EP0347307B1 - Coding method and linear prediction speech coder - Google Patents

Abstract

The method, usable in particular for low-rate speech transmission, uses vector excitation. A signal frame is represented, on the one hand, by prediction parameters, and on the other hand, by a succession of excitation vectors contained in a dictionary (20) and by amplification gains (Gk) of these vectors, the vectors adopted being determined by searching for the minimum energy of an error signal obtained by subtraction of each vector in its turn, after having submitted to a filtering, from the frame of the speech signal. Before subtraction: each frame of the speech signal is subjected to a short-term analysis filtering and to a weighted synthesis filtering, with coefficients possibly fixed in time, and the amplified vector is subjected to a long-term predictive filtering and to the same perceptual weighted synthesis filtering as the speech signal. <IMAGE>

Description

The subject of the present invention is a coding method and a speech coder of the type known as linear prediction analysis. It relates more particularly to methods and speech coders of this type with excitation by excitation vector, often designated by the English abbreviation CELP, which are to be distinguished from coding methods with analysis by linear prediction with multi-pulse excitation. (MPLPC), an example of which is given in document EP-A-0 195 487 to which reference may be made.

Vector-driven linear prediction analysis coding provides an interesting solution to the problem of speech transmission in a narrow band channel, for example, transmission between mobiles and to mobiles in a 12.5 kHz channel which reduces the bit rate available at around 8 kbits / s; in the latter case, the bit rate assigned to the transmission of the parameters representing the speech signal is reduced to about 6 kbits / s since part of the overall bit rate must be assigned to the transmission of an error correction code.

Speech coders with linear prediction and vector excitation are already known, usable with a low bit rate, usually between a quarter of a bit and a half bit per speech sample. We can in particular find an embodiment of such an encoder in the article by SCHROEDER and ATAL "Code excited linear prediction (CELP): high quality speech at very low bit rates", proc. ICASSP, March 1985.

FIG. 1 gives a schematic diagram of such an encoder 10. The speech signal is applied to this encoder via a digitization chain. In the embodiment shown in FIG. 1, the chain comprises, from a microphone 12, a low-pass filter 14 limiting the bandwidth to approximately 4000 Hz and a sampler-encoder 16. The sampler takes samples speech at a rate which is for example 8 kHz and provides successive samples, grouped by vocoder frames occupying time windows of fixed duration, for example 20 ms.

The coder 10 transforms the speech signal into a coded signal having a lower bit rate, transmitted to the transmission equipment by a multiplexer 18 which receives, for each frame, the indices k of the optimal excitation vectors _Ck , the associated gains G _k and coefficients identifying prediction parameters, for each of the constituent blocks of the frame, each occupying a sub-window.

The coder 10 shown by way of example in FIG. 1 uses analysis by synthesis: the speech spectrum in each window is modeled by a linear predictor filter whose coefficients are variable over time. The residual signal, obtained by subtraction, is subject to vector quantization using a dictionary of waveforms. In FIG. 1, the dictionary 20 contains K + 1 excitation vectors c _k (with k = 0, ..., k, ..., K) and drives an amplifier 22 of gain G _k .

Usually the excitation vectors stored in the dictionary 20 are chosen either empirically taking account of statistical data on the language, or randomly, or else from conventional binary digital codes such as the Golay codes.

The article by SCHROEDER and others mentioned above proposes for example a dictionary comprising 1024 excitation vectors each made up of 40 samples. This number of vectors is placed between the minimum below which the excitation would be poorly represented and the maximum beyond which the number of bits left free would be insufficient to transmit the parameters of the predictors.

The output of amplifier 22 is applied to a predictive synthesis filter consisting of a long-term predictor filter 24, intended to introduce the periodicity of the long-term signal, and of a short-term predictor filter 26. The output Sn of the predictor filter, which represents a synthesis of estimation of the speech signal, is applied to the subtractive input of a subtractor 28 which receives, on its additive input, the sampled and digitized speech signal Sn.

The respective transfer functions 1 / B (z) and 1 / A (z) of the filters 24 and 26 once calculated and quantified, the coding operation consists in determining the optimal sequence of innovation c _k and the gain G _k for each speech frame by a synthetic analysis process. For each of the coding sequences _Ck , the synthesis signal obtained S _k is compared to the original signal S and the difference signal obtained in the subtractor 28 is processed in a perceptual weighting filter 30 having a transfer function W (z ), whose function is to attenuate the frequencies for which the errors are less important from the perceptual point of view and on the contrary to amplify the frequencies for which the errors are more important from the perceptual point of view.

A circuit 32 searches for the coding sequence for which the energy contained in the weighted error signal e _k for a sub-window is minimal; this sequence is selected for the current block, then the optimum gain G _k is calculated.

Conventionally, the function A (z) of the short-term predictive filter 26 is of the form:

In this formula, which uses the classic notation in z, the coefficients a (i) constitute the parameters of linear prediction. Their number is generally between 8 and 16 for windows of 20 ms.

As for the transfer function B (z), it can be of the form 1-bz- ^T and involve a delay T ranging from 40 to 120 samples.

The perceptual weighting filter 30 has a transfer function W (z) which is generally of the form:

Despite its interest, the coding method which has just been described can practically not be implemented in real time, because of the enormous volume of computations required to search for the sequence of innovation, (i.e. - say the optimal excitation vector) by K + successive loop paths, each path constituting the filtering of an excitation vector, by filters with coefficients variable over time.

There is also known a CELP coding method in accordance with the preamble of claim 1 (IEEE Journal on selected areas in communications, Vol. 6, n ° 2, February 1988, pages 353-363); the present invention aims to provide a coding method with linear prediction and excitation by coding vectors of this type, which meets the requirements of practice better than those previously known, in particular in that it reduces by at least an order of magnitude the volume of calculation to be carried out for the coding of a segment.

For this, the invention notably proposes a speech coding method, with linear prediction and vector excitation, according to the characterizing part of claim 1.

Since each coding sequence consists of several equidistant pulses separated by zeros, advantageously binary, that is to say that an excitation by regular pulse sequences, or RPCELP is used, we reduce in very large proportions the duration of the search for the optimal sequence, especially if an appropriate choice is made of the characteristics of the perceptual weighting filter.

Other features of the invention are defined in claims 2 to 6.

• - la figure 1, déjà mentionnée, est un schéma de principe d'un codeur de parole à prédiction linéaire et excitation vectorielle déjà connu ;
• - la figure 2, similaire à la figure 1, est une variante du schéma montrant une constitution possible de codeur de la figure 1, susceptible d'être simplifiée pour constituer un premier mode de réalisation de l'invention ;
• - les figures 3, 4 et 5 sont des schémas montrant des évolutions successives du codeur de la figure 2 ;
• - la figure 6, similaire à la figure 5, montre de façon plus complète un mode de réalisation de l'invention réduisant encore le volume de calcul ;
• - la figure 7 montre une répartition possible de séquences de codage dans le dictionnaire ;
• - la figure 8 montre une autre constitution possible du dictionnaire.

The invention will be better understood on reading the following description of particular embodiments, given by way of nonlimiting examples. The description refers to the accompanying drawings, in which:

• - Figure 1, already mentioned, is a block diagram of a speech coder with linear prediction and vector excitation already known;
• - Figure 2, similar to Figure 1, is a variant of the diagram showing a possible constitution of the encoder of Figure 1, which can be simplified to constitute a first embodiment of the invention;
• - Figures 3, 4 and 5 are diagrams showing successive changes to the encoder of Figure 2;
• - Figure 6, similar to Figure 5, shows more fully an embodiment of the invention further reducing the computational volume;
• - Figure 7 shows a possible distribution of coding sequences in the dictionary;
• - Figure 8 shows another possible constitution of the dictionary.

In the speech coder represented diagrammatically in FIG. 2 (where the elements corresponding to those of FIG. 1 are designated by the same reference number) the perceptual weighting filter 30, placed at the output of the subtractor 28 in FIG. 1, is transferred to the two input branches of the subtractor in the form of filters 34 and 36, of transfer function 1 / A (z / y). There is thus in cascade, on the branch assigned to the original signal S (n), the filter 33 of transfer function A (z) and the filter 36 having the same transfer function as the filter 34.

où C(z/-y) est la fonction de transfert d'un prédicteur de parole à court terme, par exemple de la forme :

The filtering of all the vectors by the synthesis filter, of transfer function 1 / A (z / y) whose coefficients vary over time, represents an enormous volume of calculations. This volume is reduced very considerably according to a first aspect of the invention, by adopting a perceptual weighting filter with small number of fixed coefficients in time, chosen according to the average characteristics of the speech over a long time interval. The perceptual weighting filter then has a transfer function W (z) which can be written:

where C (z / -y) is the transfer function of a short-term speech predictor, for example of the form:

The transfer functions of components 34 and 36 in Figure 2 then become 1 / C (z / γ).

Another embodiment of the invention, which can be combined with the first, appears better by considering the successive transformations made to the circuit of FIG. 2 to achieve it.

First, as shown in FIGS. 3 and 4, the contribution of the memory in the long-term predictor filter 24, of transfer function 1 / B (z), and in the short-term predictor weighted function filter transfer 1 / A (z / y), is subtracted from the original signal having undergone the weighting to obtain a signal x _n , before the start of the search in the vector dictionary 20. This operation is carried out in the figure 3 using a subtractor 38 which receives only the memory component of the long-term predictor filter 24.

So, during the optimal vector search procedure, each vector C _k is only processed by the weighted synthesis filter 34.

We will now show how it is possible to further reduce the computational volume considerably by referring to FIG. 4. In this figure, each of the filters 34 and 36 has been shown broken down into a filter 34a or 36a of transfer function 1 / Ã (z / γ), without memory, and a filter 34b or 36b corresponding only to the contribution of the memory terms.

• r le signal résiduel après soustraction des effets du prédicteur à long terme 24,
• x le signal original dont la redondance à long terme a été écartée dans le soustracteur 38 et qui a été pondéré par W(z),
• z_k le signal synthétisé,
• x° et z° les contributions des mémoires des filtres au calcul de x et z.

During the search for the optimal vector C _k , each vector C _k amplified with the gain G _k is no longer processed except by the weighted synthesis filter without memory 1 / Ã (z / γ) which outputs a signal z (n). If we identify by a tilde the quantities without memory and if we designate by:

• r the residual signal after subtracting the effects of the long-term predictor 24,
• x the original signal whose long-term redundancy has been eliminated in subtractor 38 and which has been weighted by W (z),
• z _k the synthesized signal,
• x ° and z ° the contributions of the memories of the filters to the calculation of x and z.

• x = Hr + x°
  
• Z_k = G_k.H C_k + Z^o

We can write :

• x = Hr + x °
  
• Z _k = G _k .HC _k + Z ^o

où H est une matrice triangulaire inférieure LxL (L étant la longueur commune des séquences) dont les éléments sont tirés de la réponse impulsionnelle h(i) de 1/A(z/y), de la forme :

qui se confond avec celle de 1/Â(z/y)The filtering operation by filter 34a, without memory, is expressed above by the convolution of two finite sequences, represented by the product of a matrix and a vector:

where H is a lower triangular matrix LxL (L being the common length of the sequences) whose elements are taken from the impulse response h (i) of 1 / A (z / y), of the form:

which merges with that of 1 / Â (z / y)

The vector x 'at the input of the subtractor 28, after subtracting the memory effects, can itself be written:

The energy of the error E _k weighted for the vector of index k (with 0 ≦ k K) can be written:

• - recherche de l'indice k pour lequel le produit scalaire Pω(k) est maximum :
  
• - calcul du gain correspondant G_k :
  

The process of finding the optimal innovation sequence (vector index k C _k and gain of amplification G _k ) comprises two stages which follow from equation (6) if one takes into account the known fact (JP ADOUL et al. "Fast CELP coding based on algebraic codes", Proc. ICASSP, April 1987) that to minimize the energy E _k is to maximize a scalar product Pw:

• - search for the index k for which the scalar product Pω (k) is maximum:
  
• - calculation of the corresponding gain G _k :
  

The computation of a scalar product is obviously faster than the search for a Euclidean distance, so that the diagram of figure 3 already allows alone to reduce the volume of computation.

The next step in the process consists in eliminating the memory terms, that is to say the operations shown diagrammatically in 34a and 36a, to arrive at the constitution shown in FIG. 5.

As in the case of FIG. 2, an important simplification consists in substituting fixed synthesis filters with transfer function 1 / C (z / -y) for filters 34a and 36a with function 1 / Ã (z / γ), this which again amounts to adopting a perceptual weighting filter of the form W '(z) = A (z) / C (z / -y). There is no longer any need to perform a repetitive filtering operation by 34a, insofar as the excitation vectors are stored in the dictionary 20, on the one hand in the prefiltered state in order to apply them directly to the maximization circuit of scalar product 38, on the other hand in original form for application to the gain amplifier 22 Gk. The simplification appears immediately by a comparison with the classical processes of search for minimum.

Yet another embodiment of the invention implements a modified error evaluation criterion to be minimized. The sample frames each occupying a window are successively applied; consequently, the impulse response of the weighted synthesis filter for a frame (or a block) occurs on the next frame (or the next block). To avoid this effect, we use the damping of the filters and we apply to their input, instead of a sequence consisting only of L samples, a sequence consisting of L samples and J zeros, J being chosen so that the impulse response of the synthesis filter W (z) / A (z) is practically zero after J samples. A value J = 10 is generally sufficient for the damping of the filters to exclude the terms representative of their memory. The impulse response matrix then becomes a rectangular "strip" type matrix with (L + J) xL terms of the type:

The matrix H ^t H = R is then a symmetric Toeplitz matrix, constructed from the autocorrelation R (i) of the impulse response h (n). H ^t denotes the transpose of H.

The memory error that appears in the equation representative of x 'is then small enough to be considered as zero and equation (7) can be written:

The vector y '= r ^t H' H can be calculated precisely, only once per frame, by a filtering operation, using an adaptive filter whose coefficients are the auto-correlation terms R (i).

To implement this method in the case of a speech signal sampled at 8 kHz and the samples of which are distributed in frames of 160 samples each occupying 20 ms, it is in particular possible to split each frame, after filtering at 33 (FIG. 5) , in four blocks of L = 40 samples which are successively applied to the filter 36a, followed each time by J = 10 zeros.

A will then be calculated for each frame while k and G _k will be calculated for each block.

A particularly interesting solution in this case consists in using sequences of pulses of length L having a regular structure made up of q equidistant pulses separated by D-1 zeros, the first pulse occupying one of the positions 0 to D-1 and the number of sequences being such that all of these positions are successively occupied. It is thus possible to give a satisfactory representation of the phase information in the excitation signal. FIG. 7 shows, by way of example, four sequences (for k = 0, 1, 2 and 3) identical except that they correspond to D = 4 different phases. We can consider that the dictionary consists of a basic set of K / D sequences, with a zero phase and with three successive shifts, ie in all K sequences.

Excitation by regular excitation sequences reduces the number of operations to be performed, since many of the products to be performed are zero, one of the factors being a zero whose position is known for each sample. We can further simplify the calculations by constituting the sequences only of binary samples which can only take the values + 1, -1 (and 0), as indicated in FIG. 8. Indeed, all the sequences then contain the same energy; the search for the optimum sequence is carried out with purely scalar products and amounts to looking for the binary vector which gives the best result. It may be noted that the document EP-A-0 195 487 relates to an MPLPC coding method according to which it is necessary successively to determine an optimal phase of pulses, then to seek the optimal amplitude of all the pulses constituting sequence among discrete values , quantified for example on 3 bits.

où d_m désigne l'une des séquences (au nombre de K/D) résultant de la décimation des composantes des K vecteurs par élimination des zéros ; la séquence d_m pour 0 < k 3 3 est donnée en figure 7 à titre d'exemple.In the case of the modified criterion and of an excitation by regular sequences, and in particular in the case of sequences made up of binary samples, and with the additional condition that the autocorrelation is normalized and presents null terms whose spacing corresponds to the non-zero samples of the sequences, the terms H c _k all become equal and we have:

where d _m denotes one of the sequences (the number of K / D) resulting from the decimation of the components of the K vectors by elimination of zeros; the sequence d _m for 0 <k 3 3 is given in FIG. 7 by way of example.

If the sequences are normalized, the search procedure is limited to the search for the sequence for which the dot product P (k) = y ^t. c _k is maximum.

• - soit en adoptant un filtre fixe R tel que R(iD) soit nul pour i > 0,
• - soit en adoptant un filtre à coefficients variables, mais dont la réponse impulsionnelle finie (RIF) est tronquée pour les indices d'échantillons supérieurs à D.

The conditions necessary for the applicability of formula (11) can in particular be obtained:

• - either by adopting a fixed filter R such that R (iD) is zero for i> 0,
• - or by adopting a variable coefficient filter, but whose finite impulse response (RIF) is truncated for the sample indices greater than D.

The coder then presents the constitution of principle shown in FIG. 6. A single filtering operation is performed on the speech signal frame by the filter 33. The sequence c _k tested, in a form which no longer needs to be prefiltered , is applied to the circuit 32 for calculating the scalar product c _k ^t .y and for determining the maximum, for which an index selection order is sent at 40. The sequence c _k amplified at 22 is applied to the long-term predictor 24, shown with a single coefficient b. The term r is formed by subtracting the output of the long-term predictor 24 from the output of the filter 34 on the speech channel, in the subtractor 38. The filter 42 which receives the residue has a fixed response R (z) represented by a symmetric Toeplitz matrix.

• - détermination de la phase qui donne une valeur maximale à M(p) :
  
• - puis, parmi les vecteurs ayant la phase retenue, sélection du vecteur d_m tel que
  • y^t.C_k = M(^p)
  • c'est-à-dire :
  • d_m(i) = signe de y(p + iD)
  • pour i = 0, ..., q-1

The search for the optimal vector can then be carried out by a reduced number of multiplication and addition operations, provided that the response is truncated if the filter is variable, for example by the following approach if the excitation vectors regular are binary:

• - determination of the phase which gives a maximum value to M (p):
  
• - then, among the vectors having the selected phase, selection of the vector d _m such that
  • y ^t .C _k = M ( ^p )
  • that is to say :
  • d _m (i) = sign of y (p + iD)
  • for i = 0, ..., q-1

Once the optimal vector has been selected, the gain G _k to be retained is deduced directly from it, since ∥HC _k ∥2 is equal to a constant value q, whatever the value of k, in the case of binary vectors which all have the same standard.

This process reduces the number of calculations required in a report which is typically about three orders of magnitude compared to the conventional CELP method, regardless of the length L chosen for the speech blocks.

• - le coefficient unique b et la période T (correspondant à la périodicité du signal de parole) du filtre prédicteur à long terme 24, une ou plusieurs fois par fenêtre, les coefficients a du filtre 33, de fonction de transfert A(z), une fois par fenêtre,
• - l'indice k du vecteur optimal et le gain correspondant G_k, une fois par bloc correspondant à une sous-fenêtre de par exemple 40 échantillons.

The data to be transmitted by the multiplexer 18 will be:

• the single coefficient b and the period T (corresponding to the periodicity of the speech signal) of the long-term predictor filter 24, one or more times per window, the coefficients a of the filter 33, of transfer function A (z), once per window,
• - the index k of the optimal vector and the corresponding gain G _k , once per block corresponding to a sub-window of for example 40 samples.

The gain G _k is, for transmission, quantized in a quantizer 46. Each signal frame is split into several blocks, an intermediate memory 48 must be interposed between the components 33 and 44.

It should also be noted that, since the excitation is binary and regular, it is not very sensitive to transmission errors: an error modifying the value of a bit only modifies the vector locally. The reduced phase bits can be protected by a correction code.

Claims (6)

1. A word coding method, with linear prediction and vectorial excitation, allowing the coding of word signals put in the form of digitized samples distributed in patterns, wherein:

a signal pattern is represented on the one hand by prediction parameters and on the other hand by a succession of excitation vectors contained in a dictionary (20) and by amplification gains (G_k) of said vectors, the vectors adopted being determined by searching (32) for the minimum energy of an error signal obtained by subtracting each vector in its turn, which has previously be subjected to filtering, from the pattern of the word signal; and,

before subtraction, subjecting each word signal pattern to a short-term analytical filtering (33) and to a perceptual weighted synthesis filtering (36) and subjecting the amplified vector to a long-term predictive filtering (24) and to the same perceptual weighted synthesis filtering (34,36) as the word signal,

characterized in that all the excitation vectors are formed by the same number of equi-distant pulses separated by zeros.

2. A method according to Claim 1, characterized in that the pulses separated by zeros are binary.

3. A method according to Claims 1 or 2, characterized in that each search (32) for the minimum energy of the error signal is performed by subjecting to filtering (34,36 or 34a,36a) an assembly comprising, in addition to the real samples of a block forming a fraction of the pattern, zero samples in an adequate number for the predictive filtering pulse response corresponding to the last real sample to be substantially damped, the filtering (34,36 or 34a,36a) being performed without storing from one block to the other.

4. A method according to Claims 1 or 2, characterized in that after being subjected to a short-term analytical predictive filtering A(z), each word signal pattern is applied to the additive input of a subtracter (38) which receives at its subtractive input the contribution from the store of the long-term predictive filter (24),

the output of the subtracter is subjected to a filtering (42),

the scalar product (32) of the filtered output of the subtracter and of each unamplified sequence is calculated in its turn, while searching for that sequence for which the scalar product is the maximum.

5. A method according to Claim 4, characterized in that the filtering (42) is with fixed coefficients.

6. A word coding method, with linear prediction and vectorial excitation, allowing the coding of word signals put in the form of digitised samples distributed in patterns, wherein:

each block forming a fraction of signal pattern is represented by one of the vectors contained in a dictionary (20), by a vector amplification gain (G_k) and by prediction parameters, the vector adopted being determined by searching the minimum energy of an error signal obtained by subtracting each vector in its turn, previously subjected to a filtering, from the word signal pattern; and,

before subtraction, subjecting each word signal pattern to a short-term analytical filtering A(z), characterized in that

the result of the subtraction (38) is subjected to a perceptual weighted synthesis filtering (36) with coefficients fixed in time and

the excitation vector, stored in a precalculated form and filtered, is subjected to a filtering by a perceptual weighted synthesis filter (34) 1/C(z/y) which is fixed and without store, all the excitation vectors being formed by the same number of equi-distant pulses, separated by zeros.

Images