JP2000047695A - Encoding device and decoding device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an encoding device and decoding device which can reproduce a voice signal even if all codes are not obtained. SOLUTION: A vocal tract coefficient quantizer 202 obtains a quantization synthesis filter coefficient aq(i) from quantization prediction result (w) separated from a total code by a multiplex separating circuit 201. In a decoding system y1 of a first stage, a prediction voice is obtained based on a filter coefficient aq(i), in a decoding system y2,..., yn, after a second stage, a prediction voice of the filter coefficient aq(i) is obtained based on each parameter J2, L2,..., JN, LN. Prediction voices of each stage are synthesized by switches 204, 213,... and adders 203, 212,... and outputted.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] The present invention relates to an encoding device and a decoding device.

[0002]

2. Description of the Related Art In a mobile phone or a mobile phone, for example, 4
Coding at a low bit rate of kbit / s or less is realized. Such an encoding method is described in, for example, “Pitch Sync
hronous Innovation CELP (PSI-CELP) -PDC half-rate voice CODEC- "(Oya et al., IEICE Technical Report, R
CS93-78, 1993).

In this encoding method, the excitation source is cycled in synchronization with the pitch cycle of speech, and encoding based on CELP is performed to reduce the bit rate of encoded speech data.

Further, by encoding the pitch period, the residual, and the code length gain in units of subframes, the improvement of sound quality, the amount of calculation, and the reduction of memory requirements are realized.

[0005]

However, most of the encoding methods currently put into practical use cannot be decoded unless all pieces of encoded information for one frame are collected, and the encoding method disclosed in the above-mentioned document is not available. The same applies to the chemical conversion method.

For example, when transmitting coded information by packet communication or the like, the coded information of one frame may be divided into several packets and transmitted. Is not guaranteed to be transmitted on-time, and some packets may be lost.

[0007] For this reason, when information transmission requiring real-time performance is performed, a frame that cannot be decoded unless a packet is lost or transmitted by a predetermined time occurs. When such a frame that cannot be decoded occurs, the sound is interrupted on the telephone and an error occurs on the data communication.

The present invention has been made in view of the above-described problems, and has as its object to provide an encoding device and a decoding device that can decode even when all encoded information cannot be obtained. .

[0009]

An encoding apparatus according to the present invention comprises: a linear prediction analyzer for performing prediction based on an input signal; a linear prediction coefficient quantizer for quantizing a prediction result of the linear prediction analyzer; An adaptive excitation codebook storing an adaptive excitation code corresponding to a first parameter, a statistical excitation codebook storing a statistical excitation code corresponding to a second parameter, and an adaptive gain code corresponding to a third parameter And a gain codebook in which a statistical gain code is stored, an adaptive excitation code, a statistical excitation code, a synthesizing unit that synthesizes each gain code to output a synthesized signal, and a prediction result quantized by a linear prediction coefficient quantizer. And a comparing means for comparing the input signal and the predicted speech, and varying the first to third parameters according to the comparison result of the input signal. Explore the combination of production parameters predictive speech well approximated by a first respective optimal code
Optimizing means for setting the parameter, the second parameter and the third parameter, and a target signal setting for outputting a difference between an input signal and a predicted speech when the first to third parameters of the optimal code are used as a target signal A first coding system having means, a statistical excitation codebook storing a statistical excitation code corresponding to a fourth parameter, and a gain codebook storing a statistical gain code corresponding to a fifth parameter. A synthesis means for synthesizing a statistical excitation code, a statistical code, and outputting a synthesized signal; a synthesis filter for obtaining a predicted speech from the synthesized signal in accordance with a prediction result quantized by a linear prediction coefficient quantizer; Comparing means for comparing a target signal supplied from a quantization system with a predicted sound, and a predicted sound which best approximates an input signal according to a comparison result of the comparing means by varying first to third parameters. The combination of the generated parameters to search, the fourth parameter of each optimal code, and optimization means for the fifth parameter, fourth and fifth optimization code
One or a plurality of second codes having target signal setting means for outputting a difference between a target signal from the preceding coding system and a predicted speech when the parameter is used as a target signal of the next coding system. And a multiplexing unit that multiplexes the prediction result quantized by the linear prediction coefficient quantizer and the first to fifth parameters of each optimal code and outputs the multiplexed result as a total code. I do.

A decoding device according to the present invention decodes an audio signal coded by the coding device according to the present invention.
Demultiplexing means for separating the first to fifth parameters;
An adaptive excitation code corresponding to the first parameter, a statistical excitation code corresponding to the second parameter, and an adaptive gain code and a statistical gain code corresponding to the third parameter are combined to form a combined signal, A first decoding system for obtaining a predicted speech from the synthesized signal in accordance with the result of the generalized prediction, a statistical excitation code corresponding to the fourth parameter, and a statistical gain code corresponding to the fifth parameter to synthesize a synthesized signal. One or more second decoding systems that form and obtain predicted speech from the synthesized signal in accordance with the quantization prediction result, and synthesis that synthesizes each predicted speech obtained by the first and second decoding systems. Means.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a block diagram showing a configuration of an encoding device according to one embodiment of the present invention.

[0012] As shown in FIG.
A vocal tract analysis system x0 that performs vocal tract analysis of the input signal S, and multi-stage encoding systems x1, x2, and x that each encode the input signal.
3,..., Xn−1, xn, and the respective encoding systems x1 to xn
And a multiplexing circuit (multiplexing means) 112 for multiplexing the outputs of the above.

The vocal tract analysis system x0 includes a vocal tract analyzer (linear predictive analyzer) 101 for performing vocal tract analysis of the input signal S, and a voice for performing quantization based on the vocal tract analysis result of the vocal tract analyzer 101. Path coefficient quantizer (linear prediction coefficient quantizer) 102.

First-stage encoding system (first encoding system) x1
Are adaptive codebooks (adaptive codebooks) 103 holding adaptive code vectors corresponding to supplied coefficients I (first input parameters), and supplied coefficients K and L
A gain codebook (gain codebook) 104 holding an adaptive code gain and a statistical code gain corresponding to (third input parameter), and a supplied coefficient J (second
Statistical codebook (statistical codebook) 105 in which a statistical code vector corresponding to the input parameter is stored.
, A multiplier (synthesizing means) 106 for calculating the product of the adaptive code vector and the adaptive code gain, and a multiplier (synthesizing means) 10 for calculating the product of the statistical code vector and the statistical code gain
7, an adder (synthesizing means) 108 for adding outputs of the multipliers 106 and 107, and a synthesis filter 10 for obtaining a synthesized speech vector from the output (excitation vector) of the adder 108.
9, an adder (comparing means, target signal setting means) 110 for calculating a difference between the input signal S and the synthesized speech vector, and coefficients i, j, which minimize the weighted error from the calculated difference.
Weighted error calculation circuit (optimizing means) 1 for obtaining k and l
11 and a switch (target signal setting means) 113 for switching the output destination of the adder 110.

Encoding system of second and subsequent stages (second encoding system)
, xn-1, xn are statistical code books 114 holding statistical code vectors corresponding to the supplied coefficients J2, J3,..., JN-1, JN, respectively. .., 121 and the coefficient L2 supplied respectively,
Gain code book 115 in which statistical code gains corresponding to L3,..., L-1, LN are held.
, 122, multipliers 116,..., 123 for obtaining the product of the statistical code vector and the statistical code gain, and a synthesis filter for obtaining a synthesized speech vector from the output (excitation vector) of the multipliers 116,. 117, ...,
., 126 for obtaining a difference between the input signal S and the synthesized speech vector, and coefficients j1, k1, j2 for minimizing a weighted error from the obtained difference.
., 126 for calculating k2,..., jN-1, kN-1, jN, kN. Also, the encoding system x2 of the second stage to the (n-1) th stage
xn-1 includes switches 119,... for switching output destinations of the adders 118,.

The multiplexing circuit 112 includes a vocal tract coefficient quantizer 1
02 (quantization index w) and coefficients (indexes I, J, K, L, J1, K1,...) From the weighted error calculation circuits 111, 120,.
, JN, KN) and multiplex the same to the transmission path. It should be noted that the information may be recorded on a recording medium instead of being transmitted to the transmission path.

Hereinafter, the operation of the encoding apparatus configured as described above will be described. For example, an audio signal sampled at 8 KHz is input to the encoding device as an input signal S.
A frame composed of a predetermined number of samples (for example, 40
ms) or every subframe (for example, 10 ms).

First, in the encoding system x1 of the first stage, the selection terminal of the switch 113 is connected to the connected terminal on the weighted error calculation circuit 111 side. Thereby, the adder 11
The output of 0 is ready to be supplied to the weighted error calculation circuit 111.

The vocal tract analyzer 101 analyzes the vocal tract of the input signal S for one frame (or one subframe) to obtain a synthesized filter coefficient a (i), and a vocal tract coefficient quantizer 102
To supply. The vocal tract coefficient quantizer 102 quantizes the supplied synthesis filter coefficient a (i), supplies the quantization index w to the multiplexing circuit 112, and also converts the quantized synthesis filter coefficient aq (i) into each synthesis filter ( 109, 11
7,... 124).

The adaptive code book 103 stores an adaptive code vector eai (i = 1 to 1) corresponding to the supplied coefficient I.
M). The statistical code book 105 stores a statistical code vector esj (j = 1) corresponding to the supplied coefficient J.
To N) are output. The gain codebook 104 stores an adaptive code gain bk corresponding to the supplied coefficients K and L.
(K = 1 to A), statistical code gain gl (l = 1 to B)
Is output.

Multiplier 106 is adapted to adaptive codebook 103
Is multiplied by the adaptive code gain e from the gain codebook 104 and the adder 10
8, and the multiplier 107 supplies the statistical codebook 105
Is multiplied by the statistical code gain gl from the gain codebook 104 and the adder 10
8 The adder 108 is connected to each of the multipliers 106 and 107
Are added to obtain an excitation vector eijkl, which is supplied to the synthesis filter 109. The synthesis filter 109 obtains a synthesized speech vector vijkl from the excitation vector eijkl according to the quantized synthesis filter coefficient aq (i), and supplies this to the adder 110.

The adder 110 subtracts the synthesized speech vector vijkl from the input signal S, and obtains the obtained difference signal Eijk.
Output l. The weighted error calculation circuit 111 searches for a combination of i, j, k, and l that minimizes the weighted error based on the supplied difference signal Eijkl. The optimum combination indexes I, J, K, and L corresponding to i, j, k, and l thus obtained are output to the multiplexing circuit 112.
And supplied to each of the codebooks 103 to 105.

Next, the first-stage encoding system x1 connects the selection terminal of the switch 113 to the adder 11 of the next-stage encoding system x2.
8 is connected to the selected terminal. This allows
The output of the adder 110 is ready to be supplied to the next stage.

Then, each of the codebooks 103, 104,
105 is a vector eai, esj and gain bk, g according to the optimal combination index I, J, K, L, respectively.
Output l. Adaptive codebook 103 updates the internal state using optimal excitation vector eIJKL, which is output eijkl of adder 108 at this time.

Further, the synthesis filter 109 generates an optimum synthesized voice vIJKL using the optimum excitation vector eIJKL, generates a target signal v1 by subtracting from the input voice S, and outputs the target signal v1 via the switch 113 to the encoding system x2 of the next stage. Adder 11
8 is supplied.

Subsequently, the next-stage encoding system x2 outputs the target signal v1.
Is encoded. The encoding system x2 clears the internal state of the synthesis filter 117 every time before calculating the synthesized speech. In the coding system x2, the selection terminal of the switch 119 is connected to a selected terminal connected to the weighted error calculation circuit 120. Thus, the output of the adder 118 can be supplied to the weighted error calculation circuit 120.

The statistical code book 114 stores a statistical code vector es2m (m = 1) corresponding to the supplied coefficient J2.
To N1). The gain codebook 115 has a statistical code gain g2n corresponding to the supplied coefficient L2.
(N = 1 to B1) is output. The multiplier 215 multiplies the statistical code vector es2m by the statistical code gain g2n to generate an excitation vector esg2mn, and generates the synthesis filter 1
17. The synthesis filter 117 generates the excitation vector esg2m according to the quantized synthesis filter coefficient aq (i).
Then, a synthesized speech vector t2mn is obtained from the n
To supply.

The adder 118 subtracts the synthesized speech vector t2mn from the target signal v1 to obtain a difference signal Emn, and supplies the difference signal Emn to the weighted error calculation circuit 120 via the switch 119. The weighted error calculation circuit 120 uses the difference signal Emn to search for a combination of m and n that minimizes the weighted error. The optimum combination indices J2 and L2 corresponding to m and n obtained in this way are used as multiplexing circuits 11
2 and the respective codebooks 114 and 115.

Next, the second-stage encoding system x2 connects the selection terminal of the switch 119 to the selected terminal connected to the adder of the next-stage encoding system x3. Thus, the output of the adder 118 can be supplied to the encoding system x3 in the next stage.

Each of the codebooks 114 and 115 outputs a vector es2m and a gain g2n according to the optimum combination index J2 and L2, and generates a synthesized speech by the synthesis filter 117 using the optimum excitation vector e2 generated thereby. Then, a target signal v2 is created by subtracting the target signal v1 from the target signal v1 and supplied to the next-stage encoding system x3 via the switch 119.

Similarly, the encoding systems x3 to xn of the subsequent stages
-1 is performed, and the target signal vN-1 is supplied to the last-stage coding system xn.

The encoding system xn clears the internal state of the synthesis filter 124 every time before calculating the synthesized speech.

The statistical code book 121 stores a statistical code vector esNp (p = 1) corresponding to the supplied coefficient JN.
To NN). The gain codebook 122 stores a statistical code gain gNq corresponding to the supplied coefficient LN.
(Q = 1 to BN) is output. The multiplier 123 generates the excitation vector esgNpq by multiplying the statistical code vector esNp and the statistical code gain gNq, and generates the synthesis filter 1
24. The synthesis filter 124 calculates a synthesized speech vector tNpq from the excitation vector esgNpq.

The adder 125 subtracts the synthesized speech vector tNpq from the target signal vN-1 supplied from the preceding stage to obtain a difference signal Epq, and supplies the difference signal Epq to the weighted error calculation circuit 126. The weighted error calculation circuit 126 searches for a combination of p and q that minimizes the weighted error. The optimum combination indexes JN and LN corresponding to p and q thus obtained are stored in the multiplexing circuit 112 and each codebook 1.
21, 122.

The multiplexing circuit 112 includes the vocal tract coefficient quantizer 10
, And JN, KN are multiplexed with the quantization indices w from 2 and the indexes I, J, K, L, J2, L2,..., JN from the weighted error calculation circuits 111, 120,. It is transmitted to the transmission line as a total code C.

FIG. 2 is a block diagram showing a configuration of a decoding apparatus according to one embodiment of the present invention.

This decoding device, as shown in FIG.
Each coefficient (w, I, J, K, L, J2, L2,..., J) is calculated from the total code C supplied via the transmission line.
N, LN), and a quantized synthesis filter coefficient aq according to the quantization index w.
The vocal tract coefficient quantizer 202 for obtaining (i) and the indexes (I, J, K, L, J2, L2,..., JN, LN) from the demultiplexing circuits (demultiplexing means) 201, respectively.
A decoding system y1 for reproducing an optimal excitation vector based on
, yn.

First stage decoding system (first decoding system) y1
Is supplied with an adaptive codebook 209 that outputs an adaptive code vector corresponding to the supplied index I, an adaptive code gain 210 that outputs an adaptive code gain and a statistical code gain corresponding to the supplied indexes K and L, and A statistical code book 211 that outputs a statistical code vector corresponding to the index J, and a synthesis filter 205 that obtains a synthesized speech vector from the output (excitation vector) of the adder 206 based on the quantized synthesis filter coefficient aq (i). , An adder 203 to which the output of the synthesis filter is supplied, and a switch 204 for controlling the supply of the output of the second-stage decoding system y2 to the adder 203.

Second and subsequent decoding systems (second decoding system)
, yn are statistical code books 216,..., 220 that output adaptive code vectors corresponding to the supplied indexes J2, J3,.
And supplied indexes J2, J3,..., JN
, 221 that output statistical code gains according to the above, and multipliers 215,..., 2 that calculate the product of the statistical code vector and the statistical code gain
Based on the quantized synthesis filter coefficient aq (i) and the output (excitation vector) of the multipliers 215,...
, 218.

The decoding systems of the second to N-1 stages are
Switches 213,... For sequentially adding the outputs of the lower-order decoding system and the higher-order decoding system;
...

Hereinafter, the operation of the decoding apparatus configured as described above will be described. The demultiplexing circuit 201 calculates each coefficient (w, w, c) from the total code C supplied via the transmission line.
I, J, K, L, J2, L2,..., JN, LN), and supplies them to the corresponding vocal tract coefficient quantizer 202 and decoding systems y1 to yn.

The vocal tract coefficient inverse quantizer 202 performs quantization synthesis filter coefficients aq (i) according to the quantization index w.
, And each of the synthesis filters (205, 214,.
18).

In the first stage of the decoding system y1, the selection terminal of the switch 204 is connected to the open-side connected terminal as an initial state. As a result, the output of the second-stage decoding system y2 cannot be supplied to the adder 203.

The adaptive code book 209 outputs an optimal adaptive code vector ea corresponding to the adaptive code index I supplied from the demultiplexing circuit 201. The statistical code book 211 outputs the optimal statistical code vector es corresponding to the statistical code index J. The gain codebook 210 has an optimum adaptive code gain b and an optimum statistical code gain g corresponding to the gain code indexes K and L.
Is output.

The adder 206 adds the product of the optimal adaptive code vector ea and the optimal adaptive code gain b from the multiplier 207 and the product of the optimal statistical code vector es and the optimal statistical code gain g from the multiplier 208. An optimal excitation vector e is generated and supplied to the synthesis filter 205 and the adaptive codebook 209. The synthesis filter 205 obtains a synthesized speech S0 from the optimal excitation vector e according to the quantized synthesis filter coefficient aq (i). Also, the adaptive codebook 20
9 updates the internal state based on the optimal excitation vector e.

Subsequently, if the statistical code index J2 and the gain code index L2 have been obtained, the second stage decoding system y2 executes the following processing.

First, the second stage decoding system y2 connects the selection terminal of the switch 204 of the first stage decoding system to the selected terminal connected to the adder 212 and opens the selection terminal of the switch 213. Connected to the selected terminal on the side.

The statistical code book 216 stores the optimal statistical code vector es2 corresponding to the statistical code index J2.
Is output. The gain codebook 217 stores the optimum statistical code gain g2 corresponding to the gain code index L2.
Is output. The multiplier 215 calculates the optimal statistical code vector e.
The optimum excitation vector e2 is generated by multiplying s2 by the optimum statistical code gain g2, and supplied to the synthesis filter 214. After clearing the internal state, the synthesis filter 214 obtains a synthesized speech v2 from the optimal excitation vector e2 based on the quantized synthesis filter coefficient aq (i).

The codes (indexes I, J, K,
L, J2, L2,..., JN, LN) are obtained in the same manner as described above up to the decoding system at the stage where the decoded speech is obtained. Finally, the switch 20 up to the stage before the stage from which the sign was obtained
When the selection terminals of 3, 212,... Are connected to the selected terminals connected to the next stages, the outputs of the respective stages are sequentially added, and the output of the adder 203 becomes the sum of the synthesized voices of the respective stages. This is output as the final synthesized speech signal S.

Alternatively, if the first-stage index is not obtained, the decoding system of the second-stage or later-stage index performs decoding in the same manner as described above and outputs synthesized speech. You may make it.

As described above, in this encoding device, the output of each stage is encoded and transmitted, and therefore, even if all the encoded data cannot be reproduced by the decoding device, the reproduced code is reproduced. The audio signal can be reproduced using the encoded data. For example, as a result of an experiment using a voice signal as an input voice S, the first-stage indexes I, J, K,
Even when L was not reproduced, an audible audio signal could be reproduced.

[0052]

In the coding apparatus according to the present invention, the quantized prediction result obtained by quantizing the linear prediction result, and the first to fifth parameters of the optimum code obtained by the first and second coding systems. Are synthesized and encoded as a total code, so that it is possible to create a code that can reproduce an audio signal even if all parameters are not obtained.

Further, in the decoding device according to the present invention, even if all the parameters are not obtained, the prediction speech obtained by the first and second decoding systems is synthesized by the obtained parameters. An audio signal can be reproduced.

[Brief description of the drawings]

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

FIG. 2 is a block diagram illustrating a configuration of a decoding device according to an embodiment of the present invention.

[Explanation of symbols]

x1, x2,..., xn coding system, y1, y2,.
.., yn decoding system, 101 vocal tract analyzer, 102,
202 vocal tract coefficient quantizer, 109, 117, 124,
205, 214, 218 synthesis filter, 112 multiplexing circuit, 201 demultiplexing circuit

Claims (2)

[Claims] A linear prediction analyzer for performing prediction based on an input signal, a linear prediction coefficient quantizer for quantizing a prediction result of the linear prediction analyzer, and an adaptive excitation code corresponding to a first parameter are stored. The adaptive excitation codebook, the statistical excitation codebook storing the statistical excitation code corresponding to the second parameter, the adaptive gain code corresponding to the third parameter, and the gain codebook storing the statistical gain code. Synthesizing means for synthesizing an adaptive excitation code, a statistical excitation code, and each gain code to output a synthesized signal; and obtaining a predicted speech from the synthesized signal according to a prediction result quantized by the linear prediction coefficient quantizer. A synthesis filter, a comparing unit for comparing the input signal with the predicted voice, and varying the first to third parameters to best approximate the input signal according to a comparison result of the comparing unit. Explore the combination of parameters for generating the measurement audio,
Optimizing means for setting a first parameter, a second parameter, and a third parameter of an optimum code, respectively, and a difference between the input signal and the predicted speech when the first to third parameters of the optimum code are used. A first encoding system having target signal setting means for outputting as a target signal, a statistical excitation codebook storing a statistical excitation code corresponding to a fourth parameter, and a statistical gain corresponding to a fifth parameter A gain codebook in which a code is stored, and the statistical excitation code,
Synthesizing means for synthesizing the statistical code and outputting a synthesized signal; a synthesis filter for obtaining a predicted speech from the synthesized signal in accordance with a prediction result quantized by the linear prediction coefficient quantizer; Comparing means for comparing the supplied target signal with the predicted sound; varying the first to third parameters; and generating a predicted sound that best approximates the input signal according to the comparison result of the comparing means. Optimizing means for searching for a combination of parameters and setting them as a fourth parameter and a fifth parameter, respectively, of the optimal code, and the preceding encoding system using the fourth and fifth parameters of the optimal code. One or a plurality of second encoding systems having target signal setting means for outputting a difference between the target signal and the predicted speech as a target signal of a next-stage encoding system, and the linear prediction coefficient quantizer. And a multiplexing unit that multiplexes the prediction result quantized by the first and fifth parameters of each of the optimum codes and outputs the result as a total code. 2. A prediction based on an input signal is performed, a prediction result is quantized, an adaptive excitation code corresponding to a first parameter, a statistical excitation code corresponding to a second parameter, and a third excitation code.
The adaptive gain code and the statistical gain code corresponding to the parameters are combined to form a combined signal, and a predicted speech is obtained from the combined signal according to the quantized prediction result (quantized prediction result). The first to third parameters are varied according to the comparison result between the signal and the predicted signal, and a combination of parameters for generating a predicted speech that best approximates the input signal is searched. , A second parameter and a third parameter, and a difference between the input signal and the predicted speech when the first to third parameters of the optimal code are used is output as a target signal. The second coding system combines the statistical excitation code corresponding to the fourth parameter and the statistical gain code corresponding to the fifth parameter to output a synthesized signal,
A predicted speech is obtained from the synthesized signal in accordance with the quantization prediction result, and the first to third parameters are varied in accordance with a comparison result between a target signal supplied from a preceding coding system and the predicted speech. Then, a combination of parameters for generating a predicted speech that best approximates the input signal is searched for, and the fourth parameter and the fifth parameter of the optimal code are used, and the fourth and fifth parameters of the optimal code are used. The difference between the target signal from the preceding coding system and the predicted speech at that time is output as a target signal of the next coding system, and the quantization prediction result and the first to fifth codes of the respective optimal codes are output. A demultiplexing unit that separates the quantization prediction result and the first to fifth parameters from a total code multiplexed and encoded with the parameter, an adaptive excitation code corresponding to the first parameter, A statistical excitation code corresponding to the second parameter;
A first decoding system that combines an adaptive gain code and a statistical gain code corresponding to the parameters to form a synthesized signal, and obtains a predicted speech from the synthesized signal in accordance with the quantization prediction result; A statistical excitation code corresponding to the four parameters and a statistical gain code corresponding to the fifth parameter are combined to form a synthesized signal, and a predicted speech is obtained from the synthesized signal according to the quantization prediction result. Alternatively, a decoding apparatus comprising: a plurality of second decoding systems; and synthesizing means for synthesizing predicted voices obtained by the first and second decoding systems.