JPH05281999A - Speech encoding device using cyclic code book - Google Patents

Abstract

PURPOSE:To decrease the capacity of a memory stored with an excitation waveform and the quantity of arithmetic for an optimum excitation waveform search by cyclically accessing a code book for the excitation waveforms. CONSTITUTION:In the speech encoding device which uses a code exciting linear predictive encoding system, plural exciting waveforms are stored in the cyclic excitation code book 101. A code generated from this code book 101 generates an exciting waveform with a parameter index selecting the excitation waveform and a parameter position indicating its shift access position. F or example, 16 kind of stored exciting waveforms in 40 dimensions are cyclically accessed from 32 different positions to generate 512 excitation waveforms. When the optimum excitation waveform is searched for in the code book 101, the search is made basically by a technique used for an overlapping code book. At this time, the access is cyclically performed, so the part of recurrence processing is reducible in throughput.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device for efficiently compressing a voice waveform and transmitting or storing it.

[0002]

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

[0003] CELP is a system capable of reproducing high-quality sound even at a low bit rate, but since it is basically encoded by using an analysis method by synthesis (Analysis-by-Synthesis), the calculation amount is very large. There is a problem. Particularly, the part where the amount of calculation is large is a process in which the optimum excitation waveform is determined by performing a full search through the codebook that stores a large number of types.

Various methods have been proposed for reducing the search calculation amount of the excitation codebook. A codebook based on vector addition (for example, "Vector Sum Excited Linear") is used as a method to reduce the amount of calculation by defining the structure in the codebook.
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) or overlapping codebooks (eg "Improved Speech Quality and Efficient Vector Q
uantization in SELP ", WB Kleijn, DJ Krasinsk
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 to the overlapping codebook approach. Below is WBKleijn
The reduction of the amount of calculation using the overlapping codebook will be described according to the above-mentioned references.

The vector of the voice signal of the current analysis frame
Let s be the matrix representing the impulse response of the synthesis filter, and z be the zero-input response of the synthesis filter in the current frame.

[0007]

[Equation 1]

Here, H expresses the impulse response of the synthesizing filter as expressed by equation (2) by truncating it with R samples.

[0009]

[Equation 2]

The search for the optimum excitation waveform is performed by using the ε expressed by the equation 3 from the auxiliary vector r _i amplified by the optimum gain information.
This is the process of selecting the one that minimizes from all i (there are only codeword types).

[0011]

[Equation 3]

Expression 3 is expanded as Expression 4.

[0013]

[Equation 4]

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

Here, the shift matrix expressed by the equations 5 and 6 and the mask matrix are defined.

[0016]

[Equation 5]

[0017]

[Equation 6]

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

[0019]

[Equation 7]

Third item E _i = r _i ^T H ^T H _{i of} number 4 with a large amount of processing
Is that H ^T H is a symmetric Toeplitz matrix, so
It can be expressed recursively like Equation 8.

[0021]

[Equation 8]

The operation of the equation 8 has few elements of non-zero, and the result is 2R.
+3 multiply-add operations.

[0023]

VSELP creates 2 ^N excitation waveforms by multiplying N basis vectors by +1 or -1 and adding them. This has realized the reduction of processing amount and memory amount. For example, 512 excitation waveforms are created from 9 basis vectors. However, there are the following problems.

The number of dimensions of the vector of the excitation waveform is 40. This 40-dimensional superspace is vector-quantized by 512 representative vectors, but the vector created by addition and subtraction of 9 basis vectors is only stretched in 9-dimensional space at most, and 40-dimensional There is a problem that there is a large deviation to represent the space. This can cause deterioration in the quality of coded speech.

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

On the other hand, the overlapping codebook is
As a codebook, by shifting and accessing the excitation waveform with a long sample length, the amount of calculation and memory can be reduced. For example, by shifting the excitation waveform with a sample length of 551 by 1 and accessing it, 512 40-dimensional codebooks can be created. However, this processing method has a problem that the energy of each excitation waveform generated is not constant.

[0027]

In a speech coding apparatus using a code-excited linear predictive coding system, by providing means for cyclically accessing a codebook of excitation waveforms, the calculation amount in the encoder and the code Reduce the memory size to store the book.

[0028]

In the speech coder using the code-excited linear predictive coding method, the codebook of the excitation waveform stores one or a plurality of basis vectors, and each waveform is accessed while being cyclically shifted. Create an excitation waveform. Like recursive processing in the overlapping codebook,
The amount of calculation for optimum excitation waveform search can be reduced.

[0029]

【Example】

<Embodiment 1> FIG. 1 shows a CELP equipped with a cyclic code of the present invention.
It is a speech encoding device of. 101 is a cyclic excitation codebook of the present invention, 102 is a multiplier that amplifies an excitation waveform with gain information, 103 and 105 are adders, 104 is a long-term predictor that generates a pitch structure of speech, and 106 is a speech A short-term predictor for generating a spectral structure, 107 is an input terminal of a speech signal, 108 is a subtractor of the input speech signal and the synthesized speech signal, 109 is a perceptual weighting filter of a difference signal between the input speech signal and the synthesized speech signal, 11
0 is a perceptually weighted difference signal energy minimization determiner.

First, the method of generating the excitation waveform will be described.
After that, the reduction of the processing amount will be described. The excitation codebook 101 stores a plurality of excitation waveforms. Waveform sample length
Let (dimension number) be D and the number be N. Each stored waveform is E _I (n), I = 0, 1 ,. ． ． , N-1, n = 0,
1 ,. ． ． , D-1. The code generated from this cyclic codebook generates an excitation waveform according to Equation 9 by 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 when a is divided 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. Various methods of creating a plurality of excitation waveforms stored in the excitation codebook can be considered. For example, a random codebook in which energy is simply normalized (for the random codebook, refer to the above-mentioned MR Schroeder document) is used.

It is also possible to select an energy-normalized orthogonal basis vector of a random codebook which is orthogonal to each other. This is white Gaussia
The random codebook generated from n) is orthogonalized to all other basis vectors using the Gram-Schmidt orthogonalization method, and then the energy is normalized.

Such a white Gaussian (white Ga
The excitation waveform vector generated by cyclically shifting the orthogonal basis random codebook of ussian) is such that the angles formed by them are close to orthogonal, and the vector space of the excitation waveform expressed as a random codebook is vector-quantized. Is considered to be efficient. Further, since all the energy of the excitation waveform vector is also normalized, the information expressing the energy of the excitation waveform (expressed as gain in FIG. 1) can also be efficiently expressed.

Next, reduction of the amount of processing when searching for the optimum excitation waveform from the cyclic codebook will be described. Basically, it is based on the method used in the overlapping codebook, but the recursive processing part can be further reduced by accessing cyclically. The procedure for reducing the amount of calculation using the overlapping codebook described in the related art will be applied to a cyclic code. Here, the matrix expressed by the equation 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 generated cyclically is expressed by Equation 1, but in order to simplify the notation below, r
_i is E _index , _i (n), n = 0, 1 ,. ． ． , D-1. As an example of recursive processing, i = 0 will be described below, but i is 0, 1 ,. ． ． , D-1 can be used. In the case of the cyclic code, the right side of Equation 9 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 the equation 9 is expressed by the equation 15, which is a D + 2 sum of products operation.

[0044]

[Equation 15]

Here, b _n = a _Dn -a _n , and once H is determined, it is only calculated once. The calculation amount of recursive processing is
For overlapping codebooks, it depends on the impulse response truncation number R, and for cyclic codes it depends on the number of samples D of the analysis frame. As a concrete comparison example of the throughput, if R = D = 40, the ratio of (D + 2) / (2R + 3) is about 50%.

However, if a plurality of types (N) of codes that are the generation sources of the cyclic code are provided, the initial value E ₀ of the recursive processing of the equation 9 must be calculated a plurality of times. Therefore, the amount of processing other than the recursive calculation increases.

[0047]

As is clear from the above, in the speech coder using the code-excited linear predictive coding method, the codebook of the excitation waveform composed of the random codebook of the energy-normalized orthogonal basis is cyclically used. By accessing, the excitation waveform can be efficiently vector-quantized, and the amount of memory that stores the excitation waveform can be reduced.
The amount of calculation for optimum excitation waveform search can also be reduced.

[Brief description of drawings]

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

[Explanation of symbols]

 101 cyclic excitation codebook 102 multipliers 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 (1)

[Claims] 1. A speech coding apparatus using a code-excited linear predictive coding method, wherein a codebook of excitation waveforms is cyclically accessed.