AN ENHANCED MULTIMODAL SOUND LOCALIZATION WITH HUMANLIKE AUDITORY SYSTEM FOR INTELLIGENT SERVICE ROBOTS

Transcription

1 Internatonal Journal of Latest Research n Scence and Technology Volume 2, Issue 6: Page No.26-31,November-December ISSN (Onlne): AN ENHANCED MULTIMODAL SOUND LOCALIZATION WITH HUMANLIKE AUDITORY SYSTEM FOR INTELLIGENT SERVICE ROBOTS 1 Keun-Chang Kwa 1 Department of Control, Instrumentaton, and Robotc Engneerng, Chosun Unversty, Korea Abstract- Ths paper s concerned wth an enhanced multmodal sound localzaton wth humanle audtory system for a networ-based ntellgent servce robot, whch explots strong nformaton technology nfrastructure. The objectve of ths paper s to ntegrate several audovsual-based Human-Robot Interacton (HRI) components that can naturally nteract between human and robot through audovsual nformaton obtaned from robot camera and mcrophones n the nosy envronments or the presence of multple persons. The proposed approach s comprsed of two man stages. The frst stage performs speech recognton, sound localzaton, and speaer recognton to now whether the user calls the robot or not as well as the drecton and dentfcaton of the caller respectvely, when someone calls robot s name. In the second stage, an ntellgent robot moves forward to the specfc caller based on multple face detecton/recognton wth the ad of the nformaton dentfed by speaer recognton among multple persons. The robot platform used n ths study s WEVER, whch s a networ-based ntellgent servce robot developed n Electroncs and Telecommuncaton Research Insttute. Ths robot refers to an Ubqutous Robotc Companon (URC) that provdes necessary servces anytme and anywhere. The effectveness of the proposed approach s compared wth other multmodal methods and sound localzaton tself. Keywords Multmodal Sound Localzaton, Human-Robot Interacton, Ubqutous Robotc Companon, Speech/Speaer Recognton I. INTRODUCTION Nowadays, we have been wtnessng a rapd growth n the number and varety of applcatons of robots, rangng from conventonal ndustral robots to ntellgent servce robots. Conventonal ndustral robots perform smple tass by followng pre-programmed nstructons for humans n factores. These robots have been wdely used n many manufacturng ndustres. On the other hand, ntellgent servce robots consst of toy robots, cleanng robots, humanod robots and nformaton servce robots. In the case of the cleanng robots, the Amercan company Robot recently sold more than one mllon vacuum cleaner robots called Roomba. In Korea, the cleanng robot has drawn the attenton of the people, and further support and development are concentrated on the nformaton servce robot. Especally a networ-based ntellgent servce robot s regarded as a ubqutous robotc companon that provdes necessary servces anytme and anywhere for nformaton servce [1]. Here ths concept explots strong Informaton technology nfrastructure such as the hgh-speed nternet. In order to develop IT-based servce robot project based on ths concept as one of the next-generaton growth engne ndustres, t s necessary to develop core technologes such as Human-Robot Interacton (HRI), navgaton, operaton control, and so on. Among varous core technologes, especally HRI components play an mportant role that can naturally nteract between human and robot based on audovsual nformaton obtaned from robot camera, mcrophones, and varous sensors for networ-based ntellgent servce robots. In ths paper, we elaborate on an enhanced multmodal sound localzaton wth humanle audtory system n the nosy envronments and the presence of multple persons. Over the past few years, several studes have been completed on sound localzaton. Cho [2] has developed an audo-vsual ntegraton based on a probablstc sound localzaton system by Tme Delay of Arrval (TDOA) and face tracng by Open Computer Vson (OpenCV). Especally, when several face mages are detected wthn robot s Feld of Vew (FOV), the dsadvantage of these approach s that the robot moves forward to person wth large face mage among rectangle bounds found by face detecton. Hara [3] has proposed a robust speech nterface based on audo and vdeo nformaton fuson by Bayesan networ for humanod HRP-2. Here sound localzaton was performed by the Multple Sgnal Classfcaton (MUSIC) method extended to the broadband sgnal wth egenvalue weghtng. Human tracng ncludes face detecton by sn-color model and face tracng by ernel-based face model, respectvely. Huang [4] has presented a model-based sound localzaton system based on a model of the precedence effect of the human audtory system to cope wth echoes and reverberatons. The features of ths system are hgh tme resoluton, less computaton and correct localzaton n comparson to beamformng [5], MUSIC [6], and Maxmum Lelhood (ML) method [7]. Valn [8] has proposed a robust sound localzaton method n three-dmensonal space usng an array of eght mcrophones. Ths method s based on a frequency-doman mplementaton of a steered beamformer along wth a probablstc postprocessor. Naada [9] has used two pars of mcrophone n a ISSN:

2 humanod robot called SIG. One par s nstalled on both sdes of the head, whle the other par s placed nsde the head to record nternal sounds such as motor nose for nose cancellaton. Even though these methods have been revealed a good localzaton performance, the development of sound localzaton wth humanle audtory system remans as a matter to be researched further under robot envronments. These systems may encounter problems when the sound localzaton s performed wthout the ad of speech/speaer recognton as well as multple face recognton under the nosy envronments or the presence of multple persons. For ths purpose, we elaborate on the three man aspects for an enhanced multmodal sound localzaton. Frst we use a multmodal approach based on the ntegraton of speech recognton, sound localzaton, and speaer recognton to now whether the user calls a robot or not, as well as the poston and dentfcaton of the caller smultaneously, when someone calls robot s name. Here sound localzaton s based on Generalzed Cross Correlaton-Phase Transform (GCC- PHAT) method. The speech and speaer recognton are performed by Hdden Marov Model (HMM) and Mel- Frequency Cepstral Coeffcent-Gaussan Mxture Model (MFCC-GMM) for the networ-based ntellgent servce robots, respectvely. Second, the robot moves forward to the specfc caller based on multple face recognton wth the ad of the pror nformaton by speaer recognton among multple persons after turnng around. Here the face detecton s performed by Revsed and Modfed Census Transform (RMCT), Adaboost, and face certanty map n the step of preprocessng, detecton, and postprocessng, respectvely. Furthermore, multple face recognton has been developed by Multple Prncpal Component Analyss (MPCA) and Support Vector Machne (SVM) method. Fnally these two aspects should be performed on the concept of URC that provde varous advanced functons and servces by addng the networ to the exstng robot, enhancng moblty and user nterface. Thus the sound localzaton s performed n robot body (clent) equpped wth three mcrophones and sound board. On the other hand, the speech/speaer recognton and face detecton/recognton are performed on audovsual nformaton transmtted through wreless networ. Over the past few years, we have been developed and mplemented these HRI components for networ-based ntellgent servce robot [10][11][12][13][14][17]. The robot platform used n ths study s WEVER, whch s a URCbased ntellgent servce robot developed at Intellgent Robot Research Dvson n Electroncs and Telecommuncaton Research Insttute. The expermental results obtaned for sound localzaton database reveal that the proposed approach presented n ths paper yeld a good localzaton n comparson to the results obtaned by other multmodal approaches and sound localzaton tself. Ths paper s organzed n the followng manner. Secton 2 descrbes the developed audo- and vdeo-based HRI components for an enhanced multmodal sound localzaton wth humanle audtory system. Secton 3 presents two man stages of multmodal sound localzaton under networ-based envronments. Secton 4 covers the expermental results concernng these sound localzaton and the related HRI components. Fnally concludng comments are covered n Secton 5. Internatonal Journal of Latest Research n Scence and Technology. II. AUDIOVISUAL HUMAN-ROBOT INTERACTION COMPONENTS In ths secton, we descrbe the audovsual HRI components used n conjuncton wth multmodal sound localzaton. As mentoned before, the proposed approach s realzed by sound localzaton, speech/speaer recognton, and multple face detecton/recognton to perform humanle audtory and vsual system. In confrontng real-world envronments, t s frequently advantageous to use several audovsual technques synergstcally rather than exclusvely, resultng n constructon of multmodal sound localzaton system. 2.1 Sound localzaton As shown n Fg. 1, we use low-prced three mcrophones and sound board equpped wth the robot for sound localzaton. Here sound board wth eght channels called multmodal nterface module has already developed n ETRI. A drectonal mcrophone for speech/speaer recognton Three omndrectonal mcrophones for sound localzaton (a) Arrangement of mcrophones (b) Wever robot Fg.1 Arrangement of mcrophones and robot platform We use GCC method n the frequency doman to obtan tme delay. The sound localzaton based on GCC-PHAT n the frequency doman has the mert that ths method can estmate relatvely correct localzaton angle n the nosy and reverberant envronments. Gven that m 1 (n) and m 2 (n) are sound sources obtaned from the frst and second mcrophone, respectvely, the tme delay between m 1 (n) and m 2 (n) n the frequency doman s computed as follows 1 ( n ) 2 * m m WPHAT ( w )M ( w )M ISSN: R ( w )e where M 1 ( w ) and M 2 ( w ) are Fourer transform of m 1 (n) and m 2 (n), respectvely. W PHAT ( w ) s the nverse of * M 1( w )M 2( w ) as the frequency weghtng functon called PHAT. However, t s usually dffcult to obtan the detaled pror nowledge of nose spectra n the reverberant envronments. Thus the alternatve s calculated as followng PHAT weghtng functon W PHAT jwn dw 1 ( w ) (2) * M ( w )M ( w ) 1 2 The resultant tme delay based on GCC-PHAT s expressed as follows (1)

3 Internatonal Journal of Latest Research n Scence and Technology. D arg max R ( n ) (3) m 1 m 2 After computng tme delay, we decde the secton between mcrophones based on tme delay. Suppose that the sound wave at the mcrophone s a plan wave and the angle between mcrophones s 120, respectvely. Fnally azmuth s estmated as follows [10] 1 t v cos 30 (4) L where s the angle of sound source obtaned n the above equaton. t s tme delay between two mcrophones and v s the velocty of sound source. Moreover, L s the dstance between two mcrophones. The sound localzaton s performed through the robot s name. Ths system s effectvely combned wth speech and speaer recognton to construct humanle audtory system. 2.2 Speech/Speaer Recognton In what follows, we present the technques developed n conjuncton wth the speech and speaer recognton. On the bass of these methods, the robot can smultaneously recognze whether the user calls a robot or not as well as dentfcaton of the caller. We perform speech and speaer recognton wth drectonal mcrophone from speech sgnal transmtted through wreless networ, as shown n Fg.1. In the case of speech recognton, Korean-based spontaneous recognton s performed by HMM method. As a preprocessng process, we employ endpont detecton algorthm based on log energy and zero crossng rates as well as speech enhancement method based on Wner flter. For further detals on speech recognton, see [11]. On the other hand, speaer recognton s comprsed of four man stages ncludng on-lne speaer regstraton, feature extracton, generaton of speaer model, and text-ndependent speaer dentfcaton. After detectng speech sgnal by endpont detecton algorthm, we obtan the feature vectors of MFCC. For smplcty, we use 11 order MFCC parameters except for the frst order. Ths method s comprsed of sx stages to get MFCC: pre-emphass, frame blocng, hammng wdow to lessen dstorton, Fast Fourer Transform (FFT), trangular bandpass flter, and cosne transform. In what follows, we brefly descrbe the well-nown GMM frequently used n conjuncton wth text-ndependent speaer dentfcaton. Here the dstrbuton of feature vectors extracted from ndvdual speeches s obtaned by a Gaussan mxture densty. For a feature vector denoted as x, the mxture densty for speaer s defned as K P( x s ) w b ( x ) (5) 1 where w s mxture weghts and b s th Gaussan mxture. The densty s a weghted lnear combnaton of K Gaussan mxture parameterzed by a mean vector and covarance matrx. The Gaussan mxture s defned as the followng equaton b 1 1 T 1 ( x ) exp ( x ) ( ) ( x ) / 1 2 D / 2 ( 2 ) 2 K The mxture weghts w satsfy the constrant 1w 1. Thus the parameters of speaer s model are denoted as s w,,, 1,,K. For smplcty, dagonal covarance matrces are used to construct GMM, because dagonal matrx are more computatonally effcent than full covarance matrx for tranng [15][16]. Gven tranng speeches from a speaer, the speaer model s traned by estmatng the parameters of the GMM. The well-nown method s ML estmaton. For a sequence of T tranng vectors X x1,x2,,x T, the GMM lelhood can be expressed as p( X T s ) p( xt s ) t 1 The maxmum lelhood parameter estmaton s obtaned by usng the Expectaton-Maxmzaton (EM) algorthm. For speaer dentfcaton, a group of speaers S { 1, 2,,S } s represented by GMM s parameters. The man goal s to fnd the speaer model whch has the maxmum a posteror probablty as the followng equaton Ŝ arg max p( X ) arg max 1 S 1 S p( X )p( ) p( X ) where the second equaton s based on Bayes rule. Assumng equally lely speaers, p(x) s the same for all speaer models, the dentfcaton s smplfed as follows Ŝ arg max 1 S p( X ) Usng logarthms and the ndependence between observatons, the speaer dentfcaton s computed as T Ŝ arg max log p( x ) 1 S t1 t (6) (7) (8) (9) (10) On the other hand, on-lne speaer regstraton s performed by Unversal Bacground Model (UBM). Here each speaer s regstered by a few sentences based on the adaptaton of UBM n real-tme [12]. 2.3 Face Detecton/Recognton After performng sound localzaton, we use multple face detecton/recognton to compensate localzaton error and realze humanle vsual system. In general, face detecton/recognton s used n multmodal sound localzaton. However, t s dffcult to fnd the specfc caller among multple persons, because the robot does not now ISSN:

4 who the caller s. To solve ths problem, we employ the pror nformaton obtaned by speaer recognton. Furthermore, multple face recognton s used to dentfy multple persons shown n the robot s FOV. Here face detecton s comprsed of three steps ncludng preprocessng, detecton, and postprocessng. In the frst stage, we revse the modfed census transform to compensate the senstvty to the change of pxel values. The second stage performs Adaboost that constructs the wea classfer whch classfes the face and nonface patterns and the strong classfer whch s the lnear combnaton of wea classfers. The last stage performs face certanty map based on facal nformaton such as facal sze, locaton, rotaton, and confdence value to reduce False Acceptance Rate (FAR) wth constant detecton performance. On the other hand, we use a MPCA and SVM for face recognton[13][14]. The most representatve recognton technque frequently used n conjuncton wth face recognton s PCA. The PCA approach, also nown as egenface method, s a popular unsupervsed statstcal technque that supports fndng useful mage representatons. It also exhbts optmalty when t comes to dmensonalty reducton. The use of the MPCA n ths settng s motvated by ts nsenstvty to varaton n comparson to PCA tself. Ths method conssts of preprocessng, feature extracton, and dentfcaton. In the preprocessng, we select the face regon such as egenface, egenupper, and egentzone for multple PCA. Furthermore, a geometrc and photometrc normalzaton s used to adjust the locaton of facal features and mprove the qualty of the face mage, respectvely. In the feature extracton, the weght and edge dstrbuton vectors are obtaned. III. AN ENHANCED MULTIMODAL SOUND LOCALICATION In ths secton, we present the ntegraton of audovsual HRI components for an enhanced multmodal sound localzaton. The proposed approach s comprsed of two man stages. The frst stage uses the ntegraton of audobased HRI components ncludng sound localzaton and speech/speaer recognton to possess humanle audtory system under networ envronments. The second stage of the approach concerns the ntegraton of vdeo-based HRI components ncludng multple face detecton/recognton to have humanle vsual system and compensate the localzaton error. In what follows, we descrbe the procedure to ntegrate audo- and vdeo-based HRI components for the enhanced multmodal sound localzaton as follows [Step 1] To develop the enhanced multmodal sound localzaton, frstly the speaer and face regstraton are performed through wreless networ n real-tme. The speaer model s generated by a smple few sentences due to the adaptaton of UBM. The face mages are regstered by contnuous fve frames obtaned from robot camera. [Step 2] When the user calls the robot s name, the speech sgnal s transmtted to server through a drectonal mcrophone used for speech/speaer recognton. The use of the robot s name can effectvely nteract between Internatonal Journal of Latest Research n Scence and Technology. human and robot wth respect to human audtory system. [Step 3] The speech and speaer recognton are smultaneously performed through the transmtted sgnal n the server. Robot can smultaneously recognze whether the user calls the robot or not and who the caller s, n contrast to the conventonal ntellgent robot performed by only speech recognton. Furthermore, ths approach can appled to user-customzed servce such as daly lfe schedule and favorte TV channel selecton. On the other hand, the sound localzaton s performed wth sgnals obtaned from three mcrophones n the robot. The server nforms the robot of the nformaton concernng the start and end pont of the transmtted sgnal. [Step 4] The secton decson and azmuth are obtaned from sound localzaton based on GCC-PHAT. If the name of robot s recognzed by speech recognton, the server brngs the result of the localzaton angle computed n the robot body. Thus the robot can now the nformaton concernng the dentfcaton and drecton of user. [Step 5] The server adjusts the rotaton angle by the axs of robot wheels, f the crcle center by three mcrophone s dfferent from that of robot wheels. The robot turns around by the estmated angle. [Step 6] The robot transmts the mage shown n the robot s FOV to the server after turnng. After that, face detecton s performed as mentoned before. If the face mages are not detected n the robot s FOV, the robot requests that the user calls the robot s name agan. [Step 7] After detectng the face mages, multple face recognton s performed. The robot moves forward to the specfc caller based on face recognton wth the ad of the pror nformaton dentfed by speaer recognton among multple persons. The proposed approach outperforms the conventonal multmodal methods and sound localzaton tself, n whch the ntellgent robot can recognze whether a member of famly calls myself or not as well as the drecton and dentfcaton of the specfc member IV. EXPERIMENTAL RESULTS In the seres of experments, we report on the development and performances of multmodal sound localzaton and the related HRI components. Frstly we evaluate the performance of the presented speaer recognton system for speaer database constructed by Intellgent Robot Research Dvson n ETRI. The database s constructed by audo recordng of 20 speaers (4 females and 16 males). The data set conssts of 30 sentences for each speaer and channel. We dvde the speech data nto tranng (20 sentences 20 people 3 dstances (1, 2, 3m)) and test data sets (10 sentences 20 people 5 dstances (1,2,3,4,5m)). The speaer database set conssts of 2200 sentences. The recordng was done n an offce envronment. The audo s stored as a mono, 16bt, 16Hz, and WAV fle. Based on ths database, text-ndependent speaer dentfcaton under robot envronments was performed. The dentfcaton performance obtaned about 95% (1~3m), 87% (4m), 85%(5m) ISSN:

5 recognton rates from the varaton of several mxture sze and dstances. Furthermore, the results obtaned from the varaton of GMM sze showed a smlar performance n ths experment. If the number of famly s lmted to 10 persons for ntellgent home servce robot, the dentfcaton performance wll be further mproved. On the other hand, the sze of an mage 320 x 240 pxel array whose gray levels ranged between 0 and 255. Among varous face databases, we specfcally consdered ETRI face database constructed by Intellgent Robot Research Dvson for face recognton. Ths database conssts of 2100 samples from 105 ndvduals. There are 20 mages per subject. Here we dvded nto 5 mages for enrollment and the remanng 15 mages for test set. Fg. 2 vsualzes the results of recognton performance. As shown n Fg. 2, the expermental results revealed that the proposed method yeld a better classfcaton performance n comparson to the results produced by the egenface (L 2 and cosne smlarty) and edge dstrbuton. Recognton rate egenface(l2) egenface(cos) edge dstrbuton proposed method Fg. 2 Comparson of recognton rates In what follows, sound localzaton was performed by GCC-PHAT method as mentoned n Secton 2. The dstance and angle between mcrophones are 0.32 meter and 120 degree, respectvely. The database for sound localzaton contans 1800 sound sources recorded at every 30 degree and each dstance (2, 3, 5m). The sound source used n ths paper s the name of robot called WEVER. These speeches were recorded n the offce envronments. The error of the maxmum locaton s degree at 0 locaton pont. Furthermore, the error of the average locaton s 4.85 degree at 2 meter. However, the locaton performance became worse due to the attenuaton of sound as the dstance ncreases. The error of average locaton at 5 meter s about 15 degree. Because these locaton errors exst wthn the robot s FOV ( 25 ), the localzaton error can be compensated n the multmodal sound localzaton wth the ad of face detecton and recognton. Here the result of face detecton showed a good detecton rates n comparson to the wellnown methods. The number of false detecton for the proposed method and Vola-Jones s 3 and 78 on CMU+MIT frontal face test set, respectvely. Therefore, f the face mages exst wthn the robot s FOV after turnng around, the proposed approach could yeld a good performance close to 0 degree. Furthermore, the multmodal sound localzaton proposed n ths paper can be effectvely used for networbased ntellgent servce robots, although the expermental results obtaned by only sound localzaton showed a worse Internatonal Journal of Latest Research n Scence and Technology. localzaton performance n the nosy and reverberant envronments. V. CONCLUSIONS We have dscussed the enhanced multmodal sound localzaton wth the ad of speech/speaer recognton, sound localzaton, and multple face detecton/recognton. It has been expermentally demonstrated that the proposed approach leads to a better localzaton and humanle audtory system n comparson to the prevous multmodal methods and sound localzaton tself. The man characterstcs of ths paper can be summarzed as follows: 1) The enhanced multmodal sound localzaton was developed as the one of HRI components that s realzed by URC envronments for networ-based ntellgent servce robots. 2) Based on speech/speaer recognton, the ntellgent robot can recognze whether the user calls myself or not as well as the dentfcaton of the caller smultaneously, when he/she calls robot s name. 3) The robot can move forward to the caller based on multple face recognton wth the ad of the nformaton dentfed by speaer recognton among multple persons. The ntegraton of these complementary approaches, together wth certan audovsual HRI components, results n the enhanced multmodal sound localzaton wth humanle audtory and vsual system. REFERENCES 1. Y. G. Ha, J. C. Sohn, Y. J. Cho, and H. Yoon, Towards ubqutous robotc companon: desgn and mplementaton of ubqutous robotc servce framewor, ETRI Journal, vol. 27, no. 6, pp , J. S. Cho, M. Km, and H. D. Km, Probablstc speaer localzaton n nosy envronments by audo-vsual ntegraton, Proceedngs of the 2006 IEEE/RSJ Internatonal Conference on Intellgent Robots and Systems, pp , Bejng, Oct., I. Hara, F. Asano, Y. Kawa, F. Kanehro, and K. Yamamoto, Robust speech nterface based on audo and vdeo nformaton fuson for humanod HRP-2, Proceedngs of the 2004 IEEE/RSJ Internatonal Conference on Intellgent Robots and Systems, pp , Senda, Sep., J. Huang, T. Supaongprapa, I. Teraura, F. Wang, N. Ohnsh, and N. Suge, A model-based sound localzaton system and ts applcaton to robot navgaton, Robotcs and Autonomous Systems, vol. 27, pp , D. H. Johnson and D. E. Dudgeon, Array sgnal processng: concepts and technques, Prentce-Hall, Englewood Clffs, R. O. Schmdt, Multple emtter locaton and sgnal parameter estmaton, IEEE Trans. on Antennas and Propagaton, vol.34, no.3, pp , P. Stoca and K.C. Sharman, Maxmum lelhood method for drecton-of-arrval estmaton, IEEE Trans. on Acoustcs and Speech Sgnal Processng, vol. 38, no.7, pp , J. M. Valn, F. Mchaud, B. Hadjou, and J. Rouat, Localzaton of smultaneous movng sound sources for moble robot usng a frequency-doman steered beamformer approach, Proceedng of the 2004 IEEE Internatonal Conference on Robotcs and Automaton, pp , K. Naad, H. G. Ouno, and H. Ktano, Real-tme sound source localzaton and separaton for robot audton, In Processng IEEE Internatonal Conference on Spoen Language Processng, pp , J. Cho, J. Lee, S. Jeong, K. C. Kwa, S. Y. Ch, M. Hahn, Multmodal sound source localzaton for ntellgent servce robot, The 3 rd Internatonal Conference on Ubqutous Robots and Ambent Intellgence (URAI06), H. S. Lee, Spontaneous dalogue recognton wth large out-ofvocabulares, IEEE Asa Pacfc Conference on crcuts and systems, pp , ISSN:

6 12. S. Km, M. J, H. Km, K. C. Kwa, and S. Y. Ch, Text-ndependent speaer recognton for ubqutous robot companon, The 3 rd Internatonal Conference on Ubqutous Robots and Ambent Intellgence (URAI06), D. H. Km, H. S. Yoon, S. Y. Ch, Y. J. Cho, Face dentfcaton for human robot nteracton: ntellgent securty system for mult-user worng envronment on PC, The 15 th IEEE Internatonal Symposum on Robot and Human Interactve Communcaton (ROMAN06), pp , D. H. Km, J. Y. Lee, H. S. Yoon, H. J. Km, Y. J. Cho, and E. Y. Cha, A vson-based user authentcaton system n robot envronments by usng sem-bometrcs and tracng, IEEE/RSJ Internatonal Conference on Intellgent Robots and Systems (IROS 2005), pp , Internatonal Journal of Latest Research n Scence and Technology. 15. D. A. Reynolds and R. C. Rose, Robust text-ndependent speaer dentfcaton usng Gaussan mxture speaer models, IEEE Trans. on Speech and Audo Processng, vol. 3, no. 1, pp D. A. Reynolds, T. F. Quater, and R. B. Dunn, Speaer verfcaton usng adapted Gaussan mxture models, Dgtal Sgnal Processng, vol. 10, pp , K. C. Kwa, K. D. Ban, K. S Bae, H. J. Km, S. Y. Ch, and Y. J. Cho, Speech-based Human-Robot Interacton Components for URC ntellgent servce robots, IEEE/RSJ Internatonal Conference on Intellgent Robots and Systems (IROS 2006), Vdeo sesson, ISSN: