JPH07273660A - Gain shape vector quantization device - Google Patents

Abstract

PURPOSE:To provide a gain shape vector quantization device by which a shape vector and a gain vector are retrieved without decreasing the coding efficiency with less calculation quantity. CONSTITUTION:The device has shape vector code books 101, 102 having plural shape vectors, a multiplier 103 and an error evaluation section 105 to retrieve a shape vector in which an error between an object vector X and an object vector Y1 obtained by multiplying an optimum gain g1opt with the shape vector obtained from the shape vector code book 101 is minimized, and multipliers 106, 107, a subtractor 108 and an error evaluation section 109 generating a difference between the object vector X and a vector obtained by multiplying a tentative gain g1k with the shape vector retrieved by the shape vector code book 101 as an object vector E1 to retrieve the shape vector minimizing the error between the object vector E1 and an object vector obtained by multiplying an optimum gain g2opt with the shape vector obtained from the shape vector code book 102.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a gain shape vector quantizer for digitizing and encoding voice and image data.

[0002]

2. Description of the Related Art As a highly efficient coding method for audio signals and image signals, vector quantization is generally known in which an input signal is vectorized and quantized in vector units. Also,
As one form of the vector quantization method, there is a multistage gain shape vector quantization method, which is widely used. Gain shape vector quantization is a technique of vectorizing and quantizing the shape and size (gain) of signals such as audio signals and image signals. The two-stage (M = 2) gain shape vector quantization will be described below with reference to FIGS. 6 and 7.

In the following description, uppercase alphabets other than I to N represent vectors, lowercase alphabets and I to N represent scalars, respectively. Also,‖
‖ ² represents the Euclidean distance vector, <,> 2
It shall represent the dot product between vectors. Furthermore, the subscript in parentheses represents the index of the codebook, and the variable with * symbol represents the index of the optimal vector.
In addition, min and max represent a function that outputs the index when taking the minimum value and the maximum value, respectively.

In the encoder shown in FIG. 6, the shape vector C1 (i) obtained from the shape vector codebook 601 and the shape vector codebook 6 as shown in the following equation (1).
02 and the gain vector G (k) obtained from the gain vector codebook 603 =
A linear combination of {g1 (k), g2 (k)} with the corresponding gain is obtained as a candidate vector Y using the multipliers 603 and 604 and the adder 606.

Y = g1 (k) · C1 (i) + g2 (k) · C2 (j) (1) where 1 ≦ i ≦ I (I: number of shape vectors included in the shape vector codebook 101) 1 ≦ j ≦ J (J: number of shape vectors included in shape vector codebook 102) 1 ≦ k ≦ K (K: number of gain vectors included in gain vector codebook)

Next, in order to obtain the similarity between the target vector X input from the input terminal 607 and the candidate vector Y, the subtractor 608 subtracts the candidate vector Y from the target vector X to obtain the residual vector E, and The error evaluation unit 609 obtains the error evaluation value d of the residual vector E. Hereinafter, the Euclidean distance will be used as the error evaluation value d. In this case, the error evaluation value d is expressed by the following equation.

D = ‖E‖ ² = ‖X−Y‖ ² (2) Then, C1 (i *), at which this error evaluation value d becomes minimum,
C2 (j *) and G (k *) are searched for, and the index information i *, j *, k * is transmitted to the decoder shown in FIG.

In the decoder shown in FIG. 7, shape vector codebooks 701 and 702 similar to those in the encoder shown in FIG.
And gain vector codebook 705, index information i *, j *, k * transmitted from the encoder to C1.
(I *), C2 (j *) and G (k *) are decoded, and the output vector Z is obtained and output according to the equation (1) by the multipliers 703 and 704 and the adder 706.

As is clear from the above description, the final purpose of the gain shape vector quantization is to obtain the shape vector C1 (i *), which minimizes the evaluation value d of the error between the target vector X and the candidate vector Y, C2 (j *) and gain vector G
It is nothing but finding a combination of (k *). The most direct and sure way to achieve this goal is
A possible method is to find a candidate vector Y that can be taken from all the combinations of the shape vector and the gain vector, and search for a combination having the smallest error evaluation value d. However, such an exhaustive search method is not realistic because the amount of calculation becomes huge.

On the other hand, a method is also known in which the gains given to the M shape vectors are individually quantized, that is, scalar quantization is performed to search the M shape vectors with a realistic amount of calculation. This method will be described with reference to FIG.
First, in the search in the first stage, the candidate vector Y1 obtained by multiplying the shape vector C1 (i) obtained from the shape vector codebook 801 by the optimum gain g1opt in the multiplier 803,
The residual signal F1 with the target vector X input from the input terminal 804 is obtained by the subtractor 805, and the error evaluation value d1 for this residual signal F1 is calculated by the error evaluation unit 806. At this time, the shape vector C1 that minimizes the error evaluation value d1
Determine (i *). Next, the optimum gain g of C1 (i *)
Scalar quantization is performed for 1 opt using the gain table 807, the adder 808, and the error evaluation unit 809,
Determine g1 (k *).

Next, in the second stage, the subtractor 811 obtains the target vector E1 of the second stage by E1 = X−g1 (k *) · C1 (i *) (3), and this target vector E1 is used. A shape vector codebook 802, a multiplier 810, and
Subtractor 812, error evaluation unit 813, gain table 81
4, the shape vector C2 (j *) and the gain value g2 (n *) are determined using the subtractor 815 and the error evaluation unit 816. In this way, the shape vectors C1 (i *), C2
(J *) and the gain values g1 (k *), g2 (n *) can be determined.

By using this method, the shape vector C1 (i *) in the first stage and the shape vector C2 (j *) in the second stage can be searched independently, so that the amount of calculation required for the vector search is greatly reduced. can do. However, since the gain is scalar-quantized, the coding efficiency is reduced.

As yet another method, a method is known in which orthogonal gain processing is applied to simultaneously quantize the gains respectively given to the M shape vectors, that is, to perform the vector quantization of the gains. This method will be described with reference to FIGS. 9 and 10.

First, the shape vector quantizer is shown in FIG.
Will be explained. In the shape vector quantizer,
First, the shape vector C1 (i) obtained from the first-stage shape vector codebook 901 is added to the optimum gain g by the multiplier 903.
The error evaluation value d1 between the candidate vector Y1 obtained by multiplying by 1 opt and the target vector X input from the input terminal 904 is calculated using the error evaluation unit 403 via the subtractor 905. At this time, the shape vector C1 (i *) that minimizes the error evaluation value d1 is determined.

In the search for the shape vector C2 (j) from the shape vector codebook 902 in the second stage, the orthogonalization processing unit 908 performs orthogonalization processing so that the inner product value with C1 (i *) becomes 0. To do. The orthogonalization process is realized according to the following equation.

C2 (j) = C2 (j) − <C1 (i *), C2 (j)> · C1 (i *) / ‖C1 (i *) ‖ ² (4) where C2 (j) is This is the vector after the orthogonalization process.

The second-stage shape vector codebook 90
For 2, the shape vector search from the shape vector codebook 902 is performed using the orthogonalization processing unit 907, the multiplier 908, the subtractor 909, and the error evaluation unit 910 in the same manner as in the first stage, and the error evaluation value becomes the minimum. C2 (j *) is determined.

Here, the shape vector C1 (i *) searched from the first-stage shape vector codebook 901 by the above orthogonalization process and the vector C2 (j after the orthogonalization process).
*) Is uncorrelated with. That is, C2 (j *) has the effect of the quantization of the shape vector C1 (i *) removed, so that the target vector X is subtracted by multiplying this by the optimum gain g2opt, and the shape of the second stage It is possible to accurately evaluate the error of the shape vector searched from the vector codebook 902.

Next, the shape vector C2 (j *) corresponding to the shape vector C1 (i *) and the orthogonalized vector C2 (j *) is fixed, and the gain vector G (k *).
To decide.

FIG. 10 shows a block diagram of the gain vector quantizer. In this gain vector quantizer, the shape vector C1 (i *), C2 (j *) is multiplied by the multiplier 100.
1, 1002 multiplies each element of the gain vector G (k) obtained from the gain vector codebook 1007, adds them by the adder 1003, and subtracts the error from the target vector X input from the input terminal 1004. Determined by 1005,
This error is evaluated by the error evaluation unit 1006, and G (k *) is determined so that the error value becomes the minimum. That is, the gain vector G (k *) is determined by searching for a vector that minimizes the error evaluation formula given by the following formula.

[0021] k * = min‖X-g1 (k ) · C1 (i *) -g2 (k) · C2 (j *) ‖ ² (5) gain vector quantization using orthogonalization process thus According to the method of performing, the coding efficiency is improved as compared with the method of performing the scalar quantization of the gain. On the other hand, this method requires orthogonalization processing, and for that purpose, the calculation represented by Expression (4) must be performed for all the shape vectors obtained from the second-stage shape vector codebook 902. There is a problem that the amount of calculation is significantly increased. This problem is especially 2
This becomes remarkable when the number of shape vectors stored in the shape vector codebook 902 of the tier is large.

[0022]

As described above, in the conventional gain shape vector quantizer, the method of scalar quantizing gain reduces the coding efficiency, and the method of vector quantizing gain is orthogonalization processing. However, there is a problem that the calculation amount increases because of the above.

An object of the present invention is to perform gain vector quantization with a small amount of calculation without using orthogonalization processing, and to search for a plurality of shape vectors and gain vectors that are close to optimum without reducing the coding efficiency. It is to provide a gain shape vector quantizer capable of performing.

[0024]

In order to solve the above problems, according to the present invention, a search for an m-th stage (2≤m≤M) shape vector is performed by using a tentative gain, so that M shape vectors are searched. After deciding, the above is fixed and the gain vector is searched to realize the vector quantization of the gain.

That is, the gain shape vector quantizer according to the present invention has M shape vectors each having a plurality of shape vectors.
A shape vector codebook of stages (M ≧ 2), an M-dimensional gain vector codebook having a plurality of gain vectors, and a shape vector obtained from the shape vector codebook of the first stage are multiplied by optimum gains. A first shape vector searching means for searching the shape vector codebook at the first stage for a shape vector having a minimum error between the first candidate vector and the first target vector, and the m-1th target vector (2 ≤m≤M) and m-
The shape vector searched from the first-stage shape vector codebook is used to generate the m-th target vector using the vector obtained by multiplying the temporary gain, and the shape vector obtained from the m-th shape vector codebook is generated. Second shape vector searching means for searching the shape vector codebook of the m-th stage for a shape vector having a minimum error between the second candidate vector obtained by multiplying by the optimum gain and the m-th target vector; A linear sum of the M shape vectors searched from the M-stage shape vector codebook by the first and second shape vector search means and the corresponding gains of the gain vectors obtained from the gain vector codebook is calculated. In addition to generating M third candidate vectors, the gain vector that minimizes the error between the third candidate vector and the first target vector is used as the gain vector. And having a gain vector search means for searching the codebook

[0026]

As described above, in the present invention, when the shape vector is searched from the shape vector codebook at the m-th stage, m-
A vector obtained by multiplying the shape vector searched from the first-stage shape vector codebook by a tentative gain, that is, a gain that is considered to be close to the optimum gain, and the shape from the m-1th-stage shape vector codebook. The m-1th used when searching for a vector
The target vector is generated by using the target vector of and the target vector of. Then, after determining M shape vectors, these are fixed and the gain vector is searched from the gain vector codebook.

That is, the vector obtained by multiplying the shape vector searched from the shape vector codebook of the m-1th stage by the tentative gain represents the influence of the quantization of the shape vector. Using the m-th target vector used when searching the shape vector from the (m-1) th shape vector codebook, the m-th target required to search for the shape vector from the m-th shape vector codebook A vector is obtained.

Therefore, since the orthogonalization process is not required, the calculation amount is small, and the M optimal shape vectors and the gains are close to the optimum without the deterioration of the coding efficiency like the method of scalar quantizing the gain. Vector search is possible.

[0029]

Embodiments of the present invention will be described below with reference to the drawings. 1 and 2 are block diagrams showing the configuration of a shape vector quantization device according to a first embodiment of the present invention. FIG. 1 is a configuration of a shape vector quantization unit, and FIG. 2 is a configuration of a gain vector quantization unit. Are shown respectively. Further, in the present embodiment, in order to simplify the description, it is assumed that the shape vector codebook has a two-stage configuration (M = 2).

Since the shape vector quantizer shown in FIG. 1 generates the candidate vector by multiplying the shape vector codebooks 101 and 102 of two stages and the shape vector obtained from the shape vector codebooks 101 and 102 by the optimum gain. Multiplier 10
3 and 106, and error evaluation units 105 and 109 for evaluating the error between the generated candidate vector and the target vector.
And a multiplier 107 and a subtractor 108 for generating a target vector (first target vector) required for the shape vector search from the second-stage shape vector codebook 102.

First, the first stage shape vector codebook 101.
A method of searching for the optimum shape vector C1 (i *) from (3) will be described. A candidate vector Y1 obtained by extracting the i-th shape vector C1 (i) from the shape vector codebook 101, multiplying each element of this shape vector by the optimum gain g1opt in the multiplier 103, and the target input from the input terminal 104 An error evaluation unit 105 is used to obtain an evaluation value d1 of an error from the vector X (second target vector). If the Euclidean distance is used as the error evaluation value d1,
d1 is represented by the following equation.

D1 = ‖X−Y1‖ ² = ‖X−g1opt · C1 (i) ‖ ² (6) In order to determine g1opt so as to minimize d1, d
It is possible to obtain g1opt by partially differentiating 1 with g1opt and setting it to 0 and solving this partial differential equation. The error evaluation value d1 at that time is represented by the following equation.

G1opt = <X, C1 (i)> / ‖C1 (i) ‖ ² (7) d1 = ‖X‖ ² − <X, C1 (i)> / ‖C1 (i) ‖ ² (8) Since the first term on the right side of (Equation 8) is fixed for X, the error evaluation value d1 is minimized when the second term is maximized. That is, the optimum shape vector C1 (i *) is a shape vector given by i * which is i * = max (<X, C1 (i)> ² / ‖C1 (i) ‖ ² ) (9). Therefore, from the equation (9), the shape vector C1 (i *) to be searched can be determined from the shape vector codebook 101 in the first stage.

Next, the second stage shape vector codebook 102.
In order to determine the shape vector to be searched from, the second target vector E1 is determined. Second-stage target vector E1
Is obtained by subtracting C1 (i *) searched from the shape vector codebook 101 at the first stage by the temporary gain g1k at the multiplier 107 from the target vector X at the first stage. To be That is, E1 = X−g1k · C1 (i *) (10).

Here, the tentative gain g1k is a gain close to the optimum gain finally quantized, and in this example, the optimum gain g1opt is used. Since the optimum gain g1opt accurately represents the contribution of C1 (i *) to the target vector X, it is appropriate to use g1opt as the temporary gain g1k. Then, the expression (10) becomes E1 = X-g1opt · C1 (i *) (11).

The shape vector to be searched from the second-stage shape vector codebook 102 is 1 with E1 as the target vector.
It can be determined by performing a search similar to the step.
That is, using the shape vector C2 (j) and the target vector E1 obtained from the shape vector codebook 102, j * = max (<E1, C2 (j)> in the error evaluation unit 109 in the same manner as in Expression (9). ² / ‖C2 (j) ‖ ² ) Find j * that satisfies (12). The shape vector C2 (j *) obtained from this j * is the optimum shape vector in the second stage.

On the other hand, the gain vector quantizer shown in FIG. 2 is obtained from gain vector codebook 201 and the shape vectors and gain vector codebook 201 obtained from shape vector codebooks 101 and 102 shown in FIG. Multipliers 202 and 203 and an adder 204 for obtaining a linear sum of the gain vectors and the corresponding gains as M third candidate vectors, and a subtractor 206 for obtaining an error between these candidate vectors and the target vector. And an error evaluation unit 207 that evaluates this error.

Next, the procedure for determining the gain vector by the configuration of FIG. 2 after searching and determining the shape vectors C1 (i *), C2 (j *) in the configuration of FIG. 1 as described above will be described. To do. Shape vectors C1 (i *) and C2 (j) searched and determined by the shape vector quantizer of FIG.
*), Each element g1 (k), g2 (k) of the gain vector G (k) obtained from the gain vector codebook 201
And a vector multiplied by each of the multipliers 202 and 203,
The error with the target vector X input from the input terminal 205 is obtained by the subtractor 206, and the error evaluation value d3 is calculated as the error evaluation unit 2
When calculated as follows in 07, d3 = ‖X-g1 ( k) · C1 (i *) - find k that minimizes g2 (k) · C2 (j *) ‖ ² (13) d3 Therefore, the gain vector is determined. That is, k * = min || X | g1 (k) * C1 (i *)-g2 (k) * C2 (j *) * ² (14) is obtained, and a gain vector obtained from k * is obtained. G (k *) is the optimum gain vector.

In this way, the shape vector C1 (i
A search for *), C2 (j *) and gain vector G (k *) is achieved. The flow of the above processing is summarized in FIG.
become that way. First, in step S1, the optimum shape vector C1 (i *) is searched and determined from the first stage shape vector codebook 101 using the target vector X, and then in step S2, it is determined in step S1. Target vector E using the shape vector C1 (i *) and the tentative gain
1 is generated. Next, in step S3, the target vector E
1 is used to search for and determine the shape vector C2 (j *) from the second-stage shape vector codebook 102, and finally, in step S4, the shape vector C1 determined in step S1 is determined.
(I *) and the shape vector C2 determined in step S3
The gain vector G (k *) is determined based on (j *).

In the above-described embodiment, the provisional gain for generating the target vector E1 used when searching the shape vector from the second-stage shape vector codebook 102 is obtained from the first-stage shape vector codebook 101. Although the case where g1opt, which is the optimum gain obtained at the time of searching for the shape vector, is applied has been described, in addition to this, one or more corresponding gains of the gain vectors included in the gain vector codebook 201 are used to make a provisional change. A method of setting the gain may be used.

As an example thereof, there is a method of using the element value g1 (k) of the gain vector closest to the optimum gain g1opt. According to this method, the gain vector G (k) including the element value selected at this time is highly likely to match the finally selected gain vector G (k *), which enables efficient encoding. The effect of becoming

As another example, a plurality of target vectors of the shape vector of the second stage are generated for each of a plurality of g1 (k) from the one closest to the optimum gain g1opt as a temporary gain, and the plurality of target vectors are generated. A second stage vector search is performed for each target vector, and a gain vector search is also performed. Then, finally, a method of selecting a combination of a shape vector and a gain vector with which the overall error becomes smaller can be considered. By doing so, more efficient encoding becomes possible.

In the above embodiment, the gain shape vector quantization having the shape vector codebook of two stages has been described, but the present invention is also applicable to the structure having the shape vector codebook of the number of stages of three or more. Clearly applicable.

Further, in determining each optimum vector,
It is also possible to realize a method in which a plurality of vectors are left as candidates and the combination having the smallest error evaluation value among the combinations is made the final output. By doing so, it is possible to search for M shape vectors and gain vectors having smaller error evaluation values.

In the present invention, since the shape vector and the gain vector are searched stepwise by using the temporary gain,
The search can be realized with an extremely small amount of calculation as compared with the conventional full search method which searches for all combinations of the shape vector and the gain vector. Further, as compared with another conventional method in which the orthogonalization process is performed to determine the shape vector, the present invention can realize the vector quantization of the gain while reducing the calculation amount by the amount that the orthogonalization process is unnecessary. Therefore, high coding efficiency can be obtained.

Next, an embodiment in which the present invention is applied to speech coding will be described with reference to FIG. The same reference numerals will be given to the portions corresponding to those in FIGS. 1 and 2, and the description will focus on the differences. In the previous embodiment, the candidate vector is represented by the linear sum of the shape vector and the gain of the gain vector. In contrast, the present embodiment is different in that the candidate vector is represented by the linear sum of the gain of the gain vector and the synthesized signal obtained by passing through the weighted synthesis filter obtained by analyzing the shape vector at a certain time interval. .

In FIG. 4, the input vector S input from the input terminal 401 is used to convert the input voice into a short time period (20
(about ms). The input vector S is first analyzed by the LPC analysis section 402 to obtain an LPC coefficient, and the LPC coefficient is further analyzed by the LPC coefficient quantization section 40.
3 is quantized, and the weighting filter 404 and the weighting synthesis filters 406 and 407 are set based on the quantized LPC coefficient. Next, the input vector S is converted into the weight filter 404.
The vector obtained by subtracting the influence of the previous frame by the subtractor 405 after passing the signal is set as the target vector X.

On the other hand, the first-stage shape vector C1 (i) obtained from the shape vector codebook 101 is passed through the weighted synthesis filter 406 to obtain the weighted synthesis vector V1.
(I) is obtained, and the shape vector C2 (j) in the second stage, which is also obtained from the shape vector 102, is passed through the weighted synthesis filter 407 to obtain the weighted synthesis vector V2 (j).
Ask for. Based on these target vectors X and the weighted composite vectors V1 (i) and V2 (j), V1 (i *) is determined by the same method as in the previous embodiment, and then V2 (j *) is determined.
Can be determined. That is, i * = max (<X, V1 (i)> ² / ‖V1 (i) ‖ ² ) (15) j * = max (<E1, V2 (j)> ² / ‖V2 (j) ‖ ² (16) (here, E1 = X−g1k · V1 (i *)) i * and j * are obtained by the error evaluators 105 and 109, respectively, and the corresponding shape vectors C1 (i *) are obtained. , C
2 (j *) will be searched and determined as the optimum vector.

Next, the gain vector search method will be described with reference to FIG. The gain vector is the weighted composite vector V1 (i *), V2 (j *) and the gain vector codebook 2
It can be determined based on the gain vector G (k) obtained from 01. That is, the error evaluation unit 207 calculates the following equation and obtains the gain vector G (k *) obtained from k *.
Is determined as the optimum gain vector.

[0050] k * = min‖X-g1 (k ) · V1 (i *) -g2 (k) · V2 (j *) ‖ ² (17) as shown in the embodiments described above, the voice of the present invention When applied to coding, it is possible to obtain a shape vector and a gain vector in speech coding with a small amount of calculation, and since vector quantization of gain is applied, high coding efficiency can be realized.

In the above-mentioned embodiment, the speech coding having a two-stage structure has been described. However, it is obvious that the present invention can be applied to a structure having three or more stages.
Further, it is also possible to realize a method in which a plurality of vectors are left as candidates at the time of determining each optimum vector and the combination having the smallest error evaluation value among the combinations is made the final output. By doing so, it is possible to search for M shape vectors and gain vectors having smaller error evaluation values.

[0052]

As described above, according to the gain shape vector quantizing device of the present invention, M shape vectors and gain vectors which are close to the optimum are obtained with a small amount of calculation and without sacrificing coding efficiency. It becomes possible to search for.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a shape vector quantization unit of a vector quantization device according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a gain vector quantization unit of the vector quantization device according to the embodiment.

FIG. 3 is a flowchart showing the flow of processing in the embodiment.

FIG. 4 is a block diagram showing a configuration of a shape vector quantization unit of a vector quantization device according to another embodiment of the present invention.

FIG. 5 is a block diagram showing the configuration of a gain vector quantization unit of the vector quantization device according to the embodiment.

FIG. 6 is a block diagram showing a configuration of an encoder using gain shape vector quantization according to a first conventional example.

FIG. 7 is a block diagram showing a configuration of a decoder using gain shape vector quantization according to a first conventional example.

FIG. 8 is a block diagram showing a configuration of a gain shape vector quantization device according to a second conventional example.

FIG. 9 is a block diagram showing the configuration of an encoder using gain shape vector quantization according to a third conventional example.

FIG. 10 is a block diagram showing a configuration of a decoder using gain shape vector quantization according to a third conventional example.

[Explanation of symbols]

101 ... Shape vector codebook 102 ... Shape vector codebook 103 ... Multiplier 104 ... Input terminal 105 ... Error evaluation section 106 ... Multiplier 107 ... Multiplier 108 ... Subtractor 109 ... Error evaluation section 201 ... Gain vector codebook 202 ... Multiplier 203 ... Multiplier 204 ... Adder 205 ... Input terminal 206 ... Subtractor 207 ... Error evaluation section 401 ... Input terminal 402 ... LPC analysis section 403 ... LPC quantization section 404 ... Weighting filter 405 ... Subtractor 406 ... Weighted Synthesis filter 407 ... Weighted synthesis filter

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification number Reference number within the agency FI Technical indication location H04N 1/41 Z 7/30 (72) Inventor Susumu Kamiwa Komukai Toshiba Town, Kawasaki City, Kanagawa Prefecture No. 1 in the Corporate Research & Development Center of Toshiba Corporation (72) Inventor Emperor Amata No. 1 in Komukai Toshiba-cho, Sachi-ku, Kawasaki City, Kanagawa Prefecture In the Corporate Research & Development Center of Toshiba Corporation

Claims (3)

[Claims] 1. From M-stage (M ≧ 2) shape vector codebooks each having a plurality of shape vectors, an M-dimensional gain vector codebook having a plurality of gain vectors, and a first-stage shape vector codebook. The first candidate vector obtained by multiplying the obtained shape vector by the optimum gain and the first candidate vector
Shape vector codebook that searches the shape vector codebook at the first stage for a minimum error from the target vector, and the (m-1) th target vector (2 ≦ m ≦ M). A shape obtained by multiplying the shape vector searched from the shape vector codebook in the (m-1) th stage by a vector obtained by multiplying the temporary gain by a shape vector obtained from the shape vector codebook in the mth stage. The second obtained by multiplying the vector by the optimum gain
Second shape vector searching means for searching the shape vector codebook of the mth stage for a shape vector that minimizes the error between the candidate vector of the above and the mth target vector, and the first and second shape vectors. A linear sum of the M shape vectors searched from the M-stage shape vector codebook by the search means and the corresponding gains of the gain vectors obtained from the gain vector codebook is calculated, and the M third vectors are obtained.
And a gain vector search means for searching the gain vector codebook for a gain vector that minimizes the error between the third candidate vector and the first target vector. Gain shape vector quantizer. 2. The means for generating the m-th target vector used when searching the shape vector of the m-th step uses the optimum gain used when searching the shape vector of the (m-1) -th step as the temporary gain. The vector quantizer according to claim 1, which is characterized in that. 3. The means for generating the m-th target vector used when searching the shape vector of the m-th stage is the optimum gain used when searching the shape vector of the (m-1) -th stage and the gain vector in the gain vector codebook. The gain shape vector quantizer according to claim 1, wherein the temporary gain is set using one or more m-1 element values.