JPH05507796A - Method and apparatus for low-throughput encoding of speech - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention] Method and apparatus for low-throughput audio encoding The present invention provides a method and apparatus for low-throughput audio encoding. Concerning methods and equipment for conversion.

The present invention particularly relates to vocoders for I(F radio links or vocoders used for voice communications). applied to the production of da.

In these fields, the amount of information to be transmitted is limited by the We are gradually reaching the technical limits of our equipment. For example, the throughput is 2400 per second For communications below bits, known encoding techniques (MIC, DELTA, NLP, etc.) are no longer appropriate, and audio signals are transmitted by their waveforms. is not possible. In order to perform this communication, higher performance encoding technology of the vocoder is required. It becomes necessary to use. Therefore, most vocoders with very low throughput , vocal tract (I!condaNma OCI+1). The router's vector encoding technique is used. This modeling is done in the dictionary This is done by searching for references. However, it is very complicated to implement. This technique, which is expensive and expensive, makes it possible to obtain accurate quantization of audio signals. do not. Because the signal energy is improperly represented and therefore improperly encoded, Further difficulties arise, with the result that sudden fluctuations in the amplitude of the audio signal cannot be properly recovered. It is no longer possible to

The aim of the invention is to alleviate the above-mentioned drawbacks.

To this end, the invention aims at a method for low-throughput coding of speech. This person The method cuts the audio signal into frames of a fixed length and then models the vocal tract. Characteristics of N filters, basic period (pitch) of audio signal and voiced sound (mao) i+emenj) and energy for each frame. Then, by calculating the energy of the audio signal a specific number P times, N consecutive Characterization is performed by calculating these features for each specific interval of consecutive frames. I get kicked. Other features and advantages of the invention will appear from the description given with reference to the following accompanying drawings. It will become clear.

FIG. 1 is a flowchart illustrating the speech encoding method used by the present invention. .

Figure 2 shows the LSP coefficients of the analysis filter performed in Figure 1 for modeling the vocal tract. FIG. 2 is an explanatory diagram showing an encoding method.

FIG. 3 is a table of LSP coefficients.

FIG. 4 is a graph of frame encoding by interpolation.

FIG. 5 is a "pitch" encoding table.

FIG. 6 is a flowchart illustrating the audio signal synthesis method used by the present invention. Ru.

FIG. 7 is a graph illustrating the synthetic fill interpolation method used by the present invention.

FIG. 8 is an explanatory diagram showing an embodiment of an apparatus for carrying out the method according to the invention.

The encoding method according to the invention is based on the method of encoding sound into frames of fixed length of approximately 20 to 25 m5. After the voice signal has been segmented, it is processed P times per frame, as occurs in the normal vocal tract. The sound is generated every N consecutive frames by mutually determining the signal energy. The purpose is to determine and encode the characteristics of the voice signal.

The synthesis of the audio signal in each frame consists of a frame of encoded feature values of the audio signal. This is performed after frame removal (d'e+rt11) and decoding.

Code according to the invention applied when N=3 consecutive frames are analyzed Representative steps of the method are shown in the flowchart of FIG. This frame In the lowchart, the method describes the coefficient “LA We begin with steps 1-6 by calculating PJ, where rLSPJ is the vocal tract The line spectrum pair (LjneSpectrum) of the analysis filter that models Pxi+) is the English abbreviation of J. For example, this calculation is based on IEEE Tr! as ! clions on AcocNics, 5peech xnd Signa 1P+oces+iag ASSP-34Dec, 86 Calculation of line spectral frequencies using Ebyshev polynomials (Thecompatx tion of lin+ +pecl+sl F+eqaeneies us ing ChebHhewpolynoffiinxl+) P entitled It is explained in the paper of eter KABAL and Rswi PRAKASA. This can be done according to known methods.

Sampling the audio signal in each frame and sampling with a certain number of bits After the sample quantization, these samples undergo pre-emphasis in step 3. (Pre-emphx order ud).

Vocal tract modeling because the sampling operation makes the spectrum of the audio signal periodic. The number of samples taken into account to determine the coefficients of the filter is limited in a known way. It is a Hamming window (H) of duration equal to the duration of one frame. The sample pre-emphasized in step 3 by This window also has the advantage of reinforcing resonance. have

The coefficient k of the vocal tract modeling filter is determined in step 5 by the relationship of the following form: It is calculated based on the autocorrelation coefficient R defined as

In the above formula, 1 is an integer that changes to, for example, [1-to, and S is the pre-emphasis value. represents a signal sample that has been cissed and windowed.

For the coefficients, the calculation of LEROtl and G[IEGUEN is applied. This can be done in step 5 by using This algorithm The explanation of the program can be found in the magazine rlEEE Trsatscrioas on Aco ic+, 5peech znd Si(owl ProcessiB June “Fixed point calculation of partial correlation coefficient (A Nxed point calculation)” in 1977 J. t compIltztionof pi+cor+elxtion can be found in the paper entitled Coefficiels) J. This calculation will invert a square matrix whose elements are the coefficients R, of relation (1).

The transition from the reflection coefficient to the prediction coefficient A takes place in step 8. This transition is An algorithm known as the ev-eye o-eye n algorithm is also used, and this Lξma 1s The explanation of the on algorithm is ``Vie++er in filter design and prediction. RM5 error criterion n in filter je+ign ind, prediclion) J( J Mx+h Pt+rg, 2599614-617 f! 947)) and title found in the published paper.

Finally, the LSF coefficients of the filter are defined as It is calculated from the terms P and Q. Here, 2 in the following equation is the complex variable of these polynomials. , P (Z) = A (Z) - Z, A (Z) (2) and Q(Z)=A(Z)+Z, A(Z)'(3)1;1 Engineering αI 1β1 When e and e represent the roots of polynomials P and Q, the LSP coefficients are, by definition, The frequencies of the independent variables of these roots and g・=βF/2π (6) 1 1 e It is.

In this calculation, F represents the sampling frequency of the audio signal.

Frequencies 1 and g are stored in memory (not shown) and the above calculations It starts again for two consecutive frames of samples. 3 consecutive frames Once the parameters of the frame have been calculated and the three sets of coefficients have been stored, the method Proceed to its encoding in step 13.

To calculate the fundamental period of the signal and the fundamental period of voicing, perform steps 9 and 10. This is done in a known manner by executing:

During these steps, the audio signal is divided into two sound categories: voiced and unvoiced. are categorized. The period of the fundamental wave of the voiced sound generated by the vocal cords is In Japanese, it is treated as a pulse train called "pitch." caused by turbulence Unvoiced sounds are treated the same as white noise. Therefore, the audio signal is noticeable. When the periodicity is It recognizes voiced sounds in the opposite case and recognizes unvoiced sounds in the opposite case. This recognition is based on valid information. This is done after pre-processing of the signal in order to enhance the information and limit the non-useful information. child The preprocessing of the first low frequency wave of the signal is followed by ebssz gt) and the second low range is to perform waves. The fundamental frequency of the audio signal is 50-40 Since it varies between 0 Hertz, the first wave is e.g. A simple third-order “Butterworth” ” is done by the filter. Then, the Eversage is set according to the amplitude of the audio signal. Zero-wave samples whose level is lower than a certain predetermined threshold that can be changed at will. Make width. This Evasage reduces the harmful part to subsequent processing and at the same time makes it possible to reinforce the side of the cycle.

Finally, the second wave smoothes the results of the Eversage by removing the high frequencies. make it possible to make it clear. For this purpose, use the same butterwork as the filter in Rin1. filters are available.

The pitch and voicing calculations are performed using the tllDF function (average difference function (Ayer)). Layer BeMBnifade 0H (ereace F*nctionl) on the side. This is done in a known manner by These calculations are performed in five steps: is executed.

1. Based on the energy value, the modeling filter, and the rout of the signal passing through zero amplitude. and calculate preliminary decisions for voicing.

2. Based on the preliminary determination of voicing, low frequency energy, and internal constants. Calculating the vocalization threshold.

3. For each value of H, the function in which S(.) represents the preprocessed signal is: AMDF(kl=sUMis　-9 (2) (n-k) l (8) and calculate the maximum value of this function.

4. Compare the obtained maximum values to infer the voicing and pitch of the frame and consider it.

5. The result of the current frame in order to maintain a certain stationarity in its voiced Correct the voicing and pitch of the preceding frame accordingly.

The energy calculations performed in step 8 are performed over four subframes. . This calculation is performed for each pre-emphasized sample of one subframe. This is done by taking the base 2 logarithm of the energy sum.

The subframes within each frame have a length that is a multiple of its "pitch". are joined or superimposed on each other.

Characteristics of filter modeling and the relationship between the energy, voicing, and pitch of the audio signal Immediately after the features are obtained over three consecutive frames, the method They are encoded according to steps 13-16. Below, frame 1 and frame The filter encoding of the three frames represented by frame 2 and frame 3 is This is done in two stages starting from frame 3.

The encoding of frame 3 is of the scslsrl type.

For example, IEEE ion+nxl on 5elected stx+ incoamonictio11+, Vol, 6. Feb, 8g inside r LSP audio by N SuHmtrt and KFAYARDIN (19H) Quantization device proposal in analysis (Qasnt gr design++ in LSP 5peech xcx17si+) As explained in the paper entitled J. “Backward sequential adaptation (B!ckvxrd Sgqoe+N1tlAd*pjrtiy e) By applying the algorithm known as the J algorithm. will be carried out.

This encoding algorithm starts from the end and L They are executed in order of increasing SP coefficient. For example, a vocal tract model with 10 LSP coefficients In the case of a delt filter, the encoding of the last coefficient LSP (10) is It is performed linearly between the wave number F10MIN and '10i11AI, and llB1. B is performed over the MYY2O values linearly encoded over the cut.

The encoding of other coefficients LSP (i) (i=9.8., , , l) is as follows: This is done by comparing the coefficient LSPQ(i+1) with the maximum frequency value F, MAW. We Ru.

In this case, if LSPQ(i±11>F, MAX, then the coefficient of this coefficient Ossification is performed in one dish over Nv values and thus over 118 bits. It is performed linearly between two values F, MIN and F, MAY.

LSP(i+t) <FllAX If there is one, then it will be over the value of [over 8 bits, F, MIN and I This is performed linearly with LSPQ(i+1).

In the process of encoding frame 1 and frame 2, corresponding to frame 1 and frame 2 A good approximation of the values of the LSP coefficients is given by frame 0 (frame 0 = 3 previous frames). is obtained by interpolation between frame 3) and frame 3 of the group of frames. this In processing, frame 1 and frame 2 are not encoded directly, but are encoded. is an interpolation type that allows them to be quantized as faithfully as possible. Ru.

For each value of the odd LSF coefficient of frame l or frame 2, the encoder is the value of frame 1 and frame 1 among the three interpolations shown by the graph in Figure 4. Determine the interpolation that appears to give the best approximation of the value of system 2.

The three possible interpolation cases, namely case O, case 1 and case 2, are used for frame 1 and frame 1. For frame 2, we give the coefficient LSPQ defined in relation to Fig. 4 as follows: . (1, SPQ (frame 1) = (quantized value of LSP of frame l).

case 0 LSPQ (case 0. frame 1) = (2"LSPQ C frame 0) + LS I’Q (frame 3)’) /3LSPQ C case O, frame 2) = (LSPQ C frame O)+2’LSPQ C frame 3))/3 Case I LSPQ (case 1. frame 1) = (LSPQ C frame O) + 2”L SPQ (frame 3))/3 LSPQ (case 1. frame 2) = LSPQ (frame 3) case 2 LSPQ (case 2. frame 1) = LSPQ (frame 0) LSPQ (case 2. Frame 2) = (2$LSPQ (Frame 0) + LSPQ (Frame M3))/3 Subsequently, the method adopts the corresponding sign values from the three interpolations mentioned above. Evaluated by the function D l1fTER defined as below, Choose the interpolation that minimizes quantization error.

Rem 1)) 2+W2. (LSPQ (Case 1. Frame 2) - LSP (Frame 2))2 In the previous equation, LSPQ (case 1. frame 1) is is the value of the odd-numbered LSP coefficient of the converted frame j, and LSP (frame l ) = real value in frame j of the odd LSP coefficient to be quantized, Wt=energy value of frame 1, W2 = energy value of frame 2 It is.

Therefore, 5 codes are obtained in each of the 3 cases, i.e. 35·243 cases are obtained. The resulting code is F+, code LSPI + 3, code L SP3 + 9. Code LSP5 + 27. Code LSP7 + 81 code LSP9 be equivalent to.

This encoding occupies more than 8 bits.

Pitch and voicing encoding is performed in step 14 over three consecutive frames. It will be held in

The current types of voicedization are voiced in frame 1.2.3 and voiced in frame 1.2. ．． 6 available fields from the voiced voiced frame 0 preceding each group of 3 determined from among the

The possible types of cases that can be considered are:

Frame 1 Frame 2 Frame 3 Type 1 No voice No voice No voice Type 2 No voice No voice Yes voice Type 3 No voice Yes voice Yes voice Type 4 Yes voice No voice No voice Type 5 Yes Voice Yes Voice No voice Type 6 Yes Voice Yes Voice Yes Voice The encoding table shown in Figure 5, whose values are referred to below as the "N-chart", corresponds to the pitch. Associating all values of pitch with the nearest numerical value from that encoding table. enable.

In this case, the above six possible case types of encoding C are done as follows: Ru.

Code 0 is assigned to type 1. "N chart" value of pitch of frame 3 A code equal to is assigned to type 2. Frame 3 pitch “N cha” A code equal to 64 is assigned to type 3, to which the . A code equal to 12g where the "N chart" value of the pitch of frame 1 is appended to it. code is assigned to type 4. The "N chart" value of the pitch of frame 1 is A code equal to 192 appended thereto is assigned to type 5. type 6 is done in a very specific way. It signifies the three resulting projections. To encode, three eigenvectors (vector 1, vector 2, vector 3) By projecting a vector made by the three values of the pitch of the three frames onto the evening. What are these three vectors, namely vector 1, vector 2 and vector 3? , are approximations of the three first eigenvectors of the cross-correlation matrix. first eigenvector Since the projection onto the pitch gives the average pitch, the average pitch of frame 1.2.3 The "N chart" value closest to the average (Pl+P2+P3)/3 is used for the first projection. It is even easier to adopt it directly as code. In that case, we will respond The code is encoded by the 63 values of the encoding table.

The projection onto the second eigenvector (vector 2) is the projection onto the second eigenvector (vector 2). equal to the scalar product of the pitch of frame 1.2.3 according to rule 2), and the eigenvector of east 3 The projection onto the vector (vector 3) is the projection by the third eigenvector (vector 3). is equal to the scalar product of Rhyme 1.2.3.

whose corresponding codes are respectively by only four values and only three values of the encoding table It is possible to obtain.

The energy encoding performed in step 15 is performed for three consecutive frames. This is done in a known manner as described in patent application PR2631146. three frames Four energy values corresponding to the four subframes of each frame are encoded. death However, in order to remove redundant information in these 42 values, DaΩad IIIAY SLEMAIRE SPO[1GET1T published by In the book by ε5TII, “Data Analysis Element (Elemc++ts It is explained under the title d’tnz17sedeldonnee+)J Principal component analysis of type (one Ansly+e pst CompossIl+ es Pr1occiptle+) is performed.

Encoding is done in two stages. The first step consists in performing a base transformation. Ru. 12 energy values of magnitude consisting of 12 energy values of 3 frames The vectors are projected onto the three first principal axes determined during the principal component analysis. over 97% of the information in is contained in these three projections).

The second step consists in quantizing these three projections.

The first projection is quantized over 4 bits and the second projection is quantized over 3 bits. The third projection is quantized over two bits.

Therefore, the energy encoding obtained in this way is constant over 4+3+2=9 bits. be justified.

The frame specification (++tmBe) performed in step 16 is decomposed as shown below. Rearrange all of the following code to form one contiguous word of 54 bits It is about achieving.

1) Energy code of 3 frames over 9 bits, 2) Over 10 bits 3) the fill of frame 3 over 27 bits. 4) Filter code for frame 1 and frame 2 over 8 bits.

That is, the total is 9 + 10 + 27 + 8 = 54 bits.

For example, in the case of a frame duration of 22.5mt, this method works under these conditions. So, 54/(311,0225) = 800 bits per second Allows you to get value throughput.

The synthesis, that is, the decoding of the audio signal is performed in steps 17 to 2 of the flowchart of FIG. 8. That is, on the one hand, frame removal (step 17); Decoding the LSP coefficient values of the filter over consecutive frames (step 1 8), decoding of pitch value (step 19), decoding of voiced sound and energy value according to steps 17 to 21 with respect to (step 20); For each of the three frames based on the information obtained when executing steps 17 to 21, Steps 22 to 28 are followed for synthesizing continuous audio signals. frame removal and The decoding is performed with the frame specification defined during analysis as shown in the 70-chart of Figure 1. Proceed according to the reverse procedure of the decoding procedure. Shaping the synthesis filter is done in steps. 23, interpolation calculation of LSP coefficients over four subframes and its Calculator for converting LSP coefficients to coefficients A! It consists in doing calculations. Following the latter calculation, in step 24 the filter excitation A composite frame for the four subframes is added to which the calculation of the energy of the signal is added. A calculation of the filter gain is performed.

In order to prevent sudden transitions between different filters, in step 23 These transitions occur every quarter frame in all four steps. So If , the four interpolated filters must satisfy the relationship of the form do not have.

LSP (33τr, , TrN) = (SLP (Trn-1) (4-i) + L SP(TtN) i)/4 In the previous formula, LSP (33T "i, TtN)" is represents the interpolated filter value in subframe i of frame N.

The interpolation is performed according to the schematic diagram of FIG.

The 12 decoded energies are comparable to the energy of the pre-emphasized audio signal. Therefore, in order to obtain the energy of the excitation signal, the gain of the filter is It is necessary to separate the energy of

The gain of the filter for each subframe is calculated using the coefficient K according to the following relationship: be done.

Finally, the final step is to determine the value of the standard deviation type of energy for each subframe. (value used when calculating excitation).

Both the encoding and decoding methods according to the invention are T! x x+ Signal processing microprocessor commercially available from In+lramels A signal processing microprocessor 29 such as This can be done by a created microprogram structure. Following this structure, the audio before the signal is applied to the data bus 31 of the microprocessor 29. 30 is initially sampled. an amplifier coupled to automatic gain control device 33; Analog filter 32 filters the audio signal prior to its sampling. To the present invention The programs and data used for the execution of the method are processed by a microprocessor. In the read-only memory 35 and random access memory 34 connected to 29 will be written to. An interface circuit 36 connects it via a data transmission line 37. A microprocessor 29 is connected to a transmitting device (not shown) external to the vocoder. Connecting.

Sound formed by loudspeaker 38, output amplifier 39 and analog filter 40 The voice receiving device is connected to the microprocessor 29 via the digital AC converter 41. It is connected.

, nh・=-funih”keshireitobutu summary A method of low-throughput encoding of audio, in which audio signals are divided into frames of fixed length. After dividing the voice signal, the characteristics of the N filters that model the vocal tract as well as the voice signal are determined. In calculating all the characteristics of the fundamental period (pitch), voicing, and energy of the signal. be.

Group The melody coding is done on the one hand with respect to said filter and on the other hand with respect to the pitch and This is done regarding vocalization.

The energy of the audio signal is N (P for each frame for 11 frames). is determined over a number of times and subsequently encoded in the form of a single block.

Application to a low throughput vocoder of 800 bits per second.

Copy and Translation of Amended Sound) Submission (Article 184-8 of the Patent Law) October 27, 1992 seagull

Claims (9)

[Claims] 1. A low-throughput encoding method for audio that encodes audio into frames of fixed length. After dividing the signal, the characteristics of N filters that model the vocal tract as well as the voice In order to encode all the characteristics of the fundamental period (pitch), voicing, and energy of the signal, In order to calculate the energy of the audio signal a specific number P times for each frame, By The method, characterized in that: 2. The characteristics of the filter modeling the vocal tract are formed by LSP coefficients. A method according to claim 1, characterized in that: 3. 3. Method according to claim 1 or 2, characterized in that the number N is equal to three. 4. The encoding of the LSP coefficients is performed in the first frame and in the other two frames. 4. The process according to claim 3, characterized in that the process is performed scalarly by interpolation in the system. Method. 5. The scalar encoding of the coefficients of the third frame is called "Backward Sequential Adaptation". dsequentialAdaptive)” algorithm. 5. A method according to claim 4, characterized in that: 6. Encoding by interpolation in the other two frames results in three possible interpolations: This is done by finding the interpolation that exhibits the smallest quantization error. The method according to claim 4 or 5, characterized in that: 7. The fundamental period (pitch) and voicing encoding are performed in three consecutive frames. and the encoding includes at least one unvoiced voice in one frame. When there is a sound, it is possible to do so by direct addressing of the encoding table by its pitch value. and when the sound is voiced over three frames This is done by vector transformation of the “pitch” values that exist over the vector In voice conversion, when the sound is voiced over three frames, three A vector consisting of three values of frame pitch is one cross-correlation matrix are projected onto the three first eigenvectors of , and the three values of these three projections are sign 7. A method according to any one of claims 1 to 6, characterized in that the method is encoded. 8. The energy encoding is performed over four subframes within each frame. 8. A method according to any one of claims 1 to 7, characterized in that: 9. Apparatus for carrying out the method according to any one of claims 1 to 8, comprising: , the device is a read-only memory connected to the signal processing microprocessor 29. Microprogram structure composed of memory 34 and random access memory 35 on the one hand, the microprocessor 29 digitally samples the audio signal; on the other hand, the audio restoration device 38 is connected to the AD converter 31 for converting the sound into Apply the audio sample formed by the microprocessor to excite Connected to a DA converter for conversion into an analog signal, and further connected to an interface circuit 3 A device characterized in that it is connected to an external data transmission line 37 for 6.