EP0608174B1 - System for predictive encoding/decoding of a digital speech signal by an adaptive transform with embedded codes - Google Patents

Description

The present invention relates to a system for predictive coding-decoding of a digital speech signal by adaptive transform with nested codes.

In currently predictive transform coders used, this type of encoder being represented in figure 1, we seek to construct a synthetic signal S andn as close as possible to the digital speech signal to code Sn, resemblance in the sense of a perceptual criterion.

The digital signal to code Sn, coming from a signal of analog source speech, is subject to a process of short-term prediction, LPC analysis, coefficients of prediction being obtained by prediction of the speech signal on windows with M samples. The signal speech digital code Sn is filtered using a perceptual weighting filter W (z) deduced from the coefficients above, to get the signal perceptual pn

A long-term prediction process then makes it possible to take into account the periodicity of the residue for the voiced sounds, on all the sub-windows of N samples, N <M, in the form of a contribution p and _n , which is subtracted from the perceptual signal pn so as to obtain the signal p'n in the form of a vector P'∈R ^N.

A transformation followed by a quantification are then carried out on the aforementioned vector P ′ in order to carry out a digital transmission. The reverse operations allow, after transmission, the modeling of the synthetic signal S and _n .

In order to obtain good perceptual behavior, according to the usual criteria established by experience, it is necessary to establish a transformation process by orthonormal transform F and quantization of the vector P ', in the presence of gain values G verifying well determined properties, G = F ^T .P 'where F ^T denotes the matrix transposed from the matrix F.

A first solution, proposed by G. Davidson and A.Gersho, in the publication "Multiple-Stage Vector Excitation Coding of Speech Wave forms ", ICASSP 88, Vol.1, pp 163-166, consists in using a transformation matrix non singular V = HC where H is a triangular matrix lower and C a non-singular dictionary, constructed by learning, ensuring the invertability of the matrix transformation V for any sub-window.

In order to be able to exploit certain properties of decorrelation and ordering of vector components of coefficients of the transform G during the step of quantification, several solutions using transforms orthonormal have been proposed.

où I est le nombre de vecteurs contenu dans le corpus d'apprentissage, permet de maximiser l'expression

où K est un entier, K ≤ N. On démontre que l'erreur quadratique moyenne de la transformée de Karhunen-Loeve est inférieure à celle de toute autre transformation pour un ordre de modélisation K donné, cette transformée étant, dans ce sens, optimale. Ce type de transformée a été introduit dans un codeur prédictif par transformée orthogonale par N.Moreau et P.Dymarski, confer publication "Successive Orthogonalisations in the Multistage CELP Coder", ICASSP 92 Vol.1, pp I-61 - I-64.The Karhunen-Loeve transform, obtained from the eigenvectors of the auto-correlation matrix

where I is the number of vectors contained in the learning corpus, allows to maximize the expression

where K is an integer, K ≤ N. We prove that the mean square error of the Karhunen-Loeve transform is lower than that of any other transformation for a given modeling order K, this transform being, in this sense, optimal . This type of transform was introduced into a predictive coder by orthogonal transform by N. Moreau and P. Dymarski, confer publication "Successive Orthogonalisations in the Multistage CELP Coder", ICASSP 92 Vol.1, pp I-61 - I-64.

However, in order to reduce the computational complexity of the gain vector G, it is possible to use suboptimal transforms, such as the transform of Fourier Rapide (FFT), the transform into a discrete cosine (TCD) the discrete transform of Hadamard (DHT) or Walsh Hadamard (DWHT) for example.

matrice dans laquelle h(n) est la réponse impulsionnelle du filtre de prédiction à court terme 1/A(z) de la fenêtre courante.Another method for the construction of an orthonormal transform consists in decomposing into singular values the lower triangular Toeplitz matrix H defined by:

matrix in which h (n) is the impulse response of the short-term prediction filter 1 / A (z) of the current window.

The matrix H can then be decomposed into a sum of matrices of rank 1:

The matrix U being unitary, this can be used as an orthonormal transform. Such a construction was proposed by B.S. Atal in the publication "A Model of LPC Excitation in Terms of Eigenvectors of the Autocorrelation Matrix of the Impulse Response of the LPC Filter ", ICASSP 89, Vol.1, pp 45-48 and by E.Ofer in the publication "A Unified Framework for LPC Excitation Representation in Residual Speech Coders "ICASSP 89, Vol.1 pp 41-44.

Encoders with nested codes currently known allow data to be transmitted by theft of elements binaries normally allocated to speech on the channel transmission in a transparent manner for the encoder, which encodes the speech signal at the maximum rate.

Among this type of encoder, a 64 kbit / s encoder with scaled quantizer with nested codes has been normalized to 1986 by standard G 722 established by the CCITT. This encoder operating in the field of wideband speech (50 Hz to 7 kHz bandwidth audio signal, sampled at 16 kHz), is based on coding in two sub-bands each containing a Pulse Modulation encoder and Adaptive Differential Coding (MICDA coding). This coding technique allows you to transmit signals from wideband speech and data, if necessary, on a 64 kbit / s channel, at three different bit rates 64-56-48 kbit / s and 0-8-16 kbit / s for data.

In addition, as part of the implementation of coders excited by codes (or CELP coders) M.Johnson and T. Tanigushi described a multistage CELP coder nested. Confer the publication of the aforementioned authors entitled "Pitch Orthogonal Code-Excited LPC", Globecom 90, Vol. 1, pp 542-546.

Finally, R. Drogo De Iacovo and D. Sereno described a modified CELP type coder for obtaining codes nested or modeling the filter excitation signal LPC analysis by a sum of different contributions and using only the first of them for delivery the memory of the synthesis filter, confer the publication of these authors "Embedded CELP Coding For Variable Bit-Rate Between 6.4 and 9.6 kbit / s "ICASSP 91 Vol. 1, pp 681-684.

Coders predictive by art transformation above do not allow the transmission of data and therefore cannot fulfill the function of coders with nested codes. In addition, coded coders nested in the prior art do not use the technique of the orthonormal transform, which does not allow to tend towards or reach an optimal transform coding.

The object of the present invention is to remedy the aforementioned drawback by the implementation of a system of predictive coding-decoding of a digital speech signal by adaptive transform with nested codes.

Another object of the present invention is the implementation implementation of a predictive coding-decoding system of a digital speech and data signal allowing transmission at reduced and flexible rates.

The predictive coding system for a digital signal into a digital signal with nested codes, in which the coded digital signal consists of a speech signal encoded and, if necessary, by an auxiliary data signal inserted into the coded speech signal after coding this last object of the present invention includes a filter perceptual weighting driven by a prediction loop short term to generate a perceptual signal and a long-term prediction circuit delivering a estimated perceptual signal, this long prediction circuit term forming a long-term prediction loop allowing to deliver, from the perceptual signal and the signal estimated past excitation, a perceptual excitation signal modeled and adaptive transform circuits and quantization allowing from the excitation signal perceptual to generate the coded speech signal.

It is remarkable in that the weighting filter perceptual consists of a prediction filter to short term of the speech signal to be coded, so as to achieve a frequency distribution of the quantization noise and in that it includes a circuit for subtracting the contribution of the excitation signal passed from the signal perceptual to deliver an updated perceptual signal, the long-term prediction circuit being formed, in a loop closed, from a dictionary updated by excitement past modeled corresponding to the lowest flow to deliver an optimal waveform and gain associated with it, constituting the estimated perceptual signal. The transform circuit is formed by a module of orthonormal transform including a transformation module orthogonal adaptive and a modeling module progressive by orthogonal vectors. The modeling module progressive and long-term prediction circuit allow the delivery of indexes representative of the signal coded speech. An auxiliary data insertion circuit is coupled to the transmission channel.

The transform predictive decoding system adaptive of a digital signal coded with codes nested in which the coded digital signal consists of a signal digital coded and, where appropriate, by a data signal auxiliaries inserted in the coded speech signal after coding of the latter, is remarkable in that it comprises a data signal extraction circuit enabling, hand, data extraction for use auxiliary and, on the other hand, index transmission representative of the coded speech signal. It includes in addition to a circuit for modeling the speech signal at minimum flow and a signal modeling circuit speech at least at a rate greater than the minimum rate.

The predictive coding-decoding system for a signal speech digital by adaptive transform to codes nested object of the present invention finds application, in general, to the transmission of speech and data at flexible rates, and more specifically, audio-visual conference protocols, videophone, speakerphone, storage and transport of digital audio signals over links long distances, when transmitting with mobiles and channel concentration systems.

• la figure 2 représente un schéma de principe du système de codage prédictif d'un signal de parole par transformée adaptative à codes imbriqués objet de la présente invention,
• la figure 3 représente un détail de réalisation d'un module de prédiction à long terme en boucle fermée utilisé dans le système de codage représenté en figure 2,
• les figures 4a et 4b représentent un schéma partiel d'un codeur prédictif par transformée et un schéma équivalent au schéma partiel de la figure 4a,
• la figure 5a représente un organigramme d'un processus de transformée orthonormée construit par apprentissage,
• la figure 5b représente deux diagrammes comparatifs de valeurs de gain normalisées obtenus par décomposition en valeurs singulières respectivement par apprentissage,
• les figures 6a et 6b représentent schématiquement le processus de transformation de Householder appliqué au signal perceptuel,
• la figure 7 représente un module de transformation adaptative mettant en oeuvre une transformation de Householder,
• la figure 8a représente, pour la décomposition en valeurs singulières respectivement la construction pour apprentissage, un critère normalisé de gain en fonction du nombre de composantes du vecteur de gains,
• la figure 8b représente un schéma de principe de quantification vectorielle multiétage dans lequel le vecteur de gains G and est obtenu par combinaison linéaire de vecteurs issus de dictionnaires stochastiques,
• la figure 9 est une représentation géométrique de la prospection du vecteur de gain G dans un sous-espace de vecteurs issus de dictionnaires stochastiques,
• les figures 10a e 10b représentent le schéma de principe d'un processus de quantification vectorielle de gain par modélisations progressives orthogonales, correspondant à une projection optimale de ce vecteur de gain représentée en figure 9, dans le cas d'un seul respectivement de plusieurs dictionnaires stochastiques,
• la figure 11 représente un mode de réalisation de la modélisation de l'excitation du filtre de synthèse correspondant au débit le plus faible,
• la figure 12 représente un schéma de principe d'un système de décodage prédictif d'un signal de parole par transformée adaptative à codes imbriqués objet de la présente invention,
• la figure 13a représente un schéma de principe d'un module de modélisation du signal de parole au débit minimum,
• la figure 13b représente un mode de réalisation d'un module de transformation orthonormée inverse,
• la figure 14a représente un schéma d'un module de modélisation du signal de parole aux débits autres que le débit minimum,
• la figure 14b représente un schéma équivalent au module de modélisation représenté en figure 14a,
• la figure 15 représente la mise en oeuvre d'un filtre adaptatif de post-filtrage destiné à améliorer la qualité perceptuelle du signal de parole de synthèse S andn.

A more detailed description of the coding-decoding system which is the subject of the invention will be given below in relation to the drawings in which, in addition to FIG. 1 relating to the prior art concerning a predictive transform coder,

• FIG. 2 represents a block diagram of the predictive coding system of a speech signal by adaptive transform with nested codes which is the subject of the present invention,
• FIG. 3 represents a detail of an embodiment of a long-term closed-loop prediction module used in the coding system represented in FIG. 2,
• FIGS. 4a and 4b represent a partial diagram of a predictive coder by transform and a diagram equivalent to the partial diagram of FIG. 4a,
• FIG. 5a represents a flow diagram of an orthonormal transform process constructed by learning,
• FIG. 5b represents two comparative diagrams of normalized gain values obtained by decomposition into singular values respectively by learning,
• FIGS. 6a and 6b schematically represent the Householder transformation process applied to the perceptual signal,
• FIG. 7 represents an adaptive transformation module implementing a Householder transformation,
• FIG. 8a represents, for the decomposition into singular values respectively the construction for learning, a normalized gain criterion as a function of the number of components of the gain vector,
• FIG. 8b represents a principle diagram of multi-stage vector quantization in which the gain vector G and is obtained by linear combination of vectors from stochastic dictionaries,
• FIG. 9 is a geometric representation of the prospecting of the gain vector G in a subspace of vectors originating from stochastic dictionaries,
• FIGS. 10a and 10b represent the block diagram of a vector quantization process of gain by orthogonal progressive modeling, corresponding to an optimal projection of this gain vector represented in FIG. 9, in the case of a single respectively of several dictionaries stochastic,
• FIG. 11 represents an embodiment of the modeling of the excitation of the synthesis filter corresponding to the lowest bit rate,
• FIG. 12 represents a block diagram of a system for predictive decoding of a speech signal by adaptive transform with nested codes object of the present invention,
• FIG. 13a represents a block diagram of a module for modeling the speech signal at the minimum rate,
• FIG. 13b represents an embodiment of a reverse orthonormal transformation module,
• FIG. 14a represents a diagram of a module for modeling the speech signal at bit rates other than the minimum bit rate,
• FIG. 14b represents a diagram equivalent to the modeling module represented in FIG. 14a,
• FIG. 15 represents the implementation of an adaptive post-filtering filter intended to improve the perceptual quality of the synthetic speech signal S andn.

A more detailed description of a predictive coding of a digital speech signal by adaptive transform into a digital code signal nested will now be given in connection with the figure 2 and the following figures.

Generally we consider that the signal digital coded by implementing the coding system object of the present invention consists of a signal speech coded and if necessary by a data signal auxiliaries inserted in the coded speech signal, after coding of this digital speech signal.

Of course, the object coding system of the present invention may include, from a transducer delivering the analog speech signal, a converter analog-digital and a memory circuit input or input buffer to deliver the signal digital to code Sn.

. The coding system object of the present invention also comprises a perceptual weighting filter 11 controlled by a short-term prediction loop making it possible to generate a perceptual signal, noted

.

It also includes a prediction circuit to long term, noted 13, delivering an estimated perceptual signal, which is noted P and 1 / n.

The long-term prediction circuit 13 forms a long-term prediction loop to deliver, to from the perceptual signal and the past excitation signal estimated, noted P and 0 / n, a perceptual excitation signal modeled.

The coding system which is the subject of the invention as shown in FIG. 2 further comprises an adaptive transform and quantization circuit making it possible, from the perceptual excitation signal P _n, to generate the coded speech signal as it will be described below in the description.

According to a first particularly advantageous aspect of the coding system which is the subject of the present invention, the perceptual weighting filter 11 consists of a filter for short-term prediction of the speech signal to be coded, so as to achieve a frequency distribution of the quantization noise. The perceptual weighting filter 11 delivering the perceptual signal , the coding device according to the invention thus comprises, as shown in the same FIG. 2, a circuit 120 for subtracting the contribution of the past excitation signal P and 0 / n from the perceptual signal to deliver an updated perceptual signal, this perceptual signal updated being noted P _n .

According to another particularly characteristic advantageous coding device object of this invention the long term prediction circuit 13 is formed in a closed loop from an updated dictionary by the past excitation modeled corresponding to the lowest bit rate, this dictionary allows to deliver an optimal waveform and an estimated gain associated therewith. In Figure 2, the corresponding modeled past excitation at the lowest flow rate is noted r and 1 / n. We indicate in in addition to the optimal waveform and associated estimated gain to it constitute the estimated perceptual signal P and 1 / n delivered by the long-term prediction circuit 13.

According to another characteristic of the coding system object of the present invention, as shown in Figure 2, the transform module circuit, denoted MT, is formed by an orthonormal transform module 14, comprising a properly adaptive orthogonal transformation module said and a progressive modeling module by orthogonal vectors, noted 16.

In accordance with a particularly advantageous aspect of the coding system which is the subject of the present invention, the progressive modeling module 16 and the circuit long-term prediction 13 allow to issue indexes representative of the coded speech signal, these indexes being denoted i (0), j (0) respectively i (1), j (1) with 1 ∈ [1, L] in figure 2.

Finally, the coding system according to the invention further includes a data insertion circuit 19 auxiliaries coupled to the transmission channel, noted 18.

The operation of the object coding device the present invention can be illustrated as follows.

As indicated previously, an attempt is made to reconstruct a synthetic signal S and _n as closely as possible perceptually resembling the digital signal to be coded Sn.

The synthetic signal S and _n is of course the signal reconstructed on reception, that is to say at the decoding level after transmission as will be described later in the description.

A short-term prediction analysis formed by the analysis circuit 10 of the LPC type for "Linear Predictive Coding" and by the perceptual weighting filter 11 is carried out for the digital signal to be coded by a conventional prediction technique on windows comprising for example M samples. The analysis circuit 10 then delivers the coefficients a _i , where the aforementioned coefficients a _i are the linear prediction coefficients.

The speech signal to be coded Sn is then filtered by the perceptual weighting filter 11 of transfer function W (z), which makes it possible to deliver the perceptual signal proper, noted .

The coefficients of the perceptual weighting filter are obtained from a short-term prediction analysis on the first correlation coefficients of the sequence of the coefficients a _i of the analysis filter A (z) of circuit 10 for the current window. This operation makes it possible to achieve a good frequency distribution of the quantization noise. In fact, the perceptual signal delivered tolerates greater coding noise in high-energy areas where the noise is less audible, since it is frequently masked by the signal. It is indicated that the perceptual filtering operation is broken down into two stages, the digital signal to code Sn being filtered a first time by the filter constituted by the analysis circuit 10, in order to obtain the residue to be modeled, then a second times by the perceptual weighting filter 11 to deliver the perceptual signal .

In the process of operating the device coding object of the present invention, the second operation is to then remove the contribution from the excitement past, or estimated past excitation signal, noted p and 0 / n of aforementioned perceptual signal.

Indeed, we show that:

In this relation, h _n is the impulse response of the double filtering performed by the circuit 10 and the perceptual weighting filter 11 in the current window and r and 1 / n is the past excitation modeled corresponding to the lowest flow rate, as well as 'It will be described later in the description.

The operating mode of the long prediction circuit term 13 in closed loop is then the following. This circuit allows to take into account the periodicity of the residue for voiced sounds, this long-term prediction being performed all the sub-windows of N samples, thus that it will be described in connection with FIG. 3.

par rapport aux deux paramètres g₀ et q.The long-term closed-loop prediction circuit 13 comprises a first stage constituted by an adaptive dictionary 130, which is updated all the aforementioned sub-windows by the modeled excitation denoted r and 1 / n, delivered by the module 17 , which will be described later in the description. The adaptive dictionary 130 makes it possible to minimize the error, noted

with respect to the two parameters g ₀ and q.

Such an operation corresponds, in the frequency domain, to filtering by the transfer function filter: B (z) = 1 1-g 0 z -q '

This operation is equivalent to the search for the optimal waveform, noted f ^{j (0)} and its associated gain g ₀ in a judiciously constructed dictionary. Confer the article published by R.Rose, and T.Barnwell, entitled "Design and Performance of an Analysis by Synthesis Class of Predictive Speech Coders", IEEE Trans. on Acoustic Speech Signal Proceessing, September 1990.

The waveform of index j, noted VS j not = r 1 nq from the adaptive dictionary, is filtered by a filter 131 and corresponds to the excitation modeled at the lowest rate r and 1 / n delayed by q samples by the aforementioned filter. The optimal waveform f 1 / n is delivered by the filtered adaptive dictionary 133.

A module 132 for calculating and quantifying the prediction gain makes it possible, from the perceptual signal P _n and all the waveforms fj (0) / n, to carry out a calculation for quantifying the prediction gain, and to deliver an index i (0) representative of the number of the quantization range, as well as its associated quantized gain g (0).

A multiplier circuit 134 delivers from adaptive dictionary filtered 133, that is to say of the result filtering the waveform with index j C j / n, i.e. f j / n, and the associated quantified gain g (0), the prediction excitation at long term modeled and filtered perceptually noted P and 1 / n.

A subtractor circuit 135 then makes it possible to perform a minimization relating to e _n = | P _n - P and 1 / n |, this expression representing the error signal. A module 136 makes it possible to calculate the Euclidean norm | e _n | ² .

A module 137 makes it possible to search for the optimal waveform corresponding to the minimum value of the above-mentioned Euclidean standard and to deliver the index j (0). The parameters transmitted by the coding system object of the invention for modeling the long-term prediction signal are then the index j (0) of the optimal waveform f ^j (0) as well as the number i ( 0) of the quantization range of its associated gain g (0) quantized.

A more detailed description of the transformation module orthogonal adaptive MT of Figure 2 will given in conjunction with Figures 4a and 4b.

As part of the implementation of the predictive coding by orthonormal transform object of the present invention the method used for construction of this transform corresponds to that proposed by B.S.Atal and E.Ofer, as previously mentioned in the description.

According to the embodiment of the coding system according to the present invention, this consists in decomposing, not the short-term prediction filtering matrix, but the perceptual weighting matrix W formed by a lower triangular Toeplitz matrix defined by the relation (4):

In this relation, w (n) denotes the response of the perceptual weighting filter W (z) of the current window previously mentioned.

In Figure 4a, there is shown the partial diagram of a predictive transform coder and in FIG. 4b, the corresponding equivalent scheme in which the matrix or perceptual weighting filter W, designated by 140, has been highlighted, a perceptual weighting filter reverse 121 having, however, been inserted between the long-term prediction 13 and the subtractor circuit 120. It is indicated that the filter 140 performs a combination linear of the basic vectors obtained from a decomposition into singular values of the matrix representative of the perceptual weighting filter W.

As shown in FIG. 4b, the signal S ', corresponding to the speech signal to be coded S _{n from} which it has been subtracted the contribution of the past excitation delivered by the module 12, as well as that of the long-term prediction P and 1 / n filtered by a perceptual transfer weighting module with transfer function (W (z)) ^-1 , is filtered by the perceptual weighting filter with transfer function W (z), so as to obtain the vector P ' .

This filtering operation is written: P '= WS' and can be expressed as a linear combination of basic vectors using the decomposition into singular values of the matrix W.

As far as the embodiment of the perceptual weighting filter 140 is concerned, it is indicated that this comprises for each matrix W representative of the perceptual weighting filter a first matrix module U = (U ₁ , ..., U _N ) and a second matrix module V = (V ₁ , ..., V _N ).

• U^T désigne le module matrice transposée du module U,
• D est un module matrice diagonale dont le coefficients constituent les valeurs singulières précitées,
• U_i et V_j désignent respectivement le ième vecteur singulier gauche et le jème vecteur singulier droit, les vecteurs singuliers droit { Vj} formant une base orthonormée.

The first and second matrix modules verify the relationship: U T WV = D relationship in which:

• U ^T designates the matrix module transposed from the module U,
• D is a diagonal matrix module whose coefficients constitute the aforementioned singular values,
• U _i and V _j respectively designate the ith left singular vector and the jth right singular vector, the right singular vectors {Vj} forming an orthonormal base.

Such a breakdown makes it possible to replace the operation of filtering by convolution product by an operation filtering by a linear combination.

We indicate that the decomposition into singular values of the perceptual filtering matrix W makes it possible to obtain the two unit matrices U and V verifying the above-mentioned relation where U T WV = diag (d 1 , ..., d NOT ) with the scheduling property such that d _i ≥ d _{i + 1} > 0. The elements d _i are called singular values, and the vectors U _i and V _j , ith left singular vector, respectively jth right singular vector.

The matrix W is then decomposed into a sum of matrices of rank 1, and verifies the relation:

permet d'obtenir le vecteur P' vérifiant la relation :

   avec g(k) =

(k)d_k.The matrix V being unitary, the right singular vectors {V _i } form an orthonormal base and the signal S ', expressed in the form:

allows to obtain the vector P 'verifying the relation:

with g (k) =

(k) d _k .

By the process of decomposition into values singular, we indicate that a change on a component of excitation S 'associated with a small singular value produces a small change at the output of filter 140 and vice versa for the reverse perceptual filtering operation performed by module 121.

In order to use these properties, the unitary matrix U can be used as an orthonormal transform, checking the relation: F = [f 1 orth , ..., f NOT orth ], that is to say: fi / orth = U _i for i = 1 to N.

The weighted perceptual signal P 'then breaks down as follows: G = U T P '.

After vector quantization of the gains G, the weighted perceptual signal modeled P and is calculated in the following manner: P '= F' G = U ' G .

We indicate that the left singular vectors associated with the largest singular values play a preponderant role in the modeling of the weighted perceptual signal P '. Thus, in order to model the latter, it is possible to keep only the components associated with the K largest singular values, K <N, that is to say the K first components of the gain vector G verifying the relation: G = (g 1 , g 2 ... g K , 0, ... 0).

The short-term analysis filter circuit 10 being updated on windows of M samples, the decomposition into singular values of the matrix of W perceptual weighting is performed at the same frequency.

Decomposition processes into singular values any matrix allowing rapid processing have been developed, but the calculations remain relatively complex.

In accordance with an object of the present invention, it is, in order to simplify the processing operations mentioned above, proposed to build an orthonormal transform fixed sub-optimal, however with good properties perceptual, regardless of the current window.

In a first embodiment, such as shown in Figure 5, the orthonormal transform process is built by learning. In such a case, the orthonormal transform module can be formed by a stochastic transform submodule constructed by drawing of a Gaussian random variable for initialization, this sub-module comprising in FIG. 5 the steps 1000, 1001, 1002 and 1003 and being noted SMTS. Step 1002 can consist in applying the algorithm of the K-mean over the aforementioned vector corpus.

The SMTS sub-module is successively followed by module 1004 for building centers, a module 1005 for construction of classes and, in order to obtain a vector G whose components are relatively ordered, of a transform reordering module 1006 according to the cardinal of each class.

The aforementioned module 1006 is followed by a module of Gram-Schmidt calculation, noted 1007a, so as to obtain a orthonormal transform. The aforementioned module 1007a is associated a module 1007b for calculating the error under the conditions conventional implementation of the treatment process Gram-Schmidt.

The 1007a module is itself followed by a 1008 module test on the number of iterations, this to allow obtain an orthonormal transform performed offline by learning. Finally, memory 1009 of memory type dead allows to memorize the orthonormal transform under transform vector shape. It is indicated that the scheduling relative of the components of the gain vector G is accentuated by the orthogonalization process. When the learning-building process has converged, we obtains an orthonormal transform whose waveforms are gradually correlated with the learning corpus vectors delivered by step 1001 of transform initial.

Figure 5b shows the scheduling of components of the gain vector G, i.e. the value G normalized mean for a transform obtained on the one hand by decomposition into singular values of the matrix of perceptual weighting W, and on the other hand, by learning. The transform F obtained by this last method for those of orthonormal waveforms whose frequency spectra are bandpass and relatively ordered as a function of k, which allows to attribute to this transformed pseudo-frequency properties. A evaluation of the quality of processing in terms of energy concentration has shown that indicative, on a corpus of 38,000 perceptual vectors P ', the transformation gain is 10.35 decibels for the Karhunen-Loeve transform, and 10.29 decibels for a transform built by learning, this last therefore tending towards the optimal transform in terms of energy concentration.

As previously mentioned in the description, the orthonormal transform F can be obtained according to two different methods.

By observing that generally the waveform most correlated with the perceptual signal P is that resulting from the adaptive dictionary, it is possible to envisage carrying out an adaptive orthonormal transform F 'whose f' 1 / orth is equal to the form d optimal wave from the normalized adaptive dictionary f ^j (0), the first component of the gain vector G then being equal to the normalized long-term prediction gain g (0), which it is not necessary to recalculate, since this was quantified during this prediction.

The new dimension of the gain vector G becomes then equal to N-1, which increases the number bits per sample during quantization vector of it and therefore the quality of its modelization.

A first solution to calculate the transform F 'can then consist in making a prediction analysis at long term, to shift the transform obtained by learning up a notch, to place the long-term predictor at the first position, then apply the Gram-Schmidt algorithm, in order to obtain a new transform F '.

A second, more advantageous solution consists in using a transformation making it possible to rotate the orthonormal base, so that the first waveform coincides with the long-term predictor, that is to say: F '= TF with f ' 1 orth = f d (0) f d (0) = Tf 1 orth

In order to keep the orthogonality property, the transformation used must keep the dot product. A particularly suitable transformation is the Householder transform verifying the relation: T = I - 2 BB T B T B with B = f ^{j (0)} - | f ^{j (0)} | - f 1 / orth.

A geometric representation of the transform above is given in Figures 6a and 6b.

For a more detailed definition of this type of transformation, we can usefully refer to the Alan O.Steinhardt's publication "Householder Transforms in Signal Processing ", IEEE ASSP Magazine, July 1988, pp 4-12.

By using this transformation, it is possible to reduce the complexity of the calculations and the projection of the perceptual signal P in this new base is written: G = F ' T P = F T TP = F T P '' with P '= TP = (PB [wB ^T P]).

In this relation, w denotes a scalar equal to w = 2 / B ^T B.

It is indicated that in this embodiment of the orthonormal transform, the transformation is not applied than to the perceptual signal P, and the modeled perceptual signal P and can then be calculated by the inverse transformation.

A particularly advantageous embodiment of the orthonormal transform module itself 14 in the case where a Householder transformation is used will now be described in conjunction with FIG. 7.

As shown in FIG. 7 above, the adaptive transformation module 14 can include a Householder transformation module 140 receiving the estimated perceptual signal constituted by the optimal waveform and by the estimated gain and the signal perceptual P to generate a transformed perceptual signal P ''. It is indicated that the Householder transformation module 140 comprises a module 1401 for calculating parameters B and wB as defined previously by relation 13. It also includes a module 1402 comprising a multiplier and a subtractor making it possible to carry out the transformation proper according to the relation 14. It is indicated that the transformed perceptual signal P '' is delivered in the form of a vector of perceptual transformed signal of component P '' _k , with k ∈ [0, N-1].

The adaptive transformation module 14 such as represented in FIG. 7 also includes a plurality N of registers for memorizing orthonormal waveforms, the current register being noted r, with r ∈ [1, N]. We indicate that the aforementioned N storage registers form the read-only memory previously described in the description, each register comprising N storage cells, each component of rank k of each vector, component denoted f 1 / orth (k) being stored in a row cell correspondent of the current register r considered.

In addition, as will be seen in FIG. 7, the module 14 comprises a plurality of N multiplier circuits associated with each register of rank r forming the plurality of the previously mentioned storage registers. In addition, each rank multiplier register k receives on the one hand the component of rank k of the stored vector and on the other hand the component P '' _k of the transformed perceptual signal vector of corresponding rank k. The Mrk multiplier circuit delivers the product P '' _k .fk / orth (k) of the components of the transformed perceptual signal.

Finally, a plurality of N-1 summing circuits is associated with each register of rank r, each circuit summator of rank k, noted Srk, receiving the product of rank anterior k-1, and the product of corresponding rank k delivered by the multiplier circuit Mrk of the same rank k. The circuit summator of highest rank, SrN-1, then delivers a component g (r) of the estimated gain expressed as a vector G gain

It is indicated that the predictive coding system using the adaptive orthonormal transform constructed by learning is likely to give better results, while the transformation of Householder allows to obtain reduced complexity.

As will be observed in FIG. 2, the progressive modeling module by orthogonal vectors in fact comprises a module 15 for normalizing the gain vector to generate a normalized gain vector, denoted G _k , by comparison of the normalized value of the gain vector G with respect to a threshold value. This normalization module 15 also makes it possible to generate a signal of length of the normalized gain vector linked to the modeling order k to the decoder system as a function of this modeling order.

The progressive vector modeling module orthogonal further comprises, in cascade with the module 15 normalization of the gain vector, a stage 16 of progressive modeling by orthogonal vectors. This floor model 16 receives the normalized vector Gk and delivers the indexes representative of the coded speech signal, these indexes being denoted I (1), J (1), these indexes being representative selected vectors and their associated gain. The transmission of auxiliary data formed by indexes is performed by overwriting the parts of the allocated frame to the indices and track numbers to form the signal auxiliary data.

The operation of the standardization module 15 is the following.

où G_k = (0,g₂,g₃,...,g_k,0,...0).The energy of the perceptual signal given by | P '| 2 = | G | 2 is constant for a given sub window. Under these conditions, maximizing this energy is equivalent to minimizing the expression:

where G _k = (0, g ₂ , g ₃ , ..., g _k , 0, ... 0).

It is indicated that during such an operation, an additional way of increasing the number of binary elements per sample during the vector quantization of the vector G is to use the following normalized criterion, consisting in choosing K such that:

The gain vector thus obtained G _K is then quantified and its length k is transmitted by the coding system object of the invention in order to be taken into account by the corresponding decoding system, as will be described later in the description.

The average standard criterion according to the order of modeling K is given in figure 8a for a transform orthonormal obtained on the one hand by decomposition into values singulars of the perceptual weighting matrix W and on the other hand by learning.

A particularly advantageous embodiment of the progressive modeling module using orthogonal vectors 16 will now be given in connection with FIG. 8b. The aforementioned module allows in fact to perform a quantification vector illustration.

The gain vector G and is obtained by linear combination of vectors, noted Ψ j K = (0, Ψ 2 j , Ψ 3 j , ..., Ψ K j , 0.0 ... 0).

These vectors being derived from stochastic dictionaries, noted 161, 162, 16 L, constructed either by drawing a Gaussian random variable, or by learning. The estimated gain vector G and verifies the relation:

In this relation,  ₁ is the gain associated with the optimal vector Ψ j (1) / K from the stochastic dictionary of rank l, noted 16 l.

However, the vectors selected iteratively are generally not linearly independent and do not therefore not form a basis. In such a case, the subspace generated by the L optimal vectors Ψ j (L) / K is of dimension less than L.

In FIG. 9 the projection of the vector G on the subspace generated by the vectors optimal of rank l, respectively l-1, this projection being optimal when the aforementioned vectors are orthogonal.

It is therefore particularly advantageous to orthogonalize the stochastic dictionary of rank l with respect to optimal vector of the stage of previous rank Ψ j (l-1) / K.

Thus, whatever the optimal vector of rank l from the new dictionary or stage of corresponding rank l, it will be orthogonal to the optimal vector Ψ j (l-1) / K of previous rank, and we obtain: ψ j orth (l + 1) = ψ j orth (l) - r (l, j) ψ j (l) orth (l) √α j (l) l

In this relation, we indicate that: α j (l) l = ψ j (l) orth (l) 2 corresponds to the energy of the wave selected in step l, r (l, j) = < ψ j (l) orth (l) √α j (l) l , ψ j orth (l) > represents the intercorrelation of the optimal vectors of rank j and of rank j (l) and T K = I NOT - ψ j (l) orth (l) √α j (l) l (ψ j orth (l) ) T represents the orthogonalization matrix.

The previous operation allows to remove from the dictionary the contribution of the previously selected wave and thus imposes linear independence for any vector optimal of rank i between l + 1 and L compared to optimal vectors of lower rank.

Principle diagrams of quantification vector by orthogonal progressive modeling are given in Figures 10a and 10b depending on whether there is one or several stochastic dictionaries.

In order to reduce the complexity of the vector quantization, we indicate that the Recursive modified Gram-Schmidt can be used as well as proposed by N. Moreau, P.Dymarski, A.Vigier, in the publication titled: "Optimal and Suboptimal Algorithms for Selecting the Excitation in Linear Predictive Products ", Proc. ICASSP 90, pp 485-488.

Given the orthogonalization properties, we show that:

Given this expression, the algorithm of Recursive modified Gram-Schmidt as previously proposed can be used.

It is no longer necessary to recalculate explicitly the dictionaries at each stage of orthogonalisation.

The above calculation process can be explained in matrix form from the matrix

We indicate that Q is an orthonormal matrix, and R an upper triangular matrix whose elements of the main diagonal are all positive, which ensures the uniqueness of the decomposition.

ce qui implique que R =

.The gain vector G verifies the matrix relation:

which implies that R =

.

The upper triangular matrix R thus makes it possible to recursively calculate the gains  (k) relative to the base of origin.

} sont ordonnés de façon décroissante. Le résidu peut être modélisé de façon graduelle de la façon ci-après où  and cod / k désigne le gains associé au vecteur optimal orthogonal Ψ j(k) / orth(k) quantifié, compte tenu des relations :

G = G 1 + G 2 + G 3 (Débit le plus élevé) G = G 1 + G 2 (Débit intermédiaire) G = G 1 (Débit le plus faible)    avec 1 ≤ L₁ ≤ L₂ ≤ L.The contribution of the optimal vectors to the orthonormal base, noted: {Ψ j (l) / orth (L)} in the modelization of the gain vector G _K tends to decrease, and the gains {

} are ordered in descending order. The residue can be modeled gradually in the following way where  and cod / k denotes the gains associated with the optimal orthogonal vector Ψ j (k) / orth (k) quantified, taking into account the relationships:

G = G 1 + G 2 + G 3 (Highest speed) G = G 1 + G 2 (Intermediate flow) G = G 1 (Lowest flow) with 1 ≤ L ₁ ≤ L ₂ ≤ L.

. La transmission des données se fait alors en écrasant les parties de la trame allouées aux indices et numéros de plages j(l), i(l), pour l ∈ [L1, L2-1] et [L2,L] selon les besoins de la communication.We then obtain the gain vectors G and ¹ , G and ² , G and ³ orthogonal whose contribution in the modeling of the gain vector G is decreasing, which allows the gradual modeling of the residue r _n effectively. The parameters transmitted by the coding system which is the subject of the invention for modeling the gain vector G are then the indices j (l) of the selected vectors as well as the numbers i (l) of the ranges of quantification of their associated gains,

. Data transmission is then done by overwriting the parts of the frame allocated to the indices and track numbers j (l), i (l), for l ∈ [L1, L2-1] and [L2, L] as required. of communication.

}, il est nécessaire de coder les différents gains précités g(0) et { }. Une étude a montré que les gains relatifs à la base orthogonale {Ψ j(l)    / orth(L)} étant décorrélés, ceux-ci possèdent de bonnes propriétés pour leur quantification. En outre, la contribution des vecteurs optimaux dans la modélisation du vecteurs de gain G ayant tendance à décroítre, les gains { } sont ordonnés de façon relativement décroissantes, et il est possible d'utiliser cette propriété en codant non pas les gains précités, mais leur rapport donné par

Plusieurs solutions peuvent être utilisées pour coder les rapports précités.The previously mentioned processing process uses the recursive modified Gram-Schmidt algorithm in order to code the gain vector G. The parameters transmitted by the coding system according to the invention being the aforementioned indices, j (0) to j (L ) of the different dictionaries as well as the quantified gains g (0) and {

}, it is necessary to code the various aforementioned gains g (0) and { }. A study has shown that the gains relative to the orthogonal base {Ψ j (l) / orth (L)} being decorrelated, these have good properties for their quantification. In addition, the contribution of the optimal vectors in the modeling of the gain vectors G tending to decrease, the gains { } are ordered in a relatively decreasing order, and it is possible to use this property by coding not the aforementioned gains, but their ratio given by

Several solutions can be used to code the aforementioned reports.

As will be seen in Figure 2, the coding device object of the present invention comprises a module for modeling the excitation of the filter summary corresponding to the lowest throughput, this module being noted 17 in the aforementioned figure.

The principle diagram for calculating the excitation signal of the synthesis filter corresponding to the lowest bit rate is given in FIG. 11. An inverse transformation is applied to the gain vectors modeled G and ¹ , this inverse adaptive transformation can for example correspond to a reverse transformation of the Householder type, which will be described later in the description, in conjunction with the decoding device which is the subject of the present invention. The signal obtained after inverse adaptive transformation is added to the long-term prediction signal B '1 / n by means of a summator 171, the estimated perceptual signal or long-term prediction signal being delivered by the long-term prediction circuit 13 closed loop term. The resulting signal delivered by the adder 171 is filtered by a filter 172, which corresponds from the point of view of the transfer function to the filter 131 of FIG. 3. The filter 172 delivers the residual signal modeled r and 1 / n.

A transformative predictive decoding system adaptive to nested codes of a coded digital signal consisting of a coded speech signal, and where appropriate, by an auxiliary data signal inserted into the signal speech coded after coding of the latter will now described in conjunction with Figure 12.

According to the aforementioned figure, the decoding system comprises a circuit 20 for extracting the data signal allowing on the one hand the extraction of the data for an auxiliary use, by an output of the auxiliary data and, on the other hand , the transmission of indexes representative of the coded speech signal. It is of course understood that the aforementioned indexes are the indices i (l) and j (l), for l between 0 and L ₁ -1 previously described in the description and for l between l ₁ and L under the conditions which will be described below. As shown in addition in FIG. 12, the decoding system according to the invention comprises a circuit 21 for modeling the speech signal at the minimum bit rate, as well as a circuit 22 or 23 for modeling the speech signal at at least one flow greater than the aforementioned minimum flow.

In a preferred embodiment, such as represented in FIG. 12, the decoding system according to the invention comprises, in addition to the system for extracting data 20, a first signal modeling module 21 speech at minimum bit rate receiving signal directly coded and delivering a first estimated speech signal, noted S and 1 / n and a second module 22 for modeling the signal of speech at an intermediate rate connected to the extraction system 20 data via a circuit 27 conditional switching on actual flow criteria allocated to the speech signal and delivering a second signal estimated speech, noted S and 2 / n.

The decoding system shown in Figure 12 also includes a third modeling module 23 of the speech signal at maximum rate, this module being connected to the data extraction system 20 via of a conditional switching circuit 28 on criterion of the actual bit rate allocated to speech and delivering a third estimated speech signal S and 3 / n.

In addition, a summing circuit 24 receives the first, the second and the third estimated speech signal, and delivers at its output a resulting estimated speech signal, noted S and _n . At the output of the summing circuit 24 are connected in cascade an adaptive filtering circuit 25 receiving the resulting estimated speech signal S and _n and delivering a reconstituted estimated speech signal, denoted S and ' _n . A digital-to-analog converter 26 may be provided to receive the reconstructed speech signal and to output an audio-frequency reconstituted speech signal.

According to a particularly advantageous characteristic of the decoding device which is the subject of this invention, each of the signal modeling modules of speech at a minimum, intermediate and maximum rate, i.e. modules 21, 22 and 23 of Figure 12, includes a inverse adaptive transformation sub-module, followed by a inverse perceptual weighting filter.

The block diagram of the modeling module of the speech signal at minimum rate is given in figure 13a.

In general, the object decoding system of the present invention takes into account the constraints imposed by data transmission at the level of coding system and in particular at the dictionary level adaptive, as well as the contribution of past excitement.

The speech signal modeling circuit at minimum flow 21 is identical to that described relatively to circuit 17 of the coding system according to the invention at from an inverse adaptive transformation module similar to module 170 described in relation to the figure 11. We simply note that in Figure 13a, we have explained obtaining the perceptual signal P and 1 / n from the indexes {i (0), j (0)}, of the order of modeling K and indices i (l), j (l) for l = 1 at L1-1.

As regards the inverse adaptive transformation, an advantageous embodiment of this is shown in FIG. 13b. It is indicated that the embodiment represented in FIG. 13b corresponds to a reverse Householder type transform using elements identical to the Householder transform represented in FIG. 7. It is simply indicated that for a perceptual signal delivered by the long-term prediction circuit 13, this signal being denoted P and ¹ entering a similar module 140, the signals entering the module 1402, respectively at the level of the multipliers associated with each register, are inverted. The resulting signal delivered by the summator corresponding to the summator 171 of FIG. 11 is filtered by a filter of inverse transfer function of the transfer function of the perceptual weighting matrix and corresponding to the filter 172 of the same FIG. 11.

The speech signal modeling modules at intermediate flow or at maximum flow, module 22 or 23, are shown in Figures 14a and 14b.

Of course, it is possible for reasons of complexity to group the different models of the speech signal corresponding to the other bit rates in a single block as shown in FIG. 14a and 14b. According to the actual bit rate allocated to speech, the gain vectors modeled G and ² , G and ³ are added, as shown in FIG. 14b, by a summator 220, subjected to the process of inverse adaptive transformation in a module 221 identical to the module 210 of FIG. 13a, then filtered by the inverse weighting filter W ^-1 (z) previously mentioned, this filter being designated by 222, the filtering starting from zero initial conditions, which makes it possible to perform an operation equivalent to multiplication by the inverse matrix W ^-1 , in order to obtain a progressive modeling of the synthesis signal S and _n . Note in FIG. 14b the presence of switching devices, which are none other than the switching devices 24 and 28 shown in FIG. 12, which are controlled as a function of the actual bit rate of the data transmitted.

Finally, as regards the adaptive filter 25, a particularly advantageous embodiment is given in FIG. 15. This adaptive filter makes it possible to improve the perceptual quality of the synthesis signal S and _n obtained following the summation by the summator 24. Such a filter comprises for example a long-term post-filtering module denoted 250, followed by a short-term post-filtering module and an energy control module 252, which is controlled by a module 253 for calculating the scale factor. Thus, the adaptive filter 25 delivers the filtered signal S and ' _n , this signal corresponding to the signal in which the quantization noise introduced by the coder on the synthesized speech signal has been filtered in the places of the spectrum where this is possible. It is indicated that the diagram represented in figure 15 corresponds to the publications of JHChen and A.Gersho, "Real Time Vector APC Speech Coding at 4800 Bps with Adaptative Postfiltering", ICASSP 87, Vol.3, pp 2185-2188.

We have thus described a predictive coding system by orthonormal transform with nested codes allowing to bring new solutions in the field of nested code encoders. We indicate that in a way general, the coding system object of the invention allows wideband coding at speech / data rates of 32/0 kbit / s, 24/8 kbit / s and 16/16 kbit / s.

Claims (10)

1. System for predictive coding of a digital signal as an embedded-code digital signal, coded by embedded-code adaptive transformation, in which the coded digital signal consists of a coded speech signal and, if appropriate, of an auxiliary data signal inserted into the coded speech signal after coding the latter, a system of the type including a perceptual weighting filter (11) driven by a short-term prediction loop allowing the generation of a perceptual signal

   

   and a long-term prediction circuit delivering an estimated perceptual signal P and 1 / n, this long-term prediction circuit forming a long-term prediction loop making it possible to deliver, from the perceptual signal and from the estimated past excitation signal, a modelled perceptual excitation signal, and adaptive transform and quantization means making it possible from the perceptual excitation signal to generate the coded speech signal, characterized in that the perceptual weighting filter consists of a filter for short-term prediction of the speech signal to be coded, so as to produce a frequency distribution of the quantization noise, and in that it comprises a means (12) for subtracting the contribution of the past excitation signal P and 0 / n from the said perceptual signal to deliver an updated perceptual signal P_n, and in that the long-term prediction circuit is formed, as a closed loop, from a dictionary updated by the modelled past excitation corresponding to the lowest throughput making it possible to deliver an optimal waveform and an estimated gain associated therewith, which make up the estimated perceptual signal, and in that the transform means are formed by an orthonormal transform module including an adaptive orthogonal transformation module and a module for progressive modelling by orthogonal vectors, these means of progressive modelling and the long-term prediction circuit making it possible to deliver indices representing the coded speech signal, the said system furthermore including means (19) for inserting auxiliary data, coupled to the transmission channel.
2. Coding system according to Claim 1, characterized in that the said adaptive orthogonal transformation module includes:

   a filter producing a linear combination of the basis vectors obtained from a singular-value decomposition of the matrix representing the perceptual weighting filter.
3. Coding system according to Claim 2, characterized in that the said filter comprises, for every matrix W representing the perceptual weighting filter:

   a first matrix module U = (U₁,...,U_N) and

   a second matrix module V = (V₁,...,V_N), the said first and second matrix modules satisfying the relation: UTWV = D in which

   U^T

   denotes the matrix transpose module of the module U and where

   D

   is a diagonal matrix module whose coefficients constitute the said singular values,

   U_i and V_j denoting respectively the i^th left singular vector and the j^th right singular vector, the said right singular vectors {V_j} forming an orthonormal basis, thus making it possible to transform the operation for filtering by convolution product by an operation for filtering by a linear combination.
4. Coding system according to Claim 1, characterized in that the said orthonormal transform module is formed by:

   a stochastic transform sub-module constructed by drawing a Gaussian random variable, for initialization,

   a module for global averaging over a plurality of vectors arising from a predictive transform coder,

   a reordering module,

   a Gram-Schmidt processing module, one reiteration of the processing by the preceding modules making it possible to obtain an orthonormal transform, performed off-line, formed by learning,

   a memory of read-only memory type, making it possible to store the orthonormal transform in the form of transformed vectors.
5. Coding system according to Claim 4, characterized in that the said transform is formed by orthonormal waveforms whose frequency spectra are band-pass and relatively ordered, the first waveform of the relatively ordered orthonormal waveforms being equal to the normalized optimal waveform arising from the said adaptive dictionary and the first component of estimated gain is equal to the normalized long-term prediction gain.
6. Coding system according to Claims 2 and 5, characterized in that the said adaptive transformation module includes:

   a Householder transformation module receiving the said estimated perceptual signal p and 1 / 1 consisting of the said optimal waveform and of the said estimated gain, and the said perceptual signal to generate a transformed perceptual signal P" in the form of a transformed perceptual signal vector with component P"_k, and

   a plurality of N registers for storing the said orthonormal waveforms, the said plurality of registers forming the said read-only memory, each register of rank r including N storage cells, a component of rank k of each vector being stored in a cell of corresponding rank,

   a plurality of N multiplier circuits associated with each register forming the said plurality of storage registers, each multiplier circuit of rank k receiving, on the one hand, the component of rank k of the stored vector and, on the other hand, the component P"_k of the transformed perceptual signal vector of rank k, and delivering the product P"_k.f^k _orth(k) of the transformed perceptual signal vector components,

   a plurality of N-1 summing circuits associated with each register of rank r, each summing circuit of rank k receiving the product of previous rank k-1 delivered by the multiplier circuit of previous rank and the product of corresponding rank k delivered by the multiplier circuit of previous rank and the product of corresponding rank k delivered by the multiplier circuit of like rank k, the summing circuit of highest rank, N-1, delivering a component g(r) of the estimated gain, expressed in the form of the gain vector G.
7. System according to Claim 1, characterized in that the said module for progressive modelling by orthogonal vector includes:

   a module for normalizing the gain vector to generate a normalized gain vector Gk, by comparing the normed value of the gain vector G with respect to a threshold value, the said normalization module making it possible to generate furthermore a length signal for the normalized gain vector Gk, destined for the decoder system as a function of the order of modelling,

   a stage for progressive modelling by orthogonal vectors properly speaking receiving the said normalized vector Gk and delivering the said indices representing the coded speech signal, the said indices being representative of the selected vectors and of their associated gains, the transmission of the auxiliary data formed by the indices being performed by overwriting the parts of the frame allocated to the indices and range numbers to form the auxiliary data signal.
8. System for predictive decoding by adaptive transform of a digital signal coded with embedded codes in which the coded digital signal consists of a coded speech signal and, if appropriate, of an auxiliary data signal inserted into the coded speech signal after coding the latter, characterized in that it comprises:

   means for extracting the said data signal making it possible, on the one hand, to extract the said data with a view to an auxiliary use, and on the other hand, to transmit the indices representing the coded speech signal,

   means for modelling the speech signal at the minimum throughput,

   means for modelling the speech signal at at least one throughput above the minimum throughput.
9. Decoding system according to Claim 8, characterized in that the decoder includes, apart from the data extraction system,

   a first module for modelling the speech signal at the minimum throughput, receiving the coded signal directly and delivering a first estimated speech signal S and 1 / n,

   a second module for modelling the speech signal at an intermediate throughput connected with the said data extraction system by way of means for conditional switching by criterion of the value of the said indices, and delivering a second estimated speech signal S and 2 / n,

   a third module for modelling the speech signal at a maximum throughput, connected with the said data extraction system by way of means for conditional switching by criterion of the value of the said indices and delivering a third estimated speech signal S and 3 / n,

   a summing circuit receiving on its summing inputs the first, the second respectively the third estimated speech signal and delivering at its output a resultant estimated speech signal and which are cascaded at the output of the said summing circuit,

   an adaptive filtering circuit receiving the said resultant estimated speech signal and delivering a reproduced estimated speech signal, and a digital/ analog converter receiving the said reproduced estimated speech signal and delivering an audio frequency reproduced speech signal.
10. Decoding system according to Claim 9, characterized in that each of the minimum, intermediate or maximum throughput speech signal modelling modules comprises an inverse adaptive transformation sub-module followed by an inverse perceptual weighting filter.

Images