JPH0667696A - Speech encoding method - Google Patents

Abstract

PURPOSE:To reduce the arithmetic operation quantity by detecting the analytic output of a past speech signal which has the shortest distance to the analytic output of a current input speech signal and encoding the current input speech on the basis of this detection output. CONSTITUTION:A speech signal S(n) which is sampled, for example, at a 8kHz sampling frequency fs and converted into a digital signal is inputted to an input terminal 9. A short-period predictive inverse filter 10 filters the input signal reversely to predict a sound generated at the inner part of the throat, and its residual output r(m) is supplied to a subtracter 8; and the output of the subtracter 8 is supplied as an output difference e(n) to an energy calculation part (SIGMA()<2>) 11. Then this energy calculation part 11 calculates the energy of the output difference e(n) between the residual output r'(n) and the residual output r(n) from a short-period predictive inverse filter 10 and then selects, for example, the dynamic code vector of a dynamic code vector 1 from a terminal 12 so as to minimize the energy of the output difference e(n).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a low bit rate speech coding method, and more particularly to a speech coding method for analyzing and coding a present input speech signal by utilizing a correlation with a past speech signal.

[0002]

2. Description of the Related Art In recent years, a low bit rate, that is, 4.8
~ 9.6 kbps voice coding method, vector sum excited linear prediction (VSELP: Vector Sum Excited Linear Pr
code excitation linear prediction (CELP: Code)
Excited Linear Prediction) has been proposed.

The technical contents of this VSELP are described in the special table 2-50 by Motorola Incorporated.
2135, "Digital Speech Coder with Improved Vector Excitation Source" and "VECTOR SUM EXCITED LIN"
EAR PREDICTION (VSELP): SPEECH CODING AT 8 KBPS: I
ra A. Gerson and Jasiuk: Paper presented at the In
t.Conf.on Acoustics, Speech and Signal Processing-
April 1990 ".

The voice encoding method using this VSELP is based on the analysis by synthesis.
) Codebook search to achieve high-quality voice transmission at a low bit rate. Also, VSEL
In a speech coding apparatus (speech coder) to which a speech coding method using P is applied, a time-varying linear filter in which a pitch and noise predictor forming characteristics of an input speech signal is introduced is stored in a codebook. Speech is encoded by being excited by selecting a code vector.

Specifically, for each frame of speech,
The speech coder introduces each code vector into the filter to form a speech signal and compares it with the original input speech signal to detect its differential output. This difference output is weighted by passing through a weighting filter based on human hearing. Here, it is desired to select a code vector that produces a minimum energy weighted differential output for the current frame.

FIG. 5 is a functional block diagram showing a schematic configuration of a speech coder to which the speech coding method using VSELP as described above is applied. This speech coder includes a dynamic codebook 51 in which analysis outputs of past speech signals are stored as a plurality of code vectors, and a first fixed codebook 52 in which code vectors related to noise components are stored.
And the second fixed codebook 53, there are three codebooks in total.

The input terminal 60 is sampled at a sampling frequency f _s = 8 kHz, for example, and a band is selected by a band pass filter (BPF) (not shown).
The audio signal S (n) converted into a digital signal by the D converter is input. The audio signal S (n) is supplied to the subtractor 59. On the other hand, the subtractor 59 is also supplied with the voice S ′ (n) synthesized with each code vector selected from the above three codebooks. The output of the subtractor 59 is used as the output difference e (n) in the energy calculation unit 61 (Σ
() ² ) is supplied. This energy calculation unit 61
Calculates the energy of the output difference e (n). Then, in order to minimize the energy of the output difference e (n), for example, from the terminal 50 to the dynamic codebook 51.
Dynamic code vector of is selected. Similarly, for the first fixed codebook 52 and the second fixed codebook 53, the first fixed code vector and the second fixed code vector that minimize the energy of the output difference e (n) are selected. It

That is, in order to search each code vector from the above three codebooks, the dynamic code vector stored in the dynamic codebook 51 and the first code vector stored in the first fixed codebook 52 are searched. The output difference between the voice S '(n) and the input voice signal S (n) formed by combining the fixed code vector and the second fixed code vector stored in the second fixed code book 53. The condition is that the energy of e (n) is minimized.

First, the dynamic codebook 51 described above.
The selection of the dynamic code vector stored in is described below.

The dynamic code book 51 has, for example, 128 (128 ways) dynamic code vectors, assuming that the sampling frequency is 8 kHz and one frame of voice is composed of 40 samples. In this case, the synthesis filter 58 has 12
The operation is performed on the eight code vectors. Then, the voice S ′ (n) output from the synthesis filter 58 and the input voice S
The subtracter 59 calculates the output difference e (n) from (n) and supplies it to the energy calculator 61. This energy calculator 6
1 calculates the energy value of the output difference e (n). Then, the optimum index j _opt that limits the optimum dynamic code vector is searched from the terminal 50 so as to minimize the energy of the output difference e (n).

Since the synthesis filter 58 is generally composed of a 10th-order IIR filter,
20 multiply-add operations are performed on the sample data. Therefore, the product-sum operation is performed 20 times, for example, for 40 sample data, and further, it is repeated for 128 code vectors.

Here, the dynamic code book 5 is used.
The code vector of 1 is supplied to the adder 57 after being multiplied by the coefficient β in the multiplier 54. Further, the code vector of the first fixed codebook 52 is supplied to the adder 57 after being multiplied by the coefficient γ ₁ in the multiplier 55, and the code vector of the second fixed codebook 53 is
After the coefficient γ ₂ is multiplied by the multiplier 56, the adder 57 is added.
Is supplied to. The adder 57 adds the multiplication results from the respective multipliers 54, 55 and 56 and supplies the addition output to the synthesis filter 58.

In FIG. 5, each of the code vectors of the first fixed codebook 52 and the second fixed codebook 53 surrounded by a broken line is also subjected to the above-mentioned calculation by the synthesis filter 58. In this case, the number of each code vector is 64, and the number of repeated operations is 64.

[0014]

By the way, the above-mentioned VS
In the speech coder to which the speech coding method using ELP is applied, if all filtering is performed for each code vector of the codebook as described above, the amount of calculation becomes very large. In other words, it is necessary to calculate the distance from the original speech for all filtered vectors in the dynamic codebook search range, and the amount of calculation is considerably larger than in the analytical method. . The ratio of this part to the processing time of the entire VSELP speech coding method is considerably larger than that of the other parts.

The present invention has been made in view of the above circumstances, and an object thereof is to provide a speech coding method capable of reducing the amount of calculation.

[0016]

A speech coding method according to the present invention is a speech coding method for analyzing and coding a present input speech signal by utilizing a correlation with a past speech signal,
A step of analyzing the current input voice signal, a detection step of detecting the analysis output of the past voice signal having the shortest distance from the analysis output of the input voice signal, and based on the detection output of the detection step The above problem is solved by a coding step for coding the current input speech.

A speech encoding method according to another invention is
A coding method for searching a codebook in which an analysis output of a past voice signal is stored as a plurality of code vectors and performing coding by utilizing a correlation with a current input voice signal. And a step of directly searching for a code vector in the codebook in which the distance from the analysis output of the input speech signal is the shortest, and the current index using the index of the code vector obtained by the search is used. And a step of encoding an input voice signal.

Furthermore, a speech coding method according to another invention uses a correlation with a current input speech signal by searching a codebook in which analysis outputs of past speech signals are stored as a plurality of code vectors. An encoding method for encoding, comprising a step of analyzing a current input voice signal and a direct search for a code vector in the codebook having a shortest distance from an analysis output of the input voice signal. A second step of indirectly searching for a code vector including the code vector obtained in the first search step and having the optimum correlation with the input voice signal from the code vectors in the vicinity of the code vector. And a step of encoding the current input speech signal using the code vector index obtained in the second search step. Te to solve the above-mentioned problems.

[0019]

Since the analysis output of the past voice signal having the shortest distance from the analysis output of the current input voice signal is detected and the current input voice is encoded based on this detection output, the amount of calculation can be reduced. .

[0020]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of a voice coding method according to the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a schematic configuration of an encoding apparatus (speech coder) which is a first embodiment to which a speech encoding method of the present invention is applied.

In FIG. 1, the speech coder includes a dynamic codebook 1 in which analysis outputs of past speech signals are stored as a plurality of code vectors, and a first fixed codebook 2 in which code vectors related to noise components are stored.
And the second fixed codebook 3 in total has three codebooks. The above dynamic codebook 1
It is a codebook that changes with time, and stores a plurality of code vectors based on analysis data past a certain time. Further, noise components are stored in a fixed state in the first fixed codebook 2 and the second fixed codebook 3.

At the input terminal 9, for example, a voice signal S sampled at a sampling frequency f _s = 8 kHz and converted into a digital signal by an A / D converter (not shown).
(N) is input. The audio signal S (n) is supplied to the short-term predictive inverse filter 10 having the inverse characteristic of the above-described conventional synthesis filter 58.

The short-term predictive inverse filter 10 applies an inverse filter to the input voice and predicts a so-called deep part of the throat. The output from the short-term predictive inverse filter 10 is the residual output r (n). The residual output r (n) is supplied to the subtractor 8.

Here, the above-mentioned dynamic code bucky 1
The code vector of is multiplied by the coefficient β in the multiplier 4 and then supplied to the adder 7. The code vector of the first fixed codebook 2 is supplied to the adder 7 after being multiplied by the coefficient γ ₁ in the multiplier 5. The code vector of the second fixed codebook 3 is multiplied by the coefficient γ ₂ in the multiplier 6 and then supplied to the adder 7. Then, the residual output r ′ (n) is supplied from the adder 7 to the subtracter 8.

That is, the subtracter 8 outputs the residual output r'according to each code vector of the above three codebooks.
(n) is also supplied. The output of this subtractor 8 is the output difference e
It is supplied to the energy calculation unit (Σ () ² ) 11 as (n).

The energy calculator 11 calculates the energy of the output difference e (n) between the residual output r '(n) and the residual output r (n) from the short-term prediction inverse filter 10. Then, in the first embodiment, for example, the dynamic code vector of the dynamic code book 1 is selected from the terminal 12 so as to minimize the energy of the output difference e (n).

That is, the above dynamic codebook 1
In order to search for a dynamic code vector from the dynamic code vector stored in the dynamic codebook 1, the residual code output r ′ (n) and the residual output r (n) from the short-term prediction inverse filter 10 are used. The output difference e (n) may be searched so that the energy of the output difference e (n) is minimized.

The dynamic code book 1 has, for example, 128 (128 ways) dynamic code vectors when the sampling frequency is 8 kHz and one frame of voice is composed of 40 samples. However, the synthesis filter used in the conventional example is not used in the first embodiment. Therefore, the subtracter 8 only compares the residual output r '(n) corresponding to the 128 code vectors with the residual output r (n) from the short-term prediction inverse filter 10. In other words, only 128 different comparison operations are performed for each of the 40 sample data.

That is, the first embodiment compares the residual output r '(n) corresponding to 128 code vectors with the residual output r (n) from the short-term prediction inverse filter 10, The current input speech signal is encoded by searching and extracting the optimum dynamic code vector from the terminal 12 as the optimum index J _{opt so} as to minimize the energy of the output difference e (n).

Next, the operation of the short-term predictive inverse filter 10 will be described with reference to FIG. In FIG. 2, the number of samples per frame of the audio signal S (n) input from the input terminal 9 of FIG. 1 is N _S. As described above, in this embodiment, the sampling frequency f _s is 8 kHz,
The sample number N _S is 40, for example. The size of the dynamic code book 1 (the number of dynamic code vectors to be selected) is Nc. In the present embodiment, as described above, the number is 128 (various). Further, the input voice signal to be analyzed is S (n).
Here, n is 0 or more and less than N _S. The state of the dynamic codebook 1 at that time point (time point n), that is, the dynamic code vector to be taken out is J (n). Here, n is 0 or more and less than Nc. Then, the short-term prediction inverse filter 1 at any time
The residual waveform of 0 is r (n). Here, n is 0 or more and less than N _S.

Usually, the spectrum of the voice has an envelope as shown in FIG. 2A having only formants. When the fast Fourier transform (FFT) is applied to this envelope, the spectrum becomes as shown in FIG. 2B, and the pitch is between the peaks. When this spectrum is input to the short-term prediction inverse filter 10, a residual output r (n) having a waveform with a reduced envelope as shown in C of FIG. 2 is obtained. That is, it is a waveform like white noise in which the pitch is slightly left and the redundancy depending on the formant synthesis is lost.

The dynamic codebook 1 has a dynamic code vector J as shown in D of FIG. 2 which is very similar to the residual output r (n) shown in C of FIG.
(N) is stored. Therefore, in the first embodiment, the subtracter 8 compares and operates the residual output r (n) and the residual output r '(n) corresponding to the dynamic code vector J (n).

Here, the residual output r (n) can be regarded as an N _S -dimensional vector according to the number of samples N _S in one frame, and this is designated as r. Also, a vector obtained by extracting N _S samples from the time point of the index J in the dynamic codebook 1 is defined as c _J. In this state, the distance (approximation degree) between r and c _J is obtained, and the pitch in this frame can be extracted by using the index J _opt that _minimizes the distance. At this time, the gain of c _J does not matter.

That is, | r-tc _J | is the minimum, |
Since the pitch J in this frame can be extracted when r-tc _J | ² is the minimum,

[0035]

[Equation 1]

It becomes Here, since r is fixed, the pitch can be fixed by selecting the index J _opt that maximizes (r · c _J ) ² / | c _J | ² .

Next, the difference in the amount of calculation between the above-mentioned conventional VSELP type voice coder and the first embodiment will be specifically described. First, the calculation amount of the VSELP type voice coder will be described. In the VSELP type voice coder, sounds are synthesized using the synthesis filter 58. Therefore, the number of filtering sequences z _J (n) for synthesizing speech is

[0038]

[Equation 2]

Can be expressed as Here, h (n-i) is the impulse response at the (n-i) samples of the synthesis filter 58.

Further, z _J (n) is z _J (n) = z _J-1 (n-1) + r (-J) h (n) where 1≤n≤N-1, z _J (0) = r (-J) h
(N).

This z _J (n) is 40 in the above VSELP.
X128. Next, in the VSELP type voice coder, the power G _J of the sound resynthesized from the dynamic codebook 51 is

[0042]

[Equation 3]

It becomes Where b ′ _J (n) is the output b of the code vector from the above dynamic codebook.
_J (n) is the output after passing through the synthesis filter H (z). Also, this power G _J is

[0044]

[Equation 4]

It can also be expressed as 2 in this conventional VSELP
It becomes 0x128.

The inner product C _J of the dynamic codebook and the original voice is

[0047]

[Equation 5]

Thus, the VSELP type voice coder is 40 × 128. The impulse response h (n) of the synthesis filter is 10 × 20 because this synthesis filter is a 10th-order IIR filter and 20 samples are used.

The following table 1 shows the calculation amount of the voice coder of this embodiment with respect to the calculation amount of the conventional VSELP type voice coder.

[0050]

[Table 1]

That is, in the first embodiment, since the synthesis filter is not used, the number of filtering sequences z _J (n) for synthesizing voice is 0, and the impulse response h (n of the synthesis filter is ) Is also zero.

Further, the sound power G _J is equal to the above VSELP.
The power of the speech coder of the system is the power when the sound is synthesized from the codebook and the calculation is complicated, whereas it is the power of the dynamic codebook. This means the size of c _J described above. That is, |
c _J | ² , which is 1 × 128 in the first embodiment.

The inner product C _J of the dynamic codebook and the original voice is the product of the squares of the sizes of r and c _JJ , which is also 40 × 128 in this embodiment.

However, in the first embodiment, since the short-term prediction inverse filter 10 is used, the number of filtering sequences is calculated as 40 × 128. Conventional VSE
Since the LP system voice coder does not use the inverse filter, the value is 0.

Therefore, comparing the VSELP type voice coder and the calculation amount of the voice coder of the first embodiment,
The number of operations is 13000 to 5648, which is less than half that of the conventional VSELP type voice coder.

Next, a voice coder according to the second embodiment will be described. FIG. 3 is a functional block diagram showing a schematic configuration of a voice coder according to the second embodiment. In FIG.
This voice coder includes a dynamic codebook 21,
The first fixed codebook 22 and the second fixed codebook 23 have a total of three codebooks.

The input terminal 29 is sampled at a sampling frequency f _s = 8 kHz, for example, and the A / D (not shown) is sampled.
Audio signal S converted into digital signal by converter
(N) is input. This audio signal S (n) is applied to the subtractor 3
Supplied to zero. The subtractor 30 is also supplied with the 0-input response from the synthesis filter 33, which is 0-input at the terminal 34. This 0 input response is formed by new LPC coefficients when the 0 input passes through the synthesis filter 33. Then, the subtractor 30 subtracts the 0 input response from the audio signal S (n), and supplies the subtraction result to the short-term prediction inverse filter 31 as P (n). A clear filter signal for clearing the filter state to 0 is supplied to the short-term predictive inverse filter 31 from the input terminal 32. After filtering one frame of the input audio signal S (n), the filter state is cleared to 0. initialize. The output from the short-term prediction inverse filter 31 is the residual output r (n). The residual output r (n) is supplied to the subtractor 28.

The dynamic code vector is supplied to the adder 27 after being multiplied by the coefficient β in the multiplier 24. The code vector of the first fixed codebook 22 is supplied to the adder 27 after being multiplied by the coefficient γ ₁ in the multiplier 25. Further, the code vector of the second fixed codebook 23 is multiplied by the coefficient γ ₂ in the multiplier 26 and then supplied to the adder 27.
Then, the output of the adder 27 is supplied to the subtractor 28 as a residual output r '(n).

The subtractor 28 outputs the residual output r (n).
The difference between the residual output r '(n) and the residual output r' (n) is supplied to the energy calculator 35 as an output difference e (n). The energy calculation unit 35 calculates the residual output r ′ (n) and the short-term prediction inverse filter 3 described above.
The energy of the output difference e (n) from the residual output r (n) from 1 is calculated. Then, in the second embodiment, the terminal 36 is used so that the energy of the output difference e (n) is minimized.
The current input speech signal is encoded by searching and extracting the index J _opt of the dynamic code vector of the dynamic codebook 21 from.

In the second embodiment, the synthesis filter used in the conventional example is not used in the preceding stage of the subtractor as in the first embodiment. Therefore, in the subtractor 28, 12
The residual output r '(n) corresponding to the eight code vectors is simply compared with the residual output r (n) from the short-term prediction inverse filter 31. That is, only 128 different comparison operations are performed for each of the 40 sample data.

In the second embodiment, after filtering one frame in the first embodiment, the filter state is left as it is and the next frame is processed, that is, the filter state is retained all the time. , Is used repeatedly, after filtering one frame, the filter state is cleared to 0, and then the next frame is processed.

This is effective, for example, for matching on the audio decoder side used in the communication system together with this audio coder. Normally, there is no problem because the raw voice is input to the encoder side, but on the decoder side, the voice is resynthesized without knowing the original raw voice.
At this time, inconvenience arises if the filter state is kept held as in the first embodiment. That is,
If the 0 input response is not subtracted from the input audio signal S (n) in the state where the filter coefficient of the internal state of the previous frame is new, adverse effects will occur. Then, 0 input is passed through the synthesis filter 33, and what is subtracted from the original voice signal is input to the short-term prediction synthesis filter 31.

In the second embodiment, in order to remove the influence of the previous frame, the 0 input response is subtracted from the input voice, and the short-term prediction inverse filter 31 obtains the residual r (n). , The subtraction result P (n) can be used effectively, and since the internal state of the filter is updated with the quantized data, the same state as on the decoder side is reproduced, and matching between the encoder and the decoder can be easily achieved.

Here, in the second embodiment, the amount of calculation can be suppressed to half or less as in the first embodiment. A detailed description of the calculation amount reduction will be omitted.

Next, a voice coder which is a third embodiment of the present invention will be described. FIG. 4 is a functional block diagram showing a schematic configuration of a voice coder according to the third embodiment. In FIG. 4, this speech coder has a total of three codebooks including a dynamic codebook 41, a first fixed codebook 42 and a second fixed codebook 43.

The input terminal 48 is sampled at, for example, a sampling frequency f _s = 8 kHz, and an A / D (not shown) is sampled.
Audio signal S converted into digital signal by converter
(N) is input. This audio signal S (n) is supplied to the short-term prediction inverse filter 49 and becomes the residual output r (n).
The residual output r (n) is supplied to the subtractor 50.

The dynamic code vector is supplied to the adder 47 after being multiplied by the coefficient β in the multiplier 44. The code vector of the first fixed codebook 42 is supplied to the adder 47 after being multiplied by the coefficient γ ₁ in the multiplier 45. Further, the code vector of the second fixed codebook 43 is supplied to the adder 47 after being multiplied by the coefficient γ ₂ in the multiplier 46.
The output of the adder 47 is supplied to the subtractor 50 and the synthesis filter 52 as a residual output r '(n).

The subtractor 50 outputs the residual output r (n).
The difference between the residual output r '(n) and the residual output r' (n) is supplied to the energy calculation unit 51 as an output difference e (n). The energy calculator 51 calculates the energy of the output difference e (n). Then, in order to minimize the energy of the output difference e (n), in the third embodiment, the index J _opt of the dynamic code vector of the dynamic code book 41 is searched and extracted from the terminal 55, for example.

Next, the code vector of the index J _opt and the code vector near the code vector are passed through the synthesis filter 52 to obtain the voice S '(n). This voice S ′ (n) is supplied to the subtractor 53. The input voice S (n) is also supplied to the subtractor 53. The output of the subtractor 53 is the output difference E (n).
Is supplied to the energy calculation unit 54. The energy calculator 54 calculates the energy of the output difference E (n). And, in this third embodiment, the output difference E
The current input speech signal is obtained by searching the optimum index J ′ _opt of the optimum dynamic code vector of the dynamic codebook 41 from the terminal 55 so as to minimize the energy of (n) and extracting the optimum index J ′ _opt. Encode.

As described above, in the third embodiment, the output difference e between the residual output r (n) from the short-term prediction inverse filter 49 and the residual output r '(n) corresponding to the dynamic code vector is output.
The index J _opt of the dynamic code vector that minimizes the energy of (n) is directly searched (first search step), and the code vector of this index J _{opt and} the code vector in the vicinity thereof are synthesized by the synthesis filter 52. To synthesize the voice S '(n), and the energy of the output difference E (n) between the voice S' (n) and the input voice S (n) is minimized. The current input speech signal is encoded by indirectly searching (second search step) the vector index _J'opt .

Therefore, in the third embodiment, the index search can be roughly performed in the first search step, and the vicinity thereof can be severely searched in the second search step. Therefore, an accurate code vector index can be searched even if the amount of calculation is reduced.

[0072]

The speech coding method according to the present invention is a speech coding method in which a current input speech signal is analyzed and coded by utilizing a correlation with a past speech signal. Since the analysis output of the past speech signal that has the shortest distance from the analysis output of the signal analysis step is detected, the encoding step encodes the current input speech based on the detection output of the above detection step. The amount can be halved compared to the conventional one.

A speech coding method according to another invention is
A coding method for searching a codebook in which an analysis output of a past voice signal is stored as a plurality of code vectors and performing coding by utilizing a correlation with a current input voice signal, wherein a search step is a current step. The code vector in the above codebook that has the shortest distance from the analysis output in the step of analyzing the input voice signal is directly searched, and the coding step uses the index of the code vector obtained in the above search step to obtain the current input. Since the audio signal is encoded, the amount of calculation can be halved as compared with the conventional method.

Further, a speech coding method according to another invention searches a codebook in which analysis outputs of past speech signals are stored as a plurality of code vectors and utilizes the correlation with the current input speech signal. A coding method for performing coding, wherein the first search step directly searches for a code vector in the codebook having the shortest distance from the analysis output of the step of analyzing the current input speech signal, Of the code vector obtained in the first search step, the code vector obtained in the first search step is indirectly searched for a code vector having the optimum correlation with the input voice signal, and the code vector Since the encoding step encodes the current input speech signal using the index of the code vector obtained in the second search step, the calculation amount can be reduced as compared with the conventional case.

[Brief description of drawings]

FIG. 1 is a block diagram showing a schematic configuration of a voice coder according to a first embodiment.

FIG. 2 is a characteristic diagram of a short-term predictive inverse filter used in the first embodiment.

FIG. 3 is a block diagram showing a schematic configuration of a voice coder of a second embodiment.

FIG. 4 is a block diagram showing a schematic configuration of a voice coder of a third embodiment.

FIG. 5 is a functional block diagram of a voice coder using VSELP.

[Explanation of symbols]

 1-Dynamic codebook 10 --- Short-term prediction inverse filter 11 --- Energy calculator

Claims (3)

[Claims] 1. A voice encoding method for analyzing and encoding a current input voice signal by utilizing a correlation with a past voice signal, the method comprising: analyzing the current input voice signal; A detection step of detecting an analysis output of the past speech signal having the shortest distance from the analysis output, and an encoding step of encoding a current input speech based on the detection output of the detection step. Speech coding method. 2. A coding method for searching a codebook in which an analysis output of a past speech signal is stored as a plurality of code vectors, and coding by utilizing a correlation with a current input speech signal. The step of analyzing the current input speech signal, the step of directly searching the code vector in the above codebook that has the shortest distance from the analysis output of this input speech signal, and the index of the code vector obtained by the above search Encoding the current input speech signal using the. 3. A coding method for searching a codebook in which an analysis output of a past speech signal is stored as a plurality of code vectors and performing coding by utilizing a correlation with a current input speech signal. A step of analyzing a current input voice signal and a direct search for a code vector in the code book having the shortest distance from an analysis output of the input voice signal;
Including the code vector obtained in the first search step, and indirectly searching for a code vector having the optimum correlation with the input speech signal from the code vectors in the vicinity of the code vector. A voice encoding method comprising: a second search step; and a step of encoding a current input voice signal using the index of the code vector obtained in the second search step.