JP3290704B2 - Vector quantization method - Google Patents

Abstract

PURPOSE:To attain the miniaturization of an arithmetic circuit and economization and to perform an arithmetic operation with high accuracy by performing the arithmetic operation for retrieval with high accuracy and with small arithmetic quantity by using a DSP in form of fixed point. CONSTITUTION:Normalization parts 53, 57 are arranged at the front stages of correlation calculation parts 51, 55 in the retrieval loop of a pitch cycle and the retrieval loop of a code book, respectively, and after each element of input speech vector is normalized by the minimum power of 2 exceeding the maximum value of the absolute value at the normalization parts 53, 57, and it is supplied to the correlation calculation parts 51, 55, and correlation with a voice synthesis vector is found, and a square error can be calculated by a square error calculation parts 41, 49 based on the correlation value and the value of sum of squares of the voice synthesis vectors calculated at sum of squares calculation parts 52, 56.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vector quantization system used for encoding a signal sequence such as an audio signal and a video signal.

[0002]

2. Description of the Related Art Vector quantization is one of the techniques that are currently receiving the most attention as a technique for efficiently encoding audio signals and image signals. In particular, in the field of audio signal coding, CELP (Code Excited Linear Predic
) system or VXC (Vector Excited Coding) system is known as an excellent system to which vector quantization is applied. For details on the CELP method, see MRSc
hroeder and BSAtal's “Code-Excited Linear Pred
iction (CELP): High-Quality Speech At Very LowBit Ra
tes ”in Proc. ICASSP, 1985. pp. 937-939.

[0003] The CELP system will be briefly described below with reference to the drawings. FIG. 15 is a block diagram showing a basic configuration of a speech encoding circuit to which the CELP method is applied.
In the figure, a signal sequence of an audio signal is guided to a block extracting unit 12 via an input terminal 11, and the block extracting unit 12 divides L sample values into one frame as an input audio signal vector having a length L. Is done. Then, the audio signal vector is introduced into the LPC analysis unit 13.
The LPC analysis unit 13 performs an LPC analysis of the audio signal by an autocorrelation method or the like, and thereby extracts LPC prediction parameters {ai} (i = 1,..., P). The LPC prediction parameter is supplied to the LPC synthesis filter 19. Note that P is the prediction order.

Further, the LPC analysis unit 13 outputs an LPC
The prediction residual signal vector is output, and the LPC prediction residual signal vector is input to the pitch analysis unit 14. The pitch analysis unit 14 performs a pitch analysis, which is a long-term prediction of speech, using the LPC prediction residual signal vector, and extracts a pitch cycle Tp and a gain parameter b.
The extracted pitch period Tp and gain parameter are both supplied to the pitch synthesis filter 18.

Next, a process of generating a synthesized speech will be described. The codebook 15 stores in advance n white noise vectors of the number of dimensions (the number of elements of the vector) k. Here, k is generally selected such that L / k is an integer. Each of the white noise vectors is read out one by one from the codebook 15 and is first multiplied by the multiplier 16 by the gain parameter generated by the gain parameter generator 17. Then, the pitch synthesis filter 18 and LP
A filter operation is performed by each of the C synthesis filters 19 to generate a synthesized speech vector. The transfer function P (Z) of the pitch synthesis filter 18 and the transfer function A (Z) of the LPC synthesis filter 19 at this time are as follows.

[0006]

(Equation 1)

[0007]

(Equation 2)

[0008] The synthesized speech vector thus generated is
The comparator 20 compares the input speech vector with the target speech vector. The comparison output is a weighting filter 2
After passing through 1, the signal is introduced into the square error calculator 22, where 2
The riding distance distortion Ej is obtained. Further, the square distance distortion Ej
Is introduced into the minimum distortion search section 23i, where the minimum value is detected.

The above-described synthetic speech generation process is performed for all of the n white noise vectors, whereby the number j of the white noise vector that gives the minimum value of the square distance distortion Ej is selected. In other words, the CELP method is characterized in that vector quantization is performed by using a codebook as a driving signal of a synthesis filter in a speech synthesis process. Since the input speech vector has a length L, this process is repeated L / k times. The weighting filter 5 in the figure is used for shaping the spectrum of the error signal to reduce distortion that may be perceived by the human ear, and its transfer function is given by the following equation.

[0010]

(Equation 3)

[0011]

(Equation 4)

When an input signal is actually encoded and transmitted using the CELP method, LPC prediction parameters,
The pitch period, pitch gain parameter, codebook number, and codebook gain are each encoded and transmitted.

FIG. 16 is a block diagram showing another principle configuration example of a speech encoding circuit to which the CELP system is applied. In this figure, the same parts as those in FIG. 15 are denoted by the same reference numerals.

In FIG. 16, a weighting filter 21
Are moved from inside the search loop of the codebook 15 to outside the loop. This means that P (Z) of the pitch synthesis filter 18 is P (Z / γ),
By setting (Z) to A (Z / γ), it is possible to reduce the amount of calculation while maintaining the same function.

The initial memory in the filter operation of the pitch synthesis filter 18 and the LPC synthesis filter 19 is as follows:
The search of the codebook 15 is not affected. That is, a pitch synthesis filter 24 and an LPC synthesis filter 25 having an initial memory are newly provided. A zero input vector is output from the LPC synthesis filter 25 and supplied to a subtractor 26. This subtractor 26
Then, the zero input vector is subtracted from the weighted input speech vector output from the weighting filter 21. The comparator 20 compares the initialized weighted input speech vector output from the subtractor 26 with the output vector of the LPC synthesis filter 19. That is, a weighted input speech vector obtained by subtracting the LPC synthesis filter 25 from the output vector of the weighting filter 21 in advance becomes a target vector, and thereby the initial memories of the pitch synthesis filter 18 and the LPC synthesis filter 19 are set to zero.

At the same time, with this circuit, the filter operation performed by the synthesis filter in FIG. 15 can be represented by the product of the code vector and the next lower triangular matrix of k × k.

[0017]

(Equation 5)

Here, k is the number of dimensions (number of elements) of the code vector in the code book 15, and h (i) (i = 1,
.., K) is an impulse response of length k when the initial memory of H (Z / γ) is zero.

The speech synthesis vector output from the comparator 20 is subjected to a square error calculation section 22 to calculate the square distance distortion Ej
Is obtained, and the minimum value is obtained by the minimum distortion search unit 23.

[0020]

(Equation 6) Here, Xt is a target input vector, Cj is a j-th code vector γj is an optimal gain parameter for the j-th code vector. Finding the above Ej,
FIG. 17 shows a flowchart for obtaining the vector number j giving the minimum value.

On the other hand, as a method for improving the sound quality of the CELP method, a method called closed loop of pitch prediction or an adaptive codebook is known. For more details on this method, see WBKleijin, DJ Krasinski, and RH.Ket
chum, “Improved Speech Quality and Efficient Vect
or Quantization in CELP ”in Proc ICASSP, 1988.pp.1
55-158.

A CELP using a method called a closed loop of pitch prediction or an adaptive codebook will be briefly described below.
The method will be described. The difference between the CELP method shown in FIG. 15 and the CELP method using a closed loop of pitch prediction or a method called an adaptive codebook lies in a pitch analysis method. In the former, pitch analysis is performed using the LPC prediction residual signal vector output from the LPC analysis unit 13. On the other hand, the latter is characterized in that the pitch analysis is closed-looped as in the codebook search.

In the CELP system using a method called a closed loop or adaptive codebook for pitch prediction, a drive signal of an LPC synthesis filter is delayed by a delay device that is variable over a pitch search range a to b. Generates a drive signal vector for the pitch period j. Then, after multiplying the drive signal vector by a predetermined gain parameter, a filter operation is performed by an LPC synthesis filter,
The calculated output is compared with a weighted input vector, which is the target vector and excludes the influence of the previous frame. Then, based on the comparison result, a square distance distortion Ej of the following equation is obtained.
The drive signal vector at which the square distance distortion Ej is minimized is searched for.

[0024]

(Equation 7)

Here, j is a ≦ j ≦ b, Xt is the target vector, Bj is the drive signal vector at the pitch period j, γj is the optimal gain factor for the pitch period j, and H is the aforementioned (5) H × k lower triangular matrix given by the equation, h
(i) (i = 1,..., k) is an impulse response of length k when the initial memory of A (Z / γ) is zero. FIG. 18 is a flowchart for obtaining the above-mentioned square distance distortion Ej and obtaining the pitch period j which gives the minimum value.

[0026]

By the way, recently, a voice codec adopting the CELP system of this type includes a DSP.
(Digital signal processor) is being used. In this case, a floating-point DSP is generally used as the DSP because a wide dynamic range is required to calculate aj and bj in the CELP system arithmetic processing in consideration of the fluctuation of the input audio level. However, the floating-point DSP has a large circuit scale and is very expensive.

Therefore, an attempt has been made to use a relatively small and inexpensive fixed-point DSP. However, if this fixed-point DSP is used as it is, the calculation accuracy will be reduced, and the voice quality will be deteriorated, which is very undesirable. For example, when a codebook search is performed using a fixed-point DSP having a relatively short word length of about 16 bits, the input signal vector Xt varies depending on the input signal level. As a result, it is difficult to accurately determine the minimum value of the square distance distortion Ej. A similar fixed-point D
When a pitch period search called a closed loop or an adaptive codebook is performed by using the SP, the input signal vector Xt fluctuates according to the input signal level, and bj or Bj fluctuates with time. Therefore, a wide dynamic range is required, and as a result, it is difficult to accurately obtain the minimum value of the square distance distortion Ej.

The present invention has been made in view of the above circumstances. A first object of the present invention is to perform a calculation process for searching a code book with high precision using a fixed-point type calculation circuit. It is an object of the present invention to provide a vector quantization system which can perform an operation with a small amount of operation, thereby making it possible to reduce the size and cost of an operation circuit and perform high-precision operations at high speed.

A second object of the present invention is to make it possible to perform arithmetic processing for searching for a pitch period with high precision and with a small amount of arithmetic using a fixed-point arithmetic circuit. It is an object of the present invention to provide a vector quantization system which can reduce the size and cost of an arithmetic circuit and perform high-precision arithmetic at a high speed.

Further, a third object of the present invention is to perform a search operation of a vector pitch period using a fixed-point arithmetic circuit by considering not only the level fluctuation of the input signal but also the level fluctuation of the drive signal. An object of the present invention is to provide a vector quantization method that can be performed with higher accuracy.

[0031]

In order to achieve the first object, according to the present invention, a predetermined filter operation is performed on each of a plurality of code vectors previously stored in a code book, and the obtained operation is performed. A vector quantization method of calculating a square error between a vector and an input signal vector and searching for a code vector in which the square error is minimized from the code book using a fixed-point arithmetic circuit. In, the elements of the input signal vector before being subjected to the calculation of the square error are normalized by the minimum power of 2 exceeding the maximum value of the absolute values of these elements,
The normalized input signal vector is used for calculating a square error together with the operation vector.

According to another embodiment of the present invention, a predetermined filter operation is performed on each of a plurality of drive signal vectors generated in accordance with the pitch period, and an operation vector obtained by the above operation is obtained. A fixed-point arithmetic circuit for calculating a square error between the input signal vector and the input signal vector and searching for a drive signal vector having the minimum square error from the plurality of drive signal vectors. In the quantization method, each element of the input signal vector before being subjected to the calculation of the square error is normalized by a minimum power of 2 exceeding the maximum value of the absolute value of each element. The signal vector is used for calculating the square error together with the operation vector.

Further, in order to achieve the third object, the present invention accumulates a plurality of drive signal vectors generated in accordance with the pitch period in an adaptive codebook, and stores each of the plurality of drive signal vectors in the adaptive codebook. A predetermined filter operation is performed on each of the drive signal vectors, and a square error between the obtained operation vector and the input signal vector is calculated, and the drive signal vector that minimizes the square error is converted to the plurality of drive signals. In a vector quantization method in which an operation to search for a vector is performed using a fixed-point type arithmetic circuit, each element of an input signal vector before being subjected to the calculation of the square error is replaced by the absolute value of these elements. Normalization is performed using the minimum power of 2 exceeding the maximum value.
Each element of the drive signal vector before being subjected to the calculation of the squared error is normalized by the smallest power of 2 exceeding the maximum value of the absolute values in the elements of all the drive signal vectors, and these normalized input signal vectors And the normalized operation vector is used for calculating a square error.

Further, in the present invention, the second normalizing means converts each element of the drive signal vector read from the adaptive codebook into a minimum value exceeding the maximum value of the absolute value among the elements of all the drive signal vectors. It may be configured to normalize by a power of two, and each element of the drive signal before accumulation in the adaptive codebook is normalized by a minimum power of two exceeding the maximum value of the absolute values of these elements. May be configured.

[0035]

As a result, according to the present invention and other inventions, each element of the input signal vector exceeds the maximum value of its absolute value when searching for the codebook and when searching for the pitch. After being normalized by the minimum power of 2, it is subjected to an operation for calculating a square error together with an operation vector. For this reason, even if the level fluctuation of the input signal occurs, the above-mentioned normalization suppresses the fluctuation without losing information, and then calculates the square error. Therefore, even when a fixed-point DSP having no wide dynamic range is used, the square error comparison operation can be performed with high accuracy. Further, since the normalization is performed only on the input signal vector, it can be realized without a large increase in the amount of calculation.

According to yet another aspect of the invention, when searching for a pitch, each element of the input signal vector is normalized by a minimum power of two exceeding the maximum of the absolute values of these elements, and The minimum 2 that each element of the vector exceeds the maximum absolute value among the elements of all drive signal vectors
Are normalized by exponentiation, and then subjected to an operation for calculating a square error. For this reason, even if the input signal level fluctuates and the drive signal level fluctuates, the fluctuation is suppressed by the above-mentioned normalization without loss of information, and is then used for calculating a square error. Will be. Therefore, even when a fixed-point DSP having no wide dynamic range is used, the square error comparison operation can be performed with higher accuracy.

[0037]

【Example】

(First Embodiment) FIG. 1 is a block diagram showing a basic configuration of a speech coding circuit to which a vector quantization system according to a first embodiment of the present invention is applied.

In the figure, a signal sequence of an audio signal is inputted from an input terminal 31 and introduced into a block extracting section 32. In the block cutout unit 32, the L sample values are divided into blocks of input speech vectors having a length L as one frame, and the input speech vectors obtained by the block division are input to the LPC analysis unit 33 and the weighting filter 34.
Supplied to The LPC analysis unit 33 performs an LPC analysis of the audio signal using an autocorrelation method or the like, and extracts LPC prediction parameters {ai} (i = 1,..., P). The extracted LPC prediction parameter is the LPC synthesis filter 3
Used in 6,36,47.

The weighting filter 34 is for weighting the input speech vector, and is provided outside each of the codebook search and pitch cycle search loops. The weighting filter 34 shapes the error signal spectrum with A (Z) as A (Z / γ) by the LPC synthesis filters 36, 38, and 47, thereby reducing distortion that would be perceived by the human ear. It is possible to do. Further, by providing the weighting filter 34 outside the search loop, it is possible to reduce the amount of calculation when searching for the codebook and the pitch period. The transfer function W (Z) of the weighting filter 34 is given by the following equation.

[0040]

(Equation 8) Here, γ is (0 ≦ γ ≦ 1), and A (Z)
Is

[0041]

(Equation 9) Indicated by

Further, in the circuit of this embodiment, the initial memory used in the filter operation of the LPC synthesis filters 38 and 47 does not affect the search of the pitch period and the search of the codebook. That is, an LPC synthesis filter 36 having an initial memory is newly provided,
The LPC synthesis filter 36 creates a zero input response vector. Then, in the subtracter 35, the zero input response vector is subtracted from the weighted input speech vector output from the weighting filter 34, whereby the initial memories of the LPC synthesis filters 38 and 47 are set to zero. At the same time, this means that the filter operation of the LPC synthesis filter 38 used to search the pitch period and the filter operation of the LPC synthesis filter 47 used to search the codebook are respectively performed by the drive signal vector and k × It is possible to express the product by the product of k with the lower triangular matrix and the product of the code-vector with the lower triangular matrix of k × k. The above k × k lower triangular matrix H
Is

[0043]

(Equation 10)

Is represented as follows. Here, k is the number of dimensions (number of elements) of the drive signal vector and the code vector, and k is generally selected such that L / k is an integer. H (i) (i = 1,..., K) is an impulse response having a length k when the initial memory of A (Z / γ) is zero.

The pitch period search loop is configured as follows. That is, first, a drive signal vector corresponding to the pitch period j is accumulated in the drive signal unit 40. This drive signal vector is added by the adder 43 to the code vector multiplied by the gain parameter in the multiplier 46 and the drive signal vector multiplied by the gain parameter in the multiplier 39, and the filter drive signal obtained by this is added. e is variable-delayed by the delay unit 44. The delay unit 44 is configured such that the delay amount changes over the search range a to b of the pitch period. The drive signal vector output from the drive signal unit 40 is multiplied by a gain parameter in a multiplier 39 and then input to an LPC synthesis filter 38. The LPC synthesis filter 38 performs a filter operation for creating a synthesized speech vector based on the drive signal vector multiplied by the gain parameter.

The pitch cycle search loop has a square error calculator 41. The square error calculator 41 performs an operation for obtaining the square distance distortion Ej. The arithmetic expression is expressed as follows.

[0047]

[Equation 11]

Where j is the pitch period and a ≦ j ≦
b is set in the range. Xt is the target vector, Bj
Is a drive signal vector at pitch period j, γj is an optimal gain factor for pitch period j, and H is a k × k lower triangular matrix given by equation (10).

In the above equation (11), the square distance distortion Ej
In order to determine the code vector number j that minimizes the following equation, j that maximizes the second term of the equation may be determined. The (Xt, HBj) of the second term is divided into a weighted input speech vector from the subtractor 35 from which the influence of the previous frame has been removed and a synthesized speech vector output from the LPC synthesis filter 38. Based on this, it is calculated by the correlation calculation unit 51. In addition, {HBj} of the second term is calculated by the square sum calculator 52 based on the synthesized speech vector output from the LPC synthesis filter 38. The correlation value and the sum of squares obtained by the correlation calculator 51 and the sum of squares calculator 52 are input to the square error calculator 41. In the square error calculator 41, the square distance distortion Ej is calculated as described above, and the minimum value of Ej is searched in the minimum distortion search unit 42 based on the calculated Ej. Are supplied to the codebook 40.

That is, a multiplier 39, an LPC synthesis filter 38, a comparator 37, a
In a loop that returns to the drive signal unit 40 via the sum of squares calculation units 51 and 52, the square error calculation unit 41, and the minimum distortion search unit 42, the weighted input from which the influence of the frame before being output from the subtractor 35 is removed A pitch search is performed using a speech vector as a target vector.

By the way, a normalizing section 53 is provided before the correlation calculating section 51. The normalization unit 53 normalizes each element of the input speech vector output from the subtractor 35 by the minimum power of 2 exceeding the maximum value of the absolute value of each element. This normalization allows the input speech vector to be represented by a short word length without depending on its level fluctuation. FIG. 2 and FIG. 3 are flowcharts showing the calculation procedure when the above-described search for the pitch period is realized by the fixed-point DSP.

In the figure, first, after the value of t is initialized in step S201, the input signal vector Xt is normalized in steps S202 to S204. That is,
First, in S202, the maximum value xMAX of the absolute value of the element of the input signal vector Xt is obtained. Next, in step S203, x
Minimum power of 2 over MAX 2 ^N Is obtained, and in step S204, this 2 ^N Is used to normalize each element of the input signal vector Xt. Thus, a normalized input signal vector Xt 'is obtained.

This normalized input signal vector Xt '
Is determined, the pitch period j that gives the minimum value of the square distance distortion Ej is selected in steps S205 to S212. As the procedure, first, in step S205, HBa is obtained by a matrix-vector product operation. Then, based on this HBa, in steps S206 and S207, respectively, {HB
a ‖ and (Xt ', HBa) ² Is required. When the initial elements are obtained in this manner, the process proceeds to steps S208 to S21.
2, the other elements are recursively calculated. That is, every time the value of j is set one by one in the range of a + 1 to b in step S208, the element HBj is recursively obtained in step S209 based on the previous element HBj-1. Then, based on the HBj, at steps S210 and S211 aj = {HBj} and bj =
(Xt, HBj) ² Is required. Step S212
Then, the magnitude of the square distance distortion Ej is determined from the above calculation result. Then, HBj is obtained for all j, and when the magnitude determination of the distortion Ej corresponding thereto is completed,
The drive signal vector having the smallest square distance distortion Ej is selected according to the result of this determination.

When the optimum pitch period j is obtained by the search for the pitch period, the comparator 37 calculates the pitch period j.
The synthesized speech vector with the optimal pitch period j is subtracted from the weighted input speech vector without the influence of the previous frame, which is output from the above, thereby obtaining the weighted input without the influence of the previous frame and the influence of the pitch. A speech vector is obtained. This input speech vector is supplied to the comparator 48 of the codebook search loop.

On the other hand, the search loop of the code book is configured as follows. That is, this loop has a codebook 45, and the codebook 45 stores n code vectors composed of white noise. The code vector read from the code book 45 is multiplied by a gain parameter in a multiplier 46 and then subjected to a filter operation for creating a synthesized speech vector in an LPC synthesis filter 47. In the comparator 48 to which the synthesized speech vector created by the LPC synthesis filter 47 is input, the synthesized speech vector is subtracted from the weighted input speech vector supplied from the comparator 37 of the pitch period search loop. Is supplied to the square error calculator 49. The square error calculator 49 calculates the distortion E due to the square distance.
j is obtained, and a search for the minimum value of the square distance distortion Ej is performed by the minimum distortion search section 50. That is, in the above loop, the weighted input speech vector output from the pitch period search loop comparator 37 excluding the influence of the previous frame and the influence of the pitch is used as the target vector, An operation for selecting a code vector number j which minimizes the square distance distortion Ej of the error from the vectors is repeatedly performed.
This operation is represented by the following equation.

[0056]

(Equation 12)

Here, X is a weighted input vector excluding the influence of the previous frame and the influence of the pitch, and Cj is j
The ith code vector, γ j, is the optimal gain factor for the j th code vector, and n is the number of code vectors.

In the above equation (12), the square distance distortion Ej
In order to determine the code vector number j that minimizes the following equation, j that maximizes the second term of the equation may be determined. (Xt H, Cj) of the second term is a weighted input speech vector excluding the influence of the previous frame and the influence of the pitch output from the comparator 37 of the pitch period search loop;
It is calculated by the correlation calculator 55 based on the synthesized speech vector output from the LPC synthesis filter 47. In the second term, {HCj} is the LPC
The sum of squares calculation unit 56 calculates the sum based on the synthesized speech vector output from the synthesis filter 47. The correlation value and the sum of squares obtained by the correlation calculation unit 55 and the sum of squares calculation unit 56 are combined with the square error calculation unit 4.
9 is input. In the square error calculating section 49, the square distance distortion Ej is calculated as described above, and the minimum value of Ej is searched in the minimum distortion searching section 50 based on the calculated Ej. Are supplied to the codebook 45.

That is, the multiplier 46, the LPC synthesis filter 47, the comparator 48, the correlation
In a loop returning to the codebook 45 via the square sum calculation units 51 and 52, the square error calculation unit 49 and the minimum distortion search unit 50, the influence of the previous frame output from the comparator 37 of the pitch search loop is used. A search is made for a codebook in which the weighted input speech vector excluding the influence of the pitch and the pitch is used as the target vector.

Incidentally, a normalizing section 57 is provided at a stage preceding the correlation calculating section 55. The normalization unit 57 normalizes each element of the input speech vector output from the comparator 37 of the pitch search loop by a minimum power of 2 exceeding the maximum value of the absolute value of each element. This normalization allows the input speech vector to be represented by a short word length without depending on its level fluctuation. FIGS. 4 and 5 are flowcharts showing the operation procedure when the above codebook search is realized using a fixed-point DSP.

In the figure, first, in step S301, HC1 as an initial element is obtained by ordinary matrix-vector product operation. Then, based on the HC1, {HC1} is determined in step S302. When the initial element is obtained in this way, the operation for recursively obtaining another element is repeatedly performed. That is, every time one value of j is set in the range of 2 to n in step S303, step S30
The element HCj based on the element HCj-1 previously obtained in step 4
Is required. Then, step S is performed based on this element HCj.
At 305, {HCj} is determined.

When HCj is obtained for all js, subsequently, every time the value of t is set in the range of 1 to L / K in step S306, the input signal vector Xt is first determined in steps S307 to S309. Is performed. That is, first, in step S307, the maximum value xMAX of the absolute value of the element of the input signal vector Xt is obtained. Next, at step S308, the minimum power of 2 2 ^N exceeding xMAX Is obtained, and in step S309, this 2 ^N Is used to normalize each element of the input signal vector Xt. Thus,
A normalized vector input signal Xt 'is determined.

This normalized input signal vector Xt '
Is obtained, in step S310, Xt ' ^T H is required. In addition, this Xt ' ^T Each time H is determined one, by steps S311 ~ step S313 for each value of j for (Xt ^'T H, Cj) ² Are obtained, and these (Xt ' ^T H, Cj) ² Then, the magnitude of the distortion Ej of the square distance is determined from the above and {HCj} obtained in step S305. 'All (Xt determined based on H' ^T thus one Xt H, Cj) ² Distance distortion E due to
When j size determination is finished, the value of t returns to step S306 is incremented, normalization processing for the input vector Xt according to the step S307 ～S313 later that accordingly, Xt 'H and (Xt' ^T H, Cj) ² And the magnitude determination of the square distance distortion Ej are repeatedly performed.

As described above, according to the present embodiment, in the search loop of the pitch period and the search loop of the codebook, each element of the input speech vector is
After normalization by the smallest power of 2 exceeding the maximum value of the absolute value in each of the correlation calculation units 51 and 5,
5 to calculate the correlation with the speech synthesis vector,
Since the square error is calculated based on the correlation value and the value of the sum of squares of the speech synthesis vector, fluctuations in the input speech level can be suppressed. For example, using a fixed-point DSP Even in the case of realization, that is, in the case of performing an operation with a short word length, the square error comparison operation can be performed with high accuracy.

In both the pitch cycle search loop and the codebook search loop, the input speech vector is normalized before the correlation operation, so that the normalization process of the target vector Xt is performed for each t. You only need to do it once. For this reason, the pitch period search range a
B and the search range 1 to n of the codebook,
It can be implemented with a relatively small increase in the amount of computation. It is possible to perform a high-speed search operation. (Second embodiment)

In the present embodiment, the initialization processing described in the first embodiment is applied to the codebook search processing and the pitch cycle search processing shown in FIGS. 17 and 18, respectively. FIGS. 6 and 7 are flowcharts showing the procedure and calculation contents of the codebook search processing, and FIG. 8 is a flowchart showing the procedure and calculation contents of the pitch period search processing. 6 and FIG. 7, the same parts as those in FIG. 4 and FIG.
In the following description, the same parts as those in FIGS.

That is, in the search process of the code book, as shown in FIG. 6, the value of j is 1 in the range of 1 to n.
HCj and aj = ‖HCj
‖ Is required. Then, when HCj is obtained for all j, the normalization of the input signal vector Xt is performed in steps S307 to S309 each time the value of t is set in the range of 1 to L / K. That is, first, in step S307, the maximum value xMAX of the absolute value of the element of the input signal vector Xt is obtained. Next, step S308
The minimum power of 2 that exceeds xMAX at 2 ^N Is obtained, and in step S309, this 2 ^N Is used to normalize each element of the input signal vector Xt. Thus, a normalized vector input signal Xt 'is obtained.

This normalized input signal vector Xt '
Is obtained, then Xt ' ^T H is required. In addition, this Xt ' ^T Every time one H is obtained, (Xt ' ^T H, Cj) ² Are obtained, and these (Xt ' ^T H, Cj) ² Then, the magnitude of the square distance distortion Ej is determined from the above and {HCj} obtained earlier. 'All (Xt determined based on H' ^T and one of Xt H,
Cj) ² When the size determination of the square of the distance strain Ej is terminated by a value of t returns to step S306 is incremented, normalization processing for the input vector Xt according to the step S307 ～S313 later that accordingly, Xt 'H and (Xt' ^T H, Cj) ² And the squared distance distortion Ej
Is repeatedly determined.

On the other hand, in the pitch period search processing,
As shown in FIG. 8, every time the value of t is set one by one in the range of 1 to L / K, the input signal vector Xt is normalized in steps S202 to S204. That is, first in S202, the maximum value xMAX of the absolute value of the element of the input signal vector Xt is obtained. Next, in step S203, xMA
Minimum power of 2 over ^N 2 ^N Is obtained, and in step S204, this 2 ^N Is used to normalize each element of the input signal vector Xt. Thus, a normalized input signal vector Xt 'is obtained.

This normalized input signal vector Xt '
Is obtained, a calculation process for selecting a pitch cycle j that gives the minimum value of the square distance distortion Ej is performed. The procedure is as follows. Each time the value of j is set one by one in the range of a to b, an element HBj is obtained.
Aj = {HBj} and bj = (Xt, HBj
² Is required. Further, the magnitude of the square distance distortion Ej is determined from the above calculation result. Then, HBj is obtained for all j, and when the magnitude determination of the distortion Ej corresponding thereto is completed, the drive signal vector having the smallest square distance distortion Ej is selected according to the determination result.

As described above, also in this embodiment, the first
In the pitch loop search loop and the codebook search loop, each element of the input speech vector is normalized by the minimum power of 2 exceeding the maximum value of its absolute value, and then squared. It is supplied to an arithmetic process for calculating an error. For this reason, it is possible to suppress the fluctuation of the input voice level, and even when implementing using, for example, a fixed-point DSP, that is, when performing an operation with a short word length, the square error comparison operation can be performed with high accuracy. Can be calculated. (Third embodiment)

According to the vector quantization method of this embodiment, a normalization unit for normalizing an input signal vector and a drive signal vector output from an adaptive codebook are normally added to a pitch period search loop of a speech coding circuit. And a normalization unit for converting the normalized input signal vector and the normalized drive signal vector into a square error calculation.

FIG. 9 is a block diagram showing the basic configuration of the speech encoding circuit in this embodiment. In this figure, the same parts as those in FIG. 1 are denoted by the same reference numerals, and detailed description is omitted. That is, a normalization unit 53 is provided at a stage preceding the correlation calculation unit 51. The normalization unit 53 normalizes each element of the input speech vector output from the subtractor 35 by the minimum power of 2 exceeding the maximum value of the absolute value of each element. This normalization allows the input speech vector to be represented by a short word length without depending on its level fluctuation. The drive signal vector output from the drive signal unit 40 is multiplied by a gain parameter in the multiplier 39, and then input to the normalization unit 54a. In the normalization unit 54a, for each element of the drive signal vector output from the multiplier 39, the minimum value exceeding the maximum value of the absolute values of the elements of all the drive signal vectors stored in the drive signal unit 40 is calculated. Normalization is performed by a power of two. This normalization allows the drive signal vector to be represented by a short word length without depending on its level fluctuation. Then, the drive signal vector normalized by the normalization unit 54a is input to the LPC synthesis filter 38.

FIGS. 10 and 11 are flowcharts showing the calculation procedure when the search for the pitch period in this embodiment is realized by a fixed-point DSP. In this figure, the same parts as those in FIG. 8 are denoted by the same reference numerals.

In the figure, after the value of t is initially set in step S201, the input signal vector Xt is normalized in steps S202 to S204. That is,
First, in S202, the maximum value xMAX of the absolute value of the element of the input signal vector Xt is obtained. Next, in step S203, x
Minimum power of 2 over MAX 2 ^N Is obtained, and in step S204, this 2 ^N Is used to normalize each element of the input signal vector Xt. Thus, a normalized input signal vector Xt 'is obtained.

Next, normalization processing is performed on the drive signal vector in steps S213 to S215. That is, first, in step S213, the maximum value of the absolute value in the elements of all the drive signal vectors stored in the drive signal unit 40 is obtained. Next, the minimum power of 2 2 ^N exceeding this maximum value There determined in step S214, the the 2 ^N Thus, in step S215, each element of the drive signal vector is normalized. Thus, a normalized drive signal vector is obtained.

When the normalized input signal vector and drive signal vector are obtained, the square distance distortion E
An arithmetic process for selecting a pitch cycle j that gives the minimum value of j is performed. First, the value of j is a
The element HBj is obtained every time one is set in order in the range of .about.b. Then, based on this HBj, aj = {HBj} and bj = (Xt, HBj) ² Is required. Further, the magnitude of the square distance distortion Ej is determined from the above calculation result. Then, HBj is obtained for all j, and when the magnitude determination of the distortion Ej corresponding thereto is completed, the drive signal vector having the smallest square distance distortion Ej is selected according to the determination result.

As described above, according to this embodiment, in the pitch period search loop, each element of the input speech vector Xt and each element of the drive signal vector are normalized.
Each of them is normalized by a minimum power of 2 exceeding the maximum value of the absolute value by 4a, and then is subjected to an arithmetic process for calculating a square error. Therefore, it is possible to suppress not only the fluctuation of the input audio level but also the fluctuation of the level of the drive signal vector.
Thus, even in the case of using, for example, a fixed-point DSP, that is, in the case of performing an operation with a short word length, the square error comparison operation can be performed with higher accuracy. (Fourth embodiment)

In the vector quantization system of the present embodiment, the pitch period search loop of the speech coding circuit stores a normalizing section for normalizing an input signal vector and a drive signal section (adaptive codebook) 40. And a normalizing section for normalizing the driving signal.

FIG. 12 is a block diagram showing the basic configuration of the speech encoding circuit in this embodiment. In this figure, the same parts as those in FIG. 1 are denoted by the same reference numerals, and detailed description is omitted. That is, a normalization unit 53 is provided at a stage preceding the correlation calculation unit 51. In this normalization unit 53,
Each element of the input speech vector output from the subtractor 35 is normalized by the minimum power of 2 exceeding the maximum value of the absolute value of each element. This normalization allows the input speech vector to be represented by a short word length without depending on its level fluctuation.

The drive signal e obtained by adding the code vector multiplied by the gain parameter in the multiplier 46 and the drive signal vector multiplied by the gain parameter in the multiplier 39 by the adder 43 is After being variably delayed by the delay unit 44, it is input to the normalization unit 54b. The normalization unit 54b performs processing for normalizing each element of the drive signal e output from the delay unit 44 by a minimum power of 2 exceeding the maximum value of the absolute value of these elements. By this normalization processing, the drive signal vector stored in the drive signal unit 40 can be represented by a short word length without depending on the level fluctuation.

FIGS. 13 and 14 are flowcharts showing the calculation procedure when the search for the pitch period in this embodiment is realized by a fixed-point DSP. In this figure, the same parts as those in FIG. 8 are denoted by the same reference numerals.

In the figure, first, the value of t is initialized in step S201, and then the input signal vector Xt is normalized in steps S202 to S204. That is,
First, in S202, the maximum value xMAX of the absolute value of the element of the input signal vector Xt is obtained. Next, in step S203, x
Minimum power of 2 over MAX 2 ^N Is obtained, and in step S204, this 2 ^N Is used to normalize each element of the input signal vector Xt. Thus, a normalized input signal vector Xt 'is obtained.

Next, in steps S216 to S218, a normalization process is performed on the drive signal e. That is, first, in step S216, the maximum value eMAX of the absolute value of the element of the drive signal e is obtained. Next, the minimum power of 2 2 ^N exceeding the maximum value eMAX Is obtained in step S217,
This 2 ^N Thus, in step S218, each element of the drive signal e is normalized. Thus, a normalized drive signal e 'is obtained.

The input signal vector X thus normalized
When t 'and the drive signal e' are determined, an arithmetic process for selecting the pitch period j that gives the minimum value of the square distance distortion Ej is performed. As a procedure, first, an element HBj is obtained every time the value of j is set one by one in the range of a to b, and then aj = {HBj} and bj = (Xt, HBj) based on this HBj. ^Two Is required. Further, the magnitude of the square distance distortion Ej is determined from the above calculation result.
Then, HBj is obtained for all j, and when the magnitude determination of the distortion Ej corresponding thereto is completed, the drive signal vector having the smallest square distance distortion Ej is selected according to the determination result.

A bypass switch 58 is connected in parallel to the normalizing section 54b. The bypass switch unit 58 is opened only during a period until the drive signal vector having the smallest square distance distortion Ej is selected, and supplies the drive signal e to the normalizing unit 54b to cause the square distance distortion Ej to be supplied. Is closed after the drive signal vector with the smallest value is selected, and the drive signal e is
Is bypassed and input to the drive signal unit 40.

As described above, according to the present embodiment, in the pitch period search loop, each element of the input speech vector Xt and each element of the drive signal e are subjected to the maximum absolute value thereof by the normalizing units 53 and 54b. Is normalized by the smallest power of two that exceeds Therefore, the input audio level X
Not only can fluctuations in t be suppressed, but also fluctuations in the level of the drive signal e can be suppressed, so that, for example, when using a fixed-point DSP, the calculation is performed with a short word length. Even in this case, the comparison operation of the square error can be performed with higher accuracy.
In addition, since the normalization processing is performed on the drive signal e before being input to the drive signal unit 40, the normalization process is performed on the drive signal vector read from the drive signal unit 40, The number of processing operations can be reduced, and efficient normalization can be performed.

The present invention is not limited to the above embodiments. For example, in each of the above embodiments, the correlation calculator 5
1, a normalization unit 53 for normalizing the input signal vector is provided. This normalization unit 53 performs a normalization process on the input speech vector and then performs a correlation operation. A normalization unit may be provided in the subsequent stage to normalize the correlation calculation value, and supply the normalized correlation value to the square error calculation unit 41. In this case, the amount of calculation for normalization is increased as compared with the above embodiments, but the improvement of the arithmetic precision by normalization is sufficiently achieved.

Further, the pitch period search process and the code book search process may be executed using different DSPs. Further, in addition to the audio codec for encoding the audio signal, the present invention may be applied to a codec for encoding a video signal and a codec for encoding other signals.

In addition, the present invention also relates to the procedure for calculating the correlation and the sum of squares, the contents of the calculation, the procedure for the normalization processing, the contents of the calculation, the procedure related to the comparison of the square error, and the contents of the calculation. Various modifications can be made without departing from the spirit of the invention.

[0091]

As described above in detail, according to the present invention, a predetermined filter operation is performed on each of a plurality of code vectors stored in a code book in advance, and two operation vectors obtained by this and an input signal vector are obtained. In a vector quantization method in which an arithmetic operation of calculating a squared error and searching for a code vector in which the squared error is minimized from the codebook is performed using a fixed-point arithmetic circuit, the calculation of the squared error is performed. Before the input signal vector is subjected to the minimum 2 that exceeds the maximum value of the absolute values of these elements.
, And the normalized input signal vector is used for calculating a square error together with the operation vector.

Therefore, according to the present invention, the arithmetic processing for searching the code book can be performed with high precision and with a small amount of arithmetic operation using the fixed-point arithmetic circuit. It is possible to provide a vector quantization system that can be reduced in size and cost and can perform high-precision arithmetic at a high speed.

According to another aspect of the present invention, a predetermined filter operation is performed on each of a plurality of drive signal vectors generated according to the pitch period, and the square error between the operation vector obtained by this and the input signal vector is calculated. In a vector quantization method of performing, using a fixed-point arithmetic circuit, an operation of calculating and searching for a drive signal vector that minimizes the square error from the plurality of drive signal vectors, Each element of the input signal vector before being subjected to the calculation is normalized by the smallest power of 2 exceeding the maximum value of the absolute value of each element, and this normalized input signal vector is squared with the above-mentioned operation vector by a square error. Is calculated.

Therefore, according to the present invention, the arithmetic processing for searching for the pitch period can be performed with high accuracy and with a small amount of arithmetic using a fixed-point arithmetic circuit. It is possible to provide a vector quantization system that can be reduced in size and cost and can perform high-precision arithmetic at a high speed.

Further, according to another aspect of the present invention, a plurality of drive signal vectors generated in accordance with a pitch period are stored in an adaptive codebook, and a predetermined filter is applied to each drive signal vector stored in the adaptive codebook. An arithmetic operation is performed, and a square error between the obtained arithmetic vector and the input signal vector is calculated, and an arithmetic operation for searching a drive signal vector having the minimum square error from the plurality of drive signal vectors is fixed. In a vector quantization method executed using a decimal point arithmetic circuit, each element of an input signal vector before being subjected to the calculation of the square error is converted into a minimum 2 that exceeds the maximum value of the absolute value of each element. Each element of the drive signal vector before being subjected to the above-mentioned square error calculation is normalized by an exponentiation, and the absolute value in all the drive signal vector elements is used. Normalized by the minimum power of 2 greater than the maximum value, it is obtained by these normalized input signal vector and normalized calculation vector to provide for the calculation of the square error.

Therefore, according to the present invention, by considering not only the level fluctuation of the input signal but also the level fluctuation of the driving signal, the search operation processing of the vector pitch period using the fixed-point arithmetic circuit can be performed with higher precision. Can be provided.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a basic configuration of a speech encoding circuit to which a vector quantization method according to a first embodiment of the present invention is applied.

FIG. 2 is a flowchart showing a first half of an operation procedure in a case where a search for a pitch period is realized by a fixed-point DSP in the first embodiment of the present invention.

FIG. 3 is a flowchart showing the latter half of the calculation procedure in the case where the search for the pitch period is realized by a fixed-point DSP in the first embodiment of the present invention.

FIG. 4 is a flowchart showing the first half of an operation procedure when a search for a code book is realized by a fixed-point DSP in the first embodiment of the present invention.

FIG. 5 is a flowchart showing the latter half of the calculation procedure when the search of the code book is realized by the fixed-point DSP in the first embodiment of the present invention.

FIG. 6 is a flowchart showing a first half of a calculation procedure of a codebook search according to the second embodiment of the present invention.

FIG. 7 is a flowchart showing a latter half of a calculation procedure of a codebook search in the second embodiment of the present invention.

FIG. 8 is a flowchart showing a calculation procedure of a pitch period search in the second embodiment of the present invention.

FIG. 9 is a block diagram showing a basic configuration of a speech coding circuit to which a vector quantization method according to a third embodiment of the present invention is applied.

FIG. 10 is a flowchart showing the first half of an operation procedure when a search for a pitch period is realized by a fixed-point DSP in the third embodiment of the present invention.

FIG. 11 is a flowchart showing the latter half of the calculation procedure in the case where the search for the pitch period in the third embodiment of the present invention is realized by a fixed-point DSP.

FIG. 12 is a block diagram showing a basic configuration of a speech coding circuit to which a vector quantization method according to a fourth embodiment of the present invention is applied.

FIG. 13 is a flowchart showing a first half of an operation procedure when a search for a pitch period is realized by a fixed-point DSP in the fourth embodiment of the present invention.

FIG. 14 is a flowchart showing the latter half of the calculation procedure in the case where the search for the pitch period is implemented by a fixed-point DSP in the fourth embodiment of the present invention.

FIG. 15 is a block diagram showing a basic configuration of a speech encoding circuit to which a conventional vector quantization method is applied.

FIG. 16 is a block diagram showing a basic configuration of a speech encoding circuit to which another conventional vector quantization method is applied.

FIG. 17 is a flowchart showing a calculation procedure of a codebook search in the speech coding circuit shown in FIGS. 15 and 16;

FIG. 18 is a flowchart showing a calculation procedure of a pitch period search in the speech coding circuit shown in FIGS. 15 and 16;

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 31 ... Input terminal of audio signal 32 ... Block extraction part 33 ... LPC analysis part 34 ... Weighting filter 35 ... Subtractors 36, 38, 4
7 LPC synthesis filter 37, 48 Comparator 39, 46 Multiplier 40 Drive signal unit (adaptive codebook) 41, 49 Square error calculator 42, 50 Minimum distortion search unit 44 Delay circuit 45 Codebooks 51, 55 ... Correlation calculators 52, 56 ... 2
Multiply-sum calculation units 53, 54a, 54b, 57 ... normalization unit 58 ... bypass switch unit.

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 19/00-19/14 H03M 7/30 H04B 14/04

Claims (5)

(57) [Claims] 1. A predetermined filter operation is performed on each of a plurality of code vectors previously stored in a code book, and a square error between an operation vector obtained by this and an input signal vector is calculated. In a vector quantization method in which a search for a code vector in which the minimum is performed from the code book is performed using a fixed-point arithmetic circuit, each of the input signal vectors before being subjected to the calculation of the square error A vector quantization method, wherein elements are normalized by a minimum power of 2 exceeding a maximum value of the absolute values of these elements. 2. A predetermined filter operation is performed on each of a plurality of drive signal vectors generated in accordance with the pitch period, and a square error between the operation vector obtained by this and an input signal vector is calculated. In a vector quantization method in which a search for a drive signal vector that minimizes a squared error is performed from the plurality of drive signal vectors using a fixed-point arithmetic circuit, the calculation before calculating the squared error is performed. A vector quantization method, wherein each element of the input signal vector is normalized by a minimum power of 2 exceeding a maximum value of the absolute value of each element. 3. A plurality of drive signal vectors generated according to the pitch period are stored in an adaptive codebook, and a predetermined filter operation is performed on each of the drive signal vectors stored in the adaptive codebook. An operation of calculating a square error between the obtained operation vector and the input signal vector and searching for a drive signal vector having the minimum square error from the plurality of drive signal vectors is performed by a fixed-point arithmetic circuit. In the vector quantization method performed using, in order to normalize each element of the input signal vector before being subjected to the calculation of the square error, by a minimum power of 2 exceeding a maximum value of an absolute value of these elements. A first normalizing means, and each element of the drive signal vector before being subjected to the calculation of the square error, Vector quantization method characterized by comprising a second normalizing means for normalizing the minimum power of two above atmospheric value. 4. The second normalizing means converts each element of the drive signal vector read from the adaptive codebook into a maximum value of the absolute value of the elements of all the drive signal vectors stored in the adaptive codebook. 4. The vector quantization method according to claim 3, wherein normalization is performed by a minimum power of 2 exceeding a value. 5. The second normalizing means normalizes each element of the drive signal before storing it in the adaptive codebook by a minimum power of 2 exceeding the maximum value of the absolute values of these elements. 4. The vector quantization method according to claim 3, wherein: