JPH0527798A - Linear predictive parameter encoding method for voice - Google Patents

Abstract

PURPOSE:To enable matrix quantization and decrease the calculation quantity and storage quantity. CONSTITUTION:P line spectrum couples of a voice are used to form an mp input matrix in units of (m) frames and a distortion decision part 12 calculates the distances (distortion) between the matrices and R1 representative matrices Cij(1) (r) held in a 1st code book 11 in order and sends U representative matrices Cij(1) (ru) (u=1, 2,,U) to a 2nd-stage quantization part in increasing distortion order as candidate quantization values. The 2nd-stage quantization part adds the (r)th Cij(2) (r) among R2 representative matrixed held in a 2nd code book 14 to the (u)th quantized candidate value by a matrix adder 15 and a distortion decision part 16 selects the (u) having the least distortion between this matrix and input matrix Fij(k'). Thus, the selected U matrices in the increasing distortion order are used as 2nd-stage quantization candidate values. Similar processing is performed for respective stages.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is a code for storing linear prediction parameters representing a speech spectrum envelope characteristic for a plurality of frames in a matrix format, and expressing the parameters in the matrix format by a predetermined representative matrix for quantization. About how to turn.

[0002]

2. Description of the Related Art In a conventional speech coding system, a coefficient of a linear filter representing a speech spectrum envelope characteristic is calculated by performing a linear prediction analysis at regular time intervals, and a partial autocorrelation (P
It was converted into a parameter such as an ARCOR coefficient or a line spectrum pair (LSP), quantized, converted into a digital code, and then stored or transmitted. Details of these methods are described in, for example, "Digital Audio Processing" by Sadahiro Furui (Tokai University Press).

In this encoding, if the time interval for updating the coefficient is set long, the amount of information for storage or transmission can be reduced, but if it is set too long, it is not stored or transmitted when synthesizing speech. The estimation accuracy of the filter coefficient at a certain time decreased, which led to the deterioration of the quality of the reproduced voice. Therefore, generally, the time interval is set to about 20 milliseconds or less.

From this point of view, there is a method called matrix quantization as a method for performing more efficient coding. This is because the linear prediction analysis is performed at short time intervals, for example, at intervals of about 10 ms to 20 ms,
This is a method of collectively quantizing several sets of analysis results. The linear prediction parameters are correlated in the time direction and the parameter dimension direction, respectively. Matrix quantization is a method that makes good use of this correlation to increase the efficiency of quantization.

[0005]

However, if this method is used to directly quantize a few sets of prediction parameters, a very large amount of calculation and a large amount of memory for the codebook are required. Can not be so efficient under the typical hardware scale. It is an object of the present invention to make use of the advantage that matrix quantization is highly efficient from the viewpoint of reducing quantization distortion, and to encode highly efficient linear prediction parameters under a realistic amount of calculation and storage. To provide a method.

[0006]

According to the present invention, of the predetermined representative matrices, the first-stage representative matrix having the smallest error from the input matrix is determined, and an arbitrary number N of 2 or more is set. Among the representative matrices that are predetermined for
To minimize the error between the input matrix and the matrix obtained by adding the representative matrix of the Nth stage to the matrix that is already determined and is represented by the sum of the representative matrices of the 1st stage to the (N-1) th stage. The representative matrix of the Nth stage is determined, and the quantized value of the input matrix is set to the 1st stage to the Nth stage.
Expressed as the sum of the representative matrices up to the step.

Thus, in the matrix quantization of the linear prediction parameter, the calculation amount and the storage amount can be realized on a practical hardware scale by decomposing the quantization procedure into a set of a plurality of simple procedures. Reduce to a certain level. Moreover, a method called a delayed decision is used between each stage so as to suppress the decrease in quantization efficiency as much as possible when decomposing into a set of simple procedures. This does not decide one optimum quantized value for each stage, but leaves some quantization candidates and determines the optimum quantized value in all stages by a dynamic programming method. At this time, the larger the number of candidates to be left, the more the quantization performance can be prevented from deteriorating. On the contrary, the amount of calculation increases in proportion to the number of candidates, which is a compromise with the target hardware scale. Generally, 4 to 8 are sufficient.

As a method of further reducing the amount of calculation and the amount of storage, according to the invention of claim 3, it is divided into two stages or more, and when the number of stages is two, the second stage is used, and when there are more stages, the second stage is used. Arbitrary stages after the eye are divided into a plurality of sub-matrices, and optimal quantization is performed for each sub-matrix. Generally, there is a high correlation between the elements of each row and each column of the matrix in which the linear prediction parameters are arranged in the dimension direction and the time direction, but each of the error matrices quantized by a matrix quantizer with an appropriate number of bits. The elements follow a Gaussian distribution and are considered uncorrelated or much smaller than the original matrix. Therefore, even if the first matrix is divided into appropriate sizes and the optimum quantization is performed for each of the second and subsequent quantizations, the deterioration in quality is small. Moreover, if the delayed decision is used between the first stage and the second stage,
Further, the deterioration can be suppressed to be small.

[0009]

According to the present invention, by allocating an appropriate number of bits to a plurality of simple procedure groups, it is possible to realize a reduction in the amount of calculation while suppressing a decrease in quantization efficiency. It should be noted that when dividing the processing into a set of simple procedures, the finer the division, the more the calculation amount can be reduced, but it is unavoidable that the quantization quality gradually deteriorates.

Further, it is possible to make the allocation of the number of bits non-uniform and improve the effect, depending on the matrix size after division and the importance of each sub-matrix after division (effect on hearing, etc.).

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows an example of the configuration of a coding apparatus to which an embodiment of the speech linear prediction parameter coding method according to the present invention is applied. From the input terminal 1, the sampled and digitized audio signal s (t) is input. The linear prediction analysis unit 2 temporarily stores N samples of the audio signal s (t) in a data buffer, and then performs a linear prediction analysis on these samples to obtain a set of linear prediction coefficients a _i (k ),
(I = 1, 2, ..., P) is calculated. Here, the parameter p is called an analysis order, and a value of about 10 to 16 is generally used. The unit of N samples is called a frame update cycle or simply a frame cycle. As a result,
For N input samples, p linear prediction coefficients will be output. Therefore, here, as a unit indicating the time of the linear prediction coefficient, a time unit with N samples as a unit is expressed by k, and is called “ _i- th order linear prediction coefficient a _i (k) of the k-th frame”. To Needless to say, p linear prediction coefficients are obtained at each frame time k. Details of these treatment methods are described in the above-mentioned book by Furui.

The line spectrum pair calculation unit 3 converts p linear prediction coefficients into p line spectrum pairs, f _i (k). The reason the linear prediction coefficients are converted into line spectrum pairs in this example is that the nature of the line spectrum pairs takes advantage of the fact that there is less degradation of quality over the division of the procedure described above in the present invention. However, in the present invention, the line spectrum pair calculation unit 3 is not always necessary, and the linear prediction coefficient a _i (k) may be directly input to the buffer unit 4. In addition to line spectrum pairs,
It may be converted into any parameter that can be interconverted with a linear prediction coefficient, such as a partial autocorrelation (PARCOR) coefficient.

In the buffer section 4, p line-spectrum pairs for m frames, that is, m × p parameters f _i (k−j), i = 1, 2, ..., P j = m−1, m. -2, ..., 1, 0 are stored, and these values are supplied to the matrix quantization unit 5 when m frames have been accumulated. In FIG. 1, for simplification of description, the case of m = 4 is shown.

A set or matrix of m × p parameters quantized by the matrix quantizer 5.

[0015]

[Equation 1] Is sent to the encoding unit 6, encoded in a format suitable for the transmission path or the storage form, and transmitted or stored. here,
The reason why k'is used as the index indicating the time is because it is a time unit with m frames represented by the index k as one unit.

On the reproducing side, a matrix is formed from the above codes.

[0017]

[Equation 2] Is generated and returned to the line spectrum pair for each frame by the matrix decomposition unit 8 and then stored in the buffer unit 9.
The buffer unit 9 sends line spectrum pairs for m frames to the linear prediction coefficient conversion unit 10 frame by frame. The linear prediction coefficient conversion unit 10 converts the quantized value of the received line spectrum pair into a linear prediction coefficient a _i ^* (k) and outputs it.

FIG. 2 shows an example in which the matrix quantization in the present invention is realized by a cascade connection of simple (small-scale) matrix quantizers. The figure shows a three-stage configuration for simplicity. Here, the inputs f _i (k−j), i = 1, 2, ..., P j = m−1, m−2, ..., 1, 0 to the matrix quantization unit 5 in FIG. The side-by-side arrangement is represented as F _ij (k '). Assuming that the number of stages is N, as shown in FIG. 2, N sets of codebooks 11, 14, 18, distortion determining units 12, 16, 20,
(N-1) sets of matrix adders 15 and 19 in a continuous combination. The matrix adder does not require the first stage.

First, R stored in the first codebook 11₁Of
Each code matrix (representative matrix) C _ij ⁽¹⁾(R) in order
It is sent to the distortion determination unit 12. The distortion determination unit 12 defines in advance
The input matrix F based on the distortion measure_ij(K ') and code
Matrix C_ij ⁽¹⁾Calculate the distortion with (r) and find the most distortion
Code matrix C in ascending order of
_ij ⁽¹⁾(R_u), (U = 1, 2, ..., U) is the amount of the first stage
Candidate value

[0020]

[Equation 3] Output as. The pre-defined distortion measure here is
Although any scale may be used, usually, a scale such as measuring the Euclidean distance of a line spectrum pair or measuring the Euclidean distance after once converting into a parameter called a cepstrum is used. In the above two scales, the latter is generally said to be superior in that it is more similar to human hearing characteristics. Next, in the second-stage quantization unit, the r-th matrix of the R ₂ code matrices (representative matrix) held by the second codebook 14 is C _ij ⁽²⁾ (r)
And this is the u-th quantization candidate value in the first stage

[0021]

[Equation 4] To the matrix adder 15

[0022]

[Equation 5] Ask for. In the distortion determination unit 16, as in the first stage, first, the input matrix F _ij (k ′)

[0023]

[Equation 6] Choose u such that the distortion between and is the minimum, and let it be ur. Furthermore, with F _ij (k ')

[0024]

[Equation 7] U matrix in ascending order of distortion with

[0025]

[Equation 8] Is stored in the quantization candidate holding unit 17 as the second-stage quantization candidate value. After that, by repeating the same procedure in each stage, one matrix quantizer is constructed by the cascade connection of the N-stage small-scale matrix quantizers. In addition,
In the final stage, it is not necessary to hold U quantization candidates, and the optimum quantization value with the smallest distortion may be determined.

FIG. 3 shows an example of a structure for further reducing the calculation amount as compared with the structure of FIG. The configuration of FIG. 3 is a modification of the configuration of FIG. 2 in that the number of stages is two and the second stage is modified. The first-stage quantizing unit 39 does the same as the first-stage quantizing unit 39 in FIG.

[0027]

[Equation 9] To get Next, the quantized candidate values are sequentially sent to the second-stage quantizer 40. In the second-stage quantizing unit 40, first, the sub-matrix dividing unit 24 divides the first-stage quantizing (candidate) matrix into sub-matrices having an appropriate size.
In FIG. 3, it is shown as being divided into four for simplification. From the viewpoint of quantization quality (efficiency), the number of divided blocks and the size of the divided sub-matrix are better when the number is smaller and the size is larger. However, since the present invention aims to reduce the amount of calculation to an amount commensurate with the hardware scale, the number and size are limited by the hardware scale. Considering the progress of current hardware technology, set the number and size of sub-matrices such that the number of bits allocated to each small-scale matrix quantizer becomes about 8 to 10 bits including the first stage. If so, real-time processing is considered possible. However, if it is desired to reduce the hardware price, it may be divided into more sub-matrices. Here, as an example, it is assumed that they are divided as shown in FIG. Further, when dividing, it is not always necessary to form a sub-matrix with continuous rows and continuous columns. For example, in the first row and the 8th to 10th rows, the 1st column and the 4th column of the original matrix, You may create a 4x2 submatrix.

In the second-stage quantizer 40, the sub-matrix divider 41 also divides the input matrix F _ij (k ') in the same manner. F _ij (k ') divided into each sub-matrix and

[0029]

[Equation 10] Let p _s × m _s be the size of the s-th sub-matrix of

[0030]

[Equation 11] far. However, Σ _s p _s = p Σ _s m _s = m.

The divided u-th quantization candidate sub-matrix in the first stage is sent to each s-th matrix adder. The second and fifth codebooks store R _2s code matrixes (representative matrices) of p _s × m _s . Let this be C _ij ^(2s) (r). These are sequentially sent to the matrix adder, and Fs _ij (k ') and

[0032]

[Equation 12] The code r with the smallest distortion with is obtained. After determining the optimum quantized sub-matrix for the u-th quantized candidate in the first stage for all s, all the sub-matrices are sent to the matrix synthesizing unit 37 and returned to the original matrix form. this

[0033]

[Equation 13] It is represented by. The optimum candidate u is determined from the first-stage candidates by the distortion determination unit 38 so that the distortion with this and the input matrix F _ij (k ′) is minimized, and is used as the output of the matrix quantizer 5.

Although the description of the method of creating each codebook is omitted, a high-performance codebook can be designed by the LBG algorithm. For the LBG algorithm, see Y. Linde, A .; Buzo, R .; M.
Gray: "AnAlgorithm for Vec
tor Quantizer Design ”, IEEE
E Trans. Commum. COM-28, p. p
84-951980.

The present invention is applied to speech coding as well as
It can be applied to applications such as speech recognition and speaker recognition that use linear prediction parameters as speech feature quantities. Further, the present invention may be partially or wholly realized by hardware that is a combination of logic circuits, or may be partially or entirely realized in the form of a software program.

[0036]

The following example shows how much the present invention actually reduces the calculation amount. For example, it is assumed that a matrix in which four 10-dimensional linear prediction parameters are arranged in the time direction is created and 40 bits are given to quantize the matrix. If we try to quantize this as it is, the memory pattern is 2 ⁴⁰ (pattern), the calculation amount is 80 × 2 ⁴⁰ (times) addition / subtraction, 40 × 2 ⁴⁰ (times) multiplication, and 2 ⁴⁰ (times) comparison. Calculation is required, and it is impossible to achieve under realistic hardware scale.

The first method of the present invention, that is, FIG.
Assuming that 10 bits are divided into 4 stages by the configuration shown in FIG. 4, the storage amount when delay decision is not used, that is, when there is one candidate is 2 ¹⁰ × 4 = 4096 (pattern), and the calculation amount is 80 × 2 ¹⁰ It is x4 = 327680 (times) addition / subtraction, 40 × 2 ¹⁰ × 4 = 163840 (times) multiplication, and 2 ¹⁰ × 4 = 4096 (pieces) comparison operations, which are sufficiently realizable. Even if the delayed decision is used, addition / subtraction, multiplication, and comparison operations only increase in proportion to the number of candidates, so that 4 to 8 candidates are sufficient. If the hardware permits, the smaller the number of stages to be divided, the less the deterioration of quality due to the division.

Further, with the second method, that is, with the configuration shown in FIG. 3, 10 bits are used for the first-stage matrix quantization, and the second stage is provided with two 6 × 2 matrices as shown in FIG.
If the matrix is divided into one 4 × 4 matrix, and 10 bits are allocated to each matrix and quantization is performed, the memory capacity without delay decision, that is, in the case of one candidate is 2 ¹⁰ × 4 = 4096 (pattern ) Computation amount is (80 + 32 + 24 × 2) × 2 ¹⁰ = 163840 (times) addition / subtraction, (40 + 16 + 12 × 2) × 2 ¹⁰ = 81920 (times) multiplication, and 2 ¹⁰ × 4 = 4096 (pieces) comparison operation Therefore, even if the bit division is the same as the first method, the calculation amount is halved. In the above example, the division of bits is made the same so that the calculation amounts can be compared. However, actually, the calculation amount may be divided into smaller pieces, and the smaller the division amount is, the smaller the calculation amount is. As a special example,
If the sub-matrix is divided so that it becomes a vector of primary in the time direction and full dimension in the parameter dimension direction, vector quantization is applied to the error of the matrix quantization. If the division is performed as follows, the scalar quantization is applied to the matrix quantization error.

Furthermore, when the second method is used, not only the amount of calculation can be reduced, but also the second-stage calculation can be processed in parallel by the parallel processing processor. The most computationally intensive process in matrix quantization is the process of adding code matrices, determining distortion, and determining the optimum code. Since the first configuration is a serial connection, processing must be performed sequentially from the front, but with the second method, parallel processing can be performed, and thus processing can be performed even faster due to parallelization.

[Brief description of drawings]

FIG. 1 is a block diagram showing a general configuration example of a linear prediction parameter coding method for speech by a matrix quantization method.

FIG. 2 is a block diagram showing an example of a matrix quantizer which is a main part of the first embodiment of the present invention and which connects and encodes small-scale matrix quantizers in series.

FIG. 3 is a block diagram showing an example of a matrix quantization unit that connects two stages of quantization units in a cascade connection in the configuration of FIG. 2 and further divides the second stage into a plurality of sub-matrices for encoding.

FIG. 4 is a diagram showing an example of division into sub-matrices.

Claims (3)

[Claims] 1. A linear prediction parameter representing a speech spectrum envelope characteristic is calculated at a constant time interval called a frame, the parameter is stored for a plurality of frames and converted into a matrix format, and the parameter expressed in the matrix format ( Input matrix),
In a linear prediction parameter coding method in which a predetermined representative matrix is used for quantization, a first-stage representative matrix having a minimum error from the input matrix is determined from among the predetermined representative matrices. , The first to Nth stages that have already been determined among the predetermined representative matrices for any number N of 2 or more.
The N-th stage representative matrix is determined so that the error between the matrix obtained by adding the N-th stage representative matrix to the matrix expressed by the sum of the representative matrices up to -1 stage and the input matrix is minimized, A linear prediction parameter coding method for speech, characterized in that the quantized value of the input matrix is represented by the sum of the representative matrices of the first to Nth stages. 2. When determining the representative values of the first to (N-1) th stages, the optimal representative matrix is not determined to be only one but after leaving some candidates, When the representative matrix of the Nth stage is determined, the first stage to the Nth stage are minimized so that the error between the matrix represented by the sum of the representative matrices of the first stage to the Nth stage and the input matrix is minimized. The linear predictive parameter coding method for speech according to claim 1, characterized in that the representative matrices up to are determined. 3. When determining the representative matrix of the N-th stage,
The matrix represented by the sum of the representative matrices from the 1st stage to the (N-1) th stage and the input matrix are each divided into smaller-sized matrices, and the respective divided small-sized matrices are set to a predetermined small size. 3. A speech linear prediction parameter coding method according to claim 1, wherein the representative value of the size is represented by a representative matrix of the Nth stage and the representative value of the Nth stage is represented by a set of the matrices of the smaller size.