Journal of Computer Science

Original Research Paper

Speech Enhancement Based on Adaptive Noise Cancellation

and Particle Swarm Optimization
1
Tayseer M.F. Taha, 2Summrina Kanwal Wajid and 3,4Amir Hussain
1
Department of Computer Science, Sudan University for Sciences and Technology, Khartoum, Sudan
2
Faculty of Natural Sciences, University of Stirling, Scotland, UK
3,4
School of Computing, Edinburgh Napier University, Scotland, UK & Taibah Valley,
Taibah University, Madinah, Saudi Arabia

Article history Abstract: Speech enhancement is used in almost all modern

Received: 19-11-2018 communication systems. This is due to the quality of speech being
Revised: 24-12-2018 degraded by environmental interference factors, such as: Acoustic additive
Accepted: 18-05-2019 noise, acoustic reverberation or white Gaussian noise. This paper, explores
Corresponding Author:
the potential of different benchmark optimization techniques for enhancing
Tayseer M.F. Taha the speech signal. This is accomplished by fine tuning filter coefficients
Department of Computer using a diverse set of adaptive filters for noise suppression in speech
Science, Sudan University for signals. We consider the Particle Swarm Optimization (PSO) and its
Sciences and Technology, variants in conjunction with the Adaptive Noise Cancellation (ANC)
Khartoum, Sudan approach, for delivering dual speech enhancement. Comparative simulation
Email: [email protected] results demonstrate the potential of an optimized coefficient ANC over a
fixed one. Experiments are performed at different signal to noise ratios
(SNRs), using two benchmark datasets: the NOIZEUS and Arabic dataset.
The performance of the proposed algorithms is evaluated by maximising
the perceptual evaluation of speech quality (PESQ) and comparing to the
audio-only Wiener Filter (AW) and the Adaptive PSO for dual channel
(APSOforDual) algorithms.

Keywords: Speech Enhancement, Adaptive Noise Cancellation, Adaptive

Filters, Meta-Heuristic Algorithms, Particle Swarm Optimization

Introduction work best when the additional noise is assumed to be

constant or stationary. These assumptions, however, do
Many researchers have worked on the problem of not hold true in actual operational situations where the
noise cancellation over the past several decades properties and amplitudes of additional noise signals are
(Aggarwal et al., 2016; Fisli et al., 2018a; Mahbub et al., varying, along with external factors, such as traffic noise,
2010). Speech enhancement and noise cancellation have factory sounds and cafeteria babble. To deal with such
involved extensive applications in speech bandwidth problems, we make use of the ANC approach. In The
compression, speaker verification and speech recognition conventional ANC comprises two channels:the first
(Gorriz et al., 2009; Lin, 2003). For speech recognition captures the reference noise signal and the second
and speaker identification, signal enhancement captures the desired or primary signal source (with
techniques improve the quality of the audio signal, noise). This enables the ANC device to sense variations
which in itself is a fundamental step towards achieving in the noise amplitude easily. A number of different
correct classification. If single channel applications are algorithms have been proposed for ANC using such a
considered, spectral subtraction methods are most dual channel set-up, (Kunche and Reddy, 2016a). The
commonly used after noise estimation (Lin, 2003; Lu most commonly used methods are least mean-squares
and Loizou, 2008). In practical scenarios, however, these (LMS) and normalized LMS (NLMS) (Widrow and
techniques have their own share of limitations. They can Stearns, 1985; Gorriz et al., 2009; Mohammed, 2007; Bai
result in musical noise that might distort the signal in the and Yin, 2010). However, these methods are not ideal for
process. Furthermore, such techniques are hugely a multimodal error surface as they have a tendency to get
dependent on properties of the noise signal as, they only stuck in local optima (Ji et al., 2008).

© 2019 Tayseer M.F. Taha, Summrina Kanwal Wajid and Amir Hussain. This open access article is distributed under a
Creative Commons Attribution (CC-BY) 3.0 license.
Tayseer M.F. Taha et al. / Journal of Computer Science 2019, 15 (5): 691.701
DOI: 10.3844/jcssp.2019.691.701

Stochastic optimization algorithms have matured Their proposed method yielded better evaluation results
quite rapidly over the past few decades and one possible compared to Bayesian Non Negative Matrix
application is for solving challenging noise reduction Factorization (BNMF) (Schmidt et al., 2009) and MMSE
problems Stochastic approaches in fact, are far superior approaches (Ephraim and Malah, 1984).
to (Gentle et al., 2012). In general, there are two types of A Modified Predator-Prey Particle Swarm
stochastic algorithms, namely, heuristics and meta- optimization (MPPPSO) for noise cancellation has been
heuristics based. Heuristic means to find or to discover, recently proposed by (Fisli et al., 2018b), (Fisli and
whilst meta-heuristic is associated with (Yang, 2011). Djendi, 2018). The proposed algorithm showed good
Popular meta-heuristic optimization techniques results compared to other methods such as the Predator-
include: Particle Swarm Optimization (PSO), Prey Particle Swarm Optimization (PPPSO) and other
Accelerated Particle Swarm Optimization (APSO) and methods in the literature.
Gaussian Particle Swarm Optimization (GPSO). In The main drawback of using standard PSO is that in
particular, the PSO, a hugely popular optimization some cases, its convergence speed becomes very low. Its
technique, has been applied in a growing range of search space is also fairly limited (Kunche and Reddy,
applications. The use of PSO is not restricted to a simple 2016b). Yang (2010) however, the authors provided a
function optimization, but applied in many challenging solution to these limitations by proposing another
applications such as control systems and pattern modified form of PSO, termed the Accelerated
classification systems (Geravanchizadeh and Asl, 2010). PSO(APSO). This was shown to have a comparatively
PSO and its variants are known for their quick simpler implementation and a much faster convergence
convergence, robust global search and ease of speed. APSO was used for speech enhancement in 2014
implementation (Bai, 2010). by (Prajna et al., 2014). The authors conducted study on
Mahbub et al. (2010) considered the variation in the dual channel speech enhancement and compared the
total number of considered particles in different acoustic results of APSO with PSO. For evaluation purposes they
environments. They conducted research on different used objective measures of: speech intelligibility (FAI),
kinds of noise and voices and also under varied Perceptual Evaluation of Speech Quality (PESQ) and
operating conditions. They compared the results of PSO Signal to Noise Ratio (SNR). The noise types they
with other adaptive algorithms, namely LMS and considered were babble and factory noise, for which
NLMS. Their experiments showed that PSO outperforms APSO proved to be far superior to PSO in terms of
other techniques with respect to SNR improvement and improved speech signal quality and intelligibility.
demonstrated a satisfactory convergence rate under The key contribution of this research is to formulate
different acoustic conditions. Asl and Nezhad (2010) an ANC system based on Butterworth and Elliptic filters,
proposed a Modified PSO (MPSO) and compared it with in the form of an optimization task. Three meta-heuristic
PSO when used for adaptive filtering in the enhancement optimization techniques (PSO, APSO, GPSO) are used
to find the optimal filters coefficients, that optimize the
of speech signals. Their experimental results showed that
perceptual evaluation of speech quality (PESQ), signal
MPSO is capable of a much faster search speed when
distortion (C_sig), signal overall quality (C_ovrl) and
finding an optimal solution. Moreover, MPSO improves Likelihood Ratio (LLR,) for the noise-free audio signal
SNR to a greater extent than the simple PSO. This and the filtered signal.
improvement is more pronounced in the construction of The rest of this paper is organized as follows: Section
higher order filters. (Krohling, 2004) proposed a slightly 2 presents the background and related work. Section 2
modified MPSO technique, based on Gaussian introduces the proposed optimized speech enhancement
probability distribution. It is termed Gaussian PSO or system. Comparative results and a discussion of the
GPSO. In the standard PSO, a number of parameters, experimental set-up is presented in Section 4. Finally,
such as accelerating constants, inertia weight, maximum some concluding remarks and future work suggestions
velocity and the number of particles, need to be initially are presented in Section 5.
defined, which the GPSO does not require. The sole
variable that needs to be initially defined is the total Background and Related Work
number of swarm particles. Comparative simulation
results showed the superiority of GPSO over the Swarm systems consist of nature-based
standard PSO for the data that was considered. To the computational methods (Kennedy and Eberhart, 2001)
best of our knowledge, GPSO has never been used that are based on the behavior of a group of birds.
before for speech enhancement problems. Swarm systems can solve complex problems with
Selvi and Suresh (2016) employed a hybridization of considerable efficiency (Poli, 2008). When a group of
spectral filtering and an optimization algorithm for birds solves some given problem, it is said to be due
speech enhancement, by combining MMSE and PSO. to swarm intelligence. Other common examples are

692
Tayseer M.F. Taha et al. / Journal of Computer Science 2019, 15 (5): 691.701
DOI: 10.3844/jcssp.2019.691.701

from colonies of social insects, such as bees, termites Algorithm 1 Finding optimal solution by using PSO
or ants. This section will present a review of popular 1: For each particle in the population initializes positions
meta-heuristic algorithms, namely classical PSO and and velocities in the search space
APSO and Gaussian PSO. 2: while end criteria not reached do
3: for each particle i do
Particle Swarm Optimization and its Variants 4: Calculate velocity of the particle using
PSO is an artificial intelligence technique, quite Equation 1
5: Update the position of the particle using
commonly used for optimization purposes. It models the
Equation 2
social behavior of a group of birds (a swarm) (Lee and
6: Evaluate the fitness of each particle as in
Lee, 2013). PSO provides an approximate solution for a Equation 7
given optimization problem, using a population of 7: if fitness is better than its pBest in the history
candidate solutions (the particles are termed birds in then
this case). These birds then fly throughout the search
space in accordance with mathematical models 8: set current value as the new pBest
determining their velocity and position. One of its main 9: end if
advantages is that it can handle very large search 10: if fitness is better than its gBest then
spaces with little or no assumptions about the problem 11: set current value as the new gBest
at hand and does not require the problem to be 12: end if
differentiable. Hence it is robust enough to deal with 13: end for
problems that have some factors changing over time 14: end while
(Lee and Lee, 2013).
PSO has the ability to carry out a global search by Hence, the velocity equation is defined as follows
adjusting the positions of particles (Subha and (Wan et al., 2011):
Himavathi, 2016). The position of each particle is
determined by the current global best position and the vit +1 = vit + β1 ( Gbest − xit ) + β2 ( Pbest − xit ) (3)
personal best position.
If xit and xit represent the current position and
where, β1 and β2 are positive random number generated
velocity vector respectively for particle i, the subsequent
by a normal Gaussian distribution N(0, 1).
velocity vector and the position of the particle are
The standard PSO uses both the global best and
determined by the following equations:
personal best position of the particles (Subha and
Himavathi, 2016). The accelerated particle swarm
vit +1 = wvit + αε1 ( Gbest − xit ) + βε 2 ( Pbest − xit ) (1) optimization (APSO) algorithm is a simpler version of
the PSO algorithm, which uses the global best only.
xit +1 = xit + vit +1 (2) Thus, in the APSO, the velocity vector is generated by
the following simpler formula:
where, ε1 and ε2 are random numbers less than 1, α
and β are the acceleration constants and w is the vit +1 = vit + αε + β ( Gbest − xit ) (4)
inertia weight. Although it has numerous advantages,
PSO nevertheless has the tendency to get trapped in local where, the value of e is a random number between 0 and
minima, in some cases, converging to solutions that are 1. The position of the particles can then be updated using
far from ideal (Farooq et al., 2017). Equation 2. The next position of the particle is computed
The PSO algorithm has several parameters that are by combining Equations 2 and 4:
required to be appropriately set, in order to deliver a
good solution. The choice of these fixed parameters is xit +1 = (1 − β ) xit + β Gbest + αε (5)
known to have a considerable effect on the quality of
optimization. Much research has been conducted to find Therefore, APSO is much simpler and results in
appropriate methods which can assist in finding a faster convergence.
suitable set of these parameters. According to (Lee and Noise Cancellation using Adaptive Filters
Lee, 2013). GPSO, which is based on Gaussian The concept of ANC was first introduced by
distribution instead of a random distribution, enhances (Widrow et al., 1975). It requires a minimum of two
the convergence quality of PSO without the need for any microphones and was developed on the basis of finding
kind of parameter adjustment. orientation channel(s) that can detect features of

693
Tayseer M.F. Taha et al. / Journal of Computer Science 2019, 15 (5): 691.701
DOI: 10.3844/jcssp.2019.691.701

associated samples or references to the polluted noise. In this paper, we aim to formulate the ANC problem
An estimate of the noise is produced with the help of an in the form of an optimization task. Specifically, we
adaptive filter by utilizing the reference microphone optimize Butterworth and Elliptic adaptive filters for
output. Its output is then deducted from the primary noise cancellation. Next, we outline our proposed speech
microphone output (signal + noise). The output of the enhancement system, employing ANC based on
canceler is used to regulate the tap weights in the optimisation algorithms.
adaptive filter. With the help of an adaptation algorithm,
ANC minimizes the mean square error value of the
output. It generates an output which is the best
Proposed Speech Enhancement System
approximation of the anticipated signal in the sense of The aim of this research is to compare the
being the minimum mean square error (Taha et al., performance of PSO, APSO and GPSO algorithms for
2018). ANC removes or suppresses a noisy signal by tuning coefficients of an adaptive filter, in order to
using Adaptive-Filters and adjusting their parameters remove the noise from speech signals. This is realized by
according to an optimization algorithm, as in Fig. 1. Many determining the optimal set of filter parameters that
works are reported in the literature use Adaptive filters for optimize (PESQ), (C_sig), (C_ovrl) and (LLR) for noise-
noise reduction and cancellation (Akhaee et al., 2005), free audio signal and filtered signal.
(Kalamani et al., 2014). PESQ is a popular speech objective measure, it was
Conventional adaptive-filters include classical
recommended by ITU-T recommendation P.862
Butterworth-filters, Chebyshev-filters and Elliptic-filters.
A Butterworth filter provides the maximum flat response (Recommendation, 2001), that compares the clean signal
and its calculations are comparatively simpler than other to the degraded signal. It returns a score value ranging
forms of filters. This factor, combined with the fact between -0.5 to 4.5, the higher the value the better
that it produces impressive performance for most quality of the speech.
applications, has made it a popular choice in the field Hu and Loizou (2008) a composite measure is
of electronics-RF as well as with audio active filters introduced by combining different objective measures,
(Adrio, 2015). to determine the overall speech quality. The
An Elliptic filter (also called a Cauer filter) has ripple composite measure is obtained by combining PESQ,
in the pass-band and in the stop-band (Adrio, 2013). Ripple Weighted Spectral Slope (WSS) and (LLR) in one
levels in the pass-band and stop-band are independently measure Covrl, where:
adjustable during the design phase, as followes:
• When a ripple in the stop band approaches zero, Covrl = 1.594 + 0.805 ⋅ PESQ − 0.512 ⋅ LLR − 0.007 ⋅ WSS (6)
then the filter becomes a Chebyshev type I
• When a ripple in pass band approaches zero, then Hence, we formulate the objective function as:
the filter becomes a Chebyshev type II
• When a ripple in both, the stop and the pass-bands 1 2 1
approaches zero, then the filter becomes a C = min + + + LLR (7)
PESQ C _ Ovrl C _ Sig
Butterworth type

Primary signal
∑

Desired signal +
e(n)
∑

y(n)
Inference signal
(noise) Reference
Adaptive filter
Signal

PSO meta-heuristic
algorithm

Fig. 1: Adaptive optimised filter

694
Tayseer M.F. Taha et al. / Journal of Computer Science 2019, 15 (5): 691.701
DOI: 10.3844/jcssp.2019.691.701

Figure 2 explains the overall structure of the proposed (Rothauser, 1969), which were spoken by three
speech enhancement system. Here, the standard PSO and females and three males
GPSO are utilized to obtain the optimum solution. The • The second speech corpus used for experimenting for
APSO can be obtained in the figure by ignoring the a proposed system was an Arabic speech corpus
particle best (using the global best only). (Halabi, 2016). It is a Modern Standard Arabic
(MSA) speech corpus for speech synthesis and was
Dataset recorded in South Levantine Arabic (with a
Two different databases are used to evaluate results: Damascan accent) using a professional studio. It
contains 1813 wav files containing spoken utterances
• A noisy speech corpus for the evaluation of a speech • The Babble noise is chosen from the Signal
enhancement algorithm dataset (NOIZEUS) which Processing Information Base (SPIB) (SPIB, 2013)
is a freely available database (Hu and Loizou, 2007). and added to these clean signals at different SNRs
It has a total of 30 IEEE speech sentences for both datasets

Start

Dataset (Audio speech+ Noise)

Speech filtering

Filter selection

Butterworth filter Elliptic filter

No Optimization
algorithm

Yes
Define Pop size, Max Iteration,
Fitness function

Randomly initialize position, velocity,

Gbest, Pbest for each particle

For each particle (coefficient

of the filter)

Gbest = x Update velocity and

position

fitness x better than

fitness Gbest

Pbest = x

Evaluate the fitness function

Fitness x better f(x) = min 1/PESQ + 2 Covrl +
than fitness Pbest 1/Csig + LLR

No Termination condition
reached

Yes
Filtered speech with fixed Output the evaluated results Filtered speech with optimized Return best position
coefficients coefficients

End

Fig. 2: The overall structure of the proposed speech enhancement system

695
Tayseer M.F. Taha et al. / Journal of Computer Science 2019, 15 (5): 691.701
DOI: 10.3844/jcssp.2019.691.701

Adaptive Noise Cancellation based on Optimization And in the case of the Elliptic filter:
Algorithms
• Filter order = 2
Each particle in the search space is considered a • Peak-to-peak ripple = 0.5
possible solution representing the coefficients of the • Stop-band attenuation = 20
filter. The proposed optimized speech enhancement is • The passband edge frequency 0.99
carried out as follows:
1. Initialize positions and velocities randomly for Evaluation Measurement
each particle in the search space
2. Evaluate the fitness function for each particle using To evaluate the proposed enhancement system, the
Equation 7 objective PESQ measurement is used. PESQ is a popular
3. Find the personal best and the global best (for PSO; and widely used objective speech measure;
the global best is only for APSO) recommended by ITU-T recommendations P.862
4. Update the velocity and the position of each article (Recommendation, 2001). It compares the clean signal to
for PSO (using Equations 1 and 2), for APSO the degraded signal and returns a score value ranging
(using Equations 4, 5) and for GPSO (using from -0.5 to 4.5; the higher the value, the better the
Equation 3) quality of the speech.
5. Repeat steps 2-4 until the stop criteria are met (the
maximum no of iteration is reached or the optimal Results and Discussion
solution is found)
The performance of the proposed system was
In order to find the optimized filter co-efficients, the examined for different SNR values at (-10 db, 0 db, 5
following parameters are calculated. In the case of the db), both benchmark datasets. Further, it was compared
Butterworth-filter: to that of the state-of-the-art audio only and dual channel
speech enhancement algorithms, namely the audio only
• The cut-off frequency Wiener Filter (AW) (Scalart et al., 1996) and the dual
And in the case of the Elliptic filter: speech enhancement approach based on APSO
(APSOforDual) (Prajna et al., 2014). Matlab
• The filter order implementations of the audio only Wiener method were
• Peak-to-peak ripple in decibels used from (Loizou, 2013).
• Minimum stop band attenuation The simulation conditions for all the three
• Passband edge frequency algorithms were as follows: the population size was
set to 20, total iterations set to 50 and other
The noisy signal is filtered using these optimized parameters set as follows: α = 1.5; β = 2 and α = 0.3;
coefficients. β = 0.5 for PSO and APSO respectively.
Finally, a comparison of the speech enhancement The resulting waveforms of PSO and GPSO are
results with and without the use of optimized coefficients presented in Figs. 3 and 4 where an improved sound is
is carried out.
seen to be produced when using both Butterworth and
The noisy signal is filtered using fixed filter
Elliptic filters with optimized coefficients. The audio
coefficients with following values of parameters:
In the case of Butterworth: signal is corrupted by babble noise at 5 db SNR only.
Files were chosen randomly from the NOIZEUS
• Cut-off-frequency = 0.5 Hertz. dataset.

File#1 File#1
0.4 0.4
With noise Butterworth filtered
Audio signal amplitude
Audio signal amplitude

0.2 0.2

0 0

-0.2 -0.2

-0.4 -0.4
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Time axis ×104 Time axis ×104

696
Tayseer M.F. Taha et al. / Journal of Computer Science 2019, 15 (5): 691.701
DOI: 10.3844/jcssp.2019.691.701

File#1 File#1
0.4 0.4

Audio signal amplitude

Butterworth filtered with: PSO With no noise

Audio signal amplitude

0.2 0.2

0 0

-0.2 -0.2

-0.4 -0.4
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Time axis ×104 Time axis ×104
Fig. 3: Audio signal filtered by a PSO optimized Butterworth coefficients
File#1 File#1
0.4 0.4
With noise Elliptic filtered

Audio signal amplitude

Audio signal amplitude

0.2 0.2

0 0

-0.2 -0.2

-0.4 -0.4
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Time axis ×104 Time axis ×104

File#1 File#1
0.4 0.4
Elliptic filtered with: Gaussian PSO With no noise
Audio signal amplitude

Audio signal amplitude

0.2 0.2

0 0

-0.2 -0.2

-0.4 -0.4
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Time axis ×104 Time axis ×104

Fig. 4: Audio signal filtered by a GPSO optimized Elliptic coefficients

Table 1 shows the results of experiments conducted the methods. On the other hand, the fixed coefficient
with the NOIZUS dataset. An optimized Butterworth filter performs better than the Audio-only Wiener Filter
filter with PSO, APSO and GPSO is applied at 5db, 0db and the Dual APSO and slightly worse than the
and-10db SNRS. The averaged PESQ scores were optimized filter by PSO, APSO and GPSO.
computed for all six speech enhancement methods. The For Table 2 when the Elliptic filter is applied, the
three optimized algorithms are seen to improve the PSO outperforms all the other methods at all SNRs of
PESQ score and outperform the audio-only Winer filter 5db, 0db and -10db. Yet the optimized filter yields higher
and the Dual APSO speech enhancement algorithms. PESQ values compared to the audio-only Wiener filter
Equal scores are obtained for PSO and APSO at SNRs of and the Dual APSO speech enhancement algorithms.
5db, 0db and-10db.This trend does not remain the same We carried out experiments for the Arabic speech
for GPSO, which performs the worst at -10db among all corpus. The results are shown in Tables 3 and 4 for

697
Tayseer M.F. Taha et al. / Journal of Computer Science 2019, 15 (5): 691.701
DOI: 10.3844/jcssp.2019.691.701

different SNRs of 5db, 0db and -10db, for the case of meta-heuristic optimization techniques, namely PSO,
both Butterworth and Elliptic Filters. The APSO APSO and GPSO.
performs the best, compared to PSO and GPSO, at 0db The objective function is formulated such that the
and -10db in Table 3, when applying the Elliptic filter. PESQ, Signal distortion (Csig) and overall speech
The APSO is also seen to outperform both the PSO and quality (Covrl) measures are maximized and the Log-
APSO, at 0db and 5db. Likelihood Ration (LLR) is minimized. The algorithm
Overall, applying optimized adaptive filter searches for optimal particles over different iterations,
coefficients was found to enhance the results, compared until the optimum solution is reached or the number of
to those achieved by applying a fixed adaptive iterations is exceeded.
coefficient filer and state-of-the-art algorithms. The proposed algorithms were tested under various
levels of SNR (5db, 0db,-10db). Benchmark NOIZUS
Statistical Analysis using the t_Test
and Arabic datasets were used to evaluate the proposed
To investigate whether there are any significant techniques using PESQ as a standard evaluation metric.
differences between the means of the clean speech The proposed methods were also compared with two
signal, the filter with a fixed coefficient and the filter state-of-the-art algorithms: the audio-only Wiener Filter
with an optimized coefficient, the authors applied the t and the APSO for dual-speech enhancement algorithm.
tests to the results, at 0:05 level of significance. The null For the NOIZUS dataset and for the case of both
and alternatet hypothesis is tested for the case of the Butterworth and Elliptic filters, results in Tables 1-2,
filter with a fixed coefficient as follows: show that the PSO and APSO generally perform better
than GPSO at all levels of SNR. Furthermore, the three
H0: the clean signal did not make any difference to the proposed algorithms outperform the audio-only Wiener
signal obtained when applying a filter to it, thereby filter and the APSO for dual-channel speech
providing evidence against the alternate hypothesis enhancement algorithms, except at SNR of 5db, for the
Ha: there is a significant difference between the clean case of GPSO, which performs the worst among all
signal and the application of a filter with a fixed methods. Similarly, for the ARABIC dataset, for the case
coefficient of both Butterworth and Elliptic filters, Tables 3 and 4 show
that the performance of PSO and APSO is better than the
For the case of a filter with an optimized coefficient,
other methods in comparison with GPSO at different SNRs.
the null and alternate hypotheses are as follows:
However, at 5db SNR, for the case of PSO and APSO, it
H0: the clean signal did not make any difference to the performs the worst among all methods.
signal obtained when applying a filter, thereby Furthermore, a statistical analysis was carried on the
providing evidence against the alternate hypothesis. means of a clean speech signal, a filter with a fixed
Ha: there is a significant difference between the clean coefficient and a filter with an optimized coefficient
signal and a filter with an optimized coefficient. respectively and on the scores collected at each SNR
level. The results showed there was no statistically
The t_test result shown in Table 5 attests the significant difference at (p_0.05) amongst the
significance of the optimized filters, compared to the enhancement methods and the clean speech.
non-optimized ones and the noisy signal. For future experiments, we plan to utilize other
optimization algorithms to optimize ANC coefficients,
Conclusion and Future Work such as the Bat optimization algorithm and Artificial
This paper presents noise cancellation techniques immune systems. Intelligibility tests will also be carried
with adaptive filter coefficients optimised using three out using additional benchmark datasets.

Table 1: PESQ Comparing Filters with a fixed coefficient (Coeff), a PSO optimized coeff, an APSO optimized coeff, a GPSO
optimized coeff, Audio only Wiener Filter (AW) and APSO for Dual. The Butterworth filter is applied to an audio signal at
SNRs of 5db,0db and -10 db in Babble noise
SNR level Fixed coeff PSO APSO GPSO AW APSOforDual
5db 2.5657 2.6852 2.6852 2.7900 2.2714 2.0611
0db 2.3089 2.4194 2.4194 2.1722 1.9581 2.2785
-10db 1.7656 1.7890 1.7890 0.3118 1.2835 1.6641

Table 2: PESQ Comparing Filters with a fixed coeff, a PSO optimized coeff, an APSO optimized coeff, a GPSO optimized coeff,
Audio only Wiener Filter (AW) and APSO for Dual. The Elliptic filter is applied to an audio signal at SNRs of 5db,0db and
-10 db in Babble noise
SNR level Fixed coeff PSO APSO GPSO AW APSOforDual
5 db 2.5160 2.6015 2.5793 2.5142 2.2714 2.0611
0 db 2.2537 2.3144 2.2853 2.2537 1.9581 2.2785
-10 db 1.7018 1.8625 1.8477 1.8116 1.2835 1.6641

698
Tayseer M.F. Taha et al. / Journal of Computer Science 2019, 15 (5): 691.701
DOI: 10.3844/jcssp.2019.691.701

Table 3: PESQ Comparing Filters with a fixed coeff, a PSO optimized coeff, an APSO optimized coeff, a GPSO optimized coeff,
Audio only Wiener Filter (AW) and APSO fo rDual. The Butterworth filter is applied to an audio signal at SNRs of
5db,0db and -10 db in Babble noise, for Arabic Speech Corpus
SNR level Fixed Coeff PSO APSO GPSO AW APSOforDual
5db 1.3305 1.9657 1.9697 1.2004 0.5169 2.0611
0db 1.9401 2.7671 3.0092 2.0620 0.5417 2.2785
-10db 2.7967 2.4837 3.0899 2.8533 0.5155 1.6641

Table 4: PESQ Comparing Filters with a fixed coeff, a PSO optimized coeff, an APSO optimized coeff, a GPSO optimized coeff,
Audio only Wiener Filter (AW) and APSO for Dual. The Elliptic filter is applied to an audio signal at SNRs of 5db,0db and
-10 db in Babble noise, for Arabic Speech Corpus
SNR level Fixed Coeff PSO APSO GPSO AW APSOforDual
5db 1.2641 2.0975 2.3168 1.5805 0.5169 2.0611
0db 1.8351 2.9864 2.9945 2.4303 0.5417 2.2785
-10db 2.6339 3.6165 3.6001 2.8998 0.5155 1.6441

Table 5: The result of t_test of at the 0.05 level of significance

Dataset Alternate Hypothesis H1 p value t value Null hypothesis H0
PSO Ha : µC − µOpt ≠ 0 0.2690 -1.1054 Accept H0
Ha : µC − µFix ≠ 0 0.3799 -0.8780 Accept H 0
APSO Ha : µC − µOpt ≠ 0 0.2690 -1.1054 Accept H 0
NOIZEUS Dataset Ha : µC − µFix ≠ 0 0.3799 -0.8780 Accept H 0
GPSO Ha : µC − µOpt ≠ 0 0.3242 -0.9858 Accept H 0
Ha : µC − µFix ≠ 0 0.3799 -0.8780 Accept H 0
PSO Ha : µC − µOpt ≠ 0 0.4982 -0.6773 Accept H 0
Ha : µC − µfix ≠ 0 0.5038 -0.6684 Accept H 0
APSO Ha : µC − µOpt ≠ 0 0.4982 -0.6773 Accept H 0
Arabic Corpus Ha : µC − µFix ≠ 0 0.5038 -0.6684 Accept H 0
GPSO Ha : µC − µOpt ≠ 0 0.4955 -0.6814 Accept H 0
Ha : µC − µfix ≠ 0 0.5038 -0.6684 Accept H 0

Author’s Contributions Aggarwal, A., T.K. Rawat and D.K. Upadhyay, 2016.
Design of optimal digital fir filters using
Tayseer M.F Taha: Collected the data, carried out evolutionary and swarm optimization techniques.
the experiments, performed the data-analysis and writing
AEU – Int. J. Electr. Commun., 70: 373-385.
of the manuscript, along with designing and organizing
DOI: 10.1016/j.aeue.2015.12.012
the study.
Akhaee, M.A., A. Ameri and F.A. Marvasti, 2005.
Summrina Kanwal: Designed and organized the
paper, advised on the design the figures, provided critical Speech enhancement by adaptive noise cancellation
feedback and revised the manuscript. in the wavelet domain. Proceedings of the 5th
Amir Hussain: Conceived the original idea and International Conference on Information,
contributed to the design and implementation of the Communications and Signal Processing, Dec. 6-9,
research, along with supervision of the research. IEEE Xplore Press, Bangkok, Thailand, pp: 719-723.
DOI: 10.1109/ICICS.2005.1689142
Ethics Asl, L.B. and V.M. Nezhad, 2010. Speech enhancement
using particle swarm optimization techniques.
Proceedings of the International Conference on
References Measuring Technology and Mechatronics
Adrio, C.L., 2013. Automation, Mar 13-14, IEEE Xplore Press,
http://www.radioelectronics.com/info/rf- Changsha City, China, pp: 441-444.
technologydesign/rf-filters/elliptic-cauer-rffilter- DOI: 10.1109/ICMTMA.2010.510
Bai, L. and Q. Yin, 2010. A modified NLMS algorithm
basics.php
for adaptive noise cancellation. Proceedings of the
Adrio, C.L., 2015. IEEE International Conference on Acoustics Speech
http://www.radioelectronics.com/info/rf- and Signal Processing, Mar. 14-19, IEEE Xplore
technologydesign/rf-filters/butterworthrf-filter- Press, Dallas, TX, USA, pp: 3726-3729.
calculations-formulaeequations.php DOI: 10.1109/ICASSP.2010.5495868

699
Tayseer M.F. Taha et al. / Journal of Computer Science 2019, 15 (5): 691.701
DOI: 10.3844/jcssp.2019.691.701

Bai, Q., 2010. Analysis of particle swarm optimization Hu, Y. and P.C. Loizou, 2007. Subjective evaluation and
algorithm. Comput. Inform. Sci., 3: 180-180. comparison of speech enhancement algorithms.
DOI: 10.5539/cis.v3n1p180 Speech Commun., 49: 588-601.
Ephraim, Y. and D. Malah, 1984. Speech enhancement DOI: 10.1016/j.specom.2006.12.006
using a minimum-mean square error short-time Hu, Y. and P.C. Loizou, 2008. Evaluation of objective
spectral amplitude estimator. IEEE Trans. Acoustics quality measures for speech enhancement. IEEE
Speech Signal Processing, 32: 1109-1121. Trans. Audio Speech Lang. Process., 16: 229-238.
DOI: 10.1109/TASSP.1984.1164453 DOI: 10.1109/TASL.2007.911054
Farooq, M.U., A. Ahmad and A. Hameed, 2017. Ji, C., Y. Zhang, M. Tong and S. Yang, 2008. Particle
filter with swarm move for optimization.
Opposition-based initialization and a modified
Proceedings of the 10th International Conference on
pattern for Inertia Weight (IW) in PSO. Proceedings Parallel Problem Solving from Nature, Sept. 13-17,
of the IEEE International Conference on Innovations Springer, Dortmund, Germany, pp: 909-918.
in Intelligent Systems and Applications, Jul. 3-5, DOI: 10.1007/978-3-540-87700-4 90
IEEE Xplore Press, Gdynia, Poland, pp: 96-101. Kalamani, M., S. Valarmathy and M. Krishnamoorthi,
DOI: 10.1109/INISTA.2017.8001139 2014. Modified noise reduction algorithm for speech
Fisli, S. and M. Djendi, 2018. Blind speech enhancement. Applied Math. Sci., 8: 4447-4452.
intelligiblility enhancement by a new dual modified DOI: 10.12988/ams.2014.45365
predator-prey particle swarm optimization Kennedy, J. and R.C. Eberhart, 2001. Swarm Intelligence.
algorithm. Applied Acoustics, 141: 125-135. 1st Edn., Morgan Kaufmann Publishers Inc., San
DOI: 10.1016/ j.apacoust.2018.07.006 Francisco, CA, USA, ISBN-10: 1-55860-595-9.
Fisli, S., M. Djendi and A. Guessoum, 2018a. Modified Krohling, R.A., 2004. Gaussian swarm: A novel particle
predator-prey particle swarm optimization based swarm optimization algorithm. Proceedings of the
twochannel speech quality enhancement by forward IEEE Conference on Cybernetics and Intelligent
blind sourceseparation. Proceedings of the 2nd Systems, Dec. 1-3, IEEE Xplore Press, Singapore,
pp: 372-376. DOI: 10.1109/ICCIS.2004.1460443
International Conference on Natural Language and
Kunche, P. and K. Reddy, 2016a. Heuristic and Meta-
Speech Processing (ICN’ 18), pp: 1-6. Heuristic Optimization. In: Metaheuristic
DOI: 10.1109/ICNLSP.2018.8374372 Applications to Speech Enhancement, Kunche, P.
Fisli, S., M. Djendi and A. Guessoum, 2018b. Modified and S.M. Reddy (Eds.), Springer, pp: 17-24.
predator-prey particle swarm optimization based Kunche, P. and K. Reddy, 2016b. Metaheuristic
twochannel speech quality enhancement by forward Applications to Speech Enhancement. 1st Edn.,
blind source separation. Proceedings of the 2nd Springer, ISBN-10: 9783319316819, pp: 132.
International Conference on Natural Language and Lee, J.W. and J.J. Lee, 2013. Gaussian-distributed
Speech Processing (ICN’ 18), pp: 1-6. IEEE. particle swarm optimization: A novel Gaussian
DOI: 10.1109/ ICNLSP.2018.8374372 particle swarm optimization. Proceedings of the
Gentle, J.E., W.K. Hardle and Y. Mori, 2012. Handbook IEEE International Conference on Industrial
of Computational Statistics: Concepts and Methods. Technology, Feb. 25-28, IEEE Xplore Press, Cape
2nd Edn., Springer Science and Business Media, Town, South Africa, pp: 1122-1127.
New York, ISBN-10: 3642215513, pp: 1192. DOI: 10.1109/ICIT.2013.6505830
Geravanchizadeh, M. and L.B. Asl, 2010. Asexual Lin, C.T., 2003. Single-channel speech enhancement in
reproduction-based adaptive quantum particle variable noise-level environment. IEEE Trans. Syst.
swarm optimization algorithm for dual-channel Man Cybernet, 33: 137-143.
DOI: 10.1109/TSMCA.2003.811115
speech enhancement. Proceedings of the 4th
International Symposium on Communications, Loizou, P., 2013. Speech Enhancement: Theory and
Control and Signal Processing, Mar. 3-5, IEEE Practice. 2nd Edn., Taylor and Francis, ISBN-13:
Xplore Press, Limassol, Cyprus, pp: 1-4. 9781466504219.
DOI: 10.1109/ISCCSP.2010.5463450. Lu, Y. and P.C. Loizou, 2008. A geometric approach to
Gorriz, J.M., J. Ramırez, S. Cruces-Alvarez, C.G. spectral subtraction. Speech Commun., 50: 453-466.
Puntonet and E.W. Lang et al., 2009. A novel LMS DOI: 10.1016/j.specom.2008.01.003
algorithm applied to adaptive noise cancellation. Mahbub, U., C. Shahnaz and S.A. Fattah, 2010. An
adaptive noise cancellation scheme using particle
IEEE Signal Process. Lett., 16: 34-37.
swarm optimization algorithm. Proceedings of the
DOI: 10.1109/LSP.2008.2008584 IEEE International Conference on Communication
Halabi, N., 2016. Modern standard Arabic phonetics for Control and Computing Technologies, Oct. 7-9,
speech synthesis. Ph.D Thesis, University of IEEE Xplore Press, Ramanathapuram, India, pp:
Southampton, UK. 683-686. DOI: 10.1109/ICCCCT.2010.5670753

700
Tayseer M.F. Taha et al. / Journal of Computer Science 2019, 15 (5): 691.701
DOI: 10.3844/jcssp.2019.691.701

Mohammed, J.R., 2007. A new simple adaptive noise Selvi, R.S. and G. Suresh, 2016. Hybridization of
cancellation scheme based on ale and NLMS filter. spectral filtering with particle swarm optimization
Proceedings of the 5th Annual Conference on for speech signal enhancement. Int. J. Speech
Communication Networks and Services Research,
Technol., 19: 19-31.
May 14-17, IEEE Xplore Press, Frederlcton, NB,
Canada, pp: 245-254. DOI: 10.1109/CNSR.2007.4 DOI: 10.1007/s10772-015-9317-1
Poli, R., 2008. Analysis of the publications on the SPIB, 2013. Signal Processing Information Base (SPIB).
applications of particle swarm optimisation. J. http://spib.linse.ufsc.br/noise.html
Artificial Evolut. Applic. Subha, R. and S. Himavathi, 2016. Accelerated particle
DOI: 10.1155/2008/685175 swarm optimization algorithm for maximum power
Prajna, K., G.S.B. Rao, K. Reddy and R.U. Maheswari, point tracking in partially shaded PV systems.
2014. A new dual channel speech enhancement Proceedings of the 3rd International Conference on
approach based on Accelerated Particle Swarm Electrical Energy Systems, Mar. 17-19, IEEE
Optimization (APSO). Int. J. Intell. Syst. Applic., 6: Xplore Press, Chennai, India, pp: 232-236.
1-1. DOI: 10.5815/ijisa.2014.04.01 DOI: 10.1109/ICEES.2016.7510646
Recommendation, I.T., 2001. Perceptual Evaluation of Taha, T.M.F., A. Adeel and A. Hussain, 2018. A survey
Speech Quality (PESQ): An objective method for on techniques for enhancing speech. Int. J. Comput.
end-to-end speech quality assessment of narrow- Applic., 179: 1-14. DOI: 10.5120/ijca2018916290
band telephone networks and speech codecs. Rec. Wan, C., J. Wang, G. Yang and X. Zhang, 2011.
ITU-T P. 862. Gaussian particle swarm optimization with
Rothauser, E., 1969. IEEE recommended practice for differential evolution mutation. Proceedings of the
speech quality measurements. IEEE Trans. Audio International Conference in Swarm Intelligence,
Electroacoust., 17: 225-246. Springer, pp: 439-446.
DOI: 10.1109/IEEESTD.1969.7405210 DOI: 10.1007/978-3-642-21515-552
Scalart, P. and J.V. Filho, 1996. Speech enhancement Widrow, B., J.R. Glover, J.M. McCool, J. Kaunitz and
based on a priori signal to noise estimation. C.S. Williams et al., 1975. Adaptive noise
Proceedings of the IEEE International Conference on cancelling: Principles and applications. Proc. IEEE,
Acoustics, Speech and Signal Processing, May 9-9, 63: 1692-1716. DOI: 10.1109/PROC.1975.10036
Widrow, B. and S.D. Stearns, 1985. Adaptive Signal
IEEE Xplore Press, Atlanta, GA, USA, pp: 629-632.
Processing. Prentice-hall Englewood Cliffs, NJ.
DOI: 10.1109/ICASSP.1996.543199 Yang, X.S., 2010. Nature-Inspired Metaheuristic
Schmidt, M.N., O. Winther and L.K. Hansen, 2009. Algorithms. 1st Edn., Luniver Press,
Bayesian non-negative matrix factorization. ISBN-10: 1905986106.
Proceedings of the International Conference on Yang, X.S., 2011. Review of meta-heuristics and
Independent Component Analysis and Signal generalised evolutionary walk algorithm. Int. J. Bio-
Separation, (ASS’ 09), Springer, pp: 540-547. Inspired Comput., 3: 77-84.
DOI: 10.1007/978-3-642-005992 68 DOI: 10.1504/IJBIC.2011.039907

701