JPH03189700A - Vector quantizing system - Google Patents

Abstract

PURPOSE:To reduce the calculation quantity of all filter operations related to a search, to curtail the number of arithmetic circuit, to miniaturize the circuit and to attain the reduction in cost by executing recursively the filter operation in the case of a prescribed search by utilizing a toeplitz matrix. CONSTITUTION:The system is constituted so that a specific filter operation is shown by the product of a matrix of a code vector and a square triangular matrix are constituted of the toeplitz matrix, respectively, only an initial element in each element of the product of the matrix of the code vector and the square triangular matrix is derived by a regular matrix operation, and other each element is derived recursively, respectively, based on the element derived before by one. By this constitution, a code book is searched, and in the case of searching a pitch period, as well, it is executed similarly in its fundamental constitution. In such a manner, the calculation quantity required for executing the search of the code book and the search of the pitch period can be decreased, respectively. Accordingly, the number of arithmetic circuits can be reduced remarkably.

Description

[Detailed Description of the Invention] [Object of the Invention] (Industrial Application Field) The present invention relates to vector quantization used to compress and transmit information in signal sequences such as audio signals and video signals. .

(Prior Art) Vector quantization is one of the technologies that is currently attracting the most attention as a method for efficiently encoding audio signals and image signals. Especially in the field of audio signal encoding,
CE L P (Code Exclted Linea
r Prediction) method or VXC (Ve
ctorExClted Coding) method is known as an excellent method that applies vector quantization. For details on the CELP method, please refer to M, R, 5ch.
“Code-E” by Mr. roeder and Mr. B, S, Ata1
xcited I, 1near Prediction
(CELP):Illgh-Quality 5pee
ch At Very Low BitRates
in Proc, 1CASSP, 1985. pp, 93
7-939.

The CELP method will be briefly explained below with reference to the drawings. FIG. 6 is a block diagram showing the basic configuration of a speech encoding circuit to which the CELP method is applied. In the figure, a signal sequence of an audio signal is led to a block extraction unit 12 via an input terminal 11, and this block extraction unit 12 divides L sample values into one frame into length input audio signal vectors. be done. This audio signal vector is then introduced into the LPG analysis section 13. This LPC analysis section 13 performs LPG analysis of the audio signal using an autocorrelation method or the like, and thereby LPG prediction parameters (ai
l (i=1. . . , P) is extracted. This LP
The C prediction parameter is supplied to the LPG synthesis filter 19. Note that P is the predicted order.

Further, the LPC analysis section 13 outputs an LPG prediction residual signal vector, and this LPG prediction residual signal vector is input to the pitch analysis section 14.

The pitch analysis unit 14 performs pitch analysis, which is a long-term prediction of speech, using the LPC pre-71111 residual signal vector, and extracts the pitch period Tp and the gain parameter b. Both the extracted pitch period Tp and gain parameter are supplied to the pitch synthesis filter 18.

Next, the process of generating synthesized speech will be explained. The codebook 15 stores in advance n white noise vectors having k dimensions (number of vector elements). Here, k is generally chosen such that L/k is an integer. Each of the above white noise vectors is read out one by one from the codebook 15, and first multiplied by a gain parameter generated from a gain parameter generation section 17 in a multiplier 16. Then, the pitch synthesis filter 18 and the LPG synthesis filter 19
A filter operation is performed by each of these methods, and a synthesized speech vector is generated. Pitch synthesis filter 18 at this time
The transfer function P (Z) of and the transfer function A (Z) of the LPG synthesis filter 19 are as follows.

A (Z) −1, / (1+Σa i Z force,,,
mP (Z)-1/ (1+bZ-TP)
(2) The synthesized speech vector thus generated is compared with the input speech vector, which is a target vector, in the comparator 20. Then, the comparison output passes through a weighted filter 21 and is then introduced into a squared error calculation section 22, where a squared distance distortion Ej is determined. In addition, these two
The value of the multiplicative distance distortion Ej is introduced into the minimum distortion search unit 23i, where its minimum value is detected.

The above synthetic speech generation process is performed for n white noise vector quantities, and thereby the white noise vector number j that gives the minimum value of the squared distance distortion Ej is selected. In other words, the CELP method is characterized in that vector quantization is performed by using a codebook for the drive signal of the synthesis filter in the process of speech synthesis. Note that since the human voice vector is long, this process will be repeated L/ times. Also, the weighting in the figure is used by the filter 5 to shape the spectrum of the error signal and reduce the distortion that would be perceived by the human ear, and its transfer function is given by the following equation. .

W (Z) −H(Z/γ) /HCZ)
...(3) H (Z) -A (Z) XP (Z)
...(4) In addition, when actually encoding and transmitting a human signal using the CELP method, the LPG prediction parameter, pitch period, pitch gain parameter, codebook number, and codebook gain are each encoded. transmitted.

On the other hand, FIG. 7 is a block diagram showing another example of the basic configuration of a speech encoding circuit to which the CELP method is applied. In this figure, the same parts as in FIG. 6 are designated by the same reference numerals.

In Figure i7, the weighting filter 21 has been moved from inside the search loop of the codebook 15 to outside the loop. This changes P (Z) of the pitch synthesis filter 18 to P
(Z/7), A (Z) of LPG synthesis filter 19
By setting A (Z/γ), it is possible to reduce the amount of calculation while maintaining the same function.

In addition, the pitch synthesis filter 18 and the LPG synthesis filter 1
The initial memory in the two filter operations is prevented from affecting the search of the codebook 15. That is, the pitch synthesis filter 24 with initial memory and the LPG
A synthesis filter 25 is newly provided. This LP
A zero input vector is output from the G synthesis filter 25,
It is supplied to a subtracter 26. In this subtracter 26, the weighted human speech vector outputted from the filter 21 is subtracted from the zero input vector. Comparator 2
0, the initialized weighting output from the subtracter 26 is compared with the input speech vector and the output vector of the LPC synthesis filter 19. That is, the weighted human speech vector obtained by subtracting the LPG synthesis filter 25 from the output vector of the filter 21 becomes the target vector, and thereby the initial memories of the pitch synthesis filter 18 and the LPG synthesis filter 19 are set to zero. Ru.

At the same time, with this circuit, the filter operation performed by the synthesis filter in FIG. 6 can be expressed by the product of the code vector and the next kXk lower triangular matrix.

Here, k is the number of dimensions (number of elements) of the code vector of codebook 15, and h (1) (i-1°...
, k) is the length k when the initial memory of H(Z/γ) is zero
is the impulse response of

The speech synthesis vector output from the comparator 20 is 2
The squared distance distortion Ej of the following formula is calculated by the multiplicative error calculation unit 22,
The minimum distortion search unit 23 finds the minimum value.

Ej -1lXt-γjHcjll
(6) Here, Xt is the target input vector, Cj is the j-th code vector, and γj is the optimal gain parameter for the j-th code vector.

Find the above Ej and find the vector number j that gives its minimum value.
A flowchart for determining the is shown in FIG. In this process, it is first necessary to find HCj for each j, which requires k(k+1)/2xn multiplications. , requires 839680 multiplications. Also, in the entire flow, if L/ is set to -4, 10
It requires 48736 multiplications.

Then, when the number of samples in one frame L = 160, and the sampling frequency of human voice is 8kllz, 52 old PS
DSP (Digital Signal Pro
cesscr) also requires three.

Furthermore, a method called closed-loop pitch prediction or an adaptive codebook is known as a method for improving the sound quality of the CELP method. For details of this method, see W, B,
Kleijln, D. J., Kraslnski, and
RIl, Ketchum, “Improved 5
peach Quality andEff'1cle
nt Vector Quantization in
CELP' 1nProc ICASSP, 198
8. pp. 155-158.

Hereinafter, a CELP method using a method called closed-loop pitch prediction or an adaptive codebook will be briefly described. The difference between the CELP method shown in FIG. 7 and the CELP method using a method called closed-loop pitch prediction or an adaptive codebook lies in the method of pitch analysis. In the former, pitch analysis is performed using the LPG prediction residual signal vector output from the LPG analysis section 13. On the other hand, the latter is characterized in that the pitch analysis is a closed loop similar to the codebook search.

In the CELP method, which uses a method called closed-loop pitch prediction or adaptive codebook, the pitch period is determined by delaying the driving signal of the LPG synthesis filter with a delay device that is variable over the pitch search range a-b. Create a drive signal vector for j. After this drive signal vector is multiplied by a predetermined gain parameter, filter calculation is performed using an LPG synthesis filter, and the calculation output is compared with a weighted human vector, which is a target vector, from which the influence of the previous frame has been removed. Then, based on the comparison results, the square distance distortion is determined by the following equation, and the drive signal vector at which this square distance distortion is minimized is searched.

Ej-llXt-γjHBjll ・
...(7) However, j is a≦j≦b, Xt is the target vector, Bj is the drive signal vector at pitch period j, γj is the optimal gain factor for pitch period j, and H is the above-mentioned ( 5) A lower triangular matrix of kXk, h given by Eq.
(1) (i −1, ..., k) is A (Z/γ)
is an impulse response of length k when the initial memory of is zero.

Find the above Ej and give its minimum value.
A flowchart for determining is shown in FIG.

In this process, first, for each t and j, HBj
It is necessary to find k (k+ 1)
/ 2 X (b − a + 1) X L /
Requires k multiplications. For example = 40. L-160
, a-20, and b-147, 419840 multiplications are required. Additionally, the entire flow requires 461,312 multiplications per frame. Then, when the sampling frequency of the input audio signal is 8 kHz, a multiplication of about 23 old PS is required, and 20 MIP is required.
Even a DSP with a multiplication capacity of S requires two.

(Problems to be Solved by the Invention) As mentioned above, in the conventional circuit applying the CELP method, a large amount of calculation is required to perform a codebook search and a pitch period search called a closed loop or adaptive codebook. Therefore, a large number of arithmetic circuits such as DSPs are required to perform real-time processing, resulting in an increase in the size and cost of the circuit configuration.

Therefore, the present invention focuses on the above-mentioned circumstances and develops a vector quantization method that can reduce the amount of calculation required to search the codebook, thereby reducing the number of arithmetic circuits and making the circuit smaller and cheaper. The purpose is to provide.

Another object of the present invention is to provide a vector quantization method that can reduce the amount of calculation required to search for pitch periods, thereby reducing the number of arithmetic circuits and making the circuit smaller and cheaper. The purpose is to

Still another object of the present invention is to reduce the amount of computation required for codebook search and pitch period search, respectively, thereby further reducing the number of arithmetic circuits, resulting in further miniaturization and lower cost of the circuit. The purpose of this study is to provide a vector quantization method that can achieve quantization.

[Structure of the Invention] (Means for Solving the Problems) In order to achieve the above object, the present invention performs predetermined filter operations on each of a plurality of code vectors stored in advance in a codebook, and thereby? In the vector quantization method, the filter operation is performed using the matrix of the code vectors and the It is expressed by the product of the number of dimensions of the code vector with a square triangular matrix, and the matrix of the code vector and the square triangular matrix are each constructed by a Toeplitz matrix, and each element of the product of the matrix of the code vector and the square triangular matrix is Among them, only the initial element is determined by ordinary matrix operations, and each of the other elements is determined recursively based on the previously determined element.

In order to achieve the above-mentioned other object, another present invention performs predetermined filter calculations on each of a plurality of drive signal vectors created according to the pitch period, and the calculated vectors and input signal vectors obtained thereby are In the vector quantization method, the filter operation is performed using a matrix of the drive signal vectors and a code prepared in advance. Each element of the product of the drive signal vector matrix and the square triangular matrix is represented by the product of the number of dimensions of the vector with a square triangular matrix, and the matrix of the drive signal vector and the square triangular matrix are each constructed by a Toeplitz matrix. Among them, only the initial element is determined by ordinary matrix operations, and each of the other elements is determined recursively based on the previously determined element.

Furthermore, in order to achieve the other object described above, another present invention provides that the first filter operation for performing the codebook search is performed by multiplying a matrix of code vectors and a square triangular matrix having the number of dimensions of the code vectors. and the matrix of the code vector and the square triangular matrix are each constructed by a Toeplitz matrix, and only the initial element of each element of the product of the matrix of the code vector and the square triangular matrix is determined by ordinary matrix operation, Each of the other elements is calculated recursively based on the previously calculated element, and the second filter operation for searching the pitch period is performed by multiplying the drive signal vector matrix and the square triangular matrix. The matrix of the drive signal vector and the square triangular matrix are each constructed by a Toeplitz matrix, and only the initial element of the product of the matrix of the drive signal vector and the square triangular matrix is calculated by ordinary matrix operations. Each other element is calculated recursively based on the previously calculated element.

(Operation) As a result, according to the present invention, in the search of the codebook, the filter operation is performed recursively using the Toeplitz matrix. Therefore, it is possible to reduce the amount of calculation compared to when all filter operations related to codebook search are performed using ordinary matrix operations, and this reduces the number of arithmetic circuits, resulting in smaller circuits and lower costs. will be promoted.

According to another aspect of the present invention, in searching for a pitch period,
Filter operations are performed recursively using Toeplitz matrices. Therefore, it is possible to reduce the amount of calculation compared to the case where all filter calculations related to pitch period search are performed using normal matrix calculations, and this reduces the number of calculation circuits, resulting in smaller circuits and lower costs. will be promoted.

According to still another aspect of the present invention, filter operations are performed recursively using a Toeplitz matrix in both the codebook search and the pitch period search. Therefore, compared to the case where all filter operations related to codebook search and pitch period search are performed using normal matrix operations, the amount of calculation can be significantly reduced, and the number of calculation circuits is also reduced. Further miniaturization and cost reduction of the circuit can be achieved.

(Embodiment) FIG. 1 is a block diagram showing the basic configuration of a speech encoding circuit to which a vector quantization method is applied according to an embodiment of the present invention.

In the figure, a signal sequence of an audio signal is input from an input terminal 31 and introduced into a block extraction section 32. In this block extraction unit 32, the L sample values are divided into blocks into length input speech vectors as one frame, and the human input speech vectors divided into blocks are
The C analysis unit 33 and weighting are supplied to a filter 34.

The LPG analysis unit 33 performs LPG analysis of the audio signal using an autocorrelation method or the like, and calculates LPG prediction parameters (all
Extract (i-1,..., P). This extracted L
The PG prediction parameters are used in the LPG synthesis filter 36.

A weighting filter 34 weights the input speech vector, and is provided outside each loop of codebook search and pitch period search. This weighting is performed by the filter 34 and the LPG synthesis filters 36, 38 .
At 47, the spectrum of the error signal is shaped by changing A (Z) to A (Z/γ), thereby making it possible to reduce the distortion that would be perceived by the human ear. Furthermore, by providing the weighting filter 34 outside the search loop, it is possible to reduce the amount of calculation when searching for the codebook and pitch period. The transfer function W (Z) of the weighting filter 34 is given by the following equation.

W (Z) −A (Z/γ) /A (Z)
...(8) Here, γ is (O≦γ≦1), and A (Z) is A (Z)-1/ (1+Σ aiZ-
−(9).

Further, the circuit of this embodiment has an LPG synthesis filter 38.4.
The initial memory used in the filter operation in step 7 does not affect the pitch period search and codebook search. That is, an LPG synthesis filter 36 having an initial memory is newly provided, and the customer force response vector is created by this LPG synthesis filter 36.

Then, in the subtracter 35, the weighting is applied to the filter 3.
The weighting output from 4 is obtained by subtracting the zero input response vector from the input speech vector, thereby setting the initial memory of the LPG synthesis filter 38, 47 to zero.

At the same time, this also changes the filter operation of the LPG synthesis filter 38 used for searching the pitch period and the filter operation of the LPG synthesis filter 47 used for searching the codebook into the drive signal vector and × It is possible to express this by the product of a lower triangular matrix and the product of a code-vector and a lower triangular matrix of kXk. The above kXk lower triangular matrix H is expressed as follows. Here, k is the number of dimensions (number of elements) of the drive signal vector and the code vector, and k is generally selected so that L/k is an integer. Also, h (1) (i
−1, . . . , k) is an impulse response of length k when the initial memory of A(Z/γ) is zero.

By the way, the pitch period search loop is configured as follows. That is, first, the drive signal vector corresponding to the pitch period j is accumulated in the drive signal section 40. This drive signal vector is obtained by adding the code vector multiplied by the gain parameter in the multiplier 46 and the drive signal vector multiplied by the gain parameter in the multiplier 39 in the adder 43, and the resulting filter drive signal delayer 4
4 with a variable delay. The delay device 4
4 is configured such that the delay amount changes over the search range a-b of the pitch period. The drive signal vector outputted from the drive signal section 12 is multiplied by a gain parameter in a multiplier 39, then subjected to a filter operation to create a synthesized speech vector in an LPC synthesis filter 38, and then input to a comparator 37. Ru. In this comparator 37, the synthesized speech vector output from the LPG synthesis filter 38 is subtracted from the weighted human speech vector output from the subtracter 35. Then, the error is introduced into the square error calculation unit 41, where the distortion Ej due to the squared distance is found, the minimum value of this distortion Ej is detected by the minimum distortion search unit 42, and the detected information is used for driving. The signal is input to the signal section 40.

That is, from the drive signal section 40 to the multiplier 39, the LPG
In the loop returning to the drive signal unit 40 via the synthesis filter 38, the comparator 37, the squared error calculation unit 41, and the minimum distortion search unit 42, the weighted human input signal from which the influence of the previous frame is removed is output from the subtractor 35. A pitch search is performed using the audio vector as a target vector. Here, the distortion Ej calculated by the square error calculation section 41 is Ej-11Xt-γjHBjl+...
・Represented by (11). However, j is a pitch period and is set in the range of a≦j≦b. In addition, Xt is a target vector, Bj is a drive signal vector at pitch period j,
γj is the optimal gain factor for pitch period j, H is the (
10) is a lower triangular matrix of kXk given by Equation 10).

By the way, each element of the drive signal vector Bj is (b
j (1), b j (2), ... b j (k))". Figure 2 shows each column vector of Bj. At this time, each element is b j (+a) - b j-1(11-
1) (a+1≦j≦b.

2≦m≦k), and therefore, the drive signal matrix with the drive signal vector Bj as each column vector becomes a Toeplitz matrix. Further, the weighted LPG synthesis filter is also a lower triangular matrix and a Toeplitz matrix. Therefore, each element of HBj is (V j (1), V 02),
...Vj(k))'', the following relational expression holds true.

Vm) -h (1) e (-j)
-(12)Vj(m)-Vj-1(m-1)+h
(+a) e (-j) - (13) where m and j are 2≦m≦ and a+1≦j≦b, respectively. As is clear from this relational expression, each element of HBj can be determined recursively from the previous element.

FIG. 3 is a flowchart showing the calculation procedure of a pitch search loop constructed with this point in mind.

According to this flowchart, first, in step S41, t
After the value of is initialized, HBa is determined by matrix-vector product operation in step S42.

Then, based on this HBa, step S43. 1 in S44
lHBall and (Xt, HBa)2 are determined.

After the initial elements are obtained, step S45
- In step S49, each of the other elements is repeatedly calculated recursively. That is, in step S45, the value of j is a
Each time one is set in the range of ll-b, the element HB is set based on the previous element HB j-1 in step 54B.
j is found recursively. Then, based on this HBj, 1lHB is calculated in step S47 and step 848, respectively.
jll and (XtHBj)2 are found. Further, in step S49, the magnitude of the squared distance distortion Ej is determined based on the above calculation result. Then, when HBj is determined for all j and the magnitude determination of distortion Ej is completed, the drive signal vector with the smallest distortion Ej is selected according to the determination result.

In other words, in the above calculation loop for pitch search,
Only the initial element is determined by a normal matrix-vector product operation, and all other elements are determined recursively based on the previously determined element. In other words, the number of multiplications required for such an operation is fk (k+1) /2+k ・
(b-a) becomes l ・L/. For this reason, for example, commonly used -40, L-160, a-20,
In the case of b-147, the number of multiplications is 23,600. In addition, the total number of multiplications is 65,072 times in the entire flow, which is 14 times compared to the conventional method shown in Figure 9.
The amount of calculation is about %. Therefore, if the sampling frequency of the input audio signal is, for example, 8 kHz, the number of multiplications required is approximately 3.3 old PS. As a result, for example, if it is a 20 old PS DSP, it will be sufficient to use just one, and even cheaper SPS with less computing power will be enough.
It is also possible to use P.

When the optimum pitch period j is found by searching for the pitch period, the comparator 37 calculates the optimum pitch period j from the weighted human voice vector outputted from the subtractor 35 and excluding the influence of the previous frame. The synthesized speech vector is subtracted, thereby obtaining a weighted human speech vector that removes the influence of the previous frame and the influence of pitch.

This input speech vector is fed to a comparator 48 of the codebook search loop.

On the other hand, the codebook search loop is constructed as follows. That is, this loop has a codebook 17, and this codebook 45 stores n codevectors composed of white noise. The code vector read from the codebook 17 is multiplied by a gain parameter in a multiplier 46, and then filtered in an LPC synthesis filter 47 to create a synthesized speech vector. A comparator 4 to which the synthesized speech vector created by this LPG synthesis filter 47 is input.
At step 8, the synthesized speech vector is subtracted from the weighted human speech vector supplied from the comparator 37 of the pitch period search loop, and the error vector obtained thereby is supplied to the squared error calculation section 49. These four squared error calculation units calculate the distortion Ej due to the squared distance, and the minimum distortion search unit 50 searches for the minimum value of this distortion Ej. That is, in the above loop, the weighted human voice vector output from the comparator 37 of the pitch period search loop, excluding the influence of the previous frame and the influence of pitch, is used as the target vector, and the code in the codebook 17 is used as the target vector. Distortion Ej due to the square distance of the error from the vector
Calculation is repeated to select the code vector number j that gives the smallest value. This calculation is expressed by the following equation.

Ej-11Xt-IIHcIII
...(14) Here, X is a weighted human vector excluding the influence of the previous frame and the influence of pitch, Cj
is the jth codevector, γj is the optimal gain factor for the jth codevector, and n is the number of codevectors.

By the way, as shown in FIG. 3, the code vector Cj is cut out from the rear side of the one-digit white noise sequence U whose length is n+-1 while shifting one sample at a time, and has a length k. Consists of samples. In this case, as shown in Figure 3, Cj (Ig) = Cj-1
There is a relationship of (a+-1) (2≦j≦n, 2≦m≦k), and the codebook matrix C with these code vectors Cj as each column vector becomes a Toeplitz matrix. Also, at this time, each element of HCj is expressed as (Wj (1), Wj (2), ・
...Wj(k))'', the following relational expression holds true.

Wj (1) = h (1) u (n+1-j)
-...(15)Wj(+a)-Wj-L(
11-1)+ h (m) u (n+1-j)
・= (1B) However, m and j are each 2≦m
≦, and 2≦j≦n. As is clear from this relational expression, each element of HCj can be determined recursively from the previous element.

FIG. 5 is a flowchart showing the calculation procedure of a codebook search loop constructed with this point in mind. According to this flowchart, first, in step S51, MCI, which is an initial element, is obtained by a normal matrix-vector product operation. Then, based on this MCI, in step S52, IIH
cIII is required.

Once the initial element is determined, operations are then repeated to recursively determine other elements. That is, each time a value of j is set in the range of 2 to n in step S53, element HCj is determined in step S54 based on the previously determined element HCj-1. and,
1lHcj11 is obtained from this element HCj in restep S55. After HCj is determined for all j, each time a value of t is set in the range of 1 to L/ in step 858, XtH is determined in step S57.
is required. Furthermore, each time this XtH is obtained, (Xt"H, Cj)2 for each position of j is obtained from step 358 to step Sho, and these (Xt"H, Cj)2 and the step S55 The magnitude of the distortion Ej of the square distance is determined from 1lHcjll obtained by .

Then, all (Xt
When the determination of the magnitude of the distortion Ej of the squared distance by ``H, Cj)2 is completed, the process returns to step S56, where the value of t is incremented, and then steps 357 to S60 are performed.
The calculation and the determination of the magnitude of the distortion Ej are repeated.

In other words, even in the above calculation loop for codebook search, only the initial element is calculated using the usual matrix-vector product operation, and all other elements are recursively calculated based on the previously calculated element. Desired. In other words, the number of multiplications required for such an operation is lk (k+1) /2
+k ・(n-1) 1. For example, in the case of -40°n-1024 used for boat fishing, 41.74
The number of multiplications is 0. Also, in the entire flow, 2
The number of calculations is 50,796 times, which is approximately 24% of the amount of calculation compared to the conventional product shown in FIG. 8, and the amount of calculation is significantly reduced. Therefore, when the sampling frequency of the input audio signal is 8kHz, a multiplication of about 12.5 old PS is sufficient, and the SP required for this calculation is 20 old PS.
If you have a multiplication ability of , you can make do with just one.

As described above, with this embodiment, it is possible to significantly reduce the amount of computation in the search loop in both pitch period searches and codebook searches, compared to conventional methods. The number of required DSPs can be reduced to one. Therefore, the configuration of the encoding circuit can be downsized and the cost can be reduced.

Note that the present invention is not limited to the above embodiments. For example, in the above embodiment, a case has been described in which pitch prediction is implemented in a closed-loop method or in a CELP method called an adaptive codebook, but it can also be implemented in the same manner in a CELP method whose circuit configuration is shown in FIG. 7 above. be.

In this case, h (i) of the lower triangular matrix H of kXk in equation (10) is converted to H (Z/γ) according to equation (4).
You can replace it with

Further, in the above embodiment, the code vector Cj is divided into lengths that are cut out from the rear side of the one-digit white noise sequence U whose length is n+ and -1 while being shifted one sample at a time, as shown in FIG. Although it is composed of samples of length k, it may be composed of samples of length k that are cut out from the front side of the white noise sequence U while shifting one sample at a time. In this case, in order to recursively obtain HCj for each Cj, k (K+1)/2+(2-1)(n-1) multiplications are required, which is smaller than the number of multiplications in the previous example. The number of multiplications increases by (k-1) (n-1) times. However, compared to the number of multiplications in the conventional method, the number of multiplications can be sufficiently reduced.
By reducing the number of circuits, the circuit configuration can be made smaller and cheaper.

In addition, the circuit configurations, calculation procedures, etc. of the codebook search loop and pitch period search loop can be modified in various ways without departing from the gist of the present invention.

[Effects of the Invention] As detailed above, the present invention expresses a filter operation for searching a codebook by the product of a matrix of code vectors and a square triangular matrix of the number of dimensions of the code vector, and The code vector matrix and the square triangular matrix are each constructed by Toeplitz matrices, and only the initial element of the product of the code vector matrix and the triangular matrix is determined by ordinary matrix operations, and each other element is − Each element is calculated recursively based on the previously calculated element.

Therefore, according to the present invention, the vector quantization method can reduce the amount of calculation required to search a food book, thereby reducing the number of arithmetic circuits and making the circuit smaller and cheaper. can be provided.

In another aspect of the present invention, the filter operation for searching the pitch period is expressed by the product of a matrix of drive signal vectors and a square triangular matrix of the number of dimensions of a code vector prepared in advance, and the drive signal Construct the vector matrix and square triangular matrix using Toeplitz matrices,
Only the initial element of each element of the product of the matrix of the drive signal vector and the square triangular matrix is found by ordinary matrix operation,
Each of the other elements is calculated recursively based on the previously calculated element.

Therefore, according to the present invention, the vector quantization method can reduce the amount of calculation required to search for the pitch period, thereby reducing the number of arithmetic circuits and making the circuit smaller and cheaper. can be provided.

In still another aspect of the present invention, the first filter operation for searching the codebook is expressed by the product of a matrix of code vectors and a square triangular matrix of the number of dimensions of the code vector,
The matrix of the above code vector and the square triangular matrix are respectively constructed by Toeplitz matrices, and only the initial element of the product of the matrix of the above code vector and the square triangular matrix is determined by normal matrix operation, and each of the other The elements are determined recursively based on the previously determined element, and the second filter operation for searching the pitch period is expressed by the product of the drive signal vector matrix and the square triangular matrix. And the matrix of the drive signal vector and the square triangular matrix are each constructed by a Toeplitz matrix,
Only the initial element of each element of the product of the matrix of the drive signal vector and the square triangular matrix is found by ordinary matrix operation,
Each of the other elements is calculated recursively based on the previously calculated element.

Therefore, according to the present invention, it is possible to reduce the amount of calculation required for codebook search and pitch period search, which further reduces the number of arithmetic circuits and further reduces the size and cost of the circuit. It is possible to provide a vector quantization method that can be reduced in price.

[Brief explanation of drawings]

1 to 5 are for explaining a vector quantization method according to an embodiment of the present invention, and FIG. 1 is a block diagram showing the basic configuration of a speech encoding circuit to which the same method is applied. Figures 2 and 3 are schematic diagrams used to explain the operation of the circuit, Figures 4 and 5 are flowcharts showing the calculation procedure of the circuit, and Figures 6 and 7 are respectively diagrams of different conventional vector quantum FIGS. 8 and 9 are block diagrams illustrating the basic configuration of a speech encoding circuit to which this method is applied, and are flowcharts illustrating the operational procedures of the circuits shown in FIGS. 6 and 7, respectively. 31... Audio signal input terminal, 32... Block extraction section, 33... LPC analysis section, 34... Weighting filter, 35... Subtractor, 36.38.47...
LPG synthesis filter, 37.48... Comparator, 39°46... Multiplier, 40... Drive signal section, 41.49
... Squared error calculation section, 42.50 ... Minimum distortion search section, 44 ... Delay circuit, 45 ... Code book.

Claims (3)

[Claims] (1) A predetermined filter operation is performed on each of a plurality of code vectors stored in advance in a codebook, and the code vector that minimizes distortion between the calculated vector and the input signal vector is selected in the codebook. In the vector quantization method, the filter operation is represented by the product of a matrix of the code vector and a square triangular matrix of the number of dimensions of the code vector, and the matrix of the code vector and a square triangular Each matrix is constructed by a Toeplitz matrix, and only the initial element of each element of the product of the code vector matrix and the square triangular matrix is calculated by ordinary matrix operations, and each other element is the previously calculated element. A vector quantization method that is characterized by recursively calculating each based on . (2) Perform predetermined filter calculations on each of the plurality of drive signal vectors created according to the pitch period, and select the plurality of drive signal vectors that minimize distortion between the calculated vector and the input signal vector. In the vector quantization method, the filter operation is represented by the product of a matrix of the drive signal vector and a square triangular matrix of the number of dimensions of the code vector prepared in advance. , and the matrix of the drive signal vector and the square triangular matrix are each configured by a Toeplitz matrix, and only the initial element of each element of the product of the matrix of the drive signal vector and the square triangular matrix is determined by ordinary matrix operation, A vector quantization method characterized by recursively determining each other element based on the previously determined element. (3) Perform a predetermined first filter operation on each of a plurality of code vectors stored in advance in a codebook, and select the code vector that minimizes the distortion between the calculated vector and the input signal vector. A means for searching from the codebook and a predetermined second filter operation are performed on each of the plurality of drive signal vectors created according to the pitch period, and the distortion between the calculated vector and the input signal vector obtained by this is In the vector quantization method, the first filter operation is performed using a matrix of the code vectors and a dimension number of the code vectors. represented by a product with a square triangular matrix, and the matrix of the code vector and the square triangular matrix are each constructed by a Toeplitz matrix,
Of each element of the product of the code vector matrix and the square triangular matrix, only the initial element is determined by normal matrix operations, and each of the other elements is determined recursively based on the previously determined element, A second filter operation is represented by a product of the matrix of the drive signal vector and the square triangular matrix, and the matrix of the drive signal vector and the square triangular matrix are each configured by a Toeplitz matrix, and the matrix of the drive signal vector is A vector quantum that is characterized in that only the initial element of the product of and a square triangular matrix is found by normal matrix operations, and each of the other elements is found recursively based on the previously found element. method.