JP2004046232A - Voice decoding device and its method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reproduce voice of high quality by small information quantity in voice encoding/decoding for compressing and encoding a voice signal into a digital signal. <P>SOLUTION: In code excited linear predictor (CELP) voice encoding, the degree of voice noisiness in an encoding section is evaluated by using at least one code or encoded result out of spectral information, power information and pitch information and different excited code books 19, 20 are used in accordance with evaluation results. <P>COPYRIGHT: (C)2004,JPO

Description

The present invention relates to a voice coding / decoding method and a voice coding / decoding apparatus used when a voice signal is compression-coded / decoded into a digital signal, and particularly to a high-quality voice at a low bit rate. The present invention relates to an audio encoding method and an audio decoding method, and an audio encoding device and an audio decoding device.

Conventionally, as a high-efficiency speech coding method, Code-Excited Linear Prediction (CELP) coding is typical, and the technique is described in "Code-excited Linear Prediction (CELP): High-quality". speech at very low bit rates "(MR Shroeder and BSAtal, ICASSP '85, pp.937-940, 1985).

FIG. 6 shows an example of the overall configuration of the CELP speech coding / decoding method. In the drawing, reference numeral 101 denotes a coding unit, 102 denotes a decoding unit, 103 denotes a multiplexing unit, and 104 denotes a separating unit. The encoding unit 101 includes a linear prediction parameter analysis unit 105, a linear prediction parameter encoding unit 106, a synthesis filter 107, an adaptive codebook 108, a driving codebook 109, a gain encoding unit 110, a distance calculation unit 111, and a weighting addition unit 138. It is configured. The decoding unit 102 includes a linear prediction parameter decoding unit 112, a synthesis filter 113, an adaptive codebook 114, a driving codebook 115, a gain decoding unit 116, and a weighting addition unit 139.

In CELP speech coding, about 5 to 50 ms is defined as one frame, and the speech of the frame is divided into spectrum information and sound source information and encoded. First, the operation of the CELP speech coding method will be described. In the encoding unit 101, the linear prediction parameter analysis unit 105 analyzes the input speech S101 and extracts a linear prediction parameter that is speech spectrum information. The linear prediction parameter coding unit 106 codes the linear prediction parameter, and sets the coded linear prediction parameter as a coefficient of the synthesis filter 107.

Next, encoding of sound source information will be described. The adaptive codebook 108 stores past driving excitation signals, and outputs a time-series vector obtained by periodically repeating the past driving excitation signals corresponding to the adaptive code input from the distance calculation unit 111. The drive codebook 109 stores, for example, a plurality of time-series vectors configured by learning so that distortion between the learning speech and the encoded speech is reduced. Output a time-series vector corresponding to the code.

The time series vectors from the adaptive codebook 108 and the driving codebook 109 are weighted and added by the weighting and adding means 138 according to the respective gains provided from the gain coding means 110, and the added result is combined as a driving excitation signal. The coded voice is supplied to the filter 107 to obtain the coded voice. The distance calculation unit 111 obtains the distance between the coded speech and the input speech S101, and searches for an adaptive code, a drive code, and a gain that minimize the distance. After the above-mentioned encoding is completed, the code of the linear prediction parameter, the adaptive code for minimizing the distortion between the input speech and the encoded speech, the drive code, and the gain code are output as the encoding result.

Next, the operation of the CPEL audio decoding method will be described.

In the one-side decoding unit 102, the linear prediction parameter decoding unit 112 decodes the linear prediction parameter from the code of the linear prediction parameter, and sets it as a coefficient of the synthesis filter 113. Next, adaptive codebook 114 outputs a time-series vector corresponding to the adaptive code and periodically repeats the past excitation signal, and drive codebook 115 outputs a time-series vector corresponding to the drive code. I do. These time-series vectors are weighted and added by weighting and adding means 139 according to the respective gains decoded from the sign of the gain by the gain decoding means 116, and the addition result is supplied to the synthesis filter 113 as a drive excitation signal. Thus, an output voice S103 is obtained.

As a conventional speech coding / decoding method improved for the purpose of improving reproduced speech quality by the CELP speech coding / decoding method, “Phonetically-based vector excitation coding of speech at 3.6 kbps” (S. Wang and A. Gersho, ICASSP '89, pp.49-52, 1989). FIG. 7 in which the same reference numerals are assigned to the corresponding units as in FIG. 6 shows an example of the entire configuration of the conventional audio encoding / decoding method. In the encoding unit 101 in FIG. Codebook switching means, 119 is a first drive codebook, and 120 is a second drive codebook. Further, in the decoding means 102 in the figure, 121 is a driving codebook switching means, 122 is a first driving codebook, and 123 is a second driving codebook. The operation of the encoding / decoding method having such a configuration will be described. First, in the encoding unit 101, the voice state determination unit 117 analyzes the input voice S101, and determines which of the two voice states is voiced / unvoiced. The driving codebook switching unit 118 determines that the first driving codebook 119 is used if voiced, and the second driving codebook 120 is used if voiced, for example. The book is switched, and which driving codebook is used is encoded.

Next, in the decoding unit 102, the driving codebook switching unit 121 determines that the same driving codebook as used in the encoding unit 101 is used according to the code of which driving codebook was used in the encoding unit 101. The first drive codebook 122 and the second drive codebook 123 are switched. With this configuration, a driving codebook suitable for encoding is prepared for each state of the voice, and the quality of the reproduced voice is improved by switching and using the driving codebook according to the state of the input voice. can do.

A conventional speech coding / decoding method for switching between a plurality of driving codebooks without increasing the number of transmission bits is disclosed in JP-A-8-185198. In this method, a plurality of driving codebooks are switched and used according to the pitch period selected in the adaptive codebook. As a result, it is possible to use a drive codebook adapted to the characteristics of the input speech without increasing the transmission information.

"Code-excited linear prediction (CELP): High-quality speech at very low bit rates" (M.R.Shroeder and B.S.Atal, ICASSP '85, pp.937-940, 1985) "Phonetically-based vector excitation coding of speech at 3.6kbps" (S. Wang and A. Gersho, ICASSP '89, pp.49-52, 1989) JP-A-8-185198

As described above, in the conventional speech encoding / decoding method shown in FIG. 6, synthesized speech is generated using a single driving codebook. In order to obtain high-quality coded speech even at a low bit rate, the time-series vector stored in the driving codebook is non-noise containing many pulses. For this reason, when coding and synthesizing noise-like sounds such as background noise and fricative consonants, there is a problem in that the coded sounds emit unnatural sounds such as gritty sounds and dusty sounds. This problem can be solved by constructing the driving codebook only from noise-like time-series vectors, but the quality of the encoded speech as a whole deteriorates.

In the improved conventional speech encoding / decoding method shown in FIG. 7, a plurality of driving codebooks are switched according to the state of input speech to generate encoded speech. Thus, for example, in an unvoiced part where the input speech is noisy, a driving codebook composed of a noise-like time series vector is used, and in other voiced parts, a driving codebook composed of a non-noise time-series vector is used. Therefore, even if noise-like speech is encoded and synthesized, an unnatural sizzling sound is not emitted. However, since the decoding side uses the same driving codebook as the encoding side, it is necessary to encode and transmit information on which driving codebook is newly used, which hinders a reduction in bit rate. There was a problem.

In the conventional speech coding / decoding method for switching between a plurality of driving codebooks without increasing the number of transmission bits, the driving codebook is switched according to a pitch period selected by the adaptive codebook. However, the pitch period selected in the adaptive codebook is different from the pitch period of the actual voice, and it is not possible to judge whether the state of the input voice is noisy or non-noise based on the value alone. Does not solve the problem that the coded speech is unnatural.

The present invention has been made to solve such a problem, and an object of the present invention is to provide an audio encoding / decoding method and apparatus for reproducing high quality audio even at a low bit rate.

In order to solve the above-described problems, a speech encoding method according to the present invention uses at least one code or encoding result of spectrum information, power information, and pitch information to determine the degree of noise of speech in the encoding section. The evaluation is performed, and one of the plurality of driving codebooks is selected according to the evaluation result.

The speech encoding method according to the next invention further includes a plurality of driving codebooks having different degrees of noise of the stored time-series vectors, and a plurality of driving codes according to the evaluation result of the degree of noise of speech. Switch books.

In the speech encoding method of the next invention, the degree of noise of the time-series vector stored in the driving codebook is changed according to the evaluation result of the degree of noise of speech.

Furthermore, the speech encoding method according to the next invention includes a driving codebook storing a noise-like time-series vector, and thins out signal samples of the driving sound source according to the evaluation result of the degree of noise of the speech. A time series vector with a low degree of noise is generated.

The speech encoding method according to the next invention further comprises a first driving codebook storing a noise-like time-series vector and a second driving codebook storing a non-noise-like time-series vector. The time series vector of the first driving codebook and the time series vector of the second driving codebook are weighted and added to generate a time series vector in accordance with the evaluation result of the degree of noise of speech. .

Further, the speech decoding method of the next invention evaluates the degree of noise of speech in the decoding section using at least one code or decoding result among spectrum information, power information, and pitch information, and Accordingly, one of a plurality of driving codebooks is selected.

Furthermore, the speech decoding method of the next invention comprises a plurality of driving codebooks having different degrees of noise of stored time-series vectors, and a plurality of driving codes according to the evaluation result of the degree of noise of speech. Switch books.

In the speech decoding method of the next invention, the degree of noise of the time series vector stored in the driving codebook is changed according to the evaluation result of the degree of noise of speech.

Further, the speech decoding method according to the next invention includes a driving codebook storing a noise-like time-series vector, and thins out signal samples of the driving sound source according to the evaluation result of the degree of noise of the speech. A time series vector with a low degree of noise is generated.

Further, the speech decoding method according to the next invention is characterized in that a first driving codebook storing a noise-like time series vector and a second driving codebook storing a non-noise time-series vector are used. The time series vector of the first driving codebook and the time series vector of the second driving codebook are weighted and added to generate a time series vector in accordance with the evaluation result of the degree of noise of speech. .

Further, a speech encoding apparatus according to the next invention encodes spectrum information of an input speech and outputs the encoded spectrum information as one element of an encoding result, and the encoded spectrum information from the spectrum information encoding section. A noise evaluation unit that evaluates the degree of noise of speech in the coding section using at least one code or coding result of the spectrum information and power information obtained from The first driving codebook in which a plurality of time-series vectors are stored, the second driving codebook in which a plurality of noise-like time-series vectors are stored, and an evaluation result of the noise degree evaluation unit, A drive codebook switching unit for switching between the first drive codebook and the second drive codebook, and a time series vector from the first drive codebook or the second drive codebook. A weighted addition unit for weighting and adding according to the gain of the torquer, and using the weighted time-series vector as a drive excitation signal, based on the drive excitation signal and the encoded spectrum information from the spectrum information encoding unit. A synthesis filter for obtaining a coded voice by means of a filter, obtaining a distance between the coded voice and the input voice, searching for a drive code and a gain that minimize the distance, and using the result as a drive code and a code for the gain to obtain the coding result. And a distance calculation unit that outputs the result as

A speech decoding apparatus according to the next invention further comprises a spectrum information decoding unit for decoding spectrum information from a code of the spectrum information, and spectrum information and power obtained from the decoded spectrum information from the spectrum information decoding unit. A noise evaluation unit that evaluates the degree of noise of speech in the decoding section using at least one decoding result of the information or the code of the spectrum information, and outputs a non-noise evaluation unit; A first driving codebook in which a time series vector is stored, a second driving codebook in which a plurality of noise-like time series vectors are stored, and a first driving code A drive codebook switching unit for switching between a book and a second drive codebook, and a time series vector from the first drive codebook or the second drive codebook, respectively. A weighted addition unit that weights and adds according to the gain of the vector, and sets the weighted time-series vector as a drive excitation signal, based on the drive excitation signal and the decoded spectrum information from the spectrum information decoding unit. And a synthesis filter for obtaining a decoded speech.

A speech coding apparatus according to the present invention is a code-driven linear prediction (CELP) speech coding apparatus, wherein at least one of spectrum information, power information, and pitch information or a coding result is used. And a driving codebook switching unit that switches a plurality of driving codebooks according to the evaluation result of the noise evaluation unit.

A speech decoding apparatus according to the present invention is a code-driven linear prediction (CELP) speech decoding apparatus, wherein at least one of spectrum information, power information, and pitch information or a decoding result is used. And a driving codebook switching unit that switches a plurality of driving codebooks according to the evaluation result of the noise evaluation unit.

According to the speech encoding method and the speech decoding method, and the speech encoding apparatus and the speech decoding apparatus according to the present invention, at least one of the spectrum information, the power information, and the pitch information or the encoding result is used for the encoding. Since the degree of noise of speech in the segmentation section is evaluated and different driving codebooks are used in accordance with the evaluation result, high-quality speech can be reproduced with a small amount of information.

Further, according to the present invention, in the speech encoding method and the speech decoding method, a plurality of driving codebooks having different degrees of noise of the driving excitation stored therein are provided, and the plurality of driving codebooks are provided according to the evaluation result of the degree of noise of the speech. Since a plurality of driving codebooks are switched and used, high-quality sound can be reproduced with a small amount of information.

Further, according to the present invention, in the speech encoding method and the speech decoding method, the degree of noise of the time-series vector stored in the driving codebook is changed according to the evaluation result of the degree of noise in speech. Therefore, high-quality sound can be reproduced with a small amount of information.

Further, according to the present invention, in the speech encoding method and the speech decoding method, a driving codebook storing a noise-like time-series vector is provided. Since a time-series vector with a low degree of noise is generated by thinning out the signal samples of the vector, high-quality sound can be reproduced with a small amount of information.

Further, according to the present invention, in the speech encoding method and the speech decoding method, the first driving codebook storing the noise-like time-series vector and the second driving codebook storing the non-noise-like time-series vector. And a time series vector obtained by weighting and adding the time series vector of the first drive codebook and the time series vector of the second drive codebook according to the evaluation result of the degree of noise in speech. Is generated, a high-quality sound can be reproduced with a small amount of information.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

Embodiment 1 FIG.
FIG. 1 shows an overall configuration of a first embodiment of a speech encoding method and a speech decoding method according to the present invention. In the figure, 1 is an encoding unit, 2 is a decoding unit, 3 is a multiplexing unit, and 4 is a demultiplexing unit. The coding unit 1 includes a linear prediction parameter analysis unit 5, a linear prediction parameter coding unit 6, a synthesis filter 7, an adaptive codebook 8, a gain coding unit 10, a distance calculation unit 11, a first driving codebook 19, 2, a driving codebook 20, a noise evaluation unit 24, a driving codebook switching unit 25, and a weighting addition unit 38. The decoding unit 2 includes a linear prediction parameter decoding unit 12, a synthesis filter 13, an adaptive codebook 14, a first driving codebook 22, a second driving codebook 23, a noise evaluation unit 26, and a driving codebook switching. It comprises a unit 27, a gain decoding unit 16, and a weighting addition unit 39. In FIG. 1, reference numeral 5 denotes a linear prediction parameter analysis unit as a spectrum information analysis unit for analyzing the input speech S1 and extracting a linear prediction parameter which is spectrum information of the speech, 6 encodes the linear prediction parameter which is spectrum information, The linear prediction parameter coding units 19 and 22 serving as spectrum information coding units for setting the coded linear prediction parameters as coefficients of the synthesis filter 7 are the first drive in which a plurality of non-noise time-series vectors are stored. Codebooks 20 and 23 are second driving codebooks storing a plurality of noise-like time-series vectors, 24 and 26 are noise degree evaluation units for evaluating the degree of noise, and 25 and 27 are driven based on the degree of noise. A driving codebook switching unit that switches codebooks.

Hereinafter, the operation will be described. First, in the encoding unit 1, the linear prediction parameter analysis unit 5 analyzes the input speech S1 and extracts a linear prediction parameter that is speech spectrum information. The linear prediction parameter coding unit 6 codes the linear prediction parameter, sets the coded linear prediction parameter as a coefficient of the synthesis filter 7, and outputs the coefficient to the noise evaluation unit 24. Next, encoding of the sound source information will be described. The adaptive codebook 8 stores past driving excitation signals, and outputs a time-series vector obtained by periodically repeating the past driving excitation signals corresponding to the adaptive code input from the distance calculation unit 11. The noise degree evaluator 24 uses the coded linear prediction parameters and the adaptive codes input from the linear prediction parameter encoder 6 to calculate, for example, the spectral tilt, the short-term prediction gain, and the pitch variation as shown in FIG. The degree of noise in the coding section is evaluated, and the evaluation result is output to the driving codebook switching unit 25. According to the evaluation result of the noise degree, the driving codebook switching unit 25 determines that the first driving codebook 19 is used if the noise degree is low, and that the second driving codebook 20 is used if the noise degree is high. Switch the driving codebook to be used.

The first driving codebook 19 stores a plurality of non-noise time-series vectors, for example, a plurality of time-series vectors configured by learning such that distortion between the learning speech and the encoded speech is reduced. Have been. The second driving codebook 20 stores a plurality of noise-like time-series vectors, for example, a plurality of time-series vectors generated from random noise. Output the corresponding time series vector. Each time-series vector from the adaptive codebook 8, the first excitation codebook 19, or the second excitation codebook 20 is weighted by the weighting and adding unit 38 according to each gain given from the gain encoding unit 10. The result of the addition is supplied to the synthesis filter 7 as a drive excitation signal to obtain an encoded voice. The distance calculation unit 11 obtains the distance between the coded speech and the input speech S1, and searches for an adaptive code, a driving code, and a gain that minimize the distance. After the coding is completed, the code of the linear prediction parameter, the adaptive code for minimizing the distortion between the input voice and the coded voice, the driving code, and the gain code are output as the coding result S2. The above is the characteristic operation of the speech encoding method according to the first embodiment.

Next, the decoding unit 2 will be described. In the decoding unit 2, the linear prediction parameter decoding unit 12 decodes the linear prediction parameter from the code of the linear prediction parameter, sets the decoded parameter as a coefficient of the synthesis filter 13, and outputs the coefficient to the noise evaluation unit 26. Next, decoding of the sound source information will be described. The adaptive codebook 14 outputs a time-series vector obtained by periodically repeating a past excitation signal in accordance with the adaptive code. The noise degree evaluation unit 26 evaluates the degree of noise from the decoded linear prediction parameters input from the linear prediction parameter decoding unit 12 and the adaptive code in the same manner as the noise degree evaluation unit 24 of the encoding unit 1. Then, the evaluation result is output to the driving codebook switching unit 27. The driving codebook switching unit 27 switches between the first driving codebook 22 and the second driving codebook 23 in the same manner as the driving codebook switching unit 25 of the encoding unit 1 according to the evaluation result of the noise degree.

The first driving codebook 22 includes a plurality of non-noise time-series vectors, for example, a plurality of time-series vectors configured by learning such that distortion between the learning speech and the encoded speech is reduced. The second drive codebook 23 stores a plurality of noise-like time-series vectors, for example, a plurality of time-series vectors generated from random noise, and outputs a time-series vector corresponding to each drive code. The time series vectors from the adaptive codebook 14 and the first driving codebook 22 or the second driving codebook 23 are weighted and added to each other by the gain decoding unit 16 in accordance with the respective gains decoded from the gain codes. , And the result of the addition is supplied to the synthesis filter 13 as a drive sound source signal to obtain an output sound S3. The above is the characteristic operation of the speech decoding method according to the first embodiment.

According to the first embodiment, the degree of noise of the input speech is evaluated from the code and the coding result, and different driving codebooks are used in accordance with the evaluation result, thereby reproducing high quality speech with a small amount of information. can do.

Further, in the above embodiment, the case where a plurality of time-series vectors are stored in driving codebooks 19, 20, 22, and 23 has been described. However, if at least one time-series vector is stored, It is feasible.

Embodiment 2 FIG.
In Embodiment 1 described above, two driving codebooks are used by switching, but instead, three or more driving codebooks may be provided and used by switching according to the degree of noise. According to the second embodiment, it is possible to use a driving codebook suitable for not only two types of speech, that is, noise / non-noise, but also intermediate speech such as slightly noisy one. High quality audio can be reproduced.

Embodiment 3 FIG.
FIG. 3 where the same reference numerals are assigned to corresponding parts to FIG. 1 shows the entire configuration of a speech encoding method and speech decoding method according to a third embodiment of the present invention. Driving codebooks 29 and 31 storing the vectors are sample thinning units for setting the amplitude values of low-amplitude samples of the time series vector to zero.

Hereinafter, the operation will be described. First, in the encoding unit 1, the linear prediction parameter analysis unit 5 analyzes the input speech S1 and extracts a linear prediction parameter that is speech spectrum information. The linear prediction parameter coding unit 6 codes the linear prediction parameter, sets the coded linear prediction parameter as a coefficient of the synthesis filter 7, and outputs the coefficient to the noise evaluation unit 24. Next, encoding of the sound source information will be described. The adaptive codebook 8 stores past driving excitation signals, and outputs a time-series vector obtained by periodically repeating the past driving excitation signals corresponding to the adaptive code input from the distance calculation unit 11. The noise degree evaluator 24 uses the coded linear prediction parameter input from the linear prediction parameter encoder 6 and the adaptive code to calculate, for example, the noise of the coding section from the spectrum slope, short-term prediction gain, and pitch fluctuation. The degree is evaluated, and the evaluation result is output to the sample thinning unit 29.

The driving codebook 28 stores a plurality of time-series vectors generated from, for example, random noise, and outputs a time-series vector corresponding to the driving code input from the distance calculation unit 11. According to the noise level evaluation result, if the noise level is low, the sample decimation unit 29 converts the time series vector input from the driving codebook 28 into, for example, the amplitude value of a sample less than a predetermined amplitude value. The time series vector set to zero is output. If the noise level is high, the time series vector input from the driving codebook 28 is output as it is. The time series vectors from the adaptive codebook 8 and the sample thinning unit 29 are weighted and added by the weighting and adding unit 38 according to the respective gains provided from the gain coding unit 10, and the addition result is used as a driving excitation signal. The encoded voice is supplied to the synthesis filter 7 and is obtained. The distance calculation unit 11 obtains the distance between the coded speech and the input speech S1, and searches for an adaptive code, a driving code, and a gain that minimize the distance. After the coding is completed, the code of the linear prediction parameter, the adaptive code for minimizing the distortion between the input voice and the coded voice, the driving code, and the gain code are output as the coding result S2. The above is the characteristic operation of the speech encoding method according to the third embodiment.

Next, the decoding unit 2 will be described. In the decoding unit 2, the linear prediction parameter decoding unit 12 decodes the linear prediction parameter from the code of the linear prediction parameter, sets the decoded parameter as a coefficient of the synthesis filter 13, and outputs the coefficient to the noise evaluation unit 26. Next, decoding of the sound source information will be described. The adaptive codebook 14 outputs a time-series vector obtained by periodically repeating a past excitation signal in accordance with the adaptive code. The noise degree evaluation unit 26 evaluates the degree of noise from the decoded linear prediction parameters input from the linear prediction parameter decoding unit 12 and the adaptive code in the same manner as the noise degree evaluation unit 24 of the encoding unit 1. Then, the evaluation result is output to the sample thinning unit 31.

The drive codebook 30 outputs a time-series vector corresponding to the drive code. The sample thinning unit 31 outputs a time-series vector by the same processing as that of the sample thinning unit 29 of the encoding unit 1 according to the noise degree evaluation result. The time series vectors from the adaptive codebook 14 and the sample decimating unit 31 are weighted and added by a weighting and adding unit 39 according to the respective gains provided from the gain decoding unit 16, and the result of the addition is used as a driving excitation signal. The output sound S3 supplied to the synthesis filter 13 is obtained.

According to the third embodiment, a driving codebook that stores a noise-like time-series vector is provided, and a signal sample of a driving sound source is thinned out according to the evaluation result of the degree of noise of the speech. By generating a driving sound source with a low degree of sound, high-quality sound can be reproduced with a small amount of information. Further, since there is no need to provide a plurality of driving codebooks, there is also an effect of reducing the amount of memory required for storing the driving codebooks.

Embodiment 4 FIG.
In the above-described third embodiment, the time series vector samples are decimated / not decimated. Alternatively, the amplitude threshold value for decimating the samples may be changed according to the degree of noise. According to the fourth embodiment, it is possible to generate and use a time-series vector suitable not only for two kinds of speech, that is, noise / non-noise but also for an intermediate speech such as a little noise. As a result, high-quality sound can be reproduced.

Embodiment 5 FIG.
FIG. 4 in which parts corresponding to those in FIG. 1 are assigned the same reference numerals shows the overall configuration of a fifth embodiment of the speech encoding method and the speech decoding method of the present invention. The first driving codebooks 33 and 36 storing vectors are second driving codebooks storing non-noise time-series vectors, and 34 and 37 are weight determining units.

Hereinafter, the operation will be described. First, in the encoding unit 1, the linear prediction parameter analysis unit 5 analyzes the input speech S1 and extracts a linear prediction parameter that is speech spectrum information. The linear prediction parameter coding unit 6 codes the linear prediction parameter, sets the coded linear prediction parameter as a coefficient of the synthesis filter 7, and outputs the coefficient to the noise evaluation unit 24. Next, encoding of the sound source information will be described. The adaptive codebook 8 stores past driving excitation signals, and outputs a time-series vector obtained by periodically repeating the past driving excitation signals corresponding to the adaptive code input from the distance calculation unit 11. The noise degree evaluator 24 uses the coded linear prediction parameter input from the linear prediction parameter encoder 6 and the adaptive code to calculate, for example, the noise of the coding section from the spectrum slope, short-term prediction gain, and pitch fluctuation. The degree is evaluated, and the evaluation result is output to the weight determining unit 34.

The first driving codebook 32 stores a plurality of noise-like time-series vectors generated from, for example, random noise, and outputs a time-series vector corresponding to the driving code. The second driving codebook 33 stores, for example, a plurality of time-series vectors configured by learning such that distortion between the learning speech and the encoded speech is reduced, and is input from the distance calculation unit 11. A time series vector corresponding to the driving code to be output is output. The weight determination unit 34 determines the time series vector from the first drive codebook 32 and the second drive codebook 33 from the second drive codebook 33 according to, for example, FIG. The weight given to the time series vector of is determined. Each time-series vector from the first drive codebook 32 and the second drive codebook 33 is weighted and added according to the weight given from the weight determination unit 34. The time series vector output from the adaptive codebook 8 and the time series vector generated by the weighted addition are weighted and added by the weighting and adding section 38 according to the respective gains given from the gain coding section 10, The result of the addition is supplied to the synthesis filter 7 as a drive excitation signal to obtain an encoded voice. The distance calculation unit 11 obtains the distance between the coded speech and the input speech S1, and searches for an adaptive code, a driving code, and a gain that minimize the distance. After this coding is completed, the code of the linear prediction parameter, the adaptive code for minimizing the distortion between the input voice and the coded voice, the driving code, and the gain code are output as the coding result.

Next, the decoding unit 2 will be described. In the decoding unit 2, the linear prediction parameter decoding unit 12 decodes the linear prediction parameter from the code of the linear prediction parameter, sets the decoded parameter as a coefficient of the synthesis filter 13, and outputs the coefficient to the noise evaluation unit 26. Next, decoding of the sound source information will be described. The adaptive codebook 14 outputs a time-series vector obtained by periodically repeating a past excitation signal in accordance with the adaptive code. The noise degree evaluation unit 26 evaluates the degree of noise from the decoded linear prediction parameters input from the linear prediction parameter decoding unit 12 and the adaptive code in the same manner as the noise degree evaluation unit 24 of the encoding unit 1. Then, the evaluation result is output to the weight determination unit 37.

The first driving codebook 35 and the second driving codebook 36 output a time-series vector corresponding to the driving code. It is assumed that the weight determining unit 37 gives weights in the same manner as the weight determining unit 34 of the encoding unit 1 in accordance with the noise evaluation result input from the noise evaluation unit 26. The respective time-series vectors from the first driving codebook 35 and the second driving codebook 36 are weighted and added according to the respective weights given from the weight determining unit 37. The time series vector output from the adaptive codebook 14 and the time series vector generated by the weighted addition are combined by the weight addition section 39 in accordance with the respective gains decoded from the gain codes by the gain decoding section 16. The weighted addition is performed, and the result of the addition is supplied to the synthesis filter 13 as a driving sound source signal to obtain an output sound S3.

According to the fifth embodiment, the degree of speech noise is evaluated from the code and the coding result, and a noise-like time-series vector and a non-noise time-series vector are weighted and used in accordance with the evaluation result. Thus, high-quality sound can be reproduced with a small amount of information.

Embodiment 6 FIG.
In the above-described first to fifth embodiments, the codebook of the gain may be changed according to the evaluation result of the degree of noise. According to the sixth embodiment, a codebook having an optimum gain can be used according to the driving codebook, so that high-quality sound can be reproduced.

Embodiment 7 FIG.
In the above-described first to sixth embodiments, the degree of noise in speech is evaluated, and the driving codebook is switched according to the evaluation result. However, voiced rising and bursting consonants are determined and evaluated, respectively. The driving codebook may be switched according to the evaluation result. According to the seventh embodiment, not only the noise state of the voice but also voiced rising and burst consonants can be further finely classified, and a driving codebook suitable for each voice can be used. Can be played.

Embodiment 8 FIG.
In the above-described first to sixth embodiments, the degree of noise in the coding section is evaluated based on the spectrum tilt, the short-term prediction gain, and the pitch fluctuation shown in FIG. 2, but the magnitude of the gain value for the adaptive codebook output is used. May be evaluated.

FIG. 1 is a block diagram illustrating an overall configuration of a speech encoding and speech decoding apparatus according to a first embodiment of the present invention. 3 is a table provided for describing noise degree evaluation according to Embodiment 1 in FIG. 1. It is a block diagram which shows the whole structure of Embodiment 3 of the audio coding and audio decoding apparatus by this invention. It is a block diagram which shows the whole structure of Embodiment 5 of the audio coding and audio decoding apparatus by this invention. FIG. 15 is a schematic diagram used for describing a weight determination process according to the fifth embodiment of FIG. 4. FIG. 11 is a block diagram illustrating the overall configuration of a conventional CELP speech encoding / decoding device. FIG. 10 is a block diagram showing the overall configuration of a conventional improved CELP speech encoding / decoding device.

Claims (4)

In a code-driven linear prediction (CELP) speech decoding device that receives speech code including gain code and synthesizes speech,
A gain decoding unit that inputs the gain code and decodes the audio gain in the decoding section according to the input gain code,
Using the gain decoded by the gain decoding unit, a noise degree evaluation unit that evaluates the degree of noise of speech in the decoding section,
A drive codebook that stores a plurality of time-series vectors corresponding to a plurality of predetermined drive excitation signals,
A speech decoding device comprising: a noise level control unit that changes a degree of noise of a time-series vector output from the driving codebook according to an evaluation result of the noise level evaluation unit. In a code-driven linear prediction (CELP) speech decoding device that inputs a speech code including a code of a linear prediction parameter, an adaptive code, a drive code, and a gain code and synthesizes speech,
An adaptive codebook that stores a past driving excitation signal, inputs an adaptive code, and outputs a time-series vector corresponding to the past driving excitation signal according to the input adaptive code,
A drive codebook that stores a plurality of time-series vectors corresponding to a plurality of predetermined drive excitation signals, inputs a drive code, and outputs a time-series vector corresponding to the drive excitation signal according to the input drive code. ,
A gain decoding unit that inputs a gain code and decodes a gain of a voice in a decoding section according to the input gain code,
The gain decoding unit inputs the decoded gain, a noise degree evaluation unit that evaluates the degree of noise of the speech in the decoding section using the input gain,
The noise level evaluation unit and the time series vector output from the driving codebook are input, and the noise characteristics of the time series vector output from the driving codebook according to the evaluation result of the noise level evaluation unit are input. A noise level control unit for changing the degree of noise,
The time series vector output from the adaptive codebook, the time series vector output from the driving codebook, and the gain decoded by the gain decoding unit are input, and the time series vector output from the adaptive codebook is input. And a time-series vector output from the driving codebook are weighted using the gain, the two time-series vectors weighted using the gain are added, and a weighted addition unit that outputs an addition result,
A linear prediction parameter decoding unit that receives the sign of the linear prediction parameter, decodes the linear prediction parameter from the input sign of the linear prediction parameter, and outputs the linear prediction parameter;
A synthesis filter that inputs the linear prediction parameter output from the linear prediction parameter decoding unit and the addition result output from the weighting addition unit, and synthesizes a voice using the linear prediction parameter and the addition result. An audio decoding device, comprising: In a code-driven linear prediction (CELP) speech decoding method for inputting a speech code including a gain code and synthesizing speech,
The above-mentioned gain code is input, and the voice gain in the decoding section is decoded according to the input gain code,
Using the decoded gain, evaluate the degree of noise of the speech in the decoding section,
A plurality of time-series vectors corresponding to a plurality of predetermined drive excitation signals are stored in the drive codebook,
A speech decoding method characterized by changing a degree of noise of a time-series vector output from the driving codebook according to an evaluation result. In a code-driven linear prediction (CELP) speech decoding method for inputting a speech code including a code of a linear prediction parameter, an adaptive code, a drive code, and a gain code and synthesizing speech,
Input an adaptive code to the adaptive codebook that stores the past drive excitation signal, and output a time-series vector corresponding to the past drive excitation signal from the adaptive codebook according to the input adaptive code,
A plurality of time-series vectors corresponding to a plurality of predetermined drive excitation signals are stored in the drive codebook,
A drive code is input to the drive codebook, and a time-series vector corresponding to the drive excitation signal is output from the drive codebook according to the input drive code,
A gain code is input, and the gain of the audio in the decoding section is decoded using the input gain code,
Input the decoded gain, evaluate the degree of noise of the speech in the decoding section using the input gain,
The evaluation result and the time series vector output from the driving codebook are input, and the degree of noise of the time series vector output from the driving codebook is changed according to the evaluation result,
The time series vector output from the adaptive codebook, the time series vector output from the driving codebook and the decoded gain are input, and the time series vector output from the adaptive codebook and the driving codebook are input. Weighting the time series vector output from the above using the gain, adding the two time series vectors weighted using the gain, outputting the addition result,
Input the sign of the linear prediction parameter, decode and output the linear prediction parameter from the input sign of the linear prediction parameter,
A speech decoding method comprising: inputting the output linear prediction parameter and the output result, and synthesizing a speech using the linear prediction parameter and the addition result.

Images