Denoising real-time audio signals using matlab filter

techniques
Jayashree K C Karthik M Bharamappa Manoj Kumar P
School of ECE School of ECE School of ECE
REVA UNIVERSITY REVA UNIVERSITY REVA UNIVERSITY
Bangalore , India Bangalore , India Bangalore , India
[email protected] [email protected] [email protected]

Mahaa Santeep G Sugandha Saxena

School of ECE School of ECE
REVA UNIVERSITY REVA UNIVERSITY
Bangalore , India Bangalore , India
[email protected] [email protected]

Abstract— In this paper we analyze real time audio signals and creates major problem to the audio signals. Electrostatic
try to reduce the noise associated with the message signal under noise which is generated due to the presence of the voltages
consideration. The main drawback of noise being present in an during the design implementation and other random noises
audio signal, is that it reduces the quality of the signal that is which gets added to the signal. This proposed research will
being transmitted within the communication system. For
solve the drawbacks of various filtering techniques which
analysis purpose, white gaussian noise(awgn) is concatenated
with the audio signal under consideration and the resulting
also provides unique Knowledge to the reader. The project
noisy audio signal is subjected to the different filtering defines a comparison between three different filtering
techniques like IIR Filter, FIR Filter, Wavelet transform techniques i.e. IIR filtering, FIR filtering and wavelets
techniques. The noisy audio signal is analyzed with respect to transforms based on real time audio signals.
the different filter responses obtained on applying the foresaid
methods. A comparative study is done between these
techniques to arrive at a technique which would be the most II. BACKGROUND AND RELATED WORK
efficient one for audio signal denoising. Graps[1] came up with new analysis named Fourier
Keywords-Audio processing, Denoising, FIR filters, IIR
transform which could analyze the periodic function by
filters, wavelets, Daubechies. creating mathematical structures that vary in scale. But the
proper analysis cannot be analyzed using frequency
I. INTRODUCTION response.
Radhika Bhagat[2] has made an attempt of audio filtering
The importance of noise reduction in real-time audio signals
using extended filters like lowpass and highpass filter using
is said to be having high significance in communication.
FIR and IIR filtering techniques. They have designed
Due to this factor the noise weakens signal quality, and the
different formulas and difference equations for efficient
recognition of audio signals becomes difficult and cause
implementation of time varying filter applications.
serious difficulties for the users of electronic hearing aids. A
Er. Harpal Singh[3] has used fast Fourier transform
well-established method is filtering of the signal in the
technique for performing time domain and frequency
frequency domain or in the simplest way is analyzing the
domain analysis of the signal. Mannu Singla[4] has used
signal using different filter techniques like low pass, high
Butterworth filter and Chebyshev filters to reduce noise
pass and bandpass filters. As the voice or speech signals are
from signals with different frequency and ripple factors.
not periodic these filters distort the signal more than they
Seema rani[5] has proved more facts about FIR and IIR
reduce the noise. To attenuate noise, we need more
filters in their paper. The paper tells FIR is more stable than
advanced methods of filtering. As it requires high advanced
IIR. From the above research we can conclude that the error
methods of filtering, we go for the different filtering
of FIR filter is less compared to IIR filters that means the
techniques like Fourier transforms, wavelet transforms, and
output of FIR filter is very close to the desired value and
other methods are used in denoising the audio signals. A
FIR filter is more stable than IIR filters.
Fourier transform of a signal gives us the frequency
C Mohan Rao[6] has presented a new algorithm that is the
composition of the audio signal. The disadvantage of
Least Mean Square (LMS) in which the awgn is added to
Fourier transform is it is only valid within a certain Region
message signal and the denoising is done to reduce the noise
of Convergence (ROC). So, we go for short time Fourier
with minimum or no error efficiently. But it is sensitive to
transform (STFT), but this method used the window analysis
the scaling of its input.
approach of defined size. IIR and FIR algorithms uses the
B. Jai Shankar[7] has proposed the use wavelet
Fast Fourier transform (FFT) technique for analysis of
transformation technique to denoise audio signals by
frequency spectra and signal responses. Wavelet analysis
dividing the signal into blocks. This technique protects each
provide more detailed analysis about the signal compared to
and every unique and vital features of every individual block
other filtering approaches.
and exposes the finest detail contributed by the grouped set
This paper tells the readers about denoising of the audio
of blocks. The authors Ola Ratelli, Palle, Jorgensen (2013)
signal. As all the audio signals are continuously affected by
[8] in their book have proposed the about Discrete wavelet
the different types of real-time noise such as electrostatic
transform, its benefits and its functionalities. Priyanka
noise, thermal noise, channel noise, awng and etc. which

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

Khattar[9] have published a paper in which denoising will Where x(n) is the input signal, y(n) is the output signal
be performed using wavelet transformation by comparing
with filter coefficients a and b.
two wavelets families, Daubechies and Haar.
III. PROPOSED WORK
The flow of the project would be as shown in figure 1.

IIR Input Noisy IIR Filter

filter audio signal filter Response
signal
Input Noisy Filter
audio signal FIR Response
signal filter
Additive
wavelets White
Additive Gaussian
White noise
Gaussian
noise
Figure 2.IIR Filtering method block diagram
Figure.1 Block diagram of project C. FIR Filter Algorithm
Using basics of noise theory and audio theory, analysis of The audio signal which is to be analyzed is taken as input, to
the audio signal is carried out. The audio signal under this additive white gaussian noise is added and the resulting
consideration is concatenated with awgn. Using FFT, this noisy audio signal is obtained. White gaussian noise is
noisy signal is converted from time domain to frequency preferred as it has almost constant PSD (power spectral
domain is then analyzed in frequency domain by converting density) and for easy and precise analysis.
from time to frequency domain using Fast Fourier Now the frequency domain plot of the noisy signal is
Transform algorithm (FFT). The following step would obtained by Fast Fourier method and the signal is analyzed
involve designing of the different filters taking into with respect to the peak points to decide the cut-off
consideration the different parameters involved in filter frequency. Once the cut-off frequency is determined based
design like order of the system, cutoff frequency. The on the analysis, the filter design process is initiated by
filtering approach consists of normalization of signal, passing the essential parameters like the order, normalized
decomposition technique and reconstruction technique cut-off frequency, filter type (low pass filter) and
which we use in the filtering process to achieve our appropriate windowing technique is selected. The filter
objectives. coefficients (the numerator and denominator coefficients)
are obtained the magnitude plot of the filter is plotted and
A. IIR Filter Algorithm analyzed. Now that the filter is designed the normalized
The audio signal which is to be analyzed is taken as input, to noisy signal is passed to low pass filter to obtain the filtered
this awgn is concatenated with audio signal and the resulting signal. The filtered signal is finally plotted and used for
noisy audio signal is obtained. White gaussian noise(awgn) analysis and comparison with the other results obtained.
is preferred as it has almost constant PSD (power spectral
density) and for easy analysis of the audio signal.
D. Equations
Now the frequency domain plot of the noisy signal is
obtained by Fast Fourier Transform method and the signal is Difference Equation of FIR filter:
analyzed to decide the cut-off frequency. Once the cut-off N−1
frequency is determined based on the analysis, the filter y ( n )= ∑ h (k )x (n−k ) (2)
design process is initiated by passing the essential k=0
parameters and low pass filter is built.
The filter coefficients are obtained and the magnitude plot of
Hamming Window function:
the filter is plotted and analyzed. Now that the filter is
designed the normalized noisy signal is passed to low pass
2 πn
filter to obtain the filtered signal. The filtered signal is w ( n )=0.54−0.46 cos ⁡{ } 0 ≤ n ≤ N −1
finally plotted and used for analysis and comparison with
n−1
(3)
the other results obtained.
B. Equations
Difference Equation of IIR Filter:
Input Noisy FIR Filter
N M audio signal filter Response
y ( n )=−∑ ak ( n−k ) +¿ ∑ b k x ( n−k ) ¿ (1) signal
k=1 k=1

Additive
White
Gaussian
noise
 Figure 3. FIR Filtering method block diagram

E. Wavelets Figure.4b Block diagram of DWT decomposition technique

Wavelet transform method can be employed for analysis of
the audio signal with respect to approximation and detailed
coefficients. Wavelet transforms can be classified as,
continuous wavelet transform and discrete wavelet
transform. We are using discrete wavelet transform
technique in our project as it is more suitable for denoising T Approximate
of audio signals. The audio signal which is to be analyzed is H coefficients
taken as input, to this awgn is added and the resulting noisy R A0, A1…. An
audio signal is obtained. White gaussian noise is preferred as E
S
it has almost constant PSD (power spectral density) and for H Reconstructed
easy and precise analysis. This signal is passed through Concatenation
O
Level filter. L audio signal
The noisy signal is decomposed into two parts, detailed D Detailed
coefficients and approximation coefficients. The number of S
coefficients
levels required for decomposition generally depends upon / D0, D1… Dn
nature of the signal. H
Multilevel decomposition is done to repeat the process of
decomposition so that many lower resolution components of Figure.4b Block diagram of DWT reconstruction technique
the signal can be obtained through wavelet decomposition
trees. Wavelet thresholding
The final step is to reconstruct the original audio signal IV. RESULTS AND WAVEFORMS
without much loss of information. The audio signal is imported to MATLAB and plotted as
A construction process which involves using the wavelet shown in figure.5
coefficients and considering the levels of iteration,
successful reconstruction of the original audio signal is
obtained.

DWT has two functions wavelet and scaling function

N−1
Scaling function ∅ (t )= ∑ h [n]√ 2 ∅ (2 t−n) (4)
n=0

N−1
Wavelet function φ ( t )= ∑ g[n] √ 2∅ (2 t−n) (5)
n=0

Approximation coefficients:

1
W ∅ [ j0 , k ]= ∑ f [ n ] ∅ j , k [n] (6)
√M n 0
Figure.5 – Audio signal.

This audio signal is added with additive white gaussian

Detailed coefficients: noise and plotted as shown in figure.6

1
W φ [ j , k ]= ∑ f [ n ] φ j , k [n]
√M n
j≥j0 (7)

T
Noisy Approx. H
Input Signal Low Coeffici R
Audio Pass ents E
A0, A1, S
L-Filter …An H
Additive
White O
Gaussian High Detailed L
Noise Pass coefficie D
nts
D0, D1… S
Dn /
H
 Figure.7b pole/zero plot of IIR filter
Figure.6 Noisy audio signal
The output of IIR lowpass filter is as shown in figure 8.

The noisy signal is passed through IIR low pass filter having
magnitude & phase response is as shown in figure.7a.1 &
figure.7a.2 respectively and pole & zeroes plot is as shown
in figure.7b.

Figure.7a.1 Magnitude and phase response of IIR filter Figure8.IIR low pass filter output (denoised output)

The same noisy signal is passed through the FIR filter

having magnitude & phase response as shown in figure.9a.1
& 9a.2 respectively and pole & zeroes plot as shown in
figure.9b.

Figure.7a.2 Phase response of IIR filter

Figure.9a.1 Magnitude response of FIR lowpass filter

 Figure.9a.2 Phase response of FIR lowpass filter

Figure.11 Denoised signal from wavelet transform method

V. DISCUSSION
An analog system is said to be stable if all its poles lie in the
left half of the ‘s’ plane, but IIR filter has a both poles and
zeros as we can see in figure 7b. So IIR is said to be
unstable. IIR filters are difficult to implement as it has
delays and distortions due to large number of poles. IIR
filters are better in lower orders as IIR filters may become
unstable with higher orders. As IIR filters are unstable they
cannot have a linear phase as shown in the phase plot in
figure.7a. Theoretically we know that IIR filters depends on
present input and previous output. By the waveforms in
figure.8, we observe that noise introduced is denoised to
Figure.9b pole/zero plot of FIR filter certain limit but the output audio signal will also get
damped.
The output of FIR lowpass filter is as shown in figure.10 Due to these drawbacks we would prefer FIR over IIR as
FIR filters do not have delays and distortions. Therefore,
FIR filters are stable than IIR as FIR filters have only zeros
on the unit circle in the s plane and one pole at origin as
shown in figure.9b. FIR filters are better for higher orders
and they maintain stability. FIR filters have linear phase
characteristics as shown in phase plot of figure.9a and it’s
output depends only on the present inputs. In order to avoid
damping of the output audio signal we have tried various
FIR filters using different windowing techniques like
hamming, Kaiser and rectangular. These windows were
compared to get the Filtered audio output but Hamming
window was preferred as it had linear phase. FIR filters had
denoised to limit much greater than IIR but the audio signal
was also damped.
Due to the drawbacks we would prefer wavelets over FIR to
remove noise and get the filtered signal. The wavelets used
here is Daubechies wavelets have highest number N of
Figure.10 Output of FIR lowpass filter vanishing moments with the support width 2N-1. db wave
solves the problem of signal discontinuities and is applicable
When the same noisy signal is denoised using wavelets for continuous and discrete wavelet transforms.
transform technique, the output is as plotted in figure.11
 Process: n Efficient Audio Denoising Technique” , European Journal
of
Scientific Research, Vol.48 No.1, pp.16-28. ISSN 1450-216X.
[8] Ola Ratelli , Palle Jorgensen , ” Wavelets through a looking glass :
The world of the spectrum ” , Springer science & Business media ,
Illustrator : B. Treadway , 2013.
[9] Priya Khattar1, Dr. Amrita Rai2, Mr. Subodh Tripathi, “Audio De-
noising using Wavelet Transform”, 2016
[10] Adri E. Villanueva- Luna, Alberto Jaramillo-Nunez, Daniel Sanchez-
Lucero, Carlos M. Ortiz-Lima, “De-Noising Audio Signals Using
MATLAB Wavelets Toolbox”, Mexico.

Figure.12 Scaling function-phi and Wavelet function psi of db3 wavelet

db wave belongs to orthogonal family and it has daughter

wavelets like db1, db2 till db45. db wavelet removes noise
to get filtered output. The figure.12 shows the Scaling
function phi and Wavelet function psi of db3 wavelet. Low
pass filters are used to avoid aliasing effect and is applicable
in communication circuits as anti-aliasing filters. Bandpass
filters are used to pass only certain range of frequencies
which is applicable in wireless transmitters and receivers in
order to avoid noise.
VI. FUTURE SCOPE
We have analyzed the audio signal by three filtering
techniques that is IIR, FIR and wavelets which suppresses
the noise using MATLAB functions. In some cases, the
sound signal which we consider as noise may also contain
some important information. For example, in military,
satellite applications minute information present in the
sound signal from the base stations plays a major role. So,
the future scope of this project is to not suppress the noise
but segregate it from the audio signal and play it as store and
play it as another signal. This project can also be used to
separate the signal which have undergone stereo mixing. For
example, in a song we can separate the male voice, female
voice and background music and play all the three as
separate audio files. In this way this algorithm can be used
for generating karaoke from any given song.

REFERENCES
[1] A. GRAPS, “AN INTRODUCTION TO WAVELETS” ,22724 MAJESTIC OAK
WAY, CUPERTINO, CA, USA,1995.
[2] Radhika Bhagat, Ramandeep Kaur, “Improved Audio Filtering Using
Extended High Pass Filters”,2013
[3] Er. Mannu Singla, Er. Harpal Singh, “Review Paper on Frequency
Based Audio Noise Reduction Using Different Filters”, International
Journal of Science, 2015.
[4]Er Mannu Singla, Er Harpal Singh, “Paper on frequency based audio
noise reduction using Butterworth, Chebyshev and Elliptical filters”,
International Journal of Science, 2015.
[5] Seema rani, Amanpreet Kaur, J S Ubhi, “Comparative study of FIR and
IIR filters for the removal of Baseline noises from ECG signal”,
International Journal of Computer Science and Information Technologies,
Vol. 2 (3), 2011, 1105-1108.
[6] C Mohan Rao, Dr. B Stephen Charles, Dr. M N Giri Prasad, “A
Variation of LMS Algorithm for Noise Cancellation”, International Journal
of Advanced Research in Computer and Communication Engineering Vol.
2, Issue 7, July 2013.
[7] Shanar B. J & Duraiswamy K. (2010). “Wavelet-Based Block
Matching