International Journal of Ethics in Engineering & Management Education

Website: www.ijeee.in (ISSN: 2348-4748, Volume 1, Issue 3, March 2014)

Speech Signal Enhancement Using Wavelet Threshold Methods

P.Sunitha, Assoc. professor, ECE, Pragati Engineering College, A.P, India. [email protected]
Abstract This paper presents a Novel Approach for Enhancement of Speech Signal using different wavelet threshold methods. It is clearly observed that a signal which is obtained from the real world environment is often corrupted by means of noise. This noise effectively reduces the performance of the signal and it must be removed effectively for further processing of signal. This paper investigates the use of wavelet threshold methods i.e Minimax, SURE(Heuristic and Rigorous) and Square root-log are investigated in Enhancement of speech signal corrupted by White Gaussian noise along with performance analysis of wavelets i.e. Haar,Db10, Coif5,sym5 at different levels of decomposition and performance analysis of threshold methods. This paper also investigates the use of wavelet shrinkage methods in Enhancement of speech signal corrupted by White Gaussian at five different noise levels i.e 0dB, 5dB, 10dB, 15dB, 20dB along with performance analysis of wavelets. Further comparative analysis has been made between Daubechies, Haar, Symlet, Coiflet. Biorthognal to explore the optimal wavelet selection for Enhancement of speech signal in terms of Peak Signal to Noise Ratio (PSNR). It has been found that Daubechies wavelet higher PSNR is achieved than others. Index terms: Minimax, SURE (Heuristic and Rigorous) and Square root-log.
1

T.Sekhar, Asst. Professor, ECE ,Pragati Engineering College, A.P, India, [email protected]
parameter extraction process used in low bit rate vocoders and hence they are becoming an integral part of low-bit rate speech coding systems. Over the last three decades, many kinds of speech enhancement techniques adaptive filtering, modelbased methods. The transform-based techniques, transforms the time domain signal into other domains, suppress noise components, and apply the corresponding inverse transform to reconstruct enhanced speech signal. Discrete Fourier transformer (DFT), Discrete cosine transformer (DCT), Karhunen-Loeve transformer (KLT), and wavelet transformer (WT) are widely-known transformer methods. DFT based technique have been intensively investigated based on shorttime spectral amplitudes (STSA). A KLT-based technique, called signal subspace-based methods, decomposes the space into signal (or speech) and noise subspace by means of eigen decomposition, and then suppresses the noise component in the eigen values. DCT-based techniques are of lower computational complexity and higher frequency resolution than DFT-based methods. It is also possible to consider WTbased methods in order to simultaneously exploit the time and frequency characteristics of noisy speech signals. Adaptive filtering on the other hand, cancels the noise using adaptive filters such as the Kalman filter. A Kalman filter models noisy speech signals in terms of state space and observation equations, which present the speech production process and the noise addition model together with channel distortion, respectively. Kalman filters normally assume a white Gaussion noise distribution; however, Gibson et al. Proposed a generalization of Kalman-filtering over coloured noise signals. Finally, model-based technique classify the noise signals using an a priori speech model, such as hidden Markov and voiced/unvoiced models, and then conducts the enhancement depending on classified speech model. This method can be useful for improving the noise reduction performance for various kinds of speech signals. However, it requires extra training to build the model with intensive computation. In addition, it may exhibit model selection errors which cause significant speech quality degradation. Fundamentally, it is not easy to handle complicated speech signals with finite number of speech models. Here is presented an investigation of use of wavelet threshold methods speech signal Enhancement. Unlike Fourier 59

I. INTRODUCTION In voice communications, speech signals can be contaminated by environmental noise and, as a result, the communication quality can be affected making the speech less intelligible. Furthermore, compression of noise speech with low bit rate vocoder may result in considerable quality degradation due to frequent estimation errors of speech production model parameters required by the vocoder. This problem can be reduced significantly by speech enhancement (or noise cancellation), which may enable more pleasant voice communication by suppressing the noise components in input signals. Generally, it is assumed that the noisy speech signal is formed additively by speech and noise signals in which the noise is generated by environmental sources such as vehicles, street noise, babble, etc. Therefore, in real environments complete noise cancellation is not feasible as it is not possible to completely track varying noise types and characteristics that change with time. However, by assuming that the noise characteristics change slowly in the background noise levels producing more pleasant and intelligible speech quality. Speech enhancement techniques can help the speech model

International Journal of Ethics in Engineering & Management Education

Website: www.ijeee.in (ISSN: 2348-4748, Volume 1, Issue 3, March 2014) transducer that can be converted into an encoded digital data Transform based spectral analysis for frequency domain stream by using an analog-digital converter (ADC) and analysis, Wavelet Transform (WT) is a powerful tool for constitutes digital processing of analog signals. Alternatively, signal processing to extract the relevant time-amplitude if a digital-analog converter (DAC) is applied to digital data information from a signal for its multi resolution possibilities stream, the audio signal transmits through an amplifier and a and is suitable for application to non stationary signals, whose speaker. The process is shown symmetrically in the figure1, frequency response varies in time[1,2]. Wavelet coefficients which identifies the important steps in digital audio signal. represent a measure of similarity in the frequency content between a signal and a chosen wavelet function [2]. By wavelet analysis at high scales, extract global information from a signal called approximations, whereas at low scales can extract fine information from a signal called details [3]. In signal processing, wavelets are used for many purposes [2], such as Enhancement, detecting trends, breakdown points discontinuities in higher derivatives and self similarity in signals. Since speech signal is one dimensional and nonstationary, it is often corrupted by means of various kinds of noises present in environment like White Gaussian noise, Colored noise, Burst noise etc., In order to develop an efficient and effective technique of Enhancement of audio signal. It is Fig. 1. Shows the process of digital processing of three types of audio necessary to separate useful contents of audio from noise by signal. Part(a) represents a complete digital audio processing suitable method. An efficient method using wavelet shrinkage, comprising (from left to right) a microphone, amplifier, ADC, digital which performs well over conventional frequency selective processing material, DAC, amplifying section and speaker; an audio filter approach, is given in [4, 5]. To improve SNR, a block recognition system in (b), and a set of audio synthesis(c). thresholding estimation method in presence of transients and harmonics is suggested in [6, 7]. Heuristic argument estimator B. Recording audio signals in simulink/MATLAB removes noise effectively then Steins Unbiased Risk Estimator (SURE) by discarding purely noise coefficients in Once in the digital domain, these signals can be processed, thresholding. An efficient improvement in SNR can be transmitted or stored. We found that the Audio Device Block achieved by selec0tive smoothing at each scale of time in Simulink enables experimentation and processing of digital frequency plot thus avoiding manual selection of coefficients signals. The from Audio Device Block reads audio Data from [8]. Further to obtain minimum error and maximum signal to an Audio Device in real time. noise ratio wavelet with more varnishing moment is suggested in [8]. A. Basic Audio Theory Sound is the vibration of an elastic medium, where gaseous, liquid or solid. These vibrations are a type of mechanical wave that has the capability to simulate human ear and to create a sound sensation in the brain. In air, sound is transmitted due to pressure vibrations at a rate of change that is called frequency. The difference between the extreme values of pressure represents its amplitude. Pressure vibrations in the range of 20 Hz to 20 kHz produce the sound which is audible to the human ear and this is more receptive when it is between 1kHz to 4kHz. In physical terms, the sound is a longitudinal wave that travels through the air due to vibrations of the molecules. Similar to light, sound wave can be reflected, absorbed, diffracted, or refracted. Audio signals, which represent longitudinal variations of pressure in a medium, are converted into electrical signal by piezoelectric transducers. Transducers coverts the energy of a mechanical displacement into an electrical signal, either voltage or current. The main advantage of converting an audio signal into an electrical signal is that the signal can now be processed. An example is an audio signal obtained from the 60

From Audio Device Block Fig.2. Shows the process of identifies the main steps in a digital audio processing system based in simulink software C. Basic noise theory: Noise is defined as an unwanted signal that interferes with the communication or measurement of another signal. A noise itself is an information-bearing signal that conveys information regarding the sources of the noise and the environment in which it propagates. There are many types of sources of noise or distortions and they include: 1) Electronic noise such as thermal noise and short noise 2) Acoustic noise emanating from moving, vibrating or colliding sources such as revolving machines, moving vehicles, keyboard clicks, wind and rain.

International Journal of Ethics in Engineering & Management Education

3) Website: www.ijeee.in (ISSN: 2348-4748, Volume 1, Issue 3, March 2014) The above arrangement has used two Electromagnetic noises that can be interfere with the wavelet decomposition filters which are high pass and low transmission and reception of voice, image and data pass respectively followed by down sampling by 2 producing over the radio frequency spectrum. half input data point of high and low frequency. The high Electrostatic noise generated by the presence of a frequency coefficients (CD) and low frequency coefficients voltage, are called approximate coefficients (CA) . The signal can be Communication Channel distortion and fading and reconstructed back by inverse wavelet transform. The Quantization noise and the lost data packets due to corresponding filter bank structure for reconstruction is shown in figure 4. network congestion.

4) 5) 6)

This paper is organized by various sections, section-II describes to brief introduction to wavelet transform, Section III gives denoising scheme, section IV gives Experiments and Results, section V describes conclusion followed by References. II. DISCRETE WAVELET TRANSFORM Assume the observed signal y(t) = s(t) + n(t) Contains the original signal s(t) with additive noise n(t) as functions of time t to be sampled. Let W(.) and W-1(.) denote the forward and inverse wavelet transform. Let D(., ) denote the Enhancement operator with threshold . We intend to wavelet de noise y(t) in order to recover s^(t) as an estimate of S(t). S = W(Y) Z = D(Y,) = W-1(Z) Similar to the Fourier series expansion, the DWT maps a continuous variable (t) into a sequence of coefficients, the resultant coefficients are called discrete wavelet transform of (t). Its representation involves the decomposition of the signals in wavelet basis function (t) given by a,b(t)= [(t-b)/a] a, b R -(1) The signal S(t) can be de composed into several levels. A three level wavelet decomposition tree is shown in figure 5. III. WAVELET ENHANCEMENT SCHEME Let us assume signal S(t) is corrupted by noise n(t)as y(t)= S(t) + n(t) where n(t) is white Gaussian noise . the wavelet based de noising scheme is shown in figure 6. Where a, b are called scale and position parameters as respectively. The multi resolution analysis is given by S. Mallet and Mayer proves that any conjugate mirror filter characterizes a wavelet . The wavelet decomposition of a signal x(t) based a multi resolution theory can be obtained using filter [3], the filter based wavelet decomposition is shown in fig. 3.

61

International Journal of Ethics in Engineering & Management Education

Website: www.ijeee.in (ISSN: 2348-4748, Volume 1, Issue 3, March 2014) As seen from figure 6 the de noising scheme involves three j = j - (5) main steps. Where Nj is the length of the noisy signal at jth scale and j is (1) N-level wavelet decomposition of input noisy signal. the Median Absolute Deviation (MAD) at the jth scale given (2) Threshold estimation and thresholding of wavelet by coefficients. j = = - (6) (3) N-level inverse wavelet transform for reconstruction of de noised signal. where represent wavelet coefficient at scale j. De noising using wavelet involves thresholding, in which 3) Rigrsure: It is a soft threshold is evaluator of unbiased coefficients having less thresholding value compared to the risk. Suppose W = [1, 2,. N] is a vector consists of specific threshold value () are set to zero. This process helps the square of wavelet coefficients from small to large. Select to eliminate noise by keeping original signal characteristics. the minimum value rb(bth r) from risk vector, wich is given as, This is called hard thresholding. Another one i.e soft thresholding set the wavelet coefficients which are above the R = {ri}I = 1,2,.N = - (7) threshold. as the risk value. The selected threshold is = where, Value as well as simply shrinks or scales other coefficients which are above the threshold value [1]. b is the bth squared wavelet coefficient (coefficient at minimum risk) chosen from the vector W and is the standard deviation of the noisy signal. 4) Heursure: Threshold is selected using a combination of Sqtwolog and Rigrsure methods. If the signal to noise ratio is very small, the SURE methods estimation is poor. In such case, fixed form threshold of Sqtwolog method gives better threshold estimation [13]. Let threshold obtained from sqtwolog method is 1 and threshold obtained from Rigrsure is 2 then Heuristic SURE gives the threshold given by, = n(d) = [(di)], |di| and [0], |di| - (2) n(d) = [Sign(di).(|di| - )], |di| and [0], |di| - (3) Where di is the input and are noisy wavelet coefficient available in the detailed sub bands which are to be thresholded X is the threshold value and n(d) is the thresholded output which is used to estimate noiseless coefficients. The hard can be unstable or more sensitive to small changes in the data while soft thresholding avoid discontinuities and is more stable than hard thresholding. A. Threshold Methods 1) Minimax Criteria: this method finds threshold () using Minimax principle. It uses a fixed threshold to yield minimax performance for mean square error against an ideal procedure. The minimax principle is used in statistics to design estimators. Since the de-noised signal can be assimilated to the estimator of the unknown regression function. The minimax estimator is the option that realizes the minimum. Over a given set of functions of the maximum Mean Square Error(MSE). This procedure finds optimal thresholds. The threshold is given by where, A = and B = (log2 N)3/2 - (8) . The N is length of

The hard and soft thresholding techniques are used for denoising process. The hard and soft threshold operations with threshold are defined using (2) and (3)

wavelet coefficient vector and s is the sum of squared wavelet coefficients given as . (9) s= Threshold determination is an important problem. A small threshold may yield a result which may be noisy and large threshold can cut significant part of signal thus losing the important details of the signal. IV. RESULTS AND DISCUSSIONS This paper compares the performance analysis of wavelet threshold methods have been applied on the speech signal in English which is taken from a male speaker at a sampling frequency of 25 KHz. All the threshold methods are tested for white Gaussian noise .For performance comparison and measurement of quality denoising is calculated between speech signal S(t) and the denoised speech signal Sd(t) is given by PSNR = 10 log 10 ( smax2 /MSE) (10) Where smax is the maximum value of signal and is given by Smax = max (max( s(t), max (Sd (t))) And MSE is Mean Square Error given by MSE = Sd (t) S (t) ]2 - (11)

N>32 and N<32 (4)

Where = median and is the wavelet coefficient vector at unit scale and N is the length of signal vector. 2) Sqtwolog Criterion: the threshold values () are calculated by universal threshold (square root log) method given by, 62

International Journal of Ethics in Engineering & Management Education

Website: www.ijeee.in (ISSN: 2348-4748, Volume 1, Issue 3, March 2014)

Fig.9: Comparison of PSNR (dB) at level-1 Decomposition TABLE II COMPARISION OF PSNR (DB) AT LEVEL-2 DECOMPOSITION Method Haar Db10 Coif5 Bior !" #$%5 Mini%a& )i*r+,re He,r+,re +-t.o/o* 25!444 22! 44 24! 05 24!"02 2'!( ( 2(! 43 2'! "' 30!0 5 2'!(1 2(!2' 2"!(1 2"!02 2 !51' 24!45' 25! 02 24! 0' 2(!21' 2 !02' 2 !52' 25!5'2

Fig. 8. MATLAB Simulation Results

Fig.10:.Comparision of PSNR (dB) at level -2 Decomposition TABLE III COMPARISION OF PSNR (DB) AT LEVEL-3 DECOMPOSITION Method Mini%a& )i*r+,re He,r+,re +-t.o/o* Haar 22!4'2 1'!("4 25! 05 2(!"02 Db10 2 !0 2 23!502 2(!034 2(!2'4 Coif5 25!(( 24! ' 2 !"2 2 !(1 Bior !" 24!514 21!5'2 22! "' 23! "' #$%5 25!3"' 23!412 25!02' 24!12'

TABLE I COMPARISION OF PSNR (DB) AT LEVEL-1 DECOMPOSITION

Method
Mini%a& )i*r+,re He,r+,re +-t.o/o*

Haar
2 !"31 23!"4 2 !"24 2"!"2

Db10
30!''1 32!53 33!3 2 30!3"2

Coif5
30! ( 2"! 2 2'!(1 2"!(3

Bior !" #$%5

2(!52' 2 ! (2 2 !(52 2 !50 2"!3"' 2(!1"4 2"!323 2"!042

63

International Journal of Ethics in Engineering & Management Education

Website: www.ijeee.in (ISSN: 2348-4748, Volume 1, Issue 3, March 2014)

Fig.11: Comparison of PSNR (dB) at level-3 Decomposition TABLE IV COMPARISION OF PSNR (DB) AT LEVEL-4 DECOMPOSITION Method Mini%a& )i*r+,re He,r+,re +-t.o/o* Haar 21!("2 1"!' 2 24!502 24!142 Db10 24!51' 20! 0' 2 !0 2 24!20' Coif5 24!1( 21!1' 25!"2 23!"1 Bior !" 21!523 1"!5 3 25!" 4 23!"24 #$%5 23!'(2 22!3" 23!40' 23!424
#!N o

Fig.13. Comparative analysis of threshold methods TABLE VI COMPARATIVE ANALYSIS BETWEEN DIFFERENT WAVELET FAMILIES BY USING SQTWOLOG METHOD IN TERMS OF PSNR.
Noi+e 0e1e/ 2#N) of Noi+e +i*na/

Haar

Db10

Coif5

#$%5

1 2 3 4 5

0 dB 5 dB 10 dB 15 dB 20 dB

2"!143 24!5'' 22!102 20!1(0 1"!5""

32!"14 31!204 30!01( 2'!30" 2"!4 (

34! 0 32! 25 31!(05 30!3 2 2'!5 2

33!(12 32!354 31! 42 30! 40 2"!52

33!0 2 32!204 31! 1( 2'!224 2"!43(

Fig.12: Comparison of PSNR (dB) at level-4 Decomposition Haar: Haar wavelet, Db: Doubechies wavelet, coif: Coiflet wavelet, sym: Symlet wavelet TABLE V COMPARATIVE ANALYSIS OF THRESHOLD METHODS
#! N o Noi+e 0e1e/+ 2#N) of Noi+e #i*na/ 2#N) of denoi+ed +i*na/ ,+in* different a/*orith% MiniMa&i )i*r+,re He,r+,r e #-t.o/o*

Fig.14. Comparative analysis between different wavelet families by using sqtwolog In addition, the sqtwolog method is performed by actually performing the comprehensive study of daubechies wavelets (Db -2 to Db 10), haar wavelet, coiflet wavelets (coif-1 to coif-5),symlet wavelets(sym-2 to sym-5) but here are only the results of sym 5 wavelet. A comparative analysis has been performed between Minimaxi/Rigrsure, Heursure and sqtwolog and the results in the form of PSNR are given in table 1,2,3,4 for 5 different wavelets. It is observed as the level of decomposition is increased from level 1 to level-4, the PSNR values of noise signals go on reducing and improved in terms of PSNR for denoised signals. When Db10 is applied at all the decomposition levels, significant increase in PSNR is obtained in comparison 64

1 2 3 4 5

0 dB 5 dB 10 dB 15 dB 20 dB

2"!143 24!5'' 22!102 20!1(0 1"!5""

30!2"" 2"!"5( 2(!'14 2 !(23 2 !(02

30!5"2 2'!1"2 2"!23" 2(!531 2 !'(2

32!"(5 31!104 2'!" 5 2"!02( 2(!30

33!24 31! 0( 30!42" 2'!4"( 2"! 11

International Journal of Ethics in Engineering & Management Education

Website: www.ijeee.in (ISSN: 2348-4748, Volume 1, Issue 3, March 2014) Haar, Daubechies, Coiflet, Symlet, Biorthognal. This P.Sunitha received the B.Tech degree in demonstrates a significant improvement in signal quality and Electronics and Communication comparative analysis has been performed between Engineering from JNTU college of Minimaxi/Rigrsure, Heursure and sqtwolog and the results in Engineering , Kakinada in 2006 and the form of PSNR is given in table5 for 5 different levels of M.Tech in Digital Electronics and white Gaussian noise for sym-5 wavelet. It is observed as the Communication Systems (DECS) from level of Gaussian noise increased from 0dB to 20dB. The JNTU Kakinada. She is working towards PSNR values of noise signals go on reducing and improved in the Ph.D degree from JNTUK. terms of PSNR for denoised signals. When sqtwolog is She is currently working as an Assoc. Professor in the dept. of applied on all the 5 noisy signals significant increase in PSNR Electronics and Communication Engineering in Pragati is obtained in comparison to Minimaxi Rigrsure and Hersure. Engineering College, Andhra Pradesh, India. Her Research This demonstrates a significant improvement in signal quality. interest includes Speech processing,Signal Processing, Image In addition, the threshold method is performed by actually Processing and VLSI. performing the comprehensive study of Daubechies wavelets (Db -2 to Db 10), Haar wavelet, Coiflet wavelets (Coif-1 to Coif-5), Symlet wavelets (sym-2 to sym-5) and Biorthognal. T.Sekhar received the B.Tech degree in V. CONCLUSION In this paper the performance analysis of different wavelet threshold methods in di-noising is investigated along with 4 different levels of decomposition with white Gaussian noise. The result shows that as the level of decomposition increases the value of PSNR decreases. Thus the lower levels of decomposition can be preferred as seen from the result performance of db10 better than as compared to other wavelets. Haar wavelet is poorer as compared to others also it gives unwanted distortion in the reconstructed signal. Overall from the results, it is clear that db10 is best suitable giving minimum distortion and maximum PSNR. REFERENCES
S.Mihov, D.Doychev, R.Ivanov practical investigation of specific types of noice signals for the purpose of their hearing aid devices, processings of ICEST 2009 vo12 PP 399-402, viliko Tarnovo 2009. 2) G.Openherim J.M Poggi, M.Misiti, Y.Misiti wavelet tool box themath works, Inc., Natick Massachersetts 01760, April 2001. 3) Fundamentals of Wavelets, Theory Algorithms and Applications by J.C.Goswamy, A.K.Chan 4) Domoho D.L. and Johnson I.M., Enhancement by soft thresholding, IEEE Transaction on Information Theory, 1995,41,(3) PP.613 627. 5) D.L.Donoho and I.M.Johnstone Minimax Estimation via wavelet shrinkage Annals of Statistics, 26(3):879 921, 1998. 6) Guoshen Yu, Stephane Mallat, Emmanuel Bercy, Audio Denoising by Time Frequency Block Thresholding, IEEE transaction on signal processing. Vol. 56, No.5, May 2008. 7) S.Sreekanth, P.Dinesh Khanna, P.Uma Mallikarjuna, An Efficient Noise Reduction by using Diagonal and Non diagonal Estimation Techniques, Proceedings of the International Conference on Computationgal Intelligence, 2010. PP 393 398. 8) A. Sony John, Uday B. Desai, signal Enhancement using the wavelet Transform and Regularization, 1997, IEEE. 9) Denoising speech signals by wavelet Transform. Annual Journal of Electronics, 2009. ISSN 1313-1814. 10) Performance Analysis of wavelet Thresholding methods in Enhancement of Audio signals of some Indian Musical Instruments. ISSN:0975 5462, Vol. No. 05th May,2012. PP. 2047 2050. 1)

Electronics and Communication Engineering from Chaitanya Institute Science & Technology, Madhavapatnam, Kainada in 2010 and M.Tech in Embedded Systems (ES) from Pragati Engineering College, Surampalem in 2013. He is currently working as an Asst. Professor in the dept. of Electronics and Communication Engineering in Pragati Engineering College, Andhra Pradesh, India. He Research interest includes Embedded System, Signal Processing, and Image Processing.

65