JPH056199A - Voice parameter coding system - Google Patents

Abstract

PURPOSE:To quantize a spectrum parameter of a voice signal with a number of bits which is less than that conventionally used. CONSTITUTION:An inputted voice signal is divided into frames which are then further divided into sub frames shorter than the frames. Spectrum parameters for predetermined orders are obtained for at least one of the subframes, with respect to the voice signal. A vector quantizing means 110 quantizes a spectrum parameter of one of the subframes with the use of a previously constituted code book 120, and delivers thus quantized value. Further, an interpolating vector quantizing means 130 quantizes a spectrum parameter of at least one of the other parameters with the use of the quantized value of the spectrum parameter of the instant frame, a quantized value of a spectrum parameter of a past frame and a second code book 140, and delivers thus quantized value.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice parameter coding system for use in a voice coding system for coding a voice signal with a low bit rate, particularly with a bit rate of 8 kb / s or less.

[0002]

2. Description of the Related Art As a method for encoding a voice signal at a low bit rate of 8 kb / s or less, for example, M. Sch
roeder and B.I. "Cod by Atal
e-excited linear predictive
on: High quality speech at
Very low bit rates "(Pro
c. ICASSP, pp. 937-940, 1985.
(1), and “Improved speech quality” by Kleijn et al.
and effective vector qua
ntification in SELP "(Proc. IC
ASSP, pp. 155-158, 1988) entitled papers (CELP described in the literature 2), or the like (C od
e E xcited LP C Coding) method is known. In this method, the transmitting side first extracts a spectrum parameter indicating the spectrum characteristic of the voice signal from the voice signal for each frame (for example, 20 ms), and further divides the frame into sub-interval subframes (for example, 5 ms). Next, for each subframe, a pitch parameter representing a long-term correlation (pitch correlation) is extracted based on the past sound source signal, and the speech signal of the above-mentioned subframe is long-term predicted by this pitch parameter. Next, with respect to the residual signal obtained by long-term prediction, the error power between the signal synthesized with the signal selected from the codebook consisting of the noise signal of the predetermined type and the above-mentioned voice signal is minimized. A noise signal of one type is selected, and the index and gain representing this type and the above-mentioned spectrum parameter and pitch parameter are transmitted.

In the CELP method, LPC parameters obtained by LPC analysis are used as speech spectrum parameters, and the LPC parameters are quantized. Scalar quantization is usually used as the quantization method.
Quantizing the next LPC coefficient requires about 34 bits (1.7 kb / s) per frame, and if the number of bits is further reduced, the sound quality deteriorates. As a method for more efficiently quantizing the LPC parameter, Mori
"Transform coding by ya, et al.
of speech using a weighte
d vector quantizer, "(IEEE J. Sel. Areas, Commu
n. , Pp. 425-431, 1988) (Reference 3)
Although the vector-scalar quantization method described in U.S.A., etc. is proposed, it requires about 27 to 30 bits per frame.

[0004]

In order to further reduce the bit rate of the CELP method described above, an efficient quantization method for spectrum parameters is required. However, the number of bits required for quantization of spectrum parameters is reduced. If the frame length is made long in order to lower the frequency, it becomes difficult to express the temporal change of the spectrum well, the time distortion increases, and the sound quality deteriorates significantly.

An object of the present invention is to solve the above-mentioned problems and to provide a voice parameter coding system capable of providing good sound quality even if a spectrum parameter is quantized with a smaller number of bits than the conventional one.

[0006]

A speech parameter encoding system according to a first aspect of the present invention divides an input speech signal frame and further divides it into a plurality of subframes shorter than the frame, and the subframe of the subframe. A means for obtaining a spectral parameter of at least one predetermined order with respect to the speech signal for at least one, and a spectral parameter of one subframe is quantized by using a preconfigured first codebook, and at least another one It is characterized in that the spectral parameter of the subframe is quantized by using the quantized value of the spectral parameter in the frame, the quantized value of the spectral parameter in the past frame, and the second codebook.

A speech parameter coding system according to a second invention is the speech parameter coding system according to the first invention, wherein the first codebook or the second codebook is used.
In at least one of the codebooks, a plurality of types of candidates are output in the ascending order of quantization distortion, and in other subframes, the spectral parameter is a quantized value of the spectrum parameter in the frame and the spectrum in the past frame. Quantizing using a quantized value of a parameter and a second codebook to obtain a quantized distortion, accumulating the quantized distortion over a plurality of subframes, and selecting and outputting a candidate that minimizes the accumulated value. Is characterized by.

A speech parameter coding system according to a third aspect of the present invention comprises means for dividing an inputted speech signal into frames and further dividing into a plurality of subframes shorter than the frame, and at least one of the subframes. A means for obtaining a spectral parameter of a predetermined order for an audio signal, and a spectral parameter of one subframe are quantized using a preconfigured first codebook, and the frame is quantized in at least one other subframe. The spectral parameter is restored by a predetermined method using the quantized value of the spectral parameter and the quantized value of the spectral parameter in the past frame, and the error between the spectral parameter of the subframe and the restored value is calculated. You can choose a second codebook to make it smaller. The features.

A speech parameter coding system according to a fourth invention is the speech parameter coding system according to the second invention, wherein the first codebook or the second codebook is used.
In at least one of the codebooks, a plurality of types of candidates are output in ascending order of quantization distortion, and in other subframes, the error power between the spectral parameter of the subframe and the restoration value is accumulated over a plurality of subframes. , A candidate that minimizes the cumulative value is selected and output.

In the speech parameter coding system of the fifth invention, means for dividing the inputted speech signal into frames and further dividing into a plurality of subframes shorter than the frame, and at least one of the subframes are described above. A means for obtaining a spectral parameter in a predetermined order for a voice signal and a first codebook in which spectral parameters of one subframe are pre-configured, and a plurality of types of candidates are quantized in ascending order of quantization distortion. In another subframe, the spectrum parameter is restored by a predetermined method using the quantized spectrum parameter of the frame and the quantized spectrum parameter of the past frame, and the spectrum parameter of the subframe and the restoration are performed. Accumulate error power with the value over multiple subframes And having a means for outputting select the candidates that minimizes the cumulative value.

[0011]

The operation of the speech parameter coding system according to the present invention will be described. FIG. 1 is a diagram for explaining the action of the first and second inventions, FIG. 3 is a diagram for explaining the action of the third and fourth inventions, and FIG. 4 is a diagram for explaining the action of the fifth invention. In the following description, the LSP parameter is used as the spectrum parameter of voice.

The input voice signal is divided into frames having a predetermined time length (for example, 30 to 40 ms), and the voice signal of the frame is further divided into a plurality of subframes (for example, 5 to 8 ms) shorter than the frame. However, for at least one subframe in the frame, a well-known LP
C analysis is performed to obtain spectral parameters. Here, the frame length is 40 ms, the subframe length is 8 ms, and the subframes shown in FIG.
PC analysis shall be performed.

First, the operation of the first invention will be described with reference to FIG.

As an example, for the first subframe, L
The spectral parameters obtained by PC analysis are input from the terminal 100, and the spectral parameters obtained for the first subframe and the subframe are sequentially input from the terminal 105. As the spectrum parameter, a line spectrum pair (LSP) parameter will be used here for explanation.
The specific calculation method of LSP is “Quan” by Sugamura et al.
Tizerdesign in LSP speech
"Analysis-synthesis," (IEEE J. Sel. Areas Commu
n. , Pp. 425-431, 1988) (Reference 4)
Etc. can be referred to. The vector quantization unit 110 vector-quantizes the input LSP parameter using the codebook 120 (first codebook). The codebook 120 is pre-learned using a large amount of LSP parameter sequences for training. The learning method is, for example, “Ana by Linde, Buzo, Gray”.
lgorithm for vector quant
The paper entitled "Ization design" (IEEE
E Trans. Commun. , Pp. 84-95,
1980) (reference 5) and the like. Although various distortion measures are known for searching a code vector, the LSP squared distance is used here. The LSP squared distance is given by equation (1).

[0015]

[Equation 1]

Here, LSP (i) represents the input i-th LSP coefficient. P is the order of the LSP. LSP '
(I) is the j-th code vector included in the codebook, j = 1 to 2 ^B (B is the number of bits of the codebook)
Is. The code vector that minimizes the equation (1) is output to the interpolation vector quantization unit 130 and the terminal 155.

The interpolation vector quantization unit 130 receives the output code vector of the vector quantization unit 110 and the quantized value of the spectrum parameter of one frame past obtained by delaying the output code vector by one frame. Using the interpolation coefficient codebook 140 (codebook 2),
LS of sub-frame input from terminal 105,
Interpolation vector quantization is performed on the P sequence to restore it. That is,
The quantization distortion by the interpolation coefficient code vector is calculated by the equation (2), and the interpolation coefficient code vector A _j that minimizes the quantization distortion is selected for each subframe.

[0018]

[Equation 2]

In equation (2), LSP ^l (i) is the l-th
The i-th LSP of the input in the subframe. LSP '
_n-1 (i) and LSP ' _n (i) indicate the quantized value of the LSP in the past by one frame and the quantized value of the LSP in the current frame (the output value of the vector quantization unit 110), respectively. A
_j is the j-th code vector included in the interpolation coefficient codebook 140. The interpolation vector quantizer 130
The interpolation coefficient code vector that minimizes the equation (2) is obtained for the subframes ,, and the spectrum parameter in the subframe is restored by the equation (3),
Output from 0.

LSP ′ _n ^l (i) = LSP ′ _{n −1} (i) · A _j − (1−A _j ) LSP ′ (i) (3) A _j is ( ^L ) for the training LSP signal, In order to minimize the equation (4), learning is performed in advance using the above-mentioned document 5 and the like.

[0021]

[Equation 3]

Here, n is a frame number, and k is the number of frames of all training signals.

Next, the operation of the second invention will be described with reference to FIG.

The vector quantizer 110 uses the codebook 120 (first codebook) to output a plurality of types (M types) of code vectors in ascending order of the quantization distortion of equation (1). Next, for each of a plurality of types of code vectors, the interpolation vector quantization unit 130 uses the interpolation coefficient codebook 140 (second codebook),
A code vector that minimizes the quantization distortion of equation (2) is output. Next, the interpolation vector quantization unit 130 calculates the cumulative distortion of Expression (5) over a plurality of subframes for each of the combinations of M types of code vectors, and a first codebook that minimizes the cumulative distortion. And a code vector combination of the second codebook are output.

[0025]

[Equation 4]

Here, LSP'nk (i) is the k-th (k = 1 to M) candidate selected from the first codebook. L is the number of subframes in which distortion is accumulated.

Next, the operation of the third invention will be described with reference to FIG. Here, the components denoted by the same reference numerals as those in FIG. 1 perform the same operations as those in FIG.

The difference from the first invention is that the subframe is interpolated by a predetermined interpolation method using the quantized value of the spectrum parameter of the current frame and the quantized value of the spectrum parameter of the past frame. , ,, and the second codebook is selected so as to reduce the error between the spectrum parameter of each subframe and the above-mentioned restored value.

In FIG. 3, the interpolation unit 115 uses a quantized value of the spectrum parameter of one frame past and an output value of the vector quantization unit 110 (quantized value of the spectrum parameter of the current frame) in advance. The spectral parameters of the subframe are restored by the method described above. Here, as an example, it is assumed that restoration is performed by linear interpolation according to the equation (6).

LSP ′ ^l (i) = α ^l LSP ′ _{n −1} (i) + (1−α ^l ) LSP ′ _n (i) (6) where LSP ′ ^l (1) is restored by linear interpolation. It is the LSP of the 1st sub-frame. Further, α ¹ is a predetermined interpolation coefficient in the 1st sub-frame.

The interpolation vector quantizer 135 uses the interpolation coefficient codebook 145 (second codebook) to
The code vector A _{j is} selected for each subframe so as to minimize the error power of the following equation.

[0032]

[Equation 5]

Then, according to the equation (8), the subframe,
, And the restored value is output from the terminal 150.

LSP ″ ^l (i) = LSP ′ ^l (i) + LSP ′ _{n −1} (i) · A _j + (1 −A _j ) LSP ′ _n (i) (8) Next, a fourth invention The action of will be described with reference to FIG.

The vector quantizer 110 outputs a plurality of types (M types) of code vectors in the ascending order of the quantization distortion of the equation (1) using the codebook 120 (first codebook). Next, each of a plurality of types of code vectors is interpolated by the interpolation unit 115, and further, the interpolation vector quantization unit 135 uses the interpolation coefficient codebook 145 (second codebook) to obtain (7). The code vector that minimizes the quantization distortion of the expression is output. next,
For each of the M types of code vector combinations, the cumulative distortion over a plurality of subframes of the following formula is calculated, and the combination of the code vectors of the first codebook and the second codebook that minimizes the cumulative distortion is output. To do.

[0036]

[Equation 6]

Here, LSP ″ ^l _k (i) is the LSP of the l subframe restored using the k-th (k = 1 to M) candidate selected from the first codebook.

In the examples of the operations of the first to fourth inventions, the common interpolation coefficient codebook is constructed for all subframes, but different interpolation coefficient codebooks may be constructed for each subframe. Also, in this case, the interpolation codebooks for a plurality of subframes may be put together to form a matrix codebook.

Further, other known distance scales such as the squared distance and the weighted distance may be used for learning and searching the interpolation coefficient codebook.

The vector quantization section may use well-known vector scalar quantization.

The operation of the fifth invention will be described with reference to FIG. Here, the constituent elements denoted by the same numbers as those in FIG. 1 and FIG. 3 perform the same operations as those in FIG. 1 and FIG.

The vector quantizer 180 uses the codebook 120 (first codebook) to output a plurality of types (M types) of code vectors in ascending order of the quantization distortion of the equation (1). Next, the interpolation unit 115 performs an interpolation process on each of the plurality of types of code vectors using, for example, linear interpolation, and restores the LSP of the subframe according to the equation (6). In the cumulative distortion calculation unit 160,
For each candidate, the error power between the spectral parameter of the subframe and the above-mentioned restored value is accumulated over a plurality of subframes, and the cumulative distortion of equation (10) is calculated.

[0043]

[Equation 7]

The decision unit 165 selects the code vector that minimizes the equation (10), and outputs the candidate of the vector quantization unit from the terminal 170 and the restored value of the subframe from the terminal 170.

In the above description, the vector quantizer may use well-known vector-scalar quantization. A known method can be used for the interpolation processing. Also, other well-known distance measures such as weighted distances can be used to search the codebook.

[0046]

FIG. 5 is a block diagram showing an embodiment of the speech parameter coding system of the first, second, third, fourth and fifth inventions.

In FIG. 5, an audio signal is input from the input terminal 400, and an audio signal for one frame (for example, 40 ms) is stored in the buffer memory 410.

The subframe division circuit 420 divides the audio signal of the frame into a predetermined subframe (for example, 8 m).
s).

The LPC analysis circuit 430 includes at least one sub-frame position (eg ,, or, or, in FIG. 2).
(Any arbitrary combination is conceivable) and LSP parameters are calculated by a well-known LPC analysis as parameters representing the spectral characteristics of the audio signal. For the specific calculation method, the above-mentioned Document 4 can be referred to.

The LSP quantization circuit 440 quantizes the LSP parameter obtained in one subframe with a predetermined number of quantization bits, and outputs the obtained code 1 _k as an output terminal 450.
Output from.

First, the configuration of the LSP quantization circuit 440 according to the first invention will be described with reference to FIG.

In FIG. 6, the LSP parameter of the subframe is input from the input terminal 500. Also,
, The spectral parameters obtained for the subframes are sequentially input from the terminal 505, and the buffer memory 506
To store.

The vector quantization circuit 510 uses the codebook 520 to vector quantize the input LSP parameter by a well-known means. Codebook 520
Is preliminarily learned and configured using a large amount of LSP parameter sequences for training. Vector quantization circuit 510
Outputs the code vector that minimizes the equation (1) to the interpolation vector quantization unit 530 and the terminal 555.

The interpolation vector quantization circuit 530 inputs the output code vector of the vector quantization circuit 510 and the quantized value of the spectrum parameter one frame past by the delay circuit 545. The interpolation coefficient codebook 540 is used to interpolate and restore the LSP sequence of the subframes stored in the buffer memory 506. That is, the quantization distortion by each interpolation coefficient code vector is calculated by the equation (2), the interpolation coefficient code vector A _j that minimizes the quantization distortion is selected for each subframe, and the LSP parameter is set by the equation (3). Restore and buffer memory 535
It is stored in the terminal once and then output to the terminal 550.

The interpolation coefficient of the codebook 540 can be constructed by learning in advance by using the above-mentioned literature 5 or the like so as to minimize the equation (4) for the training LSP signal.

Next, the configuration of the LSP quantization circuit 440 in the second invention will be described with reference to FIG. In FIG. 7, the components denoted by the same numbers as in FIG. 6 perform the same operations as in FIG.

In FIG. 7, the vector quantization circuit 600 is shown.
Is a plurality of types (M types) in the ascending order of the quantization distortion of Expression (1) using the codebook 520 (first codebook).
The code vector of is output and stored in the candidate storage circuit 610. The interpolation vector quantization circuit 615 inputs each candidate from the candidate storage circuit 610, and further inputs the vector quantization value of the LSP of one frame past from the delay circuit 545,
L of each subframe stored in the buffer memory 506
Quantize SP by interpolation vector. For interpolation vector quantization, interpolation coefficient codebook 540 (second codebook)
Is used to output a code vector that minimizes the quantization distortion of equation (2). The cumulative distortion calculation circuit 620 calculates the cumulative distortion of Expression (5) for each candidate. In equation (5), L = 3. The selection circuit 625 vector-quantizes the LSP using a combination of code vectors of the first codebook and the second codebook that minimizes cumulative distortion.
Is restored and stored in the buffer memory 535, and then the terminal 5
Output from 55. The delay circuit 545 delays the LSP quantized value in the fourth subframe by one frame.

Next, the configuration of the LSP quantization circuit 440 according to the third invention will be described with reference to FIG. In FIG. 8, the components denoted by the same numbers as in FIG. 6 perform the same operations as in FIG.

In FIG. 8, the interpolation circuit 630 outputs the output value of the delay circuit 545 (the vector quantization value of the LSP of the past frame) and the output value of the vector quantization circuit 510 (the vector quantization of the LSP of the current frame). Value) and are used to restore the LSP of each subframe by a predetermined interpolation method. Various methods are known as the interpolation method, but here, as an example, it is assumed that the interpolation is performed by linear interpolation according to the equation (6).

The interpolation vector quantizer 635 uses the interpolation coefficient codebook 645 (second codebook) to
A code vector is selected for each subframe so as to minimize the error power of the equation (7). Then, according to the equation (8), the values restored for the subframes, ... Are stored in the buffer memory 535 and then output from the terminal 555.

The delay circuit 545 is used for the L of the fourth sub-frame.
The quantized value of SP is delayed by one frame.

Next, the configuration of the LSP quantization circuit 440 according to the fourth aspect of the invention will be described with reference to FIG. In FIG. 9, the components having the same numbers as those in FIG. 7 perform the same operations as those in FIG.

In FIG. 9, the vector quantization circuit 600 is shown.
Is a plurality of types (M types) in the ascending order of the quantization distortion of Expression (1) using the codebook 520 (first codebook).
The code vector of is output and stored in the candidate storage circuit 610. Interpolation circuit 630 interpolates each of a plurality of types of code vectors, and interpolation vector quantization circuit 635 further performs interpolation coefficient codebook 645.
(Second codebook) is used to output a code vector that minimizes the quantization distortion of equation (7). The cumulative distortion calculation circuit 620 calculates the cumulative distortion of Expression (9) for each of the M types of candidates. Here, L = 3. The selection circuit 625 finds a combination of code vectors of the first codebook and the second codebook that minimizes the cumulative distortion, restores the LSP at this time, stores it in the buffer memory 535, and then outputs it from the terminal 555. . The delay circuit 545 is
The LSP quantized in the fourth subframe is delayed by one frame.

The interpolation coefficient codebook may be commonly used for a plurality of types of subframes. Alternatively, an optimal interpolation coefficient codebook may be used for each subframe. In the latter case, the number of bits of the interpolation coefficient codebook can be further reduced by forming a codebook having a matrix configuration in which a plurality of subframes are collected. The method of creating the matrix codebook is described in, for example, C.I. "Matrix quantizer design" by Tsao et al.
n for LPC speech usinging
generalized Lloyd algori
thm, ”(IEEE Trans. AS
SP, pp. 537-545, 1985) (Reference 6)
Can be referred to.

Further, other well-known distance scales such as the squared distance and the weighted distance can be used for learning and searching the interpolation coefficient codebook.

Further, the vector quantizer can use well-known vector-scalar quantization.

Next, the configuration of the LSP quantization circuit 440 according to the fifth invention will be described with reference to FIG. In FIG. 10, the components designated by the same numbers as those in FIG. 6 perform the same operations as those in FIG.

In FIG. 10, the vector quantization circuit 60
0 outputs a plurality of types (M types) of code vectors in the order of decreasing quantization distortion of Expression (1) using the codebook 520 (first codebook), and stores the code vectors in the candidate storage circuit 610. The interpolation circuit 630 performs interpolation processing on each of a plurality of types of code vectors, and according to equation (6),
,, restore the LSP of the subframe. here,
Although various methods are conceivable for the interpolation processing, linear interpolation is used as an example. The cumulative distortion calculation circuit 620
For each candidate, the error power between the LSP of the subframe and the LSP restoration value obtained by the above-described interpolation processing is accumulated over a plurality of subframes, and the cumulative distortion of Expression (10) is calculated. Here, L = 3. The selection circuit 625 selects a code vector candidate that minimizes the expression (10), restores the LSP for each subframe, stores the LSP in the buffer memory 535, and then outputs the LSP from the terminal 555. The delay circuit 545 delays one candidate frame selected in the fourth subframe.

Finally, the above description will be supplemented.

First, the vector quantization circuit may use well-known vector-scalar quantization.

The interpolation circuit 630 has a plurality of interpolation positions or interpolation methods, calculates cumulative distortion for all combinations thereof, and uses interpolation positions or interpolation methods that minimize the cumulative distortion as auxiliary information for each frame. It may be transmitted with several bits. However, although this improves performance, the number of bits is slightly increased due to auxiliary information transmission.

Further, the interpolation in the interpolation circuit 630 may use other well-known methods such as logarithmic interpolation in addition to linear interpolation. Further, the optimum interpolation coefficient for each subframe may be learned by using a training signal.

Although the square distance of the LSP is used for the vector quantization of the LSP and the calculation of the cumulative distortion, other known distance measures may be used. For example, the weighted square distance of the LSP, the LPC cepstrum distance, or the like.

Although the full search type vector quantization circuit is used as the vector quantization circuit, in order to reduce the amount of calculation required to search the code vector, a tree search type, a lattice type, another stage type, or You may use the vector quantization circuit of another well-known structure. For details of these calculation amount reduction methods, see, for example, R.M. "Vector qu by Gray
paper entitled "Antiation," (IEEE A
SSP Magazine, pp. 4-29, 1984
Year) (Reference 7), etc. can be referred to.

Although LPC analysis was performed on the input speech for four subframes to calculate LSP coefficients, any method can be used to reduce the amount of calculation. For example, the LSP is calculated only for two or three subframes (when there are two, for example, or and, or and or in FIG. 2), and for other subframes, these LSPs are interpolated. The LSP may be obtained. Or L for 2 or 3 subframes
The SP may be calculated, the LSP may be quantized in one subframe, and the LSP in the other subframe may be represented by interpolation or a second codebook or a combination of the interpolation and the second codebook.

[0076]

As described above, according to the present invention, when quantizing the spectrum parameter representing the spectrum characteristic of speech, the frame is divided into subframes shorter than that, and at least one subframe is used. The spectrum parameter is obtained and quantized using the first codebook, and the spectrum parameter of another subframe is represented by an interpolation process of the quantized value of the frame and the quantized value of the past frame, or It is represented using a codebook, or using an interpolation process and a second codebook. Also, in the search of the first codebook, a plurality of types of candidates are obtained in the order of smaller quantization distortion, and interpolation processing, or a second codebook, or interpolation processing and a second codebook is used for each candidate. And restore the spectral parameters to find the cumulative distortion over multiple subframes,
The candidates are selected so as to reduce the cumulative distortion. Therefore, even if the frame length is increased to reduce the bit race, there is a great effect that the spectrum parameter can be quantized well with a small number of bits and the amount of calculation, and that the temporal change can be satisfactorily expressed.

[Brief description of drawings]

FIG. 1 is a block diagram for explaining the operation of a speech parameter coding system of the first and second inventions.

FIG. 2 is a diagram showing a relationship between frames and subframes.

FIG. 3 is a block diagram for explaining the operation of the speech parameter coding system of the third and fourth inventions.

FIG. 4 is a block diagram for explaining the operation of the speech parameter coding system of the fifth invention.

FIG. 5 is a block diagram showing an embodiment of speech parameter coding systems of the first, second, third, fourth and fifth inventions.

FIG. 6 is a block diagram showing a configuration of an LPC quantization circuit according to the first invention.

FIG. 7 is a block diagram showing a configuration of an LPC quantization circuit according to a second invention.

FIG. 8 is a block diagram showing a configuration of an LPC quantization circuit according to a third invention.

FIG. 9 is a block diagram showing a configuration of an LPC quantization circuit according to a fourth invention.

FIG. 10 is a block diagram showing a configuration of an LPC quantization circuit according to a fifth aspect of the invention.

[Explanation of symbols]

110 Vector quantizer 120 codebooks 130 Interpolation vector quantizer 140 Interpolation Coefficient Codebook 190 Delay unit 115 Interpolator 135 Interpolation vector quantizer 145 Interpolation coefficient codebook 180 Vector quantizer 160 Cumulative distortion calculator 165 Discriminator 410 buffer memory 420 subframe division circuit 430 LPC analysis circuit 440 LPC quantization circuit 506 buffer memory 510 Vector quantizer 520 codebook 530 Interpolation vector quantization circuit 535 buffer memory 540 Interpolation Coefficient Codebook 545 delay circuit 600 vector quantization circuit 610 Interpolation storage circuit 615 Interpolation vector quantization circuit 620 Cumulative distortion calculation circuit 625 selection circuit 630 Interpolation circuit 635 Interpolation vector quantization circuit 645 Interpolation Coefficient Codebook

Claims (5)

[Claims] 1. Dividing an input audio signal into frames,
Further, means for dividing into a plurality of subframes shorter than the frame, means for obtaining a spectral parameter for at least one of the subframes by a predetermined order for the audio signal, and spectral parameter for one subframe Is quantized using a first codebook configured in advance, and the spectral parameter of at least one other subframe is quantized with respect to the spectral parameter quantized in the frame and the spectral parameter quantized in the past frame. And a means for quantizing using a second codebook, a speech parameter coding method. 2. In at least one of the first codebook and the second codebook, a plurality of types of candidates are output in ascending order of quantization distortion, and in other subframes, spectrum parameters are set to the spectrum in the frame. Quantized values of parameters, quantized values of spectrum parameters in past frames, and a second codebook to obtain quantization distortion, and the quantization distortion is accumulated over a plurality of subframes, and The speech parameter coding system according to claim 1, wherein a candidate for minimizing a cumulative value is selected and output. 3. The input audio signal is divided into frames,
Further, means for dividing into a plurality of subframes shorter than the frame, means for obtaining a spectral parameter for at least one of the subframes by a predetermined order for the audio signal, and spectral parameter for one subframe Is quantized using a first codebook configured in advance, and in at least one other subframe, the quantized value of the spectral parameter in the frame and the quantized value of the spectral parameter in the past frame are used in advance. A speech parameter coding method, wherein a spectrum parameter is restored by a predetermined method, and a second codebook is selected so as to reduce an error between the spectrum parameter of the subframe and the restored value. 4. At least one of the first codebook and the second codebook outputs a plurality of types of candidates in ascending order of quantization distortion, and in other subframes, the spectrum parameter of the subframe and the candidate are output. Accumulate error power with the restored value over multiple subframes,
The speech parameter coding method according to claim 2, wherein a candidate for minimizing the cumulative value is selected and output. 5. The input voice signal is divided into frames,
Further, means for dividing into a plurality of subframes shorter than the frame, means for obtaining a spectral parameter for at least one of the subframes by a predetermined order for the audio signal, and spectral parameter for one subframe Are quantized using a preconfigured first codebook, and a plurality of types of candidates are output in the order of smaller quantization distortion. The spectrum parameter is restored by a predetermined method using the spectrum parameter, the error power between the spectrum parameter of the subframe and the restored value is accumulated over a plurality of subframes, and a candidate that minimizes the accumulated value is selected and output. And a voice parameter characterized by Goka system.