 dB. (b) APA with three different step-sizes ( =
1;  = 0:15, and  = 0:03), and VSS-APA; input signal is an AR(1) process,
SNR = 10 dB. (c) APA with three different step-sizes ( = 1;  = 0:2, and
 = 0:05), and VSS-APA; input signal is speech, SNR = 20 dB.
VR-APA is improved when the value of increases, but up to
a limit which depends on the character of the input signal and
on the SNR (e.g., around when speech is used as input
and SNR dB). In this case, its initial convergence rate
is similar to the VSS-APA initial convergence rate, but the pro-
posed algorithm achieves a lower final misalignment. Moreover,
there is no need to estimate the background noise power in the
VSS-APA; on the other hand, this parameter is required for the
PALEOLOGU et al.: VARIABLE STEP-SIZE AFFINE PROJECTION ALGORITHM 1473

Fig. 6. Misalignment of VSS-APA with four different projection orders (a) p =

1 (VSS-NLMS [16]), (b) p = 2, (c) p = 4, and (d) p = 8. Single-talk case
and exact modeling scenario L = N = 512; is chosen from Table II, SNR =
20 dB.

Fig. 7. Misalignments of APA with  = 0:2, VR-APA with  = 1, and VSS-

APA, for two different projection orders p = 4 and p = 8. Single-talk case and
exact modeling scenario, L N = = 512; is chosen from Table II, SNR =
20 dB.

Fig. 5. Misalignment of the algorithms in the single-talk case and exact mod-
eling scenario, L N= = 512 = 2 = 50
;p ;  (for VSS-APA). (a) VR-APA TABLE II
=01 =05 =1
with four different design parameters ( : ;  : ;  , and  ), =2 REGULARIZATION FACTORS FOR APA, R-PAPA, AND VSS-APA
= 20
and VSS-APA; input signal is an AR(1) process, SNR dB. ; (b) VR-APA
=01 =05 =1
with four different design parameters ( : ;  : ;  , and  : ), =15
= 10
and VSS-APA; input signal is an AR(1) process, SNR dB. (c) VR-APA
=01 =05 =1
with four different design parameters ( : ;  : ;  , and  ), =2
= 20
and VSS-APA; input signal is speech, SNR dB.

VR-APA. In the following experiments, in order to approach the

context of typical AEC applications, only the speech sequence
from Fig. 3(b) will be used as the far-end signal.
The effects of different projection orders for the APAs are portionally to the projection order (see Table II; these values
evaluated in Figs. 6 and 7. In Fig. 6, the performance of the will be used for all the following experiments, for both the APA
proposed VSS-APA is depicted, for , and . and VSS-APA). It can be noticed that there is a significant per-
Following the discussion about the regularization factor from formance improvement of the VSS-APA with over the
the end of Section II, the value of this parameter is chosen pro- case when (i.e., the VSS-NLMS algorithm proposed

Authorized licensed use limited to: Inst Natl de la Recherche Scientific EMT. Downloaded on October 21, 2008 at 13:45 from IEEE Xplore. Restrictions apply.
1474 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 16, NO. 8, NOVEMBER 2008

Fig. 8. Misalignments of APA with  =02

: , VR-APA with  =1
, and VSS-
Fig. 10. Misalignments of APA with  =02
: , VR-APA with  =1 , R-PAPA
APA. Single-talk case, exact modeling scenario, and echo path changes at time
21, L N= = 512 = 2;p for all the algorithms,  is chosen from Table II (for with the settings from [25], and VSS-APA. Single-talk case, exact modeling
APA and VSS-APA), SNR = 20
dB. scenario, and background noise variation at time 14, for a period of 14 seconds
(SNR decreases from 20 to 10 dB), L N = = 512 = 2
;p for all the algorithms,
 is chosen from Table II (for APA, R-PAPA, and VSS-APA).

a slower tracking reaction as compared to the other algorithms,

since the assumption (17) is strongly biased in this situation.
Nevertheless, as it can be noticed from Fig. 9, the tracking
capabilities of the VSS-APA are significantly improved for
larger values of the projection order, e.g., .
The background noise can also vary in AEC, and conse-
quently, its effects over the algorithms performance should be
considered. In the experiment presented in Fig. 10, the SNR
decreases from 20 to 10 dB after 14 s from the debut of the
adaptive process, for a period of 14 s. It is assumed that the new
background noise power estimate is not available for VR-APA.
Besides the previous algorithms, the R-PAPA is included for
comparison. Its parameters are set as in [25] (the same values
will be used for all the following experiments). For all the
Fig. 9. Misalignments of APA with  =02
: , VR-APA with  =1
, and VSS- algorithms, the projection order is . It can be noticed that
APA, for two different projection orders p =4
and p =8
. Single-talk case,
exact modeling scenario, and echo path changes at time 21, L N = = 512
; the proposed VSS-APA is very robust against the background
is chosen from Table II (for APA and VSS-APA), SNR dB. = 20 noise variation, while all the other algorithms are affected by
this change in the acoustic environment. The robustness of the
VSS-APA is slightly reduced when the value of the projection
in [16]). Concerning only the initial convergence rate, this im- order increases, as it can be seen from Fig. 11.
provement is not very significant when the projection order • Double-talk scenario
increases, as compared to the case with . In Fig. 7, the Maybe the most challenging situation in echo cancellation
proposed VSS-APA is compared with the APA using , is the double-talk case. Such a scenario is considered in the
and VR-APA with (these values of the parameters and simulations using the speech signals from Fig. 3. In Fig. 12,
will be used for all the following experiments), for and three algorithms are involved, i.e., APA, VR-APA, and pro-
. It can be noticed that the proposed algorithm still outper- posed VSS-APA, but without using any DTD; the value of the
forms its counterparts. Moreover, the differences between the projection order is . The R-PAPA was not included in
initial convergence rates are insignificant. this experiment; due to its nature, this algorithm is equipped
Another possible scenario in AEC is the change of the with a DTD. It can be noticed that the proposed algorithm out-
acoustic echo path. The results of such an experiment are performs by far the APA and the VR-APA. Referring to (31),
depicted in Figs. 8 and 9, where the acoustic impulse response the quantity under the square-root provides an estimate of the
was shifted to the right by 12 samples after 21 s from the debut near-end signal power (plus an estimate of the under-modeling
of the adaptive process. In Fig. 8, the projection order for all the noise power, when it is present). Nevertheless, due to the spe-
algorithms is . As expected, the proposed algorithm has cific nature of the speech signal (e.g., nonstationary character)

Authorized licensed use limited to: Inst Natl de la Recherche Scientific EMT. Downloaded on October 21, 2008 at 13:45 from IEEE Xplore. Restrictions apply.
PALEOLOGU et al.: VARIABLE STEP-SIZE AFFINE PROJECTION ALGORITHM 1475

Fig. 11. Misalignment of VSS-APA with four different projection orders (a) Fig. 13. Misalignments of APA with  =02
: , R-PAPA with the settings from
p =1 (VSS-NLMS [16]), (b) p =2
, (c) p =4
, and (d) p =8 [25], and VSS-APA. Double-talk case from Fig. 3, with Geigel DTD, exact mod-
= = 512 = 2
. Single-talk
eling scenario, L N ;p for all the algorithms,  is chosen from
= 20
case, exact modeling scenario, and background noise variation at time 14, for a
period of 14 s (SNR decreases from 20 to 10 dB), L N = = 512
;  is chosen Table II (for APA and VSS-APA), SNR dB.
from Table II.

Fig. 12. Misalignments of APA with  =02

: , VR-APA with  =1
, and VSS- Fig. 14. Misalignment of VSS-APA with four different projection orders (a)
p =1 (VSS-NLMS [16]), (b) p =2
, (c) p =4
, and (d) p =8
. Double-talk
APA. Double-talk case from Fig. 3, without DTD, exact modeling scenario,
L = N = 512 = 2
;p for all the algorithms,  is chosen from Table II (for case from Fig. 3, with Geigel DTD, exact modeling scenario L N = = 512;
APA and VSS-APA), SNR = 20
dB. is chosen from Table II, SNR = 20
dB.

the accuracy of the near-end speech power estimate is problem-
atic, especially for long double-talk periods. Consequently, a
simple DTD can be involved in practice, in order to enhance
the performance of the proposed algorithm during double-talk
periods. In Fig. 13, the previous experiment is repeated using a
Geigel DTD.
Its settings are chosen assuming a 6-dB attenuation, i.e., the
threshold is equal to 0.5 and the hangover time is set to 240
samples [25]. The algorithms chosen for comparison are APA,
R-PAPA, and VSS-APA. The VR-APA was removed from the
list because it performs similarly to the APA in double-talk situ-
ations. Moreover, due to its variable regularized nature, there is
no explicit way to “annihilate” the overall step-size parameter
(as in APA where can be set to zero, or in VSS-APA where
from (31) can be set to zero); in this case, the solution is not
to update the filter coefficients during the detected double-talk
periods, for the hangover time. It can be noticed from Fig. 13
that the performance of the VSS-APA is improved as compared
to the previous case, and it outperforms the R-PAPA. The reg-
ular APA cannot be “saved” by this procedure (it requires a more
complex DTD, e.g., [19]). In Fig. 14, the same experiment is
repeated for the VSS-APA using different values of the projec-
tion order. The robustness of the VSS-APA is slightly influenced
when the value of the projection order increases.
Several informal subjective tests were conducted in order to
evaluate the impact on the quality of the near-end speech, in the

Authorized licensed use limited to: Inst Natl de la Recherche Scientific EMT. Downloaded on October 21, 2008 at 13:45 from IEEE Xplore. Restrictions apply.
1476 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 16, NO. 8, NOVEMBER 2008

Fig. 15. Misalignments of APA with  =02

: , VR-APA with  =1
, and VSS-
Fig. 17. Misalignments of APA with  = 02
: , VR-APA with  =1
, and
APA. Single-talk case, under-modeling scenario, L ;N = 512 = 1024 =
;p
VSS-APA. Single-talk case, under-modeling modeling scenario, and echo path
changes at time 21, L = 512
;N ;p= 1024 = 8
for all the algorithms,  is
2 for all the algorithms,  is chosen from Table II (for APA and VSS-APA), = 20
SNR = 20 dB.
chosen from Table II (for APA and VSS-APA), SNR dB.

single-talk case, using a projection order ; the results are

presented in Fig. 15.
It can be noticed that the initial convergence rate is almost
the same for all the algorithms, but the VR-APA achieves the
lowest final misalignment. In terms of final misalignment, the
VSS-APA slightly outperforms the APA. Due to its nature, the
variable regularization factor of VR-APA reaches the maximum
threshold from (33), which depends on the design parameter .
In order to enhance the misalignment of APA and VSS-APA in
this situation, a larger regularization factor is required. Sim-
ulations that are not shown here indicate that by doubling the
value of from Table II, the VSS-APA and VR-APA perform
in a similar manner.
In Fig. 16, the previous experiment is repeated with an abrupt
change of the acoustic impulse response (similar to Fig. 8). The
tracking capability of the VSS-APA is slightly slower as com-
pared to the other two algorithms, due to the bias of condition
(28). Nevertheless, the tracking reaction of the proposed algo-
rithm is improved as the value of the projection order increases
Fig. 16. Misalignments of APA with  = 02
: , VR-APA with  , and =1 (e.g., in Fig. 17). Regarding the first part of the adaptive
VSS-APA. Single-talk case, under-modeling modeling scenario, and echo path
changes at time 21, L = 512
;N ;p= 1024 = 2
for all the algorithms,  is process for VR-APA (before the echo path changes in Fig. 17),
chosen from Table II (for APA and VSS-APA), SNR dB. = 20 it can be noticed that the previous value of the design param-
eter (i.e., ) could not be proper in this case. Simulations
that are not shown here indicate that by reducing the value of
case of the proposed algorithm. The results indicate only a mild this parameter, e.g., [i.e., by increasing the maximum
degradation in terms of the mean opinion score (MOS). threshold in (33)], the misalignment of VR-APA is enhanced in
• Under-modeling scenario this case. Consequently, the performance of the VR-APA is in-
The second set of simulations is performed in an under-mod- fluenced by the choice of the design parameter in a similar
eling scenario, using the entire acoustic impulse response from manner as the influence of the regularization factor over the
Fig. 3(a), while the length of the adaptive filter remains the performance of VSS-APA.
same . In this case, the expression A variation of the background noise (i.e., the SNR decreases
of the normalized misalignment is evaluated by padding the from 20 to 10 dB as in Fig. 10) is considered in Fig. 18. The
vector of the adaptive filter coefficients with zeros, i.e., behavior of APA, VR-APA, R-PAPA, and VSS-APA is evalu-
). ated in this case, for a value of the projection order . It
Several of the previous scenarios are considered. First, the can be noticed that the proposed algorithm outperforms all the
APA, VR-APA, and the proposed VSS-APA are compared in a other algorithms.

Authorized licensed use limited to: Inst Natl de la Recherche Scientific EMT. Downloaded on October 21, 2008 at 13:45 from IEEE Xplore. Restrictions apply.
PALEOLOGU et al.: VARIABLE STEP-SIZE AFFINE PROJECTION ALGORITHM 1477

Fig. 18. Misalignments of APA with  =02: , VR-APA with  =1

, R-PAPA Fig. 20. Misalignments of APA with  =02
: , R-PAPA with the settings from
with the settings from [25], and VSS-APA. Single-talk case and background [25], and VSS-APA. Double-talk case from Fig. 3, with Geigel DTD, under-
noise variation at time 14, for a period of 14 s (SNR decreases from 20 to 10 dB), modeling scenario, L = 512
;N ;p= 1024 = 2
for all the algorithms,  is
under-modeling scenario, L = 512 ;N = 1024 = 2
;p for all the algorithms, chosen from Table II (for APA and VSS-APA), SNR dB. = 20
 is chosen from Table II (for APA, R-PAPA, and VSS-APA).

algorithm in order to take into account the existence and the

nonstationarity of the near-end signal. Moreover, the case when
the under-modeling noise is present was also considered. The
variable step-size formula of the proposed algorithm resulted
in a unified manner, requiring no additional parameters from
the acoustic environment. The simulation results performed in
an AEC context sustain the theoretical findings; accordingly,
the gain is twofold. First, due to its nonparametric nature and
simplicity, the proposed VSS-APA is very suitable in practice.
Second, as compared to other APAs, it was found to be more
robust to near-end signal variations like the increase of the
background noise or double-talk. Concerning the last scenario,
the VSS-APA can be combined with a simple Geigel DTD in
order to enhance its performance. This is also a low complexity
practical solution, taking into account that in AEC applications,
more complex DTDs or robustness techniques are involved.

Fig. 19. Misalignments of APA with  =02
: , VR-APA with 
, and VSS-
APA. Double-talk case from Fig. 3, without DTD, under-modeling scenario,
L = 512 ;N = 1024 = 2
;p for all the algorithms,  is chosen from Table II
(for APA and VSS-APA), SNR = 20
dB.
Finally, the double-talk scenarios from Figs. 12 and 13 are
repeated in the under-modeling case. The results presented in
Figs. 19 and 20 indicate the good performance of the proposed
VSS-APA in these situations. Even without DTD, the proposed
algorithm is more robust as compared with the other algorithms
using DTD. This is due to the fact that its nonparametric step-
size is designed to recover the near-end signal and the under-
modeling noise in the error signal of the adaptive filter.
IV. CONCLUSION
A VSS-APA suitable for AEC applications has been derived
in this paper. The basic idea of the classical APA, i.e., to
cancel a posteriori errors, was modified within the proposed
ACKNOWLEDGMENT
=1
The authors would like to thank the Associate Editor and the
reviewers for the valuable comments and suggestions.
REFERENCES
[1] Acoustic Signal Processing for Telecommunication, S. L. Gay and J.
Benesty, Eds.. Boston, MA: Kluwer, 2000.
[2] J. Benesty, T. Gaensler, D. R. Morgan, M. M. Sondhi, and S. L. Gay,
Advances in Network and Acoustic Echo Cancellation. Berlin, Ger-
many: Springer-Verlag, 2001.
[3] Adaptive Signal Processing—Applications to Real-World Problems,
J. Benesty and Y. Huang, Eds. Berlin, Germany: Springer-Verlag,
2003.
[4] C. Breining, P. Dreiseitel, E. Haensler, A. Mader, B. Nitsch, H. Puder,
T. Schertler, G. Schmidt, and J. Tilp, “Acoustic echo control—An
application of very-high-order adaptive filters,” IEEE Signal Process.
Mag., vol. 16, no. 4, pp. 42–69, Jul. 1999.
[5] B. Widrow and S. D. Stearns, Adaptive Signal Processing. Engle-
wood Cliffs, NJ: Prentice-Hall, 1985.
[6] S. Haykin, Adaptive Filter Theory, 4th ed. Upper Saddle River, NJ:
Prentice-Hall, 2002.
[7] A. H. Sayed, Fundamentals of Adaptive Filtering. New York: Wiley,
2003.

Authorized licensed use limited to: Inst Natl de la Recherche Scientific EMT. Downloaded on October 21, 2008 at 13:45 from IEEE Xplore. Restrictions apply.
1478 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 16, NO. 8, NOVEMBER 2008
[8] K. Ozeki and T. Umeda, “An adaptive filtering algorithm using an or-
thogonal projection to an affine subspace and its properties,” Electron.
Commun. Jpn., vol. 67-A, no. 5, pp. 19–27, May 1984.
[9] S. L. Gay and S. Tavathia, “The fast affine projection algorithm,” in
Proc. 1995 IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP),
Detroit, MI, May 1995, vol. 5, pp. 3023–3026.
[10] M. Tanaka, Y. Kaneda, S. Makino, and J. Kojima, “A fast projec-
tion algorithm for adaptive filtering,” IEICE Trans. Fundamentals, vol.
E78-A, no. 10, pp. 1355–1361, Oct. 1995.
[11] F. Albu and C. Kotropoulos, “Modified gauss-seidel affine projection
algorithm for acoustic echo cancellation,” in Proc. 2005 IEEE Int.
Conf. Acoust., Speech, Signal Process. (ICASSP), Philadelphia, PA,
Apr. 2005, vol. 3, pp. 121–124.
[12] A. Mader, H. Puder, and G. U. Schmidt, “Step-size control for acoustic
echo cancellation filters—An overview,” Signal Process., vol. 80, pp.
1697–1719, Sep. 2000.
[13] H.-C. Shin, A. H. Sayed, and W.-J. Song, “Variable step-size NLMS
and affine projection algorithms,” IEEE Signal Process. Lett., vol. 11,
no. 2, pp. 132–135, Feb. 2004.
[14] J. Benesty, H. Rey, L. Rey Vega, and S. Tressens, “A nonparametric
VSS NLMS algorithm,” IEEE Signal Process. Lett., vol. 13, no. 10,
pp. 581–584, Oct. 2006.
[15] H. Rey, L. Rey Vega, S. Tressens, and J. Benesty, “Variable explicit
regularization in affine projection algorithm: Robustness issues and
optimal choice,” IEEE Trans. Signal Process., vol. 55, no. 5, pp.
2096–2108, May 2007.
[16] C. Paleologu, S. Ciochină, and J. Benesty, “Variable step-size NLMS
algorithm for under-modeling acoustic echo cancellation,” IEEE Signal
Process. Lett., vol. 15, no. 1, pp. 5–8, Jan. 2008.
[17] D. L. Duttweiler, “A twelve-channel digital echo canceler,” IEEE
Trans. Commun., vol. 26, no. 5, pp. 647–653, May 1978.
[18] T. Gänsler, M. Hansson, C.-J. Ivanson, and G. Salomonsson, “A
double-talk detector based on coherence,” IEEE Trans. Commun., vol.
44, no. 5, pp. 1421–1427, Nov. 1996.
[19] J. Benesty, D. R. Morgan, and J. H. Cho, “A new class of doubletalk
detectors based on cross-correlation,” IEEE Trans. Speech Audio
Process., vol. 8, no. 2, pp. 168–172, Mar. 2000.
[20] K. Ghose and V. U. Reddy, “A double-talk detector for acoustic
echo cancellation applications,” Signal Process., vol. 80, no. 8, pp.
1459–1467, Aug. 2000.
[21] J. C. Jenq and S. F. Hsieh, “Decision of double-talk and time-variant
echo path for acoustic echo cancellation,” IEEE Signal Process. Lett.,
vol. 10, no. 11, pp. 317–319, Nov. 2003.
[22] J.-M. Valin, “On adjusting the learning rate in frequency domain echo
cancellation with double-talk,” IEEE Trans. Audio, Speech, Lang.
Process., vol. 15, no. 3, pp. 1030–1034, Mar. 2007.
[23] J. D. Gordy and R. A. Goubran, “Statistical analysis of doubletalk
detection for calibration and performance evaluation,” IEEE Trans.
Audio, Speech, Lang. Process., vol. 15, no. 3, pp. 1035–1043, Mar.
2007.
[24] P. J. Huber, Robust Statistics. New York: Wiley, 1981.
[25] T. Gänsler, S. L. Gay, M. M. Sondhi, and J. Benesty, “Double-talk
robust fast converging algorithms for network echo cancellation,” IEEE
Trans. Speech Audio Process., vol. 8, no. 6, pp. 656–663, Nov. 2000.
[26] G. Enzer and P. Vary, “Robust and elegant, purely statistical adaptation
of acoustic echo canceler and postfilter,” in Proc. 2003 Int. Workshop
Acoustic Echo Noise Control (IWAENC), Kyoto, Japan, Sep. 2003, pp.
43–46.
[27] T. Okuno, M. Fukushima, and M. Tohyama, “Adaptive cross-spectral
technique for acoustic echo cancellation,” IEICE Trans. Fundam., vol.
E82-A, no. 4, pp. 634–639, Apr. 1999.
[28] X. Lu and B. Champagne, “Acoustic echo cancellation over a non-
linear channel,” in Proc. 2001 Int. Workshop Acoustic Echo Noise Con-
trol (IWAENC), Darmstadt, Germany, Sep. 2001, pp. 139–142.
[29] S. G. Sankaran and A. A. L. Beex, “Convergence behavior of affine
projection algorithms,” IEEE Trans. Signal Process., vol. 48, no. 4, pp.
1086–1096, Apr. 2000.
[30] C. Paleologu, S. Ciochină, and J. Benesty, “Double-talk robust VSS-
NLMS algorithm for under-modeling acoustic echo cancellation,” in
Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), Las
Vegas, NV, Apr. 2008, pp. 245–248.
Authorized
View publicationlicensed
stats use limited to: Inst Natl de la Recherche Scientific EMT. Downloaded on October 21, 2008 at 13:45 from IEEE Xplore. Restrictions apply.
[31] J. Huo, S. Nordholm, and Z. Zang, “A method for detecting echo
path variation,” in Proc. Int. Workshop Acoustic Echo Noise Control
(IWAENC), Kyoto, Japan, Sep. 2003, pp. 71–74.
[32] H. Buchner, J. Benesty, T. Gänsler, and W. Kellermann, “An out-
lier-robust extended multidelay filter with application to acoustic echo
cancellation,” in Proc. Int. Workshop Acoustic Echo Noise Control
(IWAENC), Kyoto, Japan, Sep. 2003, pp. 19–22.
Constantin Paleologu (M’07) was born in Romania
in 1975. He received the M.S. degree in telecommu-
nications networks and the Ph.D. degree in adaptive
signal processing from University Politehnica of
Bucharest, Bucharest, Romania, in 1998 and 2003,
respectively.
Since October 1998, he has been with the Telecom-
munications Department, University Politehnica of
Bucharest, where he is currently an Associate Pro-
fessor. His research interests are in adaptive signal
processing.
Jacob Benesty (M’98–SM’04) was born in 1963. He
received the M.S. degree in microwaves from Pierre
and Marie Curie University, Paris, France, in 1987,
and the Ph.D. degree in control and signal processing
from Orsay University, Paris, in April 1991.
During the Ph.D. degree (from November 1989
to April 1991), he worked on adaptive filters and
fast algorithms at the Centre National d’Etudes des
Telecommunications (CNET), Paris. From January
1994 to July 1995, he worked at Telecom Paris Uni-
versity on multichannel adaptive filters and acoustic
echo cancellation. From October 1995 to May 2003, he was first a Consultant
and then a Member of the Technical Staff at Bell Laboratories, Murray Hill,
NJ. In May 2003, he joined the University of Quebec, INRS-EMT, Montreal,
QC, Canada, as an Associate Professor. His research interests are in signal
processing, acoustic signal processing, and multimedia communications. He
coauthored the books Microphone Array Signal Processing (Springer-Verlag,
2008), Acoustic MIMO Signal Processing (Springer-Verlag, 2006), and Ad-
vances in Network and Acoustic Echo Cancellation (Springer-Verlag, 2001).
He is the Editor-in-Chief of the reference Springer Handbook of Speech
Processing (Springer-Verlag, 2007). He is also a coeditor/coauthor of the books
Speech Enhancement (Springer-Verlag, 2005), Audio Signal Processing for
Next Generation Multimedia communication Systems (Kluwer, 2004), Adaptive
Signal Processing: Applications to Real-World Problems (Springer-Verlag,
2003), and Acoustic Signal Processing for Telecommunication (Kluwer, 2000).
Dr. Benesty received the 2001 Best Paper Award from the IEEE Signal
Processing Society. He was a member of the editorial board of the EURASIP
Journal on Applied Signal Processing and was the cochair of the 1999 Interna-
tional Workshop on Acoustic Echo and Noise Control.
Silviu Ciochină (M’94) received the M.S. degree in
electronics and telecommunications and the Ph.D.
degree in communications from the University
Politehnica of Bucharest, Bucharest, Romania, in
1978.
From 1971 to 1979, he was an Assistant, and
from 1979 to 1995 a Lecturer at the University
Politehnica of Bucharest, Faculty of Electronics,
Telecommunications, and Information Technology.
Since 1995, he has been a Professor at the same
Faculty. Since 2004, he has been the Head of the
Telecommunications Department. His main research interests are in the areas
of signal processing and wireless communications, including adaptive algo-
rithms, spectrum estimation, fast algorithms, channel estimation, multi-antenna
systems, and broadband wireless technologies.