JP3052274B2 - LSP quantization method - Google Patents

Abstract

PURPOSE:To decrease both the memory capacity of a code book and the frequency of product sum arithmetic of distance calculation and to stop an increase in quantization distortion. CONSTITUTION:Each encoded frame consists of plural subframes and respective LSPs(line spectrum pair) of the subframes are represented with matrixes at every encoded frame; and the LSP matrixes are divided by a matrix divider 12 into low-range, intermediate-range, and high-range submatrixes. The low- range submatrix is quantized by a matrix quantization part 13 by using a low- range candidate code book 14 and the quantized code is outputted to a composition part 15 and also inversely quantized by an inverse quantization part 16. A difference processing part 17 subtracts the inverse quantaized value of the LSP parameter of the highest degree of the submatrix from the LSP parameter of the lowest degree of the intermediate-range submatrix; and the intermediate- range submatrix after the subtraction processing is equantized by a quantization part 18 and the high-range submatrix is equantized by a quantization part 21.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for quantizing an LSP (line spectrum pair) in a linear predictive coding, which is used in a coding method in digital communication and voice storage.

[0002]

2. Description of the Related Art High efficiency linear prediction coding techniques such as CELP are widely used for digital communication and voice storage. Here, in order to compress the redundancy of voice, a time width of several ms to several tens ms (encoded frame)
Extracts and encodes parameter information such as linear prediction coefficients, audio power, and pitch information. Among them, the linear prediction coefficient is an important parameter that largely affects the voice quality, and therefore, it is necessary to perform high-efficiency quantization, that is, quantization with a small number of bits and small distortion. For this reason, it is often converted to a parameter called LSP (line spectrum pair) that is mathematically equivalent and has high quantization efficiency.

[0003] When the predicted order is the Nth order, the LSP is also obtained as the Nth order value. Also, when the encoding frame length is long and a plurality of M encoding subframes are employed,
In order to cope with the time variation of the LSP, it is necessary to find each LSP in the M subframes. As described above, as a method of quantizing M LSPs as one code for each encoded frame, conventionally, (1) M × N individual scalar quantization is performed, and (2) M N-dimensional vector quantizations are performed. (3) (M × N) -dimensional matrix quantization,
and so on.

[0004]

However, (1),
The method (2) has a disadvantage that the quantization efficiency is low and the number of required bits is increased. In (3), the highest efficiency is obtained, but since the number of candidate matrices is enormous, the amount of memory becomes enormous, and the amount of computation required for quantization becomes too large. For example, if the prediction order (LSP order) is 10, the coding frame length is 40 ms, and the subframe length is 20 ms, the LSP of the coding frame is a (2,10) dimension composed of the respective LSPs of the two subframes. As an LSP matrix. Assuming that the number of quantization bits representing the candidate matrix is N, the memory capacity of the candidate matrix codebook is 20 × 2 ^N (words) and the input LSP
20 × 2 ^N sum-of-products operations are needed to calculate the distance to which of the candidate matrices the matrix is closer to. N = 30
In the case of (2), a memory of 20 GW and a product-sum operation of 20 G times are required, and it is difficult to realize.

To solve this problem, it is conceivable to divide the LSP matrix into a plurality of low-dimensional sub-matrices and quantize them. However, as the number of dimensions decreases, the quantization efficiency decreases as in (1) and (2). In addition, a method of sequentially quantizing only the prediction error using the correlation between the vectors has been proposed, but in this case, once a code error occurs, an error that cannot be repaired on the decoding side occurs.

An object of the present invention is to provide an LSP quantization method that requires a smaller storage capacity of a candidate matrix code memory than in the past.

[0007]

According to the first aspect of the present invention, an LSP matrix to be quantized (input LSP matrix) is determined by utilizing the fact that there is a strong correlation between LSP parameters of adjacent orders. Is divided into two or more sub-matrices in the order direction, and these divided sub-matrices are quantized first from the low-order side of the LSP. After subtracting the inverse quantization value of the highest-order LSP parameter from the lowest-order LSP parameter of the one-stage higher-order submatrix, the higher-order submatrix is quantized.

[0008] In this way, the variance of the higher-order sub-matrix is compressed, and as a result, quantization distortion is reduced. According to the third aspect of the present invention, the input LSP matrix is
Is divided into two or more sub-matrices in the order direction, and these divided sub-matrices are quantized first from the lower-order side of the LSP, and the highest-order LSP in the quantized sub-matrix is Based on the inverse quantization value of the parameter, a candidate sub-matrix codebook for quantizing the higher-order sub-matrix is selected and the higher-order sub-matrix is quantized.

According to the fifth aspect of the present invention, each candidate matrix in the candidate matrix codebook is replaced with a plurality of candidate matrices in the order corresponding to the time of the subframe, and the plurality of candidate matrices are used. The input LSP matrix is quantized, and the quantized code is represented by a first index indicating a state of the permutation of time in the candidate matrix and a second index representing a number of the candidate matrix before permutation of the time in the candidate matrix. . According to the fourth aspect of the present invention, the input LSP matrix is replaced instead of replacing the candidate matrix.

[0010]

FIG. 1 shows an embodiment of the present invention. Input terminal 1
The input LSP matrix from 1 is a matrix divider 12
In this example, in the order direction of the LSP, the signal is divided into a low-frequency sub-matrix, a medium-frequency sub-matrix, and a high-frequency sub-matrix. Let the order of LSP (predicted order) be 10,
Assuming that the coded frame is composed of two subframes, each LSP parameter is represented by a _ij (i = 1,2:
, 10: LSP order). The input LSP matrix is (2,10)
Dimensions, which are, for example, a (2,2) -dimensional low-pass sub-matrix, a (2,4) -dimensional mid-range sub-matrix,
It is divided into a (2, 4) -dimensional high-frequency sub-matrix.

The low-frequency sub-matrix (a₁₁, A₁₂, A_{twenty one},
a_{twenty two}) Is a matrix quantizer 13 for low-frequency candidate code
The distance from each low-frequency candidate matrix of the block 14 is calculated,
The code indicating the closest candidate matrix is determined and the combining unit
15. In other words, the low-frequency sub-matrix is quantized
Is done. This quantization code is also supplied to the inverse quantization unit 16.
Is dequantized, and the highest order dequantized LSP parameter is
Data a₁₂', A_{twenty two}′ Is supplied to the difference processing unit 17,
Area sub-matrix (a₁₃, A₁₄, A₁₅, A₁₆, A_{twenty three}, A
_{twenty four}, A_{twenty five}, A₂₆) The lowest order LSP parameter a in₁₃,
a_{twenty three}From the inverse quantization value a of the highest order LSP parameter ₁₂′,
a_{twenty two}'Are each reduced. This difference processed midrange
Submatrix (a₁₃-A₁₂', A₁₄, A₁₅, A₁₆, A
_{twenty three}-A_{twenty two}', A_{twenty four}, A_{twenty five}, A₂₆) Is the matrix quantizer
In 18, quantization is performed with reference to the midrange candidate codebook 19.
The result is supplied to the synthesizing unit 15.

The high-frequency sub-matrix (a ₁₇ , a ₁₈ , a ₁₉ ,
a ₁₁₀ , a ₂₇ , a ₂₈ , a ₂₉ , a ₂₁₀ ) are quantized by the matrix quantization section 21 with reference to the high-frequency candidate codebook 22, and the result is supplied to the synthesis section 15. The three quantization codes input thereto are synthesized by the synthesis unit 15 and output as a quantization code for the input LSP matrix.

The memory capacity of the low band candidate codebook 14 is 4 × 2 ¹⁰ (words), and the memory capacity of the middle band candidate codebook 19 and the high band candidate codebook 22 is 8 respectively.
× 10 ¹⁰ (words), so the total memory capacity is 2
0 × 2 ¹⁰ (words), which is significantly smaller than the conventional case where 20 × 10 ³⁰ (words) were required. Further, the product-sum operation for calculating the distance is only 20 × 2 ¹⁰ times, which is remarkably less than the conventional case.

[0014] Moreover between a ₁₂ and a ₁₃ in the LSP parameter, since there is a strong correlation between a ₂₂ and a _23, (1) the dispersion of the candidate matrix codebook difference component is reduced. (2) Since the low band LSP parameter largely affects the SNR of the decoded speech, a large number of quantization bits are normally assigned to the low band, so that the quantization distortion of the inverse quantization values a ₁₂ ′ and a ₂₂ ′ is reduced. Therefore, even if the difference is obtained, the difference error becomes small.

Therefore, by performing differential quantization as described above, quantization distortion can be greatly reduced. In the present invention, differential quantization may be applied to a high band, which is to be quantized first from a low band. However, (1) the correlation between adjacent LSP parameters is low for orders higher than 1/2 of the prediction order, and (2) the allocation of the number of quantization bits is often small for orders higher than the middle band, and the inverse quantization value Of the LSP order is relatively small, and the difference quantization is performed only for the sub-matrix corresponding to an order lower than 1/2 of the LSP order. It is effective to do.

FIG. 2 shows an embodiment of the third aspect of the present invention. Also in this case, as in the case of FIG. 1, the input LSP matrix is divided into a low-frequency sub-matrix, a medium-frequency sub-matrix, and a high-frequency sub-matrix. Matrix quantization unit 13
, And the quantized code is supplied to the synthesis unit 15 and inversely quantized by the inverse quantization unit 16.

The highest order of this inversely quantized LSP matrix
Parameter a₁₂', A_{twenty two}′
Book 19₁, 19_Two, 19_Three, 19_FourOne of
Selected with the textbook selector 24, and the selected midrange candidate
Matrix midrange sub-matrix using codebook
The quantization is performed by the quantization unit 18. That is, the parameter a_i2′ Is large
If the value is too large, the candidate code has a distribution that is biased toward the high frequency side.
A book is selected and a _i2If ′ is a small value,
A candidate codebook having a biased distribution is selected. I
Because the distribution of the shifted codebook is also narrow, the quantization distortion is small.
It becomes.

The middle band sub-matrix is quantized using the candidate codebook selected in this manner, and the quantized code is supplied to the synthesizing unit 15 and the inverse quantizing unit 2
5 is inversely quantized. According to the highest-order parameter a _i6 ′ in the dequantized middle band sub-matrix, the best one of the high band candidate codebooks 22 ₁ , 22 ₂ , 22 ₃ , 22 ₄ is selected by the codebook selector 26. , Are selected in the same manner as the codebook selector 24. Using the selected high-frequency candidate codebook, the high-frequency sub-matrix is quantized by the matrix quantization unit 21, and the quantized code is supplied to the synthesis unit 15.

In the method shown in FIG.
Since the deviation of the entire distribution is used, the difference quantum
Quantization distortion reduction effect is greater than
It works effectively even in a region with a high (LSP order). FIG. 3A
An embodiment of the invention of claim 4 is shown in FIG. Input LSP matrix
Is directly supplied to the matrix quantization unit 31,
The time order is changed by the time order changing unit 32 and the matrix is changed.
The signal is supplied to the quantization unit 31. That is, each input LSP mat
The LIX is composed of LSPs corresponding to each subframe. Previous
As shown in the example, the encoded frame consists of two subframes.
If so, the LSP parameters of the first half subframe
(A ₁₁, A₁₂, A₁₃, ..., a₁₁₀) And late subflation
LSP parameter (a _{twenty one}, A_{twenty two}, A_{twenty three}, ..., a
₂₁₀) And the input LSP matrix (a₁₁, A₁₂,
…, A₁₁₀, A_{twenty one}, A_{twenty two}, ..., a₂₁₀) Is composed
However, the LSP of the first half subframe is
Parameters and LSP parameters of the latter subframe
LSP matrix (a_{twenty one}, A_{twenty two},
…, A₂₁₀, A₁₁, A ₁₂, ..., a₁₁₀).

In the matrix quantization section 31, the input LSP
The respective distances between the matrix and the LSP matrix in which these have been temporally replaced and each candidate matrix from the candidate codebook 33 are calculated, and a code indicating the one with the smallest distance is output as a quantization code. This output code is a first index indicating a permutation state (temporal permutation state) of the input LSP matrix in the order of time;
It consists of a second index representing the number of the candidate matrix.

This is equivalent to a case where the number of candidate matrices is doubled, and the memory capacity of the candidate codebook 33 is significantly smaller than in the conventional case. FIG. 3B shows an embodiment of the invention of claim 5. In this case, the input LSP matrix is supplied to the matrix quantization unit 31, and each candidate matrix from the candidate codebook 33 is directly supplied to the matrix quantization unit 31 and the time replacement unit 34 performs the same operation as the time replacement unit 32. The temporal replacement is performed and the result is supplied to the matrix quantization unit 31. The distance between the input LSP matrix, each candidate matrix and its temporally swapped one is calculated,
The code indicating the smallest one is output as a quantization code.

When the candidate matrix codebook is represented in a fixed-point format and stored in a memory, if the dynamic range is wide, the required number of bits per word increases. Therefore, the dynamic range can be suppressed by subtracting the statistical average value from the input matrix in advance. In the case of LSP parameters, the dynamic range can be suppressed to about 1/10 by this method, and the required number of bits per word can be reduced by 3 bits or more.

The method of dividing the input LSP matrix, the number of subframes per encoded frame, the order of the LSP, and the like are not limited to the above examples. Further, the above methods may be used in combination. For example, the difference quantization method shown in FIG. 1 and the method of selecting a sub-matrix candidate codebook shown in FIG. 2 can be applied simultaneously. In the difference quantization method shown in FIG.
As described in A or B, the time may be changed. Similarly, in the sub-matrix candidate codebook selection method shown in FIG. 2, the time may be switched for the sub-matrix as described with reference to FIG. 3A or 3B. Further, with respect to the combination of the method shown in FIG. 1 and the method shown in FIG. 2, the sub-matrix may be time-swapped as described in FIG. 3A or 3B.

[0024]

According to the present invention, the memory capacity of the candidate codebook can be significantly reduced as compared with the prior art. Further, according to the first and third aspects of the present invention, it is possible to suppress the increase in quantization noise and the influence of code errors, and to greatly reduce the number of product-sum operations in distance calculation.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an example of a quantization apparatus to which the invention of claim 1 is applied.

FIG. 2 is a block diagram showing an example of a quantization device to which the invention of claim 3 is applied.

FIG. 3A is a block diagram showing an example of a quantization device to which the invention of claim 4 is applied, and FIG. 3B is a block diagram showing an example of a quantization device to which the invention of claim 5 is applied.

──────────────────────────────────────────────────続 き Continued on front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G10L 19/00-19/14

Claims (11)

(57) [Claims] An LS for matrix-quantizing LSPs (line spectrum pairs) at a plurality of different times obtained from a sound signal.
In the P quantization method, the LSP matrix to be quantized is divided into two or more sub-matrices in the order direction of the LSP, and these divided sub-matrices are quantized first from the lower-order side of the LSP. , The highest order LS in the quantized sub-matrix
An LSP quantization method, wherein the inverse quantization value of the P parameter is subtracted from the lowest-order LSP parameter of the higher-order sub-matrix, and then the higher-order sub-matrix is quantized. 2. The method according to claim 1, wherein the quantization is performed by subtracting the inverse quantization value from the highest-order LSP parameter only for a sub-matrix corresponding to an order lower than 1/2 of the LSP order. The described LSP quantization method. 3. An LS for performing matrix quantization on LSPs (line spectrum pairs) at a plurality of different times obtained from a sound signal.
In the P quantization method, the LSP matrix to be quantized is divided into two or more sub-matrices in the order direction of the LSP, and these divided sub-matrices are quantized first from the lower-order side of the LSP. , The highest L of the quantized sub-matrices
Selecting a candidate sub-matrix codebook for quantizing the higher-order sub-matrix based on the inverse quantization value of the SP parameter, and quantizing the higher-order sub-matrix. Quantization method. 4. An LSP for matrix-quantizing LSPs (line spectrum pairs) at a plurality of different times obtained from a sound signal or the like.
In the SP quantization method, an LSP matrix to be quantized is an LSP matrix in which the order of time is changed, and the LSP matrix and the LSP matrix in which time is not changed are quantized using the same candidate matrix codebook, An LSP quantization method, characterized in that the quantization codes are represented by a first index indicating a state in which the quantization codes are changed in the order of the time and a second index indicating a candidate matrix number. 5. An LS for performing matrix quantization on LSPs (line spectrum pairs) at a plurality of different times obtained from a sound signal.
In the P quantization method, each candidate matrix in the candidate matrix codebook is replaced with a plurality of candidate matrices by changing the order of the times, and an LSP matrix to be quantized is quantized using the plurality of candidate matrices. An LSP quantization method, wherein a quantization code is represented by a first index indicating the state of the permutation of the time order and a second index representing a number of a candidate matrix before the permutation of the time order. 6. A candidate sub-matrix codebook for quantizing a higher-order sub-matrix is selected based on the inverse quantization value of the highest-order LSP parameter in the quantized sub-matrix, and 2. The LSP according to claim 1, wherein the higher-order sub-matrix is quantized.
Quantization method. 7. The sub-matrix is a sub-matrix in which the order of time is changed, and this sub-matrix and a sub-matrix in which the order of time is not changed are quantized using the same candidate sub-matrix codebook, 2. The method according to claim 1, wherein the quantization code is represented by a first index representing the permutation state of the time order and a second index representing the number of the candidate sub-matrix.
The described LSP quantization method. 8. Each of the candidate sub-matrices in the candidate sub-matrix codebook is changed into a plurality of candidate sub-matrices by changing the order of the times, and the sub-matrix is quantized using the plurality of candidate sub-matrices. 2. The quantization code is represented by a first index representing the time permutation state and a second index representing a candidate submatrix number before the time permutation is performed. LSP quantization method. 9. The sub-matrix is a sub-matrix in which the order of time is changed, and this sub-matrix and a sub-matrix in which the order of time is not changed are determined using the selected candidate sub-matrix codebook. 3. The method according to claim 2, wherein the quantization code is represented by a first index representing the permutation state of the time order and a second index representing the number of the candidate sub-matrix.
The described LSP quantization method. 10. Each of the candidate sub-matrices in the selected candidate sub-matrix codebook is replaced with a plurality of candidate sub-matrices by changing the order of the times, and using the plurality of candidate sub-matrices, Quantized, and the quantization code is represented by a first index representing the permutation state of the time order and a second index representing the number of the candidate sub-matrix before the permutation of the time order. The LSP quantization method according to claim 2. 11. The LSP quantization method according to claim 1, wherein a statistical average value of the LSP input matrix is subtracted in advance.