JP3276977B2 - 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 an apparatus for efficiently compressing and transmitting or storing speech waveforms.

[0002]

2. Description of the Related Art Code excitation linear prediction coding (C
ode-Excited Linear Prediction: hereinafter referred to as CELP). (For example, "Code-Excited Linear Prediction (C
ELP): High-Quality Speech at Very Low Bit Rates "
MR Schroeder and BS Atal, Proc.IEEE Int.Con
f. on Acoustics, Speech and Signal Processing, pp.
937-940, 1985).

[0003] Although CELP is a system capable of reproducing high-quality sound even at a low bit rate, the amount of calculation is extremely large because coding is basically performed using an analysis-by-synthesis method. There is a problem. In particular, a portion having a large amount of calculation is a process of determining the optimum excitation waveform by performing a full search on a codebook in which various types of excitation waveforms are stored.

Various schemes have been proposed for reducing the amount of search calculation for the excitation codebook. As a method of reducing the amount of calculation by defining a structure in a codebook, a codebook using vector addition (for example, "Vector Sum Excited Linear
Prediction (VSELP) SpeechCoding at 8 kbps ", IA
Gerson and M. Jasiuk, Proc.IEEE Int.Conf. On Aco
ustics, Speech and Signal Processing, pp.461-463,
1990) and overlapping codebooks (eg, "Improved Speech Quality and Efficient Vector Q
uantization in SELP ", WB Kleijn, DJKrasinsk
i, and RH Ketchum, Proc.IEEE Int. Conf. on Aco
ustics, Speech and Signal Processing, pp. 155-158,
1988).

The present invention is basically an improvement on the overlapping codebook approach. WBKleijn below
A description will be given of the reduction of the amount of calculation using the overlapping codebook according to the above-mentioned reference.

The vector of the speech signal of the current analysis frame is
s, the matrix representing the impulse response of the synthesis filter is H, and the quiescent response of the synthesis filter of the current frame is z, the optimal excitation waveform vector t satisfies the relationship of Equation 1.

[0007]

(Equation 1)

[0008] Here, H is expressed as in Equation 2 by cutting off the impulse response of the synthesis filter with R samples.

[0009]

(Equation 2)

The search for the optimum excitation waveform is performed by selecting ε represented by the following equation 3 from the candidate vector r _i amplified by the optimum gain information.
Is selected from all i (there are only the types of codewords) that minimize.

[0011]

(Equation 3)

Equation 3 is developed as shown in Equation 4.

[0013]

(Equation 4)

The first term on the right side of Equation 4 is constant regardless of i. H ^T Ht of the second term is also constant regardless of i, and eventually becomes an inner product operation with the excitation waveform vector. The third item has the largest amount of processing. The recursive processing of these three items is described below.

Here, a shift matrix represented by Equations 5 and 6 and a mask matrix are defined.

[0016]

(Equation 5)

[0017]

(Equation 6)

I _k is a matrix in which the last zero element appears in the k-th row. When these matrices are used, in the overlapping codebook, the excitation waveform can be expressed recursively as shown in Expression 7.

[0019]

(Equation 7)

The third item E _i = r _i ^T H ^T H _{i of} Expression 4 with a large amount of processing
Is, since H ^T H is a symmetric Toeplitz matrix,
It can be expressed recursively as shown in Equation 8.

[0021]

(Equation 8)

The operation of Equation 8 has few elements of no zero, and eventually 2R
+3 product-sum operations.

[0023]

The VSELP generates 2 ^N excitation waveforms by multiplying N base vectors by +1 or -1 and adding them. As a result, the processing amount and the memory amount are reduced. For example, 512 excitation waveforms are created from 9 basis vectors. However, there are the following problems.

The number of dimensions of the excitation waveform vector is assumed to be 40 dimensions. This 40-dimensional super space is vector-quantized with 512 representative vectors, but the vector created by addition and subtraction of 9 base vectors is only placed on a 9-dimensional space at most, and the 40-dimensional There is a problem that the deviation is large to represent the space. This can cause quality degradation of the encoded voice.

Since it is difficult to keep the energy of each excitation waveform generated by vector addition constant, there is also a problem that a large amount of information expressing the energy of the excitation waveform is required.

On the other hand, the overlapping codebook is
By shifting and accessing the excitation waveform with a long sample length as a codebook, the amount of computation and the amount of memory are reduced. For example, by shifting the excitation waveform having a sample length of 551 one by one and accessing it, 512 code books of 40 dimensions can be created. However, this processing method has a problem that the energy of each generated excitation waveform is not constant.

[0027]

SUMMARY OF THE INVENTION In a speech coding apparatus using a code-excited linear predictive coding method, a predetermined
A cyclic code that cyclically accesses the codebook of the excitation waveform of length
Book and the synthesis filter created from the linear predictor.
The vector which is the element of the matrix obtained from the pulse response
Calculation means, which is used to search for the optimal excitation waveform.
Is calculated from the impulse response.
A vector that is an element of a matrix
The excitation waveform is further calculated by calculating the inner product of the vectors of the excitation waveform to be formed.
Recursive function that adds the value multiplied by the value of the specified sample position of
By performing the calculation, the amount of calculation in the encoder and the memory size for storing the codebook are reduced. Further, the cyclic excitation codebook comprises a random codebook in which energy is normalized.

[0028]

In the speech coding apparatus using the code excitation linear predictive coding method, the codebook of the excitation waveform stores one or a plurality of base vectors, and each waveform is accessed while being cyclically shifted. Create an excitation waveform. By recursive processing like overlapping codebook,
The amount of calculation for searching for the optimal excitation waveform can be reduced.

[0029]

【Example】

<Embodiment 1> FIG. 1 is a block diagram showing a CELP having a cyclic code according to the present invention.
Of the present invention. 101 is a cyclic excitation codebook of the present invention, 102 is a multiplier for amplifying an excitation waveform with gain information, 103 and 105 are adders, 104 is a long-term predictor for generating a voice pitch structure, and 106 is a voice A short-term predictor for generating a spectrum structure; 107, an input terminal of an audio signal; 108, a subtractor for an input audio signal and a synthesized audio signal; 109, an auditory weighting filter for a difference signal between the input audio signal and the synthesized audio signal;
0 is an energy minimization determiner of the difference signal that is weighted auditory.

First, a method of generating an excitation waveform will be described.
Thereafter, the processing amount reduction will be described. The excitation codebook 101 stores a plurality of excitation waveforms. Waveform sample length
(The number of dimensions) is D, and the number is N. Let each stored waveform be E _I (n), I = 0, 1,. . . , N-1, n = 0,
1,. . . , D-1. A code generated from the cyclic codebook generates an excitation waveform by Expression 9 using a parameter index (index) for selecting an excitation waveform and a parameter position (position) indicating the shift access position.

[0031]

(Equation 9)

Here, n = 0, 1,. . . , D-1. a% b
Indicates the remainder of dividing a by b.

For example, 512 excitation waveforms can be generated by cyclically accessing each of the excitation waveforms storing 16 types of 40-dimensional excitation waveforms from 32 different shift positions. There are various methods for creating a plurality of excitation waveforms stored in the excitation codebook. For example, a random codebook in which energy is normalized is simply used (for the random codebook, refer to the above-mentioned MR Schroeder document).

It is also possible to select orthogonal base vectors of random codebooks whose energy is normalized and which are orthogonal to each other. This is white Gaussia
The random codebook generated from n) is obtained by orthogonalizing all other basis vectors using the Gram-Schmidt orthogonalization method, and then normalizing the energy.

Such white Gaussian (white Gaussian)
(ussian) orthogonal excitation random codebook is generated by cyclically shifting the excitation waveform vector, the angle between each other is almost orthogonal, and the vector space of the excitation waveform vector space expressed as a random codebook is vector-quantized. It is considered to be efficient. Further, since all the energy of the excitation waveform vector is normalized, information expressing the energy of the excitation waveform (indicated as gain in FIG. 1) can also be efficiently expressed.

Next, a description will be given of how to reduce the amount of processing when searching for an optimal excitation waveform from the cyclic codebook. Basically, it is based on the method used in the overlapping codebook, but by performing cyclic access, the amount of recursive processing can be further reduced. The procedure for reducing the amount of computation using the overlapping codebook described in the related art will be applied to cyclic codes. Here, a matrix represented by Expression 10 is defined.

[0037]

(Equation 10)

Here, D represents the number of vector dimensions, that is, the number of samples in the analysis frame. The excitation waveform that is generated cyclically is represented by Equation 1, but in order to simplify the notation below, r
_i is Eindex, _i (n), n = 0, 1,. . . ,
D-1. Also, as an example of recursive processing, i
= 0, i is 0, 1,. . . , D-1. In the case of the cyclic code, the right side of Equation 8 is set to i = 0, and the actual operation is described as follows.

[0039]

[Equation 11]

[0040]

(Equation 12)

[0041]

(Equation 13)

[0042]

[Equation 14]

After all, by using the cyclic code, the operation of Expression 8 is represented by Expression 15 and becomes a D + 2 product-sum operation.

[0044]

(Equation 15)

[0045] Here, b _n = a _Dn - is a _n, is only calculated once H is determined. The amount of computation for recursive processing is
In the case of the overlapping codebook, it depends on the number of truncations R of the impulse response, and in the case of the cyclic code, it depends on the number of samples D of the analysis frame. As a specific comparative example of the processing amount, if R = D = 40, the ratio of (D + 2) / (2R + 3) becomes about 50%.

However, when a plurality of types (N) of codes are generated as cyclic code generation sources, the calculation of the initial value E ₀ of the recursive processing of Expression 9 must be performed a plurality of times. Therefore, the amount of processing other than the recursive operation increases.

[0047]

As is apparent from the above description, in the speech coding apparatus using the code excitation linear prediction coding method, the codebook of the excitation waveform composed of the random codebook of the orthogonal basis whose energy is normalized is cyclically transferred. By accessing, the excitation waveform can be efficiently vector-quantized, and the amount of memory for storing the excitation waveform is reduced.
The amount of calculation for searching for the optimum excitation waveform can also be reduced.

[Brief description of the drawings]

FIG. 1 is a block diagram of an embodiment of an encoder of a speech encoding device according to the present invention.

[Explanation of symbols]

 Reference Signs List 101 cyclic excitation codebook 102 multiplier 103, 105 adder 104 long-term predictor 106 short-term predictor 107 signal input terminal 108 subtractor 109 auditory weighting filter 110 energy minimization determiner

Claims (2)

(57) [Claims] 1. A speech coding apparatus using a code excitation linear predictive coding method, wherein a codebook of an excitation waveform having a predetermined length is cyclically accessed.
Access to the cyclic codebook and the synthesis created from the linear predictor
Matrix element calculated from the impulse response of the filter
A means for calculating a certain vector is provided, and calculation of an evaluation value used for searching for an optimal excitation waveform is
Vector, a matrix element obtained from the impulse response
And the waveform of the excitation waveform generated by cyclic access.
In addition to the calculation of the inner product of the vector,
A speech coding apparatus, which is executed by a recursive operation of adding values multiplied by values . 2. The cyclic codebook normalizes energy.
Excitation codebook consisting of random codebook
The speech encoding device according to claim 1, wherein: