G. Rajavali, G.Ramesh Babu / International Journal of Engineering Research and Applications                        (IJERA)...
G. Rajavali, G.Ramesh Babu / International Journal of Engineering Research and Applications                        (IJERA)...
G. Rajavali, G.Ramesh Babu / International Journal of Engineering Research and Applications                        (IJERA)...
G. Rajavali, G.Ramesh Babu / International Journal of Engineering Research and Applications                        (IJERA)...
G. Rajavali, G.Ramesh Babu / International Journal of Engineering Research and Applications                        (IJERA)...
Upcoming SlideShare
Loading in …5



Published on

IJERA (International journal of Engineering Research and Applications) is International online, ... peer reviewed journal. For more detail or submit your article, please visit www.ijera.com

  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide


  1. 1. G. Rajavali, G.Ramesh Babu / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1852-1856 A Effective Speech Enhancement By Using Robust Adaptive Kalman Filtering Algorithm G. RAJAVALI G.RAMESH BABU M.Tech Student Scholar, DECS, M-Tech, (P.hd), Assoc Professor Dept of Electronics and Communication Dept of Electronics and Communication Engineering, Engineering, SRI SIVANI COLLEGE OF ENGINEERING, SRI SIVANI COLLEGE OF ENGINEERING, Chilakalapalem, Etcherla (M); Srikakulam (Dt); Chilakalapalem, Etcherla (M); Srikakulam (Dt); A.P, India. A.P, IndiaAbstract In This paper Introduces a new technique Amount of echo. Echoless speech, measured in afor speech enhancement to enhance speech special anechoic room, sounds dry and dull to humandegraded by noise corrupted signal. When speech ear. Echo suppression is needed in big halls tosignal with an additive Gaussian white noise is the enhance the quality of the speech signal. it is crucialonly information available processing. So many not to harm or garble the speech signal. Or at least nottechniques are to improve the efficiency for very badly. Another thing to remember is that quietenhancement. Kalman filter is one of the natural background noise sounds more comfortabletechnique for speech enhancement. But in kalman than more quiet unnatural twisted noise. If the speechfilter needs to calculate the parameters of Auto- signal is not intended to be listened by humans, butRegressive model, For this a lot of matrix driven for instance to a speech recognizer, then theoperations required. In which speech signal is comfortless is not the issue. It is crucial then to keepusually modeled as autoregressive (AR) process the background noise low.and represented in the state-space domain. In thispaper presents a alternative solution for estimate Background noise suppression has manythe speech signal. In proposed technique to applications. Using telephone in a noisy environmenteliminates the matrix operations and calculating like in streets of in a car is an obvious application. Alltime by only constantly updating the first value of the speech enhancement methods aimed atstate vector X(n). The experiments results show suppressing the background noise are (naturally)that the Improved algorithm for adaptive based in one way or the other on the estimation of theKalman filtering is effective for speech background noise. If the background noise isenhancement. evolving more slowly than the speech, i.e., if the noise is more stationary than the speech, it is easy toKeywords- Kalman Filter, Speech Enhancement, estimate the noise during the pauses in speech.AR model, Noise Finding the pauses in speech is based on checking how close the estimate of the background noise is toI. INTRODUCTION the signal in the current window. Voiced sections can Speech enhancement aims to improve be located by estimating the fundamental frequency.speech quality by using various algorithms. Both methods easily fail on unstressed unvoiced orIntelligibility and pleasantness are difficult to short phonemes, taking them as background noise.measure by any mathematical algorithm. Usuallylistening tests are employed. However, since In this approach the signal is modeled as anarranging listening tests may be expensive, it has AR process. The estimation of time-varying ARbeen widely studied how to predict the results of signal model is based on robust recursive least squarelistening tests. No single philosopher’s stone or algorithm with variable forgetting factor. Theminimization criterion has been discovered so far and variable forgetting factor is adapted to a nonhardly ever will. stationary signal by a generalized likelihood ratio algorithm through so-called discrimination function, The central methods for enhancing speech developed for automatic detection of abrupt changesare the removal of background noise, echo in stationarity of signal. In this paper the signal issuppression and the process of artificially bringing modeled as an AR process and a Kalman filter based-certain frequencies into the speech signal. We shall method is proposed by reformulating and adaptingfocus on the removal of background noise after the approach proposed for control applications bybriefly discussing what the other methods are all Carew and Belanger [7]. This method avoids theabout. First of all, every speech measurement explicit estimation of noise and driving processperformed in a natural environment contains some variances by estimating the optimal Kalman gain. 1852 | P a g e
  2. 2. G. Rajavali, G.Ramesh Babu / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1852-1856After a preliminary Kalman filtering with an initial In [2] the transition matrix F and the observationsub-optimal gain, an iterative procedure is derived to matrix H are modified. They has defined asestimate the optimal Kalman gain using the propertyof the innovation sequence.II. IMPROVED KALMAN FILTERING ALGORITHMA. Conventional Kalman Filtering Method: Speech driven by white noise is All-polelinear output from the recursive process. Under theshort-time stable supposition, a pure speech can (6)establish L step AR model by It also has defined the L×1 state vector X(n) =[s(n) ........s(n-L+1) s(n-L+2)] , the L×1input vector Q(n = [s( ) n 0 …… 0] , and the 1× L observation vector R(n) = [1,v(n),...v(n-L+ 2)](1) .Finally, (3) and (4) can be rewritten into the matrix In (1) is the LPC coefficient, ω( ) is n equations bythe white Gaussian noise which the mean is zero andthe variance is. In the real environment, the speech [State equation]signal s(n) is degraded by an additive observationnoise v(n).Its mean is zero, and its variance is .This noise isn’t related to s(n) . A noisy speech (7)signal y(n) is given by [Observation equation] y(n) = s(n) + v(n)(2) (8) In this paper, it is assumed that the variance Then the recursion equation of Kalman filtering algorithm is given in Table I. In this case the is known, but in practice we need to estimate it bythe "silent segment" included in the y(n). (1) and (2) noise variance is known.can be expressed as the state equation and theobservation equation which are given by[State equation](3)[Observation equation](4) This algorithm abrogates the computation of(5) the LPC coefficient. The number of calls for the F(n) is the L X L transition matrix filtering equations is equal to the number of samplingexpressed as G is the input vector and H is the point’s n of the speech signals, so the algorithmsobservation vector. It is easy to see that the time complexity is O (Ln).conventional Kalman filtering is using the LPCcoefficient to estimate the observations of the speech B. Improved Filtering Method:signal. This part spends half the time of the whole By the recursive equations of thealgorithm. conventional Kalman filtering algorithm, we can find that (5) - (13) contain a large number of matrix operations. Especially, the inverse matrix operations 1853 | P a g e
  3. 3. G. Rajavali, G.Ramesh Babu / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1852-1856lead to an increase in the algorithm’s complexity. If or equal to U, the current speech frame can bewe can reduce the dimension of matrix or eliminate considered as noise, and then the algorithm will re-matrix operations, we can greatly reduce the estimate the noise variance. In this paper, can’tcomplexity of the algorithm. In Table I, we can find replace directly , because we do not know thethat (11) constantly pans down the value of state exact variance of background noise. In order tovector X(n), and then (12) constantly updates the first reduce the error, we use d.value s(n) of X(n). However, during the whole 2) Updating the threshold byfiltering process, only the value of s(n) is useful. Sowe can use the calculation of s(n) instead of thecalculation of the vector in order to avoid the matrix (18)inversion[4]. Furthermore, computational complexity In (15), d is used again to reduce the error.of the algorithm can be reduced to O(n). The However, there will be a large error when the noise isrecursive equations of the improved filtering method large, because the updating threshold U is notare shown in Table II. restricted by the limitation <U. It is only affected by So, we must add another limitationTABLE II before implementation (18). In order to rule out theAN IMPROVED FILTERING METHOD speech frames which their SNR (Signal-to-noise rate)PROCEDURE is high enough, it is defined that the variance of pure speech signals is the variance of the input noise speech signals, and is the variance of background noise. We calculate two SNRs and compare between them. According to [6], one for the current speech frames isC. The Adaptive Filtering Algorithm: Due to the noise changes with thesurrounding environment, it is necessary to constantly (19)update the estimation of noise. So we can get a more Another for whole speech signal isaccurate expression of noise. Here we furtherimprove the Kalman filter algorithm, so that it canadapt to any changes in environmental noise andbecome a fast adaptive Kalman filtering algorithm forspeech enhancement. The key of the fast adaptive (20)algorithm is that it can constantly update the In (19) and (20), n is the number of speechestimation of background noise. We can set areasonable threshold to determine whether the current frames, and has been updated in order to achieve aspeech frame is noise or not. It consists of two steps: higher accuracy. The speech frame is noise whenone is updating the variance of environmental noise, is less than or equal to , or is the other is updating the threshold U[5]. less than zero, and then these frames will be follow 1)Updating the variance of environmental noise by the second limitation ≤ U) [6]. However, if is larger than , the noise estimation (16) will be attenuated to avoid damaging the speech In (16), d is the loss factor that can limit the signals.length of the filtering memory, and enhance the role According to [7], this attenuation can be expressed asof new observations under the current estimates.According to [4] its formula is The implementation process for the whole algorithm can be seen in Table III. III. MATLAB/ SIMULATION RESULTS A. The Comparison between the(17) Conventional and the Fast Filtering Method The(B is a constant between 0.95 and 0.99. In this paper, simulations are done under the following conditions:it is 0.99) Before the implementation of (16), we will 1) The pure speech signals were recorded in anuse the variance of the current speech frame to anechoic chamber with 16 kHz sampling frequencycompare with the threshold U which has been and digitized.updated in the previous iteration. If is less than 1854 | P a g e
  4. 4. G. Rajavali, G.Ramesh Babu / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1852-18562) The background noises are an additive whiteGaussian noise which is produced by MATLAB.TABLE IIITHE ADAPTIVE METHOD PROCEDURE Fig.1 Pure Speech signalFunction awgn. The noise variance is assumed tobe known, and the SNR of the noise signal SNR in, isdefined by(21) where i is the total number of samples forthe speech signal. We adopt two patterns of the noisy Fig.2 The noise speechspeech as the signal samples for the simulations. Oneis the speech signal corrupted with a backgroundnoise, In this section, we will compare the simulationresult in the 3 different phases: (a) signal wave, (b)performance, (c) running time for the fast filteringmethod and the conventional Kalman filteringmethod.(a) To compare the filtering efficiency in the view ofsignal wave, the results are shown in Fig.1 and Fig.2.Each figure contains the wave of pure speech signal,and noise speech signal, filtering the result by theconventional method and filtering the result by theimproved method under the condition of SNR 10[dB]in . We can see that the improved method gives quitethe same results as the conventional method [7]-[8].(b) Compare the SNR of them in the view ofperformance: Fig.3 The conventional Kalman filter method(21)means the dimension of the transition matrix; 1855 | P a g e
  5. 5. G. Rajavali, G.Ramesh Babu / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue4, July-August 2012, pp.1852-1856 [3] Nari Tanabe, Toshiniro Furukawa , Shigeo Tsujii. Fast noise Suppression Algorithm with Kalman Filter Theory. Second International Symposium on Universal Communication, 2008.411-415. [4] Nari Tanabe, Toshiniro Furukawa ,Hideaki Matsue and Shigeo Tsujii. Kalman Filter for Robust Noise Suppression in White and Colored Noise. IEEE International Symposium on Circuits and Systems, 2008.1172-1175. [5] WU Chun-ling, HAN Chong-zhao. Square- Root Quadrature Filter. Acta Electronica Sinica, Vol.37, No.5, pp.987-992, May.2009. [6] SU Wan-xin, HUANG Chun-mei, LIU Pei- wei, MA Ming-long. Application of adaptive Kalman filter technique in initial alignmentFig.4 Improved version of speech signal by using of inertial navigation system. Journal ofproposed method Chinese Inertial Technology,Vol.18, No.1, pp.44-47, Feb.2010.(c) Compare the running time of the conventional [7] GAO Yu, ZHANG Jian-qiu. Kalman Filterwith the improved method. Table I and the improved with Wavelet-Based Unknown Measurementfiltering method in Table II. Both of them are running Noise Estimation and Its Application forunder this condition: Information Fusion. Acta Electronica Sinica, Vol.35, No.1, pp.108-111, Jan.2007.V CONCLUSION [8] XIE Hua. Adaptive Speech Enhancement This paper has presented a Robust Kalman Base on Discrete Cosine Transform in Highfiltering algorithm for speech enhancement by Noise Environment. Harbin Engineeringeliminating the matrix operation and designing a University, 2006. 332coefficient factor. It has been shown by numericalresults and subjective evaluation results that theproposed algorithm is fairly effective. Especially, theproposed method contains two-step multiplications ineach procedure so that it requires less running time,and the SNR out of this proposed method is higherwhen the speech signals are degraded by the colorednoise. It is concluded that this proposed algorithm issimpler and can realize the good noise suppressiondespite the reduction of the computational complexitywithout sacrificing the quality of the speech signal. Inthe further study, we will improve the adaptivealgorithm based on this paper to make it a moreaccurate assessment of environmental noise. On theother hand, the algorithm will be applied to theembedded-speech-recognition system at the hardwarelevel, so that it can improve the robustness of thesystem.References [1] .Quanshen Mai, Dongzhi He, Yibin Hou, Zhangqin Huang” A Fast Adaptive Kalman Filtering Algorithm for Speech Enhancement” IEEE International Conference on Automation Science and Engineering Aug, 2011 pp. 327 – 332. [2] ZHANG Xiu-zhen, FU Xu-hui, WANG Xia, Improved Kalman filter method for speech enhancement.Computer Applications,Vol.28, pp.363 365, Dec.2008. 1856 | P a g e