WO2007003505A1 - Method and device for segmenting and labelling the contents of an input signal in the form of a continuous flow of undifferentiated data - Google Patents

Abstract

The invention concerns a method for segmenting and labelling the contents of an input signal (141) in the form of a continuous flow of undifferentiated data, in accordance with at least two classes of predefined data enabling the input signal (140) to be segmented into differentiated and categorized predefined data (147), based on a corpus of learning data.

Description

 Method and apparatus for segmentation and labeling of the content of an input signal in the form of a continuous stream of undifferentiated input data

1. Field of the invention

The field of the invention is that of pattern recognition and data classification.

More specifically, the invention relates to a novel technique of segmenting a given signal composed of data of different types and undifferentiated into a plurality of classified data segments following previously modeled data classes.

The invention applies in particular, but not exclusively, to the segmentation of an audio signal into speech and music segments, and by extension, to the automated processing of audiovisual documents, or even to the indexing of sound documents. , for archival purposes for example.

An example of implementation and application of the invention are given later in this document, for a music speech segmentation system whose purpose is to segment a digital audio signal into segments of variable size and to label each of the segments. detected according to their content: speech or music.

2. State of the art

An example of a segmentation of an audio stream (10) according to two classes: a speech class (15) and a music class (16) is given in FIG. 1 which illustrates a time alignment (17) of the different speech classes and music on portions or segments (11, 12, 13, 14) of the audio stream (10).

It is indeed, within a signal of undifferentiated input data mixing speech and music, over a given duration, to identify and to distinguish on the whole of the signal the segments concerning speech and those concerning music. It is known from the prior art, segmentation techniques according to at least two distinct classes, for distinguishing on an input signal the speech segments of the music segments. These techniques of the prior art implement for the segmentation of algorithms based on models of Gaussian mixtures (or MMG) and or hidden Markov models (or MMC), which we recall here respectively the main principles, so to facilitate the understanding of the rest of this document.

2.1 Gaussian Mixture Models (MMG) MMGs are used to model a probability distribution by a sum of Gaussians and associate a weighted sum (mixture) of functions with each previously determined class. The most commonly used is the weighted sum of multidimensional Gaussian probability density functions, a Gaussian being defined as the representation of the normal law (Gaussian normal law, Laplace-Gauss law) whose density of probability is written:

- m is the average;

- σ the standard deviation (the variance used in the following description corresponds to the squaring of the standard deviation, it will be noted: σ ² ). 2.2 Hidden Markov Models (MMC)

Hidden Markov models (MMC or HMM) are regularly used in areas such as speech recognition, biological sequence analysis, and textual or musical information retrieval. and more generally pure segmentation of signals.

An MMC is defined as a double stochastic process, whose first underlying stochastic process is not directly observable (the hidden process) and whose second stochastic process produces a sequence of observations. An MMC consists of a finite number of P states. At each clock tick, a new state is reached according to a transition probability distribution that depends only on the previous state (Markovian property). Note that there may be transitions from a state to itself, which defines the notion of re-looping a state on itself. Thus, following the crossing of a transition of a hidden Markov model, an observation is emitted according to a probability distribution which depends on the current state situated upstream of the transition.

An example of a topology of a 3-state MMC 20, 21, 22 is given in FIG. 2. In this figure, each state 20, 21 and 22 contains a Gaussian mixture model (MMG) respectively denoted MMG1, MMG2. and MMG3.

The evolution of the MMC is carried out by crossing the P ₁₂ or P ₂₃ transitions, for the passage of a first state to another state of the MMC, or by re-looping on the current state, by crossing the P ₁₁ transitions. , P ₂₂ or P ₃₃ , in the example given. For more details on the MMC, the reader can refer to the doctoral thesis realized in 2004 by J. PESfQUIER at the IT research institute of Toulouse entitled "Sound indexing: research of primary components for an audiovisual structuring" .

2.3 Overview of Known Prior Art Techniques for Segmenting an Audio Signal into Speech and Music Data Segments

Most of the prior art known techniques for segmenting an audio signal into speech data and music data segments rely on Gaussian law mix model (GAM) training and Markov hidden.

In the literature, the probabilistic tools represented by Gaussian mixing models are commonly used for the segmentation of an audio signal into two distinct data classes, as described in the scientific article of E. SCHEIRER & M. STANLEY, "Construction and Evaluation of a Robust Multifile Speech / Music Discriminator" - ICASSP 97, April 21-24, Munich, Germany, and J. PINQUIER's doctoral thesis, already cited in section 2.1. The general principle of the method for estimating the parameters of the MMGs of two classes of data to be modeled (noted here Class 1 and Class 2, in the illustrative example), used in the techniques of the prior art, is explained through FIG. FIG. 3 is a flowchart for outputting two Markov chains (as shown in FIG. 4) modeling classes 1 and 2;

2 of data, according to a method proposed by J. RAZICK, D. FOHR, O. MELLA, in an article entitled "Segmentation speech / music for automatic transcription", published in the proceedings of the Journées d'études sur la parole - JEP 2004, in April 2004, in Fez, Morocco. The hidden Markov chains (MMC) thus obtained for the modeling of classes 1 and 2 are exploited by the last step of the segmentation and recognition method, which implements a Viterbi algorithm from the MMCs obtained, as illustrated through of Figure 5.

We describe below the different steps of the estimation methods of the parameters of the MMG (FIGS. 3 and 4) and segmentation (FIG. 5), with respect to FIGS.

The first step of the method for estimating the parameters of the MMGs known from the prior art is a step (31) of extracting descriptors from the input audio signal (30). A large collection of descriptors is used in the literature, among which, the modulation of energy at 4 Hz, the percentage of frames of the low energy signal, the "Spectral Rolloff Point" (English terminology commonly accepted by the technical community of the field ), the spectral centroid, the spectral flux.

Other descriptors more commonly used in the field of speech recognition are also sometimes used, for example the MFCC coefficients (for "mel frequency cepstral coefficients" in English, or in French "cepstral coefficients of frequency of mels"). .

A second step of the method for estimating MMG parameters according to the prior art is a step (32) of the allocation descriptors of extracts to the different classes to be modeled (Cl class (32i) and class C2 (32 ₂₎ in the example).

This second step (32) is facilitated by labeling (33i) (33 ₂ ), the most often manual of each class C1 and C2 to model, as shown in Figure 3.

A third step of the method for estimating the parameters of the MMGs according to the prior art comprises in particular a step (34) for learning the parameters of the Gaussian law mixture models from the training data.

This step (34 _ls 34 ₂ ) of learning takes place in two sub-steps: the first (35 _ls 35 ₂ ) is an initialization of the model by vector quantization, for example based on the algorithm LBG described in the article Y LINDE, A. BUZO, RM GRAY: "An algorithm for vector quantizer", IEEE Trans on Corn.,

January 1980, vol 28.

The second sub-step (3 O ₁ , 362) is an optimization of the mixing parameters (Gaussian averages and variances) by the classical EM algorithm

(for "Expectation Maximization" in English, or "Expected Maximization" in French), a detailed description of which is given in the doctoral thesis of

J. PINQUIER, already mentioned in paragraph 2.1.

Thus, at the output of the estimation method, a set of Gaussian mixing models defining respectively in a fourth step the states of a hidden Markov chain (37) modeling the Cl class and a Markov chain (38) respectively. hidden model C2, as shown in Figure 4.

As illustrated in FIG. 4 and more precisely, each of the classes (37) (38) of searched segments (segments of speech or of music in the example cited) is modeled by a hidden Markov chain with several states (37). _ls ... 37k ₊ i) and (38 ₁₅ ... 38 _{1 + 1} ), respectively, each state (37 _ls ... 37k ₊ i) and (38 _l5 ... 38 _{1 + 1} ) consisting of MMG learned at the stage (34 _ls 34 ₂ ) of learning, as described in J. AJMERA, I. McCOWAN, H. BOURLARD, "Speech / Music segmentation using entropy and dynamism features in a HMM classification framework", Speech communication - Elsevier - 2003.

As illustrated in FIG. 5, a last and fifth step consists in segmenting and recognizing C2 Clet classes modeled on the input audio signal (50). This last step is carried out by application of the algorithm (51) of

Viterbi on hidden Markov chains (37) (38), the latter allowing determine the optimal alignment of the acoustic form of the signal on one of Markov's two models (37) (38), that is, to determine the path in one of the hidden Markov chains that leads to the most high probability of transmitting the data form considered (speech (52) or music (53), for example). A major disadvantage of the prior art, however, lies in the "dispersive" behavior induced by the learning step and its sub-stages of initialization of the model by vector quantization, and optimization of the parameters of the mixture (means and variances of the Gaussian) by the classical EM algorithm, noted VQ + EM in the following. As illustrated in the diagram in FIG. 8.2 which presents the dispersion of the points obtained by means of a training by the algorithm VQ + EM according to the prior art, each point of FIG. 8.2 representing the distribution of the Gaussian obtained on the segments of the input audio signal by the application of the VQ + EM algorithm. Indeed, it is clear that the main interest of the application of such a learning algorithm VQ + EM is to best model each class of data individually, while one of the technical problems posed by the present invention is to identify precisely in an undifferentiated data stream, the streams of the stream respectively corresponding to classes of data well identified and separated.

Other disadvantages of the aforementioned method according to the prior art concern the extra cost in terms of calculation time and the often questionable quality of the segmentation and recognition results obtained on the input data signal, these drawbacks being due mainly to the use of the Markov chain mixing model, and the vector discretization of the global cloud of points representative of the Gaussian labels of the two classes of speech and music, as shown in Figure 8 .1.

3. Objectives of the invention

The invention particularly aims to overcome these various disadvantages of the state of the art.

More specifically, an object of the invention is to provide a technique that is more reliable in terms of the result and quality of segmentation of a signal input in the form of a set of undifferentiated data and recognition of the class of data to which each segment of the input signal belongs.

Another object of the invention is to provide such a technique which is therefore particularly efficient in terms of discrimination between speech segments and music segments, from an input signal in the form of a stream of speech. undifferentiated music and speech data.

A further object of the invention is to propose a new technique that is more economical in terms of computing time for segmentation and recognition, and therefore better suited to application areas requiring the processing of large volumes of data.

A final objective of the invention is to provide a new technique of segmentation and classification of data of different types contained in an undifferentiated input data stream, which is simple and inexpensive in terms of implementation.

4. Summary of the invention

These objectives, as well as others which will appear later, are achieved according to the invention by means of a method of segmentation and labeling of the content of an input signal (141) in the form of a continuous stream of undifferentiated input data, according to at least two predefined data classes, comprising:

a first step (141) of segmenting the input signal (140) in the form of a plurality of frames of predetermined length, and extracting at least two description data from the analysis of each frames;

a second step (142) for assigning at least two description data extracted from a set of P adjacent frames selected for the input signal (140) to at least one of the data classes to be modeled by supervised learning implementing at least one predetermined set of training data (146).

Such a method according to the invention also advantageously comprises: a third step (143) for learning a set of predetermined parameters for creating a plurality of state-machine transitions for modeling each of the data classes, starting from at least one corpus (146) predetermined learning data; a fourth step (144) for creating at least N state-transition state automata comprising P states for modeling the first of said at least two data classes, and at least M state-transitions including P states for modeling the second of said at least two data classes, so as to create at least two classes of modeled data representative of the data contained in the input signal;

a fifth step (145) for segmenting and labeling the input stream according to at least the two classes of modeled data, starting from at least N and M state - transition state - machine; to segment the input signal (140) into segments (147) of differentiated and categorized data.

Preferentially, the parameters N and M are chosen such that: N> 1 and / or M> 1.

Preferably, the third learning step (143) comprises at least:

a substep (142i) for creating a cloud (80) of points labeled and distributed according to at least the two classes (81), 82) of data to be modeled, in a space of at least two dimensions (83) , (84) respectively defined by at least two data (85), (86) calculated of description of each of the extracted frames; and

a substep (142 ₂ ) for selecting a first subset of labeled cloud points belonging to the first of the two classes to be modeled and a second subset of labeled points belonging to the second of the two classes; modeling, the first and second subsets of points being selected from the points of the cloud (80) of labeled points located in the vicinity of the boundary (87) separating the two classes (81), (82) to be modeled. Advantageously, the sub-step (142i) of creating a cloud (80) of points labeled and distributed according to at least the two classes (81), (82) of data to be modeled is a substep:

- creating a histogram (90) obtained by discretizing the space to at least two dimensions, by means of a step (91) of discretization previously determined according to at least one criterion of accuracy and,

- Labeling (9I ₁ ) (92 ₂ ) (92 ₃ ) (92 ₄ ) (92 ₅ ) (92 ₆ ) of each box (93) of the histogram (90) by at least one of the two classes ( 81) (82) with the highest number of points in the box. Preferably, the sub-step (142 ₂ ) of selection is a sub-step of pruning (101) points (102) of the cloud (80) not being located in the vicinity (87) of the boundary separating the least two classes (81) (82) to be modeled. The pruning step (101) then consists in keeping in each of the cells of the histogram only the points (103) belonging to at least one of the two classes (81) 82) and comprising at least one point in its immediate neighborhood (104) belonging to another of said at least two classes, as illustrated in Figure 10.

Thus, thanks to the learning method described in this invention, the dispersion of the points is much less important, as illustrated in Figure 8.3, since we focus here on the boundary between the two classes.

The proposed invention is therefore not content, as in prior art techniques, with modeling each class of data separately, but jointly, by focusing more on the Gaussian class boundaries, without having to model the interior any more. of the class but only its periphery according to the adjacent classes, which saves significantly in computation time and quality of result.

Preferably, the fifth step (145) of segmentation and labeling of the input stream implements a Viterbi algorithm (120) to determine the optimal alignment of the shape of the input stream (121) on the models formed by state-based automata - transitions.

Advantageously, each of the two classes (81) (82) of data (at least) is modeled by a plurality of state-transitions (110) (111) of the hidden Markov chains type, each state of which consists of at least one Gaussian (HO ₁ , ..., HOi) (H l ₁ , ..., 11 I _j ) and of which each transition (112) defines a level of probability of crossing a current state to another state or to the same state of a hidden Markov chains. Also preferentially, the input signal (60) being in the form of a continuous stream of undifferentiated input data of the speech and music type, the two modeled classes are respectively of the speech class and music class type, respectively. at least two description data then being of the average type (61) and variance (62) calculated from a plurality of instantaneous spectral streams (63) respectively associated with each of the frames of the plurality of frames of predetermined length, the streams (63) spectral being derived from the analysis of each of the frames.

Advantageously, at least one Gaussian is formed of a single pair (average, variance) (61, 62) calculated for all the P frames of predetermined length.

Also advantageously, each of the hidden Markov chains contains at least one re-looping state.

The invention also advantageously relates to a computer program product downloadable from a communication network and / or recorded on a computer readable medium and / or executable by a processor.

According to the invention, such a program advantageously comprises program code instructions for the execution of the steps of the method, as mentioned above, of segmentation and labeling of the content of an input signal in the form of a continuous stream of undifferentiated input data according to at least two classes (81) (82) of predefined data, when said program is run on a computer.

The invention also advantageously relates to a device for segmenting and labeling the content of an input signal in the form of a continuous stream of undifferentiated input data, according to at least two predefined data classes, comprising: means for segmenting the input signal in the form of a plurality of frames of predetermined length, and for extracting at least two description data from the analysis of each of the frames;

means for assigning the at least two description data, extracted from a set of P adjacent frames selected for the input signal, to at least one of the data classes to be modeled, by means of a training supervised implementing at least a predetermined set of learning data. According to the invention, such a device further comprises: means for learning a set of predetermined parameters for creating a plurality of state-machine-modeling transitions of each of the data classes, starting from at least one predetermined training data corpus;

means for creating at least N state-transition automata comprising P states for modeling the first at least two data classes, and at least M state-transition automatons comprising P states for the modeling of the at least one of the two classes of data, so as to create at least two classes of modeled data representative of the data contained in the input signal; means of segmentation and labeling of the input stream according to at least the two classes of modeled data, starting from the at least N and M state-transition automata; to segment the input signal into differentiated and categorized data segments. Finally, FIG. 13 shows the structure of a segmentation and labeling device according to the invention, which comprises a memory M (130), and a processing unit (131) equipped with a microprocessor μP, which is controlled by the Computer program Pg 132. The processing unit 131 receives (133) as input a stream (134) of undifferentiated data, from which the microprocessor μP realizes, according to the instructions of the program Pg 132, a segmentation and a labeling of the flow (134) of input data using Markov strings concealed, so as to obtain an input signal (134) segmented into segments (135) of differentiated and categorized data, for example according to the classes C1 and C2.

Such a device further comprises all the structural means for implementing the method of segmentation and labeling of an input signal, as mentioned above, which are not detailed here.

5. List of figures

Other features and advantages of the invention will appear more clearly on reading the following description of a preferred embodiment of the invention, given by way of illustrative and nonlimiting example, with reference to the appended drawings. which :

FIG. 1, already described in the description of the prior art, gives an example of segmentation of an audio stream into two Word / Music classes;

FIG. 2, also already described in relation to the prior art, gives an example of a topology of a hidden Markov model with three states; FIG. 3, already described above, presents a flowchart of the method for generating models of two data classes known from the prior art;

FIG. 4, also already described, gives an example of hidden Markov chains obtained for two classes of data, at the output of the method of FIG. 3;

FIG. 5, already already discussed above, illustrates the general principle adopted for the segmentation and recognition of classified data by application of the Viterbi algorithm on hidden Markov chain models, according to the prior art; FIG. 6 presents the principle of extracting the descriptors of the frames of the input signal by calculating the mean and variance of the spectral flux, according to the invention;

FIG. 7 gives an example of the distribution of the means and variances obtained for the "music" and "speech" classes with the method according to the invention;

FIGS. 8.1 to 8.3 respectively show an example of a Gaussian cloud labeled for two speech and music classes, the distribution of Gaussian results resulting from the application of the VQ + EM algorithm according to the prior art and, the distribution of the Gaussians obtained for this same cloud of points, by means of the method according to the invention;

FIG. 9 illustrates the principle of creating the Gaussian distribution histogram by discretization of the space 2, according to the invention;

FIG. 10 describes the principle of pruning the points of the Gaussian cloud labeled by detection of the immediate neighbors;

FIG. 11 gives an example of creation of the Markov chains for the Word and Music models, by means of the method according to the invention; - Figure 12 recalls the general principle of operation of the method according to the invention;

FIG. 13, already described above, shows the structure of a segmentation and recognition device according to the invention.

DESCRIPTION OF A PREFERRED EMBODIMENT OF THE INVENTION The present invention thus relates to a method for segmenting and labeling the content of an input signal (141) in the form of a continuous data stream. undifferentiated input, according to at least two predefined data classes, the different steps of which are described below for a preferred embodiment of the invention. In the following, we will explain each part by illustrating by an example of use in the context of a speech / music segmentation system based on the calculation of the spectral flow in an input audio signal.

A first step of the method according to the invention relates to the segmentation of the audio signal into frames with or without overlap and the extraction of one or more descriptors for each of the frames.

Preferably, a calculation of the average and the variance of the descriptor (s) extracted on a set of P adjacent frames is then carried out.

Other parameters calculated on the descriptors can of course be considered. The audio stream is preferentially segmented into segments of 32 milliseconds and for each of these segments the spectral stream corresponds to the frequency "bin" sum by "bin" frequency of the absolute value of the difference between two consecutive amplitude spectra.

From the instantaneous values of the spectral flux, we derive the mean and the variance on segments of 1 second (32 values of instantaneous spectral flux). The pairs (average, variance) thus obtained are represented by way of illustration through FIG.

In a second step of the method according to the invention, an assignment of the calculated descriptor (s) for each of the selected frames to two classes of data to be modeled, for example of the speech class and music class type, is carried out, when one seeks to distinguish between speech data and music data of an audio input signal.

This second step is done through supervised learning, based on at least one audio learning corpus containing real data. The training data used are files containing the digital data relating to the parameter of the spectral stream, calculated from the audio files.

Once all the averages and variances of the flow computed, we assign these descriptors to the 2 classes to be modeled. The body of learning here consists of two hours of music and two hours of speech. All the descriptors extracted from the music data are thus assigned to the "music" class and all the descriptors extracted from the speech data are assigned to the "speech" class.

A third step of the method according to the invention relates to the learning of the different parameters of the states of hidden Markov chains from the training data. This learning phase takes place in two sub-steps, namely:

- the creation of a histogram by discretization of the 2D space

(average and variance) by applying a discretization step chosen arbitrarily or according to at least one precision criterion previously determined, then labeling by majority vote of each of the cells of the histogram by comparing the number of pairs (average, variance) of each of the classes belonging to this box. A cloud of labeled points is thus obtained, as illustrated in FIG. 9; then

the selection of N relevant pairs (mean, variance) of the first class and of M relevant pairs of the second class of the preceding labeled point cloud. The (N + M) relevant pairs here designate the points close to the boundary between the two classes.

Optionally, the aforementioned step of selecting the relevant pairs can be carried out by the implementation of an algorithm called "pruning", the principle of which consists in keeping only the points which comprise at least one point of a another class in its immediate vicinity.

The purpose of this pruning procedure is to remove Gaussians that are considered unnecessary (that is, far from the border) in the two-dimensional space considered here. It is also in this part that the number of Gaussians to obtain the desired number of couples can be limited to save in computing time, when necessary.

In this example, we keep M = N = 64 Gaussians to model the class "speech" and the class "music". The resultant pairs are shown at the bottom of Figure 8.3.

A fourth step of the method according to the invention is then aimed at creating N hidden Markov chains at P states that model the Cl class, the P states of each of the hidden Markov chains consisting of one of the Gaussian Ns (mean torque, variance). determined in the previous step and to create M hidden Markov chains at P states modeling the C2 class, the P states of each hidden Markov chains being made up of one of the Gaussian Ms (mean torque, variance) also determined during from the previous step.

Optionally, it is possible to construct only N '(l≤N'≤ N) hidden Markov chains to model class 1 and M' (l≤M'≤ M) hidden Markov chains to model class 2.

All states of the hidden Markov chains of class 1 then consist of mixtures of the Gaussian Ns selected at the selection stage and all the states of the hidden Markov chains of class 2 consisting of mixtures of the Gaussian M selected at the selection stage as well.

Each of the classes "speech" and "music" is modeled respectively by 64 and 64 Markov chains hidden at P = 32 states, in the example presented.

The P states of the same hidden Markov chain consisting of each Gaussian determined in the third step.

For each class, from each Gaussian G [i] (l≤i <64) conserved, we create a hidden Markov chain containing as many re-looping states as elements used in the calculation of means and variances.

All the states of this chain emit their probability according to the law of the Gaussian G [i]. An example of creating 64 Markov chains for the "speech" model and 64 Markov chains for the "music" model, as illustrated in FIG. 11. Finally, a fifth step of the method according to the invention consists in carrying out the Segmentation and labeling of the audio stream using the hidden Markov strings created in the step of creating hidden Markov chains, using an algorithm to determine the optimal alignment of an acoustic form on a model of Markov. The most suitable algorithm is the Viterbi algorithm.

This procedure is illustrated in Figure 12 where C1 and C2 designate the speech and music classes.

To test and validate the significant improvement in performance of this new segmentation and labeling process according to previously modeled data classes, a learning corpus composed of two hours of instrumental music indexed manually in "music" and two hours of manually indexed speech in "speech" was used.

Another test corpus, consisting of two hours of generic music (instruments + sung voice), twenty minutes of instrumental music (different from that used during learning) and twenty minutes of speech

(different from that used during learning) was also used. The performances obtained in terms of method execution according to the invention were calculated by measuring, on the body of tests, the poorly indexed temporal segments with respect to the total time of the sequences, on the basis of a calculated error rate. as following :

_{: err} (spββch) + t _error (mUSÎc)

Error = 100 * -

• total where t _error {speech) represents the total duration of segments indexed "speech" on segments corresponding in fact to music and where t _aror (music) represents the total duration of segments indexed "music" on segments corresponding to made to speech.

The following table describes the two systems tested.

Thanks to this new segmentation process in two classes, the performance is thus significantly improved, relative to the techniques known from the prior art.

Claims

A method of segmenting and labeling the contents of an input signal (141) in the form of a continuous stream of undifferentiated input data, according to at least two predefined data classes, comprising: a first step (141) of segmenting said input signal (140) in the form of a plurality of frames of predetermined length, and extracting at least two description data from the analysis of each said frames; a second step (142) for assigning said at least two description data extracted from a set of P adjacent frames selected for said input signal (140) to at least one of said data classes to be modeled, by supervised learning implementing at least one predetermined set of training data (146); characterized in that it further comprises: a third step (143) of learning a set of predetermined parameters for creating a plurality of state-transition state machine-transitions of each of said data classes, from said at least one corpus (146) of predetermined learning data, said third learning step (143) comprising at least: a substep (142i) for creating a cloud (80) of points labeled and distributed according to the said at least two classes (81), 82) of data to be modeled, in a space of at least two dimensions (83) , (84) respectively defined by at least two calculated data (85), (86) of description of each of said extracted frames; a substep (142 ₂ ) for selecting a first subset of points of said labeled cloud belonging to the first of said at least two classes to be modeled and a second subset of labeled points belonging to the second of said classes; at least two classes to be modeled, said first and second subsets of points being selected from the points of said cloud (80) of labeled points located in the vicinity of the boundary (87) separating said at least two classes (81), (82) to be modeled; a fourth step (144) for creating at least N state-transition state automata comprising P states for modeling the first of said at least two data classes, and at least M state-transitions including P states for modeling the second of said at least two data classes, so as to create at least two classes of modeled data representative of the data contained in said input signal; a fifth step (145) of segmentation and labeling of said input stream according to said at least two classes of modeled data, based on said at least N and M state-transitioned automata; to segment said input signal (140) into differentiated and categorized data segments (147). 2. A method of segmentation and labeling of the content of an input signal according to claim 1, characterized in that said substep (142i) for creating a cloud (80) labeled points and distributed according to said at at least two classes (81), (82) of data to be modeled is a substep of: - creating a histogram (90) obtained by discretizing said space to at least two dimensions, by means of a step (91) of discretization previously determined according to at least one criterion of accuracy and, - labeling (92χ) (92 ₂ ) (92 ₃ ) (92 ₄ ) (92 ₅ ) (92 ₆ ) of each box (93) of said histogram (90) by one of said at least two classes (81) ( 82) having the greatest number of points in said box considered. 3. Segmentation and labeling process of the content of an input signal according to any one of claims 1 and 2, characterized in that said substep (142 ₂ ) selection is a substep of pruning (101) points (102) of said cloud (80) not being located in the vicinity (87) of the boundary separating said at least two classes (81) (82) to be modeled, said pruning step (lO1) consisting of in only one of said at least two classes (81) having at least one point in its immediate neighborhood (104) belonging to another of said at least two classes. 4. Segmentation and labeling process of the content of an input signal according to any one of claims 1 to 3, characterized in that said fifth step (145) of segmentation and labeling of said input stream A Viterbi algorithm (120) is used to determine the optimal alignment of the shape of said input stream (121) on models formed by said state - transition state machines. 5. Segmentation and labeling process of the content of an input signal according to any one of claims 1 to 4, characterized in that each of said at least two classes (81) (82) of data is modeled by a plurality of state - transitions (110) (111) of the hidden Markov chains type, each state of which consists of at least one Gaussian (HO ₁ , ..., HOi) (H l ₁ , ... , lll _j ) and each transition (112) defines a level of probability of crossing a current state to another state or to the same state of one of said hidden Markov chains. 6. A method of segmentation and labeling of the content of an input signal according to any one of claims 1 to 5, characterized in that said input signal (60) is in the form of a continuous stream. of undifferentiated input data of the speech and music type, said at least two modeled classes are respectively of the speech class and music class type, and in that said at least two description data are of the average type (61) and the variance (62). ) calculated from a plurality of instantaneous spectral streams (63) respectively associated with each of said frames of said plurality of frames of predetermined length, said spectral streams (63) being derived from said analysis of each of said frames. 7. A method of segmentation and labeling of the content of an input signal according to claims 5 and 6, characterized in that said at least one Gaussian is formed of a single pair (mean, variance) (61, 62) calculated for all of said P frames of predetermined length. Process for segmentation and labeling of the content of an input signal according to any one of Claims 5 to 7, characterized in that each of said hidden Markov chains contains at least one loopback state. 9. Computer program product downloadable from a communication network and / or recorded on a computer readable medium and / or executable by a processor characterized in that it comprises program code instructions for the execution of at at least one of the steps of the process of segmenting and labeling the contents of an input signal in the form of a continuous stream of undifferentiated input data, following at least two classes (81) (82) of data predefined, according to any one of claims 1 to 8, when said program is executed on a computer. Apparatus for segmentation and labeling of the content of an input signal in the form of a continuous stream of undifferentiated input data, according to at least two predefined data classes, comprising: segmentation means said input signal in the form of a plurality of frames of predetermined length, and extracting at least two description data from the analysis of each of said frames; means for assigning said at least two description data extracted from a set of P adjacent frames selected for said input signal to at least one of said data classes to be modeled, by means of supervised learning implementing at least one predetermined set of learning data; characterized in that it further comprises: - means for learning a set of predetermined parameters for creating a plurality of state machines - model transitions of each of said data classes, from said at least one a predetermined set of training data, said learning means comprising at least: means for creating a cloud of points labeled and distributed according to said at least two data classes to be modeled, in a space with at least two dimensions respectively defined by at least two computed description data of each of said extracted frames; means for selecting a first subset of points of said labeled cloud belonging to the first of said at least two classes to be modeled and of a second subset of labeled points belonging to the second of said at least two classes to be modeled said first and second subsets of points being selected from the points of said cloud of labeled points located in the vicinity of the boundary separating said at least two classes to be modeled; means for creating at least N state-transition automata comprising P states for modeling the first of said at least two data classes, and at least M state-transition automata comprising P states for the modeling of the second of the at least two classes of data, so as to create at least two classes of modeled data representative of the data contained in said input signal; means for segmenting and labeling said input stream according to said at least two classes of modeled data, based on said at least N and M state-transitioned automata; to segment said input signal into differentiated and categorized data segments.