JPH0588699A - Vector quantization system for speech drive signal - Google Patents

Abstract

PURPOSE:To encode the drive signal for a speech by using two code books whose code vectors overlap with each other and a composing filter. CONSTITUTION:While a target vector calculation part 11 generates a target vector for vector quantization from a speech signal and a filter coefficient calculation part 1 determines the coefficient of a composing filter from the speech signal, an impulse response calculation part 2 finds an impulse response and a drive signal with an integer pitch period whose code vectors overlap with a code book 3 is interpolated with a drive signal with an integer pitch period to store a drive signal with a non-integer pitch period whose code vectors overlap with a code book 5; and a composite vector is generated by a convolution part 6 with the impulse response and the drive signal with the integer pitch period, a composite vector is generated by a convolution part 7 with the impulse response and the drive signal with the non-integer pitch period, and a drive signal code is searched for by using those composite vectors and target vector.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an audio drive signal vector quantization system for encoding an audio drive signal using a codebook and a synthesis filter in which code vectors overlap.

[0002]

2. Description of the Related Art Conventionally, CELP (Cod) has been used as an encoding method capable of encoding a voice signal with high quality at a low rate of about 8 kbit / sec.
The e Excited Linear Predicion) method is known. CE
For details of the LP method, see "A Class of Analysis-b" by P. Kroon et al.
y-Synthesis Coding at RatesBetween 4.8 and 16kbits
/ s ", IEEE SAC-6, pp.353-363, February 1988.

The CELP system is classified into one of audio encoding systems for producing a synthetic speech signal by a driving signal and a synthesis filter, and is characterized by a target vector produced from the speech signal and a synthetic vector. The optimum code vector is selected from within the drive signal code book so that the waveform distortion of and becomes small. FIG. 3 is a block diagram showing the principle configuration of a drive signal vector quantization method performed on the code side of the conventional CELP method.

In the figure, reference numeral 212 denotes an input terminal, which is an input signal (for example, an audio signal) given to the input terminal 212.
Is a filter coefficient calculation unit 201 and a target vector calculation unit 2
11 respectively.

The filter coefficient calculation unit 201 predicts and analyzes the input signal to obtain the filter coefficient of the synthesis filter representing the vocal tract characteristic. The filter coefficient is given to the impulse response calculation unit 202, and the impulse response calculation unit 202 An impulse response of a finite number of samples weighted by a predetermined weighting function is obtained based on the coefficient of the synthesis filter. Also,
The target vector calculation unit 211 includes the filter coefficient calculation unit 20.
The input signal is weighted by the weighting filter determined based on the filter coefficient of 1, and the influence signal to the present input signal due to the past encoding of the influence signal calculation unit 213, which will be described later, is subtracted from the weighted input signal and driven. The target vector X used for vector quantization of the signal is calculated.

On the other hand, reference numeral 203 denotes a codebook, which sequentially takes in past drive signals V (n) generated by the drive signal generation unit 214, and also unnecessary drive signals in the codebook 203. The contents can be updated by discarding.

The drive signal data of the codebook 203 is
It is sent to the convolution unit 206, and in this convolution unit 206, the impulse response h of the impulse response calculation unit 202
A composite vector Di is calculated from (n) and a predetermined code vector based on the pitch period to be searched from the codebook 203.

Then, the composite vector of the convolution unit 206
Di is input to the pitch synthesizing unit 208, pitch synthesizing processing that gives an integer pitch period I is performed, and this is performed by the code searching unit 21.
5, a code search is performed using the synthesized vector U and the target vector X obtained by giving the pitch cycle I in the pitch synthesis unit 208.

Here, in the code search in the code search unit 215, the distortion X ^t between the combined vector U and the target vector X when the optimum gain is given. X- (X ^t U) ² / (U ^t U) minimization, that is, the value (X ^t U) ² / (U ^t This is done by searching for the code with the maximum U). Then, the information of the optimum code is output from the terminal 218 and the gain determining unit 216
Sent to.

In the gain determining unit 216, the code searching unit 2
Optimal gain (X
^t U) / (U ^t U) is used as a target value to determine the gain of the drive signal code vector, and this is given to the drive signal generation unit 214.

As a result, the drive signal generation unit 214 uses the optimum code information from the code search unit 215 and the gain of the gain determination unit 216, and multiplies the optimum code vector by the gain to generate the drive signal of the voice in the current section. .. Further, this drive signal is given to the influence signal calculation unit 213, and the influence signal calculation unit 213 gives the influence signal to the next section by the current encoding for vector quantization of the drive signal of the voice of the next section. To calculate.

Thus, according to the CELP method,
The synthesized voice signal generated from the selected drive signal code and the parameters of the synthesis filter has been confirmed on the coding side that the waveform distortion is small, so even on the decoding side, a good synthesized speech for a rate of about 8 kbit / sec. Can be generated.

Generally, in the CELP system, a plurality of codebooks having different properties are often used as the drive signal codebook, and the adaptive codebook is known as the most efficient codebook among them.

The adaptive codebook uses the coded drive signals of voices encoded in the past, and the drive signals in the section having a strong periodicity (pitch) in human voice are
Since a similar waveform repeats every bitch cycle, the adaptive codebook enables efficient coding of the drive signal.
Also, the code vector of the adaptive codebook is coded based on the pitch period, and is generated by creating a vector that repeats at this pitch period by using the past drive signal for an integer pitch period from the current sample point. It

The code vector generated in this way has a structure that overlaps the code vectors of adjacent code numbers. Conventionally, this overlap structure is used to calculate a composite vector at high speed. Recursive convolution is being considered.

The principle of high-speed convolution will now be briefly described. Now, as shown in FIG.
If the signal sequence stored in 0 is C (n) (n = ...,-2-1) and the code vector corresponding to the integer pitch period k is V _k ,
V _k and V _k-1 have the following overlapping structure. V _k-1 = (C (-k + 1), C (-k + 2), ..., C (-k + L-1), C (-k + L)) V _k = (C ( -k), C (-k + 1), C (-k + 2), ..., C (-k + L-1)) (1) Here, L represents the dimension of the vector. Also, by placing the n-th element of V _k and _{V k (n), V k} (n) = C (-k + n-1) ... (2)

It is represented by the relationship: Then, let h be the impulse response of the filter used for convolution, and let d _k be the composite vector obtained by convolving V _k and the impulse response.
_{If k} , the following relationship holds between the combined vectors of adjacent codes. d _k (n) = V _k (1) h (n) + d _k-1 (n-1) = c (-k) h (n) + d _k-1 (n-1) (3)

In the above equation, if the composite vector d _k-1 of the previous code has been calculated, then c (-k) h (n) is added to this to convolve the composite vector d _k of the next code. It shows that the calculation can be done recursively on the order of the complexity L.

Therefore, by using such a relation, the order of the amount of calculation required for convolution is L ² To L, the amount of calculation required to search for an optimal code vector can be reduced to about L / L, which is an important technique for completing speech coding in real time.

However, in such a CELP system, 8 kbi
At a rate lower than t / sec, the number of bits allocated to the drive signal becomes small, so waveform distortion of the synthesized voice signal becomes perceived as noise, and especially at a low rate of 4 kbit / sec or less, quality deterioration occurs. There is a problem in that it is not possible to perform large-scale and high-quality speech coding. As a solution to this, a method of improving the accuracy of the pitch period of the adaptive codebook has been considered.

This method is based on the 1990 International Conf
erence on Acoustics, Speech, and Signal Processing
P. Kroon's paper "PI
TCHPREDICTORS WITH HIGH TEM PORAL RESOLUTION "(PP.6
61-664) and JS Marques et al. "IMPROVED PITC.
H PREDICTION WITH FRACTIONAL DELAYS IN CELP CODIN
G "(PP.665-668).

Briefly explaining this method, the audio signal is usually sampled at 8 kHz and digitized.
For this reason, the conventional pitch period search precision is performed with integer sample precision. However, in practice, the pitch period included in human speech changes gradually and continuously, and therefore pitch search may not be performed properly due to lack of precision in pitch search, which is a cause of deterioration in synthesized speech quality. It is one. Therefore, in order to solve this, by up-sampling the past drive signal (integer sample precision) in the drive signal codebook and searching up to a pitch period of a non-integer sample precision which is finer than the integer sample precision, the drive signal is more accurately obtained. It is possible to express the pitch period of, and it is possible to synthesize a smooth uttered voice even at a low bit rate. However, in these papers,
Nothing is said about the method of fast convolution when searching for non-integer pitch periods.

With respect to this, in the CELP method of FIG. 3 described above, as a method of calculating the integer pitch period and the non-integer pitch period by high-speed convolution, the up-sampled past drive signal is sent to the convolution unit 206, and the impulse response calculation unit Convolution is performed using the result of up-sampling the impulse response h (n) from 202.

However, if this is done, all the data used for convolution will be changed from L to 2L.
The amount of calculation for fast convolution is also on the order of 2L. Therefore, in the convolution unit 206, if the pitch period candidates to be searched are the integer pitch period M types and the non-integer pitch period N types, and the dimension of the code vector is L, 2L (to obtain (M + N) combined vectors. There is a drawback in that the calculation of the order of M + N) is required and the calculation amount is remarkably increased.

[0025]

As described above, in the conventional vector coding method of the voice driving signal in the voice encoding, at the low rate of 4 kbit / sec or less, the convolution calculation amount in the convolution unit is significantly increased. Therefore, there is a problem that the optimum pitch period cannot be searched at high speed, the work efficiency is significantly reduced, and it is difficult to perform high-quality speech coding in real time.

The present invention has been made in view of the above circumstances, and it is possible to reduce the amount of convolution calculation, and to realize high-quality real-time speech coding at a low rate of 4 kbit / sec or less. The purpose is to provide a scheme.

[0027]

In order to achieve the above-mentioned object, the vector quantization system of the audio drive signal of the present invention is
Target vector generation means for generating a target vector for vector quantization from an input voice signal, impulse response generation means for determining a coefficient of a synthesis filter from the voice signal and obtaining an impulse response from the filter coefficient, code vector overlap A first codebook storing a drive signal of an integer pitch cycle, and a drive signal of a non-integer pitch cycle in which a code vector obtained by interpolation processing with respect to the drive signal of the integer pitch cycle of the first codebook is overlapped. The second codebook to be stored, the first convolution means for convoluting the impulse response of the impulse response generating means and the drive signal of the integer pitch period of the first codebook to generate a combined vector, the impulse response The impulse response of the generating means and the second Second generating a composite vector performs convolution from the drive signal of the non-integer pitch period of the codebook
And the code search means for searching the drive signal code using the combined vector from the first and second convolution means and the target vector of the target vector generation means.

Further, according to the vector quantization method of the voice drive signal of the present invention, the target vector generating means for generating the target vector of the vector quantization from the input voice signal, and the coefficient of the synthesis filter from the voice signal are determined. Impulse response generating means for obtaining an impulse response from the filter coefficient, a first codebook for storing a drive signal of an integer pitch period with overlapping code vectors, and interpolation for the drive signal of the integer pitch period of the first codebook. From a second codebook for storing a drive signal of a non-integer pitch cycle in which code vectors obtained by processing overlap, an impulse response of the impulse response generating means and a drive signal of the integer pitch cycle of the first codebook. The first convolution that performs the convolution to generate a composite vector A second convolution means for convoluting the impulse response of the impulse response generation means and the drive signal of the non-integer pitch period of the second codebook to generate a combined vector, and the first convolution means. First pitch synthesizing means for generating a synthetic vector having an integer sample pitch period with respect to the generated synthetic vector, and a synthetic vector generated by the first convolution means according to the integer sample pitch period. Delaying and adding means for delaying and adding the combined vector generated by the second convolution means to generate an added combined vector; second pitch for generating a combined vector having a pitch period of integer samples from the added combined vector Synthesizing means, synthetic vectors from the first and second pitch synthesizing means, and the target vector generating means Constitute a code search means for searching a driving signal code using the target vector.

[0029]

As a result, according to the present invention, the first codebook stores the previously encoded integer-pitch-cycle drive signal sequence having the code vector overlapping structure, and the second codebook stores The non-integer pitch-synchronized drive signal series having the overlap structure in which the drive signal series in the first codebook is shifted by the insertion processing is stored, and the drive signal series of the first and second codebooks are stored independently of each other. Since the convolution processing is performed by doing so, the amount of calculation of the convolution can be reduced, the distortion amount between the synthetic vector generated by these convolutions and the target vector can be calculated at high speed, and the non-integer pitch period is provided. Vector quantization of the drive signal can be performed at high speed.

[0030]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the principle structure of a vector quantization system for a voice drive signal according to the embodiment. In the figure, reference numeral 12 is an input terminal.
An input signal (for example, a voice signal) given to is sent to the filter coefficient calculation unit 1 and the target vector calculation unit 11, respectively.

The filter coefficient calculation unit 1 predicts and analyzes the input signal in units of frame length (usually about 20 ms) to obtain the filter coefficient of the synthesis filter representing the vocal tract characteristics, and supplies this filter coefficient to the impulse response calculation unit 2. Here, the impulse response of a finite number of samples weighted by a predetermined weighting function is obtained based on the coefficient of the synthesis filter. Subsequent processing is performed for each small section with the time length represented by the sample of the number of dimensions of the vector as one small section.

First, the target vector calculation unit 11 weights the input signal with a weighting filter determined based on the filter coefficient of the filter coefficient calculation unit 1, and the present input by past coding of the influence signal calculation unit 13 described later. The influence signal on the signal is subtracted from the weighted input signal to calculate the target vector X 1 used for vector quantization of the drive signal.

On the other hand, 3 is a codebook consisting of an adaptive codebook, which takes in the past drive signal V (n) generated by the drive signal generation unit 14 described later, and unnecessary drive in the codebook 3. By discarding the signal, its contents can be updated. In this case, the drive signal taken into the codebook 3 corresponds to an integer pitch period T = I (I is an integer), and its code vector V (n) (n =
0,1, ..., L) is defined by the following equation. (Where L
Represents the dimension of the vector. ) V (n) = V (nI) (n = 0,1, ..., I-1) (4) The drive signal data of the codebook 3 is sent to the interpolation calculation unit 4 and the convolution unit 6. ..

The interpolation calculation section 4 obtains the interpolation signal V _f (n) for the past drive signal V (n) in the codebook 3 by using, for example, an interpolation filter. In this case, a technique for obtaining the signal amplitude at a position where the non-integer sample phase is shifted from the discrete signal data by using an interpolation filter is known, for example, R.
Co-authored by E. Crochiere and LRRabiner at Prentice Hall
Multirate Digital S published by the company in 1983
Since it is described in detail in ignal Procesing, the explanation is omitted here.

Interpolation signal V obtained by the interpolation calculation unit 4
_{The f} (n) is sent to the codebook 5 and is temporarily stored therein to form an adaptive codebook having drive signal data of a non-integer pitch period. Then, the drive signal data of the codebook 5 is sent to the convolution unit 7.

The convolution unit 6 uses the impulse response h (n) of the impulse response calculation unit 2 and the code vector based on the integer pitch period searched for from the code book 3 by using the convolution method of the above-described Expression 3 to generate a composite vector D _i. To calculate.

On the other hand, also in the convolution unit 7, the impulse response h (n) of the impulse response calculation unit 2 and the codebook 5
Synthetic vector D _f from the code vectors based on non-integer pitch period to search using the convolution method of the formula 3 described above
To calculate.

Then, a synthetic vector of the convolution parts 6 and 7 is obtained.
Cutle D_i, D_fIs given to the code search unit 15 and convolution is performed.
Let the composite vector from parts 6 and 7 be a composite vector U
A code search is performed using the characteristic vector X. Coe here
Search for the value (X^t U)² / (U^t U) is the maximum
This is done by searching for the de. And the optimal code information
The information is output from the terminal 18.

In the gain determining section 16, calculation is performed based on the optimum code (X ^t U) / (U ^t U) is input from the code search unit 15, the gain of the drive signal code vector is determined using this as a target value, and this is given to the drive signal generation unit 14. As a result, the drive signal generation unit 14 uses the optimum code information from the code search unit 15 and the gain of the gain determination unit 16,
By multiplying the optimum code vector by the gain, the voice drive signal for the current small section is generated. Further, this drive signal is given to the influence signal calculation unit 13, and the influence signal calculation unit 13 transmits to the next small section by the current encoding for vector quantization of the drive signal of the voice of the next small section. The influence signal is calculated.

Therefore, in this way, the codebook 3
The drive signal of the integer pitch cycle is stored in the codebook 5, and the drive signal of the non-integer pitch cycle obtained by the interpolation calculation is stored in the codebook 5, respectively, and the impulse response h from the impulse response calculation unit 202 is separately provided for each codebook 3 and 5. Since the convolution is performed based on (n), if the pitch period candidates to be searched are M integer pitch periods and N non-integer pitch periods and the dimension of the code vector is L, (M
The amount of calculation required to obtain + N) composite vectors is
The order is L (M + N) times, the amount of calculation can be significantly reduced, and high-quality real-time speech coding can be realized even at a low rate of 4 kbit / sec or less. Next, a second embodiment of the present invention will be described with reference to FIG. In this case,
1 have the same functions as those described above, the description thereof is omitted here.

In this embodiment, a drive signal vector quantization system is shown which can also be used for pitch period search when the pitch period T to be searched and the dimension L of the code vector have a relation of T <L.

In this case, for the integer pitch period I, the pitch synthesizing unit 8 inputs the synthesis vector D _i from the convolution unit 6 and performs the pitch synthesizing process for giving the pitch period I. This processing can be performed as follows, for example.

U (n) = D _i (n) + U (nI) (5) Next, the pitch period T is a decimal, that is, a non-integer sample length F
Considering the case of, the drive signal sequence of Codebook 3 is V
(n), when a driving signal sequence codebook 5 calculated by the interpolation calculation section 4 and Vf (n), V (n ) of sampling frequency f _s from 2f _s signal sequence obtained by raising the VD ( n) becomes VD (2m) = V (m) (6) VD (2m-1) = V _f (m) (7). Then, if VD is written in order of time using V and V _f , it becomes as follows. ... V _f (n-1), V (n-1), V _f (n), V
(n), ..., V _f (-2), V (-2) V _f (-1), V (-1)

This signal sequence VD is pitched from the point of n = 0
Code vector V (n) (n = 0,1, ..., L-1) that repeats at 2F (= 2I + 1: odd number) and repeats at a fractional period F by lowering the sampling frequency to the original f _s Is obtained.

In this case, since 2F (= 2I + 1) is an odd number,
The first pitch repetition starting from V (0) in the sampling period fs is V (0) = V _f (-I), V (1) = V _f (1-I), ... It can be seen that it can be obtained only from the interpolated value V _f . And in the second iteration, the interpolated _value of V _f ,
That is, the code vector can be obtained only from V. Using this relationship, the code vector V (n) (n = 0,1
…, L-1) can be calculated by the following formula. V (n) = V _f (n) = V _f (nI) (n = 0, ..., min (L-1, I-1))… (9) V (n) = V _i (n) = V (n-2I-1) (n = I, ..., min (L-1,2I)) (10) V (n) = V (n-2F) (n = 2F, ... , L-1)… (11)

Here, it is the composite vector U that is actually used for distortion calculation with the target vector, and the composite vector U that is the convolution of V and the impulse response h of the composite filter is obtained by directly using V in the above equation. Can be calculated, but V _f in Eq. 9 and Eq. 1
The convolution calculation can be performed separately with V _{i of} 0, for example, as follows. D _f (n) = V _f (n) * h (n)… (12) D _i (n) = V _i (n) * h (n)… (13) U (n) = D _f (n) + D _i (nI) + U (n-2F) (14)

In this case, the equations 12 and 13 can be calculated by the above-described high-speed convolution for each adaptive codebook. However, since D _i is obtained by the convolution unit 6 at the time of searching for an integer pitch period, D _i in Equation 13 can be obtained by performing a search for an integer pitch and a non-integer pitch.
_It is possible to reduce the amount of calculation required for searching a non-integer pitch by the amount of calculation of _i . Since these calculations are performed once for each pitch period, the amount of calculation can be significantly reduced particularly when the number of pitch periods to be searched (codebook size) is large. From this, it is understood that the synthetic vector repeated in the non-integer pitch cycle may be obtained by inputting the synthetic vector from the convolution units 6 and 7 and performing the processing of Expression 14.

Here, the delay adder 9 and the pitch synthesizer 1
0 performs the processing of Expression 14, and the delay adder 9 delay-shifts the composite vector D _i from the convolution unit 6 by the number of samples based on the pitch period I, and the composite vector D _i of the convolution unit 7 _f is added, and an added synthetic vector is output. Further, the pitch synthesizing unit 10 is given the addition synthesizing vector of the delay adding unit 9, and performs a pitch synthesizing process for giving an integer pitch period 2F (2I + 1) to this. If the pitch period T 1 to be searched is equal to or larger than the number of dimensions L of the vector, the processing in the delay adding unit 9 and the pitch synthesizing units 8 and 10 can be omitted.

Therefore, in the vector quantization of the driving signal according to the present embodiment, in the search for the code vector representing the non-integer pitch period, the high-speed convolution method is used without increasing the sampling rate to efficiently combine the vector. Therefore, for example, if the pitch period candidates to be searched are M integer pitch periods and N non-integer pitch periods N, and the dimension of the code vector is L, when the pitch period is T ≧ L, (M + N) Since the product-sum calculation amount required to obtain each composite vector requires L times of product-sum calculation per composite vector according to the high-speed convolution method of Expression 3, (M
The synthesized vector for + N) types of pitch period candidates can be calculated by multiplying and summing L (M + N) times of products, and real-time high-quality low-rate speech coding can be realized.

[0051]

As described above, according to the present invention, it is possible to perform high-speed convolution even for the adaptive codebook search of a non-integer pitch period necessary for accurately expressing the pitch period of speech, and 4 kbit / sec. It is possible to realize high-quality speech coding in real time even at the following low rates.

[Brief description of drawings]

FIG. 1 is a block diagram showing a schematic configuration of an embodiment of the present invention.

FIG. 2 is a block diagram showing a schematic configuration of a second embodiment of the present invention.

FIG. 3 is a block diagram for explaining a conventional vector quantization method of a drive signal for audio.

FIG. 4 is a diagram for explaining a conventional vector quantization method of a drive signal for audio.

[Explanation of symbols]

1 ... Filter coefficient calculation unit, 2 ... Impulse response calculation unit,
3, 5 ... Codebook, 4 ... Interpolation calculation unit, 6, 7 ... Convolution unit, 9 ... Delay addition unit, 8, 10 ... Pitch synthesis unit, 1
DESCRIPTION OF SYMBOLS 1 ... Target vector calculation part, 12 ... Input terminal, 13 ... Influence signal calculation part, 14 ... Drive signal generation part, 15 ... Code search part, 16 ... Gain determination part.

Claims (2)

[Claims] 1. A target vector generating means for generating a target vector for vector quantization from an input voice signal, and an impulse response generating means for determining a coefficient of a synthesis filter from the voice signal and obtaining an impulse response from the filter coefficient. And a first codebook that stores a drive signal sequence of an integer pitch period in which code vectors overlap, and a code vector obtained by interpolation processing on the drive signal sequence of an integer pitch period of the first codebook is overlaid. A second codebook for storing a non-integer pitch cycle drive signal sequence to be wrapped, a convolution from the impulse response of the impulse response generating means and the integer pitch cycle drive signal sequence of the first codebook, and synthesis. First convolution means for generating a vector, said impulse From the impulse response of the answer generation means and the drive signal sequence of the non-integer pitch period of the second codebook, second convolution means for generating a combined vector; and the first and second convolution means And a code search means for searching a drive signal code by using the target vector of the target vector generation means. 2. A target vector generating means for generating a target vector for vector quantization from an input voice signal, and an impulse response generating means for determining a coefficient of a synthesis filter from the voice signal and obtaining an impulse response from the filter coefficient. And a first codebook that stores a drive signal sequence of an integer pitch period in which code vectors overlap, and a code vector obtained by interpolation processing on the drive signal sequence of an integer pitch period of the first codebook is overlaid. A second codebook for storing a non-integer pitch cycle drive signal sequence to be wrapped, a convolution from the impulse response of the impulse response generating means and the integer pitch cycle drive signal sequence of the first codebook, and synthesis. First convolution means for generating a vector, said impulse Second convolution means for convoluting the impulse response of the answer generation means and the drive signal sequence of the non-integer pitch period of the second codebook to generate a combined vector; and the second convolution means for generating the combined vector. First pitch synthesizing means for generating a synthetic vector having an integer sample pitch period with respect to the synthetic vector, and delaying the synthetic vector generated by the first convolution means according to the integer sample pitch period. And a second adding pitch for generating a combined vector having an integer sample pitch period from the added combined vector, and a delay addition means for adding the combined vector generated by the second convolution means to generate an added combined vector Combining means, combining vectors from the first and second pitch combining means, and an eye of the target vector generating means Vector quantization scheme of the audio driving signal, characterized by comprising a code search means for searching a driving signal code using a vector.