2007 IEEE International Symposiumon Signal Processing and Information Technology

Nonlinear Signal Processing for Voice DisorderDetection by Using Modified GP Algorithm and

Surrogate Data AnalysisAboozar Taherkhani, Seyyed Ali Seyyedsalehi, Arash Mohammadi, Mohammad Hasan Moradi

Faculty of Biomedical EngineeringAmirkabir University of Technology

Tehran, IranEmail: a.taherkhania c.ir, ssaIehi(_Waut.ac.ir, a.mohammad ut.ac.ir, hmorad ut.ac.ir

Methods based on nonlinear dynamics, including generalAbstract- Acoustic voice analysis is an effective, cheap and dimension (Hausdorff dimension, information dimension,

non-invasive tool that can be used to confirm the initial diagnosis correlation dimension, etc.), entropy (Kolmogrov entropy,and provides an objective determination of the impairment. The second-order entropy, etc.), and Lyapunov exponents, enablenonlinearities of the voice source mechanisms may cause the us to quantitatively describe chaotic behavior. Investigationsexistence of chaos in human voice production. Voice pathology of chaotic activities in physiologic systems suggest thatcan cause to addition colored noise to voice wave. Added noise toa chaotic signal causes reduction of the deterministic property cthaopgesy iolinardynamic measures andiae states ofand therefore increases correlation dimension of signal. pathophysiological dysfunction [3]. Poon and Merrill, [4] forSurrogate data analysis can measure this deviation and give a example, found that chaotic activity decreased incriterion for amount of noise added to the chaotic signal. By electrocardiogram (ECG) signals from patients withusing this criterion a threshold level is set to separate disordered congestive heart failure. Hornero et al [5] found that thevoice from normal voice and 95% accuracy is achieved. electroencephalogram (EEG) signals generated by

schizophrenic patients had a significantly lower correlation.INTRODOCTION dimension than the EEG signals of normal subjects. These

voice diseases are increasing dramatically, due examples suggest that chaos theory and nonlinear dynamicNowadays sc i habisand voie abuse.aItyi we methods might potentially be applied to diagnosemainly to unhealthy social habits and voice abuse. It s well physiological disorders and evaluate the effects of clinicalknown that voice diseases affect the quality of the voice treatments.register. These diseases should be diagnosed and treated at an Over the last two decades, observations in computer modelsearly stage. Acoustic voice analysis is an effective and non- of the vocal folds and nonlinear dynamic analysis of humaninvasive tool that can be used to confirm the initial diagnosis voices have established the existence of chaos in human voiceand provides an objective determination of the impairment. production. As noted by many researchers, the nonlinearitiesEarly detection and treatment of laryngeal tumors can reduce of the voice source mechanisms (eg, the nonlinear pressure-both morbidity and mortality. [1] Invasive methods need flow relation in the glottis, the nonlinear stress-strain curves ofmassive equipment and waste lots of money and time whereas vocal fold tissues, and the nonlinearities associated with vocalin the Acoustic methods, voice can be recorded for several fold collision) make this development unsurprising.times without taking much time, imposing much coast and Researchers have applied these new tools to studyingpain. Invasive methods give effective and necessary abnormal conditions associated with laryngeal pathologies, toinformation about disease, Therefore acoustic methods can be differentiate normal and pathologic voices and diagnoseused to primary detection of disease, the patient is sent to pathologies, and to assess the effects of clinical treatments.specialized centers to do invasive methods if it is necessary. Some traditional voice analysis methods, such as jitter and

According to the benefit of the acoustic voice analysis, shimmer, may be unreliable for analyzing aperiodic voices.improvement of these methods is recently considered. Roland Nonlinear dynamic methods provide informationet al. presented an acoustical feature extraction paradigm that complementary and nonredundant to existing analysisfocused on jitter, shimmer, standard deviation of fundamental methods [3].frequency, and the glottal-to-noise excitation ratio was used to In this paper we used modified standard Grassberger-analyses 120 voice samples. An improved artificial neural Proccacia (GP) algorithm that presented in [6] for estimatingnetwork (ANN) was used for classification. 80 00 of all voice the correlation dimension of a time series related to vowel /a/.samples could be classified correctly as either healthy or We apply the Surrogate data analysis and extract normalized

hoarse [2]. mean sigma deviation (nmsd). Evidences show that nmsd is aChaos has been observed in turbulence, chemical reactions, etrfauei oprsnwt orlto ieso.I

nonlinear circuits, the solar system, biological populations, Section II we first describe the methods and materials thatand seems to be an essential aspect of most physical systems. cotibreexlninsfdabs,mdfedGagrth

978-1 -4244-1 835-0/07/$25.00 ©2007 IEEE 1 171

and surrogate data analysis. In Section III the results of our For large values of R, a significant fraction of the M-experiences are brought. At last some discussions on our spheres used in the computation will typically go beyond theresults are presented in Section IV. attractor region. This "edge effect" leads to underestimation

of CM(R) for large R and finally causes CM(R) to saturate toII. MATERIALS AND METHODS unity. According to [6] a proper linear part in the

II. i. DATABASE logCm (R) versus log (R) is identified which is called the

The voice samples examined in this study were selected "scaling region" and its slope is taken to be D2.from the Disordered Voice Database [7], model 4337, version The original data set, st, is first transformed to a uniform1.03 (Kay Elemetrics Corporation, Lincoln Park, NJ), deviate, S,(t,). Note that s,(t,) ranges from 0 to 1, whichdeveloped by the Massachusetts Eye and Ear Infirmary Voice makes the vote the embeddingefrom 0 orderhicand Speech Lab. Subjects were asked to sustain the vowel la! makes the volume of the embedding space unity. In order toand voice recordings were made in a soundproof booth on a take into account the edge effects correctly, it is convenientDAT recorder at a sampling frequency of 44.1 kllz. to redefine p(R) as the number of data points within an M-

cube (instead of M-sphere) of length R around a data point.ii. ii. NONLINIER ANALIsYs This is equivalent to replacing the Euclidean norm by the

Nonlinear systems can show chaotic behavior when some maximum norm. Operationally this is done by choosingconditions accomplished. Chaos is the term used to describe randomly NC data points as centers ofM-cubes of length R. Ofthe apparently complex behavior of what we consider to be these N M-cubes, only those which are within the boundingsimple, well-behaved systems. Chaotic behavior, when looked bat casually, looks erratic and almost random. Chaos theories'. , , , , . , ~~~~~correlation sum iR S obtained by averaging the number ofenable us to categorize and understand complex behavior that CM(R) y a ghad confounded previous theories [8]. data points within the M-cubes. The imposition of the

In order to quantifying chaos, we use the embedding space requirement that an M-cube has to be within the embeddingtechnique. Unfortunately, the theoretical underpinnings of the space ensures that there are no edge effects due to limited dataembedding technique are not fully developed at present and points. However, this also means that for large values of R,some "rules of thumb" are used for them [8]. Therefore it is only a small fraction ofthe original NC M-cubes are taken intonecessary to develop these rules to achieve better result. consideration. Hence a maximum value of R, Rmax, is fixed

A. Modified GP Algorithm such that for all R < ?max the number ofM-cubes which satisfyThe GP algorithm uses the delay embedding technique for the above criterion is at least a hundredth of the total number

the calculation of correlation dimension (D2). It creates an of vectors, i.e. NV/l00. To avoid the region dominated byartificial space of dimension M with delay vectors constructed counting statistics only results from R> Rnim are taken intoby splitting a discretely sampled scalar time series s(ti) with: consideration, whereN C(R)>10, which ensures that on

X ± i1f.ifTS i51 -[ ) t+)(rf + (M-l)r)] (1) average at least ten data points are considered per center. Thismakes sure that the region Rmm < R < Rmax is not affected by

Here the delay time T is chosen suitably such that the "ingsvectors are not correlated. The relative number of pointswithin a distance R from a particular (i th) data point is given M(R) is computed for several different values ofR betweenby: Rmax and Rmin, the logarithmic slope at each point is

NI calculated and the average is taken to beD (M) . The error on1,(R)=li H(R= l-X[X` (2). .2Nv(1?)iHN R -J D,(M) is estimated to be the mean standard deviation over this

average. This error is an estimate of how well the region usedWhere Nv is the total number of reconstructed vectors and by the scheme, R im <R < R max, can be represented by a

H is the Heaviside step function. Averaging this quantity over linear scaling region. A large error signifies that those valuesrandomly selected centers Nc gives the correlation function. of R for which Cm(R) not affected by counting statistics

cW (3) and edge effects do not represent a single scaling region. It(R) = r /LPi R (3) should be noted that there often exists a critical embedding

dimension M for which R R and no significantThe correlation dimension D ()is then defined to be the cr min max

2 (M) results can then be obtained forM > Mc. Thus our algorithmscaling index of the variation of C (R) with R asR -- 0 . That fie an upe limit o M u to ch l t ar to biS, repeated. For practical implementation of the above scheme, it

R-<,(-f,O (4) is sufficient to choose NC as O.N The delay time r is~~~~iOg(ia) ~~~~~~chosen to be the value where the auto-correlation function

1172

drops by l/e. With these values, D, (M) for M= I to M = Mer is while for the corresponding ten surrogates, the curves arecomputed for a given data stream and a chi-squared fitting is represented by dashed lines. The lower panels (c) and (d) areundertaken using a simple analytical function. for red noise contamination at 20% and 50% respectively [6].

"Mat _ 1, As we can see in Fig. 1, different kind and amount of noise are

f (M) = ((Mi- 1)+1 (VArM < A distinguishable by surrogate data analysis.\M-1 (5) Voice disorder caused addition of noise to the voice. InDsalt or M >M other word pathological voice is composed of healthy voice

plus noises [3]. These noises depend on the kind and amountThe best fit value of Dsat (obtained by minimizingX2 ) is of disease. Surrogate data analysis can be used to achieve

taken to be the saturated correlation dimension with errors proper information about kind and amount of noise andcorresponding to AX2 = 1. Considering the uncertainties in the Therefore kind and amount of disease.

In order to quantify the difference between original signalcomputation and statistics of the errors inD2 (M), a more and surrogate data a quantification method is attempted bysophisticated fitting procedure is perhaps not warranted. A defining a normalized mean sigma deviation, nmsd. For thisbest fit value of D sat M implies that no saturation of the average ofsDurr (M), denoted here asr(Djr(M))r is

D2 (M) was detected. estimated using a number of realizations ofthe surrogate data.In summary, the algorithmic scheme first converts a data Then:

stream to a uniform deviate. Next, the autocorrelation functionis evaluated to estimate the time delay r . For each M, Cm(R) 1

ma /urr(M)) 2¢M d I ' 2 D2(M) -(D2"() (6)is evaluated using N = o.lNv randomly chosen centers. The nnsd (6

limits Rminand Rmax are estimated and D2(M) is computed for M=2 assMDthe region from Rmin to Rma. The process is repeated forconsecutive values of M until Rma R . T uere M is the maximum embedding dimension forD2(M) curve is fitted using function (5) which returns thesaturated correlation dimension D2st with an error estimate which the analysis is undertaken and 0j7. (M) is the[6]. standard deviation ofDurr (M).

B. Surrogate data AnalysisPhase has an important role in the chaotic signals. Chaotic

regularity (determination property of chaos) of a signal is 6 (a) h /a)related to the chaotic regularity of both phase and amplitude. FThis determination is highly related to the phase as a littledistortion in phase disturbed chaotic regularity. 4

Surrogate data analysis uses this property of chaotic signal [eto capture a criterion for amount of deterministic property of -.signal. Surrogate data generated by taking the Fourier 2 /ttransform of original data, randomizing the phases (without 2,0 /2 V t, Ni;e /changing the amplitudes), and then regenerating the surrogatedata with an inverse Fourier transform. Produced data has : rbeen removed any deterministic evolution [8]. Therefore the 6 (c) (d)correlation dimension of surrogate data is different fromoriginal data. By measuring this deference we can achieve acriterion for deterministic property of a chaotic signal. --Added noise to a chaotic signal causes reduction of the

deterministic property and therefore increase correlationdimension of signal. Surrogate data analysis can measure this 2

difference and give a criterion for amount of noise added to Red N6se 20= : RedoI Noise=0%the chaotic signal. The addition of noise to the chaotic system |,is found to decrease the difference between the D2(M) of the 2 4 6 8 2 4 6 8data and the surrogates. The effects of the addition of different Mpercentage of white and red noise on surrogate analysis of data Fig. 1. The effect of the addition of white and red noise on surrogate analysis

of data from the Rossler system. The D2(M) values for the data arefrom the Rossler system are shown in Fig. 1 [6]. The upper represented by filled circles and connected by solid lines, while for thepanels (a) and (b) are for white noise contamination at 200% corresponding ten surrogates, the curves are represented by dashed lines. Theand 500/O respectively. The D2(M) values for the data are upper panels (a) and (b) are for white noise contamination at 2000 and 5000represented~ ~by file cice and......concedb.oidlns respectively. The lower panels (c) and (d) are for red noise contamination atrepresented by illedc1rcles a(l connecte(l o SO11(1 l1nesn200% and 500% respectively [6].

1173

IV. RESULTS

Cm of vowel /a! is calculated by algorithm for IM a 2 6 F T y / *~~~~~~~~~~~~~~~~~~~~Onig-lsig.various amount of R and M. In Fig. 2 the diagram of

/, >i+log( Cm versus log( R) for various M is shown. Location

of Rml and RMax are determined by '*' where each point.min Mxn2r/Az

between them is approximated by a line. The slop of theselines is correlation dimension of signal for corresponding M.In Fig. 3, diagrams of D2 (M) according to a healthy signal

M62 j 4 6 6 'f a 9 0(black curve) with its three surrogate data (blue green reed) M On-

are shown. There is much difference between these two Fig. 4. Diagrams of D2(M) according to a disordered voice (black curve)groups of curve; in other word these two groups are easily with its three surrogate data (blue green reed) are shown. These two groupsdistinguishable. They have large nmsd (14.635). In Fig. 4, are close to each others, therefore nmsd ofthem is small (2.041)diagrams of D2(M) according to a disordered voice (black The results of the analysis on 16 disordered voices and 5curve) with its three surrogate data (blue green reed) are healthy voices are shown in Table 1 and Table2 respectively.shown. These two groups are close to each others, therefore D sat and *D;-(m)) are correlation dimension of originalnmsd of them is small (2.041). Whereas the variation rang of 2

correlation dimension is 1.6-3.4, the variation rang of nmsd is signal and average of surrogate data's correlation dimension2.041-14.635, Therefore nmsd is a better feature for respectively. Msi9 (Saturated dimension) is the dimension inclassification of pathological and healthy voice in comparison which the correlation dimension (D2 (M)) of original signalwith correlation dimension. saturates. Ms"sur is the average of saturated dimension of

surrogate data. nmsd is normalized mean sigma deviation (6).Lg(C~ g)l(R)

Table I.TBE RESULTS OF TBE ANALYSIS ON 16 DISORDERED VOICES

22 Dsat Msat (D2..(M)) (Mstsur) nmsd

3 2.3321 5.0000 2.3132 4.0000 0.4624 ', 3.8410 8.0000 3.9236 6.0000 1.6699

3.6811 7.0000 3.7693 6.0000 3.3879

1.8775 7.0000 2.0795 6.0000 3.7409

1.6843 3.0000 2.4450 7.0000 5.81291.9120 9.0000 2.2112 5.0000 8.6273

9 _B 4 2.2278 8.0000 2.1141 5.0000 3.1528

Fig. 2. The diagram of log( CM) versus log( R) for various M is shown. 1.9722 4.0000 2.3785 4.0000 2.4440

Location of R and RM are determined by '*' where each point between them 1.8973 9.0000 1.7673 3.0000 3.8098

is approximated by a line. The slop of these lines is correlation dimension of 3.6300 8.0000 3.3156 6.0000 2.9489signal for corresponding M. 2.4062 8.0000 2.2056 6.0000 0.9490

3.0555 6.0000 2.8827 7.0000 7.686834

0.2046 6.0000 1.2059 8.0000 2.9659M2 _-<

1.1502 5.0000 1.1177 6.0000 10.8247

/ >>- tt 1.6120 7.0000 1.5367 8.0000 1.6721

.i/ > 41.0848 4.0000 1.0568 4.0000 2.2814Z26

Table 11.THiE RESULTS OF TiE ANALYSIS ON 5 HEALTHY VOICES

D,t (D...(M)) Ms'sur nmsd2 r //-- 1.1929 7.0000 2.5007 5.0000 14.6349

1 '+ X:6-f_ + Y:1_673__ 1.4821 7.0000 1.9621 4.0000 11.7207

112 4 M I I I1.4170 8.0000 1.9820 7.0000 19.5827Fig. 3. Diagrams of ]9,(M) according to a healthy signal (black curve) with 119292 410000 218398 510000 913704its three surrogate data (blue green reed) are shown. There is much difference

beweteew* ruso crve 0.7938 4.0000 3.0109 5.0000 40.6419

1174

The mean values of nmsd for disordered and healthy voices oftime series generated by schizophrenic patients", IEEE Eng Med Biolare 3.902 and 19.190 respectively. If we choice nmst = Mag. 1999;3:84-90.

(3.90+19.90)/=1 .546 s diferetiaton theshod an use[6] K.P. Harikrishnana, R. Misrab__ G. Ambikac, A.K. Kembhavi, "A non-

(3.902+1 9.1l 90)/2=1 1.546 as differentiation threshold and use subjective approach to the GP algorithm for analyzing noisy time series",it for separation of disordered an healthy voices, all subjects Physica D 215 (2006) 137-145are detected correctly except dashed subject. Therefore the [7] Disordered Voice Database, Version 1.03, October 1994, Massachusettsclassification accuracy is about 95%. Eye and Ear Infirmary, Voice and speech Lab,Boston, MA, KayElemetrics Corp.

[8] Robert C. Hilborn, "chaos and nonlinear dynamics: An Introduction forV. DISCUSSION Scientists and Engineers", 2ed ed., Oxford: Clarendon, 2003, pp.1 -5 &

319-420Healthy voice has lower correlation dimension in [9] Roozbeh Behroozmand, Farshad Almasganj, Mohammad HassanMoradi, "Pathological assessment of vocal fold nodules and polyp using

comparison with disordered voice because it has much acoustic perturbation and ohase space features", ICASSP 2006regularity (determination property of chaos). Randomizing thephase in order to generate surrogate data causes distribution ofthis regularity and increase the surrogate data correlationdimension and nmsd (Fig. 3). While disordered voice has lessregularity and high correlation dimension therefore has smallnmsd (Fig. 4). This differentiation can be used to segregate thedisordered voices from healthy voices.The variation rang of correlation dimension (0.21_3.9) is

shorter than the variation rang of nmsd (0.46-40.64).Therefore surrogate data analysis is powerful method forclassification of disordered and healthy voices in comparisonwith correlation dimension.

In this paper without use of any classifier (such as neuralnetwork or support victor machine) we achieved 95%accuracy. It can be developed for classifying the kind ofdisorder by using a proper classifier. For example in [9] theeffect of nodules and polyp on voice were investigated.Acoustic, pitch and amplitude perturbation quotients, andnonlinear dynamic measures, phase space reconstruction andcorrelation dimension, were used. They showed disorderedvoice with nodules, in comparison with polyp, has lower-dimensional phase space dynamical characteristics. Surrogatedata analysis can improve their results.

ACKNOWLEDGEMENTS

We would like to thank Professor M. HashemiGholpayeghani for his critical discussions and valuablecomments that have demonstrated in chaos course andLinguistic Laboratory of Tehran University for their usefuldata.

REFRENCES

[I] Juan I. Godino-Llorente, Nicolas Saenz-Lechon, Victor Osma-Ruiz,Santiago Aguilera-Navarro, Pedro Gomez-Vilda, "An integrated tool forthe diagnosis of voice disorders", Medical Engineering & Physics 28(2006) 276_289

[2] Roland Linder, Andreas E. Albers, Markus Hess, Siegfried J. Poppl andRainer Schonweiler, "Artificial Neural Network-based Classification toScreen for Dysphonia Using Psychoacoustic Scaling of Acoustic VoiceFeatures", Journal of voice, October 2006

[3] Jiang, Jack J.; Zhang, Yu; McGilligan, Clancy, ";Chaos in voice, frommodeling to measurement, Journal of voice", 03/01/2006

[4] Poon CS, Merrill CK. "Decrease of cardiac chaos in congestive heartfailure. Nature":. 1997;389:492-495.

[5] Hornero R, Alonso A, Jimeno N, Jimeno A, Lopez M. "Nonlinear analysis

1175