JPH06230797A - Voice signal processor and telephone system - Google Patents

Abstract

PURPOSE:To reduce an arithmetic amount and to efficiently provide the approximate value of a desired correct solution without generating malfunction. CONSTITUTION:In the case of performing numerical conversion from a linear predictive coefficient to a line spectrum pair frequency in the case of voice signal encoding at an approximate value calculation part 15 by approximate calculation according to the Newton-Raphson method, the line spectrum pair frequency of a preceding frame from a preceding frame line spectrum pair frequency storage part 19 is applied as an initial value of the approximate calculation by the Newton-Raphson method, and the approximate value of the solution to be provided by the Newton-Raphson method is controlled by using an approximate value difference control part 20, infinite loop avoidance control part 21 and initial value change control part 22.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice signal processing device and a telephone device, and more particularly to a voice signal processing device and a telephone device which simplifies the numerical conversion of a linear prediction coefficient into a line spectrum versus frequency during voice encoding. It relates to the device.

[0002]

2. Description of the Related Art In mobile communication terminals such as car telephones, portable telephones and pocket type telephones, for example, a voice signal coding apparatus for compressing voice signal information in order to improve frequency utilization efficiency and reduce power consumption. Or a decryption device is used.

The linear prediction coefficient often used in the field of such speech signal coding is a coefficient of a transfer function of a formant prediction filter that removes short-term prediction redundancy of speech. This transfer function A (z) is given by the following equation (1).

[0004]

[Equation 1]

In the equation (1), a _i is a linear prediction coefficient.

The transfer function of the formant synthesis filter used in speech decoding is the transfer function A described above.
It is given by the reciprocal of (z), that is, the following equation (2).

[0007]

[Equation 2]

In a general speech coding system, a 10th-order formant filter is often used. That is, the order P = 10. Therefore, in the following description, the case of P = 10 will be described.

Further, in a general speech coding system, speech is sampled at 8 kHz and quantized, and the speech is quantized for 20 ms.
Each frame, that is, every 160 samples, is divided into one frame, the autocorrelation function of the voice in the frame is calculated for each frame, and the linear prediction coefficient a ₁ to
a ₁₀ is being calculated. Therefore, 10 linear prediction coefficients are calculated every 20 ms.

Further, in a certain kind of speech coding system, the linear prediction coefficient for each frame is further converted into a line spectrum versus frequency in order to reduce the required amount of information after coding. This line spectrum vs. frequency is disclosed, for example, by Sozo Saito, Kazuo Nakata, "Basics of Speech Information Processing", Ohmsha Co., Ltd., pp. 137 to 143 of the document published on November 30, 1981. There is. This line spectrum vs. frequency has the advantage of good quantization characteristics and good interpolation characteristics, but the drawback of using this is that the conversion method from the linear prediction coefficient is complicated. It means that the quantity will increase.

Here, some specific examples of the conversion method from the linear prediction coefficient to the line spectrum versus frequency are known,
For example, pages 246 to 2 of the above-mentioned document "Basics of Speech Information Processing"
On page 57, an example of a program for obtaining a calculation formula and LSP (pair of line spectra) is disclosed. A specific example of the above conversion method will be described below.

As a premise, it is assumed that the transfer function A (z) of the formant prediction filter is given by the following equation (3). A (z) = 1-a 1 z -1 - ··· -a 10 z -10 ··· (3) In this equation _{(3), a i (1 ≦ i ≦} 10) is a linear prediction coefficient . Using this A (z), P _A (z) and Q _B (z) are defined as follows, respectively.

[0013]

[Equation 3]

However, p _i = −a _i −a _11-i 1 ≦ i ≦ 5 (6) q _i = −a _i + a _11-i 1 ≦ i ≦ 5 (7)

The line spectrum vs. frequency is (4) above.
In the following two equations obtained by transforming the equation and the equation (5), that is, in the equations (8) and (9), ω =
It is known that there are a total of 10 solutions, each of which exists between 0.0 and ω = 0.5.

[0016]

[Equation 4]

Here, the values of p'and q'in these equations (8) and (9) are calculated by the following equations from the values of p and q. _{p '0 = q' 0 =} 1 ··· (10) p 'i = p i -p' i-1 ··· (11) q 'i = q i + q' i-1 ··· (12)

The line spectrum versus frequency characteristic is 2 if the formant synthesis filter is stable.
The solution of two equations, that is, the above equations (8) and (9), is
This means that they are staggered as shown in FIG. In FIG. 6, with respect to the frequency (angular frequency) ω on the horizontal axis, the curve representing P (ω) above is shown in A, and the curve representing Q (ω) above is shown in B, respectively. The frequency value at the zero-cross point of is the solution.

That is, as is apparent from FIG. 6, 10 solutions are 0.0 <ω ₁ <ω ₂ <ω ₃ <ω ₄ <ω ₅ <ω ₆ <ω ₇ <ω
_{When 8} <ω ₉ <ω ₁₀ <0.5, ω ₁ , ω ₃ , ω ₅ , ω ₇ , and ω ₉ are P
It is a solution of (ω) = 0, and ω ₂ , ω ₄ , ω ₆ , ω ₈ , ω ₁₀
Is the solution for Q (ω) = 0.

Since it is impossible to solve each of the five solutions of this equation by modifying the ordinary formula, it is necessary to calculate the solution approximately by another method. The Newton-Raphson method is generally well known as a method for approximating and obtaining a solution.

[0021] Newton - The Raphson method, when the appropriate first j approximation of the solution of continuous equations with x _j in the section of interest, the function value f (x _j) at x _j and its first derivative The j + 1-th approximated value x _{j + 1} obtained by calculating f ′ (x _j ) which is a function value is obtained by the following equation (13).

[0022]

[Equation 5]

[0023] In this equation (13), x _{j + 1} rather than _{x j | x j + 1 -x} j | only closer to the true solution, by repeating _{this, | x j + 1 -x j} | When is sufficiently small, the approximate value is assumed to have converged to the solution.

The Newton-Raphson method will be applied to the conversion of linear prediction coefficients into line spectrum versus frequency. ω corresponds to x _j , and P (ω) and Q (ω) corresponding to f (x _j ) are represented by the above equations (8) and (9), respectively. P ′ (ω) and Q ′ (ω) differentiated by ω are respectively

[0025]

[Equation 6]

[0026] These correspond to the above f '(x _j ). Therefore, P (ω) and Q (ω), P '(ω) and Q'
A new approximation can be calculated by calculating (ω).

The point to be noted here is that P
If there are multiple solutions within the target range, as in the equations of (ω) and Q (ω), all solutions can be efficiently used unless the initial value of the approximate value of the solution is set. It means that you cannot find it. Solution of this equation ω
_In order to calculate _{1 to} ω ₁₀ without omission, for example, a method whose procedure is shown in the flowchart of FIG. 7 can be considered. This procedure will be described below. Further, the concept of the procedure of the flowchart of FIG. 7 is shown by writing the initial values for the calculation of the above P (ω) and Q (ω) expressions in the graph of FIG. 6 with arrows.

In step S41 of FIG. 7, the linear prediction coefficients a _{1 to} a ₁₀ calculated by the external device are input.
In the next step S42, (6), (7), (1
0), (11) and (12), the linear prediction coefficients a _{1 to} a ₁₀ to (8) and the coefficients p ′ _{1 to} p ′ _{5 of the} expression (9) and
to calculate the q _'1 ~q' _5. In the next step S43, the control variable i is initialized to 1 (i = 1), and in step S44, the frequency ω is initialized to 0 (ω = 0).

Next, in step S45, the value of P (ω) is calculated by the above equation (8), and the positive / negative sign thereof is held. In the next step S46, ω is set to a certain constant value Δω
Increase by _STEP , that is, ω = ω + Δω _STEP . In the next step S47, the value of P (ω) is calculated by the above equation (8). In step S48, the above step S4
The positive / negative sign of P (ω) calculated in step 7 is used in the above step S4.
It is determined whether the sign is inverted by comparing with the sign of the value calculated in 5. When it is determined that they have not been inverted (NO), the process returns to step S46. Inverted (YE
When it is determined to be S), there is a solution between ω at that time and ω immediately before that, and the process proceeds to step S49.
In this step S49, P (ω) in the above equation (8) and P '(ω) in the above equation (14) are calculated using the value of ω at this time, and the approximate value of the new solution is obtained by the above equation (13). And set it as a new ω. In this case, the above equation (13) can be expressed as ω = ω−P (ω) / P ′ (ω).

Next, in step S50, step S49
It is determined whether or not the approximate value obtained in step 1 has converged to a solution sufficiently. If not, return to step S49,
Using this approximate value, the above equation (13) is calculated again.
By repeating this, if it is determined in step S50 that the convergence is achieved, the process proceeds to step S51. In step S51, the i-th solution ω _{i is set} to the obtained ω (ω _i = ω),
In the next step S52, i is incremented by 1 (increment, i = i + 1).

Next, in steps S53 to S60,
This is a part for calculating the solution for Q (ω) in the above equation (9) instead of calculating the solution for P (ω) in steps S45 to S52.

That is, in step S53, the above (9)
The value of Q (ω) in the equation is calculated and its positive and negative signs are retained,
In step S54, ω is increased by the constant value Δω _STEP (ω = ω + Δω _STEP ), and in step S55, the value of Q (ω) for that value ω is calculated. Next step S56
The sign of Q (ω) calculated in step S53 is compared with that calculated in step S53 to determine whether it is inverted, and if not inverted, the process returns to step S54, and if it is inverted, ω at that time and its Since there is a solution with the immediately preceding ω, the process proceeds to step S57. In step S57, Q (ω) and Q '(ω) are calculated with ω at that time, and ω = ω-Q (ω) / Q' (ω) is obtained as an approximate value of the new solution, and the new ω is obtained. And In step S58, it is determined whether or not the new approximate value has sufficiently converged to a solution. If it has not converged, the process returns to step S57, and if it has converged, the process proceeds to step S59. Ω _i = ω is set in step S59, and i is incremented by 1 in step S60 (i = i + 1)
Then, it progresses to step S61.

In step S61, it is determined whether or not all the solutions to be sought, in this example, 10 solutions have been sought, that is, i> 10. If NO, the process returns to step S45. If YES, the process proceeds to step S62, and the values of the line spectrum versus frequency of the ₁₀ solutions ω _{1 to} ω 10 are output.

The schematic configuration of the apparatus for calculating the line spectrum versus frequency from the linear prediction coefficient by such a method is as shown in FIG. 8, for example.

In FIG. 8, the linear prediction coefficient storage unit 3
1 is the linear prediction coefficients a _{1 to} a calculated by the external device.
Enter ₁₀ and keep that value. The linear prediction coefficients a _{1 to} a ₁₀ stored in the linear prediction coefficient storage unit 31 are sent to the coefficient conversion unit 12. In the coefficient conversion unit 32, from the linear prediction coefficients a _{1 to} a ₁₀ to the above (8), according to the above equations (6), (7), (10), (11) and (12),
(9) Factor p _'1 ~p' ₅ of formula, q _'1 ~q' ₅ is calculated.

The coefficient p _'1 ~p' ₅ of the calculated above by the coefficient conversion unit 32 (8) is stored and held is sent to the coefficient storage unit 33, the coefficient of the equation (9) q _'1 to q ' ₅ is sent to the coefficient storage unit 34 and stored and held. (8) Formula calculation unit 3
5 calculates the above equation (8) from the value ω held in the frequency storage unit 43 using the coefficients p ′ _{1 to} p ′ ₅ given by the coefficient storage unit 33, and obtains the above P (ω). Find the value of. The formula (9) calculating unit 36 calculates the formula (9) from the value ω held in the frequency storage unit 43 using the coefficients q ′ _{1 to} q ′ ₅ given by the coefficient storage unit 33. , The value of Q (ω) is calculated.

The sign inversion detection unit 37 determines and detects whether the sign of the value of P (ω) is inverted. The sign inversion detection unit 38 determines and detects whether the positive or negative sign of the value of Q (ω) is inverted. The approximate value calculation unit 39
From the value ω held in the frequency storage unit 43, using the coefficients p ′ _{1 to} p ′ ₅ given by the coefficient storage unit 33,
The approximate value ω of the new solution is obtained by the above equations (8), (14) and (13). The approximate value calculation unit 40 uses the above-mentioned coefficients q ′ _{1 to} q ′ ₅ given from the coefficient storage unit 34 from the value ω held in the frequency storage unit 43 to perform the above (9).
Approximate value ω of the new solution by equations (15) and (13)
Ask for. The convergence determination unit 41 determines whether the new approximate value ω obtained by the approximate value calculation unit 39 has sufficiently converged. The convergence determination unit 42 also determines whether the new approximate value ω obtained by the approximate value calculation unit 40 has sufficiently converged. The frequency increasing unit 44 increases the value of the frequency ω stored in the frequency storage unit 43 by Δω _STEP . Further, the line spectrum-to-frequency output unit 45 holds the values of ₁₀ line spectrum-to-frequency ω ₁ to ω 10 obtained by the above-mentioned Newton-Raphson method and outputs them to an external device.

[0038]

In such a method, in order to find the value of ω at which the signs of P (ω) and Q (ω) are inverted while increasing the value of the frequency ω by a constant amount Δω _STEP. P (ω) and Q including the cos function with a large amount of calculation
Since the equation (ω) needs to be calculated, Newton −
There is a drawback that the amount of processing at the stage of searching the initial value given to the Rafson method is large, and as a result, the amount of processing in the whole numerical conversion processing from the linear prediction coefficient to the line spectrum versus frequency is large.

As a measure for reducing the processing amount, Δω
By increasing the _STEP value, the above P (ω) and Q
Although it is possible to reduce the number of calculations for (ω), for example, when ω ₁ and ω ₃ are close to each other, ω ₃ is skipped and converges to ω ₅ when searching for the next initial value of ω _1. However, increasing the value of Δω _STEP causes a malfunction, and thus this measure is not a good measure.

In view of the above situation, the present invention provides a voice signal processing apparatus and a telephone apparatus capable of reducing the processing amount in the whole numerical conversion processing from a linear prediction coefficient to a line spectrum versus frequency without causing a malfunction. With the goal.

[0041]

In order to solve the above-mentioned problems, the present invention provides a speech signal processing apparatus for performing a process of converting a linear prediction coefficient used for coding a speech signal into a line spectrum versus frequency, Coefficient conversion means for calculating the coefficient of the equation showing the conversion from the linear prediction coefficient to the line spectrum pair, the coefficient storage means for holding the coefficient output from this coefficient conversion means, and one on the time axis A frequency storage means for holding the value of the line spectrum versus frequency of the previous speech frame, and a new approximate solution of the above equation using the value of the line spectrum versus frequency stored in this storage means as the initial value of the approximate value of the solution. It is characterized by having a calculating means obtained by the Rafson method.

Further, in the speech signal processing device having such characteristics, the absolute value of the difference between the new approximate value obtained by the Newton-Raphson method once and the original approximate value by the calculating means is above a certain value or more. So that the new approximation value is not restricted by the approximation value difference limiting means, and the difference value between the new approximation value obtained by the previous Newton-Raphson method and the original approximation value by the above calculating means and the current Newton-Raphson An infinite loop avoidance control means for detecting that the Newton-Raphson method enters the infinite loop and avoiding the infinite loop from the value of the difference between the new approximate value obtained by the method and the original approximate value; It is first judged whether the approximate value obtained by the calculation means has converged to a solution adjacent to the solution to be obtained, and the initial value of the Newton-Raphson method is changed according to the judgment result. It is characterized by having a value change control means.

That is, in a voice signal processing device for converting a linear prediction coefficient used for encoding a voice signal into a line spectrum versus frequency, for example, as shown in FIG. The previous frame line spectrum-frequency storage unit 19, which is a storage unit that holds the line spectrum-frequency value of the frame, and the difference Δ between the new approximation value and the original approximation value obtained by one Newton-Raphson method.
Approximate value difference limiting unit 20 which is means for limiting a new approximate value so that the absolute value of ω does not exceed a certain value Δω _MAX.
And the value of the difference Δω between the new approximation value obtained by the previous Newton-Raphson method and the original approximation value and the current Newton −
It is detected that the Newton-Raphson method enters an infinite loop from the value of the difference Δω between the new approximate value obtained by the Rafson method and the original approximate value, and in that case the difference between the new approximate value and the original approximate value Δω The infinite loop avoidance control unit 21 which is a control means for avoiding entering the infinite loop by changing the new approximate value ω so as to decrease the value of are approximations detects that omega has converged if the the solution omega _{i + 2} or omega _i-2 adjacent to the have the solution omega _i Seek better that case further it high frequency line-spectral pair frequency axis Is a control means for determining whether it is the adjacent solution ω _{i + 2} or the lower adjacent solution ω _i-2 , and in that case, changing the initial value of the Newton-Raphson method according to the determination content. An initial value change control unit 22 is provided.

According to the telephone device of another invention,
In a telephone device that encodes an input voice signal by a voice signal processing unit and then modulates it into a transmission signal, the voice signal processing device having the above-described characteristics is used as the voice signal processing unit.

[0045]

According to the above construction, P (ω) and Q (ω) are increased while increasing the value of ω by a constant amount Δω _STEP as in the conventional example.
The value of the line spectrum versus frequency of the previous frame can be used as it is as the initial value of the approximation value of the Newton-Raphson method without using the complicated procedure of searching for the value of ω at which the sign of is inverted. It is possible to complete the conversion of the linear prediction coefficient to the line spectrum versus frequency with a very small amount of processing. This is the line spectrum vs. frequency
This is because the correlation with the value of the previous frame is relatively strong. In most cases, the desired solution can be easily obtained by setting the value of the previous frame as the initial value of the Newton-Raphson method.

Also, by combining the approximate value difference limiting means, the infinite loop avoidance control means, and the initial value change control means, a linear spectrum is obtained from the linear prediction coefficient without causing a malfunction and with a smaller processing amount than the conventional example. You can perform numerical conversions to pairs.

Further, by applying to a telephone device,
The amount of calculation for converting the linear prediction coefficient into the line spectrum versus frequency can be reduced, and the configuration can be simplified and power consumption can be reduced.

[0048]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a schematic configuration of an embodiment of the present invention. In this embodiment, in order to perform numerical conversion from the linear prediction coefficient, which is the coefficient of the transfer function of the formant prediction filter for removing the short-term prediction redundancy of the speech, to the line spectrum versus frequency, the above (6), (7),
After converting the coefficients using the equations (10), (11), and (12), P (ω) = 0 and Q (ω) = in the above (8) and (9).
The solution of 0 is obtained by the Newton-Raphson method in which the approximate value calculation shown in the equation (13) is repeated.

In FIG. 1, the linear prediction coefficient storage unit 11
Are linear prediction coefficients a _{1 to} a ₁₀ calculated by an external device.
Enter and keep its value. The linear prediction coefficients a _{1 to} a ₁₀ stored in the linear prediction coefficient storage unit 11 are sent to the coefficient conversion unit 12. In the coefficient conversion unit 12, the linear prediction coefficients a _{1 to} a _{10 are} converted to the above (8), (8), according to the above equations (6), (7), (10), (11) and (12).
(9) Factor p _'1 ~p' ₅ of formula, q _'1 ~q' ₅ is calculated.

The coefficient p _'1 ~p' ₅ the calculated equation (8) by the coefficient conversion unit 12 is stored and held is sent to the coefficient storage unit 13, the coefficient of the equation (9) q _'1 to q ' ₅ is sent to the coefficient storage unit 14 and stored and held. Approximate value calculator 15
Is based on the value of the line spectrum pair of the previous frame held in the line spectrum pair frequency storage unit 19 and uses the coefficients p ′ _{1 to} p ′ ₅ from the coefficient storage unit 13 to perform the above (8).
Approximate value ω of the new solution by equations (14) and (13)
Ask for. The approximate value calculation unit 16 uses the coefficients q ′ _{1 to} q ′ ₅ from the coefficient storage unit 14 based on the value of the line spectrum pair of the previous frame held in the previous frame line spectrum pair frequency storage unit 19. Then, the above equation (9), (15)
The approximate value ω of the new solution is obtained by the equation (13). The convergence determination unit 17 determines whether the new approximate value ω obtained by the approximate value calculation unit 15 has sufficiently converged. The convergence determination unit 18 also determines whether the new approximate value ω obtained by the approximate value calculation unit 16 has sufficiently converged.

Here, the previous frame line spectrum pair frequency storage unit 19 holds the frequency value of the line spectrum pair in the immediately preceding speech frame on the time axis, and the line spectrum pair of this previous frame is held. By using the value of the frequency as the initial value of the approximate value calculation by the Newton-Raphson method, it is possible to reduce the calculation amount and improve the efficiency.

When the approximate value calculation units 15 and 16 perform the approximate value calculation by the Newton-Raphson method as described above, the approximate value difference limiting unit 20, the infinite loop avoidance control unit 21, and the initial value change control unit. 22 makes it possible to control the approximate value of the solution and efficiently obtain the approximate value of the correct solution without causing a malfunction.

That is, the approximate value difference limiting unit 20 determines that the absolute value of the difference Δω between the new approximate value obtained by the Newton-Raphson method once and the original approximate value by the approximate value calculating units 15 and 16 is constant. Limit the new approximation so that the maximum value Δω _MAX is not exceeded. This certainly exists even when the value of the line spectrum vs. frequency changes significantly compared to the previous frame, so for example, a malfunction that does not converge to the desired solution by just making the solution of the previous frame the initial value Is taken into consideration.

A specific case of this will be described with reference to FIG. When P (ω) and P ′ (ω) (or Q (ω) and Q ′ (ω), and so on) are calculated from the line spectrum frequency value ω _i (point a in FIG. 2) of the previous frame, If the value of | P '(ω) | is small, that is, if the slope of P (ω) is small, the new approximation value moves to a range where it can no longer converge to the desired solution. In FIG. 2, the approximate value (point c) obtained by the first calculation of the equation (13) by the Newton-Raphson method is far from the solution (point b) to be obtained, and the second time In some cases, the value further approaches the approximate value (point d) and the third approximate value (point e), and eventually exceeds the range of ω, that is, 0.0 <ω <0.5. In order to prevent such a situation, it is effective to limit the difference Δω between the original approximation value and the new approximation value within the range of a certain maximum value Δω _MAX . For this reason, the approximate value difference limiting unit 20 prevents the absolute value of the approximate value difference Δω from exceeding the constant value Δω _MAX .
You are limiting the new approximation.

Next, the infinite loop avoidance control unit 21 obtains the value of the difference Δω between the new approximation value obtained by the previous Newton-Raphson method and the original approximation value and the newton-Raphson method this time. From the value of the difference Δω between the new approximation and the original approximation, it is detected that the Newton-Raphson method enters an infinite loop, and when it is detected, the difference between the new approximation and the original approximation is calculated. By changing the new approximation ω so as to decrease the value Δω, control is performed to avoid entering the infinite loop.

This is because the difference Δω between the original approximate value and the new approximate value is set to a certain constant value Δω as in the approximate value difference limiting section 20.
_{In consideration of the fact} that when limiting to _MAX , there may be a case where an infinite loop state in which two approximate values (points a and b) come and go as shown in FIG. 3 may occur. is there. In order to prevent such an infinite loop situation, it is necessary to provide a means for holding the history of fluctuations in the approximate value of the Newton-Raphson method and detecting the state of falling into an infinite loop. In addition, it is also necessary to provide means for controlling the value of the new approximation so as to avoid the situation where an infinite loop is detected.
Therefore, the infinite loop avoidance control unit 21 controls to avoid entering the infinite loop.

By providing the approximate value difference limiting unit 20 and the infinite loop avoidance control unit 21, the approximate value is 0.0 <ω <
It surely converges to one of the solutions in the range of 0.5. However, this is still insufficient, and there is still the possibility of converging to adjacent solutions. In other words, there is a possibility that the approximate value will converge to the solution of ω ₃ when trying to obtain ω ₁ . Therefore, a means to determine whether the converged approximate value has really converged to the solution to be obtained, or has converged to the adjacent solution, and if it is found that the converged approximate value has converged to the adjacent solution, A means for determining whether it is a high-direction adjacent solution or a low-direction adjacent solution on the frequency axis is required. In addition, when it is found that the solution converges to the adjacent solution in the high direction on the frequency axis, the value of the line spectrum pair in the previous frame is used as it is, and the initial value of the approximate value is set to an appropriate value. It is necessary to have a means to decrease, and it is necessary to have a means to increase the initial value of the approximate value to an appropriate value when it is found that the solution converges to the adjacent solution in the lower direction on the frequency axis. Then, the approximate value of the desired solution can be reliably obtained by making the approximate value converge again using the changed initial value of the approximate value.

In consideration of such a point, the initial value change control unit 22 determines a solution adjacent to the solution ω _i for which the approximate value ω obtained due to the improper initial value is to be obtained. ω
It is detected that it has converged to _{i + 2} or ω _i-2 , and in that case it is either the higher adjacent solution ω _{i + 2} or the lower adjacent solution ω on the line spectrum vs. frequency axis.
It is determined whether _i-2 , and in that case, the initial value of the Newton-Raphson method is controlled according to the content of the determination.

By combining all the means as described above, that is, the front frame line spectrum-to-frequency storage unit 19, the approximate value difference limiting unit 20, the infinite loop avoidance control unit 21, and the initial value change control unit 22, a malfunction is caused. It is possible to perform the numerical conversion from the linear prediction coefficient to the pair of line spectra without causing the above and with a smaller processing amount as compared with the conventional example.

In the configuration of FIG. 1, the line spectrum pair frequency output unit 23 has ten line spectrum pair values ω _{1 to} ω obtained by the operation in each of the above configurations.
It is provided for storing and storing ₁₀ and outputting to an external device.

Next, the operation of the above-described embodiment will be described with reference to the flowchart shown in FIG.

In FIG. 4, in the first step S11, the linear prediction coefficients a _{1 to} a ₁₀ are input and stored in the linear prediction coefficient storage unit 11 of FIG. Step S12, the from the linear prediction coefficients a ₁ ~a ₁₀ by the coefficient conversion section 12 (8), obtained by calculating the coefficients _{p '1 ~p' 5, q} '1 ~q' 5 of (9) . In the next step S13, the above (8),
According to equation (9), the value when ω = 0, that is, P (0),
The value of Q (0) is calculated and held. Next step S
In 14, the variable i representing the solution number is initialized to 1 (i = 1).

In the next step S15, the initial value ω of the approximate value calculation of the Newton-Raphson method is used as the initial value ω of FIG.
The value of the corresponding i-th solution ω _{i in} the speech frame preceding by one frame on the time axis from the previous frame line spectrum-to-frequency storage unit 19 is substituted. In the next step S16,
The Newton-Raphson method calculation formula ((13) above)
The term of f (x) / f '(x) in the equation) is calculated. That is, depending on whether i is an odd number or an even number, Δω =
Calculate P (ω) / P '(ω), and when it is even, Δω = Q
Calculate (ω) / Q '(ω). In the next step S17, it is determined whether or not the absolute value of the calculated Δω exceeds a certain upper limit value Δω _MAX (| Δω |> Δω _MAX ). If YES (exceeds), the process proceeds to step S25, and if NO (does not exceed), the process proceeds to step S18 to replace the new approximate value ω with ω-Δω (ω = ω-Δω). The calculation in the above steps S16 and S18 corresponds to the approximate value calculation of the above equation (13).

Next, in step S19, it is determined whether the new approximate value ω has sufficiently converged to a solution. If not converged (NO), the process returns to the step S16 and the above P (ω) / P '(ω) or Δω = Q (ω) / Q' (ω)
To calculate. If converged (YES), the process proceeds to the next step S20, and whether the converged approximate value ω is actually converged to the solution ω _i to be obtained, or the adjacent solution ω _{i + 2} or ω _i- Determine if it has converged to ₂ . Specifically, the value of P (0) or Q (0) calculated in step S13, the value of i, and P '
It can be determined from the relationship between the values of (ω) or Q '(ω).

When YES is determined in the step S20, that is, when the solution converges to the adjacent solution ω _{i + 2} or ω _i-2 , the process proceeds to a step S28, and when NO, the process proceeds to a step S21. In step S21, ω _i = ω, and in the next step S22, i is incremented by 1 (i = i +
1). In the next step S23, it is determined whether i exceeds 10 (i> 10). If NO (i ≦ 10), the process returns to step S16. If YES (i> 10), the process proceeds to step S24, and the obtained ω ₁ to ω
The line spectrum to frequency value of ₁₀ is output, and the process ends.

When YES in step S17 (when the absolute value of Δω exceeds the predetermined upper limit value ω _MAX ), in step S25, the new approximation value obtained by the previous Newton-Raphson method and the original From the value of the difference Δω with the approximate value and the value of the difference Δω between the new approximate value obtained by the Newton-Raphson method and the original approximate value, Newton-
Determines whether the approximation calculation by the Rafson method enters an infinite loop. If NO in this step S25 (do not enter the infinite loop), the process proceeds to step S26 and Δω
The absolute value of is limited to the above constant value Δω _MAX (| Δω | = Δω
_MAX ), the process proceeds to step S18. If YES (into the infinite loop), the process proceeds to step S27 and the absolute value of Δω is limited to 1/2 of Δω _MAX (| Δω
| = Δω _MAX / 2), the process proceeds to step S18. That is, if the loop does not enter the infinite loop, the new approximation ω is changed so that the absolute value of the difference Δω between the new approximation and the original approximation becomes Δω _MAX. For example, by changing the new approximation value ω such that the absolute value of the difference Δω between the new approximation value and the original approximation value is half of Δω _MAX , it is possible to avoid entering the infinite loop.

Next, in step S28, when it is determined that the solution converges to the adjacent solution (YES) in step S20, the solution is either the higher adjacent solution ω _{i + 2} or the lower one on the frequency axis. Determine if it is an adjacent solution ω _i-2 . Specifically, the determination can be made by comparing the value of the previous solution ω _i-1 which has already been calculated and the magnitude.
If it is the lower adjacent solution ω _i-2 , the process proceeds to step S29, the initial value of the approximate value is increased to an appropriate value, and the process returns to step S16. If the higher adjacent solution ω _{i + 2} ,
In step S30, the initial value of the approximate value is reduced to an appropriate value, and the process returns to step S16.

Next, as an embodiment of the telephone device according to the present invention, an example of a telephone device to which the above-mentioned audio signal processing device is applied will be described with reference to FIG.

The telephone device shown in FIG. 5 is used as a mobile communication terminal such as a so-called digital cellular telephone, and the voice is converted into an electric signal by the microphone 51 on the transmitter side, and the voice encoder 52 is used. Sent to. The speech coder 52 includes a functional unit that performs the numerical conversion from the linear prediction coefficient as shown in FIG. 1 to the line spectrum versus frequency. The output signal from the voice encoder 52 is sent to the modulator 53, modulated into a high frequency signal suitable for communication, amplified by the high frequency amplifier 54, and sent out via the antenna 55. Further, a high-frequency signal sent from a telephone station, a relay device, or the like is received by an antenna 55, amplified by a high-frequency amplifier 56, and sent to a demodulation unit 57 for demodulation. Demodulator 5
The signal from 7 is sent to the speech decoder 58, and is subjected to a decoding process that restores the encoding process in the speech encoder 52. The voice signal obtained from the voice decoder 58 is sent to the speaker 59 on the receiver side and converted into an acoustic voice.

It should be noted that the present invention is not limited to the above-mentioned embodiment, and for example, the steps S20 and S can be performed.
28, etc., when it is determined that the converged solution is the higher adjacent solution on the frequency axis, the value is used as the value of ω _{i + 2} , and the calculation at the time of i + 2 is omitted. The amount of calculation can be reduced. Also, for example, the algorithm can be modified such that i is decreased from 10 to 1 instead of increasing i from 1 to 10. In addition, various modifications can be considered without departing from the scope of the present invention.

[0071]

As described above, according to the present invention,
On the time axis, the line spectrum-to-frequency value stored in the frequency storage means for holding the line spectrum-to-frequency value of the immediately preceding speech frame is converted into a line spectrum pair from a linear prediction coefficient. Since the approximate solution is the initial value of the approximate value of the solution in the calculation formula obtained by the Newton-Raphson method, the process of searching for the initial value of the approximate value of the solution given to the Newton-Raphson method is not necessary, for example. Thus, it is possible to reduce the processing amount in the whole numerical conversion processing from the linear prediction coefficient to the line spectrum versus frequency.

When proceeding with the approximation calculation by the Newton-Raphson method, the absolute value of the difference between the new approximation value obtained by one-time Newton-Raphson method and the original approximation value should not exceed a certain value. And the difference value between the new approximation value obtained by the previous Newton-Raphson method and the original approximation value by the above calculation means and the newton-Raphson method obtained this time. The infinite loop avoidance control means for detecting that the Newton-Raphson method enters the infinite loop and avoiding the infinite loop from the value of the difference between the new approximated value and the original approximated value, and the above calculating means. Initial value change control that determines whether the obtained approximate value has converged to a solution adjacent to the solution you were trying to find and changes the initial value of the Newton-Raphson method according to the result of the judgment Because it is provided with the stepped, or not converge to the desired solution, or into an infinite loop, nor malfunction such as or converge to the adjacent solution.

Further, by applying to the telephone device,
It is possible to reduce the amount of calculation, simplify the configuration, and reduce power consumption while maintaining high data compression efficiency.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an audio signal processing device according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram for explaining the reason why the difference between the original approximate value and the new approximate value must be limited to a certain constant value.

FIG. 3 is an explanatory diagram for explaining that there is a case where an infinite loop occurs.

FIG. 4 is a flowchart for explaining the operation of the audio signal processing device according to the embodiment of the present invention.

FIG. 5 is a block diagram showing an embodiment of a telephone device according to the present invention.

FIG. 6 shows Newton in a conventional audio signal processing device.
6 is a flowchart for explaining the operation of the approximate value calculation of the Rafson method.

FIG. 7 is a block diagram showing a configuration of a conventional audio signal processing device.

FIG. 8 is a Newton in a conventional audio signal processing device.
It is explanatory drawing for demonstrating the operation | movement of the approximate value calculation of the Rafson method.

[Explanation of symbols]

11 ... Linear prediction coefficient storage unit 15, 16 ... Newton-Raphson method approximation value calculation unit 17, 18 ... Approximation value determination unit 19 ... Previous frame line Spectrum-to-frequency storage unit 20 ... Approximate value difference limiting unit 21 ... Infinite loop avoidance control unit 22 ... Initial value change control unit

Claims (6)

[Claims] 1. A speech signal processing apparatus for converting a linear prediction coefficient used for coding a speech signal into a line spectrum pair frequency, wherein a coefficient of an equation showing conversion from a linear prediction coefficient to a line spectrum pair is linear. Coefficient conversion means for calculating from the prediction coefficient, coefficient storage means for holding the coefficient output from this coefficient conversion means, and frequency storage means for holding the value of the line spectrum versus frequency of the immediately preceding audio frame on the time axis. And a calculation means for obtaining a new approximate solution of the above equation by the Newton-Raphson method with the value of the line spectrum versus frequency stored in the storage means as the initial value of the approximate value of the solution. apparatus. 2. One time Newton-by the calculation means
Approximate value difference limiting means for limiting the new approximate value so that the absolute value of the difference between the new approximate value obtained by the Rafson method and the original approximate value does not exceed a certain value, and the previous Newton-Raphson method by the above calculating means. The Newton-Raphson method is infinite based on the difference value between the new approximation value obtained by the method and the original approximation value and the difference value between the new approximation value obtained by the current Newton-Raphson method and the original approximation value. An infinite loop avoidance control means that detects entering a loop and avoids entering an infinite loop, and determines whether or not the approximate value obtained by the above calculating means has converged to a solution adjacent to the solution to be obtained. 2. An audio signal processing apparatus according to claim 1, further comprising an initial value change control means for changing the initial value of the Newton-Raphson method according to the result of the determination. 3. A speech signal processing device for converting a linear prediction coefficient used for coding a speech signal into a line spectrum pair frequency, wherein the coefficient of an equation showing conversion from a linear prediction coefficient to a line spectrum pair is linear. Coefficient conversion means for calculating from the prediction coefficient, coefficient storage means for holding the coefficient output from this coefficient conversion means, frequency storage means for holding the initial value of the calculation of the line spectrum vs. frequency, and stored in this storage means. The calculation means for obtaining a new approximate solution of the above equation by the Newton-Raphson method with the value of the line spectrum versus frequency as the initial value of the approximate value of the solution, and the new approximation value obtained by the Newton-Raphson method once by this calculating means. And the approximate difference difference limiting means that limits the new approximate value so that the absolute value of the difference between it and the original approximate value does not exceed a certain value. The audio signal processing apparatus characterized by. 4. The previous Newton-based on the calculation means.
The Newton-Raphson method is calculated from the difference value between the new approximation value obtained by the Rafson method and the original approximation value and the difference value between the new approximation value obtained by the Newton-Raphson method and the original approximation value. 4. An audio signal processing apparatus according to claim 3, further comprising an infinite loop avoidance control means for detecting an infinite loop and avoiding an infinite loop. 5. In a speech signal processing device for converting a linear prediction coefficient used for coding a speech signal into a line spectrum-to-frequency conversion, a coefficient of an equation showing conversion from a linear prediction coefficient to a line spectrum pair is linear. Coefficient conversion means for calculating from the prediction coefficient, coefficient storage means for holding the coefficient output from this coefficient conversion means, frequency storage means for holding the initial value of the calculation of the line spectrum vs. frequency, and stored in this storage means. The calculation means for obtaining a new approximate solution of the above equation by the Newton-Raphson method by using the value of the line spectrum vs. frequency as the initial value of the approximate value of the solution, and the approximation value obtained by this calculation means is adjacent to the solution to be obtained. And an initial value change control means for changing the initial value of the Newton-Raphson method according to the judgment result. Audio signal processing apparatus characterized by. 6. A telephone device for encoding an input voice signal by a voice signal processing unit and then modulating it into a signal for transmission, wherein the voice signal processing unit comprises an equation showing conversion from a linear prediction coefficient to a line spectrum pair. A coefficient conversion unit that calculates a coefficient from a linear prediction coefficient, a coefficient storage unit that holds the coefficient output from this coefficient conversion unit, and a line spectrum-frequency value of the immediately preceding audio frame on the time axis. Frequency storage means, calculation means for obtaining a new approximate solution of the above equation by the Newton-Raphson method using the value of the line spectrum versus frequency stored in the storage means as an initial value of the approximate value of the solution, and the calculation means once An approximation that limits the new approximation value so that the absolute value of the difference between the new approximation value obtained by the Newton-Raphson method and the original approximation value does not exceed a certain value The value difference limiting means, the difference value between the new approximation value obtained by the previous Newton-Raphson method and the original approximation value by the calculation means, the new approximation value obtained by the newton-Raphson method this time, and the original approximation value. The infinite loop avoidance control means for detecting that the Newton-Raphson method enters the infinite loop and avoiding the infinite loop from the value of the difference from the approximate value, and the approximate value obtained by the above calculating means will be obtained. And a solution for determining whether the solution has converged to a solution adjacent to the solution, and changing the initial value of the Newton-Raphson method according to the result of the determination. .