Engineering, Technology & Applied Science Research Vol. 8, No. 4, 2018, 3228-3233 3228 www.etasr.com Siddiqui et al.: Reduced Feature Set for Emotion Based Spoken Utterances of Normal and … Reduced Feature Set for Emotion Based Spoken Utterances of Normal and Special Children Using Multivariate Analysis and Decision Trees Maria Andleeb Siddiqui Department of Software Engineering NED University of Engineering and Technology Karachi, Pakistan mandleeb@neduet.edu.pk Syed Abbas Ali Department of Computer Science & Information Technology NED University of Engineering and Technology, Karachi, Pakistan saaj@neduet.edu.pk Najmi Ghani Haider Department of Software Engineering NED University of Engineering and Technology Karachi, Pakistan najmi@neduet.edu.pk Abstract—The current paper deals with the use of multivariate data analysis and decision tree methods in order to reduce the feature set for the normal and special children speech in four different emotions: anger, happiness, neutral and sadness. Ten features were extracted, by an algorithm implemented in a previous study to classify the speech emotions of normal and special children. In the current study, the best features are selected using multivariate analysis: principal component analysis (PCA), factor analysis and decision tree. Step by step PCA is applied to reduce the feature set according to the variables that are collinear. The obtained reduced feature sets are applicable to both normal and special children samples. Experimental results revealed that PCA yields the feature set comprising pitch, intensity, formant, LPCC and rate of acceleration. Factor analysis provides three feature sets out of which the feature set comprising of Rasta PLP, MFCC, ZCR and intensity provides the best result. Decision tree yields a feature set comprising energy, pitch and LPCC. Keywords-speech emotions; PCA; factor analysis; decision tree; features I. INTRODUCTION Emotion recognition system identifies the emotional state from voice [1], therefore it is called speech emotion recognition (SER). There are four modules of SER: input, feature extraction, feature selection and classification of emotions [2]. Prosodic features, particularly pitch, intensity and duration were used in early research studies. Currently, LLD’s features such as shimmer, jitter, harmonic to noise ratio (HNR) and cepstrum have been used extensively [3, 4]. LPCC and MFCC were also accompanied in the speech feature set [5]. In [6], 40 depressed patients and 40 control subjects were used in a study for speech feature analysis. Characteristics of depressed patients were found using ANOVA analysis and the results were linked to Gaussian mixture model (GMM) and support vector machine (SVM). Autism spectrum disorder comorbid for children (ASD-CC) psychometric properties were evaluated and developed in [7]. Confirmatory factor analysis (CFA) is used for the factor structure of the Korean version of ASD-CC. In [8], ten features were extracted: frequency, pitch, intensity, rate of acceleration, formant frequencies, log power, log energy, rate of zero passages, Mel frequency cepstrum coefficient (MFCC), linear prediction cepstrum coefficient (LPCC). The extraction of frequency starts with the speech signal loading and the conversion of analog signal into numeric data. After loading, maximum and minimum frequency were set and fast Fourier transform (FFT) of windowed signal was performed as shown in (1). Followed by the cepstrum calculation the frequency is extracted as shown in (2).  (1 .* ( )) ms fft x hamming length x (1)  / 1 1 f fs ms fx   (2) where ms1=windowed signal, fft=fast Fourier transform, f=frequency, x=input signal, fs=sampling frequency, fx=input frequency. For pitch extraction, signal acquisition and signal processing are the same as in frequency extraction. FFT is applied on the processed signal. Discrete Fourier transform (DFT) is taken of the FFT signal as in (3). Then the log of FFT is calculated. After that the real cepstrum is calculated and it is the absolute value of filtered log of DFT as shown in (4). Then finally real cepstrum pitch is extracted given in (5).   dfty abs fft x (3)  10 dft log dfty (4)   _ 16 : _ rcp real ceps length real ceps (5) (dft=discrete fourier transform, rcp=real cepstrum pitch). The magnitude of the DFT signal with log10 is filtered as in (6). The conversion of the magnitude into decibels and transpose result of decibel gives out intensity as shown in (7).     20 * 10 1 : M log abs Y length x eps  (6) Engineering, Technology & Applied Science Research Vol. 8, No. 4, 2018, 3228-3233 3229 www.etasr.com Siddiqui et al.: Reduced Feature Set for Emotion Based Spoken Utterances of Normal and …  2 i mag db M (7) where M=magnitude, Y=filtered signal, x=input signal, i=intensity, mag2db=magnitude to decibels. Processing of the signal involves the time instant calculation as given in (8). The derivative of speed gives the velocity in (9). The gradient of velocity by 0.01 yields the acceleration in (10). Finally the average rate of acceleration is calculated by taking the mean of acceleration.         1 0.4 * . ^ 8 4 10.8 * . ^ 3 64.4 * . ^ 2 2 .2 * 4.4 x t t t t      (8) v=diff(x1) (9) acc=gradient(v,0.01) (10) where x1=speed, T=time instant, v=velocity, diff=derivative, acc= acceleration. For extracting the formant frequency, preprocessing involves setting the number of coefficients according to the rule of thumb for formant estimation given in (11). After that calculation of the linear prediction coefficients is carried out in (12). Then the frequencies are calculated using (13). The imaginary part of the root gives out formant frequencies. 2 / 1000 ncoeff Fs  (11)  a lpc ncoeff (12)   r roots a (13) where ncoeff=number of coefficients, Fs=sampling frequency, lpc=linear prediction cepstrum, a=linear prediction cepstrum coefficient, r=roots. Preprocessing involves the setting of the sampling rate and sampling window size for log power extraction. After that the frame size is calculated and windowing is applied. Then the average energy is extracted. Average power is yielded by dividing the average energy with window size as shown in (16). _ * windowsize sampling rate fs (14)  _ . ^ 2 Average Energy sum result (15) _ / Lp Average Energy windowsize (16) where, fs=sampling frequency, sum=sum of frame size of windowed signal, Lp=Log power. Preprocessing involves the same procedure as log power extraction. Average energy is calculated by applying windowing on signal according to window size and frame size, given in (17).  _ . ^ 2 Average Energy sum result (17) Signal acquisition and signal processing are the same as in pitch and frequency estimation given in (18).      1 2 / 2 / RZP sum abs sign y sign y windowsize  (18) where RZP=rate of zero passages, y1=maximum frequency, y2=minimum frequency. Preprocessing involves setting and analysis of frame duration and frame shift. It is followed by the setting of the pre emphasis coefficient alpha=0.97, the number of filter bank channels M=20 and lower and upper frequency limits. Then the Hertz to Mel- wrapping function is calculated. Application of DCT matrix routine and the magnitude of the spectrum are calculated in (19). The filter bank is applied to the unique part of the magnitude spectrum. Finally the calculation of cepstral liftering gives MFCC (20).            ( * ( * ) @ , 2.0 / 0 : 1 .',1, * 1 : 0.5 / , ,1 dctm N M sqrt M cos repmat N M repmat pi M M N    (19)      @ , 1 0.5 * * * 0 : 1 /CL N L L sin pi N L   (20) where dctm=secrete cosine transform matrix, N=number of coefficients, M=number of filter bank channels, L=length of channel, CL= cepstrum lifter. Preprocessing involves the estimation exponent of next high power according to signal size. Then the number of prediction paths ‘p’ is set. Calculation of the number of linear prediction of coefficients is carried out. Fourier transform is applied on X-lpc according to the number of shifts N as given in (21). The logarithm of LPC is taken and then the LPC coefficients are converted back to spectra. The number of cepstra is then set and the first and second derivative of LPCC features are estimated to have the value of coefficients in (22).    _ , x lpc lpc x p (21)  _ , lpcc fft x lpc N (22) where, Lpc=linear prediction coefficients, x=input signal, p= prediction paths, FFT= fast Fourier transform, lpcc=linear prediction cepstrum coefficients, N= number of shifts. After extracting these features, speech emotion recognition of normal and special children (SERNSC) is implemented. To make the algorithm run efficiently, dimension reduction is a valuable approach. The advantage of dimensionality reduction is that it helps to discover the grouping of features that for sure run the algorithm with improved accuracy [9]. Detection of projection subspace basis evaluation is suggested in [10]. For deduction it uses generalized hyperbolic mixture (HMMDR) fit. This method is well accepted along with discriminant analysis, model based classification and clustering analysis. Two SDR techniques are demonstrated in [11]. The relationship between partial least square (PLS) and principal component regression (PCR) is explained. Dimensionality reduction by joining features is one of the best strategies proposed so far [12]. Sparse partial least square regression (SPLSR) is investigated in depth in [13]. It is revealed that the recognition rate of SPLSR is up to 79.23% and it is superior when compared to other methods used for dimensionality reduction. In this paper, feature reduction is presented using mu dec con acc Me cep per A. rec dat of Th app loa nam Th inf cat thr fun hyb cla sco as 1) 2) em cla est 3) Em Engineerin www.etasr ultivariate ana cision tree me Seven featur nsideration fo celeration, zer el frequency c pstrum coeffi rceptual linear Classificatio The classifi cognition for n ta set is the sa the extracted he process of plied for dec ading the sam me, file path c hen, the file na formation, fil tegories are resholding fea nction for deci Features=all Function [Ca brid_Decision The test s assification fun ore position. T follows Step#1: If Result of p Set category Categories_o Else category Categories_o Step#2: After the ca motion is lab assification ac timated in (25)   ,rn cn find Step#3: If Result of p Set Emotion _output(1). Else If Resul Set Emotion _output(2). Else If Resul Set Emotio motions ng, Technology r.com alysis that con ethod. II. ME res and three or analysis: p ro crossing rat cepstrum coeff icient (LPCC) r prediction (R on Algorithm ication algori normal and sp ame as in [6]. d features is d data file prep cision making mple file spee categories and ame is added es containing also added ature extractio ision making i lfeatures_extra ategory, Categ n_making(wav_ samples were nction and the The process of position n=1 = Normal and output(1). y = Special an output(2). ategory is dec beled on th curacy. The m ) strcmp Emotion position n=1 = Angry and E lt of position n = Happy and lt of position n on = Neutra y & Applied Sci sists of PCA, ETHODOLOGY coding schem pitch, intensity te (ZCR), log fficient (MFCC ), Relative sp Rasta PLP). ithm used is pecial children In system flo done before da paration for e g of the algo ech.mat. Afte emotions wer in speech.mat g information in speech. on is perform is implemented action(wav_fil gory_%, Emot v_file) e classified en finding the f hybrid classif d Category_Pe nd Category_Pe cided as norm he category maximum sco _ / , n o p max E Emotion _Perc n=2 Emotion _Per n=3 al and Emo ience Research Siddiqui et al.: factor analysi mes are taken y, formant, ra energy, log p C), linear pred pectrum trans s speech em n (SERNC) an ow, the thresho ata file prepar each sample f orithm. It star er loading, th re extracted by t. In addition t n of emotions mat. Finally med using (2 d in (24). le) (23 tion, Emotion_ (24 by using h maximum cat fication is desc ercentage = ercentage = mal or specia according to re position is _ /Emotions o p centage = Emo rcentage= Emo otion_Percenta h V : Reduced Feat is and n into ate of power, diction sforms motion nd the olding ration. file is rts by he file y path. to this s and y, the 3). A 3) _%] = 4) hybrid tegory cribed al, the o the to be  (25) otions otions age = B. ana the mu ana flow and spe feat crit 1 t oth larg is e the und the are crit var feat pro Vol. 8, No. 4, 20 ture Set for Em _output(3). Else Set Emo Emotions _ou [End of If con Multivariate A When worki alysis is very data set invo ltivariate anal alysis. These tw w diagram of d Figure 2 resp For PCA, the eech features a ture reduction teria. In (a) if t then the com erwise it is ex ge then the co excluded from speech featu derlying factor variables. In two criteria: teria. In (a) riables should ture sets based ovided. 018, 3228-3233 motion Based Sp otion = Sad and utput (4). ndition] Analysis ing with cor valuable. Mu lves more tha lysis are used wo methods a PCA and fact pectively. Fig. 1. PC e variables ar and coding sch n: (a) Eigenva the componen mponent (varia xcluded from mponent is re the model. In ures and cod rs that have th this method t (a) communa the commun be between 0. d on the strong 3 poken Utteranc d Emotion _Pe rrelated varia ultivariate anal an one variable d in this study are performed tor analysis is CA flow diagram re the classific hemes. Two m alue criteria a nt value is grea able) is retain the model. In etained in the m n factor analysi ding schemes. he effect on th two factors ar ality criteria an nality of extr .5 and 1 while g correlation b 3230 ces of Normal a ercentage = ables, multiva lysis is used w e. Two metho y: PCA and f on MINITAB shown in Fig cation accurac methods are use and (b) Scree ater than or clo ned in the m n (b) if the slo model, otherw is, the variable . There are he communali re extracted. T nd (b) loading racted factors e in (b) the red between variab and … ariate when ods of factor B. The gure 1 cy of ed for e plot ose to model, ope is wise it es are some ity of There g plot s for duced bles is C. acc acc cla dec Engineerin www.etasr F Decision Tre In decision t curacy. The fe cording to the assification ac cision tree flow F ng, Technology r.com Fig. 2. Factor ees tree method a eatures on the e threshold of curacy feature w diagram is s ig. 3. Flow di y & Applied Sci analysis flow dia threshold is s two nodes (le f classification es are retained shown in Figur iagram of decisio ience Research Siddiqui et al.: agram set for classifi eft and right) a n accuracy. H d in the mode re 3. on tree h V : Reduced Feat cation are set Higher l. The A. use of t is th for ana redu “fac dire ana 1) rep eige oth  feat gre var eige sele exc cum pro fact is v cum incl eige form  Acc the Con Vol. 8, No. 4, 20 ture Set for Em Experimental The assumpti ed to reduce th the speech em he total consid PCA. The int alysis. Likewi duce the large ctors”. The ef ectly measure alysis. Experimental The variance resented by envalue of th ers’ as shown Factor/ Component Pitch Formant Intensity Rate of acceleration LPCC Rasta PLP MFCC ZCR Energy Power Eigenvalue C Eigenvalue is ture selection ater than 1 a riable differen envalue will ection and th cluded from mulative perce ocess of factor tors have eige very close to 1 mulative perce lude this va envalue criter mant, rate of a Scree Plot Cri Another met cording to this cliff while t nsider Figure 018, 3228-3233 motion Based Sp III. EXPERI l Design ion is that the he number of f motion recognit dered factors. tersection of th se PCA, the number of va ffects of some ed are assess l Results Based e accounted the eigenval e first compo in Table I. TABLE I. Eigenvalue 3.6049 2.2118 1.4581 1.4117 0.9818 0.2159 0.1157 0.0000 0.0000 0.0000 Criteria s one of the m n. Those com are retained i nce of eigen not serve as he variable w the model. entage of vari r selection. C envalues greate , it shows the ent of varian ariable into t ria the select acceleration (R iteria thod for facto s criterion the the base of t 4. It shows th 3 poken Utteranc IMENTAL STUD ese factors are factors that me tion. The num Factors are th hese factors is goal of facto ariables into a e underlying fa sed with the d on PCA for by a giv lue of each onent is usuall PCA STATIST Percentage o variance (proportion) 36 22.1 14.6 14.1 9.8 2.2 1.2 0.0 0.0 0.0 most commonl mponents that n the model. nvalue is so s the best cr with eigenva Therefore th iance are also Consider Table er than 1 but t 9.8% of the to nce is 96.7%. the model. A ted factors ar ROA) and LPC or selection factors having the cliff has hat the slope o 3231 ces of Normal a DY e collinear. PC easure the accu mber of compo he same consid the focus in f orial analysis smaller numb actors that can help of fac ven compone component. ly greater tha TICS of ) Cumulativ percentag 36% 58.2% 72.7% 86.9% 96.7% 98.8% 100% 100% 100% 100% ly used criteri t have eigenv If the succe o small then riterion for f alue 0.99 wil he proportion o considered i e II, the first the fifth compo otal variance an . So we can According to re pitch, inten CC. is the scree g a big slope a the trivial fac f the curve is and … CA is uracy onents dered factor is to ber of n’t be ctorial ent is The an the ve ge ia for value essive n the factor ll be and n the t four onent nd its also o the nsity, plot. are on ctors. quite sm the 2) are mu can var loa “V sch mi var int Th the bet it i job com loa   Engineerin www.etasr mall from the s e sixth factor c Experimenta The factor lo TA Variable Pitch Formant Intensity ROA LPCC Rasta PL MFCC ZCR Energy Power Variance % Var The factor lo e two extracte uch a variable n be positive o riable is stron ading result m Variables” are hemes. “Facto initab that ex riables is ex terpreted by th he assessment e communality tter for some is not bad for b for LPCC an mmunality wh ading plot of F F Figure 7 reve Rasta PLP, Z to each oth correlated. classification LPCC, pitch are in the sam ng, Technology r.com tarting of the s can be selected al Results Base oadings of each BLE II. FAC e Factor 1 0.028 t -0.739 y 0.798 -0.723 0.102 LP 0.766 0.687 0.882 -0.382 -0.331 e 3.8089 0.381 oadings for ea ed factors and is explained b or negative. La ngly influence means that the the seven sp or 1” and “Fact xplains the 1 xplained by he help of “com of the workin y. In Table II, variables. The other variable nd energy as t hich less than h Factor analysis ig. 4. Scree p eals that: ZCR, intensity her and hen So these fo n accuracy. and power sh me direction. y & Applied Sci sixth factor so d from the mod ed on Factor A h variable are CTOR LOADINGS O Factor 2 C -0.726 -0.104 -0.223 0.518 -0.310 0.010 -0.377 -0.014 0.592 -0.855 2.1763 0.218 ach variable ar the loading r by the extract arge loading r ed by the fac e variable is peech features tor 2” are the 0 variables. the extracted mmunality” as ng of the mod it is shown th e model best e es. The model they show only half variation s is shown in F plot for factor redu y and MFCC nce these fac our factors how strong cor ience Research Siddiqui et al.: o all the factors del. Analysis shown in Tabl OF VARIABLES Communality 0.528 0.556 0.686 0.791 0.107 0.588 0.614 0.778 0.497 0.840 5.9852 0.599 re calculated. result tells tha ted factor. Loa esult means th ctor whereas weakly influe s and three c extracted facto The proportio d two factors s shown in Ta del is calculat hat the model w explains powe does not do a y 10.7% and 4 of the variable Figure 5. uction lines are very ctors are str lead to the rrelation as the h V : Reduced Feat s after le II. There at how adings hat the small enced. coding ors by on of s and able II. ted by works er, and a good 49.7% e. The y close rongly best e lines  bes inte set 3) dec clas is s feat the that 91. Fig true The clas ligh pitc pre hig erro dec emo Vol. 8, No. 4, 20 ture Set for Em ROA, energy show strong c So these thre st correlation i ensity and MF for higher clas Fig Experimental The factor an cision trees. Th ssification acc shown in Figu ture in the no reduced featu Fig. The decision t energy, pitch 65%, 83.15% gure 7. The to e class energy e green cells s ssification acc ht green cells ch and LPCC dicted class a hest error rate or is shown by In this study, cision tree wer otion based 018, 3228-3233 motion Based Sp and formant a correlation. ee sets of fac is seen in set FCC. They are ssification acc g. 5. Loading l Results based nalysis for red he factor redu curacy. The cl ure 6. Energy, de of the tree ure set. 6. Decision t tree table is h and LPCC and 71.5%. T op row shows y. Predicted cl show the true curacy in the e show the nex C have high according to th e is shown by y light red cells IV. CO the two techn re used to red spoken utter 3 poken Utteranc are also in the ctors are close t 1 comprising e considered a curacy. plot of factor ana d on Decision duction is don uction is done lassification ac pitch and LPC e and hence th tree for factor red shown in Tab show classifi The confusion the classifica lasses are show e positive rate energy row is xt higher simil er similarity he accuracy in dark red cells s. ONCLUSION niques, multiv duce the ten ex ances for no 3232 ces of Normal a same direction ely correlated. g rasta PLP, Z s a reduced fe alysis Trees ne with the he with respect t ccuracy of the CC are the red hey are formul duction ble III. It is sh cation accurac matrix is show ation accuracy wn in that col column. No misclassified larity rate. En rate in true n all emotions s whereas the variate analysi xtracted featur ormal and sp and … n and . The ZCR, eature elp of to the e tree duced lating hown cy of wn in with lumn. other d. The nergy, e and . The least s and res of pecial chi and fro and utt chi exp com Fa PL fre com 3) com dev fea N [1] [2] [3] Engineerin www.etasr ildren in four d sadness). Th om two multiv d decision tr terances in U ildren. The periments are mprising of p actor analysis LP, ZCR, inten equency) and mprising of ra Decision tre mprising of en velop an exp ature sets in or Node Class 1 28.335 2 16.6665 3 45.00 4 91.65 5 83.15 6 71.5 7 55.00 Fig. 7. C S. Ramakrishn in: Speech En Applications, p S. Pahune, N. Speech and Im Recent and Inn 3, No. 2, pp. 13 B. Schuller, A Devillers, L. relevance of f user states: low 2007, Antwerp ng, Technology r.com r different em he reduced fea variate techni ee is applied Urdu language observations e: 1) PCA p pitch, intensity provides thre nsity and MF (LPC, pitch a asta PLP, ZCR ee technique nergy, pitch an periment to a rder to analyze TABLE III. sification Accura Condition 5>Accuracy>=28. 5>Accuracy>=16. 0>Accuracy>=45. 5>Accuracy>=91. 5>Accuracy>=83. 5>Accuracy>=71. 0>Accuracy>=55. Confusion matrix REFE nan, “Recognition nhancement, Mod pp. 121-138, InTe Mishra, “Emotio mage Processing novation Trends 34-137, 2015 A. Batliner, D. S Vidrascu, N. A feature type for t w level descripto p, Belgium, pp. 22 y & Applied Sci motions (anger, ature set and fe iques (PCA a d on the emo e for both n obtained fr provides a re y, formant, L e sets of redu CC), (energy, and power) ou R and MFCC s provides a r nd LPCC. Aut apply the red e the classifier DECISION TREE R acy Less T Nod .335 Nod .6665 MFC .00 POW .65 Nod .15 Nod .5 Nod .00 ENER for decision tree ERENCES n of Emotion fro deling and Recog ech, 2012 on Recognition t g: A Review”, I in Computing an Seppi, S.Steidl, T Amir, L. Kessous the automatic cla ors and functiona 253-2256, August ience Research Siddiqui et al.: , happiness, n eatures are obt and factor ana otion based sp normal and s rom demonst educed featur LPCC and RO uced features , ROA and fo ut of which th shows better re reduced featur thors are focus duced feature ’s accuracy. RESULTS Than de Great Than N de 2 Node CC ENER WER Node de 5 ENER de 6 PITC de 7 LPC RGY INTENS factor reduction om Speech: A Re gnition- Algorithm through Combina nternational Jour nd Communicatio T. Vogt, J. Wag s, V. Aharonson assification of em al”, in: INTERSP t 27-31, 2007 h V : Reduced Feat neutral tained alysis) poken special trative re set OA. 2) (rasta ormant he set esults. re set sing to s and ter Node e 3 RGY e 4 RGY CH CC SITY eview”, ms and ation of rnal on on, Vol. gner, L. n, “The motional PEECH [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] Vol. 8, No. 4, 20 ture Set for Em B. Schuller, G. comparing b INTERSPEECH Y. Zhou, Y. Su speech emotion Research Chall 28-29, 2009 S. Alghowinem Breakspear, “C INTERSPEECH K. M.Chung, D autism spectrum in Autism Spec M. A. Siddiqui for Features E Children Speec Computer Scien 2017 L. E. Aik, L. C multivariate sta AIP Conferenc Publishing, 201 ] K. Morris, P. D analysis, and di Computational ] Y. W. Lin, B equivalence of in the sufficien Intelligent Labo ] K. Mallick, S. Criterion: An Classification”, ] Y. Jingjie, X. Based on Spars pp. 465–470, 20 018, 3228-3233 motion Based Sp Rigoll, “Recogn ag of frame H, Brighton, UK, un, L. Yang, Y. n recognition”, enges in Comput m, R. Goecke, Characterizing D H, Florence, Italy D. Jung, “Validity m disorders como trum Disorders, V i, N. G. Haider, S Extraction towar ch Emotions in U nce and Network C Kiang, Z. B. M atistical methods e Proceedings, P 17 D. McNicholas, imension reductio Statistics and Dat B. C Deng, Q. S partial least squa nt dimension red oratory Systems, V Bhattacharyya, “ Efficient Dimen Procedia Techno Wang, W. Gu, se Representation” 013 3 poken Utteranc nizing interest in s and supra- pp. 1999-2002, S Yan, “Applying IEEE 9th Interna ter Science, Shan M. Wagner, J. Depressed Speec y, pp. 2534-2538, y and reliability o orbid for children Vol. 39, pp.1-10, 2 S. A. Ali, S. Hina ards Classifying Urdu Language”, I k Security, Vol. 1 Mohamed, T. W H for dimensional Perlis, Malaysia, “Clustering, clas on via generalized ta Analysis, Vol. S Xu,Y. H. Yun ares and principal duction framewor Vol. 150, pp. 58-6 “Uncorrelated Lo nsionality reduct ology, Vol. 4, pp. L. Ma, “Speech ”, Archives of Ac 3233 ces of Normal a conversational sp -segmental fea September 6-10, 2 articulatory featu ational Conferen nghai, China, Dec Epps, G. Parke ch for Classifica August 25-29, 20 f the Korean vers n (ASD-CC)”, Re 2017 a, “A: Novel App Normal and S International Jour 17, No. 7, pp. 18 Hong, “A review reduction studie Vol. 1847, No. ssification, discrim d hyperbolic mix 97, pp. 133-150, n, Y. Z. Liang, l component regr rk”, Chemometric 64, 2016 ocal Maximum M tion Method for 370-374, 2012 h Emotion Recog coustics, Vol. 38, and … peech– atures”, 2009 ures to nce on cember er, M. ation”, 013 sion of esearch proach Special rnal of 88-195, on the es”, in: 1, AIP minant xtures”, 2016 , “The ression cs and Margin r Text gnition No. 4,