JPH04352200A - Speech encoding system - Google Patents

Abstract

PURPOSE:To reduce the search arithmetic quantity and memory capacity of a code book by storing the code book with N-dimensional initial vectors and delta victors representing a reference noise string and representing code vectors in the noise string in a tree structure shape. CONSTITUTION:The initial vector C0 and (L-1) kind of delta vectors C1- CL-1 are stored in the delta code book 11 previously and the respective delta vectors are added to or subtracted from the initial vector, layer by layer to represent the code vectors C0-C1022 of 2L-1 kind of noise strings sequentially in the tree structure shape; and a zero vector is added to them, thereby representing 2<10> kind of code vectors C0-C1023. An error minimum noise string determination part 17 determines a code vector which minimizes error electric power. At this time, cross correlation and autocorrelation are represented as recurrence formulas and calculated by a cross-correlation arithmetic part 14 and an autocorrelation arithmetic part 15. Then a speech encoding part 18 outputs a code specifying this code vector to perform speech encoding.

Description

[Detailed description of the invention]

0001

[Field of Industrial Application] The present invention relates to an audio encoding system for compressing information on audio signals, and in particular, to
Analysis for encoding at a transmission rate of s
The present invention relates to a speech encoding method using by-synthesis vector quantization. A-b-S
An audio encoder based on an audio encoding method using type vector quantization, for example, CELP (Code ExcitedLi)
2. Description of the Related Art Encoders (near production) are long-awaited as they can realize information compression while maintaining voice quality in in-house communication systems, digital mobile radio systems, and the like.

0002

2. Description of the Related Art Speech can be divided into voiced and unvoiced sounds. Voiced sounds are generated from a pulse sound source caused by the vibration of the vocal cords, and the vocal tract characteristics of the individual's throat and mouth are added to form the voice. In addition, unvoiced sounds are sounds made without the vocal cords moving, and the sound source is a simple Gaussian noise train that passes through the vocal tract and becomes the voice. Therefore, as shown in Figure 5, the speech generation mechanism consists of a pulse sound source PSG that is the source of voiced sounds, a noise source NSG that is the source of unvoiced sounds, and linear predictive synthesis that adds vocal tract characteristics to the signals output from each sound source. It can be modeled by filter LPCF. It should be noted that the human voice has periodicity, and the period corresponds to the periodicity of the pulses output from the pulse sound source, and varies depending on the person and the content of speech.

From the above, if it is possible to specify the pulse period of the pulse sound source and the noise sequence of the noise source that correspond to the input voice, it is possible to encode the input voice with a code that identifies these pulse periods and the noise sequence of the noise source. Can be done. Therefore, the pulse period of the pulse sound source is identified based on the periodicity of the input audio signal using an adaptive codebook, and the pulse train with the period is input to a linear prediction synthesis filter and filter calculation processing is performed. The periodic component is removed by subtracting the filter operation result from the input audio signal. Next, a plurality of noise sequences (each noise sequence is represented by an N-dimensional code vector) is prepared in advance, and each code vector is subjected to synthesis filter processing to remove the periodic component and the reproduced signal vector. By finding a code vector that minimizes the error with the input signal vector (N-dimensional vector), it is possible to encode speech using data specifying the period and the code vector.

FIG. 6 is a block diagram of a speech encoding device using vector quantization based on A-b-S. 1 is a noise codebook that stores a plurality of randomly generated noise sequences C, for example, 1024 types (each noise sequence is expressed as an N-dimensional code vector); 2 is an amplification unit with a gain g; 3 is an amplification unit 4 is a linear prediction synthesis filter that performs calculation processing of a synthesis filter that simulates vocal tract characteristics on the output of the linear prediction synthesis filter 3; 4 is an error generation unit that outputs the error between the reproduced signal vector output from the linear prediction synthesis filter 3 and the input signal vector; This is an error power evaluation unit that evaluates the error and finds a noise sequence (code vector) that minimizes the error.

Vector quantization using A-b-S differs from normal vector quantization in that after each code vector (C) of the noise codebook 1 is multiplied by the optimal gain g, it is filtered by the linear prediction synthesis filter 3. The error signal (E) between the reproduced signal vector (gAC) obtained by filtering and the input signal vector (X) is obtained by the error generation section 4, and the power of the error signal is calculated by the error power evaluation section 5. A noise codebook is searched as an evaluation function (distance measure), a noise sequence (code vector) with the minimum error power is found, and the noise sequence (
The input signal is encoded by a code specifying the code vector).

The error power is given by the following equation |E|2=|X-gAC|2 (1). Optimal code vector C and gain g
is determined to minimize the error power shown in equation (1). Note that since the power varies depending on the loudness of the voice, the gain g is optimized to match the power of the reproduced signal to the power of the input signal. The optimal gain can be found by partially differentiating equation (1) with respect to g and setting it to zero. In other words, from d|E|2/dg=0, g is g=(XT AC)/((AC)T (AC))
It is given by (2). Substituting this g into equation (1), |E|2=|X|2-(XT AC)2/(
(AC)T (AC)) (3). The cross-correlation between the input signal X and the synthesis filter output AC is expressed as RXC,
If the autocorrelation of the synthesis filter output AC is RCC, then the cross-correlation and autocorrelation are as follows: RXC=XT AC (4)R
It is expressed as CC=(AC)T (AC) (5).

Since the code vector C that minimizes the error power in equation (3) is the one that maximizes the second term on the right side, the code vector C is calculated using the following equation C=argmax(RXC2/RCC)
(6), and the optimal gain is g=RXC/RCC from equation (2) using cross-correlation and autocorrelation that satisfy equation (6).
It is given by (7).

FIG. 7 is a block diagram of a noise codebook search processing section that encodes an input signal by finding a noise sequence (code vector) that minimizes error power. 1 is a noise codebook that stores M types of noise sequences C (each noise sequence is represented by an N-dimensional code vector), and 3 is a linear codebook of analysis order Np that performs filter calculation processing that simulates vocal tract characteristics. This is a predictive synthesis filter (LPC filter).

6 is a multiplication unit that calculates the cross-correlation RXC (=XT AC), 7 is a calculation unit that calculates the square of the cross-correlation RXC, and 8 is an autocorrelation RCC (=(AC)T (AC)).
9 is a calculation unit that calculates RXC2/RCC; 10 is a calculation unit that maximizes RXC2/RCC;
In other words, it is an error power evaluation section that determines a noise sequence (code vector) with the minimum error power and outputs a code specifying the code vector.

0010

[Problems to be Solved by the Invention] The main types of conventional codebook search processing are (1) filter processing for code vector C, (2) cross-correlation RXC calculation processing, and (3) autocorrelation RCC. There are three calculation processes. L
Assuming that the order of the PC filter 3 is Np and the dimension of vector quantization (code vector) is N, the amount of calculation required for each of (1) to (3) for one code vector is Np・N, N and N. Therefore, the amount of calculation required for codebook search per code vector is (Np+2)・N
becomes. A commonly used noise codebook 1 has about 40 dimensions and a codebook size of 1024 (N=40, M=1024), and the analysis order Np of the LPC filter 3 is 1.
Since it is about 0th order, one codebook search requires (10+2)·40·1024=480K product sums. However, K=103.

[0011] In order to perform such a codebook search every subframe (5 msec) of speech encoding, 96Mop
A huge amount of processing power called s (mega operations per second) is required, and even with the current fastest digital signal processor (allowable calculation amount of 20 to 40 Mops), it takes several chips to realize it in real time. There is a problem with it.

[0012] In addition, in order to store and hold the conventional noise codebook 1 as a table, N・M (=40・1024=
There is a problem in that a memory capacity of 40Kwords is required. Furthermore, in car phones and mobile phones, which are considered to be the applicable fields of speech encoders using A-b-S type vector quantization, miniaturization and low power consumption of the devices are essential conditions, and a large number of calculations are required. Both the size and the huge memory capacity are serious obstacles to realizing a speech encoder.

From the foregoing, it is an object of the present invention to provide a speech encoding system that can reduce the amount of calculation required for codebook search and also reduce the memory capacity required for storing the noisy codebook.

[0014]

[Means for Solving the Problems] FIG. 1 is a diagram illustrating the principle of the present invention. 11 is a delta codebook that stores and holds an initial vector C0 which is one reference noise sequence and delta vectors ΔC1 to ΔCL-1 (L=10) which are (L-1) types of delta noise sequences; 12 is a vocal tract; a linear predictive synthesis filter (LPC filter) that performs filter calculation processing that simulates characteristics;
13 is a filter output A of the initial vector obtained by performing filter calculation processing on the initial vector C0 and (L-1) types of delta vectors ΔC1 to ΔCL-1, respectively.
A storage unit that stores filter outputs AΔC1 to AΔCL-1 of C0 and (L-1) types of delta vectors, 14 a cross-correlation calculation unit that calculates cross-correlation RXC (=XT AC), and 15 an autocorrelation RCC (= (AC)T (AC)), and 17 is a minimum error noise sequence determination unit that determines a noise sequence (code vector) that maximizes RXC2/RCC, in other words, minimizes error power. Part, 18
is a speech encoding unit that encodes an input signal with data (code) specifying a noise sequence (code vector) with minimum error power.

[0015]

[Operation] Initial vector C0 which is one reference noise sequence in advance
Delta vectors ΔC1 to ΔCL-1 (L=10), which are (L-1) types of delta noise sequences, are stored in the delta codebook 11, and each of the delta vectors ΔC1 to ΔCL-1 is layered into the initial vector C0. The code vectors (code words) CO to C1022 of the (210-1) types of noise sequences are expressed in a tree structure by adding and subtracting each code vector, or the zero vector or -C is added to these code vectors.
0 vector and 210 noise sequence code vectors (
Code word) represents C0 to C1023. In this way, the initial vector C0 and (L-1) are stored in the delta codebook 11.
Type of delta vector ΔC1 to ΔCL-1 (L=10)
Just by storing 2L-1 (=210-1
=M-1) types of code vectors or 2L (=210=
M) types of code vectors can be generated, and the storage capacity of the delta codebook 11 can be set to L·N (=10·N), and the storage capacity of the conventional noise codebook is M·N (= 1
024・N).

Further, the initial vector C0 and (L-1) types of delta vectors ΔC1 to ΔCL-1 (L=10) are subjected to synthesis filter processing by the LPC filter 12 to obtain filter outputs AC0 and ( L-1) Types of delta vector filter outputs AΔC1 to AΔCL-1
(L=10) is determined and stored in the storage unit 13. Then, by adding and subtracting the filter output AΔC1 of the first delta vector to the filter output AC0 of the initial vector, filter outputs AC1 and AC2 for the code vectors C1 and C2 of the two types of noise sequences are calculated, and new By adding and subtracting the filter output AΔC2 of the second delta vector to the filter outputs AC1 and AC2 for the code vectors of each noise string calculated in
, C4; filter output AC3 to AC for C5 and C6
6, and similarly apply the filter output AΔCi-1 of the (i-1)th delta vector to the calculated filter output ACk, and apply the filter output AΔCi of the i-th delta vector to obtain 2, respectively. Filter output AC2k+1 for the code vector of the type of noise sequence,
By calculating AC2k+2, filter outputs of all code vectors are generated. In this way, the synthesis filter calculation process for each code vector C0 to C1022 can be performed using the initial vector C0 and (L-1)(L
=10) Types of delta vectors ΔC1 to ΔCL-1(L
= 10), and conventionally, filter processing requires Np・N・M (=10
What used to be 24·Np·N) times of product-sum operations can be reduced to Np·N·L (=10·Np·N) times.

Furthermore, the minimum error noise sequence determining unit 17 determines a noise sequence (code vector) with the minimum error power, and the audio encoding unit 18 outputs a code specifying the code vector, thereby generating audio. At the same time, the process of finding the code vector with the minimum error power is performed using the cross-correlation RXC (
=XTAC, T is a transposed matrix) and the ratio of the autocorrelation RCC (=(AC)(AC)) of the synthesis filter output is determined to be the maximum. Then, the synthesis filter calculation outputs AC2k+1 and AC2k+2 are expressed by a recurrence formula using the synthesis filter calculation output ACk of the previous layer and the current delta noise filter output AΔCi, as shown in the following formula: AC2k+1=ACk+AΔCi AC2k+2=ACk−AΔCi The cross-correlation RXC
(2K+2), RXC(2K+2) with the following formula RXC(2k
+1)=RXC(k)+XT (AΔCi)RXC(2
k + 2) = RXC (k) - XT (AΔCi), and the cross correlation calculation unit 14 uses the previous cross correlation RXC (k) to calculate the current cross correlation RXC (
2k+1) and RXC(2k+2). In this way, the cross-correlation between the filter outputs of all code vectors and the input signal X can be calculated by simply performing the cross-correlation calculation of the second term on the right side. In other words, what used to require M・N (=1024・N) times of product-sum calculations to obtain cross-correlation can now be done with L・N (=10・N) times. Therefore, the number of calculations can be significantly reduced.

Further, the autocorrelation calculation unit 15 calculates the current cross-correlation RCC (2k) using the autocorrelation RCC(k) of the previous layer.
+1), RCC(2k+2). In this way, the autocorrelation RCC can be calculated from the filter output AC0 of the initial vector and the filter outputs AΔC1 to AΔCL-1 of (L-1) types of delta vectors. ~(AΔCL-1)
2 and each filter output AC0, AΔC1 to AΔCL-1
The calculation can be performed using (L2-l)/2 types of cross-correlation between the two. In other words, conventionally, M・N (=1
024・N) times of product summation is now L(L
+1)·N/2 (=55·N) times of product-sum calculations, and the number of calculations can be significantly reduced.

[0019]

【Example】

Overall configuration FIG. 2 is a diagram showing the configuration of a speech encoding system according to the present invention. In the figure, 11 stores an initial vector C0 representing one reference noise sequence and delta vectors ΔC1 to ΔCL-1 (L=10) representing (L-1) types of delta noise sequences.
This is a delta codebook to be held, and the initial vector C0 and each delta vector ΔC1 to ΔCL-1 (L=10) are each expressed in N dimensions. That is, the initial vector and the delta vector are N-dimensional vectors each encoding the amplitude of N noises occurring in time series.

Reference numeral 12 denotes a linear predictive synthesis filter (LPC filter) that performs filter calculation processing that simulates vocal tract characteristics, and is composed of an IIR type filter of order Np, and is composed of an N×N square matrix A and a code vector C. A matrix calculation is performed and the code vector C is subjected to synthesis filter processing. II
The Np coefficients of the R-type filter vary based on the input audio signal and are each determined in a known manner. That is,
Since there is a correlation between adjacent samples of the input audio signal,
Find the correlation coefficient between samples, and use the correlation coefficient to calculate the
Find the partial autocorrelation coefficient called the parcoal coefficient,
The alpha coefficient of the IIR filter is determined from the filter coefficient, an N×N square matrix A is created using the impulse response sequence of the filter, and the code vector C is subjected to synthesis filter processing.

13 is a filter output AC0 obtained by performing filter calculation processing on the initial vector C0 representing the reference noise sequence and the delta vectors ΔC1 to ΔCL-1 representing (L-1) types of delta noise sequences. , AΔ
14 is a cross-correlation calculation unit that calculates cross-correlation RXC (=XT AC);
Reference numeral 5 denotes an autocorrelation calculation unit that calculates the autocorrelation RCC (=(AC)T (AC)), and 16 represents a calculation unit that calculates the ratio of the squared cross-correlation to the autocorrelation.

Since the error power |E|2 is expressed by equation (3), the code vector C that minimizes the error power is (3)
This maximizes the second term on the right side of the equation. Therefore, the calculation unit 16 includes a square calculation unit 16a and a division unit 16b, and has the following formula F(X,C)=RXC2/RCC (8
) is calculated.

17 is a minimum error noise sequence determination unit that determines a noise sequence (code vector) that maximizes RXC2/RCC, in other words, the error power is minimum; 18 is a noise sequence (code vector) that minimizes error power; This is an audio encoding unit that encodes an input signal with a code that specifies a code vector).

Structure of Code Words of Delta Codebook FIG. 3 is an explanatory diagram of the structure of code words (code vectors) in the delta codebook 11 according to the present invention. The delta codebook 11 contains 1
One initial vector C0 and (L-1) types of delta vectors ΔC1 to ΔCL-1 (L=10) are stored, and each delta vector ΔC1 to ΔCL-1 is added to the initial vector C0 for each layer. , (210-1) types of code vectors C are sequentially created in a tree structure depending on whether
0 to C1022 are expressed, and a zero vector (each element is zero) is added to these code vectors to express 210 code vectors C0 to C1023. In this way, the relationship between each code vector (code word) can be expressed as follows: C1=C0+ΔC1 C2=C0-ΔC1 C3=C1+ΔC2 (=C0+ΔC1+ΔC2)C
4=C1-ΔC2 (=C0+ΔC1-ΔC2)C5=C
2+ΔC2 (=C0-ΔC1+ΔC2)C6=C2-Δ
C2 (=C0-ΔC1-ΔC2)・・・・・・・・・・
・・・・・・・・・ C511 =C255+ΔC9(=C0+ΔC1+
ΔC2+...+ΔC9) C512 =C255-
ΔC9 (=C0+ΔC1+ΔC2+...-ΔC9)
C1021=C510+ΔC9(=C0−ΔC1−Δ
C2-...+ΔC9) C1022=C510-Δ
It is expressed as C9 (=C0-ΔC1-ΔC2-...-ΔC9), and generally C2k+1=Ck+ΔCi (9)C2k+
2=Ck-ΔCi (10) It can be expressed as a recurrence formula. That is, the delta codebook 11 contains an initial vector C0 and (L-1) types of delta vectors ΔC1 to Δ
Just by storing CL-1 (L=10), 2
It is possible to generate code vectors of L (=210) types of arbitrary noise sequences, and the storage capacity of the codebook 11 can be reduced to L・N(
=10·N), which can be significantly reduced compared to the storage capacity M·N (=1024·N) of the conventional noise codebook.

Filter processing code for code vector
Synthesis filter calculation outputs AC2k+1 and AC2k+2 for code vectors (code words) C2k+1 and C2k+2 are
Synthesis filter calculation output ACk for the code vector Ck one layer before and filter output A of the current delta vector
Using ΔCi, the following formula AC2k+1=A(Ck+ΔCi)=ACk+AΔ
Ci (11) AC2k+2=A(Ck
−ΔCi)=ACk−AΔCi (12)
However, i=1, 2,...L-1, 2i-1≦k
It can be expressed by a recurrence formula such as <2i −1. Therefore, the initial vector C0 and (L-1) types of delta vectors Δ
C1 to ΔCL-1 (L=10) are subjected to synthesis filter processing by the LPC filter 12 to obtain the initial vector filter output AC0 and (L-1) types of delta vector filter outputs AΔC1 to AΔCL-1 (L =10) and stores it in the storage unit 13, it is possible to recursively perform filter processing on the code vector of the entire noise sequence as follows.

That is, (1) By adding and subtracting the filter output AΔC1 of the first delta vector for each dimension to the filter output AC0 of the initial vector, a filter is created for the code vectors C1 and C2 of two types of noise sequences. Output AC1, AC2 can be calculated, and (2)
Newly calculated filter calculation outputs AC1, AC2
By adding and subtracting the filter output AΔC2 of the second delta vector, the filter outputs AC3 to AC6 for a total of four types of code vectors C3, C4, C5, and C6 can be calculated. 3) Filter output A calculated by applying filter output AΔCi-1 of the (i-1)th delta vector
Ck is the filter output AΔC of the i-th delta vector.
2 types of filter output AC2k by applying i
+1, AC2k+2, 2L (=21
0) can generate a filter output for the code vector of the total noise sequence.

That is, by using the tree-structured delta codebook 11 of the present invention, it becomes possible to recursively perform filter processing on each code vector using equations (11) and (12), and initial vectors C0 and ( L-1) Types of delta vectors ΔC1 to ΔCL-1 (L=10) are simply subjected to synthesis filter processing, and by adding them while changing their signs, the codes of all noise sequences can be obtained. The filter processing for the vector is obtained.

In fact, as will be described later, in the case of the delta codebook of the present invention, when calculating cross-correlation RXC and autocorrelation RCC, filter calculation outputs for all code words (code vectors) are not required, and the initial vector C0 and (L
−1) Types of delta vectors ΔC1 to ΔCL−1 (L=
Only the filter calculation processing results for 10) are required.

Therefore, the conventional synthesis filter arithmetic processing for the code vectors C0 to C1023 of the total noise sequence can be changed to the synthesis filter for the initial vector C0 and (L-1) types of delta vectors ΔC1 to ΔCL-1 (L=10). This can be reduced to arithmetic processing. Therefore, in the present invention, Np・N・M (=1024・Np・N) times of product summation are required for filter processing, but in the present invention, N・N・L (=10・
It can be reduced to Np・N) times of product summation.

Calculation of cross-correlation RXC Synthesis filter calculation outputs AC2k+1 and AC2k+2 are 1
Using the previous synthesis filter calculation output ACk and the current delta vector filter output AΔCi, (11), (12
), the cross-correlation RXC (2
k+1), RXC(2k+2) is the following formula RXC(2k
+1)=XT(AC2k+1)
=XT(ACk)+XT(AΔCi)
=RXC(k)+XT(AΔCi)
(13) RXC (2k +2)
=XT (AC2k+2) =XT
(ACk)−XT(AΔCi)=
RXC(k)-XT(AΔCi)
It can be expressed by a recurrence formula as shown in (14). Therefore, in the cross-correlation calculation unit 14, the cross-correlation RXC(k
) to calculate the current cross-correlation RXC (2k+1), RXC
(2k+2) can be calculated. In this way, it is possible to calculate the cross-correlation between the filter output and the input signal . In other words, it takes M・N (=102
According to the present invention, L
・It is possible to complete the product-sum operation N (=10・N) times, and the number of calculations can be significantly reduced.

In FIG. 2, 14a is (13), (
14) A multiplication unit that calculates the second term XT (AΔCi) on the right side of the equation, 14b is a sign adding unit that generates +1 and -1, 14c
is a multiplier that multiplies the sign ±1 and gives a sign to the second term on the right side, and 14d is the previous cross-correlation RXC(k) (the first term on the right side).
The delay unit 14e that delays storage of the item) for a predetermined period of time is (13)
, by adding the first and second terms on the right side of equation (14) to obtain the current cross-correlation RXC(2k+1), RXC(2k+2)
This is an adder that outputs .

Calculation of autocorrelation RCC Synthesis filter calculation output AC2k+1, AC2k+2 is 1
Using the synthesis filter calculation output ACk of the previous layer and the current delta vector filter output AΔCi, (11), (1
2) When expressed as a recurrence formula, the code of each noise sequence is
The autocorrelation RCC for the vector is expressed by the following equation. That is, RCC(0)=(ACO)T(ACO
) AC1=AC0+AΔC1 AC2=AC0−AΔC1 RCC(1)=(AC0)T(AC0)+(AΔC
1) T(AΔC1)+2(AC0)T(AΔC1)
RCC(2)=(AC0)T(AC0)+(AΔC1)
T(AΔC1)-2(AC0)T(AΔC1) A
C3 =AC1+AΔC2 =AC0+AΔC1+AΔ
C2 AC4 =AC1-AΔC2 =AC0+AΔ
C1-AΔC2 AC5 =AC2+AΔC2 =A
C0-AΔC1+AΔC2 AC6 =AC1-AΔ
C2 = AC0-AΔC1-AΔC2 RCC (3)
=(AC0)T(AC0)+(AΔC1)T(AΔC1
)+(AΔC2)T(AΔC2)
+2(AC0)T(AΔC1)+2(AΔC1)T(A
ΔC2)+2(AΔC2)T(AC0) RCC(4
)=(AC0)T(AC0)+(AΔC1)T(AΔC
1)+(AΔC2)T(AΔC2)
+2(AC0)T(AΔC1)-2(AΔC1)T(
AΔC2)-2(AΔC2)T(AC0) RCC(
5)=(AC0)T(AC0)+(AΔC1)T(AΔ
C1)+(AΔC2)T(AΔC2)
-2(AC0)T(AΔC1)-2(AΔC1)T
(AΔC2)+2(AΔC2)T(AC0) RCC
(6)=(AC0)T(AC0)+(AΔC1)T(A
ΔC1)+(AΔC2)T(AΔC2)
-2(AC0)T(AΔC1)+2(AΔC1)
It is expressed as T(AΔC2)-2(AΔC2)T(AC0), and generally RCC(2k+1)=RCC(k)+(AΔCi)
T(AΔCi)+2AΔCi・ACk (15)
RCC(2K+2)=RCC(k)+(AΔCi)T(
AΔCi)−2AΔCi·ACk (16)

In other words, the autocorrelation RCC(k
), the current autocorrelation of AΔCi (AΔCi)T (A
ΔCi), and AΔCi and AC0, A
By adding the cross-correlations of ΔC1 to AΔCi-1 while changing the signs, the cross-correlation RCC (2k+1), R
CC(2k+2) can be calculated. In this way, the autocorrelation RCC can be calculated from the filter output AC0 of the initial vector and (
L-1) Type of delta vector filter output AΔC1
~AΔCL-1 total L autocorrelations (AC0)2,
(AΔC1)2 to (AΔCL-1)2 and the mutual (L2-l) between each filter output AC0, AΔC1 to AΔCL-1.
)/can be calculated using two types of cross-correlation. That is,
Conventionally, calculating autocorrelation required M・N (=1024・N) times of product summation, but now it can be calculated using L(L+1)・N/2 (=55・N) times. The number of calculations can be significantly reduced. In addition, in FIG. 2, 15a is (15),
Autocorrelation (AΔCi)T(A
The autocorrelation calculation unit 15b that calculates ΔCi) is (15)
, a cross-correlation calculation unit that calculates each cross-correlation in equation (16), 15c a cross-correlation synthesis unit that adds each cross-correlation with a predetermined sign, and 15d an autocorrelation RCC of the previous layer.
(k), autocorrelation (AΔCi), T(AΔCi), and cross-correlation, and calculates equations (15) and (16). This is a delay section that stores and delays.

Overall operation Initial vector C0, which is one reference noise sequence, and (L-
1) Delta vector ΔC1 which is a type of delta noise sequence
ΔCL-1 (L=10) is stored in the delta codebook 11, and the linear LPC filter 12 uses the initial vector C
0 and (L-1) types of delta vectors ΔC1 to ΔCL
-1 (L=10) is subjected to synthesis filter processing and the filter outputs AC0, AΔC1 to AΔCL-1 (L=1
0) is determined and stored in the storage unit 13.

In this state, with i=0, the cross-correlation calculation section 14 calculates the cross-correlation RXC(0) (=XTAC0), and the autocorrelation calculation section 15 calculates the autocorrelation RCC(0)(=(AC0)T (AC0)),
Using these cross-correlation and autocorrelation, the calculation unit 16 calculates (8)
F(X,C) (=RXC2/RCC) is calculated using the formula.

The minimum error noise sequence determining unit 17 calculates the calculated F
(X, C) and the maximum value of F(X, C) up to that point Fmax(
The initial value is 0), and if F(X, C) > Fmax, update Fmax as F(X, C) → Fmax, and use a code to specify the noise sequence (code vector) that gives Fmax. updates the code up to that point.

If the above processing is performed on 2i (=20) code words (code vectors), i=1, and (13)
The cross-correlation is calculated according to the formula (k=0, i=1), the autocorrelation is calculated according to the formula (15), and the calculation unit 16 calculates the formula (8) using these cross-correlation and autocorrelation. .

The minimum error noise sequence determining unit 17 calculates the calculated F
(X, C) and the maximum value of F(X, C) up to that point Fmax(
The initial value is 0), and if F(X, C) > Fmax, update Fmax as F(X, C) → Fmax, and use a code to specify the noise sequence (code vector) that gives Fmax. updates the code up to that point.

Then, according to equation (14) (where k=0
, i=1), calculate the autocorrelation according to equation (16), and use these cross-correlation and autocorrelation to calculate equation (8) in the calculation section 16.

The minimum error noise sequence determination unit 17 calculates F(X, C) calculated in the same way and the maximum value Fm of F(X, C) up to that point.
ax (initial value is 0), F(X,C)>Fmax
If so, Fmax is updated as F(X,C)→Fmax, and the previous code is updated with a code that specifies the noise sequence (code vector) that provides the Fmax.

If the above process is performed on 2i (=21) code words (code vectors), i=2, and the same process as above is repeated, and the above process is performed on all 210 code vectors. If this is done, the speech encoder 18 outputs the latest code stored in the minimum error noise sequence determining section 17 as the speech code of the input signal.

Comparison of the conventional method and the method of the present invention in processing amount and memory capacity FIG. 4 is a chart comparing the conventional method and the method of the present invention in terms of processing amount (amount of calculations) and memory capacity. The amount of calculation is less than 1/70 of that of the conventional method, and the memory capacity is 1/100, making it possible to significantly reduce the amount of calculation and memory capacity.

[0043] In the above, each delta vector ΔC1 to Δ
By adding or subtracting CL-1 (L=10) to the initial vector C0 for each layer, (210-1) types of code vectors CO to C10 are sequentially created in a tree structure.
22 and added zero vectors to these code vectors to express 210 types of code vectors C0 to C1023. However, it is not always necessary to add zero vectors, and -1 can be added to initial vector C0 instead of zero vectors. It is also possible to express 210 types of code vectors by changing the multiplied -C0 vector.

The present invention has been explained above using examples, but
The present invention can be modified in various ways according to the gist of the present invention as described in the claims, and the present invention does not exclude these modifications.

[0045]

According to the present invention, one initial vector C0 and (L-1) types of delta vectors ΔC1~
ΔCL-1 (L=10) is stored in the delta codebook, and delta vectors ΔC1 to ΔCL-1 are added to and subtracted from the initial vector C0 for each layer to sequentially create a tree structure (210-1) Since it is configured to express different types of code words (code vectors) C0 to C1022, or to express 210 code vectors C0 to C1023 by adding a zero vector or -C0 vector to these, the delta codebook Initial vector C0 and (L-1)
Type of delta vector ΔC1 to ΔCL-1 (L=10)
Just by storing 2L-1 (=210-1
=M-1) types or 2L (=210=M) types of code vectors can be generated, and the storage capacity of the codebook can be reduced to L.
・N (=10・N), which can be significantly reduced compared to the storage capacity M・N (=1024・N) of the conventional noise codebook.

According to the present invention, the initial vector C0 and (L-1) types of delta vectors ΔC1 to ΔCL-1
(L=10), performs synthesis filter processing and filter output AC0, AΔC1 to AΔCL-1 (L=10)
Find and memorize the filter output A of the initial vector.
Filter output AΔ of the first delta vector for C0
By adding and subtracting C1, two types of codes are obtained.
Filter outputs AC1, A for vectors C1, C2
C2 is calculated, and each newly calculated filter output AC1
, AC2 is the filter output AΔC of the second delta vector.
By adding and subtracting 2, the filter outputs AC3 to AC6 for the two types of code vectors C3, C4; C5, C6 are calculated, and the (i-th
−1)th delta vector filter output AΔCi−
The filter output ΔCi of the i-th delta vector is applied to the filter output ACk calculated by applying 1 to the filter output AC2k+ for each of the two types of noise sequences.
1, AC2k+2 to calculate the code of the total noise sequence.
Since the configuration is configured to generate filter outputs for code vectors, the code vectors C0 to C1023 of each noise sequence
The synthesis filter calculation process for the initial vector C0 and (L-1) (L=10) types of delta vectors ΔC1
~ΔCL-1 (L=10) can be reduced to the synthesis filter calculation process, and conventionally, the filter process requires Np
・What used to be N·M (=1024·Np·N) times of product-sum calculations can be reduced to N·N·L (=10·Np·N) times.

Furthermore, according to the present invention, a noise string (code vector) with the minimum error power is determined, and a code specifying the noise string (code vector) is output for speech encoding. Then, the process of finding the noise sequence with the minimum error power is performed by finding a code vector that maximizes the ratio of the square of the cross-correlation RXC between the synthesis filter calculation output AC and the input signal X to the autocorrelation RCC of the synthesis filter output. Since this is reduced to the calculation, the code vector with the minimum error power can be easily found.

Furthermore, the synthesis filter calculation output AC2k+1,
Synthesis filter calculation output ACk of AC2k+2 one layer before
By using the filter output AΔCi of the current delta vector and expressing it in a recurrence formula as shown below, AC2k+1=ACk+AΔCi AC2k+2=ACk−AΔCi, the cross-correlation RXC
is expressed as a recurrence formula, and the current cross-correlation RXC(2k+1), RXC(2
k + 2), so the conventional calculation of cross-correlation, which required M・N (=1024・N) times of product summation, is now L・N (=10・N). This makes it possible to perform only one product-sum operation, and the number of calculations can be significantly reduced.

Moreover, the autocorrelation RCC(k) of one layer before
The current cross-correlation RCC(2k+1), RCC(
2k+2), the autocorrelation R
CC is the initial vector filter output AC0 and (L-1)
Types of delta vector filter outputs AΔC1 to AΔC
L-1 total L autocorrelations (AC0)2, (AΔC
1) 2 to (AΔCL-1)2 and each filter output AC0
, AΔC1 to AΔCL-1 using (L2-l)/2 types of cross-correlation, and conventionally, calculating the autocorrelation requires M・N (=1024・N) times of product summation. What used to be the case can now be reduced to L(L+1)·N/2 (=55·N) times of product-sum operations, and the number of calculations can be significantly reduced.

As described above, according to the present invention, the amount of calculation and memory capacity can be significantly reduced, resulting in miniaturization, cost reduction, and reduction in power consumption of the device.

[Brief explanation of the drawing]

FIG. 1 is a diagram explaining the principle of the present invention.

FIG. 2 is a configuration diagram of an embodiment of the present invention.

FIG. 3 is an explanatory diagram of code words of the codebook according to the present invention.

FIG. 4 is an explanatory chart comparing the conventional method and the method of the present invention with respect to the amount of calculation and memory capacity.

FIG. 5 is an explanatory diagram of a sound generation mechanism.

FIG. 6 is a diagram of a conventional speech encoding device.

FIG. 7 is a configuration diagram of a conventional noise codebook search processing section.

[Explanation of symbols]

11...Delta codebook, 12...Linear prediction synthesis filter (LPC filter), 1
3. Storage unit for storing filter output, 14. Cross correlation calculation unit, 15. Autocorrelation calculation unit, 17. Minimum error noise sequence determination unit, 18. Speech encoding unit.

Claims (8)

[Claims] 1. A code vector that minimizes the error between the input signal and the reproduced signal obtained by performing a synthesis filter process that simulates vocal tract characteristics on a code vector representing a noise string, and In a speech encoding method that encodes an input signal using a code that specifies a vector, an N-dimensional initial vector representing a reference noise sequence and a delta vector representing (L-1) types of N-dimensional delta noise sequences are determined in advance. It is characterized by storing code vectors in a delta codebook and sequentially expressing code vectors of 2L-1 types of noise sequences in a tree structure by adding or subtracting delta vectors from initial vectors for each layer. Audio encoding method. 2. The method according to claim 1, wherein code vectors of 2L types of noise sequences are expressed by adding a zero vector to the 2L-1 types of code vectors.
The audio encoding method described. 3. A code vector of 2L types of noise sequences is expressed by adding a vector obtained by multiplying an initial vector by -1 to the 2L-1 types of code vectors. The audio encoding method described in Section 1. 4. Perform synthesis filter calculation processing on the initial vector and the (L-1) types of delta vectors to obtain the filter output of the initial vector and the filter output of the (L-1) type of delta vectors, and By adding and subtracting the filter output of the first delta vector for each dimension from the filter output of the vector, the filter output for the code vector of two types of noise sequences is calculated, and the filter output for each newly calculated noise sequence is calculated. By adding and subtracting the filter output of the second delta vector to the filter output for the code vector, the filter output for the code vector of the two types of noise sequences is calculated, and similarly, the filter output for the code vector of the (i-1)th The filter output of the i-th delta vector is applied to the filter output of the code vector of each noise sequence calculated by applying the filter output of the delta vector, and the filter output is calculated for the code vector of two types of noise sequences. Assuming that a filter output is generated for the code vector of the total noise sequence by 4. The speech encoding method according to claim 1, 2, or 3, characterized in that the speech encoding method results in the following. 5. The audio encoding method according to claim 4, wherein the audio encoding method performs audio encoding by determining a code vector that minimizes the square of the difference between the reproduced signal and the input signal, and outputting a code specifying the code vector. . 6. A synthesis filter calculation output AC (A is a matrix for synthesis filter calculation, C is a code vector) minimizes the square of the difference between the reproduced signal and the input signal.
and the square of the cross-correlation between the input signal X (XT AC)2(
T is the transposed matrix) and the autocorrelation of the synthesis filter calculation output ((
6. The speech encoding method according to claim 5, wherein the method results in finding a code vector with a maximum ratio of AC)T(AC)). [Claim 7] By expressing the synthesis filter calculation output AC in a recurrence formula using the synthesis filter calculation output of one layer before and the current delta vector, the cross-correlation is expressed and calculated in a recurrence formula. The speech encoding method according to claim 6, characterized in that: 8. By expressing the synthesis filter calculation output AC in a recurrence formula using the synthesis filter calculation output of the previous layer and the current delta vector, the autocorrelation can be expressed as the filter calculation output of the initial vector and (L− 1) Calculation is performed by expressing using a total of L types of autocorrelations of filter calculation outputs of different types of delta vectors and (L2-L)/2 types of cross-correlations between the respective filter calculation outputs. The audio encoding method according to claim 6.