JPH1091193A - Voice coding method and voice decoding method - Google Patents

Abstract

(57) [Summary] [Problem] To provide a speech coding method with a small amount of calculation and high sound quality. An adaptive code vector is extracted from an adaptive code book as a pitch excitation signal, and a pitch response signal is generated by a synthesis filter.
The conversion unit 105 performs conversion on the pitch response signal using the conversion pattern extracted from Step 4 to generate a conversion response signal. An adder 109 synthesizes the pitch response signal and the converted response signal via gain multipliers 106 and 107 to generate a synthesized speech signal 131, and an adaptive code vector that minimizes distortion of the synthesized speech signal 131 with respect to the input speech signal 132 And a conversion pattern is searched from the codebooks 101 and 104, and the coefficients of the synthesis filter 102 and the codebooks 101 and 104 are searched.
And outputs an index indicating the adaptive code vector and the conversion pattern searched from, as encoding parameters.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice coding method and a voice decoding method for compressing and coding a voice signal in a digital telephone or the like.

[0002]

2. Description of the Related Art In recent years, CELP (Code Excited Linear Pr
ediction) method is often used. Regarding the CELP method, see MRSchroeder and BSAtal, “Code Exci
ted Linear Prediction (CELP): High Quality Speech
at Very Low Bit Rates, ”Proc. ICASSP, pp.937-940,1
985 (Reference 1), and WSKleijin, DJKrasinskie
t al. “Improved Speech Quality and Efficient Vecto
r Quantization in SELP, ”Proc.ICASSP, pp155-158,19
88 (Ref. 2).

[0003] The main coding parameters in the CELP system are a filter coefficient, a drive signal, and a gain. Since the encoding method is a so-called analysis-by-synthesis, that is, a method based on an analysis-synthesis method, a decoder is directly included in an encoder. The encoder performs actual decoding by using this decoder while changing the value of the above-mentioned encoding parameter (local decoding), and searches for a combination of encoding parameters in which the distortion between the decoded speech and the input speech becomes smaller. Is transmitted. The decoder performs decoding with the combination of the received coding parameters to obtain a decoded speech.

FIG. 6 shows a typical encoder based on the CELP system, and FIG. 7 shows a decoder. First, in the encoder shown in FIG. 6, the audio divided in units of frames is input to an input terminal 1060, and this is input to a linear prediction analysis unit 1.
By analyzing at 050, the coefficients of the perceptual weighting synthesis filter 1030 and the perceptual weight filter 1040 are determined. Next, the weighted input voice 1041 and the decoded voice 1031 weighted by the perceptual weight filter 1040
Is evaluated by the error evaluator 1070, and code vectors of the adaptive codebook 1011 and the noise codebook 1012 are selected so that the error, that is, the distortion of the decoded speech 1031 is minimized. Usually, these two codebooks 1011 and 1012
Are searched sequentially. That is, first, adaptive codebook 1
011 to determine the adaptive code vector, and then search the random codebook 1012 to determine the random code vector. Finally, gains to be multiplied by the adaptive code vector and the noise code vector by the gain multipliers 1021 and 1022 are obtained from the gain codebook 1013. After multiplying the adaptive code vector and the noise code vector by the gains in the gain multipliers 1021 and 1022, the adder 1023 drives the perceptually weighted synthesis filter 1030 with a drive signal obtained by the addition, whereby the decoded speech 1031 is obtained. Can be

Then, the LPC coefficients (coefficients of the perceptually weighted synthesis filter 1030) generated by the linear prediction analysis unit 1050 and quantized by the LPC quantization unit 1080, and the indexes (drives) of the adaptive codebook 1011 and the noise codebook 1012 are used. The signal (signal) and the index (gain) of the gain codebook 1013 are output as coding parameters.

On the other hand, the decoder has the same structure as a part of the encoder shown in FIG. 6, as shown in FIG. Based on the encoding parameters sent from the encoder, the adaptive codebook 2011 and the noise codebook 2012 decode the code vector and the gain codebook 2013 respectively the gain, and the LPC inverse quantization unit 2081 decodes the synthesis filter 2.
Decode the 035 coefficient. Gain multipliers 2021, 20
After multiplying the gain by 22, the adder 2023 drives the synthesis filter 2035 with a drive signal obtained by adding them, thereby obtaining a decoded voice. Normally, the decoded speech is further output through a post filter 2090 to improve the auditory quality.

Although the CELP system has high sound quality, it is known as a system requiring a large amount of calculation in an encoder. Most of the amount of calculation is spent searching for code vectors from adaptive codebook 1011 and noise codebook 1012. In this search, for all the code vectors stored in the adaptive codebook 1011 and the noise codebook 1012, the drive signal is passed through the weighted synthesis filter 1030, and then the distortion with respect to the target vector is calculated. The amount of calculation is required. The point in realizing was how to reduce the amount of calculation.

[0008] Of the codebooks 1011 and 1012, the search for the adaptive codebook 1011 is particularly easy to reduce the amount of calculation. Since the adaptive codebook 1011 generates a code vector representing a pitch component, it is possible to roughly determine which code vector is appropriate by analyzing a pitch period of an input signal or a past decoded speech. It is. By searching for the code vector thus provided and its surroundings, it is possible to search for an almost optimal code vector.

On the other hand, the search for the random codebook 1012 is difficult to reduce the amount of calculation. The noise codebook 1012 has a role of indicating a deviation from the target vector, which cannot be completely expressed by the adaptive code vector. Therefore, the stored vector is also a noise-like code vector with a small correlation, and it is difficult to search with an aim like a search of the adaptive codebook 1011.

Therefore, a so-called structured codebook in which the structure of the codebook is modified so that the amount of calculation can be reduced has been widely used. VSELP, ACELP, and the like are typical examples of performing encoding using the structured codebook. However, although the use of a structured codebook is effective in reducing the amount of calculation, it imposes restrictions on the code vectors to be stored, so that the sound quality is deteriorated compared to the case where the structure is not structured. was there.

[0011]

As described above, the CE
In the conventional speech coding method based on the LP method, there is a problem that a large number of calculations are required particularly for searching for a code vector from a noise codebook, and when the codebook is structured to reduce the calculation amount, There was a problem that the sound quality deteriorated. An object of the present invention is to provide a speech encoding method and a speech decoding method which require a small amount of calculation and have high sound quality.

[0012]

In order to solve the above-mentioned problems, a speech encoding method according to the present invention employs a past drive for driving a synthesis filter whose coefficients are determined based on an analysis result of an input speech signal. An adaptive code vector is extracted from an adaptive codebook that stores an adaptive code vector generated based on the signal and used as a pitch excitation signal, and the pitch excitation signal is passed through a synthesis filter to generate a first response signal. On the other hand, a conversion pattern codebook that stores a plurality of conversion patterns is prepared, and the first response signal is converted with the conversion pattern extracted from the codebook to generate a second response signal. Then, the first and second response signals are combined to generate a synthesized speech signal, and the adaptive code vector and the conversion pattern that reduce the distortion of the synthesized speech signal with respect to the input speech signal are converted into an adaptive codebook and a conversion pattern codebook. , And outputs at least information on the coefficients of the synthesis filter and information indicating the adaptive code vector and the conversion pattern searched from the adaptive codebook and the conversion pattern codebook as coding parameters.

In this speech coding method, when a second response signal corresponding to a convolved noise code vector according to the conventional method is obtained, the conversion stored in the conversion pattern codebook with respect to the first response signal is performed. Various conversions are applied according to the pattern, a conversion pattern that minimizes the distortion of the synthesized voice signal with respect to the input voice signal is searched, and an index indicating the conversion pattern is output.

This eliminates the need for a noise code vector filtering operation required in the conventional noise codebook search. Since the amount of calculation that is increased to search for the optimal conversion is small compared to the filtering operation, it is possible to obtain the second response signal as a whole with a significantly smaller amount of calculation than in the past.

Here, as the conversion pattern stored in the conversion pattern codebook, for example, a matrix operation is used.
In addition, when performing a calculation for searching for an optimal conversion pattern, many elements of this matrix are set to zero and some of them are non-zero, for example, each row has five or less non-zero elements. To be configured. In this way, the effect of reducing the amount of calculation is further increased.

On the other hand, in the speech decoding method of the present invention corresponding to this speech encoding method, at least a filter coefficient of a synthesis filter, an adaptive code vector, and an index indicating a conversion pattern from a coding side are input as coding parameters. An adaptive code vector extracted from an adaptive codebook storing an adaptive code vector according to a coding parameter is used as a pitch excitation signal, and the pitch excitation signal is passed through a synthesis filter whose coefficient is determined according to the coding parameter to obtain a first response signal. Generate The second response signal is generated by converting the first response signal using a conversion pattern extracted from a conversion pattern codebook storing a plurality of conversion patterns in accordance with an encoding parameter. Then, the first and second response signals are combined to generate a decoded audio signal.

The speech synthesizing method of the present invention may have a configuration without a synthesis filter. In this case, an adaptive code vector is extracted from the adaptive codebook and used as a first pitch signal, and the conversion pattern extracted from the conversion pattern codebook is used. The first pitch signal is converted to generate a second pitch signal, and the first and second pitch signals are synthesized to generate a synthesized speech signal. Then, an adaptive code vector and a conversion pattern in which the distortion of the synthesized voice signal with respect to the input voice signal is smaller are searched from the adaptive code book and the conversion pattern code book, respectively, and at least the adaptive code vector searched from the adaptive code book and the conversion pattern code book. And an index indicating the conversion pattern is output as an encoding parameter.

According to the speech decoding method of the present invention corresponding to this speech encoding method, at least an index indicating an adaptive code vector and a conversion pattern from the encoding side is input as a coding parameter, and is input from the adaptive codebook according to the coding parameter. The extracted adaptive code vector is used as a first pitch signal, and the second pitch signal is generated by converting the first pitch signal with a conversion pattern extracted from the conversion pattern codebook in accordance with an encoding parameter. Then, the first and second pitch signals are combined to generate a decoded audio signal.

[0019]

BEST MODE FOR CARRYING OUT THE INVENTION

(First Embodiment) FIG. 1 shows a configuration of a speech coding apparatus to which a speech coding method according to a first embodiment is applied. The speech coding apparatus includes an adaptive codebook 101, a synthesis filter 102, an analysis filter 103 which is an inverse filter of the synthesis filter 102, a conversion pattern codebook 104, a conversion unit 105, gain multipliers 106 and 107, and a gain codebook 10.
8, an adder 109, a linear prediction analysis unit 110, an input terminal 111 for an input audio signal 132, a subtractor 112, an audibility weighting filter 113, and an evaluation unit 114.

The adaptive codebook 101 includes a synthesis filter 102
And a plurality of adaptive code vectors generated from the past drive signals for driving. Here, the drive signal is obtained by passing the synthesized voice signal 131 through the analysis filter 103 whose filter coefficients are determined by the linear prediction analysis unit 110 that analyzes the input voice signal 132. Synthesis filter 1
In step 02, a filter coefficient is determined by the linear prediction analysis unit 110, an adaptive code vector extracted from the adaptive codebook 101 is input as a pitch excitation signal, and a pitch response signal (first response signal) is output.

The conversion pattern codebook 104 stores a plurality of conversion patterns representing conversions to be performed on the pitch response signal from the synthesis filter 102. The conversion unit 105 uses the conversion pattern extracted from the conversion pattern codebook 104. It converts the pitch response signal and outputs a converted response signal (second response signal).

The pitch response signal output from the synthesis filter 102 and the conversion response signal output from the conversion section 105 are multiplied by gains given from a gain codebook 108 by gain multipliers 106 and 107, respectively, and then added by an adder. The sum is added at 109 to generate a synthesized speech signal 131. The subtractor 112 outputs a difference signal between the synthesized voice signal 131 and the input voice signal 132. The audibility weighting filter 113 weights the difference signal with audibility.

The evaluation unit 114 includes an audibility weighting filter 113
The difference between the perceptually weighted difference signals is evaluated, and the adaptive code is set so that the power of the difference signal is minimized, that is, the distortion with the perceptual weight of the synthesized speech signal 131 with respect to the input speech signal 132 is minimized. The book 101, the conversion pattern codebook 104, and the gain codebook 108 are searched.

As a result of the search by the evaluation unit 114, the index 121 of the gain codebook 108, the conversion pattern codebook 1
An index 122, an index 123 of the synthesis filter 102, and an index 124 of the adaptive codebook 101 are obtained, and these indexes are output as coding parameters.

Next, the procedure of the speech encoding process in the present embodiment will be described with reference to the flowchart shown in FIG. [Step S1] First, the linear prediction analysis unit 110 performs linear prediction analysis on an input audio signal 132 input to the input terminal 111 with a predetermined frame length, and
2. Analysis filter 103 and auditory weight filter 113
Find the coefficient of. The coefficients of the synthesis filter 102 are usually vector-quantized and output as an index 123 of the synthesis filter 102.

[Step S2] Next, the adaptive codebook 101
Search for. That is, one adaptive code vector is searched from the adaptive codebook 101 so that the pitch excitation signal that minimizes the difference signal weighted by the perceptual weight filter 113 is output. A method of searching the adaptive codebook 101 for an adaptive code vector that is a candidate for a pitch excitation signal is well known in the art, and the method can be used in the present embodiment. This search result is output as an index of adaptive codebook 101. Further, the optimum pitch excitation signal thus obtained is passed through the synthesis filter 102 to obtain a pitch response signal.

[Step S3] Next, the conversion section 105 applies a conversion indicated by the conversion pattern stored in the conversion pattern codebook 104 to the pitch response signal to generate a conversion response signal, which is obtained in step S2. A search is made for a conversion pattern that minimizes the power of the perceptually weighted difference signal when combined with the pitch response signal, and the search result is output as an index 122 of the conversion pattern codebook 104.

[Step S4] Next, the gain multiplier 1 is set so that the perceptual weighted error from the target vector is minimized.
The gain multiplied by 06 and 107 is searched from gain codebook 108, and the search result is output as index 121 of gain codebook 108.

[Step S5] Finally, the input audio signal 1
In order to prepare for the processing of the frame next to 32, the synthesized speech signal 131 is input to the analysis filter 103 to generate a residual signal, and the content of the adaptive codebook 101 is updated using the residual signal.

Next, effects of the present embodiment will be described. In the conventional noise codebook search, it is necessary to perform a convolution operation using a filter on all the code vectors stored in the codebook, and this convolution operation has been a major cause of an increase in the amount of calculation.

On the other hand, in the present embodiment, step S2
At the end of the search of the adaptive codebook 101, the pitch excitation signal is determined. By using the pitch response signal obtained by convolving the signal with the synthesis filter 102, the amount of calculation corresponding to the conventional noise codebook search can be reduced. . That is, by performing appropriate conversion on the pitch response signal as shown in step S3, a conversion response signal corresponding to the response of the conventional noise code vector is created. By doing so, the convolution operation in the conventional random codebook search becomes unnecessary, and the amount of calculation can be greatly reduced.

Next, the conversion pattern codebook 104 and the conversion unit 105 which characterize the present invention will be described in detail. Conversion pattern of conversion applied to pitch response signal by conversion section 105, that is, conversion pattern codebook 1
The conversion pattern stored in 04 is represented by, for example, a matrix operation. If the vector representing the pitch response signal output from the synthesis filter 102 is p and the matrix representing the conversion pattern is Ai (i = 1,..., M, where M is the number of conversion patterns), the output from the conversion unit 105 The vector xi representing the converted response signal to be expressed can be expressed as xi = Aip (1).

When the evaluation unit 114 searches the conversion unit pattern codebook 104, all Ai (i = 1,...,
Xi is calculated for M) to obtain xi that minimizes distortion with respect to the target vector, and Ai at that time is set as an optimum conversion pattern, and an index i of the conversion pattern codebook 104 representing the conversion pattern is output.

The vectors xi and p are usually vectors of the dimension of the frame (or sub-frame) length. If this dimension is N, Ai becomes an N * N matrix. Therefore, if all of these elements are non-zero elements, the amount of calculation for obtaining xi increases, and the merit of reducing the amount of calculation as compared with the conventional method using a random codebook decreases. However, by limiting the number of non-zero elements, for example, by limiting the number of non-zero elements in each row to 5 or less, the amount of computation required for conversion can be significantly reduced as compared to the case of convolution operation. .

Although this seems to limit the variation of the conversion response signal output from the conversion unit 105, since an N * N redundant matrix is originally used for the conversion of the N-dimensional vector, It is not a restriction. In fact, if there are N non-zero elements, it is possible to generate an arbitrary response vector. When each row has only one non-zero element and the others are zero, the conversion for obtaining xi can be performed by N operations, and the amount of calculation can be greatly reduced as compared with the conventional method.

Explaining using mathematical expressions, the search of the conversion pattern codebook 104 in the present embodiment is performed when the target vector generated from the input speech signal 132 is r, and the evaluation value E = <r, Search for xi that maximizes xi> ² // xi | ² (2). The amount of operation required to calculate the value of the evaluation value E once is N times in the calculation of the expression (1),
Since N + 1 times in the numerator of the equation (2) and N times in the denominator, about 3N
Times.

On the other hand, in the conventional noise codebook search, when convolution by a synthesis filter is represented by a matrix H and a noise code vector is represented by ci, E = <r, Hci> ² / | Hci | ² (3) There is a need. In this method, the convolution operation of Hci requires N (N + 1) / 2 operations, and the inner product of the denominator and the numerator requires N operations, respectively.
A total of 2N + N (N + 1) / 2 calculations are required.

Now, the subframe length N = 40 (5 mse
Assuming that an 8-bit codebook is used in c), the conventional method using the noise codebook is (2 * 40 + 40 * 41/2)
* 256/5 msec = 46.1 MOPS, whereas in the present embodiment, 3 * 40 * 256/5 msec =
The conversion pattern codebook 104 can be searched at 6.1 MOPS, and the amount of calculation required for the search is 1/7 that of the conventional method.
It is reduced below.

The evaluation expressions (2) and (3) vary depending on the search method. For example, when the orthogonal search method is used, x
Instead of i, a vector xi 'obtained by making this orthogonal to the pitch response vector may be used.

Further, the converted response signal obtained by converting the pitch response signal in the present embodiment is the input voice signal 1
There is also an advantage that the response signal conforms to the characteristics of the T.32. This is because the adaptive code vector stored in the adaptive codebook 101 serving as the pitch excitation signal is created from the past drive signal, and thus reflects the properties of the already encoded input speech signal. Therefore, since the converted response signal obtained by transforming it has its properties, a response signal suitable for the input audio signal is obtained, which leads to improvement in sound quality.

In the conventional noise codebook search, a method of providing a plurality of codebooks and switching them according to the input speech signal, a method of making the code vector periodic with the pitch cycle of the input speech signal, and the like are available. Although attempts have been made to adapt to an input audio signal, it has been extremely difficult to automatically generate a conversion response signal that matches the input audio signal as in the present embodiment.

In the CELP system, usually, in order to reduce the amount of calculation, before taking the difference between the synthesized speech signal and the input speech signal,
In many cases, each signal is subjected to filtering by an audibility weighting filter. Also in the present embodiment, it is possible to make such a contrivance. The block diagram in FIG. 1 is based on the theory in order to facilitate understanding of the search mechanism, and does not limit the position of the audibility weighting filter 113.

There are various other methods for the gain search in step S4. For example, immediately after the pitch excitation signal is determined in step S2, the value of the gain to be multiplied by the gain multiplier 107 is determined, and then the process of step S3 is performed. In step S4, the value of the gain to be multiplied by the gain multiplier 106 is determined. For example, a method of determining the gain sequentially may be used. That is, the method of determining the gain in step S4 is an example, and the method is not limited to this.

Next, the configuration of the speech decoding apparatus according to this embodiment will be described with reference to FIG. Since the present invention is a speech coding method based on the analysis and synthesis method, the decoding device is the same as a decoding device (local decoder) incorporated in the coding device. That is, this speech decoding apparatus is adapted to the adaptive codebook 20.
1, synthesis filter 202, inverse filter 203 of synthesis filter 202, conversion pattern codebook 204, conversion unit 20
5, gain multipliers 206 and 207, gain codebook 208,
It comprises an adder 209, an inverse quantization unit 210, and input terminals 221, 222, 223 and 224 for coding parameters.

In FIG. 3, input terminals 221, 222,
223 and 224 include, as coding parameters, an index 121 of the gain codebook 108, an index 122 of the conversion pattern codebook 104, and an index 1 of the synthesis filter 102 output from the speech coding apparatus of FIG.
23 and the index 124 of the adaptive codebook 101 are input.

The index 121 is given to the gain codebook 208, and the gain value is read out from the gain codebook 208 to the gain multipliers 206 and 207 based on the index 121. The index 122 is provided to the conversion pattern codebook 204, and a conversion pattern is provided to the conversion unit 205 based on the index. The index 123 is provided to the synthesis filter 202 and the inverse filter 203 via the inverse quantization unit 210, and the coefficients of these filters 202 and 203 are determined. The index 124 is given to the adaptive codebook 201, and an adaptive code vector serving as a pitch excitation signal input to the synthesis filter 202 is selected.

As a result, the pitch response signal output from the synthesis filter 202 and the conversion response signal output from the conversion unit 205 are multiplied by gains in the gain multipliers 206 and 207 and then added by the adder 209. As a result, a decoded audio signal 211 is generated. This decoded audio signal 21
1 may be processed by a post-filter (not shown) so that the quality is improved audibly as needed.

(Second Embodiment) FIG. 4 shows a speech coding apparatus to which a speech coding method according to a second embodiment is applied. This embodiment is different from the first embodiment in the position of the audibility weighting filter.

In the description of the first embodiment, the position of the audibility weighting filter may be anywhere. However, when the configuration of the present embodiment is adopted, it is not equivalent to the first embodiment. This is because in the present embodiment, the pitch response signal output from the synthesis filter 102 is passed through the perceptual weight filter 141 and then converted, whereas in the first embodiment the pitch response signal is converted and then transmitted to the perceptual weight filter 113. Because it passes. In general, the order of conversion and filtering is not equivalent because they are not interchangeable. Whether they are equivalent or not is not related to the effect of the present invention, and the conversion pattern codebook 104 may be designed so as to conform to either configuration. In the present embodiment, the synthesis filter 10
2 is provided with the audibility weight filter 141 on the output side, and the audibility weight inverse filter 1 is provided on the input side of the inverse filter 103.
40, and the input audio signal is also applied to the audibility weighting filter 1
It is input to the subtractor 112 through.

(Third Embodiment) In the first embodiment, the amount of calculation can be reduced, but the amount of memory for the matrix Ai representing the transformation becomes relatively large. That is, to prepare 256 candidates even if each row has only one non-zero element, 40 * 256 =
10k words of memory are required.

Therefore, in the present embodiment, the conversion pattern codebook 104 has an overlap codebook, particularly non-zero elements have an overlap structure, and each Ai has a nonzero element in a diagonal direction. By doing so, the amount of memory has been significantly reduced.

FIG. 5 schematically shows the state. A vector having a subframe length is cut out from the i-th position of the -overlapping codebook in which the diagonal components are stored, and used as a diagonal element of Ai. Therefore, Ai and Ai + 1 have N-1 common elements, that is, overlapping components. The amount of memory at this time is N + M-1 (M is the number of conversion patterns). As described in the first embodiment, this is 0.295 k words when N = 40 and M = 256, which facilitates mounting with a lower memory.

Although the number of shifts of the overlap codebook is set to 1 here, the number of shifts can be set to a value of 2 or more. In this case, adjacent transformation matrices have overlapping elements equal to or less than N-2, and the degree of freedom increases accordingly, but the amount of memory slightly increases.

Although only the diagonal components have values as the transformation matrix and the diagonal components are overlapped here, it is not necessary to limit the components having the values only in the diagonal direction. Wrapping is also possible.

Although the CELP system normally has the synthesis filters 102 and 202 as shown in FIGS. 1 and 3, the present invention does not use a synthesis filter. In other words, when the weight of the synthesis filter is 1, Can also be applied.

When the synthesis filter is not used, the encoding side uses the adaptive code vector extracted from the adaptive codebook 101 as the first pitch signal, and converts the first pitch signal with the conversion pattern extracted from the conversion pattern codebook 104. What is necessary is just to synthesize | combine with the 2nd pitch signal and to generate | occur | produce a synthetic | combination audio signal. In this case, the encoding parameters do not include the filter coefficients of the synthesis filter. On the decoding side, on the other hand, at least an index indicating an adaptive code vector and a conversion pattern is input as an encoding parameter, and the adaptive code vector extracted from the adaptive codebook is used as a first pitch signal in the same manner as the encoding side, and the conversion pattern code The decoded voice signal may be generated by combining the first pitch signal with the second pitch signal obtained by converting the first pitch signal using the conversion pattern extracted from the book.

[0057]

As described above, according to the present invention, the filtering operation required for searching for a noise codebook in the related art is not required, the encoding can be performed with a small amount of operation, and the input speech can be obtained. There is an effect that it is easy to generate a synthesized speech signal that matches the signal properties, and the sound quality is improved.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a speech encoding device according to a first embodiment of the present invention.

FIG. 2 is a flowchart showing a processing procedure according to the embodiment;

FIG. 3 is a block diagram showing the configuration of the speech decoding apparatus according to the embodiment;

FIG. 4 is a block diagram illustrating a configuration of a speech encoding device according to a second embodiment of the present invention.

FIG. 5 is a diagram schematically showing a state of conversion applied to a first response signal in the present invention.

FIG. 6 is a block diagram showing a configuration of a conventional CELP-type speech coding apparatus.

FIG. 7 is a block diagram showing a configuration of a conventional CELP-based speech decoding apparatus.

[Explanation of symbols]

 Reference Signs List 101 adaptive codebook 102 synthesis filter 103 analysis filter 104 conversion pattern codebook 105 conversion units 106 and 107 gain multiplier 108 gain codebook 109 adder 110 linear prediction analysis unit 111 input terminal 112 ... subtractors 113, 141, 142 ... perceptual weight filter 140 ... perceptual weight inverse filter 114 ... evaluator 121 ... gain codebook index 122 ... conversion pattern codebook index 123 ... synthesis filter index 124 ... adaptive codebook index 131 ... Synthesized speech signal 132 ... Input speech signal 210 ... Dequantizer 201-204 ... Input terminal 201 ... Adaptive codebook 202 ... Synthesis filter 203 ... Analysis filter 204 ... Conversion pattern codebook 2105 ... Conversion units 206 and 207 ... Gain Multiplier 208 ... Acquisition codebook 209 Adder 210 Dequantizer 211 Decoded speech signal signal 1011 Adaptive codebook 1012 Noise codebook 1013 Gain codebook 1021, 1022, Gain multiplier 1023 Adder 1030 Perceptual weight Attached synthesis filter 1031 decoded speech signal 2035 synthesis filter 1040 perceptual weight filter 1041 perceptual weighted input voice signal 1050 linear predictive analysis unit 1060 input terminal 1070 error evaluator 1071 subtractor 1080 LPC quantizer 2011 Adaptive codebook 2012 Noise codebook 2013 Gain codebook 2021, 2022 Gain multiplier 2023 Adder 2035 Synthesis filter 2081 LPC dequantizer 2090 Post filter

Claims (7)

[Claims] An adaptive code vector is generated from an adaptive code book storing an adaptive code vector generated based on a past drive signal for driving a synthesis filter whose coefficient is determined based on an analysis result of an input voice signal. The pitch excitation signal is taken out as a pitch excitation signal. The pitch excitation signal is passed through the synthesis filter to generate a first response signal. The first response signal is obtained by using a conversion pattern extracted from a conversion pattern codebook storing a plurality of conversion patterns. Converting the signal to generate a second response signal; synthesizing the first and second response signals to generate a synthesized audio signal; and adapting the input audio signal to reduce the distortion of the synthesized audio signal. A code vector and a conversion pattern are respectively searched from the adaptive codebook and the conversion pattern codebook, and at least the coefficients of the synthesis filter and the adaptive Speech encoding method and outputting an index indicating an adaptive code vector and transformation pattern is searched from the issue book and conversion pattern codebook as an encoding parameter. 2. A conversion pattern codebook that extracts an adaptive code vector from an adaptive codebook that stores an adaptive code vector generated based on a past synthesized speech signal and that is used as a first pitch signal, and that stores a plurality of conversion patterns. Converting the first pitch signal with the conversion pattern extracted from the first pitch signal to generate a second pitch signal; synthesizing the first and second pitch signals to generate a synthesized voice signal; The adaptive code vector and the conversion pattern searched from the adaptive code book and the conversion pattern code book, respectively, for the adaptive code vector and the conversion pattern which reduce the distortion of the synthesized speech signal with respect to the adaptive code book and the conversion pattern code book. Is output as an encoding parameter. 3. The speech encoding method according to claim 1, wherein the plurality of conversion patterns stored in the conversion pattern codebook are represented by a matrix operation. 4. The speech coding method according to claim 3, wherein the matrix has rows in which the number of non-zero components is 5 or less. 5. The conversion pattern codebook according to claim 3, wherein the matrix is a diagonal matrix, and the conversion pattern codebook is configured such that diagonal components of adjacent matrices have overlapping components.
3. The speech encoding method according to claim 1. 6. An adaptive code vector extracted at least according to the coding parameter from an adaptive code book storing an adaptive code vector, at least an index indicating a filter coefficient of a synthesis filter, an adaptive code vector, and a conversion pattern is input. As a pitch excitation signal, the pitch excitation signal is passed through a synthesis filter whose coefficient is determined according to the encoding parameter to generate a first response signal, and the encoding parameter is obtained from a conversion pattern codebook storing a plurality of conversion patterns. Converting the first response signal with a conversion pattern extracted according to the following formula to generate a second response signal, and combining the first and second response signals to generate a decoded speech signal. Audio decoding method. 7. An adaptive code vector extracted from an adaptive codebook storing an adaptive code vector in accordance with the encoding parameter is inputted as at least an index indicating an adaptive code vector and a conversion pattern as a first pitch signal. Converting the first pitch signal with a conversion pattern extracted according to the encoding parameter from a conversion pattern codebook storing a plurality of conversion patterns to generate a second pitch signal; A speech decoding method comprising: synthesizing a pitch signal to generate a decoded speech signal.