JPH0783316B2 - Mass vector quantization method and apparatus thereof - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a method for encoding a signal sequence such as voice or image with a small amount of information, and is particularly effective for multiplexing when an error occurs in a code to be transmitted. The present invention relates to a vector quantization method and apparatus.

[Conventional technology]

A vector quantization method is known as a powerful method for compressing and encoding a signal sequence information. This is done by grouping a plurality of predetermined discrete signal sample values to be encoded into respective vectors, and collating each vector with a representative vector in the codebook that has been created in advance. The number of the representative vector that becomes smaller is used as the output code.

FIG. 1 shows its schematic configuration. Reference numerals 21 and 23 are codebooks, 22 is an encoder, and 24 is a decoder. Each symbol has the following meaning.

u: input vector u (i): i-th element of input vector u i = 0, 1, ..., K−1 k: vector dimension number r: code transmission rate [bit / sample] b: quantized transmission code (Kr bits) Z: code length z _l : l-th representative vector of codebook Z z (il): i-th element of representative vector z _l M: number of representative vectors z _l of codebook Z (M = 2 ^kr ) l = 0,1, ..., M-1 d ₀ : Quantization distortion The codebooks 21 and 23 have M = 2 ^kr representative vectors z _l (l
= 0,1, ..., M-1). On the transmission side, the encoder 22 refers to the codebook 21 for each input vector u, and M =
Distortion d _l represented by the square of the distance between each of the 2 ^kr representative vectors z _l and the input vector u is calculated, and the representative vector with the minimum distortion d _l is selected. The number 1 is transmitted as a transmission code b having a kr bit length. The distortion d _l is calculated by the following equation (1).

On the receiving side, the decoder 24 refers to the codebook 23 based on the received transmission code b (representative vector number), selects the representative vector corresponding to the transmission code b, and outputs it.

[Problems to be solved by the invention]

In this quantization method, if an error occurs in the transmission line, there is no distance relationship between the number of the representative vector and the value of the representative vector, so that a vector completely different from the input vector is reproduced. There are inherent drawbacks.

In order to prevent this, conventionally, it has been necessary to suppress the error rate by using an error correction code, that is, by providing a transmission code with redundancy. In this case, the bit error rate can be substantially reduced by using, for example, 50% redundant bits of the information bits. However, even if there are no code errors, it still requires the same amount of redundant bits. That is, under the condition that the total amount of information to be transmitted is the same, only 2/3 is allocated to the information bit even when no error occurs, and the quantization distortion becomes large. Practically, the bit error rate fluctuates with time, and it is difficult to change the form of the transmission code according to the situation, so the performance is sacrificed when there are no errors or when there are many errors. There was a need. Therefore, the error correction code is not always effective for the purpose of reducing the distortion under a certain amount of information. Also,
For the distance calculation (calculation of the quantization distortion d) in the vector quantization of the conventional encoder 22 shown in FIG. 1, M = 2 ^kr representative vectors z _l are stored in the codebook 21 having a memory capacity. , And the distortion d for each of the M
Therefore, there is a problem in that the calculation amount and the storage capacity of the codebook 21 increase due to the exponential function of the information amount kr per vector.

[Means and Actions for Solving the Problems]

An object of the present invention is to prevent a significant distortion in a decoded signal even when a code error occurs when compressing and coding information of a signal sequence such as a voice or an image, and further, a calculation amount required for the coding. And / or to provide a vector quantization method and apparatus with a small storage capacity.

According to the present invention, an input signal sequence is grouped into a plurality of samples to form one vector, and in a coding method in which each vector is quantized, a codebook of a plurality of systems is provided, and for each input vector Then, one representative vector is obtained from the codebook of each system so that the distortion between the input vector and the average vector of the representative vectors of each codebook is minimized, and the numbers of the obtained representative vectors are multiplexed. Output.

In this way, vector quantization is performed using a plurality of codebooks, and the transmission code is divided into a plurality, so that the quantization distortion is slightly larger when there is no code error as compared with the conventional vector quantization. When a code error occurs, it is possible to suppress an increase in distortion due to the code error.

〔Example〕

FIG. 2 shows a schematic configuration when the vector quantizer 100 according to an embodiment of the present invention is used in a transmission system. 11 and 12 and 15 and 16 are codebooks, 13 is an encoder, and 14 is an encoder. Is a multiplexer, 17 is a decoder, and 18 is an output vectorizer. That is, this embodiment shows a case where two systems of codebooks are provided. Here, codebooks 11 and 15 are X systems, codebook 1
2,16 is the Y system. The symbols used in the following description are defined as follows.

u: input vector k: number of vector dimensions, that is, number of samples forming the input vector r: transmission rate (bits / samples) b: quantized transmission code (kr bits) X, Y: codebook x (i, j) : I-th representative output vector xj from codebook X
Element y (i, m) i-th element of representative output vector ym from codebook Y u (i): i-th element of input vector u, that is, i-th sample value N: number of representative vectors of each codebook X, Y (= 2 ^{kr / 2} ) i: 0,1, ..., k-1 j: 0,1, ..., N-1 m: 0,1, ..., N-1 In this embodiment, each codebook 11 , 12 has 2 ^{kr / 2} representative vectors, and the codebooks 15 and 16 on the receiving side are also 2 respectively.
It will have ^{kr / 2} representative vectors. The codebooks 11 and 12, 15 and 16 of each system are created by using a learning sample that is statistically the same as the input to be quantized and determining a representative output point by learning. The method will be described later. Alternatively, the vector extracted from the learning sample is used as the representative output vector.

In the vector quantizer on the transmission side, the encoder 13 refers to two code lengths 11 and 12 for one input vector u,
The amount of information k · r bits assigned to the quantized transmission code b
Vector quantization is performed by 1/2. That is, the number of bits of the quantized codes b _x and b _y of each system is represented by k · r / 2. The multiplexer 14 is a quantization code for each system.
b _x and b _y are multiplexed, that is, a vector quantization code b = b _x b _y represented by k · r bits is set and transmitted to the receiving side 200. In the receiving side 200, the decoder 17, respectively with reference to the codebook 15 on the basis of the code b _x, b _y of each system included in the transmission code b, to obtain an output vector x, y of each system. The vector combiner 18 obtains the arithmetic mean of the output vectors x and y of each system,
The final output vector.

FIG. 3 shows a specific example of the configuration of the vector quantization device 100 of the present invention. Each column of the codebook 11 of the X system is a representative vector x
It consists of j and its vector number j (that is, quantization code b _x ). Similarly, each column of the codebook 12 of the Y system also has a representative vector y.
It consists of m and its vector number m (quantization code b _y ). As shown in FIG. 2, the number of representative output vectors xj, ym of the codebooks 11 and 12 of each system is 2 ^{kr / 2.} Therefore, the vector number j, m (quantization code b _{x for} one system) , b _y ) has a kr / 2 bit configuration. On the toilet bowl, in Fig. 3, vector number j, m
Are represented by 4 bits.

The encoder 13 can be functionally divided into an average vector calculation unit 131, a squared distance calculation unit 132, and a minimum distance determination unit 133.

The average vector calculation unit 131 sequentially calculates the average vector v for each combination of each representative vector xj of the X-system codebook 11 and each representative vector ym of the Y-system codebook 12.
_{j, m Vj, m} = (xj + ym) / 2 (2) is calculated. The squared distance calculation unit 132 calculates the input vector u represented by the squared distance and the average vector v _{j, m} distortion d _{j, m} d _{j, m} = | u−v _{j, m} | ² = | u− (x j + ym) / 2 | ² (3) is calculated for all combinations of (j, m). Since there are N values of j and m, respectively, N ² d _{j, m} are calculated in total. It is clear that M = N ² in FIG.

The minimum distance determination unit 133, together with each combination of the representative vectors xj, ym of each system, the input vector u corresponding to the combination.
A mean vector v _j, the distortion d _j and _m, and sequentially input _m, strained d _j, representative output vectors xj _{combinations m} takes a minimum value, ym
To decide. Then, the determined representative output vector x
The codebooks 11 and 12 of the respective systems are referred to based on j and ym, and the vector numbers j and m are output as the quantized codes b _x and b _y .

Vector number j of each line, m (quantization code b _x, b _y) are multiplexed in a multiplexer 14, the original vector quantization code b for the input vector u is obtained. In FIG. 3, X and Y
System vector numbers (quantization code) b _x = 1111, b _y = 00
It is shown that 01 is time-multiplexed like “11110001”. Frequency multiplexing or other diversity techniques may be used instead of time multiplexing.

As is apparent from the above description, in the vector quantizer using the two codebooks of the present invention shown in FIG. 3, one representative from each of the two codebooks X and Y for the input vector u. Select the vector xj, ym, and select its mean and (x
j + ym) / 2 and the distortion d of the input vector is minimized
Since j and m are determined, the selected representative vectors xj and ym
Are very likely to be chosen with close proximity to each other.
Therefore, these quantization code _{b x = j, b y =} m are multiplexed (bond), if it is transmitted as a vector quantization code b = jm, the first half of the code b received at the receiving side of FIG. 2 ( Even if one of j) or the latter half (m) contains a 1-bit error, if the other is correct, one of the decoded vectors x and y is correct, and therefore the decoded average output vector (x + y) ) / 2 is not far from the input vector u on the transmitting side. That is, the distortion of the decoded vector due to the transmission error becomes smaller. In the embodiment of the present invention described with reference to FIGS. 2 and 3, N = 2
^{Since two} codebooks that store ^{kr / 2} representative vectors are used, a memory that stores 2N representative vectors in total is sufficient. On the other hand, the conventional vector quantization shown in FIG. 1 requires a memory for storing M = 2 ^kr = N ² representative vectors.

The calculation of the squared distance d _{j, m} shown in the equation (3) can be expressed in more detail as the following equation (4). However, in order to make the formula easier to see, x / 2 and y / 2 are replaced by x and y in the following description.

FIG. 4 shows the configurations of the mean vector calculation unit 131 and the squared distance calculation unit 132 in FIG. 3 in the case of calculating the distortion d _{j, m} based on the equation (4) together with the codebooks 11 and 12. One representative vector xj, ym is read from the X codebook 11 and the Y codebook 12 for the input vector u, and the corresponding i-th element u (i), x (i, j), y (of these vectors) is read. i, m) to adder 31
Then, u (i) -x (i, j) -y (i, m) is calculated, and the addition output is squared by the square cumulative adder 32 and cumulatively added. Such two additions (subtractions) and one square cumulative addition are performed a total of three times for k elements from vector elements i = 0 to i = k−1, and as a result, the selected representative is selected. Vector xj
The distortion d _{j, m} for the set of Representative vector x
There are N ² combinations of j and ym, and 3 kN ² operations are performed to obtain the strains d _{j, m} for all of them and determine (j, m) that gives the smallest d _{j, m} among them. Need quantity,
This is to perform one subtraction and one square cumulative addition in the equation (1) that requires the conventional vector quantizer of FIG. 1 from i = 0 to i = k-1. N run ² times)
Instead, it will increase. That is, in the embodiment of the present invention shown in FIG. 4, the total memory capacity required for the codebooks 11 and 12 is significantly smaller than that shown in FIG. 1, but there is a drawback that the required calculation amount increases. An embodiment in which this point is improved will be described below with reference to FIG.

When the expression (4) is expanded, it becomes the following expression (5).

However The first term of the expanded equation (5) is the selection of the representative vector (j,
m) is irrelevant. Therefore, it is not necessary to calculate u ² (i) to determine the vector code combination (j, m) that minimizes the distortion d _{j, m} . Therefore, instead of equation (5), the combination (j, m) of codes that minimizes d' _{j, m} defined by the following equation (7) may be determined.

In equation (7), the second term F (j, m) is independent of the input vector u as defined in equation (6), and the codebook X,
It can be calculated only from the representative vector xj, ym selected from Y.
Therefore, the result of calculating F (j, m) of the formula (6) for all combinations (N, ² combinations) of (j, m) is stored in the memory in advance, and d ′ _j is calculated by the formula (7). _{, m} is calculated and the corresponding F (j, m) is read out, for example, equation (7)
The calculation amount of can be reduced.

FIG. 5 shows the mean vector calculator 13 in the embodiment of FIG.
This is an example in which the first and second distance calculating unit 132 and the codebooks 11 and 12 are configured based on the equation (7). As described above, F calculated in advance for all combinations (j, m) by equation (6)
A table memory 33 that stores the value of (j, m) is provided. The first term of equation (7) is the sum of the inner product of the vectors u and xj and the inner product of the vectors u and ym, which are calculated by the inner product calculators 34 and 35, respectively. The inner product calculation result is given to the adder 35 as a scalar value, respectively, and stored in the table memory.
It is added to the corresponding value of F (j, m) read from 33. The output of the adder 36 is the calculation result of d' _{j, m} by the equation (7). In the embodiment shown in FIG. 5, the inner product calculator 3
4,35 inner product computation 2kN times by a scalar subtraction is ^two times 2N by the adder 36. The required storage capacity of the storage device is ² kN of the code books 11 and 12 and N ² of the table memory 33.

Table I is calculated assuming that the storage capacity and the amount of operation required for the vector quantizer according to the prior art of FIG. 1 and the embodiments of FIGS. 4 and 5 according to the present invention are k = 16 and N = 16, respectively. Is shown.

In the above description, the embodiment using the squared distance is shown as a representation of the quantization distortion d _{j, m} or d ′ _{j, m} . These are the LS of the spectral envelope parameters in speech coding.
It can be applied when converting into a P parameter or a logarithmic cross-sectional area function for vector quantization. However, in order to quantize a speech waveform signal or a prediction residual signal with high efficiency, it is necessary to perform adaptive weighted distance calculation in the frequency domain.
The distance measure in that case is expressed as follows.

Here, w (i) represents the i-th element of the vector weight w and changes every time u (i) changes, but it is fixed along with u (i) while the minimum value of d is obtained. Note that w (i) is u
Regardless of (i), if it is fixed, it can be easily reduced to the case of the square distance. If the equation (8) is expanded and the first term that does not contribute to the minimum determination of distortion is omitted and only the second term and the third term are left, then d ^* is assumed. F (i, j, m) = {x (i, j) + y (i, m)} ² (10) Calculate all elements of F (i, j, m) in equation (10) in advance. In order to remember, unlike the case of formula (6),
A memory with kN ² elements is required. Furthermore, the following equation (1
In order to calculate the second term of the equation (9) defined in 1), kN ² inner product operations are required.

In order to calculate the first term of the equation (9), s (i) defined by the following equation (12) is previously calculated only once for i = 0, 1, ..., K.

s (i) = − 2w (i) u (i) (12) Further, the inner product of N elements is performed twice for s (i) to obtain G and H defined by the following equations (13) and (14). .

As a result of such preparations, d ^* _{j, m in the} equation (9) can be obtained by the following scalar operation.

d ^* _{j, m} = G (j) + E (m) + E (j, m) (15) FIG. 6 shows a configuration for executing the above calculation in the embodiment of FIG. In the table memory 33, kN ² F (i, j, m) values given by the equation (10) are calculated and written in advance. The vector quantizer has an input vector u
The weighted vector w is input together with this, and the corresponding i-th element of these vectors is calculated by the multiplier 37 as shown in Expression (12). Weighted input vector s obtained by this
(I) (i = 0, 1, ... K-1) is given to the inner product calculators 34 and 35, respectively, and the representative vectors xj, ym read from the codebooks 11 and 12 and the expressions (13) and (14), respectively. ) Is taken as the dot product. On the other hand, the weighted vector w is the vector F (i, j, m) (i read from the table memory 33 corresponding to (j, m)).
= 0,1, ... K-1) and the inner product calculator 38 calculates the inner product shown in Expression (11). The inner product calculation results from the inner product calculators 34, 35, 38 are given to the adder 36 to perform the scalar addition of Expression (15), and the value corresponding to the distortion when the input vector u is encoded into j, m d ^* _{j, m} is obtained.

Table II shows the necessary arithmetic units and storage capacities in the embodiment of FIG.
Shown in. Table II shows the formula (1) corresponding to Fig. 1 for comparison.
Also, the calculation amount and the memory capacity necessary for the vector quantization calculation in the case where the weight w is added to the formula (4) corresponding to FIG. 4 similarly to the formula (8) are also shown. The description will be omitted because it can be easily obtained in the same manner as.

In the case of the weighted distance, in comparison with, the calculation amount or the storage capacity respectively increases,
By combining and, both the calculation amount and the storage capacity can be made smaller than in the case of. That is, F (i, j, m) in equation (9)
You only need to have a part of the table. If the ratio with the table is λ (0 <λ <1), the calculation amount is 2N ² + k
+ 2kN + (3-2λ) kN ² and storage capacity is 2kN + λkN ²
Becomes By appropriately selecting λ, it is possible to interchange the calculation amount and the storage capacity so as to satisfy the given design condition.

It is also possible to further develop F (i, j, m) in the equation (10) to make each term a separate table. In this case, the calculation amount increases as it is, but the calculation of the product of x (i, j) · y (i · m) is partially or entirely omitted, and d ^* _{j, m} An approximate calculation of can be performed. At this time, it is not always possible to select a code that minimizes distortion, but the amount of computation and storage capacity can be significantly reduced.

It can also be easily applied to the case where the distance measure includes the free parameter τ that is a constant multiple of the codebook vector, as in Eq. (17).

When quantizing a residual signal of speech in the time domain, the distortion measure shown in the following equation is often used. The calculation of this distortion d _l is Using. When this distortion measure is applied to the vector quantization method of the present invention, the distortion d _{j, m} can be expressed by the following equation.

However, h (i, g) is the i-th row, g of the matrix H that performs the convolution product.
Representing the elements of the column, τ is a parameter determined so that the distortion d _{j, m} is minimized.

The definition of the distortion d _{j, m} in equations (3), (4) and (5) reflects the performance in the absence of code error. In order to make it stronger against code errors, the definition of distortion as in the following equation (17) is also effective.

d _{j, m} = μ | u−v _{j, m} | ² + (1 + μ) {| u−xj | ² + | u−ym | ² } / 4 (17) where μ is from 0 to 1. When μ is set to 1 as a parameter, it is equivalent to the distortion of the equation (3), and when μ is reduced, the redundancy of each codebook increases, and the distortion can be reduced if only one of the two systems has no error. That is, it is resistant to code errors.

Further, considering the transmission line code error, the following equation can be considered as the definition of the distortion of u with respect to xj and ym.

Here, q (f | j) indicates the probability that the vector number j is mistaken as f on the transmission path. The same applies to q (m | g).

Although these distortion measures are based on the squared distance, they can be easily applied to the above-mentioned weighted squared distance and other measures. Further, since it is assumed here that the information is equally divided into a plurality of systems, the final output vector is obtained by the arithmetic mean, but when the distribution of information is different for each system, it is obtained by the weighted average. For example, when the square distance is used in two systems, the transmission rate of each system is r (x), r (y), and v _{j, m} = {x _j · 2 ^{2r (x)} + y _m · 2 ^{2r (y)} } / {2 ^{2r (x)} +2
^{2r (y)} }. This is based on the fact that the expected value of the distortion of ^x can be approximated to 2 ^{-2r (x)} .

The procedure for creating the two codebooks 11 and 12 used in the embodiment of the present invention shown in FIGS. 2 to 6 will be described below with reference to the flowchart of FIG. This flow chart is an example in which two codebooks are alternately modified to locally minimize the distortion, but it can be easily applied to the creation of two or more codebooks.

In step S _n , a large number of vectors consisting of continuous k samples extracted at arbitrary positions from the learning sample sequence are prepared, and these are arbitrarily divided into two by the same number to obtain initial representative vectors. Make books X and Y. Alternatively, a group of vectors created by normal vector quantization learning using a learning sample may be set as a codebook X, and a group of vectors created by using an error sequence of the vector may be set as a codebook Y.

Next, in step S ₁ , a learning vector of a number sufficiently larger than the number of representative vectors is used as an input vector u, and for each input vector, a distortion d according to, for example, equation (3) is calculated for all representative vector pairs xj, ym, A set of representative vectors xj, ym having the smallest distortion d is determined. That is, each input vector is quantized by the quantization method of the present invention, and each input vector u
Determines which set of representative vectors xj, ym belongs to.

Step S calculates the sum D of the ₂ step S ₂ minimum distortion d for each input vector calculated in, codebook X if the value is the threshold or less is less than or equal to the threshold or L improvement of, Y is the objective It is determined that the vector is composed of the representative vector, and the learning of the codebook is stopped. If not, the process proceeds to the next step S ₃ , and the content ym of the codebook Y is improved while the content xj of the codebook X is fixed.

In other words, step S _3-1 In attributions representative vector of each input vector u decided in step S ₁ set (x _j, y _m) is regarded as a variable vector representative vectors ym based on attribution the variable vector ym An expression obtained by partially differentiating the sum of distortions of all input vectors u included in a vector set with a variable vector ym is set to 0, and a vector value _m obtained by solving the equation with respect to the variable vector ym is set as a new representative vector ym.
All the input vectors P (P is variable) that includes the representative vector y _m in the set of belonging vectors are expressed as u _m1 , U _m2 , ..., U _mp, and the set of vectors to which they belong is (x _f1 , ym) , (X _f2 ,
y _m ), (x _f3 , y _m ), ... (x _fp , y _m ), _m is finally given by the following equation as their center of gravity. However, fe takes any value from 0 to N-1.

Then fixed using a codebook X and the updated codebook Y for each input vector u as in step S ₁ in step S _3-2, attributable for X was fixed to the state of step S ₁ As it is, the attribution to Y is determined, and the codebook Y is updated in step _3-3 in the same manner as in step S _3-1 .

In step S ₄ , using the codebook X and the updated codebook Y, the set of representative vectors (xj, y) to which each input vector u belongs
m) are all determined. Step S ₄ is equivalent to step S ₁ .

In step S ₅ , the codebook Y is fixed while the contents of the codebook Y are fixed.
Improve the content of. That is, based on the set (xj, ym) of representative vectors to which each input vector u belongs determined in step S ₄ , the representative vector xj is regarded as a variable vector in step S _5-1 and the variable vector xj is Let 0 be the expression of the partial differentiation of the sum of the distortions of all input vectors u included in the set with respect to the variable vector xj, and solve the equation j for the variable vector xj.
Let j. All the input vectors Q (Q is variable) including the representative vector xj in the set of belonging vectors are u _j1 , u _j2 ... u _jQ
And the set of vectors to which they belong is (x _j , y _g1 ),
(Xj, y _g2 ),…, (xj, y _gQ ), j is finally given as the center of gravity of them by the following equation. However, ge is 0 to N
Takes any value up to -1.

Next, using the codebook X updated in step S _5-2 and the fixed codebook Y, for each input vector u, the attribution to Y is fixed in the state of step S ₄ and all the attributions to X are determined. In step S _5-3 , the codebook X is updated again in the same manner as step S _5-1 .

Returning to step S ₁ , the same iterative calculation as in steps S _{2 to} S ₅ is performed below, and if the sum D of distortions is less than or equal to the threshold value or the improvement rate of D is less than or equal to the threshold value in step S ₂ , Update is completed.

In the above, steps S _3-2 , S _3-3 and S _5-2 , S _5-3 may be omitted. Further, steps S _3-2 , S _3-3 and S _5-2 , S _5-3 may be alternately repeated.

FIG. 8 shows SNR of how distortion is reduced by codebook learning. The symbols of step S in the flowchart of FIG. 7 in the figure are S ₁ , S _3-1 , ●, ▲, ◆, △, ◇.
Corresponds to S _3-2 , S _5-1 , S _5-2 . Even if the set of steps S _3-1 and S _{3-2 and} the set of steps S _5-1 and S _5-2 are repeated several times, the distortion is still reduced.

FIGS. 9A and 9B are comparisons of damage ranges when there is a transmission code error. FIG. 9A shows the case of the conventional one-system quantization shown in FIG. 1, and FIG. This is the case of system quantization. Error bits are circled.
By assigning codebooks and transmission codes of multiple systems to one input vector, the transmission code unit is divided, so the probability of error in all systems is not divided into the error rate of the ordinary transmission code. It will be smaller than that. In other words, as shown by hatching in Figures 9A and 9B, 2
In the case of the system, the range affected by the 1-bit code error is half that in the case of the system.

For the sequence with the memoryless Laplace distribution of Fig. 10, r =
The performance in the case of 1 is shown. The Laplace distribution simulates a speech prediction signal as a residual signal of linear prediction. In the figure, (a) is two-system quantization according to the present invention, and (b) to (e) are conventional ones.
It is the quantization of the system. The horizontal axis is the bit error rate, and the vertical axis is SN
R. It can be seen from FIG. 10 that the two-system quantization has almost the same performance as the one-system quantization of the same number of dimensions when there is no code error, and is excellent when there is a code error.

FIG. 11 is a comparison of two systems of quantization according to the present invention with a combination of a conventional vector quantization code and an error correction code under a constant total information amount. Here, seven types of BCH codes capable of multiple error correction were used as the correction codes.
All information bits, information source bits, and error correctable bits are shown in parentheses from the left. Each of these codes has the drawback that the distortion increases due to redundant bits even when there is no error, and the distortion increases rapidly when the code error exceeds a certain level, and the quantization of the present invention is superior. There is.

FIG. 12 shows an example in which the 2-channel vector quantization of the present invention is applied to transform coding by weighted vector quantization, which is one of the powerful techniques for mid-band speech coding. The input digital audio signal S is given to the linear prediction analyzer 41, and the inverse filter is performed by using the linear prediction parameter {αi} as a filter coefficient.
The linear prediction residual signal R obtained through the speech signal S is obtained from the inverse filter 42. The residual signal R is a cosine converter 43
The DCT (discrete cosine transform) coefficient is rearranged on the frequency axis and divided into a plurality of subvectors. The vector L ₁ thus obtained is vector-quantized by the 2-channel weighted vector quantizer 100 according to the present invention with the weight of the spectrum envelope and transmitted as the waveform information B. Along with this, the pitch period Δt
The auxiliary information A including the linear prediction parameter {αi} and the signal power P is also encoded and transmitted. Receiver decoder
Reference numeral 200 decodes the representative vector set xj, ym from the waveform information code B received by referring to the codebooks 15 and 16, and outputs the average of them from the decoders 17 and 18. This average vector is given to the cosine inverse transformer 14 and the residual signal R'is decoded. On the other hand, the received parameter information A is decoded by the parameter decoder 45, of which the linear prediction parameter {αi} is given to the synthesis filter 46 as a filter coefficient. The residual signal R'is passed through a synthesis filter 46 to synthesize a voice.

In order to create two-channel quantization codebooks 11 and 12 (and 15 and 16), learning is first performed from a Gaussian random number sequence with a distortion measure weighted with a long-term average spectrum. In the vector created at this time, the power of the component corresponding to the high band on the frequency axis becomes almost zero, and the high band component of the decoded voice is lost. To alleviate this, the representative vector is replaced with a vector having the smallest weighted distance among the Gaussian random numbers in the learning sequence.

Compared to the case of using the conventional one-channel vector quantization shown in FIG. 1, the SNR for the code error in the transmission difference of FIG. 12 is shown in FIG. As the input voice S, voices of one male and one female are repeatedly used with different code error patterns,
Each SNR is an average of about 90 seconds. Except for (a), normal 1-channel vector quantization is used for waveform quantization. All the auxiliary information B and the waveform information in the case of (c) were error-corrected by the (31, 21) BCH code capable of double error correction. This is because the auxiliary information is about 20% of the total information, and the waveform distortion due to the redundant bits of the error correction code is slightly increased, but the damage is large when an error occurs in the waveform information.

From the results shown in FIG. 13, it can be seen that the channel quantization can reduce the waveform distortion more than others for an error of about 0.3% when there is no error.

Next, the case where the present invention is applied to the quantization of the linear prediction parameter {αi} of the speech coding of FIG. 12 will be described. Here, first, the linear prediction coefficient {αi} is converted into an LSP (line spectrum pair) parameter with which the stability management of the filter is easy and the interpolation characteristic is good (US Pat. No. 4,393,272). afterwards,
For this parameter, vector-scalar quantization is performed with a vector portion of 8 bits and a scalar portion of 13 bits. Spectral distortion (LPC cepstrum distance) at the time of decoding by forcibly inverting each 8-bit code corresponding to the vector portion at this time is a comparison between normal vector quantization and quantization of the present invention. Is FIG. 14. The vertical axis represents the cepstrum distance, the horizontal axis represents the inverted bit position in the code, the white bars represent normal vector quantization, and the vertical bars represent vector quantization according to the present invention. The spectral distortion is evaluated because the code error in the linear prediction parameter such as the LSP parameter is directly reflected as the quantization distortion of the coded speech. The smaller the spectrum distortion, the smaller the deterioration. From this figure,
It can be seen that the present invention, which is divided into two channels, has less influence on the spectral distortion. When there is no error, normal vector quantization is superior, but the difference is small.

〔The invention's effect〕

As described above, according to the vector quantization method of the present invention, since a single input vector is quantized by a plurality of systems using a plurality of codebooks, the unit of the quantized code is divided, and the resultant code is divided. The probability of occurrence of errors in all the divided codes in the transmission is extremely small compared to the transmission error rate of the normal quantized code without division. In other words, the range of 1-bit code error is small. Therefore, the damage on the decoding vector is reduced. On the other hand, when there is no code error, the combination of output codes is determined so that the average of the representative vectors of each system minimizes the distortion with the input, so that it is almost the same as the normal one-system quantization of the same information amount. It can be expected to have performance.

Further, the multi-system vector quantization method of the present invention has the advantage of being capable of significantly reducing the storage capacity of the codebook as shown in Tables I and II, as well as the advantage for such a code error. There is an advantage that the calculation amount of search can be reduced. In addition to this, by narrowing the search candidates to only the neighborhood of the input vector in each codebook, it is possible to reduce the amount of calculation without degrading the performance.

Therefore, by using the quantization method of the present invention for encoding a speech waveform or image encoding, improvement in performance can be expected. In particular, it is effective for applications where there is a transmission path error and information compression is required.

[Brief description of drawings]

FIG. 1 is a block diagram showing a coding device on the transmitting side and a decoding device on the receiving side for performing vector quantization using one conventional codebook, and FIG. 2 is a vector quantizing device of the present invention and its decoding device. FIG. 3 is a detailed block diagram of the vector quantizer of the present invention shown in FIG. 2, and FIG. 4 shows a configuration for performing square distance calculation in the vector quantizer of FIG. Block diagram, FIG. 5 is a block diagram showing another configuration for performing squared distance calculation in the embodiment of FIG. 3, and FIG. 6 shows a configuration for performing weighted squared distance calculation in the embodiment of FIG. A block diagram and FIG. 7 are used in the vector quantizer of the present invention.
FIG. 8 is a flow chart showing the procedure for creating one codebook by learning, FIG. 8 is a graph showing the improvement of the codebook when the learning is repeated according to the flow chart of FIG. 7, and FIGS. 9A and 9B are the conventional vector quantization method and this FIG. 10 is a diagram comparing the range of damage of code error in the vector quantization method of the invention, FIG. 10 is a graph comparing the vector quantization method of the present invention and the conventional vector quantization method with respect to the SNR with respect to the code error rate. A graph comparing the effect of the vector quantization method of the present invention with that of the prior art. FIG. 12 shows a transmission system when the vector quantization method of the present invention is applied to weighted vector quantization in the frequency domain of a speech residual waveform. The block diagram shown in Fig. 13 is for the code error of the position system shown in Fig. 12.
Fig. 14 is a graph comparing SNR with the case of using the conventional vector quantization, and Fig. 14 shows the spectrum distortion against the code error when the present invention is applied to the quantization of the linear prediction parameter in Fig. 12 by the conventional vector quantization method. The graph which compared with the case where is used is shown.

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Office reference number FI technical display location H04N 7/30

Claims (35)

[Claims] 1. A plurality of codebook storage means each having a predetermined number of representative vectors, and a set of representative vectors selected one from each of the plurality of codebook storage means for an input vector. Which of the distortion calculation means executes the calculation of the distortion between the average vector and the input vector for a plurality of different sets of representative vectors and which gives the smallest distortion among the distortions of the average vector of each set. A multi-vector quantizing device including: a minimum distortion determining means for determining whether or not; and a multiplexing means for multiplexing and outputting codes respectively representing the determined representative vectors of the set. 2. The multiplexing means is means for time-multiplexing each of the codes representing the representative vector.
The described multi-vector quantizer. 3. The multi-vector quantizer according to claim 1, wherein the distortion calculating means is means for calculating a squared distance between the input vector and the average vector as the distortion. 4. The multi-vector quantizer according to claim 1, wherein the distortion calculating means is means for calculating a weighted square distance between the input vector and the average vector as the distortion. 5. The subtraction means calculates the vector of the difference between the average vector and the input vector, and
The sum of the squares of the respective elements of the difference vector calculated by the subtraction means is calculated and output as the distortion 2
The multi-vector quantization device according to claim 1, further comprising a multiplication and addition means. 6. The distortion calculating means, for each set of representative vectors selected one by one from each of the plurality of codebooks, sum of squares of each element of the sum vector of respective representative vectors of the set. Table memory means for storing the input vector, and an inner product operation for calculating the inner product of the representative vector selected from the plurality of codebook storage means for the input vector and the input vector, respectively. And subtraction means for calculating the difference between the sum of the respective inner products from said inner product calculating means and the sum of squares corresponding to said selected set read from said table memory and outputting it as said distortion. The described multi-vector quantizer. 7. The distortion calculating means is selected from a multiplying means for multiplying the input vector and a weight vector input corresponding to each other and outputting a weighted input vector, and one from each of the plurality of codebooks. For each set of representative vectors, table memory means for storing a constant vector whose element is the square of the sum of the corresponding elements of the respective representative vectors of the set, corresponding to the set, and the plurality of codebooks. First inner product calculating means for calculating an inner product of the representative vector of the set selected from the storage means and the weighted input vector from the multiplying means, and read from the table memory means in association with the selected set. Second inner product calculating means for calculating an inner product of the constant vector and the weight vector, and an adding means for outputting a difference between the inner products from the first and second inner product calculating means as the distortion. The multi-vector quantizer according to claim 1, including: 8. The two codebooks are provided.
8. The multi-vector quantizer according to any one of 7 to 7. 9. Each of the input vectors u has k elements, and N representative vectors xj, j = 0, 1, ..., N each having k elements in the plurality of codebook storage means. An X codebook having −1 and a Y codebook having N representative vectors ym, m = 0, 1, ..., N−1 each having k elements are stored, and k is The multi-vector quantizer according to claim 1, wherein N is a positive integer and 2 or more. 10. When the i-th elements of the input vector u and the representative vector xj, ym are expressed as u (i), x (i, j) and y (i, m), respectively, the distortion calculation means For the representative vector xj, ym of the selected set (j, m), 10. The multi-vector quantizer according to claim 9, defined by 11. The distortion calculating means comprises the representative vector xj,
a table memory means for storing values obtained by calculating F (j, m) of the formula (b) for all combinations (j, m) of ym in association with the combinations (j, m), 11. The multiplex according to claim 10, wherein the distortion calculation means reads the value of F (j, m) corresponding to the selected representative vector set (j, m) from the table memory means and performs the calculation of equation (a). Vector quantizer. 12. The weight vector w, the input vector u, and the i-th element of the representative vector xj, ym, which are input, are w (i), u (i), x (i, j) and y (i), respectively. , m), the distortion calculation means calculates the distortion as follows for the representative vector xj, ym of the selected set (j, m). 10. A means for calculating a strain d' _{j, m} defined by F (i, j, m) = {x (i, j) + y (i, m)} ² (2).
The described multi-vector quantizer. 13. The distortion calculating means precalculates at least part of the calculation of a constant vector F (i, j, m) represented by the equation (b) for the selected set of representative vectors xj, ym. 13. The multi-vector quantizer according to claim 12, further comprising a table memory for storing the selected value. 14. The multiplex vector quantizer according to claim 1, wherein said multiplexing means is means for frequency-multiplexing each of said codes representing said representative vector. 15. A multi-vector quantization method in which an input signal sequence is grouped for each of a plurality of samples into one input vector, and each of the input vectors is quantized, in which (a) a plurality of predetermined representative vectors are defined. Selecting one representative vector from each of the plurality of systems of codebooks, (b) calculating the distortion between the average vector and the input vector of the selected set of representative vectors, (c) Repeating steps (a) and (b) for a plurality of different sets of representative vectors selected from the plurality of codebooks, and determining a set of representative vectors that gives the minimum distortion among the calculated distortions; ) Multiplexing each code representing the determined representative vector of the set and outputting the multiplex vector quantization method. 16. The multiple vector quantization method according to claim 15, wherein a squared distance between the input vector and the average vector is calculated as the distortion in the step (b). 17. The multi-vector quantization method according to claim 15, wherein a weighted square distance between the input vector and the average vector is calculated as the distortion in the step (b). 18. Each of the input vectors u is composed of k elements, and the plurality of codebooks are composed of N representative vectors xj, j = 0, 1, ..., N-1, each consisting of k elements. With X
A codebook and a Y codebook having N representative vectors ym, m = 0,1, ..., N-1 each consisting of k elements, where k is a positive integer and N is 2 or more. 16. The multi-vector quantization method according to claim 15, which is an integer. 19. The i-th element of the input vector u and the representative vector xj, ym, respectively, is set to u (i), x (i, j).
And y (i, m), in the step (b), the distortion is expressed by the following equation with respect to the representative vector xj, ym of the selected set (j, m): 19. The multi-vector quantization method according to claim 18, which is calculated as a distortion d' _{j, m} defined by 20. In the step (b), a constant F (j, m) defined in the equation (b) is preliminarily calculated for all pairs (j, m), and a formula (i) is used. 20. The multi-vector quantization method according to claim 19 _{, wherein the} distortion d' _{j, m} defined by 21. The step (b) calculates an inner product of the input vector u and each of the representative vectors xj, and ym, and calculates the sum of the inner products and the constant F (j, 21. The multi-vector quantization method according to claim 20 _, wherein a difference from m) is calculated as the distortion d' _{j, m} . 22. A weight vector w to be input is obtained by calculating w (i), u (i), x (i, j) and y (i) of the i-th element of the input vector u and the representative vector xj, ym, respectively. i, m), the above step (b) expresses the distortion as follows with respect to the representative vector xj, ym of the selected set (j, m). 19. The multi-vector quantum according to claim 18, which calculates a distortion d ′ _{j, m} defined by F (i, j, m) = {x (i, j) + y (i, m)} ² (b). Method. 23. In the step (b), at least a part of calculation of a constant vector F (i, j, m) defined by the equation (b) is previously performed for the selected representative vector xj, ym set. 23. The multi-vector quantization method according to claim 22 _{, wherein the} distortion d' _{j, m} defined by the equation (a) is calculated using a table of calculated results. 24. In the calculation of the equation (a) in the step (b), the step of multiplying the input vector u by the weight vector w, the multiplication result and the representative vector xj,
23. The method of claim 22, further comprising the step of calculating the inner product of each of the vector and ym. 25. The weight vector w, the input vector u, and the i-th element of the representative vector xj, ym, which are input, are w (i), u (i), x (i, j) and y (i), respectively. , m), the step (b) represents the distortion for the representative vector xj, ym of the selected set (j, m) as 19. The multi-vector quantization method according to claim 18, wherein τ is an arbitrary positive constant calculated as a distortion d _{j, m} defined by 26. The step (b) comprises the selected set (j, m).
Representative vectors xj, ym, to the distortion of the following equation _{d j, m = μ | u} -v j, m | 2 + (1-μ) {| u-xj | 2 + | u-ym | 2} 19. The multi-vector quantization method according to claim 18, wherein a distortion d _{j, m} defined by / 4 v _{j, m} = (x j + y m) / 2 is calculated, and μ is an arbitrary constant of 0 ≦ μ ≦ 1. . 27. The step (b) comprises the selected set (j, m).
For the representative vector xj, ym of The strain d _{j, m} defined by is calculated and q (f | j) and q (g |
19. The multiple vector quantization method according to claim 18, wherein m) represents the probability that the code j, m representing the representative vector xj, ym will be mistaken as f, g on the transmission line. 28. The step of creating a codebook of the code includes: (a) a step of arbitrarily distributing a predetermined number of input sample vectors to a plurality of groups to create a plurality of initial codebooks; With a large number of learning vectors as input vectors, determining a set of the representative vectors to which each learning vector belongs by sequentially quantizing with the multiple vector quantization method using the plurality of initial codebooks, (C) The contents of initial codebooks other than any one of the plurality of initial codebooks are fixed, and one arbitrary representative vector of the one initial codebook is regarded as a variable vector, and
The expression obtained by partially differentiating the sum of distortions of all the learning vectors including one representative vector in the set of belonging vectors with the variable vector is set to 0, and the vector value obtained by solving the equation for the variable vector is a new representative vector. And (d) repeating the step (c) for all other representative vectors of the one initial codebook and replacing them with a new representative vector. The step of updating the contents of the codebook, and (e) the steps (b), (c) and (d) are repeated for all other initial codebooks of the plurality of initial codebooks to update their contents. And a step of
15. The multi-vector quantization method described in 15. 29. The multiple vector quantization method according to claim 28, wherein the step of creating the plurality of codebooks includes the step of repeating the steps (b), (c) and (d) a plurality of times for the same codebook. 30. The multi-vector quantization method according to claim 28, wherein the step of creating the plurality of codebooks includes a step of repeating the steps (b), (c), (d) and (e) a plurality of times. 31. The multi-vector quantization method according to claim 15, wherein said step (d) is a step of time-multiplexing and outputting each of said codes representing said representative vector of said set. 32. The multi-vector quantization method according to claim 15, wherein said step (d) is a step of frequency-multiplexing and outputting each of said codes representing said representative vector of said set. 33. The multiple vector quantization method according to claim 31, wherein the code representing each representative vector of the determined set is a number representing each representative vector. 34. The average vector is defined by the following equation. Weighted average vector v _{j, m} v _{j, m} = {2 ^{2r (x)} x j + 2 ^{2r (y)} ym} / {2 ^{2r (x)} +
^22. The multi-vector quantization method according to claim 18, wherein 2 ^{2r (y)} }, where r (x) and r (y) are the transmission rates of the codes representing the representative vectors xj, ym, respectively. 35. If the i-th element of the input vector u is u
(I) and the g-th element of the representative vector xj, ym are respectively represented as x (g, j), y (g, m), the step (b)
Is the following equation for the representative vector xj, ym of the selected pair (j, m) In calculates the distortion d _{j, m} which is defined, h (i, g) is the i-th row of the matrix for performing a product convolution represents the elements of g columns, tau is the distortion d
19. The multiple vector embedding method according to claim 18, wherein the parameters are parameters that minimize _{j and m} .