EP0454552A2 - Method and apparatus for low bitrate speech coding - Google Patents

Abstract

The method involves dividing up the speech signal into frames of a constant length, calculating (4...10) the characteristics of N vocal duct modeling filters and the pitch, voicing and energy characteristics of the speech signal. Block coding is performed both for the filters and for the pitch and voicing. The speech signal energy is determined P times per frame for N frames and coded in a single block. Applications: low-speed 800 bits/sec vocoders.

Description

The present invention relates to a method and a device for low bit rate coding of speech.

It applies in particular to the production of vocoders for HF radio links, or of those used for voice messaging.

In these areas, the volume of information to be transmitted is increasingly coming up against the technological limits of equipment capable of carrying speech. Thus for transmissions whose bit rate is less than 2400 bits per second, the known coding techniques (MIC, DELTA, RELP, etc.) are no longer suitable, the speech signal no longer being able to be transmitted by its form of wave. To ensure these transmissions it becomes necessary to use the much more sophisticated coding techniques of vocoders. Thus, most very low speed vocoders use a vector coding technique of their digital filter to model the voice path. This modeling takes place by searching for a reference in a dictionary. However, this technique which is both very complicated and costly to implement does not make it possible to obtain a fine quantization of the speech signal. The difficulties also come from the fact that the energy of the signal is often poorly represented and therefore poorly coded, so that sudden variations in amplitude of the voice signal can no longer be restored correctly.

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

To this end, the subject of the invention is a low bit rate coding process for speech, characterized in that it consists, after having cut the speech signal into frames of constant length, to calculate the characteristics of N modeling filters of the vocal tract as well as the fundamental period (pitch), voicing and energy characteristics of the voice signal by determined intervals of N successive frames by calculating the energy of the speech signal a determined number P of times per frame to code all of these characteristics. Other characteristics and advantages of the invention will become apparent from the description given with reference to the appended drawings which represent:

Figure 1 a flowchart illustrating the speech coding method implemented by the invention.

FIG. 2 a mode of coding the LSP coefficients of the analysis filter used in FIG. 1 to model the voice path.

Figure 3 a table of LSP coefficients.

Figure 4 of frame coding paths by interpolation.

Figure 5 is a pitch coding table.

FIG. 6 is a flowchart illustrating the method for synthesizing the speech signal implemented by the invention.

FIG. 7 a graph to illustrate a mode of interpolation of the synthesis filters implemented by the invention.

Figure 8 an embodiment of a device for implementing the method according to the invention.

The coding method according to the invention consists, after having cut the speech signal into frames of constant length of approximately 20 to 25 ms, as this usually takes place in vocoders, determining and coding the characteristics of the speech signal over N successive frames by determining the energy of the signal P times per frame.

The synthesis of the speech signal on each frame then takes place by descrambling and decoding the values of the coded characteristics of the speech signal.

The representative steps of a coding method according to the invention applied to a case where N = 3 successive frames are analyzed are represented on the flow diagram of FIG. 1. On this flow diagram the method begins in steps 1 to 6, with the calculation on the first analyzed frame of the "LSP" coefficients where "LSP" is the English abbreviation for "Line Spectrum Pair", an analysis filter modeling the vocal tract: this calculation can be carried out for example by following the known method described in the article by MM. Peter KABAL and Ravi PRAKASA RAMACHANDRAN entitled "The computation of line spectral Frequencies using Chebyshev polynomials" published in IEE Transactions on Acoustics, Speech and Signal Processing ASSP-34 Dec. 86.

After sampling the speech signal on each frame and quantizing the samples over a determined number of bits, these are pre-emphasized in step 3. As the sampling operation makes the spectrum of the speech signal periodic, the number of samples taken into account for the determination of the coefficients of the vocal tract modeling filter is limited in a known manner by making the product of the pre-emphasized samples of step 3 by a HAMMING window of duration equal to that of a frame, this window also having the advantage of strengthening the resonances.

où i est un nombre entier variant de 0 à 10 par exemple, et S_i représente un échantillon de signal préaccentué et fenêtré.The coefficients k _i of the vocal tract modeling filter are calculated in step 5 from autocorrelation coefficients R _i defined by a relation of the form:

where i is an integer varying from 0 to 10 for example, and S _i represents a sample of pre-emphasized and windowed signal.

The calculation of the coefficients K _i can be carried out in step 5 by applying the known algorithm of M. LEROUX-GUEGUEN, a description of which can be found in the article of the journal IEEE Transactions or Acoustics Speech, and Signal Processing June 1977 titled "A fixed point computation of partial correlation coefficients"'. This calculation amounts to inverting a square matrix whose elements are the coefficients R _i of the relation (1).

The passage from the reflection coefficients to the prediction coefficients A _i takes place in step 8. This passage also uses an algorithm known by the name of M. Levison's algorithm, a description of which can be found in the article entitled:
"The Wiener RM5 error croterion in filter design and prediction J Math Phys, 25 pp 614-617 (1947)"

et $$\text{Q(Z⁻¹)=A(Z⁻¹)+Z⁻¹¹.A(Z)}$$

avec

Finally the LSP coefficients of the filter are calculated from two polynomials P and Q described as follows in the plane of transforms in Z, where Z is the complex variable of these polynomials, $$\text{P (Z⁻¹) = A (Z⁻¹) -Z⁻¹¹.A (Z)}$$

and $$\text{Q (Z⁻¹) = A (Z⁻¹) + Z⁻¹¹.A (Z)}$$

with

et $${\text{g}}_{\text{i}}{\text{= β}}_{\text{i}}\text{Fe/2¶}$$

If e ^jαi and e ^{j βi} denote the roots of polynomials P and Q the coefficients LSP are by definition the frequencies f _i and g _i of the arguments of these roots
is : $${\text{f}}_{\text{i}}{\text{= x}}_{\text{i}}\text{Fe / 2¶}$$

and $${\text{g}}_{\text{i}}{\text{= β}}_{\text{i}}\text{Fe / 2¶}$$

In this calculation F _e represents the sampling frequency of the speech signal.

The frequencies f _i and g _i are kept in a memory, not shown and the preceding calculations are repeated on the samples of the two frames which follow. When the parameters of three consecutive frames are calculated and three sets of coefficients have been stored, the method goes to their coding in step 13.

The calculation of the fundamental period of the signal and the voicing takes place in a known manner by performing steps 9 and 10. During these steps the speech signal is classified into two categories of sounds, voiced sounds and unvoiced sounds. Voiced sounds that are produced from the vocal cords are compared to a series of impulses whose fundamental period is called "Pitch" in English. Unvoiced sounds produced by turbulence are assimilated to white noise. Thus when the speech signal has marked periodicities, the method recognizes in step 10 for each frame a voiced sound, and a non-voiced sound otherwise. Recognition takes place after a preprocessing of the signal to reinforce useful information and limit that which is not. This preprocessing consists in carrying out a first low pass filtering of the signal, followed by bashing and a second filtering. As the fundamental frequency of the speech signal varies between 50 and 400 Hertz the first filtering is carried out for example by means of a simple "Butterworth" filter of order 3 whose cutoff frequency at 3dB can be fixed at 600 Hertz . The trimming then places the signal samples whose level is below a certain predetermined threshold at zero amplitude, possibly variable depending on the amplitude of the voice signal. This raking makes it possible to accentuate the periodic aspect of the signal while reducing the details detrimental to subsequent processing.

Finally, the second filtering makes it possible to smooth the results of the bashing by eliminating the high frequencies. To this end, a Butterworth filter identical to the first filter can be used.

• 1. A calculer une décision préliminaire de voisement à partir des valeurs de l'énergie, du filtre de modélisation et du nombre de passages par l'amplitude nulle du signal.
• 2. A calculer un seuil de voisement à partir de la décision du voisement préliminaire, de l'énergie basse fréquence et de constantes internes.
• 3. A calculer pour chaque valeur de R une fonction AMDF(k) =SOMME| (S_(n) - S_(n-k)| (8) où s(n) représente le signal prétraité, et à calculer les valeurs maximales de cette fonction.
• 4. A comparer et étudier les valeurs maximales obtenues pour en déduire le voisement et le pitch de la trame.
• 5. Et à corriger le voisement et le pitch de la trame précédente en fonction des résultats de la trame courante pour conserver une certaine stationnarité au voisement.

The pitch and voicing calculations take place in a known manner by using the AMDF (Average Magnitude Difference Function) function. They take place according to five stages which consist:

• 1. To calculate a preliminary decision of voicing from the values of the energy, the modeling filter and the number of passages by the zero amplitude of the signal.
• 2. To calculate a voicing threshold from the decision of the preliminary voicing, low frequency energy and internal constants.
• 3. To calculate for each value of R a function AMDF (k) = SUM | (S _(n) - S _(nk) | (8) where s (n) represents the preprocessed signal, and calculating the maximum values of this function.
• 4. To compare and study the maximum values obtained to deduce the voicing and the pitch of the frame.
• 5. And to correct the voicing and the pitch of the previous frame according to the results of the current frame to maintain a certain stationarity at the voicing.

The energy calculation which takes place in step 8 is executed on four subframes. This calculation takes place by taking the logarithm to base 2 of the sum of the energies of each pre-emphasized samples of a subframe.

The subframes in each frame are contiguous or overlap to have a length multiple of the "pitch".

Once the characteristics, modeling of the filter, energy, voicing and pitch are obtained for three successive frames the method proceeds to their coding according to steps 13 to 16. The coding of the filter of the three frames designated below by weft 1, weft 2 and weft 3 is done in two stages starting with weft 3.

The coding of frame 3 is of scalar type. It is carried out in application of the algorithm known under the name "Adaptive Backward Sequential" as described for example in the article of the journal IEEE on selected areas in communications, Vol. 6 feb. 88 of MM. Sugamara N and FAYARDIN N (1988) entitled "Quantizer design in LSP speech analysis".

The coding algorithm is executed in descending order of the LSP coefficients, starting with the last of the ways shown in FIGS. 2 and 3. For a voice path modeling filter with 10 LSP coefficients, for example the coding of the last LSP coefficient ( 10) takes place linearly between two frequency values F₁₀MIN and F₁₀MAX and takes place on N _V10 values coded linearly on NB₁₀ bits.

The coding of the LSP (i) other coefficients for i = 9, 8 ... 1 takes place by comparison of the coefficient LSPQ (i + 1) with a maximum frequency value F _i MAX

If LSPQ (i + 1)> F _i MAX then the coding of the coefficient is performed linearly between two values F _i MIN and F _i MAX on NV _i values and therefore on NB _i bits.

If LSP (i + 1) <F _i MAX then the coding of the coefficient is carried out linearly between F _i MIN and LSPQ (i + 1) on NV _i values and therefore on NB _i bits.

During the coding of frames 1 and 2 a good approximation of the values of LSP coefficients corresponding to frames 1 and 2 is obtained from the interpolation between frames 0 (frame 0 = frame 3 of the group of 3 previous frames) and 3 In this process, frames 1 and 2 are not coded directly, but it is the type of interpolation allowing them to be quantified as faithfully as possible which is coded.

For each of the values of LSP coefficients of odd order of frames 1 or 2, the coder determines among 3 interpolations represented by the graph of FIG. 4 which one seems to him to give the best approximation of the values of frames 1 and 2.

The three possible cases of interpolation cases 0, case1 and case 2 give for frames 1 and 2 LSPQ coefficients defined in connection with FIG. 4 as follows. (LSPQ (frame i) = Quantized value of the LSP of frame i

The method then chooses from the 3 previous interpolations the one which minimizes the quantization error, estimated by means of a function D_INTER defined below by adopting the corresponding code value.

où LSPQ(cas i, Trame j) est la valeur du coefficient LSP impair de la trame j quantifié au moyen del'interpolation du type i.The D_INTER function is defined as follows. $$\text{D\_INTER (i) = W1. (LSPQ (case i, frame 1) -LSP (Frame 1)) 2 + W2. (LSPQ (case i, Frame 2) -LSP (Frame 2)) ²}$$

where LSPQ (case i, Frame j) is the value of the odd LSP coefficient of the frame j quantified by means of type i interpolation.

LSP (frame j) = Actual value in frame j of the odd LSP coefficient to be quantified

W1 = value of the energy of frame 1

W2 = value of the energy of frame 2

We thus obtain 5 codes of 3 cases each, that is 3⁵ = 243 possible cases. The code obtained is equal to
Code LSP1 + 3.Code LSP3 + 9.Code LSP5 + 27.Code LSP7 + 81.Code LSP9

This coding takes place on 8 bits.

The pitch and voicing coding take place in step 14 on three consecutive frames.

The current voicing type is determined from six possible cases from the voices of frames 1, 2 and 3 and the voicing of frame 0 which precedes each group of frames 1, 2 and 3.

The possible types of cases considered are as follows.

A coding table represented in FIG. 5 makes it possible to associate with any pitch value a number from the table whose value subsequently designated by "N array" is the closest to the pitch.

The coding of the six types of previous possible cases then takes place as follows:

The code 0 is assigned to type 1. A code equal to the value "N. table" of the pitch of frame 3 is assigned to type 2. A code equal to 64 to which is added the value "N. table" of the pitch of frame 3 is assigned to type 3. A code equal to 128 to which is added the value "N. table" of the pitch of frame 1 is assigned to type 4. A code equal to 192 to which is added the value "N. table of the pitch of frame 1 is assigned to type 5. Coding of type 6 takes place in a very particular way by projecting the vector composed of the three values of the pitches of the three frames on the 3 vectors (Vect 1, Vect 2, Vect 3) eigen to code the three projections obtained. These three vectors Vect 1, Vect 2, Vect 3 are an approximation of the first 3 eigen vectors of the intercorrelation matrix. As the projection on the first eigen vector gives the average of the pitches it is more simple to take directly as code for the first projection the value "N.tableau" which is the closest to the average (P₁ + P₂ + P₃) / 3 of the pitches of frames 1, 2 and 3. The corresponding code is then coded on the 63 values of the coding table.

The projection on the second eigenvector (Vect 2) is equal to the scalar product of the pitches of frames 1, 2 and 3 by the second eigenvector (Vect 2) and the projection on the third eigenvector (Vect 3) is equal to the product scalar pitch of frames 1, 2 and 3 by the third eigenvector (Vect 3).

The corresponding codes can be obtained respectively on only 4 and 3 values from the coding table.

The coding of the energy which is carried out on stage 15 takes place in a known manner and described in patent application FR 2 631 146 on three consecutive frames. Four energy values corresponding to the 4 sub-fields of each of the three fields are coded. However, in order to eliminate the redundant information in these 12 values, a Main Component Analysis of the type described has the title "Data analysis elements" in the book by MM. DIDAY, LEMAIRE, POUGET and TESTU published by Dunod, is performed. Coding takes place in two stages. A first step is to make a basic change. The energy vector of dimension 12, composed of the 12 energy values of the 3 frames is projected on the first 3 main axes determined during the analysis by principal components (more than 97% of the information is contained in these 3 projections) .

The second step consists in quantifying these 3 projections, the first projection is quantized on 4 bits, the second on 3 bits and the third on 2 bits.

The energy coding thus obtained is then defined on 4 + 3 + 2 = 9 bits.

• 1) Code énergie 3 trames sur 9 bits.
• 2) Code pitch 3 trames sur 10 bits.
• 3) Code filtre trame 3 sur 27 bits.
• 4) Code filtres trames 1 et 2 sur 8 bits.
  soit au total 9 + 10+27 + 8 = 54 bits.

The framing which is carried out in step 16 consists in carrying out a grouping of all the codes to form a continuous word of 54 bits broken down as follows:

• 1) Energy code 3 frames on 9 bits.
• 2) Pitch code 3 frames on 10 bits.
• 3) Frame filter code 3 on 27 bits.
• 4) Frame filter code 1 and 2 on 8 bits.
  or in total 9 + 10 + 27 + 8 = 54 bits.

By way of example for the case of a frame duration of 22.5 ms, the method makes it possible under these conditions to obtain a bit rate per second of 54 / (3 * 0.0225) = 800 bits per second.

The synthesis, that is to say the decoding of the speech signal, takes place according to steps 17 to 28 of the flow diagram of FIG. 6, on the one hand, steps 17 to 21 for descrambling and decoding the values of the coefficients LSP of the filter (step 18), of the pitch (step 19), of the voicing and of the energy (step 20) for three consecutive frames and on the other hand, according to steps 22 to 28 which carry out the synthesis of the speech signal successively for each of the three frames on the basis of the information obtained during the execution of steps 17 to 21. Descreening and decoding follow procedures reverse to the screening and decoding procedures defined during the analysis illustrated by the flowchart of FIG. 1. The shaping of the synthesis filter consists in performing in step 23 an interpolation calculation of the LSP coefficients on four subframes and a calculation to transform the LSP coefficients into coefficients A _i . This last calculation is followed in step 24 by a gain calculation of the synthesis filter for the 4 subframes to which is added a calculation of the energy of the excitation signal of the filter. In order to avoid sudden transitions between dissimilar filters, these are done in step 23 in four steps every quarter of a frame. The four interpolated filters must then verify a relation of the form:
LSP (SS Tr _i , TrN) = (LSP (TrN-1) * (4-i) + LSP (TrN) * i) / 4
where LSP (SS Tri, Tr N) designates the value of the interpolated filter in subframe i of frame N.

The interpolation takes place according to the diagram in Figure 7.

As the 12 decoded energies correspond to the energy of the speech signal after pre-emphasis, it is necessary to obtain the energy of the excitation signal divide the energy by the gain of the filter.

The gain of the filter of each subframe is calculated using the coefficients K _i according to the relation

Filter gain

Finally the last step consists in determining the value of the standard deviation of the energy of each subframe (value used during the calculation of the excitation).

The entire coding and decoding method according to the invention can be executed by means of a microprogrammed structure formed as shown by way of example in FIG. 8 by a signal processing microprocessor 29 such as that sold by the company Texas Instrument under the designation TMS 320C25. According to this structure, the speech signal is first sampled by an analog to digital converter 30 before being applied to a data bus 31 of the microprocessor 29. An analog filter 32 coupled to an automatic gain control device 33 filters the signal speech before sampling. The programs and the data implemented for the execution of the method according to the invention are recorded in a read-only memory 34 and in a random access memory 35 connected to the microprocessor 29. An interface circuit 36 connects the microprocessor 29 via from a data line 37 to transmission devices external to the vocoder, not shown.

A speech reception device formed by a loudspeaker 38, a power amplifier 39, an analog filter 40, is connected to the microprocessor by means of a digital analog converter 41.

Claims (9)

Low bit rate coding method for speech, characterized in that after having cut the speech signal into frames of constant length, to calculate (4 ... 10) the characteristics of N modeling filters for the voice path as well that the fundamental period (pitch), voicing and energy characteristics of the voice signal by determined intervals of N successive frames by calculating the energy of the speech signal in number P determined times by frame to code all of these characteristics. Method according to claim 1, characterized in that the characteristics of the voice path modeling filters are formed by LSP coefficients. Method according to either of Claims 1 and 2, characterized in that the number N is equal to three. Method according to claim 3, characterized in that the coding of the LSP coefficients takes place scalarly on a first frame and by interpolation on the other two. Method according to claim 4, characterized in that the scalar coding of the coefficients of the third frame takes place by application of the "Backward Sequential Adaptive" algorithm. Method according to either of Claims 4 and 5, characterized in that the coding by interpolation on the two other frames takes place by searching among three possible interpolations for the one which exhibits the minimum quantization error. Method according to any one of Claims 1 to 6, characterized in that the coding of the fundamental period (pitch) and of the voicing takes place over three consecutive frames and takes place by direct addressing of a coding table by the value of the (pitch) when there is at least one unvoiced sound in a frame and by coding a pitch value obtained by vector transformation of the existing "pitch" values on the three fields when the sound is seen on the three fields, in this transformation the vector composed of the three values of the pitches of the three fields is projected on the first three eigenvectors of an intercorrelation matrix and the three values of the three projections are coded. Method according to any one of Claims 1 to 7, characterized in that the energy coding is carried out on 4 sub-frames in each frame. Device for implementing the method according to any one of claims 1 to 8, characterized in that it comprises a microprogrammed structure composed of a read-only memory 34 and a random access memory 35 connected to a microprocessor for processing the signal 29, the microprocessor 29 being connected on the one hand, to an analog digital converter 31 to convert the speech signal to digital samples and on the other hand to a digital analog converter to convert the speech samples formed by the microprocessor into signals analog to excite a device 38 for sound reproduction as well as an external data line 37 for an interface circuit 36.

Images