JP3182533B2 - High-speed operation High-performance filter bank configuration method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method of constructing a filter bank used for high-efficiency band division encoding for compressing and encoding music and audio signals.

[0002]

2. Description of the Related Art In the compression of HiHi audio digital signals conventionally used, processing of sub-band filters of bands, normalization of representative values, bit allocation, quantum coding, conversion to frame data, synchronization errors through communication paths, and synchronization errors are performed. Multi-band sub-coding is performed in the order of processing, decoding, representative value denormalization, and multi-band inverse sub-band filter processing. The actual filter bank processing is performed digitally, and the filter bank is formed by arithmetic processing using a DSP.
("Radio Technology" December 1991, pp. 156 to 161; "Radio Technology" February 1993, 54
Page to page 60)

[0003] As a method of configuring the filter bank, there are a method using an FIR filter, a method using an MDCT, and a method using a DFT bank. In the process according to FIR filters, B the number of bands, configuring maximum decimation filter bank has M times the window length band number is B, the number of multiplications and the number of additions both turned MB ² required, in proportion to the square of the number of bands As a result, the calculation amount increases.

In the MDCT method, since the filter bank can perform high-speed operation using DCT / DST, the number of multiplications is (log ₂ B) B / 2 + M · B, and the number of additions is (log ₂ B) 3B / 2. −B + 1 + M · B, and even if the number of bands is increased, the amount of calculation only increases almost in proportion to the number of bands. However, since M is limited to 2, the attenuation characteristic of the band is poor and the time axis Since cancellation of aliasing noise is performed above, there is no flexibility in filter design in the frequency domain.

Further, in the method using the DFT bank, D
One multiplication and addition of FT (FFT) and one multiplication and addition of the window are sufficient, and M is also arbitrary, so that the filter characteristics can be freely designed and the attenuation characteristics are sufficient, but in order to reconfigure the filter, Does not only require aliasing noise cancellation with adjacent bands, but requires a complex reconstruction filter.

[0006]

A condition required for a filter bank used for band division encoding for compressing and encoding music and audio signals is that the attenuation characteristic of the filter is sufficiently large (120 dB).
(The above is desirable) Cancellation of aliasing noise can be performed only between neighboring bands. Even if the number of divided bands increases, the amount of calculation does not increase so much as cannot be realized by the DSP.

The above method using the FIR filter can satisfy the above, but does not satisfy the above because the calculation amount increases in proportion to the square of the number of divided bands. Further, the method using the MDCT satisfies the above,
In the above, the width of the window is not satisfied because it is limited to twice the width of the split band. Although the above method can be satisfied with the method using the DFT bank, the above method cannot be satisfied because complicated folding cancellation by all bands is required.

It is an object of the present invention to provide a high-speed operation high-performance filter bank in band division coding that satisfies all of the above conditions.

[0009]

According to the present invention, in order to solve the above-mentioned problems, a first step of extracting N sample data for each block from an input signal, and multiplying a window by the data obtained in the first step. The second step and the second
A third step of folding and adding the data obtained in the step, and an Nth order I to the result obtained in the third step.
Performing a DCT to obtain an N-band output by performing the IDCT so that one of the odd band and the even band becomes a linear phase transform and the other becomes a Herbert transform, and a window width is obtained from the input signal. A first step of extracting a block signal having a length obtained by multiplying the magnification W with respect to the number of bands by the number N of samples by shifting N samples at a time;
A second step of multiplying the extracted block signal by a window and then folding and adding the same, one of the signals obtained in the second step to a linear phase conversion corresponding to an even band or an odd band, and the other to a Herbert conversion A third step of performing N-order IDCT processing, a fourth step of undersampling the signal obtained in the third step and transmitting it as an N-band output, and after oversampling the transmitted N-band output, Fifth, which is divided into even-numbered bands and odd-numbered bands and performs N-order DCT processing on each band
Step and the even-numbered band Nth order D obtained in the fifth step
A sixth step of performing backward addition and backward subtraction of the CT output and the Nth-order output of the odd-numbered band, and storing each of the operation outputs in a shift register having a length of W × N while shifting each block by one block; A seventh step of successively reading data stored in each shift register, multiplying the data by a window, and folding and adding the data, and outputting the N samples of the input signal and demodulating the input signal to form a high-speed operation high-performance filter bank configuration method. It is.

The attenuation characteristic of the filter is made sufficiently large,
So that the band overlaps only with the adjacent band,
In order to cancel aliasing noise, either the odd band or the even band is subjected to the Herbert transform (phase 90).
° conversion) filter and the other a linear phase filter. Further, after multiplying the windows, the amount of calculation is reduced by using DCT for each of the odd-numbered band and the even-numbered band.

[0011]

[Operation] By using a high-speed operation high-performance filter bank for band division coding as described above, a filter bank capable of canceling aliasing noise with a very large number of bands by DSP has a large number of bands. It can be easily realized in real time.

[0012]

Embodiments of the present invention will be described below with reference to the drawings. First, the fact that aliasing noise can be canceled in the present invention will be described. As a simple example, consider a two-part filterbank as shown in FIG. The input signal is input filter 1
After being divided into two by 及 び and 、, each is undersampled by 後 に and then oversampled by a factor of two, each is decoded by output filters 1 and 2, and finally both are added to form an output signal. Here, the sampling frequency is Es, the signals of the four bands are A to D, and the function of the filter curve is -1 / 2.Fs.ltoreq.f.ltoreq.1 / 2Fs.
In band 1, B ₁ (f) in band 1 and B ₁ in band 2
Assuming ₂ (f), as shown in FIG. 1, the input signal is divided into band 12 and band 2. Then 1/2
By undersampling, the characteristics are as shown in FIG. 1, and aliasing noise is generated. Then, if an output filter is applied after oversampling in order to combine the signals of the respective bands again, the filter characteristics shown in FIG. 1 are obtained. The signals of band 1 and band 2 are added to obtain the original signal as shown in FIG.

In FIG. 1, when two bands are added, aliasing noise can be canceled, and a condition that the original signal remains is required. When the input filter and the output filter of FIG. 1 have the same characteristics, the signal of each band of FIG. 1 is obtained as follows. -1/2 · Fs ≦ f ≦ -1 / 4 · Fs ( original signal A) band ^{_{1 B 1 2 (f) ·}} A + B 1 (f + 1/2 · F
s) · B ₁ (f) · C band 2 B ₂ ² (f) · A + B ₂ (f + 1/2 · F
_{s) · B 2 (f)} · C -1 / 4 ≦ f ≦ 0 ( the original signal B) band ^{_{1 B 1 2 (f) ·}} B + B 1 (f + 1/2 · F
s) · B ₁ (f) · D band 2 B ₂ ² (f) · B + B ₂ (f + 1/2 · F
_{s) · B 2 (f)} · D 0 ≦ f ≦ 1/4 · Fs ( original signal C) band ^{_{1 B 1 2 (f) ·}} C + B 1 (f + 1/2 · F
s) · B ₁ (f) · A band 2 B ₂ ² (f) · C + B ₂ (f + 1/2 · F
_{s) · B 2 (f)} · A 1/4 ≦ f ≦ 1/2 · Fs ( the original signal D) band ^{_{1 B 1 2 (f) ·}} D + B 1 (f + 1/2 · F
s) · B ₁ (f) · C band 2 B ₂ ² (f) · D + B ₂ (f + 1/2 · F
s) · B ₂ (f) · C where B ₁ (f + ／ · Fs) = B ₁ (f− ／ · Fs)
Fs), B ₂ (f + ／ · Fs) = B ₂ (f− ／ · Fs)
Fs).

As described above, of the signals in each band of each band, the first item represents the component of the original signal, and the second item represents the component of the aliasing noise. Therefore, when the condition under which the original signal returns when band 1 and band 2 are added is obtained, B ₁ ² (f) + B ₂ ² (f) = 1. Further, when a condition for canceling aliasing noise is obtained when band 1 and band 2 are added, since the second item only needs to be 0 in each band, B ₁ (f + ／ · Fs) · B ₁ (F) = − B ₂ (f + 1
/ 2 · Fs) · B ₂ (f). To satisfy this condition, it is necessary that B ₁ (f) = jB ₂ (f + Ｆ · Fs) or B ₂ (f) = jB ₁ (f + ／ · Fs). This means that assuming that the filter curves of band 1 and band 2 are the same and shifted, if one band is subjected to linear phase conversion and the other band is subjected to Herbert conversion, when both bands are added, aliasing noise is canceled. Can be realized.

In the present invention, the original signal is demodulated when the above two conditions, that is, the condition for returning the original signal and the condition for canceling the aliasing noise are satisfied by the filter function. If the number of filter bank divisions is large, dividing the odd band and the even band, when calculating the filter of each band, if the even band is a linear phase transformation, the odd band is a Herbert transformation, or the reverse combination The aliasing noise can be canceled only by the neighboring band. In the present invention, DCT is used to perform these transforms at high speed.

A procedure for dividing an input signal into sub-bands and demodulating them as an output signal will be described for a case where even-numbered bands are subjected to linear phase conversion and odd-numbered bands are subjected to Herbert transformation.

In order to construct a sub-band filter, first, a basic low-pass filter is designed, and a filter having an arbitrary characteristic can be obtained by shifting it on the frequency axis. When shifting, if it is desired to use a linear phase filter, the filter is moved by the same amount in the horizontal direction on the frequency axis. This is obtained by multiplying the impulse response (window) of the basic low-pass filter by a cosine function on the time axis. At this time, the frequency of the cosine function differs depending on the shift amount.

From this, in order to calculate one sub-band, the window of the basic low-pass filter is multiplied by a shift cosine function, and then the input signal is subjected to convolution and multiplication, and undersampled to obtain an output signal. Here, only the frequency of the cosine function differs in the calculation of each subband. Also, in order to perform undersampling, it is not necessary to perform convolution multiplication on each sample of the input signal, and it is sufficient to perform the convolution multiplication once every N times if the number of divided bands is N. FIG. 2 illustrates these procedures collectively. When an output signal of one sub-band is used,
The same calculation is repeated by shifting the input signal N times, and all the signals are added. In each band obtained in this way, since the multiplication of the window of the basic low-pass filter and the input signal is common, even if all the sub-bands are calculated, this calculation can be completed in FIG. As shown in FIG. 3, by replacing the procedure shown in FIG. 2, multiplying the input signal by the window of the basic low-pass filter first, and then multiplying by the shift cosine function, the calculation amount is reduced.

Further, the amount of calculation is reduced by utilizing the periodicity of the shift cosine function. Assuming that the number of divided bands is N, the band number is K, and the sample number on the time axis is i, the shift cosine function is cos [(2K + 1) · πi / 2
N)]. ± (2K +
1) Assume that the shift is performed by π / 2N. Section 0-N-
Assuming that the shift cosine function in 1 is a basic if cosine function, the other sections are as follows. Section N to 2N-1 (1 ≦ i ≦ N−1) cos [(2K + 1) π (2N−i) / 2N] = cos [− (2K + 1) πi / 2N + (2K + 1)
π] = − cos [(2K + 1) πi / 2N] Section −2N to −N−1 (0 ≦ i ≦ N−1) cos [(2K + 1) π (i−2N) / 2N] = − cos [(2K + 1 ) Πi / 2N] section −N to −1 (1 ≦ i ≦ N−1) cos [(2K + 1) πi / 2N] When these are put together and expressed by a basic fosine cosine function in each section, Table 1 is obtained. .

[0020]

[Table 1]

As can be seen in Table 1, this is repeated with 4N synchronization, indicating that the shift cosine function can be represented by a basic shift cosine function of length N. as a result,
The subband calculation procedure is as shown in FIG. 4, and the subband can be calculated using the basic shift cosine function instead of the shift cosine function in FIG. The only difference between the subbands is the basic shift function. After the input signal is multiplied by the window of the basic low-pass filter, folding addition is performed. At this time, the point N and the point -N are not added. Here, IDCT is a method of calculating the output signals of the sub-bands for all the sub-bands at once from the signals obtained by the folding and addition. Since the IDCT has a high-speed processing method, the number of calculations is small. As described above, the output signal of the even-numbered band can be obtained.

Furthermore, since the aliasing noise cannot be finally canceled unless the Herbert transform is applied to the odd-numbered band, the calculation procedure for the odd-numbered band is the same as that for the even-numbered band. There is. When the filter is calculated by the Herbert transform, the basic low-pass filter j
In the negative frequency direction by multiplying by −j, respectively, which means multiplying the window of the basic low-pass filter by a sine function on the time axis. The procedure for calculating the output signal of one sub-band of the odd band is shown in FIG.
The output signal of one subband is calculated using the shift sine function instead of the shift cosine function in the same procedure as shown in FIG. The procedure is shown in FIG. Further, as shown in FIG. 3, the calculation procedure is switched, and the calculation amount is reduced due to the nature of the shift sine function.

The number of divided bands is N, the band number is K, and the sample number on the time axis is i. The Herbert transform is performed only when K is an odd number, and the basic low-pass filter is ± (2K +
1) Calculate by shifting by π / 2N. At this time, the shift sine function is -sin [(2K + 1) πi / 2N].
It is represented by Each section is represented by this shift sine function as follows. Section N to 2N-1 (0 ≦ i ≦ N−1) −sin [(2K + 1) π (i + N) / 2N] = cos [(K + 1) π + (2K + 1) πi / 2N] = cos [2K + 1) πi / 2N Section 0 to N−1 (1 ≦ i ≦ N−1) −sin [(2K + 1) π (N−i) / 2N] = cos [− (K + 1) π + (2K + 1) πi / 2N] = cos [2K + 1] ) Πi / 2N] section −N to −1 (0 ≦ i ≦ N) −sin [(2K + 1) π (i−N) / N] = − cos [− (K + 1) π + (2K + 1) πi / 2
N] = − cos [2K + 1) πi / 2N] Section −2N ≦ to N−1 (1 ≦ i ≦ N−1) −sin [(2K + 1) π (−Ni) / 2N] = − cos [( K + 1) π + (2K + 1) πi / 2N] = − cos [2K + 1) πi / 2N]

The above is summarized in Table 2 and is repeated in 4N cycles.

[0025]

[Table 2]

From Table 2, it can be seen that the shift sine function is of length N
It can be seen that this is represented by the basic shift cosine function of As a result, the calculation can be performed in the same manner as the calculation procedure of the even-numbered band shown in FIG. 4. As shown in FIG. 6, the sub-band filter subjected to the Herbert transform can be calculated using the IDCT based on the basic shift cosine function. However, the method of folding and adding is different from that of FIG.

As described above, linear phase conversion is adopted for even-numbered bands, and Herbert transform is adopted for odd-numbered bands, and filters are calculated using IDCT, so that aliasing noise is canceled when demodulating to the original signal. Now, the procedure for calculating a reconstructed filter bank for demodulating the original signal will be described.

In order to demodulate the original signal from one subband signal, the window structure of the basic low-pass filter must be appropriate. The window conditions will be described. When the window is W (w) and the number of divided bands is N, at least the sections w ≦ −π / N, w ≦ −π /
At N, the gain is 0, and it is necessary that the gain is left-right symmetric with respect to w = 0, that is, W (w) = W (−w). Further, the following relationship is approximately established between the function W (w-π / N) and W (w) obtained by shifting the window W (w) by π / N. In the section 0 ≦ w ≦ π / N, {W (w)} ² + {W (w−π / N)} ² = 1 Therefore, to demodulate the original signal, it is necessary to perform subband division and reconstruction. And must use the same window. A window that satisfies such a condition can be obtained by adjusting parameters by the Remetz method.

By properly selecting the window structure in this way, by adding the demodulated signals of all the bands,
The aliasing noise can be canceled and the original signal can be demodulated. FIG. 7 shows a calculation procedure of the demodulation. The configuration of the filter is basically the same as the above-described filter calculation procedure. The window of the basic low-pass filter is multiplied by a shift cosine function, and one subband signal is convolved and multiplied. At this time, a demodulated signal by one sub-band signal is obtained by shifting and multiplying one sample at a time and adding all of them. In this convolution multiplication, since most of the multiplication with the complemented zero is performed, as shown in FIG.
A signal obtained by holding one subband signal is added every N using a signal obtained by holding, and a demodulated signal having a section length N is obtained.

Further, as seen from Table 1 above, due to the nature of the shift cosine function, the multiplication of the signal holding one subband signal and the shift cosine function is only the multiplication of the basic shift cosine function, and Can be used repeatedly. As shown in FIG. 9, a signal of a sub-band (even-numbered band) is calculated by using DCT for multiplication of the basic shift cosine function, and the calculation procedure is switched to reduce the amount of calculation. A signal having a section length N calculated by shifting every N pieces by DCT is input to the shift register. When multiplying by the window of the basic low-pass filter, the shift register determines the interval -2N to -N.
In N, the direction of the signal is read out in the section -N to 0 in the opposite direction. Then, the signal is added every N signals to obtain the original signal.

FIG. 9 shows that the demodulation of the even-numbered band of the linear phase conversion can be performed at high speed by using the DCT.
It can be calculated by the same method as the demodulation of the even-numbered band using DCT as in Table 2. If these results are added last, all the original signals can be demodulated.

Further, at the stage of input to the shift register, the signals are input to the two shift registers whose time directions are forward and reverse, and the signals of the even-numbered band and the odd-numbered band are added at this stage. Can be multiplied only once. Table 3 shows the DCT results of the even-numbered band and the odd-numbered band. The result can be expressed by using either one of the two types of shift register outputs A and B obtained by adding the results in the backward direction. In addition, the type A to which the backward direction is added is linear positive direction +
Herbert reverse direction, type B indicates linear reverse direction-Herbert forward direction.

[0033]

[Table 3]

Next, a filter bank divided into two or more N will be described with reference to a circuit block diagram. Even when the band is divided into two or more bands, the overlapping part of each band is the same as that of the two-band, and the aliasing noise is canceled under the same conditions as the two-band, and each band is alternated. In addition, the linear phase conversion and the Herbert conversion must be repeated. The calculation procedure of the filter bank is as described above.

FIGS. 10 to 12 show the filter characteristics of each band when divided into four bands. In FIG. 10, the characteristics of the filter bank are shown in the upper part. Bands of the same number with the same positive or negative frequency are the same band, and are not complex banks but real banks. Also, it is the maximum thinning filter bank, which is folded back at the dotted line in the figure. The lower part of the figure shows the characteristics of each band after division. The even band is subjected to linear phase conversion, and the even band is subjected to Herbert transformation. AP indicate the location of the band,
j represents a phase rotation of +90 degrees, and -j represents a phase rotation of -90 degrees. FIG. 11 shows the characteristics of each band that has been divided and undersampled, and it can be seen that aliasing noise has occurred. FIG. 12 shows the characteristic of each band passing through the reconstruction filter. The characteristic of the reconstruction filter is the same as that of the split filter, but the phase of the Herbert transform is inverted by 180 degrees. At this time, if the characteristics of the filter are appropriate, that is, if the above two conditions are satisfied, the aliasing noise can be canceled and the signal can be approximately demodulated to the original signal. -1 in FIG.
Indicates that the sign of the signal is inverted.

FIG. 13 is a circuit block diagram of a band division for realizing the calculation procedure shown in FIG. 4 in order to obtain the filter characteristics after division as shown in FIG. Assuming that the magnification of the width of the window with respect to the number of bands is W, a block i having a length of W × N samples is extracted from the input signal by a block extracting circuit, and W × N data are respectively extracted for the odd-numbered band and the even-numbered band. Window multiplication and folding addition circuit. Then, N-order IDCT processing is performed on the N pieces of data output from the respective circuits for the odd-numbered band and the even-numbered band. Next, an N-band output is obtained by adding the respective outputs.
14 and 15 show the processing procedure up to the window multiplication and the folding addition. As described above, since one of the window conditions may be symmetrical, the window multiplication is not performed. , The length of the left half is used to reduce the amount of calculation. Here, the case where W = 8 is shown. Then, a circuit block for performing IDCT processing on N pieces of data from the window multiplication and folding addition circuit is shown in FIG. 16, where N = 8. Here, when the IDCT input is Xk and the output is xi, Equation 1 is established.

[0037]

(Equation 1)

FIG. 17 is a circuit block diagram for reconstructing and demodulating the band output obtained in FIG. 16 into the original signal. The schematic configuration is made up of an N-order DCT, backward addition and subtraction circuits, shift registers 1 and 2, and window multiplication and addition circuits. The calculation processing procedure in this schematic configuration is shown in FIG.

FIG. 18 shows a DCT circuit portion, which divides the N-band output obtained in FIG. 13 into an odd band and an even band and performs DCT processing. DCT input is Xi,
Assuming that the output is Xk, Equation 2 holds.

[0040]

(Equation 2)

FIG. 19 shows how the outputs of the backward addition circuit and the backward subtraction circuit in FIG. 17 are input to the shift register 1 and the shift register 2, respectively. In the backward addition circuit, the even-numbered band Nth-order DCT output ODCTG
(P) The output O obtained by adding (0 ≦ P ≦ N−1) and the odd-numbered band Nth-order DCT output ODCTK (P) (1 ≦ P ≦ N−1).
UT1 (P) (0 ≦ P ≦ N−1) is calculated as follows: OUT1 (P) = ODCTG (P) (P = 0) OUT1 (P) = ODCTG (P) + ODCTK (N−
P) (1.ltoreq.P.ltoreq.N-1), and the data is shifted and input to the shift register 1 for each block (N data). In the backward subtraction circuit, the even band N
From the next DCT output ODCTG (P) (1 ≦ P ≦ N−1) to the odd band Nth DCT output ODCTK (P) (0 ≦ P ≦
N-1) is subtracted from the output OUT2 (P) (0 ≦ P ≦ N−
1) OUT2 (P) = ODCTK (P) (P = 0) OUT2 (P) = ODCTG (NP) -ODCTK
(P) The output is performed so that (1 ≦ P ≦ N−1), and the data is shifted and input to the shift register 2 for each block (N data).

FIG. 20 shows a procedure for multiplying the data read from the shift registers 1 and 2 by the window and adding the data. The read from the shift registers is stored in the shift registers 1 and 2. The data alternately and multiply by the window.

FIG. 21 shows the entire calculation processing procedure of the reconstruction filter shown in FIG. 17, and here, the window uses the length of the left half.

[0044]

As described above, according to the present invention, either the linear transfer conversion or the Herbert conversion is performed on the odd-numbered band or the even-numbered band, and the DCT is used for each band. It was possible to take a large value, cancel aliasing noise, and reduce the amount of calculation to the extent that it could be realized in real time by the DSP.

In fields such as voice and music signal compression codes, it is effective to use it for band division before coarse quantization is performed. This is because even if the signal of each band is subjected to coarse quantization and data compression, the distortion is generated only in a narrow band, so that there is little deterioration in sound quality in terms of hearing.

Further, if the structure of the window is appropriately selected,
By adding demodulated signals of all bands, aliasing noise can be canceled and the original signal can be demodulated.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram of a two-divided filter bank which is a basis of the present invention.

FIG. 2 is an explanatory diagram showing a procedure for calculating an output signal of one subband.

FIG. 3 is an explanatory diagram showing a modification procedure of the output signal calculation of FIG. 2;

FIG. 4 is an explanatory diagram showing a procedure for converting the output signal of FIG. 3 into a basic shift cosine function.

FIG. 5 is an explanatory diagram showing a Herbert transform procedure of an output signal of one subband.

FIG. 6 is an explanatory diagram showing a Herbert conversion procedure using a basic shift cosine function.

FIG. 7 is an explanatory diagram showing a demodulation procedure using one subband signal.

FIG. 8 is an explanatory diagram showing a procedure for simultaneously calculating demodulated signals having a section length N.

FIG. 9 is an explanatory diagram showing a demodulation procedure using DCT.

FIG. 10 shows filter bank characteristics in the case of four bands,
FIG. 9 is an explanatory diagram illustrating characteristics of each band after division.

FIG. 11 is an explanatory diagram showing characteristics of each band after undersampling.

FIG. 12 is an explanatory diagram showing characteristics of each band after passing through a reconstruction filter.

FIG. 13 is a schematic diagram illustrating division into odd and even bands.

FIG. 14 is an explanatory diagram showing a calculation procedure of a window multiplication and folding addition circuit for even-numbered bands.

FIG. 15 is an explanatory diagram showing a calculation procedure of an odd band window multiplication and folding addition circuit.

FIG. 16 is a schematic diagram of an IDCT circuit.

FIG. 17 is a schematic diagram of forming a reconstitution filter.

FIG. 18 is a schematic diagram of a DCT circuit. FIG. 19 is an explanatory diagram showing backward addition and backward subtraction and signal input procedures to shift registers 1 and 2;

FIG. 20 is an explanatory diagram showing a calculation procedure of a window multiplication and addition circuit.

FIG. 21 is an explanatory diagram showing a calculation procedure for forming the reconstruction filter of FIG. 17;

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) H03M 7/30 H03H 17/02

Claims (2)

(57) [Claims] A first step of extracting N sample data for each block from the input signal; a second step of multiplying the data obtained in the first step by a window;
A third step of folding and adding the data obtained in the step, and an Nth order I to the result obtained in the third step.
And a fourth step of performing DCT, wherein the IDCT is performed to obtain an N-band output so that one of the odd-numbered band and the even-numbered band is converted to the linear phase conversion, and the other is converted to the Herbert conversion to obtain an N-band output. Filter bank configuration method. 2. A first step of extracting a block signal having a length obtained by multiplying a scaling factor W with respect to the number of bands of a window width and the number of samples N by N samples from an input signal, and forming a window in the extracted block signal. A second step of folding and adding after multiplication, and N-order IDCT processing of the signal obtained in the second step so that one is subjected to linear phase conversion and the other is applied to Herbert transformation corresponding to an even band or an odd band. The third step and the third
Undersampling the signal obtained in the step
A fourth step of transmitting as a band output, a fifth step of oversampling the transmitted N-band output, dividing the output into an even band and an odd band, and performing N-order DCT processing, respectively, and the even number obtained in the fifth step. Band N
A sixth step of performing backward addition and backward subtraction of the next DCT output and the odd-numbered band Nth output, and storing each operation output in a shift register having a length of W × N while shifting each block by one block; A seventh step of sequentially reading data stored in each of the shift registers, multiplying the data by a window, and folding and adding the data, outputting N samples of the input signal and demodulating the data. Method.