Recognition and Rejection Performance in Wordspotting Systems Using Support Vector Machines

Transcription

1 Recognition an Rejection Perforance in Worspotting Systes Using Support Vector Machines Yassine Ben Aye, Doinique Fohr, Jean Paul Haton, Gérar Chollet ENST, CNRS-LTCI, 46 rue Barrault, F75634 Paris ceex 13 LORIA/INRIA-Lorraine, BP239, F54506, Vanœuvre France Abstract: - Support Vector Machines (SVM) is one such achine learning technique that learns the ecision surface through a process of iscriination an has a goo generalization capacity [6]. SVMs have been proven to be successful classifiers on several classical pattern recogntion probles [9, 11]. In this paper, one of the first applications of Support Vector Machines (SVM) technique for the proble of keywor spotting is presente. It classifies the correct an the incorrect keywors by using linear an Raial Basis Function kernels. This is a first work propose to use SVM in keywor spotting in orer to iprove recognition an rejection accuracy. The obtaine results are very proising. The Equal Error Rate (EER) for the linear kernel is about 16,34% copare to 15,23% obtaine by the raial basis function kernel. Key-Wors: - speech recognition, keywor spotting, hien Markov oel, support vector achines, raial basis function kernel, linear kernel 1 Introuction Significant progress has been ae in the evelopent of Autoatic Speech Recognition (ASR) technology for continuous speech. However, for wiesprea consuer applications, hanling spontaneous speech, as oppose to strictly prescribe coan wors an phrases, reains a challenge in the eployent of ASR technology. In particular, the characteristics of spontaneous speech heavily contribute to the acoustic isatch between speech ata use to train a syste an the speech input to a syste uring its operation. Spontaneous speech tens to contain out-of-vocabulary wors an isfluencies such as fille pauses an false starts. A recognizer ust thus be able, in first tie to spot a keywor ebee in speech, in secon tie to reject speech that oes not inclue any vali keywor. However, wor spotting an wor rejection are interrelate such that a goo wor-spotting capability necessarily iplies a goo rejection perforance. Several rejection approaches have been suggeste in previous research efforts, for rejection of putative hits in keywor spotting, for etection of out-of-vocabulary, an for utterance rejection [15, 16]. For exaple, we can fin the filler (or garbage) oels which are generally eploye to act as a sink for out-of-vocabulary events an backgroun noises. Confience easure is also suppose to represent the reliability of the hypothesis. It can be use in a post-processing proceure that rejects the less reliable hypotheses by thresholing the copute confience easure[8, 10]. In this paper, a new support vector achine base etho is propose for keywor spotting. The SVMs are gaining popularity ue to any attractive features an proising epirical perforance. The forulation eboies the Structural Risk Miniisation (SRM) principle, as oppose to the Epirical Risk Miniisation (ERM) approach coonly eploye within statistical learning ethos. SRM iniises an upper boun on the generalisation error, as oppose to ERM which iniises the error on the training ata. It is this ifference which equips SVMs with a greater potential to generalise, which is our goal in keywor spotting. The SVM can be applie to both classification an regression probles.[1, 4, 7, 13]. The organisation of this paper is as follows : Section 2 gives a brief escription of the basic principles of the SVM. The etails concerning atabase an recognition syste are given in section 3. In section 4, we present the way we use SVM for keywor spotting. Experiental results are escribe in section 5, an conclusions are given in section 6.

2 T L Q ( ( ( 2 Support Vector Machines (SVM) Support vector achines have been recently introuce as a new technique for solving pattern recognition probles [14, 17, 18]. They perfor pattern recognition between two point classes by fining a ecision, eterine by certain points of the training set, tere support vectors. This surface, which in soe feature space of possibility infinite iension can be seen as a hyperplane. It is obtaine fro the solution of the proble of quaratic prograing that epens on regularization paraeter. 2.1 The linear separable case Consier the proble of separating the set of training vectors belonging to two separate classes,!#"$ % & ')(*(+ where is a feature vector an a class label, with a hyperplane of equation :, - / The goal is to fin the hyperplane that separates the positive fro the negative exaples; the one that axiizes the argin woul generalise better as copare to other possible separating hyperplanes. A separating hyperplane in canonical for ust satisfy the following conitions :, (98;:<=% >(, -?.@0BA7')(98;:<=% C')( These can be cobine into one set of inequalities, -.10D57(FEG8H & I(*JK of a point fro the hyper- The istance L plane, 0O is,, 0M N, 0M PD RQ, - S.@0 QQ, QQ The optial separating hyperplane is given by axiizing the argin T given by the equation : U/VXW YZ\[]^ZX_!` L, 0M a.<u/b+c YZ\[]^ZX L, 0M ;H QQ, QQ To axiize the argin T, one nee to iniize : ef, g,6h The solution to the optiisation proble is given by the sale point of the Lagrange functional (Lagrangian) ij, 0M^klg,, ' k^n =, - /.10!'o(p k Where are the Lagrange ultipliers. The Lagrangian has k15 to be iniize with respect to, 0, an with 0. This proble can easily be transfore into the ual proble, an hence the solution is given by : krqb 4sIt+uIJvs k ' wx with constraints, ky5o3z{eg8h (*J krq =% 43 k k x x - x 2.2 The non-linear separable case In this case, the set of training vectors of two classes are non-linearly separable. To solve this proble, Cortex an r5o3vapnik [3] introuce non-negative variables,, which easure the iss-classification errors. The optiisation proble is now treate as a iniization of the classification error [8]. The separating hyperplane ust satisfy the following inequality :, - ;P.1065~.(2', - P.106A ')(D. {R8;: R8;: =%.( C')( The generalise optial separating hyperplane is eterine by the vector,, that iniizes the functional, ƒ, D,6h ).@ Where an C are constants. The ual proble corresponing to this case is slightly ifferent fro the linear separable case, the goal now is to axiize : subject to k!g' 3Aok!yA~ wx k! knx= x ;- zxm k q =% 43 E8g & I(*JK

3 p 2.3 Kernel support vector achines In the case where a linear bounary is inappropriate, the SVM can ap the input vector P into a high iensional space through function, where the SVM constructs a linear hyperplane in the high iensional space. Since fining the SVM solutions {- x involve the ot proucts of the saple vectors, kernel functions play a very iportant role in avoiing explicit proucing the appings, an avoiing x+ ˆ ;- zxm the curse of iensionality, so that, i.e., the ot prouct in that high iensional space is equivalent to a kernel function of the current space [2, 12]. In the linear an non-linear cases, the optial separating hyperplane efine by, q 0 an q is eterine as follows : Š, q P%.10 q where,, q ŒI k Ž ; 0 q, q, an >( ' - :N Mt 8\ >( The classification function is : O s + Nl r8\ui g, q P%.10 q O sz + PH r8 u gn Œ k q = N%.10 q p O sz M Nl y8 ui gn ŒI krq = 1 ;- P%.10qp Soe wiely use kernels are : Linear : ˆ!šl 4 - Polynoial : 1?l ƒ -).4({ Raial Basis Function (RBF) : ' 1?H œ žrnÿ'kq * h 3 Database an recognition syste 3.1 Database For training, we use 5300 sentences of the French BREF80 atabase, pronounce by 80 speakers. These sentences are recore at 16 KHz with 16 Q h bits. This is a general purpose atabase, an the sentences have no relationship with our application. The test atabase contains one hour of recoring speech of raio braocast news at 16 khz. It is segente into fragents of uration is 20s. This recore speech have been pronounce by several speakers (ifferent fro the speakers of the training atabase). In our application we choose 10 ifferent keywors. 3.2 Recognition syste The recognizer use in this work is a speaker inepenent Hien Markov Moel (HMM) syste. The oele unit is the phone, each phone is represente by 3-state, strictly left-to-right, continuous ensity HMM. A wor is represente by the concatenation of phone oels. The nuber of probability ensity function (pf) per state is eterine uring the training phase [5]. The paraeterization is base on MFCC (Mel- Frequency Cepstral Coefficients) paraeters. The user can oify this paraeterization : size of the analyzing winow, shift, nuber of triangular filters, lower an upper frequency cut-off of the filter bank, an nuber of the cepstral coefficients. Finally the elta (the first erivation) an acceleration coefficients (the secon erivation) are ae. In the following experients, the acoustic feature vectors are built as follows : 32s fraes with a frae shift of 10s, each frae is passe through a set of 24 triangular ban-pass filter resulting in a vector of 35 features, naely 11 static el-cepstral coefficients ( q is reove), 12 elta an 12 elta elta coefficients. In the recognition phase, we ajust paraeters in orer to have no eletion keywors (as consequence we obtaine a great nuber of insertion keywors). 4 SVM for keywor Spotting In this section, we escribe the way we have utilise the SVM for the keywor etection. After the recognition phase, the goal is to classify a sequence of etecte keywors into correct an incorrect keywors. For each keywor, we copute the fraes assigne to each phone state an extract the acoustic features. In orer to have goo inforation about the correctness of a wor hypothesis, we ust use the feature paraeters of each wor. In this case we choose the logarithžg š p of the acoustic observation likelihoo to provie a goo id unn \ B Q

4 classification. žg š 8 Where is the phone of the spoken utterance B > žg P žg ) phone sequence. + ª «D, an belongs to, which is the acoustic observation f sequence for the utterance. Each eleent is equivalent to f a B 0*n ;p œ n p, where an represent respectively, 8 the beginning an the ening fraes of the phone. For each phone in the spoken utterance, we co- ± \ ² žn š pute the acoustic observation likelihoo, Q using Viterbi algorith, then we obtain for each wor a feature vector of iension equals to the nuber of phones in it. Thus, for each keywor (insertion keywor, recognize correct keywor) we copute a feature vector which is use as input for the SVM. However, SVM syste nee the sae nuber of input for all keywors. For this reason, we fix the size of the input vector SVM as the largest wor size. In case of shorter wors, we coplete the feature vector with zeros. In our work the insertion keywor belongs to the class labelle -1 an the correct keywor is assigne to the class labelle +1. Thus, we classify the correct an the incorrect keywors. The SvFu package was use for these experients. It is available as freeware on : )"f"7 + {s s +œ@" œ Iœ {8 + + {s t*ģœoµb { œjž Plotting a graph of FRR versus FAR gives a Receiver Operating Characteristics (ROC) graph. Figure 1: ROC curves on test ata using a linear kernel by varying the value of C Figure 2: ROC curves on test ata using a RBF kernel by varying the value of 5 Experiental Results The atabase use in the secon phase of our experients (after the recognition phase) is copose of 600 keywors for the training ata, an 560 keywors for the test ata. In this work, we use linear an Raial Basis Function (RBF) kernels. To evaluate the perforances of our recognizers, we use two evaluation rates : ³ The False Acceptance Rate also calle False Alar Rate (FAR). It is efine by the equation : µ6"7 < + {s s +œoµ O œaž asi œ + as s MœoµB { œjž ³ The False Rejection Rate (FRR). FRR is efine by the equation : The resulting ROC curves, using linear SVM by varying the value of the paraeter, > 3z¹3š(*^3z (*(*(3z(33 are presente in figure 1.

5 The Equal Error Rate (EER) which is given by FAR=FRR, is about 16,34% obtaine by linear kernel in case 43z ( For the kernel RBF, we use several kernel paraeters an ifferent paraeters : I(*º(3z º*3z C 3z¹3š(*^3z (*(*(3z(3=3z an. We try all cobinations. We present in Table 1 best \ j results obtaine by ifferent values of the pair. Table 1: Recognition accuracy for RBF kernel SVM by varying the value of the pair (C, ) (C, ) ACC (in %) (0.01, 100) 88,96 (0.1, 200) 89,67 (1, 50) 90,56 (1, 90) 93,23 (1, 130) 92,52 (10, 145) (10, 185) 94,66 (10, 225) (100, 200) 92,23 (100, 240) 93,59 (100, 280) 92,59 As shown in Table 1, a recognition >(»=º accuracy of 94,66 % was obtaine with >(3 an. Accoring to these ¼ ½(3 obtaine results, we ecie to choose the value to achieve next experients for the RBF kernel SVM. Figure 2 presents ROC curves corresponing to the perforance obtaine using RBF kernel by varying the value of the paraeter 7 >(3, with. As entione, in our experients, we test several values of, but in orer to alleviate the figure, we choose to present only the interesting & I(O¾Iº(»=º º curves for These results eonstrate that the perforance of the RBF kernel are better than results obtaine for the linear SVM. The EER concerning the RBF kernel is about 15,23% copare to 16,34% obtaine by the linear kernel. =. 6 Conclusion This paper presents the results achieve by SVM techniques for the keywor spotting proble using linear an RBF kernels. Taking into account that is a first approach using support vector achine in keywor spotting, the results obtaine see to be very proising. In the near future, feature vector will be ajuste for each keywor in orer to have ore inforation about it. Other ifferent kernel types an paraeters will be experiente. References: [1] C. Burges. A tutorial on support vector achines for pattern recognition. In Data Mining an Knowlege Discovery, volue 2(2), [2] O. Chapelle, V. Vapnik, O. Bousquet, an S. Mukherjee. Choosing ultiple paraeters for support vector achines. In Machine Learning, volue 46, pages , [3] C. Cortes an V. Vapnick. Support vector networks. In Machine Learning, volue 20, pages 1 25, [4] T. Evgeniou an M. Pontil. Support vector achines: Theory an applications. volue 2049, pages , [5] D. Fohr, O. Mella, an C. Antoine. The autoatic speech recognition engine ESPERE : experients on telephone speech. In Proc. IEEE Int. Conf. Spoken Language Processing, [6] A. Ganapathiraju. Support Vector Machines for Speech Recognition. PhD thesis, Mississipi State University, [7] S. Gunn. Support vector achines for classification an regression. In Technical Report ISIS-1, [8] S. O. Kappari an T. J. Hazen. Wor an phone level acoustic confience scoring. In Proc. IEEE Int. Conf. on Acoustics, Speech, an Signal Processing, [9] J. Kharroubi, D. Petrovska, an G. Chollet. Cobining g s with suport vector achines for text-inepenent speaker verification. In EuroSpeech, [10] B. Maison an R. Gopinath. Robust confience annotation an rejection for continuous speech recognition. In Proc. IEEE Int. Conf. on Acoustics, Speech, an Signal Processing, [11] E. Osuna, R. Freun, an F. Girosi. Support vector achines : Training an applications. In A.I.Meo 1602, MIT, [12] E. Osuna, R. Freun, an F. Girosi. Training support vector achines : an applications to

6 face etection. In Conference on Coputer Vision an Pattern Recognition, [13] M. Pontil, R. Rifkin, an T. Evgeniou. Fro regression to classification in support vector achines. Technical Meo AIM-1649, [14] R. Rifkin, M. Pontil, an A. Verri. A note on support vector achine egeneracy. In Proceeings of the 10th International Conference on Algorithic Learning Theory, volue 1720, pages , [15] R. C. Rose an D. B. Paul. A hien Markov oel base keywor recognition syste. In Proc. IEEE Int. Conf. on Acoustics, Speech, an Signal Processing, [16] R. A. Sukkar an C. H. Lee. Vocabulary inepenent iscriinative utterance verification for nonkeywor rejection in subwor base speech recognition. In IEEE Trans. on Speech an Auio Processing, volue 4, pages , [17] V. Vapnick. The nature of satistical learning theory. In Springer-Verlag, [18] V. Vapnick. Satistical learning theory. In Johon Wiley, 1998.