EP0481895B1 - Method and apparatus for low bit rate transmission of a speech signal using CELP coding - Google Patents

Abstract

The invention relates to a method for low bit rate transmission of a digital speech signal. <??>The coding is effected by linear prediction driven by codes in order to generate a code signal, a wave form being represented by an initial vector (O) of dimension L, from a filter for synthesis by a reference wave form selected from among a dictionary of reference vectors (v), with a minimum deviation criterion min ¦¦x-H.v¦¦<2>, x representing a target vector by perceptual weighting of the initial vector (O). A dictionary (Y) factorised as a product of base vectors yi of n-ary form, corrected by a scale factor gamma i of distribution of the excitation energy, and of a dictionary G(y) of gains gk, are established in order to represent the dictionary of the reference vectors (v), vk,i = gk. gamma i.yi. The criterion is established by calculation of C(gk, gamma i.yi) = 2 gk<x¦H. gamma i.yi>-gk<2> ¦¦H. gamma i.yi¦¦<2> formed by the scalar products and perceptual energies. To the initial vector (O) is attributed the optimal reference vector vk<*>, i<*> = gk<*>.i<*>.yi<*> represented only by the index values k<*>,i<*>. <??>Application to coding and transmission of speech at a low bit rate by ternary or n-ary vectors. <IMAGE>

Description

The invention relates to a method of transmission, at low speed, by CELP coding of a speech signal and to the corresponding system.

The technique of coding speech signals according to the coding process CELP ("Code Excited Linear Prediction") is currently used and has been the subject of numerous works. This technique for coding digital samples representing the speech signal is a hybrid coding technique in which the speech signal is modeled by linear prediction filters and residues of this prediction. In general, the CELP coders, as shown schematically in FIG. 1a and 1b, exhaustively test all the elements of a list of waveforms. The waveform providing the best synthesis of the signal is retained, and its index, or characteristic address, is transmitted to the decoder. This method is called synthesis analysis. The list of waveforms stored in the encoder and the decoder is called a dictionary.
The quality of a CELP coder depends closely on the dictionary chosen, on the method of determination-modeling of the linear prediction filters used, these two parameters constituting two degrees of freedom, not independent, making it possible to adapt a particular CELP coder to the needs of 'a specific application.
Such a CELP coding technique is suitable for low bit rate coding applications (between 4 to 24 kbits / s). For a more detailed description of this type of coding, one can usefully refer to the article entitled "A robust and fast CELP coder at 16 Kbits / s" published by A. LE GUYADER, D. MASSALOUX and F. ZURCHER, CNET Lannion France, in the journal Speech Communication n ° 7, 1988.

Au cours du processus de codage, à chaque vecteur de référence vi est associée une valeur de gain adaptatif gk prise parmi un dictionnaire de valeurs de gain G, ce qui permet, suite à application du gain gk au vecteur vi pour former un vecteur v_k,i, de satisfaire au critère de distorsion minimale précité.Generally speaking, in this type of encoder, decoder, the digital signal to be analyzed, transmitted and reconstructed, is divided into blocks, or frames. Each block comprising L values is considered as a vector of a vector space of dimension L. The current excitation signal constituted by a vector v, read in the dictionary of waveforms, must minimize a criterion of perceptual distortion of the form:
min ∥χ-Hv∥ ² , in which χ denotes a target signal derived from the original signal O to be transmitted after perceptual weighting and H denotes a matrix of dimension LxL of impulse response derived from the product of the transfer functions of the synthesis filter and the perceptual weighting. It will be recalled that the object of the perceptual weighting, relative to the coding noise, analogous to white noise, is to relativize in the frequency domain the contribution of the latter to the signal actually perceived. The matrix H is a triangular matrix of the form:

During the coding process, each reference vector vi is associated with an adaptive gain value gk taken from a dictionary of gain values G, which allows, following application of the gain gk to the vector vi to form a vector v _{k , i} , to satisfy the aforementioned minimum distortion criterion.

In order to reduce the complexity of the calculations, which are very numerous as a function of the dimension L of the vectors and the bit rate of the speech signal, it has been proposed in certain works to use as a reference vector, in order to carry out the excitation signal, vectors whose components only have the values +1, O or -1, the dictionary of vectors then being constituted in the form of a dictionary of ternary vectors. Such use in a CELP-type coding process of ternary vectors of this type was mentioned in European patent application EP 0 347 307 published on December 20, 1989.
However, in such a coding process, it will be noted that all the reference vectors necessarily contain the same energy. In addition, the search for the optimum reference sequence or vector can only be reduced to the calculation of purely scalar products if the autocorrelation is itself normalized and presents null terms whose spacing corresponds to the components non-zero sequences or reference vectors.

Such an operating mode therefore does not make it possible to take into account, as a reference vector, the totality of the possibilities of combinations of the ternary values of the components of the reference vectors, the minimization of the distortion criterion not being able in all cases to be optimal.

Finally, document EP-A-0 379 296 describes a method of transmitting a low-speed speech signal, of the CELP type. This method implements a process for calculating the minimum quadratic deviation of an initial vector and a reference vector from a vector dictionary factorized into the product of two dictionaries of basic and gain vectors. The gain is separated in amplitude and in sign and the calculation process consists in finding three indices satisfying the minimum quadratic error criterion.

The object of the present invention is to remedy the abovementioned drawbacks, in order in particular to bring about a simplification of the calculations by the introduction as reference vector, into the dictionary of reference vectors, or directions, of almost all the combinations of the n-ary values of the components of the vectors, n being an odd number.

Another object of the present invention is the implementation, prior to the conventional process of applying an adaptive gain to each of the reference vectors, of a correction process by applying a scale factor, introducing the distribution of the energy of the excitation signal as a function of the frequency spectrum thereof, in order to take account of the non-uniformity of the energy distribution of the signal in the frequency domain.

Another object of the present invention is finally the implementation of a low-speed transmission method of a speech signal in which each reference vector, constituting the excitation signal, can be regenerated at the level of a decoder from the only index or address values of the optimal reference vector satisfying the minimum distortion criterion at the level of the coder, which has the effect of considerably simplifying and reducing the manufacturing costs of the aforementioned decoders.

The method for transmitting a low bit rate speech signal according to the present invention comprises a process for coding digital speech samples by code-excited linear prediction to generate a code signal, a process for transmitting the code signal and a process for decoding the received code signal. The coding process corresponds to a process in which a waveform is represented by a block of samples comprising L sample values and constituting an initial vector (o) of dimension L is represented, from a filter of synthesis, by a reference waveform selected from a dictionary of reference waveforms each forming a reference vector (v) on criterion of minimum quadratic deviation of the initial vector (o) with respect to the waveform or reference vector (v), min ∥χ-Hv∥ ² , where χ represents a target vector obtained by perceptual weighting of the initial vector (o) and H a LxL dimension matrix of impulse response from the product of the synthesis filter and the linear perceptual weighting. This process is remarkable in that this selection criterion consists in establishing a dictionary factorized into the product of a first dictionary Yn, n = 2m + 1 odd integer, m being a non-zero integer, of basic vectors yi, of form n-ary, of dimension L, of which each value of the components, yi _(j) , can take a value belonging to the set Im of consecutive relative integers between -m and m, that is to say Im = {-m, -m + 1, ..., - 1,0,1, ..., m-1, m} constituting a set with n = 2m + 1 elements, each of these basic vectors being multiplied by a scaling factor vi account of the distribution of the excitation energy in the frequency domain of the signal, and of a second dictionary G (y) of gains gk, so as to thus represent the dictionary of waveforms or reference vectors, each reference vector verifying the relation v _{k, i} = gk.vi.yi. Note that the value n / 2 corresponds to the integer division of n by 2. The minimum value of the quadratic difference ∥χ-gk.H.νi.yi∥ ² is then established by calculating the maximum of C (gk , νi.yi) = 2 gk <χ | H.νi.yi> - gk ² ∥H.νi.yi∥ ² by calculating all the scalar products <χ | H.νi.yi> and all the energies perceptual ∥Hy∥ ² , which allows to attribute to the initial vector (o) the corresponding optimal reference vector v _{k *, i *} , with v _{k *, i *} = gk * .νi * .yi *, this vector of optimal reference being represented by the only values of index k *, i * satisfying the criterion min ∥χ-gk.H.νi.yi∥ ² .

The process of transmission at low speed of a speech signal, according to the present invention, consists in transmitting, as code signal, the only values of the indices k *, i * representative of each optimal reference vector v _{k *, i *} .

The process of decoding a coded speech signal transmitted at low bit rate according to a code signal, in accordance to the object of the present invention, is remarkable in that, in order to ensure the decoding of the code signal, this process consists in discriminating the values of the indices k *, i * constituting the code signal, in decomposing the value of the index i *, representative of the optimal reference vector, in base n to regenerate the corresponding base vector yi *, to be carried out, from the value of the index i *, of the scale factor νi * and of the corresponding adaptive gain gk *, a correction of the corresponding regenerated base vector to constitute the regenerated reference vector vk *, i *. A synthesis filtering operation is performed on the regenerated reference vector vk *, i * to generate the reconstructed speech signal.

The method which is the subject of the present invention, the coding, transmission and decoding processes and the coding, transmission and decoding system and circuits allowing the implementation of this method advantageously find application in the transmission of speech signals. at low speed, especially between mobiles for example.

• la figure 2 représente au point a) d'une part les étapes de traitement dans un processus de codage conforme à l'objet de la présente invention, et au point b) d'autre part les opérations effectuées, sur les vecteurs de base, dans les étapes représentées au point a), pour des vecteurs n-aires,
• la figure 3a représente aux points 1, 2 et 3 les modules de traitement de vecteurs impulsion constituant des vecteurs de base privilégiés, dans un traitement de type récursif permettant d'engendrer un premier dictionnaire de vecteurs de base,
• la figure 3b représente successivement les opérations effectuées sur les vecteurs de base pour engendrer, de manière itérative, le premier dictionnaire de vecteurs de base précité, dans un cas particulier où n=3, les vecteurs de base étant des vecteurs ternaires,
• la figure 4 représente de manière analogue à la figure 3a, 3b un processus de calcul de la réponse impulsionnelle pour tous les vecteurs ternaires yi excitant le filtre de synthèse et le filtre de pondération perceptuelle en cascade présentant la fonction de transfert H,
• la figure 5 représente en ses différents points a), b), c) et d) des graphes représentatifs des processus de calcul des énergies perceptuelles des vecteurs ternaires à partir des réponses partielles impulsionnelles de la fonction de transfert H,
• la figure 6 représente des graphes représentatifs des processus de calcul des produits scalaires,
• la figure 7 représente un organigramme des étapes de traitement des valeurs d'indice optimal k*,i* reçus au cours du processus de décodage,
• la figure 8 représente un schéma synoptique d'un circuit de codage dans un système de transmission de parole à bas débit conforme à l'objet de la présente invention,
• la figure 9 représente un schéma synoptique d'un circuit de décodage dans un système de transmission de parole à bas débit conforme à l'objet de la présente invention.

The invention will be better understood on reading the description below and on observing the drawings in which, in addition to FIGS. 1a and 1b relating to the prior art,

• FIG. 2 represents in point a) on the one hand the processing steps in a coding process in accordance with the object of the present invention, and in point b) on the other hand the operations carried out, on the basic vectors, in the steps represented in point a), for n-ary vectors,
• FIG. 3a represents, in points 1, 2 and 3, the pulse vector processing modules constituting privileged basic vectors, in a recursive type processing making it possible to generate a first dictionary of basic vectors,
• FIG. 3b successively represents the operations performed on the basic vectors to generate, in an iterative manner, the first dictionary of the aforementioned basic vectors, in a particular case where n = 3, the basic vectors being ternary vectors,
• FIG. 4 represents, in a manner similar to FIG. 3a, 3b, a process for calculating the impulse response for all the ternary vectors yi exciting the synthesis filter and the cascade perceptual weighting filter having the transfer function H,
• FIG. 5 represents at its various points a), b), c) and d) graphs representative of the processes for calculating the perceptual energies of the ternary vectors from partial impulse responses of the transfer function H,
• FIG. 6 represents graphs representative of the processes for calculating scalar products,
• FIG. 7 represents a flow diagram of the steps for processing the values of optimal index k *, i * received during the decoding process,
• FIG. 8 represents a block diagram of a coding circuit in a low bit rate speech transmission system in accordance with the object of the present invention,
• FIG. 9 represents a block diagram of a decoding circuit in a low bit rate speech transmission system according to the object of the present invention.

The method of transmitting a low bit rate speech signal, object of the present invention, will first be described in conjunction with FIGS. 2 a and b.

According to FIG. 2 above, the method which is the subject of the invention comprises a process for coding digital speech samples by linear prediction excited by codes. This process generates a code signal. The method further includes a process for transmitting the code signal and a process for decoding the received code signal.

According to FIG. 2 above, the coding process corresponds to a process in which a waveform represented by a block of samples comprising L values of samples, or frames, constitutes an initial vector denoted by o of dimension L, this vector being represented, and the corresponding waveform, from a synthesis filter by a reference waveform, denoted v, selected from a dictionary of reference waveforms each forming an abovementioned reference vector. The selection is made on the criterion of minimum quadratic deviation of the initial vector o vis-à-vis the waveform or reference vector v, this criterion written:
min ∥χ-Hv∥ ² .

In this relation, χ represents a target vector obtained by perceptual weighting of the initial vector o and H represents a matrix of dimension LxL of impulse response resulting from the product of the synthesis filter and from the aforementioned linear perceptual weighting.

According to the method which is the subject of the present invention, the coding process is such that the selection criterion consists in establishing a dictionary factored into the product of a first dictionary Yn of basic vectors denoted yi. Each basic vector is a basic vector of n-ary form, that is to say that the components yi _(j) of these basic vectors, with j ∈ [O, L-1], can take n values discrete different. In general, each value of the components yi _(j) can take a value included in the set Im of consecutive integers between -m and m, i.e. Im = {-m, -m + 1, ... , -1,0,1, ..., m-1, m} constituting a set with n = 2m + 1 elements.

According to an advantageous characteristic of the method which is the subject of the present invention, each basic vector yi is corrected by a scale factor νi taking into account the distribution of the excitation energy in the frequency domain of the signal. It will be noted that, in the most general way, the scale factors νi are determined from of a database, experimentally, the database being constituted by recording significant speech samples over several hours for example and for several speakers of the same language of expression or of several distinct languages, l experience showing that the diversity of languages of expression only intervenes at the second level in the determination of the aforementioned scale factors νi. A more detailed description of a table of scale factors vi for ternary vectors of dimension L = 5 will be given later in the description.

It will simply be noted that according to this principle the scale factors vi are determined for each corresponding base vector yi by a process of identification of each base vector yi on a delocalized sequence of L successive recursive speech samples from the database , sorting the lowest adaptation coefficients and averaging a number u of identification or adaptation coefficients to obtain the corresponding scale factor vi associated with the aforementioned basic vector yi.

The factorized dictionary previously mentioned also consists of a second dictionary constituting the aforementioned product, this second dictionary being denoted G (y) and being formed by a dictionary of gains gk. The factorized dictionary thus constitutes a dictionary of waveforms or reference vectors. Each reference vector thus verifies the relation v _{k, i} = gk.νi.yi.

It will of course be noted, as shown in FIG. 2a, that the correction operation performed by applying the scale factor vi does not constitute a simple weighting of the components yi _(j) of each basic vector yi because each scale factor coefficient vi is representative of the distribution of excitation energy in the frequency domain of a speech signal.

par calcul de tous les produits scalaires <χ|H.vi.yi> et de toutes les énergies perceptuelles ∥H.y∥².As shown in point a) of FIG. 2, the method which is the subject of the invention then consists in establishing the minimum value of the quadratic deviation ∥χ-gk.H.νi.yi∥ ² by calculating a function denoted: $${\text{C (gk, νi.yi) = 2 gk <χ | H.νi.yi> - <figure-callout id="2" label="gk" filenames="imgf0002.png,imgf0003.png" state="{{state}}">gk</figure-callout>}}^{\text{2}}{\text{∥H.νi.yi∥}}^{\text{2}}$$

by calculating all the scalar products <χ | H.vi.yi> and all the perceptual energies ∥Hy∥ ² .

The abovementioned calculation then makes it possible to attribute to the initial vector o the corresponding optimal reference vector noted v _{k *, i *} , with V _{k *, i *} = gk * .νi * .yi *. Of course, in accordance with a particularly interesting object of the present invention, this optimal reference vector is represented by the only values of the parameters of index k *, i * satisfying the aforementioned criterion:
min ∥χ-gk.H.νi.yi∥ ² .

A more detailed description of the operations carried out at the level of each basic vector yi, these basic vectors being n-ary vectors of dimension L whose components yi _(j) have for value at most the value n / 2 or possibly -n / 2, by integer values and in increments of 1, will be given in conjunction with point b) of Figure 2.

In point b) above, we successively represented the basic vectors denoted yO, y1, yi, yK with K = (n ^L -3) / 2, each component having for value one of the values of the form n-ary . We then represented the correction by application of the scale factor νi, which, for the reasons mentioned above, does not constitute a simple weighting analogous to the adaptive application of the gain gk, to each value of the components yi _(j) of the vectors basic yi being applied the corresponding scale factor vi determined under the conditions mentioned above. On the same point b) we have finally represented the application of the adaptive gain gk, each component yi _(j) of the basic vectors yi then being multiplied by the product gk.νi.

It will of course be understood that, in the implementation of the coding process as represented in points a) and b) of FIG. 2 previously mentioned, the minimum value of the quadratic deviation min ∥χ-gk.H.νi.yi∥ ² is evaluated by selecting the corresponding gain element gk from the second dictionary G (y) making it possible to minimize the difference | g-gk * | where g checks the relation: $$\text{g =}\frac{\text{<χ | H.νi.yi>}}{{\text{∥H.νi.yi∥}}^{\text{2}}}$$

A more detailed description of the arrangement of the basic vectors yi to constitute the dictionary or first dictionary Yn of dimension L of basic vectors yi will now be given in connection with FIGS. 3a and 3b.

In general, it will be understood that the dictionary Yn of basic vectors yi of n-ary form [-n / 2, ..., O, ... n / 2] of dimension L includes all the basic vectors whose L components have for value the aforementioned n-values, with the exception of the null vector. In general, the index i of the basic vectors is taken equal to the value in base n of each base vector after transcoding the values [-n / 2 ..., O, ... n / 2] into values corresponding (0,1,2, ... n). It will thus be understood that the basic vectors yi of n-ary form are arranged as a function of their index i, this index i having for value the value in base n of each vector.

It will also be understood that the set of basic vectors yi constituting the dictionary Yn is defined from n / 2.L impulse vectors of which only one component yi _(j) of order j, with j ∈ [O, L-1 ], is equal to -1, -2, ... -n / 2. Associated with each pulse vector are the related basic vectors having values of components of order q ≤ j identical, each vector related to a pulse vector of rank q, with q = j for yi _(j) different from O, being obtained. by linear combination of the pulse vector of rank j = q and of the pulse or related vectors of rank j = q 'higher.

A more detailed description of the implementation of the dictionary of basic vectors yi in the case of ternary vectors and of the manner of generating these basic vectors will be given in connection with FIGS. 3a and 3b, basic vectors of dimension L and of n-ary shape being able to be generated according to the same principle without departing from the scope of the object of the present invention.

In FIGS. 3a and 3b, operator cells have been shown respectively making it possible to generate, from the previously defined pulse vectors and from sub-dictionaries constituted by the pulse vector considered and the related vectors corresponding to each pulse vector, the complete dictionary comprising the collection of all of the sub-dictionaries.

Each operator as represented in FIG. 3a comprises an operator called delay operator R whose transfer function is denoted Z ⁺¹ , according to the classical notation of transform into Z, a symmetrizer operator denoted Sy which has the function of multiplying the components of all the vectors presented at its entry by the value +1, by the value O then by the value -1 and a summator, noted S, receiving the output of the delay operator R and the symmetrizer Sy. The summator S receives the output of the delay operator R via a switch I, in position F, or the zero vector [0,0,0,0,0] of dimension L in position O. The operators represented in FIG. 3a are constituted by a single operator represented in 1), 2) and 3) at different stages of the processing process to generate the basic vectors yi of the abovementioned dictionary Yn.

lequel est constitué initialement par le vecteur impulsion δ L-1 précité. Le symétriseur Sy délivre un sous-dictionnaire symétrique noté $\overline{\text{DO}}$

, tel que représenté en figure 3b, et le sommateur S, lequel reçoit le vecteur impulsion δ L-2 délivré par l'opérateur de retard R, vecteur impulsion de rang q = L-2, ou le vecteur nul, et le sous-dictionnaire symétrique $\overline{\text{DO}}$

, délivre en sortie le dictionnaire D1 constitué par les vecteurs de base y0, y1, y2 et y3. On notera bien sûr, ainsi que représenté en figure 3b, qu'au vecteur impulsion δ L-2 est associé le sous-dictionnaire D1 formé par les vecteurs apparentés y1, y2, y3 au vecteur impulsion δ L-2 et par le vecteur impulsion à L-1 initial formant le vecteur de base y0, ainsi que le vecteur nul. Bien entendu, et de manière récursive ainsi que représenté au point 2) de la figure 3a, l'opérateur permettant d'engendrer les vecteurs de base yi est tel que celui-ci reçoit au niveau de l'opérateur de retard R le vecteur impulsion à L-m, au niveau du symétriseur Sy, le dictionnaire noté D m-1 formé récursivement comme le dictionnaire D1, le sommateur S tel que représenté au point 2 de la figure 3a délivrant alors à partir du vecteur impulsion δ L-m-1 précité délivré par l'opérateur de retard R ou du vecteur nul et par le sous-dictionnaire $\overline{\text{D m-1}}$

le sous dictionnaire Dm.At the start of the process of generating the basic vectors yi, as represented in point 1) of FIG. 3a, the initial pulse or pulse vector δ L-1 is present at the input of the delay operator R The symmetrizer Sy is then supplied by a sub-dictionary noted $\overline{\text{DO}}$

which is initially formed by the above-mentioned pulse vector δ L-1. Sy balancer delivers a sub-dictionary symmetrical noted $\overline{\text{DO}}$

, as shown in FIG. 3b, and the summator S, which receives the pulse vector δ L-2 delivered by the delay operator R, pulse vector of rank q = L-2, or the zero vector, and the sub- symmetric dictionary $\overline{\text{DO}}$

, outputs the dictionary D1 consisting of the basic vectors y0, y1, y2 and y3. It will of course be noted, as shown in FIG. 3b, that the pulse vector δ L-2 is associated with the sub-dictionary D1 formed by the related vectors y1, y2, y3 with the pulse vector δ L-2 and by the pulse vector to initial L-1 forming the basic vector y0, as well as the zero vector. Of course, and recursively as shown in point 2) of FIG. 3a, the operator making it possible to generate the basic vectors yi is such that the latter receives the impulse vector at the level of the delay operator R at Lm, at the symmetrizer Sy, the dictionary denoted D m-1 formed recursively like the dictionary D1, the summator S as represented in point 2 of FIG. 3a then delivering from the above-mentioned pulse vector δ Lm-1 delivered by the delay operator R or the zero vector and by the sub-dictionary $\overline{\text{D m-1}}$

the Dm dictionary.

By iteration it is thus possible and recursively to generate from the set of impulse vectors, as previously described, the related vectors and the corresponding sub-dictionaries and finally the complete dictionary. It will be noted that, in FIG. 3b, the * represented at the level of the components yi _(j) for the processing process of level m correspond to values 0, -1 or +1 when the vectors are ternary vectors. Of course, in the case of n-ary vectors the * represent values between -n / 2 and + n / 2, under the conditions previously mentioned.

It will be noted that the total ternary dictionary, sum or union of all the sub-dictionaries of intermediate level m, up to L can be obtained for the only positive or negative values of the components yi _(j) , the dictionary total can then be obtained by symmetrization through a symmetrization operator such as Sy.

In the same way, the computation of the partial response at an instant t = L-1, that is to say at a relative instant corresponding to the occurrence of the impulse vector δ L-1, of the system H constituted by the synthesis filter and the perceptual weighting filter excited by the ternary basic vectors yi can be described using the preceding operators. The partial response at time t = L-1 is denoted SL-1 (yi).

At the level of the first calculation operator, noted 1 in FIG. 4, this operator is such that the impulse responses of the system H at the relative time 0, 1, 2, L-1, that is to say the values h ₀ , h ₁ , h _L-2 , h _L-1 are applied to the aforementioned operator.

It is recalled that the operator SL-1 here also represents the addition to each element h _Lm-1 or to the value zero of all the partial responses to t = L-1 of the vectors of the symmetrized dictionary delivered by the symmetrizer Sy of level m.

We thus obtain S _{L-1 (Dm)} together of the responses t = L-1 of the vectors of Dm.

The symmetrization operator Sy multiplies the elements of S _{L-1 (Dm-1)} by +1, 0, -1 and realizes, as described above, the union of the distinct elements obtained. Finally, the last operator represented at 3 in FIG. 4 provides the response to t = L-1 of the ternary vectors yi whose first coordinate is -1.

It will be noted that the response of the linear system of the matrix H to the ternary vectors which are applied to it can therefore be carried out according to the same architecture as previously by applying the linear transformation H to each node of this architecture.

The perceptual energies of the ternary vectors can then be deduced from the only partial responses previously described at t = L-1.

Indeed,: the response of the matrix H to the excitation by a vector yi is written:

Indeed, by definition the response at the relative time t = L-1, noted SL-1 (yi) is the coordinate of order L-1 of Hyi.

et

However, we can write:

and

$${\text{∥H.yi∥}}^{\text{2}}{\text{= S}}_{\text{L-1}}{\text{(yi)}}^{\text{2}}{\text{+ ∥H.z}}^{\text{-1}}{\text{yi∥}}^{\text{2}}$$

It will be noted that y'i and y "i have the same norm and we can demonstrate, by noting z ^-1 the elementary delay operator, the following relation: $${\text{∥y'i∥}}^{\text{2}}{\text{= ∥y "i∥}}^{\text{2}}{\text{= ∥Hz}}^{\text{-1}}{\text{yi∥}}^{\text{2}}$$

$${\text{∥H.yi∥}}^{\text{2}}{\text{= S}}_{\text{L-1}}{\text{(yi)}}^{\text{2}}{\text{+ ∥Hz}}^{\text{-1}}{\text{yi∥}}^{\text{2}}$$

However, if yi belongs to Dm, z ^-1 .yi belongs to Dm-1.

An iterative process therefore makes it possible to calculate the perceptual energies for D0 then D1 then DL-1. The initial value is for D0 = δ L-1, that is to say the impulse vector previously represented in figure 3, h ₀ ² .

A schematic diagram of the numbering and calculation process of the different entities implemented by the selection criterion in accordance with the object of the present invention will be described in connection with FIGS. 5a and 5b.

In general, as shown in FIG. 5a, the basic vectors yi as already described above can be generated according to the global generation graph at the rate of 3 ° = 1 vector is generated at level 0, the vector y0, 3 ¹ are generated at level 1, the vectors y1, y2, y3 and so on, 3 ^L-1 basic vectors at level L-1.

The elementary decryption cell is represented in Figure 5b from the pulse vectors noted not-1, θ0 and θ1. It will be noted that the summation of the pulse vectors ecteurs1, θ0 and θ-1 amounts to replacing the last coordinate of the incident base vector by the component values +1, O or -1.

It will be noted that the architecture as represented in FIGS. 5a and 5b is that of a linear structure of ternary graphs. For an n-ary structure we get an n-ary graph.

It is also possible to obtain a practical calculation for the expression ∥H.yi∥ ² = SL-1 (yi) ² + ∥Hz ^-1 yi∥ ² thanks to the analogous architecture below. This architecture will be described in connection with Figures 5c and 5d.

We call E (i) the expression E (i) = ∥H.yi∥ ² .

As shown in FIG. 5c, the global graph for obtaining the energies is traversed from right to left, the initial energy E (0) being equal to SL-1 (0) ² .

The elementary cell constituting the graph represented in FIG. 5c is represented in FIG. 5d.

It will be noted that the numbering of the vectors, that is to say the assignment of their index i of basic vectors, can correspond either to a retrograde numbering, any index p of the direct numbering of a ternary vector verifying the relation corresponding in numbering p 'retrograde p' = 3 ^L -p-1. It will of course be understood that all the calculations can be carried out either with direct numbering or with retrograde numbering, the latter being preferred. It is then possible to transmit the retrograde index values for example or the direct index values on the transmission line as will be described later in the description.

It will also be noted, in accordance with previous practices in CELP-type coding, that prior to synthesis filtering each reference vector v _{k *, i *} can advantageously be weighted by a factor of predicted level, noted a. This predicted level factor a is representative of the average energy of the excitation signal estimated on at least three successive previous excitation vectors. Such an operation on the components yi _(j) of each reference vector will not be described since it corresponds to an operation known to those skilled in the art.

A more detailed description of a process for calculating scalar products of the form <2χ | H.yi> where χ = x / σ for all the basic vectors yi will now be described in conjunction with Figure 6.

It will be noted that in fact, taking into account the predicted level factor σ effectively introduced into the coding process object of the present invention, it is in fact a question of calculating the expression <2x / σ | H.yi> for all the ternary vectors yi.

The previous expression is then calculated by filtering the expression 2x / σ by the matrix transposed from the matrix H, ie H ^t .

This expression is written: $${\text{<2x / σ | H.yi> = <H}}^{\text{t}}\text{. (2x / σ) | yi>}$$

By setting x '=

The expression <x '| yi> for the ternary yi vectors can be obtained in the following way: the expression is calculated: $${\text{- <H}}^{\text{t}}\text{(2x / σ) | yi> = F (yi).}$$

The calculation process as represented by the operator in FIG. 6 allows, in a manner analogous to the calculation of the partial responses S L-1 (yi) previously described, to obtain the quantities x'0, x'Lm-1, x'L-2 and therefore the aforementioned scalar products, the null vector being replaced by the null value.

With regard to the determination and the attribution of the scale factor νi to each of the basic vectors yi it is recalled that each scale factor vi can be determined from a plurality N of frames, from a speech signal database, the scale factor νi for each base vector yi being chosen so as to make the filter residue of the aforementioned frames minimal for the frame considered. It will be recalled that several processes for determining each scale factor vi can be envisaged.

By way of nonlimiting example, in the case of basic vectors of ternary type and of dimension L = 5 the table of the scale factors vi is given below by the table of 121 values of the scale factors. The first value multiplies (-1, -1, -1, -1, -1) ..., the last (0,0,0,0, -1).
1.50, 1.66, 1.77, 1.28, 1.46, 1.36, 0.86, 2.47, 1.68, 1.51, 1.12, 1.04, 1.38, 1.86, 1.51, 4.23, 3.47, 1.96, 1.25, 2.28, 0.77, 2.50, 3.51, 0.87, 1.11, 1.16, 0.95, 1.29, 1.23, 1.85, 1.34, 1.55, 1.60, 1.51, 1.44, 1.21, 1.45, 1.95, 1.45, 1.73, 4.06, 1.73, 1.32, 1.39, 2.43, 1.38, 4.62, 1.35, 1.92, 2.15, 1.44, 2.20, 1.95, 1.07, 0.88, 1.56, 1.48, 1.33, 1.64, 1.70, 1.44, 3.33, 1.10, 1.89, 0.80, 2.07, 1.27, 1.57, 3.82, 1.28, 1.31, 1.34, 1.94, 1.86, 1.25, 1.06, 2.15, 1.39, 0.89, 1.24, 1.32, 1.17, 1.45, 0.57, 1.28, 2.00, 4.88, 2.14, 2.98, 2.24, 1.23, 1.66, 1.41, 1.82, 3.44, 1.14, 3.15, 3.91, 1.60, 0.95, 1.74, 1.50, 1.12, 2.98, 1.16, 1.23, 1.34, 1.00, 2.06, 2.52, 4.52, 1.93, 2.89, 3.21, 1.39, 2.44, 2.38, 4.55, 3.00, 2.49, 3.17

The optimal values of the indices k * and i * having been determined and numbered directly or retrogradely as described previously in the description, as regards in particular the value of the indices i, the transmission of speech at low speed is effected by the only transmission, as code signal, of the values of the indices k * and i * representative of each reference vector v _{k *, i *} .

With regard to the transmission of the aforementioned indices k * and i *, it will be noted that the transmission can be carried out using conventional transmission protocols in which a redundancy of the information transmitted is introduced in order to ensure transmission at a rate substantially zero error. It will of course be understood that the value i * can be transmitted either by direct numbering or by retrograde numbering, or according to a translated numbering whose translation table is known to the coder as well as to the decoder.

A more detailed description of the decoding process of the transmitted information, that is to say of the code signal thus transmitted in accordance with the method which is the subject of the invention, will now be given in conjunction with FIG. 7.

In accordance with FIG. 7 above, the decoding process consists in discriminating in 1000 the values of the indices k * and i * constituting the code signal and then in decomposing into 1001 the value of the index i * representative of the optimal reference vector in base n in order to regenerate the corresponding base vector yi *.

The regeneration of the base vector yi * is carried out in 1002 from the value of the index i * and the corresponding scale factor νi *, a correction of the corresponding base vector regenerated being carried out to constitute the reference vector v _{k *, i *} = νi *, yi *.

Following the above-mentioned operation, the decoding process consists in carrying out a synthesis filtering operation 1003 of the reference vector to generate the reconstructed speech signal.

It will of course be noted that, as in the case of coding process, in the coding process of the method which is the subject of the present invention, each reference vector v _{k *, i *} before the synthesis filtering is weighted by a predicted level factor σ which is estimated on at least three excitation vectors successive previous. The determination of the predicted level σ will not be described in detail since it corresponds to the level of the process of decoding the operations normally known to those skilled in the art.

A more detailed description of a system for transmitting a low bit rate speech signal in accordance with the object of the present invention will be described in conjunction with FIGS. 8 and 9.

According to FIG. 8, the coding circuit comprises a generator 1 of a first dictionary Yn of basic vectors yi of n-ary shape of dimension L, the components of these vectors, as mentioned previously, being able to take the values between -n / 2 to n / 2. It will of course be noted that the generator of the dictionary Yn can advantageously be constituted by calculating means comprising the operators as described in FIGS. 3a, 3b for example and / or a storage circuit which can be constituted by a random access memory associated with this circuit computer or by ROM. In this case, the read only memory is associated with a fast sequencer which makes it possible to carry out a successive reading of the basic vectors yi according to the indices in direct or retrograde numbering as described previously.

= νi.yi pour chaque vecteur de base yi. Un multiplexeur rapide noté MUX permet successivement de lire les valeurs correspondantes du vecteur de base corrigé $\overline{\text{yi}}$

et de délivrer cette valeur correspondante à un circuit 3 générateur d'un deuxième dictionnaire de gain adaptatif gk. De manière classique, le circuit 3 générateur du deuxième dictionnaire G(y) peut comporter avantageusement un circuit amplificateur, noté 30, relié à une table des valeurs gk constituant le deuxième dictionnaire précité. Ainsi, le circuit générateur 3 du deuxième dictionnaire G(y) délivre les vecteurs de référence v_k,i = gk.νi.yi.In addition, the coding circuit as represented in FIG. 8 comprises a circuit 2 correcting the basic vectors yi by a scale factor νi. The correction circuit can be constituted by a table of values memorized in read-only memory, this correction circuit making it possible to generate a corrected base vector noted $\overline{\text{yi}}$

= νi.yi for each basic vector yi. A fast multiplexer denoted MUX makes it possible successively to read the corresponding values of the corrected base vector $\overline{\text{yi}}$

and deliver this value corresponding to a circuit 3 generator of a second adaptive gain dictionary gk. Conventionally, the generator circuit 3 of the second dictionary G (y) may advantageously include an amplifier circuit, denoted 30, connected to a table of values gk constituting the aforementioned second dictionary. Thus, the generator circuit 3 of the second dictionary G (y) delivers the reference vectors v _{k, i} = gk.νi.yi.

It will of course be noted that the coding circuit which is the subject of the present invention also comprises an amplifier circuit 4 which makes it possible to apply to each reference vector v _{k, i} the level prediction coefficient a as defined above. in the description.

In addition, and in a conventional manner, the coding circuit object of the present invention then comprises, arranged in cascade, the synthesis filter noted 5 and the perceptual weighting filter noted 6 of transmission H as described previously in the description. A summator 7 makes it possible to receive on the one hand the original signal via a same perceptual weighting filter 6 after inversion of the difference of the signals delivered by the algebraic summator 7, allowing the application on the signal thus obtained from the minimum distortion criterion.

To this end, the coding circuit which is the subject of the present invention comprises a circuit for calculating the minimum distortion 8 which comprises a first calculating circuit 80 of the product 2 gk <x / σ | H.νi.yi> in which the expression <x / σ | H.νi.yi> denotes the scalar product of the target vector x and the perceptually reconstructed and weighted vector obtained by the product of the matrix H and the corrected base vector νi.yi. The first calculator circuit 80 delivers a first calculation result r1.

A second calculator circuit 81 makes it possible to calculate the energy of the perceptually reconstructed and weighted vector, this energy being of the form gk ² ∥H.νi.yi∥ ² .

It will be noted that the computer circuits 80 and 81 can be constituted by program modules whose calculation graphs have been explained in FIGS. 4 and 5 a) to d) respectively. The second calculation circuit 81 delivers a second calculation result denoted r2. A comparator 83 makes it possible to compare the value of the calculation results r1 and r2 which makes it possible to determine by discrimination of the values of the indices i and k, the indices i * and k * for which the criterion of minimum of the quadratic difference is satisfied . The discrimination of the indices i * and k * is carried out for example by a sorting program noted 84 in FIG. 8. The values of the indices k * and i * are then delivered, these indices being representative of the corresponding reference vector v _{k * , i *} .

FIG. 8 also shows the transmission circuit according to the subject of the present invention, this transmission circuit making it possible to deliver as a code signal representative of the speech signal the only values of the indices k * and i *. This transmission circuit does not have any particular characteristic insofar as it can in fact be constituted by a transmission system of conventional type used in the devices for transmitting speech signals by coding of CELP type of the prior art.

A more detailed description of a decoding circuit allowing the implementation of the method which is the subject of the invention is shown in FIG. 9.

In accordance with the above-mentioned figure, the decoding circuit comprises a module 10 for discriminating the values of the indices i *, k * of the code signal received, the code signal being of course transmitted according to a particular protocol which does not enter into the object of the present invention. In addition, the discrimination circuit 10 thus effecting a parallel series transformation of the information relating to the indices i *, k *, the decoding circuit comprises a circuit n-base decomposition of the value of the index i *.

It will of course be understood that in parallel the index k * is treated. For this purpose, the decoding circuit as shown in FIG. 9 comprises a table of the adaptive gain values gk denoted 11, which, on reception of the value of the index k *, makes it possible to deliver the corresponding adaptive gain value gk *. This circuit 11 can advantageously consist of a read-only memory in which the adaptive gain values gk are stored.

par décomposition en base n de la valeur de l'indice i*. Dans ce but, un circuit 14 fait correspondre à la valeur i* par transcodage des composantes en base n de la valeur de l'indice i*, la valeur [-n/2,...,0,...n/2], ce qui permet d'engendrer un vecteur de référence régénéré v_k*,i* du produit du vecteur

de base régénéré et du produit A.In addition, a generator circuit 12 of the scale factor νi * is provided. This circuit can consist of a read only memory forming a look-up table, which with the value i * makes correspond the value νi *. A multiplier circuit 12a makes it possible to generate a product coefficient A = σ.gk * .νi * from the values νi *, gk * and the predicted level coefficient σ.
As also shown in FIG. 9, the decoding circuit comprises a circuit 13 generating the regenerated base vector

by decomposition into base n of the value of the index i *. For this purpose, a circuit 14 corresponds to the value i * by transcoding the components in base n of the value of the index i *, the value [-n / 2, ..., 0, ... n / 2], which makes it possible to generate a regenerated reference vector v _{k *, i *} of the product of the vector

regenerated base and product A.

d'engendrer le signal de parole reconstruit.A synthesis filter 15 allows starting from the regenerated reference vector

to generate the reconstructed speech signal.

The operation of the decoding circuit as shown in Figure 9 can be summarized as follows according to a preferred operation.

The double multiplication carried out at the level of the multiplier 12 gives an amplitude factor noted A = σ.gk * .νi *.

• étape courante (j,t),
• si j modulo 3 vaut 0 alors v_k*,i* (L-1-t) = -A
• si j modulo 3 vaut 1 alors v_k*,i* (L-1-t) = O
• si j modulo 3 vaut 2 alors v_k*,i* (L-1-t) = A

où v_k*,i* (L-1-t) représente la composante de v_k*,i* à l'ordre L-1-t.If the index i * of the ternary vector transmitted corresponds to the retrograde numbering, we denote by:
i '= (3 ^L -3) / 2 - i * and the vector is synthesized excitation or reconstructed reference vector v _{k *, i *} as follows:

• current step (j, t),
• if j modulo 3 is worth 0 then v _{k *, i *} (L-1-t) = -A
• if j modulo 3 is worth 1 then v _{k *, i *} (L-1-t) = O
• if j modulo 3 is worth 2 then v _{k *, i *} (L-1-t) = A

where v _{k *, i *} (L-1-t) represents the component of v _{k *, i *} in the order L-1-t.

Note that j is divided by 3, integer division, and t is increased by 1, addition of 1 to an integer. The first step is initialized by j = i 'and t = O.

Of course, the current step is repeated until t = L-1 inclusive.

If on the contrary i * comes from a direct numbering, as described previously, then i '= i and the operations on j modulo 3 are carried out as previously mentioned.

We have thus described a method and a system for transmitting speech at low speed which is particularly efficient insofar as an important advantage lies in the fact that the dictionary Yn does not have to be stored at the level of the decoder. Thus only the indices of the reference vector are transmitted to the decoder, a calculation making it possible in real time to reconstruct the corresponding reference vector, which allows a saving of memory resource at the level of each decoder used. In addition and due to the processes of generation of the basic vectors, the processes of calculation of the scalar products and of the perceptual energies, it is also not necessary to memorize the basic vectors at the level of the coder, which allows a gain substantial in implementation material.

It will also be understood that the calculation algorithms described in the description of the object of the present invention make it possible to obtain a very high calculation speed by rationalizing the calculation operations used, and a simplification of the materials necessary for their implementation.

Finally, it will be noted that the method and the system for transmitting a low bit rate coded speech signal which is the subject of the present invention have been described in the case where the CELP type coding implements basic vectors of n-ary type. , the number n being in principle not limited. Of course, a preferred embodiment has been given in the case where n = 3, the basic vectors thus being ternary vectors.

However, an embodiment based on the same principle has been possible for vectors for which n = 5. The dictionary Yn is then produced from an alphabet with five symbols, the values obtained being for example, without limitation, the symbol 0, the symbol 0.5 and the symbol 1 plus the symmetrical symbols -0.5 and -1, which can be reduced to any integer values by change of scale.

In the implementation of a dictionary with five symbols, it was thus possible to carry out a method and a transmission system with variable bit rate which can reach up to 24 Kbits per second.

Claims (12)

1. A method of transmitting a speech signal at low throughput comprising a procedure for coding digital samples of speech by code excited linear prediction, in order to generate a code signal, a procedure for transmitting the code signal and a received code signal decoding procedure, the coding procedure corresponding to a procedure in which a waveform represented by a sample block comprising L sample values and constituting an initial vector (o) of dimension L is represented, on the basis of a synthesizing filter, by a reference waveform chosen from a dictionary of reference waveforms each forming a reference vector (v) relating to a criterion of minimum square deviation of the said initial vector (o) in relation to the said waveform or reference vector (v), min ∥χ-H.v∥², where χ represents a target vector obtained by perceptual weighting of the said initial vector (o) and H represents a pulse-response matrix of dimension LxL resulting from the product of the synthesizing filter and of the linear perceptual weighting, wherein the said selection criterion consists:

   - in establishing a dictionnary factorized as a product of a first dictionary Yn, with n =2m+1 being an odd integer, and m being an integer different from zero, of basis vectors yi, of dimension L, each component yi_(j) of said basis vectors having a given value belonging to the set of successive related integer values Im comprised between -m and m, with Im = {-m, -m+1, ...-1,0,1, ...m-1,m} forming a set of values comprising n=2m+1 elements, these basis vectors each being weighted by an associated scale factor νi which takes account of the distribution of excitation energy in the frequency domain of the signal and of a second dictionary G(y) of adaptive gains gk, in such a way as to thus represent the dictionary of waveforms or reference vectors, each reference vector satisfying the relation v_k,i = gk.νi.yi,

   - in establishing the minimum value of the square deviation min ∥χ-gk.H.νi.yi∥² by calculating the maximum of C(gk,νi.yi) = 2gk <χ|H.νi.yi> - gk² ∥H.νi.yi∥² by calculating all the scalar products <χ|H.νi.yi> and all the perceptual energies ∥H.y∥², this making it possible to assign to the initial vector (o) the corresponding optimal reference vector v_k*,i* with v_k*,i* = gk*.νi.yi*, this optimal reference vector being represented by the sole values of the index parameters (k*,i*) satisfying the criterion min ∥χ-gk.H.νi.yi∥².
2. The method as claimed in claim 1, characterized by said minimum value of the square deviation min ∥χ-gk.H.νi.yi∥² being evaluated by selecting the corresponding gain element gk of the second dictionary G(y) making it possible to minimize the difference |g-gk*| where g satisfies the relation: $$\text{g =}\frac{\text{<χ|H.νi.yi>}}{{\text{∥H.νi.yi∥}}^{\text{2}}}\text{.}$$

   
3. The method as claimed in one of the claims 1 or 2, characterized by said first dictionary Yn, n=2m+1, of basis vectors yi, of n-ary form [-n/2, ..., 0, ...n/2] of dimension L comprising all the basis vectors whose L components have the value of one of the values [-n/2,...,0,... n/2] with the exception of the null vector, the index i of the basis vectors being made equal to the base n value of each basis vector after transcoding of the values [-n/2 0,...n/2] into a corresponding value [0,1,2...n].
4. The method as claimed in claim 3, characterized by the set of basis vectors yi constituting the said dictionary Yn being defined from the n/2.L pulse vectors, of which a single component yi_(j) of order j with j ∈ [O,L-1] is equal to -1, -2, ..., -n/2, each pulse vector being associated with the allied basis vectors having identical component values of order q < j, each vector allied to a pulse vector of rank q with q = j for yi_(j) ≠ 0 being obtained by linear combination of the said pulse vector of rank q and of the pulse or allied vectors of higher rank q.
5. The method as claimed in one of the claims 1 to 4, characterized by for each basis vector yi, the scale factor νi associated with it is determined experimentally, from a plurality N of frames consisting of L speech-signal values and forming a database, the scale factor νi for each basis vector yi being selected in such a way as to minimize, from the relevant frame, the filtering residue from the said frames.
6. The method as claimed in claim 1, characterized by in order to ensure the transmission of the speech signal at low throughput, the transmission procedure consisting in transmitting as code signal the sole values of the indices (k*, i*) representing each reference vector v_k*,i*.
7. The method as claimed in claims 1 and 2, characterized by in order to ensure the decoding of the code signal, this method consisting:

   - in distinguishing (1000) the values of the indices k*,i* constituting the code signal,

   - in decomposing (1001) the value of the indice i*, representing the optimal reference vector to base n in order to generate the corresponding basis vector yi*,

   - in performing (1002), from the corresponding value of the index i* and of the corresponding scale factor νi*, a correction of the corresponding regenerated basis vector in order to build up the reference vector v_k*,i* = νi*.yi*,

   - in performing a synthesizing filtering operation (1003) of the reference vector in order to generate the reconstructed speech signal.
8. The method according to one of the preceding claims characterized by prior to the synthesizing filtering, each reference vector v_k*,i* being weighted (1004) by a predicted level factor a representing the average energy of the excitation signal estimated over at least three successive earlier excitation vectors.
9. A system for transmitting a speech signal at low throughput comprising a circuit for coding digital samples of speech by code excited linear prediction in order to generate a code signal, a circuit for transmitting this code signal, and a circuit for decoding the transmitted code signal, the coding circuit comprising a synthesizing filter making it possible to represent a waveform consisting of a sample block constituting an initial vector (o) of dimension L, by a reference waveform chosen from a dictionary of reference waveforms each forming a reference vector (v) relating to a criterion of minimum square deviation of the said initial vector (o) in relation to the waveform or reference vector (v), means of perceptual weighting of the said initial vector (o) in order to generate a target vector χ from the said initial vector (o), the said criterion of minimum square deviation of the initial vector (o) in relation to the said waveform or reference vector (v) being of the form min ∥χ-H.v∥² where H represents a pulse-response matrix of the dimension LxL resulting from the product of the said synthesizing filter and of the said linear perceptual weighting, characterized by, in order to implement the said selection criterion, the said coding circuit comprising:

   - means (1) for generating a first dictionary Yn, with n=2m+1 being an odd integer, and m being an integer different from zero, of basis vectors yi of dimension L, each component yi_(j) of said basis vectors having a given value belonging to the set of successive relative integer values Im comprised between -m and m, with Im = {m,-m+1, ...,-1,0,1,...,m-1,m} forming a set of values comprising n=2m+1 elements,

   - means (2) for multiplicating the said basis vectors yi by a scale factor νi, this scale factor taking into account of the distribution of the excitation energy in the frequency domain of the signal, the said correcting means making it possible to generate a corrected basis vector $\overline{\text{y}}$

   

   i = νi.yi for each basis vector yi,

   - means (3) for generating a second dictionary G(y) of adaptive gains gk, comprising multiplier means making it possible, from the said corrected basis vector yi and from the gain values gk, to generate the said reference vectors v_k,i = gk.νi.yi,

   - first means (80) of calculating the product 2gk <χ|H.νi.yi> where <χ|H.νi.yi> designates the scalar product of the said target vector χ and of the perceptually weighted reconstituted vector obtained by the product of the matrix H and of the corrected basis vector νi.yi, the said first calculation means delivering a first calculation result (r1),

   - second means (81) of calculating the energy of the perceptually weighted reconstituted vector gk² ∥H.νi.yi∥², the said second calculation means delivering a second calculation result (r2),

   - means (82) of comparing the said first and second calculation results, this making it possible to determine, by distinguishing the values of the indices i,k, the indices i* and k* for which the criterion of minimum square deviation is satisfied, the corresponding reference vector v_k*,i* with v_k*,i* = gk*.νi*.yi* being represented by the sole values of the indices k*, i* satisfying the criterion min∥χ-gk.H.νi.yi∥².
10. The system as claimed in claim 9, characterized by the transmission circuit making it possible to transmit, in the guise of a code signal representing the speech signal, the sole values of the indices k* and i*.
11. The system as claimed in claim 9, characterized by the decoding circuit comprising :

    - means (10) for distinguishing the values of the indices i*, k* of the code signal received,

    - means (11) generating a dictionary G(y) for adaptive gains gk* from the distinguished values k*,

    - means (12) generating the corresponding scale factor νi*,

    - multiplying means (12a) for generating a product coefficient σ.gk*.νi* from the values i*,gk* and from a predicted level coefficient σ,

    - means (13) of decomposing to base n the index value i*,

    - means (14) generating the regenerated basis vector ŷi* corresponding to the value i* by transcoding of the components to base n of the index value i*, each value n,...2,1,0 of the expression to base n of the index value i* being associated with respectively the value [n/2,...0,...n/2], this making it possible to generate a regenerated reference vector

    

    , a synthesizing filter making it possible on the basis of the regenerated reference vector

    

    to generate the reconstructed speech signal.
12. The system as claimed in one of claims 9 to 11, characterized by the said coding, respectively decoding circuit further comprising, upstream of the synthesizing filter, a circuit for correcting the reference vector v_k*,i*, respectively regenerated reference vector

    

    by a predicted level factor representing the average energy of the excitation signal estimated over at least three successive earlier excitation vectors.

Images