EP0341129B1 - Method and apparatus for coding speech signal energy in very low bitrate vocoders - Google Patents

Description

The present invention relates to a method and a device for coding the energy of the voice signal in vocoders at very low bit rates.

It applies in particular to the production of vocoding devices with linear prediction of the type of those described in the "THOMSON-CSF Technical Reviews" vol. 14 No. 3 September 1982 pages 715 to 731 and vol. 15 n ° 2 pages 495 to 516 published by MASSON, 120 boulevard Saint Germain 75280 Paris.

In these devices, the voice signal is divided into time slots or windows of fixed lengths of approximately 20 milliseconds in the transmission vocoders, and each signal window is analyzed to extract the parameters necessary for controlling the digital filters of the vocoders of reception. These parameters are made up of control coefficients for the reception filters, the effective mean value VME of the voice signal and an indication of the nature of the voice signal, whether or not it is heard.

To digitally code the speech signal so as to ensure very low bit rate links, typically less than 1000 bits / s, the method of coding the VME parameter consists in quantifying the VME parameter on 32 values (0 to 31) according to a scale logarithmic standardized by the NATO standard "Stanag 4198" relating to linear predictive coding of order 10 of which a description appears in the article of M TREMAIN having for title "The Government Standard Linear Predictive Coding Algorithm-LPC 10" and which is published in Speech Technology April 1982 pages 40-49.

The quantized VME signal is then coded on 11 bits during three consecutive windows. The effective mean value of the middle window is coded on 5 bits and that of each of the extreme windows is coded by a coding method. of the middle window. A description of this coding process can be found in an article published by the authors Wong D, Juang BH Gray AH in the journal IEEE Transactions on ASSP vol. 30, 1982 pages 770-780 entitled "An 800 bits / s Vector Quantization LPC Vocoder".

However, the 11-bit coding of the VME parameter limits the possibilities of reducing vocoder bit rate, in particular at low bit rates lower than 800 bits / second.

Other coding methods are also known from the German patent application DE A 2 608 244 and from the article by J. MAKHOUL et al IEEE vol. 73 n ° 11 November 1985 pages 1551-1588 New York having for title "Vector quantization in speech coding" these allowing by change of base to code the energy vector of the voice signal 'with a reduced number of bits

However, these methods also do not make it possible to obtain appreciable bit rate reductions below 800 bits / s by the fact that the calculations take place during the duration of 20 to 30 ms of each speech frame during which the signal is considered. as stationary without taking into account that the samples can in fact be correlated on several successive frames

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

To this end, the subject of the invention is a method of coding the energy of voice signals in vocoders at very low bit rates of the type consisting in analyzing the voice signals in consecutive windows, in measuring in each of the windows the energy in effective mean value (VME) of the samples of the speech signal and in quantifying said energy VME on a determined number m of levels, characterized in that it consists in constructing, in a vector space with n dimensions having for first base the vectors unitary a resulting energy vector whose components in the base are the VME energies measured respectively in n analysis windows of the vocal signal, to operate in this space a change of base having as first main axis an oriented axis corresponding to the sum of the unit vectors of the first base, to project into the new base obtained the resulting energy vector, and to code on q bits such that 2 ^q = m the component of the resulting vector projected on the main axis of the new base and on a reduced number of bits less than q, the components of the energy vector projected on the n-1 other main axes of l vector space defined in the new database.

The invention also relates to a device for implementing the above method.

• La figure 1 le principe de codage mis en oeuvre par l'invention dans un espace à deux dimensions.
• La figure 2 le principe de codage mis en oeuvre par l'invention dans un espace à trois dimensions.
• La figure 3 un tableau récapitulatif des énergies portées par les axes principaux de l'espace à trois dimensions défini dans la nouvelle base.
• La figure 4 un dispositif mis en oeuvre par l'invention pour mesurer l'énergie des échantillons du signal à l'intérieur de chaque fenêtre de signal.
• La figure 5 un dispositif de codage du paramètre VME selon l'invention.
• La figure 6 un dispositif de décodage du paramètre VME selon l'invention.
• Les figures 7 et 8 des deuxième et troisième variantes de réalisation de dispositifs de codage du paramètre VME selon l'invention.

Other characteristics and advantages of the invention will appear below with the aid of the description given with reference to the appended drawings which represent.

• Figure 1 the coding principle implemented by the invention in a two-dimensional space.
• Figure 2 the coding principle implemented by the invention in a three-dimensional space.
• Figure 3 is a summary table of the energies carried by the main axes of the three-dimensional space defined in the new database.
• FIG. 4 a device implemented by the invention for measuring the energy of the signal samples inside each signal window.
• FIG. 5 a device for coding the VME parameter according to the invention.
• FIG. 6 a device for decoding the VME parameter according to the invention.
• Figures 7 and 8 of the second and third alternative embodiments of VME parameter encoding devices according to the invention.

The method according to the invention is based on the observation that the energy contained in the speech signal varies very slowly over time, so that the energies E₀, E₁ and E₂ of the samples quantified in each signal window can be considered as strongly correlated to each other. It is indeed possible to observe, by considering only a very large number of groups of two successive windows, and by referencing the corresponding energy vector of each group in a vector space orthonormal to two dimensions where the energies E₁ and E₂ of each window represent the projections of the energy vector E of each group in the representative base of this space, the origins of the energy vectors E of all the groups being confused with that of the two-dimensional vector space, that the ends of the energy vectors E are distribute, as shown in Figure 1, in a domain (D) of the plane formed by the two vectors E₁ and E₂, substantially symmetrical with respect to a bisector E'₁ of the angle (E₁, E₂) formed by the two vectors, being very elongated in the direction of the bisector and on the other hand flattened in the direction normal to the latter.

The same observation can be made by analyzing the energy of the speech signal by groups of n successive windows. For example, on a very large number of observations we can notice, by referencing as in Figure 2 the energies E₀ E₁ and E₂ in a 3-dimensional space, that the ends of the vectors each resulting from the sum of three vectors E₀, E₁ and E₂ are all contained in a domain or "cloud" comprising three main axes of inertia.

On the axes E₀, E₁ and E₂ of the trihedron shown in Figure 2, the unit vector of the first axis of inertia has as components (3 ^-½ , 3 ^-½ , 3 ^-½ ), the unit vector of the second axis of inertia has components (-2 ^-½ , 0, 2 ^-½ ) and the unit vector of the third axis of inertia has components (-6 ^-½ , 2x6 ^-½ , -6 ^-½ ). As indicated in the table in FIG. 3, it appears that by projecting the energy of each vector E of components E₀, E₁, E₂ on each axis of inertia, that the percentages of the energies of the projected vectors are 90% for the first axis, 8% for the second and only 2% for the third. It appears that a saving in coding bits can be achieved by not coding the components E₀, E₁ and E₂ in the entire maximum coding space formed by the cube whose side lengths represent the maximum energies E _0max , E _1max and E _2max that the speech signal can take in three consecutive windows, but by coding their resultant E in the new orthonormal base formed by the 3 unit vectors forming the main axes of inertia, which only occupies a small volume of the cube defined above.

avec a=3^-½ , b=-2^-½ et c=-6^-½
et en notant par (E'₀, E'₁ et E'₂) les composantes dans la nouvelle base du vecteur résultant de l'addition des trois vecteurs E₀, E₁ et E₂, le vecteur de composantes (E'₀, E'₁ et E'₂) vérifie la relation matricielle suivante : $${\text{[E'] = P⁻¹x[E] =}}^{\text{t}}\text{P x [E].}$$

By noting by P, the matrix of the components of the 3 unit vectors such that:

with a = 3 ^-½ , b = -2 ^-½ and c = -6 ^-½
and by noting by (E'₀, E'₁ and E'₂) the components in the new base of the vector resulting from the addition of the three vectors E₀, E₁ and E₂, the vector of components (E'₀, E ' ₁ and E'₂) check the following matrix relation: $${\text{[E '] = P⁻¹x [E] =}}^{\text{t}}\text{P x [E].}$$

In this relation the matrix [E '] has for column vectors the components E'₀, E'₁ and E'₂, the matrix [E] has for column vectors the components E₀, E₁ and E₂, and ^t P denotes the matrix transposed from P.

By way of example, the preceding transformations make it possible, by limiting the values of E'₀ between 0 and 54, to code it on only 4 bits according to a linear scale comprised between these two values and by truncating the values E'₁ and E'₂ between the values -16 and +16, these can be coded respectively on 3 bits and 2 bits according to also a linear scale also included between these 2 values. The result is then the obtaining of 3 coded values (E''₀, E''₁ and E''₂) on a total of only 9 bits instead of 11 in the prior art which is sufficient to ensure good quality 800 bit / s transmissions.

On reception, the operations carried out are the reverse operations of coding. From the coded values E''₀, E''₁ and E''₂ the method determines in a first step the vector of component E'₀, E'₁ and E'₂ expressed in the base of the unit vectors of the axes principal of inertia. Then according to a second step, it multiplies the matrix P by the vector of components E'₀, E'₁, E'₂ to obtain a vector of components E₀, E₁ and E₂. Finally, according to a third step, it applies to the components E₀, E₁ and E₂ the decoding law of the linear predictive coding standard of order 10, to obtain the three effective values VME₀, VME₁ and VME₂ of the three consecutive signal windows processed.

A corresponding coding device is shown in FIGS. 4 and 5. In FIG. 4, the device for measuring the energy of the samples of the voice signal comprises an accumulator circuit 1, represented inside a closed line in dotted lines, this circuit being coupled to two registers 2 and 3 connected in series. The accumulator circuit 1 consists, in a known manner, of an accumulator register 4 and of an adder circuit 5. Each sample S _i of the voice signal is applied to a first operand input of the adder circuit 5 and is added to the content of the accumulator register 4 which is applied to the second operand input of the adder circuit 5. The accumulation of the samples S _i of a window thus takes place in the accumulator register 4 throughout the duration of the window. At the end of each window the content of the accumulator 4 is transferred to register 2 and then is loaded to the next window in register 3. In steady state, the contents of registers 3, 2 and 4 permanently indicate at the end window, the respective energies E₀ E₁ and E₂ contained in three consecutive windows for exploring the voice signal. These energy values E₀, E₁, E₂ are applied to the coding device of FIG. 5, to the corresponding inputs of an adder circuit 6. The coding device also comprises three processing channels 7, 8 and 9 shown in interior of closed dotted lines. Channel 7 comprises, an attenuator circuit 10 of attenuation ratio 3 ^-½ , a limiter stage 11 and an encoder 12. All of the elements 10, 11, 12 are coupled together, in this order, and in series at the output of the adder circuit 6. Channel 8 comprises an amplifier circuit 13 of gain 3, coupled to an attenuator circuit 15 of attenuation ratio 6 ^-½ through a circuit subtractor 14. The subtractor circuit 14 includes a first operand input, marked "+" which is connected to the output of the amplifier circuit 13 and a second operand input marked "-" which is connected to the output of the adder circuit 6 .

Channel 9 comprises an attenuator circuit 16 of attenuation ratio 2 ^-½ coupled to the output of an adder circuit 17. A switch circuit 18 applies one or the other of the signals obtained at the output of channels 8 and 9 to the input of an encoder 19 through a limiter stage 20.

The reception decoder is shown in FIG. 6. It comprises a set of three reception channels 21, 22 and 23 shown inside closed dotted lines.

The first channel 21 comprises, connected in series, an attenuation circuit 24 of attenuation ratio 3 ^-½ and two subtractor circuits 25 and 26.

The second channel 22 comprises connected in series, an attenuation circuit 27 of attenuation ratio 2 ^-½ , an adder circuit 28 and a subtractor circuit 29;
the third channel 23 comprises, connected in series with an attenuation circuit 30 of attenuation ratio 6 ^-½ , an amplifier 31 of gain 2 and an adder circuit 32.

The subtractor circuit 25 is connected by a first operand input marked "+" to the output of the attenuator circuit 24 and by a second operand input marked "-" to the output of the attenuator circuit 27. The result of the subtraction carried out by the subtractor circuit 25 is applied to a first operand input marked "+" of the subtractor circuit 26. The second operand input marked "-" of the subtractor circuit 26 is connected to the output of the attenuator circuit 30. The output of the subtractor circuit 26 supplies the energy E₀ of the first window of the voice signal. The adder circuit 28 has a first operand input connected to the output of the attenuator circuit 27 and a second operand input connected to the output of the attenuator circuit 24. The result obtained at the output of the adder circuit 28 is applied to a first operand input marked "+" of the subtractor circuit 29. The second operand input marked "-" of the subtractor circuit 29 is connected to the output of the attenuator circuit 30. The energy E₂ of the signal is obtained at the output of the subtractor circuit 29. Finally the adder circuit 32 is connected by a first operand input to the output of the amplifier 31 and by a second operand input to the output of the attenuator circuit 24 The energy E₁ of the signal is obtained at the output of the adder circuit 23.

Rather than performing a scalar coding of the vector E'₀, E'₁, E'₂ in the base of the three unit vectors of the main axes of inertia, a second variant of implementation of the method according to the invention may consist in perform, as shown in FIG. 7, a vector coding of the vector (E'₀, E'₁, E'₂), by searching for the closest vector of the vector (E'₀, E'₁, E'₂) among 2 ^N vectors whose ends would coincide with the nodes of a bounded subset of a face-centered cubic network, so as to obtain an encoding on N bits. This coding mode is carried out by the circuits of FIG. 7, which include a programmable read-only memory 33 addressed by an address counter 34, three subtracting circuits 35 to 37, three squaring circuits 38 to 40, a circuit adder 41, a comparator circuit 42, and two registers 43 and 44. The read-only memory 33 contains the three components of the 2 ^N estimated vectors (E₀, E₁ and E₂) and these are addressed by the N-bit address counter 24. Each of the components read from the memory 33 is applied respectively to a first operand input of the subtractor circuits 35 to 37. The components E₀, E₁ and E₂ of the energy of the voice signal from each of the three windows are applied respectively to the second operand inputs of the subtractor circuits 35 to 37. The results of the subtractions carried out by the subtractor circuits 35 to 37 are applied respectively to the input of the circuits squared from 38 to 40, and the results of the squared elevations are applied to the inputs of the summing circuit 41. The sums of the squares of the differences between, each component (E₀, E₁, E₂) of a vector representing the energies of the voice signal in three consecutive windows and the components E₀, E₁, E₂ of an estimated vector addressed by the address counter 34, are applied successively by the output of the summing circuit 41 to a first comparison input of a comparator circuit 42 to be compared with the content of register 43 which is applied to the second comparison input of comparator 42. On each comparison the content of register 43 is updated by the result of the summation obtained at the output of the summing circuit 41 if this result is less than the content existing in the register 43. In this way after each progression of the address counter 34, the register 43 keeps in memory the sum of the squares obtained from the summing circuit 41 which is the smallest among all the sums already made since the beginning of the addressing of the estimated vectors in the memory 33. In parallel with each update of the content of the register 43 the content of the register 44 is replaced by the address of the corresponding vector which has been read in the memory 33. The number of the VME vector coded on N bits is thus obtained directly in the register 44.

A third variant of implementation of the method according to the invention is shown in FIG. 8. As this third variant follows from the embodiment of the second variant described above, the homologous elements of FIG. 7 are represented in FIG. 8 with the same references. This third variant differs from the previous one in that the memory space 33 is divided into three memory sub-spaces 33 _a , 33 _b and 33 _c . In this case, the first N / 3 bits of the address counter 34 address the subspaces 33 _a , the following N / 3 bits address the second subspace and the remaining N / 3 bits address the subspace 33 _c . This allows to have a vector representation of the vectors in a three-dimensional space having the shape of a cubic network with centered face, by assigning to each vector a group and a sub-group of this space. By assigning the groups of the cubic network to memory space 33 _a, the subgroups to memory space 33 _b , 2 ^{N / 3} estimated energy vectors can be coded in memory space 33 _a and 2 ^{N / 3} estimated energy vectors can be coded in memory space 33 _b . The 2 ^{N / 3} remaining vectors are coded in memory space 33 _c . For N = 9, 8 groups are thus obtained, each comprising 8 subgroups of 8 vectors each.

Similarly to the device represented in FIG. 7, the energy of the voice signal with component E de, E₁ and E₂ is measured by circuits 35 to 43 relative to the energy of the corresponding estimated vectors formed successively through a multiplexer 45 by memories 33 _a , 33 _b and 33 _c .

The group, the subgroup and then the vector of the subgroup which has the energy closest to the component vector E₀, E₁ and E₂ are thus defined successively. The group numbers in the group and the vector within a sub-group are recorded respectively in the register 44 which has in FIG. 8 the form of a bank of registers composed of the registers 44 _a , 44 _b and 44 _c . AND gates 48, 49 and 50 allow the transfer of the group, sub-group and vector addresses within a group, each time the result of the comparison carried out by the comparator 42 indicates that the sum formed by the summator 41 is less than the contents of register 43.

Claims (8)

1. A process for coding the energy of voice signals in very low-rate vocoders of the type which analyse voice signals in consecutive windows, measure (1, 2, 3) in each of these windows the energy as a root mean square (RMS) of the samples of voice signal and quantify said RMS energy over a given number m of levels, characterised in that it consists of constructing in an n-dimensional vector space having as a first base the individual vectors (e₁ to e_n), a resultant energy vector of which the components in the base (e₁ to e_n) are the RMS energies measured respectively in n consecutive analysis windows of the voice signal, of making in this space a change of base having as its first principal axis an oriented axis corresponding to the sum of the individual vectors of the first base, of projecting (6, 7, 8, 9) in the new base obtained the resultant energy vector, and of coding on q bits such that 2^q = m the component of the resultant vector projected on the principal axis of the new base and, on a smaller number of bits less than q, the components of the energy vector projected on the n-1 other principal axes of the vector space defined in the new base.
2. A process according to Claim 1, characterised in that the vector space is a three-dimensional space, and in that the individual vectors carried by principal axes of the new base have as a component in the base defined by the vectors of the three-dimensional space representing the energies measured over 3 consecutive windows, respectively ( 3^-½, 3^-½, 3^-½) for the first principal axis, (-2^-½, 0, 2^-½) for the second principal axis, and (-6^-½, 2x6^-½, -6^-½) for the third principal axis.
3. A process according to Claim 2, characterised in that the coding of the resultant energy vector projected over the first, second and third principal axes have respective lengths of 4, 3 and 2 bits.
4. A device for implementing the process according to any one of Claims 1 to 3 comprising means for analysing voice signals in consecutive windows, means (1, 2, 3) for measuring in each of the windows the energy as a root mean square (RMS) of the samples of the voice signal and for quantifying said energy over a given number m of levels,
   the device being characterised in that it comprises
   matrix calculating means(6, 7, 8, 9) for projecting, in a vector base having as its first principal axis an oriented axis corresponding to the sum of the individual vectors (e₁ to e_n) of a first base, an energy vector constructed in the base (e₁ to e_n), the components of which in this base are respectively the RMS energies measured in n consecutive measurement windows of the voice signal, and coding means (12, 19) for coding on q bits such that 2^q = m the component of the energy vector projected on the principal axis of the new base and for coding on a smaller number of bits less than q the components of the energy vector projected on the n-1 other principal axes of the vector space defined in the new base.
5. A device according to Claim 4, characterised in that where n = 3 the matrix calculating means (6,7,8,9; 33 ... 44) carry out the matrix product $$\text{[E'] = P⁻¹[E]}$$

   

   where [E] is the column vector formed by the components E_o, E₁, E₂ of the energy measured over three successive windows

   

   where a = 3^-½, b = -2^-½ and c = -6^-½
6. A device according to Claim 4, characterised in that it comprises means (35 ... 44) for searching among a set of 2^N vectors, prerecorded in a memory (33), the extremities of which coincide with the nodes of a limited sub-assembly of a face-centred cube lattice, for the vector which has those components in the new base which are closest to the components E₀, E₁ and E₂ of the energies measured over three successive windows of the voice signal, so as to represent the code of the resulting vector over N bits.
7. A device according to Claim 6, characterised in that the memory (33) is organised into N/3 groups, N/3 sub-groups each of N/3 vectors.
8. A device according to any one of Claims 6 and 7, characterised in that it comprises

   - an address counter for addressing the prerecorded vectors to the memory (33)

   - subtracter circuits (35, 36, 37) for comparing the N root mean squares of the voice signals supplied by N windows with the N respective components of the vectors read from the memory (33)

   - and a decision circuit (42, 43, 44) for identifying the vector read from the memory (33) which has the read components which are closest to the N measured root mean squares of the voice signal.

Images