EP0734013A2 - Determination of an excitation vector in a CELP coder - Google Patents

Abstract

The determination method involves using an optimum vector (Copt) from a subset of excitation vectors suitable for maximising a given criterion. These vectors have component values +1 and -1 corresponding to an excitation sample sequence of a linear prediction filter. The criterion is given by the square of the ratio of the scalar product of the excitation vector with the target vector (T) to the modulus of the excitation vector after being passed through a direct linear prediction filter. The filtering process involves preselecting an excitation vector (CO) with components having the same sign as the samples corresponding to the target vector, or having the inverse sign. If the selected vector does not belong to the subset, a vector is selected from the subset (CO'-CN') which maximises the given criterion and those nearest to it in the full set (C1-CN).

Description

The present invention relates to the compression of speech signals to be transmitted over a telephone line and more particularly to the determination of an excitation vector in the context of compression according to the method of Linear Prediction Excited by Code (CELP).

FIG. 1 very schematically represents a CELP compression circuit. Such a circuit is based on a modeling of the vocal cords and of the resonance chamber constituted by the oral cavities, the throat and the larynx. Such a compression method is therefore optimized for the processing of speech signals.

The oral, throat and larynx cavities are modeled by a "linear prediction" filter 10, the transfer function of which generally comprises ten poles. The vocal cords are modeled by an excitation E treated by a comb filter 12.

A digitized speech signal S is analyzed in frames by an analysis circuit 14. At each frame, the analysis circuit 14 determines coefficients a ₁ to a ₁₀ of the transfer function of the filter 10, the step p of the filter comb 12, and a gain G applied to the excitation E at 16 at the input of the filter 12. The values a _i , p and G are calculated for each frame to take account respectively of variations in the oral cavity, the frequency spectrum of the vocal cords and the amplitude of the sound. In this way an attempt is made to obtain that the output of the filter 10 is equal to the signal S. Then, instead of transmitting the samples of the signal S, the coefficients a _i , p, and G are transmitted so that a decoder which receives these coefficients reconstruct the corresponding frames of the signal S.

Of course, the decoder must also know the excitation E to be used. The determination of the coefficients ai, p and G does not pose any particular problem. However, the procedure for finding the optimal excitation remains the heaviest in terms of computational load and it is always of great interest to simplify it even at the cost of a significant reduction in the quality of the compression.

At the origin of the CELP coding, the excitation E was selected in a table 18 (called "codebook") containing several possible excitations which represented, in fact, pieces of white noise. In this case, a control circuit 20 scans the table 18 until the difference e, formed at 22, between the current frame of the signal S and the corresponding frame at the output of the filter 10 is minimal. (Of course, instead of comparing the signal S at the output of the filter 10, one can also compare the excitation E to the frame of the signal S having undergone the reverse processing of the filters 10 and 12.)

With this technique, in addition to the coefficients a _i , p and G, the address C selecting the best excitation E in table 18 is supplied to a decoder provided with a homologous table.

Each excitation contained in the table 18 is a sequence of digital samples corresponding respectively to the samples of each of the frames of the signal to be compressed. For the compression to be of acceptable quality, it is necessary to store a relatively large number, of the order of 1000, of excitation sequences.

To limit the complexity of the search procedure, it has been proposed that each sample of an excitation sequence can only take three values, namely, 0, 1 or -1 (ternary excitation sequence). We realized that this did not change the quality of the compression in a perceptible way.

FIG. 2 represents an example of excitation sequence E which has been proposed to further reduce the complexity of the search. This excitation sequence is called binary. It includes several non-zero samples of values 1 and -1, two non-zero samples, or pulses, being separated by a constant number of zero samples, here 3. Such an excitation sequence can be represented by a binary number (or excitation code) C whose bits are associated with the pulses and correspond to the polarity of these. By doing so, the code C supplied by the control circuit 20 corresponds directly to an excitation sequence and the table 18 is deleted. The complexity is moreover reduced by the fact that the samples to be taken into account are reduced to pulses, the number of which is, in the example of FIG. 2, four times less than the total number of samples of a sequence. In addition, the structure of filters 10 and 12 is simplified.

This technique very slightly degrades the quality of the compression, but this degradation is compensated for in a simple way by a treatment intended to remove the effects of the regularity of the spacing of the non-zero samples.

Each C code is associated with an excitation vector C having as components the values 1 and -1 corresponding to the bits 0 and 1 of the code C. The terms "vector" and "code" are confused in the following description.

In order to further reduce the number of tests to minimize the error, it has been proposed to limit the number of possible vectors or excitation codes to a subset representative of a larger set. The article entitled "A Comparison of some Algebraic Structures for CELP Coding of Speech "by JP Adoul and C. Lamblin in Proc. ICASSP, 1987, describes such a method. To create a representative subset of all N-bit C codes, we form the set of values of n bits (n <N), each of which is supplemented by Nn error correction bits.

où C_i est le vecteur d'excitation essayé ; où T est un vecteur cible formé par des échantillons de la trame en cours d'analyse du signal S ayant subi le traitement inverse des filtres 10 et 12, ces échantillons étant ceux correspondant aux valeurs 1 et -1 du vecteur C_i et où F désigne la matrice représentant la fonction de transfert des filtres 10 et 12, dans laquelle on n'a conservé que les rangées correspondant aux valeurs 1 et -1 du vecteur C_i. Les notations scal(.,.) et mod(.,.) désignent respectivement le produit scalaire et le module.To find the best excitation vector C, we generally seek to maximize a selection criterion defined by: $${\text{m = scal}}^{\text{2}}{\text{(T, C}}_{\text{i}}{\text{) / mod}}^{\text{2}}{\text{(FC}}_{\text{i}}\text{),}$$

where C _i is the excitation vector tested; where T is a target vector formed by samples of the frame being analyzed of the signal S having undergone the reverse processing of filters 10 and 12, these samples being those corresponding to the values 1 and -1 of the vector C _i and where F denotes the matrix representing the transfer function of filters 10 and 12, in which only the rows corresponding to the values 1 and -1 of the vector C _i have been kept. The notations scal (.,.) And mod (.,.) Denote the scalar product and the module respectively.

The test of all the excitation vectors C _i according to this criterion represents a considerable number of calculations to be carried out between the arrivals of two frames of the signal S.

It was found that the denominator of the criterion m is approximately constant, whatever the excitation vector C _i . Thus, the criterion m is approximately maximized by maximizing the numerator. This numerator is maximized when each component of the excitation vector C _i is that of the same sign as the corresponding sample of the target vector T. In other words, an approximate optimal excitation code is obtained directly by taking for its bits the sign bits of the target vector samples (or the inverse of the sign bits).

This solution is not applicable in the case where the usable excitation codes are limited to a subset representative of a larger set obtained, for example, using an error correcting code.

An object of the present invention is to provide a method making it possible to limit the number of calculations necessary to maximize the aforementioned criterion m in the case where the usable excitation codes belong to a subset representative of a larger set.

To achieve this object, the present invention provides a method for determining an excitation vector associated with a frame of a speech signal to be compressed, said vector belonging to a subset associated with a larger set of vectors d excitation likely to maximize a criterion, and having as components values 1 and -1 corresponding to a sequence of excitation samples of a linear prediction filter. The criterion is equal to the square of the ratio between, on the one hand, the scalar product of the excitation vector by a target vector formed by samples of the frame having undergone inverse linear prediction filtering and, on the other hand, the module of the excitation vector having undergone direct linear prediction filtering. The method comprises the steps of preselecting an excitation vector having as components those of the same sign as the corresponding samples of the target vector, or those of inverse sign, and, if the preselected excitation vector does not belong to said sub- together, to select as the excitation vector the one which maximizes said criterion from the vectors of the subset which are respectively associated with the preselected vector and with the vectors closest to it in the larger set.

According to an embodiment of the present invention, the excitation vectors are associated with excitation codes, the bits of which correspond to the signs of the components of the excitation vectors, the subset of excitation codes associated with said sub- set of vectors being formed by binary values supplemented by error correction bits, each excitation code being associated with an excitation code of the subset by an error correction function. The method comprises the steps of forming the group comprising the preselected code associated with the preselected vector and the codes closest to it, so that each of these closest codes differs from the preselected code by a single bit, to be submitted the codes of this group to the error correction function to obtain a group of corrected codes belonging to the subset, and to select as excitation code, among the corrected codes, that associated with the vector which maximizes said criterion.

According to an embodiment of the present invention, the error correction bits are bits of a Hamming correcting code.

• la figure 1, précédemment décrite, illustre une méthode de compression CELP ;
• la figure 2, précédemment décrite, représente un exemple de séquence d'excitation et du code correspondant ; et
• la figure 3 illustre des traitements à effectuer selon la présente invention pour sélectionner un vecteur d'excitation optimal dans le cas où celui-ci appartient à un sous-ensemble obtenu en utilisant un code correcteur d'erreurs.

These objects, characteristics and advantages as well as others of the present invention will be explained in detail in the following description of particular embodiments, given without limitation by means of the attached figures, among which:

• Figure 1, previously described, illustrates a CELP compression method;
• FIG. 2, previously described, shows an example of excitation sequence and the corresponding code; and
• FIG. 3 illustrates the processing operations to be carried out according to the present invention for selecting an optimal excitation vector in the case where it belongs to a subset obtained using an error correcting code.

To maximize the above-mentioned criterion m, it has been found that the denominator of this criterion, the square of the module of the vector FC _i , is approximately constant, whatever the excitation vector C _i . This approximation is relatively good since the modulus of the vector C _i is constant. So, to maximize the criterion m in an approximate way, it suffices to maximize a simplified criterion which is the scalar product of the target vector T by the excitation vector C _i . This scalar product is maximal when each component (1 or -1) of the vector C _i is that of the same sign as the corresponding sample of the target vector T. We thus obtain directly an optimal excitation vector approximated Copt from the target vector T.

This method is not directly applicable in the case where the possible excitation codes belong to a subset representative of a larger set, for example when this subset is formed from values of n bits to which one adds Nn bits of an error correcting code. Indeed, it is then very likely that the excitation code found does not belong to the subset. In this case, one could consider reducing the excitation code found to an excitation code belonging to the subset by applying to it an error correction function associated with the correction code. We then come across the excitation code of the subset which is closest to the excitation code found. The effect of this "error correction" is to modify at least one bit of the excitation code, this bit possibly having in some cases a great influence on the value of the criterion m, so that the final excitation code provides bad results.

By way of example, a Hamming correcting code denoted H (N, n, 3) is used below, where 3 denotes the minimum Hamming distance separating two elements belonging to the representative subset. The Hamming distance between two values is defined as the number of bitwise differences between these two values. With this solution, we create a subset of 2 ⁿ excitation vectors of N bits.

According to the invention, a group of excitation codes is formed comprising an initial code found by maximizing the simplified criterion m as well as all the other codes obtained from it by modifying a single bit. This has the consequence, by using a Hamming correction code allowing to correct a single bit (minimum Hamming distance 3), that each of the group excitation codes is close to a code distinct from the usable subset. Next, the Hamming error correction function is applied to each of the codes in the group, which brings each of the codes in the group to the code closest to the subset. We obtain a group of "corrected" codes belonging to the subset and which "surrounds" the code initially found. Among the corrected codes, one retains as an approximate optimal code that which maximizes the complete criterion m, that is to say by calculating its numerator and its denominator.

FIG. 3 represents in diagram form the method according to the invention which has just been described. The analyzed frame of the signal S undergoes the reverse processing at 24 of that of the filters 10 and 12 of FIG. 1. A target vector T is thus obtained, of which only the samples corresponding to the pulses of the excitation sequences are kept. In 26, samples of the vector T are retained only the sign bits (or their inverses) to provide an initial excitation code C ₀ . This code C ₀ is "corrupted" at 28 to form a group of codes comprising the code C ₀ and all the other codes C ₁ to C _N obtained by modifying a single bit of the code C ₀ . Each code C ₀ to C _N undergoes an "error correction" at 30 to provide a group of corrected codes C ' ₀ to C' _N. In 32, each of the vectors associated with the corrected codes is compared with the target vector T, and the optimal associated excitation code Copt is that associated with the vector which maximizes the complete criterion m.

Generally, to obtain better results, the position of the first pulse of the excitation sequences E is variable. In the example of FIG. 2, this position can be one of the first four, which is determined by two additional bits to be transmitted to the decoder and which multiplies by four the number of excitation vectors to be tested. In this case, we start by forming, for each of the four possible positions, a target vector and an excitation vector whose components are those of the same sign as those of the corresponding target vector. One retains as approximate optimal excitation vector, among the four vectors thus obtained, that which maximizes the complete criterion m.

Claims (3)

Method for determining an excitation vector (Copt) associated with a frame of a speech signal (S) to be compressed, said vector belonging to a subset associated with a larger set of excitation vectors capable of maximizing a criterion (m), and having as components values 1 and -1 corresponding to a sequence of excitation samples (E) of a linear prediction filter (10), said criterion being equal to the square of the ratio between , on the one hand, the scalar product of the excitation vector by a target vector (T) formed by samples of the frame having undergone inverse linear prediction filtering and, on the other hand, the modulus of the excitation vector having undergone direct linear prediction filtering, characterized in that it comprises the following steps: - Preselect an excitation vector (C ₀ ) having for components those of the same sign as the corresponding samples of the target vector, or those of opposite sign; and - if the preselected excitation vector (C ₀ ) does not belong to said subset, select as excitation vector the one which maximizes said criterion (m) from among the vectors (C ' ₀ ... C' _N ) of the subset which are associated respectively with the preselected vector and with the vectors (C ₁ ... C _N ) closest to this one in the larger set. The method of claim 1, wherein the excitation vectors are associated with excitation codes whose bits correspond to the signs of the components of the excitation vectors, the subset of excitation codes associated with said subset of vectors being formed by binary values supplemented by error correction bits, any excitation code being associated with an excitation code of the subset by an error correction function; characterized in that it comprises the following stages: - form the group comprising the preselected code associated with the preselected vector (C ₀ ) and the codes (C ₁ ... C _N ) closest to it, so that each of these closest codes differs from the preselected code by a single bit; - submit the codes of this group to the error correction function to obtain a group of corrected codes (C ' ₀ ... C' _N ) belonging to the subset; and - select as excitation code (Copt), among the corrected codes, that associated with the vector which maximizes said criterion (m). Method according to claim 2, characterized in that the error correction bits are bits of a Hamming correction code.

Images