JP3319551B2 - Vector quantizer - Google Patents

Abstract

PURPOSE: To provide a vector quantizing device capable of rapidly retrieving an index of a typical vector even when the number of bits of index is large. CONSTITUTION: This device is provided with a noise vector reproducing part generating the typical vector (u) including the product vn×sn (n=0 to N-1, |sn|=1) between respective elements of an N-dimensional basic vector (v) selected from a code book 100 based on a basic vector index Ic and an N-dimensional polarity vector (s) generated based on a polarity information index Ip as the element un by a polarity multiplication part 105.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vector quantization device used for coding speech and images.

[0002]

2. Description of the Related Art Vector quantization is a method in which a block obtained by collecting a plurality of samples obtained by sampling an audio signal or the like is regarded as one point in a multidimensional vector space, and this is collectively encoded. Basically, the target vector is expressed using the vector specified by. Among such vector quantization methods, VSELP (Vector sum) is used as a vector quantization method particularly used for speech encoding at a bit rate of about 8 kbit / sec.
Excited Linear Predicion) system is known. VS
For details on the ELP method, see "VECTOR SU" by Ira A. Gerson and others.
M EXCITED LINEARPREDICTION (VSELP) SPEECH CODING A
T 8 KBPS, "Proc. IEEE Int. Conf. On Acoustics, Spe
ech and Signal Processing, pp.461-464, April 1990
(Reference 1).

One feature of the VSELP method is that when the number of bits of an index, which is an output code of vector quantization, is p, p base vectors stored in advance are used.
The point is that 2 ^p representative vectors are represented by a combination of the sum or difference of these p base vectors.

FIG. 12 schematically shows the principle of reproducing 2 ^p representative vectors from ^p base vectors. In FIG. 12, vm (n) represents the m-th base vector, ui (n) represents the representative vector of the index i, and the coefficient θim by which the base vector is multiplied takes a value of +1 or −1 depending on the values of i and m. . Coefficient θ
The representative vector u is calculated by adding
get i (n). In this way, since the optimal index can be searched using the recursive formula, if the number of bits p of the index is a value of 7 to 9, the amount of calculation required for vector quantization is suppressed to such an extent that real-time processing is possible. be able to.

However, in this conventional vector quantization method, the effect of reducing the amount of calculation required for index search is not yet sufficient, and it is difficult to increase the number of bits p of the index to 10 or more to realize real-time processing. It is.

[0006]

As described above, in the conventional vector quantization apparatus, when the number of bits of an index is large, realization of real-time processing is realized due to an increase in the amount of calculation required for a suitable index search. There is a problem that it becomes difficult. SUMMARY OF THE INVENTION It is an object of the present invention to provide a vector quantization apparatus capable of performing a suitable index search at high speed even when the number of bits of the index is very large.

[0007]

In order to achieve the above object, the present invention searches for an index corresponding to a desired representative vector based on distortion of a representative vector with respect to a target vector, and outputs a vector index for outputting the searched index. In the conversion device, an element sn of an N-dimensional polarity vector having a polarity multiplied by an element of an N-dimensional base vector as an element
A representative vector generating unit that generates a representative vector having a product of the polarity vector and a polarity information index generating unit that generates a polarity information index corresponding to the polarity of each element of the polarity vector. And a polarity vector generating means for generating a polarity vector.

That is, the representative vector generating means calculates the product vn × sn (n = 0 to N−1, |) of the element vn of the N-dimensional base vector and the corresponding element sn of the N-dimensional polarity vector.
sn | = 1) is generated as a representative vector. Here, the polarity information index means preferably sets the polarity of the element sn of the polarity vector specified by the polarity information specifying the polarity of the element sn of the polarity vector generated by the polarity information generation means to sk (k = L (p , N) (k and p are 0 ≦
k ≦ p−1, an integer satisfying 1 ≦ p ≦ N), and by associating sk with the k-th bit value bk of the polarity information, a polarity information index of p bits is generated.

In this case, the function L (p, n)
Is preferably a remainder when n is divided by p, or a maximum integer not exceeding np / N. Further, for the n-th vector element satisfying the condition of k = L (p, n) of the base vector, it is desirable to obtain a partial inner product of the target vector and the representative vector, and set the polarity of this partial inner product to sk.

In the extended vector quantization device of the present invention, a basis vector storage means for storing a plurality of N-dimensional basis vectors and one of the plurality of basis vectors stored in the basis vector storage means are selected. Vector index searching means for searching for a base vector index to perform the search. In this case, the representative vector generating means has an N-dimensional base vector selected by the base vector index searched by the base vector index searching means and an N-dimensional polarity vector having a polarity multiplied by each element of the base vector as an element. Generate a representative vector whose element is the product of the elements.

The base vector index search means sets the number of base vectors stored in the base vector storage means to I, and sets the base vector to vi (i is a base vector index, i = 0 to I-1). Where k = L of the basis vector for each basis vector Vi
(P, n) (k, p is an integer satisfying 0 ≦ k ≦ p−1, 1 ≦ p ≦ N) For the n-th vector element, the partial inner product f of the target vector and the representative vector f
calculated from k Is used to reduce the number of index candidates of the base vector to J (0
<J << I).

[0012]

According to the vector quantization apparatus of the present invention, a polarity information index corresponding to the polarity of each element sn of the N-dimensional polarity vector s is generated, and an N-dimensional polarity vector s is generated based on the polarity information index. Then, the element sn of the N-dimensional polarity vector is multiplied by the element vn of the N-dimensional base vector v to obtain vn × sn (n = 0 to N−1, | sn
By generating a representative vector u having | = 1) as an element, a representative vector u whose shape changes with respect to one base vector v according to the polarity information index can be generated very easily.

In this case, the polarity of the element sn of the polarity vector s is represented by sk (k = L (p, n) (0 ≦ k ≦ p−1, 1 ≦ p
.Ltoreq.N)), and by associating sk with the k-th bit value bk of the polarity information to form a p-bit polarity information index, the shape of the representative vector u that can be generated from one base vector v Can be limited to a maximum of 2 ^p ways.

Further, by changing the value of p of the function L (p, n), the number of possible shape changes of the representative vector u can be easily controlled. Further, the function L (p, n) is expressed as L (p, n) = n mod p or L (p, n) = floor (np / N) (floor (x) represents the largest integer not exceeding x)
By doing so, the number of elements un of the representative vector u generated using the same polarity can be made uniform. Generally, by using the function L as described above, the information of each bit bk representing the polarity information can be effectively reflected on the change in the shape of the representative vector u.

The polarity vector s that can generate a representative vector u that minimizes the shape distortion with respect to the target vector r is k
= L (p, n), calculate the partial inner product fk of the target vector r and the vector v for the n-th vector element, and sk = sign (fk) (sign (x) is a positive or negative polarity value of x) and Can be determined. In this way, sk and bk (k = 0 to p-1) can be determined with a small amount of calculation on the order of the number of dimensions N of the vector regardless of the size of the bit number p of the polarity information index. Accordingly, even if the number of bits p of the polarity information index is set to 10 or more, the optimum p-bit polarity information {b0, b1,.
.., Bp-1} (corresponding to an index of ordinary vector quantization) does not increase. Therefore, the vector quantization apparatus of the present invention uses a code having a severe restriction on the calculation amount for real-time processing. It can be seen that it is particularly suitable for a gasifier.

Further, in the present invention, there are a plurality of basis vectors v, that is, the basis vectors vi (i = 0 to I−
It is also an effective method to extend the configuration represented by 1). That is, the representative vector u is represented by two pieces of information, that is, the basis vector index Ic for selecting one optimal basis vector v from vi and the corresponding optimal polarity information index Ip. in this case,
When the following means is used in the process of selecting the base vector index, the amount of calculation required for index search can be greatly reduced. That is, in order to narrow the number I of candidates of the base vector index to a small number J of the candidates, k = L (p, n) of the base vector vi (k and p are 0 ≦ k ≦ p−1,
Is calculated from the partial inner product fk of the target vector and the representative vector for the n-th vector element satisfying the condition of 1 ≦ p ≦ N). , The number of basis vector index candidates is set to J (0 <
J << I). This is cor (i)
Is the inner product value of the optimal representative vector ui obtained from the base vector vi and the target vector r.

As described above, even if the number of bits p of the index used for vector quantization is set to a very large value of 10 or more, the vector quantization apparatus according to the present invention requires a very small amount of processing for index search. Suitable for real-time processing.

Further, according to the present invention, even if the number I of the basis vectors vi is significantly reduced as compared with the conventional method, more representative vectors can be realized as a codebook, so that the amount of memory required for storing the basis vectors can be reduced. Can be.

Further, according to the present invention, even if the polarity information is altered by a code error in the transmission medium or the storage medium, there is an effect that a representative vector with little deterioration can be reproduced. This is because 1-bit polarity information originally has a structure that only affects the polarity of some elements of the representative vector, and consequently, an error of 1-bit polarity information causes a deterioration in the shape of the representative vector. It does not reach the whole but only part of the shape is deteriorated. For this reason, the vector quantization apparatus according to the present invention has an advantage in that the polarity information is highly resistant to code errors.

[0020]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a basic configuration of a vector reproduction unit (noise vector generation unit) in the vector quantization device according to the present invention. In this figure, a codebook 100 is a noise codebook in this example, and stores a plurality (I) of N-dimensional base vectors v as noise vectors. Let the I N-dimensional basis vectors be vi (i = 0 to I−
Represented by 1). A base vector selection index Ic is input to the terminal 101, and one of the I base vectors vi is selected by the base vector selection switch 102 according to the index Ic.

On the other hand, the polarity vector generator 103 generates an N-dimensional polarity vector s from the polarity information index Ip. Polarity multiplier 105 consisting of N multipliers 104
Is to multiply N elements vn (n = 0 to N-1) of the base vector v selected by the switch 102 with N elements sn of the polarity vector s generated by the polarity vector generation unit 103, respectively. Gives the product of vn × sn as element un
Is generated.

When such a vector reproducing unit is used in a vector quantization apparatus, even when the number of bits of the representative vector index is large, an optimal representative vector index can be searched at a high speed with a small amount of calculation and suitable for real-time processing. It will be. This effect will be described using a specific example.

For example, with 20 bits of information, 2 ²⁰ (= 10
48576) Vector quantization of audio data of 8 kHz sampling is considered using a codebook expressing representative vectors (80 dimensions).

In this case, in the conventional VSELP method, 20
It can represent representative vectors of 2 ²⁰ present a combination of the sum or difference of basis vectors of the present, but in order to search a suitable representative vector from this, it is necessary 2 ¹⁹ times the distortion amount calculation loop. Such a search for the representative vector with a large number of loops requires a huge calculation amount of about 1,000 MIPS.

On the other hand, in one embodiment of the present invention, one base vector element is divided into 20 sections, and the vector element in each section is multiplied by +1 or -1 polarity to obtain a vector sum or a vector. without taking the difference, we only represent representative vectors 2 ^twenty combinations of polarity information. Since the polarity of each section is 1 bit, only one representative vector can be reproduced with 20 bits of information. The polarity of the preferred representative vector is determined by one simple vector dot product calculation (8
(0 sum of products) and 20 polarity determinations, so that the calculation amount is 1/10 MIPS or less.

In another aspect of the [0026] present invention, 2 bits to the selection of four basis vectors, a configuration that 18 bits used for polarity information, can be expressed representative vectors of 2 ²⁰ present.
In this method, the inner product of the vectors is calculated four times, and the search for a suitable basis vector requires four distortion calculation loops.
The calculation amount is about 1 MIPS.

As can be seen from the above example, the present invention can realize a 20-bit codebook even if the number of base vectors is significantly reduced as compared with the conventional method.
This has the effect of saving the amount of memory required for storing the basis vectors.

Further, according to the present invention, even if the polarity information is altered by a code error in the transmission medium or the storage medium, there is an effect that a representative vector with little deterioration can be reproduced. This is because 1-bit polarity information originally has a structure that only affects the polarity of some elements of the representative vector, and consequently, an error of 1-bit polarity information causes a deterioration in the shape of the representative vector. It does not reach the whole but only part of the shape is deteriorated. For this reason, the vector quantization apparatus according to the present invention has an advantage in that the polarity information is highly resistant to code errors.

Next, an embodiment in which the vector quantization apparatus of the present invention is applied to a speech coding apparatus will be described. FIG.
1 is a block diagram illustrating a configuration example of an encoding device when vector quantization according to an embodiment of the present invention is applied to noise component encoding of a drive signal for speech encoding.

In FIG. 1, an input signal (speech signal) is input from a terminal 201 to a synthesis filter encoding unit 202 and a weight filter unit 203. Synthesis filter encoder 2
02 extracts information of a synthesis filter representing the spectrum envelope information of the voice by analyzing the input voice signal, encodes the information, outputs the code to the multiplexing unit 208, and analyzes the input voice signal to extract the weight filter information. Is output to the weight filter unit 203, and the weighted synthesis filter information H
To the pitch component encoding unit 204 and the noise component encoding unit 205
And output to local decoding section 207.

The weight filter unit 203 receives the weight filter information, the input audio signal, and the local decoded signal from the local decoding unit 207, and outputs an N-dimensional reference audio vector x that can be processed in block units.

The pitch component coding unit 204 receives the reference speech vector x, the weighted synthesis filter information H, and the past drive signal from the local decoding unit 207, and performs past code search by an adaptive codebook search which is a known method. , A pitch vector y0 used for reproducing the pitch component of the current block is extracted from the drive signal waveform, and its index is output to the multiplexing unit 208, and the synthesized pitch vector x0 is output.

Next, a description will be given of the noise component encoding unit 205 which is a feature of this embodiment. Noise component coding section 205
Is the noise codebook 100, the reference vector correction unit 21
1, a preliminary selection unit 212, a main selection unit 213, and a noise vector reproduction unit 215.

The reference vector correction unit 211 weights the residual vector obtained by removing the influence of the pitch vector x0 from the reference speech vector x in the time inverse manner with the weighted synthesis filter information H, and outputs a corrected reference vector r. . Preselection section 212 generates reference vector r and noise codebook 1
Using 00, the number of index candidates in the codebook is narrowed down to a small number (to be J). The main selection unit 213 performs a process of narrowing down the J index candidates from the preliminary selection unit 212 with higher accuracy, and finally selecting one index candidate as the index Ic.

The noise vector reproducing section 215 is configured as shown in FIG. 1, and includes a basis vector v from the noise codebook 100 corresponding to the basis vector index Ic from the main selection section 213, and a polarity from the main selection section 213. A representative vector u obtained by element-by-element multiplication using the polarity vector s corresponding to the information index Ip is obtained as a noise vector y1 whose shape has been optimized. Further, the noise vector reproducing unit 215 outputs a synthesized noise vector x1 using the noise vector y1 and the weighted synthesis filter information H from the synthesis filter encoding unit 202.

Hereinafter, each unit of the noise component coding unit 205 will be described in detail. FIG. 3 shows a detailed configuration of preliminary selection section 212 together with noise codebook 100. Here, in order to simplify the notation, the number of dimensions of the vector is N =
6. The number of bits of the polarity information index Ip is p = 2. Further, the above function L is represented by L (p, n) = n mo
An example will be described in which d p (the remainder when n is divided by p) is used. In FIG. 3, a partial inner product calculation unit 301 calculates a partial inner product f of a representative vector vi of an index i extracted from the noise codebook 100 and a corrected reference vector r.
k (k = 0 to p-1) is obtained. The partial inner product f k is k = n mod for vector element positions n = 0 to N−1.
The inner product value is obtained only for elements that satisfy p. Therefore, in the example of p = 2 and N = 6, it can be calculated as f0 = r0 v0 + r2 v2 + r4 v4 f1 = r1 v1 + r3 v3 + r5 v5. The absolute value adding section 302 calculates the sum of absolute values of the partial inner product fk. The comparison unit 303 searches for J noise indices when cor (i) is maximized, and uses this as a preliminary selection output.

The sum of the absolute values of the partial inner products, cor (i), is expressed by u when the polarity vector s is optimally adjusted with respect to the vector vi.
It is equal to the inner product of i and r. Therefore, the preliminary selection of the index for the shape-corrected vector ui is denoted by co.
This can be performed by searching for the maximum value of r (i).

As another configuration example of the preliminary selection unit 212, when the norm of the representative vector vi stored in the noise codebook 100 is not standardized, the configuration shown in FIG. Can be modified. In the figure, the noise codebook 100 stores normalized weighting coefficients wi for each representative vector in addition to the vector elements. In the preliminary selection unit 400, the normalized absolute value addition unit 4
02, the sum of normalized absolute values based on the partial inner product fk and the normalized weighting coefficient wi J is searched for the index at which is maximized, and this is used as a preliminary selection output. As the value of the normalization weight wi, the reciprocal of the norm of the vector vi can be used.

FIG. 5 shows a processing procedure of the preliminary selection using the standardized weight. First, I, J, N, P, a vector r and a codebook are set (step S11), and i = 0
(Step S12). After calculating the partial inner product fk of the vector r and the vector vi with respect to the index i,
The sum of the absolute values of fk is calculated for all k, and the process of multiplying the sum by the normalization weight wi to calculate the normalized absolute value sum cor (i) is performed from i = 0 to I-1 (steps S13 to S1).
6). Next, J indices that maximize the normalized sum of absolute values cor are selected, and these are selected as i_opt (j) (j
= 0 to J-1) (step S17). Finally, J selected indices are output as preliminary selection results (step S18).

Next, the main selection unit 213 will be described.
FIG. 6 is a block diagram illustrating a configuration example of the main selection unit 213. The main selection unit 213 includes a partial inner product calculation unit 501, an absolute value addition unit 502, a comparison unit 503, a polarity multiplication unit 504, and an orthogonalization power calculation unit 505.
I_opt (j)
(J = 0 to J-1), and a representative vector corresponding to the index is sequentially extracted from the noise codebook 100. In the example of the figure, p = 2 and N = 6. The partial inner product calculation unit 501 generates a partial inner product fk (k = 0 to p) between the representative vector corresponding to the index fetched from the noise codebook 100 and the reference vector r.
In addition to determining -1), the polarity information bit bk and the polarity sk are determined. The polarity sk is given by sk = sign (fk) (sig
n (x) can be determined by the positive and negative polar values of x). The k-th bit bk of the polarity information is, for example, bk =
By setting (1−sk) / 2, it is possible to associate with the polarity sk.

FIG. 7 shows the polarity sk from the reference vectors r and v.
And a procedure for obtaining the polarity information bit bk. As shown in the drawing, the partial inner product fk is calculated in step S22 using the vector r and the vector v set in step S21, and the polarity sk and the bit bk are obtained from the partial inner product fk in step S23, thereby obtaining the polarity information. Can be determined.

The polarity multiplying unit 504 uses the polarity sk calculated for the vector v to calculate the vector u whose shape has been optimized by un = vn × sk (n = 0 to N-1, k = n mod p). Generate. For example, when p = 2, N = 6 and L (p, n) = n mod p, u0 = v0s0, u1 = v1s1, u2 = v2
s0, u3 = v3 s1, u4 = v4 s0, u5 = v5 s
It becomes 1.

The orthogonalization power calculator 505 calculates the vector u,
Using the synthesized pitch vector x0 and the weighted synthesis filter information H, the power pow of the orthogonalized u synthesized vector is output.

The absolute value adder 502 calculates the inner product value cor by the method described above. The comparison unit 503 calculates the inner product value cor
Then, an optimal index Ic is selected by a comparison process for each index using the power pow and the power pow, and Ic is output together with the corresponding polarity information index Ip.

FIG. 8 shows a processing procedure of the main selection unit 213. In step S36, the square value of cor corresponding to the index candidate i_opt (j) is obtained as cor2, and in step S39, the comparison between the index candidates i_opt (j) and i_opt (j-1) is performed using cor2 and pow. .

The noise index (base vector index) Ic and the polarity information index Ip thus obtained from the main selection unit 213 are input to the noise vector reproduction unit 215, and the vector v from the noise codebook 100 corresponding to the noise index Ic and the vector v A noise vector u having an optimized shape is obtained by element-wise multiplication using the polarity vector s corresponding to the polarity information index Ip.

FIG. 9 shows a processing procedure of the noise vector reproducing unit 215. First, in step S51, the index Ic
Is set and each bit bk representing the polarity information index Ip is set, and the polarity sk is obtained from bk in step S52. Next, in step S53, the polarity sk
A vector u is obtained by multiplying (k = 0 to p−1) and the vector v for each element. The vector u is set as a noise vector y1, and the noise vector reproducing unit 215 further obtains a synthesized noise vector x1 using the weighted synthesis filter information H and y1, and outputs this.

Returning to FIG. 2, the gain component coding section 206 codes gains used for the pitch component and the noise component, respectively. Specifically, pitch component encoding section 20
4 and the combined noise vector x1 from the noise component coding unit 205 are input, and the distortion of the reference vector x and the vector (g0 x0 + g1 x1) is obtained by searching for a built-in gain codebook. And the index G corresponding to the set of gains (g0, g1) that minimizes.

The local decoding unit 207 includes gains g0, g1,
A drive signal corresponding to the current block is created using the pitch vector y0 and the noise vector y1, and a local decoded signal is created using this and the weighted synthesis filter information H.

The multiplexing unit 208 includes the encoding units 204 and 2
The coding parameter information obtained in steps 05 and 206 is input, multiplexed, and output from an output terminal 209 to a transmission medium or a storage medium.

Next, an embodiment of a speech decoding apparatus for decoding transmission information from the speech encoding apparatus and outputting reproduced speech will be described with reference to FIG. In FIG. 10, coding parameter information input from an input terminal 601 is:
The separation unit 602 separates the information into information used in each decoding unit described below. The pitch component decoding unit 603 incorporates an adaptive codebook that uses a past drive signal in the same manner as the speech encoding device shown in FIG. Play the vector y0.

The noise component decoding unit 604 includes a noise codebook 605 and a noise vector reproducing unit 606.
A vector v corresponding to the decoded noise index Ic
Using the polarity information index Ip and the noise vector reproducing unit 215 of the speech coding apparatus shown in FIG. 2, a noise vector u whose shape is optimized is reproduced, and this is output as a vector y1.

The gain component decoding section 607 incorporates a gain codebook in the same manner as the speech coding apparatus, and reproduces the gains g0 and g1 based on the decoded index G. Next, the driving signals g0 y0 + g1 y1 are reproduced by using the multiplication units 608 and 609 and the addition unit 610. The synthesis filter 611 calculates a decoded audio signal using the decoded synthesis filter information and the drive signal, and outputs the signal. The post filter 612 is used as the last stage processing of the audio decoding device. The post-filter 612 determines the characteristics of the filter based on the transmitted encoding parameter information, and outputs a signal obtained by adjusting the sound quality of the decoded audio signal from the terminal 613 as a reproduced audio signal.

(Second Embodiment) FIG. 11 shows a second embodiment of the present invention.
FIG. 2 is a block diagram of a speech encoding device according to an embodiment of the present invention. The parts having the same functions as those of the speech coding apparatus according to the first embodiment shown in FIG. 2 are denoted by the same reference numerals, and the description thereof will be omitted. The difference will be mainly described. Is that the encoding unit 205 selects a plurality of sets of the index of the noise codebook 100 and the polarity information index, and the gain component encoding unit 206 finally narrows down the set of candidates to one set of Ic and Ip. This is different from the first embodiment.

In FIG. 11, noise component coding section 205
Is the reference vector correction unit 211, the noise codebook 10
0, preliminary selection unit 212 and noise vector reproduction unit 215
Consists of The preliminary selection unit 212 narrows down a large number of index candidates in the codebook to a small number (J) of index candidates using the corrected reference vector r and the noise codebook 100. Then, together with these J index candidates, the corresponding polarity information index candidates are output to the noise vector reproducing unit 215. The noise vector reproducing unit 215 has a function similar to that of the noise vector reproducing unit 215 of the first embodiment, and includes a codebook index and a polarity information index J corresponding thereto.
The combination candidates and the J vector of the noise vector combined with the corresponding noise vector are output to gain component coding section 206. The gain component coding unit 206 calculates a synthesized pitch vector x0 for each of the input J sets of candidates.
, The gain is encoded by the same method as that of the gain component encoding unit 206 of the first embodiment. Finally, a candidate having the smallest distortion with respect to the reference vector x is selected from the J sets, and the noise vector at that time is y1, the codebook index is Ic, the polarity information index is Ip, and the gains g0 and g1 are obtained. To local decoding section 207
, Ic, Ip and the index G of the gain codebook are output to the multiplexing unit 208. Other points are
Since the configuration is the same as that of the first embodiment, the description is omitted.

[0056]

As described above, the vector quantization apparatus according to the present invention requires only a small amount of processing for index search even if the number of index bits is large, enables high-speed index search, and base vectors stored as a codebook. Even if the number of is reduced, many representative vectors can be equivalently realized as a codebook.
This has the effect of reducing the amount of memory required to store base vectors.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a basic configuration of a vector quantization apparatus according to the present invention.

FIG. 2 is a block diagram showing a configuration of a speech coding apparatus according to one embodiment of the present invention.

FIG. 3 is a block diagram showing a configuration example of a preliminary selection unit in the embodiment.

FIG. 4 is a block diagram showing another configuration example of the preliminary selection unit in the embodiment.

FIG. 5 is a flowchart illustrating an example of a processing procedure of a preliminary selection unit according to the embodiment;

FIG. 6 is a block diagram showing a configuration example of a main selection unit in the embodiment.

FIG. 7 is a flowchart illustrating an example of a processing procedure for obtaining polarity information according to the embodiment;

FIG. 8 is a flowchart illustrating an example of a processing procedure of a main selection unit according to the embodiment.

FIG. 9 is a flowchart illustrating an example of a processing procedure of a noise vector reproducing unit in the embodiment.

FIG. 10 is a block diagram showing a configuration of a speech decoding apparatus according to one embodiment of the present invention.

FIG. 11 is a block diagram showing a configuration of a speech coding apparatus according to another embodiment of the present invention.

FIG. 12 is a diagram showing the principle of vector reproduction in a conventional vector quantization device.

[Explanation of symbols]

 Reference Signs List 100: noise codebook 101: base vector index input terminal 102: base vector selection switch 103: polarity vector generation unit 104: multiplier 105: polarity multiplication unit 201: audio signal input terminal 202: synthesis filter coding unit 203: weight filter Unit 204 Pitch component encoding unit 205 Noise component encoding unit 206 Gain component encoding unit 207 Local decoding unit 208 Multiplexing unit 209 Coding parameter output terminal 211 Reference vector correction unit 212 Preliminary selection unit 213: main selection unit 215 ... noise vector reproduction unit 401 ... partial inner product calculation unit 402 ... normalized absolute value addition unit 403 ... comparison unit 501 ... partial inner product calculation unit 502 ... absolute value addition unit 503 ... comparison unit 504 ... polarity multiplication unit 505: orthogonalization power calculation unit 601: input terminal 602: separation unit 03: Pitch component decoding unit 604 ... Noise component decoding unit 605 ... Noise codebook 606 ... Noise vector reproduction unit 607 ... Gain component decoding unit 608,609 ... Multiplication unit 610 ... Addition unit 611 ... Synthesis filter 612 ... Post filter 613 output terminal

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-7-253795 (JP, A) JP-A-7-170192 (JP, A) JP-A-6-209262 (JP, A) JP-A-6-209262 186998 (JP, A) JP-A-6-222796 (JP, A) JP-A-2-502135 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 19/12

Claims (6)

(57) [Claims] 1. A vector quantization apparatus for searching for an index corresponding to a desired representative vector based on distortion of a representative vector with respect to a target vector and outputting the searched index, comprising: an element of an N-dimensional base vector; A representative vector generating means for generating a representative vector whose element is a product of an N-dimensional polar vector element having a polarity multiplied by each element of the element, and a polarity information index corresponding to the polarity of each element of the polarity vector. A vector quantization apparatus, comprising: a polarity information index generating unit that generates a polarity information index; and a polarity vector generating unit that generates the polarity vector based on the polarity information index. 2. A vector quantization apparatus for searching for an index corresponding to a desired representative vector based on an error of a representative vector with respect to a target vector and outputting the searched index, comprising: an element vn of an N-dimensional base vector; Vn × sn (n = 0 to N−
1, | sn | = 1); a representative vector generation unit that generates a representative vector having elements as elements; polarity information generation unit that obtains polarity information that specifies the polarity of the element sn of the polarity vector; Element of the polarity vector
Is sk (k = L (p, n) (k and p are 0 ≦ k ≦ p
−1, an integer satisfying 1 ≦ p ≦ N), and by associating sk with the value k k of the k-th bit of the polarity information, p
A vector quantization device, comprising: a polarity information index generation unit that generates a bit polarity information index; and a polarity vector generation unit that generates the polarity vector based on the polarity information index. 3. The vector quantization according to claim 2, wherein said L (p, n) is a remainder obtained by dividing n by p, or a maximum integer not exceeding np / N. apparatus. 4. A partial inner product of the target vector and the representative vector is obtained for an n-th vector element satisfying the condition of k = L (p, n) of the base vector, and the polarity of the partial inner product is determined by the sk. The vector quantization apparatus according to claim 2, wherein: 5. A vector quantizer for searching for an index corresponding to a desired representative vector based on distortion of a representative vector with respect to a target vector and outputting the searched index, comprising: a base storing a plurality of N-dimensional base vectors; Vector storage means; base vector index search means for searching for a base vector index for selecting one of the plurality of base vectors stored in the base vector storage means; A representative vector generation unit configured to generate a representative vector whose element is a product of the selected N-dimensional base vector element and an N-dimensional polarity vector element whose polarity is a factor by which each element of the base vector is multiplied; Generate polarity information index corresponding to the polarity of each element of the polarity vector A polarity information index generating unit for generating the polarity vector, and a polarity vector generating unit for generating the polarity vector based on the polarity information index. 6. A vector quantization apparatus for searching for an index corresponding to a desired representative vector based on distortion of a representative vector with respect to a target vector and outputting the searched index, comprising: I number of N-dimensional base vectors (I is A plurality of) stored base vector storage means; a base vector index search means for searching for a base vector index for selecting one of the I base vectors stored in the base vector storage means; A representative vector that generates a representative vector having as an element a product of an N-dimensional base vector element selected by the selected base vector index and an N-dimensional polarity vector element having a polarity multiplied by each element of the base vector Generating means, and a polarity information index corresponding to the polarity of each element of the polarity vector And a polarity vector generating means for generating the polarity vector based on the polarity information index. The base vector index searching means sets the base vector to vi (i A vector index,
When i = 0 to I-1), k = L (p, n) (k, p is 0 ≦) of the basis vector vi for each basis vector Vi.
k ≦ p−1, an integer satisfying 1 ≦ p ≦ N) is calculated from the partial inner product fk of the target vector and the representative vector for the n-th vector element. Is used to reduce the number of index candidates of the base vector to J (0
<J <<I> A vector quantization apparatus comprising means for narrowing down to (I).