JP3434260B2 - Audio signal encoding method and decoding method, these devices and program recording medium - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention provides a coding method for converting an audio signal into a digital code with high efficiency and a decoding method for restoring the digital code into an audio signal. Or
It can be used for transmission and broadcasting of audio signals using communication channels.

[0002]

2. Description of the Related Art As a conventional method for encoding an audio signal with high efficiency, there is a transform encoding method shown in FIG. 1, for example. In the transform coding, the audio signal input as the discrete signal sample sequence is time / frequency transformed by the time / frequency transforming unit 11 for each input of a fixed number of samples, and is transformed into a coefficient in the frequency domain before being encoded. The processing unit 2 pre-processes the coefficients in the frequency domain, and then the quantization unit 3 quantizes the coefficients. Frequency domain weighted interleaved vector quantization
The ave Vector Quantization (TWIN VQ) method applies to this example.

The TWIN VQ method uses weighted interleaved vector quantization as the final quantization method in the quantization unit 3. Further, the vector quantization can be performed more efficiently as the distribution of the input coefficients is smaller, so that the coefficient is flattened in two steps in the preprocessing by the preprocessing unit 2 in FIG. There is. In the first-stage flattening, the overall coefficient variation is flattened by normalizing the coefficients in the frequency domain with the spectrum obtained by the linear prediction method. In the second-stage flattening, the coefficients in the frequency domain are normalized for each of the small bands divided into equal widths on the Bark scale, which is one of the frequency scales,
The flattening is performed more finely than the linear prediction spectrum.

The Bark scale is characterized in that if the band is divided into equal widths, the widths are perceptually almost equal. Subbands having a uniform width on the Bark scale are almost equal in width perceptually, but as shown in FIG. 2, the subbands divided in this way are smaller in band on higher frequencies on the linear scale. The width is getting wider. Therefore, when the coefficient group in the frequency domain is divided so as to have the same width on the Bark scale, the smaller band located at the higher frequency contains more coefficients.

The reason why the flattening of the second stage is performed in units of Bark scale is to effectively distribute a limited amount of information in consideration of hearing. The operation of flattening by normalization for each sub-band on the Bark scale is performed in the expectation that the coefficients in the sub-band are stationary, but the sub-band located at a high frequency contains many coefficients. Therefore, there are many cases where the small band is not stationary as shown in FIG. This leads to a reduction in the efficiency of vector quantization, resulting in deterioration of the sound quality of the decoded audio signal. Such a problem is particularly likely to occur when an audio signal including many tone components in the high frequency band is input.

Regarding the TWINVQ method, Iwagami et al., "Frequency Domain Weighted Interleaved Vector Quantization (TwinVQ)"
Musical Encoding by "The Journal of the Institute of Electronics, Information and Communication Engineers Vol.J80-A,
Details are described in pp.830-837 (1997). In the encoding of the form shown in FIG. 1, there is also an encoding method that uses adaptive bit allocation scalar quantization as a quantization method. In such an encoding method, the coefficient in the frequency domain is divided into small bands, and optimal bit allocation is performed for each small band. As a method of dividing the small band, there is a case of dividing it into equal widths on the Bark scale for the purpose of auditory optimization.In this way, when it is not steady in the small band located at a high frequency as in the case of TWINVQ. However, there is a problem that the quantization efficiency decreases.

As a coding method for solving such a problem, when the input signal is converted into a frequency domain signal and is flattened (normalized) for each small band, the normalized bandwidth is changed according to the shape of the spectral distribution. It is disclosed in Japanese Patent Application Publication No. 7-336232 that such changes are adaptively made. In this method, the normalization bandwidth is reduced in the small band containing the tone component,
By widening the normalization bandwidth in other regions, the number of flattening small bands is reduced as a whole, and the coding efficiency is improved. However, according to this method, when the tone components exist sparsely, the fine normalization width may be applied to the flat portion near the tone components, and the coding efficiency may decrease. Further, since the normalization information must also be encoded and attached for each tone component, if a large number of dispersed tone components exist, the code amount of the normalization information increases accordingly.

Japanese Patent Application Publication No. 7-168593 discloses that the tone component and the other components are separated and encoded in order to improve the encoding efficiency. In this method, normalized as a spectrum and toe NNaru component signal of a group of spectral near each spectrum of each local maximum, for encoding, the information of the position and size of each local maximum spectrum by encoding Must be attached. Therefore, if the tone Ingredient there are many, it would require a lot of information about the position and size, which may cause that prevents the improvement of the coding efficiency.

Japanese Patent Application Laid-Open No. 7-248145 discloses that pitch components in which tone components are arranged at equal intervals are separated and encoded respectively. The positional information of the pitch component is given by the fundamental frequency of the pitch and requires a small amount of information, but there is a problem that the tone component cannot be separated well for sounds that have a non-integer overtone structure such as metal sounds. .

[0010]

SUMMARY OF THE INVENTION An object of the present invention is to provide a coding method capable of coding an audio signal with good efficiency even for an input sound containing a large number of tone components in the conversion coding. It is an object of the present invention to provide a decoding method, an apparatus using the method, and a recording medium in which the method is recorded as a computer-executable program.

[0011]

According to the present invention, an audio signal coding method for coding an input audio signal sample includes the following steps: (a) inputting a fixed number of input audio signal samples Time / frequency conversion is performed for each sample to obtain frequency domain coefficients,
(b) The frequency domain coefficient is divided into one or more bundled coefficient pieces to generate a coefficient piece group, (c) the strength of each coefficient piece of the coefficient piece group is calculated, and (d) A coefficient fragment sequence of at least two systems is generated by classifying the coefficient fragment group into one of at least two groups based on their intensities, and the classification information is encoded and output as a classification information code. e) Coding the coefficient strip sequence of at least two systems and outputting it as a coefficient code.

According to the present invention, a decoding method for inputting a digital code and outputting audio signal samples comprises:
It includes the following steps: (a) Decode the input digital code to obtain a group of coefficient strips of multiple systems, and (b) Decode the input digital code to obtain the classification information of the coefficient strips. A plurality of groups of coefficient small pieces are combined into a frequency domain coefficient consisting of one group of coefficient small pieces arranged in sequence, and (c) the above step (b)
The frequency domain coefficient obtained in (1) is subjected to frequency / time conversion, and the result is output as an audio signal.

Alternatively, the decoding method comprises the steps of: (a) decoding the input digital code to obtain a group of coefficient pieces each coefficient piece consisting of a plurality of coefficients in the frequency domain; Decode the code, obtain the classification information of the coefficient small pieces, classify the coefficient small piece group into the coefficient small piece string of multiple systems based on the above classification information, (c) decode the input digital code,
While obtaining the normalization information of the coefficient small piece group, each of the coefficient small piece sequences classified into a plurality of systems in the step (b) is denormalized by the normalization information, and (d) obtained in the step (b). Based on the classification information of the coefficient pieces, the coefficient piece strings of the multiple systems denormalized in the step (c) are sequentially arranged in an array of one system before the classification, and are restored to the frequency domain coefficient. The frequency domain coefficient obtained in step (d) is frequency / time transformed, and the result is output as an audio signal.

According to the present invention, an encoding device for inputting audio signal samples and outputting a digital code comprises:
A time / frequency conversion unit that obtains a frequency domain coefficient by time / frequency converting the input audio signal for each fixed number of inputs;
A coefficient small piece generation unit that divides the frequency domain coefficient from the time / frequency conversion unit into a small piece group in which adjacent coefficients are bundled; and an intensity calculation unit that calculates the strength of each small piece group from the coefficient small piece generation unit. , The coefficient small pieces are divided into at least two groups based on the relative magnitude of the strength of the small piece group calculated by the strength calculation unit, and the small piece group generated by the coefficient small piece generation unit based on the information of this grouping. Are classified into at least two systems, the classification information is coded and output as a digital code, and each of the coefficients classified into at least two systems by the coefficient classification class is coded, and the result is digitalized. And a quantizer for outputting as a code.

Alternatively, the encoding device may include a time / frequency transforming unit for time / frequency transforming an input audio signal for each fixed number of inputs to obtain a frequency domain coefficient, and a frequency domain coefficient from the time / frequency transforming unit. A coefficient small piece generation unit that divides adjacent coefficients into a small piece group, a strength calculation unit that calculates the strength of each small piece group that is generated by the coefficient small piece generation unit, and a small piece group that is calculated by the strength calculation unit. The coefficient pieces are divided into at least two groups on the basis of the relative magnitude of the intensity, and the piece group generated by the coefficient piece generation unit is classified into at least two systems based on this grouping information. The coefficient small piece classification unit that encodes information and outputs it as a digital code and the strength of each of the coefficient small piece groups that are classified into at least two systems by the coefficient small piece classification unit are normalized, and the normalized information is encoded. Conclusion Is output as a digital code, and a group of at least two normalized coefficient pieces from the flattening section is used to classify the coefficient pieces using the grouping information. A coefficient synthesizing unit for re-synthesizing into a small piece group, and a quantizing unit for quantizing the coefficient small piece group re-synthesized by the coefficient synthesizing unit and outputting the result as a digital code.

According to the invention, a decoding device for inputting a digital code and outputting audio signal samples comprises:
Dequantizer that decodes the input digital code to obtain a group of coefficient small pieces of multiple systems, and decodes the input digital code to obtain the classification information of coefficient small pieces, and based on this information, the group of coefficient small pieces of multiple systems. A coefficient synthesizing section for synthesizing and sequentially restoring the frequency domain coefficients of one system, and a frequency / time converting section for frequency / time converting the frequency domain coefficients restored by the coefficient synthesizing section and outputting the result as an audio signal; Including and

Alternatively, the decoding apparatus decodes the input digital code and obtains the coefficient small piece group, and the input digital code to obtain the classification information of the coefficient small piece, and based on this information. A coefficient small piece classifying unit that classifies the coefficient small piece group into a plurality of systems, and decodes the input digital code to obtain the normalization information of the coefficient small piece group, and the coefficient small piece classification unit uses this information to classify the coefficient small piece group into a plurality of systems. Based on the classification information of the coefficient small pieces obtained by the inverse flattening unit and the coefficient small piece classification unit for denormalizing each of the It includes a coefficient synthesizing unit for sequentially arranging in one system and restoring the frequency domain coefficient, and a frequency / time converting unit for frequency / time converting the frequency domain coefficient from the coefficient synthesizing unit and outputting the result as an audio signal.

[0018]

BEST MODE FOR CARRYING OUT THE INVENTION According to the present invention, after converting an input signal into coefficients in the frequency domain, a series of coefficients are bundled into bands of about 100 Hz width to form coefficient pieces, and each coefficient piece is adjusted according to its strength. Are classified into at least two groups, for example, a high level group and a low level group. For example, as shown in FIG. 3A, a modified discrete cosine transform (MDC)
T) When the magnitude of a coefficient in the frequency domain, such as a coefficient, is fluctuating, adjacent ones of the frequency domain coefficients (FIG. 3B) are bundled into a coefficient small piece as shown in FIG. 3C. Are classified into a group G ₀ having a small coefficient size and a group G ₁ having a large coefficient size, as shown in FIG. 3D. Independent processing is performed on the high-strength group G ₁ and the low-strength group G ₀ . The independent processing after grouping is not only the method of performing the quantization processing separately but also the method of performing the flattening independently and then performing the vector quantization by combining the two groups G ₀ and G ₁ into one. There is.

Since the shape of the coefficient pieces belonging to each group after classification is often caused by the same sound source, the variation of the intensity within the group is small. For this reason,
In the independent processing after grouping, for example, if processing is performed for each small band at equal intervals on the Bark scale, it is possible to perform quantization with high efficiency while keeping allocation of the information amount that is acoustically optimum. The grouping is not limited to two, but three
You may divide into two or more.

As described above, according to the present invention, the coefficient pieces are classified into a plurality of groups, each group is flattened and then coded, and the classification information is coded. This classification information is based on the Japanese patent application publication 7-1685 mentioned above.
Since it is easier to compress than the position information required by 93,
The amount of information can be suppressed to a small amount, and encoding can be efficiently performed. First Embodiment FIG. 4 shows a first embodiment of the present invention.

The processing units 11 to 18 in FIG.
Then, the audio signal x, which is a discrete sample sequence, is input, and the encoded bit sequence C is output. The processing units 31 to 36 in FIG. 4 are the decoding unit 30, which inputs the encoded bit string C and outputs the audio signal x which is a discrete sample string. Time / frequency conversion unit 11 The input audio signal x is input to the time / frequency conversion unit 11 as a sequence of discrete samples, and time / frequency conversion is performed for each fixed number N of input samples, and converted into N frequency domain coefficients. To do. The time / frequency conversion method includes discrete cosine transform (DCT) and modified discrete cosine transform (MD).
CT) can be used. When the modified discrete cosine transform is used as the transform method, N of each N inputs and the N 2 × N input audio samples immediately before that are transformed to obtain N frequency domain coefficients. obtain. A window function such as a Hamming window or a Hanning window may be applied immediately before performing the time / frequency conversion processing. In particular, when the modified discrete cosine transform is selected as the transform method, it is desirable to apply the window represented by the following formula W to the input sample x.

W (i) = 0.5 {1-cos [2π (0.5 + i) / N]}, i = 0,1, ..., N-1 (1) The above processing is performed by the modified discrete cosine transform. Taking the case as an example, it is expressed by the following equation.

[0023]

[Equation 1] Here, i represents an input sample number, k represents a frequency number, and x represents an input sample. Coefficient small piece generation unit 12 The coefficient in the frequency domain obtained by the time / frequency conversion unit 11 is input to the coefficient small piece generation unit 12. The coefficient piece generator 12, the coefficients of the input frequency domain, Ru coefficient pieces columns and to <br/> of one system is divided into coefficient pieces grouped by the M. As a result, the engagement number E is configured as follows.

E (q, m) = X (q · M + m) (3) where q = 0,1, ..., Q-1, and m = 0,1, ..., M for each q. -1, where q is a number representing a coefficient piece, m is a number representing a coefficient in the coefficient piece, and Q is the number of coefficient pieces. The coefficient fragment size M can take any integer value of 1 or more,
The effect is high when the frequency width is set to, for example, about 100 Hz. For example-sampling frequency of the input signal is 48kH
If z, set M = 8. Further, here, the value of M is described as being constant for all coefficient small pieces, but the coefficient small pieces do not necessarily have to have the same size, and the value of M may be set individually.

The coefficient small pieces generated by the coefficient small piece generation unit 12 are input to the coefficient small piece classification determining unit 13 and the coefficient small piece classification unit 14. -Coefficient small piece classification determining unit 13 FIG. 5 shows a detailed structure of the coefficient small piece classification determining unit 13.
The coefficient piece classification determining unit 13 inputs the coefficient piece and outputs the classification information. That is, the input coefficient small pieces are input to the coefficient small piece strength calculation unit 3-1, and the strength I is calculated for each coefficient small piece by the following equation.

[0026]

[Equation 2] 1 coefficients piece row lineages band divider 3-2, a plurality of coefficient Small
It is divided into small bands for each piece . Each coefficient pieces strength in that each subband represents _{I sb (i sb, q sb} ) and. However, i _sb is
The sub-band number is shown, and q _sb shows the sub-piece number in the sub-band. The number of coefficient pieces included in one small band is an arbitrary number of 2 or more and is given by Q _sb (i _sb ) .

[0027] i_sb， Q_sbThe relation between q and q is expressed as
To be done.

[0028]

[Equation 3] The coefficient segment strengths band-divided by the band segmentation unit 3-2 are the threshold value determination unit 3-3, the segment classification determination unit 3-4, and the separability calculation unit 3-.
Passed to 5. In the threshold value determining unit 3-3, the band dividing unit 3-
The maximum value and the minimum value of the coefficient small piece intensity received from 2 are obtained for each small band, and the threshold value T for small piece classification is determined based on this value as follows.

T _sb (i _sb ) = αI _sb (i _sb , q _max ) + (1−α) I _sb (i _sb , q _min ) (7) where q _min gives the minimum value of I _sb Coefficient piece number,
q _max is the coefficient fragment number that gives the maximum value of I _sb , a constant α
Is 1 ≧ α> 0. The value of α is set to about 0.4. The threshold value T _sb determined here is passed to the small piece classification determination unit 3-4. The small piece classification determination unit 3-4 compares the coefficient strength I _sb received from the band division unit 3-2 with the threshold value T _sb received from the threshold value determination unit 3-3 to determine the classification of coefficient small pieces. .
The classification information G of the coefficient small pieces is determined by the following equation for each of q = 0, 1, ..., Q-1.

G (q) = 0 for I _sb (i _sb , q _sb ) ≦ T _sb (i _sb ) = 1 for I _sb (i _sb , q _sb )> T _sb (i _sb ) (8) where The determined classification information G (q) of the coefficient pieces is passed to the separation degree calculation unit 3-5 and the classification information output unit 3-7. The separability calculation unit 3-5 also includes the coefficient fragment strength I _sb received from the band division unit 3-2 and the coefficient fragment classification information G (q) received from the fragment classification determination unit 3-4. And I _sb for each small band G (q) = 0 and G (q)
It is divided into two groups of = 1 and the degree of separation is calculated from the strength of each group. Before calculating the degree of separation, the intensities of the two classified groups are obtained. The intensity I _G0 of the group with G (q) = 0 is _calculated by the following equation.

[0031]

[Equation 4] The intensity I _G1 of the group with G (q) = 1 is _calculated by the following equation.

[0032]

[Equation 5] The degree of separation D _sb is determined from I _G0 and I _G1 by the following equation. D _sb (i _sb ) = I _G1 (i _sb ) / I _G0 (i _sb ) (11) The separation degree D _sb (i _sb ) for each subband i _sb determined here is
It is passed to the small piece classification use / non-use decision unit 3-6.

The small piece classification use / non-use determination unit 3-6 determines the use / non-use of the small piece classification for each small band based on the separation degree determined by the separation degree calculation unit 3-5. When the degree of separation D _sb exceeds the threshold value Dt, the small piece classification use flag F _sb (i _sb ) is set to 1. Otherwise, small piece classification use flag F
Set _sb (i _sb ) to zero. The small piece classification use flag F determined here is passed to the classification information output unit 3-7. In the classification information output unit 3-7, the coefficient fragment classification information G (q) received from the fragment classification determination unit 3-4 is sent to the fragment classification use flag F _sb received from the fragment classification use / non-use determination unit 3-6. _Redetermine for each small band based on (i _sb ). If the value of F _sb (i _sb ) is 0, all the values of the classification information G (q) of the coefficient fragment belonging to the i _sb- th small band are set to 0. When the value of F _sb (i _sb ) is 1, the classification information G (q) of the coefficient fragment belonging to the i _sb- th small band is left unchanged. Note that G using F _sb
Re-determination of (q) is not always necessary, but if this re-determination by F _sb is performed, it is possible to set G (q) of the coefficient fragment with small fluctuation of the coefficient size to 0, The classification information G (q) can be encoded with higher efficiency.

The coefficient piece classification information G (q) re-determined by the classification information output section 3-7 is used as the output of the coefficient piece classification determining section 13 in FIG. 4, and this information is used as the coefficient piece classifying section 14 and the coefficient piece class. It is passed to the classification information compression unit 15. Coefficient small piece classifying unit 14 The coefficient small piece classifying unit 14 receives the coefficient small piece generated by the coefficient small piece generating unit 12 and the coefficient small piece classification information G (q) determined by the coefficient small piece classification determining unit 13 as input, and receives all coefficient small pieces. For the coefficient small pieces, the group E _g0 of G (q) = 0 and the group E _{g1 of} G (q) = 1
Classify into.

The coefficient fragment classifying unit 14 is the group E._g0And E_g1
Size S₀, S₁Has a memory (not shown) for storing
Be present. In addition, the role of the counter of the small piece number
It also has a memory (not shown). Figure 6
3 is a flow chart of processing of the coefficient segment classifying unit 14. Coefficient piece
In the classification unit 14, first, S of the memory₀, S₁, Q to 0
To rear. Next, the small piece number q in the memory is set to the coefficient small piece E (q, m).
If q is smaller than the number Q of
Go to step S3, otherwise E_g0(S₀, m) to glue
E _g0 And then E_g1(S₁, m) to group E_g1 As each
Size S₀, S₁And output with (step S
2).

In step S3, the classification information G of coefficient small pieces
It is determined whether or not the value of (q) is 1, and if it is true, the process proceeds to step S6, and if false, the process proceeds to step S4. In step S4, the small piece E (q, m) indicated by the memory counter q is added to the small piece group E _g0 as in the following equation. E _g0 (S ₀ , m) = E (q, m), m = 0, 1, ..., M-1 In step S5, the group size S _{0 of the} memory is incremented by 1 and the process proceeds to step S8. Proceed to.

In step S6, the small piece E (q, m) indicated by the memory counter q is added to the small piece group E _g1 as in the following equation. E _g1 (S ₁ , m) = E (q, m), m = 0,1, ..., M-1 In step S7, the group size S _{1 of the} memory is incremented by 1 and the process is performed in step S8. Proceed to. Step S8
Then, the memory counter of the small piece number q is incremented by 1, and the process proceeds to step S2.

In this way, the small piece groups E _g0 and E _g1 classified by the coefficient small piece classifying unit 14 and their sizes S
₀ and S ₁ are the first and second quantizers 16 and 17, respectively.
Passed to. -Coefficient small piece classification information compressing unit 15 In the coefficient small piece classification information compressing unit 15, a series of coefficient small piece classification information G (q), q = 0, determined by the coefficient small piece classification determining unit 13.
Compressed 1, ..., Q-1 and compressed coefficient piece classification information G (q) ^*
To the multiplexing unit 18.

Since the appearance probabilities of the coefficient segment classification information G (q) values 0 and 1 are usually biased, any lossless compression coding using this can be used, but Huffman coding or arithmetic coding is possible. It is particularly efficient to use entropy coding such as. In addition, the effect can be obtained by using the run length coding. Further, as shown in FIG. 7, coefficient piece classification information G (q), q = 0, 1, ..., Q-1 is divided into several blocks, and coefficient piece classification information having a value of 1 in each block. If there is no G (q), set the F _G flag represented by 1 bit to 0,
The block is represented only by the flag F _G , and if there is the coefficient piece classification information G (q) with a value of 1, the flag F _G is set to 1 and then the flag F _G = 0 is added to the beginning of the block. Then
The number of bits may be reduced as a whole by representing the coefficient piece classification information G (q) in each block by 1 bit. Further, the coefficient fragment classification information with the reduced number of bits may be applied to the above-mentioned Huffman coding or arithmetic coding, for example. First Quantization Unit 16 The first quantization unit 16 encodes the coefficients forming the small piece group E _g0 classified by the coefficient small piece classification unit 14.

The small piece group E _g0 is converted into a single coefficient sequence as shown in the following equation before being encoded. C ₀ (s · M + m) = E _g0 (s, m), s = 0, 1, ..., S ₀ , m = 0,1, ..., M-1 A method of dividing the coefficients forming the coefficient sequence C ₀ into some small blocks and adaptively assigning the number of quantization bits to each small block to perform scalar quantization may be used (method A), or the coefficient sequence C may be used. _The coefficient that constitutes ₀ is divided into several small blocks, the optimum quantization width is determined for each small block, scalar quantization is performed, and then entropy coding such as Huffman coding or arithmetic coding is used. Alternatively, the coefficient sequence C ₀ may be collectively vector-quantized (method C), or the coefficient sequence C ₀ may be interleaved vector-quantized (method D).

The information quantized here is the method A or C.
Alternatively, when D is used, the quantized index In _E0 is converted into a bit string by converting it into a binary string with the necessary minimum number of bits and then passed to the multiplexing unit 18. When the method B is used, the bit string is directly passed to the multiplexing unit 18. Further, the size S ₀ of the small piece group E _g0 input from the coefficient small piece classifying unit 14 is also converted into a binary string by a predetermined number of bits and then passed to the multiplexing unit 18. Second Quantization Unit 17 The second quantization unit 17 encodes the coefficients forming the small piece group E _g1 classified by the coefficient small piece classification unit 14. The coding is performed in the same procedure as the first quantizing unit, but the coding method does not necessarily have to match the first quantizing unit.

The small piece group E _g1 is converted into a single coefficient sequence as shown in the following equation before being encoded. C ₁ (s ・ M + m) ＝ E _g1 (s, m), s ＝ 0, 1, ..., S ₁ , m ＝ 0,
As the 1, ..., M-1 encoding method, the coefficient forming the coefficient sequence C ₁ is divided into several small blocks, and the number of quantization bits is adaptively assigned to each small block for scalar quantization. The method may be used (method A), the coefficient constituting the coefficient sequence C ₁ is divided into several small blocks, the optimum quantization width is determined for each small block, scalar quantization is performed, and then Huffman coding is performed. Alternatively, a method of performing entropy coding such as arithmetic coding may be used (method B), or the coefficient sequence C ₁ may be vector-quantized collectively (method C), or the coefficient sequence C ₁ may be interleaved vector quantum. It may be converted (method D).

The information encoded here is the method A or C.
Alternatively, when D is used, the quantization index In _E1 is converted into a binary string with the minimum necessary number of bits, converted into a bit string, and then passed to the multiplexing unit 8. When the method B is used, the bit string is directly passed to the multiplexing unit 18. Further, the size S ₁ of the small piece group E _g1 input from the coefficient small piece classifying unit 4 is converted into a binary string with a predetermined number of bits, and is then passed to the multiplexing unit 18.

Whichever method is used, the coding method used by the second quantizer 17 does not have to be the same as the coding method used by the first quantizer 16. Rather, by paying attention to the difference in the characteristics of the coefficient segment groups E _g0 and E _g1 given to the first and second quantizers 16 and 17, and adopting different encoding methods suitable for each, the amount of code is reduced. Alternatively, distortion due to code error can be reduced. Multiplexing Unit 18 In the multiplexing unit 18, the coefficient fragment classification information compressing unit 15, the first
Input information G (q) from the quantizer 16 and the second quantizer 17
^* , In _E0 , In _E1 are all output as an output bit string.
The bit string output from the multiplexing unit 18 becomes the output of the encoding unit 10 and is passed to the demultiplexing unit 31 of the decoding unit 30.

The decoding unit 30 will be described below. Demultiplexing unit 31 In the demultiplexing unit 31, the bit string output from the encoding unit 10 is input, and the bit string In _E0 and the 2 inverse quantizer 33
To the bit string In _E1 and the bit string G (q) ^* to the coefficient segment classification information restoring unit 34, and each of them is divided into the first dequantization unit 32,
The second dequantization unit 33 and the coefficient segment classification information restoration unit 34
Send to. First Dequantization Unit 32 The first dequantization unit 32 reconstructs the bit string output from the demultiplexing unit 31 and outputs the coefficient segment group E _g0 and its size S ₀ . The size S ₀ is restored by converting a bit string representing a size that has been converted into a binary number with a predetermined number of bits into an integer.

The bit string representing the small piece group E _g0 is the first
After the coefficient sequence C ₀ ^q is restored by the reverse processing of the quantization method A, B, C, or D used in the quantization unit 16, the small piece group E _g0 ^q is restored as in the following equation. E _g0 ^q (s, m) = C ₀ ^q (s · M + m), s = 0,1, ..., S ₁ -1, m = 0,
1, ..., M-1 Note that the superscript ^q such as C ₀ ^q and E _g0 q is the encoding unit 10.
Since a quantization error occurs due to being subjected to the quantization processing in the first quantization unit 16 of, the C ₀ ^q and E _g0 ^q restored in the inverse quantization unit 32 are quantized with _respect to the original C ₀ and E _g0. It indicates that the conversion error is included. The same applies to the superscript q of other symbols. Second Dequantization Unit 33 The second dequantization unit 33 restores the bit string output from the demultiplexing unit 31 and outputs the coefficient segment group E _g1 and its size S ₁ . The size S ₁ is restored by converting a bit string representing a size that has been converted into a binary number with a predetermined number of bits into an integer.

The bit string representing the small piece group E _g1 is the second bit string.
After the coefficient sequence C ₁ ^q is restored by the inverse processing of the quantization method A, B, C, or D used in the quantizer 17, the small piece group E _g1 ^q is restored as in the following equation. E _g1 ^q (s, m) ＝ C ₁ ^q (s ・ M ＋ m), s ＝ 0,1, ..., S ₁ -1, m ＝ 0,
1, ..., M-1 ・ Coefficient small piece classification information restoring unit 34 In the coefficient small piece classification information restoring unit 34, the bit string output from the demultiplexing unit 31 is reversibly used in the coefficient small piece classification information compressing unit 15. Restoration is performed by the reverse processing of the compression encoding method to restore the coefficient segment classification information G (q), q = 0, 1, ..., Q-1. Of course, if different encoding methods are used in the first and second quantizers 16 and 17 in the encoder 10, the decoding used in the first and second inverse quantizers 32 and 33 of the corresponding decoding 30. The methods are also different. -Coefficient synthesizing unit 35 In the coefficient synthesizing unit 35, the small piece group output from the first inverse quantizing unit 32 and the second inverse quantizing unit 33 is converted into the coefficient small piece classification information G output from the coefficient small piece classification information restoring unit 34. Based on (q), it is reconstructed into one system and the coefficient in the frequency domain is output.

FIG. 8 shows a processing procedure in the processing of the coefficient synthesizing unit 35 until the coefficient small piece group E ^q is obtained. Step S
In step 1, the initial values of S ₀ , S ₁ , and q are set to 0, and in step S2, it is determined whether q <Q. If YES, the coefficient piece classification information G (q) is 1 in step S3. Judge whether N
If it is O, the coefficient piece E _g0 ^q (S ₀ , m) is changed to E ^q (q,
m), S ₀ is incremented by 1 in step S 5, q is incremented by 1 in step S 8, and the process returns to step S 2. If the determination in step S3 is YES, then in step S6, E _g1 ^q (S ₁ ,
m) is defined as E ^q (q, m), S ₁ is incremented by ₁ in step S7, q is incremented by 1 in step S8, and the process returns to step S2. If it is determined in step S2 that q is not smaller than Q, the process is terminated and the coefficient segment group E ^q (q, m), q = 0, 1, ..., Q-
1, m = 0,1, ..., M-1 is obtained.

The coefficient small piece group E ^q is reconstructed into the coefficient X ^q in the frequency domain as in the following equation in the reverse procedure of the coefficient small piece generating unit 12. X ^q (q · M ＋ m) ＝ E ^q (q, m), q ＝ 0, 1, ..., Q-1; m ＝ 0, 1,
..., M-1 ・ Frequency / time conversion unit 36 In the frequency / time conversion unit 36, a series of coefficients X ^q (q · M + m) in the frequency domain output from the coefficient synthesis unit 35 are frequency / time converted to perform audio Generate and output signal x ^q .

The frequency / time conversion method includes inverse discrete cosine transform (IDCT) and inverse modified discrete cosine transform (IM
DCT) can be used. When the inverse modified discrete cosine transform is used as the transform method, N input coefficients are transformed to obtain 2N time domain samples. After multiplying this sample by the window function W represented by the following equation, the first half N samples of the current frame and the second half N samples of the immediately preceding frame are added together to obtain N samples.

W (i) = 0.5 [1−cos [2π (0.5 + i) / N]}, i =
0,1, ..., N-1 The above processing can be expressed by the following equation using the inverse modified discrete cosine transform as an example.

[0052]

[Equation 6] ^xq (i) = ^Zt-1 (i + N) + Z (i), i = 0,1, ..., N-1 where ^xq (i) is the output audio sample signal. Second Embodiment FIG. 9 shows a second embodiment of the present invention. Processing unit 1 in FIG.
1, 12, 13, 14, 15, 19, and 20 are encoding units 1
0 is input, an audio signal which is a discrete sample sequence is input, and an encoded bit sequence is output. Processing unit 31,
Decoding units 30 and 34 to 36 to 40 input the encoded bit string and output an audio signal which is a discrete sample string.

The same numbers are assigned to the processing units corresponding to those in the first embodiment. Since the processing units 11 to 15 in the encoding unit 10 of the second embodiment perform the same processing as the corresponding ones in the first embodiment, detailed description will be omitted. Figure 1
0 is referred to in the description of frequency domain coefficient flattening in the embodiments described below. In line A, the frequency domain coefficient from the time / frequency conversion unit 11 is the coefficient fragment generation unit 12
Shows the state defined by the coefficient small piece E (q, m), and rows B and C show the coefficient small piece of G (q) = 1 and the coefficient of G (q) = 0 determined by the coefficient small piece classification determining unit 13. Show pieces separately, row D
And E are the coefficient fragment classifying unit 14
Shows that the coefficient pieces are output separately in two continuous systems, that is, two coefficient piece groups E _g0 and E _g1 are shown. The processing of the coefficient pieces indicated by these rows A to E is the first.
This is the same as the case of the embodiment.

The coefficient segment groups E _g0 , E _g1 (rows E, D) and their sizes S ₀ , S ₁ from the coefficient segment classifying unit 14 are input to the flattening / synthesizing unit 20, and the coefficient segment classifying unit 1
The coefficient fragment classification information G (q) from 3 is also flattened / synthesized by the flattening / synthesis unit 20.
Entered in. The flattening / synthesizing unit 20 sequentially determines the coefficient small pieces in each coefficient small piece group for each original small band based on the coefficient value L ₀ = L
₀₀ , L ₀₁ , L ₀₂ , L ₀₃ , L ₀₄ , L ₀₅ , L ₀₆ (row E) and L ₁ = L
It flattens by normalizing with ₁₀ , L ₁₁ , L ₁₃ and L ₁₅ (row D), and the coefficient small pieces (rows G and F) of these two flattened groups are made into coefficient small piece classification information G (q). Based on the same, it is incorporated into the original frequency position on the same frequency axis to form a series of flattened frequency domain coefficient sequences e (q, m) (row H), which are provided to the vector quantization unit 19. Further, coefficient piece flattening information L ₀ , L ₁ used for flattening is encoded and given to the multiplexing unit 18 as L ₀ ^* , L ₁ ^* . The representative value L ₀ and / or L ₁ of the coefficient small pieces in the same small band is determined by the small band.
There is a possibility that the system numerical values that are more distant in frequency than the bandwidth of (that is, # of one or more small bands apart) may have significantly different values, and if they are normalized together, there is little improvement in flatness. Is. Vector Quantization Unit 19 The vector quantization unit 19 vector-quantizes the frequency domain coefficients given by the flattening / coefficient synthesis unit 20 and sends the coding index In _e to the multiplexing unit 18. Weighted interleaved vector quantization is preferable as a vector quantization method. The multiplexing unit 18 is the vector quantization unit 1.
The coding index In _e from 9 is the compressed classification information G (q) ^* from the coefficient fragment classification information compression unit 15, and the flattening / synthesis unit 2
The coefficient piece flattening information L ₀ ^* and L ₁ ^{* from} _{0 are} multiplexed and transmitted to, for example, the decoding unit 30.

The decoding unit 30 in the second embodiment will be described below. Vector dequantization unit 37 In the vector dequantization unit 37, the vector quantization index In _e received from the demultiplexing unit 31 is reproduced with reference to the codebook, and the coefficient group e in the flattened frequency domain is regenerated. ^q
(q, m) is obtained and sent to the coefficient fragment generation unit 38. Coefficient small piece generating unit 38 In the coefficient small piece generating unit 38, the coefficient small piece generating unit 1 of the first embodiment is used.
2 (FIG. 4), the coefficient group e ^q (q, m) in the flattened frequency domain is flattened into small coefficient pieces e ^q (q), q = 0,
Divide into 1, ..., Q-1. -Coefficient small piece classification unit 39 In the coefficient small piece classification unit 39, the coefficient small piece classification information restoration unit 34
According to the coefficient fragment classification information G (q) = 0 or 1 from FIG.
In the same manner as the coefficient segment classifying unit 14 in, the flattened coefficient segment e ^q (q) is flattened into the coefficient segment group e _g0.
Classify into ^q (size S ₀ ) and e _g1 ^q (size S ₁ ).

The inverse flattening / synthesizing unit 40 divides the flattening coefficient small piece groups e _g0 ^q and e _g1 ^q into flattening information L _g = (L ₀ , L ₁ ), L ₀ = L ₀₀ , L ₀₁ , L ₀₂ , L ₀₃ , L ₀₄ , L ₀₅ ,
L ₀₆ ; L ₁ = L ₁₀ , L ₁₁ , L ₁₃ , L ₁₅ Inverse flattening, that is E _g0 ^q
= e _g0 ^q L ₀ , E _g1 ^q = e _g1 ^q L ₁ is calculated, and coefficient pieces are sequentially sampled from E _g0 ^q or E _g1 ^q according to the classification information G (q) = 0 or 1,
Coefficient pieces EA (q) of the entire band are obtained by sequentially arranging them on the same frequency axis. The frequency / time conversion unit 36 converts the whole band coefficient small piece EA (q) into the time domain signal X and outputs it.

FIG. 11 shows a configuration example of the flattening / synthesizing unit 20 and the inverse flattening / synthesizing unit 40 in the second embodiment described with reference to FIG.
Shown in A and 11B. The coefficient segment group E _g0 and its size S ₀ output from the coefficient segment classifying unit 14 are input to the first flattening unit 21, and the coefficient segment group E _g1 and its size S ₁ output from the coefficient segment classifying unit 14 are output. Is input to the second flattening unit 22. First Flattening Unit 21 The first flattening unit 21 flattens the coefficient small piece group E _g0 output from the coefficient small piece classifying unit 14 by using the coefficient small piece classification information G (q) as auxiliary information. Coefficient fragment group E
_{The g0} flattening process is a process of _obtaining a representative value for each of a plurality of coefficient small pieces (small bands), and normalizing the coefficients forming all the coefficient small pieces of the small band by the representative value.

When the entire processing of the encoding unit 10 and the decoding unit 30 is executed by a computer program,
Handling all the coefficient small pieces by defining the positions on the linear frequency axis includes many processing steps that can be commonly used in the encoding and decoding processes, and makes the configuration of the encoding and decoding programs easy. Here, an example is shown in which the coefficient small pieces of the obtained coefficient small piece group E _g0 are returned to the original frequency position for flattening, and then returned to the group of continuous coefficient small pieces.
However, this method requires a larger amount of calculation and a larger memory capacity for processing than a method that does not restore the frequency position described later. The same applies to the second flattening portion 22.

FIG. 12 shows a structural example of the first flattening portion 21. In the frequency band restoration unit 21-1, the coefficient small pieces E _g0 (s, m), s = 0,1, ..., S ₀ forming the input small coefficient piece group E _g0.
On the basis of the coefficient small piece classification information G (q) is expanded into a coefficient small piece group EA over the entire band (see row C in FIG. 10) and passed to the small band dividing unit 21-2. Fig. 13 shows the coefficient fragment group E _g0 (s,
The restoration procedure by the frequency band restoration unit 21-1 for m), s = 0, 1, ..., S ₀ is shown.

In step S1, the values of q and S are initialized to 0, and in step S2 it is determined whether the coefficient piece classification information G (q) from the coefficient piece classification unit 13 is 0. If 0, in step S3, the s-th coefficient piece E of the coefficient piece group E _g0.
g0 (s, m) is arranged as the q-th coefficient piece EA (q) in the entire band (q = 0,1, ..., Q-1), and q and s are each advanced by one step. If G (q) = 0 in step S3, the coefficient 0 is set as the q-th coefficient piece EA (q) in the entire band in step S4.
Insert (M). In step S6, it is determined whether q is smaller than Q, and if it is smaller, the process returns to step S2 and the same procedure S is performed.
2, S3, S4 and S5 are repeated. If q is not smaller than Q in step S6, the restoration of the coefficient segment group E _{g0 to} the entire band is completed.

The small band division unit 21-2 divides the coefficient small piece group EA expanded in all bands into small bands. The divided bandwidth of the small band may be constant over the entire band, or may be widened as the frequency increases. The band-divided coefficient small piece group is provided to the small band representative value calculation unit 21-3 and the normalization unit 21-5. The small band representative value calculation unit 21-3 calculates a representative value for each of the divided small bands. The representative value may be the maximum absolute value of the coefficient within the sub-band, or may be the square root of the average of the powers of the coefficients within the sub-band that are greater than zero. The calculated representative value is given to the small band representative value encoding unit 21-4.

The small band representative value encoding unit 21-4 encodes the small band representative value. First, the representative value of the small band is scalar-quantized to obtain the quantization index L ₀ ^* . Those with a quantization index of 0 are not encoded, and only those with a quantization index of greater than 0 are output to the multiplexing unit 18 as coefficient flattening information. As another method, the representative value group may be subjected to interleave vector quantization. Further, the quantized representative value L ₀ of the small band is passed to the normalization unit 21-5.

The normalizing unit 21-5 quantizes the coefficient small piece group E _g0 divided into the small bands given from the small band dividing unit 21-2 by the small band representative value encoding unit 21-4. Normalize with the representative value of the small band. The normalized or flattened coefficient small piece group e _g0 is passed to the coefficient small piece group restoring unit 21-6. In the coefficient small piece group restoring unit 21-6, the entire band coefficient small piece group normalized by the reverse processing of the frequency band restoring unit 21-1,
It is restored to the flattened coefficient small piece group and is used as the output of the first flattening unit 21. Second flattening unit 22 The second flattening unit 22 is configured in the same manner as the first flattening unit 21, and the coefficient small piece group E _g1 given from the coefficient small piece classifying unit 14 is converted into coefficients by the same processing. Piece classification information G
Flatten using (q) as auxiliary information. The procedure is first
Same as the method in the flattening unit 21, but the frequency band restoring unit
21-1 and the part corresponding to the coefficient fragment group restoration unit 21-6,
The processing when the value of the coefficient fragment classification information G (q) is 0 and when it is 1 is exchanged. It should be noted that although there is a small band in which the coefficient small piece group E _g1 does not exist, flattening by the second flattening unit 22 is not performed in such a small band. This applies to all the processes by the second flattening unit 22 described below. -Coefficient synthesizing unit 23 In the coefficient synthesizing unit 23, the coefficient small piece groups flattened by the first flattening unit 21 and the second flattening unit 22 are used as the first embodiment.
The coefficients are combined in the same manner as in the coefficient combining unit 35 to obtain the flattened frequency domain coefficients.

The inverse flattening / synthesizing unit 40 shown in FIG.
Coefficient small piece group e received from the piece classification unit 39_g0 ^qAnd e_g1 ^qTo
Decoded coefficient piece flattening information L₀, L₁Inverse flat with
The coefficients of these two groups that have been inversely flattened
Piece E_g0 ^q, E_g1 ^qAccording to the coefficient fragment classification information G (q)
Frequency domain coefficient of E^qCombine to (q, m) and output. -First inverse flattening portion 41 FIG. 12 shows the first inverse flattening portion 41 in FIG. 11B.
A configuration corresponding to the first flattening portion 21 is shown in FIG. Reverse
Flattening information L input from the multiplexing unit 31₀ ^*, L ₁ ^*use
And flattening coefficient small piece group e_g0 ^qReverse flatten. Immediately
Then, as shown in FIG. 14, the frequency band restoration unit 41-1 inputs
Flattening coefficient small group e_g0 ^qThe flattening staff
A few small pieces e_g0 ^q(s), s = 0,1, ..., S₀Is the coefficient fragment classification information G
Based on (q), expand to a group of coefficient small pieces EA (q) over the entire band
To be done. This coefficient fragment group EA (q) is given to the small band division unit 41-2.
available.

The small band division unit 41-2 divides the coefficient small piece group EA (q) expanded into the entire band into small bands. The bandwidth of the small band may be constant over the entire band, or may be wider as the frequency is higher. The band-divided coefficient segment group is provided to the denormalization unit 41-5. Small band representative value decoding unit 41
-4, the input coefficient small piece flattening information L ₀ ^* is decoded by the decoding method corresponding to the coding method in the small band representative value coding unit 21-4 (FIG. 12), and the small band A representative value L ₀ is obtained.

In the denormalization unit 41-5, the flattening coefficient small piece group e _g0 ^q divided into the small bands given from the small band dividing unit 41-2.
Is inversely normalized with the small band representative value L ₀ decoded by the small band representative value decoding unit 41-4. The coefficient small piece group restoring unit 41-6 restores the denormalized coefficient small piece group to a coefficient small piece group by a process reverse to that of the frequency band restoring unit 41-1 and outputs the output E _g0 ^{q of} the first inverse flattening unit 41. And Second de-flattening unit 42 The second de-flattening unit 42 of FIG. 11B is also configured similarly to the first de-flattening unit 41 shown in FIG. 14, and the flattening information L input from the demultiplexing unit 31. _The flattening coefficient small piece group e _g1 ^q is inverse flattened using the small band representative value L ₁ obtained from ₁ ^* . The procedure is the same as the method in the first inverse flattening unit 41, but in the parts corresponding to the frequency band restoration unit 41-1 and the coefficient small piece group restoration unit 41-6, the value of the coefficient small piece classification information G (q) is The processing in the case of 0 and the processing in the case of 1 are exchanged. It should be noted that although there are some small bands in which the flattening coefficient small piece group e _g1 ^q does not exist, the processing by the second inverse flattening unit 26 is not executed in such small bands. This applies to all the processes by the second inverse flattening unit 26 described below.

The frequency / time conversion unit 36 uses the frequency / time conversion unit of FIG.
The frequency domain coefficient X ^q = E ^q (q, m) from the inverse flattening / synthesizing section 40 is converted into the time domain signal x ^q and output in the same manner as the time converting section 36. The flattening portion 21 (or 2 in FIG. 11A)
As an example of 2), the one shown in FIG. 12 shows the case where the coefficient small pieces are restored to the entire band, then flattened by normalization, and returned to the coefficient small piece group again. An example of the configuration of the flattening unit 21 when directly normalizing without restoring the band is shown. In this example, the coefficient fragment group given together with the size S ₀ from the coefficient fragment classification unit 14
E _g0 is the coefficient fragment classification determining unit 1 in the small band dividing unit 21-2.
It divides based on the classification information G (q) from 3 (row E), and the correspondence between those small bands and the classification information G (q) is obtained. The small band representative value calculation unit 21-3 may use the maximum absolute value of coefficient values or the root mean square of coefficient values other than zero for each small band. The small band representative value is coded by the small band representative value coding unit 21-4, and the coded representative value L ₀ ^* is given to the multiplexing unit 18 as coefficient small piece flattening information and is also quantized by decoding. The small band representative value L ₀ is supplied to the normalization unit 21-5, and the small band coefficient small pieces are normalized to generate the flattening coefficient small piece group e _g0 . The second flattening portion 22 can be similarly configured.

FIG. 16 shows a configuration example of the first deflattening unit 41 in the decoding unit 30 corresponding to the configuration of FIG. In this example, the flattening coefficient small piece group e _g0 ^q from the coefficient small piece classifying unit 39 (FIG. 9) is divided into small bands associated with the coefficient small piece classification information G (q) in the small band dividing unit 41-2. , To the denormalization unit 41-5. On the other hand, the small band representative value decoding unit 41-4
Is the coding coefficient small piece flattening information L ₀ ^* from the demultiplexing unit 31 ^.
To obtain the small band representative value L _0, and give it to the denormalization unit 41-5. The denormalization unit 41-5 denormalizes the flattened coefficient small piece group e _g0 ^q by the small band representative value L ₀ corresponding to each small band to generate an inverse flattened coefficient small piece group E _g0 ^q. ,Output.

17A and 17B show another configuration example of the flattening / synthesizing unit 20 and the inverse flattening / synthesizing unit 40 in FIG. In the flattening / synthesizing unit 20 of the encoding unit 10, the first
The flattening information calculation unit 21A divides the given coefficient small piece group E _g0 (row E in FIG. 10) into small areas, and the representative values L ₀₀ , L ₀₁ , L ₀₂ ,. .. is calculated and given to the flattening information combining unit 23A as flattening information L ₀ (= L ₀₀ , L ₀₁ , L ₀₂ , ...), and the encoded L ₀ ^{* is given} to the multiplexing unit 18. give. The small region is formed by grouping input coefficient small pieces that belong to the same small band when they are expanded on the frequency axis. The small band is set in advance. The representative value may be, for example, the maximum absolute value of the coefficients in the small area, or the average absolute value of the coefficients excluding 0. Similarly, the second flattening information calculation unit 22A receives the given coefficient segment group E _g1.
(FIG. 10 line D) is divided into small areas of the same size as the first flattening information calculation unit 21A, and the representative value L ₀₁ ,
L ₁₁ , ... are calculated and given to the flattening information combiner 23A as flattening information, and the encoded L ₁ ^* is multiplexed 1
Give to eight.

The flattening information combining unit 23A is provided with the flattening information L ₀₀ , L ₀₁ , ... From the first flattening information calculating unit 21A and the flattening information L ₁₀ from the second flattening information calculating unit 22A. , L
₁₁ , ... are given, and the first flattening information calculation unit 21 is determined depending on whether the classification information G (q) is 0 or 1 for q = 0, 1 ,.
Flattening information from A, or the second flattening information calculation unit 22A
By obtaining the flattening information from (1), and arranging them in order on the same frequency axis (that is, q = 0, 1, ...) By one sequence of flattening information (I in FIG. 10). To get

On the other hand, the coefficient synthesizing unit 24A uses the small piece group E.
Given _g0 and E _g1 , the same frequency is obtained from E _g0 or E _g1 depending on whether G (q) is 0 or 1 in the same procedure as the flattening information combining unit 23A combines flattening information. All bands (that is, q = 0,1, ..., Q-
1) Obtain a series of coefficient pieces E (q, m). The series of coefficient small pieces is the same as the coefficient small piece series generated by the coefficient small piece generating unit 12 (FIG. 9), and the coefficient synthesizing unit 24A may be omitted by using this.

The flattening unit 25 responds to the coefficient small piece E from the coefficient synthesizing unit 24A (or the coefficient small piece generating unit 12) by the flattening information sequence from the flattening information system convenience unit 23A.
The flattening coefficient over the entire band (line H in FIG. 10) is obtained by dividing each value. The obtained flattening coefficient is given to the vector quantizer 19 in FIG. The inverse flattening / synthesizing unit 40 in the decoding unit 30 performs the reverse process of the flattening unit 20 (FIG. 17A) in the encoding unit 10, as shown in FIG. 17B. That is, the first and second flattening information reproducing units 41A,
42A is the flattening information L ₀ ^* , L ₁ ^* from the demultiplexing unit 31A ^.
And the representative values L ₀ and L ₁ of the small area are decoded and the flattening information combining unit 4
Give to 3A. The flattening information combining unit 43A sets the flattening information L ₀ and L ₁ to 1 over the entire band according to the coefficient segment classification information G (q).
It is connected to the system and given to the inverse flattening unit 45. Coefficient synthesizer 4
4A is given the flattening coefficient small piece groups e _g0 ^q , e _g1 ^q from the coefficient small piece classification unit 39 (FIG. 9), and the coefficient small piece classification information G
Based on (q), combine e _g0 ^q and e _g1 ^q over the entire band, and
The system flattening coefficient small piece e ^q (q, m) is set, and the inverse flattening unit 45 reverses the given 1-system full-band flattening coefficient e ^q (q, m) according to 1-system full-band flattening information. Flatten and frequency domain coefficients
E ^q (q, m) is generated and given to the frequency / time conversion unit 36 (FIG. 9). Third Embodiment FIG. 18 shows a third embodiment of the present invention. In the third embodiment, in the configuration of the second embodiment shown in FIG. 9, a flattening unit 29 is provided in the encoding unit 10 between the time / frequency conversion unit 11 and the coefficient small piece generation unit 12, and a decoding unit is provided. Inverse flattening section 4 at 30
9 is a provided et the configuration between the inverse flattening and combining unit 40 and a frequency / time conversion unit 36. Flattening unit 29 In the flattening unit 29, the coefficient group in the frequency domain given from the time / frequency conversion unit 11 is flattened, and the coefficient fragment generation unit 12
Send to. As a flattening method, for example, normalization using a linear prediction spectrum is desirable. In that case, the linear prediction coefficient LP used for generating the linear prediction spectrum is encoded and sent to the multiplexing unit 18 as auxiliary information LP ^* . Other processing is the same as that of the embodiment of FIG. Inverse flattening unit 49 In the inverse flattening unit 49, a linear prediction spectrum is generated from the linear prediction coefficient LP obtained by decoding the linear prediction coefficient information LP ^* obtained from the demultiplexing unit 31, and thereby the inverse The frequency domain coefficient obtained by inversely flattening the coefficient group E ^q (q, m) sent from the flattening / synthesizing unit 40 is output to the frequency / time transforming unit 36. The operation of the other parts is similar to that of the embodiment of FIG.

In the above description, when the number of samples is not required when quantizing the first and second coefficient segment groups E _g0 and E _g1 , it is possible to omit the determination of the group sizes S ₀ and S ₁ . Although the coefficient pieces are classified into two groups in the above, they may be classified into three or more groups.
The width of the coefficient small piece was set to about 100 Hz, but it can be set to an appropriate value at about 200 Hz or less, and it may be possible to narrow the width on the low frequency side. Further, it is not always necessary to carry out the division of coefficient small pieces as described above over the entire band, and it is also within the scope of the present invention to carry out only for a part of the bands.

In the third embodiment shown in FIG. 18, the first and second flattening units 21 and 22 of the flattening / synthesizing unit 20 and the first and second inverse flattening units of the inverse flattening / synthesizing unit 40. 41,
Reference numeral 42 may have the same configuration as the flattening portion and the inverse flattening portion shown in FIGS. 12 and 14, respectively, or may be the same as the flattening portion and the inverse flattening portion shown in FIGS. Further, the flattening / synthesizing unit 20 and the inverse flattening unit 40 in FIG. 18 may be replaced with the configurations shown in FIGS. 17A and 17B, respectively. Further, the configuration in which the flattening unit 29 shown in FIG. 18 is provided between the time / frequency conversion unit 11 and the coefficient segment generation unit 12 can be applied to the first embodiment shown in FIG.

FIG. 19 shows a configuration in which the encoding method and the decoding method according to the present invention are implemented by a computer. The computer 50 includes a CPU 51, a RAM 52, a ROM 53, an input / output interface 54, which are connected to each other via a bus 50. It includes a hard disk 55. ROM
A basic program for operating the computer 50 is written in 53, and a program for executing the above-described encoding method and decoding method according to the present invention is stored in the hard disk 55 in advance. For example, when encoding C
The PU 51 loads the coding program from the hard disk 55 into the RAM 52, processes the audio sample signal input from the interface 54 according to the coding program, codes the audio sample signal, and outputs the coded signal from the interface 54. At the time of decoding, the decoding program is loaded from the hard disk 55 to the RAM 52, the input code is processed according to the decoding program, and the audio sample signal is output. As the program for executing the encoding / decoding method according to the present invention, the program recorded in the external disk device 57 connected to the internal bus 58 via the drive device 56 may be used. The recording medium in which the program for executing the encoding / decoding method according to the present invention is recorded may be a magnetic recording medium, an IC memory, a compact disc, or any other type of recording medium.

[0076]

As described above, according to the present invention, the frequency domain coefficients are sequentially bundled for each of a plurality of coefficients to form a plurality of coefficient pieces, and the coefficient pieces are classified into a plurality of groups according to the strength of each coefficient piece. Since coding is performed for each classified group, coefficient pieces in the same group have good flatness, and thus coding is performed efficiently. Utilizing this invention,
It is possible to efficiently encode even a music signal in which a strong tone component such as metallic sound is mixed in the high range.

[Brief description of drawings]

FIG. 1 is a block diagram showing a general form of a transform coding method.

FIG. 2 is a diagram showing an example of the tendency of the amplitude of frequency domain coefficients.

FIG. 3 is a diagram showing a method of the present invention.

FIG. 4 is a block diagram showing a functional configuration of a first embodiment of the present invention.

FIG. 5 is a block diagram showing a detailed functional configuration of a coefficient segment classification determining unit 13 in the first, second and third embodiments.

FIG. 6 is a diagram showing a processing flow of a coefficient segment classifying unit 14 in the first, second and third embodiments of the present invention.

FIG. 7 is a schematic diagram showing an operation of a coefficient segment classification information compression unit 15 in the first, second and third embodiments of the present invention.

FIG. 8 is a diagram showing a processing flow of a coefficient synthesizing unit 35 in the first, second and third embodiments of the present invention.

FIG. 9 is a block diagram showing a functional configuration of a second embodiment of the present invention.

FIG. 10 is a diagram for explaining the flattening of frequency domain coefficients in the second and third embodiments.

11A is a block diagram showing a configuration example of a flattening / synthesizing unit 20 in FIG. 9, and FIG. 11B is a block diagram showing a configuration example of an inverse flattening / synthesizing unit 40 in FIG.

FIG. 12 is a block diagram showing a detailed functional configuration of a first flattening unit 21 in second and third embodiments of the present invention.

FIG. 13 is a diagram showing a processing flow of the frequency band restoration unit 21-1 in the flattening unit in the second and third embodiments of the present invention.

FIG. 14 is a block diagram showing a functional configuration example of a first inverse flattening unit 41 in FIG. 11B.

FIG. 15 is a block diagram showing another functional configuration example of the first flattening unit 21 in FIG. 11A.

FIG. 16 is a block diagram showing another functional configuration example of the first inverse flattening unit 41 in FIG. 11B.

17A is a block diagram showing another functional configuration example of the flattening / synthesizing unit 20 in FIG. 9, and FIG. 17B is a block diagram showing another functional configuration example of the inverse flattening / synthesizing unit 40 in FIG.

FIG. 18 is a block diagram showing a functional configuration of a third embodiment of the present invention.

FIG. 19 is a block diagram showing the configuration of a computer for implementing the encoding and decoding of the present invention by a program.

─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Kazuaki Chikira 2-3-1, Otemachi, Chiyoda-ku, Tokyo Within Nippon Telegraph and Telephone Corporation (72) Inventor Taketoshi Mori 2-chome, Otemachi, Chiyoda-ku, Tokyo No. 3 No. 1 in Nippon Telegraph and Telephone Corporation (56) Reference JP-A-7-168593 (JP, A) JP-A-7-336231 (JP, A) JP-A-8-186500 (JP, A) JP H8-834797 (JP, A) JP 10-91196 (JP, A) International publication 94/028633 (WO, A1) (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 19/00 G10L 19/02

Claims (22)

(57) [Claims] 1. An audio signal coding method for coding an input audio signal sample, comprising the steps of: (a) time / frequency converting the input audio signal sample for every fixed number of input samples. to obtain a frequency domain coefficients Te, (b) the frequency domain coefficients is divided into coefficient pieces bundled by multiple and coefficient pieces sequence of one system, further engagement of the one system
The sequence of several small pieces is divided into a plurality of small bands for each coefficient small piece , (c) the strength of each coefficient small piece is calculated, and (d) each coefficient small piece in each small band is set to their strength. Generate at least two sequences of coefficient fragments by classifying into at least two groups based on
The classification information is encoded and output as a classification information code, and (e-1) is each small band of the coefficient small piece string of at least two systems.
For each, obtain the representative value of the strength of the coefficient piece in the small band,
By the value corresponding to the representative value, all the
Together collectively normalized number pieces, encoding normalization information, and outputs the result as the normalization information code, the coefficient pieces (e-2) Factor pieces sequence of at least the above normalized two systems, recombining the coefficient pieces sequence of one strain of sequences prior to classification based on the classification information, the quantizing coefficient piece column in resynthesized one system in (e-3) the step (e-2), and outputs the result as engaging number code. 2. The audio signal encoding method according to claim 1 , wherein said step (e-1) restores at least two systems of coefficient small piece sequences separately to respective full bands, and sets a representative value for each small band. Is obtained, normalized by the representative value thereof, and output as a coefficient small piece sequence flattened. 3. The audio signal encoding method according to claim 1 or 2 , wherein the step (e-1) obtains a representative value of coefficient segment strength in each sub-band, and quantizes the representative value.
With normalizing the subband by the quantized representative value, you the quantization representative value and the normalization information. 4. An input audio signal sample is encoded.
The audio signal encoding method for encoding is as follows.
Includes steps: (a) Input audio signal samples over a fixed number of input samples
Time / frequency conversion is performed for each rule to obtain frequency domain coefficients, and (b) the above frequency domain coefficients are bundled into multiple coefficient pieces.
Divide into a series of coefficient small pieces for one system, and
The sequence of several small pieces is divided into a plurality of small bands for each coefficient small string, and (c) the strength of each coefficient small piece is calculated, and (d) each coefficient small piece in each small band is
Classify into at least two groups based on strength
To generate at least two systems of coefficient fragment sequences,
The classification information is encoded and output as a classification information code. (E-1) For each small band of the coefficient small piece string of at least two systems, a value representative of the coefficient small piece strength in the small band is obtained as flattening information. , (e-2) the flattening information of said at least two systems attached to the entire band, the flattening coefficient pieces upper row Symbol combined 1 by normalizing the coefficients pieces row of the one system by flattening information system Then, ( e-3 ) the above - mentioned one-system flattening coefficient piece sequence is encoded and output as a coefficient code. 5. The method of claim 1 or 4 audio signal encoding method as claimed, the coding of the classification information in step (d) are performed by lossless compression. 6. The audio signal encoding method according to claim 1 or 4, wherein said step (e-3) is encoded by adaptive bit allocation quantization. 7. The audio signal encoding method according to claim 1 or 4, wherein said step (e-3) is the entropy coding after scalar quantization. 8. The audio signal encoding method according to claim 1 or 4, wherein said step (e-3) is encoded by vector quantization. 9. Enter a digital code, a decoding method for outputting an audio signal samples, comprising the steps of: (a) decoding the input digital code, comprising a plurality of coefficients of frequency domain coefficients coefficient pieces rows of small pieces are arranged one system
And (b) decoding the input digital code to obtain the classification information of coefficient pieces, and dividing the coefficient piece sequence of one system into a plurality of small bands for each coefficient piece based on the classification information.
Then, the coefficient is divided into a series of coefficient pieces of multiple systems, and (c) the input digital code is decoded to correspond to the above multiple systems.
Together to obtain the normalization information sequence, each the small band of each of the classified coefficients pieces columns plurality of systems at the step (b)
Said plurality of channels of corresponding inverse normalization than <br/> the normalization information of the normalization information sequence, based on the classification information of the coefficient pieces obtained in (d) the step (b), the step (c ), The coefficient debris strings of multiple systems denormalized in) are sequentially arranged in the array of one system before classification, and are restored to frequency domain coefficients. (E) The frequency domain coefficients obtained in step (d) above are
Time conversion is performed and the result is output as an audio signal. 10. The decoding method according to claim 9 , wherein
In the step (c), the coefficient small piece strings of the plurality of classified systems are restored to the original full bands in accordance with the coefficient small piece classification information, and denormalized by the normalization information for each small band made up of a plurality of coefficient small pieces . To do. 11. The decoding method according to claim 9 or 10 , wherein in the step (d), the restored frequency domain coefficient is denormalized with a spectral outline and used as a frequency domain coefficient. 12. A decoding method according to claim 9, wherein, decoding of the upper <br/> Symbol classification information is performed by decoding the lossless compression coding. 13. A decoding method according to claim 9, wherein said step (a) decoding the adaptive bit allocation quantization code. 14. The decoding method according to claim 9 , wherein in step (a), an entropy code is decoded for at least one series of coefficient pieces to obtain scalar quantized coefficients. 15. The decoding method according to claim 9 , wherein
It said step (a) for decoding the vector quantization code. 16. A coding device for inputting audio signal samples and outputting a digital code, comprising a time / frequency conversion unit for time / frequency converting input audio signals for each fixed number of inputs to obtain frequency domain coefficients. When the time / frequency domain coefficients from the frequency converting unit, and the coefficient piece generator for dividing the coefficient pieces column 1 lines bundled adjacent coefficients together, the coefficient pieces row of the one system, each of the plurality of coefficients pieces Swath of
A small band division unit that divides the band into regions, an intensity calculation unit that calculates the strength of each of the coefficient small pieces generated by the coefficient small piece generation unit, and a coefficient small piece in each of the small bands, the relative strength of those small pieces. based on the specific large and small, divided into least two groups,
Based on this grouping information, the one-system coefficient piece sequence generated by the coefficient-piece generation unit is classified into at least two system coefficient piece sequences , and the classification information is encoded and output as a digital code. For each small band of the coefficient small piece classification unit and the coefficient small piece row of the at least two systems,
Obtain the representative value of the strength of the coefficient fragment in the small band, and calculate the representative value.
A representative value calculating unit for outputting as a digital code, and the coefficient pieces column of said at least two systems, each the subband
Are normalized by the corresponding ones of the above representative values.
Generate a flattened and flattened sequence of at least two coefficient fragments
The flattening section and the flattened coefficient strip of at least two systems
Based on the classification information, the coefficient of one system of the array before classification is small.
A coefficient combining part for recombining into pieces column, the coefficients pieces sequence of one system that recombines with coefficient combination unit as well as quantized audio characterized in that it comprises a quantization unit for outputting the result as a digital code Signal encoder. 17. A input audio signal time every fixed number of input / frequency domain coefficients to obtain a time / frequency conversion unit, the frequency domain coefficients, 1 system a bundle of adjacent coefficients together
A coefficient pieces generator for dividing the coefficient pieces columns of integration, the coefficient pieces row of the one system, the zonules of each of a plurality of coefficients pieces
A small band division unit that divides into small areas, an intensity calculation unit that calculates each intensity of the coefficient small pieces, and coefficient small pieces in each of the above small bands , based on the relative magnitude of those intensities. Divide into at least two groups,
The coefficient pieces row of the one system of information of the grouping based on at least classify the coefficient pieces row of two systems, the coefficient piece classification unit for outputting a digital code classification information by encoding of the at least two systems For each small band of the coefficient small piece string,
Flatten the representative value representing the strength of the coefficient fragment in the small band
It is obtained as information, and the flattening information is used as a digital code.
And flattening information calculating unit that outputs Te, the flat corresponding coefficient pieces column of said at least two systems
Based on the above classification information,
As a report, by using the flattening information combining unit that is connected to all bands and the flattening information that is set as the single line,
Generates a series of flattening coefficient pieces by flattening a small piece sequence
An audio signal coding apparatus comprising: a flattening section for quantizing, and a quantizing section for quantizing the one-system flattening coefficient piece sequence and outputting the result as a digital code. 18. The encoding apparatus according to claim 16 or 17, wherein, by normalizing the spectral envelope over the entire band of an input audio signal to frequency domain coefficients from said time / frequency conversion unit, a frequency domain coefficients flattened
And a second flattening unit for encoding the spectrum outline information and outputting the result as a digital code. 19. A decoding device for inputting a digital code and outputting audio signal samples, wherein the decoding device decodes the input digital code and outputs a plurality of coefficients in a frequency domain.
The inverse quantization unit coefficients pieces to obtain a coefficient pieces row of one line arranged consisting decodes the input digital code, the obtained classification information of the coefficient pieces, coefficient pieces of the one system based on this information Columns
Dividing into multiple bands for each coefficient fragment and multiple systems
A coefficient piece classification unit for classifying the coefficients pieces column of the input digital codes to decode, with obtaining <br/> normalization information sequence corresponding with said plurality of channels, the plurality of normalization information sequence
The coefficient pieces columns of the plurality of channels by the normalization information, the
An inverse flattening unit that performs denormalization for each small band and the coefficient small piece sequence obtained by the coefficient small piece classifying unit are used to classify the coefficient small piece sequences of multiple systems that have been denormalized by the inverse flattening unit. A frequency synthesizer that sequentially arranges the signals in the previous one system and restores frequency domain coefficients, and frequency / time transforms the frequency domain coefficients from the coefficient synthesizer, and outputs the result as an audio signal.
An audio signal decoding device comprising: a time conversion unit. 20. A decoding apparatus according to claim 19, decodes the input digital code, along with obtaining a spectral envelope over the entire band, in the spectral envelope, an input to the <br/> frequency / time conversion unit O, characterized in that it comprises a second inverse flattening unit for inversely normalizing the frequency-domain coefficients to be
Dio signal decoding device . 21. The ode according to claim 1.
Computer to execute each step of the audio signal encoding method
Computer read recording program for
Recordable media. 22. The auto according to any one of claims 9 to 15.
Implement each step of the audio signal decoding method on a computer.
Computer reading recording program for running
Removable recording medium.