SOUND CLASSIFICATION IN A SMART ROOM ENVIRONMENT: AN APPROACH USING GMM AND HMM METHODS

Transcription

1 SOUND CLASSIFICATION IN A SART ROO ENVIRONENT: AN APPROACH USING G AND H ETHODS ichel VACHER, Jean-François SERIGNAT and Stéhane CHAILLOL CLIPS - IAG, Team GEOD, UR CNRS-INPG-UJF 5524, 385, rue de la Bibliothèque, BP 53, 3804 Grenoble cedex 9, France, ichel.vacher@imag.fr Corresonding author: ichel VACHER Because of cost or convenience reasons, atients or elderly eole would be hositalized at home and smart information systems would be needed in order to assist human oerators. In this case, osition and hysiologic sensors give already numerous informations, but there are few studies for sound use in atient's habitation. However, sound classification and seech recognition may greatly increase the versatility of such a system: this will be rovided by detecting short sentences or words which could characterize a distress situation for the atient. Analysis and classification of sounds emitted in atient's habitation may be useful for atient's activity monitoring. Gs and Hs are well suited for sound classification. Until now, Gs are frequently used for sound classification in smart rooms because of their low comutational costs, but Hs should allow a finer analysis: indeed the use of 3 states Hs should allow better erformances by taking into account the variation of the signal according to time. For this framework a new sound corus was recorded in exerimental conditions. This corus includes 8 sound classes useful for our alication. The choice of needed acoustical features and the two aroaches are resented. Then an evaluation is made with the initial corus and with additional exerimental noise. The obtained results are comared. At the end of this framework a segmentation module is resented. This module has the ability of extracting isolated sounds in a record by the means of a wavelet filtering method which allows the extraction in noisy conditions. Key words: Gaussian mixture model; Hidden arkov model; Background noise; Sound classification in smart rooms; Wavelet transform.. INTRODUCTION It is well known that ageing is emerging as an imortant concern for Euroean countries. In this context the central challenge of health and long-term care olicies is to rovide full access to high-quality services for all, while ensuring the financial sustainability of these services. Progress in aids and assistive technologies might be a cost-efficient way to suort the suly of informal care and care rovisions. In this way seech analysis and sound classification can give interesting information by taking into account distress calls from the atient and fall sounds. Therefore sound classification and seech analysis can give information on the atient and may hel the decision-making by the medical monitoring system []. The medical monitoring system, described in [], uses 2 kinds of information. Information issued from the medical sensors, the actimeter and door contacts are analysed in order to detect a difference in the behaviour and state of the atient. Information given by the sound analysis system [2] may be analyzed in the same manner but can also be used to detect critical or distress situation. It will be the case when a sentence like Hel me, Doctor quickly is recognized or when a scream, an obect fall or a glass breaking is classified. The imlementation of the sound analysis system must meet 2 different aims: - the real-time ability of the system, - a good recision for seech recognition and sound classification. The real-time ability is achieved if sounds and seech are detected on the flow and not missed. Concerning seech recognition and sound classification, results may be known some seconds after the sound event but it is very imortant that neither false nor missed alarm occurs. In this aer we will only discuss sound classification; Gaussian

2 2 ichel VACHER, Jean-François SERIGNAT, Stéhane CHAILLOL ixture odel (G) [3] and Hidden arkov odel (H) [4], [5], [6] based methods are often used in this area. The G method is easy to imlement while the H method takes into account the shae of the signal. For this framework, the ALIZE library was used as well for the G method than for the H method.. Short Overview of the Sound Analysis System The aim of our global roect is to obtain useful sound information and to transmit it through network to a medical suervising alication in a medical centre. The habitat we used for exeriments is a 30m 2 aartment situated at the TIC laboratory inside the Faculty of edicine of Grenoble. It is equied with various sensors, esecially microhones in every room (hall, toilet, shower-room, living-room) []. The entire tele-monitoring system is comosed of three comuters which exchange information through a local network (see Figure ). This system is designed for the surveillance of the elderly, convalescent ersons or regnant women. Its main goal is to detect serious accidents or falls or faintness at any lace of the aartment. Each time a sound event is analysed, a message is sent to the Data Fusion PC, notifying occurrence time of detection, most robable sound class or recognized sentence, localization of the emitting source. From this and from other data obtained from localisation and hysical sensors, the Data Fusion PC can send an alarm if necessary. Figure. edical Telemonitoring System The sound analysis system has been divided in four modules as shown in Figure 2. The first module is the detection module in charge of extraction of audio events from the signal flow. Extracted signals are then transmitted to the segmentation stage, which switches them to the classification module in case of life sounds or to the RAPHAEL recognition module [2] in case of seech. At the end, the obtained information will be send to the data fusion system, which will resond to the question: Is the atient in a normal or a distress situation? Figure 2. Sound analysis system

3 Sound Classification in a Smart Room Environment: an Aroach using G and H ethods 3 2. G AND H ETHODS G and H methods are well suited for sound classification [8], [9]. H-based methods are used in the acoustical stage of seech recognition systems, rosodic analysis [0] and isolated word recognition []. The imlementation of these two methods in this framework uses the ALIZE library [3]. It was necessary to add some secific extensions for our alication. 2. Gaussian ixture odels - Gs The classification with a G-based method suoses that the acoustical arameter reartition for a sound class may be modelled with a sum of Gaussian distributions. This method evolves in two stes: a training ste and a classification ste as shown in Figure 3. The acoustic re-rocessing stage in charge of feature extraction will be described in subsection 4.. During the training ste and for each sound class the characteristics of each Gaussian model are estimated, the number of these models N will be discussed later in subsection 4.2. Parameters of the Gaussian distribution m ( m < N) are for the sound class k ( k 8) : the likelihoodπ k, m, the mean vector µ and the inverse covariance matrix k, m Σ k, m. During the initialisation ste, these arameters are initialized by the mean of the arbitrary artition of the training corus in N equal sized arts. This ste is followed by a second ste including 2 iterations of the E algorithm (Exectation aximisation) on 20% of the corus (randomly drawn). The last ste is made of 2 iterations of the E algorithm on the full corus. In the classification ste the likelihood of each frame of the signal is calculated for each sound class (see Equation ). A frame is a vector of d acoustical features. The global likelihood X ω ) for one class ( k is the geometrical average of the likelihoods of the n frames as exressed in Equation 2, and the signal belongs to the class for which likelihood is maximal. N ( x i ω k ) = π k, m..ex( A,, ) d / 2 / 2 i k m (2π ) det( ) () m= With n k, m Τ A i k, m = ( xi µ k, m). Σk, m.( xi µ k, 2, m ( X ω ) = ( ω ) (2) k x i i= k ) Figure 3. Block diagram of the G and H methods: Training and Classification 3

4 4 ichel VACHER, Jean-François SERIGNAT, Stéhane CHAILLOL 2.2 H Classification of Sounds In the context of audio signal encoding, the inut signal can be decomosed into transient, tonal and residual comonents as described by Daudet in [2]. We choose then to use 5 states H as shown in Figure 4, the states q0 and q4 corresonding to the silent art at the beginning and at the end of the signal. There is no transition ossibility from q0 to q4 because they reresent the same state. The states q, q2 and q3 are related to the three comonents of the signal. A transition is ossible from each state qi to a state q if is more or equal to i, excet from q0 to q4. For any sounds, some of the 3 states may be emty. P i denotes the transitional robabilities from the state qi to the state q. P P 00 P 22 P 4 P P P 0 P 02 P 2 q0 q q2 q3 q4 P 23 P 34 P 03 P 24 Figure 4. H state transitions An examle is given in Figure 5 in the case of a scream; it is the result of the training state described in section 2.3. The first state q0 has a very short duration and is not visible because of the scale of the figure. The state q is short and made of the establishing art of the signal. The state q2 is corresonding to the highest energy art and q3 to the decreasing energy art. During q3 resonant frequencies are decreasing too. Figure 5. H states of a scream The classification ste uses a Viterbi algorithm [3] to estimate the class k of a sound X between m n models, a sound being reresented by a sequence of vectors of n comonents X. First member of Equation 3 can be simlified because of the equal robability of each sound class. k = arg max ( X ) P( ) = arg max ( X ), 0 < m (3)

5 Sound Classification in a Smart Room Environment: an Aroach using G and H ethods 5 The robabilities are estimated by using a Viterbi method. The Viterbi algorithm is a forward robability method of best ath estimation. For simlification all the sums are relaced by a maximum function, and then the estimated robability for the best artial ath q i from initial state q to the state q i, after emission of the first vectors X of X must be exressed by Equation 4, ( q i, X ) = max ( q ) ( q, X, ), 0 < n (4) k k k where ( qi, X ) denotes the robability of the artial ath from the state q to the state q i of the model, X the first ( ) vectors of X, th X, q the state q when the ) x the th vector of X, X the sequence of the first vectors of ( comonents is reached. Considering the equal robability of each sound class the equation can be written as in Equation 5. n ( q, X ) = max ( q, X ) P( q q, ) ( x q (5) i [ ] ) k i k i k The robabilities of each vector are evaluated through Gaussian models, each state qk or qi being modelled by a G in conunction with the robability of transition. The robability P X ) is then estimated for each model using Equation 6. ( n n ( X ) = ( qf, X ) Where q F denotes the final state. (6) The signal belongs to the sound class q associated to the model value. 2.3 H Training and Automatic Labelling q for which the robability has the highest Indexed corus 8 sound classes (Short art of the recorded corus) st ass (bootstra) nd th 2 to n ass 2 to n H State Feature Extraction Generation q0...q4 (LFCC) G Training VITERBI Alignement 8 sound classes wav Indexation data SA format Probabilities of transition Udate Udate Convergence? No Yes End Figure 6. H Training and Automatic Labelling System The global H training and automatic labelling system is resented in Figure 6. The most used method for H training is the Baum-Welch algorithm [3], as it is not available in the ALIZE library we have develoed a H training system using the Viterbi algorithm. This system, described in Figure 6, requires an indexed starting corus. Thus a short art, around 0%, of the corus was manually indexed by examining the sectrogram of each sound; indexing marks refer to H state transitions. Training is erformed for each sound class searately. 5

6 6 ichel VACHER, Jean-François SERIGNAT, Stéhane CHAILLOL At the beginning of the first stage, segmentation data are used in order to extract the 5 states q0 q4 from the initialisation corus. Then acoustic re-rocessing ste is initiated and roduces for all the sound classes the corresonding acoustical features. It is then ossible to obtain a G model for the 5 states of all the classes. The same method as in section 2. is used. The Viterbi algorithm is then alied on the entire corus using these G models. The outut of this algorithm is the best ath across each sound wave and then, for each frame of the sound wave, the corresonding H state. Indexation data for the full corus and robabilities of transition are then extracted from these oututs. From the second ass, the same rocess is started again but G models may then be evaluated from the full corus. After n iterations of the rocess the convergence is reached if the indexation data remains quite constant. That requires between 20 and 50 stes. At the end of the final training ste, the indexation values for the initial indexed corus have changed because of the otimisation rocess. 3. SOUND DATABASE Each sound roduced in an aartment is characteristic of a normal atient s activity (door sla, dishes ), a ossible distress situation (obect fall, scream ) or a atient s hysiology (cough ). Sounds related to the atient s hysiology are not yet taken into account because of the difficulty in recording such sounds. Therefore a new corus, adated to this framework, was recorded; this corus is made of 8 everyday sound classes related to two categories: Normal sounds related to a usual activity, Critical sounds related to the ossibility of a distress situation for the atient and thus giving very imortant information to be sent to the remote monitoring system. A small share of the corus consists of sounds extracted from a receding corus recorded at the time of former studies [7]. New sounds have been recorded in the CLIPS laboratory using omni-directional wireless microhones (SENNHEISER ew500). Some sounds were obtained from the Web [8]. Some characteristics of this corus are given on Table. Each sound is recorded in one file; the samling rate is 6 khz. Because of the use of an H classifier, each file begins and ends with a silence art of minimal duration 32 ms. The average RSB of the corus is +27 db. With this corus we have generated a noised corus with 4 levels of signal to noise ratio (SNR=+8 db, +7 db, +22 db, +26 db). The noise was recorded in an aartment. The original corus and the noised corus have been used for the classification tests. Table. Everyday sound corus Class of Sound Former Corus New Records Internet Number of Files Average Duration of one Sound (ms) Dishes Sounds 45% 50% 5% Door Lock - 00% Door Sla 40% 60% Glass Breaking 40% 50% 0% Obect Falls - 00% Ringing Phone 5% 70% 5% Screams 80% - 20% Ste Sounds 0% 60% 30% Entire Corus 29% 6% 0%

7 Sound Classification in a Smart Room Environment: an Aroach using G and H ethods 7 Some examles of sounds are shown in Figure 7. The sectra are very different but in each case high frequency comonents must be taken into account. In case of the door sla sound, there are two arts: in the first art sound is like a decreasing white noise, in the second art some resonant frequency are imortant. The synthetic ringing bell is constituted of discrete and regularly saced frequencies. The scream is very similar to voice signal with a high number of harmonics. Resonant frequencies are imortant during all the dishes sound, imact between a cu and a saucer. Door Sla Ringing Bell Scream Dishes Sound Figure 7. Some examles of sounds with the corresonding sonograms Key Lock 4. FEATURES AND ODEL SELECTION G and H classification methods are not erformed directly on the signal, but use extracted acoustical arameters which are synthetic reresentations of the time signal. Analysis window width for each frame was set to 6 ms with an overla of 8 ms. For a samling rate of 6 khz the analysis window is comosed of 2 8 samles, an integer ower of two being required by Fast Fourier Transform analysis. This width is a comromise between the time recision of state transitions and with frequency analysis constraints. A great number of the signals being as short as 86 ms, it might be imossible to use a wider analysis frame. 4. Features Acoustical arameters classically used in seech/seaker recognition are: FCC (el Frequency Cestral Coefficients), LFCC (Linear Frequency Cestral Coefficient) and LPC (Linear Predictive Coefficients). FCC are frequently used in seech recognition because of their characteristics that are very similar to human hearing mainly thanks to the logarithmic el frequency scale. H H 0 f 0 Linear Scale (LFCC) f 0 f 0 f f n EL Scale (FCC) f Figure 8. LFCC and EL Triangular Filter Resonse As discussed in section 3, the bandwidth of sound signals is very large and includes frequently high frequency comonents. The comuting stes for the LFCC and FCC arameters are: re-emhasis and windowing; FFT of the analysis frame signal; triangular filtering; logarithmic calculus of the filtered coefficients and inverse cosine transform. The inverse cosine transform is obtained according to Equation 6. 7

8 8 ichel VACHER, Jean-François SERIGNAT, Stéhane CHAILLOL As shown in Figure 8, the bandwidth is constant over the sectrum for LFCC but larger in high frequencies for FCC because of EL logarithmic scale. In our study, it is imortant to use comonents allowing an equal sensitivity over the full bandwidth as allowed by LFCC. All the 24 coefficients are considered in order to take into account the full bandwidth. π n [ ] = E[ m] cos = c n m 0 ( m ) 2, 0 n < (6) Normalized energy is not used as additional arameter, this arameter being too deendent of exerimental recording conditions. Derivatives of first ( delta ) and second order ( delta-delta ) of LFCC arameters are referred. The total amount of used arameters is then Number of odels in the Case of the G-based ethod The Bayesian Information Criterion (BIC) is used in this aer in order to determine the otimal number of Gaussian models [4]. This criterion is well suited for Gaussian mixture as roved by Roeder and Wassermann [5]. BIC criterion selects the model trough the maximization of integrated likelihood (). Where BIC = L + ν ln( ) (7) m, K 2 m, K m, K n L m, K is logarithmic maximum of likelihood, equal to log (,, θ ) ~ f x m K ( f is the integrated likelihood), m is the model and K the comonent number of the model, ν m, K is the number of free arameters of the m model and n is the number of frames. The minimum value of BIC indicates the best model. 50,000 Sublot a: Sound Classes C, C3, C7, C9 C 50,000 Sublot b: Sound Classes C2, C4, C5, C6 C2 BIC Coefficient 20,000 C3 C7 C9 BIC Coefficient 25,000 C4 C5 C Number of Gaussian odels Number of Gaussian odels Figure 9. BIC Coefficient Evolution for the 8 sound classes, G Evaluation with 24 LFCC features in conunction with derivatives of first and second order Table 2. Corresondence between number and class of sounds Number Class of Sounds Duration Number Class of Sounds Duration C Door Sla 6 min 20 s C5 Screams 3 min 30 s C2 Glass Breaking 2 min 53 s C6 Obect Falls 2 min 3 s C3 Ringing Phone 5 min 7 s C7 Dishes Sounds 3 min 40 s C4 Ste Sounds 44 s C9 Door Lock min s

9 Sound Classification in a Smart Room Environment: an Aroach using G and H ethods 9 The BIC criterion has been used first for the sound class and for the seech class in noiseless conditions, for 4, 5 and 24 Gaussian models in case of 24 LFCC arameters in conunction with derivatives of first and second order. The results of the Figure 9 are given for a number of Gaussian models between 4 and 24 in case of C, C3, C7, C9 sound classes (sublot a) and C2, C4, C5, C6 sound classes (sublot b). According to the BIC criterion, erformances will be otimal when the log-likelihood of the observations, given the Gs, is minimal. As it aears on these 8 curves, the otimal Gaussian number is different for each sound class. In sublot a, the criterion is minimal between 2 and 20 models for the bell ringing class (C3), the records of this class are very heterogeneous (old style bells, synthesised bells ). Curves are very flat for C, C7 and C9 but increase below 6 models. In sublot b, curves are slowly increasing excet for ste sounds (C4), so a number of 2 Gaussian models seem to be accetable. The C4 class could be neglected because these sounds are very low level and not often detected in the real environment. We have decided to use 2 Gaussian models, which may be a good comromise between classification erformances and calculus consumtion (real time constraints). 4.3 Number of odels in the Case of the H-based ethod Since the likelihood of one frame of signal is evaluated using Gs, the number of models may be chosen in the same conditions with the BIC criterion. In the Figure 0 and for the sound classes C5 (screams) and C7 (dishes) the BIC coefficient is reresented as function of the Gaussian number for the three states q, q2 and q3. The curve shaes are not very different as the receding related to G configuration. We then chose the same number of Gaussian for H evaluation. 50,000 Sublot a: Sound Classe C7, q, q2, q3 5,000 Sublot b: Sound Classe C5, q, q2, q3 q q BIC Coefficient q2 q3 BIC Coefficient q2 q3 30, Number of Gaussian odels 0, Number of Gaussian odels Figure 0. BIC Coefficient Evolution for C5 and C7 sound classes, states q, q2 and q3, 24 LFCC features 5. CLASSIFICATION EVALUATION Training is made with original sounds but testing is made with original sounds and sounds mixed with exerimental noise at 4 different RSB levels. The tests use a cross validation rotocol, training is achieved with 90% of the original sound corus and each of the 0% remaining files is evaluated at these 4 RSB levels and for the original sounds. The sound classification erformances are evaluated through the Classification Error Rate (CER), which reresents the ratio between badly classified sounds and the total number of sounds to be classified. The number of Gaussian is fixed to 2 for the G-based method and for the H-based method. The samling rate is 6 khz, the bandwidth is then 8 khz. The analysis window width is 6 ms with an overla of 8 ms. Results are shown in Table 3. We can observe that in all cases (excet at SNR = +7 db) results are the best for the H-based method against the G-based method, even if the differential / acceleration coefficients are used for the G-based method and not for the H-based method. We can conclude that Hs allow a finer analysis by taking into account the temoral shae of the signal. Then the use of 3 states Hs allows better erformances. 9

10 0 ichel VACHER, Jean-François SERIGNAT, Stéhane CHAILLOL Best results are reached by the use of 24 LFCC arameters with derivatives of first and second order and the H-based method. The Classification Error Rate is.7% for the original corus and below 6% at SNR = +22 db. These values are good according to former results [2] and other results in the literature [6], [7]. Table 3. Classification Error Rate (%) SNR 24 LFCC only 24 LFCC and [db] G H G H Original H SOUND SEGENTATION The roosed global sound recognition system is comosed of two modules: the first is the segmentation system and the second is the H classification system yet resented. A segmentation system must be able to detect the beginning and the end of each isolated sound in a flow and that may be achieved by the way of a H segmentation model with only one class of sounds. This class includes all the sounds of the revious classes. H state transitions are shown in Figure. A ossible transition has been added from q4 to q0 which are actually the same state (the silent state). P P 00 P 22 P 4 P P P 0 P 2 P P q0 q q2 q3 q4 P 02 P 24 P 03 P 40 Figure. H State Transitions for Segmentation Training is oerated in the same conditions than in subsection 2.3 but with a unique model over the comlete corus after a bootstra ste using a short labelled art of the corus. The number of Gaussian mixtures is 4; a greater value is not needed because we must only make the distinction between silence and one of the 3 sound states. We recorded in real conditions a dedicated corus in our laboratory using the same omni-directional wireless microhones (SENNHEISER ew500). This test corus is made of 0 sound wave files; so each wave file contains a sequence of about 0 sounds of all the sound classes excet ste sounds. The total duration of the corus is 9 minutes. The total amount is 29 sounds and the average RSB is +28 db. Table 4. Classification Results (Number of files) C C2 C3 C5 C6 C7 C9 Classified as C 8 Classified as C2 7 Classified as C3 4 Classified as C5 25 Classified as C6 3 Classified as C7 4 4 Classified as C9 5

11 Sound Classification in a Smart Room Environment: an Aroach using G and H ethods All the 29 sounds are correctly detected; there is no missed or false sound detection. The results after the classification ste are given in Table 4. We can notice that one door sla sound is classified as obect fall and that 4 glass breaking sounds are classified as dishes sounds. This denotes the great similarity of these sounds because they are roduced in a similar manner. An imact between a cu and a saucer is not very different of an imact between a cu and the floor even if the cu is broken at the end. So the recognition error rate is 3.9%. Segmentation results are very oor in noise conditions. So we studied an algorithm of noise reduction in the arts of the signal corresonding to the silent art. In order to not degrade the sound signal with filtering artefacts we try not to modify this art, indeed the noise is not stationary and not known at any time. The roosed method suoses that at least one frame of signal containing only noise may be isolated at any moments before the sound signal; these frames of noise will be used for reference. Figure 2. Wavelet Noise Reduction The Discrete Wavelet Transform (DWT) of each frame of noise is calculated on 256 samle windows; wavelets are distributed over 9 wavelet coefficients. For each coefficient level, the maximal absolute value is memorised. The rocessing of the signal to be segmented is described in Figure 2; it is oerated frame after frame. During the first ste, the absolute value of each wavelet is comared to the memorised threshold at the corresonding level value; the number of wavelets below the threshold is n. An attenuation coefficient is then evaluated as function of n; this coefficient is linearly decreasing between.0 (n = 29) and 0.2 (n = 256). Inverse DWT is then oerated after alying the same attenuation over all of the wavelets. This system is at this time in the course of evaluation and only the first results are available. The final results should be resented at the time of the conference. 7. CONCLUSION In this aer we have resented a comarison between two methods for sound classification in the framework of a medical remote monitoring alication. Analysis and classification of sounds emitted in atient's habitation may be useful for atient's activity monitoring. An adated sound corus was recorded in exerimental conditions and used for evaluation urose; this corus includes 8 sound classes which are useful for this alication. G-based methods are frequently used for sound classification in smart rooms because of their low calculus consumtion, but H-based methods should allow a finer analysis: indeed the use of 3 states Hs should allow better erformances by taking into account the temoral shae of the signal. The two aroaches are resented like the needed acoustical features. Then an evaluation is made with the initial corus and with additional exerimental noise in order to comare these two methods. In the same noise conditions, H results are always the best. Best results are achieved with the original corus (SNR = +28 db), the Classification Error Rate is below 2%. At +7 db, the CER is below 0% with 24 LFCC arameters and their derivatives of first and second order. However the time consumtion is very imortant in the case of the H algorithm and his imlementation in a real-time system will involve to greatly otimize the algorithm and to use fast rocessors.

12 2 ichel VACHER, Jean-François SERIGNAT, Stéhane CHAILLOL At the end of this framework a segmentation module is resented. This module has the ability of extracting isolated sounds in a record by the means of a wavelet filtering method which allows the extraction in noisy conditions. We are working to add the ossibility of seech segmentation in order to extract at once seech and sounds from a wave record. REFERENCES. RIALLE, V., LAY, J.B., NOURY, N., BAJOLLE, L., Remote monitoring of atients at home: A software Agent aroach, Comuter ethods and Programs in Biomedicine, Vol. 72, Issue 3, , VACHER,., SERIGNAT, J.-F., CHAILLOL, S., ISTRATE, D., POPESCU, V., Seech and Sound Use in a Remote onitoring System for Health Care, Lecture Notes in Comuter Science, Artificial Intelligence, Text Seech and Dialogue, Brno, Czech Reublic, Vol. 488, Sringer,. 7-78, BONASTRE, J.-F., ALIZE: A software toolkit for Seaker Recognition, htt:// AJERA J., CCOWAN L., BOURLARD H., Seech/music segmentation using entroy and dynamism features in a H classification framework, Seech Communication 2003, Vol. 40, , LEFEVRE S., AILLARD B., VINCENT N., A two level classifier for audio segmentation, IEEE International Conference on Pattern Recognition, ICPR 02 Proceedings, Vol. 3, Aug REYES-GOEZ. J., ELLIS D. P., Selection Parameter Estimation and Discriminative Training of Hidden arkov odels for General Audio odeling, IEEE International Conference on ultimedia and Exo, ICE 03 Proceedings, Vol., , July ISTRATE D., CASTELLI E., VACHER., BESACIER L., SERIGNAT J.-F., Information Extraction From Sound for edical Telemonitoring, IEEE Transactions on Information Technology in Biomedicine, Vol. 0, NO. 2, , Aril, PINQUIER, J., SENAC, C., ANDRE-OBRECHT, R., Seech and music classification in audio documents, IEEE International Conference on Acoustics, Seech and Signal Processing, ICASP 2002, Vol. 4,. 464, YAADA, T., WATANABE, N., Detection using non-seech odels and H Comosition, Worksho on Hands-free Seech Communication, Tokyo, Jaan, VICSI, K., SZASZAK, G., Prosodic Cues for Automatic Phrase Boundary Detection in ASR, Lecture Notes in Comuter Science, Artificial Intelligence, Text Seech and Dialogue, Brno, Czech Reublic, Vol. 488, Sringer, , CARUNTU, A., TODOREAN, G., A Comarative Study of the ethods Used in Isolated Word Recognition, Seech Technology and Human- Comuter Dialogue, , DAUDET, L., TORRESANI, B., Hybrid reresentations for audiohonic signal encoding, Journal of Signal Processing, Secial issue on Image and Video Coding Beyond Standards, Vol. 82(), , Nov BOITE, R., BOULARD, H., DUTOIT, T., HANCQ, J., LEICH, H., Traitement de la arole, Presses olytechniques et universitaires normandes, Lausanne, , 2000, ISBN SCHWARZ, G., Estimating the dimension of a model, Annals of Statistics, Vol. 6, , ROEDER, K., WASSERANN, L., Practical bayesian density estimation using mixtures of normals, Journal of the American Statistical Association, Vol. 92, , COWLING,., SITTE, R., Analysis of seech recognition techniques for use in a non-seech recognition system, IEEE Transactions on Seech and Audio Processing, Vol. 0, , DUFAUX, A., BESACIER, L., ANSORGE,., PELLANDINI, F., Automatic Sound Detection and Recognition for Noisy Environment, Eusico 2000, Tamere, Finland, Se Bruitage, Bruitages Gratuits : Sound-Fishing.net, htt:// Nov