JPH0561499A - Voice encoding/decoding method - Google Patents

Abstract

PURPOSE:To curtail the arithmetic quantity by calculating a time inversion auditory sense weighing input sound signal vector and multiplying it by each pitch prediction residual vector of an adaptive code book generating a correlation value of both of them. CONSTITUTION:An arithmetic means 21 calculates a time inversion auditory sense weighting input sound signal vector <1>AAX from an input sound signal vector AX subjected to auditory sense weighing. Also, a multiplying part 22 multiplies the time inversion auditory sense weighting input sound signal vector <t>AAXC and each pitch prediction residual vector P of an adaptive code book 1 and generated a correlation value t(AP)AX of both of them. Subsequently, a filter arithmetic part 23 derives a self-correlation value <t>(AP)AP of a vector AP after auditory sense weighting reproduction of each pitch prediction residual vector P of the adaptive code book 1. Moreover, an evaluating party 10 selects an optimal pitch prediction residual vector Popt and gain bopt for minimizing power of an error signal E to the input sound signal vector AX subjected to auditory sense weighting, based on both correlation values.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coding / decoding method, and more particularly to a highly efficient speech coding / decoding method for compressing information of a speech signal by using vector quantization.

In recent years, a vector quantization method for compressing information while maintaining the quality of a voice signal has been used in a corporate communication system, a digital mobile radio system and the like. This vector quantization method is a codebook. Prediction weighting is applied to each signal vector in the (codebook) to create a reproduced signal, and the error power between the reproduced signal and the input audio signal is evaluated to determine the number (index) of the signal vector with the smallest error. However, there is an increasing demand for a method that is a more advanced version of this vector quantization method in order to further compress audio information.

[0003]

2. Description of the Related Art FIGS. 12 and 13 show a high-efficiency speech coding system called CELP (Code Excited LPC) using vector quantization. Among them, FIG. 12 shows sequential optimization CELP. Called, Figure 13 is a joint optimization CELP
The method is called. The symbols P, X, Y, C, and E used in the following description each mean a vector (however, in the drawings, they are shown by vector-specific symbols).

In FIG. 12, an adaptive codebook 1a stores an audio signal while adaptively changing an N-dimensional pitch prediction residual vector corresponding to N samples whose pitch period is delayed by one sample. Also, the fixed codebook 2 has 2 ^m of code vectors of non-periodic components other than the periodic components in the adaptive codebook 1a generated by using N-dimensional white noise corresponding to the same N samples. Only the pattern is fixed in advance.

First, a scalar A = 1 / A '(Z) (where A'is included in each pitch prediction residual vector P of the adaptive codebook 1a).
(Z) indicates a perceptual weighted linear prediction analysis filter) and a gain of 5 is added to the pitch prediction vector AP generated by perceptual weighting by the perceptual weighted linear prediction reproduction filter 3
Is multiplied by a gain b to obtain a pitch prediction reproduction signal vector bA
Generate P.

The pitch prediction reproduction signal vector bAP and the perceptual weighting filter 7 represented by A (Z) / A '(Z) (where A (Z) represents a linear prediction analysis filter) are perceptually weighted. And the input speech signal vector AX
Y is obtained by the subtraction unit 8, and the evaluation unit 10 calculates the following formula for each frame so that the power of the pitch prediction error signal vector AY becomes a minimum value: P = argmin (| AY | ² ) = argmin (| AX− bAP | ² ) ..., the optimum pitch prediction residual vector P is selected from the codebook 1a, and the optimum gain b is selected.

Further, each code vector signal C of the fixed codebook 2 of white noise is similarly perceptually weighted by the linear predictive reproduction filter 4 and generated with the gain 6 to the code vector AC after perceptual weighting reproduction. The gain g is multiplied to generate the linear prediction reproduction signal vector gAC.

Then, the linear prediction reproduction signal vector g
An error signal vector E between AC and the pitch prediction error signal vector AY is obtained by the subtraction unit 9, and the power of this error signal vector E is expressed by the following formula: C = argmin (| E | ² ) = argmin (| AY -GAC | ² ) ... The evaluation unit 11 selects the optimum code vector C from the codebook 2 for each frame and the optimum gain g so that the value becomes the minimum value.

The adaptation (update) of the adaptive codebook 1a is
The optimum driving sound source signal bAP + gAC is obtained by the addition unit 12,
This is a perceptual weighting linear prediction analysis filter (A '(Z))
This is performed by returning the signal to bP + gC in 13 and further delaying it by one frame in the delay device 14 as the adaptive codebook (pitch prediction codebook) of the next frame.

Thus, the sequential optimization CE shown in FIG.
In the LP method, the gains b and g are controlled separately, whereas in the simultaneous optimization CELP method shown in FIG.
P and gAC are added by the addition unit 15 and AX ′ = bAP +
gAC is obtained, and the subtraction unit 8 further obtains an error signal vector E with the perceptually weighted input speech signal vector AX from the filter 7 in the same manner as the above equation, and the evaluation unit 16 minimizes the power of this vector E. Code vector C
Is selected from the fixed codebook 2 and optimum gains b and g are obtained.
Are controlled simultaneously.

In this case, from the above equation, C = argmin (| E | ² ) = argmin (| AX-bAP-gAC | ² ) ...

The adaptation of the adaptive codebook 1a in this case is similarly performed for AX 'corresponding to the output of the adder 12 in FIG. Further, the filters 3 and 4 may be provided in common after the addition unit 15, and in this case, the inverse filter 13 is not necessary.

By the way, the actual codebook search is performed in two steps, that is, the search for the adaptive codebook 1a and the search for the fixed codebook 2. In the pitch search of the adaptive codebook 1a, the above equation is used. Even is done as shown in the equation.

That is, in the above equation, the gain g for minimizing the electric power of the vector E is obtained by partial differentiation: 0 = δ (│AX-bAP│ ² ) / δb = 2 ^t (-bAP) ( From AX-bAP), b = ^t (AP) AX / ^t (AP) AP. However, " ^t " shows a transposed matrix.

Therefore, in the pitch period optimization algorithm shown in FIG. 14, the auditory weighted input speech signal vector AX and each pitch prediction residual vector P of the adaptive codebook 1a are passed through the auditory weighted linear prediction reproduction filter 4. the resulting code vector AP is multiplied by the multiplication unit 41 generates a correlation value ^t (AP) AX of both, after perceptual weighting reproduction pitch prediction residual vector AP autocorrelation value ^t (
AP) AP is calculated by the multiplication unit 42.

Then, in the evaluation section 10, both correlation values ^t (A
P) AX and ^t (AP) Select the optimum pitch prediction residual signal vector P and gain b that minimizes the power of the error signal vector E = AY with respect to the perceptually weighted input signal vector AX according to the above equation based on AP ..

If the gain b is obtained for each pitch prediction residual signal vector P so as to minimize the above equation, and the quantization for this gain is performed in open loop, the ratio of the correlation value, It is equivalent to maximizing ( ^t (AX) AP) ² / ^t (AP) AP.

That is, Therefore, the second term on the right side should be maximized.

[0019]

In such a pitch search of the adaptive codebook 1a, the impulse response of the auditory weighting reproduction filter is convolved by the filter 4 with respect to each pitch prediction residual signal vector P of the adaptive codebook 1a. Therefore, the dimension of each of M (M = 128 to 256) pitch prediction residual signal vectors of the adaptive codebook 1a is set to N (normally N = 40).
, 60), the order of the perceptual weighting filter 4 is set to N _P (II
In the case of an R type filter, if N _P = 10), the multiplication unit 4
The calculation amount for 1 is the sum of the calculation amount N × N _P required for the perceptual weighting filter for each vector and the calculation amount N required for the inner product calculation of the vector.

In order to determine the optimum pitch vector P, this calculation amount is required for all of the M pitch vectors included in the adaptive codebook 1a, which causes a problem that the calculation amount becomes huge. there were.

Since the adaptive codebook 1a is updated by directly feeding back the optimum driving excitation signal of the past frame, the code vector component from the fixed codebook 2 is also included, as shown in FIG. The signal is fed back, and an aperiodic noise component that is not preferable for the adaptive codebook is superimposed, and the quality of the coded speech is deteriorated especially in the voiced sound section having a strong pitch period as a property of the driving excitation signal. There were also problems.

Therefore, according to the present invention, in the CELP-type speech coding / decoding system for performing long-term prediction by pitch period search by such an adaptive codebook, the calculation amount of pitch period search is minimized and the fixed codebook is used. The purpose is to prevent non-periodic noise components from leaking into the adaptive codebook.

[0023]

FIG. 1 is a block diagram of an optimum pitch vector P and a gain b of an adaptive codebook 1 in a speech coding system according to the present invention for solving the above problems. This is a conceptual illustration of the optimization algorithm, which corresponds to the improvement of the optimization algorithm of the conventional example shown in FIG.

According to the present invention, the adaptive codebook 1 is a sparse codebook having all zeros except for predetermined elements, and the optimum pitch vector b _opt P _opt selected by pitch search from the adaptive codebook 1 is sparse circuit. After being sparsed by 17, the optimum code vector g _opt C _opt selected by the codebook search from the fixed codebook 2 is added to the delay unit 14
It is updated by delaying by 1 frame and giving. The sparse circuit 17 can be made sparse based on the fixed threshold Th or the adaptive threshold Th according to the average signal amplitude of a predetermined number of samples.

Further, the calculating means 21 for calculating the time-reversed auditory-weighted input speech signal vector ^t AAX from the auditory-weighted input speech signal vector AX, the time-reversed auditory-weighted input speech signal vector ^t AAX and the adaptive codebook 1 are used. A multiplication unit 22 that multiplies each pitch prediction residual vector P by each other to generate a correlation value ^t (AP) AX between them, and a vector AP after the auditory weighting reproduction of each pitch prediction residual vector P of the adaptive codebook 1 A filter operation unit 23 for obtaining an autocorrelation value ^t (AP) AP, and an optimum pitch prediction residual vector P _opt that minimizes the power of the error signal E with respect to the perceptually weighted input speech signal vector AX based on both correlation values. And an evaluation unit 10 that selects the gain b _opt .

In contrast to the coding side as shown in FIG. 1, the decoding side of the present invention has the same sparse adaptive codebook 1 and fixed codebook 2 as the coding side, as shown in FIG. An optimal code vector b _opt P _opt obtained by multiplying the optimally selected pitch prediction residual vector P _opt in the adaptive codebook 1 by the optimal gain b _opt is provided by providing the sparsification circuit 17 and the delay unit 14. A code vector X which is sparsed by the sparsification circuit 17 and is added with an optimum code vector g _opt C _opt obtained by multiplying the optimum selected code vector C _opt of the fixed codebook 2 by an optimum gain g _opt. Is obtained through the linear predictive reproduction filter 200 and is given to the sparsification circuit 17. Also in this case, the sparse circuit 17 is made sparse by a constant threshold Th.
Alternatively, the adaptive threshold Th according to the average signal amplitude can be used as a reference.

Further, the sparsification circuit 17 in FIGS. 1 and 2 is as shown in FIGS. 3 and 4, respectively.
Optimal Pitch rather than to vector b _opt P _opt, the optimum pitch vector b _opt P _opt and the optimal code vector g _opt C _opt and may be provided with respect to the added combined value, in this case, Threshold value T corresponding to the ratio of the power of the optimum code vector g _opt C _opt to the total power
After the h is generated by the threshold value calculation circuit 18 and given to the sparsification circuit 17 to be sparsified, it is sent to the delay device 14.

In such a CELP system, as shown in FIG.
As shown in (a), the calculating means 21 may be configured to multiply the transposed matrix ^t A of the FIR auditory weighting filter matrix.

Alternatively, as shown in FIG. 5 (b), the calculating means 21 rearranges the input signals in reverse order on the time axis, performs IIR auditory weighting filter processing (1 / A '(Z)), and then repeats the time. It can also be configured by rearranging in reverse on the axis.

[0030]

First, in the CELP type speech coding system of the present invention shown in FIG. 1, since the adaptive codebook 1 is updated by the sparsified optimum driving excitation signal, the pitch prediction residue that is always stored. It is in a sparse (decimated) state in which the difference signal vector is zero except for a predetermined sample.

One autocorrelation value ^t (AP) AP to be given to the evaluation unit 10 is calculated in the same manner as in the conventional example shown in FIG. 14, but the correlation value ^t (AP) AX is the auditory sense. The weighted input audio signal vector AX is calculated by the calculating means 21.
Adaptive codebook 2 with sparse structure after conversion to ^t AAX
Of the pitch prediction residual signal vector P
Since it is obtained by applying to the sparse circuit 17, the multiplication is performed in a form in which the advantage of the adaptive codebook 1 sparsed by the sparse circuit 17 is used as it is (that is, in the form in which the portion of which the sample value is “0” is not multiplied). This can be performed and the amount of calculation can be reduced. This can be applied in exactly the same manner to both the sequential optimization method and the simultaneous optimization CELP method, and further to the pitch orthogonal optimization CELP method in which both are combined.

The sparse circuit 17 compares the signal amplitude of each sample with a threshold value, and replaces the sample value with zero for sample points that do not exceed the threshold value, thereby leaking aperiodic components into the adaptive codebook 1. Can be prevented.

Further, the sparsification circuit 17 is provided with an optimum pitch vector b _opt as shown in FIGS. 3 and 4, respectively.
The optimal pitch vector b, not for P _{opt
It is} provided for a value obtained by adding _opt P _opt and the optimum code vector g _opt C _opt, and the sparse circuit 17 calculates a threshold Th corresponding to the ratio of the power of the fixed codebook 2 to the total power. If the circuit 18 generates the signal, supplies it to the sparsification circuit 17 to make it sparse, and then sends it to the delay device 14, it is possible to further suppress the leakage of the aperiodic component into the adaptive codebook 1.

That is, when the voice is a voiced sound, a signal of a certain fixed period (pitch period) becomes large, and an amplitude difference from other aperiodic components becomes large. Conversely, when unvoiced,
The western part of the pitch period is almost eliminated, and the non-periodic component becomes dominant and the amplitude difference disappears.

Therefore, the threshold value of the sparse circuit 17 is determined by the difference between the pitch periodic component and the aperiodic component, that is, the signal power difference between the adaptive codebook and the fixed codebook (the ratio of the fixed codebook 2 power to the total power). By making it adaptively variable, leakage of non-periodic components into the adaptive codebook 1 can be reduced.

Further, in contrast to the encoding side as shown in FIG. 1, the decoding side of the present invention is the optimum selection from the adaptive codebook 1 notified from the encoding side as shown in FIG. An optimum code vector b _opt P _opt obtained by multiplying the pitch prediction residual vector P _opt by an optimum gain b _opt ,
Again the optimal code vectors obtained by multiplying the optimum gain g _opt optimally selected code vector C _opt fixed codebook 2 was informed from the encoding side g _opt C _opt
The adaptive codebook 1 is updated by obtaining the reproduction signal of the code vector X obtained by adding and through the linear prediction reproduction filter 200.

In this case, the calculation means 21 is
As shown in Fig. 5 (b), the input signals are rearranged in the reverse order on the time axis, and IIR auditory weighting filter processing (1 / A '
(Z)) and then rearranging them again on the time axis, the transposed matrix ^t of the FIR auditory weighting filter matrix as shown in FIG.
Compared to the case where it is configured by multiplying A, IIR and F
The amount of calculation is reduced due to the difference in IR.

[0038]

FIG. 6 shows an embodiment of the sparsification circuit 17 shown in FIGS. 1 and 2. In this embodiment, as shown in FIG. For a sample point having a value of, the input value is output as it is, and when it is less than or equal to the threshold Th, the input value is replaced with zero and sparsified.

Therefore, the sparsification circuit 17 in this case becomes a circuit having a center clipping characteristic as shown in FIG. 9B, and there are two possible methods for realizing such a center clipping circuit.

First, in the embodiment shown in FIG. 7, the value of each sample point of the input signal (optimum pitch prediction residual signal) is ranked from the one having the larger absolute value (signal amplitude), and the value is desired from the higher order. Up to the number of samples (corresponding to the constant threshold Th) are output as they are, and the other sample points are replaced with zero. As a result, the number of "non-zero sample points" (degree of sparseness) that directly affects the calculation amount of pitch search can be accurately controlled.

On the other hand, in the embodiment shown in FIG. 8, the average signal amplitude V _AV per predetermined sample is calculated for the input signal, and the value V _AV is multiplied by the coefficient λ to determine the threshold Th = V _AV · λ. However, the center clipping is performed using this threshold Th. In this case, the degree of sparseness of the adaptive codebook 1 changes somewhat depending on the nature of the input signal.
As compared with the embodiment described above, the amount of calculation required for ranking the sample points is unnecessary, and thus the amount of calculation can be reduced.

FIG. 9 shows an embodiment of the threshold value calculation circuit 18 shown in FIGS. 3 and 4. In this embodiment, the threshold value Th is fed back to the adaptive codebook in the optimum drive excitation signal. The power (electric power) | b _opt P _opt | ² and | g _opt C _opt | ² of each component of the pitch vector and the code vector is calculated (by vector inner product calculation), and the component power of the code vector | g _opt The ratio (k _C ) of C _opt | ² to the whole is calculated by the following equation. k _C = (| g _opt C _opt | ² / (| b _opt P _opt | ² + | g _opt C _opt | ² )) ^1/2 where 0 ≦ k _C ≦ 1.

[0043] Then, the threshold Th = λ / k _C as shown as the k _C of the function f (k _C) in FIG. 9 is determined. However, λ is a constant determined by experience.

On the other hand, the component power of the pitch vector | g
_The ratio (k _P ) of _opt C _opt | ² to the whole is as follows. k _p = ( _{│b opt} P _opt │ ² / ( _{│b opt} P _opt │ ² + │g _opt C _opt │ ² )) ^1/2 where k _C ² + k _p ² = 1.

Now, considering k _C and k _p ,
These values have a complementary relationship, and when the ratio of the pitch component in the driving sound source is large and the adaptive codebook can sufficiently follow the pitch periodicity of the input signal, the value of k _C becomes small. Grows on the contrary. Therefore, only the pitch component remains, other signal components are clipped, and the non-periodic component returned to the adaptive codebook is suppressed.

On the contrary, when the ratio of the pitch component is small and the adaptive codebook cannot follow the pitch periodicity of the input signal, the value of k _C becomes large and the threshold Th becomes small.
The optimum driving excitation signal component is directly returned to the adaptive codebook.

In this way, it is possible to control the feedback amount of the aperiodic component of the driving excitation signal used for updating the adaptive codebook according to the tracking state of the adaptive codebook.

As a method of evaluating the ratio of the pitch vector and the code vector, instead of the above k _C , a weighting synthesis filter is applied to each component, and k _C ′ is calculated as follows. You may ask. k _C '= (| g _opt AC _opt | ² / (| b _opt AP _opt | ² + | g _opt AC _opt | ² )) ^1/2

Further, such a threshold value Th can also be obtained from the value of k _{C by} a table lookup method.

FIG. 10 shows an embodiment of the arithmetic means used in the speech coding system according to the present invention shown in FIG. 5 (a), which is a FIR (finite impulse response) auditory weighting filter matrix. A and this matrix A
Assuming that the transposed matrix ^t A of N is a N-dimensional matrix that matches the codebook dimension number N shown in FIG. 10 (a), in the case of the CELP method shown in FIG. 1, the weighted input signal vector AX is shown in FIG. 10 (b). If it is as shown in, the transposed matrix ^t A is added to the weighted input signal vector AX.
Time-reversed auditory weighted input signal vector ^t A multiplied by
AX becomes as shown in FIG. 10 (c). In the figure, * indicates a multiplication code.

FIG. 11 shows an embodiment of the arithmetic means used in the speech coding system according to the present invention shown in FIG. 5 (b). First, the CELP system shown in FIG. 1 is used. In this case, if the weighted input signal vector AX is as shown in FIG. 11 (a) (the same as that shown in FIG. 10 (b)), the result obtained by reversing this on the time axis is shown in FIG.
It is the vector (AX) _TR shown in (b).

Then, this vector (AX) _TR is _converted into an IIR (infinite impulse response) type auditory weighted linear prediction reproduction filter A of the auditory weighting filter function 1 / A '(Z).
As a result, A (AX) _TR becomes as shown in FIG. 11 (c), for example.

In this case, the matrix A is a matrix obtained by returning the transposed matrix ^t A shown in FIG. 10 (a), and therefore, in order to restore the above A (AX) _TR , the rearrangement is performed in reverse on the time axis. When done, it becomes as shown in FIG. 11 (d), which is the same as the time-reversed auditory-weighted input signal vector ^t AAX shown in FIG. 10 (c).

Thus, the embodiment of FIG. 10 and FIG.
It can be seen that the example of FIG.

Although the filter matrix A is an IIR filter in the embodiment shown in FIG. 11, an FIR filter may be used. However, the use of FIR filters, but the total number of multiplications in the same manner as in the example of FIG. 10 is a N ^2/2 (and 2N moving operation), in the case of using an IIR filter, for example, 10-order linear prediction analysis If 1 then
It only requires 0N multiplications and 2N move operations.

[0056]

As described above, according to the present invention,
A sparse codebook with all zeros except for certain elements is used as the adaptive codebook, and the optimum pitch vector is updated by sparseizing with a sparser circuit to give the correlation value to the evaluation unit. The time-reversed auditory weighted input speech signal vector is calculated from the weighted input speech signal vector and multiplied with each pitch prediction residual vector of the adaptive codebook to generate the correlation value between them. It is possible to carry out multiplication in encoding and decoding while making the best use of the advantage of (3) as it is, and to reduce the amount of calculation.

Further, as a threshold for sparsification, of the driving excitation signals used for updating the adaptive codebook according to the tracking state of the adaptive codebook, the non-periodic component is fed back by the code vector from the fixed codebook. Since it is variable to control the amount, the periodicity is maintained more than the conventional one, and as a result, the encoded / decoded voice quality is applied to the voice having a drive source with a strong pitch periodicity such as voiced sound. Can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram conceptually showing an optimization algorithm for pitch search in a speech coding system according to the present invention.

FIG. 2 is a block diagram conceptually showing a reproduction algorithm of a voice decoding system according to the present invention.

FIG. 3 is a block diagram conceptually showing an optimization algorithm when the pitch search of the speech coding method according to the present invention is executed by another sparsification.

FIG. 4 is a block diagram conceptually showing a reproduction algorithm when it is executed by another sparsification of the speech decoding system according to the present invention.

[Fig. 5] Fig. 5 is a diagram conceptually showing a configuration example of a calculation means used in a voice encoding / decoding system according to the present invention.

FIG. 6 is a graph diagram conceptually illustrating an embodiment of a sparsification circuit used in a voice encoding / decoding system according to the present invention.

FIG. 7 is a graph diagram conceptually illustrating center clipping in the signal amplitude order of the sparsification circuit used in the speech encoding / decoding system according to the present invention.

FIG. 8 is a graph diagram conceptually illustrating center clipping due to an averaging threshold of a sparsification circuit used in a voice encoding / decoding system according to the present invention.

FIG. 9 is a diagram showing an embodiment of a threshold value calculation circuit used in the voice encoding / decoding method according to the present invention.

FIG. 10 is a diagram for explaining an embodiment of a calculation means used in the present invention.

FIG. 11 is a diagram for explaining another embodiment of the calculating means used in the present invention.

FIG. 12 is a block diagram schematically showing a general sequential optimization CELP method.

FIG. 13 is a block diagram schematically showing a general joint optimization CELP method.

FIG. 14 is a block diagram conceptually showing a conventional pitch search optimization algorithm.

FIG. 15 is a block diagram for explaining problems of the conventional method.

[Explanation of symbols]

 1 sparse (pitch period) adaptive codebook 2 fixed codebook 10 evaluation unit 14 frame delay unit 17 sparse circuit 18 threshold value calculation circuit 21 calculation unit 22 multiplication unit 23 filter calculation unit In the drawings, the same reference numerals indicate the same or corresponding portions.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Yoshihiro Sakai 1015 Kamiodanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture, Fujitsu Limited (72) Inventor Yoshinori Tanaka, 1015, Kamiodanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture, Fujitsu Limited

Claims (7)

[Claims] 1. Encoding is performed by using two codebooks, an adaptive codebook (1) and a white noise fixed codebook (2), to perform a pitch search / codebook search to find an optimum driving excitation signal. In the CELP-type speech coding method to be performed, the adaptive codebook (1) is a sparse codebook of all zeros except for predetermined elements, and the optimum pitch vector (b _opt P _opt ) is sparsified by a sparsification circuit (17), and then the fixed codebook
The input speech signal is updated by adding the optimum code vector (g _opt C _opt ) selected by the codebook search from (2) and delaying it by one frame with the delay device (14), and further weighted perceptually. Vector (A
X) time reversal auditory weighting input speech signal vector (
calculating means (21) for calculating ^t AAX) and the time-reversal auditory-weighted input speech signal vector ( ^t AA
X) and each pitch prediction residual vector (P) of the adaptive codebook (1) are multiplied to obtain a correlation value (( ^t (AP) AX) of the two.
And the autocorrelation value (( ^t (AP) A) of the vector (AP) after auditory weighting reproduction of each pitch prediction residual vector (P) of the adaptive codebook (1). A filter calculation unit (23) to be obtained and an optimum pitch prediction residual vector (P _opt ) that minimizes the power of the error signal (E) with respect to the perceptually weighted input speech signal vector (AX) based on both correlation values. And an evaluation unit (10) for selecting a gain (b _opt ), and a speech coding method. 2. The sparsification circuit (17) has a constant threshold value (Th).
The speech coding method according to claim 1, wherein the sparse conversion is performed with reference to. 3. The speech coding according to claim 1, wherein the sparsification circuit (17) performs sparsification based on an adaptive threshold (Th) corresponding to an average signal amplitude of a predetermined number of samples. method. Wherein said sparse circuit (17), said optimum rather than to the pitch vector (b _opt P _opt), said optimum pitch vector (b _opt P _opt) and the optimum code vector (g _opt C _opt ) and the optimum code vector (g
a threshold value (Th) corresponding to the power ratio of ( _opt C _opt ) is generated by the threshold value calculation circuit (18), given to the sparsification circuit (17) to be sparsified, and then sent to the delay device (14). The voice encoding system according to claim 1, wherein 5. The voice code according to claim 1, wherein the calculating means (21) is for multiplying a transposed matrix ( ^t A) of the FIR auditory weighting filter matrix. Method. 6. The computing means (21) rearranges the input signals in reverse on the time axis, and IIR auditory weighting filter processing (1
/ A '(Z)), and then reversely rearranged again on the time axis. 5. The speech encoding system according to claim 1, wherein 7. The same sparse adaptive codebook as the encoding side
(1), fixed codebook (2), sparsification circuit (17), delay device (1
4) and a calculating means (21), and the optimum gain obtained by multiplying the optimum selected pitch prediction residual vector (P _opt ) in the adaptive codebook (1) by the optimum gain (b _opt ). code·
The vector (b _opt P _opt ) is sparsed by the sparsification circuit (17), and the optimum selected code vector (C _opt ) of the fixed codebook (2) is multiplied by the optimum gain (g _opt ). Optimal code vector (g _opt C _opt )
7. The speech decoding system according to claim 1, wherein the code vector (X) obtained by adding and is obtained through a linear prediction reproduction filter (200) to obtain a reproduction signal.