JPH07271397A - Voice encoding device - Google Patents

Abstract

PURPOSE:To provide a voice encoding device in which a high precision preselection of driving vectors is accomplished, the number of candidates of the driving vectors is reduced while keeping the high quality of encoded voices and the amount of computations for a final selection is reduced. CONSTITUTION:The device is provided with a driving vector generating section 101, a synthetic filter 102 which receives the generated driving vectors and generates synthetic voice vectors, a preselection section 104 which selects at least one driving vector from the generated driving vectors and a final selection section 105 which selects an optimum driving vector from the selected driving vector by the section 104. The section 104 selects the driving vector which makes the size of the inner product value of a target vector obtained from an inputted voice and the synthetic voice vector larger than the value of the driving vector that is weighted by a weighting function in which the driving vector is made as a parameter.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a linear predictive analysis type speech coding apparatus, in particular, a plurality of driving vectors such as CELP are input to a synthesis filter, and the obtained synthesized speech vector and the input speech are weighted under a perceptual weight. The present invention relates to a speech coder that searches a codebook for a drive vector that minimizes distortion by comparing the above.

[0002]

2. Description of the Related Art CELP (Code Excit) is used as a method for encoding telephone band voice at a transmission rate of about 4 kbps.
The ed Linear Prediction) method is one of the effective methods. The processing by the CELP method is roughly divided into a processing for obtaining a voice synthesis filter that models a vocal tract from input speech divided into frame units, and a processing for obtaining a drive vector corresponding to an input signal of this filter. Of these,
The latter requires a process called codebook search that compares the synthesized speech with the input speech by passing multiple drive vectors stored in the codebook one by one through a speech synthesis filter, and this processing requires a large amount of calculation. And The present invention relates to reduction of calculation amount in codebook search.

Regarding the CELP method, for example, MRSchr
oeder and BSAtal, "Code ExcitedLinear Prediction
(CELP): High Auality Speech at Very Low Bit Rate
s ", Proc. ICASSP, pp.937-940, 1985 and WSKlei
jin, DJKrasinski et al. "Improved Speech Quality
and Efficient Vector Quantization in SELP ", Proc.I
CASSP, pp.155-158, 1988.

FIG. 9 is a block diagram showing an outline of a speech coding apparatus based on the CELP method. First, the codebook search will be described. Two systems are prepared as drive vectors. These are adaptive codebook 901 and random codebook 90.
Stored in 2. The adaptive codebook 901 is a variable codebook that stores past drive vector groups, whereas the noise codebook 902 is a fixed codebook that stores a plurality of fixed patterns. The linear prediction analysis unit 907 analyzes the input speech vector R input to the terminal 906, obtains the characteristics of the synthesis filter 908, and then selects one optimal driving vector from each of the adaptive codebook 901 and the noise codebook 902. .

Specifically, the drive vectors extracted one by one from the codebooks 901 and 902 are combined into a synthesis filter 908.
Then, the obtained output (synthesized speech vector) is compared with the input speech vector R, and the codebooks 901 and 902 are searched for the drive vector that produces the optimum synthesized speech vector that is closest to the input speech vector R.

Next, a method for searching the drive vector will be described using mathematical expressions. The i-th drive vector extracted from the codebooks 901 and 902 is yi, and this is the synthesis filter 90.
When the synthesized speech vector obtained through 8 is denoted by Yi and the input speech vector is denoted by R, the difference between Yi and R expressed by the following equation is 2
It is general to search for a drive vector yi that minimizes the sum of multiplications, E = | R−αYi | ² (1). However, α is the optimum gain when yi is selected, and is given by the gain circuits 903 and 904 in FIG. The optimum gain α can be obtained by setting the expression obtained by partially differentiating this expression by α to zero. By substituting this into the equation (1) and rearranging, R = | R | ² − (R, Yi) ² / | Yi | ² (2). Since the first term of this equation is a constant that does not depend on the driving vector, searching for the optimal driving vector under the optimal gain α is the second term of the equation (2); (R, Yi) ² / | Yi | ² This is equivalent to using (3) as an evaluation expression and finding yi that maximizes this. Since this evaluation formula is calculated for all the candidates, the codebook search is the part that requires the most calculation amount in the entire CELP method.

Therefore, using a simplified evaluation formula, N candidates (1 <N <M) are selected from the codebook consisting of M driving vector candidates, and the above-mentioned N candidates are selected from the above candidates. A method has been proposed that uses an evaluation formula to narrow down the optimal candidates to one. This method is described in detail in JP-A-5-10.
0697. The following evaluation formula is obtained from M drive vectors obtained from the codebook; E = (R, Yi) ² (4) The first half of the work to select N is called preselection, and N candidates are calculated. The work in the latter half of narrowing down to one using (3) is called main selection. According to this method, the main selection, which requires a large amount of calculation, only needs to be performed for N selected in the preliminary selection, and the main selection is performed for all M candidates of the codebook without performing the preliminary selection. The amount of calculation can be reduced significantly compared to the case.

It is considered that this conventional pre-selection method is based on the assumption that the power of Yi corresponding to the denominator of equation (3) is almost constant. However, in reality power | Yi |
Considering ² is that the power of the drive vector yi is not always constant, and even if it is constant, the gain of the synthesis filter depends on the drive vector and does not become a constant.
There is a problem in that the assumption that Yi | ² is constant is unreasonable, which causes a decrease in the accuracy of preselection.

In recent years, the noise codebook proposed in the classical CELP method has been used less frequently, and the structured code is used in order to reduce the calculation amount and the memory amount and obtain a coded voice with higher sound quality. Often, a random codebook is used. For example, based on the pitch information obtained by the overlapping codebook and the adaptive vector search unit, which extracts a drive vector for one frame from one long noise signal while overlapping it with the previous drive unit vector. There are pitch-synchronized codebooks that are made periodic, and adaptive density codebooks that use one drive vector by changing the number of 0s sandwiched between samples.

Due to the structure of these structured codebooks, it is difficult to keep the power of the drive vector constant.
In such a situation, the use of the conventional pre-selection method, which assumes that | Yi | ² is constant, reduces the accuracy of pre-selection and consequently deteriorates the quality of coded speech. Further, if the number of candidates to be left in the preliminary selection is increased in order to maintain the quality of the coded voice, there is a problem that the amount of calculation in the main selection increases.

[0011]

As described above, when the preselection of the drive vector is performed by using the structured codebook, the power of the drive vector is not constant. Therefore, the conventional preselection method uses the drive vector. There was a problem that the selection accuracy of was not necessarily good.

The present invention enables highly accurate preselection of drive vectors and can reduce the number of drive vector candidates to be passed to the main selection while maintaining the quality of the encoded speech, and most of the calculation amount required for encoding is reduced. It is an object of the present invention to provide a speech coding apparatus capable of reducing the calculation amount in this selection.

[0013]

According to the present invention, a drive vector generating means for generating a drive vector, a synthesizing means for receiving a drive vector generated by the drive vector generating means and generating a synthesized voice vector, and a drive Speech coding having preselection means for selecting at least one drive vector from the drive vectors generated by the vector generation means, and main selection means for selecting an optimum drive vector from the drive vectors selected by the preselection means The essence of the apparatus is that the preselection means is configured to preselect drive vectors by an evaluation formula weighted by a weighting coefficient based on the power of drive vectors.

That is, in the first aspect of the invention, the magnitude of the inner product value of the target vector and the synthesized speech vector obtained from the input speech divided into the predetermined unit period is the driving vector generated by the driving vector generation means as a parameter. The pre-selection unit is configured to select the drive vector that makes the value weighted by the weighting function larger.

When this is expressed by a mathematical expression, M driving vectors yi (i = 1, 1, which are generated from the driving vector generating means).
, M) are input to the synthesizing means to search for a driving vector that outputs the synthesized speech vector closest to the target vector R, the weighting coefficient W (yi) having yi as a parameter is used in the preliminary selecting means. Preliminary selection evaluation formula; E = W (yi) (R, Yi) ² (5) N y (1 <N <M) yi for increasing the value are selected as preliminary selection candidates. However, Yi is an output obtained by inputting yi to the voice synthesizing means.

In the second invention, the drive vector generating means has a codebook, and when one drive vector specified by a predetermined index is cut out from the codebook and is generated, it is divided into predetermined unit periods. The value obtained by weighting the size of the inner product value of the target vector obtained from the input speech and the synthesized speech vector by a weighting function using the past drive vector group and the index stored in the codebook of the drive vector generation means as parameters The preselection means is configured to select at least one drive vector to be increased.

When this is expressed by a mathematical expression, when a drive vector group stored in the codebook is C and an index is i, when a drive vector that outputs a synthesized voice vector closest to the target vector R is searched for, preselection is performed. In the means, a preselection evaluation formula using a weighting coefficient W (C, i) having C and i as parameters; E = W (C, i) (R, Yi) ² (6) Drive for increasing the value of Select a vector as a preliminary selection candidate.

Further, in the third invention, after obtaining the target vector and the optimum synthesized speech vector obtained from the input speech divided into the predetermined unit period, the synthesized speech vector generated by the synthesizing means is made into the optimal synthesized speech vector. A driving vector that obtains an orthogonalized vector that has been orthogonalized, and increases the value obtained by weighting the size of the inner product value of the orthogonalized vector and the target vector with a weighting factor using the driving vector generated by the driving vector generation means as a parameter The pre-selection means is configured to select.

Describing this using mathematical expressions, the target vector R
Under the condition that the optimum synthesized speech vector X approximating R is already obtained, M driving vectors yi generated from the driving vector generating means are driving vectors for outputting the second synthesized speech vector approximating R. When searching from among (i = 1, ... M), the synthesized speech vector Yi of yi is orthogonalized to X to obtain an orthogonalized vector Yvi, and then y
Preliminary selection evaluation formula using weighting factor W (yi) with i as a parameter; E = W (yi) (R, Yvi) ² (7) N (1 <N <M) increasing the value of Select yi as a preliminary selection candidate.

[0020]

The evaluation formula in the preselection according to the present invention; E = W (yi) (R, Yi) ² (8) is a conventional preselection evaluation formula which is not multiplied by the weight W (yi) based on the following grounds. It can be said that the accuracy is higher than the equation (4).

The denominator of the expression (3), which is the evaluation expression of this selection, is the power of the synthetic speech vector, and | Yi | ² = G (yi) ² using the drive vector yi and the gain G (yi) of the synthetic filter. It can be expressed as | yi | ² (9). G (yi) takes different values depending on yi, but if the shape of the spectrum of yi is almost the same, it can be assumed to be a constant value. In practice, the codebook is often composed of noise sequences and the like, and the distribution of the spectra does not differ greatly between drive vectors. Therefore, it is realistic to assume that G (yi) is a constant value, and by setting G = G (yi) as a constant, equation (9)
Can be written as | Yi | ² = G ² | yi | ² (10). This expression shows that the power of the synthetic speech vector Yi can be estimated by multiplying the power of the driving vector yi by the square of the constant G which is the gain of the synthetic filter. Here, assuming that the power of the drive vector included in the codebook is a constant value in advance, equation (9) further becomes | Yi | ² = G ² y ² (11), and | Yi | ² is a constant. Become. As a result, formula (3)
In order to compare the size of the evaluation formulas of, it is possible to make an approximate evaluation using only the molecule. It is considered that the conventional method mainly compares the magnitudes of the evaluation formulas with the denominator | Yi | ² as a constant based on this assumption. However, as described in the section of the prior art, the structured codebook Considering the recent situation in which is used, this assumption causes the accuracy of the preselection evaluation formula to decrease. On the other hand, in the evaluation formula of the present invention, the weight coefficient W (yi) is set to the reciprocal 1 of the power of the drive vector yi.
When / │Yi | ² is set, the power of the drive vector can be included in the evaluation formula, and the accuracy of the evaluation formula is improved accordingly. Further, when it is difficult to obtain the reciprocal of the power, even if the estimated value is used, the accuracy is better than that of the conventional method using the constant. Also, the conventional method can be considered as a special case where W (yi) = 1 in the present invention.

By the way, the number N of candidates to be selected in the preliminary selection is closely related to the precision of the simplification of the evaluation formula, and if the simplification is performed with high precision, N can be a small value, and as a result, it is required in the main selection. The amount of calculation is also small. If over-simplification is impaired in accuracy, N must be increased in order to maintain the quality of coded speech, resulting in an increase in the amount of calculation in this selection. In other words, it is considered that the point of preliminary selection is how to simplify the evaluation formula without lowering the accuracy.

When the evaluation formula based on the present invention is used, the accuracy of the preliminary selection is improved as described above, so that the number of candidates to be passed to the main selection can be reduced while maintaining the quality of the coded speech, and the calculation amount of the main selection can be reduced. Can be reduced. The calculation amount in this selection occupies most of the calculation amount of the entire encoding device, and as a result, the calculation amount of the entire encoding device is significantly reduced.

[0024]

Embodiments of the present invention will be described below with reference to the drawings.

(Embodiment 1) FIG. 1 shows the configuration of a speech coding apparatus according to an embodiment of the present invention. This speech coding apparatus is roughly divided into a drive vector generation unit 101, a synthesis filter 102, a preliminary selection unit 104, and a main selection unit 108. The input voice vector R is input to the input terminal 103. The preliminary selection unit 104 includes a weighting factor derivation unit 105, an evaluation formula calculation unit 106, and an evaluation unit 107.

The drive vector yi generated by the drive vector generation unit 101 is passed through the synthesis filter 102,
A synthetic speech vector Yi is obtained. Further, the drive vector yi is also input to the weighting factor deriving unit 105, and the weighting factor W
(Yi) is obtained. In the evaluation formula calculation unit 106, the weight coefficient W with the synthesized voice vector Yi and yi as parameters
(Yi) and an evaluation expression consisting of the input speech vector R; E = W (yi) (R, Yi) ² value, that is, the magnitude of the inner product value of R and Yi is W (yi)
The value weighted by is calculated and output. Evaluation part 1
At 07, a plurality of drive vectors yi that make the value of the evaluation expression E larger among the drive vectors yi are obtained, and the index i thereof is output as a preliminary selection candidate.

The evaluation formula E has the effect of improving accuracy as compared with the conventional preselection evaluation formula in which the weighting coefficient W (yi) is not multiplied. Regarding the specific method of determining W (yi),
The third and subsequent embodiments will be described.

The preliminary selection candidates thus selected by the preliminary selection unit 104 are narrowed down to one candidate by the main selection unit 108,
The optimum drive vector X is available as output 109. However, as an exception, for the purpose of using delayed decision and the like, although the candidates are narrowed down even in the main selection, a plurality of candidates may remain.

(Embodiment 2) FIG. 2 shows the configuration of a speech coding apparatus according to another embodiment of the present invention. The present embodiment differs from the embodiment of FIG. 1 in that the input speech vector R is passed through the deconvolution operation unit 201 and then input to the evaluation formula calculation unit 106.

If the matrix representing the filtering by the synthesis filter 102 is H, then Yi = Hyi.
Therefore, the calculation of the inner product included in the evaluation formula E can be expressed as (R, Yi) = R ^t Hyi = (H ^t R, yi). This indicates that if H ^t R is calculated once before the search for the drive vector is started, the inner product value (R, Yi) can be obtained only by calculating the inner product of this value and yi when the drive vector is searched. . Therefore, it is not necessary to perform the filtering calculation during the search, and the calculation amount can be further reduced.

In FIG. 2, the input speech vector R is input to the deconvolution operation unit 201, and the deconvolution input speech vector H ^t R is generated. This convolutional input speech vector H ^t R is input to the evaluation formula calculation unit 106. on the other hand,
The drive vector yi is directly input to the evaluation formula calculation unit 106 and the weighting factor derivation unit 105. The evaluation formula calculation unit 106 calculates (H ^t R, yi), which is (R, Yi).
Therefore, the same preliminary selection candidate i as in FIG. 1 is obtained from the preliminary selection unit 104. In this case, the calculation amount is reduced as compared with the first embodiment by the amount that the filtering operation by the synthesis filter is not performed. This calculation amount reduction method can be applied to the following embodiments.

(Embodiment 3) FIG. 3 shows the configuration of a speech coder according to a third embodiment of the present invention. In the present embodiment, the drive vector generation unit 101 is composed of an index generation unit 301, an adaptive codebook 302, and a drive vector cutout unit 303. The adaptive codebook 302 stores past drive vector groups. In the drive vector cutout unit 303, the index generation unit 301
The drive vector yi is cut out from the past drive vector group stored in the adaptive codebook 302 and output based on the pitch period corresponding to the index i output from. The drive vector yi passes through the synthesis filter 102 and is input to the evaluation formula calculation unit 106 as a synthesized voice vector.

On the other hand, the past drive vector group C and the index i stored in the adaptive codebook 302 are input to the weighting coefficient deriving unit 105, and the weighting coefficient W (C, i) having these values as parameters is output. It The evaluation formula calculation unit 106 calculates and outputs an evaluation formula; E = W (C, i) (R, Yi) ² from the drive vector Yi, the weighting coefficient W (C, i) and the input voice vector R. The subsequent processing is the same as that in the first embodiment.

4A and 4B show examples of the adaptive codebook 302 and the weighting coefficient W (C, i). Weighting factor deriving unit 105
When the past drive vector group input from the adaptive codebook 302 has the waveform shown in FIG. 4A, the weights shown in FIG. 4B are calculated from this waveform based on the reciprocal of the average power of each block. Create a function W (C, i).

When the index i is input, the corresponding pitch cycle is determined, and when the pitch cycle is determined, the drive vector cutout unit 303 cuts out the drive vector from which time of the drive vector group stored in the adaptive codebook 302. Whether to come or not is decided. The weighting factor W (C, i) is calculated by using the cut-out time of this drive vector as a parameter in FIG.
It is obtained from the weighting factor graph of (b). Adaptive codebook 3
Since the contents of 02 do not change during the search of the drive vector,
If the weighting factor graph of FIG. 4B is created before the search is started, the weighting factor W (C,
i) is obtained by table lookup.

According to the present embodiment, even though the adaptive codebook 302 that changes with the input speech vector R is used in the drive vector generation unit 101, the weighting coefficient can be easily obtained during the search within the frame without requiring calculation. effective.

(Embodiment 4) FIG. 5 shows the configuration of a speech coding apparatus according to this embodiment. In the present embodiment, the drive vector generation unit 101 is obtained from the adaptive codebook 501 that stores past drive vector groups, the noise codebook 502 that stores a plurality of fixed vectors, and these codebooks 501 and 502. A gain circuit 503 for multiplying a vector by a gain
504. Further, a weighted synthesis filter 506, which is a recursive filter using a prediction coefficient obtained by performing a linear prediction analysis of the input speech and a perceptual weighting filter, is used as the voice synthesis unit, and is further weighted by the perceptual weighting filter as a target vector. The weighted input speech vector R obtained by subtracting the zero input response of the weighted synthesis filter in the internal state immediately after the end of the previous frame processing is used from the input speech of the current frame.

First, the adaptive codebook 501 is searched for a drive vector. Since the random codebook 502 is not used at this time, it is considered separately from the circuit. Preliminary selection unit 10
4, the weighted input speech vector R is first fetched, then the driving vector xi from the adaptive codebook 501 and the synthesized speech vector Xi obtained by passing this through the weighted synthesis filter 506. At this time, the gain of the gain circuit 503 is fixed to a constant (usually 1). In the preliminary selection unit 104, an evaluation formula using the weighting factor W (yi); E = W (xi) (R, Xi) ² (12) is calculated, and several driving vectors are calculated in descending order of this value. Leave as a preliminary selection candidate. How many drive vectors to leave as candidates for preliminary selection depends on the size of the adaptive codebook 501, the required coded speech quality, and the like, but 4 to
Sufficient quality is often obtained with about 16 candidates. The following equation is used as the weighting coefficient W (xi). This is the reciprocal of the power of the synthetic speech vector. Here, xi (n) represents the nth element of xi, and L represents the frame length.

[0039]

[Equation 1] In this way, the preliminary selection candidates selected by the preliminary selection unit 104 are narrowed down to one candidate by the main selection unit 108, and the first optimum drive vector X is obtained as an output. However, as an exception, for the purpose of using delayed decision and the like, although the candidates are narrowed down even in the main selection, a plurality of candidates may remain.

Next, an orthogonalization search is performed on the random codebook 502. At this time, adaptive codebook 501 is not used, and therefore it is considered separately from the circuit. In the pre-selection unit 104, the weighted input speech vector R previously fetched and the optimum driving vector x searched from the adaptive codebook 501 are passed through the weighted speech synthesis filter 506 to obtain the synthetic speech with the smallest distortion. The vector X is held, and the driving vector yi obtained from the random codebook 502 and the synthetic speech vector Yi obtained by passing the driving vector yi through the weighted speech synthesis filter 506 are input. The gain of the gain circuit 504 is usually fixed to a constant (normally 1).
The preselection unit 104 preselects several evaluation formulas for preselection in the orthogonalization search using the weighting coefficient W (yi); E = W (yi) (R, Yvi) ² (14) having a large value. Leave as a candidate. Yv
i is a synthetic speech vector Yi orthogonalized to X, and is specifically calculated by Yvi = Yi-{(Yi, X) / | X | ² } (15). The weight coefficient W (yi) is given by the reciprocal of the power of the drive vector as shown in the following equation.

[0041]

[Equation 2] Since the drive vector in the random codebook 501 is fixed,
It is not necessary to calculate equation (15) for each frame, (1) have it as table data in advance, (2) calculate it only once when the encoder is initialized, (3) the power of the drive vector itself It is possible to omit the calculation of the weighting coefficient W (yi) itself by a method such as designing in advance. (1) (2)
The method of (1) can be used regardless of the type if it is a fixed codebook, but it requires a memory for storing W (yi) by the number of drive vectors. The method (3) does not require a memory, but depending on the structure of the codebook, it may be impossible to use it because it is difficult to arrange the power in advance.

Finally, the preliminary selection candidates selected by the preliminary selection unit 104 are narrowed down to one candidate by the main selection unit 108, and the second optimum drive vector yi is obtained.

(Embodiment 5) In the present embodiment, E = W (yvi) (R, Yvi) ² is used as the evaluation formula of the preliminary selection in the orthogonal search in the fourth embodiment. This evaluation formula is the same as formula (12) when yvi and its synthesis filter output Yvi are replaced with xi and its synthesis filter output Xi. Therefore, the evaluation of the general preselection described above is essentially performed. It is an expression. In the evaluation formula for preselection of the orthogonalization search, yi is approximated to yi as a parameter of the weighting coefficient, but yvi is used in this embodiment, so the accuracy is improved accordingly. effective.

(Embodiment 6) In this embodiment, an overlapping codebook is used instead of the random codebook in the fourth embodiment. The overwrapping codebook is an example in which the method (3) described in the fourth embodiment cannot be used in the calculation of W (yi).
(Yi) can be calculated.

FIG. 6 shows the characteristics of the driving vector of the overlapping code length. The kth drive vector yk is
This is a form in which S elements are newly added to the head from the head of the (k-1) th drive vector yk-1 to the LS elements. Assuming that the power Qk-1 of yk-1 has already been obtained, it can be seen that Qk can be easily obtained from the relationship of the following equation.

[0046]

[Equation 3] In other words, if Qk-1 is given, the power of Qk is obtained by calculating the power of 2S elements and adding and subtracting twice, and W (yk) is obtained by taking the reciprocal of Qk. Actually S =
Since 2, L = 80, the power calculation for one drive vector needs to calculate the power of 80 elements when the overlapping structure is not used, whereas it is 4 elements when the overlapping structure is used. This reduces the calculation amount. For the sake of caution, if there is a memory, it is often practical to use the methods (1) and (2) described in the fourth embodiment.

(Embodiment 7) In this embodiment, a method for efficiently calculating W (yi) at the time of adaptive codebook search in Embodiment 4 by applying the method described in Embodiment 6 will be described.

FIGS. 7 and 8 show the adaptive codebook and the drive vector obtained from it. A past drive vector group is stored in the adaptive codebook, and the pitch cycle T
Is shorter than the frame length L (T <L), the section of length T cut out from the adaptive codebook is repeated as shown in FIG. 7 until the frame length is reached. When T> L, the section before T by L is taken out as it is as shown in FIG.

When calculating the power of a drive vector obtained from such an adaptive codebook, when T <L, the power P _T for one cycle is calculated, and this is repeated the number of times (2 in the figure). Just add them. However, in the end part (the part of T0 in the figure), a section less than one cycle occurs,
This part needs to be calculated separately. Alternatively,
A method may be considered in which this section is not included in the power calculation. That is, this is a method in which the average power per sample is regarded as P _T / T and (P _T / T) × L is regarded as the drive vector power. When the power is obtained in this way, this reciprocal becomes the weighting coefficient W (yi). Furthermore, P _T and P
_{Since T + 1} has a relationship of P _{T + 1} = P _T + C _{T + 1} ² (18), if W (yi) is calculated at the period T, then W (yi of T + 1 is calculated using the above equation. ) Can be easily obtained. When L <T, as is clear from FIG. 3, it is exactly the same as the overlapping codebook with the shift amount S = 1. However, since the content of the code length changes, the methods (1) and (2) described in the fourth embodiment cannot be used, and the equation (1) described in the sixth embodiment is used.
The method of utilizing the relationship of 7) is effective.

[0050]

As described above, according to the present invention, since the power of the drive vector is taken into consideration in the evaluation formula at the time of preselection of the drive vector, the precision of the preselection is improved as compared with the conventional method. As a result, the number of drive vector candidates to be passed to the main selection can be reduced while maintaining the quality of the coded speech, and the amount of calculation in the main selection can be reduced. Since the calculation amount in this selection occupies most of the calculation amount of the entire encoding device, as a result, the effect of greatly reducing the calculation amount of the entire encoding device can be expected.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a speech coding apparatus according to a first embodiment.

FIG. 2 is a block diagram showing the configuration of a speech encoding apparatus in which the amount of calculation is reduced by using deconvolution according to the second embodiment.

FIG. 3 is a block diagram showing a configuration of a speech coding apparatus using an adaptive codebook according to a third embodiment.

FIG. 4 is a diagram showing a method of obtaining a weighting coefficient in the third embodiment.

FIG. 5 is a block diagram showing a configuration of a speech coding apparatus according to a fourth embodiment.

FIG. 6 is a diagram showing the structure of an overlap codebook.

FIG. 7 is a diagram showing a structure of an adaptive codebook.

FIG. 8 is a diagram showing a structure of an adaptive codebook.

FIG. 9 is a schematic diagram of a CELP coding method.

[Explanation of symbols]

101 ... Driving vector generation unit 102 ... Synthesis filter 103 ... Input terminal 104 ... Preliminary selection unit 105 ... Weighting coefficient derivation unit 106 ... Evaluation formula calculation unit 107 ... Evaluation unit 108 ... Main selection unit 109 ... Optimal driving vector 201 ... Deconvolution operation Unit 301 ... Index generating unit 302 ... Adaptive codebook 303 ... Driving vector cutout unit 501 ... Adaptive codebook 502 ... Noise codebook 503 ... Gain circuit 504 ... Gain circuit 506 ... Weighted synthesis filter

(72) Inventor Susumu Kamiwa 1 Komukai Toshiba-cho, Saiwai-ku, Kawasaki-shi, Kanagawa Inside the Toshiba Research and Development Center, Inc. (72) Inventor Masahiro Oshikiri Komukai Toshiba-cho, Saiwai-ku, Kawasaki-shi, Kanagawa Incorporated company Toshiba Research and Development Center

Claims (3)

[Claims] 1. A drive vector generating means for generating a drive vector, a synthesizing means for inputting the drive vector generated by the drive vector generating means to generate a synthesized voice vector, and a drive vector generating means for generating the synthesized voice vector. And a pre-selection unit for selecting at least one drive vector from the drive vectors, and a main selection unit for selecting an optimum drive vector from the drive vectors selected by the pre-selection unit. The value obtained by weighting the size of the inner product of the target vector obtained from the input voice divided into unit periods and the synthesized voice vector with a weighting function having the drive vector generated by the drive vector generation means as a parameter is made larger. A speech coding apparatus characterized in that a driving vector to be selected is selected. 2. A drive vector generating means having a codebook, and extracting and generating one drive vector designated by a predetermined index from the codebook, and the drive vector generated by the drive vector generating means. A synthesizing means for inputting and generating a synthesized voice vector, a pre-selecting means for selecting at least one drive vector from the drive vectors generated by the drive vector generating means, and an optimum one from the drive vectors selected by the pre-selecting means. The main selection means for selecting a different driving vector, and the preliminary selection means sets the magnitude of the inner product value of the target vector obtained from the input voice divided into a predetermined unit period and the synthesized voice vector to the sign. The drive vector group stored in the book and the value weighted by the weighting function with the index as a parameter are A speech coding apparatus characterized by selecting at least one drive vector to be made larger. 3. A drive vector generating means for generating a drive vector, a synthesizing means for inputting the drive vector generated by the drive vector generating means to generate a synthesized voice vector, and a drive vector generating means for generating the synthesized voice vector. And a pre-selection unit for selecting at least one drive vector from the drive vectors, and a main selection unit for selecting an optimum drive vector from the drive vectors selected by the pre-selection unit. After obtaining the target vector and the optimum synthesized speech vector obtained from the input speech divided into unit periods, the orthogonalized vector obtained by orthogonalizing the synthesized speech vector generated by the synthesizing means with respect to the optimal synthesized speech vector, The drive vector generation means calculates the magnitude of the inner product of the orthogonalized vector and the target vector. A speech coding apparatus, characterized in that a driving vector that increases a value weighted with a weighting coefficient using the generated driving vector as a parameter is selected.