ELECTRONICS AND ELECTRICAL ENGINEERING

ISSN 1392 – 1215 2007. No. 2(74)

ELEKTRONIKA IR ELEKTROTECHNIKA

ELECTRONICS
T 170
ELEKTRONIKA

Modified Adaptive Filtering Algorithm for Noise Cancellation in Speech

Signals
V. R. Vijaykumar, P. T. Vanathi
Department of Electronics and Communication Engineering, PSG College of Technology,
Coimbatore-641004, India, tel.: -0091-9442014139, e-mail: [email protected], [email protected]
P. Kanagasapabathy
Madras Institute of Technology, Chennai, India, email: [email protected]

Introduction In general, noise [2] that affects the speech signals

can be modeled using any one of the following:
Speech is a very basic way for humans to convey
information to one another with a bandwidth of only 4 1. White noise,
kHz; speech can convey information with the emotion of a 2. Colored noise,
human voice. The speech signal has certain properties: It is 3. Impulsive noise.
a one-dimensional signal, with time as its independent
variable, it is random in nature, it is non-stationary, i.e. the White noise
frequency spectrum is not constant in time. Although
human beings have an audible frequency range of 20Hz to White noise is a sound or signal consisting of all
20 kHz, the human speech has significant frequency audible frequencies with equal intensity. At each
components only up to 4 kHz. frequency, the phase of the noise spectrum is totally
The most common problem in speech processing is uncertain: It can be any value between 0 and 2π, and its
the effect of interference noise in speech signals. value at any frequency is unrelated to the phase at any
Interference noise masks the speech signal and reduces its other frequency. When noise signals arising from two
intelligibility. Interference noise can come from acoustical different sources add, the resultant noise signal has a
sources such as ventilation equipment, traffic, crowds and power equal to the sum of the component powers. Because
commonly, reverberation and echoes. It can also arise of the broad-band spectrum, white noise has strong
electronically from thermal noise, tape hiss or distortion masking capabilities.
products. If the sound system has unusually large peaks in
its frequency response, the speech signal can even end up Colored noise
masking itself.
One relationship between the strength of the speech Any noise that is not white can be termed as colored
signal and the masking sound is called the signal-to-noise noise. Colored noise has frequency spectrum that is limited
ratio, expressed in decibels. Ideally, the S/N ratio is greater within a range unlike white noise which extends over the
than 0dB, indicating that the speech is louder than the entire spectrum.
noise. Just how much louder the speech needs to be in There are different types of colored noise (brown
order to be understood varies with, among other things, the noise, pink noise, orange noise etc.) depending upon the
type and spectral content of the masking noise. gradation in the Power Spectral Density (PSD) of the
The most uniformly effective mask is broadband noise. Colored noise can be generated by passing white
noise. Although, narrow-band noise is less effective at noise through a filter with required frequency response.
masking speech than broadband noise, the degree of
masking varies with frequency. Impulsive noise
High-frequency noise masks only the consonants, and
its effectiveness as a mask decreases as the noise gets Impulsive noise refers to sudden bursts of noise with
louder. But low-frequency noise is a much more effective relatively high amplitude. This type of noise causes click
mask when the noise is louder than the speech signal, and sounds in the signal of interest.
at high sound pressure levels it masks both vowels and Impulsive noise is generally modeled as contaminated
consonants. Gaussian noise, as indicated in equation (1).

17
 η0 ( n ) = η g ( n ) + b ( n ) η w ( n ) , (1) e(k) is then obtained by subtracting the digital filter output,
y(k), from the contaminated signal.
where ηg (n) and ηw (n) are zero-mean Gaussian noises and
b (n) is a Bernoulli random variable. The ratio of the + e(k)
d(k) = s(k)+n(k)
variances of the two Gaussian noises decides the impulsive
character of the noise generated. ∑
_
Adaptive Filtering
Adaptive Filter
Adaptive filtering can be considered as a process in x(k) y(k)
which the parameters used for the processing of signals
changes according to some criterion. Usually the criterion W(k)
is the estimated mean squared error or the correlation [1].
The adaptive filters are time-varying since their parameters
are continually changing in order to meet a performance
requirement. In this sense, an adaptive filter can be Fig. 2. Adaptive filter as a noise canceller
interpreted as a filter that performs the approximation step
on-line. Usually the definition of the performance criterion
requires the existence of a reference signal that is usually The filter output y(k) and the error e(k)[2] are given
hidden in the approximation step of fixed-filter design. by equations (2) and (3).
The general set up of adaptive filtering environment
[2,6] is shown in Fig. 1, where k is the iteration number, y ( k ) = X T ( k )W ( k ) , (2)
x(k) denotes the input signal, y(k) is the adaptive filter
output, and d(k) defines the desired signal. The error signal
e(k ) = d (k ) − y (k ) . (3)
e(k) is calculated as d(k)-y(k). The error is then used to
form a performance function or objective function that is
required by the adaptation algorithm in order to determine Adaptive AFA algorithm
the appropriate updating of the filter coefficients. The
minimization of the objective function implies that the In many applications of noise cancellation the
adaptive filter output signal is matching the desired signal changes in signal characteristics could be quite fast. This
in some sense. requires the utilization of adaptive algorithms, which
converge rapidly. From this point of view the best choice is
the recursive least squares (RLS) algorithm. Unfortunately
this algorithm has high computational complexity and
e(k)
d(k)
+ stability problems. The adaptive filtering with averaging
(AFA) [4] is used for noise cancellation in above
∑ conditions. The algorithm [4] described above is
_ summarized in equations (4)–(7):

Filter y ( k ) = W T ( k − 1) X ( k ) , (4)
x(k) y(k)
W(k)
e(k ) = d (k ) − y (k ) , (5)
Adaptive
Algorithm N
xe(k ) = ∑ x ( m − i ) e(m) . (6)
m =1

Fig. 1. General setup of adaptive filter For filter of N-th order

1
Wi ( k + 1) = Wi ( k ) + xei ( k ) , (7)
Adaptive Noise Canceller
kγ
where 0 < i < N and 1 2 < γ < 1 .
Adaptive filter is widely used as noise canceller. In an
adaptive noise canceller (Fig. 2) [3] two input signals, d(k)
and x(k), are applied simultaneously to the adaptive filter. The averaging here does not create additional burden
The signal d(k) is the contaminated signal containing both since the terms W (k ) and X (k )e(k ) can be recursively
the desired signal, s(k), and the noise n(k), assumed computed from their past values. Second, the algorithm
uncorrelated with each other. The signal, x(k), is a measure does not use the covariance matrix, so there is no need of
of the contaminating signal which is correlated in sole way covariance estimate. This implies low computational
with n(k), x(k) is processed by the digital filter to produce complexity and escape from stability issues compared to
an estimate y(k), of n(k). An estimate of the desired signal, the RLS [5] algorithm.

18
Proposed Modification in AFA algorithm White Gaussian noise was generated and added to the
original speech signal. The SNR of the signal corrupted
We propose a modification in the existing AFA with noise was 7.776 dB. A linear combination of the
algorithm to improve the performance in terms of signal to generated noise and the original signal is used as the
noise ratio. By introducing a variance factor in the AFA primary input for the filter. Fig. 4 shows the original
algorithm, the performance of the algorithm can be speech corrupted by white noise
improved. The modified update equation is shown in The denoised speech signal using the original AFA
equation (9): algorithm is shown in Fig. 5. The output Signal to noise
1 1 ratio of the signal denoised with original AFA algorithm
( g
X (k)e(k)) var( g X (k)e(k)) was 28.5101 dB.
W (k + 1) = W (k) + k + k . (9)
2 2
The proposed modification was simulated and tested
for noise cancellation in speech signal corrupted by white
noise. The simulation results are shown in the following
section.

Simulation and results

The parameters of clean speech sample considered for

testing of the algorithms were: duration 2 seconds, PCM
22.050 kHz, 8 bit mono sample recorded under laboratory
conditions. Fig. 5. Denoised signal using original AFA algorithm

Fig. 6 shows the signal denoised with proposed

modification included in the AFA algorithm. The figure
shows the output of modified AFA algorithm. The output
signal to noise ratio of the signal was 30.4568 dB as
compared to 28.5101 dB of the original AFA algorithm.

Fig. 3. Original speech signal

The filter order was fixed at 12 for all cases of noise

cancellation in speech .The recorded sentence “A quick
brown fox jumps over the lazy dog” was used as the clean
speech. This sentence is conventionally used as a
benchmark for speech processing. The above sentence Fig. 6. Denoised signal using modified AFA algorithm
contains all the alphabets of the English language. Hence
the variability of effect of noise on speech with frequency Table 1 shows the output signal to noise ratio of the
of the signal is accounted. The original speech signal is denoised speech signal using the original and modified
shown in Fig. 3. AFA algorithms.

Table 1. Output SNR of denoised signal

Algorithm Output SNR in dB
AFA 28.5101
Modified AFA 30.4568

Conclusions

The proposed modification in the existing adaptive

filtering with averaging algorithm was simulated and tested
for noise cancellation in speech signals. When tested with
white noise, the modified AFA algorithm showed an
Fig. 4. Speech signal corrupted by white noise

19
improved performance of around 2 dB compared to the 4. Georgi Iliev, Nikola Kasabov. Adaptive filtering with
original AFA algorithm. For future work we planned to averaging in noise cancellation for voice and speech
test this modified algorithm for colored and impulsive recognition. – Department of Information Science, University
of Otago, 2001.
noise. The study and comparison of adaptive algorithms
5. Diniz P. Adaptive Filtering Algorithms and Implementation
can be extended to the use of multi-dimensional adaptive Issues. – Kluwer Academic Publishers, USA, 2002.
filtering techniques; for applications like noise cancellation 6. Xiao Hu, Ai-qun Hu, Li Zhao. A Robust Adaptive Speech
in images. Further, the modified algorithms proposed can Enhancement System // IEEE Int. Conference on Neural
be optimized to have lower complexity. Networks and Signal processing. – Nanjing, China, December
14-17, 2003.
References 7. Ikeda S., Sugiyama A. An adaptive noise canceller with low
signal distortion for speech codecs // IEEE Trans. Signal
1. Widrow Bernard, Samuel Stearns D. Adaptive Signal Processing. – Vol. 47, Mar. 1999. – P. 665–674.
Processing. – Pearson Education, Delhi, 2004. 8. Julie E. Greenberg. Modified LMS Algorithms for Speech
2. Monson Hayes H. Statistical Digital Signal Processing and Processing with an Adaptive Noise Canceller // IEEE Trans.
Modelling. – John Wiley & Sons Inc, Kundli, 2002. On Speech and Audio Processing. – Vol. 6, No. 4, July 1998.
3. Emmanuel Ifeachor C., Jervis Barrie W. Digital Signal
Processing – A practical approach. – Pearson Education, Submitted for publication 2006 12 10
Delhi, 2004.

V. R. Vijaykumar, P. T. Vanathi, P. Kanagasabapathy. Modified Adaptive Filtering Algorithm for Noise Cancellation in Speech
Signals // Electronics and Electrical Engineering. – Kaunas: Technologija, 2007. – No. 2(74). – P. 17–20.
Adaptive filtering techniques are one of the important techniques used for noise cancellation in speech and biomedical signals. The
Least Mean Squares (LMS) algorithm is one of the widely used algorithms in many adaptive signal processing environments. The
adaptive filtering algorithm with averaging (AFA) algorithm is an improvement over the widely used Least Mean Squares (LMS)
algorithm and has an improved performance. In this paper, we propose a modification in the AFA algorithm with improved
performance for speech signal processing. The proposed modification was implemented in Matlab and was tested for noise cancellation
in speech signals. The simulation results showed that modification has improved performance in terms of signal-to-noise ratio compared
to the original adaptive filtering algorithm. Ill. 6, bibl. 8 (in English; summaries in English, Russian and Lithuanian).

В. Р. Вияйкумар, П. Т. Ванати, П. Канагасабапати. Модифицированный адаптивный алгоритм фильтрования для

удаления шума в сигналах речи // Электроника и электротехника. – Каунас: Технология, 2007. – № 2(74). – C. 17–20.
Методы адаптивной фильтрации являются одними из самых популярных при устранении шумов из речи и биомедицинских
сигналов. Алгоритм наименьших квадратов широко используется во многих сферах обработки адаптивного сигнала. Алгоритм
фильтрации адаптивного сигнала с применением усреднения (AFA) является более преимущественным по сравнению с
методом наименьших квадратов. Предложена модификация (AFA) алгоритма, позволяющая более эффективно обработать
речевые сигналы. Модификация внедрена применяя программу Matlab и тестирована устраняя шумы из речевых сигналов.
Результаты моделирования показали, что при модификации алгоритма его эффективность по сравнению с оригинальным
адаптивным алгоритмом фильтрации в соотношении сигнал-шум, улучшилась. Ил. 6 , библ. 8 (на английском языке; рефераты
на английском, русском и литовском яз.).

V. R. Vijaykumar, P. T. Vanathi, P. Kanagasabapathy. Modifikuotas adaptyvus filtravimo algoritmas triukšmams iš kalbos

signalų pašalinti // Elektronika ir elektrotechnika. – Kaunas: Technologija, 2007. – Nr. 2(74). – P. 17–20.
Adaptyvaus filtravimo metodai yra vieni iš dažniausiai taikomų triukšmui iš kalbos ir biomedicininių signalų šalinti. Mažiausių
kvadratų algoritmas yra plačiai naudojamas daugelyje adaptyvaus signalo apdorojimo sričių. Adaptyvaus signalo filtravimo, naudojant
vidurkinimą, algoritmas (AFA) yra pranašesnis, palyginti su mažiausių kvadratų metodu. Siūloma AFA algoritmo modifikacija,
leidžianti efektyviau apdoroti kalbos signalą. Pasiūlytoji modifikacija įgyvendinta naudojant Matlab programą ir testuota šalinant
triukšmą iš kalbos signalų. Modeliavimo rezultatai parodė, jog modifikuoto algoritmo našumas, lyginant su originaliu adaptyviu
filtravimo algoritmu pagal signalo ir triukšmo santykį, pagerėjo. Il. 6, bibl. 8 (anglų kalba; santraukos anglų, rusų ir lietuvių k.).

20