JP3053464B2 - Audio coding device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coding apparatus for compressing information of a speech signal, and more particularly to a speech coding apparatus for compressing information of 4-16 kb / s.
In order to perform encoding at a processing speed of s, Analysis
-CELP (Code Excited Linear Predictio) adopting -by-Synthesis (analysis by synthesis) type vector quantization (hereinafter referred to as "analysis / synthesis type coding")
n) It relates to a speech encoding device represented by an encoder.

[0002] The demand for a speech encoder using such a speech encoding device is increasing as it can realize information compression while maintaining speech quality in an in-house communication system or a digital mobile radio system.

[0003]

2. Description of the Related Art FIG. 5 is a block diagram showing a basic configuration of a conventional CELP speech coder using an analysis / synthesis type coding system. The codebook 51 is a codebook of white noise, and is composed of code vectors of size M (for example, 1024), and each code vector is composed of _ND- dimensional (for example, 40-dimensional) time-series samples.

In analysis / synthesis coding, unlike normal vector quantization, after each code vector C in the codebook 51 is multiplied by a gain g, a certain vocal tract characteristic of an analysis order Np (for example, 10th order) is obtained. Weighted filtering (filtering matrix A) is performed by the linear prediction synthesis filter 52 (gAC). This reproduced signal vector gA
An input signal vector X is added to C, and their difference (X-gAC) is obtained as an error signal vector E. Using the power of the error signal vector E as an evaluation function (distance scale), the error evaluation device 53 searches for an optimal code vector of the codebook 51 so as to minimize the power.

That is, first, the mean square error (power) of the error signal vector E is obtained by the following equation (1). | E | ² = | X−gAC | ² (1) This is partially differentiated with g, and set to 0 to obtain g that minimizes | E | ² .

D | E | ² / dg = 2 (─AC) ^T (X-gAC) = 0 Therefore, the optimal gain g is expressed as g = (X ^T AC) / ((AC) ^T (AC)) Is determined.

[0007] This is substituted into equation ^{(1) | E | 2 =} | X | 2 - (X T AC) 2 / ((AC) T (AC)) From this equation, | X | ² is the input signal Represents the power and is irrelevant to the change in the value of the code vector C.
To minimize E | ² , (X ^T AC) ² / ((AC) ^T
It can be seen that the code vector C that maximizes the value of (AC)) may be selected.

Here, the cross-correlation value between X and AC is represented by Rx
c, If the autocorrelation value of AC is Rcc, that is, if Rxc = X ^T AC (2) Rcc = (AC) ^T (AC) (3), the power of error signal vector E is minimized. The optimum code vector C and gain g to be obtained are obtained by the following equation (4).

C = argmax (Rxc ² / Rcc) (4) g = Rxc / Rcc (5) Equation (4) shows that 1024 code vectors of which the value of (Rxc ² / Rcc) is maximum are Represents the code vector C selected from the above, and the equation (5) expresses the gain g based on the value of the code vector C thus selected.
Represents

Therefore, the optimum code vector and gain are determined as minimizing the error power shown in equation (1). In practice, however, the cross-correlation values and auto-correlation values in equations (2) and (3) are obtained. Based on the correlation value, a code vector (index) is selected by Expression (4), and a gain based on the selected code vector is determined by Expression (5).

FIG. 6 is a diagram showing the arithmetic functions necessary for actually implementing the codebook search based on the above equations (4) and (5). That is, in the cross-correlation calculation device 54,
A cross-correlation value Rxc between the input signal vector X and the code vector AC after the weighting filtering is obtained.
Rxc ² is obtained by squaring with the squaring arithmetic unit 55. On the other hand, in the self correlation operation unit 57 obtains a code vector AC autocorrelation values Rcc the weighted filtering, the division unit 56, the square value Rxc ² ratio of the cross-correlation value Rxc for autocorrelation values Rcc (Rxc ² / Rxc). In the error evaluation device 53, the ratio (R
xc ² / Rxc)
And further calculates an optimum gain g based on the selected code vector C.

FIG. 7 is a diagram showing the convolution operation of the code vector C _k (N-dimensional) and the weighting filtering matrix A (Nth order) according to the conventional method. FIG.
(A) is a time series sample x (i) (i = 0, 1,...) Of the code vector C _k read from the conventional codebook 51.
N) and the elements h (i) (i = 0, 1,... N) of the weighting filtering matrix A, FIG. 7B shows the results of their convolution operation, and FIG. Indicates the number of multiply-accumulate operations to be performed. In the conventional method, the operation must be performed on all the time-series samples, and as shown in FIG. 7C, (N + 1) ×
(N + 2) × 1/2 (820 assuming N = 39) product-sum operations were required.

[0013]

In the above-described codebook search processing, the processing with a large amount of calculation is mainly performed by weighting filtering processing on the code vector, calculation of the cross-correlation value Rxc, and calculation of the auto-correlation value Rcc. There are three calculations. The amount of computation required for each, from the order of the linear prediction synthesis filter 52 is Np, the dimension of the vector quantization codebook 51 is N _D, a Np × N _D, N _D and N _D, per each code vector calculation amount required for the codebook search becomes _{(Np + 2) × N D} . Codebook usually used, codebook size 1024 in 40-dimensional (N _D
= 40, M = 1024), and the analysis order Np of the linear predictive synthesis filter is of the order of 10. Therefore, one codebook search requires (10 + 2) × 40 × 1024 ≒ 48.
This requires a product-sum operation of 0K. In order for such a codebook search to be performed for each subframe (5 ms) of speech coding, a huge processing capability of 96 Mops (million operations / second) is required, and the current highest speed DSP (permissible operation) Even if the amount is 20 to 40 Mops), there is a problem that several chips are required to realize the processing.

Further, in order to store and hold such a codebook as a table, a memory having a capacity capable of storing N _D × M time-series samples (for example, 40 × 1024 = 40K words) is required. Become. In addition, for mobile phones such as automobiles, which are considered to be the application areas of speech encoders using analysis / synthesis coding, miniaturization and low power consumption are essential conditions for commercialization. Both the huge amount of computation and the huge memory capacity described above have been serious obstacles to the realization of a speech encoder as a commercial product.

The present invention has been made in view of the above circumstances, and it is an object of the present invention to provide a speech coding apparatus that shortens the processing time of codebook search and reduces the storage capacity of a codebook .

[0016]

To achieve the above object , a gain is multiplied by each code vector of a white noise codebook.
Weighting filtering by linear prediction synthesis filter
To generate a linear predictive playback signal,
And the error signal vector
Code vector that minimizes the value of
And a CELP-type speech coding device that selects the gain
And the codebook stores the time in each time-series sample.
Indicating position information and amplitude information corresponding to the time.
A data sequence, wherein one of the data sequences is a frame
Specified as the start position, time-series samples for one frame
Reading means for reading out as one code vector
And a speech encoding device characterized by having
Is done.

[0017]

In the speech coding apparatus according to the present invention, information on a plurality of time-series samples of each code vector stored in the codebook includes not only amplitude information but also position information corresponding to time. Let Therefore, when time series samples are decimated as in a sparse codebook, zero-value amplitude information has to be stored even for the decimated time series samples, whereas the There is no need to store any information on the time series samples that have been drawn, and the amount of codebook storage can be reduced.

[0018]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram illustrating the principle of a speech encoding apparatus according to the present invention.
is there. As shown in FIG. 1, position information and amplitude information
A code base consisting of multiple time-series samples
Codebook 1 having a plurality of vectors C, and each code of codebook 1
Reading means for reading vectors in a predetermined reading order
3, the linear prediction synthesis filter 2, the error evaluation means 4 Tona
You.

Each code vector C of the codebook 1 has a gain g
To obtain a signal gC. In the linear prediction synthesis filter 2,
Weighting filtering is performed on the signal gC to perform linear prediction.
Generate the measured and reproduced signal gAC. In addition, the input signal vector
The difference (X-gAC) between X and the linear prediction reproduction signal gAC is
It is obtained as an error signal vector E. Error evaluation device 4
Uses the power of this error signal vector E as an evaluation function.
Code vector of codebook 1 that minimizes the power of
Search for a file.

FIG. 2 is a diagram for explaining the principle of creating a codebook used in the speech encoding / decoding system of the present invention. FIG. 2 (a) shows that the frame pointer 21 of the code vector reading means determines the code vector C which was specified last time.
It represents _K-1, FIG. 2 (c) represents a code vector C _K frame pointer 21 has identified this time. FIG. 2 (b) represents the delta vector [Delta] C _K is the difference between the code vector C _K-1 and the code vector C _K. The frame pointer 21
While specifying a predetermined number of samples (for example, 4) along the data sequence of the time-series samples, it moves in the predetermined direction by one sample at a time, for example.

The code vector C _{K-1 in} FIG. 2A is, for example, a time-series sample x whose amplitude is equal to or larger than a predetermined threshold value.
(i) (location information i = 3, 11, 19, 36)
As amplitude information, -1.9, 5.8, -8.
2.1.2. Time-series samples having an amplitude of 0 exist at positions other than i = 3, 11, 19, and 36. However, since the code vector has both amplitude information and position information for each time-series sample, the code vector becomes i = 3. 3,1
It is not necessary to have information on positions other than 1, 19, and 36. That is, in the related art, the relative position of each sample is known by storing the amplitude information of all the time-series samples in a time-series order. Therefore, it is necessary to store the time-series samples having the amplitude of 0. However, the codebook of the present invention has position information for each time-series sample, so that it is not necessary to store a time-series sample having an amplitude of 0.

By adding the delta vector ΔC _K to the previous specific code vector C _K−1 ,
_K is obtained. Delta vector [Delta] C _K includes a time series sample -x counteract last specific code vector C _K-1 of the time-series samples x (3) (3) (amplitude 1.9), the last specific code vector C _K-1 Missing time series sample x (27)
(Amplitude-2.5), the current specific code vector C _K obtained by adding the delta vector ΔC _K to the previous specific code vector C _K-1 is a time-series sample x
(i) (position information i = 11, 19, 36, 27), and as amplitude information, 5.8, -8.2, 1..
2, -2.5.

FIG. 3 is an example of a codebook (delta codebook) 22 created according to the principle shown in FIG. That is,
The time-series samples are arranged along the moving direction of the frame pointer 21. For example, when the frame pointer 21 moves from the current specific code vector _CK to the next specific code vector _{CK + 1} , the sequence samples x (3) is left when it should be removed from the vector C _{K + 1,} the next specific code vector C _K when _{+ 1} to be newly added to the sequence samples x (27) are arranged so that the right end . The pair at both ends of the arrow shown in the figure is the delta vector Δ
It represents a C _K.

Each code vector of the present codebook 22 is
It is generated by adding a delta vector (including adding a negative value) to a code vector generated before that. In the delta codebook configured in this manner, all codevectors can be configured only by holding the initial codevector (Co) and the delta vector (ΔCi) having a small number of time-series samples. Can be reduced in the memory capacity for storing the data.

In the conventional codebook, the number of dimensions N _D , the size M
Must store all the time series samples of the code vectors, N _D xM (e.g. 40 × 1024 ≒ 40K
(Word) of memory. In contrast, in the present invention,
When the time series number of samples of the delta vector [Delta] C _K and Nd (e.g. 2), time series sample number of the first code vector C1 is N _D, thereafter chronological number of samples sequentially applied to the new A Nd ÷ 2 In addition, since two word numbers of position information and amplitude information are required for one time-series sample, the memory capacity of the codebook of the present invention is:
_{(N D + Nd ÷ 2 ×} (M-1)) × sufficient if only two words. By applying the numerical values of the above embodiment, (4 + 2 ÷ 2
× (1024-1)) × 2 ≒ 2K words, and the memory capacity can be significantly reduced.

Further, a sparse codebook which sets a predetermined threshold value and clips the value whose amplitude is smaller than that of the sparse codebook to zero most of the time-series sample values has been conventionally proposed.
In the conventional sparse codebook, since the sample position cannot be specified, all the samples must be stored in the codebook. Accordingly, a filtering operation described below has been performed on unnecessary zero-value time-series samples.

On the other hand, in the case of the delta codebook of the present invention, 1
Since the position information and the amplitude information are provided for one time-series sample, the amount of information to be stored for each time-series sample increases.However, it is not necessary to store the zero-value time-series sample due to sparseness. If the number of sequence samples becomes equal to or more than half of the number of all time-series samples, for example, 40, it is effective in reducing the storage amount of the codebook and the calculation amount of search processing described later.

In the code book 22, the frame pointer 2
The number of time series sample 1 to identify (corresponding to N _D) is 4, not limited to this, and the number of time series sample frame pointer 21 to identify, for example, 6,
Various numbers are possible, such as 8,10. Also, the current specific code vector CK is _converted to the next specific code vector C
The number Nd of time-series samples to be removed when the frame pointer 21 moves to _{K + 1} and the number of time-series samples newly added, that is, the number Nd of time-series samples of the delta vector ΔC _{K + 1} is two. The number is not limited, and various numbers such as four and six are possible.

Further, according to the code vector storage method of the code book of the present invention, it is possible to specify a code vector by moving a frame pointer with respect to a data sequence of a series of time-series samples. the number of time series sample that, by adjusting the amount of movement of the frame pointer, frame length for any code vector (dimension number N _D), it is possible to set a time series number of samples (N _D) of delta vectors.

Next, a description will be given of a code vector search process, that is, an error evaluation process, performed using the code book configured as described above. The error evaluation process for the code vector AC obtained by applying the weighting filtering to the code vector specified this time using the codebook of the present invention with respect to the input speech signal is performed using the above-described formulas (2) and (3). That is, assuming that C _k = C _k−1 + 置 く C _k , the cross-correlation value Rxc (k) and the auto-correlation value Rcc (k) relating to the specific code vector at this time are represented by the following equations (6) and (7). As shown in
It is expressed in the form of the sum of the correlation value at the time of the previous error evaluation of the specific code vector and the correlation element of the delta vector.

[0031] _{Rxc (k) = (AC k} ) T X = (C k-1 + △ C k) T A T X = Rxc (k-1) + △ C k T A T X (6) Rcc (k ) = (AC _k ) ^T (AC _k ) = (C _k−1 + ΔC _k ) ^{TA T} A (C _k−1 + ΔC _k ) = Rcc (k−1) + 2ΔC _k ^{T AT} AC _k-1 + △ C _k ^T A ^T A △ C _k (7) Means for calculating “△ C _k ^T A ^T X” in the above equation (6)
This is referred to as a correlation element calculation unit.

Further , “Rxc (k)” of the equation (6) is calculated.
Is referred to as correlation element calculating means (reference numeral 54 in FIG. 6).
See). According to the present invention, the code vector is updated by moving the frame pointer on a series of time-series samples divided and stored in the codebook into position information and amplitude information. C _k is determined. Therefore, if the second and subsequent terms of the above equations (6) and (7) are calculated using this delta vector △ C _k , the work of updating the cross-correlation value Rxc (k) and the auto-correlation value Rcc (k) can be performed. Can be easily realized.

The search processing for each code vector requires a weight filtering operation (calculation of the product AC of the matrix A and the vector C) for each code vector. Generally, this weighting filtering operation is performed by an IIR type filter having about tenth order coefficients, but equivalent processing can be realized by convolution with a vector having the impulse response of the filter as an element.

Further, in the method of the present invention, when the code vector is, for example, a code vector C _k as shown in FIG. 4A, and the elements of the weighting filtering matrix A are h (i) (i = 0, 1,. · N), the results of the convolution operations are as shown in FIG. 4B, and the number of product-sum operations performed by this operation is as shown in FIG.
As shown in FIG. The number of times indicated in FIG. 4 (c) shows the average value when the code vectors of N _D dimension took any location information, the 80th position Based on a value of the above embodiment.

As described above, the amount of convolution of the code vector C _k and the weighting filtering matrix A is greatly reduced in the method of the present invention as compared with the conventional method shown in FIG.

As described above, the search processing is performed by calculating the error between the reproduced signal of each code vector and the input audio signal and evaluating the error. It can be reduced to the maximum value search processing of a function having a value as an element. If this correlation value calculation uses a delta vector,
As shown in Expressions (6) and (7), it is possible to recursively update by adding a correlation element of a delta vector to a correlation value of a specific code vector last time.

[0037]

As described above in detail, according to the present invention, information relating to a plurality of time-series samples of each code vector stored in a codebook is not limited to amplitude information, but corresponds to time.
When time-sequential samples are decimated like a sparse codebook, zero-value amplitude information must be stored for the decimated time-sequence samples as well. In contrast to this, it is not necessary to store any information on the decimated time-series samples, and the storage amount of the codebook can be reduced. In addition, since information about time series samples of zero values is removed from the codebook, the amount of calculation based on each codevector in the codebook, which is performed in the process of searching for the optimal codevector, is significantly reduced. Is done.

Further , the reading means operates the command specified last time.
Some of the time series samples that make up the code vector
Some of the time series samples of the code vector to be read
So that the code book has few
Many code vectors can be prepared with the amount of moly,
Regarding the amount of calculation at the time of reading,
Regarding calculation, the calculation result is
The amount of calculation can be reduced.

Further, the time series samples are arranged in the codebook in accordance with the movement of the frame pointer, so that a part of the previous specified time series sample can be shared with the current specified time series sample. Thereby, the storage amount of the codebook can be reduced. In other words, read data so that they partially overlap
As a result, the codebook can store many code vectors with a small amount of storage.
Can be prepared.

Furthermore, this is because the delta vector ΔC is added to the code vector C _k-1 specified by the frame pointer last time.
Adding _K indicates that the frame pointer becomes the code vector C _k specified this time, and the delta vector Δ
By the use of C _K, it is possible to shorten the amount of calculation in the process of searching for the optimum code vector. That is, read
The amount of computation required to calculate cross-correlation and auto-correlation
As a result, the calculation result for the overlapping part can be diverted
Therefore, the amount of calculation can be reduced.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating the principle of a speech encoding apparatus according to the present invention.

FIG. 2 is a diagram illustrating the principle of creating a codebook used in the method of the present invention.

FIG. 3 is a diagram showing a specific example of a codebook.

FIG. 4 is a diagram showing a convolution operation between a code vector and a filtering matrix according to the method of the present invention.

FIG. 5 is a diagram illustrating a basic configuration of a conventional CELP speech encoder.

FIG. 6 is a diagram illustrating a calculation function of a code search.

FIG. 7 is a diagram showing a convolution operation between a code vector and a filtering matrix according to a conventional method.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Codebook 2 Linear prediction synthesis filter 3 Reading means 4 Error evaluation means C code vector A Filtering matrix g Gain X Input signal vector E Error signal vector

──────────────────────────────────────────────────続 き Continuing on the front page (72) Yoshiki Tanaka, Inventor Fujitsu Limited (1015) Uedanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture (72) Inventor Masako Kato 1015, Kamiodanaka, Nakahara-ku, Kawasaki City, Kanagawa Prefecture 56) References JP-A-62-21200 (JP, A) JP-A-63-221400 (JP, A) JP-A-63-271400 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 19/00 G10L 19/08 G10L 19/10 JICST file (JOIS)

Claims (5)

(57) [Claims] 1. Codebook, readout means, multiplication means, linear
CELP type speech code having predictive synthesis filter and error means
An encoding apparatus, wherein the codebook includes a plurality of time-series samples arranged in a predetermined order.
Is configured, the time series sample position information, the amplitude information is not 0
And the reading means comprises a plurality of time-series samples from the codebook.
When pulls are sequentially read by the frame pointer and read
A code vector is output based on the sequence samples, and the multiplying means outputs the code vector output from the reading means.
Multiply the gain vector by the gain
And the linear prediction synthesis filter outputs the output of the multiplication means.
The gain code vector is filtered and re-
Outputting a raw signal vector, wherein the error means includes an input signal vector and the reproduced signal vector;
Frame pointer position and
CELP-type speech coding characterized by determining
apparatus. 2. The reading means according to claim 1, wherein
Time series samples identified by the
The time series server specified by the frame pointer used
Used so that they partially overlap with the sample
The CELP-type speech coding apparatus according to claim 1. 3. A code book, a reading means, and a linear predictive synthesis
Filter, auto-correlation function calculation means, cross-correlation function calculation
CELP-type speech coding apparatus having error means
The codebook is composed of a plurality of time-series samples arranged in a predetermined order.
Is configured, the time series sample position information, the amplitude information is not 0
And the reading means comprises a plurality of time-series samples from the codebook.
When pulls are sequentially read by the frame pointer and read
Outputting a code vector based on the sequence sample, wherein the linear prediction synthesis filter outputs
Filter and play back the code vector
A signal vector, and the autocorrelation function calculating means outputs the signal vector of the reproduced signal vector.
An autocorrelation value is calculated, and the cross-correlation function calculating means calculates the
Calculating a cross-correlation value of the input signal vector, wherein the error means calculates a ratio between the auto-correlation value and the cross-correlation.
Determining the position of the frame pointer that maximizes the rate
A CELP-type speech coding apparatus characterized by the above-mentioned. 4. The reading means according to claim 1, wherein
Time series samples identified by the
The time series server specified by the frame pointer used
Used so that they partially overlap with the sample
The CELP-type speech coding apparatus according to claim 4, wherein 5. The reading means according to claim 1, wherein said reading means has read the last time.
The difference between the series sample and the time series sample read this time is
Output as a delta vector, and the autocorrelation function calculating means outputs the delta vector,
Autocorrelation value is calculated recursively by using autocorrelation values
And the cross-correlation function calculating means calculates the delta vector,
The cross-correlation value is calculated recursively using the previous cross-correlation value.
The CELP-type sound according to claim 4, wherein the sound is emitted.
Voice coding device.