JP2002118517A - Orthogonal transform apparatus and method, inverse orthogonal transform apparatus and method, transform coding apparatus and method, and decoding apparatus and method - Google Patents

Abstract

(57) [Summary] [PROBLEMS] To provide a transform coding apparatus and method capable of arbitrarily determining a break of an overlap and enabling complete signal reproduction by overlap addition. SOLUTION: A linear / nonlinear prediction analysis unit 3 performs a linear / nonlinear prediction analysis process on an audio signal input from an input terminal 2 and outputs a prediction residual. Stationarity estimator 7
Estimates the stationarity of the audio signal. The block length determining unit 8 determines a block length at the time of MDCT based on the result estimated by the stationarity estimating unit 7. MDCT unit 5
Performs MDCT processing on the time series sample M of the prediction residual input via the buffer 4 with the block length determined by the block length determining unit 8 to generate MDCT coefficients.
The quantization unit 6 quantizes the MDCT coefficients generated by the MDCT unit 5.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to an orthogonal transformation apparatus and method for performing orthogonal transformation while overlapping input time-series samples. The present invention also relates to an inverse orthogonal transform apparatus and method for performing orthogonal transform on orthogonal transform coefficients obtained by performing orthogonal transform while overlapping time series samples. The present invention also relates to a transform coding apparatus and method using the above-described orthogonal transform apparatus and method. The present invention also relates to a decoding device and a method to which the above-described inverse orthogonal transform device and method are applied.

[0002]

2. Description of the Related Art In a digital encoding method for time-series samples such as audio and image signals, orthogonal transform such as fast Fourier transform (FFT), discrete cosine transform (DCT), and improved discrete cosine transform (Modified DCT: MDCT) is performed. Various proposals have been made for use.

In particular, the MDCT performs orthogonal transform while overlapping time-series samples, thereby reducing noise generated at seams between blocks as compared with DCT, and orthogonally transforms and compresses and encodes an audio signal. In recent years, the transform coding method has become a very popular method.

[0004] MDCT and IM which is an inverse transform of MDCT
The definition of DCT is as in the following equations (1) and (2).

[0005]

(Equation 1)

[0006]

(Equation 2)

In the above equations (1) and (2), x is an input signal, y is an MDCT coefficient, xｘ is an inverse MDCT output, M is a block length, h is a window function for forward conversion, and f is an inverse conversion window function. It is a window function.

By substituting equation (2) into equation (1), the following equation (3) is obtained.

[0009]

(Equation 3)

From equation (3), after MDCT, IMDC
It can be seen that the time series signal xｘ (m) obtained by T contains an aliasing component. This aliasing component can be completely canceled by selecting an appropriate window function h (m), f (m) and converting it into a time-series signal with 50% overlap.

FIG. 6 is a diagram for explaining the MDCT and IMDCT algorithms. FIG. 6 shows an operation for an arbitrary adjacent block in the time-series sample x (m). In the figure, the block j-1 and the block j are both M in length, and each block has an overlap of 50%. There is wrap. By windowing this block with a window function h (m) and performing a linear forward transformation, the M / 2 point M
DCT coefficients are obtained. The above is the operation of the MDCT forward conversion. In IMDCT, M / 2 time-series samples xｘ (m) are obtained by performing a linear inverse transform of MDCT coefficients, performing windowing with a window function f (m), and performing overlap addition on adjacent blocks. .

[0012] In an audio coding method, particularly in a transform coding method, selection of a block length of orthogonal transform is one factor that characterizes sound quality. Generally, the longer the orthogonal transform block length, the higher the frequency resolution, and the shorter the orthogonal transform block length, the higher the time resolution. Therefore, a long block length is more advantageous in terms of efficiency and the like for an input with a small time variation of a signal, and a shorter block length is better for a signal with a large time variation. For example, if MDCT is performed with an excessively long block length for the input of attack music having a large time variation of the signal, a pre-echo or a post-echo occurs in the reproduced sound due to insufficient time resolution, resulting in deterioration of sound quality. It is. For this reason, it is conceivable to adopt a method of adaptively changing the block length based on the characteristics of an input signal into high-efficiency coding, and a method that actually adopts the method has been proposed.

However, when the block length is to be changed using the definition formulas (1) and (2),
Care must be taken to completely match x ~ (m) and x (m) because it is necessary to cancel aliasing occurring in the time domain. For example, [1]: “Takashi Mochiz
uki.Perfect Recontsruction Conditions for Adaptive
Blocksize MDCT.IEICE Trans.fundamentals, vol.E77-
A, No.5, pp.894-899, May1994. ”, By selecting a window that cancels the aliasing,
The block lengths are mixed with the definition expression of expression (2). Figure 7 using the method of [1] is an example of switching the block length from _{M 1} to _{M 2 (M 1 <M 2} ). 7, the frame j-2, a j-1 at block length M _1, block length has shifted from M ₁ to M ₂ in frame j.

[0014]

By the way, in the frame j in which the block length shifts in FIG. 7, the first half (M2-M1) / 4 coefficients of the window are 0, and the effective range of the window is 3 ( M ₂ −M ₁ ) / 4, which is shorter than the block length M _{2 of the} MDCT. Therefore, 3 (M ₂ −
Since MDCT is performed with an unnecessarily long block length for input samples of M ₁ ) / 4, it cannot be said that it is a good idea in terms of efficiency. In addition, when the pre-processing is performed in blocks in the time domain before the MDCT is performed, the phase changes, so that the processing of the pre-processed sample becomes troublesome.

[0015] Therefore, in the frame j the block length is changed from the M ₁ to M ₂ as shown in FIG. 8, if to match the number of samples that overlap the front and rear frames, the scope and MDCT window The block lengths match. Furthermore, if M ₂ is an integer multiple of M ₁ , there is no change in phase due to switching of the block length.
Pre-treated samples are easier to handle.

However, as long as the MDCT is performed using the above defined equations (1) and (2), the IMDCT is obtained using the symmetrical time windows unless the frames before and after are overlapped by 50%. Aliasing included in x ~ (m) cannot be canceled,
The original time-series samples cannot be restored with an overlap such as frame j in FIG.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide an orthogonal transformation apparatus and method capable of arbitrarily determining a break of an overlap. It is another object of the present invention to provide an inverse orthogonal transform apparatus and method for inversely transforming orthogonal transform coefficients obtained by such an orthogonal transform apparatus and method.

Further, the present invention has been made in view of the above-mentioned circumstances, and a transform coding apparatus and method for arbitrarily determining a break of an overlap and enabling complete signal reproduction by overlap addition, and It is an object of the present invention to provide a decoding device and method.

[0019]

SUMMARY OF THE INVENTION In order to solve the above-mentioned problems, an orthogonal transform apparatus according to the present invention performs orthogonal transform while overlapping input time-series samples. In the orthogonal transform, the orthogonal transform is performed by arbitrarily determining a sample position α as a boundary where aliasing occurs at the time of the inverse orthogonal transform in a range of 0 ≦ α <M.

In order to solve the above problem, an orthogonal transform method according to the present invention is directed to an orthogonal transform method for performing orthogonal transform while overlapping input time-series samples.
When orthogonal transformation is performed on the time-series M samples, the orthogonal transformation is performed by arbitrarily determining a sample position α as a boundary where aliasing occurs at the time of inverse orthogonal transformation in a range of 0 ≦ α <M.

In order to solve the above problem, an inverse orthogonal transform apparatus according to the present invention is directed to an inverse orthogonal transform apparatus for inverse orthogonal transforming orthogonal transform coefficients obtained by performing orthogonal transform while overlapping time series samples. It is characterized by arbitrarily determining a sample position α as a boundary where aliasing occurs at the time of orthogonal transformation within a range of 0 ≦ α <M, and inversely orthogonally transforming the orthogonally transformed orthogonal transformation coefficient.

In order to solve the above problem, the inverse orthogonal transform method according to the present invention is directed to an inverse orthogonal transform method for inverse orthogonal transforming orthogonal transform coefficients obtained by orthogonally transforming time-series samples while overlapping. It is characterized by arbitrarily determining a sample position α as a boundary where aliasing occurs at the time of orthogonal transformation within a range of 0 ≦ α <M, and inversely orthogonally transforming the orthogonally transformed orthogonal transformation coefficient.

In order to solve the above-mentioned problems, a transform coding apparatus according to the present invention is a transform coding apparatus for orthogonally transforming an input signal and compressing and encoding the input signal. Predictive analysis means for outputting a prediction residual, characteristic determining means for determining characteristics of the input signal for each predetermined sample, and determining a block length at the time of orthogonal transformation based on the characteristics determined by the characteristic determining means. The block length determining means and the block length determined by the block length determining means arbitrarily determine a sample position α as a boundary at which aliasing occurs at the time of inverse orthogonal transformation in a range of 0 ≦ α <M, and perform the prediction analysis. The orthogonal residual is applied to the time series M samples to generate orthogonal transform coefficients while overlapping the prediction residuals output from the means as input time series M samples. And a quantizing means for quantizing the orthogonal transform coefficients generated by the orthogonal transform means.

Therefore, it is possible to easily realize transform coding such as switching the orthogonal transform block length in accordance with the properties of the input signal and quantizing the orthogonal transform output coefficients.

In order to solve the above-mentioned problems, a transform coding method according to the present invention is a transform coding method for orthogonally transforming an input signal and compressing and encoding the input signal, wherein the input signal is fetched by a predetermined sample and prediction analysis is performed. A prediction analysis step of outputting a prediction residual by using the above method, a characteristic determination step of determining characteristics of the input signal for each predetermined sample, and determining a block length at the time of orthogonal transformation based on the characteristics determined in the characteristic determination step. The block length determining step and the block length determined in the block length determining step arbitrarily determine a sample position α as a boundary where aliasing occurs at the time of inverse orthogonal transformation in a range of 0 ≦ α <M, and perform the prediction analysis. While overlapping the prediction residual output from the process as an input time series M samples, the time series M samples are subjected to orthogonal transform processing to generate orthogonal transform coefficients. And a quantization step of quantizing the orthogonal transformation coefficients generated in the orthogonal transformation step.

In order to solve the above problem, the decoding apparatus according to the present invention sets a sample position α, which is a boundary at which aliasing occurs at the time of inverse orthogonal transform, to 0 ≦ α with a block length determined according to the characteristics of an input signal. A decoding device for decoding quantized data obtained by quantizing orthogonal transform coefficients obtained by orthogonally transforming the input time-series M samples while overlapping the input time-series M samples by arbitrarily determining in the range of <M. Inverse quantization means for inversely quantizing the data, and inverse orthogonal transform of the orthogonal transform coefficient obtained by inverse quantization by the inverse quantization means, with a block length determined according to the characteristics of the input signal. And an inverse orthogonal transform unit.

In order to solve the above-mentioned problems, the decoding method according to the present invention sets a sample position α, which is a boundary at which aliasing occurs at the time of inverse orthogonal transform, to 0 ≦ α with a block length determined according to the characteristics of an input signal. A decoding method for decoding quantized data obtained by quantizing orthogonal transform coefficients obtained by performing orthogonal transform while arbitrarily determining the M samples in the input time series while overlapping each other in the range of <M. An inverse quantization step of inversely quantizing the data, and an inverse orthogonal transform of the orthogonal transform coefficient obtained by the inverse quantization in the inverse quantization step, with a block length determined according to the characteristics of the input signal. And an inverse orthogonal transformation step.

[0028]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Some embodiments of the present invention will be described below with reference to the drawings. First, first
Is an encoder 1 shown in FIG. 1 and is a specific example of a transform encoding device that compresses and encodes an audio signal sampled at 16 KHz input from an input terminal 2 using MDCT described later. .

In FIG. 1, an encoder 1 performs a linear / nonlinear prediction analysis process on the audio signal input from an input terminal 2 to output a prediction residual, and a linear / nonlinear prediction analysis unit 3 for outputting the prediction residual; , A block length determining unit 8 that determines a block length at the time of MDCT based on the result estimated by the continuity estimating unit 7, and a block length determining unit 8 that determines the block length. The MDCT processing is performed on the time series M samples of the prediction residual inputted through the buffer 4 with the set block length and the MDC
An MDCT unit 5 that generates a T coefficient and a quantization unit 6 that quantizes the MDCT coefficient generated by the MDCT unit 5 are provided.

The linear / non-linear prediction analysis unit 3 takes in, for example, 1024 samples of the audio signal, performs linear or non-linear prediction, and outputs a prediction residual to the buffer 4. Further, the linear / nonlinear prediction analysis unit 3 outputs the obtained analysis parameters from the output terminal 9. For example, the audio signal is subjected to LPC analysis in the 16th order. The LPC coefficients are converted to LSP, and after quantizing the LSP, the LSP
Is interpolated between frames, and the LCP is calculated using the interpolated coefficients.
Find the residual. Further, an optimum pitch lag in the LPC residual is obtained, an optimum gain at this lag and ± 1 point is calculated, and vector quantization is performed. Then, a pitch inverse filter is constituted by the quantized pitch gain, and a pitch residual is obtained using the inverse filter.

The stationarity estimating section 7 estimates stationarity of the audio signal. With an excessive signal, if the block length of the MDCT is too long, a sufficient time resolution cannot be obtained, so that a pre-echo or a post-echo occurs in the reproduced sound. Therefore, it is desirable to shorten the MDCT block length for such a signal. On the other hand, for a quasi-stationary signal with a small time variation of the signal, by increasing the MDCT block length, the number of bits used for normalization and analysis parameters can be reduced. Bits can be allocated. For this reason, the encoder 1 shown in FIG. 1 switches the block length between two stages, for example, LONG and SHORT according to the properties of the input signal. At this time, it is the continuity estimator 7 that determines the characteristics of the input signal. Specifically, a change in the frame power of the input signal from the frame immediately before and a change from the frame immediately before the LSP are obtained, and a frame determined to have a large change compared to the previous frame based on a certain threshold value is flagged. I will make up. Then, when the flag is not set from the current frame to several frames before and after the current frame, the signal is determined to be a quasi-stationary signal with little time variation of the signal.

If the continuity estimating unit 7 determines that the continuity is high, the block length determining unit 8
The block length of T is lengthened (LONG), and is shortened in the case of an excessive signal (SHORT). The block length information determined by the block length determining unit 8 is output via the input terminal 11.

For example, FIG. 2 is a plot of a sample sequence of a certain audio signal. It can be seen that there is a sharp change in the signal level near the middle point in the figure. When such a signal is input to the encoder 1 of FIG. 1, it is desirable to select a short block length at a sample position where there is a sudden change.

The MDCT unit 5 arbitrarily determines a sample position α as a boundary where aliasing occurs at the time of IMDCT within the range of 0 ≦ α <M based on the block length determined by the block length determination unit 8, and performs linear / non-linear processing. While overlapping the prediction residuals output from the prediction analysis unit 3 as input time-series M samples, the time-series M samples are subjected to MDCT processing to generate MDCT coefficients.

The quantization section 6 quantizes the MDCT coefficients generated by the MDCT section 5 and outputs the index from the output terminal 10. A specific example of the quantization in the quantization unit 6 will be described below. When the prediction residual output from the linear / nonlinear prediction analysis unit 3 is the pitch residual, the MDCT
After performing normalization, the MDCT coefficients generated by the unit 5 are quantized using three types of quantization units of two-dimensional eight bits, four-dimensional eight bits, and eight-dimensional eight bits. Bit allocation is determined using weights calculated from only parameters used for analysis and normalization. This allows
Compared with the method of obtaining the optimal bit allocation for each coefficient and then performing quantization, since parameters such as position information are not required, it is possible to increase the ratio of the number of bits that can be allocated to the quantization of the coefficient. . Then, the quantization unit 6 converts the quantized data as the index into 6
Output at a rate of kbps to 32 kbps.

Next, a schematic operation of the encoder 1 having the above configuration will be described. An audio signal sampled at 16 KHz is input to the encoder 1 via the input terminal 2. The linear / nonlinear prediction analysis unit 3 captures 1024 samples of the audio signal, performs linear or non-linear prediction, supplies the prediction residual to the buffer 4, and supplies the prediction residual to the continuity estimation unit 7. Stationarity estimator 7
Estimates the stationarity of the audio signal based on the supplied prediction residual, and supplies the estimation result to the block length determination unit 8. The block length determining unit 8 determines the MDCT block length to be, for example, either 1024 or 2048 based on the characteristics of the input sample estimated by the stationarity estimating unit 7. The block length is selected to be 1024 in a section where it is determined that a higher time resolution is required, and 2048 is selected in a section in which a signal has little temporal change and can be regarded as quasi-stationary. Thereafter, based on the determined block length, the MDCT unit 5 takes out a sample from the buffer 4 and performs MDCT. M
The DCT coefficient is quantized by the quantization unit 6, and its index is output as quantized data from the output terminal 10 at a rate of 6 kbps to 32 kbps. The analysis parameters obtained by the linear / nonlinear prediction analysis unit 3 are output to an output terminal 9.
Output from The block length information obtained by the block length determining unit 8 is output from the output terminal 11.

Next, a decoder 20 shown in FIG. 3 will be described as a second embodiment. The decoder 20 receives the analysis parameter, index, and block length information output from the encoder 1 shown in FIG. 1, and reproduces an audio signal based on the information.

The decoder 20 inputs an inverse quantization unit 22 for inversely quantizing the index received via the input terminal 21 and an MDCT coefficient obtained by inverse quantization by the inverse quantization unit 22 to an input. Based on the block length information received via the terminal 23, an IMDCT unit 24 for inverse MDCT, and an audio signal based on an output time-series sample from the IMDCT unit 24 and an analysis parameter received via the input terminal 25 And a synthesizing unit 26 for synthesizing, and outputs an audio signal from an output terminal 27.

Next, a schematic operation of the decoder 20 having the above configuration will be described. The inverse quantum unit 22 includes an input terminal 21
, The index supplied from the encoder 1 is inversely quantized. The IMDCT unit 24 includes the inverse quantization unit 22
The MDCT coefficient obtained in step (1) is calculated using the block length determined from the block length information supplied through the input terminal 23 by the IMD.
CT. The synthesizing unit 26 synthesizes and reproduces an audio signal based on the analysis parameters supplied via the input terminal 25 and the time-series parameters from the IMDCT unit 24.

As described above, the encoder 1 of FIG. 1 is used as the first embodiment, and the decoder 2 of FIG. 3 is used as the second embodiment.
0 has been described. Hereinafter, specific examples of the orthogonal transform device and the inverse orthogonal transform device of the present invention will be described.

Hereinafter, the MDCT unit 5 constituting the encoder 1 shown in FIG. 1 will be described as a specific example of the orthogonal transform apparatus of the present invention.
Further, the IMDCT unit 24 constituting the decoder 20 of FIG. 3 will be described in detail as a specific example of the inverse orthogonal transform device of the present invention. As described with reference to FIG. 8, the MDCT unit 5 uses a symmetrical time window in the conventional MDCT performed using the definition formulas (1) and (2), and 50% overlap with frame, so IM
Since the aliasing included in x エ イ (m) obtained by performing the DCT cannot be canceled, the original time-series samples can be restored by the overlapping method such as the frame j in FIG. It was done to solve the problem of being unable to do so.

Therefore, in the MDCT unit 5, even when the block length is switched as shown in FIG. 8, the above equation (1) and (1) are used so that the original time-series samples can be completely restored. The definition of MDCT and IMDCT in the expression 2) is generalized, and the following expressions (4) and (5) are introduced.

[0043]

(Equation 4)

[0044]

(Equation 5)

In the above equations (4) and (5), x is an input signal, y is an MDCT coefficient, x ブ ロ ッ ク is an inverse MDCT output, M is a block length, h is a window function for forward transform, and f is an inverse transform. Window function,
α is an aliasing boundary (0 ≦ α ≦ M).

In the above equations (4) and (5), a parameter α is newly introduced.
These parameters determine a sample position that is a boundary of aliasing in the samples x to (m) subjected to the IMDCT by the unit 24. If α = M / 2, the above equation (1)
This is the same as the normal MDCT defined by equation (2).

By substituting equation (5) into equation (4), the following equation (6) is obtained.

[0048]

(Equation 6)

Here, ξ (l) is defined as follows.

[0050]

(Equation 7)

When the above equation (6) is rewritten using the above ξ (l), the following equation (7) is obtained.

[0052]

(Equation 8)

Here, ξ (l) is the following equation (8).

[0054]

(Equation 9)

Since 0 ≦ r <M and 0 ≦ m <M in the above equation (6), only the terms satisfying the following equation remain.

[0056]

(Equation 10)

Therefore, the following equation (9) is obtained.

[0058]

[Equation 11]

In the equation (9), what appears in the second term is an aliasing component, and the aliasing is α
The one with the opposite polarity occurs after the second sample. Therefore, by appropriately selecting the windows f (m) and h (m) and matching the aliasing boundary between adjacent frames,
Aliasing can be canceled.

Next, conditions for restoring the original sample will be described. Let the block length in frame j be M
_j , the aliasing boundary is α _j , and the forward transform window is h
_{If j} (m) and the inverse transform window are f _j (m), the conditions for canceling the aliasing and completely restoring the original sample are as follows (10), (11), (12)
It is an expression.

[0061]

(Equation 12)

[0062]

(Equation 13)

[0063]

[Equation 14]

Here, the generalized MDCT unit 5
Shows an example of the block length switching performed in step (a). For simplicity, it is assumed that the forward conversion window h (m) and the inverse conversion window f (m) are the same. Further, except for the block where the switching is performed, normal MDCT (α = M / 2) is performed, and the window is symmetric. That is, when 0 ≦ m <M, the above-mentioned window is set as the following equation.

[0065]

(Equation 15)

At this time, if the following equation is satisfied, the condition for restoring the original sample is satisfied.

[0067]

(Equation 16)

[0068] Under such conditions, consider a switch of block lengths from _{M 1} shown in FIG. 8 _M 2 to _(M 1 <M _2). First, the aliasing boundary α of the j-th frame for which block length switching is performed must satisfy the following expression (13) from the condition of the above expression (10).

[0069]

[Equation 17]

The window used in the frame whose block length is M1 is h _s (m), and the window used in the frame whose block length is M2 is h.
_l (m). The window h _t (m) of the j-th frame is calculated by using
From the condition of the expression (1), the following expression (14) must be satisfied.

[0071]

(Equation 18)

If the conditions of the above equations (13) and (14) are satisfied, it is clear that the above equation (12) is satisfied. Therefore, even in the switching block,
It can be seen that the original time-series samples can be completely restored.

Next, [2]: "Masahiro Iwatare, Takao Nishitani, Akihiko Sugiyama. A Study on MDCT Method and High-Speed Algorithm. IEICE Technical Report, Vol. CAS90-9 DSP90-13, pp. 49-54, 1990. Is MD
A high-speed calculation of CT has been proposed. Using this algorithm, generalized MD defined by the above equations (4) and (5) is used.
CT can also be calculated at high speed. The procedure is described below.

First, the forward conversion will be described. Xh (m) as follows First, define the x 2 _(m).

[0075]

[Equation 19]

The rearrangement operation of the above equation (15) is as follows.
It is equivalent to the equation (11) in [2], and actually α =
If M / 2, it will be exactly the same. When the above equation (4) is rewritten using x ₂ (m), the following equation (16) is obtained.

[0077]

(Equation 20)

This equation (16) is equivalent to equation (12) in [2].
It has the same form as the formula. In [2], a high-speed operation is realized by modifying this equation, and the equation (1) in [2] is realized.
By performing the operation of Expression (15) of the present application instead of 1) and performing the same operation as in [2], the calculation of Expression (4) is applied to the high-speed operation algorithm proposed in [2]. It turns out that it is possible. The calculation procedure is summarized as follows.

First, the input signal xh with the window for forward conversion applied
(m) is rearranged according to the above equation (16).

Next, according to the following equation (17), x
Create x ₃ (m) from ₂ (m).

[0081]

(Equation 21)

Next, exp (−j · (2πm / M)) is applied to x ₃ (m).
To produce a complex signal z ₁ (m) shown in the following equation (18).

[0083]

(Equation 22)

Next, an M / 2-point FFT is performed on z ₁ (m).
Z ₂ (k) shown in the following equation (19) is obtained.

[0085]

(Equation 23)

Finally, MDCT coefficients are extracted from the output of the FFT as in the following equation (20).

[0087]

(Equation 24)

In the case of the inverse transform, the high-speed operation of [2] can be applied as in the case of the forward transform. In the case of the inverse transform, the calculation may be performed in the same procedure except for the polarity inversion and rearrangement of the last time series sample.

That is, the coefficients are rearranged as in the following equation (21).

[0090]

(Equation 25)

Next, exp (-j · (2πk / M)) is applied to y ₂ (k).
To produce a complex signal z ₁ (k) shown in the following equation (22).

[0092]

(Equation 26)

Next, an inverse FFT of M / 2 points is performed on z ₁ (k) to obtain z ₂ (m) shown in the following equation (23).

[0094]

[Equation 27]

From the output of the inverse FFT, the following (24)
Extract x ₀ ~ (m) as in the equation.

[0096]

[Equation 28]

Finally, the polarity inversion and rearrangement of x ₀ to (m) are performed to obtain the output x to (m) of the IMDCT of the following equation (25).

[0098]

(Equation 29)

Here, the number of input points will be described. When the block length is switched by the method of the present invention, there are cases where the block length M of the frame to be subjected to the conventional MDCT is a power of 2, but does not become a power of 2 in the frame to be switched. For example, this is a case where the block lengths and the aliasing boundaries of the j-1 frame and the j + 1 frame are respectively represented by the following equations.

[0100]

[Equation 30]

[0101]

(Equation 31)

Then, the block length M _j of the j-th frame is given by the following equation (26) from the condition of the above equation (10).

[0103]

(Equation 32)

At this time, if a <b, M _j is
It becomes like the following formula.

[0105]

[Equation 33]

That is, it does not become a power of two. In this example
In the j-frame, an FFT or IFFT that is not a power of 2 in the above equations (19) and (23) must be performed. Normally, FFT and IFFT are based on the premise that the number of points is a power of two, and the number of points that does not apply to this cannot be calculated. Therefore, when such an FFT device is used, the above calculation becomes impossible.

By the way, when P is an odd number and Q is an integer, P × 2
^As a method of realizing the ^Q- point FFT and IFFT, there are a fast Fourier transform method and a fast inverse Fourier transform method applied by the present applicant. By using this method, high-speed operation can be performed even in the case of the above example.

Among them, the fast Fourier transform method uses a complex number data of P × 2 ^Q points as input data when P is an odd number and Q is an integer, performs a fast Fourier transform on the input data, and obtains a P × 2 ^Q Output the complex data of point. Specifically, the N points that make up the array x are odd Q
Extract P point data for each N / Q divided by
Performing discrete Fourier transform on the data of the Q point, the twist coefficient by multiplying the Q-point discrete Fourier transform coefficients obtained, the multiplication result from the back to the sequence x, 2 ^Q in areas divided into P It performs point fast Fourier transform.

Next, FIG. 4 shows an example of frame switching when a conventional method is used as a method for switching the MDCT block length. A short block length is selected for frame j + 2 at frame j + 1, and a long block length is selected for frame j + 4 at frame j + 3. As is apparent from FIG. 4, 25 frames between frame j + 1 and frame j + 2.
It can be seen that the phase is shifted by 6 samples, and similarly, there is a shift of 256 samples between frame j + 2 and frame j + 3. In the frames j + 2 and j + 4, the processing prior to the MDCT (linear / non-linear prediction) needs to consider the deviation, so that special processing such as changing the block length of the preprocessing is required.

FIG. 5 shows an example of how to take the window of the MDCT block when the signal of FIG. In the frame j + 2, a short block length is selected, and in other cases, a long block length is selected. Unlike the case of FIG. 4, no phase shift occurs, so the MDCT
No earlier need for exceptional processing.

According to the above-described embodiment, a phase shift occurs when the MDCT block length is switched in the transform coding in which the MDCT is performed on the prediction residual signal or the like after the preprocessing is performed. The block length of the pre-processing can be kept constant without any problem. Further, in the conventional method, even when a phase shift occurs in a frame to be switched, the block length can be switched without causing a phase shift.

[0112]

According to the orthogonal transform apparatus and method according to the present invention, it is possible to arbitrarily determine the break of the overlap.

According to the inverse orthogonal transform apparatus and method according to the present invention, the orthogonal transform coefficients obtained by the above orthogonal transform apparatus and method can be inversely transformed.

The apparatus and method for transform coding according to the present invention
It is possible to arbitrarily determine the break of the overlap and to enable complete signal reproduction by overlap addition.

The decoding apparatus and method according to the present invention can decode coded data coded by the above-described transform coding apparatus and method.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of an encoder according to a first embodiment of the present invention.

FIG. 2 is a diagram in which a sample sequence of a certain audio signal is plotted.

FIG. 3 is a block diagram illustrating a configuration of a decoder according to a second embodiment of the present invention.

FIG. 4 is a diagram for explaining a specific example of a conventional method of switching a block length.

FIG. 5 is a diagram for explaining a specific example of block length switching in the present invention.

FIG. 6 is a diagram for explaining an MDCT algorithm.

FIG. 7 is a diagram showing a specific example of block length switching in a conventional method.

FIG. 8 is a diagram illustrating a specific example of block length switching in which there is no 0 in a window.

[Explanation of symbols]

1 encoder, 3 linear / nonlinear predictive analysis section, 5M
DCT section, 6 quantization section, 7 stationary estimation section, 8 block length determination section, 20 decoder, 22 inverse quantization section, 24
IMDCT section

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Masayuki Nishiguchi 6-35 Kita Shinagawa, Shinagawa-ku, Tokyo Sony Corporation F-term (reference) 5D045 DA02 5J064 BA12 BA13 BA16 BB03 BB05 BC02 BC08 BC16 BD01 5K041 AA01 CC01 CC02 EE11 HH09

Claims (18)

[Claims] 1. An orthogonal transform apparatus for orthogonally transforming an input time-series sample while overlapping the time-series sample. When orthogonally transforming a time-series M sample, a sample position α which is a boundary at which aliasing occurs at the time of inverse orthogonal transform is set to 0. An orthogonal transformation device, wherein the orthogonal transformation is performed arbitrarily determined within a range of ≦ α <M. 2. The orthogonal transform apparatus according to claim 1, wherein the sample position α is adjusted between adjacent frames. 3. The orthogonal transform apparatus according to claim 2, wherein the sample position α is matched between adjacent frames by appropriately selecting a window function. 4. The orthogonal transform apparatus according to claim 3, wherein said window function does not include a zero component. 5. An orthogonal transform method for orthogonally transforming input time-series samples while overlapping the time-series samples. When orthogonally transforming time-series M samples, a sample position α, which is a boundary at which aliasing occurs at the time of inverse orthogonal transform, is set to 0. An orthogonal transformation method characterized by arbitrarily determining an orthogonal transformation in a range of ≦ α <M. 6. An inverse orthogonal transform apparatus for inversely orthogonally transforming orthogonal transform coefficients obtained by orthogonally transforming time-series samples while overlapping the time-series samples, wherein a sample position α at which aliasing occurs at the time of inverse orthogonal transform is set to 0 ≦ α. An inverse orthogonal transform apparatus, which performs inverse orthogonal transform on orthogonal transform coefficients that have been arbitrarily determined and orthogonally transformed in a range of <M. 7. An inverse orthogonal transform method for performing inverse orthogonal transform on orthogonal transform coefficients obtained by orthogonally transforming time-series samples while overlapping the time-series samples, wherein a sample position α that is a boundary at which aliasing occurs at the time of inverse orthogonal transform is 0 ≦ α. <Inverse orthogonal transformation method characterized by performing inverse orthogonal transformation on orthogonal transformation coefficients that have been arbitrarily determined and orthogonally transformed in the range of M. 8. A transform coding apparatus for orthogonally transforming an input signal and compressing and encoding the input signal, comprising: a predictive analysis unit that captures the input signal by predetermined samples, performs a predictive analysis, and outputs a prediction residual; Characteristic determining means for determining a characteristic for each predetermined sample; block length determining means for determining a block length at the time of orthogonal transformation based on the characteristic determined by the characteristic determining means; and a block determined by the block length determining means A sample position α which is a boundary at which aliasing occurs at the time of inverse orthogonal transformation is arbitrarily determined within a range of 0 ≦ α <M, and the prediction residual output from the prediction analysis unit is input to the input time series M
Orthogonal transform means for performing orthogonal transform processing on the time series M samples to generate orthogonal transform coefficients while overlapping as samples, and quantizing means for quantizing the orthogonal transform coefficients generated by the orthogonal transform means. A transform encoding device, comprising: 9. The transform coding apparatus according to claim 8, wherein said orthogonal transform means matches a sample position α of said time-series M samples to be subjected to an orthogonal transform process between adjacent frames. 10. The transform code according to claim 9, wherein said orthogonal transform means matches a sample position α of said time-series M samples to be subjected to an orthogonal transform process between frames by appropriately selecting a window function. Device. 11. The transform coding apparatus according to claim 10, wherein said window function does not include a zero component. 12. The transform coding apparatus according to claim 8, wherein said characteristic judging means judges the continuity of said input signal for each predetermined sample. 13. The block length determining means determines the block length when the characteristic determining means determines quasi-stationarity with a small time variation of the signal, and determines that the signal temporal variation is large by the characteristic determining means. Than the above block length,
13. The transform encoding device according to claim 12, wherein the length is made longer. 14. An apparatus according to claim 8, wherein said input signal is a speech signal and / or an audio signal. 15. The quantized data is from 6 Kbps to 32 Kbps.
The transform coding apparatus according to claim 8, wherein the output is performed at a rate of kbps. 16. A transform encoding method for orthogonally transforming an input signal and compressing and encoding the input signal, wherein a prediction analysis step of capturing the input signal by predetermined samples, performing a prediction analysis and outputting a prediction residual, A characteristic determining step of determining a characteristic for each predetermined sample; a block length determining step of determining a block length at the time of orthogonal transformation based on the characteristic determined in the characteristic determining step; and a block determined in the block length determining step A sample position α, which is a boundary at which aliasing occurs at the time of inverse orthogonal transformation, is arbitrarily determined within a range of 0 ≦ α <M, and the prediction residual output from the prediction analysis step is input time series M
An orthogonal transformation step of performing orthogonal transformation processing on the time-series M samples to generate orthogonal transformation coefficients while overlapping as samples, and a quantization step of quantizing the orthogonal transformation coefficients generated in the orthogonal transformation step. A transform encoding method comprising: 17. An input time series in which a block length determined according to characteristics of an input signal and a sample position α serving as a boundary at which aliasing occurs at the time of inverse orthogonal transform is arbitrarily determined within a range of 0 ≦ α <M. What is claimed is: 1. A decoding apparatus for decoding quantized data obtained by quantizing orthogonal transform coefficients obtained by orthogonally transforming M samples while overlapping each other, comprising: inverse quantization means for inversely quantizing the quantized data; A decoding device comprising: an inverse orthogonal transform unit that performs an inverse orthogonal transform on an orthogonal transform coefficient obtained by inverse quantization by a quantization unit with a block length determined according to the characteristics of the input signal. . 18. A block length determined according to characteristics of an input signal, and a sample position α serving as a boundary at which aliasing occurs at the time of inverse orthogonal transform is arbitrarily determined within a range of 0 ≦ α <M to obtain an input time series. A decoding method for decoding quantized data obtained by quantizing orthogonal transform coefficients obtained by orthogonally transforming M samples while overlapping each other, comprising: an inverse quantization step of inversely quantizing the quantized data; An inverse orthogonal transformation step of performing an inverse orthogonal transformation on the orthogonal transform coefficient obtained by the inverse quantization in the quantization step with a block length determined according to the characteristics of the input signal. .