JP2947788B1 - High-speed encoding method and apparatus for speech and audio signals and recording medium - Google Patents

Abstract

An object of the present invention is to provide audio and audio signal encoding that can obtain high-quality reproduced audio with a low bit rate, a small amount of memory, and a small amount of computation. SOLUTION: In a process of calculating a distortion between a target audio signal and a reproduced audio signal, a periodic gain for periodicization is set to 1, a period for periodicization is approximated by an integer sample value, and a synthesis filter is provided. Alternatively, the impulse response of the synthesis filter considering the auditory weight is obtained, the impulse response is truncated at a length equal to or less than half of the period of the period, and the correlation matrix of the impulse response matrix is calculated using the truncated impulse response. Calculating the approximate value of the power of the reproduced sound signal using the correlation matrix of the impulse response matrix calculated from the truncated impulse response, assuming that the periodicization is replaced with the approximation process using the integer sample values described above. I do.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to predictive coding for synthesizing speech by driving a filter representing a spectral envelope characteristic of speech, music, or an acoustic signal with a sound source vector to thereby reduce a signal sequence of the speech with a small amount of information. The present invention relates to a high-efficiency audio encoding method and apparatus for digital encoding and a recording medium.

[0002]

2. Description of the Related Art In order to efficiently use radio waves in digital mobile communication and to efficiently use communication lines and storage media for voice or music storage services, a high-efficiency voice coding method is used. At present, as a method for efficiently encoding speech, original speech is divided into sections at a fixed interval of about 5 to 50 milliseconds called frames or subframes (hereinafter, frames or subframes are collectively referred to as frames). A method has been proposed in which the audio of one frame is divided into two pieces of information, namely, a linear filter characteristic representing an envelope characteristic of a frequency spectrum and a drive excitation signal for driving the filter, and each is encoded. ing.

In this method, as a method of encoding a driving excitation signal, there is known a method of encoding by separating a periodic component considered to correspond to a pitch period (fundamental frequency) of voice and other components. I have. As an example of the coding method of the driving excitation information, code driving linear prediction coding (Code
-Excited Linear Prediction (CELP). For details of the above technology, see the literature MRSchroeder and BSA
tal, "Code-Excited Linear Prediction (CELP): High Q
uality Speech at Very Low Bit Rates ", IEEEProc. IC
ASSP-85, pp. 937-940, 1985.

FIG. 4 shows a configuration example of the above-mentioned encoding method. For the speech input to the input terminal, a linear prediction parameter representing a frequency spectrum envelope characteristic of the input speech is calculated in the linear prediction analysis unit 1-1. The obtained linear prediction parameters are encoded in the linear prediction parameter encoding unit 1-2 and sent to the linear prediction parameter decoding unit 1-3. In addition, when distortion calculation is performed using spectral information of input speech, for example, by taking into account auditory characteristics in distortion calculation, the linear prediction parameter is also sent to the distortion calculation unit 1-6.

[0005] The linear prediction parameter decoding section 1-3 reproduces a synthesis filter coefficient from the received code and sends it to the synthesis filter 1-5. When the auditory characteristics are considered in the distortion calculation, the decoded linear prediction parameter may be used in the distortion calculation instead of using the linear prediction parameter before quantization in the distortion calculation unit 1-6. The details of the linear prediction analysis and examples of encoding of the linear prediction parameters are described in, for example, "Digital Speech Processing" by Sadahiro Furui (Tokai University Press). Here, the linear prediction analysis unit 1-1, the linear prediction parameter encoding unit 1-2, the linear prediction parameter decoding unit 1-3, and the synthesis filter 1-5
May be replaced with a non-linear one.

[0006] The drive excitation vector generation section 1-4 generates a drive excitation vector candidate having a length of one frame and sends it to the synthesis filter 1-5. Driving sound source vector generation section 1-4
Is roughly divided into a portion corresponding to the pitch cycle (fundamental frequency) of the voice and a portion corresponding to the other aperiodic components. The adaptive codebook 1-10 is a portion corresponding to the pitch period, and is a drive excitation vector of the immediately preceding past stored in the buffer (a drive excitation vector of one to several frames just before quantization) c (t-1 ) Is cut out at a length corresponding to a certain cycle, and the cut-out vector is repeated until the length of the frame is reached, thereby outputting a time-series vector candidate corresponding to a sound cycle component. As the “certain period”, a period in which the distortion in the distortion calculator 1-6 is reduced is selected, and the selected period generally corresponds to a pitch period of a voice in many cases.

The fixed codebook 1-11 is a portion corresponding to a component other than the basic period of speech, and stores a predetermined number of candidate vectors according to the number of bits for encoding independently of the input speech. Then, a time-series code vector candidate having a length of one frame is output therefrom. The fixed code vector candidates output from the fixed codebook 1-11 are periodicized as necessary by the periodicization unit 1-16 at a period specified by the periodic code (generally corresponding to the pitch period as described above). You. The term “periodization” refers to applying a comb filter having a tap at a specified period position, or repeating a vector cut out from the head of the vector at a length corresponding to the specified period, similarly to the adaptive codebook. When the period is a non-integer sample value, the convolution of the sampling function is used to realize the non-integer sample period. Note that the periodicization unit 1-16 may not be applied to a case where a speech itself has no pitch component or a small pitch component, such as a consonant section or a non-speech section.

Adaptive Codebook 1-10 and Periodizing Unit 1-1
6 are output to the multiplication unit 1-
13 and 1-14, respectively, the weight creating unit 1-12
Are multiplied by the weights generated in step (1), and added in the adder 1-15 to obtain a driving sound source vector candidate c. Note that the fixed codebook 1- 1 is used without using the adaptive codebook 1-10.
In particular, when encoding a signal having a small pitch periodicity such as a consonant part or a non-voice section, a configuration not using the adaptive codebook 1-10 may be used in order to save bits. Many.

The synthesis filter 1-5 is a linear filter that uses the output of the linear prediction parameter decoding section 1-3 as a filter coefficient, and outputs a reproduced sound candidate y with a driving excitation vector candidate c as an input. Generally, the order of the synthesis filter 1-5, that is, the order of the linear prediction analysis, is generally about 10 to 16 order. As described above, the synthesis filter 1-5 may be a non-linear filter.

The distortion calculator 1-6 includes a synthesis filter 1-
The distortion between the output y of the reproduced voice candidate y and the input voice x is calculated. The calculation of this distortion is often performed taking into account the coefficients of the synthesis filter or unquantized linear prediction coefficients, for example, perceptual weighting.

The codebook search control unit 1-8 selects a periodic code, a fixed code, and a weight code that minimize the distortion between each reproduced voice candidate y and the input voice x, and determines the driving excitation vector in the frame. decide. At this time, ideally, an optimal combination of a periodic code, a fixed code, and a weight code that minimizes distortion should be selected. However, an enormous amount of processing is required, and processing is performed in a realistic time. It is difficult. So in practice, the period codes → weighting code of an adaptive codebook component g _a → or order of the fixed code → Weight Codes of the fixed codebook component g _r, the provisional value of the period codes → weighting code of an adaptive codebook component g _a → decide the order of the fixed code → weighting code (both the fixed codebook component g _r and the adaptive codebook component g _a), is often regarded as optimal combination with a result of deciding the order.

When each code is determined in order as described above,
When searching for the optimal codebook component of the optimal periodic code and weight code, the search is performed assuming that there is no fixed codebook (assuming that the fixed codebook component is zero). FIG. 5 shows a configuration example in which the fixed code and the weight code are determined after the adaptive codebook components of the periodic code and the weight code are determined. For simplification of the drawing, FIG. 5 omits a part related to the linear prediction among the units described in FIG.

[0013] destination weight previously determined in the adaptive codebook output v _a which is determined or the provisional value of the weight, multiplied by the multiplication unit 2-13 generates an adaptive code vector c _a. Subtracting the vector through the c _a synthesis filter 2-18 from the input speech x. This is called the reference speech x _r of the fixed codebook component, when determining the fixed code and weighting code, as not adaptive codebook component, a playback sound generated by the components of the fixed codebook, the reference speech x _r The optimum value is searched so as to minimize the distortion. In the case where the adaptive codebook is not used, the input voice and the reference voice are the same.

FIG. 5 shows an adaptive codebook and a fixed codebook.
Although the search method is a configuration using one codebook, it can be generalized to a case where three or more codebooks are used, for example, when a fixed codebook has a two-stage configuration. Simply put, a reference speech is created by subtracting the reproduction speech component obtained from the already determined or provisionally determined codebook output from the input speech, and there is no codebook other than the codebook to be searched from now on. Specifically, a code that minimizes distortion from the reference voice is searched for. In the above-described procedure for determining codes in order, instead of deciding only one code in order, some candidates are left at an intermediate stage, and finally the most suitable combination is selected from those combinations. In some cases, this is called a delayed decision.

The excitation code (periodic code, noise code, weight code) determined by codebook search control section 1-8 in FIG. 4 and linear prediction parameter code output from linear prediction parameter coding section 1-2. Is sent to the code sending section 1-9 and stored in a storage device or sent to the receiving side via a communication path according to the form of use.

A problem in such a CELP method is that a large amount of arithmetic processing is required for distortion calculation for selecting a driving excitation vector candidate, particularly for distortion calculation for searching a fixed codebook. It is becoming. To address this problem, a pulse-driven CELP or a CELP scheme using a sparse codebook has been proposed. In each case,
A vector having a value in only some elements (sample points) of the fixed code vector output from the fixed codebook and having a value of 0 in other elements (sample points) is used as a fixed code vector.

One example of the former is a pulse-driven CE.
The LP does not store the fixed codebook as a vector pattern of a frame length, but puts several pulses having a height of 1 in an appropriate position in a frame, for example, four pulses in a frame of 80 samples. This is a method of setting a fixed code vector by setting up. By adopting such a pulse-type driving sound source and devising the calculation order in the distortion calculation, the number of calculation processes can be reduced as compared with the conventional CELP method. Hereinafter, an example of distortion calculation in the pulse driving type method will be described with reference to the drawings.

FIG. 6 shows an example of a configuration for calculating distortion in consideration of auditory weighting in the configuration for searching for a fixed codebook in FIG. 6, when searching fixed codebook (vector shape fixed codebook) generally weights g _r is searched as may take any value, or determines the optimum fixed code base vector, or a suitable After searching for a weight code after narrowing down the candidates for the fixed code vector to a small number,
The weight codebook portion is omitted, and the weight is simply expressed as g. g
May take any value for each fixed code vector. The auditory weighting is configured in the form of an auditory weight filter using unquantized linear prediction parameters or quantized synthetic filter coefficients.

The reproduced speech candidate y _r output from the synthesis filter 3-1 is passed through the perceptual weighting filter 3-2, also with the reference speech x _r which is passed through the perceptual weighting filter 3-3 strain Is calculated. Here, the auditory weight filter 3
Since −2 and 3-3 usually use the same filter coefficient,
The hearing weight filters 3-2 and 3-3 are provided by a distance calculator 3-4.
Is equivalent even if it is inserted as one filter after, but from the viewpoint of the processing amount, as shown in FIG.
It is often divided into two places just before 4.

[0020] In the distance calculating section 3-4 measures the distance between the auditory weighted reference speech x _w and auditory weighted reproduced speech candidate y _w. The distance scale at this time is, for example,

(Equation 1) Such a distance scale may be used. A driving sound source vector that minimizes the distance scale of the equation (1) is selected. The auditory weight filters 3-2 and 3-3 are filters for performing distortion calculation to reduce noise in the reproduced voice by using human auditory characteristics, and need not always be used.

In the configuration shown in FIG. 6, the driving sound source vector candidates are combined with the synthesis filter 3-1 and the auditory weighting filter 3-2.
In order to execute the operation of passing through the filter at high speed, it is preferable to combine these two filters into one auditory weighted synthesis filter having equivalent filter characteristics. In order to make an equivalent one filter, for example, it can be expressed by an FIR filter that uses an impulse response from an input of the synthesis filter 3-1 to an output of the auditory weight filter 3-2 as a filter coefficient. At this time, since the synthesis filter is an IIR filter, in order to realize equivalent filter characteristics with the FIR filter, an FIR filter having the number of taps of the frame length is required.

The frame length (the number of samples or the vector length) is N, the impulse response (coefficient sequence of the FIR filter) is h ₀ , h ₁ , h ₂ ,..., H _N−1, and these impulse responses are Using the following impulse response matrix H
Is defined.

(Equation 2) Further, the function of the periodic unit 3-7 can also be represented by a matrix, and this matrix is represented by P.

Using the above matrices H and P, minimizing equation (1) on the premise that the weight value g can take an arbitrary value is equivalent to maximizing equation (3) below. Is equivalent to

(Equation 3) The search for the codebook by maximizing the above equation (3) is based on the fact that (HP) ^t (HP) and _xw ^t (HP) need only be calculated once at the beginning of the search for one frame. This is because a high-speed search can be realized when a pulse-driven or sparse codebook is used. In the above equation, the symbol t represents transposition of a vector and a matrix. Also,
The denominator in the equation (3) corresponds to the power (sum of squares of the vector) of the combined signal.

FIG. 7 shows an example of a configuration of a method for quickly obtaining the value of the above equation (3). The hearing weighted synthesis filter impulse response calculation unit 4-1 calculates the impulse response of the hearing weighted synthesis filter unit 3-9 in FIG. FIG.
So although periodic unit 3-7 is between the fixed codebook synthesis filter, in the calculation of HPV _r in the configuration of FIG. 7, H
Since the value of P is calculated first, the periodicization unit 4-2 is arranged next to the impulse response calculation unit 4-1 to periodicize the impulse response h.

In the correlation matrix calculation section 4-5, the periodicization section 4-
The correlation matrix (HP) ^t (HP) is calculated using the periodicized impulse response, which is the output of 2. The convolution unit 4-4 calculates x _w ^t (HP) using the impulse response that has been similarly periodized and the reference speech x _w weighted with the auditory sense. The final distance scale calculation unit 4-7 uses the output matrix of the correlation matrix calculation unit 4-5, the output vector of the convolution unit 4-4, and the fixed code vector output from the fixed codebook 4-6 to obtain ( 3) Calculate the value of equation.

[0026]

A problem when performing the distortion calculation using the configuration of FIG. 7 is when the frame is long. For example, if the sampling frequency is 8 kHz and the frame length is 10 ms, the frame length is 80 points, and the size of the correlation matrix (HP) ^t (HP) is 80 × 80.
Becomes For this reason, a matrix calculation having 6400 elements must be performed, which requires a large amount of memory and arithmetic processing.
In order to realize high-efficiency voice encoding at a low bit rate, the above-mentioned frame must be lengthened. Therefore, low-bit-rate voice encoding with a small amount of computation is realized using the method of FIG. That was impossible.

One way to solve this problem is as follows:
There is an “audio signal encoding method” (Japanese Patent Application No. 9-040404) filed by the present inventors. FIG. 8 shows a configuration example of this method. In this method, first, an impulse response of an auditory weighted synthesis filter used for distortion calculation is truncated halfway to obtain a finite-length FIR filter. The reference voice is once subjected to an inverse filter of the synthesis filter, and then to the FIR filter truncated to the above-mentioned finite length, and the truncation distortion of the impulse response is also superimposed on the reference voice. Figure 6
In the configuration of FIG. 7, the driving sound source vector is pitch-periodized, but in the configuration of FIG. 8, a pitch inverse filter is inserted on the reference voice side.

By using this method, the correlation matrix H
The size of _f ^t H _f can be made very small. For example, assuming that the impulse response is terminated by 5 taps, the correlation matrix only needs to be calculated in a matrix of size 5 × 5. With this method, even if the frame is long, encoding can be realized with a very small amount of processing.Furthermore, the distortion between the reference speech with the impulse response censored and the synthesized speech with the censored distortion also superimposed. Since the calculation is performed, the distortion component due to the truncation is canceled between the two, and as a result, there is an advantage that even if the impulse response is terminated with a short tap, the sound quality degradation can be suppressed to a small extent.

However, in the method shown in FIG. 8, quality deterioration cannot be avoided, albeit little, at the cost of realizing high-speed operation. The main factor of the quality deterioration is that the pitch period on the original drive sound source vector side is put into the reference sound side as a pitch period inverse filter. The truncation of the impulse response itself has little effect on quality. Therefore, it is desired to realize a high-speed calculation by terminating the impulse response while keeping the pitch period on the driving sound source vector side. However, simply returning the pitch-periodization unit to the driving sound source vector side, even if the impulse response h is truncated to a finite length, Ph whose pitch-period of the impulse response does not become a finite length, realizes high-speed operation. Can not.

An object of the present invention is to provide a low bit rate, low memory amount within a range permitted by an inexpensive processor, and a small amount of operation, so that high quality reproduced sound can be obtained by voice or music or other processing. It is an object of the present invention to provide a method and apparatus for digitally encoding an audio signal and a recording medium.

[0031]

According to the present invention, a synthesis filter is driven by a driving sound source vector created by using a time-series vector extracted from a codebook using a vector periodicized at a cycle corresponding to a basic cycle of speech. In the encoding to determine a driving sound source vector such that the distortion between the target audio signal and the reproduced audio signal is minimum or similar to the minimum, in the process of calculating the distortion, , The periodization gain of the periodization is set to 1, the period for the periodization is approximated by an integer sample value, and the impulse response of the synthesis filter or the synthesis filter in consideration of the auditory weight is obtained. The period is truncated to a half or less of the period, the correlation matrix of the impulse response matrix is calculated using the truncated impulse response, and the period is calculated by the integer sampling. It assumes that replace approximation by value, by using a correlation matrix of the censored impulse response from the calculated impulse response matrix, and calculates the approximate value of the power of the acoustic signal reproduced.

[0032]

DESCRIPTION OF THE PREFERRED EMBODIMENTS §1. Overview In the present invention, the pitch period variation range indicated by the periodic code is obtained by approximating both the numerator and the denominator pitch periodic matrix P or only the denominator pitch periodic matrix in the equation (3) by the integer sample point pitch. In consideration of the above, by setting the truncation order of the impulse response within a certain range, high-speed distortion calculation using a matrix of a small size can be realized while the pitch periodization unit is kept on the side of the driving sound source vector. .

§2. Embodiments Embodiments of the present invention will be described below with reference to figures and tables. In the present embodiment, the encoding method according to the present invention is specifically executed by a personal computer. That is, the encoding method described below uses a semiconductor memory (ROM, ROM, etc.) as a control program for controlling the CPU (central processing unit) of the personal computer.
RAM, etc.) or other recording media (magnetic disk, etc.)
Is stored in Then, the CPU encodes audio based on the control program.

First, the period gain of the pitch period is set to 1. Further, in (3), replace the pitch period of matrix P in the denominator in the approximate matrix P _a. Pitch cycle of matrix P of the molecule may or may not replaced be replaced by a P _a, but should not replace good in terms of sound quality. Here, the approximate matrix P _a pitch period of matrices, when the pitch period is a non-integer sample values, for example (in the case of 8kHz sampling corresponds to approximately 173Hz) 46.25 Sample was, integers, for example 46 samples Is a pitch periodic matrix created by approximating

In general, the pitch period of a non-integer sample period requires a process of convolving a sampling function with a past signal sequence. However, when the integer sample value is used and the period gain is 1, the past sample value is used. May simply be repeated or simply added. The following is an example of a pitch periodic matrix in which the periodic gain is set to 1 and the period is approximated to an integer sample value.

(Equation 4) Further, the impulse response h is truncated to a finite length in the same manner as described in the above-mentioned “Audio Signal Coding Method” (Japanese Patent Application No. 9-040404), that is, the impulse response before the truncation is regarded as 0. .

An impulse response matrix created using the truncated impulse response is represented by H _f , and is replaced with H of the numerator and denominator. Then, equation (3) becomes the following equation (4)
It is represented by an expression.

(Equation 5) Equation (4) can be regarded as an approximation of equation (3),
(3) Instead of searching for a code vector v _r to maximize, to search for a code vector which maximizes the equation (4). The denominator of the equation (4) corresponds to the power of a signal synthesized by approximating the pitch period period with an integer sample value.

In equation (4), the impulse response
By shortening the cutting order, the correlation matrix H _f ^tH_fThe size of the non
Can always be smaller. For example, 5 taps on the impulse response
If the correlation matrix is assumed to be truncated, the correlation matrix has a size of 5 × 5.
It is only necessary to calculate the matrix. Correlation matrix H_f ^tH_fCalculate
After that, P_a ^tH_f ^tH_fP_aIs calculated. At this time,
The censored order of the lus response is less than half the pitch period. "
Under the circumstances, very simple processing_f ^tH_fTo P_a ^tH_f
^tH_fP_aIs required. "Truncation of impulse response
The number is less than half the pitch period ''
Cycle is 20 samples (8 kHz sampling
Is equivalent to 400 Hz).
Corresponds to the case where the order of cutting is 10th order (10 taps) or less
You.

The human pitch frequency is at most 400H
It is said to be equal to or less than z. For example, if the impulse response is terminated by 5 taps, this condition is always satisfied. The truncation order of the impulse response has a small effect on the quality degradation of the reproduced sound up to about 3 taps, so it is preferable to set the order to about 3 to 10 taps from the viewpoint of the quality of the reproduced sound and the amount of arithmetic processing.

[0039] FIG. 1 shows a schematic diagram showing that ^{_{P a t H f t H f}} P a can be calculated by a simple process from H _f ^t H _f. This figure assumes that the frame length is 20 samples and the pitch period is 12
Samples, in the case of a 5 to truncation order of the impulse response is an example of a matrix ^{_{P a t H f t H f}} P a. Actually, when the present invention is applied, the effect is great when the frame length is as long as 80 points, for example, and as described above, the human pitch frequency does not exceed 400 Hz. Not even 12 samples. Therefore, the example of FIG. 1 has different conditions from the example when the present invention is actually applied. However, the example under the above conditions is shown in order to simplify the drawing and to simplify the description for easy understanding. . Further, the matrix ^{_{P a t H f t H f}} P a is which is always a symmetric matrix, triangular matrix portion of the upper right are omitted. The value at the i-th row and the j-th column is the same as the value at the j-th row and the i-th column. The same symbol indicates that the value is the same. For example, elements are denoted with m ₃₄ all have the same value. The features of this matrix are listed below.

Feature 1: Band-like from upper left to lower right
And the elements between the bands have a value of zero. Feature 2: Shift the band left or right by the pitch period.
And the values match exactly. Feature 3: m_ijIn terms of the portion represented by, 0 ≦ i
≦ 4, i ≦ j ≦ 4, the values are independent, otherwise m_i4(0
≦ i ≦ 4) are merely arranged diagonally. k_ijso
The same applies to the portions described. Feature 4: m_ijAnd k_ijHave a close relationship with k_ij= M_ij+ M_{4- (ji), 4} (0 ≦ i ≦ 4, i ≦ j ≦
4) If you use the above features 1-4,
Correlation matrix H of pulse response _f ^tH_f, Ie a 5 × 5 matrix
m_ijBy calculating only (0 ≦ i ≦ 4, i ≦ j ≦ 4), P
_a ^tH_f ^tH_fP_aIs H_f ^tH_fIt can be easily converted from
Call

FIG. 2 is a diagram schematically showing a case where the scale of the matrix is slightly increased and the frame length is larger than twice the pitch length. In FIG. 2, it is assumed that the frame length is 40 samples, the pitch period is 18 samples, and the cutoff of the impulse response is 5 taps. In FIG. 2, the triangular matrix portion at the upper right is omitted because it is a symmetric matrix similarly to FIG.
Also, when writing a matrix, the index is usually written such that it grows from left to right, and grows from top to bottom. The index is written so that the index becomes larger from the top to the left. The element at the i-th row and the j-th column of this matrix is represented by φ _ij for convenience.

The above features 1 to 4 also apply to the example of FIG. In this example, since the frame length is more than twice and less than three times the pitch period, three bands of the matrix are formed, and the value between the bands expressed by vertical stripes is 0. When the band is shifted to the left or right by an integral multiple of the pitch period, the values completely match. For example, a φ _4,20 = φ _4,38. A region represented by an oblique lattice pattern, for example, φ _ij (18 ≦
i ≦ 22, i ≦ j ≦ 22) and φ _ij (36 ≦ i ≦ 39, i
≦ j ≦ 39) is a submatrix φ _ij (0 ≦ i ≦
4, i ≦ j ≦ 4), it can be easily calculated from the law of the feature 4 and its extension. For example, φ _18,18 = φ _0,0 + φ _4,4 φ _20,21 = φ _2,3 + φ _3,4 φ _36,38 = φ _0,2 + φ _20,22 = φ _0,2 + 2 × φ _{2 , 4} . It should be noted here that the area after the nth cycle of the pitch, in this example, φ _ij (36 ≦ i ≦ 39, i ≦ j ≦ 3
9) corresponds to the third cycle. In this case, the following term is n-1 times (3-1 = 2 times in this example). Similarly, when the pitch is longer than the fourth cycle, such as when the frame is long and the pitch cycle is short, it is similarly multiplied by n-1.

If these features are used, it becomes possible to obtain the value of equation (4) for calculating the distance with a very small amount of calculation. Conventionally, for example, when the frame length is 80 samples, the value of 80 × What required 80 matrix calculations,
For example, it can be realized by 5 × 5 matrix calculation and simple matrix conversion processing. In FIGS. 1 and 2, all elements of the matrix are displayed for easy understanding, but actually, for example, only a 5 × 5 matrix is stored in the memory,
If the denominator of equation (4) is expanded and expressed using only elements of a 5 × 5 matrix, the memory area for storing the matrix can be significantly reduced.

However, when the denominator is expanded, the conventional method uses 1
Since the number of terms is increased by the expansion of the expression in the processing that only needs to refer to the table (matrix) twice, the processing amount for calculating the correlation matrix can be greatly reduced, but the number of operations for calculating the denominator slightly increases. If it is not necessary to extremely reduce the memory area, such as 5 × 5, for example, a memory area of 5 × (frame length) is secured and the value of the matrix is substituted. , A very small amount of calculation can be achieved. For example, in FIG.
Only the value of _ij (0 ≦ i ≦ 39, 36 ≦ j ≦ 40) is stored. At this time, the portion longer than the frame (j = 40) is calculated assuming that the frame is extended.
Other elements of the matrix (eg, 0 ≦ i ≦ 39, 0 ≦ j ≦ 3
5) can be substituted by referring to the stored memory area.

FIG. 3 shows a configuration example of a high-speed distortion calculation method according to the present invention. First reference speech x _r is converted to quantized inverse filter of the synthesis filter by (decoded) synthesis filter coefficients as a (synthetic inverted filter) 8-3, ideal (not quantized) excitation vector e . Finite tap length FIR type auditory weighted synthesis filter coefficient calculator 8-
In step 1, the impulse response of a perceptually weighted synthesis filter obtained by combining the synthesis filter and the perceptual weight filter is obtained, and the impulse response is terminated at a finite tap length equal to or less than half the pitch period, for example, 5 taps. The FIR type synthesis filter using filter coefficients obtained here, through the ideal excitation vector e, obtain the target speech x _f where truncation distortion is superimposed.

The truncated impulse response is periodicized by the pitch periodicizing filter 8-2 and sent to the convolution unit 8-4. The pitch period at this time may be performed at a non-integer sample period using a sampling function or may be approximated to an integer sample point, but the quality of the reproduced sound is better when the non-integer sample period is maintained. . The convolution unit 8-4 and the distance scale numerator calculation unit 8-9 calculate the numerator of the equation (4).

On the other hand, the truncated impulse response is
Is sent to the matrix calculation unit 8-5, and the correlation matrix H_f ^tH_fIs calculated
It is. The correlation matrix is, for example, an impulse response with 5 taps.
If you censor the answer, just calculate the 5x5 matrix
Is fine. In the pitch period correlation matrix conversion unit 8-7, the period
The pitch period specified by the sign is an integer sample period
, And P_a ^tH_f ^tH_fP_aTotal
Calculate. At this time, P _a ^tH_f ^tH_fP_aIs originally (frame
Length) × (frame length), for example, an 80 × 80 matrix
However, because of the aforementioned characteristics, P_a ^tH_f ^tH_fP_aPart of
For example, only the 80 × 5 sub-matrix is stored in the memory and stored.
(Frame length) x (file
Is equivalent to referencing the matrix of
As a result, the amount of processing and the amount of memory can be significantly reduced.

As described above, the denominator of equation (4) is
Open and calculate, H_f ^tH_fJust by referring to_a ^tH_f
^tH_fP_aTo obtain the same result as the reference
The period-to-phase matrix conversion unit 8-7 may be omitted. Again
In the raw signal power (denominator) calculator 8-8, the matrix P_a ^tH_f ^t
H_fP_aOr H_f ^tH_fAnd the denominator v of equation (4)_r ^tP_a ^t
H_f ^tH_fP_av_{r '}In other words, the pitch period
The power of the reproduced signal approximated by the pull value is calculated. In addition,
When a time series pulse with an amplitude of 1 is used as a fixed codebook
Contains the matrix P_a ^tH_f ^tH_fP_aOr H_f ^tH_fSee elements of
And add the value of the read element to the denominator by a simple process.
Can be calculated.

Even when the real value whose amplitude is not 1 is taken, when the pulse-type driving method has a value of 0 at most sample positions of the fixed code vector v _r and a non-zero value at a small number of sample positions, the present invention Very fast distance calculations can be performed using the method. Distance calculator 8-10
Then, the value of equation (4) is calculated from the calculated numerator and denominator values to obtain the distortion. Since the distortion at this time approximates the pitch period matrix of the denominator to the integer sample period,
Although it is an approximate value of the original distortion, if only the denominator is approximated by an integer sample period (the numerator is not approximated), the quality of the reproduced sound due to the approximation hardly deteriorates.

The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and a design change or the like may be made without departing from the gist of the present invention. Even if there is, it is included in the present invention.

By applying the present invention, 4 kbit / s speech coding was designed and realized in the form of a computer program. The sampling frequency is 8 kHz, the frame length of the linear prediction is 20 milliseconds (160 samples), the frame length of the driving sound source vector is 10 milliseconds (80 samples), the lower limit of the allowable range when searching for the pitch period is 18 samples, and the impulse The response truncation order was set to 6. As the configuration of the fixed codebook, a pulse train having an amplitude of 1 was used. When the pitch period is expressed by a non-integer number of samples, the resolution is 1 /
Accuracy of 3 samples. In this case, the present embodiment is implemented in the form of a computer program to confirm the effect, but may be implemented in a signal processing processor or in the form of dedicated hardware.

As a result of examining the processing time required for encoding, it was found that real-time processing could be sufficiently performed by a popular personal computer. In addition, as a result of examining the subjective quality when the reproduced sound is actually heard, ITU-TG. 726 (32 kbit / s AD
PCM). Therefore, G. 72
Compared to the 6 system, the same quality can be realized at a bit rate of 1/8. Also, the PSI-C bit rate used in mobile phones is 3.45 kbit / s.
Compared to the ELP method, higher quality can be realized with a processing amount of 1/4 or less. As described above, it has been confirmed that when the present invention is used, high-quality speech coding can be realized with a very small amount of processing and a small amount of memory, and the quality deterioration due to the approximate calculation is very small.

The present invention realizes an approximate calculation method of distortion which has very little influence on the quality deterioration of reproduced sound. However, when it is desired to calculate a non-approximate distortion, the method according to the present invention is used. And the order in which the approximate value of distortion is small ((4)
It is also possible to narrow down to several candidates (in the order of larger values of the equation), calculate non-approximate distortion for each of those candidates, and finally determine one code vector. The first step in selecting an optimum one in such two-step selection is called preliminary selection. The present invention is also effective as a method for preliminary selection.

[0054]

According to the present invention, high-quality reproduced sound can be obtained with a low bit rate, a small amount of memory, and a small amount of computation.

[Brief description of the drawings]

1 is a schematic view for explaining the characteristics of the pitch period of the correlation matrix used in the invention ^{_{P a t H f t H f}} P a.

Figure 2 is a schematic diagram for explaining the characteristics of the pitch period of the correlation matrix ^{_{P a t H f t H f}} P a to be used in the present invention, is a diagram for explaining in more generalized to FIG.

FIG. 3 is a block diagram illustrating a configuration example of a device that performs high-speed distortion calculation using the present invention.

FIG. 4 is a block diagram illustrating a configuration example of a conventional CELP-type speech encoding device.

FIG. 5 is a block diagram illustrating a configuration example of a fixed codebook search device in a CELP-type speech coding device.

FIG. 6 is a block diagram illustrating a configuration example of an apparatus that calculates a composite distortion.

FIG. 7 is a block diagram illustrating a configuration example of a device that performs distortion calculation using an impulse response.

FIG. 8 is a block diagram illustrating an example of a configuration of a high-speed distortion calculation device that uses the method already applied for by the present inventors to terminate an impulse response and use a pitch inverse filter.

[Explanation of symbols]

8-1 ... Finite tap length FIR type auditory weighted synthesis filter coefficient calculation unit 8-2 ... Pitch periodicization filter 8-3 ... Synthesis inverse filter 8-4 ... Convolution unit 8-5 ... Correlation matrix calculation Unit 8-6: Fixed codebook 8-7: Pitch periodic correlation matrix conversion unit 8-8: Reproduction signal power (denominator) calculation unit 8-9: Distance scale numerator calculation unit 8-10: Distance Scale calculator 8-11 FIR type synthesis filter

────────────────────────────────────────────────── ─── Continued on the front page (58) Fields surveyed (Int. Cl. ⁶ , DB name) G10L 3/00-9/20 H03M 7/30 JICST file (JOIS)

Claims (15)

(57) [Claims] An audio signal is reproduced by driving a synthesis filter by using a drive sound source vector created by using a time-series vector extracted from a codebook using a vector that is periodicized at a cycle corresponding to a basic cycle of speech, In a coding method for determining a driving excitation vector such that a distortion between a target audio signal and the reproduced audio signal is minimum or similar to the minimum, a periodical gain of periodicization is included in a process of calculating the distortion. Is set to 1 and the period for the periodization is approximated by an integer sample value, and the impulse response of the synthesis filter or the synthesis filter in consideration of the auditory weight is obtained. Calculate the correlation matrix of the impulse response matrix using the truncated impulse response. And an approximate value of the power of the reproduced audio signal is calculated using the correlation matrix of the impulse response matrix calculated from the truncated impulse response. . 2. The audio signal encoding method according to claim 1, wherein a signal obtained by subjecting a target audio signal to an inverse filter of a synthesis filter is subjected to an FIR filter having a coefficient of the truncated impulse response. Generating a target sound signal on which the truncation distortion of the impulse response is superimposed, and determining the driving sound source vector by regarding the target sound signal on which the truncation distortion is superimposed as a new target sound signal. Encoding method. 3. The audio signal encoding method according to claim 2, wherein the truncation order of the impulse response is set as short as about 3 to 10 taps. 4. The audio signal encoding method according to claim 1, wherein the pitch period correlation matrix is calculated only from the correlation matrix of the impulse response matrix calculated from the truncated impulse response. Using the features that can be used, the expression of the reproduced signal power is expressed only by the elements of the correlation matrix of the impulse response matrix calculated from the truncated impulse response, and the correlation of the impulse response matrix calculated from the truncated impulse response is calculated. A method for encoding an audio signal, comprising calculating the power of a reproduction signal while referring to elements of a matrix. 5. The audio signal encoding method according to claim 1, wherein the elements of the pitch-periodic correlation matrix have the same value at positions shifted by the period of the period. Utilizing the feature that the values match at the position where the row or column is shifted by the period, the matrix of (frame or subframe length) x (the truncation order of the impulse response) is stored in the memory, and the pitch period correlation matrix A method for encoding an acoustic signal, wherein the reference is performed by referring to the memory. 6. A sound filter is reproduced by driving a synthesis filter by using a driving sound source vector created by using a time-series vector extracted from a codebook using a vector that is periodicized at a cycle corresponding to a basic cycle of speech, In a coding apparatus for determining a drive excitation vector such that a distortion between a target audio signal and the reproduced audio signal is minimum or similar to the minimum, the distortion calculation means for calculating the distortion includes: Setting means for setting the quantization gain to 1, period approximation means for approximating the period for the periodization with an integer sampled value, response means for obtaining an impulse response of a synthesis filter or a synthesis filter in consideration of auditory weights, Truncating means for truncating the response with a length equal to or less than half the period of the periodicization, and using the truncated impulse response to form an impulse response matrix A correlation matrix calculating means for calculating the correlation matrix, and the periodicity is replaced by an approximation process using the integer sample values, and the correlation matrix is reproduced using the correlation matrix of the impulse response matrix calculated from the truncated impulse response. A sound signal approximating unit for calculating an approximate value of the power of the sound signal. 7. An audio signal encoding apparatus according to claim 6, wherein a signal obtained by subjecting a target audio signal to an inverse filter of a synthesis filter is subjected to an FIR filter using the truncated impulse response as a coefficient. A driving sound source vector determining unit that determines a driving sound source vector by generating a target sound signal on which the truncation distortion of the impulse response is superimposed, and regards the target sound signal on which the truncation distortion is superimposed as a new target sound signal. A coding device for an audio signal. 8. The audio signal encoding apparatus according to claim 7, wherein the truncation order of the impulse response is set as short as about 3 to 10 taps. 9. The audio signal encoding device according to claim 6, wherein said audio signal approximation means includes an impulse response in which a pitch periodic correlation matrix is calculated from the truncated impulse response. Using the features that can be calculated only from the correlation matrix of the matrix, the expression of the reproduced signal power is expressed only by the elements of the correlation matrix of the impulse response matrix calculated from the truncated impulse response, and calculated from the truncated impulse response An audio signal encoding apparatus, wherein the power of a reproduced signal is calculated with reference to the elements of the correlation matrix of the impulse response matrix. 10. The audio signal encoding apparatus according to claim 6, wherein the audio signal approximation unit includes a position in which an element of a pitch periodic correlation matrix is shifted by a period of the period. By taking advantage of the feature that takes the same value in, that is, the value matches at the position where the row or column is shifted by the period, a matrix of (frame or sub-frame length) x (discontinuation order of impulse response) is stored in the memory. An audio signal encoding apparatus, wherein reference to a pitch period correlation matrix is performed by referring to the memory. 11. A computer generates a time series vector extracted from a codebook using a driving sound source vector created by using a vector that has been periodicized at a period corresponding to a basic period of speech, and drives a synthesis filter to generate an acoustic signal. A medium in which a program for reproducing and functioning as a coding apparatus for determining a target excitation signal and a driving excitation vector such that a distortion between the reproduced audio signal is minimum or similar to the minimum, In the process of calculating the distortion, the computer sets the periodic gain of the periodicization to 1, a period approximating unit that approximates the period for the periodicization by an integer sample value, and considers a synthesis filter or an auditory weight. Response means for obtaining the impulse response of the synthesized filter, and striking the impulse response with a length equal to or less than half the period of the periodicization. A truncating means, a correlation matrix calculating means for calculating a correlation matrix of an impulse response matrix using the truncated impulse response, and the truncated impulse response on the assumption that the periodicization is replaced by an approximation process using the integer sampled values. A medium on which is recorded a program for functioning as acoustic signal approximating means for calculating an approximate value of the power of a reproduced acoustic signal using the correlation matrix of the impulse response matrix calculated as above. 12. The recording medium according to claim 11, wherein the computer applies a FIR filter having a coefficient of the truncated impulse response to a signal obtained by applying an inverse filter of a synthesis filter to a target acoustic signal. A target sound signal on which the truncation distortion of the impulse response is superimposed, and the target sound signal on which the truncation distortion is superimposed is regarded as a new target sound signal to function as a driving sound source vector determining means for determining a driving sound source vector. Recording a program for the computer. 13. The recording medium according to claim 12, wherein the cutoff order of the impulse response is set to be as short as about 3 to 10 taps. 14. The recording medium according to claim 11, wherein said acoustic signal approximation means comprises: a correlation matrix of an impulse response matrix in which a pitch periodic correlation matrix is calculated from a truncated impulse response. Using the features that can be calculated from only the impulse response calculated from the truncated impulse response, the expression of the reproduction signal power is expressed using only the elements of the correlation matrix of the impulse response matrix calculated from the truncated impulse response. A recording medium for calculating the power of a reproduction signal while referring to elements of a correlation matrix of a matrix. 15. The recording medium according to claim 11, wherein the acoustic signal approximation unit sets the elements of the pitch periodic correlation matrix to have the same value at a position shifted by the period of the period. Taking advantage of the feature that values match at positions where rows or columns are shifted by a period, a matrix of (frame or subframe length) × (discontinuation order of impulse response) size is stored in memory, and pitch periodization is performed. A recording medium characterized in that reference of a correlation matrix is performed by referring to the memory.