International Journal of Scientific & Engineering Research, Volume 4, Issue 10, October ISSN

Transcription

1 Iteratioal Joural of Scietific & Egieerig Research, Volume 4, Issue, October-3 48 A ovel Noise estimatio by histogram equalizatio ad optimal filterig for robust speech ehacemet M.Shravai, P.Chadra Sekhar, Ch. Gaapathy Reddy 3 Post graduate studet, ECE Dept, G.Narayaamma istitute of techology ad sciece, Hyderabad, A.P, Idia, Assistat Professor, ECE Dept, G.Narayaamma istitute of techology ad sciece, Hyderabad, A.P, Idia, 3 Professor,ECE Dept,G.Narayaamma istitute of techology ad sciece,hyderabad,a.p,idia srv5mg@yahoo.com, chadrasekharpaseddula@gmail.com, gaapathi7898@yahoo.co.i Abstract: Geerally, Speech ehacemet aims to improve speech quality by usig various algorithms. The objective of ehacemet is improvemet i overall perceptual quality of degraded speech sigal usig audio sigal processig techiques. I earlier, there are so may algorithms proposed for speech ehacemet. But they are ot able to ehace the speech effectively by reducig the oise compoets. Recetly a ew mathematical algorithm called Empirical mode decompositio (EMDF) method was proposed. Though this algorithm ehaces the speech effectively the take to process this ehacemet is too high. Maily this is because of the IMF evaluatios for the complete speech samples. To overcome this issue this paper proposes a Histogram based speech ehacemet techique. The histogram proposed i this work estimates the oise compoets cotamiated with the clea speech samples, a optimal filterig is proposed to filter those estimated oise samples. I. Itroductio Speech ehacemet plays a importat role i umerous applicatios such as hearig aids speech codig cell phoes automatic recogitio of speech sigals by machies ad may more. Speech sigals from the ucotrolled eviromet may cotai degradatio compoets alog with the required speech compoets. Degradatio compoets iclude back groud oise reverberatio ad speech from other speakers. Therefore the degraded speech compoets eed to be processed for the ehacemet. Speech ehacemet algorithms improve the quality ad itelligibility of speech by reducig or elimiatig the oise compoet from the speech sigals. Improvig quality ad itelligibility of speech sigals reduce listeer s exhaustio; improve the performace of hearig aids speech coders ad may other speech processig systems. I most speech ehacemet algorithms it is assumed that a estimate of oise spectrum is available. Noise estimate is critical part ad it is importat for speech ehacemet algorithms. Performace of speech ehacemet algorithms depeds o correct estimatio of oise. Simple approach to estimate the oise spectrum of the sigal usig a Voice Activity Detector (VAD) [,,3,4]aother approach to estimate the oise usig differet oise estimatio algorithms Noise estimatio algorithms that cotiuously track the oise spectrum. If the VAD approach is coservative, the it will attempt to reduce false alarms for silece detectio, which results i less frequet oise power updates. I highly o-statioary eviromets, the oise power must be tracked eve durig speech activity. Noise estimatio techiques which operate i the short- Fourier trasform (STFT) domai are very popular, icludig ewer oise estimatio systems such as the miimum statistics (MS) [5] ad the improved miima cotrolled recursive averagig (IMCRA) [6]. These techiques estimate the oise spectrum based o the observatio that the oisy sigal power decays to values characteristic of the cotamiatig oise durig speech pauses. The mai challege faced by these techiques is trackig the oise power durig speech segmets. This would result i poor estimates durig log speech segmets with few pauses. Speech ehacemet systems such as the optimally modified log-spectral amplitude (OMLSA) estimator [7] require a oise estimate to suppress oise ad ehace the oisy speech.. I [], speech ehacemet i car iterior oise is achieved by usig a speech aalysis sythesis approach, based o a harmoic oise model, as post processig after a traditioal log-spectral amplitude speech estimatio system. This system is sesitive to accurate pitch estimatio ad voiced/uvoiced speech frame classificatio. Recetly a ew method for aalyzig oliear ad o-statioary data has bee developed. The key part of the method is the empirical mode decompositio [8-] method with which ay complicated data set ca be decomposed ito a fiite ad ofte small umber of itrisic mode fuctios. This decompositio method is adaptive, ad, therefore, highly efficiet. Sice the decompositio is based o the local characteristic scale of the data, it is applicable to oliear ad o-statioary processes. The mai problem associated with the implemetatio of EMDF to ehace the speech is it takes too much, as well as it is also ot applicable to those oisy speech sigals which are cotamiated with the oise havig same power 3

2 Iteratioal Joural of Scietific & Egieerig Research, Volume 4, Issue, October-3 48 spectral desity at low frequecies with highly ostatioary eviromets. To overcome this issue this paper proposes a histogram [] based oise estimatio which gives a effective PSD characteristics ad optimal filterig based speech ehacemet. The rest of the paper is orgaized as follows: Sectio II gives the basic details about the backgroud of empirical mode decompositio ad histogram extractio. The proposed histogram based oise estimatio ad the kalma filterig for speech ehacemet is illustrated i sectioiii. The performace evaluatio of the proposed approach is illustrated i sectioiv; fially the coclusios are illustrated i sectiov. II.Back groud This sectio gives the basic details about the Empirical mode decompositio ad the histogram evaluatio. This sectio is orgaized uder two parts. The first part gives the details about the Empirical mode decompositio which is used recetly for speech ehacemet. O the other had the secod part gives the basic details about the histogram equalizatio, applied o the speech sigal cotamiated with various types of oises. A.EMDF EMD is a method of breakig dow a sigal without leavig the domai. It ca be compared to other aalysis methods like Fourier Trasforms ad wavelet decompositio. The process is useful for aalyzig atural sigals, which are most ofte oliear ad o-statioary. This parts from the. Idetify all the local extrema i the test data.. Coect all the local maxima by a cubic splie lie as the upper evelope. 3. Repeat the procedure for the local miima to produce the lower evelope. The upper ad lower evelopes should cover all the data betwee them. Their mea is m. The differece betwee the data ad m is the first compoet h : X(t) m = h Ideally, h should satisfy the defiitio of a IMF, for the costructio of h described above should have made it symmetric ad havig all maxima positive ad all miima egative. After the first roud of siftig, a crest may become a local maximum. New extrema geerated i this way actually reveal the proper modes lost i the iitial examiatio. I the subsequet siftig process, h ca oly be treated as a proto-imf. I the ext step, it is treated as the data, the h m = h After repeated siftig up to k s, h becomes a IMF, that is h (k ) m k = h k The, it is desigated as the first IMF compoet from the data: c = h k At the ed of the decompositio, the data s(t) will be represeted as a sum of IMF sigals plus a residue sigal, s(t) = c i (t) + r (t) assumptios of the methods we have thus far leared (amely that the systems i questio be LTI, at least i approximatio). The EMD method is a ecessary step to reduce ay give data ito a collectio of itrisic mode fuctios (IMF) to which the Hilbert spectral aalysis ca be applied. A IMF is defied i= The fially obtaied s(t) gives the completely deoised sample of origial speech, though it is efficiet to ehace the speech sample it takes too much to ehace as well as it is ot applicable all types of oise cotamiated sigals. To over this problem a ovel speech ehacemet techique is as a fuctio that satisfies the followig proposed i this paper ad the complete details are requiremets: provided i further sectios.. I the whole data set, the umber B.Histogram of extrema ad the umber of zerocrossigs I geeral, a histogram is a graphical represetatio must either be equal or differ at most by oe. of the distributio of data. It is a estimate of the probability distributio of a cotiuous variable. A. At ay poit, the mea value of the evelope histogram is a represetatio of tabulated defied by the local maxima ad the frequecies, show as adjacet rectagles, erected evelope defied by the local miima is zero. Therefore, a IMF represets a simple oscillatory over discrete itervals (bis), with a area equal to the frequecy of the observatios i the iterval. The height of a rectagle is also equal to the frequecy mode as a couterpart to the desity of the iterval, i.e., the frequecy divided by simple harmoic fuctio, but it is much more geeral: istead of costat amplitude ad frequecy i a simple harmoic compoet, a IMF ca have variable amplitude ad frequecy alog the axis. The procedure of extractig a IMF is called siftig. The siftig process is as follows: the width of the iterval. Cosider a speech sigal {x} ad let i be the umber of occurreces of pauses i. The probability of a occurrece of a pause of level i i the speech is p x (i) = p(x = i) = i i < L 3

3 Iteratioal Joural of Scietific & Egieerig Research, Volume 4, Issue, October L beig the total umber of pauses i the image, beig the total umber of occurreces i the speech, ad p x (i)beig i fact the histogram for occurrece value i, ormalized to [,]. Let us also defie the cumulative distributio fuctio correspodig to p x as cdf x (i) = p x (i) j= which is also the speech's accumulated ormalized histogram. We would like to create a trasformatio of the form y = T(x) to produce a ew speech {y}, such that its CDF will be liearized across the value rage, i.e. cdf x (i) = ik for some costat K. The properties of the CDF allow us to perform such a trasform (see Iverse distributio fuctio); it is defied as y = T(x) = cdf x (x) Notice that the T maps the levels ito the rage [,]. I order to map the values back ito their origial rage, the followig simple trasformatio eeds to be applied o the result: y = y. (max{x} mi{x}) + mi {x} i Figure: Histogram plot values spaig a widow of several hudreds of millisecods ad take as a estimate of the oise spectrum the value correspodig to the Maximum of the histogram values. This is doe separately for each idividual frequecy bi. The histogram based oise estimatio is summarized [3] as follows. Compute the power spectrum of a oisy speech y(λ, k) ².. Smooth the oisy psd usig st order recursio. P(λ,k) = αp(λ, k) +( α) y(λ, k) ² Where α is smoothig costat. 3. Compute the histogram of D part PSD estimates P(λ, k) &P(λ-, k) P(λ-, k), P(λ-D,k) usig say l bis 4. Let C = [C, C, ---- C l ] be the couts i each of the 4 bis i the histogram ad S = [S, S, S l ] deote the correspodig ceters of the histogram bis. 5. Let C max be the idex of the Maximum Cout C max III. Proposed Approach A. Estimatio of oise I a geeral mathematical sese, a histogram is a fuctio that couts the umber of observatios that fall ito each of the disjoit categories kow as bis, whereas the graph of a histogram is merely oe way to represet a histogram as show i figure-. = arg Max (C i ) for I< i < l. The take a estimate of Histogram based oise estimatio algorithms are the oise psd deoted by N max (λ, k) the value motivated by the observatio that the Most frequet correspodig to the maximum of the histogram N max value (that is the Histogram maximum) of eergy (λ, k) = P( C max ). values i idividual frequecy bads correspods to 6. Smooth the oise estimate N max (λ, k) usig st the oise level of the specified frequecy bad, that is order recursio the oise level correspods to the maximum of the r²(λ, k) = αmd² (λ-, k)+( αm) N max (λ, k) histogram of eergy values. I some cases, the Where r² (λ, k) is the smoothed estimate of the oise histogram of spectral eergy values may cotai two psd ad αm is a smoothig costat. modes lst a low eergy mode correspodig to the B. Optimal Filterig speech abset ad low eergy segmets of speech ad d a high eergy mode correspodig to the If we assume that the speech ad oise sigals are (oisy) voiced segmets of speech. The oise idepedet ad statioary (eve though it is oly estimate is obtaied based o the histogram of part approximately true), we ca use the o-causal power spectrum values [] that is for each i comig optimal filter ([6],[5],[4]) to fid a gai factor g w. frame, st costruct the histogram of power spectrum g w = r sx = r x r r x r x The equality r sx =r s comes from the idepedece of origial y s ad y from y x =y s +y. Usig the certaity equivalece priciple ([7]) y=f(x) y =f(x ). We get: g w = r s r r s I practice we wat to assure the o-egativity of g w. I.e.halfwave rectificatio 3

4 Iteratioal Joural of Scietific & Egieerig Research, Volume 4, Issue, October Fially we get r s =g w r x = max {,r x r } r x Fially the obtaied r s gives the deoised speech sigal. The performace of the proposed approach is illustrated i ext sectio. IV.Performace Evaluatio This sectio gives the complete details about the performace of the proposed approach. The performace uder this sectio is evaluated for various types of oises. To test the proposed approach we have cosidered the three types of oisy speech sigals. Those are babble oisy speech, restaurat oisy speech ad car oisy speech. Now each source sigal has, samples ad havig the samplig frequecy of 6, Hz. The accuracy of the recovered sigal r() compared to the desired speech sigal ca be measured by the sigal to oise ratio which is give by r SNR = log r 5 - = 5 r e r x r s Where r e is the error sigal. The followig figures (- 6) give the performace evaluatio of the proposed (d) approach. Oroigal sample.4 Spectogram for origial speech Oroigal sample (e) Figure3: restaurat oise at 5db, : Spectrogram, : Histogram of origial sample, (d):, (e): Spectrogram of deoised sample Oroigal sample x Spectogram for origial speech Spectogram for deoised speech Spectogram for origial speech x x (d) (e) Figure: Babble oise sigal at 5db, : Histogram of babble at 5db, : Spectrogram of babble at 5db, (d):, (e): Spectrogram of deoised sample The above figures from figure are the test samples belogig to babble oisy speech. The histogram of this sample is show i figure 3 with bis Spectogram for deoised speech.5 x (d) (e) Figure4: car oisy speech at 5db, : Spectrogram, : Histogram of car oisy speech, (d):, (e): Spectrogram of deoised sample Spectogram for deoised speech 3

5 Iteratioal Joural of Scietific & Egieerig Research, Volume 4, Issue, October The table give below shows the SNR values Processig. Amsterdam, The Netherlads: Elsevier, obtaied foe the above tested samples. Nov., vol. 8, pp [8] P. Fladri et al., Detredig ad deoisig with TABLEI. Segmetal SNR for various types of oisy speech empirical mode decompositios, i Proc. Eur. samples Sigal Process. Cof. (EUSIPCO), 4, pp. 58 Sample db Seg SNR 584. Car oise 5db 6.34 [9] K. Khaldi et al., Speech ehacemet via Babble oise 5db.646 EMD, i Proc. EURASIP J. Adv. Sigal Process., Restaurat oise 5db.763 8, vol. 8, p. 8. [] Y. Kopsiis ad S. McLaughli, Developmet Table I shows the average segmetal SNR obtaied of EMD-based deoisig methods ispired by for various oise types ad at a 5dB oise level. The wavelet thresholdig, IEEE Tras. Sigal Process., proposed approach cosistetly achieves a higher vol. 57, o. 4, pp , Apr. 9. improvemet i the segmetal SNR. Its advatage is [] Auradha R. Fukae, Shashikat L. Sahare, more sigificat i o-statioary oise Noise estimatio Algorithms for Speech eviromets. Ehacemet i highly o-statioary Eviromets, IJCSI Iteratioal Joural of V. Coclusios Computer Sciece Issues, Vol. 8, Issue, March This paper proposes a ovel speech ehacemet techique to reduce the oise compoets ISSN (Olie): cotamiated i clea speech samples. This proposes [] Hirsch, H., Ehrlicher, C., (995) Noise a histogram based oise estimatio method to estimatio Techiques for robust speech recogitio. estimate the o-statioary oise compoets, a Proc. IEEE Iterat. Cof. o Acoust. Speech Sigal optimal filterig cocept to remove those estimated Proc.pp [3] P. C. Loizou, Speech oise compoets. The performace of this techique Ehacemet: Theory ad Practice st ed. Boca was evaluated usig speech cotamiated with car Rato, FL. CRC, 7 iterior oise, babble oise, ad restaurat oise [4] M. Hayes, Statistical Digital Sigal Processig coditios. Whe compared to a IMCRA, EMDF ad Modelig, Joh Wiley ad Sos, 996. systems, this method was show to give improved [5] Z. Hrdia, Statistick a radiotechika, skriptum performace at suppressig backgroud oise uder ˇCVUT, Praha, 996. the preseted oisy coditios. [6] M. Sambur, Adaptive Noise Cacelig for Speech Sigals, IEEE Tras. o Acoustics, Speech Refereces ad Sig. Proc., October 978. [] Soh J. ad Kim N.(999), Statistical Model [7] J. ˇ Stecha a V. Havlea, Moder ı teorie based voice activity detectio, IEEE Sigal ˇr ıze ı, skriptum ˇCVUT, Praha, 996. Proc.Lett.6(), -3. [] Tayer S. ad Ozer H. (), Voice Activity Detectio i No statioary Noise, IEEE Speech Audio Procc.8 (4), pp [3] Shriivasa K. ad Gersho A. (993), Voice Activity Detectio for Cellular Network, Prec. IEEE Speech Codig Workshop pp [4] Haigh J. ad Maso J.(993) Robust Voice Activity Detectio usig Ceptral Features, Prec. IEEE TENCON 3-34A. [5] R. Marti, Noise PSD estimatio based o optimal smoothig ad miimum statistics, IEEE Tras. Speech Audio Process., vol. 9, o. 5, pp. 54 5, Jul.. [6] I. Cohe, Noise spectrum estimatio i adverse eviromets: Improved miima cotrolled recursive averagig, IEEE Tras. Speech Audio Process., vol., o. 5, pp , Sep. 3. [7] I. Cohe ad B. Berdugo, Speech ehacemet for o-statioary oise eviromets, i Sigal 3