JP3335650B2 - Audio coding method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech encoding system for encoding a speech signal at a low bit rate, particularly at a high quality of 8 kb / s or less.

[0002]

2. Description of the Related Art As a method of encoding a speech signal at a low bit rate of 8 kb / s or less, for example, M. Sch
roeder and B.R. "Cod by Atal
e-excited linear predictic
on: High quality speech at
very low bit rates ”(Pro
c. ICASSP, pp. 937-940, 1985
), And "Improved speech quality" by Kleijn et al.
and efficient vector qua
ntization in SELP "(Proc. IC
ASSP, pp. 155-158, 1988) (Reference 2).
eExcited LPC Coding) is known. In this method, the transmitting side extracts a spectrum parameter representing a spectrum characteristic of a voice signal from a voice signal for each frame (for example, 20 ms), further divides the frame into subframes (for example, 5 ms), and The parameters (delay parameter and gain parameter) in the adaptive codebook are extracted based on the sound source signal, the pitch of the speech signal of the subframe is predicted by the adaptive codebook, and the residual signal obtained by pitch prediction is determined in advance. An optimal excitation code vector is selected from an excitation codebook including a predetermined type of noise signal, and an optimal gain is calculated. The method of selecting the excitation code vector is to select one type of noise signal so as to minimize the error power between the signal synthesized from the selected noise signal and the residual signal. Then, an index and a gain representing the type of the selected code vector, the spectrum parameters and the parameters of the adaptive codebook are transmitted. Description on the receiving side is omitted.

[0003]

In the conventional methods described in the above-mentioned references 1 and 2, the size of the sound source codebook needs to be sufficiently large (for example, 10 bits) in order to obtain good sound quality. Therefore, an enormous amount of computation was required to search for the sound source codebook. Furthermore, the amount of memory is enormous (for example, in the case of 10 bits and 40 dimensions, the amount of memory is 40 k words), and it has been difficult to realize compact hardware. Further, when the frame length and the sub-frame length are increased to reduce the bit rate, and the number of dimensions is increased without reducing the number of bits of the codebook, there is a problem that the amount of calculation increases significantly.

[0004] As a method of reducing the size of the code book, for example, B.I. "Multipl by Juang et al.
e stage vector quantizati
onfor speech coding ”(Pro
c. ICASSP, pp. 597-600, 1982
As described in a paper entitled “Year” (Reference 3) and the like, a multi-stage vector quantization method in which a code book is divided into multiple stages and configured to search each code book independently is known. In this method, since the codebook is divided into a plurality of stages, the size of the codebook per stage is reduced to, for example, B / L bits (where B is the total number of bits and L is the number of stages). , The amount of computation required for codebook search is L × 2 ^{B / L} compared to one B-bit stage in the entire L stages.
To be reduced. Also, the amount of memory required to store the codebook is similarly reduced. However, in this method, since the codebook of each stage is independently learned and searched, there is a problem that the performance is greatly reduced as compared with one stage of B bits.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems and to provide a relatively small amount of calculation and a small amount of memory, thereby achieving an 8 kb / s
It is an object of the present invention to provide a speech coding system with good sound quality at s or less, especially at 4.8 kb / s or less.

[0006]

[0007]

According to the present invention, an input audio signal is divided into frames of a predetermined time length, spectrum parameters representing a spectrum envelope of the audio signal are obtained, and the frame is determined in advance. Dividing into time-length subframes, obtaining parameters in the adaptive codebook based on past excitation signals for each subframe, searching for an optimal code vector from a preconfigured excitation codebook, In the speech coding method for quantizing the sound source codebook, the excitation codebook is configured by a multistage cascade connection of a plurality of codebooks whose dimensionality is shorter than the subframe length, and at least one stage has a code vector in the order of smaller quantization distortion. Output a plurality of candidates, accumulate distortion over the entire subframe, and output a set of candidates for minimizing the accumulated distortion. The audio signal, wherein the quantizing Ri.

[0008]

The operation of the speech coding system according to the present invention will be described.

FIG. 1 is a block diagram showing the configuration of the invention related to the present invention . In the figure, an audio signal is divided into frames (for example, 30 ms), and an LPC parameter is obtained as a spectrum parameter by a known LPC analysis in a spectrum parameter calculation unit 150. In addition, the audio of the frame is converted to a sub-frame
Divided into

Next, adaptive codebook section 160 applies a delay T, as a parameter to the audio signal for each subframe as a parameter based on the excitation signal calculated in the past subframe.
Find the gain β. The method of calculating the parameters of the adaptive codebook can be referred to the above-mentioned references 1, 2 and the like. Then, the prediction value by the adaptive codebook is subtracted by the following equation.

[0011]

(Equation 1)

A feature of the invention related to the present invention is that the sound source codebook is constituted by a plurality of cascade connections of a plurality of different codebooks, and at least one stage outputs a plurality of codebook candidates in ascending order of distortion. It is in. Hereinafter, for simplicity, it is assumed that a sound source codebook is configured by cascade connection of two-stage codebooks. The sound source codebook search unit 170 uses the codebooks 180 ₁ and 180 ₂ that have been learned in advance to construct e _w
A candidate for a code vector representing (n) is output. This operation will be described below. Hereinafter, it is assumed that M types of code vector candidates are output in the first stage. First, in the first stage, the quantization distortion is obtained using the codebook 1 by the following equation.

[0013]

(Equation 2)

Here, γ _1j , c _1j (n), h _w (n), N
Are the optimal gain, the j-th code vector of codebook 1 (j = 1 to 2 ^B1 : B1 is the number of bits of codebook 1), the perceptual weighting impulse response obtained from the spectral parameters, and the number of samples representing the subframe length. It is. Here, the method of obtaining the auditory weighting impulse response can be referred to the above-mentioned documents 1 and 2. The symbol * indicates convolution integration. N is the number of samples indicating the subframe length.
In the first stage, a plurality of types (for example, M types) of code vectors are obtained in ascending order of the quantization distortion of Expression 2, and the quantization distortion of each candidate at this time is stored. An error signal is calculated for each candidate by the following equation.

[0015]

(Equation 3)

Here, e _w ′ _k (n) indicates an error signal calculated for the k-th candidate.

[0017] In the next second step, determine the optimum code vector minimizing the following equation with respect to each candidate of the first stage using a codebook 180 _2, the quantization distortion at that time, candidates for the corresponding first stage To obtain the cumulative distortion.

[0018]

(Equation 4)

Here, l = 1 to 2 ^B2 : B2 indicates the number of bits of the codebook 2. B = B1 + B2: B is 2
This is the number of bits in the entire codebook of the column.

Next, a combination of first-stage and second-stage candidates for minimizing the cumulative distortion is selected, and the selected code vector is output.

As a measure for selecting a combination of candidates, a combination of candidates that minimizes the second-stage quantization distortion shown in Expression 4 may be selected in addition to the cumulative distortion. The number of cascade connections can be set arbitrarily.

The codebooks 1 and 2 may be composed of Gaussian random number signals as in the above-mentioned documents 1 and 2, or may be constructed by learning using a large amount of training signals in advance. For the latter configuration method, reference can be made, for example, to Japanese Patent Application No. 2-42956 (Document 4). This concludes the description of the operation of the invention relating to the present invention .

In order to further improve the performance, the codebook at each stage may be learned so as to be optimal for the entire codebook under the condition that a plurality of types of candidates are output.

Next, the operation of the present invention will be described with reference to FIG. In FIG. 2, components denoted by the same reference numerals as those in FIG. 1 perform the same operations, and thus description thereof will be omitted. In the present invention, the structure of the sound source codebook and the search method are related to the present invention.
Invention differs so that this point will be described. The number of dimensions of the sound source codebook is shorter than the subframe length.
In the following, as an example, the number of dimensions is 1/1 / subframe length N.
2 N / 2. Therefore, the number of dimensions of the codebook 210 ₁ and the codebook 210 ₂ is N / 2. In addition, the number of stages of the cascade connection is set to 2 as an example below.

First, for the first half of the N / 2 samples of the subframe, quantization distortion is obtained by the following equation using the codebook 1 in the first stage.

[0026]

(Equation 5)

Here, γ _1j , c _1j (n) and h _w (n) are the optimum gain and the j-th code vector of codebook 1 (j = 1 to 2 ^B1 : B1 is the number of bits of codebook 1). ), The auditory weighting impulse response obtained from the spectrum parameter. In the first stage, a plurality of types (for example, M types) of code vectors are obtained in ascending order of the quantization distortion of Expression 5, and the quantization distortion of each candidate at this time is stored. Also,
An error signal is calculated for each candidate by the following equation.

[0028]

(Equation 6)

Next, in the second stage, an optimal code vector for minimizing the following equation is determined for each candidate in the first stage using the code book 210 _2, and the quantization distortion at that time is determined by the corresponding candidate in the first stage. To obtain the cumulative distortion.

[0030]

(Equation 7)

Here, l = 1 to 2 ^B2 : B2 indicates the number of bits of the codebook 2. B / 2 = B1 + B2: B
Is the number of bits in the entire subframe.

Next, for the N / 2 samples in the latter half of the subframe, the following equation is calculated for each code vector candidate determined above.

[0033]

(Equation 8)

Here, s _wk (n) is an influence signal which affects the k-th candidate of the code vector in the latter half of the subframe.

Next, M kinds of candidates are output in ascending order of the quantization distortion in the following equation using the codebook 1, and the quantization distortion at that time is accumulated.

[0036]

(Equation 9)

An error signal is calculated for each candidate by the following equation.

[0038]

(Equation 10)

Next, in the second stage, an optimal code vector for minimizing the following equation is obtained for each of the first-stage candidates using the codebook 210 _2, and the quantization distortion at that time is determined by the corresponding first-stage candidate. To obtain the cumulative distortion.

[0040]

[Equation 11]

Next, the cumulative distortion in the entire subframe is obtained, a combination of candidates in the entire subframe that minimizes the cumulative distortion is selected, and the selected code vector is output. In the above description, the same codebook is used in the first half and the second half of the subframe. However, different codebooks can be used in the first half and the second half.

The operation of the present invention has been described above.

[0043]

FIG. 3 is a block diagram showing an example of a speech coding apparatus for implementing a speech coding method according to the present invention .

In the figure, on the transmitting side, an input terminal 400
And input an audio signal from one frame (for example, 30 m
The audio signal of s) is stored in the buffer memory 410.

The LPC analysis circuit 430 uses LSP as a parameter representing the spectral characteristic of the audio signal of the frame.
The parameters are calculated from the audio signal of the frame using a well-known LPC.
The analysis is performed and a predetermined order P is calculated. The specific calculation method is described in “Quantizer design in LSPs” by Sugamura et al.
"peech analysis-synthesis"
And the like (Reference 5).

Next, the LSP quantization circuit 440 quantizes the LSP parameter with a predetermined number of quantization bits,
The obtained code l _k is output to the multiplexer 560, which is decoded to further obtain the linear prediction coefficient a _i ′ (i =
1 to P) and output to the weighting circuit 500 and the impulse response calculation circuit 470. For the method of encoding the LSP parameter and converting the LSP parameter into the linear prediction coefficient, reference can be made to the above-mentioned reference 5. Also, LSP
May be used to more efficiently encode the vector-scalar quantization or the vector quantization. As a specific method of the former, for example, “T
transform coding ofspeech
using a weighted vector q
A paper entitled "Uniqueizer" (IEEE J. Se
l. Areas Commun. Pp. 425-43
1, 1988) (Document 6).

The sub-frame division circuit 450 divides the audio signal of the frame into sub-frames. Here, for example, the subframe length is set to 7.5 ms.

The weighting circuit 500 weights the perceptual weight of the signal divided into subframes by using the method described in the above-mentioned documents 1 and 2.

The subtracter 490 subtracts the output of the synthesis filter 581 from the signal weighted by the auditory sense and outputs the result.

The impulse response calculation circuit 470 calculates the impulse response h _w (n) of the synthesis filter weighted by the auditory sense.
Is calculated for a predetermined number L of samples. The specific calculation method can be referred to the aforementioned documents 1, 2, 3, and the like.

The adaptive codebook 510 receives the input signal v (n) of the synthesis filter 581, and receives a weighted impulse response h from the impulse response output circuit 470.
_w (n), a weighting signal is input from the subtractor 490, pitch prediction is performed based on long-term correlation, and delay M and gain β are calculated as pitch parameters. The order of the adaptive codebook is 1. The method of calculating the delay M and the gain β in the first-order adaptive codebook can be referred to the above-mentioned reference 2.

Further, the residual signal is calculated and output according to equation (1). In Expression 1, v (n) is a past sound source signal, which is an input signal of the synthesis filter 581. h _w (n) is a weighted impulse response obtained by the impulse response calculation circuit 470.

Here, it is assumed that the sound source codebook can be represented by a two-stage cascade connection of two different codebooks 540 ₁ and 540 ₂ . FIG. 4 shows the operation of the sound source codebook search circuit 530. In FIG. 4, the quantization distortion calculation circuit 605 of Equation 2 receives the signal e _w (n) from the terminal 600 and the weighted impulse response h _w (n) from the terminal 602.
Is input, and the quantization distortion of Expression 2 is calculated using the codebook 540 ₁ . Here, Equation 2 may be directly calculated,
Equation 2 may be modified and calculated by a correlation operation. For the latter method, reference can be made to the aforementioned reference 4.

Next, the M candidate selection circuit 610 selects M candidate code vectors in ascending order of the quantization distortion of Expression 2 and outputs them to the subtractor 615, and outputs the quantization distortion of each candidate to the cumulative distortion calculation circuit 625. I do.

The subtractor 615 calculates the error signal e ′ _kw (n) for each candidate by using _equation (3).

The quantization distortion calculation circuit 620 of Expression 4 finds an optimal code vector for the error signal for each candidate by using Expression 4 using Codebook 2, and accumulates the quantization distortion at this time. Output to the distortion calculation circuit 625.

The cumulative distortion calculation circuit 625 selects and outputs a combination of quantization distortion accumulated in the entire two stages or a combination of code vector candidates for minimizing the second stage quantization distortion.

The sound source code book may be composed of Gaussian positive random number signals (random number code book), or may be constructed by previously learning training signals (learning code book), as described in the section of operation. May be. Alternatively, the learning code book and the random number code book may be used.

Returning to FIG. 3, gain code book 55
0 is the adaptive codebook 510 and the sound source codebook 53
The vector of the gain of 0 is quantized. Here, in the gain codebook 555, a gain code vector that minimizes the following equation is selected and output.

[0060]

(Equation 12)

Here, β ′, γ ₁ ′, and γ ₂ ′ are the gain of the adaptive codebook, the gain of the codebook 1,
It is a gain code vector corresponding to the gain of codebook 2. The gain codebook 550 calculates a sound source signal by the following equation, and outputs it to the synthesis filter 581.

[0062]

(Equation 13)

The synthesis filter 581 receives v (n), obtains perceptually weighted synthesized speech for one subframe by the following equation, and inputs a sequence of 0 to the filter for the other subframe to obtain a weighted response signal sequence. And outputs a weighted response signal sequence for one subframe to the subtractor 490.

[0064]

[Equation 14]

However,

[0066]

(Equation 15)

Here, δ is a coefficient for determining the degree of hearing weighting, and is selected to be 0 <γ <1.

Multiplexer 560 combines and outputs the output code sequences of LSP quantizer 440, adaptive codebook 510, excitation codebook search circuit 530, and gain codebook 550.

This concludes the description of examples of the present invention related to the present invention .

In the present invention, the sound source codebook search circuit and the operations of the codebook 1 and the codebook 2 correspond to the present invention.
Since these are different from the related invention, these will be described.
FIG. 5 is a block diagram illustrating the configuration of the sound source codebook search circuit according to the present invention. The number of dimensions of the sound source codebook is shorter than the subframe length. In the following, as an example, the number of dimensions is set to N / 2, which is 1/2 of the subframe length N. Therefore, the codebook 740 ₁ and the codebook 74
The number of dimensions of O ₂ is N / 2. In addition, the number of stages of the cascade connection is set to 2 as an example below. In the figure, terminal 700
From the first half of the subframe.
At this time, the switch 703 is moved down. The quantization calculation circuit 705 of Expression 5 calculates quantization distortion by Expression 5 using Codebook 1. The M candidate selection circuit 710 obtains a plurality of types (for example, M types) of code vectors in ascending order of the quantization distortion, and outputs the quantization distortion of each candidate at this time to the cumulative distortion calculation circuit 725.

The subtractor 715 calculates an error signal for each candidate according to equation (6).

Next, the number 7 quantization distortion calculation circuit 720, and with reference to codebook 740 _2, selects a code vector that minimizes the number 7 for each candidate, accumulated quantization distortion at this time Output to the distortion calculation circuit 725.

The cumulative distortion calculation circuit 725 calculates the quantization distortion accumulated in two stages for each candidate.

Next, the subtractor 704 receives the N / 2 samples of the latter half of the subframe from the terminal 701, and calculates Equation 8 for each of the code vector candidates obtained above. The switch 703 is tilted upward, and the quantization distortion calculation circuit 705 and M
The candidate selection circuit 710, the quantization distortion calculation circuit 720, and the cumulative distortion calculation circuit 725 repeat the above operation, and calculate the cumulative distortion of the entire subframe in the cumulative distortion calculation circuit 725, and calculate the code vector of the code vector for minimizing the cumulative distortion. The combination is determined and output from terminal 730.

The sound source code book may be composed of Gaussian random number signals (random number code book), or may be constructed by previously learning training signals (learning code book). May be. Alternatively, the learning code book and the random number code book may be used.

The description of the present invention has been completed.

The number of cascade connections of the sound source code book is 2
Any number of stages above the stage can be used.

The learning of the sound source codebook may be performed for each stage using a training signal, or may be performed so as to simultaneously optimize and minimize distortion in a plurality of stages. Good. Further, learning may be performed in consideration of outputting a plurality of types of candidates in each stage. For example, in the case of the second stage, after the first stage is optimized, the second stage is optimized together with the first stage, and these optimizations are repeated until the distortion reduction is saturated.

The codebooks 1 and 2 may be full search type codebooks, or may be tree search type codebooks in order to further reduce the amount of calculation required for search. A method of constructing a tree search type codebook is described in, for example, R. Gray
A paper entitled "Vector quantification" (IEEE ASSP Magazine, pp. 146-64).
4-29, 1984) (Reference 7).

In the embodiment of the present invention, in the first half of the subframe, a plurality of types of candidates may be selected in the code vector selection in the codebook 2 to calculate the cumulative distortion. Further, different codebooks may be used for the first half and the second half of the subframe.

In the above embodiment, the adaptive codebook gain and the first and second codebook gains were not simultaneously optimized. However, the adaptive codebook, the first codebook, and the second codebook were not optimized simultaneously. By performing simultaneous optimization on the book gain, the characteristics are further improved. This synchronization optimization is very effective when applied when selecting code vectors of codebook 1 and codebook 2. As a method of the simultaneous optimization, for example, after finding the delay and the gain β of the adaptive codebook, when searching for the code vector c _1j (n) and the gain γ ₁ of the codebook 1, for each code vector, Β and γ ₁ are simultaneously optimized by solving the following equation to minimize it.

[0082]

(Equation 16)

Next, when the code vector of the code book 2 is selected, the gain of the adaptive code book, the code book 1 and the code book 1 are set so as to minimize the following equation for each code vector.
Simultaneously optimize the codebook 2 gain.

[0084]

[Equation 17]

Further, in order to reduce the amount of calculation, when searching for codebooks 1 and 2, simultaneous optimization of gains is not performed. Simultaneous gain optimization is performed between the adaptive codebook and the adaptive codebook from Equation (16), and then a code vector is selected in the codebook2. A configuration for performing optimization may be adopted.

Further, in order to further reduce the amount of computation, after the code vector set of the codebooks 1 and 2 that minimizes the cumulative distortion is selected in the cumulative distortion calculating circuit, the gain β of the adaptive codebook is selected. And the three kinds of gains γ ₁ , γ ₂ of the _first and _second codebooks can be simultaneously optimized from Expression 17.

As another method for performing the gain synchronization optimization, a code vector is orthogonalized to an adaptive code book at the time of search of the code book 1, and then search is performed.
At the time of searching the codebook 2, a configuration may be adopted in which the codevector is orthogonalized to the adaptive codebook and the codevector selected in the codebook 1, and the search is performed. As the orthogonalization method, for example, I.I. Gerson
"Vector sum excited"
linear prediction (VSELP) s
paper titled "peech coding at 8 kb / s" (Proc. ICASP, pp. 461-46).
4, 1990) (Literature 8).

In the embodiment, the optimum gain is obtained for each codebook of each stage.
One gain may be used, or a codebook of all stages may be collectively used as one gain. By doing so, the number of bits used to transmit the gain can be reduced, so that the bit rate can be further reduced.

In the above embodiment, the K parameter and the LSP parameter are encoded as spectral parameters, and the LPC analysis is used as the analysis method. However, other well-known parameters such as L
PC cepstrum, cepstrum, improved cepstrum,
Generalized cepstrum, mel cepstrum and the like can also be used. In addition, an optimal analysis method can be used for each parameter.

The LPC coefficient obtained for the frame is given by LS
A configuration may be used in which interpolation is performed linearly or non-linearly for each subframe on P or a linear prediction coefficient, and the adaptive codebook, codebooks 1, 2 are searched using the interpolated coefficients. With such a configuration, the sound quality is further improved.

The LSP coefficients can be encoded more efficiently by a vector quantization or a vector-scalar quantization by a well-known method, and furthermore, by a matrix quantization of LSPs of a plurality of subframes collectively. Can be. The method of vector-scalar quantization can be referred to, for example, the above-mentioned reference 4.

In order to reduce the amount of calculation, the weighting circuit 500 may be arranged before the sub-frame division circuit 450, and the synthesis filter 581 may calculate a weighted synthesized signal by the following equation.

[0093]

(Equation 18)

Here, δ is a weighting coefficient for determining the degree of hearing weighting.

In the embodiment, the adaptive codebook parameter is determined to be one before searching the sound source codebook. However, for example, a plurality of types of delay parameter candidates are output, and each of the delay parameter candidates is output. , A combination of the adaptive codebook and the excitation code vector that minimizes the accumulated distortion between the adaptive codebook and the excitation codebook may be selected for each subframe. This increases the amount of computation but improves performance.

To further improve the performance, LS
Instead of uniquely determining at least one of the P coefficient, the parameter of the adaptive codebook, the code vector of the sound source codebook, and the gain codebook in a subframe, a delayed decision (delay determination method) can be adopted. For example, when performing a delayed decision on the parameters of the adaptive codebook and the code vector of the excitation codebook, in the subframe, a plurality of types of candidates are obtained in the adaptive codebook and the excitation codebook, and for each candidate combination, It is also possible to adopt a configuration in which the cumulative distortion is calculated over a plurality of subframes (for example, L subframes), and a combination of an adaptive codebook parameter and an excitation code vector that minimizes the cumulative distortion over the entire L subframes. With such a configuration, the operation amount is further increased, but the performance is further improved.

[0097]

As described above , according to the present invention, since the codebook size can be further reduced, good characteristics can be obtained with a smaller operation amount and memory amount. Therefore, according to the present invention, 8 to
At a bit rate of 4.8 kb / s, there is a great effect that encoded and reproduced speech with good sound quality can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram for explaining an operation of a speech coding system according to the present invention .

FIG. 2 is a block diagram for explaining the operation of the speech coding system according to the present invention.

FIG. 3 is a block diagram showing an example of an apparatus for implementing a speech coding method according to the present invention .

FIG. 4 is a block diagram showing a configuration of a sound source codebook search circuit according to the invention related to the present invention .

FIG. 5 is a block diagram showing a configuration of a sound source codebook search circuit according to the present invention.

[Explanation of symbols]

150 spectral parameter calculation unit 160 the adaptive codebook section 170, 200 excitation codebook searching unit 180 _1, 210 ₁ codebook 1 180 _2, 210 ₂ codebook 2 410 buffer memory 430 LP analysis calculation circuit 440 LSP quantization circuit 450 subframes Dividing circuit 470 Impulse response calculation circuit 490, 615, 704, 715 Subtractor 500 Weighting circuit 510 Adaptive codebook 530 Sound source codebook search circuit 540 ₁ , 740 ₁ Codebook 1 540 ₂ , 740 ₂ Codebook 2 550 Gain codebook 560 Multiplexer 605 Equation 2 quantization distortion calculation circuit 610,710 M candidate selection circuit 620 Equation 4 quantization distortion calculation circuit 625,725 Cumulative distortion calculation circuit 703 Switch 705 Equation 5 quantization distortion calculation circuit 720 Quantization distortion calculation circuit of Expression 7

 ────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-1-205638 (JP, A) Ozawa, Kumagaya "Study on efficient vector quantization of LSP parameters" 1990 IEICE Autumn National Convention Lecture Papers, P1-132

Claims (1)

(57) [Claims] An input audio signal is divided into frames of a predetermined time length, spectrum parameters representing a spectrum envelope of the audio signal are obtained, and the frame is divided into sub-frames of a predetermined time length. A speech coding method for obtaining parameters in an adaptive codebook based on a past excitation signal for each subframe, searching for an optimal code vector from a previously configured excitation codebook, and quantizing the audio signal. In the sound source codebook, the number of dimensions is configured by a multistage cascade connection of a plurality of codebooks shorter than the subframe length, and at least one stage outputs a plurality of candidate code vectors in ascending order of quantization distortion and outputs the entire subframe. Accumulates the distortion in and outputs a set of candidates for minimizing the accumulated distortion to quantize the audio signal. A speech coding method characterized by: