JPH09244698A - Voice coding/decoding system and device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a highly subjective voice quality by using a meridional generalized cepstrum as the expression parameter of a voice spectrum. SOLUTION: A meridional generalized cepstrum in which a power parameter γis set to other than γ0 or γ=-1 is made to be a parameter. The synthetic filter 1 constituted based on drivings of filter coefficients calculated from the parameter obtained by the analysis of a meridional generalized cepstrum analyzing method and an auditory sensation weighting filter 2 are applied to a CELP. In this case, the auditory sensation weighting filter 2 is realized by using a B(Z). Thus, spectral poles and zeros are not only expressed in a form in which the auditory characteristic of the human race is concidered and a voice synthesis is achieved with the synthetic filter having simple constitution but also the voice synthesis of higher quality is made possible because the auditory sensation weighting filter 2 and the filter coefficients of the post-filter are selected in such a way by using the meridional generalized cepstrum analysis.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coding / decoding system, and more particularly to a speech coding / decoding system and apparatus using a mel generalized cepstrum. In this specification, “voice coding / decoding” is a concept including voice coding, voice decoding, and voice code decoding.

[0002]

2. Description of the Related Art Conventionally, in a speech coding method, a speech generation model composed of two main elements, an excitation source model (sound source model) and a speech model (vocal tract model), is used to encode speech. . When using such a speech generation model to improve subjective speech quality, encoding,
Performs processes such as "noise shaping (or perceptual weighting)" and "post-filtering" to improve the subjective performance by controlling the spectral shape of the quantization noise generated by decoding and the spectral shape of the speech signal. In general, these processes improve the subjective performance in consideration of auditory masking characteristics. In the "noise shaping" process, the energy of the quantization noise is reduced in a portion where the energy of the speech spectrum is small, and the energy of the quantization noise is increased in a frequency axis so that the energy of the speech spectrum is increased in a portion where the energy is large. This is the method of controlling above.

The "post-filtering" process is an operation on a decoded audio signal, which is relatively SN.
Energy in the valley portion of the speech spectrum where the ratio becomes small
Is a method of performing filtering so as to emphasize a mountain portion of a speech spectrum having a relatively small SN ratio and a relatively large SN ratio, and any of the above processes is effective in reducing an auditory noise feeling. In the above-described conventional speech coding / decoding method, all-pole type speech spectrum model based on the linear prediction method is often used, and the above-mentioned "noise shaping", "post-filtering" processing and the like are also linear prediction. It is often performed based on the all-pole spectrum obtained by the method. FIG. 15 illustrates CELP coding using a coding method in which a linear prediction coefficient corresponding to a spectral envelope is extracted from the input speech by a linear prediction analysis method and is driven by an excitation vector stored in a codebook to synthesize speech. The configuration of the method is shown. In the figure, 1 is a synthesis filter, 2 is a perceptual weighting filter, 3 is an adaptive codebook, 4 is a noise codebook, 5 is error minimization, and the synthesis filter 1 uses a linear prediction coefficient a (m). ,

[0004]

[Equation 1] , And the perceptual weighting filter 2 is

[0005]

[Equation 2] Is defined by

The adaptive codebook 3 is used to express the periodicity of speech and stores past excitation sources.
Further, the noise codebook 4 is used for expressing a variable portion of speech that cannot be expressed by the adaptive codebook, and the vectors of the codebooks 3 and 4 are analyzed by synthesis (Analysis by Synthesis).
(AbS)) It is determined by the closed loop search by the method, the adaptive codebook 3 and the noise codebook 4 are searched in this order, and then the gain is determined by vector quantization. That is, the vectors from the adaptive codebook 3 and the noise codebook 4 are input to the synthesis filter 1, the difference between the output of the synthesis filter 1 and the input speech is obtained, and the error data is input to the auditory weighting filter 2. Then, the outputs of the adaptive codebook 3 and the noise codebook 4 are updated by error minimization 5 so that the output of the perceptual weighting filter 2 is minimized. On the other hand, the decoder is
6, a synthesis filter 1, an adaptive codebook 3, a noise codebook 4 and a post filter 6 are provided. Based on the transmitted information, the synthesis source 1 is driven by the excitation source, and the post filter 6 is passed. This will reduce auditory noise. Here, the post filter 6 is generally

[0007]

(Equation 3) With a transfer function such as A (z / δ1) / A (z / δ
The formant is emphasized in 2), and the slope of the spectrum is corrected in (1-μz ^-1 ).

[0008]

The CELP coding method based on such linear prediction is (1) the expression parameter of the spectrum can be obtained with a small amount of calculation, and (2) the LSP (Line
Parameters can be effectively quantized by using the Spectrum Pair) expression, etc. (3) LSP (Li
There is an advantage that the stability of the synthesis filter can be guaranteed by using the ne Spectrum Pair) expression and the like. However, when the speech spectrum is represented by the all-pole model, it is extremely difficult to accurately represent the speech spectrum including zeros such as nasal sounds, and there is a drawback that high quality synthetic speech cannot be obtained. . Therefore, it is considered necessary to use a spectral model that can represent not only the poles of the spectrum but also the zeros in order to perform the operation for improving the subjective performance more effectively. , The present inventors
As disclosed in -38 (1994-09), there is proposed a CELP coding method using a mel-cepstrum as a parameter, which can represent the poles and zeros of the spectrum with equal accuracy. In addition, when the mel cepstrum is used, there is an advantage that the auditory characteristics of the human frequency axis having fine resolution in the low frequency range and coarse resolution in the high frequency range can be considered.

However, in the speech synthesis using the spectrum model capable of expressing the zero point of the spectrum as described above, although the poles and zeros of the spectrum can be expressed with the same accuracy, the human being can be more accurately expressed. It is not possible to obtain synthetic speech that reflects the auditory characteristics. That is,
Since phonological characteristics such as nasal sounds are characterized by the zeros of the spectrum, it is necessary to represent the zeros, but human auditory characteristics are more sensitive to the poles (peaks of the spectrum) than to the zeros (valleys of the spectrum). , By increasing the representation accuracy of the poles higher than the representation accuracy of zero, it is possible to achieve spectral representation that is more suitable for human auditory characteristics. However, when a spectrum is expressed using a mel cepstrum, human auditory characteristics can be introduced on the frequency axis, but there is a limit to the improvement of sound quality because weighting cannot be applied to the expression precision of the poles and zeros of the spectrum. There was a problem.

[0010]

(Equation 4) The present invention has been made in order to solve the above-mentioned problems, and by using a spectral representation by a mel-generalized cepstrum that can consider human auditory characteristics while accurately expressing a speech spectrum, high subjective speech quality It is an object of the present invention to provide a speech coding / decoding system and a device therefor that can obtain

[0011]

In order to solve the above-mentioned problems, the speech coding / decoding system according to the present invention is characterized in that a mel generalized cepstrum is used as a speech spectrum expression parameter. Further, a voice coding / decoding device that realizes the above voice coding / decoding system is CELP, AD.
A voice code / decoding system that implements a voice coding / decoding method using some kind of voice spectrum expression including PCM and MBE
In the decoding device, parts relating to the representation of the speech spectrum (speech spectrum analysis unit, synthesis filter, noise shaping filter, post filter, spectral parameter quantization unit, etc.) are replaced with those by the mel generalized cepstrum. It is characterized by taking a configuration.

[0012]

BEST MODE FOR CARRYING OUT THE INVENTION The present invention will be described below in detail based on the embodiments shown in the drawings. The voice encoding / decoding method according to the present invention is based on ADPCM (Adaptive Pulse Cod).
e Modulation), APC (Adaptive Predictive Coding),
CELP (Code Excited Linear Prediction), MBE
(MultiBand Excitation Vocoder), ATC (Adaptive T
ransformation coding, etc., it can be applied to each speech coding / decoding method that uses a spectrum representation of speech, but here, an example applied to the CELP coding method will be described. (Mel Generalized Cepstrum Analysis) In the speech coding method according to the present invention, the spectrum estimated by the mel generalized cepstrum analysis is used instead of the all-pole spectrum 1 / A (exp (jω)) estimated by the linear prediction method. It is a thing. The spectrum estimated by the mel generalized cepstrum analysis has mel generalized cepstrum coefficients c = [c (0) ,. . . c (M)] T

[0013]

(Equation 5) It is expressed as However,

[0014]

(Equation 6)

[0015]

(Equation 7)

[0016]

(Equation 8) It is.

It is guaranteed that H (z) obtained by this analysis is always stable, and the phase characteristic ω ^* of the first-order all-pass function Z ^{* -1} = exp (-jω) is the sample. It is known that if α is selected to be 0.31 when the digitization frequency is 8 kHz, the Mel scale, which represents the auditory characteristics with respect to the pitch of human sound, can be approximated well. The spectral model of the above [Formula 5] is an all-pole model with (α, γ) = (0, −1), and (α,
γ) = (0,0) with cepstrum model, and | α | <
When 1 and γ = 0, it is equivalent to the mel cepstrum model. Further, the spectrum model can be continuously changed depending on the values of a and γ. The characteristics of the spectral model based on the values of α and γ can be summarized as follows. Regarding α, a uniform frequency resolution is obtained when α = 0, and a high resolution is obtained in a low frequency range when α> 0. Regarding γ, when γ = −1, the pole of the spectrum is emphasized, and when −1 <γ <0, γ = −
It has an intermediate property between 1 and γ = 0. When γ = 0, the poles and zeros of the spectrum are expressed in the same way.

(Structure of Synthesis Filter) Next, the structure of the synthesis filter will be described. As described above, the transfer function H (z) obtained by the analysis is always stable, and its spectrum continuously changes depending on the values of α and γ.
If a synthesis filter for arbitrary α and γ can be realized, it is possible to select the value of γ that gives the most preferable synthesized sound. Here, since the mel generalized cepstrum is not easy to handle because each coefficient value changes as a whole due to the change in gain, the gain is calculated from H (z),
Consider expressing H (z) by a parameter that does not depend on gain. [Equation 5] with K as the gain,

[0019]

[Equation 9] And transform. However,

[0020]

(Equation 10)

[0021]

[Equation 11] Also,

[0022]

(Equation 12)

[0023]

(Equation 13) It is. Since the gain of the filter D (z) is always 1, the mel generalized cepstrum coefficient c ′ of D (z) is

[0024]

[Equation 14] To be satisfied. this is,

[0025]

(Equation 15) Is equivalent to At this time, c'and c are

[0026]

(Equation 16) Is established. [Equation 11] is M when γ = 0
LSA (Mel Log Spectrum Approximation) filter,
For other γ, MGLSA (Mel Generalized)
It can be realized with high approximation accuracy as a Log Spectrum Approximation) filter. In particular, γ is a natural number and γ
= -1 / n, the synthesis filter D
(Z) is

[0027]

[Equation 17] However,

[0028]

(Equation 18) Since it can be expressed as follows, 1 / C ′ (z ^* ) is connected in n stages. However, C '(z ^* ) cannot be directly realized because it has a delay-free loop. So C '(z ^* )

[0029]

[Equation 19] However,

[0030]

(Equation 20)

[0031]

(Equation 21)

[0032]

(Equation 22)

[0033]

(Equation 23)

[0034]

(Equation 24)

[0035]

(Equation 25)

[0036]

(Equation 26) It transforms like. The transformation from the second line to the third line on the right side of [Equation 19] is given by [Equation 15].

[0037]

[Equation 27] Was used. The calculation of A ⁻¹ c ′ in [Equation 25] is
The following relational expression

[0038]

[Equation 28] Can be calculated recursively. Where filter B (z ^* )

[0039]

(Equation 29) In other words, it is realized by a simple structure of 1 / B (z ^* ) n-stage connection as shown in FIG. Also, voice analysis,
For voice synthesis, voice recognition, etc., γ =
Since a good result is obtained at -1/3, this value will be used also when introducing the Mel generalized cepstrum into the CELP coding method. In this case, the synthesis filter has a 3-stage cascade structure of 1 / B (z ^* ). The basic synthesis filter 1 / B (z ^* ) at each stage has the block configuration shown in FIG.

(Structure of Auditory Weighting Filter and Post Filter) Synthesis filter [Equation 36] based on the linear prediction method,

[0041]

[Equation 30] Auditory weighting filter, [Equation 31],

[0042]

(Equation 31) Post filter, [Equation 32],

[0043]

(Equation 32) In order to apply to the mel generalized cepstrum code system, simply substituting as in [Equation 33] gives

[0044]

[Equation 33] Synthesis filter, [Equation 34],

[0045]

(Equation 34) Auditory weighting filter, [Equation 35],

[0046]

(Equation 35) Post filter, [Equation 36],

[0047]

[Equation 36] Therefore, the mel-generalized cepstrum code system becomes quite complicated.
7],

[0048]

(37) Auditory weighting filter, [Equation 38],

[0049]

(38) Post filter, [Equation 39],

[0050]

[Equation 39] With this, it becomes possible to construct a coding system having a simple filter structure very easily.

(CE by Mel Generalized Cepstrum Analysis
LP method) Next, a synthesis filter, an auditory weighting filter, and a post filter configured on the basis of driving with the above-described mel generalized cepstrum analysis method and the filter coefficients obtained from the parameters obtained by the analysis method are used. To C
An example applied to the ELP will be described based on the form example shown in the drawings. Figure 3 shows CE by mel generalized cepstrum analysis.
FIG. 4 is a diagram showing a basic configuration of a CELP encoder when LP encoding is realized, and FIG. 4 is a diagram showing a basic configuration of a CELP decoder based on the Mel generalized cepstrum.
Here, the auditory weighting filter 2 and the post filter 6
Are realized by using B (z ^* ) and 1 / B (z ^* ), respectively. As a result, the poles and zeros of the spectrum can be expressed in a form that takes human auditory characteristics into consideration, and not only can speech synthesis be achieved by a synthesis filter with a simple configuration, but also the mel generalized cepstrum analysis can be used. Since the filter coefficients of the weighting filter and the post filter are selected as described above, higher quality speech synthesis can be performed.

FIG. 5A is a diagram showing a state of perceptual weighting and post-filtering in the voice encoding / decoding system according to the present invention, and FIG. 5B is a view for comparing the present invention with the conventional method. Speech code using linear prediction method (LPC) /
It is a figure which shows the state of perceptual weighting and a post-filtering in a decoding system. [Equation 2] and [Equation 3] are used for the weighting filter and the post filter of the linear prediction method, and each parameter is (δ1, δ2) = (0.
9, 0.4), (δ3, δ4) = (0.5, 0.8),
It was set to μ = 0.5. The speech spectrum estimated by the mel generalized cepstrum analysis shown in FIG.
As is clear from comparison with the linear prediction method shown in (b), not only the poles of the spectrum but also zero are expressed,
It can be seen that the resolution is particularly high in the low frequency range. Further, since the auditory weighting and post-filtering processing of the encoding system according to the present invention is performed via the mel generalized cepstrum, the estimated spectrum information is well preserved, and the auditory performance can be expected to be improved. .

However, in the configurations shown in FIGS. 3 and 4, the coefficients of the perceptual weighting filter and the post filter are B (z ^* ).
Alternatively, it is 1 / B (z ^* ), and since the auditory weighting filter has no variable, the strength of auditory weighting cannot be adjusted. In addition, although there is a slope correction coefficient μ in the post filter, it is not possible to adjust the strength of post filtering with this coefficient. Therefore, by adjusting each filter, high-quality speech considering human auditory characteristics can be adjusted. Can't get Therefore, a method of adjusting the perceptual weighting and the strength of post-filtering by expanding the bandwidth of the z ^* region will be described below.

[0054] (z ^* area bandwidth extension in) bandwidth extension in the z ^* region, [Expression 2], for the z → z / beta as A (z / β) [Expression 3] Instead, the bandwidth is expanded by replacing z ^* → z ^* / β.
B (z ^* ) of [Equation 29], that is, c ′ (z ^* / β) obtained by bandwidth-extending c ′ (z ^* ) in the z ^* region is

[0055]

(Equation 40) However,

[0056]

[Equation 41] It can be expressed as. When [Equation 40] is transformed in the same manner as [Equation 19],

[0057]

(Equation 42) However,

[0058]

[Equation 43] Becomes Generally, b (0) ≠ 0 and c ^* (z ^* /
The gain of β) is not 1, so if we derive the gain from [Equation 42],

[0059]

[Equation 44] However,

[0060]

[Equation 45]

[0061]

[Equation 46] Also,

[0062]

[Equation 47] Therefore, the relation between b and b ′ is given by the following equation.

[0063]

[Equation 48] As described above, the bandwidth expansion operation in the z domain is performed by the following recursive expression from c ′ to [Equation 43].

[0064]

[Equation 49] [Equation 50] is calculated by, and [Equation 5] is calculated using [Equation 48].
1] is required. This factor, by a factor of B (z ^*) the number 51, can be implemented by B (z ^*) with exactly the same configuration.

[0065]

[Equation 50]

[0066]

(Equation 51) Furthermore, [Equation 43] is

[0067]

[Equation 52] It is also possible to transform b to obtain b ^* ′ from b through the matrix A ⁻¹ FA. At this time, the matrix A
The calculation of ⁻¹ FA can be performed recursively as in [Equation 49]. Perceptual weighting and post-filtering are performed with filters B (z ^* ), 1 / B (z ^* ), B (z ^* / β),
By combining 1 / B (z ^* / β), its strength and shape can be adjusted, and one example thereof is shown in FIGS.
Shown in

FIG. 6 shows an example of the structure of the perceptual weighting filter, and FIG. 7 shows an example of the structure of the post filter. The state of perceptual weighting by these filters is shown in FIGS. 8 (a) to 8 (d). Moreover, the states of the spectra of the post-filtering are shown in FIGS. In the state of the spectrum of the post-filtering shown in FIG. 9, the slope correction coefficients from (a) to (d) are respectively μ
= 0.5, 0.4, 0.3, 0.2. FIG.
Each spectrum in (a) and FIG. 9 (a) is equivalent to the weighting and post-filtering spectrum shown in FIG. 5 (a). Therefore, as is clear from FIGS. 8A to 8D and FIGS. 9A to 9D, by reducing β without lowering the resolution in the low frequency range, the auditory weighting is strengthened and the post-filtering is performed. Can be weakened by increasing β.

FIG. 10 is a diagram showing a running spectrum when the bandwidth extension of the z ^* region is introduced, and (a)
Is an input speech waveform, (b) is a running spectrum by mel generalized cepstrum analysis, and (c) is a conventional method (LP
C) Running spectrum by analysis. Also,
(B) and (c) i), ii), and iii) respectively show the estimated spectrum, the perceptual weighting, and the state by post-filtering with respect to the input speech. However, due to the relation of the figure, the perceptual weighting ii) is drawn as the spectrum of the inverse filter. To obtain this running spectrum, each parameter is β = 0.8 for auditory weighting and β = μ = 0.3 for post filtering.
Further, the same values as those in FIG. 5B were used as the parameters for obtaining the running spectrum by the conventional method analysis shown in (c). Further, among the ones shown in (b) and (c), the auditory weighting and post-filtering spectra ii) and iii) are drawn in a scale twice as large as the estimated spectrum i). As is clear from this figure, it can be seen that the spectrum represented in the coding system according to the present invention has high resolution in the low frequency range.

FIG. 11 is a diagram showing another embodiment of the encoder system in which the speech coding / decoding system according to the present invention is applied to CELP. A voice coding / decoding system according to the present invention;
That is, in the case of using the method based on the mel generalized cepstrum analysis, the position of the auditory weighting filter is changed in the encoder, so that the encoding is performed with a simpler configuration than that of the embodiment shown in FIG. The whole system can be constructed. Further, in order to adjust the strength of perceptual weighting, it is possible to simplify the structure of the synthesis filter by adopting the structure shown in FIG.

[0071]

EXAMPLE An example of the above-described embodiment will be shown below. FIG.
3 (a) is a table showing speech analysis conditions, and FIG. 3 (b) is a table showing bit allocation of the coding system. The adaptive codebook represents an 8-bit 256 state with a non-integer lag. The noise codebook is composed of Gaussian noise. The code length search is performed in the order of the adaptive codebook and the noise codebook, and the noise codebook search is performed after orthogonalizing each vector of the noise code length to the vector determined by the adaptive codebook. . Further, to ensure the stability of the synthesis filter, the mel generalized cepstrum was performed using a line spectral representation in the z ^* domain of C '(z ^* ). In both the conventional method and the present invention, a subjective evaluation experiment using the equivalent Q value was performed under the same conditions for the configurations other than the voice analysis and the filter based on the above configurations.

(Results of Objective Evaluation Experiment) FIG. 14 shows the results of the above evaluation experiment. As shown in the figure, an average Q improvement of about 2.3 dB in subjective performance is achieved.

[0073]

As described above, according to the present invention, a high subjective voice quality is obtained by using the spectral representation by the mel-generalized cepstrum capable of considering the human auditory characteristics while accurately expressing the speech spectrum. It is possible to improve the voice quality of the voice encoding / decoding system.

[Brief description of drawings]

FIG. 1 is an explanatory diagram of the configuration of one embodiment of the present invention.

FIG. 2 is a block diagram showing 1 / B (z ^* ) of a basic synthesis filter of each stage.

FIG. 3 is a diagram showing a basic configuration of a CELP encoder when CELP encoding is realized by a mel generalized cepstrum analysis.

FIG. 4 is a diagram showing a basic configuration of a CELP decoder based on mel generalized cepstrum analysis.

5A is a diagram showing a state of perceptual weighting and post-filtering in a voice encoding / decoding system according to the present invention, and FIG. 5B is a comparison diagram of the present invention and a conventional example.

FIG. 6 is a diagram showing an example of the configuration of a perceptual weighting filter.

FIG. 7 is a diagram showing an example of a configuration of a post filter.

8A to 8D are diagrams showing how auditory weighting is performed by a filter.

9A to 9D are diagrams showing a spectrum of post-filtering.

FIG. 10 is a diagram showing a running spectrum when the bandwidth extension of the z ^* region is introduced, (a) is an input speech waveform, (b) is a running spectrum by mel generalized cepstrum analysis, and (c) is. It is a running spectrum by a conventional method (LPC) analysis.

FIG. 11 is a block diagram showing a voice coding / decoding method according to the present invention as CEL.
It is a figure which shows the other embodiment of the encoder system applied to P.

FIG. 12 is a diagram showing a configuration example for adjusting the strength of perceptual weighting.

13A is a diagram showing speech analysis conditions, and FIG. 13B is a diagram showing coding system bit allocation.

FIG. 14 is a diagram showing a result of an evaluation experiment.

FIG. 15 is a CELP code using a coding method in which a linear prediction coefficient corresponding to a spectrum envelope is extracted from input speech by a linear prediction analysis method and is driven by an excitation vector stored in a codebook to synthesize speech. The figure which shows the structure of the conversion system.

FIG. 16 is an explanatory diagram of a configuration of a conventional decoder.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (71) Applicant 596045638 Kazuto Koishida 2-2 Suzukageso, 10 Kayajuku, Nakahara-ku, Kawasaki-shi, Kanagawa 2 (71) Applicant 000003104 Toyo Communication Equipment Co., Ltd. 2 Otani, Samukawa-cho, Takaza-gun, Kanagawa Prefecture 1-1-1 (72) Inventor Satoshi Imai 5-24-3 Tamagawa Gakuen, Machida, Tokyo (72) Inventor Takao Kobayashi 1-4-3-201 Minamiyamada, Tsuzuki-ku, Yokohama-shi, Kanagawa (72) Inventor Tokuda Keiichi Tokyo 2739-12 Naruse, Machida-shi, Tokyo River 72 Nishinase 204 (72) Inventor Kazuto Koishida 2-2 Suzukageso, 10 Kayajuku, Nakahara-ku, Kawasaki-shi, Kanagawa Yusuke Furukawa Kanagawa-cho, Takaza-gun, Kanagawa 2-1-1 Kotani Toyo Communication Equipment Co., Ltd.

Claims (9)

[Claims] 1. A speech code using a spectral representation of speech /
A speech coding / decoding method characterized in that, in the decoding method, a mel-generalized cepstrum whose power parameter γ is set to a value other than γ = 0 or γ = -1 is used as a parameter. 2. The adaptive codebook, the random codebook, the error minimization, the synthesis filter, and the auditory weighting filter are provided on the coding side, and the adaptive codebook, the random codebook, and the synthesis on the decoding side. In a CELP speech coding / decoding system including a filter and a post filter, the CELP speech coding / decoding device is characterized in that the coefficient of the synthesizing filter uses a parameter by mel generalized cepstrum analysis. 3. The CELP speech coding / decoding apparatus according to claim 2, wherein the coding and decoding side synthesis filter is 1.
/ B (z ^* ) is connected in n stages, and the auditory weighting filter on the code side is B (z ^* ),
Further, the CELP speech coding / decoding device is characterized in that 1 / B (z ^* ) is used as a post filter on the composite side. 4. The CELP speech coding / decoding device according to claim 2, wherein the synthesizing filter on the coding and decoding side is 1.
/ B (z ^* ) is connected in n stages, and the auditory weighting filter is B (z ^* ), B (z ^* ), 1 / B.
A CELP speech coding / decoding device having a configuration including a series connection of (z ^* / β). 5. The CELP speech coding / decoding device according to claim 2, wherein the coding and decoding side synthesis filter is 1.
/ B (z ^* ) n-stage connection, and the post filter includes a series connection of B (z ^* / β) and 1 / B (z ^* ) in series. / Decoding device. 6. A CELP speech coder / decoder according to claim 3, 4 or 5, wherein n = 3. 7. The CELP speech coder / decoder according to claim 2, wherein the synthesizing filter on the code side is 1 / B (z
^CE ), wherein the auditory weighting filter is configured by B (z ^* ) and the output difference between the synthesis filter and the auditory weighting filter is minimized. Speech coding / decoding device. 8. The CELP speech coding / decoding apparatus according to claim 2, wherein the decoding side synthesis filter is 1 / B (z
^* ) And 1 / B (z ^* / β) are connected in series and the post filter is B (z ^* ) B (z ^* ) / B (z ^*
/ Β), and is configured to minimize the error in the output difference between the synthesis filter and the post filter. 9. The CELP speech coder / decoder according to claim 2, wherein the mel generalized cepstrum is C ′ (z ^* ).
Quantization using a line spectral representation in the z ^* domain of
A CELP coding / decoding device characterized by interpolating.