JPH0934498A - Acoustic signal encoding method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To decrease the selective calculating amount of a noise exciting signal. SOLUTION: Matrices H and H<t> H=R using the impulse response of a synthesizing filter as elements are determined by a calculation part 31 after selecting a pitch exciting signal (p), and ψ is determined by another calculation part 32 from H, R, signal (p) and input voice (s), and preliminary selection part 33-37 calculate aj ψ(lj ) for vector elements ψ(i) of ψ about eight pulse positions lj and their polarities aj in groups 0-4, and in the sequence from greater value of aj ψ(lj ), for example, three pulse positions are preliminarily selected, and a candidate preparing part 38 prepares a noise exciting signal (c) of five pulses as one selection each from the three positions among group 0-4, and for each (c), ψc is calculated and also distortion for (s) is calculated so that the optimum noise exciting signal is selected.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to reproducing an acoustic signal such as voice or music by driving a synthesis filter of a pitch excitation signal expressing a pitch component and a noise excitation signal expressing a noise component. The present invention relates to a method of encoding an acoustic signal by utilizing the method.

[0002]

2. Description of the Related Art VSELP, CS are available as audio signal encoding methods for driving a synthesis filter using two excitation signals.
-CELP and other methods are known. These are I.
A. Gerson and MA Jasiuk: “Vector Sum Excited L
inear Prediction (VSELP) Speech Coding at 8kb / s ",
Proc. IEEE ICASSP '90, pp.461-464 (1990) (Reference 1)
Or A. Kataoka, T. Moriya and S. Hayashi: “An
8-kbit / s Speech CoderBased on Conjugate Structure
CELP ", Proc. IEEE ICASSP '93, pp.592-595 (1993)
(Reference 2) It is disclosed in other references.

A conventional example of this type of speech coding method will be described with reference to FIG. However, in the figure, only the main signal flow is shown for the sake of simplicity of description. First,
The sampling value series of the input speech waveform input from the input terminal 1 is supplied to the filter coefficient determination unit 2, where the filter coefficient is calculated by linear prediction analysis or the like.
The filter coefficient calculated in the filter coefficient determination unit 2 is then supplied to the filter coefficient quantization unit 3, the filter coefficient is quantized here, and the quantized filter coefficient is supplied to the synthesis filter 4 and synthesized there. The filter coefficient of the filter 4 is set.

The excitation signal for driving the synthesis filter 4 is composed of two excitation signals. One of the excitation signals is the output of the pitch codebook 11 and the other is the output of the noise codebook 21. The pitch codebook 11 is a component extracted in a plurality of pitch periods with respect to the past excitation signal for the synthesis filter 4, and the selected pitch period component candidate is multiplied by the pitch gain in the pitch gain multiplication unit 12 and output. . The noise codebook 21 includes a plurality of noise waveform components, and the selected noise waveform candidate is multiplied by the noise gain in the noise gain multiplication unit 22 and output. Synthesis filter 4
Is driven by the sum of the output of the pitch gain multiplication unit 12 and the output of the noise gain multiplication unit 22 in the addition unit 8.

In the distortion calculating section 5, the difference between the input audio signal input through the input terminal 1 and the synthetic audio signal output from the synthesizing filter 4 is taken in the subtracting section 9, and the difference which is the distortion. Each excitation component candidate in both codebooks 11 and 21 is selected so that is minimized, and at the same time, an optimum gain is set for each excitation component candidate. The code output unit 6 includes the quantized filter coefficients supplied from the filter coefficient quantization unit 3 and the gains of the selected candidate gain units 12 and 22 of the codebooks 11 and 21 selected by the distortion calculation unit 5. Are encoded and output. These codes are transmitted or stored via the output terminal 7. Normally, after selecting the pitch excitation signal,
Select the noise waveform excitation signal.

ACELP is known as a speech coding method in which the noise waveform component from the noise codebook 21 is expressed by a plurality of pulses. This is the reference R.
Salami, C. Laflamme and JP. Adoul: "ACELP Speech
Coderat 8kbit / s with 10msFrame: A Candidate for CC
ITT Standardization ", Proc. IEEE Workshop on Speec
h Coding, pp. 23-24 (1993) (Reference 3) and other references. In this method, a noise excitation signal having a time length of 5 ms is represented by a pulse train of 4 or 5 constant amplitudes, and the noise excitation signal is specified by the position of these pulses and their polarities. It eliminates the memory required for storage.

A conventional method for selecting a pulse as a noise excitation signal will be described with reference to FIG. The pitch excitation signal selected in the same manner as in the case shown in FIG. 3 is input from the terminal 13 to the adding section 8, and the noise excitation creating section 14 supplies the combination of a plurality of pulses as a noise excitation signal candidate to the adding section 8. , Its output is the synthesis filter 4 to which the filter coefficient from the quantizer 3 of FIG. 3 is set through the terminal 15.
And the difference between the output and the input audio signal is passed to the distortion calculator 5.
The noise excitation generator 14 is controlled so as to minimize the power of the error signal supplied to the pulse position and the pulse position where the error power is minimized and each polarity thereof are encoded and output.

However, since this method represents a noise excitation waveform with a combination of a plurality of pulses, there are a large number of noise excitation signal candidates, and in order to determine the optimum noise excitation signal candidate, a large number of arithmetic processings are performed. Need. That is, FIG.
In this case, it is necessary to perform the processes of noise excitation signal generation, synthesis filter, and error power calculation for all noise excitation signal candidates, and therefore many arithmetic processes are required.

Further, a method is known in which, when specifying a noise excitation signal, each noise excitation signal candidate is orthogonalized to a pitch excitation signal determined in advance and selected. This method can eliminate the influence of the pitch excitation signal determined in advance when the noise excitation signal is specified, and enables the noise excitation signal to be specified with higher accuracy. This is IA Ger
son and MA Jasiuk; “Vector Sum Excited Linear
Prediction (VSELP) Speech Coding at 8kb / s ", Proc.
IEEE ICASSP '90, pp.461-464 (1990) (reference 4) and other references.

The orthogonal selection of the conventional noise excitation signal will be described with reference to FIG. 5, and the case where the noise excitation signal is selected using the pitch excitation signal selected similarly to FIG. The filter coefficient is set to the synthesis filters 4a and 4b from the terminal 15, the pitch excitation signal selected from the terminal 13 is input to the synthesis filter 4a, and the noise excitation signal composed of a plurality of pulses selected from the noise codebook 17 is synthesized. Filter 4
b. The synthesized signals from the synthesis filters 4a and 4b are orthogonalized by the orthogonalization unit 18 and input to the distortion calculation unit 19, and the noise codebook 17 is selected so that the error power with respect to the voice signal from the terminal 1 is minimized. Done.

Since the synthetic noise component is uncorrelated with the synthetic pitch component in the orthogonalization unit 18, the distortion calculation unit 19
The noise excitation signal that minimizes the error power between the orthogonalized synthesized noise component and the noise component in the input speech is selected. When the selected pitch excitation signal is p, the noise excitation signal candidate is c, and the matrix having the impulse responses of the synthesis filters 4a and 4b as elements is H, the synthesized noise component C _{0 in} the output of the orthogonalization unit 18 is the above-mentioned document. As shown in FIG.

[0012] C ₀ = Hc - ^{[(p t H t Hc) /} (p t
H ^t Hp)] Hp H ^t denotes the transpose of H. The synthetic noise component C ₀ and the distortion D of the time series vector s of the input voice signal are obtained by the following equation. D = {s-γ _c C ₀ } ^t {s-γ _c C ₀ } The noise excitation signal candidate c that minimizes the distortion D is equivalent to the candidate c that maximizes the following expression.

D _su = (Ψc) ² / {(cRc) (pR
^{p) - (pRc) 2}} R = H t H Ψ = ( selection of ^{p t Rp) s t H- (} s t Hp) p t R optimum noise excitation signal to obtain the c that D _su is maximum Become. These are shown in the above-mentioned reference 4. The optimal noise excitation signal is selected by calculating D _su for all seat excitation signals c.

[0014]

For example, the duration (frame length) of the noise excitation signal is set to 5 ms, each noise excitation signal is represented by 5 pulses, and as shown in FIG. It is said that each of them can take only one of eight fixed pulse positions in the frame,
Any one of the eight pulse positions is taken out from each of the groups 0 to 4 to obtain a noise excitation signal composed of five pulses. In that case, each pulse has a polarity of either +1 or -1. Since the noise excitation signal is created in this way, the number thereof is remarkably large, and it is necessary to calculate the formula D _su for each noise excitation signal, and the amount of calculation is remarkably increased and the time also becomes long. .

According to the present invention, in a sound signal encoding method for driving a synthesis filter using two excitation signals, a pitch excitation signal expressing a pitch component and a noise excitation signal expressing a noise component, the noise excitation signal is used as a noise excitation signal. An object of the present invention is to provide an acoustic signal encoding method that can select a noise excitation signal that can be executed with a small amount of calculation, using a noise excitation signal composed of a plurality of pulses.

[0016]

According to the invention of claim 1, a candidate of the noise excitation signal is preliminarily used by utilizing the values obtained from the noise excitation signal, the input acoustic signal, the pitch excitation signal and the synthesis filter coefficient. , And select the distortion between each synthetic acoustic signal and the input acoustic signal obtained by driving the synthesis filter with these selected candidates of each noise excitation signal, and select the noise excitation signal that minimizes the distortion. Select one.

According to the invention of claim 2, a noise excitation signal,
Using the values obtained from the input acoustic signal, the pitch excitation signal, and the synthesis filter coefficient, a plurality of candidates for the position and amplitude of the pulse forming the noise excitation signal are preliminarily selected, and the pulse position is selected. The distortion between each synthetic acoustic signal obtained by driving the synthetic filter with each noise excitation signal and the input acoustic signal is obtained, and one noise excitation signal that minimizes the distortion is selected.

More specifically, two excitation signals, a pitch excitation signal expressing a pitch component and a noise excitation signal expressing a noise component, are used to drive a synthesis filter to reproduce an acoustic signal, Encoding by selecting a pitch excitation signal and a noise excitation signal with minimum distortion with the input acoustic signal, in which case the noise excitation signal is represented by a plurality of pulses, and a matrix representing the characteristics of the synthesis filter, In a method of selecting a noise excitation signal that maximizes the result of calculation using a pitch excitation signal, an input acoustic signal, and a noise excitation signal, in the invention of claim 1, the term corresponding to the numerator of the calculation is calculated. , Pre-select a predetermined number of noise excitation signals in descending order of the calculation result, and select the pre-selected noise excitation signal that gives the maximum result of the above calculation. .

According to the second aspect of the invention, the noise excitation signal in the first aspect of the invention is created by selecting one pulse position from each of a plurality of groups each having a plurality of predetermined positions on the frame where the pulse can be taken. Select and create by giving polarity, pre-selection is performed for the pulse position in each group, the calculation corresponding to the numerator of the maximum calculation is performed, and a predetermined number is selected from the calculation result in each group in descending order. The selected pulse position is selected and the noise excitation signal generated by selecting it from each group is determined so as to maximize the result of the above calculation.

[0020]

For simplicity, the sampling frequency of the input voice is 8000 Hz, the duration time (frame length) of the noise excitation signal is 5 ms, the noise excitation signal is expressed by 5 pulses, and The ~ fourth pulse can exist only at the position on the frame defined for each of the 0th to 4th groups shown in Fig. 6. That is, the noise excitation signal is formed by selecting pulse positions one by one from these 0th to 4th groups and further setting the polarity of each pulse. The pitch excitation signal is selected, for example, as shown in FIG.
As described above. The present invention is characterized by the selection of the noise excitation signal, and this selection seeks the one that maximizes the calculation of D _su described in the section of the prior art.
In the invention of FIG. 1, first, as shown in FIG. 1A, the calculation of Ψ of the numerator of D _su is performed by a matrix H having the impulse response of the synthesis filter as an element, the time series vector s of the input speech signal, and the pitch excitation signal (Vector) p and (S ₁ ). Next, the product of this Ψ and all noise excitation signals (vectors) c obtained from FIG. 6 is obtained (S ₂ ). Let i (i) be the i-th element (i = 0, 1, 2, ... 39) of the vector of Ψ,
When the polarities of the five pulses of the noise excitation signal c are a _{0 to} a ₄ and the pulse positions are l ₀ to l ₄ , Ψc is calculated by the following equation.

Ψc = a ₀ ψ (l ₀ ) + a ₁ ψ (l ₁ ) +
... + a ₄ ψ (l ₄ ) This calculation is executed for each noise excitation signal. These Ψ
For example, about 30 pieces are preselected in descending order of the value of c (S ₃ ). The D _su is calculated for each of the preselected noise excitation signals (S ₄ ), and the maximum D is calculated.
_The noise excitation signal that has become _su is selected (S ₅ ). Preliminary selection is performed in this way, but if the value of D _su is large, ψc of the numerator is also considered to be large, and there is no possibility of dropping the optimum noise excitation signal by appropriately selecting the number even in this preselection. .

Next, an embodiment of the invention of claim 2 will be described.
This invention is also characterized by the selection of the noise excitation signal,
Also in this embodiment, as shown in FIG.
Calculate (S₁). This operation is the same as that of FIG. 1A.
U. Next, for each group 0 to 4 in FIG.
8 pulse positions l in this example_j(0,
About 1, ..., 7)_j, L_jThe element of Ψ using
a_jψ (l_j) Is calculated (S_Two). Each of these groups
Every a_jψ (l_j1) from the calculation result of 1)
Multiple a_j, L_jTake out (S_Three). These are taken out
A_j, L_jTake out one from each group,
A possible noise excitation signal c is created (S _Four). like this
From the noise excitation signal that can be made from the one shown in FIG.
A preliminary selection has been made. This preselected miscellaneous
D using sound excitation signal_suIs calculated (C_Five), The performance
Select the noise excitation signal with the maximum calculation result
(S₆).

Functionally, the embodiment shown in FIG. 1B is as shown in FIG. That is, H and R are calculated by the R, H calculation unit 31 by the synthesis filter coefficient from the terminal 15, and from these, the input sound s from the terminal 1 and the pitch excitation signal p from the terminal 13, Ψ is calculated by the Ψ calculation unit 32. It is calculated. The values of a _j ψ (l _j ) in the groups 0 to 4 are calculated by the preliminary selection units 33 to 37, and a _j and l _j are respectively calculated in descending order of the value of a _j ψ (l _j ) in each group, for example. Select three at a time.

In the noise excitation signal candidate creation unit 38, a preliminary selection unit
A selected in each of 33 to 37 _j, L_jOne by one
Take out all of them and make a noise
As excitation signal candidates, Ψ calculator
Ψ obtained in 32 is multiplied by the Ψc calculation unit 39, and each Ψc
The multiplication result of R, p and the noise excitation signal candidate creation unit 38
The distortion is calculated by the distortion calculating unit 41 according to the created corresponding c.
D_suIs calculated, and the noise that maximizes the calculation result is
Select the excitation signal candidate in the optimum excitation signal candidate selection unit 42
Output.

In the invention of claim 2, as in the case of the invention of claim 1, the fact that Ψc is large means that the constituent element a
_{Since 0} ψ (l ₀ ), a ₁ ψ (l ₁ ), ..., A ₄ ψ (l ₄ ) should also be large, a _j from each group
Even if the one with a large value of ψ (l _j ) is preselected, the optimum noise excitation signal can be correctly selected.

[0026]

As described above, according to the invention of claim 1.
For example, calculate Ψc and preselect c that has a large value.
And D only for the preselected c_suTo calculate
D for all c_suThe amount of calculation is smaller than
It becomes. According to the invention of claim 2, a_jψ (l_jPlay)
A in each group_j, L _jReduce the number of
The noise excitation signal c created from this_suCompute
Because D for all c_suComputation amount rather than computing
Is less. For example, 3 pulse positions from each group
When preselecting
Is 3^Five= 243, if you do not make this preliminary selection
Number 8^FiveDistortion that is significantly less than the number of pieces and has a large amount of processing
Calculate (D_suThe number of operations is reduced accordingly, and the effect is
Remarkable.

FIG. 7 shows the amount of calculation required for noise excitation signal selection and the quality of reproduced voice when the number of pulse position selections from each group is changed. When the number of selections is 1, there is one noise excitation signal to be created, which is the selected noise excitation signal. When the number of selections is 8, it is a conventional method that does not perform preliminary selection, and in this case, the amount of calculation is significantly large. The reproduced voice quality shows that the higher the value, the better. Therefore, originally, the larger the number of selections, the better the quality should be, but there was not much change. This can be said to indicate that this preliminary selection is done fairly correctly.

Although the present invention has been applied to the coding of audio signals in the above, it can also be applied to the coding of acoustic signals.

[Brief description of drawings]

FIG. 1 is a flow chart showing an essential part of an embodiment of the invention of claim 1, and B is a flow chart showing an essential part of an embodiment of the invention of claim 2.

FIG. 2 is a block diagram functionally showing a main part of an embodiment of the invention of claim 2;

FIG. 3 is a block diagram functionally showing a conventional CELP encoding method.

FIG. 4 is a block diagram functionally showing a part of a conventional ACELP encoding method.

FIG. 5 is a block diagram functionally showing a conventional method of orthogonalizing a noise excitation signal with respect to a pitch excitation signal and selecting the noise excitation signal.

FIG. 6 is a diagram showing an example of pulse positions of each group of noise excitation signal action.

FIG. 7 is a diagram showing the relationship between the number of preliminary selections from a group, the amount of calculation, and the reproduced voice quality.

Claims (2)

[Claims] 1. A synthesizing filter in which a quantized filter coefficient obtained from an input acoustic signal is set to a pitch excitation signal expressing a pitch component and a noise excitation signal expressing a noise component.
The audio signal is encoded by utilizing the time-series vector component composed of the individual excitation signals to reproduce the audio signal by driving for each frame, and the noise excitation signal is represented by a small number of pulses, and each pulse In the acoustic signal encoding method for encoding the position and the amplitude of the noise excitation signal, the noise excitation signal, the input acoustic signal, the pitch excitation signal, and the value obtained from the synthesis filter coefficient are used to reserve the candidate of the noise excitation signal. A plurality of selected, further, the distortion between each synthetic acoustic signal and the input acoustic signal obtained by driving the synthetic filter with the candidate of each of the noise excitation signals selected here is obtained, and the distortion is minimized. A method for encoding an acoustic signal, characterized in that only one noise excitation signal is selected. 2. A synthesizing filter in which a quantized filter coefficient obtained from an input acoustic signal is set to a pitch excitation signal expressing a pitch component and a noise excitation signal expressing a noise component.
The audio signal is encoded by utilizing the time-series vector component composed of the individual excitation signals to reproduce the audio signal by driving for each frame, and the noise excitation signal is represented by a small number of pulses, and each pulse In the acoustic signal encoding method for encoding the position and amplitude of the pulse, the noise excitation signal is formed by using a value obtained from the noise excitation signal, the input acoustic signal, the pitch excitation signal, and the synthesis filter coefficient. A plurality of candidates for the position and amplitude of each of them are preliminarily selected, and each synthesized acoustic signal and input acoustic signal obtained by driving the synthesis filter with each of the noise excitation signals configured by the pulse positions selected here. And a noise excitation signal that minimizes the distortion is selected, and an acoustic signal coding method is characterized.