text
stringclasses 10
values |
|---|
ARCHIVESOFACOUSTICS Vol. 42,No. 2, pp. 213-222(2017)Copyrightc⃝2017by PAN-IPPT DOI:10. 1515/aoa-2017-0024 Automatic Genre Classification Using Fractional Fourier Tr ansform Based Mel Frequency Cepstral Coefficientand Timbral Feature s Daulappa Guranna BHALKE,Betsy RAJESH,Dattatraya Shankar BORMANE Deptartmentof Electronics&Telecommunication JSPM's Rajarshi Shahu Collegeof Engineering SPPU,Pune,India; e-mail:bhalkedg2000@yahoo. co. in,bet syrajesh@gmail. com,bdattatraya@yahoo. com (received October12,2015;accepted February1,2017 ) Thispaperpresentsthe Automatic Genre Classificationof In dian Tamil Musicand Western Musicusing Timbraland Fractional Fourier Transform(Fr FT)based Mel F requency Cepstral Coefficient(MFCC) features. Theclassifiermodelfortheproposedsystemhasbe enbuiltusing K-NN(K-Nearest Neighbours) and Support Vector Machine(SVM). Inthiswork,theperforma nceofvariousfeaturesextractedfrom musicexcerptshasbeenanalysed,toidentifytheappropria tefeaturedescriptorsforthetwomajor genresof Indian Tamilmusic,namely Classicalmusic(Carna ticbaseddevotionalhymncompositions)& Folkmusicandforwesterngenresof Rockand Classicalmusic fromthe GTZANdataset. Theresults for Tamilmusichaveshownthatthefeaturecombinationof Sp ectral Rolloff,Spectral Flux,Spectral Skewnessand Spectral Kurtosis,combinedwith Fractional M FCCfeatures,outperformsallotherfeature combinations,toyieldahigherclassificationaccuracyof9 6. 05%,ascomparedtotheaccuracyof84. 21% withconventional MFCC. Ithasalsobeenobservedthatthe Fr FTbased MFCCeffiecientlyclassifiesthe twowesterngenresof Rockand Classicalmusicfromthe GTZAN datasetwithahigherclassification accuracyof96. 25%ascomparedtotheclassificationaccurac yof80%with MFCC. Keywords:featureextraction;Timbralfeatures;MFCC;Fractional Fo urier Transform(Fr FT);Frac-tional MFCC;Tamil Carnaticmusic. 1. Introduction Digitaltechnologyhascompletelyrestructuredthe musicindustry,becauseofwhichconsumershaveac-cesstothousandsofmusictracksstoredlocallyon theirsmartphonesandmillionsofrecordsinstantly availablethroughcloud-basedmusicservices. Recent technologicaladvanceshelpusersinteractwithmu-sicbydirectlyanalyzingthemusicalcontentofaudio files(Shaoelal.,2005). Thevastamountofavail-ablemusicaldatabasescreatestheneedforreliable methodsofsearchingandorganizingthem,andde-mandsnovelmethodsofdescription,indexing,search-ing,andinteraction( Benetos,Kotropoulos,2010; Scaringellaetal.,2006). MIR(Music Information Retrieval)dealswiththeautomaticanalysisofmusic signalsandusesvariouscharacteristicsthatbestde-scribethemusiccontent( Scaringellaetal.,2006). Genreinformationisonesuchcharacteristicthatcan helpdescribemusiccontentandisafundamentalcom-ponentof MIR(Fuetal.,2011). Themusicof Indiahasveryancientrootsandhas existedformanymillennia. Differentmusicalforms likethe North Indian Hindustani,South Indian Car-natic,Ghazals,diverseformsoffolkmusic,filmmusic and Indo-westernfusionmusiccontributetothe In-dianmusic(Baguletal.,2014). The Tamillanguage of Tamil Nadu,South India,hasanantiquitygoing beyondtheperiodofatleast B. C. 1250(Tamil Mu-sic,2011). Tamilmusicisclassifiedintotwomaingen-res,namely(1)Tamil Classical-structuredmusic, sungtoarhythmiccycleortala(calledthe Carnatic stylemusic)(Kumaretal.,2014)and(2)Tamil Folk-ruralmusiccomposedincolloquialstyle. Mostde-votionalsongscalledas'Keeerthanai'or'Kriti'and arecomposedinclassicalstyle. The'Keertanai'con-sistof(1)'Pallavi'(firstsectionofthesongindicating thetheme),(2)'Anupallavi'(formsthechorusalong with Pallavi,tobe repeated)and(3)'Charanams' (footofthe song)( Ashok Narayanan, Prabhu, 2003). Inthiswork,variousoriginalcompositionsof Tamil'gospel Keertanai'andpatrioticcompositions
|
214 Archivesof Acoustics-Volume42,Number2,2017 havebeenemployedforthepurposeofclassification. Thesedevotionalandpatrioticmusicalcompositions havebeencomposedonclassicalragas(melodies)and tala(rhythm),mostlysungasasolo(Vedanayagam Sastriar,2011). Tamilfolk music is knownforthe tala intrica-ciesandtheancientfolkmusicwasbasedonclas-sicalragasormelodieslike Manji,Sama,Navaroz, and Kalyani. Instrumentaccompanimentsare Nad-haswarams(atypeofflute),drumsormelamandcym-balsor Kaimani. Folkmusichascontinuedtoevolve overtheyears. Thepresentdayfolkmusiccanbe classifiedas“Naatupurapaadalgal'(ruralfolkmusic) and“Gaanapadalgal'(urbanfolkmusic). Thefollow-ingsectiondiscussesthevariousmethodologiesthat havebeenemployedinthepreviousyearsforauto-maticgenreclassificationofbothwesternand Indian music. 1. 1. Literaturesurvey Tzanetakis,Cook(2002)havedonepioneering workinautomatedgenreclassificationofwesternmu-sicanddeducedthreeessentialfeaturesformusical contentnamelytimbraltexture,rhythmandpitchcon-tentfeatures. Statisticalpatternrecognitionclassifier s havebeentrainedforreal-worldmusicdata. Frames andwholesongswereusedforwhich,aclassification accuracyof61%hasbeenachievedfortenwesternmu-sicalgenres. Theresultscloselymatchedthereported resultsforhumanmusicalgenreclassification. Further, Liand Tzanetakis(2003)havedescribedthefactors in Automatic Genre Classificationandstudiedtheper-formanceof Support Vector Machinesand LDA(Lin-ear Discriminant Analysis)classifiers. LDAwasused tofinddiscriminativefeaturetransformaseigenvec-tors,tocapturebothintraclassandinterclasssepa-ration. Mengetal. (2007)proposedthetemporalinte-grationofshorttimefeaturesusing Multivariateau-toregressivemodels(MAR). Theideahasbeentoex-tractasummarizedpowerofeachfeaturedimension independentlyinfourspecifiedfrequencybands. MAR hasbeenusedfortemporalfeatureintegrationsince ithasthepotentialofmodelingbothtemporalcor-relationsanddependenciesamongfeatures. Lietal. (2010)extractedfeaturesfrom Daubechieswaveletco-efficientshistogram(DWCH)andhaveobservedthat timbralfeaturescombinedwith MFCCyieldhighac-curacy. Toextractmorepowerfulfeatureslike DWCH and OSC(Octave-based Spectral Contrast)subband analysishasbeenperformedwherethepowerspectrum wasdecomposedintosubbandsandfeatureswereex-tractedfromeachsubband. Limetal. (2012)proposeda Music-Genre Classi-fication Systembasedon Spectro-Temporal Features and Feature Selection. Themean,variance,minimumandmaximumvalues,spectralmodulationflatness, crest,contrastandvalleyfeatureswereestimatedand Support Vector Machine(SVM)wasusedasaclassi-fier. Themethodhasprovedtohavehigheraccuracy atalowerfeaturedimensionforthe GTZANand IS-MIR2004databases. Chenetal. (2012)improvedclassificationaccuracy usingwaveletpackagetransform(WPT),since WPT performsawaveletdecompositionthatoffersaricher signalanalysis. Abestbasisalgorithmselectionhas beenperformedusingtop-downsearchstrategy. Mel-frequency Cepstral Coefficients(MFCC)andlogener-giesextractedfromthedecompositioncoefficientswere usedtobuildthe SVMclassifierwithresultingaccu-racyof89. 03%. Baniyaetal. (2014)derivedafeature setthatincludedhigherordermomentsofskewness, kurtosisandcovarianceoffeaturesinadditiontotheir meanandvariances,resultinginimprovementofclassi-ficationaccuracyto85. 15%. Rosneretal. (2014)pro-posedtheclassificationofgenresbasedon Music Sepa-rationinto Harmonicand Drum Components. They haveemployedco-training(semi-supervisedlearning) to SVM-basedclassification,whichenabledthe SVM tolearnfromasmalltrainingset,whichlaterhelped toclassifyunlabelleddataiteratively. Nagavietal. (2011),intheoverviewofclassifica-tionandretrievalsystemsof Indianmusic,haveveri-fiedthat PCD(Pitchclass Distribution),toneprofiles andspectralprofilessufficientlydiscriminateragasau-tomatically. Thepitchclassorchromaisarepresen-tationofpitchesfromalloctaves,mappedtoasingle octave. Kiniet al. (2011) have classified bhajan and qawwalisub-genresof North Indiandevotionalmu-sicwithtimbralfeatures,tempoandmodulationspec-traoftimbralfeatures. Theyachieved92%accuracy byapplying10foldcrossvalidationoftempoestima-tions,featuresummariesofmean-varianceandenve-lopemodulationwith Support Vector Machine(SVM) and Gaussian Mixture Model(GMM). Rao(2012) workedontheextractionofmetadatafor Hindustani Classicalmusicusingfactualinformationthataccom-paniesmusicona CD,suchascomposer,genre,artist andothersemanticlabelssuchasmood. Pitchdetec-tion,rhythmdetectionandmelodyestimationthrough motif(repetitivephrase)identificationandoscillation (gamakas),wereemployedforclassificationofragasof Hindustanimusic. Bhalkeet al. (2015) had proposed Fractional Fourier Transform(Fr FT)based MFCCfeaturesfor discriminating musicalinstruments and havefound thattheinterclassvariationwasgreatlymaximisedand intra-classvariationwasminimisedbytheuseofthe chirplikekernelbasisfunctionofthe Fr FT. Variousfeaturessuchas Timbral,Temporal,Spec-tral,Waveletand MFCChavebeenproposedinthe pastyearsformusicgenreclassification. Also,ithas
|
D. G. Bhalke,B. Rajesh,D. S. Bormane-Automatic Genre Class ification Using Fractional Fourier Transform... 215 beenobservedthat Timbraland MFCCfeatureshave madeasignificantcontributiontogenreclassification (Bhalkeetal.,2014;Fuetal.,2011;Ghosaletal., 2012). Inaddition,Fractional Fourier Transformmakes useofchirpdecompositionwhichisespeciallysuitable formusicsignalsinwhichtime,frequencyandphase informationplayavitalrole( Ashok Narayanan, Prabhu,2003;Bhalkeetal.,2016). Ithasfurtherbeenobserved,thatverylittlework hasbeendoneon Indiangenresandtherehasbeen nopreviousworkon South Indian Tamilmusic. Thus, thenextsectionsofthepaperpresentanovelfeature extractionschemeforautomaticmusicgenreclassifi-cationof Indian Tamilmusic,thatinclude Fr FTbased MFCCfeaturesor Fractional MFCC(Fr MFCC)using twoclassifiersnamely KNNand SVM. Thecompara-tiveresultsofthetwoclassifierswiththeproposedand previousmethodshavebeentabulatedandpresented. Alsotheresultshavebeentestedforwesternmusic genrestoo. Thesectionsthatfollowincludeproposedsystem, featureextractionanddatabasedetailsin Sec. 2,ex-perimentalresultsin Sec. 3,conclusionandfuturework in Sec. 4,acknowledgementsin Sec. 5andfinallythe listofreferencesin Sec. 6. 2. Proposedsystemandfeatureextraction Variousmethodologieshavebeenusedfor Auto-maticgenreclassification. Thetwomajorstagesin genre classification are (1) Feature extraction and (2)Classification. The firststageis toextractthe meaningfulandrelevantfeaturesfromaudiothatcould sufficientlydiscriminatethemusicgenres. Thesecond stepistrainingasuitableclassifierwithextractedfea-turevaluesandthentestingitwithnewsamples. Fig. 1. Trainingphase. Fig. 2. Testingphase. 2. 1. Framingandsegmentation Tobeabletodiscriminatesufficientlybetweenthe genres,30-secondclipsweretakenfromthesongs. Itis assumedthatanyrandomlyoccurringsignalisstation-ary,andthusthepropertiesremaininvariantfor10ms to20ms. Thisassumptionmakesitpossibleforsignal processingtechniquestoapplytotheshortstation-arysignals. Thus,the30-secondexcerptswerefurther framedinto20msframeswith50%overlapping. 2. 2. Windowing Tomaintaincontinuityofthefirstandlastpoints intheframe,a Hammingwindowismultipliedwith eachframe. Since,the Hammingwindowisasmooth windowandreducesthesizeofthesidelobes,ithas beenemployedforwindowing. Forasignalframegiven by Xs(n),where n= 0,1,...,N-1,thewindowed signalisgivenby Xs(n)∗W(n). The Hammingwindow W(n)isdescribedas W(n) = 0. 54-0. 46cos(2πn N). (1) 2. 3. Featureextraction Meaningfulandrelevantfeatures,thatcouldsuffi-cientlydiscriminatethemusicgenreareextractedfrom theaudioclippingsinthisphase. Thetimbral,rhyth-micandpitchfeaturesextractedareasfollows: Table1. Listoffeatures. Feature number Feature class Number of features Featuresused 1-6Timbral (Spectral)6Mean&Std. Dev. of Spectral Centroid, Spectral Rolloff, Spectral Flux 7-32MFCC26Mean&Std. Dev. of MFCCfeatures 33-40Statistical8Mean&Std. Dev. of Spectral Skewness, Spectral Kurtosis, Flatness,Entropy 41-66Fractional MFCC26Mean&Std Dev. of Fractional MFCC 67-68Temporal2Meanof Zero Crossing Rateand Root Mean Square Energy 2. 3. 1. Timbralfeatures Thetimbralfeatureshavebeenobtainedfromthe frequencydomainofthesignal. Thesignalhasbeen firsttransformedintothefrequencydomainandvar-
|
216 Archivesof Acoustics-Volume42,Number2,2017 iousspectralfeatureshavebeenextractedfromthe spectrum. Spectral Centroid: The Spectral Centroidgivesthe measureofthebrightnessofasound. Thecentroid ofaspectralframecanbedefinedastheaveragefre-quencyweightedbyamplitudes,anddividingitbythe sumoftheamplitudesandcanbegivenby Spectral Centroid =N∑ k=1k M[k] N∑ k=1M[k],(2) where M[k]isthemagnitudeofthe FFTatfrequency binkand Nisthenumberoffrequencybins. Centroid findsthisfrequencyforagivenframeandthenfinds thenearestspectralbinforthatfrequency. Spectral Roll-off: The Spectral Roll-offisameasure ofthespectralshape. Itisdefinedasthefrequencybin Mbelowwhichthe85%ofthemagnitudedistribution isconcentrated Rolloff=M∑ k=1M[k] = 0. 85N∑ k=1k M[k]. (3) Spectral Flux:The Spectral Flux gives the rate ofchangeofthepowerspectrumandindicateshow quicklythepowerchangesfromframetoframe. The spectralfluxhasbeencalculatedbycomparingthe powerspectrumofoneframewiththepowerspectrum ofthepreviousframeandisgivenby: F=||M[k]-Mp[k]||,(4) where Mp[k]denotesthe FFTmagnitudeofthepre-viousframeintime. 2. 3. 2. Mel-frequency Cepstralcoefficients(MFCC) Mel Frequency Cepstral Coefficient (MFCC) is shorttimepowerspectralrepresentationofasignal. It providesusefulinformationregardingpsychoacoustic propertyofhumanauditorysystem. Blockschematic of MFCC is shownin Fig. 3. Thisfeature extrac-tionconsistsofpre-processing,pre-emphasis,framing, windowing,FFT,Triangularmelbandpassfilterand DCT. Inpreprocessingthesilencepartofthesignalis removedusing ZCRandenergyfeatures. Thishelps toreducethecomputationalcomplexityofthesys-tem. pre-emphasisisdonetoboostthehighfrequency Fig. 3. Blockschematicof Melfrequencycepstral coefficients. componentsusingfirstorderhighpassfilter. Framing isdonewith20msdurationwith10msoverlapping. Thenthesignaliswindowedwithhammingwindow. Further,thesignalistransformedintospectraldo-mainandpassedthrough24melfrequencytriangular bandpassfilters. Logvaluesofthesespectralcompo-nentshavebeenobtained. Discrete Cosine Transform (DCT)oftheselogvalueshavebeentakentodecor-relatethesignal. Thirteenmostsignificant MFCCco-efficienthavebeenobtainedforeachframe. Statistical valuessuchasmeanvalueofthesecoefficientshave beencomputedandusedasfeaturevector. 2. 3. 3. Fr FTbased MFCCfeatures Thetime-frequencyrepresentationofasignalis onaplanewithtwoorthogonalaxes,wherethetime axisisrepresentedhorizontallyas x(t)andthefre-quencyaxisisrepresentedvertically. Theconventional Fourier Transform X(ω)of a signal x(t)is repre-sentedalongthefrequencyaxis. The Fourier Trans-form F[x(t)] =X(ω),employsthe Fourier Transform operator FT,whichrotatesthetimeaxisanticlock-wiseby π/2radians. The Fractional Fouriertrans-formoperator Fr FTlikewiserotatesthesignalbyan anglethatisanon-multipleof π/2radians(Ashok Narayanan,Prabhu,2003). Thus,thesignalisrep-resentedinaplanethatmakesanangle' α'tothe timeaxiswhere α=a∗π/2. Thevalueof'a'canlie anywherebetween0and1. Forthiswork,thevalue a= 0. 98wasfoundtoyieldbetterresults. Fr FTuseslinearchirpsasabasisfunctionandthus, itoffersagreatdealofflexibilityintheprocessingof audiosignals(Bhalkeetal.,2016). Fig. 4. Blockschematicsof Fractional Mel Frequency Cepstral Coefficients(Fr MFCC). Fr FTisrepresentedby Fα. Fr FThasfollowing properties: (1)F0=I,zerorotationoridenticaloperatorcorre-spondstotimedomain, (2)Fπ/2=F,Correspondsto Fouriertransformop-erator, (3)F2π=I,2πrotation, (4)FαFβ=Fα+β,additivityofrotation. Fr FTofasignal x(t)withorder αisgivenby Fα(u) whichisrepresentedby Fα(u) =∞∫-∞x(t)K∝(t,u)dt,(5)
|
D. G. Bhalke,B. Rajesh,D. S. Bormane-Automatic Genre Class ification Using Fractional Fourier Transform... 217 where Kα(t,u)isatransformationkernelandisgiven by Eq. (6) Kα(t,u)= (√ 1-jcotα 2π) ej((t2+u2)/2)cotα-jutcscα, ifαisnotmultipleof 2π, δ(t-u),ifαismultipleof 2π, δ(t+u),ifα+πismultipleof 2π. (6) The Fr FTbased MFCCfeaturesarecalculatedby preprocessing,framing(with50%overlap),window-ing(Hammingwindowofsize1024),Fractional Fourier Transform,positioningona Melfilterbank,calculation of Logenergy,energycompactionby DCTandfinally calculationof Fractional MFCCfromthesignal. The Meanand Standard Deviationofthe Fr MFCCvalues, alongtheframes,foratexturewindowiscalculated andtakenasfeaturevectors. 2. 3. 4. Temporalfeatures Temporalfeaturesgivetheevolutionofthesignal overaperiodoftime. Thetemporalfeaturesthathave beenextractedare: Zero Crossing Rate (ZCR): Itisdefinedasthe numberoftimesasignalcrossesthe X-axis. ZCRgives anideaofthefrequencyofthesignalandisgivenby: ZCR=1 NN-1∑ 0|sgn[m(n)]-sgn[m(n-1)]|,(7) where Nisthetotalnumberofsamples, m(n)and m(n-1)isthesignalat n-thand( n-1)-thsample respectively. 2. 3. 5. Energyfeatures Root Mean Square Energy: Theglobalenergyof thesignal x[n]hasbeencomputedsimplybytaking therootaverageofthesquareoftheamplitude,also calledroot-mean-squareenergy(RMS) RMS=√1 Nn∑ i=1x2 i,(8) wherexiistheamplitudeofthesignal. 2. 3. 6. Statisticalfeatures Entropy:The Entropygivesadescriptionoftheinput curvepandindicateswhetheritcontainspredominant peaksornot. Itiscalculatedusing Shannonsentropy basedontheequation: H(X) =-n∑ i=1p(xi)logbp(xi),(9) wherepindicatesthecurveand bisthebaseofthe algorithm. Spectral skewness: The spectral skewness is the thirdcentralmomentandisameasureofthesym-metryofthedistribution. Apositivevalueindicatesapositivelyskeweddistributionwithfewvalueslarger thanthemeanandthushasalongtailtotheright. Anegativelyskeweddistributionhasalongertailto theleft. Skewnessisgivenby µ=∫ (x-µ1)3f(x)dx,(10) whereµisthemeanofthedistribution. Spectralkurtosis: Thespectralkurtosisisdefinedas thefourthstandardisedmomentandisdefinedas kurtosis=µ4 σ4,(11) whereµisthemeanand σisthevariance. Kurtosis givesthesharpnessofthepeaks. Spectral flatness: The spectral flatness indicates whetherthedistributionissmoothorspiky. Itiscal-culatedastheratiobetweenthegeometricmeanand thearithmeticmean: flatness=√ N-1∏ n=0x(n) N-1∑ n=0x(n) N. (12) 2. 4. Classification Afterthefeatureextraction,thesummarisedfea-turevalueswerefedtoclassifiersformodelingandpre-diction. Inthiswork,twoclassifierswereused:KNN (K-Nearest Neighbour)and SVM. KNN (K-Nearest Neighbours): The KNN is asimplealgorithmthatstoresallavailableclassla-belsanddecidestheclassofatestsamplebasedon a similarity measure (e. g., distance functions) ( Fu etal.,2011). KNNisusedinstatisticalestimation andpatternrecognitionasanonparametrictechnique whichdoesnotmakeanyassumptionsabouttheun-derlyingdatadistribution( Scaringellaetal.,2006). The KNNisalazylearningalgorithmthatdoesnot usethetrainingsamplestoperformanygeneralisation andtheentiredataisusedforthetestingphasewith-outdiscardingany. Sointhe KNNalgorithmthereis aminimalcostinvolvedinthetrainingbutahighcost involvedintestingbothintermsoftimeandmemory sincealldatapointsareutilizedandstoredfordeci-sionmaking. Variousdistancefunctionsareusedfor measuringandareasfollows; Euclidean:√k∑ i=1(xi-yi)2,(13) cityblock:k∑ i=1|xi-yi|,(14) Chebychev: max{|xi-yi|},(15)
|
218 Archivesof Acoustics-Volume42,Number2,2017 wherexiandyiaretwoinstancesandthedistancebe-tweenthemisdefinedby d(xi,yi). Thestandard Eu-clidean,cityblockand Chebychevdistancesaregiven by Eqs. (13),(14)and(15),respectively. The KNN classifierisaverysimplealgorithmthatworkswellfor realworlddatawheretheclassesmaybelinearlysep-arableornot. Thevalueof Kandthedistancemetric aloneneedtobetuned. Support Vector Machines(SVM): The Support Vector Machine is a supervised learning algorithm thatisusedforclassificationandregressionanaly-sis. The SVMbuildsamodelwithatrainingsetthat ispresentedtoitandassignstestsamplesbasedon themodel. An SVMmodelrepresentspointsofsam-plesinspace,mappedinsuchawaythatthesam-plesoftheseparatecategoriesareaswideaspossible (Scaringellaetal.,2006)Thechallengeistofindthe optimalhyperplanethatmaximisesthegap. Newsam-plesarethenmappedintothatsamespaceandclass predictionsaremadeasbelongingtoeithercategory basedonwhichsideofthegaptheyfallon. Theper-formanceofthe SVMisgreatlydependentonitskernel functions(linear,polynomialorexponential). Forthe purposeofthisexperiment,themorepopular Radial Basis Function(RBF)kernelhasbeenchosen. RBFis asquaredexponentialkernel,capableofhandlingcom-plexdataandismoreflexiblesinceitgivesaccessto allinfinitelydifferentiablefunctions. The RBFkernel fortwosamples xandx′isdefinedby K(x,x′) = exp( ∥x-x′∥2 2σ2),(16) where∥x-x′∥2isthesquared Euclideandistancebe-tweenthefeaturevectorsand σisafreeparameterand theparameter γ= 1/2σ2. 2. 5. Dataset Inthiswork,twotypesofdatabaseswereusedfor thepurposeofclassification: (1)Tamil Genres:Thedatabasehasbeenformed withclippingsfromcommerciallyavailable Tamil music CDs. Theclassicaldatabaseisfromde-votional songs composed in Carnatic style by Vedanayagam Sastriar,Subramanya Bharathiar and other devotional singers, and the Folk databaseisformedfrom Folksongsbypopular folksingerslike Dr. Pushpavanam Kandaswamy andothers. Tamil Classicalstyledevotionalmusic-103 songexcerpts, Tamil Folkmusic-113songexcerpts. Thetrainingsetforbothgenrescomprisedof70 songseach,whereasthetestingdatahad43Folk songsand33classicalsongs. (2)Western Genres:Thedatasetcontains10gen-res,eachrepresentedby100tracks,whichare each of 30-second duration from the GTZAN databasethatisavailableonline. Thetracksareall 22050Hz Mono16-bitaudiofilesin. wavformat. 100songseachof Rockand Classicalhavebeen chosenforthepurposeofclassification. Rockmu-sichasafastrhythmandbeatlikethe Tamil Folk music. Thenumberofsongsforthetrainingset andthetestsetforthe Rockand Classicalgenres werechosentobe60and40respectively. Sincethechorusofasongismoredescriptiveof thegenre,the30-secondexcerptsweretakenfromthe middleofeachsong,approximately2minutesafterthe beginningofeachpieceforboththe Tamilandwestern genres. 3. Experimentalresults Theresultsobtainedfromexperimentationhave beendiscussedinthissection. The30-secondsongex-cerptsofboth Tamilandwesterngenreswereframed into20msframeswitha50%overlapsothatonefea-turehasbeenobtainedevery10ms. Timbral,Spectral Shape,Temporal,MFCC,Fractional MFCCand Sta-tisticalfeatureswereextractedfromtheexcerpts. The statisticalvaluesofthe Meanand Standarddeviation werecalculatedfromthetemporalsummarizationof thefeaturevalues,alongthe20msframes. Thefea-turesweresortedtoidentifythebestfeaturedescrip-torsforthe Tamilandwesterngenres. The Meanand Standarddeviationofthefeatureswerefedasfeature inputvectorstotheclassifiers. K-NNand SVMclas-sifierswereusedfor Tamilgenreswhereasonly SVM wasemployedforthewesterngenres. Theresultswere observedasfollows: (1)Tamilgenres:Thefeaturesthatcontributedfor efficientdiscriminationof Tamilmusicwere Spec-tral Roll-off,Flux,Skewness,Kurtosisand Fr FT based MFCC. Therespectivegraphsofthecon-tributingfeatureshavebeenshownbelow. For this experiment the value of K= 2gave ahigheraccuracythanothervalues. The Radial Basis Function(RBF)wasusedasthekernelfunctionfor Fig. 5. Spectral Roll-offvaluesof Classicaland Folkgenres.
|
D. G. Bhalke,B. Rajesh,D. S. Bormane-Automatic Genre Class ification Using Fractional Fourier Transform... 219 Fig. 6. Spectral Fluxvaluesfor Classicaland Folkgenres. Fig. 7. Spectral Skewnessfor Classicaland Folkgenres. Fig. 8. Spectral Kurtosisvaluesfor Classical and Folkgenres. Fig. 9. Fractional MFCCvaluesfor Classical and Folkgenres. the SVMclassifier,sinceitefficientlyhandlesmulti-classproblems. Thevalueforgammaofthe RBFker-nelwaschosentobe0. 5andthecostfunction C= 5, forthepurposeofthisexperiment. Theresultshave beendisplayedin Table2. Table2. Accuraciesfordifferentcombinationsoffeatures with KNNand SVMclassifiersfor Tamilgenres. Feature Set Classifier%Accuracy Feature Set1: KNN66. 23 Spectral Rolloff +Flux +Skewness +Kurtosis +MFCCSVM(RBFKernel) 83. 50 Feature Set2: KNN70. 45 Spectral Rolloff +Flux +Skewness +Kurtosis +Fractional MFCCSVM(RBFKernel) 96. 05 Table3. Confusionmatrixforfeatureswith MFCC(Tamil genres)and Fr MFCC(Tamilgenres). Numberofsongs=76Predicted Classical Folk MFCC(Tamilgenres) Actual Classical32 (TP)1 (FN)33 Folk2 (FP)41 (TN)43 3442 Fr MFCC(Tamilgenres) Actual Classical21 (TP)12 (FN)33 Folk0 (FP)43 (TN)43 2155 TP-True Positivegivesthenumberofcorrectpredictionsth at Classicalsongis“Classical”, FN-False Negativeisthenumberofincorrectpredictionsth at a Classicalsongis“Folk”, FP-False Positiveisthenumberofincorrectpredictionsth at a Folksongis“Classical”, TN-True Negativenumberofcorrectpredictionsthata Folk songis“Folk”. Table4. Performance Measuresfor Feature Set1 and Feature Set2(Tamilgenres). Performance measure Feature Set1 (with MFCC) [%]Feature Set2 (with Fr MFCC) [%] Accuracy(TP+TN)/ (TP+TN+FP+FN)84. 2196. 05 Precision (TN/(FN+TN))78. 0097. 61 Recall (TN/(FP+TN))95. 34100. 00 Accuracy:Theproportionofthetotalnumberofcorrectpre-dictions. Precision:Theproportionofthepredictedpositivecasest hat werecorrect. Recall:The proportionofpositive casesthatwerecorrectl y identified.
|
220 Archivesof Acoustics-Volume42,Number2,2017 (2)Western Genres: Fromtheexperimentation,it wasobservedthatthe Fr FTbased MFCCwhen combinedwithotherspectralfeaturessignificantly increasedtheaccuracyofwesterngenresaswell. Butitwasalsoobservedthatthefeaturecombi-nationforaccurateclassificationofwesterngenres wasdifferentfromthatofthe Tamilgenres. While spectralskewnessandkurtosisgreatlycontributed toimprovetheclassificationaccuracyof Tamil genres,thesetwofeaturesdidnotmakeanycontri-butiontotheclassificationofthetwowesterngen-resofrockandclassicalmusic. Thetwofeatures alongwithspectralcentroidinfactreducedthe classificationaccuracy. Thefeaturecombination ofspectralroll-off,spectralfluxand Fractional Table5. Accuraciesfor Western Genresof Rock and Classical. Feature Set Classifier%Accuracy Spectral Centroid +Rolloff+Flux +Skewness +Kurtosis +Fr MFCC SVM (RBFkernel)75. 00 Feature Set1: Spectral Rolloff +Flux+MFCC80. 00 Feature Set2: Spectral Rolloff +Flux +Fractional MFCC96. 25 Table6. Confusion matrix for Featuresetwith MFCC (westerngenres)and Fr FTbased MFCC(westerngenres). Numberofsongs=80Predicted Classical Rock MFCC(westerngenres) Actual Classical40 (TP)14 (FN)54 Rock0 (FP)40 (TN)40 4054 Fr FTbased MFCC(westerngenres) Actual Classical40 (TP)3 (FN)43 Rock0 (FP)40 (TN)40 4043 TP-True Positivegivesthenumberofcorrectpredictionsth at Classicalsongis“Classical”, FN-False Negativeisthenumberofincorrectpredictionsth at a Classicalsongis“Rock”, FP-False Positiveisthenumberofincorrectpredictionsth at a Rocksongis“Classical”, TN-True Negativenumberofcorrectpredictionsthata Rock songis“Rock”. MFCCyieldedthehighestaccuracyof96. 25% comparedtoanaccuracyof80%forthesamefea-tureswith MFCC. With Fr FTbased MFCC,77 songsoutof80songswereclassifiedcorrectly. The misclassificationsweredueto3Classicalsongsbe-ingmisclassifiedas Rock. Table7. Performance Measuresfor Feature Set1 and Feature Set2(westerngenres). Performance measure Feature Set1 (with MFCC) [%]Feature Set2 (with Fr MFCC) [%] Accuracy(TP+TN)/ (TP+TN+FP+FN)85. 1096. 38 Precision (TN/(FN+TN))74. 0793. 02 Recall (TN/(FP+TN))100. 00100. 00 Accuracy:Theproportionofthetotalnumberofcorrectpre-dictions. Precision:Theproportionofthepredictedpositivecasest hat werecorrect. Recall:The proportionofpositive casesthatwerecorrectl y identified. 4. Conclusionandfuturework Inthispaper,anovelfeatureextractionschemefor automaticgenreclassificationof Indian Tamilmusic andwesternmusic,usingcombinationof Fr FTbased Fractional MFCCfeatureswith Timbralfeatureshave beenproposed. Since,Fractional Fourier Transform makesuseofchirpdecompositionthatishighlysuit-ableformusicsignalprocessing,theproposed Fr FT based MFCCfeaturesefficientlyclassifiesthe Tamil genresandwesterngenreswithhigheraccuracycom-paredtotheconventional MFCCfeatures. For Tamil music,thefeaturecombinationof Spectral Rolloff, Spectral Flux,Spectral Skewnessand Spectral Kurto-sis,whencombinedwith Fractional MFCCfeatures, outperformsallotherfeaturecombinations,toyield aclassificationaccuracyof96. 05%withan SVM(RBF kernel)classifier. Ithasalsobeenobservedthatthe Fr FTbased MFCCalongwith Spectral Roll-offand Spectral Fluxefficientlyclassifiesthewesterngenres (Rockand Classical)fromthe GTZANdatasetwitha higherclassificationaccuracyof96. 25%ascompared totheclassificationaccuracyof80%with MFCC Acknowledgments Wewishtothank JSPM's Rajarshi Shahu Col-legeof Engineering,Savitribai Phule Pune University, Pune,Indiaforprovidingthelabfacilitiesandguid-ance.
|
D. G. Bhalke,B. Rajesh,D. S. Bormane-Automatic Genre Class ification Using Fractional Fourier Transform... 221 References 1. Ashok Narayanan V., Prabhu K. M. M. (2003), The Fractional Fourier Transform:theory,implemen-tationand error analysis, Microprocessors and Mi-crosystems,27, 10, 511-521, doi: 10. 1016/S0141-9331(03)00113-3. 2. Bagul M.,Soni D.,Saravana Kumar K. (2014), Recognitionofsimilarpatternsinpopular Hindi Jazz songs by music data mining, International Confer-ence on Contemporary Computing and Informat-ics (IC3I), pp. 1274-1278, November 27-29, doi: 10. 1109/IC3I. 2014. 7019799. 3. Baniya B. K.,Ghimire D.,Lee J. (2014),Anovel approachofautomaticmusicgenreclassificationbased ontimbraltextureandrhythmiccontentfeatures,16th International Conferenceon Advanced Communication Technology(ICACT),pp. 96-102. 4. Benetos E.,Kotropoulos C. (2010),Non-Negative Tensor Factorization Appliedto Music Genre Classifi-cation,IEEETransactionson Audio,Speech,and Lan-guage Processing,18,8,1955-1967. 5. Bhalke D. G,Rao C. B. R.,Bormane D. S. (2014), Musical Instrument classification using higher order Spectra,International Conferenceon Signal Processing and Integrated Networks(SPIN),pp. 40-45,February 20-21,doi:10. 1109/SPIN. 2014. 6776918. 6. Bhalke D. G.,Rao C. B. R.,Bormane D. S. (2016), Automatic musical instrument classification using Fractional Fourier Transform based-MFCC features and counter propagation neural network, Journal of Intelligent Information System, 46,3,425-446,doi: 10. 1007/s10844-015-0360-9. 7. Chen S-H.,Chen S-H.,Truong T-K. (2012),Auto-maticmusicgenreclassificationbasedonwaveletpack-agetransformandbestbasisalgorithm,IEEEInterna-tional Symposiumon Circuitsand Systems(ISCAS), pp. 3202-3205,May20-23. 8. Chen S-H.,Chen S-H.,Guido R. C. (2010),Music genreclassificationalgorithmbasedondynamicframe analysisandsupportvectormachine,IEEEInterna-tional Symposiumon Multimedia(ISM),pp. 357-361, December13-15. 9. Fu Z.,Lu G.,Ting K. M.,Zhang D. (2011),ASur-veyofaudio-basedmusicclassificationandannotation, Multimedia IEEETransactions, 13,2,303-319. 10. Gaikwad S.,Chitre A. V.,Dandawate Y. H. (2014), Classification of Indian classical instruments using spectralandprincipalcomponentanalysisbasedcep-strumfeatures,International Conferenceon Electronic Systems,Signal Processingand Computing Technolo-gies(ICESC),pp. 276-279,January9-11. 11. Ghosal A.,Chakraborty R.,Chandra Dhara B., Saha S. K. (2012),Musicclassificationbasedon MFCC variantsandamplitudevariationpattern:ahierarchicalapproach,International Journalof Signal Processing, Image Processingand Pattern Recognition, 5,1,131-150. 12. Jothilakshmi S.,Kathiresan N. (2012),Automatic musicgenreclassificationfor Indian Music,Interna-tional Conferenceon Softwareand Computer Applica-tions(ICSCA2012),IPCSIT,Vol. 41,pp. 55-59,IAC-SITPress,Singapore. 13. Kini S.,Gulati S.,Rao P. (2011),Automaticgenre classificationof North Indiandevotional music,Na-tional Conferenceon Communications(NCC),pp. 1-5, January28-30,doi:10. 1109/NCC. 2011. 5734697. 14. Krishnaswamy A. (2003),Applicationofpitchtrack-ingto South Indianclassicalmusic,[in:]Proceedings of IEEEInternational Conferenceon Acoustics,Speech and Signal(ICASSP'03),Vol. 5,pp. V-557-60,April 6-10,doi:10. 1109/ICASSP. 2003. 1200030. 15. Kumar V.,Pandya H.,Jawahar C. V. (2014),Iden-tifying Ragasin Indianmusic,22nd International Con-ferenceon Pattern Recognition(ICPR),pp. 767-772, August24-28,doi:10. 1109/ICPR. 2014. 142. 16. Li T.,Ogihara M. (2006),Towardintelligentmusic informationretrieval,IEEETransactionson Multime-dia,8,3,564-574. 17. Li T.,Tzanetakis G. (2003),Factorsinautomatic musical genre classification of audio signals, IEEE Workshopon Applicationsof Signal Processingto Au-dioand Acoustics,pp. 143-146,October19-22. 18. Lim S-C.,Lee J-S.,Jang S-J.,Lee S-P.,Kim M. Y. (2012),Music-genre classification system based on spectro-temporalfeaturesandfeatureselection,IEEE Transactionsin Consumer Electronics, 58,4,1262-1268. 19. Meng A.,Ahrendt P.,Larsen J.,Hansen L. K. (2007),Temporalfeatureintegrationformusicgenre classification, IEEE Transactions in Audio, Speech, and Language Processing, 15,5,1654-1664. 20. Nagavi T. C., Bhajantri N. U. (2011),Overview of automatic Indian music information recognition, classification and retrieval systems, International Conference on Recent Trends in Information Sys-tems (Re TIS), pp. 111-116, December 21-23, doi: 10. 1109/Re TIS. 2011. 6146850. 21. Rao P. (2012),Audiometadataextraction:Thecase for Hindustani classical music, International Con-ference on Signal Processing and Communications (SPCOM), pp. 1-5, July 22-25, doi: 10. 1109/SP-COM. 2012. 6290243. 22. Rosner A.,Schuller B.,Kostek B. (2014),Classi-ficationofmusicgenrebasedonmusicseparationinto harmonicanddrumcomponents,Archivesof Acous-tics,39,4,629-638,doi:10. 2478/aoa-2014-0068. 23. Salamon J.,Gome E. (2012),Melodyextractionfrom polyphonicmusicsignalsusingpitchcontourcharacter-
|
222 Archivesof Acoustics-Volume42,Number2,2017 istics,IEEETransactionson Audio,Speech,and Lan-guage Processing,20,6,1759-1770. 24. Scaringella N.,Zoia G.,Mlynek D. (2006),Auto-maticgenreclassificationofmusiccontent:asurvey, IEEESignal Processing Magazine, 23,2,133-141. 25. Shao X., Maddage M. C., Changsheng Xu, Kankanhalli M. S. (2005),Automaticmusicsumma-rizationbasedonmusicstructureanalysis,IEEEInter-national Conferenceon Acoustics,Speech,and Signal Processing,2, ii/1169-ii/1172, March 18-23, doi: 10. 1109/ICASSP. 2005. 1415618. 26. Tamil Music, http://www. carnatica. net/tmusic. htm (accesson February2011). 27. Tzanetakis G., Cook P. (2002),Musical genre classificationofaudiosignals,IEEETransactionson Speech and Audio Processing, 10, 5, 293-302, doi: 10. 1109/TSA. 2002. 800560. 28. Vedanayagam Sastriar,http://www. sastriars. org(ac-cesson February2011).
|
README.md exists but content is empty.
- Downloads last month
- 11