基于不确定性线性判别分析的维吾尔语语音情感识别-塔什甫拉提·尼扎木丁.pdf

《基于不确定性线性判别分析的维吾尔语语音情感识别-塔什甫拉提·尼扎木丁.pdf》由会员分享，可在线阅读，更多相关《基于不确定性线性判别分析的维吾尔语语音情感识别-塔什甫拉提·尼扎木丁.pdf（7页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、Journal of Southeast University(English Edition) V0133，No4，PP437443 Dec2017 ISSN 1003-7985Emotion recognition of Uyghur speechuslnlg uncertain 1inear OlScrlmlnant analysis l - l Tashpolat Nizamidinl2 Zhao Lil Zhang Mingyan91 Xu Xinzhoul Askar Hamdulla2(1 Key Laboratory of Underwater Acoustic Signal

2、Processing of Ministry of Education，Southeast University，Nanjing 210096，China)(2 School of Information Science and Engineering，Xinjiang University，Ummqi 830046，China)Abstract：To achieve efncient and compact lowdimensionalfeatures for speech emotion recognition， a novel featurereduction method using

3、uncertain linear discriminant analysis isproposedUsing the same principles as for conventional lineardiscriminant analysis(LDA)， uncertainties of the noisy ordistoned input data are employed in order to estimate maximallydiscriminant directionsnle effectiveness of the proposeduncertain LDA(ULDA)is d

4、emonstrated in the Uyghurspeech emotion recognition taskThe emotional features ofUyghur speech，especially，the fundamental frequency andformantare analyzed in the collected emotional dataThenULDA is employed in dimensionality reduction of emotionalfeatures and better performance is achieved compared

5、withother dimensionality reduction techniquesThe speechemotion recognition of Uyghur is implemented by feeding thelowdimensional data to support vector machine f SVM 1based on the proposed ULDAThe experimental results showthat when employing an appropriate uncertainty estimationalgorithmuncertain LD

6、A outperforms山e conventional LDAcounterpart on Uyghur speech emotion recognitionKey words：Uyghur language；speech emotion corpus；pitch；formant；unceaain linear discriminant analysis f ULDA)DoI：103969Jissn10037985201704008q peech is oneo“he moSt effective waysofh岫anUcomputer interaction in the era of a

7、rtificial intelligenceThereforein the humancomputer interaction systemin order to make the machine understand humanemotionthe identification of the emotional state in thespeech signal becomes increasingly importantSpeechemotion recognition(SER)involves several differentfields，including speech signal

8、 processing，pattern recognition，machine learning，psychology，and so onFor SERit is generally regarded as the default methodfor capturing paralinguistic 1nformation to generate a single highdimensional representation of an utterance from aset of underlying lowlevel acoustic descriptorsPreviousReceived

9、 2017-05-17，Revised 2017-08-30Biographies：Tashpolat Nizamidin(1988一)，male，graduate；Zhao Li(corresponding author)，male，doctor，professor，zhaoliselledu cnFoundation item：The National Natural Science Foundation of China(No61673108，61231002)Citation：Tashpolat Nizamidin，Zhao Li，Zhang Mingyang，et a1Emo。tio

10、n recognition of Uyghur speech using uncertain linear discriminantanalysis【JJournal of Southeast University(English Edition)，2017，33(4)：437443 DOI：103969jissn10037985201704008investigations consistently demonstrate the usefulness ofthis technique when applied to a range of different SERproblemsDimen

11、sionality reduction is frequently used in the preprocessing stage to make the input data more suitable formodelingLinear discriminant analysis(LDA)21 is oneof the simplest and most popular transforms to enhanceclass separability for multidimensional observationsConventional LDA assumes that each cla

12、ss follows a normal distribution and classes share the same covariancestructureAlthough these assumptions do not generallyhold in practice，the conventional approach and its variants have been found useful in many applications including automatic speech and speaker recognitionWhen thedimensionality o

13、f the data becomes comparable with thenumber of samples per class，the sample covariance esti-mation becomes unstableRegularization and Bayesianestimation of covariance modeIs have been discussed inexiting 1iterature to overcome this issueIt is also possibleto obtain a nonlinear class separation usin

14、g subclass discriminant analysis and the kernel trick in LDA Wheneach class is composed of several partitionssubclass discriminant analysis aims to maximize the distance betweenthe class means and the subclass means in the same classat the salTle timeCompared to the principal component analysis(PCA)

15、13，classdependent dimensionality reduction isexpected to be more effective in modeling classesTheextension of LDA includes heteroscedastic LDAquadratic discriminant analysisand mixture discriminant analysisA distance preserving dimensionality reduction transform maps the D-dimensional data samples t

16、o a d-dimen-sional space(d2In this paper，we address the task of finding linear discriminant directions，using a probabilistic descriptioninstead of a pointestimation for an observationWeachieve such a probabilistic description by using socalledobservation uncertaintiesIn this approach，the feature万方数据

17、438 Tashpolat Nizamidin，Zhao Li，Zhang Mingyang，Xu Xinzhou，and Askar Hamdullaextraction process outputs the pointestimation of a featurevector along witll an uncertaintyThe pointestimation isassumed to form a Gaussian mean，while the corresponding variance is set as the estimated uncertaintyThroughout

18、 this paper，we note this process as an uncertain observationAccordinglywe utilize uncertain LDA f ULDAl”1 to account for the observation uncertainties in estimating scatter matrices for LDABased on the Uyghur speech emotion database，we setup a benchmark for Uyghur speech emotion recognitionwhich inv

19、olves a set of SER tasks in various trainingtestconditions Additionally，the dimensionality reductiontechnique ULDA f uncertain linear discriminant analysis)is appliedWe provide the complete data descriptionsystem architecture，experimental setup and evaluationperformanceThese can be used as a full re

20、ference forUyghur speech emotion recognition research1 AlgorithmDimensionality reduction techniques are widely appliedin speech emotion recognition research，such as PCALDA，locality preserving projections(LPP)1 51，local dis-criminant embedding(LDE)61，graphbased Fisher anal-ysis(GbFA)1 and so onIt is

21、important to note thatthese methods do not solve recognition and hypothesistesting problems directly，and they are used as a pre-processing stage to reduce dimensionaiitVA conventionalSER task requires a large number of features；henceweshould use an efficient dimensionality reduction techniqueto deal

22、 with this highdimensional caseIn this paperweapplied uncertain linear discriminant analysis(ULDA)todimensionality reduction11 Conventional LDALet X=xI，x2，xL be a set of L samples(features)，each sample belonging to one of K classes and partitio-ning the data into clusters CI，C2，CfThe conven-tiona】LD

23、A aims at finding a linear transforrnation ofthose features that can maximize the separability of theclustersEach class is assumed to be Gaussian distributedand has the same Gaussian covariance structureIn orderto find the discriminant directions，we first calculate thesample mean，l and class mean p女

24、as1 Lp=x， (1)p 2了己f 【l Jp女2 x，l Ecwhere CI is the cardinality of class kNext，the within-class Sw and betweenclass scatters SB are given asrS。=(p。一p)(p。一p)1k=JThe optimization problemthe FisherRap criteriont 1。3is then solved by maximizingaStWI，W2,-,W“K_1，硼伊叫器措) (5)where命，(i=1，2，K一1)is the i-th eigen

25、vector corresponding to the i-th eigenvalue A。，which is obtained bysolving(SBAiSw)命，=0The optimal projection matrixWis formed by putting d eigenvectors(dK一1)corresponding to the largest eigenvalues together and the newrepresentation of features is given byY=W“X (6)where Y 2yl，Y2，YL12 Uncertain LDAUL

26、DA 1s proposed for the case that the input data isavailable in the form of posterior distributions as茎=x1，点2，兰L，where each兰fR”is described by a respective probability density function正，(兰f)with，as theavailable informationThe conventional LDA can dealwith this type of data by only using the first-ord

27、er statis-tics as xf=肛f=E(兰，)and continue to calculate betweenand withinclass scatter matrices SB and SwULDA is developed in such a way that uses the secondorder statisticsof工l I，namely，三f=COV(工)when calculating the expected scatter matrices SB and SwThe eigenvectors cal-culated by employing expecte

28、d scatter matrices aredeemed to be more representative of variability in Xcompared to those obtained by considering X=E(茎I DThisclaim is tested by the application of speech emotion recognition in this paperIn the following，we describe howto find expected scatter matricesFor the sake of tractability，

29、we assume that posteriordistributions of observations are Gaussian as xf|，N(pt f，三)，which can be fully described by the firstand secondorder statisticsIn the following derivations，we rely onthe fact that the sum of Gaussian variables is anotherGaussian variableBy applying this principle，we obtain笆f，

30、肌lc。 善。三)(8)Nextwe find the distribution for the sample mean deviation 6-m，and the class mean deviation 6-as亟肚=兰，一胜。j亟。l，N(E(6-。)，coy(6-。，6-,k)E(6-腩)=p 肛，l E ccov(甄愚)=乏一南孙卉磊亟I=丛一肛=翘“三，(9)I，一N(E(6-。)，cov(6-。，6-k)、l_I，。上rp。一L，I_IT、，肛一r，L、JI肛一X，L女C。=wS万方数据Emotion recognition of Uyghur speech using unce

31、rtain linear discriminant analysis 439E(亟。)cov(曼t，蠡)=F 1亡乙弘I一了乙弘f E c L，=i三一2L Gwhere we need to take into account the correlation be一tween the mean and each sampleTo calculate Sw andSB，we just need to apply the linearity of the expectationoperator so thatleading to=1 f EC膏(兰#一坦t)(XEq。亟：l=一丝。)7)=cov

32、(亟舭驯+Eq。旧亟：)k=1 IECK；1k=1(笆。一丝)(丝。一楚)7)=E(亟。，亟：)=C。l(cov(亟。，点。)+E亟。】E(亟：)(12)Sw=Sw+Sw+cov(q舭=荟K等荟。三，C女I cov(亟女，亟。)=CKKI=1三，+L一2L C。 磊。乏lBy removing uncertainties，ie。，setting最=0，the posterior description of features becomes a Dirac delta func一 tion centered on xf and hence SB=SB and Sw=SwBy using Sw

33、and SB in Eq(5)，the ULDA transform isfound to be命+and the lowdimensional uncertain observations!=11，艺2，艺are obtained by Eq(6)with W+replacing W+All卫fR“(ducoflt)3J一0w-1H万方数据442 Tashpolat Nizamidin，Zhao Li，Zhang Mingyang，Xu Xinzhou，and Askar HamdullaTab2 Acoustic descriptorsDescriptor NumberIntenseLou

34、dnessMFCCl12LSP 0-7ZCRProbability of voicingF0F0 envelopeMax，min，and rangeRelative position of max and minArithtactic meanLinear regression coefficients and correspondingappoximate errorStandard deviation，skewness，kurtosisQuartiles and interquartile ranges】9Tab4 Average recognition accuracy on Uyghu

35、r speech emotion databaseAlgorithms AccuracvPCA+SVMLDA+SVMProposed ULDA+SVM4208441748 755 ConclusionsIn this paper，we present a Uyghur emotional databasewith basic emotions to analyze the emotional states in Uyghur speechThe emotional utterances are produced byUyghur native speakersThen，we propose a

36、 novel approach for coping with observation uncertainties in deri-ving an optimal linear discriminant feature transformThis so-termed uncertain LDA takes the probabilistic de-scription of observations into account in finding the mostdiscriminant directionsCompared to the existing algorithms，the prop

37、osed ULDA can effectively represent theemotional features in Uyghur speechOur experiments indicate that by employing fin appropriate uncertainty deftnition and a reliable uncertainty estimator，the Uyghurspeech emotion recognition can be improved further whenequipped with uncertain LDAReferences1】E1

38、Ayadi M，Kamel M S，Karray FSurvey on speechemotion recognition：Features，classification schemes，and databasesJ助ttern Recognition，201 1，44(3)：572587DOI：101016ipatcog201009020【2】Chu D，Liao L Z，Ng M K，et a1Incrementallinear discriminant analysis：A fast algorithm and comparisons f J11EEE Transactions on N

39、eural Networks and Learning Systems，2015，26(11)：27162735DOI：101109TNNLS20152391201【3Quan C，Wan D，Zhang B，et a1Reduce the dimensionsof emotional features by principal component analysis forspeech emotion recognition【C】Proceedings ofthe 2013IEEEsICEnternationat Symposium on System IntegrationKobeJapan

40、2013：222226DOI：101 109sii20136776653f41 Saeidi R，Astudillo R F，Kolossa DUncertain LDA：Ineluding observation uncertainties in discriminative transforms、j、IEEE Tnonsactions on Pattern Analysis andMachine Intelligence，2016，38(7)：14791488DOI：101016ipatcog20100902051 Soldera J，BehaiRe C A R，Scharcanski J

41、Customized orthogonal locality preserving projections with softmarginmaximization for face recognitionj1IEEE Transactionson Instrumentation and Measurement，2015，64(9)：24172426DOI：101109TIM20152415012【6】Zhou Y，Peng J，Chen C L PDimension reduction usingspatial and spectral regularized local discrimina

42、nt embedding for hyperspectral image classification【J】1EEE乃诅nsactions on Geoscience and Remote Sensing201553f21：108210957“W，Du QLaplacian regularized collaborative graphfor discriminant analysis of hyperspectral imagery【J1EEE Trnsactions on Geoscience and Remote Sensing、2016，54(12)：70667076DOI：10110

43、9tgrs2016259484881 Burldaardt F，Paeschke A，Rolfes M，et a1A database ofGerman emotional speechCProceedings of the 2005INTERSPEECHLisbonPortugal2005：1s11一、5209McGilloway S，Cowie R，Douglas-Cowie E，et a1Approaching automatic recognition of emotion from voice：Arough benchmarkC】Proceedings of the 2000 1SC

44、AWorkshop on Speech and Emotion：A Conceptuaf Framework for ResearchNewcastle，Northem Ireland，UK，2000：207212【10】Ablimit M，Eli M，Kawahara TPartly supervised Uyghur morpheme segmentation【C】Oriental Committee南r the Coordination and Standardization of Speech Databases and Assessment Techniques WorkshopKy

45、oto，Japan2008：7l一7611Pan S，Tao J，Li YThe CASIA audio emotion recognition method for audiovisual emotion challenge 201 1CAffective Computing and Intelligent InteractionFourth International ConferenceMemphisUSA201 l：388395DOI：1010079783胡2245718 5012】Eyben F，Wollmer M，Schuller BOpensmile：111e munich ve

46、rsatile and fast opensource audio feature extractorC1ACMnternational Conference on MultimediaFirenzeItaly，2010：1459146213】Xu X Z，Deng J，Zheng W M，et a1Dimensionality reduction for speech emotion features by multiscale kernels【C】Proceedings of Annual Conference of the International Speech Communicati

47、on AssociationDresdenGermany2015：15321536【14Wu S，Falk T H，Chan W YAutomatic speech emotionrecognition using modulation spectral featuresJ印eechCommunication，2011，53(5)：768785DOI：101016jspecom201008013眨832436万方数据Emotion recognition of Uyghur speech using unceaain linear discriminant analysis 443基于不确定性

48、线性判别分析的维吾尔语语音情感识别塔什甫拉提尼扎木丁1，2 赵 力1 张明阳1 徐新洲1 艾斯卡尔艾木都拉2(1东南大学水声信号处理教育部重占实验室，南京210096)(2新疆大学信息科学与工程学院，乌鲁木齐830046)摘要：为了在语音情感识别中获得高效、紧凑的低维特征，提出了一种新的基于不确定线性判别分析的特征约简方法用与传统LDA相同的原则，在最大判别方向的估计中引入带噪声或失真输入数据的不确定性在维吾尔语语音情感识别任务上验证了不确定性判别分析的有效性在该情感数据上，分析了维吾尔语的语音情感特征，着重对维吾尔语语音的基音频率和共振峰频率进行了详细分析利用不确定性线性判别分析对特征维数进行了降维研究，获得了比其他的常用降维技术更好的结果通过不确定性线性判别分析获得的低维数据供给支持向量机，实现了维吾尔语的语音情感识别实验结果表明，采用适当的不确定性估计算法时，在维吾尔语音情感识别任