NOISE ESTIMATION USING STANDARD DEVIATION OF THE FREQUENCY MAGNITUDE SPECTRUM FOR MIXED NON-STATIONARY NOISE

Transcription

1 DOI: /jct ESTIMATION USING STANDARD DEVIATION OF THE FREQUENCY MAGNITUDE SPECTRUM FOR MIXED NON-STATIONARY A. Indumath 1 and E. Chandra 1 Department of Master of Computer Applcatons, Dr. SNS Rajalakshm College of Arts and Scence, Inda E-mal: endhumath@gmal.com Department of Computer Scence, Bharathar Unversty, Inda E-mal: crcspeech@gmal.com Abstract Nose estmaton and suppresson s very mportant for mprovng the qualty of speech sgnal. Noses exst n almost all places. In realty, more than one nose degrades the speech sgnal. It s hard to fnd and supress varous types of nose that affect the speech qualty. Ths paper proposed a method for nose estmaton of mxed non-statonary nosy speech sgnal. Ths method uses Spectral propertes of the nosy speech sgnal to detect the frequency regons of nose sgnal. Hghest frequency of speech sgnal s calculated and t s consdered as the threshold value for separatng nose sgnal and clean speech sgnal. Usng Spectral subtracton, Standard devaton of nose spectrum s subtracted wth nosy spectrum to acqure enhanced speech sgnal. Performance of the method s evaluated usng SNR and Spectrogram. The man focus of ths paper s to propose an ndependent method whch estmates the nose of any type and nature. Keywords: Nose Estmaton, Spectral Subtracton, Standard Devaton, SNR, Spectrogram 1. INTRODUCTION Speech s the natural mode of communcaton for human. They express ther mnds through speech. Speech sgnals have rch and temporal spectral varaton that conveys words, ntenton, expresson, emoton, gender and speaker dentfcaton [3]. In real tme stuaton, the speech receved by human s not the clear speech. The qualty of speech s degraded by varous background noses. Nose exts everywhere wth dfferent causes and wth dfferent features. Some of the noses are Addtve nose whch gets added to the proposed sgnal lke Background nose due to Bogus sounds, Termnal nose durng Sgnal capture and Envronment nose due to machne, Industry, Transport, Reverberaton and Echo. Many research n recent years have been mplemented varous technque to reduce the nose. It s dffcult to fnd the factors for nose and hard to cancel the specfc nose. Ths paper ntroduces a method to suppress all background noses nvarable of nose type.. RELATED WORK In recent studes, several flter desgns have been mplemented n Speaker recognton to suppress and elmnate the unwanted background nose, as well as to mprove speech qualty [1]-[4]. Implemented a technque for removal of hgh frequency nose for speech enhancement usng Frequency Response Maskng (FRM) based on desgnng low complexty, narrow transton bandwdth and lnear phase Fnte Impulse Response (FIR) flters [1]. Nose spectrum s calculated from speech absence frames through Voce Actvty Detector (VAD) or Mnmum Statstc methods [5]. Many Nose estmaton algorthms are proposed for strange non-statonary nose sgnals [5-7]. Martn [5] proposed an algorthm for estmatng the nose by trackng the varous nose level usng Mnmum Statstcs (MS) [5]. Cohen proposed an algorthm for nose estmaton by trackng the nose only regon of the nosy speech spectrum called Mnma Controlled Recursve Algorthm (MCRA) [6]. R. Sundarrajan and C.L. Phlpos proposed a method for comparng the rato of the nosy speech sgnal to the local mnmum aganst a threshold [7]. Doblnger [8] proposed a method for estmatng the nose spectrum for the speech sgnal, the man drawback of ths algorthm s nose estmaton ncreases smultaneously wth ncrease n nosy speech power. Cohen and Berdugo proposed a method for nose estmaton. In ths, nose estmaton s updated contnuously by averagng the past spectral values of the nosy speech wth tme and frequency dependent smoothng factors [9]. Hresh H.G and Ehrlcher C [13] proposed a method based on estmatng a hstogram of past spectral values, whch are compared aganst a Threshold. Threshold s based on the past nose estmaton. Ths algorthm fals to adapt when the nose estmaton suddenly ncrease. Marc Karam et al. proposed a method for Nose removal for real tme data usng Spectral subtracton [14]. Spectral Subtracton [10]-[1] s a tradtonal approach for reducng addtve background nose n Sngle channel system. Spectral Subtracton wth Partal Dfferental Equaton (PDE) method has been proposed by [15]. In ths method, nput speech sgnal s enhanced usng PDE and the qualty of speech sgnal s mproved through Spectral Subtracton. M. Thrumara Chellapand and P. Kablan [16] proposed a method to mprove the qualty of speech for moble applcaton by reducng acoustc nose usng Spectral Subtracton and Lnear Predcton Codng (LPC). Poornaprya G et al. [17] proposed a method to solate the nose from nosy speech sgnal for moble phone users. Wavelet based Spectral Subtracton technque s mplemented to enhance the speech sgnal by reducng the nterferences. Israel Cohen, Baruch Berdugo [18] presented a speech estmator OM-LSA and a nose estmaton method for Nonstatonary envronment. Ln. L, E. Ambkarajan, W.H. Holmes [19] proposed a technque usng audtory flter bank for denosng the speech sgnal whch s sutable for Non-statonary and colored nose envronment. 118

2 ISSN: (ONLINE) ICTACT JOURNAL ON COMMUNICATION TECHNOLOGY, DECEMBER 015, VOLUME: 06, ISSUE: 04 To overcome the drawback of Short tme Spectral attenuaton (STSA) Kotta [0] proposed a post processng method to detect nose domnaton regon whch are attenuated usng a SNR Based rule. 3. SYSTEM OUTLINE Spectral Subtracton and Lnear Predcton suts good for eradcaton of Background nose and resdual Nose [0]. Block dagram of our proposed system s shown below. Nose Estmaton Wndowng and Segmentng FFT Spectral Magntude Lnear Predcton Spectral Subtracton IFFT Clear Speech Fg.1. Block dagram of proposed method Phase Informaton Speech s non-statonary sgnal whch s affected by varous envronmental noses. Propertes of speech sgnal changes rapdly over tme. Speech sgnal s represented as, s(n) = x(n) + d(n) (1) where, x(n) s clean speech sgnal, d(n) s nose sgnal and s(n) s Nosy speech sgnal. In general, nose propertes change qut rapdly over tme n non-statonary envronment. Nosy speech sgnal remans nvarant for a short perod of tme [13]. The speech sgnal s dvded nto the frames by usng Hammng wndow. In wndowng, the begnnng and end of the sgnal s attenuated n the calculaton of the spectrum. Hence overlappng s performed to recover a porton of each prevous frame that s lost due to wndowng. Overlappng ensure better temporal contnuty n the transform doman. An overlap of half the wndow sze s typcal [6]. Thus nput speech sgnal s segmented nto frames wth overlap of 1/ of the frame sze. The tme doman of framed speech sgnal s denoted as, s (n) = s 1(n) + s (n) + s 3( n) + +s (n) () where, n ranges over 1-56 samples and denotes the frame number correspondng to the tme-doman frame. Phase value s calculated and stored for later. FFT s performed to obtan the magntude frequency response of each frame. In FFT, the sgnal wthn a frame s consdered as perodc, and contnuous when wrappng around. Lnear Predcton s one of the most mportant tools n speech processng. LP s a useful method for estmatng parameters of recorded speech sgnal. The output samples are predcted as a lnear combnaton of flter co-effcent and prevous samples. For any nosy speech sgnal, back ground nose s to be suppressed. Spectral subtracton s commonly used algorthm for speech enhancement. The basc prncple of spectral subtracton s subtractng nose spectrum from the nosy speech spectrum, whch results n clean speech spectrum. IFFT s used to reconstruct the spectral speech sgnal and transforms Frequency doman sgnal nto Tme doman enhanced speech sgnal. 4. PROPOSED ESTIMATION ALGORITHM Nose suppresson s an mportant factor of modern communcaton systems. Pre-processng of nosy speech sgnal mproves the performance of speech communcaton system for sgnals that was corrupted by statonary envronment nose, through mprovng the speech qualty or ntellgblty. It s dffcult to suppress non statonary nose wth pre-processng. The proposed method presents a Nose estmaton technque wth Spectral subtracton whch well suted for rapdly varyng nose n non-statonary envronment. Nose estmaton and Spectral Subtracton cannot be successfully mplemented n tme doman. Hence, tme doman sgnal s converted nto frequency doman sgnal usng FFT. Nosy speech sgnal s segmented, wndowed and half overlapped. Spectral magntude of nosy speech sgnal s obtaned through FFT. Spectral magntude of the nosy speech sgnal s fltered usng Lnear Predcton. Hghest frequency of speech sgnal s calculated from nosy speech sgnal and t s treated as Threshold value. Ths value helps to calculate the frequency range of Nose sgnal from Nosy speech sgnal. In proposed method, Nose s estmated by usng mean and Standard devaton. As the characterstcs of sgnal changes rapdly n non-statonary sgnal, the slowly changng mean nterferences wth the calculaton of the standard devaton leads to naccurate value [14]. Ths problem s overcome by dvdng the sgnal nto small sectons and calculatng the statstcs for each secton ndvdually and Standard devaton for each of the secton s averaged to produce a sngle accurate value. The mean value of sgnal s calculated from, 1 N 1 x 0 (3) N where, μ s mean value of sgnal, x s a sgnal, s an ndex whch vares from 0 to N-1. Sum of the values n sgnal x s calculated as, x 0 + x 1 + x + x 3 + x N-1 (4) Ths resultant value s dvded by N to derve the mean of the sgnal. x (5) 119

3 The above expresson descrbe how far the th sample devaton from the mean. Standard Devaton of the sgnal s found by summng the dervatons of all the ndvdual samples and then dvdng by the number of samples N. The Standard Devaton s calculated from, 1 N 1 x 0 (6) N 1 1 N 1 x 0 (7) N 1 where, σ s Standard Devaton, N s number of samples. Mean value descrbes what s beng measured and Standard Devaton represents nose and other nterference n sgnal. 5. SPECTRAL SUBTRACTION Spectral subtracton s used to subtract estmated Standard Devaton of the nose spectrum from the spectrum of nosy speech. The sgnal receved s the sum of clean speech and nose. s(n) = x(n) + d(n) (8) where, s(n) s nosy speech sgnal, x(n) s clean speech sgnal and d(n) s nose sgnal. Once t s framed and transformed nto frequency doman t s represented as, s (k) = x (k) + d (k) (9) where, k ranges over 1-56 samples and ranges over the number of frames. Magntude spectrum of nosy speech s(n) s represented as, M (k) = S (k) (10) Lkewse magntude of each frame s calculated. Fg.. Spectral subtracton Ths magntude and nose estmate whch s already estmated through proposed method s used for Spectral subtracton. where, Nosy Spectrum E k M M(k) k N k f M k N k 0 f M k N k E (k) represents the estmated clean spectrum M (k) represents the Nosy spectrum (11) N (k) represents nose estmate of our proposed method Spectral subtracton s restrcted to use postve magntude spectrum. Thus clean spectrum of each frame s estmated. Estmated clean magntude spectrum E (k) s combned wth phase spectrum θ (k) C N(k) E (k) Estmated Nose j k k E k e Estmated Clean Spectrum. (1) Accordng to Boll [10], the phase value of nosy speech spectrum s suffcent to use for an estmaton of clean speech sgnal after Spectral subtracton. w w s y (13) Inverse FFT of C (k) results tme doman frames to reconstruct enhanced speech sgnal. 6. PERFORMANCE EVALUATION Nose sgnal have dfferent propertes and the way t affects the speech sgnal also vares. Hence the performance of the proposed method s tested wth varous Nosy speech samples. Whte nose, Pnk nose and Factory nose are consdered as background noses whch were taken from the X-9 database. The Fg.3 shows the temporal results of Nosy speech, Clean speech and Estmated Speech sgnal. Ths Fgure says Clean Speech Sgnal and Estmated Speech sgnal resembles same. Fg.3. Nosy speech, clean speech and estmated speech sgnal The performance of proposed method s evaluated through () SNR, () Spectrogram, () Perceptual Evaluaton of Speech Qualty, (v) Tme and Frequency Sgnal to Nose Rato. 6.1 SIGNAL TO RATIO (SNR) As speech s processed n frames, SNR s computed for each segment. SNR value for th segment s defned as defned as, M 1 n0 n s n0 SNR 10log (14) M 1 n n where, S and n are speech and nose sample n the th segment. Averages of these segments are calculated as SNR value of speech sgnal. SNR value for Nosy Speech sgnal (Pnk, Whte and Babble) are compared wth ts correspondng clean speech sgnal. The Table.1 shows that the SNR value for clean speech sgnal s hgh whch results ncrease n hearng ablty and ntellgblty of speech sgnal. 10

4 ISSN: (ONLINE) ICTACT JOURNAL ON COMMUNICATION TECHNOLOGY, DECEMBER 015, VOLUME: 06, ISSUE: 04 Type of Nose Nosy Speech Wth Whte Nose Nosy Speech Wth Pnk Nose Nosy Speech Wth Factory Nose Nosy Speech Wth Multple Nose 6. SPECTROGRAM Table.1. Performance Comparson Nose (db) SNR Before Nose Estmaton SNR After Proposed Nose Estmaton Spectrogram s a vsual qualty comparson of speech sgnal [16].The spectrogram of nosy speech sgnal and enhanced speech sgnal s shown n the Fg.4. rato s measured usng orgnal speech sgnal and processed sgnal. It s calculated through, Nm Nm N 1 x n 10 M 1 nnm SNR seg log (15) M m0 N 1 xn xˆ n nn where, x(n) s the orgnal sgnal, xˆ n s the enhanced sgnal, N s the frame length and M s the number of frames n the sgnal. The Segmental SNR n terms of frequency doman yelds the frequency weghted segmental SNR. f wsnr seg s gven by, 10 M 1 fwsnr seg M m0 m k F m, j B log j 1 j 10 F m, j Fˆ m, j k E j 1B j (16) where, B j s the weght placed on the j th frequency band, k s the number of bands, M s the total number of frames, F(m, j) s the flter bank ampltude of the orgnal sgnal and F ˆ m, j s the flter bank ampltude of resultant speech sgnal. 7. CONCLUSION Ths paper proposed a new method for nose estmaton whch suts for nose suppresson Spectral Subtracton algorthm. Nose s estmated by fxng the threshold value whch separates nose range from speech sgnal range. Unlke other methods, Nose estmaton dd not depend on specfc nose and t suts for any specfc nose. From the expermental results and performance of SNR and Spectrogram, nose estmaton method effectvely supports for nose suppresson. REFERENCES Fg.4. Spectrograms of Nosy Speech, Clean Speech and Estmated Speech 6.3 PERCEPTUAL EVALUATION OF SPEECH QUALITY (PESQ) The PESQ s the most complex objectve measure to compute the Qualty of speech. Ths evaluaton method focus on end-to end behavour ncludng the effects of flterng, lstenng frequency and the perceved loudness. The PESQ provdes a speech qualty score based on the comparson between the orgnal speech sgnal and the sgnal degraded by the network under test. Psychoacoustc and Cogntve models are used to perform ths test. 6.4 TIME AND FREQUENCY SIGNAL TO RATIO The segmental sgnal to nose rato s evaluated ether n terms of tme or frequency doman. Segmental sgnal to Nose [1] K.P. Hymavathy and P. Janardhanan, Nose Flterng n Speech Usng Frequency Response Maskng Technque, Internatonal Journal of Emergng Trends n Engneerng and Development, Vol., No. 3, pp , 013. [] Supavt Muangjaroen and Thaweesak Yngthawornsuk, A Study of Nose Reducton n Speech Sgnal Usng FIR Flterng, Proceedngs of Internatonal Conference on Advances n Electrcal and Electroncs Engneerng, pp , 01. [3] T.L. Kumar and K.S. Rajan, Nose Suppresson n Speech Sgnals Usng Adaptve Algorthms, Internatonal Journal of Engneerng Research and Applcatons, Vol., No. 1, pp , 01. [4] Rajeev Aggarwal, Ja Karan Sngh, Vjay Kumar Gupta, Sanjay Rathore, Mukesh Twar and Anubhut Khare, Nose Reducton of Speech Sgnal Usng Wavelet Transform wth Modfed Unversal Threshold, Internatonal Journal of Computer Applcatons, Vol. 0, No. 5, pp , 011. [5] R. Martn, Nose power spectral densty estmaton based on optmal smoothng and mnmum statstcs, IEEE 11

5 Transactons on Speech Audo Processng, Vol. 9, No. 5, pp , 001. [6] Israel Cohen, Speech Enhancement Usng a Noncausal a Pror SNR Estmator, IEEE Sgnal Processng Letters, Vol. 11, No. 9, pp , 004. [7] Sundarrajan Rangachar and C.L. Phlpos, A Nose- Estmaton Algorthm for Hghly Non-Statonary Envronments, Speech Communcaton, Vol. 48, No., pp. 0-31, 006. [8] Gerhard Doblnger, Computatonally Effcent Speech Enhancement by Spectral Mnma Trackng n Subbands, Proceedngs of European Conference on Speech Communcaton and Technology, pp , [9] I. Cohan and B. Berdugo, Nose Estmaton by Mnma Controlled Recursve Averagng for Robust Speech Enhancement, IEEE Sgnal Processng Letters, Vol. 9, No. 1, pp. 1-15, 00. [10] S. Boll, Suppresson of Acoustc Nose n Speech Usng Spectral Subtracton, IEEE Transactons on Acoustcs Speech and Sgnal Processng, Vol. 7, No., pp , [11] P. Lockwood and J. Boudy, Experments wth a Nonlnear Spectral Subtracton (NSS), Hdden Markov Models and Projecton for Robust Recognton n Cars, Speech Communcaton, Vol. 11, No. -3, pp. 15-8, 199. [1] Hdetosh Nakashma, Yoshfum Chsak, Tsuyosh Usagawa and Masanao Ebata, Spectral Subtracton based on Statstcal Crtera of the Spectral Dstrbuton, IEICE Transactons on Fundamentals of Electroncs, Communcatons and Computer Scences, Vol. E85-A, No. 10, pp. 83-9, 00. [13] H.G. Hrsch and C. Ehrlcher, Nose Estmaton Technques for Robust Speech Recognton, Proceedngs of Internatonal Conference of Acoustcs, Speech, and Sgnal Processng, Vol. 1, pp , [14] Marc Karam, Hasan F. Khazaal, Heshmat Aglan and Clston Cole, Nose Removal n Speech Processng Usng Spectral Subtracton, Journal of Sgnal and Informaton Processng, Vol. 5, No., pp. 3-41, 014. [15] D. Deepa and A. Shanmugam, Spectral Subtracton Method of Speech Enhancement Usng Adaptve Estmaton of Nose wth PDE Method as a Preprocessng Technque, ICTACT Journal of Communcaton Technology, Vol. 1, No. 1, pp. 1-6, 010 [16] M. Thrumara Chellapand and P. Kablan, Evaluaton of Speech Enhancement n Nosy Condtons Usng a Spectral Subtracton and Lnear Predcton Combnaton, ICTACT Journal of Communcaton Technology, Vol. 3, No. 4, pp , 01. [17] G. Poornaprya, R. Shanmugasundaram and N. Santhyakumar, Clatter Dmnshng for Moble Telephony, Proceedngs of Internatonal Conference of Innovatons n Informaton, Embedded and Communcaton Systems, pp. 1-4, 015. [18] Israel Cohen and Baruch Berdugo, Speech Enhancement for Non-statonary nose envronment, Sgnal processng, Vol. 81, No. 11, pp , 001. [19] L. Ln, E. Ambkarajah and W.H Holmes, Speech Enhancement for Non-statonary Nose Envronment, Proceedngs of Asa Pacfc Conference on Crcuts and Systems, Vol. 1, pp , 00. [0] Kotta monahar and Preet Rao, Speech Enhancement n Non-statonary Nose Envronment Usng Nose Propertes, Speech Communcaton, Vol. 48, No. 1, pp , 006. [1] Sundarrajan Rangachar, Phlpos C. Lozou and Y Hu, A Nose Estmaton Algorthm wth Rapd Adaptaton for Hghly Non-statonary Envronments, Proceedngs of IEEE Internatonal Conference on Acoustc, Speech and Sgnal Processng, Vol. 1, pp , 004. [] Soo-Jeong Lee and Soon-Hyob Km, Nose Estmaton based on Standard Devaton and Sgmod Functon Usng a Posteror Sgnal to Nose Rato n Non-statonary Nosy Envronments, Internatonal Journal of Control, Automaton and Systems, Vol. 6, No. 6, pp ,