JPH03130800A - Voice encoding system - Google Patents

Abstract

PURPOSE:To perform filter arithmetic recurrently, and to decrease a calculation quantity and perform real-time processing by setting the matrix of a driving signal to a Toeplitz matrix and using the characteristics of this matrix. CONSTITUTION:The signal series of a voice signal is inputted to an input terminal 101 and a block segmentation part 102 outputs a voice vector by putting sampled values in one frame and inputs it to an LPC analysis part 104 and a weighting filter 105. The analysis part 104 analyzes the voice signal and the filter 105 weights an extracted LPC prediction parameter. Then when a pitch period search is made, the LPC filter driving signal outputted by an adder 118 is inputted to a switch 115. The matrix of this driving signal is so set as to obtain the Toeplitz matrix, whose Toeplitz characteristics are utilized to find a pitch period recurrently by the filter arithmetic. Thus, the calculation quantity is reduced and the real-time processing is carried out.

Description

[Detailed Description of the Invention] [Object of the Invention] (Industrial Application Field) The present invention relates to vector quantization used for information compression and transmission of digital signal sequences such as audio signals, It is an encoding method.

(Prior art) Vector quantization is one of the technologies that is currently attracting the most attention as a method for efficiently encoding audio and image signals.
It is. In particular, in the field of audio coding, CE
L P (Code ExcitedLinear
Prediction method or V
tor-・1xcited coding) method is
It is known as an excellent method that applies vector quantization. For details on the CELP method, see M, R, 5c.
Mr. hroeder and Mr. B, S, Ata1… “
Code-Excited Linear Pre-d
ictron (CELP): High-Qual
ity 5peech At Very Low = Bi
tRates 1n Proc, ICASSP,
1985. pp.

937-939, the CELP method will be briefly explained below with reference to this drawing. FIG. 7 is a block diagram showing the basic configuration of the CELP system. In the same figure, a signal sequence of an audio signal is inputted manually from an input terminal 1, and a block extraction section 2 extracts L sample values into one.
frame, and the input audio signal vector of length L is 3
The signal is output from the LPG analyzer 4 and input to the LPG analyzer 4. The LPG analysis unit 4 performs LPG analysis of the audio signal using an autocorrelation method etc., and calculates the LPC prediction parameter (α, +(i-1...
・Extract P)l. Here, P is the prediction order. In addition, the LPG predicted residual signal vector is outputted by 18 and inputted to the pitch analyzer 21.The pitch analyzer 21 uses the LPC predicted residual signal vector to perform pitch analysis, which is a long-term prediction of speech, and performs pitch analysis. Extract the period (TP) and gain parameter (b).

The extracted LPG measurement parameters, pitch period, and gain parameters are each input to the LPG synthesis filter 4.
and is used when the pitch synthesis filter 23 generates synthesized speech.

Next, the process of generating synthesized speech will be explained. The code book 17 contains n white noise vectors of dimension k (number of vector elements). Here k is generally L/
K is chosen to be an integer.

By multiplying the third white noise vector in this code book 17 by the gain parameter 22 and passing it through the pitch synthesis filter 23 and the LPG synthesis filter 4,
24, a synthesized speech vector is generated. At this time, the transfer function p (z) of the pitch synthesis filter 23 and the transfer function A (Z) of the LPG synthesis filter 14 are as follows.

P(Z) 41/ (1+b z-" ) (2) This generated synthesized speech vector is input to the squared error calculation unit 19 together with the input speech vector which is the target vector, where the synthesized speech vector and the input speech vector are The Euclidean distance i1t!Ej from In other words, the CELP method is characterized in that it performs vector quantization by using a code book for the drive signal of the synthesis filter in the process of speech synthesis.In addition, since the human speech vector has length L, , this process will be repeated L/ times. Also, the weighting in the figure indicates that the filter 5 is used to shape the spectrum of the error signal and reduce the distortion that would be perceived by the human ear. The transfer function is given by the following equation.

H(Z) -A(Z) -P(Z)
(4) When actually using the CELP method as an encoder, LPC prediction parameters, pitch period, pitch gain parameters, code book number, code...
The book gain will be encoded and transmitted to the decoder.

Figure 8 is functionally a C which performs exactly the same processing as Figure 7.
FIG. 2 is a block diagram of the ELP method. In FIG. 8, the weighting filter 5 has been moved out of its position within the loop of the conventional code book search. This is the pitch synthesis filter 23 (7) P (Z) woP ('
), LPG synthesis filter 4 (7) By replacing γ A (Z) with A(z-), it is possible to reduce the amount of calculation while maintaining the same function γ .

In addition, the pitch synthesis filter 23 and the LPG synthesis filter 14
The initial memory in the filter operation is made to not affect the search of the code book by the generation of synthesized speech. That is, by providing a pitch synthesis filter 25 and an LPG synthesis filter 7 having initial memories, and subtracting the zero person vector outputted by 8 from the weighted human voice vector outputted by 6 in advance, and using it as the target vector. , pitch synthesis filter 23 and LP
It is possible to set the initial memory of the G synthesis filter 14 to zero. Also, at the same time, is this conventionally generated synthetic speech?
What was previously performed in the filter operation of a synthesis filter that receives a code vector as input can now be expressed as the product of the code vector and the next kXk lower triangular matrix.

Here, k is the number of dimensions (number of elements) of the code vector of code book 17, h(1), i-1...
k is an impulse response of length γ when the initial memory of H(''-) is zero.

From the above, the vector outputted by 11 in FIG. 8 is set as the target vector, and the square error calculation unit 19 calculates
j is determined, and the minimum distortion search unit 20 determines the minimum value.

E j -If Xt-7,HC,II (j-1,2
,=-n)(e)J Here, Xt is the target input vector 2, C1 is the j-th code vector, and γj is the optimal gain parameter for the j-th code vector.

FIG. 9 shows a flowchart for determining this Ej and determining the vector number j that gives the minimum value.

In this process, it is first necessary to obtain HCj for each j, which requires k(k+1)/2·n multiplications, and the commonly used k -40,...n-10
24 requires 839,680 multiplications.

Furthermore, in the entire flow, if L/ is set to -4, 1048736 multiplications are required per frame.

Then, when the number of samples in one frame is L-180 and the sampling frequency of the input audio is 13KHz, it is 52MIP.
This requires about S multiplications, and even a DSP with a multiplication capacity of 20 old PSs requires three times.

Further, as a method for improving the sound quality of the CELP method, a method called closed-loop pitch prediction or adaptive code book is known. For details on this method, see W.
B., K.I., D., J., Krasinskian.
d RH, XetchuIIl, ”lll1p
roved 5peech Quality and E
frlclent Vector Quantizat
ion In CELP ”in Proc ICA!
l;SP, (L988.ppL55- ++ 158)
It is stated in

Referring to the drawings below, we will briefly explain CELP using a method called closed rubbization of pitch prediction or adaptive code book.
I will explain about it. FIG. 10 is a block diagram showing its basic configuration. The difference between CELP based on a method called a pitch prediction closed loop or applied code book and the CELP shown in FIG. 7 lies in the difference in pitch analysis. In FIG. 7, pitch analysis is performed using the LPC prediction residual signal vector output from the LPG analysis section. On the other hand, CELP shown in Figure 10 is characterized in that the pitch analysis is performed in a closed loop similar to the code book search. In this CELP, the LP output from 18
The G synthesis filter drive signal passes through a delay device 13 that is variable over the pitch search range a-b to create a drive signal vector for pitch period j. As a result, a pitch search is performed using a weighted human vector, which excludes the influence of the frame before being output, as a target vector.

At this time, the following Ej is found by the evaluation section using square distance distortion.

Ej-11X1-γjHBj 11 (a≦j≦b)
(7) Here, X7 is the target vector, Bj is the pitch period j
The drive signal vector γ when , is the optimal gain factor for pitch period j, H is given by equation (5),
. . . k is an impulse response of γ length k when the initial memory of AC"-) is zero.

FIG. 11 shows a flowchart for determining the pitch period which determines this E and gives its minimum value. In this process, it is first necessary to find HBj for each t and j, and k(k+1)/2-(b-a+1)・L/
It requires -40,L multiplications and is commonly used.
-180,a -20,b 419 in case of -147
Requires 840 multiplications. Furthermore, the entire flow requires 461312 multiplications per frame. Then, when the input audio sampling frequency is 8K) Iz, approximately 23 MIPS of multiplication is required, and 20 MIPS
Even a DSP with a multiplication capacity of

(Problems to be Solved by the Invention) As mentioned above, the code book search and the pitch period search called a closed-loop or adaptive code book in the CELP method require a large amount of calculation, and DSP (Digital Signal Pr
Therefore, the present invention aims to provide an audio encoding method that can eliminate this problem and reduce the amount of calculation to the point where real-time processing can be performed using a DSP or the like. The purpose is

[Structure of the Invention] (Means for Solving the Problems) The present invention provides a pitch that minimizes distortion between a vector obtained by performing a filter operation on a drive vector and an input vector in pitch pre-DI called a closed-loop or adaptive code book. When searching for a period, this is an audio encoding method in which the driving signal matrix is set to be a Tebritz matrix, and then a filter operation is recursively determined by using the Tebritz characteristic.

(Operation) In the present invention, in Pittsch prediction called a closed-loop or adaptive code book, the driving signal matrix is made into a Toeplitz matrix, and by utilizing the Tebritz characteristic, it is possible to obtain the filter operation recursively and at high speed. Therefore, the number of multiplications can be reduced.

(Example) Hereinafter, one embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing the basic configuration of a speech coding system to which a pitch parameter search method according to an embodiment of the present invention is applied. In the same figure, a signal sequence of an audio signal is input through an input terminal 101, and a block extraction unit 102 sets L sample values as one frame and has a length L.
The input signal is output as a speech vector, and is input to the LPG analysis section 104 and the weighting filter 105, respectively. The LPG analysis unit 104 performs LPG analysis of the audio signal using an autocorrelation method or the like, and calculates LPG prediction parameters...
- Extract 1ail (i-1...P). Here P is the prediction order. This extracted LPC prediction parameter is the LPG synthesis filter 1 (17, 109, 11
Used in 4. The weighting filter 105 uses the LPG prediction parameters extracted by the LPC analysis unit 104 (
at), it is moved out of the original code book search and pitch period search loop for the purpose of weighting the input signal vector. This is LPG synthesis filter 107, 109, 114
By changing A (Z) to A (R), the error signal Sγ vector is shaped to reduce the amount of calculation while maintaining the function of reducing distortion that would be perceived by the human ear. The weighting allows the transfer function W (Z) of the filter to be given by the following equation.

W(Z) −A (') /A(Z) (0≦7≦
1) (8) γ In addition, the initial memory in the filter calculation of the LPG synthesis filters 109 and 114 is designed to influence the pitch period search or code book search in the generation of synthesized speech. i.e. initial memory with zero value 1
The initial memory of the LPG synthesis filter is It is possible to reduce the value to zero. Also,
At the same time, this means that the filter operation of an LPG synthesis filter that takes a drive signal vector as an input in a pitch period search or the filter operation of an LPC synthesis filter that takes a code vector as an input in a code book search is a drive signal vector or a code vector, respectively. and the next k
This allows it to be expressed as a product of Xk and a lower triangular matrix.

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

i-1, . . . are impulse responses of length k when the initial memory of A(''-> is zero γ).

In the pitch period search, the LPG synthesis filter drive signal e output from the adder 11111 is first input to the switch 115, and when the pitch period j to be searched there is greater than or equal to the number of dimensions of the code vector, the LPG synthesis filter drive signal e is input to the delay device 116.
, if it is less than K, it is input to the waveform coupler, and a drive signal vector for the pitch period j is created over the pitch search range a-b. This process will be explained in detail using diagrams. The counter -H increments the pitch period j over the pitch period search range a to b, and sends the value to the drive signal section 112. Switch 115. Output to delay device 11B. When the pitch period j is greater than or equal to 1, the drive signal vector Bj is changed from the past drive signal vector e created by the delay device as shown in (1) in FIG.
make. That is, e −(e (-b), e (-b+1)= ・−
e (-1)) (11)-(e (-D, e
(−j+1), ・−e (−1 to −j+)) t
(j=to, ni+1. . . . , b) (12) Here, B is the drive signal vector at period j. Further, t represents transposition. Next, when the pitch period is less than j, the past drive signal (e (-
P), e (-p+t)-++ e (-1)) is combined with the past drive signal vector e, and the e'
Create a driving signal vector from That is, B j =
(e(-j), e(-j+1l-e(-L)e(-P)
e(-P+1)-e(-P+ to -j-1))'
(13) (j=a, a+1...K-1) By doing this, each element of the drive vector Bj is (bj (1), bj (2),...b j (k)
) T, b −(Ill) −b −(o+−1
) (a+1≦j≦b, 2≦J J−
1 m≦k), and the drive signal matrix B having the drive signal vector B as a column vector can be made into a complete Teblitt matrix as shown below.

In the search for the pitch period, a pitch period that minimizes the error is determined as a weighted human vector target vector that is output from the adder 10B and excludes the influence of the previous frame. Distortion due to the square distance of the error at this time.

Ej is determined by the following formula.

E, -11X -7, HB, If (a≦j≦b)
(15) J tJ J Here, Xt is the target vector, Bj is the drive signal vector at pitch period J, γj is the optimal gain factor for pitch period j, and H is given by equation (10).

In calculating Equation (15), if we use the properties that the driving signal matrix is a Tepli-Knots matrix and that the weighted LPG synthesis filter is also a lower triangular matrix and a Teblitz matrix, the calculation of HBj, that is, the filtering can be performed recursively as shown in the following equation.

V , (1) -h (1) e (-D
(1B)V-(n)-V, (
m-1)+h(m)e(-j)J J-1 (2≦m≦K) (17) (B+l≦j≦b) However, (V, (1), V-(2), -V-(K))T is an element of HBjkoJ3.

As a result, if you follow the flowchart in Figure 3, HB
Only a is found by ordinary matrix-vector product operation, and HB, (
all≦j≦b) is HB, more recursive J
The required number of multiplications is IK(K+
1)/2+K(ba)J ・Reduced to L/K.

- K -40, L -180, a used for boat fishing -
In the case of 20°b-147, there are 23,600 multiplications.

Also, in the entire flow, there are 85072 multiplications,
The calculation amount is about 14% of the conventional method shown in Figure 9, and when the input audio sampling frequency is 8KHz, it is 3JMIP.
A multiplication of about S is sufficient.

In FIG. 1, when the optimum pitch frequency Mj is determined by such pitch/frequency search, the output of the LPG synthesis filter 109 is obtained from the weighted human speech vector outputted from the adder 106 and excluding the influence of the previous frame. The pitched synthesized speech vector with the optimum pitch period j is subtracted, and a weighted input speech vector excluding the influence of the previous frame and the pitch is output.

Next, in the code book search, the weighted human speech vector, which is the output of the adder 131 and which excludes the influence of the previous frame and the influence of pitch, is used as the target vector.
Synthesized speech is generated by the code vectors in the code book 117. Then, the code vector number j whose distortion Ej due to the squared distance of the error is the smallest is selected. This process is expressed by the following equation.

E, -11X -7,HC,It (1≦j≦n)
(14) tJJ where X is a weighted human vector excluding the influence of the previous frame and pitch, Cj is the jth code vector, γj is the optimal gain factor for the jth code vector, and n is the This is the number of code vectors.

When C0 is composed of independent white noise, it takes a large amount of calculation to find E and to find the code book number that gives its minimum value. Requires.

Therefore, in the present invention, in order to reduce the amount of calculation, Cj
As shown in FIG. 4, it is assumed that a sample of length k is cut out by shifting one sample from the rear of a one-digit white noise sequence U of length n+ -1.

At this time, as shown in Figure 4, C, (m)--C
, (m-1) (Z≦j≦n, 2≦m≦k) -
1, and the code book matrix C with the code vector C1 as each column vector has the characteristic that it is a Templitz matrix. At this time, IC.

Let each element of (W, (1), W, (2),...
・W, (k))TshikoJJ When this happens, the following relational expression holds true and can be found recursively.

Wj (1) = h (1) U (n+1-j)
(2≦m≦k)W, (m)−Wj−1(+1−1)(
2≦j≦n) 10h (m) U (n+1−D Therefore, according to the flow chart in Fig. 5, H
Only Cl is determined by ordinary matrix-vector product operation, and HCo
(Z5j*n) is HC, more J
Since it is possible to find J-1 recursively, the number of operations required is tll (K+1)/2
+K·(n-1)l.

In the case of k-40, n-1024 used for boat fishing, the number of multiplications is 41740 times. Further, the total number of calculations is 250,796 times in the entire flow, which is about 24% of the calculation amount according to the flowchart of FIG. 8, and when the sampling frequency of the input audio is 8 KHz, only about 12.5 MIPS of multiplication is required.

Conversely, the code vector is shifted by one sample from the front of the one-digit white noise sequence of -1 to length n+, and
It is possible to cut out a sample of This method requires (K-1)(n-1) more multiplications than the one described above. In addition, the code book search using this method is not only possible in CELP, which is called a closed-loop first pitch prediction or adaptive code book, but also in CELP with the configuration shown in FIG.
This is possible by replacing (1) with H (-) according to equation (4).

γ In FIG. 1, the pitch period given from register -10 can also be determined in frame units by a known method such as the autocorrelation method and given to the waveform combiner 130.

FIG. 6 shows a block diagram showing the basic configuration of the audio encoding system according to the embodiment in this case. Furthermore, when the pitch period j is less than J, the waveform combiner 130 can combine the zero vector with the past drive signal vector e and create a drive signal vector from it. In this case, the amount of calculation can be further reduced.

[Effects of the Invention] As explained above, according to the present invention, in pitch prediction called a closed-loop or adaptive code book,
Let the drive signal matrix be a Tebritz matrix and use its characteristics 4
By doing this, the filter operation can be found recursively, and by using the code book matrix as a Toeplitz matrix in the code book search, the filter operation can be found recursively, which reduces the amount of calculation. It has the advantage of

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing one embodiment of the present invention. FIG. 2 is a diagram for explaining the present invention in detail, FIG. 3 is a flowchart for explaining the calculation means of one embodiment of the present invention, and FIG. 4 is a diagram for explaining one embodiment of the present invention. figure,
FIG. 5 is a flowchart for explaining the calculation means of one embodiment of the present invention, FIG. 6 is a block diagram for explaining another embodiment of the present invention, and FIG. 7 is a block diagram for explaining the conventional technique. . FIG. 8 is a block diagram for explaining the conventional technology. This is a flowchart for 1.101...Input terminal, 2,102...Block cutting section. 3... Output terminal, 4,104... LPG analysis section. 5.105... Weighting is a filter, 6... Output terminal. ? , L07.109.114...I, PC synthesis filter. 8...Output terminal, 9...LPG synthesis filter. 10... Output terminal, 11... Output terminal. 12,112... Drive signal section, 13,116...
delay device. t4...LPG synthesis filter, 15...output terminal. 16...Output terminal, 17,117...Code book. 18... Output terminal, 19.119... Squared error calculation unit. 20.120...Minimum distortion search unit, 103...Pitch analysis unit. 106.131, 132...Adder, 21...Pitch analysis section. 22...Gain parameter. 23... Pitch synthesis filter section, 24... Output terminal. 25... Pitch synthesis/filter section. 26...Evaluation section based on square distance distortion.

Claims (1)

[Claims] In an audio encoding method that searches for a pitch period so that the distortion between a drive signal vector obtained by performing a filter operation on a drive vector and a target input vector is minimized, the matrix of the drive signal is set to be a Toeplitz matrix, A speech encoding method characterized in that the filter operation is recursively obtained by utilizing the Toeplitz characteristic of the Toeplitz matrix.