EP0428445B1 - Method and apparatus for coding of predictive filters in very low bitrate vocoders - Google Patents

Description

The present invention relates to a method and a device for coding predictor filters with very low bit rate coders.

Among the low-speed speech counting methods, the best known method is that of linear predictive coding LPC1 0, where LPC1 0 is the abbreviation in English for "Linear predictive coding, order 10". According to this method speech synthesis takes place by exciting by means of a periodic signal or by a noise source a filter whose function is to give the frequency spectrum of the signal a waveform close to that of the signal. original speech.

The major part of the bit rate, which is 2400 bits per second, is devoted to the transmission of the filter coefficients. For this, the bit stream is split into 22.5 millisecond frames comprising 54 bits, 41 of which are used to adapt the transfer function of the filter.

A known method of flow reduction is described in the document "variable rate speech compression by encoding subsets of the PARCOR coefficients", P. Papamichalis, Trans. on ASSP, Vol ASSP-31, n ° 3, June 1983. In this document, we group the frames by packets of successive frames and we associate a predictor filter to each frame. Only the set of coefficients minimizing a predetermined cost function is chosen for the transmission.

Another known method for reducing bit rates as described for example in US Pat. No. 4,852,179 consists in compressing the 41 bits associated with a filter into 10 to 12 bits which represent the number of a predefined filter, belonging to a dictionary of 2 ¹⁰ to 2 ¹² different filters, this filter being the one closest to the original filter. This method has however a first major drawback which is to require the construction of a dictionary of filters whose content depends closely on the set of filters used to constitute it by conventional data techniques ("clustering") and in this way this method is not perfectly suited to the actual sound recording conditions. A second drawback of this method is that, for its implementation, it requires a very large memory size to store the dictionary (2 ¹⁰ to 2 ¹² packets of coefficients). Correlatively the computation times become important because it is necessary to search in the dictionary for the filter closest to the original filter. Finally, this process does not make it possible to satisfactorily reproduce stable sounds. This is due to the fact that even for a stationary sound the LPC analysis never selects in practice twice the same original filter but chooses successively in the dictionary of close but distinct filters.

As in television where the reconstruction of a colored image depends essentially on the quality of the luminance signal and not that of the chrominance signal which can therefore be transmitted with a lower definition, it also appears sufficient in synthesis speech to reproduce only the contour of the energy of the voice signal, its coloring (voicing, spectrum form) being of less importance for its reconstruction. Therefore, in the known methods of speech synthesis the process of finding spectra based on the evolution of the minimum distance which separates the spectra of the original speech (of the speaker) and synthetic speech is not fully justified.

For example, different copies of the sound "A" spoken by different speakers, or recorded under different conditions may have a high spectral distance but will always remain "A" that can be recognized as such, and if there is ambiguity, translated by a possibility of confusion with a close sound, the listener can always correct himself thanks to the context. In fact, experience shows that by devoting no more than thirty bits to the coefficients of the predictor filter instead of 41, the quality of reproduction remains satisfactory even if a trained listener can perceive a slight difference between the synthesized sounds. with predictor coefficients defined on 30 or 41 bits. On the other hand, since the transmission takes place at a distance and the recipient does not therefore have the possibility of making this difference, it appears sufficient that the listener can correctly recognize the synthesized sound.

Also it seems important that in the stable parts of the signal (vowels) the predictor filter remains stable and is as close as possible to the original predictor filter. On the other hand in unstable parts (transition, unvoiced sound) the predictor transmitted does not need to be a faithful copy of the original predictor.

The object of the invention is to overcome the aforementioned drawbacks.

To this end, the subject of the invention is a method and a device for coding filters for predicting very low bit rate vocoders as claimed in claims 1 and 9.

• - la figure 1 un schéma de principe d'un synthétiseur de parole de l'art connu ;
• - la figure 2 une mise sous forme de tableaux des quatre codages possibles des filtres prédicteurs du vocodeur selon l'invention ;
• - la figure 3 un organigramme pour illustrer le calcul de l'erreur de prédiction des filtres prédicteurs mis en oeuvre par l'invention ;
• - la figure 4 un graphe de transformation des coefficients de réflexion des filtres prédicteurs ;
• - la figure 5 la loi de quantification des coefficients de réflexion des filtres transformés par le graphe de la figure 3 ;
• - la figure 6 un dispositif pour la mise en oeuvre du procédé selon l'invention.

Other characteristics and advantages of the invention will appear below on reading the following description made with reference to the appended drawings which represent:

• - Figure 1 a block diagram of a speech synthesizer of the known art;
• - Figure 2 a table layout of the four possible codings of predictor filters of the vocoder according to the invention;
• - Figure 3 a flowchart to illustrate the calculation of the prediction error of the predictor filters used by the invention;
• - Figure 4 a graph of transformation of the reflection coefficients of the predic filters teurs;
• - Figure 5 the law of quantification of the reflection coefficients of the filters transformed by the graph of Figure 3;
• - Figure 6 a device for implementing the method according to the invention.

où S_in est un échantillon n du signal dans la trame i et W_n désigne la fenêtre de pondération. A la deuxième étape référencée 6 le calcul des coefficients de réflexion du filtre prédicteur en treillis correspondant aux coefficients Ri(k) précédent est effectué en application d'un algorithme standard par exemple, de l'algorithme connu de LEROUX-GUEGUEN ou SCHUR. Acette étape, les coefficients R_ik sont transformés en coefficients K_ij où j est un entier positif prenant les valeurs successives de 1 à 10. A la troisième étape portant la référence 7 les coefficients k dont les valeurs sont comprises par définition entre -1 et +1 sont transformés en des coefficients modifiés qui évoluent entre "-l'infini" et "+I'infini" et qui tiennent compte du fait que la quantification des coefficients k doit être fidèle lorsqu'ils ont une valeur absolue proche de 1 et une valeur qui peut être plus grossière lorsqu'ils sont voisins de 0 par exemple. Chaque coefficient K_ij est par exemple transformé suivant une relation de la forme

dont le graphe est représenté à la figure 4 ou encore suivant les relations (L_ij=K_ij|1- | K_ij|) ; (L_ij=arc cos K_ij) ; (L_ij=arc sin K_ij) ou encore en application de la méthode de calcul des coefficients LSP décrite dans l'article de George S. Kang an Lawrence, J. Fransen du Naval Research Laboratory Washington DC 20375 1985 ayant pour titre "Application of line spec- trum pairs to low bit rate speech encoder". A la quatrième étape représentée en 8 les coefficients L_ij sont quantifiés suivant n_j bits chacun de façon non uniforme en tenant compte de la répartition des coefficients pour donner une valeur L_ij suivant une loi de répartition représentée par l'histogramme des L_ij de la figure 5. A l'étape 5 les valeurs de L_ij sont à leur tour utilisées pour calculer des coefficients K_ij suivant la relation

The speech synthesizer shown in FIG. 1 comprises, in a known manner, a predictor filter 1 coupled by its input E ₁ to a periodic signal generator 2 and to a noise generator 3 through a switch 4 and an amplifier to variable gain 5 connected in series. The switch 4 couples the input of the predictor filter 1 to the output of the periodic signal generator 2 or to the output of the noise generator 3 depending on whether or not the sound is to be reproduced. The amplitude of the sound is controlled by the amplifier 5. The filter 1 restores a speech signal on its output S as a function of prediction coefficients applied to its input E ₂ . Unlike what is shown in FIG. 1, the speech synthesizers to which the method and the coding device of the invention apply must include three predictor filters 1 adapted to each group of three successive 22.5 ms frames of the speech signal according to the stable or non-stable state of the sound to be synthesized. This organization makes it possible, for example, to reduce the bit rate from 2400 bits per second to 800 bits per second, by grouping the frames in packets of 3 x 22.5 = 67.5 milliseconds of 54 bits in which 30 to 35 bits are used to describe for example the 10 predictor coefficients of the three successive filters necessary for the implementation of the LPC10 coding method described above, and two bits among these are used to define the configuration to be given to the three filters to be generated according to the nature whether or not the speech signal to be generated is stable. In the table of FIG. 2 where the four possible configurations of the three filters are recorded, in state 00 of the two configuration bits corresponds a first configuration where the three predictor filters are identical for the three frames of the voice signal. For the second configuration, the configuration bits have the value 01 and only the first two filters of frames 1 and 2 are identical. In the third configuration, corresponding to configuration bits 10, only the last two filters of frames 2 and 3 are identical. Finally in the fourth configuration, corresponding to the configuration bits 11, the three filters of frames 1 and 3 are different. Naturally, this configuration mode is not unique and it is equally possible, while remaining within the scope of the invention, to define the number of frames in a packet by any number. However, for the sake of convenience, this number may be between 2 and 4 inclusive. In these cases, of course, the number of possible configurations can be extended to a maximum of 8 or 16. The definition of the filters is established according to steps 5 to 10 of the method represented by the flowchart of FIG. 3. According to a first step of the method bearing the reference 5 on the flowchart the autocorrelation coefficients R _{i, k} of signal are calculated according to a relation of the form:

where S _in is a sample n of the signal in frame i and W _n denotes the weighting window. In the second step referenced 6, the calculation of the reflection coefficients of the lattice predictor filter corresponding to the preceding coefficients Ri (k) is carried out in application of a standard algorithm for example, of the known algorithm of LEROUX-GUEGUEN or SCHUR. In this step, the coefficients R _ik are transformed into coefficients K _ij where j is a positive integer taking the successive values from 1 to 10. In the third step bearing the reference 7 the coefficients k whose values are by definition between -1 and +1 are transformed into modified coefficients which evolve between "-finite" and "+ infinite" and which take into account the fact that the quantification of the coefficients k must be faithful when they have an absolute value close to 1 and a value which can be coarser when they are close to 0 for example. Each coefficient K _ij is for example transformed according to a relation of the form

whose graph is represented in FIG. 4 or according to the relations (L _ij = K _ij | 1- | K _ij |); (L _ij = arc cos K _ij ); (L _ij = arc sin K _ij ) or also in application of the method of calculating the LSP coefficients described in the article by George S. Kang an Lawrence, J. Fransen of the Naval Research Laboratory Washington DC 20375 1985 entitled "Application of line spec- trum pairs to low bit rate speech encoder ". In the fourth step represented at 8 the coefficients L _ij are quantified according to n _j bits each in a non-uniform manner taking into account the distribution of the coefficients to give a value L _ij according to a distribution law represented by the histogram of L _ij of Figure 5. In step 5 the values of L _ij are in turn used to calculate coefficients K _ij according to the relation

pour p=1,2,...10. avec

These values K _ij represent the quantified values of the prediction coefficients from which the coefficients of a predictor A _{i (z)} can be deduced by recurrence relations defined as follows:

for p = 1.2, ... 10. with

ou encore

avec

Finally in the last step represented in 10 the calculation of the energy of the prediction error is carried out by applying the following relation

or

with

To complete the algorithm, it suffices to test the four different configurations described above by inserting between the first and second step of the process an additional step taking into account the possible configurations to finally retain only the configuration for which the total prediction error. obtained is minimal (summed over the three frames).

avec j variant de 0 à 10.In the first configuration the same filter is used for the three frames. A fourth unique dummy filter is then used for the progress of steps 2 to 6 which is calculated from the coefficients R _4j given by the relation

with j varying from 0 to 10.

et l'algorithme du procédé revient en fait à considérer les trois trames comme une seule trame de durée trois fois supérieure.The total prediction error is then equal to

and the algorithm of the method in fact amounts to considering the three frames as a single frame of duration three times greater.

The coefficients L1 to L10 can then be quantified with, for example 5.5,4,4,4,3,2,2,2,2 bits respectively, or 33 bits in total.

• R₅,_j = R_1j ⁺ R₂,_j où j prend successivement les valeurs de 1 à 10 pour les deux premières trames et R_3,j (j variant de 1 à 10) pour la dernière trame.

According to the second configuration, in which the same filter is used for frames 1 and 2, the algorithm is executed with values of the autocorrelation coefficients R _5j and R _3j defined as follows:

• R ₅ , _j = R _1j ⁺ R ₂ , _j where j successively takes the values from 1 to 10 for the first two frames and R _{3, j} (j varying from 1 to 10) for the last frame.

The prediction error is equal to E ₅ ² + E ₃ ² which amounts to considering that frames 1 and 2 are grouped into a single frame of double duration, frame 3 remaining unchanged. It is then possible to quantify the coefficients L ₁ to L ₁₀ on frames 1 and 2 with respectively 5,4,4,3,3,2,2,2,0,0 bits (25 bits in total, the coefficients L ₉ and L ₁₀ not being transmitted), and their variation to obtain those of the third frame using 3,2,2,1,0,0,0,0,0,0 bits respectively (8 bits in total) , or 33 bits for the three frames.

The fact of not transmitting the coefficients L ₉ and L ₁₀ is not annoying since in this case the configuration corresponds to predictors which evolve and whose coefficients are of decreasing importance as a function of their rank.

In the third configuration, where the same filters are used for frames 2 and 3, the same method as in the second configuration is used by grouping the coefficients R _ij of frames 2 and 4 such that R _6j = R _2j + R _3j . The same quantification method is used but by coding the predictor of frames 2 and 3 and the differential for frame 1.

Finally for the last configuration where all the filters are different, it must be considered that the three frames are decoupled and that the total error is equal to E ₁ 2 + E ₂ ² + E ₃ ² . In this case the coefficients L ₁ to L ₁₀ of frame 2 will be quantified with respectively 4,4,3,3,3,2,2,0,0 bits or 21 bits, as well as the differences for the first frame with 2 , 2,1,1,0,0,0,0,0,0 bits or 6 bits as well as the differences for frame 3 (6 additional bits). This last configuration corresponds to a coding of 21 + 6 + 6 = 33 bits.

The device for implementing the method which is shown in FIG. 6 includes a device 11 for calculating the 10 autocorrelation coefficients for each frame coupled to delay elements formed by three frame memories 12 ₁ to 12 ₃ for memorizing the coefficients R _ij calculated in the first step of the process. It also includes a device 13 for calculating the coefficients K _ij and L _ij according to the second step of the method. A data bus 14 conveys the values of the coefficients L _ij (i = 1 to 3, j = 1 to 10) and the values of the coefficients R _io representing the energies where i = 1 to 3. The data bus 14 connects the elements delay 12 ₁ to 12 ₃ and the calculation device 13 has four calculation chains referenced from 15 ₁ to 15 ₄ . The calculation chains 15 ₁ to 15 ₃ respectively comprise a summing device, respectively 16 ₁ to 16 ₃ which is connected to the delay elements 12 ₁ to 12 ₃ to calculate the coefficients R _4j , R _5j and R _sj according to the 4 configurations described previously. The outputs of the summing devices 16 ₁ to 16 ₃ are connected to devices 17 ₁ to 17 ₃ respectively for calculating the coefficients L _4j , K _4j ; K _{5d, 5d} L and K and L _{6 days 6 days.} The coefficients _4j L L L _{5d 6d} are respectively transmitted to January ₁₈ quantizing devices to 18 ₃ to calculate the coefficients L _{ij in} accordance with the fourth step of the method. These coefficients are applied to total error calculation devices referenced respectively from 19 ₁ to 19 ₃ to respectively provide total prediction errors E ₄ ² , E ₅ ² + E ₂ ² and finally E ₁ ² + E ₆ ² for each of the configurations 1 to 3 described above. The calculation chain 15 ₄ comprises connected to the data bus 14 a separate quantification device 18 ₄ of the coefficients L _ij . The coefficients L _ij obtained at the output of the quantization device 18 ₄ are applied to a total error calculation device 19 ₄ to calculate the total error according to the relation E ₁ ² + E ₂ ² + E ₃ ² defined above. Each of the outputs of the total error calculation devices 19 ₁ to 19 ₄ of the calculation chains 15 _i to 15 ₄ are applied to the respective inputs of a total search device of minimum 20. On the other hand, each of the outputs of the quantification device 18 ₁ to 18 ₄ , supplying the coefficients L _ij , are applied to a switching device 21 controlled by the output of the minimum total error search device 20 to select coefficients L _ij to be transmitted which correspond to l minimum total error calculated by the device 20. In this example, the device output comprises 35 bits, 33 bits representing the values of the coefficients L _ij obtained at the output of the switching device 21 and 2 bits representing one of the four configurations possible indicated by the minimum total error finding device 20.

It goes without saying that the invention is not limited to the examples which have just been described and that it can receive other variant embodiments depending in particular on the coefficients which are applied to the filters which may be different from the coefficients L _ij defined above and of the number of these coefficients which can be different from 10. It is also clear that the invention can still be applied for definitions of packets of frames comprising numbers different from three frames or filtering configurations different from four and that these variants must naturally lead to total numbers of quantization bits different from (33 + 2) bits with a distribution by different configuration.

Claims (9)

1. Method for coding predictive filters for very low bit-rate vocoders of the type in which the voice signal is partitioned into binary frames of specified duration, and in which the frames are grouped (12₁ ... 12₃) into packets of successive frames and a predictive filter (1) is associated respectively with each frame contained in a packet, characterized in that it consists moreover in defining, with regard to each packet of frames, possible configurations to be accorded to the filters of the packet depending on the voiced or unvoiced nature of the signal, in computing (17,18) for each configuration the prediction coefficients and the energy (19) of the prediction error, retaining (20) only the configuration and the prediction coefficients for which the total prediction error, summed over the frames of the packet, is a minimum, and in quantizing the coefficients of each predictive filter (5 ... 9) as a function of the configuration retained.

2. Method according to Claim 1, characterized in that the number of frames in a packet is between 2 and 4 inclusive (12₁ ... 12₃).

3. Method according to Claims 1 and 2, characterized in that the number of configurations is 4, 8 or 16 in number.

4. Method according to Claim 3, characterized in that it consists in limiting the choice of the configurations to four,

a first configuration in which the predictive filters are identical,

a second and a third configuration in which only two predictive filters are identical

and a fourth configuration in which the three predictive filters are different.

5. Method according to any one of Claims 1 to 4, characterized in that it consists, when computing the prediction coefficients, in calculating in each frame the autocorrelation coefficients R_¡,_k of the sampled voice signal, and in applying the Leroux-Gueguen or Schur algorithm to determine the reflection coefficients of each predictive filter.

6. Method according to any one of Claims 1 to 5, characterized in that the reflection coefficients L_i,j of the filters are 10 in number and are coded over a total length of 33 bits irrespective of the configuration.

7. Method according to Claim 6, characterized in that the reflection coefficients L₁ to L₁₀ of the fit- ters have respective lengths:

(5, 5, 4, 4, 4, 3, 2, 2, 2, 2) bits in the first configuration

(5, 4, 4, 3, 3, 2, 2, 2, 0, 0) bits and (3, 2, 2, 1, 0, 0, 0, 0, 0, 0) bits in the second and third configurations

(4, 4, 3, 3, 3, 2, 2, 0, 0) bits for coding the intermediate frame (frame 2) in the fourth configuration and (2, 2, 1, 1, 0, 0, 0, 0, 0, 0) bits for the other two frames (frame 1) (frame 3) in the fourth configuration.

8. Method according to Claim 7, characterized in that the reflection coefficients of the filters are determined through the relation

9. Device for implementing the method according to any one of Claims 1 to 8, characterized in that it includes a device (11) for computing the autocorrelation coefficients for each frame, coupled to frame memories (12₁, 12₂, 12₃) for storing the autocorrelation coefficients, a device (13) for computing the reflection coefficients of the predictive filters, coupled via its inputs to the frame memory (12₁ - 12₃) and via its outputs to facilities (15₁ ... 15₃) for computing prediction errors, so as to select by means of a minimum total error search device (20), the reflection coefficients of the filters to be transmitted which correspond to the minimum total error.

Images