1、語音情感識別中的分層特征與序貫識別專業(yè)名稱：計算機(jī)應(yīng)用技術(shù)申請人：羅梅指導(dǎo)教師：黃漢明教授論文答辯委員會i席o。紅塵彬委員：菠疊c鲴篷塑!c盥叢!L‘曼磚，『蠢StratifiedFeaturesandSequentialClassificationonSpeechAffectiveRecognitionName：L盟M亟SpecialField：Supervisor：HuangHanmingTechnologyofComputerAp

2、plication——AbstractGrade：2皿Interests：SpeechSignalProcessingDuetospeechaffectiverecognitionhasfarrangingapplicableprospects，SOemotionalspeechrecognitionresearchhasimportantsignificanceInthispaper，fourbasicemotionmodes：ang

3、ry，happyneutralandsadareconsidered，researchesarecarriedoutontheemotionalfeaturesextractionandselection，andalsoontherecognitionalgorithmsThecompletedreseachescanbesummarizedasthefollowing：(1)Animprovedmedianfiltersmoothin

4、galgorithmisproposedtoamelioratethesbortcomesofthetypicalmediansmoothingalgorithmforpitchfrequencyfeatureextractionTheresultsofexperimentsshowthattheimprovedmedianfiltersmoothingmethodismuchmoreeffectivetemove”outliers”t

5、hantheclassicmedianfiltersmoothingmethodFurthermore，insomeextent，thisnewsmoothingalgorithmsuppressestheinfluenceofnoisesontheirneighborhood，makingtheestimatedfundamentalfrequencytrajectorymuchmoreclosertotherealtrajector

6、yandachievingamoredesirablesmoothingeffect(2)Anewfearture～wavelet—fractalfeatureisextractedbasingonwavelettheoryandfractaltheory，extendingtheclassicemotionalfeaturesextractedfromspeechsignalsThesequencingwavelet—fractalf

7、eaturesaremodeledbyGaussianMixedModeI(GMM)TheexperimentresultsshowthatthisnewsequencingwaveletfractalfeatureisverygooddiscriminativetoneutralandsademotionsSothisnewproposedfeaturecanbeconsideredasanoptionalfeatureinspeec

8、haffectiverecognition(3)ThegoodnessofcorrectrecognitionratearegreatlyimpactedbydifferentemotionalfeaturesandclassiryingmethodsInthispapertwoclassifyingmethods：SVMandGMMareimplementedtobuildrecognitionmodelsrespectivelyba

9、singonfull—wavespeechfeaturesandsequencingwaveletffactalfeatures，inordertofindefficaciousemotionalfeaturescorrespondingtodifferentpairsofemotionalmodes(4)Astratifiedandsequentialclassificationsschemeisconductedforfourbas

10、icemotionalmodesThisisimplementedbasingontheemotionaldimensiontheorybyappropriateblendingthebasicemotionalmodesandformulatedcorrespodingcomparisonsofthecombinedemotionalmodeswhileappropriateanddifferentemotionalfeaturesa

11、reutilizedThewholerecognitionprocessisdividedintotwostepsThefirststepusesoneortwoclassifierstodistinguishtwoblendedemotionalmodes(angryandhappyVSneutralandsad)ThesecondstepUSeSappropriateanddiiffercntclassifierstodisting

12、uishbetweenbasicemotionalmodes：angryandhappy，neutralandsadThissequentialclassificationmethodeffectivelyreducestheerrorrateofinvigoratingemotionalmodesanddepressingemotionalmodes，raisingthetotalcorrectrecognitionrateofthe