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

net/publication/264554879

Fast Adaptive Filtering Algorithm for Acoustic

Noise Cancellation

Conference Paper · July 2012

CITATIONS READS

2 51

5 authors, including:

Madjid Arezki Abderrahmane Namane

Saad Dahlab University Saad Dahlab University
31 PUBLICATIONS 67 CITATIONS 18 PUBLICATIONS 62 CITATIONS

SEE PROFILE SEE PROFILE

Patrick. L Meyrueis Daoud Berkani

University of Strasbourg National Polytechnic School of Algiers
161 PUBLICATIONS 403 CITATIONS 99 PUBLICATIONS 246 CITATIONS

SEE PROFILE SEE PROFILE

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

Real time detection of 1D barcode in blurry, low contrast and low resolution. images. View project

Vector Quantization (Information theory) View project

All content following this page was uploaded by Ahmed Benallal on 21 January 2016.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue are added to the original document
and are linked to publications on ResearchGate, letting you access and read them immediately.
Proceedings of the World Congress on Engineering 2012 Vol II
WCE 2012, July 4 - 6, 2012, London, U.K.

Fast Adaptive Filtering Algorithm for

Acoustic Noise Cancellation
M. Arezki, A.Namane, A. Benallal, P. Meyrueis and D. Berkani

 large number of fast RLS (FRLS) algorithms have been

Abstract— In this paper, we propose an algorithm based on developed over the years [4]-[6], but, unfortunately, it seems
the adaptive filtering which we can use for the noise that the better a FRLS algorithm is in terms of computational
cancellation. This algorithm is stable numerically and very efficiency, the more severe is its problems related to
powerful when the input signal is stationary with additive noise
on the output signal. It has a high convergence speed
numerical stability. The FRLS algorithm shows a complexity
comparable with that of RLS algorithm and reduced of O(L). Several numerical solutions of stabilization, with
computational complexity close to NLMS algorithm. The stationary signals, are proposed in the literature [7]. A new
proposed algorithms allow a greater choice of compromise technique using reduced size predictors of order P, where
between these performances criteria and the computational P<<L, was developed in [8], [9]. Fast RLS algorithms that
complexity. use this technique are often called Fast Newton Transversal
Index Terms— NLMS, Fast RLS, FTF-type algorithm, Filter (FNTF) Algorithms. Fast RLS algorithms that use this
Estimation, Adaptive Filtering, ANC. technique are often called Fast Newton Transversal Filter
(FNTF) Algorithms. Several fast RLS algorithms including
I. INTRODUCTION [10], [11] use the technique, and they are able to produce
fast stable RLS algorithms that have a complexity of
A COUSTIC noise cancellation (ANC) techniques are
usually applied in applications where a reference signal
that is correlated with the noise at the primary signal is
(2L  4P)  2L . The computational complexity of these
algorithms approach that of NLMS, and ought to be a very
easily obtained [1]. In these techniques, the reference signal attractive choice for implementation.
is uncorrelated with the clean speech signal. These In many applications of noise cancellation, the changes of
techniques make use of noise reference input and attempt to the signal characteristics can be made quickly. This requires
subtract the noise component from the noisy speech signal. the use of adaptive algorithms, which converge rapidly. Our
The primary microphone picks up the noisy speech signal main objective is to evaluate the adaptive performances of
while a set of secondary microphone measure a signal the algorithms in real situations with criteria measured by the
consisting mainly of noise. The signal from the reference mean square error MSE (n) and the gain G SNR according to
input is fed to an adaptive filter which estimates the noise on SNR In . The proposed algorithms provide better results
the primary input and simply subtracts it from the
transmitted speech [2]. A number of filter structures and when making the noise less audible at exit of the system of
adaptation algorithms have been evaluated in the literature. noise cancellation.
There are two major classes of adaptive algorithms [3]. One
II. ACOUSTIC NOISE CANCELLATION
is the Normalized Least Mean Square (NLMS) algorithm,
which has a computational complexity of O(L), L is the finite Conventional ANC employs two input signals to reduce
impulse response (FIR) filter length. The other class of the noise at the output of the system as illustrated in Fig 1.
adaptive algorithm is the Recursive Least Squares (RLS) The primary input is the noise-corrupted signal d n . The
algorithm has an impressive performance. The main reference input xn , is measure of background noise alone
drawback with the RLS algorithm is its complexity O(L2). A which is in some way correlated with the noise in the
primary. The output a priori error  L, n of this system at time
Manuscript received March14, 2012; revised April16, 2012. This work
was supported in part by the LATSI Laboratory of the Department of n is:
Electronics, University of Blida Algeria and the LSP Laboratory,
University of Strasbourg, France.
 L,n  d n  yˆ n (1)
M. Arezki is with the Department of Electronics, University of Blida,
B.P. 270 Route de Soumâa Blida 09000 Algeria. (phone: +213771204390;
fax: +21325433850; e-mail: [email protected]).
A.Namane is with the Department of Electronics, University of Blida,
Algeria (e-mail: [email protected]).
A. Benallal is with the Department of Electronics, University of Blida
Algeria. (e-mail: [email protected]).
P. Meyrueis is with the LSP Laboratory, University of Strasbourg,
ENSPS, Bd. Sébastien Brant – B.P. 10413 ILLKIRCH 67412 France. (e-
mail: [email protected]).
D. Berkani, is with the Department of Electrical and Computer
Engineering of ENP Algiers, 10 Avenue Pasteur, B.P. 182 El Harrach - Fig 1: Principle of noise cancellation
16000 Algiers Algeria. (e-mail: [email protected]).

ISBN: 978-988-19252-1-3 WCE 2012

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the World Congress on Engineering 2012 Vol II
WCE 2012, July 4 - 6, 2012, London, U.K.

where yˆ n  w TL,n 1 x L,n is the model filter output, by  x,n  (1   )  x,n1   x n2 , where  is a forgetting factor
x L,n  x n , ..., x n L1  (   1 / L ). The computational complexity of the NLMS
T
is a vector containing the last L

samples of the input signal x n , w L,n  w1,n , ...,wL,n T is
algorithm is 2L multiplications per sample.
The RLS algorithm [3], for which minimizes a
the coefficient vector of the adaptive filter and L is the filter deterministic sum of squared errors. Fast versions of these
length. The signal d n is the contaminated signal (Fig 2) algorithms are derived from the RLS by the introduction of
containing both the clean speech s n and the noise y n , forward and backward predictors. The FRLS algorithm
assumed to be uncorrelated with each other. The primary shows a complexity of O(L). The adaptation gain is given
input signal from the model is: by:
~
d n  yn  sn (2) g L, n  R L1,n x L,n   L, n k L, n (6)
    
y n  h Tn x L,n
RLS FRLS
(3) n

where 
h n  h1,n , ...,hL,n T represents the unknown
R L,n  
i 1
n i
x L,i x TL,i   R L, n 1  x L, n x TL, n (7)

system impulse response vector. The purpose of this system

where R L, n is an estimate of the correlation matrix of the
is to enable the system to control the filter until ŷ n is as
input signal vector and  denotes the exponential forgetting
close to y n as possible. ~
factor (0 1). The variables  L, n and k L, n respectively
The error signal  L, n can be used to adapt the adaptive
indicate the likelihood variable and normalized Kalman gain
filter w L,n 1 using some algorithm for filter adaptation.
vector. Several numerical solutions of stabilization, with
Several different algorithms for filter adaptation have been stationary signals, are proposed in the literature [7]. The
proposed. The filter is updated at each instant by feedback computational complexity of the stabilized FRLS (NS-
of the estimation error proportional to the adaptation gain, FRLS) algorithm is 8L per sample, and is stable, with the
denoted as g L, n , and according to assumption of a white Gaussian input signal, under the
following condition [7]:
w L,n  w L,n1  g L,n  L,n (4)
  1 1 / 2L (8)
The different algorithms are distinguished by the
adaptation gain calculation. B. Proposed Adaptive Algorithms
The numerical stabilization of the FRLS algorithm limits
III. PROPOSED SYSTEM the range of the forgetting factor (8) and consequently their
convergence speed and tracking ability. And the resulting of
A. Adaptive NLMS and FRLS Algorithms
that algorithm has an 8L complexity.
The LMS Algorithms derived from the gradient [3], for In [11], they propose more complexity reduction. The
which the optimization criterion corresponds to a modified and simplified FTF-type (M-SMFTF) algorithm
minimization of the mean-square error. For the normalized [11], derived from the SMFTF algorithm [10], where the
LMS (NLMS) algorithm, the adaptation gain is given by: adaptation gain is obtained only from the forward prediction
 variables and using a recursive method to compute the
g L, n  x L, n (5) likelihood variable. The computational complexity is 6L and
L  x,n  c 0
its algorithm is stable, with assumption of a white Gaussian
where  is referred to as the adaptation step and c 0 is a input signal, under the following condition [11]:
small positive constant used to avoid division by zero in   1 1 / L (9)
absence of the input signal. The stability condition of this
By a method of extrapolation [9], the autocorrelation
algorithm is 0<  <2 and the fastest convergence is obtained
matrix of order L is built starting from an estimate of the
for  = 1 [12]. The power of input signal is estimated autocorrelation matrix of order P (P<<L). In this case, it is
not anymore necessary to propagate prediction vectors of
order L in the prediction part of the FRLS algorithm. If we
denote P the order of the predictor and L the size of adaptive
filter, the M-SMFTF algorithm can be easily used with
reduced size prediction part. The computational complexity
of the reduced size predictor M-SMFTF (RM-SMFTF)
algorithm is (2L  4P)  2L and it is stable, with the
assumption of a white Gaussian input signal, under the
following condition [11]:
  11 / P (10)
Fig 1: Principle of noise cancellation

ISBN: 978-988-19252-1-3 WCE 2012

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the World Congress on Engineering 2012 Vol II
WCE 2012, July 4 - 6, 2012, London, U.K.

IV. SIMULATIONS introducing a linear gain variation on the primary input

The noise reference input pass through the adaptive filter signal. In Fig 3, we give the evolution, in decibels, of the
and output ŷ n is produced as close a replica as possible of mean square error MSE (n) (15). It shows that better
y n . The filter readjusts itself continuously to minimize the performances in convergence speed are obtained for the
proposed algorithm. It is observed that the proposed
error between y n and ŷ n during this process. The system
algorithm converges much faster and tracks better the
output is: variation of the system than both NS-FRLS and NLMS
 L,n  s n  y n  yˆ n (11) algorithms.
B. Noise Reduction
We can get the following equation of expectations: We define the input signal to noise ratio segmental [13],
which is calculated on screens of a few milliseconds:
  
E ( L,n ) 2  E (s n  y n  yˆ n ) 2  (12)
 N 1 
M 1

 s 2
i  kN



Assuming that x n and s n are not correlated and have zero 1 i 0
SNR In  10 log1 0  N 1  (16)
means, we can write: M k 0  
  y i2 kN 
    
E ( L, n )2  E (sn )2  E ( yn  yˆ n )2  (13)  i 0 

After convergence of the adaptive filter, explicitly, where N is the size of the screen and M the number of
screens. Same manner, we define the output signal to noise
w L,n  h n  ŷn  y n   L,n  s n (14) ratio segmental:

 N 1 
s
A. Tracking Ability and Convergence
 2
i  kN

M 1
 

In the absence of the speech signal ( s n =0). The 1 i 0
SNR Ou t  10 log1 0 N 1  (17)
 s 
expression (13) becomes: M k 0  2 
 sˆi  kN
 
   
i  kN
E ( L, n )2  E ( yn  yˆ n )2 (15)  i 0 
One of the qualifying criteria of the noise reduction
The input signal x n used in our simulation is a white
algorithms is the increase measurement in signal to noise
Gaussian noise, with mean zero and variance equal to 0.32. ratio. This criterion, noted G SNR and expressed in dB, is
The impulse response of the system represents a real impulse
response measured in a car and truncated to 256 samples. defined as the difference between SNR o u t and SNR In for
We compare the convergence speed and tracking capacity of each screen, averaged on the whole of the screens:
the (NLMS, NS-FRLS, M-SMFTF and RM-SMFTF)
G SNR  SNR out  SNR In (18)
algorithms. The filter length is L=256, the NLMS (  =1)
and NS-FRLS (   11 / 3L ) algorithms are tuned to obtain We generate noisy speech signals with different SNR In
fastest convergence. The forgetting factor  are starting from files from clean speech signal alone and noise
respectively chosen for M-SMFTF and RM-SMFTF alone. We synthesize the noisy speech signals by adding the
algorithms to   1 1 / L and   1 1 / P . The noise to the clean speech signal so as to reach the desired
nonstationarity of the system to be modelled is simulated by levels of SNR In . In our simulations, we took in the case of
the NS-FRLS algorithm,   1 1 / 10L to ensure numerical
stability.
The noisy speech signal with SNR In = 30dB represents a
case which we can say without noise (Fig 4). We observe
(Fig 5a), that the output error of the system, which
represents the estimated speech signal (  L,n  ŝ n ), is almost
confused with the original speech signal except for the zones
of silence (very weak power). We can say in these zones of
silence, in absence of noise the NS-FRLS algorithm adapts
less than the other algorithms. The curves (b) and (c) of
Fig.5, which represent the temporal evolutions SNR o u t and
G SNR with SNR In = 30dB, confirm well that the NS-FRLS
algorithm adapts less than the other algorithms. In Fig 6,
Fig 3: Comparative performance of the algorithms,
with SNR In = - 5dB, we notice that the speech signal is
L=256. M-SMFTF:  =0.9961,  =0.985, c a =0.5, E0=1; relatively drowned in the noise. Fig 7 gives the temporal
RM-SMFTF: P=32,  =0.9688,  =0.9985, c a =0.5, E0=0.2; evolutions SNR o u t and G SNR . We can say that the NS-FRLS
NS-FRLS:  =0.9987, E0=1; NLMS:  =1 algorithm is slightly above the proposed algorithms.

ISBN: 978-988-19252-1-3 WCE 2012

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the World Congress on Engineering 2012 Vol II
WCE 2012, July 4 - 6, 2012, London, U.K.

Fig 8 represents the gain G SNR as a function of SNR In .

The original speech signal is combined with noise for
SNR In between - 10dB and 20dB by steps of 5dB. We
observe that, the gain G SNR decrease with SNR In , the
system has an overall tendency to not to remove the noise
when SNR In is strong and especially for the NS-FRLS
algorithm. It is not necessarily a handicap; values beyond Fig 6 : Noisy speech signal with SNR I n = - 5dB
15dB correspond relatively to slightly disturbed conditions,
and attenuate the noise does not systematically improve
comfort of listening. We notice that the NS-FRLS algorithm
is placed slightly above the proposed algorithms.

V. CONCLUSION
From these performances criteria, measured by the mean
square error MSE (n) and the gain G SNR , we have noticed
that the proposed algorithms provide better results when
making the noise less audible at exit of the system of noise
cancellation. The simulations have shown that, the estimate
of the noise have sensibly improved the adaptive
performances in terms of noise reduction without distorting
the speech signal. The NLMS algorithm is definitely less Fig 7 : Temporal evolutions of SNR O u t and G S NR with SNR I n =-5dB.
powerful than the other algorithms, and that the proposed
algorithms allow a greater choice of compromise between L=256. M-SMFTF:  =0.9961,  =0.985, c a =0.5, E0=1;
these performances criteria and the computational RM-SMFTF: P=32,  =0.9688,  =0.9985, c a =0.5, E0=0.2;
complexity. NS-FRLS:  =0.9996, E0=0.5; NLMS:  =1.

Fig 4 : Noisy speech signal with SNR I n = 30dB Fig 8 : G S NR  f SNR I n 

L=256. M-SMFTF:  =0.9961,  =0.985, c a =0.5, E0=1;
RM-SMFTF: P=32,  =0.9688,  =0.9985, c a =0.5, E0=0.2;
NS-FRLS:  =0.9996, E0=0.5; NLMS:  =1.

REFERENCES
[1] S. M. Kuo and D. R. Morgan, Active Noise Control Systems, New
York: Wiley, 1996.
[2] M.Arezki, P.Meyrueis, N.Javahiraly, “Specific signal processing
method for sound detected by an intrinsic optical fiber sensor”, Proc.
SPIE 7675, 2010, 76750Q, Defense, Security, and Sensing, Orlando-
USA, April 5-9, 2010.
[3] S. Haykin, Adaptive Filter Theory, 4th ed., NJ, Prentice-Hall, 2002.
[4] L. Ljung, M. Morf and D. Falconner, “Fast calculation of gain
matrices for recursive estimation schemes,” Int. J. Control., vol. 27,
1978, pp.1-19.
[5] G. Caryannis, D. Manolakis and N. Kalouptsidis, “A fast sequential
algorithm for least squares filtering and prediction,” IEEE Trans.
Acoust. Speech Signal Process, vol. 31, 1983, pp.1394–1402.
[6] J. Cioffi and T. Kailath, “Fast RLS Transversal Filters for adaptive
filtering,” IEEE press. On ASSP, vol.32, no.2, 1984, pp.304-337
[7] M. Arezki, A. Benallal, P. Meyrueis, A. Guessoum and D. Berkani,
“Error Propagation Analysis of Fast Recursive Least Squares
Algorithms,” Proc. 9th IASTED International Conference on Signal
and Image Processing, Honolulu, Hawaii, USA, August 20–22,
Fig 5 : Temporal evolutions: (a) MSE(n) ; (b) SNR O u t ; (c) G S NR 2007, pp.97-101.
L=256. M-SMFTF: =0.9961,  =0.985, c a =0.5, E0=1; [8] G.V. Moustakides and S. Theodoridis, “Fast Newton transversal
filters - A new class of adaptive estimation algorithms,” IEEE Trans.
RM-SMFTF: P=32,  =0.9688,  =0.9985, c a =0.5, E0=0.2; Signal Process, vol.39, no.10, 1991, pp.2184–2193.
NS-FRLS:  =0.9996, E0=0.5; NLMS:  =1.

ISBN: 978-988-19252-1-3 WCE 2012

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
Proceedings of the World Congress on Engineering 2012 Vol II
WCE 2012, July 4 - 6, 2012, London, U.K.

[9] P. P. Mavridis and G. V. Moustakides, “Simplified Newton-Type

Adaptive Estimation Algorithms,” IEEE Trans. Signal Process,
vol.44, no.8, 1996.
[10] A. Benallal and A. Benkrid, “A simplified FTF-type algorithm for
adaptive filtering,” Signal processing, vol.87, no.5, 2007, pp.904-
917.
[11] M.Arezki, A. Benallal, P. Meyrueis, and D. Berkani, “A New
Algorithm with Low Complexity for Adaptive Filtering”, IAENG
Journal, Engineering Letters, vol.18, Issue 3, 2010, pp.205-211.
[12] D.T.M. Slock, “On the convergence behaviour of the LMS and the
NLMS algorithms,” IEEE Trans. Signal Processing, vol.42, 1993,
pp.2811-2825.
[13] M. Bradstein and D. Ward, Eds., Microphone arrays: Signal
processing techniques and applications. New York: Springer, 2001.

ISBN: 978-988-19252-1-3 WCE 2012

ISSN: 2078-0958 (Print); ISSN: 2078-0966 (Online)
View publication stats