JPH1020889A - Voice coding device and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce an operation amount when a noise code book is retrieved, a memory amount required for the noise code book and to improve tone quality of a sound part by structuring the noise code book, sharing plural pieces of pulses with noise code vectors of adjacent indexes and making so that only a few or a piece of pulse is different. SOLUTION: A square error minimization means 21 selects the noise code vector 25 that a square error between a target vector of a noise code vector component and the noise code vector after passing through a linear predictive synthetic filter becomes minimum from the noise code book 24 to output it by using the target vector 22 of the noise code vector component and the impulse response 23 of the linear predictive filter. At this time, the noise code book 24 is structured so that only a piece of pulse is different between the noise code vectors of the adjacent indexes. For instance, in the index 2, 2a, 2b, 2c are the same as 1b, 1c, 1d of the index 1, and only 2d is different.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a CELP-type speech coding apparatus and a recording medium on which the software is recorded.

[0002]

2. Description of the Related Art Conventionally, a CELP-type speech coding apparatus using a sparse noise source (a noise source configured by raising several pulses having a non-zero amplitude in one frame) has been proposed in the Proceedings of the Acoustical Society of Japan (Heisei Heisei). March 2006) 305 to 3rd
Page 06 and the 2nd Annual Meeting of the Acoustical Society of Japan (March 1995)
It is disclosed on pages 43 to 244. These CE
The noise code vector generation technique in the LP-type speech coding apparatus will be described with reference to FIG.

In FIG. 8, a square error minimizing means 11 is shown.
Is the target vector 12 of the noise code vector component
(Subtracting the zero input response of the linear prediction synthesis filter and the adaptive code vector component from the input speech after the hearing weighting)
A noise code vector 15 that minimizes an error between the noise code vector component of the synthesized speech and a noise code vector component obtained by passing the noise code vector through a linear predictive synthesis filter and multiplying by an appropriate gain is selected from the noise code book 14. Output. The calculation of the noise code vector passed through the linear prediction synthesis filter is performed by convolving the impulse response 13 of the linear prediction synthesis filter with each code vector of the noise code vector.

When the noise code vector is orthogonalized with respect to the adaptive code vector, the target vector of the entire sound source (a linear predictive synthesis filter based on the input speech after hearing weighting) is used without using the target vector of the noise code vector component. ), And the adaptive code vector is used to minimize the square error between the target vector and the synthesized speech.

[0005]

However, in the above conventional noise codebook, the calculation of the error evaluation function performed by the square error minimizing means is independent for each noise code vector stored in the noise codebook. Therefore, there is a problem that the amount of calculation is large and the memory capacity used is large.

SUMMARY OF THE INVENTION The present invention solves such a conventional problem. The present invention aims to reduce the amount of computation for searching for a random codebook, reduce the amount of memory required for the random codebook, and improve the sound quality of voiced parts. And a recording medium embodying the same.

[0007]

According to a first aspect of the present invention, a noise code book is structured, and noise code vectors of adjacent indexes share a plurality of pulses (the same position and the same amplitude). The noise code vector corresponding to the N-th index is changed to the N-th index of the noise codebook by making only a small number or one of the pulses different, or the noise code vector is composed of n or more pulses. Is expressed by the sum of the pulse of the position and amplitude described in (1) and the pulse of the position and amplitude described in the (N-1) th to (N-n + 1) th indices. And the amount of memory required for the random codebook is reduced. Since the availability of some of the results of the function calculation, it is possible to reduce the amount of calculation as compared with the case of performing the error evaluation function calculating independently for random code vectors for each index. Also, since only the position and amplitude of a pulse whose index is different from the previous noise code vector need be stored in the codebook, the amount of memory required for the noise codebook can be reduced.

A second invention for achieving the above object is:
A plurality of noise codebooks shown in the first invention are prepared, and the plurality of noise codebooks are selectively used according to the position of a pitch peak existing in an adaptive code vector, thereby improving the voice quality of a voiced part. By switching the noise codebook, phase information included in one frame (subframe) can be removed, and the sound source quantization performance of voiced parts can be improved. Further, the amount of calculation can be reduced as in the first invention.

A third invention for achieving the above object is:
In the noise codebook according to the first aspect, the position of the pulse is not represented by an absolute position in a frame (subframe), but is represented by a relative position where the pitch peak position of the adaptive code vector is 0. Thereby, the effect of the second invention can be obtained with the same memory amount as the first invention.

[0010] A fourth invention for achieving the above object is:
When stochastically generating pulse positions of a sparse noise codebook, by increasing the probability of occurrence near the pitch pulse position of the adaptive code vector and increasing the variations near the pitch pulse position, the voice quality of voiced parts is increased. Can be improved.

[0011]

BEST MODE FOR CARRYING OUT THE INVENTION
In a CELP type speech coding apparatus, a noise codebook is composed of elements describing a combination of a pulse position and an amplitude, and a noise code vector stored in the noise codebook is generated between noise code vectors of adjacent indexes. This is a speech coding apparatus structured so that only a small number of pulses are different, and has the effect of reducing the amount of memory required to store a random codebook and the amount of computation when searching for a random codebook. Have.

According to a second aspect of the present invention, the noise code vector of the speech coding apparatus according to the first aspect is structured such that only one pulse differs between noise code vectors of adjacent indexes. It is composed of n or more pulses and is at the N-th index, and has the effect of reducing the amount of memory required to store the random codebook and the amount of computation when searching for the random codebook.

The invention according to claim 3 of the present invention provides a CELP
In the speech coding apparatus, the random codebook is composed of elements describing the combination of the position and the amplitude of the pulse, and the random code vector is composed of n or more pulses.
The random code vector corresponding to the n-th index is obtained by adding the pulse of the position and the amplitude described in the N-th index of the random codebook and the (N−1) th to (N−n +
1) A speech encoding device represented by the sum of the position and amplitude pulses described in the index, which reduces the amount of memory required to store the random codebook and reduces This has the effect of reducing the amount of calculation.

According to a fourth aspect of the present invention, there is provided the speech encoding apparatus according to the second or third aspect, wherein the content of each index comprises only the position and amplitude of one pulse. Since the calculation of the error evaluation function can be performed by reusing a part of the calculation result of the error evaluation function performed on the index of, the operation amount required for searching the random codebook can be reduced.

According to a fifth aspect of the present invention, there is provided the speech coding apparatus according to the fourth aspect, wherein the noise code book to be used is switched and used according to the pitch peak position of the adaptive code vector. One frame (sub-frame) by switching and using books
This has the effect of removing the phase information present in the voiced part and improving the quantization efficiency of the sound source of the voiced part.

In the invention according to claim 6 of the present invention, the position of the pulse described in the content of each index is relative to the pitch peak position of the adaptive code vector when the pitch peak position of the adaptive code vector is 0. The speech coding apparatus according to any one of claims 1 to 5, wherein the speech coding apparatus is described by a position, and can reduce a memory amount necessary for storing a random codebook and a calculation amount required for searching for a random codebook. In addition, it has an effect of improving the voice quality of the voiced part.

According to a seventh aspect of the present invention, the position of a pulse is determined by a random number generation function in which the probability of occurrence of a position corresponding to the pitch peak position of an adaptive code vector is higher than the probability of occurrence of other positions. The speech coding apparatus according to any one of claims 1 to 6, further comprising a determined noise codebook, wherein the variation of the noise code vector near the pitch peak position is increased to improve the voice quality of the voiced part. It has the effect of being able to.

[0018] The invention according to claim 8 of the present invention is the invention according to claim 1.
8. A recording medium on which a program that implements the audio encoding device according to any one of claims 1 to 7 by software is recorded, and the audio encoding device according to any one of claims 1 to 7 is a magnetic disk such as a personal computer, This has the effect that it can be realized in a system using a recording medium such as a magnetic disk or a ROM cartridge.

Hereinafter, an embodiment of the present invention will be described with reference to FIG.
This will be described with reference to FIG. (Embodiment 1) FIG. 1 shows a noise code vector generation unit of a CELP type speech coding apparatus according to Embodiment 1 of the present invention, in which a noise code vector stored in a noise codebook is a noise code of an adjacent index. The vectors are structured such that only a small number of, for example, one pulse is different. In FIG. 1, reference numeral 21 denotes a square error minimizing means.
The target vector 22 of the noise code vector component, the impulse response 23 of the linear prediction synthesis filter, and the noise codebook 2
4, the noise code vector 25 that minimizes the square error between the target vector of the noise code vector component and the noise code vector after passing through the linear prediction synthesis filter is selected. Output. Reference numeral 22 denotes a target vector of a noise code vector component obtained by subtracting the zero input response of the linear prediction synthesis filter and the adaptive code vector component from the input speech after the hearing weighting.
Reference numeral 23 denotes an impulse response of the linear prediction synthesis filter in which the impulse response of the linear prediction synthesis filter is weighted by auditory sense. Reference numeral 24 denotes a noise codebook structured so that only one pulse differs between noise code vectors of adjacent indexes. For example, for index 2, 2
a, 2b, 2c are the same as 1b, 1c, 1d of index 1, and only 2d is different. Reference numeral 25 denotes a noise code vector that minimizes the square error of the noise code vector component from the target vector.

The operation of the noise code vector generation unit of the CELP type speech coding apparatus configured as described above is shown in FIG.
This will be described with reference to FIG. The square error minimizing means 21 maximizes (Equation 1) by using the target vector 22 of the noise code vector component and the impulse response 23 of the linear prediction synthesis filter, thereby obtaining a linear relationship with the target vector of the noise code vector component. The noise code vector 25 that minimizes the square error with the noise code vector after passing through the prediction synthesis filter is selected from the noise codebook 24 and output.

[0021]

(Equation 1) (Subscript t represents transposition of matrix) X: Target vector (L is frame (subframe) length) H: Impulse response convolution matrix C: Noise code vector

[0022]

(Equation 2) an: nth pulse amplitude pn: nth pulse position

[0023]

(Equation 3)

[0024]

(Equation 4)

[0025]

(Equation 5)

Here, the numerator term of (Equation 1) is represented by (Equation 2), and the denominator term of (Equation 1) is represented by (Equation 3). Further, since the noise codebook is structured so that only one pulse is different between the noise code vectors of adjacent indexes, (Equation 2) is (Equation 4).
(Equation 3) can be expressed by (Equation 5), and as shown in (Equation 4) and (Equation 5), there is a part that can be used recursively when calculating these equations. (Equation 4) (Equation 5)
Where p _N indicates the position of the newly added pulse by increasing the index by one, and p _Nn indicates the position of the deleted pulse by increasing the index by one. _{p N-n + 1 ~p N} -1 indicates the position of the common pulse to the noise code vector of the noise code vector and the current index of the previous index. If the noise code book is not structured such that only one pulse is different between the random code vectors of the adjacent indexes,
Since it cannot be expressed as (Equation 3) (Equation 4) (Equation 5),
It is necessary to calculate for all the terms included in (Equation 2) and (Equation 3) for each code vector, which increases the amount of calculation.

Although the case where only one pulse differs between code vectors of adjacent indexes has been described here, the same effect can be obtained when two or three pulses are used. However, the effect of the reduction in the amount of computation is greater as the value of (number of pulses different between code vectors of adjacent indexes) / (total number of pulses included in the code vector) is smaller. However, the effect of reducing the amount of memory is lost.

Further, when maximizing (Equation 1), a plurality of numerator components (Equation 2) are selected in advance (preliminary selection), and a denominator component (Equation 3) is calculated only for the selected one. By maximizing Equation 1, it is possible to further reduce the amount of computation. Also in this case, the calculation amount can be reduced by calculating the molecular component by (Equation 4) rather than by (Equation 2).

(Embodiment 2) Next, Embodiment 2 will be described. FIG. 2 shows a random code vector generation unit using a random codebook according to an embodiment of the present invention, in which the random code vector is composed of n or more, for example, n pulses, and the random code vector corresponding to the N-th index is , The pulse of the position and the amplitude described in the N-th index of the noise codebook and the (N−1) th to (N−n + 1)
It is represented by a combination of the position and amplitude pulses described in the third index.

In FIG. 2, reference numeral 31 denotes a noise codebook which outputs the position and amplitude of one pulse to the buffer 32 and the squared error minimizing means 35 in the order of index, and 32 denotes N which is output from the noise codebook 31. The position and the amplitude of the pulse of the index number are buffered, and (N−n) th to (N
-1) A buffer for outputting the position and amplitude of the pulse of the first index to the square error minimizing means 35. The buffer 33 subtracts the zero input response of the linear predictive synthesis filter and the adaptive code vector component from the input speech after the auditory weighting. The target vector of the noise code vector component, 34 is the impulse response of the linear prediction synthesis filter obtained by weighting the impulse response of the linear prediction synthesis filter with auditory weight, and 35 is the position and amplitude of the pulse of the Nth index input from the noise codebook 31. , (N−n) th to (N−1) th pulse positions are input from the buffer 32, and the target vector 33 of the noise code vector and the impulse response 34 of the linear prediction synthesis filter are input, and the noise code vector component Between the target vector 33 and the noise code vector after passing through the linear prediction synthesis filter Squared error minimization means multiplication error selects and outputs the noise code vector 36 that minimizes, 36
Is a noise code vector finally output from the square error minimizing means 35.

The operation of the noise code vector generator configured as described above will be described with reference to FIG. The noise codebook 31 stores the position and amplitude of one pulse for each index, and outputs the position and amplitude of the pulse to the buffer 32 and the square error minimizing means 35 in order from the first index. .

The buffer 32 buffers the pulse position and amplitude stored at the N-th index output from the noise codebook 31. In the buffer 32, the positions and amplitudes of the pulses stored in the (N−n) th to (N−1) th indexes of the random codebook 31 are buffered. The position and amplitude of the pulse stored at the third index is added to the buffer. Then, (N-
The amplitude and the position of the pulse stored in the (n) th to (N−1) th indexes are output to the square error minimizing means 35.

The square error minimizing means 35 outputs the noise codebook 3
(Equation 4) and (Equation 5) by using the position and amplitude of the N-th index pulse input from No. 1 and the (N-n) th to (N-1) -th index pulses input from the buffer 32. ) Is calculated to evaluate (Equation 1), and a noise code vector 36 represented by the index N of the noise codebook that maximizes (Equation 1) is output from all the indexes of the noise codebook 31. The noise code vector 36 is n
It is composed of pulses of position and amplitude stored in the Nth to (N-n + 1) th indexes.

FIG. 3 shows an example of the random codebook 31. In the example of FIG. 3, the random code vector represented by index 1 is
This is a vector in which a pulse of amplitude 1 exists in the fourth sample (all other samples are zero), and the random code vector represented by index 2 is
A vector in which pulses of amplitudes 1 and -2 exist at the 0th sample, and a noise code vector represented by index 3 are pulses of amplitudes 1, -2 and 3 at the 4th, 50th and 33rd pulses, respectively. Is a vector in which When n = 3, the noise code vector represented by index 4 is a vector in which pulses of amplitude −2, 3 and 5 exist at the 50th sample, the 33rd sample, and the 14th sample, respectively. The resulting noise code vector is a vector composed of three pulses of the position and amplitude stored in the indexes (N−2) to N. In this example, the noise code vector represented by index 1 or 2 does not have three pulses (the noise code vector represented by index 1 or (n-1) does not have n pulses). When assigning a vector having three (n) pulses to all indices, the effective index range may be limited to three or more (n or more). In this case, the first two (n)
An extra memory is required for the index for storing the position and amplitude of the minute pulse.

It is possible to improve the resistance to transmission line errors by devising the assignment of the index (Gray code).

Third Embodiment Next, a third embodiment will be described. FIG. 4 shows a noise code vector generation unit of a speech coding apparatus according to Embodiment 3 of the present invention, in which a plurality of noise codebooks shown in Embodiment 1 or 2 are prepared and a pitch peak existing in an adaptive code vector is provided. Has a configuration in which these plurality of random codebooks are properly used in accordance with the position of.

In FIG. 4, reference numeral 41 denotes an adaptive code vector selected and output from the adaptive code book, and reference numeral 42 denotes a noise code book selector 43 which receives the adaptive code vector 41, detects a pitch pulse position in the adaptive code vector, and A pitch pulse position detector 43 which receives the pitch pulse position output from the pitch pulse position detector 42 and outputs information on a noise code book to be selected to a switch 45; A random codebook 45 structured as shown in the first or second embodiment and optimized in accordance with the positions of different pitch pulses, 45 is to be selected from the random codebook selector 43 A switch 46 for switching the noise codebook 44 to be used based on the information in the noise codebook and connecting to the square error minimizing means 48;
Is the target vector of the noise code vector component obtained by subtracting the zero input response of the linear prediction synthesis filter and the adaptive code vector component from the input speech after the hearing weighting, and 47 is the linear prediction synthesis filter obtained by weighting the impulse response of the linear prediction synthesis filter with the hearing weight , The impulse response 48 of the noise code vector component, the impulse response 47 of the linear prediction synthesis filter, and the noise code vector stored in the noise codebook 44 input via the switch 45 are used as inputs. A square error minimizing means for selecting and outputting a noise code vector 49 that minimizes the square error between the target vector of the vector component and the noise code vector after passing through the linear prediction synthesis filter;
The noise code vector finally output from the squared error minimizing means 48.

The operation of the noise code vector generation unit of the speech coding apparatus configured as described above will be described with reference to FIG. In FIG. 4, a pitch pulse position detector 42 detects a position of a pitch pulse existing in an adaptive code vector using an input adaptive code vector.
The position of the pitch pulse can be determined by maximizing the normalized cross-correlation between the impulse train arranged in the pitch cycle and the adaptive code vector. Further, by minimizing the error between the impulse train arranged in the pitch cycle that has passed through the synthesis filter and the adaptive code vector that has passed through the synthesis filter, the impulse train can be obtained with higher accuracy. The pulse position is represented by the number of samples from the beginning of a frame (or subframe) to the first pitch pulse.

The noise codebook selector 43 selects one of a plurality of noise codebooks 44 considered to be optimal from among a plurality of noise codebooks 44 based on the pitch peak position of the adaptive code vector.
The control information for switching the switch 45 is transmitted to the switch 4
5 is output.

The switch 45 controls a plurality of random codebooks 4 according to the control information input from the random codebook selector 43.
Noise codebook and squared error minimizing means 48
And connect.

The square error minimizing means 48 maximizes (Equation 1) using the target vector 46 of the noise code vector component and the impulse response 47 of the linear prediction synthesis filter, thereby obtaining the target of the noise code vector component. The noise code vector 4 in which the square error between the vector and the noise code vector after passing through the linear prediction synthesis filter is minimized
9 is selected from the random codebook 44 connected via the switch 45 and output.

Since the configuration shown in the third embodiment is effective for voiced parts, the voiced speech of the CELP type speech coding apparatus having a configuration in which the sound source is divided into voiced and unvoiced parts and used in different modes. It can be applied to parts. Also,
By providing a pitch period processing unit in this noise code vector generator, the sound quality of voiced parts can be further improved.

(Fourth Embodiment) Next, a fourth embodiment will be described. FIG. 5 shows a noise code vector generation unit of the speech coding apparatus according to the fourth embodiment. In the noise code book according to the first to third embodiments, the position of a pulse is represented by an absolute position in a frame (subframe). , And is represented by a relative position where the pitch peak position of the adaptive code vector is 0.

In FIG. 5, reference numeral 51 denotes an adaptive code vector selected and output from the adaptive code book, and 52 receives the adaptive code vector 51 as an input, detects a pitch pulse position in the adaptive code vector, and sends it to a pulse position converter 53. The pitch pulse position detector 53 for outputting the pitch pulse position outputted from the pitch pulse position detector 52 and the noise codebook 54
A pulse position converter which receives only the pulse position and pulse amplitude output from the input, converts only the pulse position, and outputs the converted pulse amplitude to the buffer 55 and the square error minimizing means 58; Embodiment 2 in which the pulse position is relatively represented by setting the pulse position to 0
The noise code book 55 described above buffers the pulse position (pulse position converted by the pulse position converter 53) and the amplitude of the N-th index pulse output from the noise code book 54 via the pulse position converter 53. And (N-
A buffer for outputting the pulse position (pulse position converted by the pulse position converter 53) and the amplitude of the (n) th to (N-1) th indices to the square error minimizing means 58; A target vector of a noise code vector component obtained by subtracting the zero input response of the linear prediction synthesis filter and the adaptive code vector component from the input speech; 57, an impulse response of the linear prediction synthesis filter obtained by weighting the impulse response of the linear prediction synthesis filter with an auditory weight; Inputs the position and amplitude of the pulse of the N-th index from the random codebook 54 via the pulse position converter 53, and (N
The positions of the (n) th to (N-1) th pulses are input from the buffer 55 via the pulse position converter 53, and the target vector 56 of the noise code vector and the impulse response 57 of the linear prediction synthesis filter are input. Square error minimizing means for selecting and outputting a noise code vector 59 that minimizes the square error between the target vector 56 of the noise code vector component and the noise code vector after passing through the linear prediction synthesis filter; The noise code vector finally output from the power error minimizing means 58.

The operation of the noise code vector generator configured as described above will be described with reference to FIG. FIG.
In, the pitch pulse position detector 52 detects the position of the pitch pulse existing in the adaptive code vector using the input adaptive code vector 51. The position of the pitch pulse can be obtained by maximizing the normalized cross-correlation between the impulse train arranged in the pitch period and the adaptive code vector. Further, by minimizing the error between the impulse train arranged in the pitch cycle that has passed through the synthesis filter and the adaptive code vector that has passed through the synthesis filter, the impulse train can be obtained with higher accuracy. The pulse position is represented by the number of samples from the beginning of a frame (or subframe) to the first pitch pulse.

The pulse position converter 53 sets the pulse position stored in the noise codebook 54, which is relatively expressed assuming that the pitch pulse position is 0, as the absolute value of the start of the frame (or subframe) as 0. Convert to representation. Specifically, the pulse position output by the pitch pulse position detector 52 is added to the pulse position input from the noise codebook 54. Then, the pulse position and the pulse amplitude (the amplitude is input as it is) converted into an absolute expression in which the beginning of the frame (or the subframe) is 0 are output to the buffer 55 and the square error minimizing means 58. .

The buffer 55 stores the position of the pulse (pulse position converter 53) stored at the N-th index output from the noise codebook 54 via the pulse position converter 53.
And the amplitude of the converted pulse position are buffered. The buffer 55 buffers the positions and amplitudes of the pulses stored in the (N−n) th to (N−1) th indexes of the noise codebook 54 via the pulse position converter 53, New pulse position converter 5
The position and amplitude of the pulse stored in the N-th index of the random codebook 54 via 3 is added to the buffer. Then, the amplitude and position of the pulse stored in the (N−n) th to (N−1) th indexes of the noise codebook 54 via the pulse position converter 53 are output to the square error minimizing means 58. I do.

The square error minimizing means 58 includes the N-th index pulse input from the noise codebook 54 via the pulse position converter 53 and the (N-n) -th to (N-) pulses input from the buffer 55. (Equation 4) and (Equation 5) are calculated using the pulse position and amplitude of the (1) th index, and (Equation 1) is evaluated. ) Is output, and the noise code vector 36 represented by the index N of the noise codebook that maximizes the value is output. The noise code vector 59 is composed of n pulses, and N
It is generated by the pulse of the position and the amplitude stored in the (th)-(N-n + 1) th indices.

The noise codebook used in the fourth embodiment is the same as that in the second embodiment shown in FIG. 3, except that the pulse position value is a relative value with the pitch pulse position being 0. It is shown. Also, since the configuration shown in Embodiment 4 is effective for voiced parts, it is applied to a voiced part of a CELP-type speech coding apparatus having a configuration in which a sound source is divided into modes for voiced and unvoiced parts and used properly. can do. Further, by providing a pitch period processing unit in the noise code vector generation unit, the sound quality of voiced parts can be further improved.

Further, the audio encoding apparatus described in the first to fourth embodiments can be realized by recording as software on a recording medium such as a magnetic disk, a magneto-optical disk, and a ROM cartridge.

(Fifth Embodiment) Next, a fifth embodiment will be described. FIG. 6 shows a pulse position codebook generator according to Embodiment 5 of the present invention. When stochastically generating a pulse position of a sparse noise codebook, the probability of occurrence near the pitch pulse position of the adaptive code vector is increased. It is characterized by doing.

In FIG. 6, 61 is the minimum value of the generated pulse position, 62 is the maximum value of the generated pulse position, 63 is the position of the predetermined pitch pulse, and 64 is the probability distribution representing the occurrence probability distribution of the pulse position. A function 65 is a pulse position generator that sequentially outputs pulse positions by using the pulse position minimum value 61, the pulse position maximum value 62, the pitch pulse position 63, and the probability distribution function 64 as input, and 66 is an output from the pulse position generator 65. It is a pulse position codebook constituted by the obtained pulse positions.

The operation of the pulse position codebook generator configured as described above will be described with reference to FIGS. In FIG. 6, a minimum pulse position value 61 is input to a pulse position generator 65 for defining a minimum value of a pulse position appearing in a pulse position codebook 66.

The maximum pulse position value 62 is input to a pulse position generator 65 for defining the maximum pulse position appearing in the pulse position codebook 66.

The pitch pulse position 63 is input to the pulse position generator 65 in order to utilize the tendency that a pulse is likely to be formed near the pitch pulse position in a voiced part when a pulse sound source is applied to a noise sound source. You.

The probability distribution function 64 expresses, as a probability distribution function, a tendency that a pulse is likely to be formed near a pitch pulse position in a voiced part when a pulse sound source is applied to a noise source. The pulse position is input to a pulse position generator 65 to generate a pulse position.

The pulse position generator 65 has a minimum pulse position value 61, a maximum pulse position value 62, and a pitch pulse position 63.
And the probability distribution function 64 as inputs, the pulse position minimum value 6
Positions between 1 and the pulse position maximum value 62 are generated according to the distribution of the probability distribution function 64, and the generated pulse positions are output to the pulse position codebook 66. Here, the probability distribution function is a function having a distribution center at the pitch pulse position 63, and is as shown in FIG. The probability distribution function is prepared in advance by performing speech coding using a pulse sound source having a sufficient code vector and obtaining statistics.

Note that the codebooks of the first to fourth embodiments can be created using the pulse position codebook shown in the fifth embodiment.

[0059]

As described above, according to the present invention, by structuring a sparse noise codebook, the amount of computation required for searching for a random codebook and the amount of computation required for storing the noise codebook can be reduced. Becomes possible. Also, by expressing the pulse position using a relative position from the pitch pulse position existing in the adaptive code vector, it is possible to improve the sound quality of the voiced part by using the phase information existing in one pitch waveform. The effect that it can be obtained is obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram of a noise code vector generation unit of a CELP-type speech coding apparatus according to Embodiment 1 of the present invention.

FIG. 2 is a block diagram of a noise code vector generation unit of a CELP-type speech coding apparatus according to Embodiment 2 of the present invention;

FIG. 3 is a list showing an example of a random codebook according to Embodiment 2 of the present invention.

FIG. 4 is a block diagram of a noise code vector generation unit of a CELP-type speech coding apparatus according to Embodiment 3 of the present invention.

FIG. 5 is a block diagram of a noise code vector generation unit of a CELP-type speech coding apparatus according to Embodiment 4 of the present invention.

FIG. 6 is a block diagram of a pulse position codebook generator according to a fifth embodiment of the present invention.

FIG. 7 is a schematic diagram of a function representing an occurrence distribution of pulse positions used in a pulse position codebook generator according to Embodiment 5 of the present invention.

FIG. 8 is a block diagram of a noise code vector generation unit of a general CELP type speech coding apparatus.

[Explanation of symbols]

11, 21 square error minimizing means 12, 22 target vector of noise code vector component 13, 23 linear prediction synthesis filter impulse response 14, 24 noise codebook 15, 25 noise vector 31 noise codebook 32 buffer 33 noise code vector component Target vector 34 linear prediction synthesis filter impulse response 35 square error minimization means 36 noise code vector 41 adaptive code vector 42 pitch pulse position detector 43 noise codebook selector 44 noise codebook 45 switch 46 target of noise code vector component Vector 47 linear prediction synthesis filter impulse response 48 square error minimizing means 49 noise code vector 51 adaptive code vector 52 pitch pulse position detector 53 pulse position converter 54 noise codebook 55 buffer 56 noise code Target vector of vector component 57 linear prediction synthesis filter impulse response 58 square error minimizing means 59 noise code vector 61 minimum pulse position 62 maximum pulse position 63 pitch pulse position 64 probability density function 65 pulse position generator 66 pulse position code Book

Claims (8)

[Claims] In a CELP speech coding apparatus, a noise codebook is composed of elements describing a combination of a position and an amplitude of a pulse, and a noise code vector stored in the noise codebook is used for an index of an adjacent index. A speech coding apparatus structured so that only a small number of pulses are different between noise code vectors. 2. The speech coding apparatus according to claim 1, wherein only one pulse is structured to be different between noise code vectors of adjacent indexes. 3. In a CELP speech coding apparatus, a noise codebook is composed of elements describing a combination of a pulse position and an amplitude, a noise code vector is composed of n or more pulses, and an N-th index is used. Is a pulse of the position and amplitude described in the N-th index of the random codebook and (N-1)
A speech encoding device represented by the sum of the position and amplitude pulses described in the (N−n + 1) th index from the (n) th index. 4. The speech encoding apparatus according to claim 2, wherein the content of each index comprises only the position and amplitude of one pulse. 5. The speech coding apparatus according to claim 4, wherein a noise codebook to be used is switched according to a pitch peak position of the adaptive code vector. 6. The pulse position described in the content of each index is described by a relative position from the pitch peak position of the adaptive code vector, where the pitch peak position of the adaptive code vector is 0. 6. The speech encoding device according to any one of claims 1 to 5. 7. A noise code book in which a pulse position is determined using a random number generation function in which the probability of occurrence of a position corresponding to the pitch peak position of an adaptive code vector is higher than the probability of occurrence of another position. Item 7. The speech encoding device according to any one of Items 1 to 6. 8. A recording medium storing a program that implements the speech encoding device according to claim 1 by software.