JP3472974B2 - Acoustic signal encoding method and acoustic signal decoding 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 coding method and a decoding method for converting an audio signal such as voice or music into a frequency domain for efficient quantization.

[0002]

2. Description of the Related Art A method of quantizing a sound signal such as voice or music in the frequency domain when it is encoded with a small number of bits is well known. For the conversion, DFT (Discrete Fourier Transform), DCT (Discrete Cosine Transform) MDCT (Modified Discrete Cosine Transform), etc. are used. It is also known that linear prediction analysis is effective for the purpose of flattening the frequency domain coefficient before quantization. An acoustic signal conversion coding method and a decoding method (Japanese Patent Application No. 7-52389) are examples of methods of combining these techniques to realize high-quality coding for a wide range of acoustic signals. This process is illustrated
3 shows.

The acoustic signal, which has been converted into a digital signal, is overlapped by N pieces for each N input samples (one frame) in the frame division processing 11 to extract an input sequence of the past 2 × N samples, and 2 × N pieces thereof are extracted. In the time windowing process 12, a 2N window function (time window) is applied to the sample sequence of.
For example, a Hanning window is used as the window function W (n).

The signal x (n) multiplied by this window function is an example.
For example, Nth-order MDCT (Modified Discrete)
te Cosine Transform: Modified Discrete Co
Sine transformation) In step 13, the modified discrete cosine transformation is performed and the
Wavenumber domain coefficient (the sun at each point on the frequency axis
Pull value) y (k). Also obtained by windowing process 12
The obtained signal x (n) is linearly predicted in the linear prediction analysis process 14.
Analyzed, P-order prediction coefficient α₀，… ， Α_pIs sought
It This prediction coefficient α₀，… ， Α_pIs an example of quantization processing 15
For example, it is converted into LSP parameters and then quantized
Index I showing the outline of the vector_pIs obtained. This
Example, the LSP parameter quantized by the quantization processing 15
The linear prediction coefficient α
₀，… ， Α _pPower spectrum envelope (power transfer function)
The square root of is calculated as the approximation of the amplitude envelope of MDCT coefficients.
It This calculated spectral envelope gives the MDCT coefficient
Is divided by the flattening process 17. Also the spectrum outline
For the spectrum envelope calculated in calculation process 16,
The weight calculation processing 18 calculates a weighting coefficient according to the auditory characteristic.
And using this weighted coefficient, the flattening process 17
The factorized MDCT coefficient v (k) is the weighted quantization process.
Quantization with the weighted auditory sense
Decks I_MIs output.

In the decoding method, the indexes I _P and I _M are inversely quantized in an inverse quantization process 21, and an LSP parameter and a flattening coefficient v ^ (k) are obtained, and the LSP parameter is a spectrum outline calculation. In process 22, the square root of the spectral envelope characteristic is calculated, the flattening coefficient v ^ (k) is divided by the result of the inverse flattening process 23, and the division result x ^ (n) is obtained in the inverse MDCT process 24. The inverse modified discrete cosine transform is performed to form a time domain signal x ^ (n), and the time domain signal x ^ (n) is one frame (N samples).
The window function 25 multiplies the window function on the 2N samples extracted by overlapping N pieces for each of the samples. The 2 × N samples multiplied by the window function are superposed 26
Then the first half N samples of the current frame and the second half N of the previous frame
The sample and the sample are added together, and the N samples are used as the reproduced sound signal of the current frame.

Although MDCT has an advantage that no frame boundary noise is generated, the signals in the time domain are folded back by the operations of the transform and the inverse transform, and the original waveform in the time domain is not reproduced. Therefore, it cannot be combined with a linear prediction filter or a synthesis filter. Further, in order to flatten the MDCT coefficient with the spectrum envelope of the linear prediction, it is necessary to obtain the reciprocal of the spectrum envelope characteristic. In the encoding process, it is necessary to multiply the reciprocal of the square root of the spectrum envelope characteristic for each transform coefficient sample, and in the decoding process, it is necessary to divide the sample of the transform coefficient by the reciprocal number of the square root of the spectrum envelope characteristic.
Especially in the decoding process, the amount of calculation such as division has been a problem.

For DCT (discrete cosine transform), M
Since the signal in the time domain is not folded back equivalently like the DCT, the operation of the normal inverse filter in the time domain becomes the reversible operation by the operation of the synthesis filter. This process is shown in FIG. In this case, the input digital acoustic signal is divided into every N samples in the frame division processing 31, and this is subjected to the inverse filtering in the inverse filtering processing 32 to obtain a linear prediction residual waveform, and the DCT is applied to this residual waveform. DCT is performed in the processing to obtain frequency domain coefficients, and the frequency domain coefficients are subjected to auditory weighted quantization in the weighted quantization processing 34, and an encoded code is output. In the decoding process, the input code is inversely quantized by the inverse quantization process 35 to reproduce the frequency domain coefficient, which is the inverse DCT process 3
In step 6, the inverse discrete cosine transform is performed to form a residual waveform in the time domain, and the residual waveform is passed through a linear prediction synthesis filter in synthesis filter processing 37 to reproduce an acoustic signal.

By using the DCT as described above, it is possible to obtain a coefficient that has been flattened by transforming the signal that has passed through the inverse filter of the linear prediction analysis into the frequency domain. The original spectral envelope characteristic can be reproduced by passing it through a synthesis filter. However,
In DCT, discontinuous noise at the frame boundary may be a problem.

[0009]

SUMMARY OF THE INVENTION It is an object of the present invention to provide an acoustic signal coding method and an acoustic signal decoding method that reduce the amount of processing and improve prediction efficiency while suppressing noise at frame boundaries.

[0010]

According to the encoding method of the present invention, by overlapping with the previous frame and applying a window function of twice the update length N of the frame, the sample x (i),
(I = 0, ..., 2N-1) is created, and N / in x (i)
2 length samples x (i), (i = 0, ..., N / 2−
1) is temporally inverted and then subtracted from samples x (i), (i = N / 2, ..., N-1) of length N / 2 in x (i), respectively, and x (i) Medium N / 2 length sample x
(I), (i = N + N / 2, ..., 2N-1) are temporally inverted, and then x (i) having a length of N / 2 in x (i),
(I = N, ..., N + N / 2−1) is added respectively to create y (i) at N points, and proximity prediction (for example, partial autocorrelation prediction) and long-term prediction (for example, for y (i)) are created. Pitch prediction) or both linear prediction analysis,
Prediction residual signals z (i), (i = 0, ..., N-1) are created through y (i) in an inverse filter using the prediction coefficient as a coefficient, and z (i) is subjected to cosine transform in the frequency domain. Coefficient v
(I), (i = 0, ..., N-1) is created, and its v
Quantize (i) to obtain the encoded output.

In this case, it is determined whether the input signal is a voice having pitch periodicity or general music, and when the determination is highly likely to be a voice, the overlap over the frame of the window function is reduced (zero overlap). , The window function can be overlapped over the frame when the possibility of music is strong. According to the decoding method of the present invention, the frequency domain coefficients v ^ (i), (i = 0, ..., N) created by inverse quantization.
-1) is subjected to inverse cosine transform to reproduce residual signal z ^ (i),
(I = 0, ..., N−1) is created, and linear prediction analysis (backward prediction) of either or both of near-field prediction and long-term prediction is performed, or the input code is decoded to calculate the prediction coefficient. A signal y ^ (i), (i =) is generated (forward prediction) and is passed through z ^ (i) to a linear prediction synthesis filter using the coefficient.
0, ..., N-1), and the first half of N in y ^ (i) is N /
2 length samples y ^ (i), (i = 0, ..., N / 2
-1) is multiplied by -1, inverted in time, extended before the frame of y ^ (i), and the latter half of Y ^ (i) has a length of N / 2 y ^ (i). , (I = N / 2, ..., N-1) is expanded after the frame of y ^ (i) that is temporally inverted, and samples x ^ (i), (i = 0, ..., N-1) ), And a window function of twice the frame length N is applied to this x (i), and the waveforms of the latter half and the first half of the preceding and following frames are superimposed to obtain a reproduced acoustic signal.

As described above, in the conventional transform coding that also uses linear prediction analysis, MDCT and time domain filter operation cannot coexist. However, in the present invention, MDCT is subjected to a preprocessing step and a DCT step. And
Further, the MDCT is divided into an inverse DCT step and a post-processing step, and inverse filter processing and synthesis filter processing are applied to signals in the middle of the steps, and noise at the frame boundary is suppressed, It is possible to reduce the amount of calculation and distortion by predicting the time domain.

[0013]

1A shows an embodiment of an encoding method according to the present invention. This is the same as the conventional MDCT-based encoding method in that the N-point frequency domain coefficients are obtained and quantized from 2N input waveforms having N overlaps. That is, by dividing the input signal to the waveform of frame division processing N points one by duplicate 2N points 11, relative to the waveform of each 2N points, window function 2N points length in windowing processing 12 w _in (n)
2N point samples x (i), (i = 0, ..., 2N
-1) is obtained.

In the present invention, preprocessing 41 replaces the samples. To make this easier to understand, consider the case of normal MDCT. The sequence multiplied by the window function is x
(N), this x (n) is the MD defined below.
When CT is applied, the coefficient Y (k) at N points is obtained. Y (k) = Σ _{i = 0} ^2N-1 x (i) cos {π (2i + 1 + N) (2k +
1) / (4N)} (1) The transform function of this MDCT, that is, cos {π (2i + 1 +
N) (2k + 1) / (4N)} is, for example, N = 32, as shown in FIGS. 5A, 5B, 5C, and 5D for k = 0, k = 1, k = 2, and k = 31. . In the first half of these curves (0 to N, i / N = 0 to 1), N / 2 (i
/N=0.5) is an odd symmetric function, and in the latter half (N to 2N, i / N = 1 to 2), 3N / 2 (i / N =
It is an even symmetric function centered on 1.5). In the present invention, using this property, MDCT is not directly performed but divided into pre-processing and cosine transformation for processing. Samples in the time domain by MDCT are i / N = 0 to 0.5, i / N =
It is multiplied by 1.5 to 2, respectively. And M
Since the DCT is a product-sum operation of x (i) and the function of FIG. 5 as shown in the equation (1), in the preprocessing 41, N / 2 (i / N = 0 to 0 from the beginning of each frame). .5) sample x (i), (i = 0, ..., N / 2−1) is temporally inverted and then sample x (i) of length N / 2 next,
(I = N / 2, ... , N- 1: i / N = 0.5 to 1), and also the sample of the last N / n length of the frame x
(I), (i = N + N / 2, ..., 2N-1: i / N =
1.5 to 2) is temporally reversed, and immediately before that, N / 2
Samples of length x (i), (i = N, ..., N + N / 2
-1: i / N = 1, ..., 1.5)
The pre-processing 41 is performed.

That is, the folding process of the preprocessing 41 can be expressed by the following equation. y (i) = x (N / 2 + i) -x (N / 2-1-i), (i = 0, ..., N / 2-1) y (i) = x (N / 2 + i) + x (5N / 2-1-i), (i = N / 2, ..., N-1) (2) The following N-point DCT (discrete cosine transform) is performed on the folding result y (i). And normal MDCT
It is the same as the coefficient v (k).

V (k) = Σ _{i = 0} ^N−1 y (i) cos {π (2i + 1 + 2N) (2k + 1) / (4N)} (3) In the present invention, an inverse filter is applied to y (i). In process 42, a prediction residual signal z (i) is obtained through a linear prediction inverse filter in which α _j , (j = 1, ..., P) are p-th prediction coefficients, that is, the following equation is calculated.

Z (i) = y (i) + Σ _{j = 1} ^p α _j y (i−j), (i = 0, ..., N−1) (4) The prediction coefficient α _j is y (i ) Is obtained by linear prediction analysis. Further, at the beginning of the frame, that is, at i <p, the last (pi) point of y ^ (i) of the previous frame may be used instead of the current frame y ( i).

Next, the same cosine transform as in equation (3) is performed on z (i) by the DCT process 33, and the frequency domain coefficient v
(I) is obtained, and this v (i) is weighted and quantized 19
Quantize with audio weights. In other words, MDCT preprocessing 4
1 and the DCT process 33, and the inverse filter process 42 is performed during the process. The prediction residual signal has a substantially flat spectrum envelope, and when the cosine transform (DCT) is applied to z (i), a substantially flat coefficient v (i) is obtained in all bands. For this reason, the normal MDCT coefficient flattening process 17 (FIG. 3A) is not required even when quantizing. The inverse filter processing 42 is proximity prediction using the linear prediction coefficient α _i in the above example, but may be long-term prediction such as pitch prediction or both. Prediction coefficient α _i
Alternatively, they may be separately quantized and transmitted (forward prediction), or may be estimated from past synthesized waveforms (backward prediction). For quantization of the DCT coefficient, quantization on a distance scale with a weight of the spectrum envelope or adaptive bit allocation quantization according to the spectrum envelope is preferable.

An embodiment of the decoding method of the present invention will be described with reference to FIG. 1B. In decoding, 2N-point waveforms are created by inverse transformation from N-point coefficients, and N frames are superposed on the preceding and following frames and synthesized, which is the same as the conventional MDCT-based decoding method. The flattened DCT coefficient v ^ (k) is reproduced from the code input in the inverse quantization process 21. In the present invention, for this coefficient v ^ (k), the inverse D
In the CT process 36, the inverse DCT is performed by the calculation of the equation (5) to reproduce the residual signal z ^ (i).

Z ^ (i) = Σv ^ (k) cos (π (2i + 1 + 2N) (2k + 1) / (4N)) (5) Σ is the next reproduction residual signal z from k = 0 to N−1. ^
For (i), the synthesis filter processing 44 performs the linear prediction synthesis filter processing by the calculation of the equation (6).

[0021] y ^ (i) = z ^ (i) -Σ j = 1 p α j y ^ (i-j), (i = 0, ..., N-1) ... (6) It should be noted that the beginning of the frame, That is, when i <p, the last (p-i) point of y ^ (i) of the previous frame may be used instead of the current frame y ^ (i). Post-processing 4 from this y ^ (i)
At 5 the signal x (i) at 2N points is reproduced.

In this post-processing 45, the first half N of each frame is processed.
Samples of length y ^ (i), (i = 0, ..., N /
2-1) is multiplied by -1, and the order is inverted in time to extend before the frame of y ^ (i), and the sample y ^ (N ^) of the latter half of each frame has a length of N / 2. i), (i = N /
2, ..., N−1) are temporally reversed in order, and y ^
After the frame of (i), it is expanded to make x ^ (i).
That is, x ^ (i) is as follows.

When i = 0, ..., N / 2, x ^ (i) = -y ^ (N / 2-1-i) i = N / 2, ..., 3N / 2-1, x ^ (i) = Y ^ (i−N / 2) i = 3N / 2, ..., 2N−1 and x ^ (i) = y ^ (5N / 2−1−i) (7) Next, in the windowing process 25 The window function w _out (i) is applied to x ^ (i), and then, in the superposition process 26, the first half x ^ (i) of the current frame is superposed with the latter half x ^ (i) of the previous frame, and the output waveform , That is, to obtain a reproduced sound signal. It should be noted that it is if there is a following relationship between the window function of the input w _in (i) and the output of the window function w _out (i).

W _out (i) win _in (i) + w _out (2N-1-i) win _in (2N-1-i) = 1, (i = 1, ..., 2N-1) (8) Another embodiment of the encoding method of the present invention switches the window function according to an input signal. _{w in (i) = sin (} iπ / (2N)), i = 0, ..., because it has a 2N-1 ... (9) is of 50% if the window is defined by the overlap, the frame also in a steady music Boundary noise hardly occurs. However, when the change is relatively fast or when the pitch period is clear, the effect of linear prediction and pitch prediction in the time domain becomes larger and the distortion becomes smaller when the overlap is reduced.
If the length of the overlapping portion is M, the window will be as follows.

[0025] _{w in (i) = 0 0} <i <N / 2M / 2 w in (i) = sin ((i-N / 2 + M / 2) π / (2M)) N / 2M / 2 <<N / 2 + M / 2 w in (i) = 1 N / 2 + M / 2 <3N / 2M / 2 w in (i) = sin ((i-3N / 2 + 3M / 2) π / (2M) in) 3N / 2-M / 2 <i <3N / 2 + M / 2 w in (i) = 0 3N / 2-M / 2 <i < the case of a window of 2N ... (10) equation (9) M = N 2A, the overlap at M = N / 2 is as shown in FIG. 2B, and at M = 0 the overlap is zero as shown in FIG. 2C. If M is small, pitch prediction can be performed and voice distortion can be reduced, but frame boundary noise may occur in stationary sounds. Therefore, in the case of voice input, M =
It is set to 0, and in the case of music input, various inputs can be dealt with by adaptively adjusting the overlap M such that M = N and selecting the most preferable overlap. At this time, if a code designating the overlap is sent as auxiliary information, it can be reproduced by the decoder. That is, in the decoding method, the overlap of the window functions over the frames is changed according to the code that specifies the overlap in the input code.

Since the inverse filter and the quantization process after applying the window can be efficiently performed in common regardless of the input, the configuration can be made more compact than having a plurality of encoding systems and decoding systems for switching.

[0027]

According to the present invention, it is possible to maintain the continuity at the frame boundary by the superposition in the time domain.
While utilizing the characteristics of DCT, time domain prediction and filter processing can be performed, and quantization distortion can be reduced. Further, the spectrum envelope in the decoder and the division for each MDCT coefficient can be replaced with the calculation of the synthesis filter, and the calculation amount can be reduced.

[Brief description of drawings]

FIG. 1A is a diagram showing a process of encoding of the present invention, B
FIG. 6 is a diagram showing a decoding process of the present invention.

FIG. 2 is a diagram showing various examples of overlapping states of window functions for explaining a third embodiment of the present invention.

3A is a diagram showing a processing step of a conventional encoding method based on MDCT, and FIG. 3B is a diagram showing a processing step of the decoding method.

4A is a diagram showing a conventional encoding process based on a DCT and a synthesis filter, and FIG. 4B is a diagram showing a decoding process thereof.

FIG. 5 is a diagram showing an example of transform coefficients of MDCT.

Front Page Continuation (72) Inventor Kazunaga Ikeda 3-19-2 Nishishinjuku, Shinjuku-ku, Tokyo Inside Nippon Telegraph and Telephone Corporation (72) Inventor Satoshi Miki 3-19-3 Nishishinjuku, Shinjuku-ku, Tokyo Nihon Telegraph and Telephone Corporation (56) Reference JP-A-4-44099 (JP, A) JP-A-6-232824 (JP, A) (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 19/08 G10L 19/04

Claims (4)

(57) [Claims] 1. A coding method for transforming an input signal of a frame unit consisting of a plurality of (N) samples into a frequency domain and quantizing the same, in a window function having a length twice that of N by overlapping with a previous frame. X (i), (i = 0, ..., 2N−
The process of creating 1) and the sample x of length N / 2 in the sample x (i)
(I), (i = 0, ..., N / 2−1) is temporally inverted
It was above SL samples x (i) the length in N / 2 of x (i),
Subtract from (i = N / 2, ..., N-1) respectively, x
Samples of length N / 2 in (i) x (i), (i = N +
N / 2, ..., the length in the 2N-1) a temporally inverted x (i) N / 2 of x (i), (i = N, ..., N + N / 2-
1) and y (i), (i =
0, ..., N-1), that is, i = 0, ..., N / 2-1, y (i) = x (N / 2 + i)
−x (N / 2−1−i) i = N / 2, ..., N−1, and y (i) = x (N / 2 + i)
+ X (5N / 2−1−i) is calculated, and y (i) is subjected to linear prediction analysis of either proximity prediction or long-term prediction or both, and y is applied to the inverse filter using the prediction coefficient as a coefficient. Prediction residual signal z by passing through (i)
(I), (i = 0, ..., N-1), and the frequency domain coefficients v (i), (i = 0, ...) By the cosine transform of the prediction residual signal z (i). N-1) is created, and the frequency domain coefficient v (i) is quantized to obtain a coded output. 2. A process for determining whether the input signal is voice having pitch periodicity or general music, and when the determination is highly likely to be voice, the overlap of the window function over frames is reduced. The method of encoding an acoustic signal according to claim 1, wherein the overlap of the window function over the frame is increased when the possibility of music is strong. 3. A method of reproducing a voice acoustic signal in units of frames from a code quantized by converting it into a frequency domain, wherein a frequency domain coefficient v ^ (i), which is created by inverse quantization,
(I = 0, ..., N-1) is processed by inverse cosine transform to generate a reproduction residual signal z ^ (i), (i = 0, ..., N-1), which is close prediction or long-term prediction. Either or both linear prediction analysis (backward prediction) is performed, or an input code is decoded to create a prediction coefficient (forward prediction), and a linear prediction synthesis filter using the coefficient is used to reproduce the residual signal. z ^
By passing through (i), samples y ^ (i), (i = 0,
, N-1), and the sample y ^ (i), (i = 0, ..., N / 2-1) of the first half length N / 2 in the sample y ^ (i). Multiply by -1, inverted in time, extended before the frame of sample y ^ (i) above, and the second half length in sample y ^ (i)
N / 2 samples y ^ (i), (i = N / 2, ..., N
-1) is expanded after the frame of y ^ (i) which is temporally inverted, and x ^ (i), that is, x ^ (i) = -y ^ (at i = 0, ..., N / 2. N / 2-1-
i) i = N / 2, ..., 3N / 2-1, and x ^ (i) = y ^ (i-
N / 2) i = 3N / 2, ..., 2N-1 and x ^ (i) = y ^ (5N /
2-1-i), and a process of multiplying x ^ (i) by a window function that is twice the frame length N and superimposing it on the waveforms of the preceding and following frames. Acoustic signal decoding method. 4. The method of decoding an acoustic signal according to claim 3, wherein the overlap of the window functions over the frames is changed in accordance with the information designating the overlap in the input signal.