JPH028900A - Voice encoding and decoding method, voice encoding device, and voice decoding device - Google Patents

Abstract

PURPOSE:To improve the quality of a reproduced voice by finding a sound source signal consisting of multiple pulses in each subframe according to the sound source signal of a past subframe by utilizing pitch correlations between subframes and in a subframe. CONSTITUTION:A subframe division part 355 divides a voice signal into subframes of predetermined time length (including at least plural pitches) equal to or shorter than a frame section. Pitch prediction is carried out among frames by using a sound source signal which is found in a past subframe and stored in a sound source pulse memory 305 to remove the pitch correlation between subframes and then pitch prediction is performed even in the subframe to re move the pitch correlation in the subframe, thereby finding the sound source signal consisting of multiple pulses. Consequently, there is not a large increase in the amount of sent information and the sound quality can be improved with a sent information and the sound quality can be improved with a relatively small arithmetic quantity.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a speech encoding/decoding method, a speech encoding device, and a speech decoding device, and in particular to a method for efficiently encoding and decoding an audio signal at a low bit rate. The present invention relates to a speech encoding/decoding method, a speech encoding device, and a speech decoding device for encoding.

[Conventional technology]

A multi-pulse encoding method is known as a method for transmitting audio signals at a low bit rate, for example, about 16 kb/s or less. These represent the sound source signal as a combination of multiple pulses (multipulse), represent the characteristics of the vocal tract with a digital filter, and transmit the information on the sound source pulse and the filter coefficients after determining them for each fixed time interval (frame). ing. For details of this method, see for example Araseki,
“M” by Ozawa, Ono+ Ochiai
ulti-pulse Excited 5peech
CoderBased on Maximun+C
rosscorrelation 5earch A1
gorithn+”, (GLOBECOM 83.
IEEE Global Telecommunica
tion, lecture number 23.3.1983) (Reference 1
)It is described in. This method outputs a good audio signal after decoding by separately representing the vocal tract information and the sound source signal, and by using a combination of a plurality of pulse trains as a means to represent the sound source signal.

Here, the basic concept for obtaining a pulse train representing a sound source signal will be explained using FIG. 3.

An audio signal divided into frames is inputted from an input terminal 800 in the figure. Spectral parameters determined from the audio signal of the current frame are input to the synthesis filter circuit 820. An initial multipulse is generated in the sound source calculation circuit 810 and inputted to the synthesis filter circuit 820 to obtain a synthesized speech waveform as an output. A subtracter 840 subtracts the synthesized speech waveform from the input signal. This result is input to a weighting circuit 850, and the weighting error power in the current frame is obtained as an output. Then, the amplitude and position of a specified number of multipulses are determined in the sound source calculation circuit 810 so that this weighting minimizes the error power.

[Problem to be solved by the invention]

However, in the above-mentioned conventional technology, the sound quality was good when the bit rate was sufficiently high and the number of sound source pulses was sufficient, but there was a problem in that the sound quality deteriorated as the bit rate was lowered.

In order to improve this problem, a pitch prediction multi-pulse method has been proposed that utilizes the single periodicity (pitch correlation) of each pitch of a multi-pulse sound source.

For details of this method, see Japanese Patent Application No. 58-139022.
The detailed description can be found in the No. 1 specification (Reference 2), so the explanation will be omitted here. In the pitch prediction multi-pulse method of Document 2, pitch prediction is performed using the pitch correlation of the sound source only within a frame interval, and the pitch correlation between frames is not used, so the sound quality improvement effect of pitch prediction is not so much. In particular, if the frame interval is set to about 10 to 20 m5 to avoid long transmission delays, male speakers with a long pitch period (typically, the pitch period is 8 to 10 m5)
5th place), there was a problem in that pitch prediction had almost no effect.

On the other hand, in order to improve the above-mentioned problems, a pitch prediction multi-pulse method has been proposed in which the pitch of a sound source is predicted not only within a frame interval but also between frames, and between frames. For example, regarding this method, reference can be made to Japanese Patent Application No. 1982-273936 (Document 3), so the explanation will be omitted here. Since a spectral envelope synthesis filter is required, there is a problem in that the amount of calculation to obtain the filter increases, and the amount of information for transmitting four types of filter coefficients increases in total.

The purpose of the present invention is to provide audio encoding and decoding that can reproduce better audio than ever before even when the bit rate is high or even when the bit rate is lowered, and that can be realized with a relatively small amount of calculation. An object of the present invention is to provide a method, a speech encoding device, and a speech decoding device.

[Means to solve the problem]

In the audio encoding/decoding method of the present invention, on the transmitting side, a discrete audio signal is input and divided into predetermined frame sections, and a spectral parameter representing a spectral characteristic and a pitch parameter representing a pitch are extracted from the audio signal. The audio signal is divided into subframe sections including at least a plurality of the pitches, and the sound source signal consisting of multi-pulses is divided into the subframes by pitch prediction between subframes and within the subframe based on past sound source signals. On the receiving side, based on the past sound source signal, the pitch parameter and the sound source signal are used to predict the pitch between subframes and within the subframe, and the driving sound source signal is calculated, and the driving sound source signal is calculated using the spectral parameter. and outputs a synthesized audio signal that satisfactorily represents the audio signal.

Furthermore, the audio encoding device of the present invention divides an input audio signal into predetermined frame sections, and performs parameter calculation to obtain and encode a spectral parameter representing a spectral characteristic and a pitch parameter representing a pitch from the audio signal. circuit, a subframe dividing circuit that divides the audio signal into subframe sections including at least a plurality of the pitches, and performs pitch prediction over the subframes based on past sound source signals, and further performs pitch prediction within the subframes. and a multiplexer circuit that combines and outputs the output of the parameter calculation circuit and the output of the sound source pulse calculation circuit.

Furthermore, the audio decoding device of the present invention includes: a demultiplexer circuit that inputs, separates, and decodes a code representing a spectral parameter of an audio signal, a code representing a pitch parameter, and a code representing a sound source signal; and the decoding device. a pitch recovery filter that performs inter-subframe pitch prediction based on the converted pitch parameters and past sound source signals, and further calculates a driving sound source signal by predicting the pitch within the subframe using the decoded sound source signal; The present invention is characterized by comprising a synthesis filter circuit that synthesizes an audio signal that satisfactorily represents the audio signal using the audio source signal and the decoded spectral parameter.

[Effect]

According to the present invention, an audio signal is divided into subframes of a predetermined time length (including at least a plurality of pitches) equal to or shorter than a frame interval, and a sound source signal obtained in a past subframe is After performing pitch prediction between frames and removing pitch correlation between subframes, perform pitch prediction within subframes using . Furthermore, in the present invention, 1 m
A voice encoding/decoding method and a voice encoding device that use a pitch recovery filter and a set of spectral envelope synthesis filters, and can significantly improve sound quality without significantly increasing the amount of transmitted information and with a relatively small amount of calculation. The present invention is characterized in that it can realize an audio decoding device and an audio decoding device.

〔Example〕

Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an embodiment of the speech encoding/decoding method, speech encoding device, and speech decoding device according to the present invention, and FIG. FIG. 2 is a block diagram for explaining a pitch prediction method.

In the audio encoding and decoding processing according to the present invention, on the transmitting side, a discrete audio signal is input and divided into predetermined frame sections, and a spectral parameter representing spectral characteristics and a pitch parameter representing pitch are extracted from the audio signal. The audio signal is divided into subframe sections including at least a plurality of the above-mentioned pitches, and the sound source signal consisting of multi-pulses is divided into the subframes by pitch prediction between subframes and within the subframe based on the past sound source signal. On the reception side, based on the past sound source signal, the driving sound source signal is obtained by pitch prediction between subframes and within the subframe using the pitch parameter and the sound source signal, and using the spectral parameter. and outputs a synthesized audio signal that satisfactorily represents the audio signal.

First, explanation will be given using FIG. 2. The audio signal for each frame section (for example, 20 m5) is collected by the subframe dividing unit 355.
The frame is divided into subframe sections that are equal to or shorter than the frame section. From the audio signal of the frame, a coefficient a+ (here, M-th order) of the spectral envelope synthesis filter 320 is determined by a well-known method (eg, LPC analysis method). Further, by a well-known method (eg, autocorrelation method), the coefficients of the pitch recovery filter 325 and the pitch period T8 (here, for simplicity, it is assumed to be first-order) are determined.

Here, the spectral envelope synthesis filter 320. The transfer characteristics of the pitch recovery filter 325 are expressed by the following equations (1) and (21), respectively.

H, (z) = The sound source pulse memory 305 stores the sound source signal one pitch period before the past subframe. First, the switch 280 is turned to the 1 side, and the past multi-pulse train is input to the pitch reproduction filter 325 to reproduce the pitch of the current subframe. Next, the switch 290 is turned to the 1 side, and the signal is output to the spectrum envelope synthesis filter 320. This output is output to subtracter 330. The subtracter 330 extracts the spectral envelope synthesis filter 32 from the audio signal in the subframe section.
The output of 0 is subtracted to obtain e and (n), which are output to the auditory weighting filter 350. Here, the filter 350 can use the same configuration as the weighting circuit described in Document 2 above. The perceptual weighting is performed using 2 of the output of the filter 350.
The root mean value E is output to the sound source pulse calculation section 310.

Next, flip the switches 280 and 290 to the 2 side. The sound source pulse calculation unit 310 performs intra-subframe pitch prediction to minimize the value of E using, for example, the method described in Reference 1 or the same method as the sound source calculation unit in Reference 2.3. A predetermined number of multi-pulses with large amplitudes are obtained using this method. Thereafter, the signal 'K(n) regenerated by the large amplitude multi-pulse is transferred to the pitch regeneration filter 3.
25. is obtained using a spectral envelope synthesis filter 320, and subtracted in a subtracter 330 to obtain a residual e,(n),
This is output to the filter 350 for auditory weighting. For auditory weighting, the root mean square value of the output of the filter 350 is calculated and output to the small amplitude sound source calculation section 315.

Next, flip the switch 290 to the 3 side. Small-amplitude multi-pulses do not have strong pitch periodicity like large-amplitude multi-pulses, and are considered to occur almost randomly. Therefore, after obtaining the large amplitude multi-pulse, the pitch recovery filter 32
5 is separated and a predetermined number of small amplitude multipulse trains are obtained using only the spectral envelope synthesis filter 320.

For how to obtain multipulses, reference can be made to the above-mentioned document 1. Note that instead of a small amplitude multi-pulse train, a codebook with a predetermined number of dimensions and number of bits is created as a more efficient representation.
Among these codebooks, the optimal one may be selected so as to minimize the root mean square value. The codebook may be created in advance by training on learning data, or a random number sequence having a certain statistical distribution (for example, Gaussian distribution) may be arranged with various phases.

The former method of creating a codebook is, for example, Makhou
“Vector Quantizatio” by Mr. 1 et al.
n in 5peech Coding” (Proc.
, IEfiE, PP, 1551-1588.198
5 years) (Reference 4). Furthermore, the latter method is described, for example, in “Code-excited linear pr.
edition (CBLP): high-qua
lityspeech at very low bi
t rates” (Proc, ICASSP+pp
, 26-29.1985) (Reference 5). Also, regarding the codebook selection method and gain calculation method using the error minimization criterion, see the above-mentioned document 4.5.
etc. can be referred to.

On the other hand, in order to greatly reduce the amount of calculation while keeping the characteristics substantially the same, the following configuration is also possible.

Now, the cross-correlation function between the perceptually weighted audio signal x,(n) obtained for each subframe and the perceptually weighted impulse response of the synthesis filter, 4(n), is φ
, and so on. The method of obtaining this is detailed in the above-mentioned document 2, so the explanation will be omitted here.

Pitch prediction is performed between subframes using the sound source pulse memory 305 and the autocorrelation function Rhh (m) of the impulse response, and the cross-correlation function φxh is modified as shown in the following equation.

φ' xh (m) =φ,, (m) −Σg,L
-1・Rhh (l m m6t' l ) (
31 Here, gi'-' and m1L-' each indicate the amplitude 9 position of the i-th pulse of the multi-pulse train of the past (l-1) subframe stored in the sound source memory. Next, by predicting the pitch within the subframe using Φ xh(m) and Rt, h(m), the amplitude g11 position m of the large amplitude multi-pulse train in the current subframe is determined.

Find the specified number of . Regarding this method, reference can be made to the above-mentioned document 2 and the like. Next, the signal 'K(n) is -
Once reproduced, °K (n) is subtracted from the subframe audio signal to obtain the residual e, (n), and then the small amplitude multipulse train or the codebook index and gain are obtained using the same method as described above. seek.

In this way, the transmission information on the transmitting side of the present invention is the amplitude 1 position of the large amplitude multi-pulse train and the amplitude 1 position of the small amplitude multi-pulse train.
These are the position (or the index and gain of the codebook when a codebook is used), the coefficients of the pitch recovery filter, and the coefficients of the spectral envelope synthesis filter.

Next, referring to FIG. 1, the speech encoding/decoding system is comprised of a speech encoding device and a speech decoding device, both of which are connected via a suitable transmission path.

The audio encoding device includes a parameter calculation circuit that divides an input audio signal into predetermined frame sections, obtains and encodes a spectral parameter representing a spectral characteristic and a pitch parameter representing a pitch from the audio signal, and encodes the audio signal. a subframe dividing circuit that divides the signal into subframe sections including at least a plurality of the pitches;
a sound source pulse calculation circuit that performs pitch prediction over subframes based on past sound source signals, further predicts pitch within subframes, and obtains and encodes a sound source signal consisting of a multipulse train; It has a multiplexer circuit that combines and outputs the output of the sound source pulse calculation circuit.

The audio decoding device includes a demultiplexer circuit that manually separates and decodes a code representing a spectral parameter, a code representing a pitch parameter, and a code representing a sound source signal of an audio signal, and a demultiplexer circuit that manually separates and decodes a code representing a spectrum parameter of an audio signal, a code representing a pitch parameter, and a code representing a sound source signal, and a pitch recovery filter that predicts the pitch between subframes based on the sound source signal and further calculates a driving sound source signal by predicting the pitch within the subframe using the decoded sound source signal; and a synthesis filter circuit that synthesizes an audio signal that satisfactorily represents the audio signal using the obtained spectral parameters.

Audio encoding and decoding processing is performed as follows.

In FIG. 1 showing an embodiment of the present invention, an input terminal 50
A discrete audio signal x (n) from 0 is input. The spectral and pitch parameter calculation circuit 520 calculates the spectral parameter ai of the spectral envelope synthesis filter 610, which represents the spectral envelope of the audio signal in the divided frame sections (for example, 20 m5), using the well-known LPG analysis method. Further, the coefficient and pitch period T of the Pisochi reproduction filter 605 are determined by a well-known autocorrelation method. As a specific method, for example, the parameter in the above-mentioned document 3,
You can refer to the pitch calculation circuit.

A parameter quantizer 525 quantizes the determined parameters, coefficients, and pitch period. The quantization method may be scalar quantization as shown in Japanese Patent Application No. 59-272435 (Reference 6), or vector quantization as shown in Reference 4. The dequantizer 530 dequantizes and outputs the quantized result.

The subframe division circuit 515 divides the frame audio signal into subframes that are shorter than the frame or have the same length as the frame. Here, using the pitch period,
At least a plurality of pitch periods should be included, and if the pitch period is short, the subframe length should be about 10 m5,
If the pitch period is long, it is the same as the frame length.

A weighting circuit 540 weights the signal using the audio signal divided into subframes and the dequantized spectral parameter. For the weighting method, reference can be made to the weighting circuit 200 in Document 6.

The impulse response calculation circuit 550 calculates an impulse response of a perceptually weighted synthesis filter using the dequantized spectral parameter, pitch period, and coefficient. For a specific method, refer to the weighting circuit in Document 2.

The autocorrelation function calculation circuit 560 calculates the autocorrelation function R(m) of the impulse response and outputs it to the cross-correlation function correction circuit 575 and the sound source pulse calculation circuit 580. For the method of calculating the autocorrelation function, reference can be made to the autocorrelation function calculation circuit 180 in Document 2 and Document 6.

The cross-correlation function calculation circuit 570 calculates a cross-correlation function Φxh (
m) is calculated and output to the cross-correlation function correction circuit 575.

In the cross-correlation function correction circuit 575, the sound source pulse memory 5
81, input the past sound source signal and autocorrelation function R(m) by one pitch period, perform pitch prediction between subframes according to equation (3), correct the cross-correlation function, and obtain Φ'xh. Output (m).

In the sound source pulse calculation circuit 580, φ', h(m) and Rh
Using h(m), the amplitude and position of a specified number of large-amplitude multi-pulse trains are determined by intra-frame pitch prediction. Regarding the pulse train calculation method, reference can be made to the sound source pulse calculation circuit in Document 2.

A quantizer 585 quantizes the amplitude and position of the multi-pulse train and outputs a code. The specific method is described in the above document 6.
etc. can be referred to. This output is dequantized by a dequantizer 590 and passed through a pulse generator 60 and a pitch recovery filter 605 . By passing the signal through a spectral envelope synthesis filter 610, a synthesized speech signal "K(n)" based on the large-amplitude sound source pulse is obtained.

Here, the pitch reproduction filter 605. The spectral envelope synthesis filter 610 is similar to the pitch recovery filter 325 in FIG.
, operates in the same way as the spectral envelope synthesis filter 320.

The subtracter 615 subtracts the synthesized speech signal x (n) from the speech signal 'K(n) to obtain a residual signal ex
A small amplitude sound source signal is calculated for (n).

In the small-amplitude sound source calculation circuit 620, the small-amplitude sound source signal is expressed using a predetermined number of multi-pulse trains or a codebook, as described in FIG. 2. here,
A multi-pulse train will be used. By the method described in connection with FIG. 2, one amplitude position of a small amplitude multi-pulse train is calculated, quantized and encoded, and a code is output. The adder 625 decodes these to generate a small amplitude sound source pulse train and outputs it to the adder 625.
The small amplitude sound source pulse train which is the output of the small amplitude sound source calculation circuit 620 is added and output to the sound source pulse memory 581.

The multiplexer 630 has a code representing nine amplitude positions of the multi-pulse train which is the output of the quantizer 585, a code representing one amplitude position of the small amplitude multi-pulse train which is the output of the small amplitude excitation calculation circuit 620, and a parameter quantizer 525. The spectral parameters, pitch periods, and codes representing the coefficients are combined and output.

On the other hand, on the receiving side, the demultiplexer 710 separates and outputs a code representing one amplitude position of the large-amplitude multipulse train, a code representing the spectral parameter, pitch period, and coefficient, and a code representing the small-amplitude excitation signal.

The source pulse decoder 720 decodes one amplitude position of a large amplitude multi-pulse.

The spectrum and pitch parameter decoder 750 functions in the same way as the inverse quantizer 530 on the transmitting side, and decodes and outputs the spectrum parameter, pitch period T, and coefficients.

The small amplitude excitation decoder 745 decodes and outputs one amplitude position of the small amplitude multi-pulse train.

The pulse generator 725 generates a sound source signal based on the large-amplitude multi-pulse train.

The pitch recovery filter 730 inputs the output of the pulse generator 725, the pitch period T, and the coefficient S, and uses the output of the sound source pulse memory 735 to predict the pitch between subframes and reproduces the pitch of the sound source signal. Search adder 7
Output to 40.

An adder 740 combines the sound source signal with a small amplitude sound source decoder 745.
The driving sound source signal is obtained by adding the output signals of the driving sound source signal, and the synthesis filter circuit 760 is driven.

The sound source pulse memory 735 operates in the same way as the sound source pulse memory 581 on the transmission side, and stores the output of the adder 740.

The synthesis filter circuit 760 uses the driving sound source signal and the decoded spectrum parameters to obtain and output a synthesized speech waveform for each subframe.

The embodiment described above is an example of the method described as a method for reducing the amount of calculations described in relation to FIG. 2, but the basic method described in relation to FIG. 2 may also be used. .

However, such a configuration increases the amount of calculations.

As the small amplitude sound source signal, in addition to the small amplitude multi-pulse train, as described in connection with FIG. 2, a codebook may be used. In this case, instead of transmitting one amplitude position of the small amplitude multi-pulse train, the index and gain of the codebook are transmitted.

Furthermore, in order to further reduce the amount of calculation, the small amplitude sound source signal may also be obtained while performing intra-frame pitch prediction in the same way as the large amplitude multi-pulse. In this case, there is no need to regenerate the signal once to obtain the residual e2(n), so the pulse generator 600. Pitch recovery filter 605. Spectral envelope synthesis filter 610
．． The small amplitude sound source calculation circuit 620 becomes unnecessary. However, since it is thought that there is little pitch correlation for a small amplitude signal at -i, it is thought that such a configuration reduces the amount of calculation but degrades the characteristics.

Furthermore, as a method for calculating multi-pulses, various well-known methods can be used in addition to the method shown in the above-mentioned document 1. One example of this is the “A
5tudy on Pulse 5earch AI
gorithmsfor Multi-pulse 5
peech Coder Realization”
(IEEE JSAC, pp. 133-141.198
6) (Reference 7) can be referred to.

In addition, as a method for calculating the pitch period, in addition to the method shown in the above embodiment, for example, as shown in equation (4) below, the past sound source signal v (n), a pitch reproduction filter, and spectral envelope synthesis can be used. It is also possible to search for a position T that minimizes the error power E between the signal reproduced by the filter and the input audio signal x (n) of the current subframe, and find the coefficient S at that time.

E=Σ((x (n) −b −v (n−T) *h
(n) )*W (n) )”
(41 Here, h(n) is the impulse response of the spectral synthesis filter, and w (n) is the impulse response of the auditory weighted circuit.

Furthermore, a linear deviation τ may be allowed in the pitch period T of the subframe. For specific methods, see O
“2.4kbps pitch pre
diction multi-pulse 5peec
h coding” (proc, IEEE ICA
SSI”88.9.S4.9.1988) (Reference 8
) can be referenced. However, in this case, in addition to T, it is also necessary to transmit τ as pitch information.

〔Effect of the invention〕

According to the present invention, a sound source signal consisting of multi-pulses is obtained for each subframe based on sound source signals of past subframes using pitch correlation between subframes and within a subframe, and the sound source signal is efficiently and Because the calculations are performed in the correlation domain, and a pair of pitch recovery filters and spectral envelope synthesis filters are used, reproduction can be performed with a relatively small amount of calculation compared to conventional technology. This has the great effect of greatly improving the sound quality of the voice. Moreover, this effect is more pronounced when the bit rate is low.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of an embodiment of a speech encoding/decoding method, speech encoding device, and speech decoding device according to the present invention, and FIG. 2 is a method for interframe and intraframe pitch prediction according to the present invention. FIG. 3 is a block diagram showing a conventional example of a pulse train search method. 280, 290...Switches 305, 581.735...Sound source pulse memory 31
0, 580...Sound source pulse calculation (part) circuit 315, 620...Small amplitude sound source calculation (part) circuit 320, 610...Spectrum envelope synthesis filter 325, 605.730... Pitch reproduction filter 3
30 ..... Subtractor 350 ..... Auditory weighting is filter 355
, 515... Subframe division (part) circuit 520... Spectrum and pitch parameter calculation circuit 525... Parameter quantizer 530,
590...Inverse quantizer 540...Weighting circuit 550...
... Impulse response calculation circuit 560 ...
・・Autocorrelation function calculation circuit 600゜625゜760 ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ ・ Cross correlation function calculation circuit Cross correlation function correction circuit Excitation pulse calculation circuit Quantizer Pulse generator Adder Multiplexer Demultiplexer Excitation pulse decoder Small amplitude excitation decoding Figure 2: Pitch parameter decoder synthesis filter circuit

Claims (3)

[Claims] (1) On the transmitting side, a discrete audio signal is input and divided into predetermined frame sections, a spectral parameter representing a spectral characteristic and a pitch parameter representing a pitch are obtained from the audio signal, and the audio signal is divided into predetermined frame sections. Divide into subframe sections containing at least a plurality of pitches, and calculate a sound source signal consisting of multipulses in the subframe section by predicting the pitch between subframes and within the subframe based on past sound source signals, and on the receiving side, A driving sound source signal is obtained by pitch prediction between subframes and within a subframe using the pitch parameter and the sound source signal based on the past sound source signal, and the sound signal is well expressed using the spectral parameter. A speech encoding/decoding method characterized by outputting a synthesized speech signal. (2) a parameter calculation circuit that divides an input audio signal into predetermined frame sections, obtains and encodes a spectral parameter representing a spectral characteristic and a pitch parameter representing a pitch from the audio signal; A subframe dividing circuit divides the pitch into subframe sections including at least a plurality of pitches, predicts the pitch over the subframe based on past sound source signals, and further predicts the pitch within the subframe to generate the sound source signal consisting of a multi-pulse train. 1. A speech encoding device comprising: an excitation pulse calculation circuit that calculates and encodes an excitation pulse; and a multiplexer circuit that combines and outputs the output of the parameter calculation circuit and the output of the excitation pulse calculation circuit. (3) a demultiplexer circuit that inputs, separates, and decodes a code representing a spectrum parameter of an audio signal, a code representing a pitch parameter, and a code representing a sound source signal; and the decoded pitch parameter and past sound source signal. a pitch recovery filter that performs inter-subframe pitch prediction based on the decoded excitation signal and further calculates a driving excitation signal by predicting the pitch within the subframe using the decoded excitation signal; and the driving excitation signal and the decoded spectral parameter. and a synthesis filter circuit for synthesizing an audio signal that satisfactorily represents an audio signal using the following.