MXPA96003416A - Ha coding method - Google Patents

Abstract

The present invention relates to a voice coding method, comprising the steps of: generating short-term prediction coefficient, based on an input speech signal, providing first and second codebooks formed of classified parameters representing the short-term prediction coefficients, the first and second codebooks relating to at least a plurality of characteristic parameters of the input speech signal, selecting one of the first and second codebooks, based on a density value of the input speech signal, and quantify the short-term prediction coefficients, using the codebook select

Description

"SPEECH CODING METHOD" SPECIFICATION TECHNICAL FIELD 5 This invention relates to a speech coding method for encoding short-term prediction residuals or parameters that represent short-term prediction coefficients of the input speech signal Y)., By means of vector or matrix quantification.

ANTECEDENTS OF THE TECHNIQUE There are a variety of coding methods known to encode the audio signal, including the speech signal and the acoustic signal, exploiting the statistical properties of the audio signal in the domain . of time and in the frequency domain, and the psychoacoustic characteristics of the human auditory system.

These coding methods can be classified more or less into coding in the time domain, coding in the frequency domain and analysis / synthesis coding. Whether in multi-band excitation (MBE), 25 single-band excitation (SBE), harmonic excitation, sub-band coding (SBC), linear predictive coding (LPC), discrete cosine transformation (DCT), modified DCT (MDCT) or fast Fourier transformation (FFT), as examples of high efficiency coding for 5 speech signals, various information data such as spectral amplitudes or parameters thereof, such as LSP parameters, alpha-parameters or k -parameters are quantified, having usually adopted a scalar quantification. * .. If with this scalar quantization, the bit rate is decreased v.gr., from 3 to 4 kbps to further increase the efficiency of quantization, quantization noise or distortion are increased, thus elevating the difficulties in practical use. In this way, it is currently practiced to group different data provided for coding, such as the time domain data, the frequency domain data or the filter coefficient data to a vector or to group these vectors through the multiple frames 0 towards a matrix, and to effect the vector or matrix quantization instead of individually quantifying the different data. For example, in linear code excitation prediction coding (CELP) the LPC residues are directly quantified by vector quantization or matrix as the time domain waveform. In addition, the spectral envelope in the MBE coding is quantified in a similar manner by vector or matrix quantization. 5 If the bit rate is reduced further, it is not feasible to use enough bits to quantify the parameters specifying the envelope of the spectrum itself or the LPC residues, thereby deteriorating the quality of the signal. ! & - In view of the foregoing, an object of the present invention is to provide a speech coding method capable of providing satisfactory quantization characteristics even with a smaller number of bits. 15 EXHIBITION OF THE INVENTION With the speech coding method according to the present invention, a first code book and a second codebook classifying the parameters representing the short-term prediction values related to a reference parameter consisting of one or a combination of a plurality of parameters characteristic of the input speech signal.

The short-term prediction values are generated based on the input speech signal. One of the first and second code books related to the reference parameter of the input speech signal is selected, and the short-term production values are quantized by reference to the codebook selected to encode the input speech signal. The short-term prediction values are short-term prediction coefficients or short-term prediction errors. Characteristic parameters include the density values of the speech signal, the intensity of density, the frame power, the speech discrimination flag / without voice and the gradient of the signal spectrum. The quantification is the quantification of the vector or the quantification of the matrix. The reference parameter is the value of the density of the speech signal. One of the first and second code books is selected depending on the magnitude relationship between the value of the density of the input speech signal and the pre-graduated density value. In accordance with the present invention, the short-term prediction value, generated based on the input speech signal, is quantified by reference to the selected codebook to improve the quantization efficiency.

BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 is a schematic functional diagram showing a speech coding device (encoder) as an illustrative example of a device for carrying out the speech coding method in accordance with the present invention. Figure 2 is a circuit diagram for illustrating a modifier circuit that can be used for f a density detection circuit shown in Figure 1. Figure 3 is a functional diagram to illustrate the method for forming a codebook (method of training) used for quantification of 15 vector.

BEST WAY TO CARRY OUT THE INVENTION r The preferred embodiments of the present invention will be explained below. Figure 1 is a schematic functional diagram showing the constitution for carrying out the speech coding method in accordance with the present invention.

In the present speech signal encoder, the speech signals supplied to an input terminal 11 are supplied to a linear prediction coding analysis (LPC) circuit 12, a reverse filtering circuit 5 and a calculation circuit 23 of perceptual weighting filter. The LPC analysis circuit 12 applies a Hamming window to an input waveform signal, with a length within the order of 256 samples of the signal of * "" input waveform as a block, and calculate the linear prediction coefficients or the alpha-parameters by a self-correlation method. The frame period, as the data output unit, comprises, eg, 160 samples. If the frequency of sampling fs is v.gr., of 8 kHz, the frame period is equal to 20 milliseconds. The alpha-parameters of the LPC analysis circuit 12 are supplied to an LSP conversion circuit 13 for conversion into line spectral pair (LSP) parameters. That is, the alpha-parameters, which are 0 as direct type filter coefficients, become v.gr., ten, that is, five pairs of LSP parameters. This conversion is carried out using e.g., the Newton-Raphson method. The reason why the alpha-parameters are converted into LSP parameters is that the LSP parameters are higher than the alpha-parameters in interpolation characteristics. The LSP parameters of an LSP conversion circuit 13 are quantized by vector by an LSP vector quantizer 14. During this time, the difference between tables can be found first before carrying out the quantification of the vector. Alternatively, multiple LSP parameters for multiple frames are grouped together to perform matrix quantization. For this quantification, 20 milliseconds correspond to a table, and the LSP parameters calculated every 20 milliseconds are quantified by vector quantization. To carry out vector quantization or matrix quantization, a male voice code book 15M or a female voice codebook 15F is used by switching between them with a change switch 18, in accordance with the density. The quantization output of the LSP vector quantizer 14, which is the index of the LSP vector quantization, is provided, and the quantized LSP vectors are processed to an alpha-conversion circuit 17 for the conversion of the LSP parameters to the alpha -parameters as coefficients of the direct type filter. Based on the LSP output to the alpha-conversion circuit 17, the filter coefficients of a perceptual weight synthesis filter 31 for linear code excitation coding (CELP) coding are calculated. An output of a so-called dynamic codebook 5 (density codebook also called an adaptive codebook) 32 for linear code excitation prediction coding (CELP) is supplied to an adder 34 through a multiplier 33 of coefficient designated to multiply a gain g0. On the other? ^ - part, an output of a so-called stochastic code book (noise code, also called a probabilistic book code) is applied to the adder 34 through a coefficient multiplier 36 designed to multiply a gain g ^. The output of the sum of adder 34 is applies as an excitation signal to the perceptual weighting synthesis filter 31. In the dynamic code book 32 the past excitation signals are stored. These excitation signals are read in a period of density and are multiply by the gain g0. The signal of the resulting product is added by adding 34 to a signal from the stochastic codebook multiplied by the gain q. The resulting sum signal is used to excite the perceptual weighting synthesis filter 31. In addition, the sum output from the adder 34 is fed back to the dynamic codebook 32 to form an IIR filter class. The stochastic codebook 35 is configured so that the changeover switch 35S, switches between the codebook 35M for male voice 5 and the codebook 35F for female voice to select one of the codebooks. The coefficient multipliers 33, 36 have their respective gains g0, gi, controlled in response to the book outputs of the gain code 37. An output of the filter f ^. 31 of perceptual weighting synthesis that supplies a subtraction signal to an adder 38. An output signal of the summing machine 38 is supplied to a circuit 39 for minimizing waveform distortion (Euclid distance). Based on an output of the circuit 39 By minimizing the waveform distortion, the reading of the signal from the respective codebooks 32, 35 and 37 is controlled to minimize an output of the adder 38, i.e. the shape distortion of weighted wave. In a reverse filtering circuit 21, the input speech signal from the input terminal 11 is counter-filtered by the alpha-parameter from the LPC analysis circuit 12 and is supplied to a density sensing circuit 22 for detection of the density. He change switch 16 or change switch 35S is changed in response to density detection results from density detection circuit 22 for selective switching between the code book for the male voice and the codebook for the female voice In the perceptual weighting filter calculation circuit 23, the perceptual weighting filter calculation is carried out on the input speech signal from the input terminal 11, using an output of the LPC analysis circuit 12. The resulting perceptual weighted signal is supplied to an adder 24 which is also fed with an output of a zero input response circuit 25 as a subtraction signal. The zero input response circuit 25 synthesizes the response of the previous frame by a weighted synthesis filter and sends a synthesized signal. This synthesized signal is subtracted from the perceptual weighted signal by canceling the filter response of the previous frame "remaining" in the perceptual weighting filter 31 to produce a signal required as a new input for a decoder. An output of the adder 24 is supplied to the adder 38, where an output of the perceptual weight synthesis filter 31 is subtracted from the addition output. In the above-described encoder, assuming that an input signal from the input terminal 11 is x (n), the LPC coefficients, ie, the alpha-parameters are alphai and the prediction residuals are res (n). With the number of orders for analysis of P, 1 < i < P. The input signal x (n) is counter-filtered by the reverse filtering filter 21 in accordance with equation (1): P H (z) = 1 + alfaiz-1 i = l ... (1) to find the prediction residues (n) within the scale, eg, of 0 <n <; N-1, where N represents the number of samples corresponding to the length of the frame as a coding unit. For example, N = 160. Then, in the density detection circuit 22, the prediction residue res (n) obtained from the reverse filtration circuit 21 is passed through the low pass filter (LPF) to derive resl (n). Said LPF usually has a cutoff frequency within the order of 1 kHz in the case of the sampling clock frequency fs of 8 kHz. Then, the fresl (n) resln auto-correlation function is calculated in accordance with equation (2): N-i-1 fres] _ (i) =] _j resl (n) resl (n + i) N = 0 (2) where Lmin < i < Lmax Usually, Lmj_n is equal to 20 and Lmax equals 147, approximately. The density is found following the number i that gives a maximum value of the self-correlation function number i. which provides a maximum value by appropriate processing, is used as the density for the current frame. For example, assuming density more specifically, the density delay of the square k'th is P (k). On the other hand, density reliability or density resistance is defined by equation (3): PMk) = Fresl < P (k)) / fresl (0) ... (3) That is, the intensity of the auto-correlation, normalized by fresl (°) 'is defined as above. In addition, with the usual code excitation linear prediction coding (CELP), the frame power Ro (k) is calculated by equation (4): N-l i = 0: 4) where k represents the number of the box. Depending on the values of the density delay P (k), the density intensity Pl (k) and the power of the Rg (k) box, the quantization table for . { alfaj_} or the quantization box formed by converting the alpha-parameters into line spectral pairs (LSPs) are changed between the code book for the male voice and the liter of code for the female voice. In the embodiment of Figure 1, the quantization table for the vector quantizer 14 used to quantize the LSPs changes between the codebook for the male 15M voice and the codebook for the female 15F voice. For example, if P ^ n represents the threshold value of the density delay P (k) used to make the distinction between the male voice and the female voice and Plth and ^ oth represent respective threshold values of the density intensity Pl ( k) to discriminate the density reliability and power of the RQ frame (1). (i) a first codebook, e.g., the codebook for the male 15M voice, is used for P (k) Pth 'pl (k) > plth and R0 < k > Roth'- (ii) a second code book, eg, the book of code for the 15F fememine voice, is used for P (k) < Pth 'pl (k) > plth and Ro < k > > Roth'- and (iii) a third code book is otherwise used. Even though a different codebook from the 35M codebook can be used for the male voice and , < "" "'the 35F codebook for the female voice as the third codebook, it is also possible to use the 35M codebook for male voice or the 35F codebook for female voice as the third codebook. of threshold previously mentioned can be exemplified, eg, by Pth = 45, plth ~ ° -7 and Rg (k) = (full scale - 40 dB). Alternatively, the code books can be changed by keeping the n past frames of the density delays P (k), finding an average value of P (k) a through these boxes n and discriminating the average value with the pre-graduated threshold value Pth- It will be noted that these tables n are selected so that Pl (k) > Plth 'and Rg (k) > RQth 'that is, so that the tables are pictures with voice and exhibit high density reliability.

Still alternatively, the density delay P (k) satisfying the aforementioned condition can be supplied to the rectifier circuit shown in Figure 2 and the output of the resulting rectifier circuit 5 can be discriminated by the threshold value Pth to change the reference books. code. It will be noted that an output of the rectifier circuit of Figure 2 is obtained by multiplying the input data with 0.2 by a multiplier 41 and adding the signal of? '"resulting product by an adder 44 to an output data delayed by a frame by the delay circuit 42 and multiplying by 0.8 by a multiplier 43. The output state of the regulator circuit is maintained unless the delay is supplied P (k), density, the input data. In combination with the change described above, the code books can also be changed depending on the speech / voiceless discrimination, the density intensity value Pl (k) or the value of the power of the frame Ro (k). In this way, the average value of the density is extracted from the stable density section and the discrimination is made whether the input voice is the male voice or the female voice to change between the book code for the male voice and the code book for the female voice. The reason is that, since there is a deviation in the frequency distribution of the vowel's format between the male voice and the female voice, the space occupied by the vectors to be quantified decreases, that is, the variety decreases. of vector, changing between the masculine voice and the feminine voice especially in the portion of the vowel, allowing this way a satisfactory training that is to say, to learn to reduce the error of quantification. It is also possible to change the stochastic codebook in the CELP coding in accordance with the aforementioned conditions In the embodiment of Figure 1, the switch 35S is changed in accordance with the conditions previously mentioned for select a 35M codebook for the male voice and a 35F codebook for the female voice as the codebook 35 stochastic. To learn the code book, the training data can be classified under the same rule as that for coding / decoding so that the training data is brought to the low optimum, eg, the so-called LBG method. That is, referring to Figure 3, the signals of the training set 51 consisting of 5 speech signals for continuous training, eg, several minutes, are supplied to a line 52 line spectral calculation circuit (LSP). ) and a density discrimination circuit 53. The LRP calculation circuit 52 is equivalent, e.g., to the LPC and alpha analysis circuit 12 to the LSP conversion circuit 13 of Figure 1, while the density discrimination circuit 53 is equivalent to the circuit 21 of counterfiltration and the density detection circuit 22 of Figure 1. The density discrimination circuit 53 discriminates the density delay P (k), the density intensity Pl (k) and the power of the square Ro (k) by means of the previously mentioned Pth 'plth and Roth threshold values for classification according to the aforementioned conditions (i), (ii) and (iii). Specifically, the discrimination in which at least the male voice under condition (i) and the voice ends under condition (ii) is sufficient. Alternatively, the density delay values P (k) of the past n voice frames with high density reliability can be retained, and a mean value of the values of P (k) of these n frames can be found and discriminated by the value Threshold Pt. An output of the rectifier circuit of Figure 2 can also be discriminated by the threshold value Pt.

The LSP data from the LSP calculation circuit 52 is sent to a training data classification circuit 54 where the LSP data is classified into training data for the male voice 5 and training data for the voice. 56 female, depending on the discrimination output of the density discrimination circuit 53. This training data is provided to the training processors 57, 58 where the training is carried out from * 0- compliance with eg, called the LBG method to formulate the 35M code book for the male voice and the 35F code book for the female voice. The LBG method is a method for training the codebook proposed in the article by Linde, Y., Diver, A. and Gray, R.M., "An 5 Algorithm for vector Quantizer Design", in IEEE Trans. Comm., COM-28, pages 84 to 95, January 1980. Specifically, it is a technique to design an optimal vector quantifier locally for a source of information, whose probabilistic density function has not been known, or with the help from a training string call. The code book 15M for the male voice and the codebook 15F for the female voice, formulated in this way, are selected by changing the change switch 16 at the time of quantization of the vector by the voter quantizer 14 shown in FIG. Figure 1.

This changeover switch 16 is controlled for the change depending on the results of the discrimination by the density detection circuit 22. The index information, such as the quantization output of the vector quantizer 14, ie, the codes of the representative vectors are sent as data to be transmitted, while the quantized LSP data of the output vector is converted by LSP into a conversion circuit 17 in alpha-parameters that are fed to a perceptual weighting synthesis filter 31. This perceptual weighting synthesis filter 31 has the characteristics 1 / A (z) as shown by equation (5): w: A (z; 1+ Z ^ alphaiZ-1 i = l ... (5) where W (z) represents the perceptual weighting characteristics Among the data to be transmitted in the CELP coding described above, there is the index information for the dynamic codebook 32 and the stochastic codebook 35, the index information of the gain codebook 37 and the density information of the density detection circuit 22 in addition to the index information of representative vectors in the vector quantizer 14. Since the density values or the Dynamic Code Book Index are 5 parameters inherently required to be transmitted, the amount of the transmitted information or the transmission rate is not increased. if the parameters are not going to be inherently transmitted, such as the density information, it is used as a reference base to go between the code book for the male voice and the book of code for the female voice, it is necessary to transmit separate code switching information. It will be noted that the discrimination between the male voice and the female voice need not coincide with the The sex of the speaker always and when the selection of the code book has been made under the same rule as that for the classification of the training data. Therefore, the name of the code book for male voice and the code book for female voice is only the name of convenience. In the present embodiment, the codebooks are changed depending on the density value by exploiting the fact that there is a correlation between the density value and the spectral envelope configuration.

The present invention is not limited to the aforementioned embodiments. Although each component of the arrangement of Figure 1 manifests as hardware, it can also be implemented by a software program using a so-called digital signal processor (DSP). The code book on the short range side of quantization of band division vector or partial code book such as a codebook for part of the vector quantization of multiple caps can be switched between multiple code books for male voice and for female voice. In addition, the array quantization can also be carried out instead of the vector quantization by grouping the multiple frame data together. In addition, the speech coding method according to the present invention is not limited to the linear prediction coding method employing code excitation but can also be applied to a , "variety of speech coding methods wherein the voice portion is synthesized by sinusoidal wave-0 synthesis and the voiceless portion is synthesized based on the noise signal." As for use, the present invention is not limited to the transmission or registration / reproduction but can be applied to a variety of uses, such as speech modification of density conversion, synthesis 5 of regular speech or noise suppression.

Industrial Applicability As will be apparent from the foregoing description, a speech coding method according to the present invention provides a first codebook and a second codebook formed by classifying the parameters representing the short-term prediction values related to a reference parameter comprising one or a combination of a plurality of characteristic parameters of the input speech signal. The short-term prediction values are then generated based on the input speech signal and one of the first and second codebooks is selected in relation to the reference parameter as the input speech signal. The short-term prediction values are coded having reference to the selected codebook to encode the input speech signal.

/ - This improves the quantification efficiency. For example, the quality of the signal can be improved without increasing the bit rate of transmission or the bit rate of transmission can be further decreased while suppressing deterioration in signal quality.

Claims (31)

1. CLAIMS: 1. A method of speech coding comprising: generating values related to short-term prediction based on an input speech signal; provide a first codebook and a second codebook formed by classifying parameters representing the values related to the short-term prediction in relation to a reference parameter and producing data based on the classified parameters, the reference parameter comprises one or a combination of a plurality of characteristic parameters of the input speech signal; selecting one of the first and second codebooks in relation to the reference parameter of the input speech signal; and quantifying the values related to the short-term projection by referring to the codebook selected to encode the input speech signal.
2. 2. The speech coding method according to claim 1, wherein the values related to the short-term prediction are short-term prediction coefficients.
3. 3. The speech coding method according to claim 1, wherein the values related to the short-term prediction are short-term prediction errors.
4. 4. The speech coding method according to claim 1, wherein the characteristic parameters are the density value of a speech signal, the intensity of density, the power of the frame and a discrimination flag with speech / without voice and the i3 * "gradient of the signal spectrum 5. The speech coding method according to claim 1, wherein the values related to the short-term prediction are quantized in vector to encode the speech output. 6. The speech coding method according to claim 1, wherein the values related to the short-term prediction are quantized in matrix to encode the input speech signal 7. The speech coding method of 20 according to claim 1, wherein the reference parameter is the density value of the speech signal and wherein one of the first codebook and the second codebook is selected depending on the magnitude relationship of the density value of the 5-input speech signal and a pre-graduated density value. 8. A speech coding device comprising: a short-term prediction means for generating short-term prediction coefficients based on the input speech signals; a plurality of codebooks formed by classifying the parameters specifying the short-term prediction coefficients with respect to the reference parameters, the reference parameters J * "" being the combination of one or more of a plurality of parameters characteristic of the speech signals; a selection means for selecting one of the codebooks in relation to the reference parameter of the input speech signals; and 15 a quantization means for quantifying the short-term prediction coefficients by reference to the code book selected by the means selected; wherein the improvement is that an excitation signal is brought to optimum using a quantized value for the quantization means. The speech coding device according to claim 8, wherein the characteristic parameters include a density value of speech signals, density intensity, frame power, speech discrimination flag / voiceless and the gradient of the signal spectrum. 10. The speech coding device according to claim 8, wherein the quantizing means quantizes the vector and the short-term prediction coefficients. 11. The speech coding device according to claim 8, wherein the quantizing means quantizes the matrix and the short-term prediction coefficients. 12. The speech coding device according to claim 8, where the reference parameter is an intensity value of the speech signals the selection means selects one of the codebooks in response to the relative magnitude of the density value of the input speech signals and the density value undergraduate The speech coding device according to claim 8, wherein the codebooks include a codebook for a male voice and a codebook for a female voice. 14. A method of speech coding comprising: generating short-term prediction coefficients based on input speech signals; providing a plurality of codebooks formed by classifying parameters specifying the short-term prediction coefficients with respect to the reference parameters, the reference parameters 5 being the combination of one or more of the characteristic parameters of the speech signals; selecting one of the codebooks in relation to the reference parameters of the input speech signals; j / P * 'quantify the short-term prediction coefficients by referring to the selected codebook; and optimizing an excitation signal using a quantized value of the short-term prediction coefficients. 15. The speech coding method according to claim 14, wherein the characteristic parameters include a density value of speech signals, density intensity, 20 frame power and a voice discrimination / voiceless flag and the gradient of the signal spectrum. 16. The speech coding method according to claim 14, wherein the short-term prediction coefficients are quantified in 25 vector for encoding input speech signals; 17. The speech coding method according to claim 14, wherein the short-term prediction coefficients are quantized in matrix to encode the input speech signals. 18. The speech coding method according to claim 14, wherein the reference parameter is a density value of the speech signals and wherein one of the codebooks is selected in response to the relative magnitude of the speech. density value -W > of the input speech signals, and the pre-graduated density value. 19. The speech coding method according to claim 14, wherein the codebooks include a codebook for a voice 15 male and a code book for a female voice. 20. A speech coding device comprising: a short-term prediction means for generating short-term prediction coefficients based on the input speech signals; a first plurality of codebooks formed by classifying parameters specifying the short-term prediction coefficients with respect to the reference parameters, the reference parameters being the combination of one or more of the characteristic parameters of the speech signals; a selection means for selecting one of the codebooks in relation to the reference parameters of the input speech signals; a quantization means for quantifying the short-term prediction coefficients by reference to the codebook selected by the selection means; a second plurality of code books formed on the basis of the training data classified with respect to the reference parameters, the reference parameters being the combination of one or more of the characteristic parameters of the speech signals, one of the second plurality of codebook being selected as the codebook of the first plurality of codebooks that is selected by the 'means of selection; and a synthesis means for synthesizing on the basis of the quantized value from the quantizing means, an excitation signal related to the sending of selected codebook of the second plurality of codebooks: the excitation signal is brought to optimum in response to an exit from the synthesis medium. 21. The speech coding device according to claim 8, wherein the characteristic parameters include a density value of the speech signals, the density intensity, the frame power, a voice discrimination / voiceless flag and the gradient of the signal spectrum. 22. The speech coding device according to claim 20, where the means of quantification quantifies the vector of the short-term projection coefficients. 23. The speech coding device according to claim 20, wherein the quantizing means quantizes the matrix of the short-term transmission coefficients. 24. The speech coding device according to claim 20, wherein the reference parameter is a density value of the speech signals and wherein the selection means selects one of a first plurality of codebooks in response to the relative magnitude of the density value of the input speech signals and the pre-graduated density value. 25. The speech coding device according to claim 20, wherein each of the first plurality of codebooks and the second plurality of codebooks includes a codebook for male voice and a codebook for speech feminine 26. A speech coding method comprising: 5 generating short-term prediction coefficients based on the input speech signals; provide a plurality of code books formed by classifying the parameters specifying the short-term prediction coefficients with respect to the J r- reference parameters, the reference parameters being the combination of one or more of the characteristic parameters of the speaks; selecting one of the first plurality of codebooks in relation to the reference parameters of the input speech signals; quantify the short-term prediction coefficients by referring to the book of the selected code; provide a second plurality of code books formed on the basis of training data classified with respect to the reference parameters, the reference parameters being the combination of one or more of the characteristic parameters of the speech signals, one of the second plurality of codebooks is selected with the codebook selection of the first plurality of code; and synthesizing, on the basis of the quantized value of the short-term prediction coefficients, an excitation signal related to the sending of the codebook selected from the second plurality of codebooks to bring the excitation signal to the optimum. 27. The speech coding method according to claim 26, wherein the characteristic parameters include a density value of the speech signals, the density intensity, the frame power, and a speech discrimination flag / without voice and a gradient of the signal spectrum. 28. The speech coding method according to claim 26, wherein the short-term prediction coefficients are quantized in the vector to encode the input speech signals. 29. The speech coding method according to claim 26, wherein the short-term prediction coefficients are quantized in matrix to encode the input speech signals. 30. The speech coding method according to claim 26, wherein the reference parameter is a density value of the speech signals and wherein one of the first plurality of codebooks is selected in response to the magnitude relative to the density value of the input speech signals and the pre-graduated density value. The method of speech coding according to claim 26, wherein each of the first plurality of codebooks and the second plurality of codebooks includes a codebook for male voice and a codebook for speech feminine