ANALYSIS OF VARIOUS SPEECH ENHANCEMENT METHODS

Transcription

1 ANALYSIS OF VARIOUS SPEECH ENHANCEMENT METHODS Assistant professor, Department of Electrical Engineering, Annamalai University ABSTRACT Speech signal is constantly disturbed by the unwanted occurrence of background noise. Due to the noise presence, the original speech signals are affected and not clear to listen mainly in the field of communication and speech processing. Hence efficient noise reduction techniques are needed in order to extract the preferred speech signal from its corrupted. That is, removal of background noise from the original. Therefore, Speech enhancement is very necessary to develop speech quality, clearness, accuracy and robustness. By using various methods and algorithms of Speech enhancement are often used in the speech coding, synthesis and recognition of speech region. This paper presents the review of various speech enhancement methods along with the noise performance estimators. Keywords Spectral Subtraction, SNR, Speech Enhancement, Filtering, VAD, MFCC [1] INTRODUCTION Speech is the special identity of the human beings gifted by the nature. Speech helps to deliver the thoughts and messages are freely to others for better understood. Speech frequency normally ranges between 3 Hz to 4 KHz depending upon the personality. Normally human beings have an audible frequency range of 20 Hz to 20 KHz. The most common problem in speech processing is the effect of interfering of noise in the speech signals[1].the noise masks the speech signal reduces the quality and the speech is greatly affected by presence of surroundings noise. This makes the listening task hard for both circuitous and straight spectators and gives poor performance in some of speech processing. Restraining or reducing such backdrop noise and improving the perceptual quality and intelligibility of a speech without disturbing the speech signal quality is very tough to hold. Speech enhancement system can be categorize by several opinion like adaptive or non-adaptive, frequency or time domain and the quantity of channels etc.[6] for better removing of noise. Hence the ultimate goal of speech enhancement differs according to particular applications, to increase clearness and boost the overall speech quality, and to improve the performance of voice communication [2]. This paper presents an analysis of various speech enhancement methods and performance evaluation. 129

2 ANALYSIS OF VARIOUS SPEECH ENHANCEMENT METHODS This paper is organized as follows, Section II describe speech enhancement, Section III represent types of speech enhancement method, Section IV performance evaluation, and Section V deals the conclusion. [2] SPEECH ENHANCEMENT Speech enhancement aspires at improving the quality speech in communication system in noisy environments. Quality of the enhanced signal is measured by the clarity, unclear character, and the stage of residual noise in that signal [27]. And intelligibility of the enhanced signal is measured by the listener s talents. There are various speech improvement methods proposed for noise reduction and to improve the speech quality and clearness. Only one algorithm is not adequate for all the types of noise present in the surroundings. Hence speech enhancement algorithms are created based on the applications. The speech enhancement systems can be classified in different types. For a sample, single channel speech enhancement, multichannel speech enhancement and model based speech enhancement and so many (22). The single channel speech enhancement is very less pricey and make easy to process. Multi-channel assists to eliminate noise in an effective manner and it is high difficult to process. Input speech Segment/ Windowing DCT/FFT Noise Estimation Level of Spectrum Spectral gain calculator Inverse DCT/FFT Output speech Fig.1.Block diagram of speech enhancement The block diagram of speech enhancement is show in figure.1. Generally hamming window is used for speech is better. Then apply Fourier transform either Discrete or Fast Fourier transform of segmented and windowed. FFT is optimum for speech enhancement. Noisy signal find and send to noise estimation block. This noise estimation block is used for calculate the overall noise in the original speech. Then if noise estimate is too low, unnecessary residual noise will be audible else too high, speech will be unclear. Enhancement spectral gain block improve speech spectrum is generate and apply the inverse Fourier transform; it gives a clean speech signal.

3 [3] TYPES SPEECH ENHANCEMENT 1. Spectral subtraction method 2. Spectral Subtraction with over subtraction 3. Nonlinear spectral subtraction 4. Multiband spectral subtraction method 5. Iterative spectral subtraction method 6. Eigen value spectral subtraction method 7. Nonlinear weighted noise subtraction method 8. Minimum Mean Square Error (MMSE) method 9. Modified spectral subtraction method 10. Least mean square (LMS) 11. Inverse Fourier spectral subtraction method 12. Linear predictive coding (LPC) 13. Magnitude averaging 14. Signal subspace method 15. Adaptive averaging 16. Cepstral smoothing of spectral filter gains 17. Voice activity detection (VAD) 18. Minimal tracking algorithm 19. Minimum statics noise cancellation 20. Adaptive noise cancellation 21. Short time spectral amplitude (STSA) 22. Power spectral algorithm 23. Beoruti spectral subtraction 24. Partial differential equation (PDE) 25. Spectral subtraction based on perception qualities 26. Beam forming algorithm 27. Hidden Markov Model (HMM) algorithm 28. Minima controlled recursive algorithm 29. Karhuenen-Loeve transform (KLT) method 30. Improved controlled recursive algorithm Spectral subtraction method The estimated noise spectrum is subtracted from the noisy speech input in order to obtain clean speech. If moreover much noise is eliminated then some information might be missing. The noise spectrum is estimated during the stage when only noise is present. 131

4 ANALYSIS OF VARIOUS SPEECH ENHANCEMENT METHODS Speech +Additive noise Magnitude of FFT X (k) =Y (k)-d (k) Inverse FFT Clean speech Fig 2.Block diagram of spectral subtraction method That is y(u) = x(u) +d(u) Where y (u) is noisy speech, x (u) is clean speech and d (u) denote uncorrelated additive noise. In frequency domain it can be written as Y(k) = X(k)+ D(k) k is index of frequency. Estimated clean speech spectrum is obtained as X (k) = Y(k) - D (k) And D (k) is average magnitude of noise spectrum. Spectral Subtraction with over subtraction The basic spectral subtraction is slightly modified and executed by subtracting the overestimated noise range and carefully consider about the minimum spectral base value. Then, X (k) 2 = Y(k) 2- D (k) 2 if Y(k) 2 is greater than ( (over subtraction factor (O)+ Spectral base value (N) 2* D (k) 2 ) else C D (k) 2. N controls the noise (28). Where noise N is too small, the speech is audible, and then N is too large, subject to spectral subtraction reduces. Non linear Spectral Subtraction Lock wood and boudy has proposed this nonlinear spectral subtraction method. This method processed subtraction process in nonlinear way with depends on the over subtracted factor frequency. Thought behind of this nonlinear is, does not affect the entire spectral parameters are in similarly. X(k) = Y(k) -F(k)* nl(k) F (k) is frequency dependent factor, nl (k) is nonlinear function. nl (k) = max( D (k) ). Multi band Spectral Subtraction Speech spectrum is divided in to N overlapping bands; independently apply the spectral subtraction method to the each band. By using time or frequency domain, the speech signal has to split in to various bands. The estimate of the clean speech is

5 X (k) 2 = Y(k) 2- O*AB D (k) 2 Where O is the over subtraction factor and AB is the additional band. Iterative Spectral Subtraction In this technique, the output of the spectral subtraction method is used as the input signal of the next iteration process. Then spectral subtraction is done, the background noise is transform to scrap noise. Again this scrap noise to be taken and estimate by spectral subtraction method, finally good speech retrieved, this process is iterative spectral subtraction. Iteration period is more important in this method. Minimum Mean Square Error Ephram and malah proposed this system, that MMSESTSA using models for the sharing of spectral components of speech and noise signal [33]. That is, to measure the degree of resemblance between two signals. The main target of this system is to minimize the mean square error between short time spectral magnitude of the clean speech and enhanced speech. It is defined as, MMSE= 1/A *(r(x) -i(x) *i(x)) Where A is length of renovated signal, i(x) is input1 signal and r(x) is reconstructed speech. The block diagram illustrate the MMSESTSA method, Input signal Input Filtering Nosiness Identified Factors Restructured signal Fig 3. Block diagram of MMSA method Adaptive LMS filtering Digital filter and adaptive method are the two basic components are used in this lms adaptive filter technique [15]. This filtering is very effective in real time usage. In this technique, Digital filter gives output with respect of input signal and adaptive algorithm is correcting the coefficients of digital filter. It is also Robustness and simple technique. Noise +Signal input LMS program Blare New Filter 131

6 ANALYSIS OF VARIOUS SPEECH ENHANCEMENT METHODS Fig 4. Block diagram of adaptive filtering Toward revise the filter coefficients multiply the μ error by the learning rate limit, and then multiply the results by the filter input and add this result to the values of the prior filter coefficients. Linear Predictive Coding LPC is mainly used in audio, speech and signal processing to signifies the spectral surround of a speech digital signal in compacted way. This LPC is very much useful for training good quality of speech with low bit rate and provide accurate speech. LPC analyzes the speech signal by estimating the formants, removing their possessions from the speech signal, and approximation the concentration and frequency of the outstanding noise(5)(6). The method of removing the formants is called inverse filtering, and the remaining signal later than the subtraction of the filtered modeled signal is called the remainders. The result is to convey every sample as a linear mixture of earlier samples. This equation is called linear predictor or linear predictive coding. Input Frame jamming Windowing Correlation study LP Analysis Fig 5. Block diagram of linear predictive coding Output Signal Subspace Method The approach engages the use of a signal dependent convert to decompose a noisy signal into two split subspaces, the signal in addition noise subspace, and the noise-only subspace (5).The change employed to perform by using the Karhuenen-Loeve transform (KLT) method. Speech can only period the signal plus noise subspace, called the signal subspace, while noise can extent the entire Euclidean gap. Only the signal subspace is used to estimating the clean signal. The KLT apparatus which represent the noise only subspace are worthless, while the components which represent the noisy signal are adapted by a gain function. The enhanced signal received from the inverse KLT. Hence here the plan is to improve the quality, even as minimizing any loss in intelligibility. Noise KLT Block Gain sum Inverse KLT Improved Signal

7 Fig 6. Block diagram of signal subspace method Adaptive noise cancellation Adaptive noise cancellation is great and another technique of estimating signals despoiled by additive noise or intrusion. Here the technique is to obtain an approximation of the noise signal and subtract it from the degraded signal. It assumes no need of priori facts about the noise or signal [23]. This method also used in multi-channel speech enhancement purely based on the availability of a supporting channel, said as reference path, where correlated sample or reference of the pollute noise is present. ANC cancels the primary unwanted noise by commencing anti-noise. It also improves the signal to noise ratio of the expected signal. Voice activity detector VAD is commonly used in single channel speech enhancement methods to detecting presence and absence of speech in a noisy speech signal. It indicates and gives the values of zeros as speech pauses and one as a speech activity in each frame [25]. The noise is assumed to be short term inactive, so that noise from silent frames can be used to remove noise from speech frames. VAD algorithms take out features the short-time power, zero crossing, LPC from the input signal and compares against a threshold value, usually determined during speech missing time [8].VAD cannot well approximate the noise level in non-stationary and short SNR setting. Beamforming In multiple input and single output application beamfoming is used. It is the multichannel speech enhancement filtering technique, and it enhances the desired speech signal and suppressed noise single (1)(9). More than two microphones are arranged in some desired shape then beamformer senses the filters output and amplifies the signals depending on the direction of arrivals. There are two types of beamformers are fixed and adaptive. Mel Frequency Cepstral Coefficients (MFCC) MFCC are the coefficients obtained from the Mel-frequency Cepstrum account. The Mel Frequency Cepstrum (MFC) can be defined by the short time power continuum of speech signal which is calculated as the linear cosine transform of the log power spectrum on a nonlinear Mel scale of frequency(18)(15). The divergence between the MFC and Cepstral analysis is that the MFC records frequency components using a Mel scale based on the human ear perception sound as a substitute of a linear scale. The Mel-frequency Cepstrum Coefficient (MFCC) technique is often used to create the fingerprint of the sound files. MFCC are based on the known dissimilarity of the human ear s vital bandwidth frequencies with filters output. The following formula is used to compute the Mels for an exacting frequency: mel( f ) = 2595*log10(1 + f / 700). There are two types of MFCC (i) stochastic models and (ii) template models. MFCC can be observed as the typical features in speaker as well as speech recognition. MFCC is also ever more 133

8 ANALYSIS OF VARIOUS SPEECH ENHANCEMENT METHODS detection uses in music information such as genre classification, audio similarity quantify and so on. It gives high accuracy results for clean speech. Input Speech Framing DFT LOG MEL Frequency warping Contrary DFT Mel Cepstrum MFCC output Fig 7. Block diagram of MFCC method An MFCC is also uses filter to wrap the frequency spectrum onto the Mel-scale that is related to the human ear recognizes sound. [4] Performance of Evaluation [4.1] Signal-to-Noise Ratio (SNR) Signal-to-Noise Ratio (SNR)-to-noise ratio is a compute used in signal processing that evaluates the level of a desired wave to the level of background noise.signal-to-noise ratio is sometimes used to refer to the part of useful information to extraneous data in a discussion or replace. Signal-to- noise ratio is defined as the power percentage between a significant background noises. SNR=10 * log 10 mean (input 2 ) Mean (input 2 -Enhanced 2 ) [4.2] Mean Opinion Score Measure (MOS) The mean opinion score (MOS) presents an arithmetical measure of the quality of human speech. The system uses skewed tests (out looked scores) that are mathematically averaged to obtain a quantitative indicator of the system performance. MOS is determined, a number of listeners rate the quality of test words by male and female speakers. A listener gives each sentence a rating as follows: (1) Bad (2) Poor (3) Fair (4) Good (5) Excellent. The MOS is the arithmetic mean of all the individual scores, desired by the value from bad to best. [5] CONCLUSION In this paper, various algorithms for speech enhancement were reviewed. Each algorithm has their own benefits and losses. The purpose of enhancing the signal is to give a clear signal removed from noise, and then only the signal is used in various applications. The principle of speech enhancement can be realized by based on the type of purpose. This paper presents an approaching to different types of speech

9 enhancement techniques based upon the needs. Hence in future, use any one among these types of enhancement technique from that concluded which one is better. ACKNOWLEDGMENTS The author thanks the authorities of Annamalai University for providing the required facilities in order to complete this portion of work. REFERENCES 1. Kaladharan N, Speech Enhancement by Spectral Subtraction Method International Journal of Computer Applications ( ) vol-95, no.13, June Sunita Dixit et al, Review on Speech Enhancement Techniques International Journal of Computer Science and Mobile Computing, Vol.3 Issue.8, August- 2014, pg Sunnydayal. V, N. Sivaprasad, T. Kishore Kumar A Survey on Statistical Based Single Channel Speech Enhancement Techniques I.J. Intelligent Systems and Applications, 2014, 12, G Ramesh Babu et al, Combination of Beam forming and Kalman filter Techniques for Speech Enhancement International Journal of Computer Science & Communication Networks,Vol 1(3), Dr. G. Ramesh Babu et al Speech Enhancement Using Beamforming International Journal Of Engineering And Computer Science ISSN: Volume 4 Issue 4 April 2015, Page No Nandini Garg, JyotiGupta Review on Speech Enhancement using Signal Subspace method International Journal of Application or Innovation in Engineering & Management (IJAIEM) Volume 2, Issue 5, May 2013 ISSN Rajdeep Kaur, Jyoti Gupta Speech Enhancement Using LPC Analysis-A Review International Journal of Application or Innovation in Engineering & Management (IJAIEM) Volume 2, Issue 5, May Kavita Sharma, Prateek Haksar Speech Denoising and Speech Enhancement Using Wavelet Filter IOSR Journal of Engineering (IOSRJEN) Vol. 2 Issue 1, Jan.2012, pp Savita Hooda and Smriti Aggarwal Review of MMSE Estimator for Speech Enhancement International Journal of Advanced Research in Computer Science and Software Engineering Volume 2, Issue 11, November Pankaj Bactor, Anil Garg Different Techniques for the Enhancement of the Intelligibility of a Speech Signal International Journal of Engineering Research and Development Volume 2, Issue 2 (July 2012), PP Paurav Goel, Anil Garg Review of Spectral Subtraction Techniques for Speech Enhancement IJECT Vol. 2, Issue 4, Oct. Dec Soumasunderaswari.D, Prashanthini.K A survey on various multichannel speech enhancement algorithms International Journal of Advanced Research in Computer and Communication Engineering Vol. 4, Issue 1, January Anuprita P Pawar, Kirtimalini.B.Choudhari Enhancement of Speech in Noisy Conditions International Journal of Advanced Research in Electrical, Electronics and Instrumentation Engineering Vol. 2, Issue 7, July

10 ANALYSIS OF VARIOUS SPEECH ENHANCEMENT METHODS 14. S.B.Magre, P.V.Janse, R.R.Deshmukh A Review on Feature Extraction and Noise Reduction Technique International Journal of Advanced Research in Computer Science and Software Engineering, ISSN: X, Volume No.4, Issue No.2, pp: , Subha S, V.Murugan Reverberation Suppression of Noisy Speech Signal International Journal of Innovative Research in Computer and Communication Engineering, Vol.2, Special Issue 1, March Mrs.Vishakha V. Jadhav, Prof. Vijay M. Sardar Improved Performance Based Method for Text Independent Speaker Identification International Journal of Innovative Technology & Adaptive Management (IJITAM) ISSN: , Volume-1, Issue-6, March, Nainesh B Patel, Prof. Hardik N Patel Single channel speech enhancement techniques for removal of additive noise IOSR Journal of Electronics and Communication Engineeringe-ISSN: ,p- ISSN: Volume 9, Issue 2, Ver. III (Mar - Apr. 2014), PP Ganga Prasad, Surendar A Review of Different Approaches of Spectral Subtraction Algorithms for Speech Enhancement Current Research in Engineering, Science and Technology (CREST) Journals, Vol 01 Issue 02 April Lalchhandami and Rajat Gupta Different Approaches of Spectral Subtraction Method for Speech Enhancement International Journal of Mathematical Sciences, Technology and Humanities 95 (2013) ISSN Shishir Banchhor, Jimish Dodia, Darshana Gowda GUI Based Performance analysis of speech enhancement techniques International Journal of Scientific and Research Publications, Volume 3, Issue 9, September ISSN Suma. M. O, Madhusudhana Rao. D, Rashmi. H. N and Manjunath B.S Speech Enhancement using Spectral Subtraction International Journal of Advanced Electrical and Electronics Engineering, (IJAEEE) ISSN (Print), , Volume-2, Issue-4, Isiaka A. Alimi, Michael O. Kolawole Enhancement of Speech Communication Technology Performance Using Adaptive-Control Factor Based Spectral Subtraction Method JTIT, 2/ Ching-Ta Lu1, Kun-Fu Tseng and Chih-Tsung Chen Reduction of Musical Residual Noise Using Hybrid-Mean Filter International Journal of Signal Processing,Image Processing and Pattern Recognition Vol. 6, No. 4,August, Arata Kawamura, Weerawut Thanhikam, and Youji iguni Single Channel Speech Enhancement Techniques in Spectral Domain International Scholarly Research Network ISRN Mechanical Engineering Volume Ekaterina Verteletskaya, Boris Simak Noise Reduction Based on Modified Spectral Subtraction Method IAENG International Journal of Computer Science,, 38:1, IJCS 26. Anuradha R. Fukane, Shashikant L. Sahare Different Approaches of Spectral Subtraction method for Enhancing the Speech Signal in Noisy Environments IJSER, Volume 2, Issue 5, May Paurav Goel, Anil Garg Review of Spectral Subtraction Techniques for Speech Enhancement International Journal of Electronics & Communication Technology, Vol. 2, Issue 4, Oct. - Dec Radu Mihnea Udrea, Dragos Nicolae Vizireanu, Claudia Cristina Oprea, Ionut Pirnog An Improved Multi-band Speech Enhancement Method for Colored Noise Estimation and Reduction IJAT, vol 3 no 3 & 4, year Ekaterina Verteletskaya, Boris Simak Enhanced spectral subtraction method for noise reduction with minimal speech distortion IWSSIP th International Conference on Systems, Signals and Image Processing.

11 30. D. Deepa, A. Shanmugam Spectral Subtraction Method of Speech Enhancement using Adaptive Estimation of Noise with PDE method as a preprocessing technique ICTACT Journal of Communication Technology, March 2010, Issue: Ganga Prasad Surender A Review of Different Approaches of Spectral Subtraction Algorithms for Speech Enhancement Current Research in Engineering, Science and Technology (CREST) Journals. 137