EP2005756A2

EP2005756A2 - Restrained vector quantisation

Info

Publication number: EP2005756A2
Application number: EP07731724A
Authority: EP
Inventors: Stéphane RAGOT; Cyril Guillaume
Original assignee: France Telecom SA
Current assignee: Orange SA
Priority date: 2006-03-21
Filing date: 2007-03-09
Publication date: 2008-12-24
Also published as: WO2007107659A2; KR20090005027A; US8285544B2; KR101370018B1; JP4981122B2; JP2009530940A; US20100228808A1; CN101467459B; WO2007107659A3; CN101467459A

Abstract

The invention relates to a method for generating a dictionary for the vector quantisation of a signal. Said method comprises a step (10) of the statistical analysis of driving vectors representing the signal determining a finished set of code vectors (CITER) representing said driving vectors. The inventive method is characterised in that it also comprises a step (3) of modifying the finished set of code vectors in order to impose a minimum distance between the code vectors modified two-by-two, the set of said modified code vectors forming the dictionary.

Description

QUANTIFICATION VECTORIAL CONSTRAINT

The present invention relates to quantization dictionaries for coding and decoding signals such as audio, video signals, and more generally multimedia signals, for storage or transmission.

More precisely, the present invention proposes a solution to the problem of unstructured statistical vector quantization integrating a priori knowledge of a distortion threshold.

A very widespread solution for the processing of particularly digital signals is the vector quantization (QV) which, in general, represents an input vector by a vector of the same dimension chosen in a finite set. This finite set is called the reproduction alphabet or dictionary or directory, and its elements are code vectors, code words, exit points, or representatives. In vector quantization, a block of n samples of a signal is treated as a vector of dimension n. The vector is encoded by choosing, in a finite dictionary, the code vector that "most" resembles it. For example, an exhaustive search is made among all elements of the dictionary to select the dictionary element that minimizes a measure of distance between that element and the input vector.

In general, vector quantization dictionaries are designed from the analysis of samples of the signal to be coded forming a training sequence, using statistical methods such as generalized Lloyd-Max algorithms (GLA: Generalized Lloyd-Max Algorithm) or LBG (for Linde, Buzo and Gray). Such a type of algorithm is described for example in Gersho A., Gray R. M, "Vector Quantization and Signal Compression", Kluwer Academy Publishers, 1992.

From the training sequence, and from an initial dictionary, the dictionary is built iteratively. Each iteration includes a step of classifying the training vectors to build quantization regions of the training sequence according to the rule of closest neighbor, and a step of improving the dictionary by replacing the old code vectors by the centroids of the regions, according to the so-called centroid rule.

To avoid convergence towards a local minimum of this deterministic iterative algorithm, so-called stochastic relaxation variants (SKA: Stochastic K-means algorithm) have been proposed by introducing a random part in the centroid construction step and / or in that of class construction. Such a modification is described for example in Kovesi, B .; Saoudi, S .; Boucher, J. M .; Reguly, Z., "A robust robust stochastic algorithm for vector quantizer design for nonstationary channels," IEEE International Conference on Acoustics, Speech, and Signal Processing, Vol. 1, pp. 269-272, 1995.

The dictionaries, or statistical vector quantizers, thus obtained have no particular structure, which makes their exploration expensive in calculations and greedy in memory. In fact, the use, particularly in real time, of such dictionaries obtained by conventional statistical vector quantization is often limited to small coding and / or low bit rates of the order of 4 to 10 bits per vector.

In general, statistical vector quantization exploits the shape of the probability distribution of the signal to be processed and this is reflected in the distribution of code vectors of the dictionary of a statistical vector quantizer. The construction algorithms used in vector quantization are aimed at minimizing mean distortion; they therefore tend to place a lot of code vectors in high density areas. This concentration of centroids in areas of high density of the signal to be encoded is not always relevant from a perceptual point of view. Thus, it is possible that two code vectors of a zone of high density are so close that the distortion generated by the substitution of one by the other is not perceptible. Moreover, for low-resolution quantizers, that is to say with few code vectors, this concentration is done to the detriment of zones less dense of the probability distribution of the signal to be encoded which are poorly represented, that is to say poorly coded.

To overcome the constraints of size and dimension, several statistical vector quantization variants were developed, in particular to try to remedy the lack of structure of the dictionary and thus reduce complexity at the expense of quality. However, the performance-complexity trade-off is improved, which increases the range of resolutions and / or dimensions over which vector quantization can be effectively applied. Many constrained vector quantizer schemes have been proposed in the literature. We can cite in particular the tree vector quantizer, the multi-stage vector quantizer, the vector quantizer "cartesian product" or the vector quantizer "gain / orientation", a particular case of the Cartesian product vector quantizer. As in unstructured vector quantization, the dictionaries obtained by constrained methods always have a large amount of code vectors close to each other and in particular code vectors whose differences are not discernible by the receiver. This leads to a non-optimal quality. Another approach, called algebraic vector quantization, has also been proposed. Algebraic vector quantization uses strongly structured dictionaries, derived from regular networks of points or error correcting codes. A simple example of a regular network is the cubic network, formed for example of a set of points with integer coordinates. Thanks to the algebraic properties of their dictionaries, the algebraic vector quantizers are simple to implement and do not have to be stored in memory. The exploitation of the regular structure of these dictionaries allows the development of optimal and fast search algorithms and mechanisms for associating the index with a corresponding code vector, for example by a formula. The algebraic vector quantization techniques developed in the state of the art lead to dictionaries that are subsets of regular networks.

One of the main parameters of a regular network is its minimum quadratic distance between two points on which the code vectors are placed. To facilitate indexing, the dictionary is generally chosen as the intersection of the network with a regular polytope, such as a hypercube or hypersphere, or with its surface. The points outside this intersection are then outliers. For a given bit rate, that is to say a fixed number of points of the intersection, limiting the number of outliers leads to increasing the minimum distance between two code vectors.

While algebraic vector quantizers are less complex to implement and require less memory, they are only optimal for encoding signals with uniform distribution. It should also be noted that the operations of indexing the code vectors, or numbering, and the reverse decoding operations require more calculations than in the case of statistical vector quantizers, for which these operations are performed by simple reads of tables.

Finally, another form of vector quantization combines algebraic quantization and statistical quantization. This is the case of the fine-coarse vector quantizer "Fine-Coarse Vector Quantization" described in the publication "Moayeri, N., Neuhoff D. L,; Stark, WE, "Fine-coarse vector quantization" IEEE Trans. on Information Theory, Vol. 37, Issue 4, pp. 1503-1515, July 1991 ". This quantization comprises two quantifiers. The first high-resolution, or fine-resolution, quantizer is generally well structured and uncomplicated to implement, while the second, coarse-resolution quantizer is an unstructured statistical vector quantizer. The first floor serves as a pre-calculation to speed up search in the second floor. The quantizer "Fine-coarse" suffers from the same disadvantages as those indicated above for the statistical vector quantization. Indeed, the link between the fine-resolution quantizer and the quantizer Coarse resolution is a simple look-up table and therefore, even if the high-resolution vector quantizer is a finite subset of a regular network with a minimal distance between code vectors, this constraint is not imposed on the vector quantizer with coarse resolution that is constructed by conventional techniques of statistical vector quantization.

It thus appears that statistical vector quantization results in dictionaries whose distribution of points is poorly adapted, some of these points being redundant because they are indistinguishable at the level of the receiver. Conversely, the algebraic vector quantization results in dictionaries whose distribution is suitable only for signals with a uniform probability distribution and results in degrading the quality of the coding for the other signals.

The invention has the particular advantage of defining a method for obtaining a dictionary by statistical vector quantization adapted to the signal to be encoded and taking into account the perceptual thresholds.

To this end, the subject of the invention is a method for generating a vector quantization dictionary of a signal, of the type comprising a step of statistical analysis of training vectors representative of the signal determining a finite set of vectors of a signal. code representing said drive vectors, characterized in that the method further comprises a step of modifying said finite set of code vectors to impose a minimum distance between the modified code vectors two by two, these modified vectors forming said dictionary. Thus, the invention results in a dictionary adapted to the signal because of a step of statistical analysis of the drive vectors but with a minimum distance between the code vectors, which makes it possible to take into account the fact that vectors too close code are indistinguishable at the level of perception. According to other characteristics of the invention, the statistical analysis step comprises a sub-step of creating an initial dictionary, a substep of classification of the training vectors from the initial dictionary for forming quantization regions and a sub-step for determining a centroid for each region, said centroids being said code vectors. Such an embodiment corresponds to a particular implementation of statistical analysis. Advantageously, said substep of creating an initial dictionary is made from said training vectors and comprises replacing each training vector with a rounded vector chosen on a regular network of points and the selective elimination of vectors. rounded according to their frequency of appearance. In a particular embodiment of the invention, the statistical analysis step further comprises a sub-step of calculating the distortion of the initial dictionary and a substep of comparing this distortion with a tolerance threshold, said statistical analysis step being repeated if said comparison is negative. Such an iterative approach makes it possible to obtain an initial dictionary of better quality.

Alternatively, the modifying step includes replacing each of said code vectors with a neighbor code vector selected on a regular network of points. Such an embodiment makes it possible to ensure a minimum distance between the code vectors. Advantageously, the method further comprises a step of eliminating duplicates among said modified code vectors, which has the effect of reducing the size of the dictionary.

In a particular embodiment, said duplicate elimination step includes a sub-step of searching for duplicates by comparing the modified code vectors with each other and a sub-step of deleting duplicates.

Advantageously, the method comprises a step of merging the modified code vectors with the drive vectors having at least said minimum distance between them and with said modified code vectors in order to improve the distribution of the code vectors of the dictionary.

In a particular embodiment, this melting step includes a substep of replacing each training vector with a rounded vector chosen over a regular network of points and a substep of combining modified code vectors and rounded training vectors. Thus, the rounded training vectors that are combined with the modified code vectors also have a minimum distance of two to two.

In yet another variation, said merging step further comprises scheduling the rounded training vectors according to their occurrence frequency and selecting a finite number of these vectors to combine them with the modified code vectors. In one embodiment, said combining sub-step includes adding to said modified code vectors a finite number of rounded training vectors separate from said modified code vectors.

In one variant, the regular network of points has a step corresponding to a sensitivity level of a recipient receiver of said signal, which makes it possible to impose a minimum distance between two code vectors corresponding to a perceptual reality.

In yet another embodiment, the method further comprises an optimization phase of said dictionary.

For example, the optimization phase comprises a step of calculating the distortion of the dictionary and a step of comparing this distortion with a tolerance threshold, said steps of statistical analysis and modification being repeated if said comparison is negative.

Alternatively, the optimization phase comprises at least one iteration of the method implemented from the dictionary formed of said modified code vectors.

Moreover, the subject of the invention is also a computer program comprising code instructions for implementing the method as mentioned above, as well as a device comprising means such as, for example, a working memory and a device. processor, to execute this computer program and thus generate a vector quantization dictionary within the meaning of the invention. The present invention also relates to encoding and decoding methods, and a decoder using a dictionary obtained according to the invention.

The invention will be better understood in the light of the description which follows, given by way of non-limiting example with reference to the appended figures in which:

FIG. 1 is a flowchart of the method of the invention;

FIGS. 2 and 3 are representations of a dictionary at different stages of the method of the invention; FIG. 4 is a flowchart of an optimization method;

- Figures 5 and 6 are block diagrams of an encoder and a decoder implementing the dictionary obtained by the method of the invention.

With reference to FIG. 1, a method for obtaining a vector quantization dictionary according to the invention will now be described.

In the embodiment described, this method is implemented on training vectors representative of a digitized audiophonic signal such as a speech signal.

This method firstly comprises a step 10 of statistical analysis of the training vectors. In the embodiment described, step 10 corresponds to the application of the generalized Lloyd-Max algorithm.

Step 10 begins with a sub-step 12 of creating an initial dictionary marked CO. This dictionary is created from the analysis of the training vectors or by other means known per se such as a random selection of a finite number of the training vectors, or so-called "splitting" algorithms, "LBG algorithm", or other. This initial dictionary CO is of a determined size.

In the embodiment described, the method then comprises a substep 14 for initializing common variables and in particular for initializing an ITER variable corresponding to an iteration number and a DITER variable corresponding to a value of distortion. Step 10 then comprises a substep 16 of classification of the training vectors with respect to the initial dictionary CO to form classification regions and a sub-step 18 of determining a centroid for each region. More precisely, during the sub-step 16 of classification, for each training vector, the whole of the current dictionary, that is to say of the CO dictionary modified by successive iterations, is scanned, and one chooses the dictionary code vector minimizing quadratic error with drive vector. Sub-step 18 is performed as follows: for each class or region Vi, the new centroids are calculated.

We define Vi the set of training vectors xj satisfying Q (xj) = yi, where Q () is the quantization function and yi is a code vector of the dictionary. Let Ni be the number of vectors of Vi, and Vi (j, k) be the kth element of the jth vector of Vi.

We have, for the kth element of the centroid yi:

In a substep 20, a distortion value for the current dictionary is calculated.

During this substep 20, the drive vectors are quantized with the current dictionary provided by the preceding steps. The quadratic error between the drive vectors and their quantized version is the distortion measure. The substeps 16, 18 and 20 are then repeated iteratively until a maximum number of ITERMAX iterations is reached or the distortion is less than a tolerance threshold, the tests being carried out at a later time. sub-step 22.

As indicated above, in the example, this step 10, in particular with the repetition of substeps 16, 18, 20 and 22, corresponds to the application of the generalized Lloyd-Max algorithm. At the end of step 10, the method thus delivers a CITER dictionary formed by an iterative statistical analysis of the training vectors.

A two-dimensional representation of the CITER dictionary is given in Figure 2 in which the code vectors are represented by a scatter plot. In this representation, the gray dots correspond to the training vectors or vectors of the signal and the black dots to the vectors of the dictionary. It can be seen that the distribution of the CITER dictionary vectors follows the density of the drive vectors and that many vectors are close to each other. The method then comprises a step of modifying the CITER dictionary to impose a minimum distance between the two-to-two code vectors.

This modification step comprises the replacement of each of the code vectors by a neighboring vector chosen on a regular network of points such as a Cartesian grid. These modified vectors are also called rounded vectors. In the embodiment described, the neighboring vector chosen is the vector placed on the regular network closest to the code vector. Depending on the embodiments and types of vectors, criteria other than the distance between the vectors may be retained to determine the neighboring vector of a code vector.

Thus, step 30 delivers a new dictionary C obtained by rounding the CITER dictionary obtained in step 10, on the regular network.

The rounding of the code words on the regular network is done in the following way, considering the pace of the network:

where i = 1 ... M, k = 1 ... n and [.jindicates the rounding to the nearest integer. The method then comprises a step 32 of eliminating duplicates among the modified code vectors forming the dictionary C.

Step 32 comprises a substep of searching for duplicates performed on the set of dictionary C by comparing each of the vectors code to the other code vectors of the dictionary and then a sub-step of deleting duplicates.

At the end of this step 32, the dictionary C may comprise fewer code vectors than the CITER dictionary. The method then comprises a step 34 of melting the modified dictionary with the rounded training vectors in order to complete the dictionary if its size has been reduced.

This step 34 begins with a substep 36 of replacing each training vector with a neighboring vector, called a rounded training vector, chosen on a regular network of points. Advantageously, the same network as that of step 30 is used. The rounded training vectors are then ordered according to their frequency of occurrence in descending order, advantageously after eliminating the duplicates. The method then comprises a sub-step 38 of combining the dictionary C of modified code vectors with the rounded training vectors. During this sub-step, the dictionary C is completed by the addition of the rounded training vectors not present in the dictionary, advantageously following the order of decreasing occurrence frequencies.

Thus, the dictionary C "is constructed in such a way that a minimum distance is guaranteed between the centroids forming the code vectors of the dictionary so that it is called constraint, In the example, the parameters to be quantified are taken from a logarithmic scale, a constant step in this domain is equivalent to a logarithmic step in the field of energies, which corresponds to the sensitivity of the human ear.by choosing the pitch p as a function of a perceptual threshold, it is therefore possible by means of the method of the invention, generating a statistical vector quantization dictionary taking into account the perceptual thresholds Moreover, by combining the dictionary with the rounded training vectors, a better distribution of the code vectors is obtained. FIG. 3 shows the dictionary C "in the form of a two-dimensional point cloud. It will be noted in FIG. 3 that the dictionary C" obtained according to the method described above covers a large area of the drive vectors. and has a minimum distance between the two-to-two code vectors.

The number of apparent points of the dictionary in FIG. 3 is smaller than that of FIG. 2 even if the two dictionaries actually have the same number of code vectors. Indeed, Figures 2 and 3 show only the two-dimensional projection of vectors of larger size. Since the code vectors of the dictionary illustrated in FIG. 3 are rounded on a regular network, several vectors are reduced to a single point, which is not the case for the dictionary illustrated in FIG.

Advantageously, the method further comprises a phase 40 for optimizing the dictionary C "whose flowchart is shown with reference to FIG. 4.

The optimization starts with the reception of the dictionary C "as obtained at the end of the process described with reference to FIGS. 1 to 3.

This optimization phase consists, in the embodiment described, in taking over the same processing steps by applying them to the code vectors of the dictionary C.sub.H However, in the optimization phase described, at each iteration, the classification , the determination of the centroids, the modification of the code vectors and the elimination of the duplicates are carried out.

Thus, the phase 40 begins with a classification step 42 applied to the code vectors of the dictionary C "followed by a step 44 of determining the centroids.These steps 42 and 44 are similar to the steps 16 and 18 described above.

Subsequently, a step 46 of modification by rounding on the regular network is performed. This step 46 is followed by a step 48 of replacing the duplicates. Finally, the optimization phase comprises a step 50 of measuring the distortion and a test 52 to determine if the distortion is less than a tolerance threshold or if a maximum number of iterations has been reached. Such an optimization phase makes it possible to further improve the quality of the dictionary.

Of course, other embodiments are possible.

Alternatively, the initial dictionary is created by pruning rounded training vectors on the regular network. That is to say, the creation of this initial dictionary comprises the replacement of the drive vectors by neighboring vectors selected on the regular network and then the selective elimination, according to their frequency of appearance, of the vectors of rounded training. In yet another variant, the optimization phase consists in repeating the method of FIG. 1 identically, using as initial dictionary the dictionary obtained previously.

In another variant, the melting step is implemented between the modified code vectors and the unrounded training vectors. In such an embodiment, however, the drive vectors used are chosen to ensure the maintenance of the minimum distance between the code vectors of the dictionary.

The method of generating a dictionary as described above can be implemented through a program for any type of computer and computer.

With reference to FIGS. 5 and 6, an encoder and a decoder using dictionaries obtained according to the method of the invention are described.

These encoders and decoders are part of a global hierarchical subband audio coding and decoding system that operates at three possible bit rates: 8, 12 or 13.65 Kbit / s. In practice the encoder always operates at the maximum rate of 13.65 kbit / s, while the decoder can receive the heart at 8 kbit / s, and one or two enhancement layers, respectively 12 or 13.65 kbit / s.

The encoder 60 shown in FIG. 5 comprises two channels receiving the input signal. The first channel comprises a filter unit 62 able to extract a low band BF from 0 to 4000 Hz of the signal. The second channel comprises a similar unit 64 capable of extracting a high band HF spreading from 4000 to 8000Hz. These units 62 and 64 comprise, for example, filters and decimation modules made in a conventional manner.

The low band signal is then coded according to a conventional CELP coding such as a CELP encode of 8 to 12 Kbps in a coding unit 66.

The high band signal is coded in particular by using a vector quantization in a unit 70. In the embodiment described, the unit 70 is a parametric coding unit in which the coded parameters are time and frequency envelopes of the corresponding signal respectively. to the root mean square (rms) value per subframe or subband of the high band signal. These envelopes are passed in a logarithmic domain as follows: where σ is the value of the envelope.

For each frame of the input signal of the time envelope and frequency envelope parameters are extracted in parametric analysis modules 72 and 74. These parameters in the logarithmic domain are then jointly quantized in a vector quantization module 76. using dictionaries generated according to the invention.

More precisely, in the embodiment described, the coding of the parameters is carried out by a vector quantization called Cartesian product with suppressed average.

The mean μ of the time envelope vector is calculated and then this average is quantized. The quantized mean is noted μq. The temporal envelope and frequency envelope vectors are then centered on μq before being separately quantized and multiplexed.

The quantization of the mean μ is performed by a uniform scalar quantization whose pitch p = 0.99657 is the perceptual threshold, that is to say the step of the regular network used to generate the dictionary. The Cartesian product vector quantization is performed using the dictionary obtained according to the invention and in which the stored code vectors are words with a minimum distance of p = 0.99657, which corresponds to a step of 6 dB. The two channels HF and BF are then multiplexed in a multiplexer 80 to form the output signal of the encoder 60.

The corresponding decoder 100 is described with reference to FIG.

This decoder firstly comprises a demultiplexer 102 separating the corresponding channels at the low band and high band portions of the signal. The channel corresponding to the low-band signal is introduced into a CELP decoder 104. This CELP decoder also provides excitation parameters calculated in a conventional manner per se.

The channel corresponding to the high band signal is introduced into a decoding unit 1 10 using in particular a reverse vector quantization 1 12.

The output of this module 1 12 is supplied to a temporal shaping module 114. This module 1 14 also receives a synthetic excitation signal generated by a module 116 from the CELP excitation parameters supplied by the decoder CELP 104 The unit 1 10 finally comprises a frequency shaping module

1 18.

Symmetrically to the coding of the parameters, the decoding is carried out by denormalization of the temporal and frequency envelopes dequantized by the dequantized average. Finally, the two paths are processed in processing modules 120 and 122 before being recombined with each other in a mixer 124.

The architecture and operation of such an encoder and decoder are conventional in themselves and will not be described in more detail here. A detailed description is given in ITU-T, COM 16, D214 (WP 3/16), "High level description of the scalable 8-32 kbit / s algorithm submitted to the Qualification Test by Matsushita, Mindspeed and Siemens," Q.10 / 16, Study Period 2005-2008, Geneva, 26 July - 5 August 2005 (Source: Siemens).

Claims

A method of generating a vector quantization dictionary (C ") of a signal, of the type comprising a step (10) of statistical analysis of training vectors representative of the signal determining a finite set of code vectors ( CITER) representatives said drive vectors, characterized in that the method further comprises a step (30) of modifying said finite set of code vectors to impose a minimum distance between the modified code vectors two to two, these vectors of modified code forming said dictionary.

2. Method according to claim 1, characterized in that the step (10) of statistical analysis comprises a sub-step (12) for creating an initial dictionary (CO), a substep (16) classification said drive vectors from the initial dictionary for forming quantization regions and a sub-step (18) for determining a centroid for each region, said centroids being said code vectors.

3. Method according to claim 2, characterized in that said substep (12) for creating an initial dictionary is made from said training vectors and comprises replacing each training vector with a chosen rounded vector. on a regular network of points and the selective elimination of vectors rounded according to their frequency of appearance.

4. Method according to any one of claims 1 to 3, characterized in that the statistical analysis step further comprises a sub-step (20) for calculating the distortion of the initial dictionary and a substep (22). ) of comparing this distortion with a tolerance threshold, said step (10) of statistical analysis being repeated if said comparison is negative.

The method according to any one of claims 1 to 4, characterized in that the modifying step (30) comprises replacing each of said code vectors of the finite set with a neighbor code vector selected on a network. regular points.

6. Method according to any one of claims 1 to 5, characterized in that it further comprises a step (32) for eliminating duplicates among said modified code vectors.

7. Method according to claim 6, characterized in that said step

(32) duplicate elimination comprises a substep of finding duplicates by comparing the modified code vectors with each other and a sub-step of removing duplicates as a function of their occurrence frequency.

8. Method according to any one of claims 1 to 7, characterized in that it further comprises a step (34) of merging the modified code vectors with drive vectors having at least said minimum distance between them and with said modified code vectors.

9. Method according to claim 8, characterized in that the melting step comprises a substep (36) for replacing each driving vector with a rounded vector chosen on a regular network of points and a substep (38). ) combination of the modified code vectors and the rounded training vectors.

The method according to claim 9, characterized in that said melting step (36) further comprises scheduling the rounded entrainment vectors according to their occurrence frequency and selecting a finite number of these vectors for the combine with the modified code vectors.

A method according to any of claims 9 and 10, characterized in that said combining sub-step (38) comprises adding to said modified code vectors a finite number of rounded training vectors distinct from said modified code vectors.

12. Method according to any one of claims 3, 5 or 9, characterized in that said regular network of points has a step corresponding to a sensitivity level of a receiver to said signal.

13. Method according to any one of claims 1 to 12, characterized in that it further comprises a phase (40) for optimizing said dictionary formed of said modified code vectors.

14. Method according to claim 13, characterized in that the optimization phase (40) comprises a step (50) for calculating the distortion of the dictionary and a step (52) for comparing this distortion with a tolerance threshold, said steps of statistical analysis and modification being repeated if said comparison is negative.

15. Method according to claim 13, characterized in that the phase

(40) optimization comprises at least one iteration of the method implemented from the dictionary formed of said modified code vectors.

Vector quantization dictionary (C ") comprising code vectors representing training vectors representative of a signal, characterized in that it is obtained by a method according to any one of claims 1 to 15.

17. Computer program for generating a vector quantization dictionary (C ") of a signal, characterized in that it comprises code instructions for the implementation of a method according to one of any of claims 1 to 15 when executed on a computer of a computer.

18. Device generating a vector quantization dictionary (C "), characterized in that it comprises means for executing the computer program according to claim 17.

19. A method for coding a signal comprising at least one signal analysis step to obtain representative parameters, characterized in that at least one of said representative parameters is coded by vector quantization using a dictionary. generated by a process according to any one of claims 1 to 15.

Signal encoder (60) comprising at least signal analysis means (72, 74) for obtaining representative parameters, characterized in that it further comprises means (76) coding at least one of said representative parameters by vector quantization using a dictionary generated by a method according to any one of claims 1 to 15.

21. A method for decoding a signal comprising at least one step of processing parameters representative of said signal, characterized in that at least one of said representative parameters is decoded by inverse vector quantization using a dictionary generated by a process according to any one of claims 1 to 15.

22. Signal decoder (100) comprising at least means for processing parameters representative of said signal, characterized in that it further comprises means (1 12) for decoding at least one of said representative parameters by inverse vector quantization. using a dictionary generated by a method according to any one of claims 1 to 15.