KR100316304B1 - High speed search method for LSP codebook of voice coder - Google Patents

Abstract

The present invention relates to a method for reducing the amount of computation required in a codebook search without degrading the spectral distortion (SD) performance in quantizing LSP coefficients of a speech coder for a multimedia service.

Since the speech coder according to the present invention quantizes an error signal that is a difference between the LSP coefficient vector from which the DC component is removed and the vector predicted by the first order predictor, the quantization of the LSP coefficients cannot use ordering. So the present invention, the first estimator is configured as a target vector and the error code book by giving the added P _DC to the target vector and the error code book having the sequence properties while in use. Each of the three codebooks reorganized to have an ordering property selects one specific column used to determine the search range, and sorts the entire codebook in descending order based on the element values of the selected column. In addition, the search range of the codebook is reduced by comparing the element values of the specific column arranged in each codebook with the element values before and after the corresponding column in the target vector, and obtaining Elm only for vectors satisfying the ordering property. The final search range of the codebook is determined through forward and reverse comparison.

Description

High speed search method for LSP codebook of voice coder

The present invention relates to an algorithm for processing the LSP (Line Spectral Pairs) coefficient quantization method of G.723.1, which is an encoder used to compress a speech signal at a low speed in a multimedia application.

It is very important to efficiently quantize the LPC (Line Predictive Coding) coefficient, which represents the short-term correlation between speech signals, for high quality speech coding in a low speed speech encoder. In G.723.1, LPC coefficients are transmitted every transmission frame (30msec). The transfer function of the LPC synthesis filter by these LPC coefficients is given by the following equation.

In Equation 1, {a _i } is an LPC coefficient, and p is an order of the LPC coefficient.

LPC coefficients are often quantized after being transformed into LSP coefficients that are mathematically equivalent to the LPC coefficients, have good quantization characteristics, and are easy to examine the stability of the synthesis filter. To obtain the LSP coefficients from the LPC coefficients, the following two polynomials are generated using the regression equation of the analysis filter A (Z).

Where q _i is the LSP coefficient value, Is the root of P (z) and Q (z). _i = q has a relationship cosw _i, where w _i is the La line spectral frequencies (LSF).

The polynomials of P (z) and Q (z) above have the following important characteristics:

1) All roots of P (z) and Q (z) are on a unit circle.

2) All roots of P (z) and Q (z) are staggered.

Based on the above two characteristics, some important properties of LSP coefficients are as follows.

1) Synthetic filter is stable when LSP coefficient meets the following order quality.

0 <ω ₁ <ω ₂ <.... <ω _p <π

2) The LSP coefficient represents the formant frequency and formant bandwidth of the voice. That is, the closer the LSP coefficient is, the sharper the formant region is.

3) The interpolated LSP coefficient also satisfies the ordering quality.

4) The LSP coefficient has a small dynamic range, which is advantageous for quantization.

To efficiently quantize these LSP coefficients, G.723.1 uses Predictive Split Vector Quantization (PSVQ).

1 is a block diagram of an LSP quantization scheme of a conventional G.723.1 speech coder.

As shown in FIG. 1, the order of the predictor is first order, and p is the LSP coefficient vector. If p is subtracted from the average vector (P _DC ), the DC component of the LSP coefficients, then p 'is obtained, and the decoded LSP vector ( ) And the first predictor coefficient b (= 12/32), the prediction vector p 'and the LSP error vector e are obtained as follows.

N is the current frame. The LSP error vector (e) is divided into three subvectors having 3, 3, and 4 dimensions, and each subvector is vector quantized into 8 bits, where the optimal code vector value for each subvector is Choose to minimize the reference E _{l, m} .

, 0≤m≤2, 1≤ l ≤256

W _m is a weighting matrix for the m th subvector, and is obtained from an unquantized LSP vector p. And the quantized LSP vector p _{l, m} can be represented by the following equation.

, 0≤m≤2, 1≤ l ≤256

Substituting Equation 8 into Equation 7 can be expressed as follows.

, 0≤m≤2, 1≤ l ≤256

According to Equation 9, since E _{l, m} is expressed as an expression for e _m and e _{l, m} , the amount of computation can be greatly reduced when E _{l, m} is calculated for the actual 256 codebooks. Where e _m is a target vector for codebook searching and e _{l, m} corresponds to the l th error code vector of the m th subvector. Therefore, we use these values to transmit the codebook index l through the channel that minimizes E _{l, m} .

The prior art as described above has a problem in that most of the calculation amount required when quantizing the LSP coefficients is used when searching for an optimal code vector for three subvectors.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and an object of the present invention is to provide a fast search method that reduces the amount of search computation by reducing the range of code vector to be searched using the order property of LSP coefficient values.

As a technical idea for achieving the object of the present invention, in the method of reducing the amount of computation required for codebook search in the quantization of LSP coefficients of a G.723.1 speech coder, the difference between the LSP coefficient vector (P) and the prediction vector and the mth sub A first step of obtaining an error criterion (E _{l, m} ) by using the sum of the l th error code vector of the vector and the DC component mean vector (P _DC ) of the LSP coefficients, and the m th subvector of the first step. a second step of precomputing the sum of the l-th error code vector and the DC component mean vector (P _DC ) of the LSP coefficients to create a new codebook for the three subvectors, and the specific code used to determine the search range in the new codebook. A third step of selecting a column and sorting in descending order based on element values of the selected column, and a fourth step of determining a search range by using an order property of a given target vector and aligned code vectors. The invention as claimed is presented.

1 is a block diagram of an LSP quantization scheme of a conventional G.723.1 speech coder.

2 is a block diagram of an LSP quantization scheme considering ordering properties according to the present invention.

3 is a flowchart illustrating a fast code vector search method according to the present invention.

4 is a conceptual diagram illustrating three codebooks for three subvectors of 3, 3, and 4 dimensions according to the present invention.

5A is a block diagram of a forward comparison method when selecting an optimal code vector according to the present invention.

5B is a block diagram of a backward comparison method when selecting an optimal code vector according to the present invention.

<Description of the symbols for the main parts of the drawings>

1: Split Vector Quantizer (SVQ)

2: 1st order predictor p: LSP vector

p _DC : Average vector of DC components of LSP coefficients

p ': p-p _DC

Decoded LSP vector of previous frame

e: LSP error vector : Quantized LSP Vector

r: p-

Hereinafter, with reference to the accompanying drawings with respect to the configuration and operation of the embodiment of the present invention will be described.

The most researched speech coder is the CELP (Code Excited Linear Prediction) structure. This method can achieve good sound quality at data rates of around 4.8kbps and is standardized in various applications through various international standardization organizations. In particular, there is an increasing demand for a codec that provides high sound quality at low data rates for video conferencing over mobile communication and the Internet. Among these CELP series vocoders, G.723.1 was developed as a voice standard vocoder for multimedia services.

G.723.1 has a dual data rate of 5.3 / 6.3kbps and is used in internet phones and electronic answering machines, which are commercially available as separate communication, and provide excellent sound quality compared to low data rates. In addition, it is more applicable than other vocoder standards because it uses two data rates for optimal transmission environment.

2 is a block diagram of an LSP quantization scheme considering ordering properties according to the present invention.

3 is a flowchart illustrating a fast code vector search method according to the present invention.

As shown in FIG. 2 and FIG. 3, the present invention provides a method for reducing the amount of computation required for codebook searching in L. coefficient quantization of a G.723.1 speech coder, the difference between an LSP coefficient vector and a prediction vector, an error code vector, and an LSP. The first step of obtaining an error criterion using the sum of the DC component mean vectors of the coefficient values, and the third sub vector by previously calculating the sum of the error code vector and the DC component average vector of the LSP coefficient values in the first step. A second process of creating a new codebook for the fields, a third process of selecting a particular column used to determine the search range in the new codebook and sorting in descending order based on element values of the selected column, and a code vector aligned with a given target vector. The fourth step is to determine the search range using the ordering properties of the fields.

A schematic process of the present invention is as follows.

First then sorted in descending order according to the sum of the pre-P _DC average vector codebooks for the three sub-vectors to the vector code values within a specified heat element size. In addition _{, since} the optimal code vector value that minimizes E _{l, m} in three subvectors has a value almost similar to that of the target vector, it can be assumed that this value has an ordering quality. Under this assumption, the element values of a particular column of codebooks sorted in descending order are compared with the element values adjacent to a particular column in the target vector. So calculate the number of E _{l, m} value computation for only the code vector to satisfy the order, and the nature or not the code vectors are omitted E _{l, m} the value calculated. When searching each codebook, if the search target is reduced by using the ordering quality of the LSP vector, a large amount of computation can be reduced. This method is particularly effective for SVQ (Split Vector Quantization), and is very effective for predeictive SVQ structures that quantize prediction error values.

Specific processes used in the present invention are as follows. First, in G.723.1, the target vector e _m and the corresponding codebook correspond to error vectors and thus have no ordering. Thus, Equation 7 is changed into Equation 10 below to form a target vector having an ordering quality.

, 0≤m≤2, 1≤ l ≤256

In Equation 10, E _{l, m} is and You can get it by using the value, where If γ _m , then P _m did not remove the P _DC value. Since is predicted from the error value, the displacement is small. Thus, γ _m has an ordering property and becomes a target vector of codebook search. Is c _{l, m} , this value is the existing error codebook. Is the result of adding P _DC with ordering _, so c _{l, m} will have ordering again Since and P _DC are both fixed values, they are precomputed to replace the existing codebook. The values of E _{l, m are} calculated using r _m and c _{l, m} , and the codebook index l that minimizes E _{l, m} is transmitted through the channel. Referring to FIG. 2 configured to use the result of Equation 10, the structure is substantially the same as that of FIG. 1 showing a conventional method, but the process of subtracting P _DC which is a DC component of the LSP coefficient value from the LSP vector p is omitted. The codebook for the frames is replaced with a new codebook plus P _DC value in advance. Then both the target vector and the corresponding codebook are ordered.

4 is a conceptual diagram illustrating three codebooks for three subvectors of 3, 3, and 4 dimensions according to the present invention.

As shown in FIG. 4, for three codebooks for three subvectors of 3, 3, and 4 dimensions, the first codebook is based on the element values of the first column, and the second codebook is based on the element values of the second column. , And the third codebook sorts the codevectors in descending order based on the element values in the first column. The ordered codebooks and target vectors maintain the ordering quality because they contain DC components. Therefore, by using the following equations (11) and (12), the codebook is searched with the concept of excluding code vectors that violate the order quality by comparing the element values of a specific column of the codebook with the element values before and after the corresponding column in the target vector. Determine the range.

r _{n + l} c _{l, n} , 1≤ l ≤256, 0≤n≤8

r _nl <c _{l, n} , 1≤ l ≤256, 1≤n≤9

Here, as shown in Equation 11, comparing element values of the nth column of a code vector with elements of the (n + l) th column of the target vector to determine whether the order property is satisfied is called forward comparison. As shown in Equation 12, comparing the element value of the nth column of the code vector with the element value of the (n-1) th column of the target vector is called backward comparison.

4 illustrates the fast search method described above. f1, f2, f3 and b1, b2 represent the code vector element value and target vector element value used for the forward and reverse comparison, respectively. Here, since each codebook is sorted in descending order, only the starting point that satisfies the ordering property is obtained in the forward comparison, and all the rest satisfy the ordering property of the forwarding order. In the reverse comparison, an endpoint satisfying the ordering quality is obtained. The codebook search to find the actual E _{l, m} values is performed only in the range of the start and end points.

FIG. 5 is a flowchart illustrating a process of obtaining a specific starting point and an ending point of a set of code vectors satisfying an ordering quality for a given target vector according to the present invention. FIG. 5A shows a forward comparison and FIG. 5B shows a backward comparison, respectively. .

As shown in FIG. 5A, the process of determining a codebook search starting point includes calculating a LSP vector p (S100), calculating a target vector r = p-p ′ (S110), and initializing the variable i to 0. In step S120, comparing and determining the sizes of r _{n + 1} and C _{j + 16, n} (S130), and in step S130, the size of r _{n + 1} is smaller than C _{j + 16, n.} In the case of increasing the variable i by 64 (S140), and when the size of r _{n + 1} is greater than C _{j + 16, n} storing the variable i (S150), variable j is the variable stored variable i Initializing to a value (S160), comparing and determining the size of r _{n + 1 and} the size of C _{j + 16, n} (S170), and the size of r _{n + 1} in the step (S170) is C _j Increasing the variable j by 16 when it is smaller than _{+ 16, n} (S180) _; storing the variable j when the size of r _{n + 1} is larger than C _{j + 16, n} (S190), and variable k Initializing the variable to the stored variable j value (S2) 00), comparing the size of r _{n + 1 with} the size of C _{k + 4, n} (S210), and in step S210, the size of r _{n + 1} is smaller than C _{k + 4, n.} In the case of increasing the variable k by 4 (S220), and if the size of r _{n + 1} is greater than C _{k + 4, n} storing the variable k (S230), and the variable m variable variable k Initializing to a value (S240), comparing and determining the size of r _{n + 1 and} the size of C _{m + 1, n} (S250), and the size of r _{n + 1} in the step (S250) is C _m Increasing the variable m by 1 when it is smaller than _{+ 1, n} (S260) _; storing the variable m + 1 when the size of r _{n + 1} is larger than C _{m + 1, n} (S270); It consists of a step (S280) of setting the m + 1 calculated through the above process as a starting point.

As shown in FIG. 5B, the process of setting a codebook search end point may include calculating an LSP vector p (S300) and a target vector r = p −. Calculating (S310), initializing the variable i to 257 (S320), comparing and determining the sizes of r _n-1 and C _{j-16, n} (S330), and the process (S330) in the case size is C _j-16, is less than _n the process of reducing the variable i by 64 as the (S340), and the magnitude of _{r n-1 C j-16} , greater than _n variable i of r _n-1 Storing (S350), initializing the variable j to the stored variable i value (S360), comparing and determining the size of r _{n-1 and} the size of C _{j-16, n} (S370); In step S370, when the size of r _n-1 is smaller than C _{j-16, n,} reducing the variable j by 16 (S380), and the size of r _n-1 is C _{j-16, n} If a larger process of storing the variable j (S390), and the process (S400) to initialize the variable k to the variable of the stored variable j value, comparing the size of the size and C _{k-4, n} of r _n-1 is determined In step S410, and the size of r _n-1 in the step (S410) is smaller than C _{k-4, n} Is a step of reducing the variable k by 4 (S420), storing a variable k when the size of r _n-1 is greater than C _{k-4, n} (S430), and storing the variable m as the stored variable k. Initializing to a value (S440), comparing and determining the size of r _{n-1 and} the size of C _{m-1, n} (S450), and the size of r _n-1 in the step (S450) is C _m Reducing the variable m by 1 when smaller than _{-1, n} (S460); storing the variable m-1 when the size of r _n-1 is larger than C _{m-1, n} (S470); Comprising the step (S480) to set the m-1 calculated through the above process as the end point.

The codebook search range is determined by the start point and the end point obtained by the above processes of the flowchart. In the reverse comparison, since the first codebook is not reversely compared, the endpoint is always 256.

As can be seen from the above description, the present invention replaces the codebook for the target vector of G.723.1, the corresponding codebook, and the respective subframes with a new codebook plus a DC component in advance, so that the corresponding codebook has ordering. By reducing the range of codevectors to be searched using the ordering properties of the LSP coefficients, the codebook search computations having a large amount of computations are reduced.

Claims (5)

A method for reducing the amount of computation required for codebook searching in LSP coefficient quantization of a speech encoder for processing a speech signal, A first step of obtaining an error criterion using a difference between an LSP coefficient vector and a prediction vector and a sum of an error code vector and a DC component mean vector of the LSP coefficient values; A second step of generating a new codebook for three sub-vectors by calculating a sum of an error code vector and a DC component average vector of LSP coefficient values in advance in the first step; A third step of selecting a specific column used to determine the search range in the new codebook and sorting in descending order based on element values of the selected column; And a fourth process of determining a search range by using an order property of a given target vector and aligned code vectors. The method of claim 1, wherein the first process comprises: a first step of obtaining a difference between an LSP coefficient vector and a prediction vector; A second step of obtaining a sum of an error code vector and a DC component mean vector of LSP coefficient values; And a third step of determining a codebook index for minimizing an error criterion using the difference and the sum. The method according to claim 1, wherein the second step is performed by calculating a new sum of the error code vector used in the first step and the average DC component vector of the LSP coefficient value every time the error criterion is calculated. A fast search method for an LSP codebook of a speech coder, comprising adding a DC component mean vector value of a value to an existing codebook to form a new codebook. The method of claim 1, wherein the position of a specific column used to determine the search range in each of the three newly configured codebooks is the first column in the first codebook, the second column in the second codebook, and the first column in the third codebook. The first stage of selection, And a second step of sorting the entire codebook in descending order based on the columns selected from each of the three codebooks. The method of claim 1, wherein the fourth process performs a process of determining a search range by forward and reverse comparison of the codebooks arranged in the second and third processes.