JP3230380B2 - 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,
In particular, the present invention relates to a CELP-type audio encoding device that encodes an audio signal at a low bit rate of about 8 to 4 kb / s with high quality.

[0002]

2. Description of the Related Art In recent years, the digitization of automobile telephones and cordless telephones using radio waves as a medium has been rapidly advancing. Since a radio wave has a small frequency band that can be used for this type of telephone, it is important to develop a coding method for a low bit rate audio signal in order to reduce the occupied band. Examples of this type of coding system having a bit rate of about 8 to 4 kb / s include, for example, Eyecast Prop. 85, published in the United States in 1985 (ICASSP pro
ceedings 85) Schroeder and Atal, Code Excited Linear Prediction, pages 937-940: High Quality Speech Atrobitrates (M. Schroeder and
B. S. Atal, "Code-excited l
earprediction: High quali
ty speech at lowbit rate
s ") (Reference 1) and the like.
e Excited LPC Coding) is known.

[0003] In the CELP which is the first conventional speech coding apparatus described in Document 1, a coding process is performed on the transmitting side in the following procedure. First, every frame (for example, 20m
In s), a short-term prediction code representing a frequency characteristic of the voice is extracted from the voice signal to be coded (short-term prediction). Next, the frame is further divided into sub-frames (for example, 5 ms) of a small section. For each sub-frame, a pitch parameter indicating a long-term correlation (pitch correlation) is extracted from a past sound source signal, and the speech signal of the sub-frame is predicted for a long time by the pitch parameter. The long-term prediction uses a delay code representing the pitch correlation using an adaptive codebook composed of a subframe-length excitation signal (adaptive code vector) obtained by delaying the past excitation signal by delay samples corresponding to each delay code. Is determined by the following procedure.
That is, the delay code is changed by the size of the adaptive codebook, and an adaptive code vector corresponding to each delay code is extracted. A composite signal is generated using the extracted adaptive code vector, and error power with respect to the audio signal is calculated. An optimal delay code that minimizes the calculated error power, an adaptive code vector corresponding to the optimal delay code, and its gain are determined.

Next, a synthesized signal generated from an excitation code vector extracted from a noise signal (excitation codebook), which is a kind of quantization code prepared in advance, and a residual signal obtained by the long-term prediction are described. A sound source code vector that minimizes error power and its gain are determined (sound source codebook search). An index indicating the type of the adaptive code vector and the excitation code vector determined in this way and an index indicating the type of the gain and the spectrum parameter of each excitation signal are transmitted.

A specific method of searching for a delay code of an adaptive code vector and a quantization code of an excitation code vector is performed in the following procedure. First, a signal z [n] is calculated by performing weighting on audibility and subtracting a past influence signal from the input audio signal x [n]. Next, calculated by the above short-term forecast,
The synthesis filter H composed of the quantized and inversely quantized spectral parameters is driven by the code vector ej [n] of the quantization code j to calculate the synthesized signal Hej [n]. Next, in the following equation, the signal z [n] and the signal He are
A quantization code j that minimizes the error power E of j [n] is obtained.

[0006]

Here, N _s is a subframe length, H is a matrix for realizing a synthesis filter, and g _ej is a code vector e.
j, respectively. In practice, equation (1) is expanded as follows.

[0008]

The numerator Cj in the equation (2) is a cross-correlation, a denominator Gj
Is the autocorrelation, which can be calculated by the following equations.

[0010]

The calculation of the auto-correlation Gj and the cross-correlation Cj is executed after the signal Hej [n] is calculated by driving the synthesis filter, that is, filtering. then,
Since the calculation of the filtering process is performed over the size of the codebook, the amount of calculation (the number of times of product sum) of the calculation for the frame to be processed becomes very large.

Japanese Patent Application No. 22858/1990 discloses a second conventional speech coding apparatus for reducing the amount of computation during long-term prediction.
No. 1 (Reference 2), “Digital voice coder and method for obtaining parameters used in the coder”
Is well known. This method realizes a long-term prediction with a small amount of calculation without deteriorating accuracy by performing preliminary selection of a delay code by an open loop and searching for a code near the delay code determined by the preliminary selection by a closed loop. are doing.

FIG. 5 is a block diagram showing a conventional CELP type speech coding apparatus including first and second speech coding apparatuses. Referring to FIG. 5, the speech coding apparatus shown in FIG. And a transmission line 3 connecting the encoding unit 1 and the decoding unit 2 to each other.

The encoding unit 1 includes a buffer circuit 11 for storing an audio signal input from an input terminal TI, an LPC analysis circuit 12 for extracting an LPC coefficient as an audio spectrum parameter, and a parameter quantization circuit for quantizing the LPC coefficient. Circuit 13, a weighting circuit 14 for weighting the audibility of an audio signal, an adaptive codebook 15 for storing past sound sources, and a long-term prediction circuit 16 for searching for an adaptive code vector that is a delay code representing a pitch correlation. An excitation codebook 17 which is a codebook in which excitation codes of subframe lengths representing long-term prediction residuals are stored; an excitation codebook search circuit 18 for determining an optimal excitation codevector from the excitation codebook; Gain that stores the parameters representing the gain terms of the vector and sound source code vector And Dobukku 19, the gain codebook searching circuit determines the quantization gain of the adaptive code vector and the excitation code vector from the gain codebook 40
And a multiplexer 4 for combining and outputting a code sequence
1 is provided.

The sound source codebook 17 uses the vector quantization (V) described in Japanese Unexamined Patent Application Publication No. 2-22955 or Japanese Unexamined Patent Application Publication No. 2-22956 as a noise codebook described in Reference 1.
Q) A learning codebook learned by an algorithm may be used.

The decoding unit 2 is a demultiplexer 21 for decoding the supplied transmission code into predetermined code sequences.
And adaptive codebook 2 identical to adaptive codebook 15
2, a sound source codebook 23 that is the same as the sound source codebook 17, a gain codebook 24 that is the same as the gain codebook 19, a synthesis filter 25 that reproduces a sound signal from the generated sound source and the sound synthesis filter, and an output terminal. TO
And an output terminal TO for audio output.

Referring to FIG. 6 showing a block diagram of the configuration of the long-term prediction circuit 16, the conventional long-term prediction circuit 16 includes a delay code variable circuit 161 for changing a delay code by the size of an adaptive codebook, and a delay code variable circuit. The adaptive code vector e corresponding to the delay code d set in 161
Adaptive code vector generation circuit 162 for generating _d (n) from past signals stored in codebook 166
When the adaptive code vector e _d (n) and a synthetic signal for receiving the weighted adaptive code vector H · e _d synthesis filter 163 to generate a (n), synthesized signal H and the audio signal stored in the audio buffer An evaluation function calculation circuit 164 for calculating an evaluation function representing an error power of e _d (n);
An optimal delay code determination circuit 165 that determines an optimal delay code CD using an evaluation function corresponding to all variable delay codes d; a past excitation signal, a residual signal, and a weighting signal for adaptive codebook search; Alternatively, a codebook 166 which is a buffer for accumulating a signal such as an audio signal, and a buffer which accumulates an audio signal in a section where encoding processing is performed, are searched for a delay code which minimizes error power with respect to this audio signal. An audio buffer 167 is provided.

Next, with reference to FIGS. 5 and 6, a description will be given of the flow of processing of the conventional audio encoding circuit. First, the encoding unit 1 inputs an audio signal from an input terminal TI and sends it to a buffer 11. Store. The LPC analysis circuit 12 performs a short-term prediction analysis using the audio signal of a certain sample stored in the buffer 11, and calculates an LPC coefficient of the audio signal. The LPC coefficient obtained by the LPC analysis circuit 12 is quantized by the parameter quantization circuit 13, and the quantized code CL of the LPC coefficient is sent to the multiplexer 41, and is dequantized and used for the subsequent encoding processing.

On the other hand, the audio signal stored in the buffer 11 is converted into a perceptually weighted signal SW by the weighting circuit 14 using the quantized / dequantized LPC coefficients. The signal is supplied to the circuit 18 and the gain codebook search circuit 40, respectively, and used for the subsequent codebook search.

Next, a codebook search for the signal SW is performed using each of the adaptive codebook 15, the sound source codebook 17, and the gain codebook 19. First, a long-term prediction is first performed by the long-term prediction circuit 16, and an optimum delay code CD representing the pitch correlation is determined as described later.
The delay code CD is transferred to the multiplexer 41, and a corresponding adaptive code vector is generated. Next, after subtracting the influence of the adaptive code vector, the sound source codebook search circuit 18 performs a sound source codebook search,
The quantization code CS is determined, an excitation code vector is generated, and the quantization code CS is transferred to the multiplexer 41. After obtaining the adaptive code vector and the excitation code vector, the gain codebook search circuit 40 calculates the gain of these two excitations and transfers the code CG to the multiplexer 41. The multiplexer 41 combines these codes CL, CD, CS, and CG and converts them into a transmission code CT, and transfers this code CT to the decoding unit 2 via the transmission path 3.

The decoding unit 2 uses the demultiplexer 2 to convert the transmission code CT input from the transmission path 3 into codes CL, CD,
Decomposes into CS and CG respectively. The filter coefficient is decoded from the code CL corresponding to the LPC coefficient and transferred to the synthesis filter 25. Adaptive codebook 1 from delay code CD
5 to generate an adaptive code vector. An excitation code vector is generated from the quantization code CS corresponding to the excitation using the excitation codebook 17. The gain of the adaptive code vector and the excitation code vector is calculated from the code CG corresponding to the gain, and each excitation is multiplied by a gain term to generate an input signal of the synthesis filter. Finally, the synthesis signal is synthesized by the synthesis filter 25 using the input signal and output from the terminal TO.

Next, the operation of the long-term prediction circuit 16 will be described with reference to FIG.
1 sets the delay code d by changing the delay code corresponding to the signal SW by the size of the adaptive codebook. The delay code d is preferably a code representing a decimal point delay, but may be a code representing an integer delay. Next, the adaptive code vector generation circuit 1
62 is an adaptive code vector e corresponding to the delay code d.
_d (n) is generated from the codebook 166. Next, the synthesis filter 163 generates an adaptive code vector e _d (n) is the zero-state synthesis signal for receiving the weighted adaptive code vector H · e _d (n). Evaluation function calculation circuit 16
4 is a weighted adaptive code vector H · e _d (N), the cross-correlation C _d of the zero input response subtraction signal z (n) in the encoding section is stored, the autocorrelation G _d is calculated respectively in the audio buffer 167 , The evaluation function C _d ² corresponding to the error power
/ _Gd is calculated. After the processes of the adaptive code vector generation circuit 162, the synthesis filter 163, and the evaluation function calculation circuit 164 are performed on all delay codes to be changed by the delay code variable circuit 161, the optimum delay code determination circuit 16
In step 5, the delay code d that maximizes the evaluation function C _d ² / G _d is determined as the optimum delay code CD.

[0023]

The above-mentioned first conventional speech coding apparatus requires a large amount of computation when searching for a long-term prediction codebook, and realizes speech coding with good sound quality at a low bit rate. There is a drawback that it is difficult.

Even in the conventional second speech coding apparatus in which the amount of calculation is reduced, the amount of calculation is considerably large, and further reduction in the amount of calculation is required for realizing a speech coding apparatus having a low bit rate and good sound quality. There was a problem that reduction was required.

It is an object of the present invention to reduce the amount of computation in long-term prediction and to realize a speech encoding device with a low bit rate of about 4 kb / s and good sound quality.

[0026]

SUMMARY OF THE INVENTION A speech encoding apparatus according to the present invention analyzes a speech signal having a predetermined frame length, extracts a spectrum parameter of the speech signal, and generates a corresponding short-term prediction code. Adaptive codebook storage means for further dividing the frame length into predetermined subframe lengths and storing past excitation signals of this subframe length unit; and the excitation signal read from the adaptive codebook storage means. An optimal delay code representing a pitch correlation of the voice signal by calculating a delay value that minimizes a weighted square error as an evaluation scale by delaying within a predetermined pitch period and a gain for the delayed past sound source. A long-term prediction means for searching for an adaptive code vector that is a long-term prediction, and an excitation code that is a quantization code indicating a residual signal after the long-term prediction. A sound source code book storage means for storing vector, in the speech coding apparatus and a sound source code book search means for determining the optimum excitation code vector from the excitation codebook, said long-term prediction means, before
The past sound for searching for an adaptive codebook storage means
Stores a signal including a source signal, the residual signal, and a weighted signal
A codebook is a buffer that, before the delay code
A thinning-out delay code that is a processing-target delay code reduced by making it variable while thinning out within the codebook size.
Variable sign thinning variable means for generating a signal , and said thinning delay
Code and inputs the adaptive code corresponding to the thinning-out delay code.
Code vector candidates from the codebook.
Adaptive code vector generating means, and the adaptive code vector
Weighted adaptive code that is a zero-state synthesized signal
Synthesis filter means for generating a code vector and encoding processing
That stores the audio signal of the coding section that is the section where
A buffer, the weighted adaptive code vector and the sound
Zero in the coding interval stored in the voice buffer
Coding from cross-correlation with input response subtracted signal and auto-correlation
An evaluation function corresponding to error power of an audio signal with respect to the audio signal.
And evaluation function calculation means to calculate the number, the weighting 2 Noayama
Before maximizing the evaluation function to correspond to the minimization of the difference
Optimum delay for determining a thinning-out delay code as the optimal delay code
And an extension code determining means .

[0027]

FIG. 1 is a block diagram showing the configuration of a long-term prediction circuit 16A according to a first embodiment of the present invention. In addition to the adaptive code vector generation circuit 162, the synthesis filter 163, the evaluation function calculation circuit 164, the optimum delay code determination circuit 165, the code book 166, and the audio buffer 167, the adaptive code vector generation circuit 162 is used instead of the delay code variable circuit 161. There is provided a delay code uniform thinning-out variable circuit 168 for changing the delay code of the code book while thinning it out uniformly.

Next, the operation of this embodiment will be described with reference to FIG.
Reference numeral 8 designates a delay code d to be processed by changing the delay code while thinning it uniformly within the size of the codebook, and generates an adaptive code vector e _d (n) by the adaptive code vector generation circuit 162 for each delay code d. , Synthesis filter 1
The adaptive code vector H · _ed (n) weighted by 63
Generation processing, evaluation function C by evaluation function calculation circuit 164
Calculation processing of _d ² / G _d is performed. As the uniform thinning method, only odd-numbered delay codes are extracted. The adaptive code vector generation processing, the weighted adaptive code vector generation processing, the evaluation function calculation processing, and the optimum delay code determination processing are common to those in the related art, and thus description thereof is omitted.

After performing each of the above processes on all the delay codes d to be changed, the optimum delay code determination circuit 165 determines the delay code _d that maximizes the evaluation function C _d ² / G _d in the same manner as before. It is determined and output as the optimum delay code CD.

As described above, in this embodiment, the amount of calculation can be reduced by thinning out the delay codes at the time of long-term prediction.

Next, the configuration of the long-term prediction circuit 16B according to the second embodiment of the present invention will be described with reference to FIG. Then, the difference between this long-term prediction circuit 16B and the long-term prediction circuit 16A of the first embodiment is that the delay code is changed while thinning out the delay code of the adaptive codebook non-uniformly instead of the delay code uniform thinning-out variable circuit 168. In other words, a variable delay code non-uniform thinning variable circuit 169 is provided.

The operation of this embodiment is the same as that described above except that the delay code non-uniform thinning-out variable circuit 169 changes the delay code while thinning it non-uniformly within the size of the codebook and sets the delay code d to be processed. Is similar to the first embodiment.

As the non-uniform thinning-out method, a delay code having a small delay extracts all codes, and a delay code having a large delay extracts only even-numbered delay codes.

Next, the configuration of the long-term prediction circuit 16C according to the third embodiment of the present invention will be described with reference to FIG. Then, the difference between this long-term prediction circuit 16B and the long-term prediction circuit 16B of the second embodiment is that, in addition to the delay code non-uniform thinning-out variable circuit 169, the output side of each of the synthesis filter 163 and the audio buffer 167. In other words, sample thinning circuits 170 and 180 for thinning the input signal to 1 / D (arbitrary integer) are provided.

The operation of this embodiment is similar to that of the second embodiment, except that the delay code non-uniform thinning-out variable circuit 169 changes the delay code while thinning it out non-uniformly within the size of the codebook and changes the delay code to be processed. d, the adaptive code vector generation circuit 162 generates an adaptive code vector e _d (n) for each delay code d, and the synthesis filter 163 generates a weighted adaptive code vector H · _ed (n). Then, the weighted adaptive code vector H · _ed (n) is 1 / D (for example, D =
Thinning code vector H · e _d performs sample thinning 2)
(N) ′ is generated. The sample thinning circuit 180 thins out the sample of the zero input response subtracted signal z (n) in the coding section stored in the audio buffer 167 to 1 / D in the same manner, and thins out the zero input response subtracted signal z (n). 'Is generated.
The evaluation function calculation circuit 164 calculates these thinning code vectors H
The second embodiment calculates a cross-correlation C _d and an auto-correlation G _d between e _d (n) ′ and a decimation zero input response subtracted signal z (n) ′, respectively, and calculates the second embodiment from the cross-correlation and auto-correlation C _d and G _d. Similarly, the evaluation function C _d ² / G _d is calculated. The execution of the sample thinning is performed by downsampling using a low-pass filter or by simple thinning. The other processes are performed in the same manner as in the second embodiment.

In the present embodiment, a further reduction in the amount of calculation is realized by using a signal from which samples have been thinned out during execution of the correlation calculation.

Next, the construction of the long-term prediction circuit 16D according to the fourth embodiment of the present invention will be described with reference to FIG. Then, the difference between the long-term prediction circuit 16D and the long-term prediction circuit 16C of the third embodiment is that M optimum delay codes CD based on the evaluation function C _d ² / G _d of the output of the evaluation function calculation circuit 164 are used as a reference. , And a final delay code determination circuit 172 that determines an optimal delay code CD from these M delay code candidates instead of the optimal delay code determination circuit 165. .

The operation of the present embodiment will be described. The evaluation function evaluation function C _d obtained as a result of the same processing as in the third embodiment for all of the variable delay codes d of the delay code non-uniform thinning circuit 169 will be described. ² / _Gd is the delay code candidate decision circuit 1
71. The delay code candidate determination circuit 171 uses the evaluation function evaluation function C _d ² / G _d as a reference for M (for example, M =
5) Preliminarily select the delay code candidates D1 to DM (= 5). The final delay code determination circuit 172 determines an optimum delay code CD from among the delay code candidates D1 to D5 by a search method similar to the related art.

In this embodiment, the precision is maintained by performing the same search as in the prior art using the thinning of the samples for the preliminary selection of the optimum delay code and using the delay code corresponding to the preselected delay code candidate. The amount of calculation can be reduced as it is.

Although the embodiments of the present invention have been described above, the present invention is not limited to the above embodiments, and various modifications can be made.

For example, although the cross-correlation ² / auto-correlation is used as the evaluation function, the same effect can be obtained with the cross-correlation ² . Although the weighted adaptive code vector obtained by combining the adaptive code vectors is used, the same effect can be obtained by using the adaptive code vector itself without performing the combining process. Further, instead of using the signal stored in the audio buffer as the zero-state subtraction signal, an input audio signal, a residual signal, or a weighting signal for audibility may be used. Furthermore, the same effect can be obtained by using a past input speech signal, a residual signal, or a weighting signal on audibility instead of using a past sound source signal as a codebook. Furthermore, although the method of filtering with the synthesis filter H is used for calculating the cross-correlation and the auto-correlation, the same effect can be obtained by a method using an approximate expression. This approximation method is described, for example, in ICASP86 Proceedings 86, published in the United States in 1986.
Transco and Atal, Efficient Procedures for Finding the Optimum Innovation in Stochastic Coder, pp. 375-2378 (IM. Transco, B .;
S. Atal, "EFFICIENTPROCEDUR
ES FOR FINDING THE OPTIMU
MINNOVATION IN STOCASTIC
CODERS "), etc. Also, instead of narrowing down the optimal delay code to one, a plurality of candidates are obtained and the main selection is performed in the next step, and the simultaneous optimal search is performed in the subsequent codebook search. The effect can be obtained.
Although the description has been made using the analysis circuit, the same effect can be obtained by using another analysis method such as the BURG method for extracting a spectrum parameter. Although the description has been made using the LPC coefficient, it is apparent that a similar effect can be obtained with other spectral parameters such as the PARCOR coefficient and the LSP coefficient. Furthermore, it is needless to say that a multi-stage sound source codebook search circuit can be applied instead of a single-stage sound source code book search circuit without departing from the gist of the present invention.

[0042]

As described above, the speech coding apparatus of the present invention includes the delay code thinning-out variable means for changing the delay code while thinning out the delay code. Since the number of target delay codes is reduced, the amount of calculation in adaptive codebook search is reduced.

[Brief description of the drawings]

FIG. 1 is a block diagram of a long-term prediction circuit showing a first embodiment of a speech encoding apparatus according to the present invention.

FIG. 2 is a block diagram of a long-term prediction circuit showing a second embodiment of the speech coding apparatus of the present invention.

FIG. 3 is a block diagram of a long-term prediction circuit showing a third embodiment of the speech coding apparatus of the present invention.

FIG. 4 is a block diagram of a long-term prediction circuit showing a fourth embodiment of the speech coding apparatus of the present invention.

FIG. 5 is a block diagram illustrating an entire speech encoding apparatus.

FIG. 6 is a block diagram illustrating an example of a conventional speech encoding device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Encoding part 2 Decoding part 3 Transmission line 11 Buffer circuit 12 LPC analysis circuit 13 Parameter quantization circuit 14 Weighting circuit 15,22 Adaptive codebook 16 Long-term prediction circuit 17,23 Sound source codebook 18 Sound source codebook search circuit 19, Reference Signs List 24 gain codebook 40 gain codebook search circuit 41 multiplexer 21 demultiplexer 25 synthesis filter 161 delay code variable circuit 162 adaptive code vector generation circuit 163 synthesis filter 164 evaluation function calculation circuit 165 optimal delay code determination circuit 166 codebook 167 audio buffer 168 Delay code uniform thinning variable circuit 169 Delay code non-uniform thinning variable circuit 170, 180 Sample thinning circuit 171 Delay code candidate determination circuit 172 Final delay code determination circuit

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-4-301900 (JP, A) JP-A-5-11799 (JP, A) JP-A-5-61499 (JP, A) 158497 (JP, A) (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 19/00-9/14 H04B 14/04 H03M 7/30

Claims (5)

(57) [Claims] 1. A speech analyzing means for analyzing a speech signal having a predetermined frame length, extracting a spectrum parameter of the speech signal and generating a corresponding short-term prediction code, and further comprising a sub-frame length having a predetermined frame length. And an adaptive codebook storing means for storing the past sound source signal of this subframe length unit, and the sound source signal read from the adaptive codebook storing means is delayed by a predetermined pitch cycle as an evaluation scale. Calculate a delay value that minimizes a weighted square error and a gain for the delayed past sound source, search for an adaptive code vector that is an optimal delay code representing a pitch correlation of the voice signal, and perform long-term prediction. Long-term prediction means, and excitation codebook storage means for storing an excitation code vector which is a quantization code indicating the residual signal after the long-term prediction And a sound source codebook searching means for determining an optimum sound source code vector from the sound source codebook, wherein the long-term predicting means is used for searching for the adaptive codebook storing means.
Including the past sound source signal, the residual signal, and the weighting signal
A codebook which is a buffer for accumulating signals, and a delay code to be processed which is reduced by changing the delay code while thinning it within the size of the codebook.
Delay code thinning variable means for generating a certain thinning delay code
And the thinning-out delay code is input and corresponds to the thinning-out delay code.
The candidate of the adaptive code vector to be
Adaptive code vector generating means for generating from said adaptive
A candidate for a code vector is input and
Synthesis filter means for generating a matching adaptive code vector
And the audio signal of the coding section, which is the section where the coding process is performed,
Stored in the audio buffer, the weighted adaptive code vector and the audio buffer.
Zero input response subtraction in the stored coding interval
Before the coded speech signal from the cross-correlation with the signal and the auto-correlation
Calculate the evaluation function equivalent to the error power for the voice signal
Evaluation function calculating means; and the evaluation function corresponding to minimization of the weighted square error.
The decimation delay code that maximizes the function is replaced with the optimal delay code
An audio encoding apparatus comprising: an optimal delay code determining means for determining a signal . 2. A speech encoding apparatus according to claim 1, wherein said delay code thinning-out variable means includes a uniform thinning means for thinning out said delay code uniformly for every predetermined number. 3. The non-uniform thinning-out means according to claim 1, wherein said delay code thinning-out variable means comprises a non-uniform thinning-out means for thinning out said delayed codes in accordance with a predetermined non-uniform thinning-out target code determining method. Audio coding device. 4. The long-term predicting means thins out each of the speech signal and the weighted adaptive code vector corresponding to the delayed code, which are supplied to an evaluation function calculating means for calculating the weighted square error, at a predetermined rate. The speech encoding apparatus according to claim 1, further comprising first and second sample thinning means. 5. A delay candidate determining means for selecting a predetermined number of delay candidates which are candidates for the adaptive code vector, an optimum delay code corresponding to the adaptive code vector from the delay candidates. 2. The speech encoding apparatus according to claim 1, further comprising: a final delay code determination unit that determines the final delay code.