IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO.

8, NOVEMBER 2009 1591

A Class of Sparseness-Controlled Algorithms

for Echo Cancellation
Pradeep Loganathan, Andy W. H. Khong, Member, IEEE, and Patrick A. Naylor, Senior Member, IEEE

Abstract—In the context of acoustic echo cancellation (AEC), it [5]–[7], data reusing techniques [8], [9], partial update adap-
is shown that the level of sparseness in acoustic impulse responses tive filtering techniques [10], [11] and subband adaptive filtering
can vary greatly in a mobile environment. When the response is (SAF) schemes [12]. These approaches aim to address issues in
strongly sparse, convergence of conventional approaches is poor.
Drawing on techniques originally developed for network echo
echo cancellation including the performance with colored input
cancellation (NEC), we propose a class of AEC algorithms that can signals, time-varying echo paths and computational complexity,
not only work well in both sparse and dispersive circumstances, to name but a few. In contrast to these approaches, sparse adap-
but also adapt dynamically to the level of sparseness using a new tive algorithms have been developed specifically to address the
sparseness-controlled approach. Simulation results, using white performance of adaptive filters in sparse system identification.
Gaussian noise (WGN) and speech input signals, show improved For sparse echo systems, the NLMS algorithm suffers from slow
performance over existing methods. The proposed algorithms
achieve these improvement with only a modest increase in compu-
convergence [13].
tational complexity. One of the first sparse adaptive filtering algorithms for NEC
is proportionate NLMS (PNLMS) [2] in which each filter co-
Index Terms—Acoustic echo cancellation (AEC), network echo efficient is updated with an independent step-size that is lin-
cancellation (NEC), sparse impulse responses, adaptive algo-
early proportional to the magnitude of that estimated filter co-
rithms.
efficient. It is well known that PNLMS has very fast initial con-
vergence for sparse impulse responses after which its conver-
I. INTRODUCTION gence rate reduces significantly, sometimes resulting in a slower
overall convergence than NLMS. In addition, PNLMS suffers
from slow convergence when estimating dispersive impulse re-
CHO cancellation in telephone networks comprising
E mixed packet-switched and circuit-switched components
requires the identification and compensation of echo systems
sponses [13], [14]. To address the latter problem, subsequent im-
proved versions, such as PNLMS++ [14], were proposed. The
PNLMS++ algorithm achieves improved convergence by alter-
with various levels of sparseness. The network echo response nating between NLMS and PNLMS for each sample period.
in such systems is typically of length 64–128 ms, characterized However, as shown in [15], the PNLMS++ algorithm only per-
by a bulk delay dependant on network loading, encoding, and forms best in the cases when the impulse response is sparse or
jitter buffer delays [1]. This results in an “active” region in highly dispersive.
the range of 8–12 ms duration and consequently, the impulse An improved PNLMS (IPNLMS) [15] algorithm was pro-
response is dominated by “inactive” regions where coefficient posed to exploit the “proportionate” idea by introducing a
magnitudes are close to zero, making the impulse response controlled mixture of proportionate (PNLMS) and non-propor-
sparse. The echo canceller must be robust to this sparseness tionate (NLMS) adaptation. A sparseness-controlled IPNLMS
[2]. This network echo cancellation (NEC) issue is particularly algorithm was proposed in [16] to improve the robustness
important in legacy networks comprising packet-switched and of IPNLMS to the sparseness variation in impulse responses.
circuit switched components whereas in pure packet-switched Composite PNLMS and NLMS (CPNLMS) [17] adaptation was
networks NEC is not normally required. proposed to control the switching of PNLMS++ between the
Traditionally, adaptive filters have been deployed to achieve NLMS and PNLMS algorithms. For sparse impulse responses,
NEC by estimating the network echo response using algorithms CPNLMS performs the PNLMS adaptation to update the large
such as the normalized least-mean-square (NLMS) algorithm. coefficients and subsequently switches to NLMS, which has
Several approaches have been proposed over recent years to better performance for the adaptation of the remaining small
improve the performance of the standard NLMS algorithm in taps. The -law PNLMS (MPNLMS) [18] algorithm was
various ways for NEC. These include Fourier [3] and wavelet [4] proposed to address the uneven convergence rate of PNLMS
based adaptive algorithms, variable step-size (VSS) algorithms during the estimation process. As proposed in [18], MPNLMS
uses optimal step-size control factors to achieve faster overall
Manuscript received September 09, 2008; revised May 29, 2009. First pub- convergence until the adaptive filter reaches its steady state.
lished nulldate. The associate editor coordinating the review of this manuscript With the development of hands-free mobile telephony in re-
and approving it for publication was Dr. Jingdong Chen.
P. Loganathan and P. A. Naylor are with the Department of Electrical and cent years, another type of echo, acoustic echo, seriously de-
Electronic Engineering, Imperial College London, London SW7 2AZ, U.K. grades user experience due to the coupling between the loud-
(e-mail: [email protected]; [email protected]). speaker and microphone. For this reason, effective acoustic echo
A. W. H. Khong is with the School of Electrical and Electronic En-
gineering, Nanyang Technological University, Singapore 639798 (e-mail:
cancellation (AEC) [19] is important to maintain usability and
[email protected]). to improve the perceived voice quality of a call. Although sparse
Digital Object Identifier 10.1109/TASL.2009.2025903 adaptive filtering algorithms, such as those described above,
1558-7916/$26.00 © 2009 IEEE
1592 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 8, NOVEMBER 2009

their conventional methods. As will be shown, the proposed

sparseness-controlled algorithms achieve fast convergence for
both sparse and dispersive AIRs and are effective for AEC.

II. REVIEW OF ALGORITHMS FOR ECHO CANCELLATION

Fig. 1 shows a Loudspeaker–Room–Micro-
phone system (LRMS) and an adaptive filter
deployed to cancel
acoustic echo, where is the length of the adaptive filter
assumed to be equal to the unknown room impulse response
Fig. 1. Adaptive system for acoustic echo cancellation in a Loud- and is the transposition operator. Defining the input
speaker–Room–Microphone system (LRMS). signal and
as the unknown
impulse response, the output of the LRMS is given by

(1)

where is additive noise and the error signal is given by

(2)

Several adaptive algorithms such as those described below have

been developed for either AEC or NEC.
Fig. 2. Room impulse responses obtained, at 20 kHz sampling frequency, when Many adaptive algorithms can be described by (2) and the
the distances between a loudspeaker and a microphone are (a) 75 cm and (b) 185 following set of equations:
2 2
cm, with room dimensions 3 5 3.5 m.  represents the sparseness measure
defined in Section III-A.
(3)
(4)
have originally been developed for NEC, it has been shown in
[20] that such algorithms give good convergence performance where is a step-size and is the regularization parameter. The
in the AEC system as illustrated in Fig. 1. diagonal step-size control matrix is introduced here to
The time variation of the near-end acoustic impulse response determine the step-size of each filter coefficient and is dependent
(AIR) may arise due to, for example, a change in temperature on the specific algorithm.
[21], pressure and changes in the acoustic environment. It is also
well known that the reverberation time of an AIR is proportional A. NLMS, PNLMS, and MPNLMS Algorithms
to the volume of the enclosed space and inversely proportional
The NLMS algorithm is one of the most popular for AEC due
to the absorption area [22]. For an outdoor environment, the
to its straightforward implementation and low complexity com-
reverberation time is reduced significantly due to the lack of
pared to, for example, the recursive least squares algorithm. For
reflections from any enclosure. The outdoor environment refers
NLMS, since the step-size is the same for all filter coefficients,
here to a typical urban area or a rural area with sparsely placed
with being an identity matrix.
acoustically reflecting objects. The sparseness of the AIR of an
One of the main drawbacks of the NLMS algorithm is that
outdoor environment is significantly greater than typical indoor
its convergence rate reduces significantly when the impulse re-
environments and equally, if not more, variable.
sponse is sparse, such as often occurs in NEC. The poor perfor-
Variation in the sparseness of AIRs can also occur in AEC
mance has been addressed by several sparse adaptive algorithms
within an enclosed space. Consider an example case where the
such as those described below that have been developed specifi-
distance between a loudspeaker and the user using, for ex-
cally to identify sparse impulse responses in NEC applications.
ample, a wireless microphone is varying. Fig. 2 shows two AIRs
The PNLMS [2] and MPNLMS [18] algorithms have been
obtained in the same room, at 20 kHz sampling frequency, when
proposed for sparse system identification. Diagonal elements
(a) cm and (b) cm, with room dimensions
of the step-size control matrix for the PNLMS and
3 5 3.5 m. As can be seen, the sparseness of these AIRs
MPNLMS algorithms can be expressed as
varies with the loudspeaker-microphone distance. Hence, algo-
rithms developed for mobile hands-free terminals are required
to be robust to the variations in the sparseness of the acoustic
path. (5)
In this paper, we propose a class of algorithms that are ro-
bust to the sparseness variation of AIRs. These algorithms com-
pute a sparseness measure of the estimated impulse response
at each iteration of the adaptive process and incorporate it into (6)
LOGANATHAN et al.: A CLASS OF SPARSENESS-CONTROLLED ALGORITHMS FOR ECHO CANCELLATION 1593

where is specific to the algorithm. The parameter

in (6) prevents the filter coefficients from
stalling when at initialization and , with a
typical value of 0.01, prevents the coefficients from stalling
when they are much smaller than the largest coefficient.
The PNLMS algorithm achieves a high rate of convergence
by employing step-sizes that are proportional to the magnitude
of the estimated impulse response coefficients where elements
are given by

(7) Fig. 3. Acoustic impulse responses.

Hence, PNLMS employs larger step-sizes for “active” coeffi-

cients than for “inactive” coefficients and consequently achieves shown that, although the IPNLMS algorithm has faster conver-
faster convergence than NLMS for sparse impulse responses. gence than NLMS and PNLMS regardless of the impulse re-
However, it is found that PNLMS achieves fast initial conver- sponse nature [15], we note from our simulations that it does
gence followed by a slower second phase convergence [18]. not outperform MPNLMS for highly sparse impulse responses
The MPNLMS algorithm was proposed to improve the with the above choices of .
convergence of PNLMS. It achieves this by computing the
optimal proportionate step-size during the adaptation process. III. CHARACTERIZATION OF FRAMEWORK
The MPNLMS algorithm was derived such that all coefficients FOR ROBUST CONVERGENCE
attain a converged value to within a vicinity of their optimal In this section, we quantify the degree of sparseness in AIRs.
value in the same number of iterations [18]. As a consequence, We provide an illustrative example to show how the sparseness
for MPNLMS is specified by of AIRs varies with the loudspeaker–microphone distance in an
enclosed space such as when the user is using a wireless micro-
(8)
phone for tele/video conferencing. This serves as a motivation
with and is a very small positive number chosen as a for us to develop new algorithms which are robust to the sparse-
function of the noise level [18]. It has been shown that ness variation of AIRs in the next Section. In addition, we also
is a good choice for typical echo cancellation. The positive bias demonstrate how the choice of in (6) affects the step-size of
of 1 in (8) is introduced to avoid numerical instability during the each filter coefficient for PNLMS.
initialization stage when . A. Variation of Sparseness in AIRs
It is important to note that both PNLMS and MPNLMS suffer
from slow convergence when the unknown system is dis- The degree of sparseness for an impulse response can be
persive [14], [13]. This is because when is dispersive, quantified by [16], [23]
in (6) becomes significantly large for most .
As a consequence, the denominator of in (5) is large, (10)
giving rise to a small step-size for each large coefficient. This
causes a significant degradation in convergence performance for It can be shown [16], [23] that . In the extreme
PNLMS and MPNLMS when the impulse response is dispersive but unlikely case when
such as can occur in AIRs.
(11)
B. IPNLMS Algorithm and

The IPNLMS [15] algorithm was originally developed for where and , then . On
NEC and was further developed for the identification of acoustic the other hand, when , then . In reality
room impulse responses [20]. It employs a combination of pro- and hence is time-varying and depends on factors
portionate (PNLMS) and non-proportionate (NLMS) adapta- such as temperature, pressure and reflectivity [21]. As explained
tion, with the relative significance of each controlled by a factor in Section I, the sparseness of AIRs varies with the location
such that the diagonal elements of are given as of the receiving device in an open or enclosed environment. We
show below how can also vary with the loudspeaker–mi-
crophone distance in an enclosed space.
Consider an example case where the distance between
(9) a fixed position loudspeaker and the talker using a wireless
microphone is varying. Fig. 3 shows two AIRs, generated using
where is defined as the -norm and the first and second the method of images [24], [25] with 1024 coefficients using
terms are the NLMS and the proportionate terms, respectively. It room dimensions of 8 10 3 m and 0.57 as the reflection
can be seen that IPNLMS behaves like NLMS when coefficient. The loudspeaker is fixed at 4 9.1 1.6 m in the
and PNLMS when . Use of a higher weighting for LRMS while the microphone is positioned at 4 8.2 1.6 m
NLMS adaptation, such as or , is a favor- and 4 1.4 1.6 m giving impulse responses as shown in
able choice for most AEC/NEC applications [15]. It has been Fig. 3(a) and (b) for m and m, respectively.
1594 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 8, NOVEMBER 2009

Fig. 5. Convergence of the PNLMS for different values of  using WGN input
signal. Impulse responses in Fig. 3(a) and (b) are used as sparse and dispersive
AIRs, respectively. [ = 0 3 SNR = 20
: ; dB].

Fig. 4. Sparseness measure against the distance between loudspeaker and mi-
crophone a. The impulse responses are obtained from the image model using a
2 2
fixed room dimensions of 8 10 3 m.

Fig. 4 illustrates how of such AIRs varies with . For each

loudspeaker–microphone distance , the microphone is directly
in front of the loudspeaker. As can be seen, reduces with
increasing , since for increasing , the sound field becomes
more diffuse. Since varies with , we propose to incor-
porate into PNLMS, MPNLMS, and IPNLMS in order
to improve their robustness to the sparseness of AIRs in AEC.
Since is unknown during adaptation, we employ to
estimate the sparseness of an impulse response, where at each ()
Fig. 6. Magnitude of q n for 0   0 1 against the magnitude of
l L
sample iteration ()
normalized tap coefficients h n in PNLMS.

(12) at a more uniform rate. This provides a good convergence per-

formance for PNLMS for a dispersive AIR. On the other hand, a
lower will increase the degree of proportionality hence giving
good convergence performance when the AIR is sparse. As a
B. Effect of on Step-Size Control Matrix for PNLMS consequence of this important observation, we propose to incor-
As explained in Section II-A, the parameter in (6) was orig- porate into for both PNLMS and MPNLMS as described
inally introduced to prevent freezing of the filter coefficients in the next section.
when they are much smaller than the largest coefficient. Fig. 5
shows the effect of for both sparse and dispersive AIRs on the IV. CLASS OF SPARSENESS-CONTROLLED ALGORITHMS
convergence performance of PNLMS measured using the nor-
malized misalignment defined by We propose to improve the robustness of PNLMS,
MPNLMS, and IPNLMS to varying levels of sparseness
of impulse response such as encountered in, for example,
(13)
AEC. As will be shown in the following, this is achieved by
incorporating the sparseness measure of the estimated AIRs
A zero mean white Gaussian noise (WGN) sequence is used as into the adaptation process. We will discuss these approaches
the input signal while another WGN sequence is added to conceptually and with simulation results on both WGN and
give an SNR of 20 dB. Impulse responses as shown in Fig. 3(a) speech. For an analytical prospective, the reader is referred to
and (b) are used as sparse and dispersive AIRs, and [26].
. It can be seen from this illustration that, for a sparse ,
we desire a low value of while, for a dispersive unknown A. Proposed SC-PNLMS and SC-MPNLMS Algorithms
system , we desire a high value of . This is due to the
resulting effect of how different values of affect the step-size In order to address the problem of slow convergence in
control element as illustrated in Fig. 6. It can be observed PNLMS and MPNLMS for dispersive AIR, we require the
that a higher value of will reduce the influence of the propor- step-size control elements to be robust to the sparseness
tional update term meaning that all filter coefficients are updated of the impulse response. Several choices can be employed
LOGANATHAN et al.: A CLASS OF SPARSENESS-CONTROLLED ALGORITHMS FOR ECHO CANCELLATION 1595

TABLE I
SPARSENESS-CONTROLLED ALGORITHMS

Fig. 7. Variation of  against sparseness measure  (n) of impulse response.

to obtain the desired effect of achieving a high when

is small when estimating dispersive AIRs. We consider an
example function

(14)

The variation of in PNLMS for the exponential function

is plotted in Fig. 7 for the cases where and . It can
be noted that a linear function also achieves
our desired condition. We have tested this case and found it to
perform worse than the more general form of (14), so we will
not consider it further.
It can be seen that low values of are allocated for a large
range of sparse impulse responses such as when . As
a result of small values in using (14), the proposed sparse-
ness-controlled PNLMS algorithm (SC-PNLMS) inherits the
proportionality step-size control over a large range of sparse im-
pulse response. When the impulse response is dispersive, such
as when , the proposed SC-PNLMS algorithm in-
herits the NLMS adaptation control with larger values of .
As explained in Section III-B and Fig. 6, this gives a more uni-
form step-size across . Hence, the exponential function de- step-size across , as in NLMS, when the estimated AIR is
scribed by (14) will achieve our overall desired effect of the ro- dispersive. In addition, we note that when
bustness to sparse and dispersive AIRs. and hence, to prevent division by a small number or zero,
The choice of is important. As can be seen from Fig. 7, a can be computed for in both SC-PNLMS and
larger choice of will cause the proposed SC-PNLMS to inherit SC-MPNLMS. When , we set as described
more of PNLMS properties compared to NLMS giving good in [15]. The SC-PNLMS algorithm is thus described by (2)–(7),
convergence performance when AIR is sparse. On the other (12), and (14), whereas SC-MPNLMS is described by (2)–(6),
hand, when the AIR is dispersive, must be small for good con- (8), (12) and (14) with , as summarized in Table I.
vergence performance. Hence, we show in Section VI-A that a
good compromise is given by , though the algorithm is B. SC-IPNLMS Algorithm
not very sensitive to this choice in the range of . We choose to incorporate sparseness-control into the
Incorporating in a similar manner for the MPNLMS IPNLMS algorithm (SC-IPNLMS) [16] in a different manner
algorithm, the resulting sparseness-controlled MPNLMS algo- compared to SC-PNLMS and SC-MPNLMS because, as can be
rithm (SC-MPNLMS) inherits more of the MPNLMS proper- seen from (9), two terms are employed in IPNLMS for control
ties when the estimated AIR is sparse and distributes uniform of the mixture between proportionate and NLMS updates.
1596 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 8, NOVEMBER 2009

TABLE II
COMPLEXITY OF ALGORITHMS’ COEFFICIENTS UPDATE—ADDITION
(A), MULTIPLICATION (M), DIVISION (D), LOGARITHM (LOG),
AND COMPARISON (C)

Fig. 8. Magnitude of q (n) for 0   0

l L 1 against the magnitude of
coefficients h (n) in SC-IPNLMS and different sparseness measures of eight
systems.

the computation of the sparseness measure . Given that

The proposed SC-IPNLMS improves the performance of the in (10) can be computed offline, the remaining
IPNLMS by expressing for as -norms require an additional additions and multiplica-
tions. The SC-PNLMS and SC-MPNLMS algorithms addi-
tionally require computations for (14). Alternatively, a look-up
table with values of defined in (14) can be computed for
. Segment PNLMS (SPNLMS) is proposed in
(15) [27], to approximate the -law function in MPNLMS using
line segments. Since computation is already available
As can be seen, for large when the impulse response is from IPNLMS in (9), SC-IPNLMS only requires an additional
sparse, the algorithm allocates more weight to the proportionate additions, multiplications, and 1 division. As we
term of (9). For comparatively less sparse impulse responses, shall see, the increase in the complexity is compromised by the
the algorithm aims to achieve the convergence of NLMS by ap- algorithm’s performance. Consequently, the tradeoff between
plying a higher weighting to the NLMS term. An empirically complexity and performance depend on the design choice for
chosen weighting of 0.5 in (15) is included to balance the per- a particular application.
formance between sparse and dispersive cases. In addition, nor-
VI. SIMULATION RESULTS
malization by is introduced to reduce significant coefficient
noise when the effective step-size is large for sparse AIRs with We present simulation results to evaluate the performance
high . of the proposed SC-PNLMS, SC-MPNLMS and SC-IPNLMS
Fig. 8 illustrates the step-size control elements for algorithms in the context of AEC. In addition, we show an
SC-IPNLMS in estimating different unknown AIRs. As can example case of how SC-IPNLMS can be employed in NEC.
be seen, for dispersive AIRs, SC-IPNLMS allocates a uniform Throughout our simulations, algorithms were tested using a
step-size across while, for sparse AIRs, the algorithm zero mean WGN and a male speech signal as inputs while
distributes proportionally to the magnitude of the co- another WGN sequence was added to give an SNR
efficients. As a result of this distribution, the SC-IPNLMS of 20 dB. We assumed that the length of the adaptive filter
algorithm varies the degree of NLMS and proportionate adap- is equivalent to that of the unknown system. Two
tations according to the nature of the AIRs. In contrast, in receiving room impulse responses for AEC simulations
standard IPNLMS the mixing coefficient in (9) is fixed have been used as described in Fig. 3. The sparseness measure
a priori. The SC-IPNLMS algorithm is described by (2)–(4), of these AIRs are computed using (10) giving (a)
(12) and (15), as specified in Table I. and (b) , respectively.

V. COMPUTATIONAL COMPLEXITY A. Effect of on the Performance of SC-PNLMS for AEC

The relative complexity of NLMS, PNLMS, SC-PNLMS, SC-PNLMS was tested as shown in Fig. 9 for different
IPNLMS, SC-IPNLMS, MPNLMS, and SC-MPNLMS in values in (14) to illustrate the time taken to reach 20 dB
terms of the total number of additions (A), multiplications (M), normalized misalignment using a WGN sequence as the input
division (D), logarithm (Log), and comparisons (C) per iteration signal. A step-size of was used in this experiment. We
for adaptation of filter coefficients is assessed in Table II. The see from the result that, for each case of , the SC-PNLMS
additional complexity of the proposed sparseness-controlled has a higher rate of convergence for a sparse system compared
algorithms, on top of their conventional method, arises from to a dispersive system. This is due to the initialization choice
LOGANATHAN et al.: A CLASS OF SPARSENESS-CONTROLLED ALGORITHMS FOR ECHO CANCELLATION 1597

0
Fig. 9. Time to reach 20-dB normalized misalignment level for different Fig. 10. Convergence of the SC-PNLMS for different values of  using WGN
values of  in SC-PNLMS using WGN input signal. Impulse response input signal with an echo path change at 3.5 s. Impulse response is changed from
in Fig. 3(a) and (b) used as sparse AIR and dispersive AIR, respectively. Fig. 3(a) to (b) and  = 0 3 SNR = 20
: ; dB.
[ = 0 3 SNR = 20
: ; dB].

of , where most filter coefficients are initialized

close to their optimal values. In addition, a smaller value of
is favorable for the dispersive AIR, since SC-PNLMS performs
similarly to NLMS for small values. On the contrary, a higher
value for is desirable for the sparse case. It can be noted that
SC-PNLMS behaves like NLMS for . It can also be seen
that a range of good value for is . Fig. 10 shows
the performance of SC-PNLMS with an echo path change
introduced from Fig. 3(a) to (b) at 4.5 s, for and
. We observe from this result that the convergence rate of
SC-PNLMS is high when is small for a dispersive channel.
This is because, as explained in Section IV-A, the proposed
algorithm inherits properties of the NLMS for a small value.
For a high , the SC-PNLMS algorithm inherits properties of
PNLMS giving good performance for sparse AIR before the Fig. 11. Relative convergence of NLMS, PNLMS, and SC-PNLMS using
echo path change. As can be seen, a good compromise of is WGN input signal with an echo path change at 3.5 s. Impulse response is
given by . changed from that shown from Fig. 3(a) to (b) and   = =
 = 0 3 SNR = 20
: ; dB.

B. Convergence Performance of SC-PNLMS for AEC

Fig. 11 compares the performance of NLMS, PNLMS, and dispersive AIR, the SC-PNLMS maintains its high convergence
SC-PNLMS using WGN as the input signal. The step-size rate over NLMS and PNLMS giving approximately 4-dB im-
parameter for each algorithm is chosen such that all algorithms provement in normalized misalignment compared to PNLMS.
achieve the same steady-state. This is achieved by setting Fig. 12 shows simulation results for a male speech input
. An echo path signal where we used the same parameters as in the case of
change was introduced from Fig. 3(a) to (b) while for the WGN input signal. As can be seen, the proposed SC-PNLMS
SC-PNLMS algorithm is set to 6. It can be seen from Fig. 11 algorithm achieves the highest rate of convergence, giving
that the convergence rate of SC-PNLMS is as fast as PNLMS convergence as fast as PNLMS and approximately 7-dB im-
for sparse and much better than PNLMS for dispersive, there- provement during initial convergence compared to NLMS for
fore achieving our objective of improving robustness to varying the sparse AIR. For dispersive AIR, SC-PNLMS performs
sparseness. This is because SC-PNLMS inherits the beneficial almost the same as NLMS with approximately 4-dB improve-
properties of both PNLMS and NLMS. It can be seen from ment compared to PNLMS.
the result that SC-PNLMS achieves high rate of convergence
similar to PNLMS giving approximately 5-dB improvement in C. Convergence Performance of SC-MPNLMS for AEC
normalized misalignment during initial convergence compared Fig. 13 illustrates the performance of NLMS, MPNLMS, and
to NLMS for a sparse AIR. After the echo path change, for a SC-MPNLMS using WGN as the input signal. As before, the
1598 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 8, NOVEMBER 2009

Fig. 14. Relative convergence of NLMS, MPNLMS, and SC-MPNLMS

using speech input signal with echo path changes at 58 s. Impulse response is
Fig. 12. Relative convergence of NLMS, PNLMS and SC-PNLMS using
changed from that shown in Fig. 3(a) to (b) and  =03 : ; =
speech input signal with echo path changes at 58 s. Impulse response is
changed from that shown in Fig. 3(a) to (b) and  = 03: ; =  = 0 25 SNR = 20
: ; dB.
 = 0 1 SNR = 20
: ; dB.

For dispersive AIR, the SC-MPNLMS algorithm achieves

an improvement of approximately 4 dB compared to both
NLMS and MPNLMS. It is also noted that NLMS achieves
approximately 7-dB better steady-state performance than the
MPNLMS-based approaches for this example with speech
input. This is attributed in [4] to sensitivity to eigenvalue spread
of the speech signal’s autocorrelation matrix.

D. Convergence Performance of SC-IPNLMS for AEC

For SC-IPNLMS performance comparison, we used
in order
to attain same steady-state performance. Proportionality con-
trol factors have been used for both
IPNLMS and SC-IPNLMS. It can be seen from Figs. 15 and 16
that by using both WGN and speech input signals, SC-IPNLMS
achieves approximately 10-dB improvement in normalized
misalignment during initial convergence compared to NLMS
Fig. 13. Relative convergence of NLMS, MPNLMS, and SC-MPNLMS using for the sparse AIR. For a dispersive AIR, the SC-IPNLMS
WGN input signal with an echo path change at 3.5 s. Impulse response is
changed from that shown from Fig. 3(a) to (b) and  =03: ; = achieves a 5-dB improvement compared to NLMS. For a
 = 0 25 SNR = 20
: ; dB. speech input, the improvement of SC-IPNLMS over IPNLMS
is 3 dB for both sparse and dispersive AIRs. On the other hand,
the improvement of SC-IPNLMS compared to NLMS are 10
step-sizes were adjusted to achieve the same steady-state mis- dB and 6 dB for sparse and dispersive AIRs, respectively.
alignment for all algorithms. This corresponds to
. We have also used E. Convergence Performance of the Algorithms for Various
for SC-MPNLMS. As can be seen from this result, the Airs With Different Sparseness in AEC
SC-MPNLMS algorithm attains approximately 8-dB improve- We extracted eight different impulse responses from the
ment in normalized misalignment during initial convergence set of AIRs with sparseness measure as
compared to NLMS and same initial performance followed shown in Fig. 4. The time taken to reach 20 dB normalized
by approximately 2-dB improvement over MPNLMS for the misalignment is plotted against for NLMS, PNLMS,
sparse AIR. After the echo path change, SC-MPNLMS achieves SC-PNLMS, IPNLMS, and SC-IPNLMS in Fig. 17, and for
approximately 3-dB improvement compared to MPNLMS and NLMS, MPNLMS, and SC-MPNLMS in Fig. 18. As before, all
about 8-dB better performance than NLMS for dispersive step-sizes have been adjusted so that the algorithms achieve the
AIR. As shown in Fig. 14, with speech signal as the input, same steady-state normalized misalignment. These correspond
the proposed SC-MPNLMS algorithm achieves approximately to
10-dB improvement during initial convergence compared to and .
NLMS and 2 dB compared to MPNLMS for the sparse AIR. A zero mean WGN was used as an input signal while another
LOGANATHAN et al.: A CLASS OF SPARSENESS-CONTROLLED ALGORITHMS FOR ECHO CANCELLATION 1599

0
Fig. 17. Time to reach the 20-dB normalized misalignment against different
sparseness measures of eight systems for NLMS, PNLMS, SC-PNLMS,
Fig. 15. Relative convergence of NLMS, IPNLMS, and SC-IPNLMS using IPNLMS, and SC-IPNLMS.
WGN input signal with an echo path change at 3.5 s. Impulse response is
changed from that shown from Fig. 3(a) to (b) and  =  =
03
: ; = 0 7 SNR = 20
: ; dB.

0
Fig. 18. Time to reach the 20-dB normalized misalignment against dif-
ferent sparseness measures of eight systems for NLMS, MPNLMS, and
SC-MPNLMS.

F. Convergence Performance of SC-IPNLMS for NEC

Fig. 16. Relative convergence of NLMS, IPNLMS, and SC-IPNLMS using
speech input signal with echo path changes at 58 s. Impulse response is
We provide additional simulations to illustrate the per-
changed from that shown in Fig. 3(a) to (b) and  =  = formance of SC-IPNLMS in the context of sparse adaptive
03
: ; = 0 8 SNR = 20
: ; dB. NEC, such as may occur in network gateways for mixed
packet-switched and circuit-switched networks. Fig. 19 shows
two impulse responses, sampled at 8 kHz comprising a 12-ms
WGN sequence w(n) was added to achieve an SNR of 20 dB. It active region located within a total duration of 128 ms. The
can be seen that when the AIRs are sparse, the speed of initial sparseness of these impulse responses computed using (12) are
convergence increases significantly for each algorithm. This is (a) and (b) respectively. As before,
because many of the filter coefficients are initialized close to we used a WGN input signal while another WGN sequence
their optimum values since during initialization, . is added to give an SNR of 20 dB. Fig. 20 shows the perfor-
In addition, the sparseness-controlled algorithms (SC-PNLMS, mances of NLMS, IPNLMS, for and , and
SC-MPNLMS and SC-IPNLMS) give the overall best per- the proposed SC-IPNLMS algorithm with .
formance compared to their conventional methods across the An echo path change was introduced using impulse responses
range of sparseness measure. This is because the proposed as shown from Fig. 19(a) to (b) at 3.5 s. We can see from the
algorithms take into account the sparseness measure of the result that the performance of IPNLMS is dependent on .
estimated impulse response at each iteration. More importantly, a faster rate of convergence can be seen for
1600 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 8, NOVEMBER 2009

[2] D. L. Duttweiler, “Proportionate normalized least mean square adap-

tation in echo cancellers,” IEEE Trans. Speech Audio Process., vol. 8,
no. 5, pp. 508–518, Sep. 2000.
[3] A. W. H. Khong, J. Benesty, and P. A. Naylor, “Stereophonic acoustic
echo cancellation: Analysis of the misalignment in the frequency do-
main,” IEEE Signal Process. Lett., vol. 13, no. 1, pp. 33–36, Jan. 2006.
[4] H. Deng and M. Doroslovacki, “Wavelet-based MPNLMS adaptive al-
gorithm for network echo cancellation,” EURASIP J. Audio, Speech,
Music Process., 2007.
[5] R. H. Kwong and E. Johnston, “A variable step size LMS algorithm,”
IEEE Trans. Signal Process., vol. 40, no. 7, pp. 1633–1642, Jul. 1992.
[6] C. Rusu and F. N. Cowan, “The convex variable step size (CVSS) al-
Fig. 19. Sparse impulse responses, sampled at 8 kHz, giving (a)  (n) = 0:88
gorithm,” IEEE Signal Process. Lett., vol. 7, no. 9, pp. 256–258, Sep.
and (b)  (n) = 0:85, respectively.
2000.
[7] J. Sanubari, “A new variable step size method for the LMS adaptive
filter,” in Proc. IEEE Asia-Pacific Conf. Circuits Syst., 2004, pp.
501–504.
[8] B. A. Schnaufer and W. K. Jenkins, “New data-reusing LMS algo-
rithms for improved convergence,” in Proc. 27th Asilomar Conf. Sig-
nals, Syst., Comput., 1993, pp. 1584–1588.
[9] K. A. G. Robert, A. Soni, and W. K. Jenkins, “Low-complexity data
reusing methods in adaptive filtering,” IEEE Trans. Signal Process.,
vol. 52, no. 2, pp. 394–405, Feb. 2004.
[10] A. W. H. Khong and P. A. Naylor, “Selective-tap adaptive algorithms in
the solution of the non-uniqueness problem for stereophonic acoustic
echo cancellation,” IEEE Signal Processing Lett., vol. 12, no. 4, pp.
269–272, Apr. 2005.
[11] P. A. Naylor and A. W. H. Khong, “Affine projection and recursive
least squares adaptive filters employing partial updates,” in Proc. 38th
Asilomar Conf. Signals, Syst. Comput., Nov. 2004, vol. 1, pp. 950–954.
[12] K. A. Lee and S. Gan, “Improving convergence of the NLMS algorithm
using constrained subbands updates,” IEEE Signal Process. Lett., vol.
11, no. 9, pp. 736–739, Sep. 2004.
[13] A. Deshpande and S. L. Grant, “A new multi-algorithm approach to
sparse system adaptation,” in Proc. Eur. Signal Process. Conf., 2005.
[14] S. L. Gay, “An efficient, fast converging adaptive filter for network echo
cancellation,” , vol. 1, pp. 394–398, Nov. 1998.
[15] J. Benesty and S. L. Gay, “An improved PNLMS algorithm,” in Proc.
0
Fig. 20. Relative convergence of NLMS, IPNLMS for = 0:5 and 0:75 0 IEEE Int. Conf. Acoust. Speech Signal Process., 2002, vol. 2, pp.
and SC-IPNLMS using WGN input signal with an echo path change at 3.5 s. 1881–1884.
Impulse response is changed from that shown in Fig. 19(a) to (b) and  = [16] A. W. H. Khong and P. A. Naylor, “Efficient use of sparse adaptive fil-
 = 0:3;  = 0:7; SNR = 20 dB. ters,” in Proc. Signals, Syst. Comput., ACSSC’06. 40th Asilomar Conf.,
Oct. 2006, pp. 1375–1379.
[17] M. A. M. Nekuii, “A fast converging algorithm for network echo can-
cellation,” IEEE Signal Process. Lett., vol. 11, no. 4, pp. 427–430, Apr.
SC-IPNLMS compared to NLMS and IPNLMS both at initial 2004.
convergence and also after the echo path change. [18] H. Deng and M. Doroslovacki, “Improving convergence of the
PNLMS algorithm for sparse impulse response identification,” IEEE
Signal Process. Lett., vol. 12, no. 3, pp. 181–184, Mar. 2005.
VII. CONCLUSION [19] E. Hänsler, “The hands-free telephone problem—An annotated bibli-
ography,” Signal Process., vol. 27, no. 3, pp. 259–271, Jun. 1992.
We have presented a class of sparseness-controlled algo- [20] R. Ahmad, A. W. Khong, and P. A. Naylor, “Proportionate frequency
rithms which achieves improved convergence compared to domain adaptive algorithms for blind channel identification,” in Proc.
IEEE Int. Conf. Acoust. Speech Signal Process., May 2006, vol. 5, pp.
classical NLMS and typical sparse adaptive filtering algo- V29–V32.
rithms. We have incorporated the sparseness measure into [21] G. W. Elko, E. Diethorn, and T. Gänsler, “Room impulse response vari-
PNLMS, MPNLMS, and IPNLMS for AEC to achieve fast ation due to thermal fluctuation and its impact on acoustic echo cancel-
lation,” in Proc. Int. Workshop Acoust. Echo Noise Control, 2003, pp.
convergence that is robust to the level of sparseness encoun- 67–70.
tered in the impulse response of the echo path. The resulting [22] H. Sabine, “Room acoustics,” Trans. IRE Professional Group Room
SC-PNLMS, SC-MPNLMS, and SC-IPNLMS algorithms take Acoust., vol. 1, no. 4, pp. 4–12, Jul. 1953.
[23] J. Benesty, Y. A. Huang, J. Chen, and P. A. Naylor, “Adaptive al-
into account the sparseness measure via a modified coefficient gorithms for the identification of sparse impulse responses,” in Se-
update function. lected Methods for Acoustic Echo and Noise Control, E. Hänsler and
It has been shown that the proposed sparseness-controlled al- G. Schmidt, Eds. New York: Springer, 2006, ch. 5, pp. 125–153.
[24] J. B. Allen and D. A. Berkley, “Image method for efficiently simu-
gorithms are robust to variations in the level of sparseness in lating small-room acoustics,” J. Acoust. Soc. Amer., vol. 65, no. 4, pp.
AIR with only a modest increase in computational complexity. 943–950, Apr. 1979.
Moreover, we have shown that these proposed algorithms have [25] P. M. Peterson, “Simulating the response of multiple microphones to
a single acoustic source in a reverberant room,” J. Acoust. Soc. Amer.,
same or faster convergence in NEC. vol. 80, no. 5, pp. 1527–1529, Nov. 1986.
[26] K. Wagner and M. Doroslovacki, “Towards analytical convergence
analtsis of proportion-type NLMS algorithms,” in Proc. IEEE Int.
REFERENCES Conf. Acoust. Speech Signal Process., 2008, pp. 3825–3828.
[1] J. Radecki, Z. Zilic, and K. Radecka, “Echo cancellation in IP net- [27] H. Deng and M. Doroslovacki, “Proportionate adaptive algorithms for
works,” in Proc. 45th Midwest Symp. Circuits Syst., 2002, vol. 2, pp. network echo cancellation,” IEEE Trans. Signal Process., vol. 54, no.
219–222. 5, pp. 1794–1803, May 2006.
LOGANATHAN et al.: A CLASS OF SPARSENESS-CONTROLLED ALGORITHMS FOR ECHO CANCELLATION 1601

Pradeep Loganathan received the M.Eng. degree in research was mainly on partial-update and selective-tap adaptive algorithms
information systems engineering from Imperial Col- with applications to mono- and multichannel acoustic echo cancellation for
lege London, London, U.K., in 2007. hands-free telephony. He has also published works on acoustic blind channel
Since 2007, he has been with Imperial College identification for speech dereverberation. His other research interests include
London as a Research Postgraduate. His research in- speech enhancement and blind deconvolution algorithms.
terests are mainly in the area of adaptive algorithms,
both in time and frequency domains, with applica-
tions to single-channel acoustic echo cancellation.

Patrick Naylor (M’89–SM’07) received the B.Eng.

degree in electronics and electrical engineering from
the University of Sheffield, Sheffield, U.K., in 1986
and the Ph.D. degree from Imperial College London,
Andy W. H. Khong (M’06) received the B.Eng. London, U.K., in 1990.
degree from Nanyang Technological University, Since 1989, he has been a Member of Academic
Singapore, in 2002 and the Ph.D. degree from De- Staff in the Communications and Signal Processing
partment of Electrical and Electronic Engineering, Research Group, Imperial College London, where he
Imperial College London, London, U.K., in 2005. is also Director of Postgraduate Studies. His research
He is currently an Assistant Professor in the interests are in the areas of speech and audio signal
School of Electrical and Electronic Engineering, processing and he has worked in particular on adap-
Nanyang Technological University, Singapore. Prior tive signal processing for acoustic echo control, speaker identification, multi-
to that, he a Research Associate (2005–2008) in channel speech enhancement, and speech production modeling. In addition to
the Department of Electrical and Electronic Engi- his academic research, he enjoys several fruitful links with industry in the U.K.,
neering, Imperial College London. His postdoctoral U.S., and in mainland Europe.
research involved in developing signal processing algorithms for vehicle Dr. Naylor is an Associate Editor of IEEE TRANSACTIONS ON AUDIO, SPEECH
AND LANGUAGE PROCESSING and a member of the IEEE Signal Processing So-
destination inference as well as the design and implementation of acoustic
array and seismic fusion algorithms for perimeter security systems. His Ph.D. ciety Technical Committee on Audio and Electroacoustics.