JPH08179800A - Speech coding device - Google Patents

Abstract

(57) [Summary] [Object] An object of the present invention is to reduce the ROM capacity in a coding device based on the CELP method. To achieve the above object, the present invention uses not only the forward order of code vectors stored in the stochastic codebook but also the reverse order vector as a code vector at the time of searching the stochastic codebook.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coder for efficiently transmitting data by coding acoustic information as a digital signal with a small amount of information, transmitting it, and decoding it.

[0002]

2. Description of the Related Art In the field of digital mobile communications such as mobile phones, a low bit rate voice compression coding method is required to cope with an increase in subscribers, and research and development are progressing in each research institute. . In Japan, VSE with a bit rate of 11.2 kbps developed by Motorola
An encoding method called LP was adopted as a standard encoding method for digital mobile phones (digital mobile phones equipped with the same method were released in Japan in the fall of 1994).
Furthermore, a coding method called PSI-CELP with a bit rate of 5.6 kbps developed by NTT Mobile Communication Network Co., Ltd. has been adopted as a standardization method for the next mobile phone, and is currently in the stage of product development. All of these methods are CEL
P (Code Exited Linear Pred)
iction: MRSchroeder "High Quality Speech
At Low Bit Rates ”Proc.ICASSP'85 pp.937-940)), which is an improvement of the method that separates speech into source information and vocal tract information, and the source information is encoded. The input speech is coded by the index of a plurality of sound source samples stored in the book and the LPC (linear prediction coefficient) is coded for vocal tract information, and the vocal tract information is taken into consideration when coding the sound source information. And a comparison method (AbS: Analysis by Sy
It is characterized by adopting the N. Thesis).

Here, a basic algorithm of the CELP method will be described. FIG. 2 is a functional block diagram of a CELP type encoding device. First, the LPC analysis unit 22 performs autocorrelation analysis and L
The LPC coefficient is obtained by performing PC analysis, the obtained LPC coefficient is encoded to obtain the LPC code, and the obtained LPC code is decoded to obtain the decoded LPC coefficient. Next, in the adder 25, the sound source samples stored in the adaptive codebook 23 and the stochastic codebook 24 are taken out, the optimum gains for them are obtained, and the respective sound sources whose powers have been adjusted by the optimum gains are added to obtain a synthesized sound source To get Further, in the LPC synthesis unit 26, the synthesized sound source obtained in the addition unit 25 is filtered by the decoded LPC coefficient obtained in the LPC analysis unit 22 to obtain a synthesized sound. Further, the comparison unit 27 includes an adaptive codebook 23 and a stochastic codebook 24.
Adder 25, LPC synthesizer for all sound source samples
The distance between many synthetic sounds obtained by operating 26 and the input voice is calculated, and the index of the sound source sample at the smallest of the obtained distances is obtained. The parameter coding unit 28 obtains a gain code by coding the optimum gain, and collects the LPC code, the excitation sample index, the gain code, etc., in the transmission path.
Send to 29. Also, a synthetic excitation is created from the gain code and the index, stored in the adaptive codebook 23, and at the same time the old excitation sample is discarded. Further, the LPC synthesis unit 26 performs perceptual weighting using a linear prediction coefficient, a high-frequency emphasis filter, and a long-term prediction coefficient (obtained by performing long-term prediction analysis of input speech). Further, the excitation search using the adaptive codebook and the stochastic codebook is performed in a section (called a subframe) in which the analysis section is further divided.

[0004]

In order to mount the encoding / decoding device on a small device such as a mobile phone, it is desirable that the ROM capacity is small. However, the conventional method based on the CELP method requires a large number of code vectors for the fixed sound source. Some devices have been devised so that they can be used as code vectors while shifting long vectors, but since they are the same vector, the sound quality is worse than when many different code vectors are used. It had the problem of being lost. In addition, since calculation for creating a code vector must be performed each time it is created, there is a problem that a large amount of calculation is required.

[0005]

SUMMARY OF THE INVENTION In order to solve the above problems, the present invention codes not only the forward order of code vectors stored in a probabilistic codebook but also the reverse order of vectors in a codebook search. Used as a vector.

[0006]

With this configuration, a vector in the reverse order of a certain code vector has a large difference from the original code vector, which does not lead to deterioration in sound quality, and can be realized only by using the vector sample pointer in the reverse direction. The amount is almost the same as the conventional one, so that the size of the codebook can be halved with almost no deterioration in sound quality and without increasing the calculation amount.

[0007]

Embodiments of the present invention will be described below with reference to FIGS. 1 and 2 by taking the case of application to the CELP system as an example. Since the basic algorithm of the CELP method in the embodiment is the same as that of the conventional example (FIG. 2), the description thereof is omitted.

Among them, in the probabilistic codebook search, as shown in FIG. 1, two code vectors of forward order and reverse order are generated from one code vector and each is used as a search candidate. Then, the index of the optimum code vector and the coded 1 bit in the forward / reverse order are transmitted as a code. The search using the reverse order vector is a process that uses the sample pointer of the code vector in reverse, and thus can be realized with almost the same calculation amount as the conventional calculation amount. Therefore, this process reduces the number of vectors to 2
Since it can be doubled, it is 1 of the conventional codebook size.
Similar coding performance can be obtained with a stochastic codebook of size / 2.

In learning the code vector of the stochastic code length according to the present invention, processing is performed so that the code vector does not become close to left-right symmetry, or processing that does not become close to other vectors when reversed. This will improve the sound quality. An example of a specific method will be described below. (1) LBG algorithm (reference “IEEE TRANSACTIONS
ON COMMUNICATIONS, VOL.COM-28, NO.1, JANUARY 1980 "p
VQ (Vector Quantization: Vector Q) according to p.84-95
uantization) In the case of codebook creation, there is a method of adjusting the population vector in advance before creation.
The first half power and the second half power of each element of the vector are obtained, and when the power of the second half is larger than that of the first half, conversion is performed in reverse order for all the vectors. By this process, a centroid group with large power in the first half is obtained. (2) In codebook learning by sequential adaptation, there is a method of performing the same learning in reverse order. The processing procedure is described below. (1) The input speech is encoded and the optimum excitation and optimum gain of the adaptive codebook and the stochastic codebook are obtained. (2) An ideal sound source is obtained by applying an inverse filter of the synthesis filter of the LPC synthesis unit to the input speech, and an ideal stochastic codebook sound source is obtained by subtracting the optimum sound source of the adaptive codebook from the obtained ideal sound source. , The optimal sound source of the stochastic codebook is learned by the following (Equation 1).

[0010]

[Equation 1]

(3) The processes of (1) and (2) are repeated by inputting all the learning voice data. At this time, the learning coefficient is gradually reduced to converge.

[0012]

As described above, according to the present invention, since the reverse vector of a certain code vector has a large difference from the original code vector, it does not lead to the deterioration of the sound quality, and the vector sample pointer is used in reverse. The calculation amount is almost the same as the conventional one because it can be realized only by
The size of the codebook can be halved with almost no deterioration in sound quality and without increasing the calculation amount.

[Brief description of drawings]

FIG. 1 is a diagram showing a state of code vector reverse ordering according to the present invention.

FIG. 2 is a block diagram of the basic functions of the CELP method.

[Explanation of symbols]

 21 input speech 22 LPC analysis section 23 adaptive codebook 24 stochastic codebook 25 addition section 26 LPC synthesis section 27 comparison section 28 parameter coding section 29 transmission path

Claims (1)

[Claims] 1. An LPC analysis means for performing LPC (Linear Prediction Coding) analysis on an input speech signal (input speech) to obtain a linear prediction coefficient, an adaptive codebook in which past synthesized sound sources are stored, A stochastic codebook in which fixed excitations are stored, a sound source adding means for extracting a sound source from two codebooks and multiplying it by gains to obtain a synthetic sound source, and a linear prediction coefficient for the synthetic sound source By comparing the distance between the synthesized voice and the input voice with the LPC synthesis means for obtaining the synthesized voice by applying the used filter, 2
And a means for searching the most suitable one of the sound sources stored in one codebook, and in the probabilistic codebook search, the forward and reverse vectors of the stochastic codebook code vector are used as the sound source candidates. A speech coding apparatus characterized by: