JPH03189699A - Parameter quantizing system - Google Patents

Abstract

PURPOSE:To reduce the necessary calculation quantity, to curtail the number of arithmetic circuit, to miniaturize the circuit and to attain the reduction in cost by deriving the quantizing value of plural parameters by executing a tree search from a denominator value and a numberator value of an arithmetic expression for showing the parameteter. CONSTITUTION:In an encoding circuit shown in the figure, the calculation of a quantizing value of a gain parameter (rj) for searching a pitch period by a tree search is executed by a gain parameter calculating part 114 and a quantizing part 115, In the case of executing the calculation, the quantizing value is derived by executing the tree search as to a large/small relation of a quantization table r(i) of the value (rj) and the value (rj). The value (rj) is given by (rj)=(Xt, HB)/II HBj II, therefore, when it is deformed, and replaced with (rj)=bj/aj, the tree search is executed by positive/negative of delta=bj - r(i) aj, and delta = delta aj. That is, from a denomenator value (bj) and a numberator (aj), the quantizing value of the parameter (rj) can be derived. In such a manner, the number of circuits can be curtailed by a portion of the calculation quantity required for a division.

Description

[Detailed Description of the Invention] [Object of the Invention] (Field of Industrial Application) The present invention relates to parameters in an analysis-synthesis type encoding system used for compressing and transmitting information on a digital signal sequence such as an audio signal. It concerns the quantization method.

(Prior art) Analysis-synthesis encoding is a method that assumes a signal generation model and extracts and encodes parameters that minimize the error signal between the generated value and the observed value, and is currently receiving the most attention. This is one of the currently used encoding methods. In particular, in the field of audio encoding, CE L P (Code Excit
ed Linear Pre-former method or V
X C (Vector-Excited Codin
g) method is known as an excellent analysis-synthesis encoding method. For details of the CELP method, please refer to SM, R.
,5 Mr. chroeder and Mr. B, S, ^ta1...
“Code-Exclted Linear Pred
iction (CELP): Hlgh-Quality
Speech At Very Low Bit Ra
tes In Proc.

ICASSP, 1985. pp. 937-939.

The CELP method will be briefly explained below with reference to the drawings. FIG. 5 is a block diagram showing the basic configuration of the CELP method. In the figure, a signal sequence of an audio signal is introduced into a block extraction unit 12 via an input terminal 11, and is divided into long input audio signal vectors in this block extraction unit 12, with L sample values as one frame. Ru.

This input audio signal vector is then input to the LPG analysis section 13. The LPG analysis section 13 performs LPG analysis of the audio signal using an autocorrelation method, etc., and as a result, the LPG
Prediction parameters (ail(i-1, . . . , P)) are extracted. This LPG prediction parameter is supplied to the LPG synthesis filter 19. Note that P is the prediction order.

Further, the LPG analysis section 13 outputs an LPC prediction residual signal vector, and this LPG prediction residual signal vector is input to the pitch analysis section 14.

The pitch analysis unit 14 performs pitch analysis, which is a long-term prediction of speech, using the LPG prediction residual signal vector, and extracts the pitch period Tp and gain parameter b. Both the extracted pitch period and gain parameter are supplied to the pitch synthesis filter 18.

Next, the process of generating synthesized speech will be explained. The codebook 15 stores n white noise vectors with k dimensions (number of vector elements).

Here k is generally chosen such that L/k is an integer.

Each of the above white noise vectors is read out one by one from the codebook 15 and multiplied by the gain parameter generated from the gain parameter generation section 17 in the multiplier 16. Then, predetermined filter calculations are performed on each of the pitch synthesis filter 18 and the LPG synthesis filter]9, and
A synthetic speech vector is generated. At this time, the transfer function P (Z) of the pitch synthesis filter 18 and the transfer function A (Z) j of the LPG synthesis filter 19 are as follows.

P (Z)-1/ (1+bZ-”) ・・・(
2) The synthesized speech vector thus generated is sent to the comparator 2
0 and is compared with the input speech vector, which is the target vector. Then, the comparison output is passed through a weighted filter 21 and then introduced into a squared error calculation section 22, where the Euclidean distance Ej between the synthesized speech vector and the input speech vector is determined. The minimum distortion search unit 23 searches for the minimum value of Ej. This process is repeated for all n white noise vectors, thereby selecting the white noise vector number j that gives the minimum value of Ej.

In other words, the CELP method is characterized in that vector quantization is performed by using a codebook for the drive signal of the synthesis filter in the process of speech synthesis. Note that since the input speech vector is of length, this process is L
It will be repeated / times. In addition, the weighting in the figure indicates that the filter 21 is used to shape the spectrum of the error signal to reduce distortion that would be perceived by the human ear, and its transfer function is given by the following equation. .

W (Z)-H(Z/7)/H(Z)-(3)H
(Z) -A (Z) XP (Z) ・(4
) Note that when an input signal is actually encoded and transmitted using the CELP method, the LPG prediction parameter, pitch period, pitch gain parameter, codebook number, and codebook gain are each encoded and transmitted.

On the other hand, FIG. 6 is a block diagram showing another basic configuration of a speech encoding circuit to which the CELP method is applied. In this figure, the same parts as in FIG. 5 are given the same reference numerals.

In FIG. 6, weighting filter 21 has been moved from inside the search loop of codebook 15 to outside the loop. This is the pitch synthesis filter 18(7)P (Z)
P (Z/7), A of the LPG synthesis filter 19 (
By setting Z) to A (Z/γ), it is possible to reduce the amount of calculation while maintaining the same function.

In addition, the pitch synthesis filter 18 and the LPG synthesis filter 1
The initial memory in the filter operation of 9 does not affect the codebook search. That is, a pitch synthesis filter 24 and an LPG synthesis filter 25 having an initial memory are newly provided. A zero-input vector is output from this LPG synthesis filter 25 and supplied to a subtracter 26. In this subtracter 26, the weighted human speech vector outputted from the filter 21 is subtracted from the zero input vector. The comparator 20 compares the weighted human voice vector output from the subtracter 26 with the synthesized voice vector output from the LPG synthesis filter 19. That is, the comparator 20 detects the difference between the synthesized speech vectors output from the LPG synthesis filter 19, using the weighted human speech vector from which the 0-input vector has been subtracted as a target vector. Therefore, as a result, the initial memories of the pitch synthesis filter 18 and the LPG synthesis filter 19 are set to zero.

At the same time, this circuit performs a filter operation to generate a synthesized speech vector using the code vector and the next kxk
can be expressed as a product of the lower triangular matrix.

Here, k is the number of dimensions (number of elements) of the code vector of codebook 15, and h (1) (i −1,...
・k) is an impulse response of length k when the initial memory of H(Z/γ) is zero.

The speech synthesis vector output from the comparator 20 is 2
The multiplicative error calculation section 22 calculates the Euclidean distance MEj of the following equation, and the minimum distortion search section 23 detects its minimum value.

Ej-11Xt-γjHcjll...(6
) Here, Xt is the target input vector, Cj is the j-th code vector, and γj is the optimal gain parameter for the j-th code vector.

Find the above Ej and find the vector number j that gives its minimum value.
A flowchart for determining is shown in FIG.

By the way, in the above process, H(Z/γ
) and the optimum gain parameter γj, the characteristics of the encoder will generally be better if the vector number j is selected using the same quantization value used in the decoder. However, the gain parameter γj is γj − (Xt H, Cj ) / II HCj I
Since t-(1>), finding the quantized value requires one division and quantization for each γj (n pieces).
When performing the above operation using a processor such as Al Slgnal Processor, D steps are used for division and Q steps are used for quantization.
Assuming that steps are required, (D+Q)X
A calculation amount of n steps is required.

For example, in the case of D-50, Q-25, n=1024, 76,800 steps are required. In addition, in the entire flow of Fig. 7, if k-40°L/ is set to -4, then
The number of steps per frame is 330,716, the number of samples per frame is L - 160, and the sampling frequency of the input audio signal is 8 kl (when z, 18
．． The number of steps of about 5 MIPS is required.

On the other hand, as a method for improving the sound quality of the CELP method, a method called a closed-loop pitch prediction method or an adaptive codebook is known. Details of this method can be found in V.B., D.J., and Xrasl.
nskl and RH, Ketchui.

”Improved 5peech Quality
nd Efficient Vector Quanti
zation In 5ELP'' In Pro
c JCASSP, 198B, pp155-...1
58.

The CELP method, which employs a method called closed-loop pitch prediction or an adaptive codebook, will be briefly described below. The difference between the CELP method, which employs a method called a closed-loop pitch preparation or applied codebook, and the general CELP method, the configuration of which is shown in FIG. 6, lies in the method of pitch analysis. That is, in the method shown in FIG. 6, pitch analysis is performed using the LPG prediction residual signal vector output from the LPC analysis section 13. On the other hand, in the CELP method that employs a method called closed-loop pitch prediction or adaptive codebook, pitch analysis is performed in a closed-loop manner in the same way as codebook search.

In the CELP method, which is called a closed-loop pitch prediction or adaptive codebook, the drive signal for the LPG synthesis filter is delayed by a delay device that varies the pitch search range from a to b, thereby determining the drive signal for the pitch period Mj. Create a signal vector. Then, after multiplying this drive signal vector by a predetermined gain parameter, L
A synthesized speech vector is generated by filter calculation using a PC synthesis filter, and this synthesized speech vector is compared with a weighted human speech vector excluding the influence of the previous frame using a comparator. Then, the distortion Ej due to the squared distance which is the comparison output
, and further searches for the minimum value of this distortion Ej, and selects a drive signal vector that gives this minimum distortion Ej. In this way, a pitch search is performed using the weighted human voice vector, which excludes the influence of the previous frame, as the target vector. At this time, the above squared distance distortion Ej is Ej-11Xt-γjHBjll (8
). Here, Xt is the target vector, Bj
is the drive signal vector at pitch period j, γj is the optimal gain parameter for pitch period j, and H is the above-mentioned (
5) A lower triangular matrix h(
1) (i-1,-,k) is an impulse response of length k when the initial memory of A (Z/γ) is zero. FIG. 8 shows a flowchart for determining the square distance distortion Ej and determining the pitch period j that gives the minimum value.

By the way, in this process, as in the case of searching the codebook, the same quantization values used in the decoder are used for the parameter of H (Z/γ) that provides H and the optimal gain parameter γj. In general, the characteristics of the encoder will be better if the pitch period is selected according to the pitch period. However, the above gain parameter γj is γj - (Xt H, Bj )/1IHBj II...
- Since it is given by (9), in order to obtain the quantized value, one division and quantization are required for each gain parameter γj (a≦j≦b).

Therefore, when implementing the calculation to obtain the above quantized value using a DSP, etc., D steps are used for division, and Q steps are used for quantization.
Assuming that steps are required, (D+Q)X (b-
The number of calculation steps is required for a+1)xL/. For example, D-50°Q-25, ni-40, L-160, a-
For 20°b-147, 38400 steps are required. Further, in the entire flow, 103,472 steps are required per frame, and therefore, when the sampling frequency of the input audio signal is 8 kHz, the number of steps approximately 5.2 old PS is required.

(Problems to be Solved by the Invention) As described above, in the conventional encoding circuit applying the CELP method, the calculation of the quantized value of the codebook gain in the codebook search and the pitch calculation called closed-loop or adaptive codebook. Pitch in period search
Calculating each gain quantization value requires a large amount of calculation. For this reason, a large number of arithmetic circuits such as DSPs are required to perform real-time processing, resulting in an increase in the size and cost of the circuit configuration.

Therefore, the present invention focuses on the above-mentioned circumstances and reduces the amount of calculation required to calculate the quantized values of parameters such as codebook gain and pitch gain. The purpose of this invention is to provide a parameter quantization method that can be made smaller and lower in price.

The purpose is to provide

[Structure of the Invention] (Means for Solving the Problems) In order to achieve the above object, the present invention repeatedly calculates parameters expressed by an arithmetic expression including division every time a variable included in the above arithmetic expression is changed. However, in the parameter quantization method that selects the optimal parameter from among the multiple parameters obtained by this, a tree search is performed from the denominator value and numerator value of the above calculation formula to find the quantized value of the multiple parameters. This is what I did.

The present invention also provides a parameter quantization method using the above-mentioned tree search, in order to obtain a quantized value of a gain parameter for pitch search, and to obtain a quantized value of a gain parameter for codebook search. It is also characterized by its application to each.

(Operation) As a result, according to the present invention, the quantized values of parameters such as codebook gain and pitch gain are obtained by tree search from the denominator value and the numerator value. Therefore, even though the calculation formula for calculating the above parameters includes division, it is possible to calculate them without using division, thereby reducing the amount of calculation by the amount of calculation required for the above division. can do. Therefore, the number of arithmetic circuits such as SP required to calculate the quantized value of the parameter is reduced, and thereby the circuit configuration is made smaller and the price is reduced.

(Embodiment) FIG. 1 is a block diagram showing the basic configuration of a speech encoding circuit to which a parameter quantization method according to an embodiment of the present invention is applied.

In the figure, a signal sequence of an audio signal is input to a block extraction section 102 via an input terminal 101. In the block extraction section 102, the input audio signal is divided into length-length input signal audio vectors, with L sample values as one frame.

This divided input speech vector is processed by the LPG analysis section 10.
3. The LPG analysis unit 103 performs LPG analysis of the input speech vector using an autocorrelation method, etc., and thereby calculates the LPG prediction parameter (at) (i=
1. ..., p) are extracted. Note that P is the predicted order. The extracted LPG prediction parameters are quantized by quantization section 104 and then input to LPG synthesis filter 105 . This LPG synthesis filter 105 creates a zero input response vector.

The input speech vector outputted from the block extraction section 102 is also supplied to a weighted filter 106 . The weighting filter 106 is provided outside of the codebook search and pitch period search to reduce the computational complexity of these search loops. This weighting is determined by the filter 106 as W CZ')-(Z/γ)/A (Z)
...(lO), and here, in order to reduce the distortion that would be perceived by the human ear, the input speech vector is weighted according to the LPG prediction parameter (al). will be carried out. Note that A (Z) in the above equation 10) is A (Z) = 1/1 + Lai Z-' .
1ll). This weighted input speech vector is input to a subtracter 107. This subtracter 107 subtracts the zero input response vector created by the LPG synthesis filter 105 from the weighted human voice vector. This applies to each LPG synthesis filter 1 of the pitch period search loop and codebook search loop, which will be described later.
This is to set the initial memory of 09.121 to zero. At the same time, this also means that the filter operation of the LPG synthesis filter that takes the drive signal vector as input in the pitch period search or the filter operation of the LPG synthesis filter that takes the code vector as input in the codebook search is the drive signal vector or code vector, respectively. and the next kX
It is possible to express it by multiplying k with a lower triangular matrix.

Here, k is the number of dimensions (number of elements) of the drive signal vector and code vector, and is generally selected so that L/k is an integer. Also h (f) (f −1.

..., k) is an impulse response of length k when the initial memory of A (Z/γ) is zero.

Now, the encoding circuit of this embodiment includes a drive signal food book 108, an LPG synthesis filter 109, and a multiplier 110 for multiplying by a gain parameter for pitch period search.
, a comparator 111 , a square error calculation section 112 , and a minimum distortion search section 113 . The drive signal codebook 108 stores drive signal vectors corresponding to pitch period j.

This drive signal vector is created as follows. That is, in the multiplier 116, the optimal drive signal vector of the previous frame output from the drive signal codebook 108 of the previous frame is multiplied by the gain parameter for searching for the optimal pitch period of the previous frame. Also multiplier 1
At step 18, the code vector output from the codebook 120 is multiplied by a gain parameter for codebook search. And these multipliers 116°11
The vectors output from 8 are mutually added in an adder 117 to become a filter drive signal, and the signal is passed through a delay device 119. This delay device 119 has a pitch period search range a −
The delay amount is configured to be variably set in accordance with b. Therefore, in the delay device 119, the pitch period j (a≦j≦b) is determined based on the filter drive signal.
A drive signal vector is created according to.

The drive signal vector output from the drive signal codebook 108 is subjected to filter calculation to generate a synthesized speech vector in the LPG synthesis filter 109, and then
A multiplier 110 multiplies the signal by a gain parameter for pitch period search, and then supplies the signal to a comparator 111. In this comparator 111, the synthesized speech vector output from the multiplier 110 is subtracted from the weighted human speech vector from which the influence of the previous frame has been removed, which is output from the subtracter 107, and the difference vector is 2 Multiplicative error calculation unit 11
2 will be introduced. The square error calculation unit 112 calculates the square distance distortion Ej from the vector of the difference, and the minimum value of the square distance distortion Ej is detected by the minimum distortion search unit 113.

In other words, in this pitch period search loop, the weighted human voice vector from which the influence of the previous frame has been removed is set as the target vector, and the drive signal vector when the squared distance distortion is minimized, that is, the optimal pitch period, is searched. It is done. At this time, the square distance distortion Ej is determined by the following calculation formula.

Ej-11Xt-γjHBjll...(1
3) Here, xt is the target vector, Bj is the pitch period j
The drive signal vector when , γj is a gain parameter for searching for the optimum pitch period for pitch period j, and H is the kxk lower triangular matrix given by the above equation (12).

FIG. 2 is a flowchart showing the above pitch period search procedure. In the figure, first, in step S21, MAX-0 is initialized. Further, in step S22, the value of t is set to one of 1 to L/.

Next, in step S23, HBa is calculated by a normal matrix-vector product operation, and from this calculated HBa, aa-11HBal is calculated in steps S24 and S25, respectively.
l and ba - (Xt, HBa) are determined.

Then, γa is determined from the above aa and ba in step S2B, and further, the square distance strain Ea is determined in step S27. The calculation formula for Ea is expressed by the following formula.

Ea -2ba Xγa -7a'' xaaWhen the square distance distortion Ea is determined, in step S28 Ea>
It is determined whether the value is MAX or not, and if the result of this determination is that Ea is larger than the value of MAX, then in step S29 TP -
a, MAX-Ea, 7=7a is set, and after this setting, the process moves to step S30.

On the other hand, if Ea is smaller than the value of MAX in step S28, the process directly proceeds to step S30.

Once the initial element HBa is determined, each of the other elements is then repeatedly calculated recursively in steps S30 to S37. That is, in step S30, the value of j is a+
Each time one is set in the range of 1 to b, step S3
1, HBj is recursively determined based on the previous element HBj-1. Then, from this obtained HBj, step S32. aj-llHBjll respectively in S33
and bj - (Xt, HBj) are calculated, and γj is determined from these values in step S34.

The method for calculating the quantized value of γj will be explained in detail later.

Subsequently, in step S35, Ej =
The squared distance distortion Ej is calculated using the formula 2bj xγj−γj2xaj.

Based on this calculated value, in step S3B, the above Ej
>MAX, and if YES, step S37 TP-j, MAX-Ej.

γ−γj is set, and after this setting, the process returns to step S30. On the other hand, if the result of the determination in step 52g is NO, the process directly returns to step S30. Above step S3
The processes from 0 to S37 are repeatedly executed until the value of j becomes b.

Then, for one value of t, all squared distance distortions E
When the calculation of j is completed, the process returns to step S22 and the value of t is changed to the next value, and in this state, the above process for calculating Ej is repeated.

Then, the search for the pitch period ends when the calculation of Ej is completed for all the values of t (1 to L/K).

By the way, when searching for the pitch period j that gives the minimum value of the squared distance distortion Ej,
In general, the characteristics of the encoder will be better if the pitch period is selected using the same quantization value used in the decoder for the parameter of (Z/γ) and the gain parameter γj for searching for the optimal pitch period. Become. However, the gain parameter γj for pitch period search is γj = (Xt, HBj)/IIHBj II
...(14), so in order to find the quantized value, we need 1 for each gain parameter γj.
Requires multiple divisions and quantization.

However, the encoding circuit of this embodiment uses a tree search quantization method to obtain the quantized value of the gain parameter γj for pitch period search without performing division. In other words, when calculating the gain parameter γj for pitch period search by tree search, γj
It is possible to obtain the quantized value by performing a tree search for the magnitude relationship between the quantization table γ(1) (i-1, . . . , L) and the gain parameter γj. However, L-2M
M is the number of quantization bits. That is, δ−γj−γ
(1) ... The quantized value can be obtained by performing a tree search based on the sign of δ given by (15).

Further, (Xt, HBj) in the above equation (14) is expressed as bj, II HBj II is expressed as aj, and thereby the gain parameter γj is expressed as γj - bj /aj...(
By replacing (6) with (6), the main search can be performed based on the sign of ρ given by the following equation.

ρ-bj-γ(1) ・aj...
(17) −bj−(γj−δ)・aj −bj−bj+δ・aj −δ・aj ...(18)
This is because, as can be seen from equation (17), when aj is always positive, the positive and negative signs of δ and ρ match.

In other words, the (15th
) By replacing the judgment statement of equation (17),
The quantized value of the gain parameter γj for pitch period search can be directly determined from the denominator value and numerator value.

In the encoding circuit shown in FIG. 1, the gain parameter calculation unit 114 and the quantization unit 115 calculate the quantized value of the gain parameter γj for pitch period search using this tree search. The gain parameter calculation unit 114 includes:
A weighted human voice vector from which the influence of the previous frame has been removed, which is output from the subtractor 107, that is, a target vector;
The synthesized speech vectors output from the LPG synthesis filter 109 are respectively input, and based on these vectors, calculations are performed to obtain quantized values by the above-mentioned main search.

Figure 4 shows the gain parameter γj for this pitch period search.
A flowchart for finding the quantized value of by tree search is shown. In the figure, first, steps S61, S62,
In S63, L-2'i -L/2, m-i, respectively
is initialized, and further in step S64 m −m /
2 is set. Then, in step S85, ρ-bj-γ(1)·aj is calculated, and from this calculation result, in step 986, ρ>0
It is determined whether or not. If ρ is determined to be positive in this determination, it is set to i−i+m in step S67, while ρ
If it is determined that 1-1-m is negative, in step 368
is set to Then, in step S69, it is determined whether or not the number has reached m-5l, and the processes from steps 364 to S89 are repeated until the number has reached m-1.

When m-1 is reached, ρ-b is again determined in step S70.
A calculation of j-γ(1)·aj is performed, and based on the calculation result, it is determined in step S71 whether ρ>0. If ρ is determined to be positive in this determination, it is set to j−j+1 in step S72,
On the other hand, if it is determined that ρ is negative, then j−
j-1.

Subsequently, in step S74, the calculation σ-γ(1)·aj-bj is performed, and based on the result of this calculation, it is determined in step S75 that 1σ1≦1ρ1.

If the result of this determination is YES, it is set to ij in step 37B, and then γj - γ(1) is set in step S77. On the other hand, the answer in step S75 is N.

If it is determined that γj-7(1) is set, the process directly proceeds to step S77. In this way, the calculation for determining the quantized value of the gain parameter γj is completed.

Therefore, if the quantized value of the gain parameter γj for pitch period search is determined in this way, the following effects can be obtained. In other words, if Q steps are required for quantization, the number of steps required to obtain the quantized value of the gain parameter γj is Q・(ba+1)・L/, and the number of steps is equal to the amount of calculation required for division. will be reduced. For example, Q-25, ni-40, L-1
In the case of 60 degrees a-20, b-147, the number of steps is 12,800, and the number of calculation steps per frame is 77,872 in the entire flow of the pitch collection search loop. Therefore, the number of calculation steps is 75% compared to the conventional method shown in the flowchart in Figure 8.
When the sampling frequency of the input audio signal is 8 kllz, the number of steps is approximately 3.9 MIPS.

On the other hand, the encoding circuit of this embodiment includes a codebook 120, an LPG synthesis filter 121, a multiplier 122, a comparator 123, a square error calculation section 124, and a minimum distortion search section 1.
It has a codebook search loop consisting of 25 codebooks. In the codebook 120, n code vectors composed of white noise are stored in advance. The code vector output from the codebook 120 is
It is input to the LPG synthesis filter 121. This LPG synthesis filter 121 performs a filter operation on the code vector to generate a synthesized speech vector,
The synthesized speech vector obtained by this operation is multiplier 1
After being multiplied by a second gain parameter for code vector search in step 22, the signal is input to a comparator 123. In this comparator 123, the comparator 11 of the pitch period search loop
The weighted human speech vector from which the influence of the previous frame and the influence of pitch have been removed, which is output from the multiplier 122, is compared with the synthesized speech vector which is output from the multiplier 122, and the error is calculated as a squared error. 124. This squared error calculation unit 124 calculates 2 from the above error.
A squared distance distortion Ej is determined, and the minimum value of this squared distance distortion Ej is detected by the minimum distortion search unit 125.

That is, in this search loop, the weighted human voice vector from which the influence of the previous frame and the influence of pitch output from the comparator 111 of the pitch period search loop have been removed is set as a target vector, and the codebook 1 is
Among the 20 code vectors, the code vector number j whose distortion Ej due to the squared distance of the error is the smallest
The calculation for selecting is performed repeatedly. The calculation formula is expressed by the following formula.

Ej-11Xt-γjHcjll...(1
9) Here, X is a weighted human voice vector from which the influence of the previous frame and the influence of pitch have been removed,
In other words, the target vector, Cj is the j-th code vector,
γj is the optimal gain factor for the j-th code vector, n is the number of code vectors, and H is a lower triangular matrix with the number of dimensions kXk of the code vector given by the above equation (I2).

FIG. 3 is a flowchart showing the above pitch period search procedure. In the figure, first, in step S41, MAX-0 is initialized. Subsequently, in step S42, MCI, which is an initial element, is obtained by a normal matrix-vector product operation, and in step S43, al-IIHcIII is obtained from the above-mentioned HCI.

Once the initial element is determined, operations are then repeated to recursively determine other elements. That is, each time one value of j is set in the range of 2 to n in step S44, the previously calculated H C is set in step S45.
Element HCj is determined based on j-1. Then, aj-11HCjll is obtained from this element HCj in step S4B. After HCj is determined for all j, in step S47, each time a value of t is set in the range of 1 to L/, X is calculated in step S48.
t''H is determined. Each time this Xt''H is determined, bj-(XtH, C1)2 for each position of j is determined in steps S49 to S54.

Then, in step S51, a quantized value γj of the gain parameter is determined from this bj and the above aj, and further, in step S52, a squared distance distortion Ej is calculated using the following arithmetic expression.

Ej -2bj xγj -γj 2xaj Then, based on this calculated value, it is determined in step S53 whether Ej > MAX, and if it is determined as YES, step S54
TEIN-jSMAX-Ejγ-γj are respectively set. On the other hand, if the determination in step 353 is No, the process directly proceeds to step S54, where the IN-j is determined.

MAX-Ej and γ-γj are respectively set. When the determination of the magnitude of the squared distance distortion Ej due to all bj determined based on one XtH is completed, step S4 is performed.
Returning to step 7, the value of t is incremented, and then the calculation in steps S48 to S54 and the determination of the magnitude of the squared distance distortion Ej are repeated.

Thus, the search for the optimal codebook with the minimum squared distance distortion Ej is completed.

By the way, in the above codebook search, when searching for the pitch period j that gives the minimum value of the squared distance distortion Ej, the parameter of H (Z/γ) that gives H and the optimal code are used. Generally, the characteristics of the encoder will be better if the pitch period is selected using the same quantization value used in the decoder as the gain parameter γj for book search.

However, the gain parameter γ for this codebook search is
j is γj − (Xt, HCj) / If HCj
Since it is given by It-(20), obtaining the quantized value requires one division and quantization for each gain parameter γj.

However, the encoding circuit of this embodiment uses a tree search quantization method to obtain the quantized value of the gain parameter γj for codebook search without performing division. That is, when calculating the gain parameters γ,j by tree search, the magnitude relationship between the quantization table γ(1)(i-1,...,L) of γj and the gain parameter γj is calculated by tree search. It is possible to determine the quantized value. However, L-2' and M are the number of quantization bits. In other words, δ−γj−γ(1) ...(2
The quantized value can be obtained by tree search based on the sign of δ given in 1).

Further, (Xt, HCj of the above equation (17)
) is expressed as bj, If HCj II is expressed as aj, and thereby the gain parameter γj is expressed as γj-bj/aj...(22
), tree search can be performed using the sign of ρ given by the following equation.

ρ-bj-γ(1)・aj...(2
3) −bj−(γj−δ)・aj −bj−bj+δ・aj −δ・aj ...(24)
This is because, as can be seen from equation (20), when aj is always positive, the positive and negative signs of δ and ρ match.

That is, the (18th
) By replacing the judgment statement of equation (20),
The quantized value of the gain parameter γj for codebook search can be directly determined from the denominator value and numerator value.

In the encoding circuit shown in FIG. 1, the calculation of the quantized value of the gain parameter γj for codebook search by tree search is performed as follows:
This is performed by a gain parameter calculation section 126 and a quantization section 127. The gain parameter calculation unit 126 includes a weighted human voice vector from which the effects of the previous frame and pitch have been removed, output from the comparator 111, that is, a target vector, and a synthesized voice vector output from the LPG synthesis filter 121. are respectively input, and based on these vectors, the above-mentioned tree search is performed to obtain the quantized value. The procedure for calculating the quantized value of the gain parameter γj for the codebook search using this tree search is the same as the procedure explained for the quantization of the gain parameter for the pitch period search (Fig. 4), so the explanation will be omitted. do.

In this way, the gain parameter γj for codebook search
When finding the quantized value of the gain parameter γj, it is also found by tree search from the numerator and denominator, so if Q steps are required for quantization, the number of steps required to obtain the quantized value of the gain parameter γj is QXrl step, and the amount of calculation required for division is reduced. For example, Q=25. 256 for n-1024
00 steps. In addition, in the entire flow, k-
40, if L/ is set to -4, 27951 per frame
The number of calculation steps is 6.

Therefore, the number of calculation steps is reduced to 84% compared to the conventional method shown in the flowchart of FIG. 7, and the sampling frequency of the input audio signal is reduced to 84%.
At kHz, the number of steps is approximately 14.0 old PS.

As described above, in this embodiment, the quantized values of both the gain parameters for pitch period search and the gain parameters for codebook search can be determined directly by tree search from the denominator and numerator. Therefore, although the equations for each gain parameter described above include division, the quantized value can be obtained without performing this division. Therefore, the amount of calculation can be reduced by the amount that the above-mentioned division can be omitted, and thereby the number of arithmetic circuits such as DSP required to calculate the quantized value of the gain parameter can also be reduced.

Therefore, the circuit configuration can be made smaller and cheaper.

It should be noted that the present invention is not limited to the above-described embodiments, and includes, for example, a procedure for calculating a quantized value of a gain parameter, a circuit configuration thereof, a procedure for searching a pitch period, a circuit configuration thereof, and a codebook. The search procedure, its circuit configuration, etc. can also be modified in various ways without departing from the gist of the present invention.

[Effects of the Invention] As detailed above, according to the present invention, parameters expressed by an arithmetic expression including division are repeatedly calculated each time a variable included in the arithmetic expression is varied, and a plurality of In the parameter quantization method that selects the optimal parameter from among the parameters, the codebook The amount of calculation required to calculate the quantized values of parameters such as gain and pitch gain can be reduced.
Thereby, it is possible to provide a parameter quantization method that can reduce the number of arithmetic circuits, thereby making the circuit smaller and cheaper.

[Brief explanation of drawings]

1 to 4 are for explaining a parameter quantization method according to an embodiment of the present invention, and FIG. 1 is a block diagram showing the basic configuration of a speech encoding circuit to which the same method is applied. Figures 2 to 4 are flowcharts showing the calculation procedure and contents of the same circuit, and Figures 5 and 6 are flowcharts showing the calculation procedure and contents of the circuit, respectively.
The figures are block diagrams showing the basic configuration of speech encoding circuits that apply different conventional parameter quantization methods.
FIGS. 7 and 8 are flowcharts showing the operational procedures of the circuits shown in FIGS. 5 and 6, respectively. 101... Audio signal input terminal, 102... Block extraction section, 103... LPG analysis section, 104...
Quantization unit, 105...LPG synthesis filter, 106.
...Weighting is a filter, 107... Subtractor, 108.
・・Drive signal code book, 109°121...LP
G synthesis filter, 110,116°118.122...
・Multiplier, 111, 123... Comparator, 112, 12
4... Square error calculation section, 113.125... Minimum distortion search section, 114... Gain parameter calculation section for pitch period search, 115... Quantization section for pitch period search, 119. ...Delay device, 120...Codebook, 1
26... Gain parameter calculation unit for codebook search, 127... Quantization unit for codebook search.

Claims (3)

[Claims] (1) The quantized value of the parameter expressed by the arithmetic expression including division is repeatedly calculated every time the variables included in the above arithmetic expression are varied, and the optimal quantized value of the multiple parameters obtained by this is calculated. A parameter quantization method for selecting quantized values, characterized in that the quantized values of the plurality of parameters are determined by performing a tree search from the denominator value and numerator value of the arithmetic expression. (2) Repeatedly calculate the quantized value of the gain parameter for pitch period search expressed by the arithmetic expression including division every time the variable included in the above arithmetic expression is varied, and thereby obtain the quantization value of the multiple gain parameters. In the parameter quantization method for selecting an optimal quantization value from among the quantization values, the quantization value of the plurality of gain parameters is determined by performing a tree search from the denominator value and the numerator value of the arithmetic expression. Parameter quantization method. (3) Repeatedly calculate the quantized value of the gain parameter for codebook search expressed by the arithmetic expression including division each time the variable included in the above arithmetic expression is varied, and thereby obtain the quantization value of the multiple gain parameters. In the parameter quantization method for selecting an optimal quantization value from among the quantization values, the quantization value of the plurality of gain parameters is determined by performing a tree search from the denominator value and the numerator value of the arithmetic expression. Parameter quantization method.