JP3297752B2 - Vector quantization method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vector quantization method in which a plurality of sample values of a speech waveform and a spectrum envelope parameter are combined into a vector, represented by one code, and quantized.

[0002]

2. Description of the Related Art When compressing and encoding an audio signal or image including speech, for example, a speech waveform signal is subjected to vector quantization, and an identification code (index) obtained by the vector quantization is encoded. High-efficiency encoding for output may be performed.

The vector quantization is a mapping (mapping) Q of an input vector x existing in a k-dimensional Euclidean space R ^k to an output vector y . However, one of the output vectors y is selected from a set of N reproduction vectors of Y = { y _1, y ₂ ,..., Y _N }. That is, the output vector y, y = Q (x) where expressed as y ∈Y ······ (1). Here, the set Y is called a codebook,
N (level) code vectors y _1, y ₂ ,.
has y _N. This N is called a codebook size.

For example, an N-level k-dimensional vector quantizer has a subspace of an input space consisting of N regions (cells). The N cells are represented by {R ₁ , R ₂ ,.
Is represented by _N}, for example, as shown in FIG. 5, because the cell R _i is, as a representative vector y _i, can be said that the set of input vectors x such as pick, Ri = Q-1 (yi ) = {X} Rk: Q (x) = yi} (2) Here, 1 ≦ i ≦ N.

[0005] The sum of all the divided cells matches the original k-dimensional Euclidean space R ^k , and these cells have no overlapping parts. Expressing this as an equation,

[0006]

(Equation 1)

[0007] Therefore, the cell division {R _i } corresponding to the output set Y uniquely determines the vector quantizer Q.

The above-mentioned vector quantizer can be divided into a coder (encoding device) C and a decoder (decoding device) De. The coder C maps the input vector x to an index i. Here, the index i is selected from N sets of I = {1, 2,..., N}, and i = C ( x ) where i∈I... (4).

The decoder De maps the index i to a corresponding reproduction vector (output vector) y _i . However, the reproduction vector y _i is the same as the codebook Y
Selected. This can be expressed by the following equation (5). y _i = De (i) where y _i ∈Y (5)

The operation of the above-mentioned vector quantizer is represented by the equations (1), (2),
Expressions (3), (4) and (5) can be expressed as in the following expression (6). y = Q ( x ) = De (i) = De (C ( x )) (6)

[0011] The above index i is expressed in binary 2
The bit rate Bt of the vector quantizer and the resolution b of the vector quantizer, which is a transmission rate, are as follows: Bt = log ₂ N (bit / vector) (7) b = Bt / k (bit / Sample) ... (8)

Next, a description will be given of a distortion measure which is an evaluation scale of an error. The distortion measure d ( x , y ) is a measure indicating the degree of mismatch (error) between the input vector x and the output vector y . The distortion measure d ( x , y ) is

[0013]

(Equation 2)

## EQU1 ## Where x _i and y _i are vectors
It is the i-th element of x and y .

That is, the performance of the above-described vector quantizer is defined by the total average distortion given by: Da = E [d ( x , y )] (10) Here, E is an expected value.

Usually, the above equation (10) is the average of many samples,

[0017]

(Equation 3)

It can be expressed as follows. However, it is {x _n} is the input vector _{sequence, y n = Q (x n} ). M is the number of samples.

Next, the LBG algorithm used for generating the codebook of the vector quantizer will be described. Originally, the distortion measure and the PDF (Prob
Without knowing the ability density function, it was difficult to design a codebook for a specific vector quantizer. However, by using training data, a codebook for a vector quantizer can be designed without knowing the probability density function. For example, dimension k, codebook size N and training data x
When (n) is determined, an optimal codebook can be created from these factors, and the method is an algorithm called LBG. That is, assuming that training data of a certain type represents audio PDF, a codebook of a vector quantizer can be generated by optimizing the training data. .

The features of the LBG algorithm are Nearest-Neighbor Condition (neighboring condition: optimal dividing condition) which is a condition for division and Centroid Condition (centroid condition: representative point condition) which is a condition for determining a representative point. Or the center of gravity condition). In other words, the problem is how to determine the division and the representative point. The optimal division condition is an optimal coding condition, specifically, an optimal coder condition when a decoder is provided.
The representative point condition is an optimal decoder condition when a coder is given.

In the above optimum resolution conditions, the cell R _j when the representative point is given, Rj = {x: d ( x, yj) ≦ d (x, yi) for all i ≠ j, i, j∈ I｝ ... (12) In the above equation (12), the j-th cell R _j
Is the input signal such that the j-th representative point y _i is closest
It is a set of x . That is, when a given input signal is given, a set of inputs x that searches for the nearest representative point determines the space _Rj that constitutes that representative point. this is,
In other words, selecting the code vector closest to the current input from the code book means the operation itself of the vector quantizer and the operation of the coder itself.

As described above, if the decoder is determined, an optimum coder (which gives the minimum distortion) is required. The coder C ( x ) is expressed as follows: C ( x ) = j iff d ( x , y _j ) ≦ d ( x , y _i ) for all i∈I (13) Here, iff means “only when”. This means that the index j is output when the distance between the input signals x and y _j is shorter than any other y _i . That is, it is the optimum coder to search for the nearest representative point and output the index.

[0023] because the representative point conditions, when the space that Ri has been determined, i.e., when the coder is determined, optimal output vector y _i is located in the center of gravity of the space of the i-th cell R _i The center of gravity is used as a representative vector. This y _i is expressed as y _i = cent (R _i ) ∈R _i (14)

[0024] However, the center of gravity of the R _i cent (R _i) is defined as follows. yc = cent (Ri) if E [d (x, y c) | x ∈Ri] ≦ E [d (x, y) | x ∈Ri] and for all y ∈Ri ····· (15) . This equation (15) shows that the above-mentioned y _c is a representative point in the space of R _i because the input signal x and the y
_This means that the expected value of the distortion with _c is minimized. The optimal code vector y _i minimizes the distortion in R _i . Therefore,
When the coder is determined, the optimal decoder outputs the representative point of this space, and can be expressed by the following equation (16).

[0025] _{D e (i) = cent (} R i) ····· (16) typically has an average value of the input vector x (the weighted average value or simple average) as the representative point.

When the optimum division condition and the representative point condition, which are conditions for determining the division and the representative point, are determined, the LBG algorithm is executed according to the flowchart shown in FIG.

First, in step S21, initialization is performed. Specifically, the distortion D _{−1 is set} to infinity, and the number of repetitions n is set to “0” (n = 0). Also, Y
₀ the initial codebook, threshold epsilon, be defined as the maximum number of repeated n _m.

In step S22, the above-mentioned step S21
, The initial codebook Y ₀ is given, and all the training data given here are encoded under the above-mentioned optimal division condition. That is, mapping is performed on the initial codebook.

In step S23, a distortion calculation for calculating the sum of squares of the distance between the input data and the output data is performed.

In step S24, it is determined whether or not the reduction rate of the distortion obtained from the previous distortion D _n-1 and the current distortion D _n obtained in step S23 is smaller than a threshold value ε, or the number of repetitions n It is determined whether or not a predetermined maximum number of repetitions _nm has been reached. If YES is determined, the execution of the LBG algorithm ends, and N
If O is determined, the process proceeds to the next step S25.

In step S25, when an inappropriate initial codebook is set in step S21, there is a code vector to which no input data is mapped. Usually, a code vector to which no input data is mapped is moved near the cell having the largest distortion.

In step S26, a new center of gravity is obtained by calculation. Specifically, the average value of the training data existing in a given cell is calculated, this is used as a new code vector, and the code vector is updated.

When the process proceeds to step S27, the process proceeds to step S2.
2, the flow is repeated until YES is determined in step S24.

By executing this flow, it can be understood that the LBG algorithm converges in a direction to reduce the distortion between the input vector and the output vector, and terminates the operation at a certain point.

[0035]

The conventional LB
In the vector quantizer trained by the G algorithm, there is no relation between the Euclidean distance of the code vector and the Hamming distance of the index. Therefore, there is a possibility that a completely unrelated code vector may be selected due to a code error in the transmission path.

On the other hand, a method of designing vector quantization in consideration of a transmission line code error has been proposed, but has a disadvantage that characteristics when no error occurs are deteriorated.

Accordingly, the present invention has been made in view of the above circumstances, and has as its object to provide a vector quantization method which does not cause deterioration of characteristics at the time of no error and has a certain strength against a transmission line error. And

[0038]

According to the present invention, there is provided a pitch extracting apparatus, comprising: a plurality of codes each having a magnitude relationship between a distance between code vectors in a codebook and a Hamming distance when a code vector index is expressed in binary. In a vector quantization method of searching for an optimum code vector for an input M-dimensional vector using a codebook including M pieces of M-dimensional code vectors and outputting an index of the searched code vector, The above problem is solved by using, as a distance, a distance obtained by weighting with a weight matrix used for defining a distortion measure.

In the vector quantization method according to the present invention, an optimum code vector is searched for an input M-dimensional vector by using a code book including a plurality of M-dimensional code vectors, and the searched code vector is searched for. In a vector quantization method that outputs an index of
Some of the bits of the binary data representing the index are protected by an error correction code, and the magnitude relation between the Hamming distance of the remaining bits and the distance between the code vectors in the codebook is matched. Solve.

Further, according to the vector quantization method of the present invention, an optimum code vector is searched for an input M-dimensional vector by using a code book including a plurality of M-dimensional code vectors, and the searched code vector is obtained. In a vector quantization method that outputs an index of
When protecting some bits of the binary data representing the index with an error correction code and matching the magnitude relation between the Hamming distance of the remaining bits and the distance between the code vectors in the code book, the code vector The above problem is solved by using a distance obtained by weighting with a weight matrix used for defining a distortion measure as a distance between the two.

[0041]

The vector quantization method according to the present invention uses a distance obtained by weighting with a weight matrix used for defining a distortion measure as a distance between code vectors in a code book. The magnitude relation between the code vector index and the Hamming distance when the code vector index is expressed in binary is matched.

In the vector quantization method according to the present invention, some bits of the binary data representing the index of the code vector are protected by an error correction code, and the Hamming distance of the remaining bits and the code vector Match each magnitude relationship with the distance.

In the vector quantization method according to the present invention, some bits of the binary data representing the index of the code vector are protected by an error correction code, and the Hamming distance of the remaining bits and the code vector When matching the magnitude relations with the distance, a distance obtained by weighting with a weight matrix used for defining a distortion measure is used as the distance of the code vector.

[0044]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the vector quantization method according to the present invention will be described below with reference to the drawings.

A feature of the first embodiment of the vector quantization method according to the present invention is that the magnitude relation between the distance between code vectors in the codebook and the Hamming distance when the index is expressed in binary is matched. This is a vector quantization method that is resistant to transmission line errors.

Here, the creation of a general initial codebook which is the basis of the above codebook will be described.

The above-mentioned LBG merely optimizes the center of gravity in the cell by fine adjustment, but does not change the mutual positional relationship. Therefore, the quality of the above codebook based on the initial codebook is
Determined only by the way the initial codebook is created. In the first embodiment, a splitting algorithm is used to create an initial codebook.

To create an initial codebook by the splitting algorithm, first, a representative point of all training data is obtained from an average of all training data. Next, the representative points are slightly shifted to two representative points. At this time, the above-described LBG is performed, and then the two representative points are slightly shifted from each other and divided into four representative points. When this LBG and convergence are repeated many times, the representative point becomes
2, 4, 8,... 2 ⁿ This operation is expressed by the following expression (17).

Y _{(N / 2) + i} = modify ( y _i , L) (17) where 1 ≦ i ≦ N / 2 and L represents the L-th element of y _i .

Therefore, the creation of the initial codebook by this splitting algorithm is based on the code vector Y = ｛ y ₁ ,
y _2, from ··· y N / _2}, the above equation (17), a method of obtaining the N-level initial codebook.

Here, the right side of the above equation (17), modify (
y _i , L) is the L of (y ₁ , y ₂ ,..., y _L , y _k )
Is to be transformed, and (y ₁ , y ₂ ,.
··· y _L + ε ₀ , y _k ). That is, the L-th element of the code vector y _i is shifted by a small amount ε 0 (in other words, + ε 0 is added to the L-th element of the code vector y _i ).
Is added).

Then, using the modified code vector y _L + ε ₀ as a new start, training is performed by the LBG, and division is performed.

In the creation of the initial codebook by the splitting algorithm described above, the Euclidean distance is closer to the end of the division, and the first embodiment of the present invention is established by utilizing this property. . This will be described below with reference to FIG. FIG. 1 shows a state in which one representative point obtained from an average of training data in one cell becomes a representative point (eight) of eight divided cells by repeating LBG and convergence. A is 1
1, B in FIG. 1 shows two, C in FIG. 1 shows four, and D in FIG. 1 shows eight changes and directions of the division.

Here, the representative point of D in FIG.y _Threeas well asy
₇Is the C in FIG.y'_ThreeIs a division. the abovey
_ThreeIs expressed in binary as "11",y _Threeas well asy ₇
Are “011” and “111”. this thing
Is the above equation (17)y _{(N / 2) + i}Wheny _iAnd the index
Only the polarity (1 or 0) of the MSB (most significant digit) of the
It is shown that. Therefore,y _{(N / 2) + i}Wheny _iNo
The distance between code vectors is expected to be quite close.
You. In other words, the more the division progresses, the more the code
The moving distance of the vector is reduced. This is the index
Even if the upper bit of the
That it would n’t be so big
Show. Therefore, the upper bits of the index are incorrectly
The effect is relatively small.

Here, it is more convenient for the subsequent processing to make the upper bits important, so that the MSB and the LSB (least significant digit) of the bit string of the index expressed in binary in the codebook are exchanged. Table 1 shows the eight indices of D in FIG. 1 together with the code vector.
Table 2 shows that B and LSB are interchanged.

[0056]

[Table 1]

[0057]

[Table 2]

In Table 2, for example, the code vector
Ley _ThreeAnd code vectory ₇Is, in decimal notation,
Code vectors corresponding to “6” and “7”y ₀Wheny
_FourCorresponds to “0” and “1”. The code vectory
_ThreeAnd code vectory ₇And code vectory ₀Wheny _Four
Can be seen from each other as can be seen from FIG.
The code vectors that are closest to each other.

Therefore, the difference between the LSB “0” and “1” in the binary representation of the index is the difference between “0” and “1” in the decimal representation, “2” and “3”, and “4”. The difference between “5”, “6” and “7”, for example, “11”
0 "is" even accidentally 111 ", the code vector y ₃ is y
_7, also incorrectly "000""001", the code vector y ₀ is only err to y _4. These code vectors are code vectors that are closest to each other on D in FIG. That is, even if the LSB side of the index is incorrect, the error in the distance of the code vector corresponding to the index is small.

In the binary data of the above index, L
This means that the magnitude relationship between the Hamming distance on the SB side and the distance between the code vectors is matched. Therefore, only by protecting the MSB side of the binary data of the index with the error correction code, for example, the influence of an error on the transmission path can be minimized.

Next, a second embodiment of the vector quantization method according to the present invention will be described. The second embodiment is a method in which the Hamming distance is considered when training a vector quantizer.

First, before entering into the description of the second embodiment, a vector quantizer is matched to a communication path, and the communication system shown in FIG. A vector quantization method will be described. In the communication system shown in FIG. 2, the input vector x input from the input terminal 21 to the vector quantizer 22 is mapped by the mapping unit 22a to output y _i . Then, the number (index) i is transmitted from the encoder 22 b as binary data to the decoder 24 via the communication path 23. The decoder 24 dequantizes the transmitted index and outputs information from an output terminal 25.

The probability that the index i changes to j before an error is added to the index i through the transmission line 23 and the index i is supplied to the decoder 24 is the probability P (j
| I). That is, the probability P (j | i) is the probability that the reception index is j when the transmission index is i. This probability P (j | i) is expressed by a binary symmetric communication channel (a communication channel of binary data) with a bit error rate (bit error rate) e.

[0064]

(Equation 4)

Can be expressed as follows. Here, d _ij is the Hamming distance when the transmission index i and the reception index j are represented in binary, and S is the number of digits (the number of bits) when the transmission index i and the reception index j are represented in binary.

Under the condition that a channel error occurs at the probability P (j | i) shown in the above equation (18), the cell division {Ri}
Given the center of gravity (representative point) y _u is given by

[0067]

(Equation 5)

Is obtained. In this equation (19), | R _i |
Is the number of training vectors belonging to the subspace R _i. The calculation for obtaining a typical representative point is an average obtained by dividing the sum of the training vectors x belonging to a certain subspace by the number thereof. In the above equation (19), the sum of the average of the training vectors x belonging to all the subspaces For P (u |
This means that a weighted average weighted by the error probability i) is output. Therefore, it can be said that the above equation (19) is a weighted average at the center of gravity weighted with the probability that the transmission index i changes to the reception index u.

Further, the code book ｛ y _i : i = 1, 2,
Optimal partition R _u when · · · N} is given,

[0070]

(Equation 6)

Can be expressed by That is, the above equation (20) is
When the encoder outputs the index u,
And the distortion measure d (x
_, y _j) And take the weighted average
The input vector that chooses the index uxSet of
Subspace. At this time,
condition is,

[0072]

(Equation 7)

Can be expressed as follows. As described above, an optimal codebook is generated for the bit error rate e. This is a codebook created in consideration of the bit error rate e, and when no error occurs (no error). Is degraded compared to the conventional vector quantization method.

Therefore, the present inventor has proposed a vector quantization method as a second embodiment of the vector quantization method in which the Hamming distance is taken into consideration when training the vector quantizer and the characteristic does not deteriorate even in the case of no error. Considered.

More specifically, the bit error rate e is set to 0.5, which is a value with no reliability of the communication path. That is,
P (u | i) and P (i | u) are both constants.
Then, the index is in an unstable state in which it does not know where to go to the cell. To avoid this, it is best to output the center point of the cell on the decoder side. This means that y _u converges at one location (the center of gravity of the entire training set) in the above equation (19). On the encoder side, all input vectors x are mapped to the same code vector as shown in equation (21). In other words, the state of the codebook is a state where the energy level is high enough to be deformable in any way.

Here, the bit error rate e is set to 0.5 to 0
If the bit error rate e is finally reduced to 0 by gradually decreasing the structure until the bit error rate e becomes 0, a partial space that covers the entire original training data x is created. That is, in the course of the LBG training, the influence of the Hamming distance of the index of the adjacent cell is affected by the P
(I | j). In particular,
At the representative point represented by the expression (19), the representative point of another cell affects the update of its own representative point while performing weighting according to the Hamming distance. Thus, the process of gradually reducing the error rate e from 0.5 to 0 is as follows.
It corresponds to the process of gradually removing heat and cooling.

FIG. 3 is a flow chart of the second embodiment, that is, the flow of processing of the vector quantization method in which the Hamming distance is taken into account during training of the vector quantizer and the characteristics do not deteriorate even when there is no error. This will be described using FIG.

First, in step S11, initialization is performed. Specifically, the distortion D _-1 is infinite, the number of repetitions n is "0" (n = 0), and the bit error rate e is 0.49.
And Also, Y ₀ is the initial codebook, ε is the threshold, n
_m is defined as the maximum number of repetitions.

In step S12, the above-mentioned step S11
, The initial codebook Y ₀ is given, and all the training data given here are encoded under the above-mentioned optimal division condition. That is, mapping is performed on the initial codebook.

In step S13, a distortion calculation for calculating the sum of squares of the distance between the input data and the output data is performed.

In step S14, it is determined whether or not the reduction rate of the distortion obtained from the previous distortion D _n-1 and the current distortion D _{n in} step S13 is smaller than the threshold value ε, or whether the number of repetitions n is It is determined whether or not a predetermined maximum number of repetitions _nm has been reached. If YES is determined, the process proceeds to step S15. If NO is determined, the process proceeds to step S16.

In step S15, it is determined whether or not the bit error rate e has become 0. If YES is determined, the flow ends. If NO is determined, the flow proceeds to step S19.

In step S16, when an inappropriate initial codebook is set in step S11, there is a code vector to which no input data is mapped. Usually, a code vector to which no input data is mapped is moved near the cell having the largest distortion.

In step S17, a new center of gravity is obtained by calculation based on the above equation (19).

When the process proceeds to step S18, the above-described step S1 is performed.
Returning to step 2, this flow is repeated until YES is determined in step S15.

In step S19, α (for example, α = 0.01) is reduced from the bit error rate e for each flow until it is determined in step S15 that the bit error rate e = 0.

By executing this flow, in the second embodiment, a codebook optimized at the error rate e = 0 is finally formed, and the vector quantization characteristic at the time of no error according to the second embodiment is obtained. Will hardly occur.

In the index represented by W bits, for example, upper g bits are protected by error correction and lower W-bits are protected.
When g bits are not subjected to error correction, the lower W-g
Reflecting only the Hamming distance of the bit by the above equation (18),
P (i | j) may be obtained. That is, the same high-order g bits have the same Hamming distance, and P (i | j) = 0 if the high-order g bits differ even by one bit. That is, since the upper g bits are protected by error correction, it is assumed that the upper g bits of the index have no error.

Next, a third embodiment of the vector quantization method according to the present invention will be described. The third embodiment is a method of giving a desired structure to the initial codebook of N points.
If an initial codebook having a similar relationship between the Hamming distance and the Euclidean distance is created, the structure will not be destroyed even after training with a normal LBG.

The creation of the initial codebook in the third embodiment is to update the representative point every time one sample of training data is input. Usually, as shown in FIG. 4, the training input data x belonging to the cell of m _j
Representative point is updated by the only m _j, m _{j + 1} and
m _{j new} such as m _{j + 2} is,

[0091]

(Equation 8)

Is updated so that

That is, once all the training data x have been searched (scanned), the same scan is performed with a smaller α, and the α is further reduced and finally converged to 0, thereby obtaining the initial codebook. Create

[0094] In this third embodiment, the training input data x, to affect to all the cells in the neighborhood, the m _j not only m j _{+ 1} and m j _{+ 2,} etc. is also x Is reflected. For example, for m _{j + 1} ,

[0095]

(Equation 9)

Is as follows. In equation (23), f
(J + 1, j) is, for example, f (j + 1, j) = P (j
+1 | j), which returns a value proportional to the reciprocal of the Hamming distance between j and j + 1.

When the above equation (23) is rewritten more generally,

[0098]

(Equation 10)

Is as follows. Here, C in the above equation (24)
( X ) returns the index u of the cell with the centroid closest to the input x . The above C ( x ) is

[0100]

[Equation 11]

Is defined as Here, as an example of the function of f, f (j, C ( x )) = P (j | C ( x )) can be used. In the third embodiment, an initial codebook is created by such an updating method, and the above-described LBG is performed.

Therefore, in the third embodiment according to the present invention, if an initial codebook of N points having a similar relationship between the Hamming distance and the Euclidean distance is created, the ordinary L
The structure does not collapse even if trained with BG.

[0103]

The vector quantization method according to the present invention uses a distance obtained by weighting with a weight matrix used for defining a distortion measure as a distance between code vectors in a code book. Since the magnitude relation between the distance and the Hamming distance when the index of the code vector is expressed in binary is matched, the influence of a code error on the transmission path can be suppressed. Further, the vector quantization method according to the present invention protects some bits of the binary data representing the index of the code vector with an error correction code, and calculates the difference between the Hamming distance of the remaining bits and the distance of the code vector in the code book. Since the magnitude relations are matched, the influence of a code error on the transmission path can be suppressed. Further, the vector quantization method according to the present invention,
When some bits of the binary data representing the index of the code vector are protected by an error correction code, and the magnitude relation between the Hamming distance of the remaining bits and the distance of the code vector in the code book is matched, the above code vector is used. Since the distance obtained by weighting using the weight matrix used for defining the distortion measure is used as the distance, the influence of a code error in the transmission path can be suppressed without causing the characteristic degradation at the time of no error.

[Brief description of the drawings]

FIG. 1 is a schematic diagram for explaining a first embodiment of a vector quantization method according to the present invention.

FIG. 2 is a diagram for explaining a communication error of a general communication system used for explaining a second embodiment of the vector quantization method according to the present invention.

FIG. 3 is a flowchart for explaining a second embodiment of the vector quantization method according to the present invention;

FIG. 4 is a schematic diagram for explaining a third embodiment of the vector quantization method according to the present invention.

FIG. 5 is a schematic diagram for explaining a conventional vector quantizer.

FIG. 6 is a flowchart for explaining an LBG algorithm.

Continuation of the front page (56) References JP-A-2-18600 (JP, A) JP-A-59-153346 (JP, A) JP-A-1-205638 (JP, A) JP-A-63-285599 (JP, A) , A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 19/00-19/02 H03M 7/30 H04B 14/04

Claims (3)

(57) [Claims] 1. The distance between code vectors in a code book.
Binary representation of separation and code vector index
Using a codebook consisting of a plurality of M-dimensional code vectors in which the magnitude relation with the Hamming distance at the time is matched.
The optimal code vector for the input M-dimensional vector.
A vector quantization method for searching for a vector and outputting an index of the searched code vector, wherein a distance between the code vectors is defined as a distortion measure.
Weighted by the weight matrix used
A vector quantization method, characterized in that the distance is used . 2. The method according to claim 1, wherein a codebook comprising a plurality of M-dimensional code vectors is used.
In the vector quantization method for searching for an optimal code vector and outputting an index of the searched code vector, some bits of the binary data representing the index are protected by an error correction code, and the Hamming distance of the remaining bits is And a distance between code vectors in the code book. Wherein the distance between the code vectors, wherein which comprises using the distances determined are weighted by weighting matrix used in the definition of distortion measure
Item 4. The vector quantization method according to Item 2 .