JP3174781B2 - Diffusion sound source vector generation apparatus and diffusion sound source vector generation method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain good synthesized voice with small calculation amount for codebook survey and with small ROM capacities by supplying the same vector with the code vector generated by an algebraic codebook as a voice source vector. SOLUTION: An information storage part 1 for voice source stores voice source information of generating information of the like of code vector similar to an algebraic codebook and outputs the voice source information as a voice source vector. A diffused vector storing part 2 stores diffused information of a random progression or the like, and outputs the diffused information as a diffused vector. A convolution part 3 executes the convolution by inputting the voice source vector and the diffused vector and outputs them as a diffused voice source vector. Thus, a vector component of drive voice source diffused source has a random property, and a voice quality of output voice can be improved. Further, as the diffused voice source vector is formed with the voice source information of the information storage part 1 and the diffused vector of the diffused vector storing part 2, a ROM capacity can be reduced.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to an apparatus and a method for generating an excitation vector used in a speech encoding / decoding apparatus.

[0002]

2. Description of the Related Art In the field of mobile communications such as digital cellular phones, there is a demand for a low bit rate (up to about 8 kbps) voice compression coding method in order to cope with an increase in the number of subscribers. In Japan, VSELP and PSI
A voice coding method called -CELP has been adopted and put into practical use as a voice coding standard method for full-rate and half-rate digital mobile phones. Internationally, a coding system called CS-ACELP has been adopted as an international standard voice coding system of 8 kbps, and the ITU-T recommendation G.264. 729 (1995). All of these speech encodings are CELP ("CODE-EXCITED LI
NEAR PRIDICTION (CELP): HIGH QUALITY SPEECH AT VERY
LOW RATES ", Manfred R. Schroeder, Bishnu S. Atal,
Proc., ICASSP'85, pp.937-940).

Here, the basic algorithm of the CELP system will be described. The CELP system is a system in which voice information is separated and coded into sound source information and vocal tract information. The sound source information is coded by a plurality of code vector indexes stored in a codebook, and the vocal tract information is coded by L.
A method of encoding a PC (Linear Prediction Coefficient) and comparing the input speech with vocal tract information when encoding the sound source information (Abs: Analysis by Sy)
It is characterized in that it employs (theory). In general, in CELP, an input speech is divided into intervals (called frames) at certain time intervals, and LPC analysis is performed, and an adaptive codebook and a probabilistic codebook are divided into subdivided frames (called subframes). Is performed.

FIG. 5 is a functional block diagram of a conventional CELP speech decoder. Parameter decoding unit 152
Is a speech code (LPC code, probabilistic codebook index, probabilistic codebook coding gain, adaptive codebook index, transmitted from a conventional CELP speech coding apparatus (FIG. 6: described later). (The coding gain of the adaptive codebook) through the transmission unit 151. Next, the LPC code is decoded to obtain decoded LPC coefficients, the coding gain of the stochastic codebook is decoded to obtain the decoding gain of the stochastic codebook, and the coding gain of the adaptive codebook is decoded. Obtain the adaptive codebook decoding gain. Further, the index of the probabilistic codebook, the decoding gain of the probabilistic codebook, the index of the adaptive codebook,
The decoding gain of the adaptive codebook is output to driving excitation generating section 155, and the decoded LPC coefficient is output to LPC synthesizing section 156.

Driving excitation generating section 155 first reads an adaptive code vector based on the index of the adaptive code book from adaptive code book 153, and multiplies the obtained adaptive code vector by a decoding gain of the adaptive code book to obtain a driving excitation. Obtain an adaptive codebook component. Next, the probabilistic codebook based on the index of the probabilistic codebook is read out from the probabilistic codebook 154, and the obtained probabilistic codevector is multiplied by the decoding gain of the probabilistic codebook to generate a driving excitation probabilistic codebook. Get the ingredients. Further, a driving excitation is obtained by adding the driving excitation adaptive codebook component and the driving excitation stochastic codebook component, and the obtained driving excitation is output to LPC synthesis section 156 and adaptive codebook 153. Here, the old code vector in adaptive codebook 153 is updated with the above-mentioned driving excitation inputted from driving excitation generating section 155.

[0006] The LPC synthesizing section 156 includes a driving sound source generating section 1
55. The parameter decoding unit 1
LPC synthesis is performed based on the decoded LPC coefficients obtained from 52, and the output is sent to the output section as digital output sound 157.

FIG. 6 is a functional block diagram of a conventional CELP type speech coding apparatus. The LPC analysis unit 112 first performs an autocorrelation analysis and a linear prediction analysis on a certain frame in the digital input speech 111 to perform L prediction.
Calculate the PC coefficient and quantize the LPC coefficient to obtain the LPC
Code and outputs it to the parameter coding unit 123,
The code is decoded to obtain decoded LPC coefficients. Next, the impulse response of an auditory weighting filter having characteristics such as pitch emphasis and high-frequency emphasis is obtained and output to the auditory weighting section 113, and the perceptual weighting LPC synthesis filter is obtained. The impulse response is obtained, and the perceptual weighting LPC reverse order synthesis unit A
114, an audibility weighted LPC synthesis unit A116, an audibility weighted LPC reverse order synthesis unit B119, and an audibility weighted LPC synthesis unit B121.

The perceptual weighting section 113 performs perceptual weighting filtering on the input speech data for each subframe, subtracts the zero input response of the perceptual weighting LPC synthesis filter from the output result, and searches for the sound source in the adaptive codebook. A target signal that is sometimes referred to is obtained and output to the perceptual weighting LPC reverse order synthesis unit A114 and the subtraction unit 118.

A perceptual weighting LPC reverse order synthesis unit A114
, Time-reverse-orders the target signal obtained by the auditory weighting section 113 and converts the obtained reverse-ordered signal to the LPC analysis section 11.
The impulse response given by 2 is synthesized by a perceptually weighted LPC synthesis filter having coefficients, and its output signal is again time-reversed and output to the comparison unit A117 as a time-reverse-combined output of the target signal.

The adaptive codebook 115 includes an adaptive codebook updating unit 1
24, and stores the past drive sound source information as the audibility weighting LPC synthesis unit A1.
16, is referred to as an adaptive code vector by the comparison unit A117 and the adaptive codebook updating unit 124.

The perceptual weighting LPC synthesizing unit A 116 reads out the adaptive code vector from the adaptive code book 115 and applies the LPC analysis unit 11 to the read out adaptive code vector.
Weight L with the impulse response obtained as a coefficient
The signal is synthesized by the PC synthesis filter, and the result is compared by the comparison unit A117.
Output to

The comparing unit A117 firstly receives the adaptive codebook 11
5, the square value of the inner product of the adaptive code vector read directly from No. 5 and the time-reverse synthesized output of the target signal obtained by the perceptual weighting LPC reverse order synthesizing unit A114 is obtained. The power of a signal obtained by subjecting the adaptive code vector to perceptual weighting LPC synthesis is obtained, and the square value of the inner product is divided by this power to obtain a reference value for adaptive codebook search, and the reference value is the largest. The index of the adaptive code vector read at this time and the optimal gain by which the code vector is multiplied are calculated and output to the subtraction unit 118 and the parameter encoding unit 123. This series of processing is called adaptive codebook search.

A subtraction unit 118 subtracts a signal obtained by multiplying an output signal obtained by subjecting the code vector searched for by the adaptive codebook search to a perceptual weighting LPC synthesis by a gain from the target signal obtained by the perceptual weighting unit 113, The result of the subtraction is output to the perceptual weighting LPC reverse order synthesis unit B119 as a target signal to be referred to when searching the stochastic codebook.

Audience weighting LPC reverse order synthesis section B119
Subtracts the time-reversed order of the target signal for excitation search of the probabilistic codebook generated in the subtraction unit 118, performs LPC synthesis with perceptual weighting, and time-reverse-orders the output signal again to obtain the sound source of the probabilistic codebook. A time-reverse composite output of the search target signal is obtained and output to the comparison unit B122.

The stochastic codebook 120 stores a plurality of code vectors, and these code vectors are converted by the perceptual weighting LPC synthesizing unit B121, the comparing unit B122, and the adaptive codebook updating unit 124 as stochastic code vectors. Referenced.

An auditory weighting LPC synthesizing unit B121 synthesizes the stochastic code vector read from the stochastic codebook 120 with an auditory weighting LPC synthesizing filter having the impulse response obtained from the LPC analyzing unit 112 as a coefficient.
The combined signal is output to comparison section B122.

As shown in (Equation 1), the comparing unit B122 firstly outputs the i-th stochastic code vector V (i, n) directly read from the stochastic codebook 120 and the auditory weighting LP
The square value of the inner product of the target signal and the time-reverse-combined output r (n) determined by the C-reverse-order combining unit B119 is determined, and then the squared value of the combined signal S (i, n) received from the auditory weighting LPC combining unit B121 The power is calculated, and the square value of the inner product is divided by the power to obtain a reference value std (i) of the probabilistic codebook search, and the probabilistic code read when the reference value is maximized. An index representing a vector number and an optimal gain to be multiplied by the probabilistic code vector are calculated and output to the parameter encoding unit 123. This series of processing is called a probabilistic codebook search.

[0018]

(Equation 1)

The parameter encoding unit 123 first encodes an optimal gain by which the adaptive code vector obtained from the comparing unit A117 is multiplied and an optimal gain by which the probabilistic code vector obtained by the comparing unit B122 is multiplied. The coding gain of the adaptive codebook and the coding gain of the probabilistic codebook are obtained, and then the obtained coding gain of the adaptive codebook and the coding gain of the probabilistic codebook are decoded to obtain the adaptive codebook. The decoding gain and the decoding gain of the probabilistic codebook are obtained respectively, and the coding gain of the adaptive codebook, the coding gain of the probabilistic codebook, the index of the adaptive codebook, the index of the probabilistic codebook, the LPC Output the code to the transmission unit 125,
The decoding gain of the adaptive codebook, the decoding gain of the probabilistic codebook, the index of the adaptive codebook, and the index of the probabilistic codebook are output to the adaptive codebook updating unit 124.

Adaptive codebook updating section 124 receives an input from parameter encoding section 123, first reads an adaptive code vector based on the index of the adaptive codebook from adaptive codebook 115, and reads the read adaptive code vector. Is multiplied by the decoding gain of the adaptive codebook to obtain a driving excitation adaptive codebook component, and then reads a probabilistic code vector based on the index of the probabilistic codebook from the probabilistic codebook 120 and reads the read probabilistic codebook. Multiply the code vector by the decoding gain of the probabilistic codebook to obtain the driving excitation probabilistic codebook component,
The driving excitation adaptive codebook component and the driving excitation probabilistic codebook component are added to obtain a driving excitation, and the obtained driving excitation is output to adaptive codebook 115. Here, the old code vector in adaptive codebook 115 is updated by the above-mentioned driving excitation input from adaptive codebook updating section 124.

The stochastic codebook used in CELP includes a random codebook and an algebraic codebook (“8KBIT / S ACELP CODIN”).
G OF SPEECH WITH 10 MS SPEECH-FRAME: A CANDIDATE
FORCCITT STANDARDIZATION ": R. Salami, C. Laflamme, J
-P. Adoul, ICASSP'94, pp. II-97 to II-100, 1994). Each will be briefly described.

"Noise codebook" is the most classic codebook.
It stores a random number sequence created from random numbers.
Since the nature of the codebook is random, high quality synthesized speech can be obtained. However, since all code vectors must be stored in advance, a large RO
M capacity is required. In addition, in the codebook search, perceptually weighted LPC synthesis is performed for all code vectors, so that a large amount of calculation is required.

Each code vector of the "algebraic codebook" is composed of four pulses having a magnitude of 1 (amplitude is +1 or -1), and the position of each pulse is determined by an index operation. It is characterized by that. Therefore, a codebook ROM is not required. However, since the sound is composed of a small number of pulses, the sound quality of the synthesized sound (particularly, the unvoiced part) is deteriorated. The greatest feature of the algebraic codebook is that the codebook search can be performed with a small amount of calculation. The biggest feature of the algebraic codebook is that V (i,
Since only four pulses of magnitude 1 exist in n), the value of four samples in r (n) is added to the numerator of (expression 1) (addition: 3 times), and the addition is performed. Square the result (multiplication: 1)
By storing the autocorrelation matrix of the impulse response of the perceptual weighting LPC synthesis filter obtained in advance in the RAM, the denominator of (Equation 1) is added at most 15 times (quadruple loop). The actual calculation amount is further reduced by using the property of (1) and the symmetry of the autocorrelation matrix.).

[0024]

When the noise codebook is used, the sound quality of the synthesized sound can be improved because the perceptually weighted LPC synthesis filter can be driven by a sound source having random characteristics, but the sound quality of the synthesized sound is improved. Since the calculation amount increases, the calculation amount of the codebook search increases, and the ROM capacity increases because all the code vectors are stored in advance.

On the other hand, when an algebraic codebook is used, (Equation 1)
The amount of calculation of the numerator and denominator is reduced, so the amount of calculation in codebook search is reduced, and the ROM capacity is reduced because the code vector need not be stored as it is, but the perceptual weighting LPC synthesis filter is driven by a small number of pulses. Therefore, the sound quality of the synthesized sound (especially the unvoiced part) is deteriorated.

The present invention provides a speech coding apparatus and a speech decoding apparatus which have an advantageous effect that the amount of calculation for searching for a codebook is small, the ROM capacity is small, and good synthesized speech can be provided. It is an object of the present invention to provide a sound source vector generation device and method used for the above.

[0027]

In order to solve this problem, a spread speech vector generating apparatus according to the present invention comprises: a sound source vector supply means for supplying a sound source vector;
And a convolution means for convolving the stored diffusion vector and the excitation vector to output a diffusion excitation vector, and a code vector generated by an algebraic codebook. It was used as the above excitation vector. Further, the present invention is expanded.
The method for generating a scattered voice vector includes a step of supplying a sound source vector.
Floor and a diffusion vector for diffusing the sound source vector.
Supplying, the diffusion vector and the sound source vector
And outputting a diffused sound source vector by convolving
The code vector generated by the algebraic codebook is
It was used as a vector.

[0028]

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described with reference to FIGS.

(Embodiment 1) The embodiment of the present invention
This will be described with reference to FIGS.

FIG. 1 is a block diagram showing a main part of a speech decoding apparatus according to the present invention. In FIG. 1, 1 is a sound source information storage unit, 2 is a diffusion vector storage unit, and 3 is a convolution unit. The excitation information storage unit 1 stores, for example, excitation information such as generation information of the same code vector as an algebraic codebook, and outputs the excitation information as an excitation vector. The diffusion vector storage unit 2 stores diffusion information such as a random number sequence,
The spread information is output as a spread vector. The convolution unit 3 performs convolution by inputting the sound source vector and the diffusion vector, and outputs the result as a diffusion sound source vector.

FIG. 2 shows a CELP-type speech decoding apparatus in which a main part of the speech decoding apparatus shown in FIG. 1 is used in place of a conventional stochastic codebook section. The excitation information storage unit 1 stores the same code vector generation information as the algebraic codebook, and the spreading vector storage unit 2 stores a random number sequence as a spreading vector. The parameter decoding unit 4 first transmits the speech code (LPC code, spread excitation index, coding gain of the spread excitation vector, index of the adaptive codebook, coding gain of the adaptive codebook) generated by the coding apparatus. 5 to decode the LPC code to obtain decoded LPC coefficients. Next, the decoding gain of the spreading excitation vector is decoded to obtain the decoding gain of the spreading excitation vector, and the coding gain of the adaptive codebook is decoded to obtain the decoding gain of the adaptive codebook. Further, the spread excitation index is output to the convolution unit 3, the adaptive codebook index and the decoding gain, and the diffusion gain of the spread excitation vector are output to the driving excitation generation unit 7, and the decoded LPC coefficients are output to the LPC synthesis unit 8. Output.

The convolution unit 3 first reads out the sound source vector based on the diffused sound source index from the sound source information storage unit 1, then reads out the diffusion vector stored in the spread vector storage unit 2, and further reads out the read out sound source vector. The convolution of the vector and the diffusion vector is performed to generate a diffusion sound source vector, which is output to the driving sound source generation unit 7. Driving excitation generating section 7 first reads an adaptive code vector based on the index of the adaptive code book from adaptive code book 6, multiplies the obtained adaptive code vector by the decoding gain of the adaptive code book, and generates a driving excitation adaptive code book. Then, the component is obtained, and the spread excitation vector obtained by the convolution unit 3 is multiplied by the decoding gain of the spread excitation vector to obtain the drive excitation spread excitation vector component. next,
The obtained driving excitation adaptive codebook component and the driving excitation diffusion excitation vector component are added to obtain a driving excitation, and the obtained driving excitation is output to LPC synthesis section 8 and adaptive codebook 6. here,
The old code vector in the adaptive codebook 6 is updated with the driving excitation input from the driving excitation generation unit 7. LPC
The synthesizing unit 8 synthesizes the driving sound source input from the driving sound source generating unit 7 with an LPC synthesis filter having the decoded LPC coefficient obtained from the parameter decoding unit 4 to obtain an output sound 9.

In this embodiment, since the spreading vector storage unit 2 stores a random sequence as a spreading vector, the driving sound source spreading sound source vector component has a random property, and the output sound (particularly, Silent part)
Sound quality can be improved. Further, since the diffused sound source vector can be created from the sound source information in the sound source information storage unit 1 and the diffusion vector in the spread vector storage unit 2, the ROM capacity is small.

Although the present embodiment has been described with an example in which the excitation information storage unit 1 stores the same code vector generation information as the algebraic codebook, other codebook generation information or other codebooks themselves are stored. Can also be implemented in the same manner. Further, in the present embodiment, an example has been described in which a random number sequence is stored in the diffusion vector storage unit 2. However, a case where a sequence obtained by other learning or a sequence obtained by knowledge is used can be similarly performed. It is.

Although the speech decoding apparatus according to the present embodiment is of the CELP type, it can be applied to other speech decoding apparatuses such as the VOCODER type.

(Embodiment 2) The embodiment of the present invention
This will be described with reference to FIGS.

FIG. 3 is a block diagram showing a main part of a speech coding apparatus according to the present invention. 3, reference numeral 11 denotes a diffusion vector storage unit, 12 denotes a digital filter, 13 denotes a convolution unit, and 14 denotes a sound source information storage unit. The diffusion vector storage unit 11 stores diffusion information such as a random number sequence,
The spread information is output as a spread vector.

The digital filter 12 filters and outputs the input signal and outputs filter information for determining the characteristics of the filter itself. FIG. 3 shows an output of an impulse response or a coefficient as the filter information. The excitation information storage unit 14 stores, for example, excitation information such as generation information of the same code vector as the algebraic codebook, and outputs the excitation information as an excitation vector.

In FIG. 3A, the convolution unit 13 performs convolution by inputting a diffusion vector and an impulse response or a coefficient, and outputs a new impulse response or a new coefficient. In FIG. In addition to the function of a), the sound source vector and the diffusion vector are input, convolution is performed, and the resultant is output as a diffusion sound source vector.

FIG. 4 shows a CELP-type speech coding apparatus in which the main part of the speech coding apparatus of FIG. 3 is used in place of the conventional stochastic codebook section. The excitation information storage unit 14 stores the same code vector generation information as the algebraic codebook, and the spreading vector storage unit 11 stores a random number sequence as a spreading vector. Reference numeral 15 denotes input voice, which is digital input voice data.

The LPC analysis unit 16 calculates an LPC coefficient by performing an autocorrelation analysis and a linear prediction analysis on a certain frame in the input speech 15, quantizes the LPC coefficient to obtain an LPC code, and performs parameter coding. Part 17
And decodes the LPC code to obtain a decoded LPC coefficient. Next, the impulse response of an audibility weighting filter having characteristics such as pitch emphasis and high frequency emphasis is obtained and output to the audibility weighting unit 18, and the impulse response of the audibility weighting LPC synthesis filter is obtained.
It outputs to PC reverse order synthesis part A19 and audibility weighting LPC synthesis part A20. Audience weighting LPC reverse order synthesis unit A1
9 and the perceptual weighting LPC synthesis unit A20 each include a digital filter.

The perceptual weighting unit 18 performs perceptual weighting filtering on the input speech data for each subframe, subtracts the zero input response of the perceptual weighting LPC synthesis filter from the output result, and refers to the result when searching for a sound source in the adaptive codebook. Obtain target signal and weight LPC
The data is output to the reverse synthesis unit A19 and the subtraction unit 18. The perceptual weighting LPC reverse order synthesizing unit A19 includes the perceptual weighting unit 18
, And synthesizes the inverted signal with a perceptually weighted LPC synthesis filter having the impulse response given by the LPC analysis unit 16 as a coefficient. A time inverse composite output of the signal is obtained and output to the comparison unit A21.

The adaptive codebook 22 includes an adaptive codebook updating unit 23
, And the past driving sound source information is stored as the perceptual weighting LPC synthesis unit A20,
The comparing unit A21 and the adaptive codebook updating unit 23 refer to the adaptive codebook as an adaptive code vector. The perceptual weighting LPC synthesis unit A20 first reads out the adaptive code vector from the adaptive codebook 22, and synthesizes the read out adaptive code vector with a perceptual weighting LPC synthesis filter having the impulse response obtained from the LPC analysis unit 16 as a coefficient. , And outputs the result to the comparison unit A21. Next, the hearing weighting LP
The impulse response of the C synthesis filter is output to the convolution unit 13.

First, the comparison unit A21 obtains the square value of the inner product of the adaptive code vector directly read from the adaptive codebook 22 and the time-inverse combined output of the target signal obtained by the perceptual weighting LPC reverse-order combining unit A19. Next, the power of a signal obtained by subjecting the adaptive code vector received from the perceptual weighting LPC synthesis unit A20 to perceptual weighting LPC synthesis is obtained,
By dividing the square value of the inner product by this power, a reference value for adaptive code vector search is obtained. The index of the adaptive code vector read when the reference value becomes the maximum and the optimal gain by which the adaptive code vector is multiplied are calculated, and the subtraction unit 24 and the parameter encoding unit 1
7 is output.

The subtraction unit 24 performs perceptual weighting LPC synthesis of the adaptive code vector searched for in the adaptive codebook search from the target signal obtained by the perceptual weighting unit 18, and applies the adaptive codebook obtained by the comparison unit A21 to the synthesized signal. Is subtracted from the signal obtained by multiplying by the optimum gain of. The result of the subtraction is output to the perceptual weighting LPC reverse order synthesis unit B25 as a target signal to be referred to when searching for a diffusion sound source vector.

The diffusion vector storage section 11 stores a random number sequence. The convolution unit 13 has a hearing weight LP
The convolution operation of the impulse response of the perceptual weighting LPC synthesis filter received from the C synthesizing unit A20 and the diffusion vector read from the diffusion vector storage unit 11 is performed, and the calculation result is used as a new impulse response, and the perceptual weighting LPC
Output to the reverse order synthesis unit B25 and the perceptual weighting LPC synthesis unit B26. Both the perceptual weighting LPC reverse order synthesizing unit B25 and the perceptual weighting LPC synthesizing unit B26 include digital filters.

The perceptual weighting LPC reverse order synthesis unit B25
The target signal at the time of searching for the diffused sound source vector generated by the subtraction unit 24 is time-reversed, and the inverted signal is synthesized by an audibility weighting LPC synthesis filter having the new impulse response obtained from the convolution unit 13 as a coefficient. The output signal is time-reversed again and output to the comparison unit B27 as a time-reverse synthesized output of the target signal.

Since the excitation information storage unit 14 stores the same code vector generation information as the algebraic codebook, it can generate a code vector composed of a small number (here, four) of pulses. The generated code vector is referred to as a sound source vector by the auditory weighting LPC synthesis unit B26, the comparison unit B27, and the convolution unit 13. The audibility weighting LPC synthesis unit B26 synthesizes the sound source vector read from the sound source information storage unit 14 with an audibility weighting LPC synthesis filter having the new impulse response obtained from the convolution unit 13 as a coefficient, and the autocorrelation matrix of the synthesized signal And outputs it to the comparison unit B27.

The comparing section B27 firstly outputs the sound source information storing section 1
The sound source vector read directly from No. 4 and the audibility weighting L
The square value of the inner product of the target signal obtained by the PC reverse order synthesis unit B25 and the time reverse synthesis output is obtained. Next, with reference to the autocorrelation matrix received from the perceptual weighting LPC synthesis unit B26, the power of the signal obtained by subjecting the sound source vector to perceptual weighting LPC synthesis is obtained, and the square value of the inner product is divided by this power. Thus, a reference value for searching for a diffusion sound source vector is obtained. A diffusion excitation index indicating the number of the excitation vector read when the reference value becomes maximum, and an optimum gain by which the diffusion excitation vector is multiplied are calculated and output to the parameter encoding unit 17.

The parameter encoding unit 17 first encodes the optimal gain by which the adaptive code vector obtained by the comparison unit A21 is multiplied and the optimal gain by which the spread excitation vector obtained by the comparison unit B27 is multiplied. The coding gain of the codebook and the coding gain of the spreading excitation vector are obtained. Next, the obtained coding gain of the adaptive codebook and the coding gain of the spread excitation vector are decoded to obtain the decoding gain of the adaptive codebook and the decoding gain of the spread excitation vector, respectively. Further, the coding gain of the adaptive codebook, the coding gain of the spreading excitation vector, the LPC code, the index of the adaptive codebook, and the spreading excitation index are output to the transmission unit 28,
The decoding gain of the adaptive codebook and the index of the adaptive codebook are output to the adaptive codebook updating unit 23, and the decoding gain of the diffusion excitation vector and the diffusion excitation index are convolved with the convolution unit 13.
Output to

The convolution unit 13 first reads out a sound source vector based on the diffusion sound source index from the sound source information storage unit 14, and then reads out a diffusion vector stored in the diffusion vector storage unit 11. Next, a convolution operation of the read-out excitation vector and the diffusion vector is performed to generate a diffusion excitation vector, and the diffusion excitation vector is multiplied by a decoding gain to obtain a driving excitation diffusion excitation vector component. Output. The adaptive codebook updating unit 23 first converts the adaptive codebook based on the index of the adaptive codebook received from the parameter encoding unit 17 into the adaptive codebook 2.
2 to obtain a driving excitation adaptive codebook component by multiplying the obtained adaptive code vector by the decoding gain of the adaptive codebook. Next, a driving excitation is generated by adding the obtained driving excitation adaptive codebook component and the driving excitation diffusion excitation vector component input from the convolution unit 13, and the generated driving excitation is output to the adaptive codebook 22. Here, the old code vector in the adaptive codebook 22 is updated by the driving excitation input from the adaptive codebook updating unit 23.

In the present embodiment, an example has been described in which the excitation information storage unit 14 stores the same code vector generation information as the algebraic codebook, but other codebook generation information or another codebook itself is stored. Can also be implemented in the same manner. Further, in the present embodiment, an example in which a random number sequence is stored in the diffusion vector storage unit 11 has been described. However, a case where a sequence obtained by other learning or a sequence obtained by knowledge is used can be similarly performed. It is. Although the speech coding apparatus according to the present embodiment is of the CELP type, it can be applied to other speech coding apparatuses such as a VOCODER type.

[0053]

As described above, according to the present invention, the following advantages can be obtained: the amount of calculation for codebook search can be reduced, the ROM capacity can be reduced, and a better synthesized sound can be provided. Can be

[Brief description of the drawings]

FIG. 1 is a block diagram showing a main part of a speech decoding apparatus according to an embodiment of the present invention;

FIG. 2 is a block diagram showing a CELP-type speech decoding apparatus according to an embodiment of the present invention;

FIG. 3A is a block diagram showing a main part of a speech coding apparatus according to an embodiment of the present invention; FIG. 3B is a block diagram showing a main part of the speech coding apparatus according to one embodiment of the present invention;

FIG. 4 is a block diagram showing a CELP-type speech encoding apparatus according to an embodiment of the present invention;

FIG. 5 is a block diagram showing a conventional CELP-type speech decoding device.

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

[Description of Code] 1 sound source information storage unit 2 diffusion vector storage unit 3 convolution unit 4 parameter decoding unit 5 transmission unit 6 adaptive codebook 7 driving excitation generation unit 8 LPC synthesis unit 9 output sound 11 diffusion vector storage unit 12 digital filter Reference Signs List 13 Convolution unit 14 Sound source information storage unit 15 Input speech 16 LPC analysis unit 17 Parameter encoding unit 18 Perception weighting unit 19 Perception weight LPC reverse order synthesis unit A 20 Perception weight LPC synthesis unit A 21 Comparison unit A 22 Adaptive codebook 23 Adaptive code Book update unit 24 Subtraction unit 25 Perception weighting LPC reverse order synthesis unit B 26 Perception weighting LPC synthesis unit B 27 Comparison unit B 28 Transmission unit

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-2-280200 (JP, A) JP-A-2-282800 (JP, A) JP-A-6-130994 (JP, A) JP-A-6-130994 195098 (JP, A) JP-A-9-34498 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 19/00-19/14 H03M 7/30 H04B 14/04

Claims (4)

(57) [Claims] 1. A sound source vector supply means for supplying a sound source vector, a means for storing a diffusion vector for spreading the sound source vector, and a convolution of the stored diffusion vector and the sound source vector to generate a spread sound source vector. and a convolution unit outputs, the use of <br/> code vectors generated by the algebraic codebook as the excitation vector
Diffusion sound source vector generator characterized by the following . Wherein the step of supplying the excitation vector, the sound
Supplying a diffusion vector for spreading the source vector, and convolving the diffusion vector and the excitation vector to output a diffusion excitation vector , wherein the code vector generated by the algebraic codebook is converted to the excitation vector. A method for generating a diffuse sound source vector, wherein the method comprises: 3. The stored diffusion vector may be learned or
Is a sequence determined by one of the findings.
2. The diffused sound source vector generation device according to claim 1, wherein: 4. The method according to claim 1, wherein the supplied diffusion vector is learned or learned.
Is a sequence determined by one of the findings.
3. The method according to claim 2, wherein