Primitives-based evaluation and estimation of emotions in speech

Size: px
Start display at page:

Download "Primitives-based evaluation and estimation of emotions in speech"

Transcription

1 Speech Commuicatio 49 (27) Primitives-based evaluatio ad estimatio of emotios i speech Michael Grimm a, *, Kristia Kroschel a, Emily Mower b, Shrikath Narayaa b a Uiversität Karlsruhe (TH), Istitut für Nachrichtetechik (INT), Kaiserstraße 2, 7628 Karlsruhe, Germay b Uiversity of Souther Califoria (USC), Speech Aalysis ad Iterpretatio Laboratory (SAIL), 374 McClitock Aveue, Los Ageles, CA 989, USA Received 3 March 26; received i revised form 2 December 26; accepted Jauary 27 Abstract Emotio primitive descriptios are a importat alterative to classical emotio categories for describig a huma s affective expressios. We build a multi-dimesioal emotio space composed of the emotio primitives of valece, activatio, ad domiace. I this study, a image-based, text-free evaluatio system is preseted that provides ituitive assessmet of these emotio primitives, ad yields high iter-evaluator agreemet. A automatic system for estimatig the emotio primitives is itroduced. We use a fuzzy logic estimator ad a rule base derived from acoustic features i speech such as pitch, eergy, speakig rate ad spectral characteristics. The approach is tested o two databases. The first database cosists of 68 seteces of 3 speakers cotaiig acted emotios i the categories happy, agry, eutral, ad sad. The secod database cotais more tha utteraces of 47 speakers with authetic emotio expressios recorded from a televisio talk show. The estimatio results are compared to the huma evaluatio as a referece, ad are moderately to highly correlated (.42 < r <.85). Differet scearios are tested: acted vs. authetic emotios, speaker-depedet vs. speaker-idepedet emotio estimatio, ad geder-depedet vs. geder-idepedet emotio estimatio. Fially, cotiuous-valued estimates of the emotio primitives are mapped ito the give emotio categories usig a k-earest eighbor classifier. A overall recogitio rate of up to 83.5% is accomplished. The errors of the direct emotio estimatio are compared to the cofusio matrices of the classificatio from primitives. As a coclusio to this cotiuous-valued emotio primitives framework, speaker-depedet modelig of emotio expressio is proposed sice the emotio primitives are particularly suited for capturig dyamics ad itrisic variatios i emotio expressio. Ó 27 Elsevier B.V. All rights reserved. Keywords: Emotio estimatio; Emotio expressio variatios; Emotio recogitio; Emotio space cocept; Fuzzy logic; Ma machie iteractio; Natural speech uderstadig; Speech aalysis. Itroductio I recet years, automatic recogitio of emotios from speech ad other modalities has achieved growig iterest withi the huma machie iteractio research commuity. This iterest has merit, sice emotio recogitio is a essetial part of the road map to make commuicatio betwee humas ad computers more huma-like. Moreover, automatic assessmet of affective speech cotiues to gai importace i the cotext of speech data miig. * Correspodig author. Tel.: ; fax: address: grimm@it.ui-karlsruhe.de (M. Grimm). A search query o a speech archive may be located by the affective state of the target speaker i additio to, or istead of, just the sematic cotet. There is a large body of literature o the classical approach to emotio recogitio. Cowie et al. (2) give a excellet comprehesive review. Other examples of relevat work iclude (Dellaert et al., 996; Batlier et al., 2; Oudeyer, 23; Nwe et al., 23; Ververidis et al., 24). They all treat the emotio recogitio problem as a multiple classificatio task of several emotioal categories such as agry, happy, ad sad; or simply, egative ad o-egative. However, emotio psychology research has show that, as a alterative to categories, /$ - see frot matter Ó 27 Elsevier B.V. All rights reserved. doi:.6/j.specom.27..

2 788 M. Grimm et al. / Speech Commuicatio 49 (27) emotios ca also be described as poits i a multidimesioal emotio space. Cowie ad Corelius (23) give a review of the differet cocepts. The multi-dimesioal descriptio beefits from a greater level of geerality. Additioally, it allows for describig the itesity of emotios. These properties are ecessary for a aalysis of the iter- ad itra-speaker emotio expressio variability. I this paper, we take oe step beyod curret emotio recogitio algorithms ad propose a method for evaluatig ad automatically estimatig these emotio primitives that determie the locatio of a emotio i the multi-dimesioal emotio space from the speech sigal. Our approach cotributes to a importat challege i automatic emotio recogitio, amely recogizig emotios ot oly from acted speech of professioal speakers but also from spotaeous speech of o-professioal speakers. A icreasig umber of recet studies are based o spotaeous speech of aïve subjects (Douglas-Cowie et al., 23; Yu et al., 24; Vidrascu ad Devillers, 25; Schuller et al., 26). For these atural emotios, a descriptio usig just oe category label is ot sufficiet. I fact, the emotio space cocept allows for a more adequate descriptio of these emotios. I particular, gradual emotio trasitios, ad chages i the itesity of a emotio ca easily be described. Furthermore, speaker-depedet variability i the expressio of emotios, i.e., the spectrum of actually commuicated emotios ad the similarity of opposite emotios withi this rage, ca be characterized. These properties are crucial for aalyzig emotios i spotaeous, atural speech. Describig emotios by attributes alog bipolar axes was origially proposed by Wudt (896). Although a geeral emotio descriptio framework itself is still uder discussio i the emotio psychology commuity (see Scherer, 25 for istace), the cocept of descriptio by attributes has bee sice pursued i various forms. However, there has bee oly very limited research o automatic emotio recogitio withi the multi-dimesioal emotio space framework. Yu et al. (24) divide the 2D emotio space of valece ad arousal ito three ad five levels, respectively. They thereby trasform the task of determiig the cotiuous values of the emotio attributes to a more coveiet multiple classificatio task. For the LDC CallFried corpus, they achieve recogitio rates betwee 54% ad 67%, depedig o the umber of classes used. Valece ad arousal are classified separately. Vidrascu ad Devillers (25) report a recogitio accuracy of 82% o the twolevel classificatio of valece ito positive ad egative values. Their study is based o a large corpus of a medical call ceter. Fragopaagos ad Taylor (25) also motivate their choice of the activatio evaluatio space by emotio psychology. They divide the emotio space ito four regios for classificatio based o activatio (positive/egative) ad evaluatio (positive/egative), respectively. Tested o their ow database, geerated through a Wizard-of-Oz experimet, they report a average recogitio rate of 48.5% if oly acoustic features are used as iput to a artificial eural et (ANN). Combiig these features with facial expressio aalysis or emotioal saliece aalysis of the words or both improved the results by.3%, 2.5%, ad 2.%, respectively. I the case of separate classificatio, they report a average of 73.5% for activatio ad 64% for evaluatio. The results are improved by up to 6% by usig additioal iformatio chaels. Thus it ca be summarized that usig emotio dimesios as motivated by emotio psychology is a promisig step toward improvig the state-of-the-art i emotio recogitio. However, to our kowledge there is o previous study o directly estimatig the cotiuous-valued emotio primitives. We address this problem i this paper. I geeral, there are several ways to represet emotios i a multi-dimesioal emotio space. They ca be distiguished by the umber ad meaig of their basic etities (Cowie ad Corelius, 23; Kehrei, 22; Schröder et al., 2). The so-called dimesios are actually descriptive, geeric attributes of a emotio that fuctio as costituets. These costituets will be referred to as primitives i this paper. Note that these primitives are ot regarded as meta-features of emotio categories but as a fully complemetary descriptio of emotios. Two-dimesioal represetatios iclude oe primitive that describes the appraisal (or valece, evaluatio) takig values from positive to egative. The other emotio primitive describes the activatio (or arousal, excitatio), ad is sometimes motivated by the actio tedecies of emotios. Three-dimesioal represetatios additioally iclude a primitive defiig the apparet stregth of the perso, which is referred to as domiace (or power). This third dimesio is ecessary to distiguish ager from fear for istace, sice the domiace (or the ability to hadle a situatio) is the oly discrimiatig elemet i this case. We chose the combiatio of the followig three emotio primitives (Kehrei, 22): Valece (V) positive vs. egative, Activatio (A) excitatio level high vs. low, ad Domiace (D) apparet stregth of the speaker, weak vs. strog. Our study cosists of the followig mai parts. () We itroduce a robust ad efficiet huma emotio assessmet method to produce the three-dimesioal emotio refereces, which provides quick-ad-easy assessmet of authetic emotios i atural speech. (2) We propose a rule-based fuzzy logic method to estimate the cotiuous values of the emotio primitives from acoustic features derived from the speech sigal. (3) We assess our emotio recogitio method based o primitives by comparig the results with covetioal categorical classificatio. (4) We fially show how emotio primitives are well suited for capturig the speaker-depedet variability i emotio expressio. The rest of the paper is orgaized as follows.

3 M. Grimm et al. / Speech Commuicatio 49 (27) Table Databases used for this study Descriptio Laguage Emotio type No. speakers No. seteces Avg. o. se./speaker No. evaluators EMA Am. Eglish Acted VAM I Germa Authetic VAM II Germa Authetic Sectio 2 itroduces the data we use. Sectio 3 describes the huma evaluatio of emotioal speech i terms of the three emotio primitives. Sectio 4 presets details of estimatig the three-dimesioal emotio primitives from speech usig a rule-based fuzzy logic classifier. Sectio 5 shows the results ad provides a compariso betwee the results of real-valued primitives estimatio ad discrete emotio classificatio. Sectio 6 details how the speaker depedet variability preset i expressed emotios ca be described i terms of the emotio primitives. Sectio 7 provides coclusios ad outlies future work. 2. Data For this study we use two databases. The first corpus, called the EMA Corpus, cotais speech with acted emotios i America Eglish. The secod corpus, called the VAM Corpus, 2 cotais spotaeous speech with authetic emotios that was recorded from guests i a Germa TV talk-show. Table summarizes the key facts about both databases: laguage, emotio elicitatio type (acted or atural), umber of speakers ad seteces, average umber of seteces per speaker, ad umber of evaluators. The two databases are deliberately chose to cotai two differet emotio productio styles. While the spotaeous speech database is used to push the applicatio orieted research o authetic emotios, the acted speech database is used to provide a compariso with state-ofthe-art emotio categorizatio. The use of these two differet databases was also partly motivated by our goal to explore if the proposed methods hold good for two differet laguages ad across atural ad acted emotioal speech. We however test the emotio primitives estimator oly with speech of the same laguage that has bee used for traiig. Nevertheless, similar recogitio results for both laguages, Eglish ad Germa, may imply cross-cultural robustess of the proposed method. speakers read 2 seteces, ad the male speaker read 28 seteces. These recordigs cosist of (4) seteces, each of them repeated 5 times i 4 differet emotios. Each block cosisted of 4 seteces that were radomized withi the block. Each repetitio for a give emotio was block-wise; the subjects produced all seteces withi a give block i the same emotio. This was repeated for each of the four emotios, i a radom order of emotios (Lee et al., 25). All seteces are i Eglish, spoke by ative speakers of America Eglish. As described i (Grimm et al., 26a), the EMA corpus was evaluated by four ative speakers of America Eglish. For each setece, the evaluators assiged oe of the category labels from amog happy, agry, sad, eutral, ad other to the utterace. The average huma recogitio rate of the acted emotios was 8.8%. Happy emotio was most poorly recogized (76.6%). This was due to the fact that several seteces that were iteded to be happy were perceived as eutral emotios. See Table 2 for the cofusio matrix, give as a average of all three speakers i the database. Similar results were reported by Bulut et al. (22). To assess the iter-evaluator agreemet, we used the parameter j derived from the Kappa statistics (Carletta, 996), j ¼ P A P : ðþ P This parameter describes the level of iter-evaluator agreemet j 2 [, ]. P A represets the proportio of the evaluators that assiged the same class label, ad P corrects for their agreemet by chace. We foud a moderate iterevaluator agreemet of j =.48 betwee the four evaluators, which is a typical value for such categorical emotio assessmet by humas (cf. Vidrascu ad Devillers, 25) Natural speech corpus The secod database, the VAM Corpus, cosists of recordigs of ivited guests i a Germa TV show called 2.. Acted speech corpus The EMA Corpus (Lee et al., 25) cotais 68 seteces of emotioal speech, produced by oe professioal (f) ad two o-professioal (f/m) speakers. The female The acroym EMA stads for electromagetographic articulatory study. However, the articulatory data were aalyzed i a differet study. 2 The acroym VAM is the abbreviatio of the talk-show title Vera am Mittag. Table 2 Cofusio matrix of emotio class labelig of EMA corpus, i percet, by four huma listeers (j =.48) Agry Happy Neutral Sad Other Agry Happy Neutral Sad

4 79 M. Grimm et al. / Speech Commuicatio 49 (27) Vera am Mittag. 3 This show is broadcasted Moday through Friday o Free-TV with a regular duratio of oe hour. Each show cotais five dialogues betwee two or three guests, moderated by a achorwoma. The speakers mostly discuss persoal problems or family issues i a spotaeous uscripted fashio. The first part of this corpus, VAM I, was first used i (Grimm ad Kroschel, 25a). The secod part, VAM II, cotais seteces from additioal speakers i the talk-show that were evaluated after the iitial experimet was reported. I total, the VAM database cotais 2 emotioal utteraces from 47 speakers (m/36f). All sigals were recorded usig a samplig frequecy of 6 khz ad 6 bit resolutio. The dialogues were maually segmeted at the utterace level. Each utterace cotaied at least oe itermediate phrase. The video stream was ot aalyzed i this study. The speakers were selected by a prelimiary evaluatio durig the data segmetatio ad selectio step to guaratee that each speaker showed both eutral expressios ad at least some emotioal deviatio from the eutral state. The emotios covered i the spotaeous speech corpus are summarized i Sectio 3.2, after itroducig the evaluatio method. 3. Primitives-based emotio evaluatio Evaluatio of the emotios cotaied i the speech data was doe through huma listeer tests. A popular, ad widely used tool for the huma evaluatio of emotios i a multi-dimesioal emotio space is the Feeltrace tool developed by Cowie et al. (2). This istrumet allows for time-cotiuous ad value-cotiuous assessmet of emotios i the activatio evaluatio space. The method is based o Plutchik s cocept of defiig emotios as positios withi a circle, wherei the agle determies the character of the emotio, ad the distace from the origi determies the itesity of the emotio. We did ot use this istrumet sice () it is restricted to a two-dimesioal emotio space that has bee show ot to be adequate for distiguishig certai emotios such as fear ad ager, for istace, see (Cowie et al., 2), (2) a square space (or a cube i 3D) is more appropriate for our chose primitives, sice valece is a bipolar rather tha a agular-type periodic etity (Russell ad Mehrabia, 977), ad (3) time-cotiuous evaluatio was ot well suited for our utterace-level uits. The evaluatio method described below builds upo our prelimiary work reported i (Grimm ad Kroschel, 25b,c). The ovel aspects reported here iclude a more ituitive scalig ad orietatio of the axes. Additioally, the evaluatio tool was exteded to iclude elicitatio of the evaluator s backgroud such as laguage comprehesio capabilities, ad self-evaluatio of his/her persoality 3 Eglish: Vera at oo; Vera is the ame of the talk-show host. with respect to hadlig emotios. However, we could ot observe ay statistically sigificat differece i the evaluatio of emotios by humas of differet cultural backgroud or differet self-evaluatio. Sectio 3. describes the utterace-based assessmet method. Sectio 3.2 cotais the primitives-based evaluatio results o the acted ad spotaeous speech databases, respectively. 3.. Evaluatio method For the evaluatio of emotios i the 3D emotio space of valece, activatio, ad domiace, we propose to use the self assessmet maikis (SAMs) proposed origially by Lag (98). This istrumet cosists of a array of five images per primitive (see Fig. ). These images allow us to avoid the use of categorical labels for emotios. Evaluatig emotios usig SAMs is fast ad very ituitive. Note that the SAMs origiate from self-assessmet, however i our case, the speech was ot evaluated by the speakers themselves. For each utterace i the database, 6 6 N, the evaluator k, 6 k 6 K, chooses 3 values ^x ðiþ ;k oe for each emotio primitive i 2 {valece, activatio, domiace}. The selectio of the icos is mapped to iteger values {,2,3,4,5} ad the trasformed to uity space [,+]. For ituitive comprehesio of the primitives, the axes are orieted from egative to positive (valece), calm to excited (activatio), ad weak to strog (domiace). Although it ca be assumed that each evaluator assesses the emotioal cotet of a utterace to the best of his/her kowledge, the assessmet does ot ecessarily reflect the emotio truly felt by the speaker. There is a umber of iput- ad output-specific issues, as Fragopaagos ad Taylor call it (Fragopaagos ad Taylor, 25). Both the expressio ad perceptio of emotios are subject to several iflueces, such as display rules ad cogitive effects. From a sigal processig viewpoit, these iflueces ca be modeled as sigals with superimposed oise o top of the hidde true emotio. Assumig a ubiased esemble of evaluators, the hidde emotio ca best be deter- Fig.. Self assessmet maikis (Fischer et al., 22). This evaluatio tool is used for a text-free, three-dimesioal assessmet of emotio i speech.

5 M. Grimm et al. / Speech Commuicatio 49 (27) mied by estimatig it from the combied assessmet results of several evaluators. I (Grimm ad Kroschel, 25c), two differet methods to merge the evaluatio results of several evaluators were discussed. We choose the evaluator weighted estimator (EWE), x EWE;ðiÞ ¼ X K P K r ðiþ k¼ rðiþ k ^xðiþ ;k : k k¼ This estimator averages the idividual evaluators resposes, ad takes ito accout that each evaluator is subject to a idividual amout of disturbace durig evaluatio. This is doe by itroducig evaluator-depedet weights r ðiþ k, r ðiþ k P N ¼ ^xðiþ ;k P N N ¼^xðiÞ ;k x ðiþ P N N ¼ xðiþ ð2þ ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi r ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 : 2 P N ¼ ^xðiþ ;k N P N ¼^xðiÞ ;k P N ¼ xðiþ N P N ¼ xðiþ These evaluator-depedet weights measure the correlatio betwee the listeer s resposes, f^x ðiþ ;k g ¼;...;N, ad the average ratigs of all evaluators, fx ðiþg ¼;...;N, where x ðiþ ¼ K X K k¼ ^x ðiþ ;k : The assessmet quality is determied by calculatig the stadard deviatio r ðiþ of the evaluatios, vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u r ðiþ X K 2 ¼ t ^x ðiþ ;k xewe;ðiþ : ð5þ K k¼ A compariso of the EWE ad a maximum likelihood estimator, as well as a discussio of itrisic evaluatio errors due to emotio space quatizatio by the SAMs ca be foud i (Grimm ad Kroschel, 25c). It was show that the EWE yields up to 2% better results tha the maximum likelihood estimator. It coverges ito the maximum likelihood estimator i the case of equal weights for all estimators. Corrected by the emotio space quatizatio error, the actual evaluatio error was foud to be ðiþ ¼ r ðiþ bðkþ; ð6þ where bðkþ 2ð ; pffiffi Š is a costat bias, ad depeds o the umber of evaluators K. ð3þ ð4þ 3.2. Evaluatio results The described method for evaluatig emotio primitives was applied to both the EMA ad the VAM database. For each setece, the EWE estimate was calculated accordig to (2). With K = 8, 7, ad 6 evaluators for the 3 corpora, respectively, we use a comparatively high umber of evaluators for this task. Most other studies ivolve 2 evaluators (Vidrascu ad Devillers, 25; Vidrascu ad Devillers, 25; Schuller et al., 25) or, at most, 4 5 evaluators (Yu et al., 24; Fragopaagos ad Taylor, 25; Lee ad Narayaa, 23). The iter-evaluator agreemet was measured by determiig the stadard deviatios r ðiþ of the assessmet ad the correlatio coefficiets r (i) usig Eqs. (5) ad (3), respectively. The stadard deviatio, o the oe had, measures the suitability of a particular setece for our task. A low stadard deviatio idicates that the emotioal expressio is perceived by all huma listeers similarly. The iter-evaluator correlatio, o the other had, measures the agreemet amog the idividual evaluators ad thus focuses o the more geeral evaluatio performace. The average results for each database are reported i Table 3. O average, the stadard deviatio was betwee.28 ad.38 for each primitive. Thus, the stadard deviatio was slightly above.25, i.e. half the distace betwee two SAMs, idicatig good evaluatio results. There was o sigificat differece betwee the database cotaiig acted emotios ad the databases cotaiig authetic emotios. Note that the stadard deviatio icludes the quatizatio error due to the discretizatio of the SAMs i the emotio primitive space. The iter-evaluator correlatio was moderate to high with values i the rage of The correlatio was i geeral greater for the EMA database tha for the VAM database. This result is probably due to the more stereotypical ature of the emotios portrayed by the actors. Furthermore it could be observed that the valece primitive yields a smaller iter-evaluator correlatio tha activatio or domiace. I particular, for the VAM database cotaiig authetic emotios from talk show dialogues this result might be due to the fact that the distributio of valece values was arrower tha the distributios of activatio or domiace, ad thus evaluators deviatios by the same amout resulted i a smaller correlatio coefficiet. Table 3 Average stadard deviatio r ad correlatio coefficiet r for the emotio primitives evaluatio of the EMA corpus ad the VAM I/II corpus by huma listeers, averaged over all speakers ad all seteces Stadard deviatio r Correlatio coefficiet r Valece Activatio Domiace Valece Activatio Domiace EMA VAM I VAM II

6 792 M. Grimm et al. / Speech Commuicatio 49 (27) Occurrece Valece VAM I+II Corpus Activatio All correlatio coefficiets were statistically sigificat (p <.). Fig. 2 shows the histogram of the emotios i the VAM talk show database. It has to be oted that a large percetage of the utteraces i this database cotais eutral or egative speech with high activatio ad domiace values. This distributio is probably due to the ature of the topics discussed i the talk show, which iclude family problems, paterity questios ad friedship issues. The restrictiveess of recordig a wide spectrum of emotios is a itrisic problem i spotaeous speech processig. Moreover, averagig assessmet results aturally teds to result i a more Gaussia-like distributio. We addressed the problem of uequally distributed emotios i the database by usig a rule-based emotio primitives estimator that is ot iflueced by a priori probabilities, cf. Sectio 4. We calculated the stadard deviatios for each setece to discard a few outliers: All utteraces that had bee evaluated with a stadard deviatio r ðiþ > :5 for ay of the emotio primitives i 2 {V,A,D} were ot used for the further study. I may of these cases the utteraces were too log, ad cotaied more tha oe, coflictig emotios. The remaiig utteraces were all evaluated with a deviatio of oe SAM or less. Thus the VAM I database was reduced to 49 utteraces (98.2%), the VAM II database was reduced to 489 utteraces (94.2%), ad the EMA database was reduced to 64 seteces (9.3%), respectively. The resultig ew average stadard deviatios were margially smaller tha the oes reported above. For compariso: Cowie et al. report similar stadard deviatios usig the Feeltrace tool o a differet evaluatio task ad three evaluators (Cowie et al., 2). From Cowie et al. (2, Fig. 4), it ca be iferred that the stadard deviatio of their chose primitives, evaluatio ad activatio, was i the rage of.2.3. It ca be summarized that () the SAMs are well suited for evaluatig emotios i speech, (2) the iter-evaluator correlatio o activatio ad domiace is higher tha o valece, ad (3) the iter-evaluator correlatio o acted emotios is slightly higher tha the oe o authetic emotios. 4. Primitives-based emotio estimatio I this sectio, we focus o automated emotio estimatio from speech. Specifically, we describe a fuzzy logic Domiace Fig. 2. Histogram of emotios i VAM corpus. iferece system for primitives-based automated emotio estimatio. Fuzzy logic leds itself to cotiuous-valued estimates of emotios i spotaeous atural speech. Such cotiuous-valued emotio estimates are ecessary to automatically assess temporal dyamics i emotio, or to tackle the problem of a speaker-depedet variability i emotio expressio. The emotio estimator described below builds upo our previous work (Grimm ad Kroschel, 25a; Heradez, 25). The prelimiary results reported were based o a fractio of our database, ad a smaller umber of evaluators tha i the preset study. Fuzzy logic was chose because the ature of liguistic emotio class labels is iheretly fuzzy ad vague. Fuzzy logic trasforms crisp values ito fuzzy values usig membership grades. The crisp values that are extracted from the acoustic speech sigal are processed as liguistic variables. For istace, the mea value of the pitch is processed as a high, medium, or low mea pitch value. While the idea of applyig fuzzy logic to the problem of emotio recogitio has bee previously discussed with other objectives (Lee ad Narayaa, 23; Huag ad Akagi, 25), fuzzy logic has ot bee used to estimate cotiuous values of emotio primitives yet. We cosider this approach i this paper. A alterative would be multidimesioal, kerel-based regressio methods (Schölkopf ad Smola, 22), which we will aalyze i the future. Sectio 4. describes the pre-processig ad the acoustic feature extractio. Sectio 4.2 details the proposed estimatio method, ad it describes how the rule system is derived from acoustic features. 4.. Pre-processig ad feature extractio All sigals were sampled at 6 khz samplig rate ad a resolutio of 6 bit. They were processed at the utteracelevel. The acoustics of emotioal speech have bee studied for may years. I geeral, the differeces of the prosodic characteristics betwee emotioally loaded ad eutral speech have bee aalyzed ad reported (Murray ad Arott, 993; Base ad Scherer, 996; Cowie et al., 2). The major acoustic speech features cosidered iclude fudametal frequecy f ( pitch ), speakig rate, itesity, ad voice quality. For example, Murray ad Arott state that agry speech is slightly faster, has a very much higher pitch average, much wider pitch rage, ad higher itesity (Murray ad Arott, 993). Some of these characteristics ca be related directly to physiological chages i the vibratio of the vocal chords. The umber of features extracted from the speech sigal varies sigificatly from approximately basic features such as mea values ad rage i pitch ad itesity (Lee et al., 2) to 276 i the case of systematic applicatio of fuctioals to a set of basic trajectories (Schuller et al., 26). This spectrum results from the fact that it is still uclear which features are suited best, ad that the

7 M. Grimm et al. / Speech Commuicatio 49 (27) feature set is highly depedet o the data ad the classificatio task. We chose M = 46 acoustic features that were derived from the pitch ad the eergy cotour of the speech sigal, as well as features related to the speakig rate ad spectral characteristics. This is i accordace with most studies i this field. The emotioally colored prosody of the utterace is thus described i terms of statistics, such as mea value, stadard deviatio, ad percetiles. The followig features were extracted from the speech sigal: Pitch related features: f mea value, stadard deviatio, media, miimum, ad maximum, 25% ad 75% quatiles, differece betwee f maximum ad miimum, differece of quartiles. These features related to the fudametal frequecy f describe the itoatio ad speakig melody. They capture mootoe speech or highly acceted syllables, for example. The pitch was estimated usig autocorrelatio method sice it was show to give good results i a wide rage of applicatios (Nagel, 25). Speakig rate related features: ratio betwee the duratio of uvoiced ad voiced segmets, average duratio of voiced segmets, stadard deviatio of duratio of voiced segmets, average duratio of uvoiced segmets, ad stadard deviatio of duratio of voiced segmets. These features describe the temporal characteristics i the prosody. They might reveal whether the speech souds urged or relaxed, for example. Itesity related features: itesity mea, stadard deviatio, maximum, 25% ad 75% quatiles, ad differece of quartiles. Itesity related features are used to capture the eergy i speakig, ad helps to discrimiate shoutig from sad or depressed speech, for example. Spectral features: mea value ad stadard deviatio of 3 Mel frequecy cepstral coefficiets (MFCC). The MFCCs are very commo i automatic speech recogitio (ASR). While the short-term statistics are very useful for phoeme recogitio, the log-term statistics idicate voice quality ad are thus ofte icluded i the feature set for automatic emotio recogitio. A pricipal compoet aalysis (PCA) was applied to the feature set to reduce the umber of features usig a eigevalue threshold of.. However, the estimatio results were best whe all features were used. The described features form the basis of the rule system i the fuzzy iferece emotio estimator. Each feature m, 6 m 6 M, is related to each of the emotio primitives x (i) :¼ x EWE,(i), i 2 {V,A,D}, to be estimated Rule-based fuzzy logic emotio estimatio The fuzzy logic classifier cosists of the three compoets fuzzificatio, iferece, ad defuzzificatio (Kroschel, 24). Util the last step of defuzzificatio, the emotio primitives x (i), i 2 {V,A,D}, will therefore be represeted by the fuzzy, liguistic variables x ðv Þ! B ðv Þ l 2 B ðv Þ ¼fegative; eutral; positiveg x ðaþ! B ðaþ l 2 B ðaþ ¼fcalm; eutral; excitedg x ðdþ! B ðdþ l 2 B ðdþ ¼fweak; eutral; strogg: The membership fuctios of these fuzzy variables are depicted i Fig. 3. The three emotio primitives are estimated separately. I the followig, we briefly summarize the three elemets of the fuzzy iferece system reported i (Grimm ad Kroschel, 25a; Heradez, 25). Fig. 4 shows a example of the fuzzy logic iferece system. It is based o a example of two features ad iteds to give a compact overview o the idividual elemets of the fuzzy logic estimator Fuzzificatio I the fuzzificatio step, each feature m is trasformed from a crisp value v m to three fuzzy variables A j, j =, 2,3, where A j 2 A ¼flow; medium; highg: ð8þ This reflects the fact that, for example, the absolute value of the average fudametal frequecy is ot relevat, but it is importat to distiguish betwee low, medium ad high pitch average. These geeralized terms A are applied to all features, although whe talkig about a idividual feature we would rather use more specific terms for descriptio. The degree of membership l j,m of each liguistic variable A j is determied by the value of the membership fuctio l Aj;mðaÞ at the poit of the crisp feature value, l j;m ¼ l Aj;mðv m Þ: ð9þ The membership fuctios of the iput features, l Aj ;mðv m Þ, have the same piecewise liear shape as the membership fuctios of the emotio primitives depicted i Fig. 3. This shape is commo i fuzzy logic systems (Kroschel, 24), ad it has bee foud to be well suited for emotio represetatio, too (Heradez, 25). The edges of the membership fuctios are determied by the % ad 9% quatiles of the distributios of the feature values, Q = Q (m) ad Q 9 = Q 9 (m). Thus for feature m the membership fuctios are defied as follows: 8 >< ; v m < Q v l ;m ¼ m Q 5 ; Q Q Q 5 6 v m < Q 5 ðþ >: ; v m P Q 5 μ B (x (V) ).5 Valece.5.5 x (V) egative eutral positive μ B (x (A) ).5 Activatio.5.5 x (A) calm eutral excited μ B (x (D) ).5 Domiace ð7þ.5.5 x (D) Fig. 3. Membership fuctios of the emotio compoets. weak eutral strog

8 794 M. Grimm et al. / Speech Commuicatio 49 (27) Fuzzificatio Implicatio ( V ) ( V ) ( V ) ( V) μaj.8 () μ () μ B.8 B2.8 () μ () B3.8 ( ) ρ V > 8 Q v 8 Q 5 Q ( V ) () ( V ) () ( V ) () ( V) () μ μ μ μ Aj.9 B.9 B2.9 B3.9 ρ ( V) 9 < Qv 9 Q 5 Q Aggregatio max Accumulatio Defuzzificatio ( V ) ( V ) ( V ) ( V ) ( V) μ () μ () μ () μ B B2 B3 B () μb () max ( ^x V ) =.24 Fig. 4. Fuzzy logic for emotio primitives estimatio: fuzzificatio, iferece (implicatio, aggregatio, accumulatio) ad defuzzificatio for valece, a example usig two features. 8 ; v m < Q v >< m Q ; Q Q 5 Q 6 v m < Q 5 l 2;m ¼ v m Q 9 ; Q Q 5 Q v m < Q 9 >: ; v m P Q 9 8 >< ; v m < Q 5 v l 3;m ¼ m Q 5 ; Q Q 9 Q v m < Q 9 >: ; v m P Q 9 ; ðþ ð2þ where Q 5 is a abbreviatio for (Q + Q 9 )/2. Fig. 4 (top left) shows the fuzzificatio of a crisp feature value (v 8 ) ito membership grades of the fuzzy variables egative, eutral, positive for valece (l,8,l 2,8,l 3,8 ) Iferece The rule base is derived from the correlatio qm ðiþ betwee the acoustic features m, with 6 m 6 M, ad the emotio x ðiþ attested by huma listeers (cf. Sectio 3), P N q ðiþ m ¼ ¼ v ð m; v m Þ x ðiþ xðiþ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffir ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi P ; ð3þ 2 N ¼ð v m; v m Þ 2 xðiþ x ðiþ P N ¼ where v m ¼ NP N ¼ v m;, adx ðiþ ¼ NP N ¼ xðiþ. Thus for each liguistic iput variable A j 2 A, oe rule is formulated to lik it to each liguistic output variable B l 2 B, IF v m is A j THEN x ðiþ is B ðiþ l : ð4þ The sig of the correlatio coefficiet q ðiþ m thereby determies which variable pairs (j, l) are related to oe aother, l ðiþ m ¼ 2 þðj m 2Þsigðq ðiþ m Þ; j m ¼ ; 2; 3; m ¼ ;...; M: ð5þ For example, we derive the followig rules from q ðaþ 8 ¼ :8 ad q ðv Þ 9 ¼ :4, respectively: If the pitch rage (m =8)is high (j 8 = 3) the the activatio (i = A) is excited ðl ðaþ 8 ¼ 3Þ. Or, if the average pause duratio betwee cosecutive words (m = 9) is high (j 9 = 3) the the valece (i = V)isegative ðl ðv Þ 9 ¼ Þ. That is why i Fig. 4 a low feature value of feature 8 is implied to egative, while a low feature value of feature 9 is implied to positive. The absolute value jq ðiþ m j determies the importace of the rule ad is defied as the rule weight. This way the rules are geerated i a automatic way. The expert kowledge is reflected i the fact that features which are highly correlated with the emotio primitives are give large impact i the rule base.

9 M. Grimm et al. / Speech Commuicatio 49 (27) Applyig the rules to each acoustic feature, each fuzzy iput yields a degree of support for each fuzzy output variable. This degree of support is the membership grade of the feature assiged to the appropriate fuzzy variables of the emotio primitives, multiplied with the rule weight. I the aggregatio step, the degrees of support of all acoustic features are fused usig a maximum operator. This maximizatio has bee foud to be superior to sum aggregatio (Heradez, 25). I Fig. 4 the aggregatio is foud i vertical directio. It ca be applied before or after implicatio as it gives the same result for the chose operators. The implicatio draws the actual coclusio ad scales the output membership fuctios by the appropriate aggregated degree of support usig a multiplicatio (product implicatio). By this, the output membership fuctios depicted i Fig. 3 are scaled to the appropriate level determied by the rules. Still, the emotio primitives are described by the values of the fuzzy variables, cf. (7). I the accumulatio step, the three scaled membership fuctios of the fuzzy variables B ðiþ l, l =,2,3, are accumulated usig a maximum operator. For valece, for example, this accumulatio fuses the three fuzzy variables egative, eutral, ad positive that were used to scale the three output membership fuctios ito oe curve for valece. Thus the result is oe cotiuous membership fuctio l ðiþ B ðaþ describig the fuzzy value of valece, activatio, ad domiace for i 2 {V,A,D}, respectively. Fig. 4 shows the accumulatio i the bottom row: the three output membership fuctios resultig from the implicatio are depicted i thi lies, while the accumulatio result is depicted i a bold lie. Aother elaborate example of this iferece system is described i (Grimm ad Kroschel, 25a) Defuzzificatio The last step i the fuzzy logic emotio estimator, defuzzificatio, trasforms the fuzzy values to crisp values. We use the cetroid method, R ^x FL;ðiÞ ¼ a lðiþ B ðaþ da R lðiþ B ðaþ da : ð6þ The defuzzificatio is show i Fig. 4, bottom right, for the same sample values. The result is oe crisp, real-valued umber per emotio primitive. The crisp emotio estimates are ormalized by a costat factor c =.63 to accout for the restricted iterval of possible values. This restrictio results from the shape of the membership fuctios l ðiþ B l. The cetroid method has bee show to give better results tha the mea of maximum or bisector method, respectively (Heradez, 25). 5. Results The fuzzy logic emotio estimator was applied to the EMA ad the VAM databases. The rule base was costructed for male ad female speakers separately, ad for all speakers joitly. I total, we defied 3 differet scearios, as itemized i Table 4. The umber of speakers ad the umber of seteces used i each sceario are stated as #Sp. ad #Se., respectively. For the emotio estimatio test usig the described fuzzy logic method we geerated the rule base from all available utteraces, depedig o the sceario, ad the tested with each of the utteraces sequetially. Due to the large database size ad the ature of the rule base i the classifier, which was determied i a geeric way from the correlatio betwee the idividual acoustic features ad the emotio primitives, we foud that there was o differece i the results if the tested utterace was excluded from the traiig set. The followig sectios discuss these results. Sectio 5. describes several aspects of the estimatio results, for example the impact of the idividual features, the emotio type, ad the speaker depedecy. Sectio 5.2 compares the results of the real-valued emotio primitives estimatio to the results of a classical discretized emotio classificatio task. 5.. Estimatio results The automatic estimatio of emotio primitives was assessed by calculatig the estimatio error e ðiþ ¼ x EWE;ðiÞ ^x FL;ðiÞ ð7þ for each utterace i the database, 6 6 N, ad for each emotio primitive i 2 {V, A, D} separately. The mea error for each sceario is reported i Table 4. O average, the estimatio error was betwee.6 ad.28 for the differet scearios. These errors are comparable to the stadard deviatio i the huma evaluatio of emotios i the emotio space, cf. Table 3 ad Eq. (6). They are i the rage of half the distace betwee two evaluatio maikis ad thus otably small. Sice these results are based o a large umber of samples, N >, they ca be regarded as statistically sigificat. The correlatio coefficiets were also used as a meas for assessig the estimatio results. For all scearios, the correlatio coefficiet was foud to be positive, ad for most of the scearios we foud fairly high correlatio i the rage of Agai, for the EMA database the correlatio coefficiet was i geeral higher tha for the VAM database. Usig separate classifiers for male ad female speakers, or icreasig the database size by joiig VAM I ad II, did ot improve the correlatio sigificatly Impact of idividual features The rakig of the rules with respect to the rule weights jq ðiþ m j was database ad speaker-geder depedet. The highest correlatio betwee a idividual acoustic feature ad the emotio was foud to be the 25% quatile of the

10 796 M. Grimm et al. / Speech Commuicatio 49 (27) Table 4 Mea error ad correlatio to referece for the automated emotio primitives estimatio ( Estimatio colums) of the EMA corpus ad the VAM I/II corpus, respectively Sceario Selectio Database #Sp. #Se. Estimatio Evaluatio Mea error Mea correlatio Mea error Mea correlatio All VAM I All VAM II All VAM I + II Male VAM I Female VAM I Male VAM II Female VAM II Male VAM I + II Female VAM I + II Female EMA() Female EMA(2) Male EMA(3) All EMA Maual results of the huma evaluatio are added for compariso ( Evaluatio colums). pitch (m = 6) for the male speakers i the VAM I database with q ðv Þ 6 ¼ :7; q ðaþ 6 ¼ :89; ad q ðdþ 6 ¼ :9. Other features of high correlatio to emotio primitives icluded the f media ad the stadard deviatio of the 3rd ad 3th MFCC, respectively. I geeral, it was observed that all features had a ozero rule weight, ad thus at least partly cotribute some iformatio about the emotio. Although there is some agreemet with the feature rakig foud i other studies (Schuller et al., 25), it has to be suspected that the rakig strogly depeds o the data used Natural emotios vs. acted emotios I geeral, the error was higher for the atural speech database VAM (.7 i sceario 6, to.28 i scearios 4 ad 5) tha for the acted speech database EMA (.6 i sceario, to.23 i sceario ). The error i recogizig acted emotios (.9 i sceario 3) was approximately 2% below the error i recogizig authetic emotios (.24 i sceario 3), whe all speakers were used. Thus, acted emotios yielded better recogitio results. The result of the huma evaluatio of the acted emotios (EMA) also gave higher iter-evaluator agreemet (.66.8) tha for the spotaeous, atural emotios i the VAM corpus (.4.65). For these stereotype emotios the machie recogitio eve outperformed the huma evaluatio i terms of error ad correlatio Impact of the database The two modules VAM I ad II are comparable i the umber of speakers ad seteces. The mea error of the estimatio was similar for the two modules VAM I ad II (.27 ad.23). This was ot the case for the evaluatio (.2 ad.5), which gave a smaller error for VAM II. However, this might be due to the differet set of evaluators or due to more explicit emotioal cotet which led to a smaller iter-evaluator deviatio. The correlatio coefficiet of the estimatio was higher for VAM I tha for VAM II (.7 ad.43 i scearios ad 2, respectively). The same tedecy was foud for the evaluatio (.65 ad.56). This discrepacy might be due to the differet a priori distributios of the emotios i the two database modules. While VAM I has a very arrow distributio of emotio primitives, i particular for valece, VAM II has a much wider distributio. For example, the variace for valece i VAM I was oly 67% of the variace i VAM II. Sice the variace of the distributio cotributes reciprocally to the correlatio coefficiet (cf. Eq. (3)), the correlatio coefficiet for VAM II is itrisically smaller tha for VAM I. Whe we compared the separate modules to the joit database (scearios 3) we foud that the mea error for the joit database was betwee the results of the two modules. The same observatio was made whe oly male (scearios 4, 6, 8) or oly female speakers (scearios 5, 7, 9) were aalyzed. Thus we could ot make the observatio that a larger database automatically yielded better results. However, the advatage gaied from usig the larger joit VAM database i terms of more thorough traiig of the rule base might have bee over-compesated by differet emotio expressio styles of the differet speakers Geder depedecy For the scearios comparig geder-specific versus ogeder-specific rule bases, we could ot observe cosistet tedecies despite the fact that male ad female speakers express their emotios differetly (Schröder et al., 2). Oly for the male speakers i VAM II, a geder-depedet rule base gave remarkable improvemets from.23 to.7 i average error values (scearios 2, 6, 7). For all other scearios of VAM I (scearios, 4, 5) or VAM I + II (scearios 3, 8, 9), the mea error was approximately the same, ad idepedet from usig separate estimators for both male ad female speakers or oe joit estimator.

11 M. Grimm et al. / Speech Commuicatio 49 (27) This result might be caused by the method the rules i the rule system are derived, which are all based o the same feature set for all scearios. Those features that might ideed have differet values depedig o the emotio ad the geder were ruled out by features depedig o the emotio oly. This ca easily happe i the case of greater rule weights Speaker depedecy For the EMA database we foud that the estimators usig oly oe speaker to build the rule system (scearios,, 2), whe tested o that particular speaker s speech, achieved better results tha the estimator usig all three speakers (sceario 3). This coicides with previous work idicatig that speaker-depedet traiig of the estimator achieves the most accurate emotio classificatio results Compariso with respect to the emotio primitives The mea errors ad correlatio coefficiets metioed i Table 4 cotai the average values for each of the emotio primitives valece, activatio, ad domiace. We observed that for each of the scearios, the error i the valece dimesio was greater tha the error i either the activatio or domiace dimesio. For sceario 3 (all speakers of VAM I + II) for istace, the idividual mea values of the estimatio error are.34,.9, ad.2, for valece, activatio, ad domiace, respectively. Whe observig the correlatio betwee the emotio estimates ad the emotio referece we observed a similar discrepacy. For each sceario, the correlatio coefficiet was smaller for valece tha for activatio or domiace. I sceario 3 for istace, the correlatio coefficiet for valece was.34, while it was.73 ad.7 for activatio ad domiace, respectively. Thus valece was more difficult to estimate automatically usig our feature set tha activatio or domiace. Note that the better results obtaied for activatio ad domiace are i accordace with the iter-evaluator correlatio, cf. Table Compariso to the maual results of the evaluatio doe by listeers The estimatio results achieved with the automated method ca be compared to the evaluatio results of the huma listeers. While for classical emotio categorizatio we could simply compare cofusio matrices, it is less ituitive to compare emotio assessmet results withi the emotio space approach i terms of real-valued emotio primitives. I Table 4, the last two colums recall the evaluatio results of Table 3, corrected for the emotio space quatizatio (cf. Sectio 3.) ad more detailed for the idividual scearios. The evaluatio measures oly compare the evaluators amogst themselves, i.e. we do ot have a groud truth for the emotio. Therefore, the compariso betwee the estimatio error ad the evaluatio error, as well as the compariso of the respective correlatio coefficiets ca oly be a rough oe. We ca see that the estimatio performs i the same rage as the huma evaluatio. I most cases the huma agreemet is still higher tha the machie recogitio. However, a compariso of the huma evaluatio performace ad the machie recogitio with respect to the database modules is difficult due to the differet set of evaluators. It ca be summarized that the fuzzy logic system for emotio estimatio is well suited for this task, sice a small estimatio error ad a moderate to high correlatio to the emotio referece ca be observed. Estimatio results for acted emotios i performed speech yield slightly higher results tha for authetic emotios expressed i urehearsed atural speech Comparig emotio primitives estimatio ad categorical classificatio The proposed method to estimate emotio primitives from acoustic features derived from the speech sigal was show to yield low errors ad a high correlatio to the referece. Sice most previous studies target discrete emotio categories istead of cotiuous-valued emotio primitives, it is importat to compare the estimatio errors from the fuzzy logic estimator to the performace achieved with categorical recogitio. To facilitate this compariso, we aalyzed the EMA corpus, sice each setece i this corpus has a defied emotio category label. The recogitio rates for the EMA corpus may serve as a rule of thumb for the VAM corpus, for which strict emotio categorizatio could ot be applied due to the lack of objective emotio class labels. Toward eablig this compariso, we coducted a straightforward classificatio task. We used the emotio primitives estimates as a iput to a k-earest eighbor (knn) classifier. The knn classifier estimates the a posteriori probability P(Q x) of the emotio class Q 2 {agry, happy,eutral,sad} give the emotio primitives x =(x (V), x (A),x (D) ) T for a local volume elemet i the 3D emotio space from give traiig data (Kroschel, 24). Depedig o the feature set ad the data, this classifier achieved results comparable to support vector machies ad liear discrimiat classifiers (Lee ad Narayaa, 25; Hammal et al., 25), ad i some cases outperformed these other classifiers (Dellaert et al., 996; Yu et al., 22). The experimet was doe usig leave-oe-out (LOO) cross-validatio. We tested k 2 {,3,5,7,9} as a parameter of the knn classifier, ad applied the classificatio to both the idividual speakers of the EMA database (speaker-depedet task) ad the combied set of all seteces across all speakers i the database (speaker-idepedet task). The best recogitio rate was achieved usig k = 7 for the speaker-depedet, ad k = 9 for the speaker-idepedet task. For these parameters, the average recogitio rate was 83.5% for the speaker-depedet, ad 66.9% for the speaker-idepedet task. The cofusio matrices are

arxiv: v1 [cs.dl] 22 Dec 2016

arxiv: v1 [cs.dl] 22 Dec 2016 ScieceWISE: Topic Modelig over Scietific Literature Networks arxiv:1612.07636v1 [cs.dl] 22 Dec 2016 A. Magalich, V. Gemmetto, D. Garlaschelli, A. Boyarsky Uiversity of Leide, The Netherlads {magalich,

More information

Natural language processing implementation on Romanian ChatBot

Natural language processing implementation on Romanian ChatBot Proceedigs of the 9th WSEAS Iteratioal Coferece o SIMULATION, MODELLING AND OPTIMIZATION Natural laguage processig implemetatio o Romaia ChatBot RALF FABIAN, MARCU ALEXANDRU-NICOLAE Departmet for Iformatics

More information

E-LEARNING USABILITY: A LEARNER-ADAPTED APPROACH BASED ON THE EVALUATION OF LEANER S PREFERENCES. Valentina Terzieva, Yuri Pavlov, Rumen Andreev

E-LEARNING USABILITY: A LEARNER-ADAPTED APPROACH BASED ON THE EVALUATION OF LEANER S PREFERENCES. Valentina Terzieva, Yuri Pavlov, Rumen Andreev Titre du documet / Documet title E-learig usability : A learer-adapted approach based o the evaluatio of leaer's prefereces Auteur(s) / Author(s) TERZIEVA Valetia ; PAVLOV Yuri (1) ; ANDREEV Rume (2) ;

More information

'Norwegian University of Science and Technology, Department of Computer and Information Science

'Norwegian University of Science and Technology, Department of Computer and Information Science The helpful Patiet Record System: Problem Orieted Ad Kowledge Based Elisabeth Bayega, MS' ad Samso Tu, MS2 'Norwegia Uiversity of Sciece ad Techology, Departmet of Computer ad Iformatio Sciece ad Departmet

More information

Consortium: North Carolina Community Colleges

Consortium: North Carolina Community Colleges Associatio of Research Libraries / Texas A&M Uiversity www.libqual.org Cotributors Collee Cook Texas A&M Uiversity Fred Heath Uiversity of Texas BruceThompso Texas A&M Uiversity Martha Kyrillidou Associatio

More information

Fuzzy Reference Gain-Scheduling Approach as Intelligent Agents: FRGS Agent

Fuzzy Reference Gain-Scheduling Approach as Intelligent Agents: FRGS Agent Fuzzy Referece Gai-Schedulig Approach as Itelliget Agets: FRGS Aget J. E. ARAUJO * eresto@lit.ipe.br K. H. KIENITZ # kieitz@ita.br S. A. SANDRI sadra@lac.ipe.br J. D. S. da SILVA demisio@lac.ipe.br * Itegratio

More information

CONSTITUENT VOICE TECHNICAL NOTE 1 INTRODUCING Version 1.1, September 2014

CONSTITUENT VOICE TECHNICAL NOTE 1 INTRODUCING  Version 1.1, September 2014 preview begis oct 2014 lauches ja 2015 INTRODUCING WWW.FEEDBACKCOMMONS.ORG A serviced cloud platform to share ad compare feedback data ad collaboratively develop feedback ad learig practice CONSTITUENT

More information

Application for Admission

Application for Admission Applicatio for Admissio Admissio Office PO Box 2900 Illiois Wesleya Uiversity Bloomig, Illiois 61702-2900 Apply o-lie at: www.iwu.edu Applicatio Iformatio I am applyig: Early Actio Regular Decisio Early

More information

HANDBOOK. Career Center Handbook. Tools & Tips for Career Search Success CALIFORNIA STATE UNIVERSITY, SACR AMENTO

HANDBOOK. Career Center Handbook. Tools & Tips for Career Search Success CALIFORNIA STATE UNIVERSITY, SACR AMENTO HANDBOOK Career Ceter Hadbook CALIFORNIA STATE UNIVERSITY, SACR AMENTO Tools & Tips for Career Search Success Academic Advisig ad Career Ceter 6000 J Street Lasse Hall 1013 Sacrameto, CA 95819-6064 916-278-6231

More information

Management Science Letters

Management Science Letters Maagemet Sciece Letters 4 (24) 2 26 Cotets lists available at GrowigSciece Maagemet Sciece Letters homepage: www.growigsciece.com/msl A applicatio of data evelopmet aalysis for measurig the relative efficiecy

More information

part2 Participatory Processes

part2 Participatory Processes part part2 Participatory Processes Participatory Learig Approaches Whose Learig? Participatory learig is based o the priciple of ope expressio where all sectios of the commuity ad exteral stakeholders

More information

Speech Emotion Recognition Using Support Vector Machine

Speech Emotion Recognition Using Support Vector Machine Speech Emotion Recognition Using Support Vector Machine Yixiong Pan, Peipei Shen and Liping Shen Department of Computer Technology Shanghai JiaoTong University, Shanghai, China panyixiong@sjtu.edu.cn,

More information

VISION, MISSION, VALUES, AND GOALS

VISION, MISSION, VALUES, AND GOALS 6 VISION, MISSION, VALUES, AND GOALS 2010-2015 VISION STATEMENT Ohloe College will be kow throughout Califoria for our iclusiveess, iovatio, ad superior rates of studet success. MISSION STATEMENT The Missio

More information

Speech Recognition at ICSI: Broadcast News and beyond

Speech Recognition at ICSI: Broadcast News and beyond Speech Recognition at ICSI: Broadcast News and beyond Dan Ellis International Computer Science Institute, Berkeley CA Outline 1 2 3 The DARPA Broadcast News task Aspects of ICSI

More information

Analysis of Emotion Recognition System through Speech Signal Using KNN & GMM Classifier

Analysis of Emotion Recognition System through Speech Signal Using KNN & GMM Classifier IOSR Journal of Electronics and Communication Engineering (IOSR-JECE) e-issn: 2278-2834,p- ISSN: 2278-8735.Volume 10, Issue 2, Ver.1 (Mar - Apr.2015), PP 55-61 www.iosrjournals.org Analysis of Emotion

More information

Module 12. Machine Learning. Version 2 CSE IIT, Kharagpur

Module 12. Machine Learning. Version 2 CSE IIT, Kharagpur Module 12 Machine Learning 12.1 Instructional Objective The students should understand the concept of learning systems Students should learn about different aspects of a learning system Students should

More information

Mandarin Lexical Tone Recognition: The Gating Paradigm

Mandarin Lexical Tone Recognition: The Gating Paradigm Kansas Working Papers in Linguistics, Vol. 0 (008), p. 8 Abstract Mandarin Lexical Tone Recognition: The Gating Paradigm Yuwen Lai and Jie Zhang University of Kansas Research on spoken word recognition

More information

Learning Methods in Multilingual Speech Recognition

Learning Methods in Multilingual Speech Recognition Learning Methods in Multilingual Speech Recognition Hui Lin Department of Electrical Engineering University of Washington Seattle, WA 98125 linhui@u.washington.edu Li Deng, Jasha Droppo, Dong Yu, and Alex

More information

WHEN THERE IS A mismatch between the acoustic

WHEN THERE IS A mismatch between the acoustic 808 IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 14, NO. 3, MAY 2006 Optimization of Temporal Filters for Constructing Robust Features in Speech Recognition Jeih-Weih Hung, Member,

More information

2014 Gold Award Winner SpecialParent

2014 Gold Award Winner SpecialParent Award Wier SpecialParet Dedicated to all families of childre with special eeds 6 th Editio/Fall/Witer 2014 Desig ad Editorial Awards Competitio MISSION Our goal is to provide parets of childre with special

More information

OPTIMIZATINON OF TRAINING SETS FOR HEBBIAN-LEARNING- BASED CLASSIFIERS

OPTIMIZATINON OF TRAINING SETS FOR HEBBIAN-LEARNING- BASED CLASSIFIERS OPTIMIZATINON OF TRAINING SETS FOR HEBBIAN-LEARNING- BASED CLASSIFIERS Václav Kocian, Eva Volná, Michal Janošek, Martin Kotyrba University of Ostrava Department of Informatics and Computers Dvořákova 7,

More information

Word Segmentation of Off-line Handwritten Documents

Word Segmentation of Off-line Handwritten Documents Word Segmentation of Off-line Handwritten Documents Chen Huang and Sargur N. Srihari {chuang5, srihari}@cedar.buffalo.edu Center of Excellence for Document Analysis and Recognition (CEDAR), Department

More information

On March 15, 2016, Governor Rick Snyder. Continuing Medical Education Becomes Mandatory in Michigan. in this issue... 3 Great Lakes Veterinary

On March 15, 2016, Governor Rick Snyder. Continuing Medical Education Becomes Mandatory in Michigan. in this issue... 3 Great Lakes Veterinary michiga veteriary medical associatio i this issue... 3 Great Lakes Veteriary Coferece 4 What You Need to Kow Whe Issuig a Iterstate Certificate of Ispectio 6 Low Pathogeic Avia Iflueza H5 Virus Detectios

More information

Intra-talker Variation: Audience Design Factors Affecting Lexical Selections

Intra-talker Variation: Audience Design Factors Affecting Lexical Selections Tyler Perrachione LING 451-0 Proseminar in Sound Structure Prof. A. Bradlow 17 March 2006 Intra-talker Variation: Audience Design Factors Affecting Lexical Selections Abstract Although the acoustic and

More information

Evaluation of Teach For America:

Evaluation of Teach For America: EA15-536-2 Evaluation of Teach For America: 2014-2015 Department of Evaluation and Assessment Mike Miles Superintendent of Schools This page is intentionally left blank. ii Evaluation of Teach For America:

More information

USER ADAPTATION IN E-LEARNING ENVIRONMENTS

USER ADAPTATION IN E-LEARNING ENVIRONMENTS USER ADAPTATION IN E-LEARNING ENVIRONMENTS Paraskevi Tzouveli Image, Video and Multimedia Systems Laboratory School of Electrical and Computer Engineering National Technical University of Athens tpar@image.

More information

Human Emotion Recognition From Speech

Human Emotion Recognition From Speech RESEARCH ARTICLE OPEN ACCESS Human Emotion Recognition From Speech Miss. Aparna P. Wanare*, Prof. Shankar N. Dandare *(Department of Electronics & Telecommunication Engineering, Sant Gadge Baba Amravati

More information

AP Statistics Summer Assignment 17-18

AP Statistics Summer Assignment 17-18 AP Statistics Summer Assignment 17-18 Welcome to AP Statistics. This course will be unlike any other math class you have ever taken before! Before taking this course you will need to be competent in basic

More information

Statewide Framework Document for:

Statewide Framework Document for: Statewide Framework Document for: 270301 Standards may be added to this document prior to submission, but may not be removed from the framework to meet state credit equivalency requirements. Performance

More information

Entrepreneurial Discovery and the Demmert/Klein Experiment: Additional Evidence from Germany

Entrepreneurial Discovery and the Demmert/Klein Experiment: Additional Evidence from Germany Entrepreneurial Discovery and the Demmert/Klein Experiment: Additional Evidence from Germany Jana Kitzmann and Dirk Schiereck, Endowed Chair for Banking and Finance, EUROPEAN BUSINESS SCHOOL, International

More information

Author's personal copy

Author's personal copy Speech Communication 49 (2007) 588 601 www.elsevier.com/locate/specom Abstract Subjective comparison and evaluation of speech enhancement Yi Hu, Philipos C. Loizou * Department of Electrical Engineering,

More information

Grade 6: Correlated to AGS Basic Math Skills

Grade 6: Correlated to AGS Basic Math Skills Grade 6: Correlated to AGS Basic Math Skills Grade 6: Standard 1 Number Sense Students compare and order positive and negative integers, decimals, fractions, and mixed numbers. They find multiples and

More information

Lip reading: Japanese vowel recognition by tracking temporal changes of lip shape

Lip reading: Japanese vowel recognition by tracking temporal changes of lip shape Lip reading: Japanese vowel recognition by tracking temporal changes of lip shape Koshi Odagiri 1, and Yoichi Muraoka 1 1 Graduate School of Fundamental/Computer Science and Engineering, Waseda University,

More information

International Journal of Computational Intelligence and Informatics, Vol. 1 : No. 4, January - March 2012

International Journal of Computational Intelligence and Informatics, Vol. 1 : No. 4, January - March 2012 Text-independent Mono and Cross-lingual Speaker Identification with the Constraint of Limited Data Nagaraja B G and H S Jayanna Department of Information Science and Engineering Siddaganga Institute of

More information

Modeling function word errors in DNN-HMM based LVCSR systems

Modeling function word errors in DNN-HMM based LVCSR systems Modeling function word errors in DNN-HMM based LVCSR systems Melvin Jose Johnson Premkumar, Ankur Bapna and Sree Avinash Parchuri Department of Computer Science Department of Electrical Engineering Stanford

More information

On-Line Data Analytics

On-Line Data Analytics International Journal of Computer Applications in Engineering Sciences [VOL I, ISSUE III, SEPTEMBER 2011] [ISSN: 2231-4946] On-Line Data Analytics Yugandhar Vemulapalli #, Devarapalli Raghu *, Raja Jacob

More information

Likelihood-Maximizing Beamforming for Robust Hands-Free Speech Recognition

Likelihood-Maximizing Beamforming for Robust Hands-Free Speech Recognition MITSUBISHI ELECTRIC RESEARCH LABORATORIES http://www.merl.com Likelihood-Maximizing Beamforming for Robust Hands-Free Speech Recognition Seltzer, M.L.; Raj, B.; Stern, R.M. TR2004-088 December 2004 Abstract

More information

Modeling function word errors in DNN-HMM based LVCSR systems

Modeling function word errors in DNN-HMM based LVCSR systems Modeling function word errors in DNN-HMM based LVCSR systems Melvin Jose Johnson Premkumar, Ankur Bapna and Sree Avinash Parchuri Department of Computer Science Department of Electrical Engineering Stanford

More information

Curriculum Design Project with Virtual Manipulatives. Gwenanne Salkind. George Mason University EDCI 856. Dr. Patricia Moyer-Packenham

Curriculum Design Project with Virtual Manipulatives. Gwenanne Salkind. George Mason University EDCI 856. Dr. Patricia Moyer-Packenham Curriculum Design Project with Virtual Manipulatives Gwenanne Salkind George Mason University EDCI 856 Dr. Patricia Moyer-Packenham Spring 2006 Curriculum Design Project with Virtual Manipulatives Table

More information

Affective Classification of Generic Audio Clips using Regression Models

Affective Classification of Generic Audio Clips using Regression Models Affective Classification of Generic Audio Clips using Regression Models Nikolaos Malandrakis 1, Shiva Sundaram, Alexandros Potamianos 3 1 Signal Analysis and Interpretation Laboratory (SAIL), USC, Los

More information

An Empirical Analysis of the Effects of Mexican American Studies Participation on Student Achievement within Tucson Unified School District

An Empirical Analysis of the Effects of Mexican American Studies Participation on Student Achievement within Tucson Unified School District An Empirical Analysis of the Effects of Mexican American Studies Participation on Student Achievement within Tucson Unified School District Report Submitted June 20, 2012, to Willis D. Hawley, Ph.D., Special

More information

The Use of Statistical, Computational and Modelling Tools in Higher Learning Institutions: A Case Study of the University of Dodoma

The Use of Statistical, Computational and Modelling Tools in Higher Learning Institutions: A Case Study of the University of Dodoma International Journal of Computer Applications (975 8887) The Use of Statistical, Computational and Modelling Tools in Higher Learning Institutions: A Case Study of the University of Dodoma Gilbert M.

More information

Class-Discriminative Weighted Distortion Measure for VQ-Based Speaker Identification

Class-Discriminative Weighted Distortion Measure for VQ-Based Speaker Identification Class-Discriminative Weighted Distortion Measure for VQ-Based Speaker Identification Tomi Kinnunen and Ismo Kärkkäinen University of Joensuu, Department of Computer Science, P.O. Box 111, 80101 JOENSUU,

More information

Voice conversion through vector quantization

Voice conversion through vector quantization J. Acoust. Soc. Jpn.(E)11, 2 (1990) Voice conversion through vector quantization Masanobu Abe, Satoshi Nakamura, Kiyohiro Shikano, and Hisao Kuwabara A TR Interpreting Telephony Research Laboratories,

More information

also inside Continuing Education Alumni Authors College Events

also inside Continuing Education Alumni Authors College Events SUMMER 2016 JAMESTOWN COMMUNITY COLLEGE ALUMNI MAGAZINE create a etrepreeur creatig a busiess a artist creatig beauty a citize creatig the future also iside Cotiuig Educatio Alumi Authors College Evets

More information

The Good Judgment Project: A large scale test of different methods of combining expert predictions

The Good Judgment Project: A large scale test of different methods of combining expert predictions The Good Judgment Project: A large scale test of different methods of combining expert predictions Lyle Ungar, Barb Mellors, Jon Baron, Phil Tetlock, Jaime Ramos, Sam Swift The University of Pennsylvania

More information

Lecture 10: Reinforcement Learning

Lecture 10: Reinforcement Learning Lecture 1: Reinforcement Learning Cognitive Systems II - Machine Learning SS 25 Part III: Learning Programs and Strategies Q Learning, Dynamic Programming Lecture 1: Reinforcement Learning p. Motivation

More information

IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 3, MARCH

IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 3, MARCH IEEE TRANSACTIONS ON AUDIO, SPEECH, AND LANGUAGE PROCESSING, VOL. 17, NO. 3, MARCH 2009 423 Adaptive Multimodal Fusion by Uncertainty Compensation With Application to Audiovisual Speech Recognition George

More information

South Carolina English Language Arts

South Carolina English Language Arts South Carolina English Language Arts A S O F J U N E 2 0, 2 0 1 0, T H I S S TAT E H A D A D O P T E D T H E CO M M O N CO R E S TAT E S TA N DA R D S. DOCUMENTS REVIEWED South Carolina Academic Content

More information

Learning Methods for Fuzzy Systems

Learning Methods for Fuzzy Systems Learning Methods for Fuzzy Systems Rudolf Kruse and Andreas Nürnberger Department of Computer Science, University of Magdeburg Universitätsplatz, D-396 Magdeburg, Germany Phone : +49.39.67.876, Fax : +49.39.67.8

More information

A New Perspective on Combining GMM and DNN Frameworks for Speaker Adaptation

A New Perspective on Combining GMM and DNN Frameworks for Speaker Adaptation A New Perspective on Combining GMM and DNN Frameworks for Speaker Adaptation SLSP-2016 October 11-12 Natalia Tomashenko 1,2,3 natalia.tomashenko@univ-lemans.fr Yuri Khokhlov 3 khokhlov@speechpro.com Yannick

More information

Design Of An Automatic Speaker Recognition System Using MFCC, Vector Quantization And LBG Algorithm

Design Of An Automatic Speaker Recognition System Using MFCC, Vector Quantization And LBG Algorithm Design Of An Automatic Speaker Recognition System Using MFCC, Vector Quantization And LBG Algorithm Prof. Ch.Srinivasa Kumar Prof. and Head of department. Electronics and communication Nalanda Institute

More information

A study of speaker adaptation for DNN-based speech synthesis

A study of speaker adaptation for DNN-based speech synthesis A study of speaker adaptation for DNN-based speech synthesis Zhizheng Wu, Pawel Swietojanski, Christophe Veaux, Steve Renals, Simon King The Centre for Speech Technology Research (CSTR) University of Edinburgh,

More information

AGS THE GREAT REVIEW GAME FOR PRE-ALGEBRA (CD) CORRELATED TO CALIFORNIA CONTENT STANDARDS

AGS THE GREAT REVIEW GAME FOR PRE-ALGEBRA (CD) CORRELATED TO CALIFORNIA CONTENT STANDARDS AGS THE GREAT REVIEW GAME FOR PRE-ALGEBRA (CD) CORRELATED TO CALIFORNIA CONTENT STANDARDS 1 CALIFORNIA CONTENT STANDARDS: Chapter 1 ALGEBRA AND WHOLE NUMBERS Algebra and Functions 1.4 Students use algebraic

More information

Maximizing Learning Through Course Alignment and Experience with Different Types of Knowledge

Maximizing Learning Through Course Alignment and Experience with Different Types of Knowledge Innov High Educ (2009) 34:93 103 DOI 10.1007/s10755-009-9095-2 Maximizing Learning Through Course Alignment and Experience with Different Types of Knowledge Phyllis Blumberg Published online: 3 February

More information

Probabilistic Latent Semantic Analysis

Probabilistic Latent Semantic Analysis Probabilistic Latent Semantic Analysis Thomas Hofmann Presentation by Ioannis Pavlopoulos & Andreas Damianou for the course of Data Mining & Exploration 1 Outline Latent Semantic Analysis o Need o Overview

More information

Australian Journal of Basic and Applied Sciences

Australian Journal of Basic and Applied Sciences AENSI Journals Australian Journal of Basic and Applied Sciences ISSN:1991-8178 Journal home page: www.ajbasweb.com Feature Selection Technique Using Principal Component Analysis For Improving Fuzzy C-Mean

More information

Speaker recognition using universal background model on YOHO database

Speaker recognition using universal background model on YOHO database Aalborg University Master Thesis project Speaker recognition using universal background model on YOHO database Author: Alexandre Majetniak Supervisor: Zheng-Hua Tan May 31, 2011 The Faculties of Engineering,

More information

SARDNET: A Self-Organizing Feature Map for Sequences

SARDNET: A Self-Organizing Feature Map for Sequences SARDNET: A Self-Organizing Feature Map for Sequences Daniel L. James and Risto Miikkulainen Department of Computer Sciences The University of Texas at Austin Austin, TX 78712 dljames,risto~cs.utexas.edu

More information

Expressive speech synthesis: a review

Expressive speech synthesis: a review Int J Speech Technol (2013) 16:237 260 DOI 10.1007/s10772-012-9180-2 Expressive speech synthesis: a review D. Govind S.R. Mahadeva Prasanna Received: 31 May 2012 / Accepted: 11 October 2012 / Published

More information

Rule Learning With Negation: Issues Regarding Effectiveness

Rule Learning With Negation: Issues Regarding Effectiveness Rule Learning With Negation: Issues Regarding Effectiveness S. Chua, F. Coenen, G. Malcolm University of Liverpool Department of Computer Science, Ashton Building, Ashton Street, L69 3BX Liverpool, United

More information

Python Machine Learning

Python Machine Learning Python Machine Learning Unlock deeper insights into machine learning with this vital guide to cuttingedge predictive analytics Sebastian Raschka [ PUBLISHING 1 open source I community experience distilled

More information

Integrating simulation into the engineering curriculum: a case study

Integrating simulation into the engineering curriculum: a case study Integrating simulation into the engineering curriculum: a case study Baidurja Ray and Rajesh Bhaskaran Sibley School of Mechanical and Aerospace Engineering, Cornell University, Ithaca, New York, USA E-mail:

More information

AUTOMATIC DETECTION OF PROLONGED FRICATIVE PHONEMES WITH THE HIDDEN MARKOV MODELS APPROACH 1. INTRODUCTION

AUTOMATIC DETECTION OF PROLONGED FRICATIVE PHONEMES WITH THE HIDDEN MARKOV MODELS APPROACH 1. INTRODUCTION JOURNAL OF MEDICAL INFORMATICS & TECHNOLOGIES Vol. 11/2007, ISSN 1642-6037 Marek WIŚNIEWSKI *, Wiesława KUNISZYK-JÓŹKOWIAK *, Elżbieta SMOŁKA *, Waldemar SUSZYŃSKI * HMM, recognition, speech, disorders

More information

A Comparison of Two Text Representations for Sentiment Analysis

A Comparison of Two Text Representations for Sentiment Analysis 010 International Conference on Computer Application and System Modeling (ICCASM 010) A Comparison of Two Text Representations for Sentiment Analysis Jianxiong Wang School of Computer Science & Educational

More information

Learning From the Past with Experiment Databases

Learning From the Past with Experiment Databases Learning From the Past with Experiment Databases Joaquin Vanschoren 1, Bernhard Pfahringer 2, and Geoff Holmes 2 1 Computer Science Dept., K.U.Leuven, Leuven, Belgium 2 Computer Science Dept., University

More information

Seminar - Organic Computing

Seminar - Organic Computing Seminar - Organic Computing Self-Organisation of OC-Systems Markus Franke 25.01.2006 Typeset by FoilTEX Timetable 1. Overview 2. Characteristics of SO-Systems 3. Concern with Nature 4. Design-Concepts

More information

Mathematics process categories

Mathematics process categories Mathematics process categories All of the UK curricula define multiple categories of mathematical proficiency that require students to be able to use and apply mathematics, beyond simple recall of facts

More information

Unvoiced Landmark Detection for Segment-based Mandarin Continuous Speech Recognition

Unvoiced Landmark Detection for Segment-based Mandarin Continuous Speech Recognition Unvoiced Landmark Detection for Segment-based Mandarin Continuous Speech Recognition Hua Zhang, Yun Tang, Wenju Liu and Bo Xu National Laboratory of Pattern Recognition Institute of Automation, Chinese

More information

Using focal point learning to improve human machine tacit coordination

Using focal point learning to improve human machine tacit coordination DOI 10.1007/s10458-010-9126-5 Using focal point learning to improve human machine tacit coordination InonZuckerman SaritKraus Jeffrey S. Rosenschein The Author(s) 2010 Abstract We consider an automated

More information

Rote rehearsal and spacing effects in the free recall of pure and mixed lists. By: Peter P.J.L. Verkoeijen and Peter F. Delaney

Rote rehearsal and spacing effects in the free recall of pure and mixed lists. By: Peter P.J.L. Verkoeijen and Peter F. Delaney Rote rehearsal and spacing effects in the free recall of pure and mixed lists By: Peter P.J.L. Verkoeijen and Peter F. Delaney Verkoeijen, P. P. J. L, & Delaney, P. F. (2008). Rote rehearsal and spacing

More information

ScienceDirect. A Framework for Clustering Cardiac Patient s Records Using Unsupervised Learning Techniques

ScienceDirect. A Framework for Clustering Cardiac Patient s Records Using Unsupervised Learning Techniques Available online at www.sciencedirect.com ScienceDirect Procedia Computer Science 98 (2016 ) 368 373 The 6th International Conference on Current and Future Trends of Information and Communication Technologies

More information

Probability and Statistics Curriculum Pacing Guide

Probability and Statistics Curriculum Pacing Guide Unit 1 Terms PS.SPMJ.3 PS.SPMJ.5 Plan and conduct a survey to answer a statistical question. Recognize how the plan addresses sampling technique, randomization, measurement of experimental error and methods

More information

Reducing Features to Improve Bug Prediction

Reducing Features to Improve Bug Prediction Reducing Features to Improve Bug Prediction Shivkumar Shivaji, E. James Whitehead, Jr., Ram Akella University of California Santa Cruz {shiv,ejw,ram}@soe.ucsc.edu Sunghun Kim Hong Kong University of Science

More information

CS Machine Learning

CS Machine Learning CS 478 - Machine Learning Projects Data Representation Basic testing and evaluation schemes CS 478 Data and Testing 1 Programming Issues l Program in any platform you want l Realize that you will be doing

More information

Why Pay Attention to Race?

Why Pay Attention to Race? Why Pay Attention to Race? Witnessing Whiteness Chapter 1 Workshop 1.1 1.1-1 Dear Facilitator(s), This workshop series was carefully crafted, reviewed (by a multiracial team), and revised with several

More information

Assignment 1: Predicting Amazon Review Ratings

Assignment 1: Predicting Amazon Review Ratings Assignment 1: Predicting Amazon Review Ratings 1 Dataset Analysis Richard Park r2park@acsmail.ucsd.edu February 23, 2015 The dataset selected for this assignment comes from the set of Amazon reviews for

More information

DOMAIN MISMATCH COMPENSATION FOR SPEAKER RECOGNITION USING A LIBRARY OF WHITENERS. Elliot Singer and Douglas Reynolds

DOMAIN MISMATCH COMPENSATION FOR SPEAKER RECOGNITION USING A LIBRARY OF WHITENERS. Elliot Singer and Douglas Reynolds DOMAIN MISMATCH COMPENSATION FOR SPEAKER RECOGNITION USING A LIBRARY OF WHITENERS Elliot Singer and Douglas Reynolds Massachusetts Institute of Technology Lincoln Laboratory {es,dar}@ll.mit.edu ABSTRACT

More information

The Evolution of Random Phenomena

The Evolution of Random Phenomena The Evolution of Random Phenomena A Look at Markov Chains Glen Wang glenw@uchicago.edu Splash! Chicago: Winter Cascade 2012 Lecture 1: What is Randomness? What is randomness? Can you think of some examples

More information

American Journal of Business Education October 2009 Volume 2, Number 7

American Journal of Business Education October 2009 Volume 2, Number 7 Factors Affecting Students Grades In Principles Of Economics Orhan Kara, West Chester University, USA Fathollah Bagheri, University of North Dakota, USA Thomas Tolin, West Chester University, USA ABSTRACT

More information

PROJECT MANAGEMENT AND COMMUNICATION SKILLS DEVELOPMENT STUDENTS PERCEPTION ON THEIR LEARNING

PROJECT MANAGEMENT AND COMMUNICATION SKILLS DEVELOPMENT STUDENTS PERCEPTION ON THEIR LEARNING PROJECT MANAGEMENT AND COMMUNICATION SKILLS DEVELOPMENT STUDENTS PERCEPTION ON THEIR LEARNING Mirka Kans Department of Mechanical Engineering, Linnaeus University, Sweden ABSTRACT In this paper we investigate

More information

Success Factors for Creativity Workshops in RE

Success Factors for Creativity Workshops in RE Success Factors for Creativity s in RE Sebastian Adam, Marcus Trapp Fraunhofer IESE Fraunhofer-Platz 1, 67663 Kaiserslautern, Germany {sebastian.adam, marcus.trapp}@iese.fraunhofer.de Abstract. In today

More information

Knowledge Transfer in Deep Convolutional Neural Nets

Knowledge Transfer in Deep Convolutional Neural Nets Knowledge Transfer in Deep Convolutional Neural Nets Steven Gutstein, Olac Fuentes and Eric Freudenthal Computer Science Department University of Texas at El Paso El Paso, Texas, 79968, U.S.A. Abstract

More information

Automatic intonation assessment for computer aided language learning

Automatic intonation assessment for computer aided language learning Available online at www.sciencedirect.com Speech Communication 52 (2010) 254 267 www.elsevier.com/locate/specom Automatic intonation assessment for computer aided language learning Juan Pablo Arias a,

More information

12- A whirlwind tour of statistics

12- A whirlwind tour of statistics CyLab HT 05-436 / 05-836 / 08-534 / 08-734 / 19-534 / 19-734 Usable Privacy and Security TP :// C DU February 22, 2016 y & Secu rivac rity P le ratory bo La Lujo Bauer, Nicolas Christin, and Abby Marsh

More information

Phonetic- and Speaker-Discriminant Features for Speaker Recognition. Research Project

Phonetic- and Speaker-Discriminant Features for Speaker Recognition. Research Project Phonetic- and Speaker-Discriminant Features for Speaker Recognition by Lara Stoll Research Project Submitted to the Department of Electrical Engineering and Computer Sciences, University of California

More information

Operational Knowledge Management: a way to manage competence

Operational Knowledge Management: a way to manage competence Operational Knowledge Management: a way to manage competence Giulio Valente Dipartimento di Informatica Universita di Torino Torino (ITALY) e-mail: valenteg@di.unito.it Alessandro Rigallo Telecom Italia

More information

University of Groningen. Systemen, planning, netwerken Bosman, Aart

University of Groningen. Systemen, planning, netwerken Bosman, Aart University of Groningen Systemen, planning, netwerken Bosman, Aart IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from it. Please check the document

More information

QuickStroke: An Incremental On-line Chinese Handwriting Recognition System

QuickStroke: An Incremental On-line Chinese Handwriting Recognition System QuickStroke: An Incremental On-line Chinese Handwriting Recognition System Nada P. Matić John C. Platt Λ Tony Wang y Synaptics, Inc. 2381 Bering Drive San Jose, CA 95131, USA Abstract This paper presents

More information

Major Milestones, Team Activities, and Individual Deliverables

Major Milestones, Team Activities, and Individual Deliverables Major Milestones, Team Activities, and Individual Deliverables Milestone #1: Team Semester Proposal Your team should write a proposal that describes project objectives, existing relevant technology, engineering

More information

CHAPTER 4: REIMBURSEMENT STRATEGIES 24

CHAPTER 4: REIMBURSEMENT STRATEGIES 24 CHAPTER 4: REIMBURSEMENT STRATEGIES 24 INTRODUCTION Once state level policymakers have decided to implement and pay for CSR, one issue they face is simply how to calculate the reimbursements to districts

More information

The Effect of Extensive Reading on Developing the Grammatical. Accuracy of the EFL Freshmen at Al Al-Bayt University

The Effect of Extensive Reading on Developing the Grammatical. Accuracy of the EFL Freshmen at Al Al-Bayt University The Effect of Extensive Reading on Developing the Grammatical Accuracy of the EFL Freshmen at Al Al-Bayt University Kifah Rakan Alqadi Al Al-Bayt University Faculty of Arts Department of English Language

More information

THE PENNSYLVANIA STATE UNIVERSITY SCHREYER HONORS COLLEGE DEPARTMENT OF MATHEMATICS ASSESSING THE EFFECTIVENESS OF MULTIPLE CHOICE MATH TESTS

THE PENNSYLVANIA STATE UNIVERSITY SCHREYER HONORS COLLEGE DEPARTMENT OF MATHEMATICS ASSESSING THE EFFECTIVENESS OF MULTIPLE CHOICE MATH TESTS THE PENNSYLVANIA STATE UNIVERSITY SCHREYER HONORS COLLEGE DEPARTMENT OF MATHEMATICS ASSESSING THE EFFECTIVENESS OF MULTIPLE CHOICE MATH TESTS ELIZABETH ANNE SOMERS Spring 2011 A thesis submitted in partial

More information

A Web Based Annotation Interface Based of Wheel of Emotions. Author: Philip Marsh. Project Supervisor: Irena Spasic. Project Moderator: Matthew Morgan

A Web Based Annotation Interface Based of Wheel of Emotions. Author: Philip Marsh. Project Supervisor: Irena Spasic. Project Moderator: Matthew Morgan A Web Based Annotation Interface Based of Wheel of Emotions Author: Philip Marsh Project Supervisor: Irena Spasic Project Moderator: Matthew Morgan Module Number: CM3203 Module Title: One Semester Individual

More information

Learning Structural Correspondences Across Different Linguistic Domains with Synchronous Neural Language Models

Learning Structural Correspondences Across Different Linguistic Domains with Synchronous Neural Language Models Learning Structural Correspondences Across Different Linguistic Domains with Synchronous Neural Language Models Stephan Gouws and GJ van Rooyen MIH Medialab, Stellenbosch University SOUTH AFRICA {stephan,gvrooyen}@ml.sun.ac.za

More information

VIEW: An Assessment of Problem Solving Style

VIEW: An Assessment of Problem Solving Style 1 VIEW: An Assessment of Problem Solving Style Edwin C. Selby, Donald J. Treffinger, Scott G. Isaksen, and Kenneth Lauer This document is a working paper, the purposes of which are to describe the three

More information

OPAC and User Perception in Law University Libraries in the Karnataka: A Study

OPAC and User Perception in Law University Libraries in the Karnataka: A Study ISSN 2229-5984 (P) 29-5576 (e) OPAC and User Perception in Law University Libraries in the Karnataka: A Study Devendra* and Khaiser Nikam** To Cite: Devendra & Nikam, K. (20). OPAC and user perception

More information

Transfer Learning Action Models by Measuring the Similarity of Different Domains

Transfer Learning Action Models by Measuring the Similarity of Different Domains Transfer Learning Action Models by Measuring the Similarity of Different Domains Hankui Zhuo 1, Qiang Yang 2, and Lei Li 1 1 Software Research Institute, Sun Yat-sen University, Guangzhou, China. zhuohank@gmail.com,lnslilei@mail.sysu.edu.cn

More information

Proceedings of Meetings on Acoustics

Proceedings of Meetings on Acoustics Proceedings of Meetings on Acoustics Volume 19, 2013 http://acousticalsociety.org/ ICA 2013 Montreal Montreal, Canada 2-7 June 2013 Speech Communication Session 2aSC: Linking Perception and Production

More information

DYNAMIC ADAPTIVE HYPERMEDIA SYSTEMS FOR E-LEARNING

DYNAMIC ADAPTIVE HYPERMEDIA SYSTEMS FOR E-LEARNING University of Craiova, Romania Université de Technologie de Compiègne, France Ph.D. Thesis - Abstract - DYNAMIC ADAPTIVE HYPERMEDIA SYSTEMS FOR E-LEARNING Elvira POPESCU Advisors: Prof. Vladimir RĂSVAN

More information