JP3063087B2 - Audio encoding / decoding device, audio encoding device, and audio decoding device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to an audio encoding / decoding system for encoding and decoding an audio signal efficiently at a low bit rate and an apparatus therefor.

(Prior Art) As a method of transmitting an audio signal at a low bit rate, for example, about 16 kb / s or less, a multi-pulse encoding method or the like is known. In these, the sound source signal is represented by a combination of a plurality of pulses (multi-pulse), the characteristics of the vocal tract are represented by a digital filter, and the information of the sound source pulse and the coefficient of the filter are represented by:
It is obtained and transmitted for each fixed time section (frame). For details of this method, see, for example, (Araseki, Ozawa, Ono
and Ochial, “Mulit-pulse Excited Speech Coder Bas
ed on Maximum Cross-correlation Search Algorith
m ", (GLOBECOM'83, IEEE Global Telecommunication, 23.
3,1983)). In this method, a good voice signal is output after decoding by separating and expressing the vocal tract information and the sound source signal, respectively, and using a combination (multi-pulse) of a plurality of pulse trains as means for expressing the sound source signal. . The basic concept of obtaining a multi-pulse train representing a sound source signal will be described with reference to FIG. An audio signal of a section divided for each frame is input from an input terminal 800 in the figure. The spectrum parameter from which the audio signal of the current frame is obtained is input to the synthesis filter circuit 820. A sound source calculation circuit 810 generates an initial sound source multi-pulse train,
By inputting to 820, a synthesized speech waveform is obtained as an output. The subtractor 840 subtracts the synthesized speech waveform from the input signal. The result is input to the weighting circuit 850, and the weighted error power in the current frame is obtained as an output. Then, the amplitude and position of the sound source multi-pulse train are obtained in the sound source calculation circuit 810 so as to minimize the weighted error power.

On the other hand, a pitch prediction multi-pulse method for improving the sound quality of the method of Reference 1 by performing pitch prediction using a pitch parameter representing a fine structure of pitch is described in Japanese Patent Application No.
Since it is described in the specification of Japanese Patent Application No. 58-139022 (Reference 2), the description is omitted here.

(Problems to be Solved by the Invention) However, in the conventional method of Document 1, the sound quality is good when the pit rate is sufficiently high and the number of sound source pulses is sufficient, but when the bit rate is reduced, the sound quality is reduced. Had declined. In particular, the conventional method has a drawback that in the case of an input signal having a high pitch frequency, for example, when a female voice is input, the reproduced voice is deteriorated. This means that when the pitch frequency is high, many pitch waveforms are included in the frame of the pulse calculation, and in order to reproduce this pitch waveform well, compared to a speaker with a low pitch frequency, , Because more multi-pulses are required. Therefore, for this reason, it has been difficult to greatly reduce the transmission bit rate without deteriorating the sound quality, that is, to significantly reduce the number of pulses in one frame.

On the other hand, in the conventional method of Document 2, although pitch prediction is performed using a pitch parameter based on a correlation for each bitch, a multi-pulse and pitch prediction are used regardless of a large-amplitude excitation signal or a small-amplitude excitation signal. The sound source signal. A large amplitude excitation signal is considered to have a high correlation for each pitch, while a small amplitude excitation signal is considered to have a low correlation.
In order to further improve the sound quality by this method, the role of the small-amplitude multi-pulse train or the small-amplitude sound source signal among the multi-pulse trains representing the sound source signal is more important.
This is especially important for consonant audio signals.
In the conventional method, as a multi-pulse train expressing a sound source signal, only a set number is obtained and transmitted in descending order of amplitude. Therefore, in the conventional example, a sufficient number of small-amplitude pulses cannot be obtained due to the preset upper limit of the information amount, the approximation of the sound source signal is not sufficient, and there is a limit in the quality of the reproduced sound. Could not be improved. This was particularly remarkable when the bit rate was low.

An object of the present invention is to provide an audio encoding / decoding method and an apparatus capable of reproducing sound better than before even with a relatively small amount of operation even at a high bit rate or at a reduced bit rate. Is to do.

(Means for Solving the Problems) A speech encoding / decoding system according to the present invention inputs a discrete speech signal and represents a pitch parameter representing a fine structure of a pitch of the speech signal and a spectrum of the speech signal. Spectral parameters are determined, the voice signal is classified into a predetermined type, and a multi-pulse sequence and a codebook obtained by pitch prediction of a sound source signal of the voice signal according to the type are obtained. Expressed using a pulse train and transmitted, restore the sound source signal using the codebook, the multi-pulse train and the pitch parameter or the multi-pulse train and the pitch parameter according to the type of the audio signal, and restore the spectral parameter. To output a synthesized speech signal that satisfactorily represents the speech signal.

A speech coding apparatus according to the present invention includes a pitch parameter calculation circuit that obtains and encodes a pitch parameter representing a fine structure of a pitch from an input discrete speech signal sequence, and a code that obtains a spectrum parameter representing a short-time spectrum characteristic. A spectral parameter calculation circuit for converting the audio signal into a predetermined type and outputting a determination code, and a sound source signal of the audio signal according to a section using the pitch parameter and the spectrum parameter. An excitation signal calculation circuit represented by a multi-pulse sequence determined by pitch prediction and a code book of one type selected from a plurality of code books, the pitch parameter, the spectrum parameter, the discrimination code, the multi-pulse sequence, and the code book. And a multiplexer circuit for combining and outputting.

A speech decoding apparatus according to the present invention comprises a code representing a sound source multi-pulse train representing a sound source signal, a code representing a pitch parameter representing a fine structure of a pitch of a speech signal, and a code representing a spectrum parameter representing a short-time spectrum characteristic of the speech signal. And a demultiplexer circuit that inputs and separates and decodes a code representing a sound source signal and a discrimination code, and restores a multi-pulse train among the decoded sound source signals and reproduces a pitch using the decoded pitch parameter. A pitch reproduction circuit for obtaining the generated sound source signal, a sound source signal restoration circuit for selecting one type from a plurality of types of codebooks according to the discrimination code and restoring a driving excitation signal using an output of the pitch reproduction circuit, A sound signal using the driving sound source signal or the reproduced sound source signal and the decoded spectrum parameter. And a spectral envelope filter circuit formed.

(Function) In the present invention, in the pitch prediction multi-pulse coding method of the above-mentioned document 2, in order to express a sound source signal more effectively than a conventional method with a small amount of transmission information, an audio signal for each frame is predetermined. Vowel part, consonant part or voiced part, unvoiced part, and as a sound source signal for representing the voice signal, a vowel or voiced part is a pitch prediction multipulse method, and a consonant or unvoiced part is a pitch prediction multipulse. It is characterized by the use of a code and a law.

The vowels and consonants can be determined using well-known methods, for example, parameters such as the power of the current frame, the difference in power from the previous frame, the change in spectrum from the previous frame, and the pitch. On the other hand, a parameter such as a pitch gain can be used to determine a voiced or unvoiced portion.

The operation of the present invention will be described with reference to FIG. The upper part of FIG. 2 is a block diagram of a circuit for obtaining a large-amplitude multi-Muhals for expressing a sound source signal by pitch prediction. An audio signal of a section divided for each frame is input from an input terminal 300 in the figure. From the code input terminal 305, a code indicating whether the current frame is a vowel section or a consonant section is input. The pitch parameter obtained from the audio signal of the current frame is input to the pitch reproduction filter 315. Spectral parameters obtained from the audio signal of the current frame are input to the spectral envelope filter circuit 320. A driving sound source signal is obtained by generating an initial sound source multi-pulse train in the multi-pulse sound source 310 and inputting it to the pitch reproduction filter 315. By inputting the driving sound source signal to the spectrum envelope filter circuit 320, a synthesized speech waveform is obtained as an output. The subtractor 340 subtracts the synthesized voice waveform from the input signal. The result is input to the weighting circuit 350, and the weighted error power in the current frame is obtained as an output. And to minimize this weighted error power,
In the multi-pulse sound source 310, the amplitude and position of the sound source multi-pulse train are obtained. The switch 380 is always connected to the upper side when obtaining a multi-pulse train. Switch 380
Is connected to the lower side when the code input from the code input terminal 305 indicates that the current frame is a consonant part after the multi-pulse train is obtained, and the obtained multi-pulse train, the pitch reproduction filter 315, and the spectral envelope The synthesized speech waveform (n) obtained by the filter circuit 320 is subtracted by a subtractor 2
Output to 20. When the code input from the code input terminal 305 indicates that the current frame is a vowel portion, the switch remains connected to the upper side, and a multi-pulse is output as a sound source signal.

The above is a case where a sound source is represented by a pitch-predicted multi-pulse sequence. When a pitch-predicted multi-pulse sequence and a codebook are used to represent a sound source signal, the following processing is performed in addition to the above processing.

The small-amplitude sound source calculation circuit 230 calculates a small-amplitude sound source signal representing the feature of the small-amplitude component of the sound source signal for each of the small sections (for example, about 5 msec.) Obtained by further dividing the frame section. Here, this small-amplitude sound source signal has almost random phase characteristics and is considered to be almost close to a noise signal. In order to encode such a signal very efficiently, a vector quantization technique of preparing and encoding a codebook (codebook) of a plurality of small amplitude excitation signals prepared in advance can be used. Regarding vector quantization, for example, RMGray, “Vector quantization for spe
ech coding and recognition "(J. Acoustical Soc. Amer
ica, vol.80, Suppl.1, Q1, 1986), so the description is omitted here.

Hereinafter, the operation of the small-amplitude sound source signal circuit 230 will be described.
The reproduced signal (n) is converted by the subtracter 220 into the original audio waveform x (n).
The time division circuit 240 divides the residual signal e (n) resulting from the subtraction into the small sections shorter than the frame uniformly in time. A plurality of codebooks (codebooks) 250 are prepared and prepared in advance. For each of the small sections divided by the time division circuit 240, one of the codebooks is input and the gain is set through the gain circuit 255. After matching, a spectrum envelope filter 260 using the same spectral parameters as the spectrum envelope filter 320
To synthesize the combined residual signal (n).

The subtracter 270 subtracts the combined residual signal (n) from the input residual signal e (n). This result is input to the weighting circuit 280, and a weighted error power is obtained as an output. Here, the weighting circuit 280 performs the same operation as the weighting circuit 350 in the upper part of FIG. Then, an optimal one is selected from the codebook 250 so as to minimize the weighted error power, and the index is output.

Next, an actual method for expressing a small-amplitude sound source signal using a codebook and selecting the codebook will be described below using equations. As a codebook selection method, calculation is performed so as to minimize the error power E defined by the following equation.

Here, e (n) is the input error signal in FIG.
Is a gain to be multiplied by the gain circuit 255, and (n) is a residual signal reproduced by one selected codebook and a spectrum envelope filter. w (n) is a weighting filter (the weighting circuit 280 in FIG.
Is the same as 3) shows the impulse response. When the expression (1) is minimized with respect to g, the expression (2) is obtained.

Here, _W (n) = (n) * w (n) = n (n) * h
(N) * w (n) (3a) e _W (n) = e (n) * w (n) (3b) The symbol * represents convolution. The denominator of equation (2) is
Autocorrelation of _W (n) (strictly covariance), then the molecules are cross-correlation _W (n) and e _W (n). Further, n (n) in the expression (3a) is a signal represented by one selected type of code in the code book.

Also, h (n) indicates the impulse response of the spectrum envelope filter circuit 260.

At this time, the error power E can be written as the following equation. The codebook that minimizes E may be selected so as to maximize the second term of Expression (4), that is, maximize g.

As a method for further greatly reducing the amount of calculation for selecting a codebook, the following configuration can be considered. The multipulse train representing the sound source signal is searched for using the cross-correlation. This Determination detailed in the Documents 1 and 2 is omitted described here, by using a cross-correlation function [Phi _xh 'after determined pitch prediction multi-pulse train, a significant amount of calculation than the method described above After the reduction, the codebook can be selected.

In the method described below, the signal _W
Since (n) does not need to be reproduced, the amount of calculation can be greatly reduced while maintaining the characteristics almost the same as in the above-described method. The derivation method will be described below. First, Φ _xh ', _W (n) can be written as follows.

Φ _xh ′ = Σe _W (n) h _W (n) (5) _W (n) = n (n) * h _w (n) (6) Substituting equation (6) into equation (2), Using the expression, the following modifications are possible.

Here, Φ _xh ′ is a cross-correlation function after _{obtaining a} multi-pulse train by pitch prediction, and R _hh (0) is the power of the impulse response of a filter composed of a cascade connection of the spectral envelope filter 260 and the weighting circuit 280. R _nn (0) is the power of the signal n (n) represented by the code when one type of code is selected from the code book 250. The numerator of equation (7) is the cross-correlation function between Φ _xh 'and the signal n (n) represented by the selected code. As in the case of the above equation (2), it is preferable to select a codebook that maximizes g in the equation (7).

Note that the codebook 250 may be created by training in advance using residual signals of a sound source after obtaining a predetermined number of large-amplitude pitch prediction multi-pulse trains, or may be created by, for example, Gaussian characteristics. A noise signal having a statistical property may be created by using a plurality of noise signals with variously changing phase characteristics. MR for the latter method
Shroeder and BSAtal: “Code-Excited linear predi
ction (CELP): high-quality speech at very low bit
rates, ", Proc, ICASSPvol.1, paper no.25.1.1, Ma
rch, 1985.

The transmission information on the transmitting side is the amplitude and position of the large-pulse multipulse whose pitch is predicted, the index and gain of the cobook of the small-amplitude excitation signal, the pitch parameter, and the spectrum parameter.

(Embodiment) In FIGS. 1A and 1B showing an embodiment of the present invention, a discrete audio signal x (n) is input from an input terminal 500. FIG. The time division circuit 510 divides the input audio signal into frames that are temporally uniform (for example, every 20 msec.).
The pitch parameter calculation circuit 515 calculates a pitch parameter representing the fine structure of the pitch. As a calculation method, a method as shown in the aforementioned reference 2 is used. The quantizer 516 quantizes the obtained pitch parameter. The inverse quantizer 518 performs inverse quantization using the result of quantization and outputs the result. The spectrum parameter calculation circuit 520 obtains a spectrum parameter representing the spectrum of the audio signal in the divided section by a well-known LPC analysis method.

The obtained spectral parameters are quantized by a spectral parameter quantizer 525. As a quantization method, scalar quantization as described in Japanese Patent Application No. 59-272435 (Reference 5) or vector quantization as described in Reference 4 may be performed. The inverse quantizer 530 inversely quantizes using the result of the quantization and outputs the result. The weighting circuit 540 weights the divided audio signal using the dequantized spectral parameters. The weighting method is described in the weighting circuit of the above-mentioned reference 5.
200 can be referenced. Impulse response calculation circuit 5
50 calculates the impulse response using the dequantized pitch parameter and the dequantized spectral parameter. The specific method can be referred to the above-mentioned document 2. The autocorrelation function circuit 560 calculates the autocorrelation function of the impulse response and outputs the calculated function to the sound source pulse calculation circuit 580. The calculation method of the autocorrelation function can be referred to the autocorrelation function calculation circuit 180 of the aforementioned document 2. The cross-correlation function calculation circuit 570 calculates a cross-correlation function between the weighted signal and the impulse response, and outputs the result to the sound source pulse calculation circuit 580.
The specific method can be referred to the above-mentioned document 2.

The determination circuit 575 determines whether the current frame is, for example, a vowel section or a consonant section, and outputs a determination code indicating the result to the excitation pulse calculation circuit 580. For the discrimination, as shown in the section of the operation, for example, well-known parameters such as spectrum change, power, and power change can be used.

When the output of the discrimination circuit 575 is a code indicating a vowel, the sound source pulse calculation circuit 580 obtains a predetermined number (L1) of multi-pulses by pitch prediction. For the calculation method of the multi-pulse train, reference can be made to the sound source pulse calculation circuit 210 of Reference 2.

If the output of the discrimination circuit 575 is a code indicating a vowel, the small amplitude excitation signal is not calculated, and the calculation for the excitation signal ends here. So in this case the quantizer
585, pitch reproducer 600, pitch reproduction filter 605, spectrum envelope filter 610, subtractor 615, and small amplitude sound source calculation circuit 620 do not operate.

The quantizer 585 quantizes the source multipulse train and outputs a code. This output is inversely quantized by an inverse quantizer 590, and a multi-pulse is reproduced by a pulse generator 600. The pitch reproduction filter 605 receives the reproduced multi-pulse and the pitch parameter dequantized by the dequantizer 518 as inputs and outputs a sound source signal whose pitch has been reproduced. By passing the sound source signal and the spectrum parameter output from the inverse quantizer 530 through a spectrum envelope filter 610, a synthesized speech signal (n) based on the sound source pulse is obtained.

When the output of the discriminating circuit 575 indicates a consonant, a multi-pulse train whose pitch is predicted by L2 (L2 <L1) is obtained by the above-described configuration to obtain a composite signal (n).

Further, the subtracter 615 subtracts the synthesized audio signal (n) from the audio signal x (n) to generate a residual signal e.
(N) is output to the small amplitude sound source calculation circuit.

As described in the operation section, the small-amplitude sound source calculation circuit 620 converts a small-amplitude sound source signal of a small section (for example, 5 msec.) Divided into sections shorter than a frame from a plurality of codebooks. Express using the optimal one.

A code indicating whether the current frame is a vowel section or a consonant section, an index and a gain of a codebook representing a small amplitude excitation signal, a code obtained by quantizing a multi-pulse train output from the quantizer 585, a quantizer 516 The code obtained by quantizing the pitch parameter, which is the output of, and the code obtained by quantizing the spectrum parameter, which is the output of the quantizer 525, are input to the multiplexer 630. However, when the current frame is a vowel section, the index and the gain of the codebook representing the small amplitude excitation signal are not input.
The multiplexer 630 combines and outputs the above codes.

On the other hand, on the receiving side, the demultiplexer 710 includes a code of the multi-pulse train, a code of the pitch parameter, a code of the spectrum parameter, a code of the spectrum parameter, a discrimination code indicating whether the current frame is a vowel section or a consonant section, If the frame is a consonant section, the index representing the small amplitude excitation signal and the sign of the gain are separated and output. The excitation pulse decoder 720 decodes the amplitude and position of the multi-pulse. The spectrum parameter decoder 750 performs the same function as the inverse quantizer 530 on the transmission side. The small-amplitude excitation decoder 730 has the same codebook as the small-amplitude excitation calculation circuit 620 on the transmitting side, and when receiving a code indicating that the current frame is a consonant section, the received index is To select and output a code representing a small amplitude excitation signal. When a code indicating that the current frame is a consonant section is received, the gain circuit 735 determines the amplitude of the small-amplitude excitation signal using the received code of the gain. The pitch parameter decoder 745 performs the same function as the inverse quantizer 518 on the transmission side. The pulse generator 725 generates a sound source signal based on the multi-pulse train. A pitch reproduction filter 755 reproduces a synthesized excitation signal whose pitch has been reproduced by using the obtained excitation signal and the decoded pitch parameter as inputs. The adder 740 adds the sound source signal obtained by regenerating the pitch and the output signal of the gain circuit 735 when receiving a code indicating that the current frame is a consonant section, obtains a driving sound source signal, and obtains a spectrum envelope filter. Drive circuit 760. The spectrum envelope filter circuit 760 obtains and outputs a synthesized speech waveform using the driving excitation signal and the decoded spectrum parameters.

The configuration described above is only one configuration of the present invention, and various modifications are possible.

In order to further reduce the amount of calculation for obtaining the small-amplitude sound source signal, as described in Equations (5) to (7) in the operation section, after calculating the large-amplitude multipulse by pitch prediction, A code book can be selected using the cross-correlation function Φ _xh ′. In this case, as described in the above-mentioned operation, the signal _W (n) does not need to be reproduced at the time of codebook selection.
The amount of calculation can be greatly reduced as compared with the configuration shown in FIG.

For the consonant part, different codebooks may be created in advance in accordance with the properties of the consonant (for example, burstiness, friction, etc.), and these may be switched and used.

Further, as a method of calculating a multi-pulse,
Various known methods can be used in addition to the method shown in FIG.

Further, other well-known parameters (a line spectrum pair, a cepstrum, a mel cepstrum, a logarithmic cross-sectional area ratio, etc.) can also be used as the spectrum parameter. Further, as a method of quantizing the spectrum parameter, vector quantization or the like can be used in addition to scalar quantization. Reference 3 can be referred to for the vector quantization.

Although the frame length is fixed, the frame length may be variable. (Effect of the Invention) According to the present invention, the sound source signal is set to a predetermined type (for example, a vowel portion, a consonant portion, or a voiced portion) as compared with the conventional example. , Unvoiced part), and according to the classification, the vowel or voiced part has a relatively small number of sound sources of pitch prediction multi-pulse trains. By using a certain small-amplitude sound source signal together with a codebook, it can be represented by a very small amount of transmission information. Therefore, even if the bit rate is the same as that of the conventional method, there is a great effect that a better reproduced audio signal can be obtained not only in the vowel part but also in the consonant section as compared with the conventional method. Further, this effect becomes more remarkable when the bit rate is reduced.

[Brief description of the drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of a speech encoding / decoding method and apparatus according to the present invention, and FIG. 2 is a block diagram showing the operation of the present invention. FIG. 3 is a block diagram showing a conventional example of the multi-pulse train search method. In the figure, 510,240: time division circuit, 515: pitch parameter calculation circuit, 520: spectrum parameter calculation circuit, 516,5
25,585 …… Quantizer, 518,530,590 …… Dequantizer, 54
0,350,280,850 ... weighting circuit, 550 ... impulse response calculation circuit, 560 ... autocorrelation function calculation circuit, 570 ... cross-correlation function calculation circuit, 575 ... discrimination circuit, 580 ... sound source pulse calculation circuit, 600,725 ... pulse Generator, 755,315,605
…… Pitch reproduction filter, 610,760,320,260 …… Spectrum envelope filter, 820 …… Synthesis filter circuit, 620,230
…… Small amplitude sound source calculation circuit, 630 …… Multiplexer, 710
…… Demultiplexer, 720 …… Sound source pulse decoder, 730
…… Small amplitude excitation decoder, 740 …… Adder, 745 …… Pitch parameter decoder, 750 …… Spectrum parameter decoder, 310 …… Multi pulse excitation, 810 …… Excitation calculation circuit,
615,340,220,270,840 ... Subtractor, 735,255 ... Gain circuit, 380 ... Switch.

Claims (3)

(57) [Claims] 1. A speech encoding / decoding device comprising a speech encoding device and a speech decoding device, wherein the speech encoding device comprises a pitch parameter calculating means (51).
5) a spectrum parameter calculating means (520), a discriminating means (575), a sound source calculating means, and a multiplexer (630). The pitch parameter calculating means (515) calculates a pitch parameter representing a pitch parameter from an input discrete voice signal. The spectrum parameter calculation means (520) determines and encodes a spectrum parameter representing a short-time spectrum from the input discrete speech signal, and the discrimination means (575) converts the input discrete speech signal into a predetermined type. When the discriminating means (575) indicates a predetermined classification, the sound source calculating means performs the first processing. When the discriminating means (575) indicates any other classification, the sound source calculating means performs the first and second processing. Performing a first process, wherein a first synthesized signal is reproduced based on a pitch parameter, a spectrum parameter, and a multi-pulse sequence, and And a multi-pulse train at which the signal becomes optimal is coded as multi-pulse information. The second processing is to obtain a difference signal between the determined first combined signal and the input discrete signal, and to obtain a pitch parameter, a spectrum parameter, and a code. The second combined signal is reproduced based on the book and the gain, and the code vector and the gain of the codebook when the second combined signal is optimal with the difference signal are determined and encoded. The output of the parameter calculating means (515), the spectrum parameter calculating means (520), the discriminating means (575) and the sound source calculating means are combined and output. (730), pulse generation means (720, 72
5), a pitch reproducing means (755), an adding means (740), and a spectrum envelope synthesis filter (760). The demultiplexer (710) converts a gain, a discrimination code, an index, multi-pulse train information, a spectrum parameter, When the discrimination code is a predetermined one, the small-amplitude excitation signal restoring means (730) adds a gain to the code vector indicated by the index and outputs the code vector. The pulse generating means (720, 725) ) Generates a multi-pulse train based on the multi-pulse train information, the pitch reproducing means (755) outputs a synthesized sound source signal whose pitch is reproduced based on the multi-pulse train and the pitch parameter, and the adding means (740) Small amplitude excitation signal decoding means (730),
The output of the pitch reproduction means (755) is added to output a sound source signal, and the spectrum envelope synthesis filter (760) synthesizes a sound signal based on the sound source signal output from the addition means and the spectrum parameter. apparatus. 2. A pitch parameter calculating means (515), a spectrum parameter calculating means (520), and a discriminating means (57).
5) A speech coding apparatus comprising a sound source calculation means and a multiplexer (630), wherein the pitch parameter calculation means (515) obtains and encodes a pitch parameter representing a pitch parameter from an input discrete speech signal, and encodes the spectrum parameter. The calculating means (520) obtains and encodes a spectrum parameter representing a short-time spectrum from the input discrete audio signal, and the determining means (575) classifies the input discrete audio signal into a predetermined type, The means performs first processing when the determination means (575) indicates a predetermined classification, and performs first and second processing when the determination means (575) indicates another classification. Reproduces a first synthesized signal based on a pitch parameter, a spectrum parameter, and a multi-pulse train, and generates a multi-pulse when the first synthesized signal is optimal with an input discrete voice signal. The second processing determines a difference signal between the determined first combined signal and the input discrete signal, and determines the difference signal based on the pitch parameter, the spectrum parameter, the codebook, and the gain. Is reproduced, and the code vector and the gain of the code book when the second synthesized signal is optimal with the difference signal are determined and coded. The multiplexer (630) includes a pitch parameter calculating means (515), An audio encoding device that combines and outputs the outputs of the parameter calculation means (520), the determination means (575), and the sound source calculation means. 3. A demultiplexer (710), a small-amplitude sound source signal restoring means (730), a pulse generating means (720, 725), a pitch reproducing means (755), an adding means (740), and a spectrum envelope synthesis filter (760). A voice decoding device, wherein a demultiplexer (710) separates a gain, a discrimination code, an index, multipulse train information, a spectrum parameter, and a pitch parameter from an input signal and outputs the separated signal; When the discrimination code is a predetermined one, the gain is added to the code vector indicated by the index and output. The pulse generating means (720, 725) generates a multi-pulse train based on the multi-pulse train information, (755) outputs a synthesized sound source signal whose pitch is reproduced based on the multi-pulse train and the pitch parameter, and the adding means (740) , Small amplitude excitation signal decoding means (730),
A speech decoding device that adds the output of the pitch reproducing means (755) to output a sound source signal, and the spectrum envelope synthesis filter (760) synthesizes a sound signal based on the sound source signal output by the adding means and the spectrum parameter.