JPH02287400A - Vector quantization system for predicted residual signal - Google Patents

Abstract

PURPOSE:To decrease the computation quantity searching code vectors by selecting one code book from plural pieces of the code books and quantizing a predicted residual signal by using this code book as well as a synthetic filter and input signal. CONSTITUTION:The code book 700 is constituted of plural pieces of the code books corresponding one to one to plural pieces of the preset synthetic filters. One code book can be selected from plural pieces of the code books by the code book selection information outputted from a code book selecting section 9. The synthetic filters 8 are independently inputted with all of N-pieces of the vectors outputted from the selected code book and form N-pieces of the synthetic vectors associated to the code book by a code number. The residual signal code selecting section 6 selects the code relating to the min. distortion among N-pieces of distortions as a residual signal code. The vector quantization of the predicted residual signal is thus completed. The computation quantity is decreased in this way.

Description

[Detailed Description of the Invention] [Object of the Invention] (Field of Industrial Application) The present invention relates to vector quantization used for information compression and transmission of digital signal sequences such as audio and images.
In particular, the present invention relates to a vector quantization method for a prediction residual signal, which is an encoding method in which a synthesis filter is driven by the prediction residual signal to obtain a reproduced signal.

(Prior Art) Vector quantization is one of the technologies currently attracting the most attention as a means of compressing and encoding information on digital signals. This is a plurality of discretized signal sample values X to be encoded.

X 2 , 0. ', Xn as one vector and X""
(X 1.

X 2 , ..., Xn) vector ""yit'y1 in the codebook created in advance
2"・-1yln) (f-1,-,k) (where k is t near - the number of vectors in Dobuk) represents the discretized signal sample values X, , X2..., xn.

One method to efficiently apply this vector quantization to the encoding of audio and image signals is to vector quantize the prediction residual signal obtained by predicting the input signal, and synthesize it using the quantized signal. An encoding method is known that drives a filter and obtains a reproduced signal.

FIG. 6 is a principle diagram showing an example of the above encoding method,
In the field of audio coding, CELP (Code Excided Line
ar PredIctlon) or V X C (Vec
This is a method known as a tor exclusive code method. For details on the CELP method, please refer to M, R, 5 chroeder and B, S, Ata1.
Mr.'s "C0DE-EXCITED LINEARPRE"
DICTION (CELP) :Hl(iH-GυAL
ITY 5PEECHAT VERY LOW BIT
RATES”, 1nProc, IcAs5P 8
5. pp. 937-939.1984 (Reference 1).

This method is characterized in that the code for the quantized value of the predicted residual signal (residual signal code) is selected by matching the reproduced signal level. In other words, the vectors in the codebook are used as residual signal candidates and are input to the synthesis filter actually used in the decoding section to create a reproduced signal, and the input signal and the reproduced signal are matched to obtain the reproduced signal with the least distortion. The code of the vector in the codebook to be reproduced is used as the residual signal code. Therefore, there is an advantage that vector quantization of the residual signal can be performed with the least distortion between the input signal and the reproduced signal for a given codebook.

As is clear from FIG. 6, the CELP method is configured to perform vector quantization of the residual signal using a fixed codebook for any input signal.

If the prediction parameter extraction unit completely extracts the power spectrum information of the input signal with infinite accuracy and parameters on the infinite side, the prediction residual signal obtained by pre-DIing the input signal will be a perfect white noise signal. Therefore, a white noise codebook consisting of a small number of fixed vectors can sufficiently express the residual signal.

However, in actual encoding devices, the input signal is highly non-stationary, and how efficiently the power spectrum information is extracted by the prediction parameter extractor depends heavily on the shape of the power spectrum of the input signal. , in many cases, the prediction residual signal is not a completely white noise signal, but a colored signal that depends on the input signal spectrum.

Therefore, in order to vector quantize residual signals whose whiteness differs depending on the shape of the power spectrum of the input signal with little distortion, when using a fixed codebook, the total number of vectors (number of codes) in the codebook must be ) M must be made very large to accommodate various types of residual signals. As a result, when performing vector quantization, the optimal vector and its code cannot be determined unless distortion calculation and distortion comparison are performed for M types of vectors, so the amount of calculation required for vector quantization is essentially within the codebook. increases in proportion to the increase in the number of vectors M, and M
If it is very large (for example, 1000 or more), it becomes impractical.

In addition, if the vector M is made smaller using a fixed codebook, the number of elements (number of dimensions) of the vector becomes smaller under the condition that the transmission rate of the residual signal is constant, and the amount of calculation required for vector quantization of the residual signal can be realized. However, it is known that the compression efficiency of vector quantization itself decreases with the number of dimensions, and the quantization error in the reproduced signal increases, resulting in a reduction in the reproduction signal at a low bit rate. There is a problem that the quality of the product deteriorates.

(Problem to be Solved by the Invention) As described above, the conventional predictive residual signal vector quantization method generates a residual signal that minimizes the distortion between the input signal and the reproduced signal for a given codebook. However, since the codebook used is fixed for any input signal, it is possible to efficiently express residual signals with different whiteness depending on the shape of the spectrum of the input signal. Therefore, there is a problem that prediction residual signal vector quantization of sufficient quality cannot be realized at low bit rates.

The present invention has been made in view of the above, and an object thereof is to realize vector quantization of a prediction residual signal with high efficiency and high quality even when applied to a low bit rate encoding device. The purpose is to provide a difference signal vector quantization method.

[Structure of the Invention] (Means for Solving the Problem) In order to achieve the above object, the predictive residual signal vector quantization method of the present invention is provided with a predetermined number of different codebooks in advance, and input Select at least one codebook from the plurality of codebooks based on the frequency characteristics of the synthesis filter obtained from predictive analysis of the signal, and use the selected codebook, the synthesis filter, and the input signal. This method is characterized by quantizing the prediction residual signal.

(Function) In the prediction residual signal vector quantization method of the present invention, prediction parameters representing the shape of the spectrum of the synthesis filter obtained by linear predictive analysis (LPC analysis) of the input signal for each block are used to set the prediction parameters in advance. M (M ≧2
) of the CC spectrum patterns of the synthesis filters, and select one codebook from M codebooks set in advance based on the classification information, Conventional known XC using the selected codebook
The vectors in the codebook are synthesized using the synthesis filter in the same manner as in the method, to create a synthesized vector, and a signal obtained by removing the influence of the past quantized residual signal from the input signal,
Residual signal vector quantization is performed by searching for a code of a vector that has the minimum distortion with respect to the composite vector.

The M codebooks are classified and weighted offline based on the M spectral patterns.
The training sequences are created independently using the LBG algorithm.

For the LBG algorithm, Y, LINDE, A, 8070, R, M, GRA.
'An Al-gorithm f'or V by Y
ector Quantlzer Design,
(IEEE Trans, C0N-28, PP, 84-
95. Jan, 1980).

A method for creating the above M training sequences will be described below.

First, a residual signal is obtained by block-by-block predictive analysis from a sufficiently long signal sequence (original training sequence) whose statistical properties are similar to the input signal to be encoded. Next, this residual signal is divided into vectors by a predetermined number of vector dimensions.

Next, the spectral shape (H) of the original training sequence of the current block obtained by the predictive analysis is classified into one of the M0 spectral patterns H, H, ..., HMe, and the classification information 1 (i- 1,...M), the prediction residual signal for each block is converted into Me residual signal sequences T1 (i-1
, 2, ... 2M0). M obtained at this time
The residual signal sequences T1 become M0 training sequences.

M codebooks B, (i-1, 2,...M)
is created as follows using M training sequences Ti (i-CC 1, 2, . . . , M >).

Based on the classification information i of the spectrum pattern, Tt
Bl from T2, B2 from T2, B from TMc
Create each Me independently. In this case, for example, if the LBG algorithm used to create BI from TI is LBGl, then the squared error distortion calculated by LBG is calculated by applying a weight Hl of a predetermined spectral pattern determined by the corresponding classification information l. is calculated. Also, LBGl
The centroid (average value) vector within each cluster calculated by is also determined so that the distance from the training vector within the cluster is minimized at the level weighted by the weight Hl.

The created M codebook B1 has a weight H
Since it is designed with a total scale in which l is added, the codebook of the residual signal is more suited to the shape of the spectrum than the white noise codebook.

Therefore, it is less than the number of vectors NF required to obtain good quality with a conventional fixed codebook.

Each M has a number of vectors N (N≦NF/Mo). With this codebook, it is possible to perform predictive residual signal vector quantization with quality comparable to or better than conventional methods.

Moreover, by simply selecting M codebooks through a very low-complexity process of classifying the spectral pattern of the synthesis filter found for each block of the input signal, the vector search range required for vector quantization of the residual signal is can be reduced to the conventional N /NP (≦1/Mo),
This has the effect of making it possible to significantly reduce the amount of calculation while maintaining high quality.

(Example) Hereinafter, the present invention will be explained in detail using the drawings.

FIG. 1 is a block diagram showing the principle of the residual signal vector quantization method according to the first embodiment of the present invention.

In the figure, a prediction parameter is extracted from an input signal input from an input terminal in a prediction parameter extraction section 1. The codebook selection unit 9 classifies the synthesis filter by matching the synthesis filter determined from the prediction parameter information extracted by the prediction parameter extraction unit with a plurality of preset synthesis filters, and classifies the synthesis filter. Output codebook selection information based on the information.

The codebook 700 is composed of a plurality of preset synthesis filters and a plurality of codebooks in one-to-one correspondence. The code book is configured such that one code book can be selected from among the food books using the above-mentioned correspondence.

The synthesis filter 8 independently inputs all N vectors output from the selected codebook and creates N synthesized vectors that are related to the codebook by code number.

The matching unit 5 generates a signal x obtained by subtracting the influence of the quantization value of the past residual signal from the input signal x (n> using the same synthesis filter as that used in the synthesis filter 8).
(n), and the N resultant vectors and x
(n) and obtain N distortions independently. The distortion obtained at this time is related to the composite vector by the code number.

The residual signal code selection section 6 selects the code related to the smallest distortion among the N distortions found by the matching section 5 as the residual signal code. This completes the prediction residual signal vector quantization.

2 and 3 are block diagrams showing an example of an encoding device and a decoding device to which the predictive residual signal vector quantization method according to the first embodiment shown in FIG. 1 is applied. .

First, the encoding device shown in FIG. 2 will be explained in detail. In FIG. 2, an input signal inputted from an input section is cut out into a predetermined sequence length by a block partitioning section 11. The prediction parameter extraction unit 12 uses the input signal cut out by the block extraction unit 11 to extract coefficient information (prediction parameters) of a prediction filter that expresses the shape of the spectrum of the input signal.

As a method for extracting prediction parameters, for example, a linear prediction analysis method is used, or an optimal pre/1111 filter is selected from among the prediction filter candidates using a plurality of preset prediction filter candidates and the cut-out input signal. Although there is a method of selecting , other known methods may also be used. The quantization unit 13 quantizes the prediction parameters extracted by the prediction parameter extraction unit 12. The normalized gain calculation unit 16 generates a prediction residual signal by passing the input signal extracted by the block partitioning unit 11 through a prediction filter created based on the prediction parameters quantized by the quantization unit 13. . For example, when the predictive filter is a P-order linear filter A (Z) expressed in two transformation domains,
The prediction residual signal e (n) is obtained, for example, by the following equation.

Here, n represents the discretized time, and S (n) represents the cut out input signal. Also (al) (i-1
, ..., P) are filter coefficients of a prediction filter uniquely determined from the quantized prediction parameters.

Next, one or more normalized gains are calculated from the calculated prediction residual signal. As the normalization gain, for example, there is a method in which the square root of the total average of the squared amplitudes of the prediction residual signal is obtained for a predetermined sample interval as the normalization gain, but any other normalization gain may be used. A calculation method may also be used. Further, if the power variation of the prediction residual signal is small due to the statistical properties of the input signal, the configuration may be such that the prediction residual signal is not normalized.

The quantization unit 17 quantizes the normalized gain of the single unit or coefficient obtained by the normalized gain calculation unit 16.

The codebook selection unit 100 uses the pre-A11 parameters quantized by the quantization unit 13 to create a synthesis filter H, which is an inverse filter of the prediction filter, and a burn number of the synthesis filter, which is set in advance by a predetermined method. The following M (
M ≧2) composite filters H, H2, . . . , HM. The synthesis filter H is classified into one of the most similar ones in the set synthesis filters H, H2,..., HMe, and the Outputs codebook selection information for selecting a codebook designed using the method described in .

As a method for classifying synthesis filters, when the synthesis filter is an all-pole linear filter expressed by the reciprocal of the filter in equation (1), for example, Itak
ura-8afto d4tortion mcas
Known as uro.

d (a,!,) −(a as ) R(a)(” ”1)t(i
-1, 2, ..., M) There is a method using (3) (t represents the transpose of the matrix).

Here, a represents the coefficient of the synthesis filter H determined based on the prediction parameters, and a-11I(at r a21 +・+, a
) (4). Also, the meeting is M mentioned above. represents the filter coefficient of the first synthesis filter among the filters H1゜H, . . . , HMe, and k−(@ , 'a , .
(5) Let it t2. Further, R(2) is a constant PxP symmetric matrix and can be uniquely determined from the prediction parameters.

Itakura-9alto distortron
For details on measure, see Itakura, F.
, 'MaxiIlum Prediction Re
51dualPrinciple App! 1ed t
o 5peech Recognltron.

IEEE Trans, ASSP, 23. PP,
l37-72. Ped, 1975 (Reference 3).

In addition, as a distortion measure for classification of other synthesis filters, square error distortion (i-1, 2, ..., M) (B
) may be used. Or, the above filter coefficients are converted into parameters, tog-area r
atto.

It is also possible to convert it into an equivalent parameter such as an LSP parameter or an impulse response, and use the square error distortion between the parameters at the level of the parameter.

Other distortion measures include a composite signal sequence G obtained by driving the synthesis filter or its inverse filter by a predetermined number of samples with a predetermined signal sequence, and a composite signal sequence G obtained by driving the M synthesis filters or its inverse filter, respectively. squared error distortion d (G, (1; 1)-1G-Gl 12(i-1,2
,...,'M) A method using (7) can also be used.

As a method for creating the Mc synthesis filters, for example, the following method can be considered. First, from a long input signal sequence with statistical properties similar to the input signal to be encoded, prediction parameters are calculated for each block in the same manner as for the input signal sequence to be encoded as described above. Obtain a large number of predictive parameters.

Then, considering each prediction parameter as one vector,
Using the obtained large number of vectors as training vectors, a codebook having predetermined M reproduction vectors is created using, for example, the LBG algorithm (see Reference 2). The code numbers in this codebook correspond to the pattern numbers of the synthesis filters, and the synthesis filters H, H2, etc. are uniquely set based on the reproduction vectors corresponding to the code numbers.
・■ The coefficient of HMe can be determined.

For example, the synthesis filter H is expressed as the reciprocal of the filter in equation (1) above, and the training vector is expressed as a'' (al
! B2 r ..., a ), one reproduction vector LB in the created code book
The distortion scale used in the G algorithm is as described above (
Equation 3) and Equation (B) can be used.

Furthermore, as a simple method, M distortion filters can be created by scalar quantizing the prediction parameters or equivalent parameters obtained by converting the prediction parameters, without quantizing them all at once.

As mentioned above, the codebook 110 is composed of codebooks with different Mc, and the codebook selection section 10
Based on the codebook selection information output as 0, one codebook designed most suitable for the residual signal classified according to the spectral envelope of the input signal can be selected.

An example of how to create the M codebooks is shown below.

The M codebooks are created offline.

That is, M codebooks are created using a codebook creation algorithm such as the LBG algorithm (see Reference 2) independently from M training sequences classified and weighted based on M spectral patterns. .

Figure 4 shows M codebooks B, B2゜I..., BM. is classified based on the spectral shape H1.
This is a block diagram showing how it is created.

First, M training sequences T, T2.

..., TM. An example of how to create is shown below.

First, prepare a signal with similar statistical properties to the input signal to be encoded (for example, various audio signals if the input signal is an audio signal) for a sufficiently long time (typically several minutes), and then This is the original training series.

Next, from the original training sequence, the spectral shape H and residual signal for each block are determined in the same manner as for the input signal. At this time, the residual signal of the original training sequence is normalized and vectorized using the same method as performed by the encoder.

Next, M from the above residual signal. training series T
,T,...TM. make. This process is performed using M spectra shape turns H (iml
, 2. .

For example, when Hl out of M spectral shape patterns is optimal for the spectral shape H of the current block, the residual signal obtained for the current block is classified into the first training sequence Tt.

By performing the above classification process on sufficiently long original training sequences, M
Training sequences 'r, T2°I . . . TMe are created.

The final step in creating M codebooks is to create M codebooks using M training sequences.

What must be noted here is that the M food books are created independently based on the spectral shape classification information i. That is, the training sequence T
codebook B1 from l, B from T,...
BMe independently from T 2 2
MeCreate. At this time, for example, from T to B
The LBG algorithm used to create 1 is LBG
.

Then, the squared error distortion calculated by LBG, is weighted and calculated by the corresponding spectral pattern H,. Similarly, the centroid within the cluster (central vector within the cluster) calculated by LBG is also determined at a level weighted by the spectral pattern H1 so that the sum of distortions with each training vector within the cluster is minimized. Ru.

To express this more specifically, the centroid V of the cluster is determined by solving the following simultaneous equations.

Table V (k) R(k, j) k=1 hh --! - Hair v(1)(k)R05(k,j)N
-1 to m-1 (j-1, 2,..., L) <8
) two places, Ruc, j)-+:, h(n-k)h(n-j:+h
h, 1 (k, j-1, 2,..., L)
(9), L represents the dimension of the vector, ■(k) represents the th element value of the Sen(III) antroid, and V (k)
represents the th element value of the m-th vector of N vectors in this cluster, and h (n) represents the impulse response of the spectral pattern H1.

This concludes the explanation of the creation of M codebooks.

Since the created M codebooks Bl are each designed using a total scale with added weights, they become codebooks for the residual signal that are more suited to the shape of the spectrum than the white noise codebook.

Therefore, Mc codes each having a number of vectors NS (N ≦N, /Mc) that is an order of M smaller than the number NF of vectors required to obtain good quality with a conventional fixed codebook. The book allows prediction residual signal vector quantization of the same or higher quality than conventional methods.

Moreover, by simply selecting M codebooks through a very low-complexity process of classifying the spectral pattern of the synthesis filter found for each block of the input signal, the vector search range required for vector quantization of the residual signal is can be reduced to N 2 /Np (≦1/M) compared to the conventional method, which has the effect of making it possible to significantly reduce the amount of calculation while maintaining high quality.

When a codebook to be used is selected, the synthesis filter 20 applies the synthesis filter only once to each of the N vectors in the selected codebook independently to create a synthesis vector. The synthesis filter is 1/
(1+Z a Z-') In country l, N vectors are V - (V (1), V,
<2).

s f
i--”1 (L)l (i-1,2,-,N5)
(CL is the number of dimensions of the vector, and N is the number of codes), the composite vector U1 can be calculated using the following equation.

An example of quantization is shown in detail.

Now, G represents the normalized gain calculated and quantized for the current block of the input signal cut out in the block. Here, to simplify the explanation, one normalization gain will be used within the block.

The composite vector U1 obtained at this time is associated with the selected codebook by the code i, and is held in the memory until it is necessary to perform the residual signal vector quantization of the current block.

In this embodiment, the matching section 21 is configured to include the functions of both the matching section 5 and the residual signal code selection section 6 shown in FIG.

The matching unit 21 uses the extracted input signal, the selected one codebook, and the synthesis filter 120.
Vector quantization of the prediction residual signal is performed using the synthesis filter used in .

Below, the prediction residual signal vector performed here is the vector in the selected codebook as U − (Ul(1),
U, (2),..., Ul(L)1(i-1,2
,...,N) (L is the number of dimensions of the vector, N is the number of codes).

First, the input signal x (n) is divided into L samples, and K
(K-N/L, N is the number of samples to be encoded in the block) input vectors x (m-1°(m) . . . , K) are created. Here, X represents the m-th input vector. Input vector x (1) − (x (1), x (2), −
Consider selecting an optimal code for X(L)) from the codebook.

Influence signal x (L), x (1) (2), = -, x that affects the first input vector 1c (1) of the current block at the reproduced signal level of the past quantized residual signal
(1) (L) is f' f f' (n=1. ..., L)
It is expressed as (11). Here, x (n) is a past quantized value of x (n) and is a known value. By subtracting this from χ(1), we get (1)=U) (1) (,.

)χ　　　　K　　　Xr get.

υ1 created by the above X(1) and synthesis filter 20
(i-1, ..., NS), for example, using the following formula, the distortion (iml, ..., N)
Find (13). At this time, if the one with the smallest distortion is di) (1≦1(1)≦N), then the residual signal code to be transmitted is l. At this time,(
Since the quantized value of 1) is v U , the quantized value of the input vector is x (n) = U (1) (n) + x (1) (n
) fi f (n=1...., L) (14
) can be found.

mm2. using exactly the same procedure. By setting L+1≦n≦2L, for the next second input vector
Residual signal code 1 (2) can be obtained using equation (14). This operation is repeated until 9(N) of equation (14) is determined, and the residual signal vector quantization of the current block is completed. The procedure of residual signal vector quantization by codebook selection in this embodiment is summarized in the flowchart of FIG. In this example, the creation of the composite vector is
By performing this outside of the code search routine, residual signal vector quantization can be achieved with one-half the amount of calculation compared to a configuration in which the composite vector is determined internally. It has a further improved configuration.

The encoding unit 22 encodes a predetermined number of residual signal codes selected by the matching unit 21.

The encoding unit 14 encodes the prediction parameters quantized by the quantization unit 13.

The encoding unit 18 encodes the quantized single or multiple normalized gains.

The multiplexer 23 combines the coded residual signal code, the coded prediction parameter, and the coded normalization gain information and outputs the combination to the transmission path.

This concludes the explanation of the encoding device shown in FIG. 2.

Next, the third decoding device will be explained. In FIG. 3, the demultiplexer 24 is the multiplexer 23
The information transmitted from the controller is separated into encoded normalized gain, encoded prediction parameter, and encoded residual signal code information and output. Decoding unit 25
, 26. and 27 are the demultiplexer 2, respectively.
The information separated in step 4 is used to decode one or more quantized normalized gains, quantized prediction parameters, and a predetermined number of residual signal codes.

The codebook selection section 1000 is the codebook selection section 10
Since it has the same function as 0, the explanation will be omitted here.

Codebook 1100 has the same M as codebook 110.
The codebook is configured such that the same codebook as the codebook selected by the encoding device is selected from the M codebooks based on the selection information of the codebook selection unit 1000. There is. Then, a vector corresponding to the decoded residual signal code is output from the selected codebook.

The synthesis filter 200 multiplies the vector output from the codebook 1100 by the decoded normalization gain to generate a quantized prediction residual signal, and also generates a synthesis signal reproduced from the quantized prediction parameters. By inputting the quantized prediction residual signal to a filter, a reproduced signal within the block is generated. buffer 20
0 combines the reproduced signals block by block and outputs an output reproduced signal.

This concludes the description of the decoding device shown in FIG. 3.

When applying the present invention to encoding an input signal having a pitch such as an audio signal, the synthesis filter used in this embodiment,
By using, as a synthesis filter, a cascade of pitch filters representing the fine structure of the spectrum of an input signal, a high-quality encoding device can be realized.

In this embodiment, the configuration was explained in which the normalized gain is calculated for each block, but the normalized gain G(1) (m-1°..., K) is calculated from the prediction residual signal in units of vector dimensions. Further improved residual signal vector quantization may be performed using a configuration in which calculation is performed. In this case, equation (10) and (1
3) Equations are (i-1,..., N ; j-
1,..., P; n-1,...,'L)
(15) and (i-1,...,N)
(1G).

Further, it is also possible to improve the subjective quality by shaping the noise signal included in the reproduced signal using a known method.

For example, if the synthesis filter determined from the prediction parameters is
(z), the synthesis filter used in the encoder is H(z/γ) (0 x γ x 1), and the synthesis filter used in the decoder is H(z). can be shaped into the shape of the spectrum of the input signal.

The present invention can also be combined with other compatible fast search methods to further reduce the amount of computation required for code search.

For example, the present invention can be combined to reduce the amount of calculation required for vector quantization by using a model that has been thinned out using a predetermined method as a residual signal for generating a reproduced signal.
Demonstrate its effectiveness. In that case, it is necessary to create M codebooks by weighting the spectral patterns used in the LBG algorithm in a manner that matches the thinned out model.

[Effects of the present invention] As explained above, the residual signal vector quantization method according to the present invention performs the residual signal vector quantization at the reproduction level using a codebook suitable for the shape of the spectrum of the input signal. It is possible to provide residual signal vector quantization with better quality than that of quantization, and in addition, by selecting one codebook from multiple codebooks by classifying the spectral shape of the input signal, it is possible to select a small number of vectors from many vectors in multiple codebooks. This has the effect of greatly reducing the amount of calculation required to search for an optimal code vector.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the principle of a predictive residual signal vector quantization method according to a first embodiment of the present invention. 2 and 3 are block diagrams showing an example of applying the predictive residual signal vector quantization method according to the first embodiment shown in FIG. 1 to an encoding/decoding device, and FIG. is a block diagram showing the creation of M codebs in the predictive residual signal vector quantization method according to the first embodiment, and FIG. 5 is a flowchart showing the procedure for determining the residual signal code according to the first embodiment. , FIG. 6 is a block diagram showing the principle of the conventional method. 1.12.--Predicted parameter extraction unit 5.21.
. . . Matching section 6 . . . Residual signal code selection section 7.700, 110.1100. :r-Dobuk 8.
120.1200...Synthesis filter 9.100
．． 1000...Codebook selection section 11...
...Block cut section 13.17...Quantization section 14.18.22...Encoding section 25.26.2
7...Decoding unit 16...Normalization gain calculation unit 23...Multiplexer 24...Demultiplexer 200...Buffer

Claims (1)

[Claims] (1) In the prediction residual signal vector quantization method, which is an encoding method in which a synthesis filter is driven by a prediction residual signal obtained by predicting an input signal to obtain a reproduced signal, a predetermined number of different codebooks are used. At least one codebook is selected from the plurality of codebooks based on the frequency characteristics of the synthesis filter, and prediction is performed using the selected codebook, the synthesis filter, and the input signal. A predictive residual signal vector quantization method characterized by quantizing the residual signal.