Speech Enhancement Using
Recursive Least Square Based on
Real-time adaptive filtering algorithm
Paper ID : 247
[Link] Venkateswarlu,
Institute of Aeronautical Engineering,
Hyderabad,Telangana State, INDIA.
Insert Author
Headshot
Image Here
Contents
Abstract
Introduction
Existing Method
Proposed Method
Methodology
Adaptive Algorithm
Recursive Least Square(RLC) Algorithm
Simulation Results
Conclusion and Future Scope
References
Questions
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Abstract
In real time environment a speech signal is often corrupted
and losses its characteristics either by natural disturbances
or anything.
The main aim of our proposed algorithm is to increase the
speech intelligibility and quality.
In order to do that a filter has been developed using
Recursive Least Squares (RLS) algorithms and Least
Mean Square (LMS).
Real-time adaptive filtering algorithms are one of the best
methods used for the speech enhancement methods.
In this research work we have proposed the recursive least
square which is under adaptive filtering method for the
enhancement of the speech signal.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Abstract
Initially we compare the performance of noise cancellation of the
proposed Recursive least square which uses objective evaluations that
is based on wavelet based speech enhancement like Signal to noise
ratio Loss,
Signal to Noise ratio and Mean Squared Error.
Based on the Objective and Subjective evaluation, it was found that this
algorithm clearly in increases the intelligibility and removes the
corrupted noise in the waveforms.
There are different types of filters like Kalman filter,
Wiener filter,
Spectral subtraction,
notch filter and wavelet methods.
The performance of every filter depends upon the intelligibility and
quality of the speech signal.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Abstract
The reduction or augmentation in the SNR ratio is
the main aim of the most methods.
Adaptive filtering is a technique which uses
certain predefined criterion like the estimated
mean squared error or the correlation has to be
considered for the analyses of the waveform.
In this adaptive filter we use weighted
coefficients and it is updated itself by an adaptive
algorithm.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Introduction
The main aim of all algorithms in the signal processing
is to improve the intelligibility and quality of the speech
signal without affecting the quality of the signal.
In order to achieve better noise cancellation estimation
of the noise in a particular signal is much important.
Removing noise is a major issue in recent decades
because it is very difficult to predict the spectral
properties of non-stationary noise.
In the process of enhancing the speech the noise
estimation is the issue which is complicated
if the noise power is more than that of the speech
power ,then the speech content is removed and treated
as noise.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Introduction
Speech processing has spread its wings almost in many
applications in telecommunication systems, medical equipment
signal processing, hearing devices ATM processing it is most
complicated and interested research area.
To improve the noisy speech many different algorithms have been
developed by many researches, but still it is a challenge area as
the features of noise varies depending upon the time and noise.
We have many enhancement algorithms which are proposed by
researches like Kalman filter, Wiener filter, Spectral subtraction and
signal subspace and wavelet methods.
Spectral subtraction is the initial approach for a signal corrupted by
noise.
It can be used in magnifying speech quality which are interrupted
by the environmental issues or any other, but it is effected by
musical noise but in silence period it does not remove noise.
From the noisy signal magnitude it undergoes subtraction of the
magnitude spectrum then approximating the spectrum of the
original signal.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Introduction
Researcher’s made many changes to the kalman filter, but did not reach
the expected point and also computational characteristics are more.
The paper has solved problem of colored noises and de-noising the
random in addition to coefficient estimation.
We considered the colored noise is also an autoregressive process.
So we found that its AR coefficients of the signal and variance which is
determined by prediction using linear estimation in similar way.
In our observation, to solve the stated problem Kalman filter based method
is used with preprocessing called digital expander of noisy speech signal
which is proposed in order to reconstruct the signal and the addictive
noise is modeled as the auto regressive (AR) process.
The auto regressive (AR) speech model parameters which are estimated
by varying are based on linear prediction coefficient estimation (LPC).
In this paper we found coefficient estimation and also de-noising of
coloured signal.
The noise is also an auto regressive process.
Its variants and coefficients by LPC are estimated in the similar way.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Existing Method
There are different methods individually and also combined with other
filters to study and improve the performance of speech signal.
In noisy environment the discrete wavelet transform combined with the
Kalman filter used for signal speech enhancement.
On fixed point digital signal processing (DSP) is implemented to improve
the quality for Texas instruments and intelligibility of the signal.
To filter the noisy speech signal we apply Kalman filter. And then the
discrete wavelet transform is conducted to the filtered frames.
Then, to track the variation of noisy signal, the noisy discrete wavelet
coefficients is truncated for soft thresholding according to the input noise
level, and then to get the enhanced signal we apply inverse discrete
wavelet coefficient.
To reduce the white and the colored noise we proposed speech
enhancement is adopted properly from the noisy speech.
The TIMIT database is applied with proposed method of simulation results
are compared with the discrete wavelet transform approach.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Existing Method
Wavelet based methods for de-noising of speech signals are widely used in speech signals
which are corrupted with background noise.
A filter is eliminates noise by low pass filtering which distorts the noisy features in the speech
signal, the methods which uses wavelets for large variety of signals are proves to providing
good performance.
Wavelet transform which is based on speech enhancement is applied in many applications.
There are number of issues for a successful noise suppression application to be treated for
the real time speech signal corrupted by additive noise to be successful.
The DWT thersholding is previously approached for speech enhancements which are
corrupted by Gaussian white noise of real time speech signals.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Proposed Method
Noise cancellation can be used by Fixed or Adaptive filters.
The prior information of noise and signal for creation of fixed filters. i.e.
To design a filter we have to know the signal and noise which passes
frequencies in the signal and which rejects the frequency band
occupied by the noise.
Then, impulse response automatically adjusted by the Adaptive
filters, and the little or no knowledge of signal or noise characteristics
for their design.
The main aim is to separate the noise from a signal of an adaptive
filter and to improve the signal to noise ratio.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Proposed Method
There are two inputs for Adaptive Noise Canceller(ANC) one is primary input and
reference input. The primary input which is corrupted by noise receives a signal from
the signal source which is uncorrelated with signal.
The reference input which is uncorrelated with signal but correlated with noise
receives a noise.
To produce the close estimation input signal the noise is passed through the adaptive
filter i.e. y(n).
To produce the estimation of error e(n), the estimated noise is subtracted from the
corrupted signal. Over many years the adaptive filters from the researchers have
gained attention. As a result, various algorithms have been developed which are
computationally efficient.
y(t) = s(t) + n(t) -- (1)
S= W(Y) -------------------------------2
Z= D(Y, S)--------------------------------3
S= W-1----------------------------------4
Ψa,b(t)= [(t-b)/a] a, b ]-----------------------------------------------5
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Methodology
The proposed method is the RLS algorithm for the effective noise
cancellation and block diagram is as shown below.
The details of every step as shown
RLS method acts as the best solution
for the mean square error solutions.
Here, the segments of the
decomposition are applied to
every frame of the DWT and then
segmented to obtain the noise signal.
After that we will obtain the mother wavelet as variance of the minimizing
error and maximizing the signal to noise ratio in between the original and
the reconstructed signal.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Methodology
The proposed method is the RLS algorithm for the effective
noise cancellation and block diagram is as shown below.
The details of every step have been explained in below.
RLS method acts as the best solution for the mean square
error solutions.
Here, the segments of the decomposition are applied to every
frame of the DWT and then segmented to obtain the noise
signal.
After that we will obtain the mother wavelet as variance of the
minimizing error and maximizing the signal to noise ratio in
between the original and the reconstructed signal.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Adaptive Algorithm
In adaptive noise cancellation there are basically two types of algorithms (i) Lease mean square(LMS)
(ii)Recursive least square
Generally the least mean square algorithm is one of the most basically used algorithm because of its
error coefficients and minimum mean square error.
Let us consider the output be y(n). FIR filter can be calculated from Eq.
Where N is the order of filter used and n is the no. of iterations n=1,2,…. The output error signal is
calculated using
Here we use current weight value w(n) and obtain the required value
Error value can be estimated by the tap weight and eq(8) right hand side is the tap adjustment that is a[pplied to w(n).
Error value can be estimated by the tap weight and eq(8) right hand side is the tap adjustment that is a[pplied to w(n).
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Adaptive Algorithm
The RLS algorithm has the high performance and computational
complexity and it also has some stability issues in spite of all these it
can be considered as one of the finest algorithm for the noise
cancellation. Here the filter tap weight vector is updated and then
given by Eq(11).
Where λ is a small positive constant, but smaller than 1. The filter tap
weights are calculated by the vector of previous iterations and also
the current input vector as well. This is shown by the Eq(14) and the
error signal is given by the
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Adaptive Algorithm
Where λ is a small positive constant, but smaller than 1.
The filter tap weights are calculated by the vector of previous
iterations and also the current input vector as well.
This is shown by the Eq(14) and the error signal is given by the
Eq(15)
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Recursive Least Square(RLC)
Algorithm
There are certain situations in speech enhancement that the signal and its characteristics
change rapidly and cannot be estimated then in that situation the algorithm used has to
be very active in predicting the changes and in that case recursive least square algorithm
has considered to be one of the fastest algorithm.
The Proposed algorithm repeatedly finds out a coefficient which reduces the cost
function of the linear least squares.
The main advantage of the RLS is that it has got the ability to adjust its coefficients
automatically without our intervention.
This process performs an exact minimization of the sum of squares of desired signal.
Figure: Design of RLS Algorithm
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Simulation Results
We tabulate the values of both error signal and the noisy
signal by selecting the type of signal at the top right.
Here we have selected noisy signal and have made some
settings in the cursors measurements and
signal statistics here we note the observations which are
displayed at the right side or by placing the cursor on the signal.
Signal statistics Bi-level measurements
Measurement Value Measurement Value
Max 1.781e+00 High 9.685e-01
min -1.534e+00 Low -9.542e-01
Peak to peak 3.315e+00 Amplitude 1.923e+00
Mean -3.812e=0.3 +Edges 9
+Rise time 25.505 s
+Slew rate 61.330(/ks)
Median -2.040e-02 -Edges 10
-Falltime 23.232 s
-Slew rate -71.021(/ks)
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Simulation Results
Figure 4: Noisy signal analysis of RLS Adaptive filter
Figure 5: Noisy signal analysis of RLS Adaptive filter
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Simulation Results
Figure: Analysis of RLS Adaptive filter for performance Analysis
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Simulation Results
TABLE : SNR LOSS SCORE
SNR Loss Degree of SNR Loss Expected improvement with
directional microphones
0-3 dB Normal/near normal May hear better thannormal
3-7 dB Mild SNR loss hear in noise
May hear almost as wellas
7-15 dB Moderate SNR loss normal hear in noise
>15 dB Severe SNR loss Directional microphones help.
Consider array microphones
Maximum SNR improvement is
needed.
Tableof: Noise
Type Nlms and RLS Algorithms Observations
SNR(DB)
Pink
Noise -5 0 5 10
MSE 0.0023 9.7817x10-4 5.4487x10-4 3.4010x10-4
SNR 7.6840 10.9904 13.3827 15.3756
SNR 0.6932 0.6368 0.5873 0.5428
LOSS
NLMS
MSE 2.6419x10-4 2.6378x10-4 2.6350x10-4 2.6310x10-4
SNR 16.3881 16.3946 16.3990 16.4054
RLS SNR 0.4475 0.4535 0.4533 0.4522
LOSS
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Simulation Results
Performance Results of NMLS and RLS algorithm based on
MSE
Figure: Performance Results of NMLS and RLS algorithm based on SNR
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
Conclusion
In our project we have studied and performed a new algorithm based on
Recursive least square adaptive filtering method.
This method removes the noise levels in a particular signal and helps us to
analyses the characteristics of the signal easily.
The speech signal intelligibility and quality of the signal is increased without
affecting the quality of the signal.
The performance of the system has been increased when compared to all other
algorithms.
The mean square error (MSE), SNR ratio etc., have been used for the
experimental analysis.
We know that as the SNR increases the output quality of the image or speech is
also increased.
So using MATLAB code we have performed simulations of various waveforms
observed the values and calculated the MSE and SNR.
By the results we have understood that the proposed method is effective in
reduction of noise.
Finally, in the real-time test of speech de-noising the proposed method has been
successfully implemented and reveals that the proposed algorithm has
significantly improved the speech intelligibility.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
References
[1] Shabeera Bee S, Pramod JT, Jilani SAK. Real time speech denoising using Simulink
and beagle bone black. International Journal of Core Research in Communication
Engineering. 2015 April;2 Texas-instruments. TMS320DSP/BIOS.5 (42): User‟Guide,
spru423I, Août 2012.
[2] Ramesh Babu G. Modified Kalman Filter-based Approach in Comparison with
Traditional Speech Enhancement Algorithms from Adverse Noisy Environments.
International Journal on Computer Science and Engineering (IJCSE). 2011 Feb; 3(2).
[3] Almajai, Milner B. Visually derived wiener filters for speech enhancement. IEEE Trans.
Audio, Speech, and Language Processing. 2011 Aug; 19( 6): 1642–51.
[4] Jianfen Maa and Philipos C. Loizou, “SNR loss: A new objective measure for predicting
the intelligibility of noise-suppressed speech,” Speech Communication, Vol. 53, No. 3,
March 2011, pp.340–354.
[5] Komal R. Borisagar and Dr. [Link], “Simulation and Comparative Analysis of LMS
and RLS Algorithms Using Real Time Speech Input Signal,” Global Journal of Researches
in Engineering, Vol. 10, No. 5, October 2010, pp. 44-47.
[6] P. Krishnamoorthy and S.R. Mahadeva Prasanna, “Temporal and Spectral Processing
Methods for Processing of Degraded Speech: A Review,” IETE Technical Review, Vol. 26
, No. 2, March2009.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
References
[7] Raj Kumar Thenua and S.K. Agarwal, “Simulation and performance analysis of adaptive filter in noise
cancellation,” International Journal ofEngineering Science and Technology, Vol. 2, No. 9, 2010, pp.
4373-4378.
[8] Ben Jebara S. A perceptual approach to reduce musical noise phenomenon with wiener denoising
technique. IEEE Int. Conf. Acoustics, Speech and Signal Processing, (ICASSP).2006 May; 3:3.
[9] Tabibian S, Akbari A, Nasersharif B. A new wavelet thresholding method for speech enhancement
based on symmetric kullback leiblerdivergence. in Proc. 14th Int. CSI Computer Conference, CSICC);
2009 Oct.p.495–500.
[10] Claesson,Rossholm A. Notch filtering of humming GSM mobile telephone noise. in Proc [11] M. V.
Butz, K. Sastry, and D. E. Goldberg. Strong, stable, and reliable fitness pressure in XCS due to
tournament selection. Genetic Programming and Evolvable Machines, 6(1):53–77,2005.
[12] [Link] Venkateswarlu, [Link] Kumar, A Usha Sree , K. Chaitanya “Effect of feature normalization
objective improvement of over Noisy Single-channel Speech Enhancement with Neural Networks”,
TEST Engineering & Management (TESTEM), Volume- 83, Issue-1, May June 2020, pp. 8121-
8131,DOI: 10.35940/ijitee.A9110.119119.
[13] [Link] Venkateswarlu, [Link] Kumar, [Link] Kumar, Aannam Karthik, and [Link]
“Implementation of Area optimized Low power Multiplication and Accumulation”, International Journal
of Innovative Technology and Exploring Engineering (IJITEE) ISSN: 2278-3075, pp. 2928- 2932
,Volume-9, Issue-1, November 2019. Published By: Blue Eyes Intelligence Engineering & Sciences
Publication, Retrieval Number: A9110119119/2019©BEIESP DOI:10.35940/ijitee.A9110.119. .
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.
References
[14] [Link] Venkateswarlu, [Link] and .[Link] Kumar “Performance on Speech Enhancement in
Objective Quality Measures Using Hybrid Wavelet Thresholding”, International Journal of Engineering
and Advanced Technology (IJEAT), Volume 6, Issue 8(2019) ,pp.3523-3533 published on August
2019. DOI: 10.35940/ijeat.F9343.088619.
[15] [Link] Venkateswarlu, [Link] Rami Reddy and .[Link] Prasad “Speech Enhancement in Terms
of Objective Quality Measures with Effect of Adjustable Window Shape Parameters”, WJTER,
Volume 2, Issue 1(2017) 347-359 published on December 2017.
[16] [Link] Venkateswarlu, [Link] Rami Reddy and .[Link] Prasad “Speech Enhancement in Terms
of Objective Quality Measures with Effect of Adjustable Window Shape Parameter on Speech
Enhancement Techniques”, Cromejournals based –Journal of Signal and Image Research published
on July 2016.
[17] [Link] Venkateswarlu, [Link] Rami Reddy & [Link] Prasad, “Speech Enhancement in terms of
Objective Quality Measures Based on Wavelet Hybrid Thresholding the Multi-taper Spectrum”, published in
IJAREEIE, Volume5, Issue 1, pp.201-219, January 2016, DOI:10.15662/IJAREEIE.2015.0501036.
[18] Vinyl Ho Oquiño , S. China Venkateswarlu , Rogelio B. Paguntalan, “Development of Yoghurt Incubator for
Small-Scale Production using PIC16F877A Microcontroller”, published in Global Journal of Science Frontier
Research: Agriculture and Veterinary, publisher Global Journals Inc. (USA), , Volume 16, Issue 1, pp.11-18,
publication date.2016, Online ISSN: 2249-4626 & Print ISSN: 0975- 5896.
[19] [Link] Rami Reddy, [Link], [Link] Venkateswarlu, “Telugu Speech Enhancement in Terms of Objective
Quality Measures Using Discrete Wavelet Transform using Hybrid Thresholding”, published in
JAREEIE,Volume3,Issue8,August 2014, ISSN(online):2278- 8875,ISSN(Print):2320- 3765..
[20] [Link] Venkateswarlu, [Link] Rami Reddy,[Link] Prasad, “Speech Enhancement using Bolls Spectral Subtraction Method
Based on Gaussian Window" published in Global Journal of Researches in Engineering –GJRE Volume 14 Issue 6 Version
[Link] 2014.
[Link] Venkateswarlu Professor of ECE, Your logo here
IARE-Hyderabad.