JP6220701B2 - Sample sequence generation method, encoding method, decoding method, apparatus and program thereof - Google Patents

Description

  In the signal processing technology such as sound signal encoding technology, the present invention generates a sequence in which the frequency domain sample sequence derived from the sound signal is expanded and contracted in the frequency domain of the sample points in the frequency domain sample sequence. Related to technology.

  As an encoding method for sound signals of low bits (for example, about 10 kbit / s to 20 kbit / s), adaptive encoding for orthogonal transform coefficients in the frequency domain such as DFT (Discrete Fourier Transform) and MDCT (Modified Discrete Cosine Transform) is available. Are known. For example, MEPG USAC (Unified Speech and Audio Coding), a standard technology, has a TCX (transform coded excitation) coding mode, in which MDCT coefficients are normalized for each frame and variable after quantization. Long encoding is performed (for example, see Non-Patent Document 1).

  A configuration example of a conventional TCX-based encoding device is shown in FIG. Hereinafter, each part of FIG. 1 will be described.

<Frequency domain converter 11>
A time domain sound signal is input to the frequency domain converter 11. The sound signal is, for example, an audio signal or an acoustic signal.

  The frequency domain transform unit 11 converts an input time domain sound signal into N frequency MDCT coefficient sequences X (0), X (1),..., X (N− Convert to 1). N is a positive integer.

  The converted MDCT coefficient sequences X (0), X (1),..., X (N−1) are output to the envelope normalization unit 14.

<Linear prediction analysis unit 12>
The linear prediction analysis unit 12 receives a sound signal in the time domain.

The linear prediction analysis unit 12 generates linear prediction coefficients α ₁ , α ₂ ,..., Α _p by performing linear prediction analysis on the sound signal input in units of frames. Further, the linear prediction analysis unit 12 encodes the generated linear prediction coefficients α ₁ , α ₂ ,..., Α _p to generate a linear prediction coefficient code. An example of the linear prediction coefficient code is LSP (Line Spectrum Pairs). p is an integer of 2 or more.

Further, the linear prediction analysis unit 12 generates quantized linear prediction coefficients ^ α ₁ , ^ α ₂ ,..., ^ Α _p that are linear prediction coefficients corresponding to the generated linear prediction coefficient code.

The generated quantized linear prediction coefficients ^ α ₁ , ^ α ₂ ,..., ^ Α _p are output to the power spectrum envelope sequence generation unit 13. The generated linear prediction coefficient code is output to the decoding device.

<Power Spectrum Envelope Sequence Generation Unit 13>
The quantized linear prediction coefficients ^ α ₁ , ^ α ₂ ,..., ^ Α _p generated by the linear prediction analysis unit 12 are input to the power spectrum envelope sequence generation unit 13.

The power spectrum envelope sequence generation unit 13 uses the quantized linear prediction coefficients ^ α ₁ , ^ α ₂ ,..., ^ Α _p to smooth the power spectrum envelope sequence ^ W _γ defined by the following equation (P1). Generate (0), ^ W _γ (1), ..., ^ W _γ (N-1). Where exp is a real number and exp (•) is an exponential function with the Napier number as the base, j is an imaginary unit, and σ ² is the predicted residual energy. γ is a positive constant of 1 or less, and the amplitude unevenness of the power spectrum envelope sequence W (0), W (1),..., W (N-1) defined by the following formula (P1 ′) A coefficient for blunting, in other words, a coefficient for smoothing the power spectrum envelope sequence.

The generated smoothed power spectrum envelope sequences ^ _Wγ (0), ^ _Wγ (1),..., ^ _Wγ (N−1) are output to the envelope normalization unit 14.

<Envelope normalization unit 14>
The envelope normalization unit 14 includes the MDCT coefficient sequence X (0), X (1),..., X (N-1) generated by the frequency domain conversion unit 11 and the smoothing output from the power spectrum envelope sequence generation unit 13. The power spectrum envelope sequence ^ _Wγ (0), ^ _Wγ (1), ..., ^ _Wγ (N-1) is input.

The envelope normalization unit 14 normalizes each coefficient X (i) of the MDCT coefficient sequence with the square root of each value ^ _Wγ (i)) of the smoothed power spectrum envelope sequence, thereby obtaining a normalized MDCT coefficient sequence X _N. (0), X _N (1), ..., X _N (N-1) are generated. That is, X _N (0) = X (i) / sqrt (^ W _γ (i)) [i = 0, 1,..., N−1]. Here, sqrt (•) represents the square root of •.

The generated normalized MDCT coefficient sequences X _N (0), X _N (1),..., X _N (N−1) are output to the encoding unit 15.

Here, in order to realize quantization that audibly reduces distortion, the envelope normalization unit 14 performs a smoothed power spectrum envelope sequence ^ W _γ that is a power spectrum envelope sequence in which the power spectrum envelope is blunted. Using (0), ^ W _γ (1), ..., ^ W _γ (N-1), MDCT coefficient sequences X (0), X (1), ..., X (N-1) Normalized.

As a result, the generated normalized MDCT coefficient sequences X _N (0), X _N (1),..., X _N (N-1) are inputted MDCT coefficient sequences X (0), X (1), ..., which has no amplitude gradient or amplitude irregularity as large as X (N-1), but has a similar magnitude relationship to the power spectrum envelope sequence of the input sound signal, that is, the coefficient side corresponding to a low frequency This region has a slightly large amplitude and a fine structure resulting from the pitch period.

<Encoding unit 15>
The encoding unit 15 receives the normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1) generated by the envelope normalization unit 14.

The encoding unit 15 generates a code corresponding to the normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1).

Codes corresponding to the generated normalized MDCT coefficient sequences X _N (0), X _N (1),..., X _N (N−1) are output to the decoding device.

A sequence of integer values obtained by dividing the coefficients of normalized MDCT coefficient sequences X _N (0), X _N (1), ..., X _N (N-1) by gain (global gain) g and quantizing the result A code obtained by encoding the quantized normalized coefficient series X _Q (0), X _Q (1),..., X _Q (N−1) is an integer signal code. In the technique of Non-Patent Document 1, the encoding unit 15 determines a gain g such that the number of bits of the integer signal code is equal to or smaller than the allocated bit number B, which is the number of bits allocated in advance, and as large as possible. To do. Then, the encoding unit 15 generates a gain code corresponding to the determined gain g and an integer signal code corresponding to the determined gain g.

The generated gain code and integer signal code are output to the decoding apparatus as codes corresponding to the normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1).

  As described above, in the conventional encoding based on TCX, the normalized MDCT coefficient sequence is encoded after normalizing the MDCT coefficient sequence using the smoothed power spectrum envelope sequence in which the power spectrum envelope is blunted. This encoding method is employed in the above MPEG-4 USAC and the like.

M. Neuendorf, et al., “MPEG Unified Speech and Audio Coding- The ISO / MPEGStandard for High-Efficiency Audio Coding of all Content Types,” AES 132ndConvention, Budapest, Hungary, 2012.

  The power spectrum envelope obtained using the linear prediction coefficient expresses the original spectrum with a resolution of approximately (number of signal samples) / (linear prediction order). This resolution is uniform in the frequency domain. That is, the power spectrum envelope sequence used for normalization of MDCT coefficient sequences in encoding based on conventional TCX such as MPEG USAC is discretized at equal intervals on the frequency axis in the frequency domain, in other words, at a uniform resolution in the frequency direction. It was the envelope value (hereinafter also referred to as “linear discretization”).

  In normal voice and music signals, energy is often concentrated in a specific frequency region (for example, low frequency region), and the power spectrum envelope tends to change greatly in the frequency region where energy is concentrated. When using envelope values that are discretized with uniform resolution in the entire frequency domain, the resolution in the frequency direction is insufficient in the frequency domain where energy is concentrated, and the resulting power spectrum envelope sequence is the amplitude of the original power spectrum envelope. May not be expressed with sufficient accuracy. When the MDCT coefficient sequence is normalized using such a power spectrum envelope sequence, the difference between the MDCT coefficient sequence and the power spectrum envelope sequence in the portion where the resolution is insufficient increases, and the variation in the value of the normalized MDCT coefficient sequence is increased. Since it becomes large, encoding efficiency may fall.

  Here, if the prediction order is increased, the resolution can be increased even by linear discretization, but there is a possibility that the amount of parameter information increases and the coding efficiency decreases. Further, if the order is increased only for a specific frame, the processing may become complicated due to the continuity of processing between frames.

  In signal processing of a sound signal, not limited to encoding processing, signal processing is performed by using a sample sequence obtained by discretizing a sound signal with a non-uniform resolution in the frequency direction as a sample sequence in the frequency domain derived from the sound signal. Accuracy may be improved.

  In view of such a technical background, an object of the present invention is to provide a technique for improving coding efficiency with a small increase in calculation amount. It is another object of the present invention to provide a technique for improving the accuracy of signal processing with a small increase in the amount of computation in signal processing other than encoding.

  A sample sequence generation method according to an aspect of the present invention (hereinafter referred to as a first sample sequence generation method) uses a frequency domain sample sequence derived from a sound signal for each predetermined time interval as ξ (0), ξ (1 ), ..., ξ (N-1) are defined by the following equations using a predetermined N × N sparse matrix U: ~ ξ (0), ~ ξ (1), ..., ~ ξ ( N-1) has a sample sequence generation step,

The sparse matrix U is a matrix that includes non-zero values only in the neighboring components of the diagonal component, and the number of nonzero elements included in the upper triangular matrix is greater than the number of nonzero elements included in the lower triangular matrix. The sample point sequence corresponding to ξ (0), ξ (1),..., Ξ (N-1) is a linear discretized sample point sequence in which the frequency intervals of adjacent sample points are uniform.

  A sample sequence generation method (hereinafter referred to as a second sample sequence generation method) according to an aspect of the present invention uses a frequency domain sample sequence derived from a sound signal for each predetermined time interval as ξ (0), ξ (1 ), ..., ξ (N-1) is defined by the following equation using a predetermined non-negative matrix U of M × N, which is defined by the following formula: ~ ξ (0), ~ ξ (1), ..., ~ ξ A sample sequence generation step for generating (M-1),

The non-negative matrix U has a component of the i-th row and k-th column as U [i, k], and for each i = 1, 2,.

Where g ₁ <g ₂ <... <g _N is satisfied, and ξ (0), ξ (1), ..., ξ (N-1) or ~ ξ (0), ~ ξ (1), ... , ~ ξ (M−1) are sample point sequences that are equally spaced in the frequency direction.

  A sample sequence generation method according to an aspect of the present invention (hereinafter referred to as a third sample sequence generation method) uses a frequency domain sample sequence derived from a sound signal in a predetermined time interval as ~ η (0), ~ η ( 1), ..., ~ η (N-1), using a predetermined N × N sparse matrix V, η (0), η (1), ..., η (N -1) has a sample sequence generation step to generate

The sparse matrix V is a matrix including non-zero values only in the neighboring components of the diagonal component, and the number of nonzero elements included in the upper triangular matrix is greater than the number of nonzero elements included in the lower triangular matrix. In many cases, the sample point sequence corresponding to η (0), η (1),..., Η (N−1) is a linear discretized sample point sequence in which the frequency intervals of adjacent sample points are equal.

In the encoding method according to an aspect of the present invention, the sequence corresponding to the power of the sample sequence obtained by converting the sound signal for each predetermined time interval into the frequency domain by the first sample sequence generation method is the ξ (0 ), ξ (1), ..., ξ (N-1), and ~ ξ (0), ~ ξ (1), ..., ~ ξ (N-1) are expanded and contracted pseudo power spectrum series ~ Y (0) , ~ Y (1),..., ~ Y (N-1) to generate a stretched pseudo power spectrum sequence, and p is a positive integer, the stretched pseudo power spectrum series ~ Y (0), ~ Y ( 1), ..., ~ Y (N-1) is transformed into time domain ~ X (0), ~ X (1), ..., ~ X (N-1) of telescopic linear prediction coefficient _{^ β 1, ^ β 2,} ..., a linear prediction analysis step of generating a ^ beta _p, the quantization telescopic linear prediction coefficient _{^ β 1, ^ β 2,} ..., corresponding to the ^ beta _p Stretching power spectrum envelope series ~ W (0), ~ W (1), ..., ~ W (N-1), which is a frequency domain series And the third sample sequence generation method, the stretchable power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W (N-1) is converted into ~ η (0), ~ η ( 1), ..., ~ η (N-1), and η (0), η (1), ..., η (N-1) are power spectrum envelope sequences W (0), W (1), ..., The inverse stretch conversion step generated as W (N-1) and the frequency spectrum sample sequence X (0) using the power spectrum envelope sequence W (0), W (1), ..., W (N-1) ), X (1),..., X (N-1) is normalized to generate an normalized frequency domain sample sequence, and the normalized frequency domain sample sequence is encoded. And generating a code corresponding to the normalized frequency domain sample sequence.

The decoding method according to an aspect of the present invention generates quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p by decoding an input stretched linear prediction coefficient code, where p is a positive integer. And a stretched power spectrum envelope sequence that is a frequency domain sequence corresponding to the quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p ~ W (0), ~ The stretched power spectrum envelope sequence ~ W (0), ~ W (1) is generated by the stretched power spectrum envelope series generating step for generating W (1), ..., ~ W (N-1) and the third sample string generating method. ), ..., ~ W (N-1) as ~ η (0), ~ η (1), ..., ~ η (N-1), and η (0), η (1), ..., η Inverse expansion / conversion conversion step for generating (N-1) as a power spectrum envelope sequence W (0), W (1), ..., W (N-1), and an input sample sequence of normalized frequency domain _{X N (0), X N} (1), ..., and decodes the code corresponding to X _{N (N-1),} the difference in the normalized frequency domain Pull string _{X N (0), X N} (1), ..., a decoding step of generating a X _{N (N-1),} the power spectrum envelope sequences W (0), W (1 ), ..., W (N -1) is used to denormalize the normalized frequency domain sample sequence X _N (0), X _N (1),..., X _N (N-1) to obtain a frequency domain sample sequence. Envelope denormalization step for generating X (0), X (1),..., X (N-1).

  A sample string generation device (hereinafter referred to as a first sample string generation device) according to an aspect of the present invention uses a frequency domain sample string derived from a sound signal for each predetermined time interval as ξ (0), ξ (1 ), ..., ξ (N-1) are defined by the following equations using a predetermined N × N sparse matrix U: ~ ξ (0), ~ ξ (1), ..., ~ ξ ( Including a sample sequence generator for generating (N-1)

The sparse matrix U is a matrix that includes non-zero values only in the neighboring components of the diagonal component, and the number of nonzero elements included in the upper triangular matrix is greater than the number of nonzero elements included in the lower triangular matrix. The sample point sequence corresponding to ξ (0), ξ (1),..., Ξ (N-1) is a linear discretized sample point sequence in which the frequency intervals of adjacent sample points are equal. .

  A sample sequence generation device according to an aspect of the present invention (hereinafter referred to as a third sample sequence generation device) uses a sample sequence in a frequency domain derived from a sound signal for each predetermined time interval as ~ η (0), ~ η (1),..., ~ Η (M-1), using a predetermined N × N sparse matrix V, η (0), η (1),. Including a sample sequence generator for generating (N-1)

The sparse matrix V is a matrix including non-zero values only in the neighboring components of the diagonal component, and the number of nonzero elements included in the upper triangular matrix is greater than the number of nonzero elements included in the lower triangular matrix. In many cases, the sample point sequence corresponding to η (0), η (1),..., Η (N−1) is a linear discretized sample point sequence in which the frequency intervals of adjacent sample points are equal.

In the encoding device according to one aspect of the present invention, the sequence corresponding to the power of the sample sequence obtained by converting the sound signal for each predetermined time interval into the frequency domain is the ξ (0), ξ (1),. , ξ (N-1) and ~ ξ (0), ~ ξ (1), ..., ~ ξ (N-1) are expanded and contracted pseudo power spectrum series ~ Y (0), ~ Y (1), ... , ~ Y (N-1), a stretched pseudo power spectrum sequence generation unit that is a first sample string generating device, and p is a positive integer, the stretched pseudo power spectrum series ~ Y (0), ~ Y ( 1), ..., ~ Y (N-1) is transformed into time domain ~ X (0), ~ X (1), ..., ~ X (N-1) of telescopic linear prediction coefficient _{^ β 1, ^ β 2,} ..., a linear prediction analysis unit for generating a ^ beta _p, the quantization telescopic linear prediction coefficient _{^ β 1, ^ β 2,} ..., corresponding to the ^ beta _p A stretchable power spectrum envelope sequence generating unit for generating stretchable power spectrum envelope sequences ~ W (0), ~ W (1), ..., ~ W (N-1), which are frequency domain sequences, Power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W (N-1) as ~ η (0), ~ η (1), ..., ~ η (N-1), Second sample sequence generation for generating η (0), η (1),..., Η (N-1) as power spectrum envelope sequences W (0), W (1),. The inverse stretch / conversion transforming unit, which is a device, and the power spectrum envelope sequence W (0), W (1),..., W (N-1), and the frequency domain sample sequence X (0), X (1) ,..., X (N-1) is normalized to generate an normalized frequency domain sample sequence, and the normalized frequency domain sample sequence is encoded and the normalized And an encoding unit that generates a code corresponding to the frequency domain sample sequence.

The decoding device according to an aspect of the present invention generates quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p by decoding an input stretched linear prediction coefficient code, where p is a positive integer. And a stretched power spectrum envelope sequence that is a frequency domain sequence corresponding to the quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p ~ W (0), ~ Stretched power spectrum envelope sequence generation unit for generating W (1), ..., ~ W (N-1) and the stretched power spectrum envelope series ~ W (0), ~ W (1), ..., ~ W (N -1) as ~ η (0), ~ η (1), ..., ~ η (N-1), and η (0), η (1), ..., η (N-1) as power spectrum Inverse expansion / conversion transform unit, which is a second sample sequence generation device that generates as envelope sequences W (0), W (1),..., W (N-1), and an input sample sequence in a normalized frequency domain X _N (0), X _N (1), ..., code corresponding to X _N (N-1) is decoded and normalized frequency domain sample sequence X _N (0), X _N (1) ,…, X _N (N-1) and a frequency domain sample sequence X _N normalized using the power spectrum envelope sequence W (0), W (1),..., W (N-1) By denormalizing (0), X _N (1), ..., X _N (N-1), the frequency domain sample sequence X (0), X (1), ..., X (N-1) An envelope denormalization unit for generating

  Coding efficiency can be improved. Alternatively, the accuracy of signal processing can be improved.

The block diagram for demonstrating the example of the conventional encoding apparatus. The block diagram for demonstrating the example of the encoding apparatus of 1st embodiment and 2nd embodiment. The flowchart for demonstrating the example of the encoding method of 1st embodiment and 2nd embodiment. The figure for demonstrating the example of the conversion matrix U. FIG. The figure for demonstrating the example of the conversion matrix U. FIG. The block diagram for demonstrating the example of the decoding apparatus of 1st embodiment and 2nd embodiment. The flowchart for demonstrating the example of the decoding method of 1st embodiment and 2nd embodiment. The block diagram for demonstrating the example of the encoding apparatus of 3rd embodiment. The flowchart for demonstrating the example of the encoding method of 3rd embodiment. The block diagram for demonstrating the example of the decoding apparatus of 3rd embodiment. The flowchart for demonstrating the example of the decoding method of 3rd embodiment. The block diagram for demonstrating the example of the encoding apparatus of 4th embodiment. The flowchart for demonstrating the example of the encoding method of 4th embodiment. The block diagram for demonstrating the example of the decoding apparatus of 4th embodiment. The flowchart for demonstrating the example of the decoding method of 4th embodiment. The figure for demonstrating the example of the conversion matrix U. FIG. The figure for demonstrating the property of the gravity center position in the conversion matrix U of FIG. The figure for demonstrating the example of the transformation matrix V. FIG.

[Technical background]
First, the technical background will be described using the encoding process described in the related art as an example.

  In an example of the present invention, when a power spectrum envelope sequence is used, a power spectrum envelope sequence that is a discrete coefficient sequence is generated by discretization using nonlinear resolution in the frequency direction. At this time, it is discretized with a fine resolution in a frequency region where the amplitude variation of the power spectrum envelope is large, and is discretized with a coarse resolution in a frequency region where the amplitude variation of the power spectrum envelope is small. Thereby, the variation in the value of the normalized MDCT coefficient sequence can be reduced, and the encoding efficiency can be increased.

  For example, the discretization interval in the frequency region where energy is concentrated is made smaller than the discretization interval in other frequency regions. In other words, the resolution of the frequency region where energy is concentrated is set higher than the resolution of other frequency regions.

  Conventionally, the power spectrum envelope is expressed by discretization with linear resolution in the frequency direction. That is, when F = 0Hz to F = 100Hz are all frequency regions, N = 10, and discretization with 11 sample points, each of the 11 frequencies in the table below is used as a sample point. The power spectrum envelope is expressed by a series of power spectrum envelope values corresponding to the sample points.

つまり、隣接するサンプル点間の周波数領域での間隔が均等（上述の例では10Hz間隔）であるようなサンプル点列を用いてパワースペクトル包絡を表現していた。

That is, the power spectrum envelope is expressed using a sample point sequence in which the intervals in the frequency domain between adjacent sample points are uniform (in the above example, 10 Hz intervals).

  In this way, the sample point sequence for expressing the power spectrum envelope sequence is discretized so that the intervals in the frequency domain between adjacent sample points are uniform (in the above example, 10 Hz intervals). This is called “linear discretization”.

  On the other hand, in an example of the present invention, the power spectrum envelope is expressed by discretization with nonlinear resolution in the frequency direction. For example, each of the 11 frequencies in the following table is set as a sample point, and the power spectrum envelope is expressed by a series of power spectrum envelope values corresponding to these 11 sample points.

この例では、低周波数領域の方が隣接するサンプル点間の周波数領域での間隔が狭く、高周波数領域ほど隣接する離散化サンプル点間の周波数領域での間隔が広くなっている。例えば、最も低周波数領域のサンプル点の間隔、言い換えればインデックス0に対応するサンプル点とインデックス1に対応するサンプル点との周波数領域での間隔は1Hzであるが、最も高周波数領域のサンプル点の間隔、言い換えればインデックス9に対応するサンプル点とインデックス10に対応するサンプル点との周波数領域での間隔は30Hzとなっている。

In this example, the interval in the frequency region between adjacent sample points is narrower in the low frequency region, and the interval in the frequency region between adjacent discrete sample points is wider in the higher frequency region. For example, the interval between the sample points in the lowest frequency region, in other words, the interval in the frequency region between the sample point corresponding to index 0 and the sample point corresponding to index 1 is 1 Hz, but the sample point in the highest frequency region The interval, in other words, the interval in the frequency domain between the sample point corresponding to the index 9 and the sample point corresponding to the index 10 is 30 Hz.

  In this way, discretizing a sample point sequence for expressing a power spectrum envelope sequence so that intervals in the frequency direction between adjacent sample points are not uniform is called “nonlinear discretization”.

  In the following, in order to distinguish the difference in resolution in the frequency direction, a sequence of sample points that are equally spaced in the frequency direction is also referred to as a “linear discretization sample point sequence”. The series is also referred to as a “nonlinear discretized sample point sequence”. The frequency intervals between adjacent sample points in the linear discretized sample point sequence are uniform, but the frequency intervals between adjacent sample points in the non-linear discretized sample point sequence are unequal. Each sample point included in the linear discretized sample point sequence is also referred to as “linear discretized sample point”, and each sample point included in the nonlinear discretized sample point sequence is also referred to as “nonlinear discretized sample point”.

  Also, the power sequence of the input sound signal corresponding to each sample point of the linear discretized sample point sequence is also called “power spectrum sequence”, and the input sound corresponding to each sample point of the non-linear discretized sample point sequence The signal power sequence is also referred to as a “stretching pseudo power spectrum sequence”.

  In the above example, the interval between the sample points of the non-linear discretized sample point sequence is narrower in the low frequency region, but it is not always necessary to have this property. For example, the middle frequency region is lower than the low frequency region. May be narrow. In short, in the non-linear discretized sample point sequence, the interval between adjacent sample points may be non-uniform.

[First embodiment]
(Encoding of the first embodiment)
A configuration example of the encoding apparatus of the first embodiment is shown in FIG. As shown in FIG. 2, the encoding device according to the first embodiment includes a frequency domain transform unit 21, a stretched pseudo power spectrum sequence generation unit 22, a linear prediction analysis unit 23, and a stretched power spectrum envelope sequence generation unit 24. The inverse expansion / conversion conversion unit 25, the envelope normalization unit 26, and the encoding unit 27 are provided, for example. FIG. 3 shows an example of each process of the encoding method of the first embodiment realized by this encoding apparatus.

  Hereinafter, each part of FIG. 2 will be described.

<Frequency domain converter 21>
A time domain sound signal is input to the frequency domain converter 21. Examples of sound signals are voice digital signals or acoustic digital signals.

  The frequency domain transform unit 21 converts the input time domain sound signal into N frequency MDCT coefficient sequences X (0), X (1),..., X (N− 1) (step E1). N is a positive integer.

  The converted MDCT coefficient sequences X (0), X (1),..., X (N−1) are output to the envelope normalization unit 26.

  Unless otherwise specified, the subsequent processing is performed in units of frames.

  Each sample point corresponding to the MDCT coefficient sequence X (0), X (1),..., X (N−1) here is a linear discretization sample point. That is, the frequency intervals of adjacent sample points in the sample point sequence corresponding to the MDCT coefficient sequence X (0), X (1),..., X (N−1) are equal. In other words, when i = 0, 1,..., N−2, the interval between the frequency corresponding to the index i in the MDCT coefficient sequence and the frequency corresponding to the index i + 1 in the MDCT coefficient sequence is equal.

<Expandable pseudo power spectrum sequence generation unit 22>
MDCT coefficient sequences X (0), X (1),..., X (N−1) converted by the frequency domain conversion unit 21 are input to the expansion / contraction pseudo power spectrum sequence generation unit 22.

First, the expansion / contraction pseudo power spectrum sequence generation unit 22 is a power spectrum sequence which is a sequence including the square value (power) of each coefficient of the MDCT coefficient sequence X (0), X (1),..., X (N-1). Y (0), Y (1), ..., Y (N-1) are generated. That is, Y (i) = X (i) ² (i = 0, 1,..., N−1).

  The expansion / contraction pseudo power spectrum sequence generation unit 22 performs interpolation or linear conversion of the power spectrum sequence Y (0), Y (1),. 0), ~ Y (1), ..., ~ Y (N-1) are generated (step E2).

  The generated stretched pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1) is output to the linear prediction analysis unit 23.

  Here, the sample point sequence corresponding to the power spectrum sequence Y (0), Y (1),..., Y (N-1) is a linear discretized sample point sequence, and the stretched pseudo power spectrum sequence ~ Y (0) , ~ Y (1), ..., ~ Y (N-1) are sample point sequences that are nonlinear discretized sample point sequences.

  In other words, the frequency intervals corresponding to the indexes 0, 1,..., N-1 in the power spectrum series Y (0), Y (1),. Also, the frequency intervals corresponding to the indices 0, 1, ..., N-1 in the expansion / contraction pseudo power spectrum series ~ Y (0), ~ Y (1), ..., ~ Y (N-1) are uneven. It is an interval.

  When generating the expansion / contraction pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1) by interpolation, the expansion / contraction pseudo power spectrum series generation unit 22 performs the following processing, for example. Let f be the frequency between adjacent sample points in the sample point sequence corresponding to the power spectrum series Y (0), Y (1),..., Y (N-1). Then, an interpolated curve is obtained assuming that the power spectrum value at f follows the sinc function (sinc (f) = sin (f) / f). Then, by setting the value corresponding to each sample point (frequency) of the non-linear discretized sample point sequence in the curve as the stretch pseudo power spectrum value, the stretch pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1) is generated. In this case, it is assumed that the nonlinear discretized sample point sequence is given in advance.

  When obtaining the expansion / contraction pseudo power spectrum sequence ~ Y (0), ..., ~ Y (N-1) by linear conversion, the expansion / contraction pseudo power spectrum series generation unit 22 generates the power spectrum series Y (0), Y (1), ..., Y (N-1) is multiplied by a predetermined transformation matrix U from the left to expand and contract the pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1 ) Is generated.

  In other words, the expansion / contraction pseudo power spectrum sequence generation unit 22 generates expansion / contraction pseudo power spectrum sequences ~ Y (0), ~ Y (1), ..., ~ Y (N-1) defined by the following equations.

  Here, the transformation matrix U is a matrix that approximates the mapping from the linear discretized sample point sequence to the nonlinear discretized sample point sequence. This conversion is to convert a sample point sequence equally spaced in the frequency direction to a sample point sequence of unequal intervals in the frequency direction, so to speak, it expands or contracts the frequency interval between adjacent sample points. This is referred to as “stretch conversion”.

  The frequency of a non-linear discretization sample point with a non-linear discretization sample point sequence after stretch conversion (hereinafter referred to as post-stretch frequency) is the stretch of the linear discretization sample points of the linear discretization sample point sequence before stretch conversion. It can be approximated by a weighted sum of the frequencies of one or more linear discretization sample points having a frequency close to the rear frequency. In other words, the post-stretching frequency can be approximated by a weighted sum of the frequencies of one or more linear discretization sample points in the vicinity of the linear discretization sample point of the frequency closest to the post-stretching frequency.

  Each row of the transformation matrix U corresponds to each nonlinear discretization sample point of the nonlinear discretization sample point sequence after the stretch transformation, and each column of the transformation matrix U corresponds to each linear discretization sample point of the linear discretization sample point sequence . That is, each row of the transformation matrix U is a series of weights for the frequency of each linear discretization sample point for expressing the frequency of the nonlinear discretization sample point corresponding to each row.

  Originally, if the scaled frequency is approximated by a weighted sum of the frequencies of one or more linear discretization sample points in the vicinity of the linear discretization sample point of the frequency closest to the scaled frequency, the weight is set to a negative value. Also good. However, if the transformation matrix U is configured so as to include negative values, in order to accurately generate the stretchable pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1) Post-processing is required. Therefore, in the present invention, for example, all elements of the transformation matrix U are set to non-negative values (that is, U is set to a non-negative value matrix). As a result, it is possible to generate a highly accurate expansion / contraction pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1) without post-processing.

  Furthermore, since the weight multiplied by the frequency of the discretized sampling point at a frequency distant from the post-stretching frequency is assumed to be a small value, even if the small value element is regarded as 0, the stretched pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1) have little effect on the accuracy. Therefore, in each row of the transformation matrix U, only the neighboring elements in the column corresponding to the linear discretization sample point having the frequency closest to the frequency of the nonlinear discretization sample point corresponding to each row are set to non-zero values, and the remaining elements are set to 0. For example, it may be set. Here, an element having a value other than 0 in the transformation matrix U is referred to as an “element corresponding to expansion / contraction”. It can be said that the transformation matrix U is a band matrix (sparse matrix) that has a non-zero value only in the vicinity of the corresponding element due to expansion and contraction and the other components are zero, for example.

The conversion using all values of the matrix may increase the amount of calculation, but by making the conversion matrix U sparse in this way, the expansion / contraction pseudo power spectrum can be obtained with a small amount of calculation. A starting sample point of a non-zero element in the matrix may be stored separately, and a small number of operations may be performed only from the sample point.
Thus, by making the transformation matrix U a non-negative matrix or a sparse matrix, the expansion / contraction pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N -1) can be obtained. When the transformation matrix U is a sparse matrix, it may include a negative value element.

  The transformation matrix U can be obtained in advance by learning or the like from the correlation between the non-linear discretized sample point sequence and the linear discretized sample point sequence. A method for obtaining the transformation matrix U will be described later.

The degree of expansion / contraction of the interval between adjacent sample points in the non-linear discretized sample point sequence can be set arbitrarily. The following generalized logarithm can be used as an example of reducing the interval between sample points in the low frequency region and expanding the interval between sample points in the high frequency region. Note that “interval between adjacent sample points”, also called “discretization width”, is an example of a function representing the relationship with the frequency ω _i corresponding to the sample point of index i of the linear discretization sample point sequence S _λ (ω _i ) can be used. For example, the following relationship holds between the frequency S _λ (ω _i ) and the frequency ω _i corresponding to the sample point of the same index i in the non-linear discretized sample point sequence.

S _λ (ω _i ) is a function called a generalized logarithmic function. λ is a constant that determines the nonlinear expansion / contraction degree, that is, the expansion / contraction degree of the interval between the sample points of the nonlinear discretized sample point sequence. Λ, which is the degree of expansion / contraction conversion, can be designed according to the nature of the input signal. When λ = 0, S _λ (ω _i ) is a logarithmic function.

  An example of the relationship between the transformation matrix U and λ is shown in FIG. 4 (a), 4 (b), and 4 (c), the vertical axis represents the index corresponding to the row of the transformation matrix U, and the horizontal axis represents the index corresponding to the column of the transformation matrix U. The value of each element is indicated by color. White represents an element having a value of 0, and black represents an element having a value greater than 0. The transformation matrix U in the case of λ = 1 is a matrix in which nothing is transformed as a linear discretized sample point sequence, and is a diagonal matrix in which only the diagonal component has a value of 1. From these figures, as λ increases, the shape of the curve formed by the non-zero components in the transformation matrix U, in other words, the shape of the curve that approximates an element having a value greater than 0, is bent more nonlinearly below the diagonal line. I understand that.

  Here, the vertical axis is the frequency of the sample points of the linear discretization sample point sequence, the frequency increases from the top to the bottom of the vertical axis, the horizontal axis is the frequency of the nonlinear discretization sample point sequence, and the horizontal axis Consider a two-dimensional plane that is defined as a frequency that increases from left to right.

  The shape of the curve formed by the non-zero components in the transformation matrix U is mapped on this two-dimensional plane with the frequency of the linear discretization sampling point and the frequency of the nonlinear discretization sampling point corresponding to the index of each element of the transformation matrix U. It can be said that it corresponds to a curve obtained by interpolating the point sequence at that time. Hereinafter, this curve is also referred to as a “stretching curve”. In other words, the expansion / contraction curve is a curve obtained by interpolating a point sequence in which “elements corresponding to expansion / contraction” are mapped. Here, the linear discretization sample point corresponding to the element index is a linear discretization sample point corresponding to the element column. The non-linear discretization sample point corresponding to the element index is a non-linear discretization sample point corresponding to the row of the element.

  In the above example, an example of conversion in which the resolution in the low frequency region is higher than the resolution in the high frequency region is shown. In other words, an example of transformation is shown in which the interval between nonlinear discrete sample points in the low frequency region is narrower than the interval between nonlinear discrete sample points in the high frequency region. However, this is only an example.

  Since the resolution or the degree of expansion and contraction corresponds to the slope of the sample point where the non-zero element of the transformation matrix exists, for example, as a non-linear discretization sample point sequence, the resolution in the frequency region intermediate between the low frequency region and the high frequency region is A non-linear discretized sample point sequence in which a region that is higher than a resolution of other frequency regions and a region that is lower than the center may be used. In this case, the transformation matrix U is a transformation matrix in which, for example, only neighboring components on the stretch curve as shown in FIG. 5 are non-zero.

  In any case, the transformation matrix U is an M × N matrix, and the expansion / contraction curve in the transformation matrix U corresponds to the component at the upper left corner in a two-dimensional plane with the row as the horizontal axis and the column as the vertical axis. The curve is monotonically decreasing from [1,1] toward the point [M, N] corresponding to the lower right component. In other words, the transformation matrix U has the following properties.

The center of gravity g _i of the i-th row of the transformation matrix U is

It is defined as U [i, k] represents the (i, k) element of the transformation matrix U. Assume that the row and column indices start at 1. Then, sequence g _1, g ₂ of the center of gravity of each row of the transformation matrix U, ..., g _N satisfies the relationship of _{g 1 <g 2 <... <} g N. M and N are integers of 3 or more. In this embodiment, M = N.

In addition, since energy is often concentrated in a low frequency region or a medium frequency region in an audio signal and an acoustic signal, the frequency interval between nonlinear discrete sampling points in the low frequency region or the medium frequency region is set to a non-linear region in the high frequency region. A stretchable pseudo power spectrum corresponding to a non-linear discretized sample point sequence narrower than the frequency interval between the discretized sample points may be used. In this case, in each row of the transformation matrix U, the distance between the non-zero element of each row that is farthest from the diagonal element and the diagonal element (this distance is how far the column index is separated) Represents the expansion / contraction distance of the index of each row, the average of the expansion / contraction distance with respect to the index of the first half of all indexes tends to be larger than the average of the expansion / contraction distance with respect to the index of the rear half. The non-zero component takes a positive value. Alternatively, the transformation matrix U is a sparse matrix in which the number of non-negative components included in the upper triangular matrix is smaller than the number of non-negative components included in the lower triangular matrix. FIG. 16 shows an example of the transformation matrix U. In this example, N = M = 16. Figure 17 is a center of gravity g _i of the transformation matrix U in FIG. 16, the row number and the horizontal axis shows a plot on a two-dimensional plane having a longitudinal axis the value of g _i. As shown in FIG. 17, the value of g _i increases monotonically as the line number increases.

  Expanded pseudo power spectrum series ~ Y (0), ~ Y (1), ..., ~ Y (N-1) is a non-linear discretization of the power spectrum of the signal converted from the input sound signal to the frequency domain. Equivalent to.

  In this way, the nonlinear discrete discretization is such that the discretization interval in the frequency domain where the energy of the signal converted from the input sound signal is converted to the frequency domain is narrower than the discretization interval in other frequency domains. The power spectrum is expressed based on the sampling points of the conversion interval.

<Linear prediction analysis unit 23>
The linear prediction analysis unit 23 receives the expansion / contraction pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1) generated by the expansion / contraction pseudo power spectrum series generation unit 22.

The linear predictive analysis unit 23 uses the stretched pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1) to define ~ X (0), ~ X (1), ..., ~ X (N-1) are subjected to linear prediction analysis to generate stretched linear prediction coefficients β ₁ , β ₂ , ..., β _p, and the generated stretched linear prediction coefficients β ₁ , β ₂ , ..., β _p is encoded and the stretched linear prediction coefficient code and the quantized stretched linear prediction coefficient ^ β ₁ , ^ β ₂ ,…, which is the quantized stretched linear prediction coefficient corresponding to the stretched linear prediction coefficient code ^ β _p is generated (step E3).

The generated quantized stretch linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p are output to the stretch power spectrum envelope sequence generation unit 24.

  The generated stretched linear prediction coefficient code is transmitted to the decoding device.

In order to generate the stretched linear prediction coefficient code, the linear prediction analysis unit 23 first samples points corresponding to the stretched pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1). The time domain corresponding to ~ Y (0), ~ Y (1), ..., ~ Y (N-1) is obtained by assuming that the frequency intervals of , X (0), ~ X (1), ..., ~ X (N-1) are obtained. Then, the linear prediction analysis unit 23 performs linear prediction analysis on the obtained stretch correlation function signal sequence ~ X (0), ~ X (1), ..., ~ X (N-1) to obtain stretch linear prediction. Coefficients β ₁ , β ₂ ,..., Β _p are generated. Then, the linear prediction analyzer 23, the generated elastic linear prediction coefficient beta _1, beta _2, ..., by encoding the beta _p, generates a telescopic linear prediction coefficient code. As a result, quantized stretch linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p corresponding to the stretch linear prediction coefficient code are also obtained.
The expansion / contraction linear prediction coefficient is assumed that the intervals in the frequency direction of the sample points of the expansion / contraction pseudo power spectrum series ~ Y (0), ~ Y (1), ..., ~ Y (N-1) are equal. It is a linear prediction coefficient corresponding to the signal of the time domain at the time.

  The generation of the expansion / contraction linear prediction coefficient code by the linear prediction analysis unit 23 is performed by, for example, a conventional encoding technique. The conventional encoding technique is, for example, an encoding technique in which a code corresponding to the linear prediction coefficient itself is a prediction coefficient code, a code corresponding to the LSP parameter by converting the linear prediction coefficient into an LSP parameter, and a prediction coefficient code. Encoding techniques for converting linear prediction coefficients into PARCOR coefficients and using codes corresponding to the PARCOR coefficients as prediction coefficient codes. In these conventional encoding techniques, the linear prediction coefficient is replaced with the expansion / contraction linear prediction coefficient, and the expansion linear prediction coefficient code that is a code corresponding to the expansion / contraction linear prediction coefficient is obtained.

<Elastic Power Spectrum Envelope Sequence Generation Unit 24>
The expansion / contraction power spectrum envelope sequence generation unit 24 receives the quantized expansion / contraction linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p generated by the linear prediction analysis unit 23.

Stretch power spectrum envelope sequence generating unit 24, the quantization telescopic linear prediction coefficient _{^ β 1, ^ β 2,} ..., ^ β by conversion to the frequency domain to _p, stretching the power spectral envelope sequence ~ W (0), ~ W (1),..., W (N-1) are generated (step E4).

  The generated expansion / contraction power spectrum envelope sequences ~ W (0), ~ W (1), ..., ~ W (N-1) are output to the inverse expansion / conversion conversion unit 25.

The expansion / contraction power spectrum envelope sequence generation unit 24 uses the expansion / contraction linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p to generate expansion / contraction power spectrum envelope sequences ~ W (0), ~ W (1), ..., ~ W (N-1), for example, a stretched unsmoothed power spectrum envelope sequence defined by the following equation (2) ~ _Wo (0), ~ _Wo (1), ..., ~ _Wo ( N-1) or the stretched and smoothed power spectrum envelope sequence ~ _Wγ (0), ~ _Wγ (1), ..., ~ _Wγ (N-1) defined by the following equation (3).

In other words, an example of a stretched power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W (N-1) is an stretched unsmoothed power spectrum envelope series defined by equation (2) ~ W _o (0), ~ W _o (1), ..., ~ W _o (N-1) or the stretched and smoothed power spectrum envelope sequence defined by equation (3) ~ W _γ (0), ~ W _γ ( 1), ..., ~ W _γ (N-1).

Here, the correction coefficient γ is a predetermined constant equal to or less than 1, and the stretched and unsmoothed power spectrum envelope sequence ~ _Wo (0), ~ _Wo (1), ..., ~ _Wo (N-1) Is a coefficient that smoothes the unsmoothed power spectrum envelope sequence ~ _Wo (0), ~ _Wo (1), ..., ~ _Wo (N-1). is there. It may be considered that it is the same as γ used for smoothing the power spectrum envelope sequence in the conventional encoding process that does not perform expansion / conversion, that is, γ in the above formula (P1).

  Note that the sampled point sequence corresponding to the stretchable power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W (N-1), in other words, the stretchable power spectrum envelope series ~ W (0), ~ W A sequence of frequencies corresponding to indexes 0, 1,..., N-1 in (1),..., ~ W (N-1) is a non-linear discretized sample point sequence.

  In this way, the expansion / contraction power spectrum envelope sequence generation unit 24 discretizes the envelope obtained by smoothing the power spectrum envelope of the frequency domain sample sequence derived from the sound signal for each predetermined time interval at unequal intervals in the frequency direction. The expansion / contraction power spectrum envelope sequence that is the sequence is generated.

In this example, the sample sequence in the frequency domain derived from the sound signal for each predetermined time interval is the MDCT coefficient sequence X (0), X (1),..., X (N−1). A sample sequence in the frequency domain other than the MDCT coefficient sequence X (0), X (1), ..., X (N-1) may be used as the frequency domain sample sequence derived from the sound signal for each predetermined time interval. Good.
The stretchable power spectrum envelope series ~ W (0), ~ W (1), ..., ~ W (N-1) is expressed with fine resolution where the frequency interval of the sample points is narrow. It is also possible to express fine changes in amplitude irregularities. On the contrary, when the frequency interval of the sample points is wide, it is expressed with a coarse resolution, so that only a rough change of the power spectrum envelope is expressed. In general, an audio-acoustic signal has a large change in power spectrum envelope in a portion where energy is concentrated, and a small change in power spectrum envelope in other portions. Therefore, a nonlinear discrete sample point sequence in which the frequency interval between adjacent sample points in the frequency region where energy is concentrated is narrower than the frequency interval between adjacent sample points in other frequency regions is used. A limited sample is obtained by setting the power sequence of the input sound signal in the corresponding frequency domain as a stretched pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1) It is possible to obtain a discrete series that more accurately expresses the change in the amplitude unevenness of the power spectrum by the point sequence. In other words, a discrete sequence that expresses the power spectrum more accurately can be obtained by performing nonlinear discretization so that the frequency region in which energy is concentrated is expressed with a finer resolution than the other frequency regions.

  When the MDCT coefficient sequence is normalized using the spectrum envelope value calculated based on the expansion / contraction pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1) thus obtained, Since the change in the size of the generalized MDCT coefficient sequence becomes small, it becomes possible to efficiently encode.

<Reverse expansion / conversion conversion unit 25>
The inverse expansion / conversion conversion unit 25 receives the expansion / contraction power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W (N-1) generated by the expansion / contraction power spectrum envelope series generation unit 24.

  The inverse expansion / conversion conversion unit 25 converts the expansion / contraction power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W (N-1) into power corresponding to the linear discretized sample point sequence by interpolation or linear conversion. It converts into the spectrum envelope series W (0), W (1),..., W (N-1) (step E5).

  The converted power spectrum envelope sequences W (0), W (1),..., W (N-1) are output to the envelope normalization unit 26.

  When obtaining the power spectrum envelope series W (0), W (1),..., W (N−1) by interpolation, the inverse expansion / conversion conversion unit 25 performs, for example, the following processing. Similar to the expansion / contraction pseudo power spectrum sequence generation unit 22, the inverse expansion / conversion conversion unit 25 interpolates the expansion / contraction power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W (N-1) by the sinc function. Obtained curve (envelope that smoothly connects the expansion and contraction power spectrum envelope). Then, the power spectrum envelope sequence W (0), W (1),..., W (N− Get as 1).

  When obtaining the power spectrum envelope sequence W (0), W (1),..., W (N-1) by linear transformation, the inverse expansion / conversion conversion unit 25 performs the expansion / contraction power spectrum envelope sequence ~ W (0), ~ W ( 1), ..., ~ W (N-1) is multiplied by a predetermined transformation matrix V from the left to multiply the power spectrum envelope sequence W (0), W (1), ..., W (N-1 ) Is generated.

  In other words, the inverse expansion / conversion conversion unit 25 generates power spectrum envelope sequences W (0), W (1),..., W (N−1) defined by the following equations.

  Here, the transformation matrix V is a matrix that approximates the inverse transformation of the transformation matrix U, and is a matrix that approximates the mapping from the non-linear discretized sample point sequence to the linear discretized sample point sequence. This conversion is to convert a sample point sequence with unequal intervals into a sample point sequence with equal intervals, and to increase or decrease the interval between sample points so as to have a reverse relationship to the above-mentioned “stretch conversion”. For this reason, it will be referred to as “inverse stretch conversion”. However, the transformation matrix V does not mean an inverse matrix of the transformation matrix U.

  The transformation matrix V is, for example, a band matrix (sparse matrix) that has a non-zero value only in the vicinity of the corresponding element by inverse expansion and contraction, and the other components are zero.

  A linear discretization sample point having a linear discretization sample point sequence after inverse stretching transformation (hereinafter referred to as “frequency after inverse stretching”) is a nonlinear discretization sampling point of the nonlinear discretization sample point sequence before inverse stretching transformation. Can be approximated by a weighted sum of the frequencies of one or more nonlinear discretized sampling points having a frequency close to the frequency after inverse stretching. In other words, the frequency after inverse expansion / contraction can be approximated by a weighted sum of the frequencies of one or more nonlinear discretization sample points in the vicinity of the nonlinear discretization sample point having a frequency closest to the inverse expansion / contraction frequency.

  Each row of the transformation matrix V corresponds to each linear discretization sample point of the linear discretization sample point sequence after inverse stretching transformation, and each column of the transformation matrix V corresponds to a nonlinear discretization sample point of the nonlinear discretization sample point sequence . That is, each row of the transformation matrix V is a series of weights with respect to the frequency of each non-linear discretization sample point for expressing the frequency of the linear discretization sample point corresponding to the row.

  From the above property, each row of the transformation matrix V is set to a non-zero value only for the neighboring elements in the column corresponding to the non-linear discretization sample point of the frequency closest to the frequency of the linear discretization sample point corresponding to each row, and the remaining elements Is set to 0, for example. This is because the influence of the non-linear discretization sample point having a frequency different from the frequency of the linear discretization sample point is extremely small and can be regarded as zero.

  The above-mentioned “element corresponding to inverse expansion / contraction” corresponds to a non-linear discretization sample point used to approximate the frequency after inverse expansion / contraction in the element of the row of the transformation matrix V corresponding to the frequency after reverse expansion / contraction. Points to a column element. That is, the above-mentioned “elements corresponding to inverse expansion / contraction” refers to the column corresponding to the nonlinear discretization sample point of the frequency closest to the frequency after inverse expansion / contraction corresponding to each row among the elements of each row of the transformation matrix V. Neighboring elements.

The transformation curve in the transformation matrix V is the point [1,1] corresponding to the upper left component in a two-dimensional plane assuming that the transformation matrix V is an N × M matrix and the row is the horizontal axis and the column is the vertical axis. To a point [N, M] corresponding to the component at the lower right corner. In other words, the transformation matrix V has the following properties.
The centroid g _i ′ of the i-th row of the transformation matrix V is

It is defined as V [i, k] represents the (i, k) element of the transformation matrix V. Assume that the row and column indices start at 1. Then, the transformation matrix center of gravity series g ₁ of each row of the _{V ', g 2', ...} , g M ' is g _1' satisfy the relationship of _{<g 2 '<... <g} M'.
However, the expansion / contraction curve in the conversion matrix V is a sparse matrix in which only components along the expansion / contraction curve (inverse expansion / contraction curve) in the opposite direction to the expansion / contraction curve of the conversion matrix U have non-zero values. In other words, the stretching curve of the transformation matrix U and the stretching curve of the transformation matrix V are points [1,1] corresponding to the upper left component in a two-dimensional plane with the horizontal axis as the row and the vertical axis as the column. The shape is almost line-symmetric with respect to a straight line connecting the points [N, M] corresponding to the component at the lower right end.

  For example, when the transformation matrix U is a sparse matrix in which the number of nonzero components included in the upper triangular matrix is less than the number of nonzero components included in the lower triangular matrix, the inverse transformation matrix V is The sparse matrix is such that the number of non-zero components included in the upper triangular matrix is greater than the number of non-zero components included in the lower triangular matrix. At this time, in each row of the transformation matrix V, the distance between the non-zero element of each row that is farthest from the diagonal element and the diagonal element (this distance is how far the column index is separated) Represents the expansion / contraction distance of the index of each row, the average of the expansion / contraction distance with respect to the front half index of all the indexes tends to be smaller than the average of the expansion / contraction distance with respect to the rear half index. FIG. 18 shows an example of the transformation matrix V. In this example, N = M = 16.

Note that the transformation matrix V may be a non-negative matrix similarly to the transformation matrix U. Also in this case, in the transformation matrix V, the center of gravity of each row satisfies the above-described property g ₁ '<g ₂ '<...<g _M '.

<Envelope normalization unit 26>
The envelope normalization unit 26 includes an MDCT coefficient sequence X (0), X (1),..., X (N-1) converted by the frequency domain conversion unit 21 and a power spectrum envelope sequence converted by the inverse expansion / conversion conversion unit 25. W (0), W (1),..., W (N-1) are input.

The envelope normalization unit 26 uses the power spectrum envelope series W (0), W (1),..., W (N-1) to generate MDCT coefficient sequences X (0), X ( 1),..., X (N-1) are normalized to generate a normalized frequency domain sample sequence (step E6). In this example, the normalized frequency domain sample sequence is a normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1).

  The generated normalized frequency domain sample sequence is output to the encoding unit 27.

For example, the envelope normalization unit 26 sets each coefficient X (i) of the MDCT coefficient sequence X (0), X (1),..., X (N-1) as i = 0, 1,. Is divided by the square root of each envelope value W _γ (i) of the power spectrum envelope series W (0), W (1),..., W (N-1) to obtain a normalized MDCT coefficient sequence X _N (0) , X _N (1),..., X _N (N−1) coefficients X _N (i) are generated. That is, X _N (i) = X (i) / sqrt (W (i)) where i = 0, 1,. Here, sqrt (x) represents the square root of x, where x is a real number.

  The power spectrum envelope sequences W (0), W (1),..., W (N-1) are expanded power spectrum envelope sequences ~ W (0), ~ W (1), ..., ~ W (N- It is derived from 1).

  Therefore, the envelope normalization unit 26 uses the MDCT coefficient sequence X ((frequency domain sample sequence) based on the stretched power spectrum envelope sequences ~ W (0), ~ W (1), ..., ~ W (N-1). It can be said that the normalized frequency domain sample sequence is generated by normalizing 0), X (1),..., X (N-1).

<Encoding unit 27>
The normalized frequency domain sample sequence generated by the envelope normalization unit 26 is input to the encoding unit 27. In this example, the normalized frequency domain sample sequence is a normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1).

  The encoding unit 27 encodes the normalized frequency domain sample sequence and generates a code corresponding to the normalized frequency domain sample sequence (step E7).

  The generated code is output to the decoding device.

  For example, the encoding unit 27 generates a code corresponding to a frequency domain sample sequence normalized as in the conventional case.

That is, each coefficient of the normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N-1) is divided by a gain (global gain) g, and an integer value obtained by quantizing the result by a series quantized normalization haze coefficient sequence _{X Q (0), X Q} (1), ..., a code obtained by coding X _Q (N-1) and the integer signal code. The encoding unit 15 determines a gain g such that the number of bits of the integer signal code is equal to or smaller than the allocated bit number B, which is the number of bits allocated in advance, and as large as possible. Then, the encoding unit 27 generates a gain code corresponding to the determined gain g and an integer signal code corresponding to the determined gain g. In this case, the gain code and the integer signal code are codes corresponding to the normalized frequency domain sample sequence.

[Example of how to find transformation matrix U and inverse transformation matrix V]
A transformation matrix U that implements a linear transformation from a linear discretization sample point sequence to a nonlinear discretization sample point sequence, and an inverse transformation matrix V that implements a linear transformation from a nonlinear discretization sample point sequence to a linear discretization sample point sequence. Can be obtained, for example, by learning in advance. Here, the U and V elements are all non-negative. This is because the vectors before and after the conversion are power spectra or their envelopes, and are all positive values.

  Using linear transformation, the power spectrum sequence Y (0), Y (1), ..., Y (N-1) can be expanded and contracted pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N The method of obtaining -1) has an advantage that unintended conversion can be prevented as compared with the case of obtaining a stretchable pseudo power spectrum by interpolating with a sinc function or the like.

  First, as learning data, a power spectrum sequence (Ym (0),..., Ym (N-1)) (m = 1) expressed based on a linear discrete sample point sequence consisting of M linear discrete sample points , 2, ..., M), and a stretched pseudo power spectrum sequence (~ Ym (0), ...,) expressed based on a non-linear discretized sample point sequence obtained by interpolating this with a sinc function or the like A set ~ Y of ~ Ym (N-1)) (m = 1,2, ..., M) is prepared.

  Then, the conversion matrix U is learned by updating each component of U so that the distance between UY and ~ Y is minimized using a band matrix U in which appropriate initial values are set in advance.

As the distance, since the Itakura Saito distance between the square value of the envelope of the linear prediction coefficient and the original power spectrum is minimized, the Itakura Saito distance may be used as a distance measure during learning. That is, in learning of the expansion / contraction transformation matrix, learning is performed using a function defined by the following equation as an objective function. D _IS (a | b) represents the Itakura Saito distance from b based on a.

  The inverse transformation matrix V may be learned so that the Itakura Saito distance between Y and VUY is minimized, that is, a function defined by the following equation is used as an objective function.

  This optimization problem can be solved by an auxiliary function method, for example. For example, until the predetermined condition is satisfied, the U and V elements can be updated by the following update formula to approximate the optimal solution.

Here, U _{i, j} ^(p) and V _{i, j} ^(p) are respectively the (i, j) component of the transformation matrix V and the (i, j) of the inverse transformation matrix V obtained by the p-th iteration. Represents an ingredient. Y _{j, k} = Y _k (j), that is, the j-th element in the k-th pseudo power spectrum sequence of the learning data Y is represented. Similarly, ~ Y _{i, k} = ~ Y _k (i), which represents the i-th element in the k-th pseudo power spectrum sequence of learning data ~ Y.

  From this update formula, elements with an initial value of 0 in the transformation matrix U and inverse transformation matrix V remain 0 after learning, so there is no need to calculate and optimization within this constraint is possible. It is. Also, since U and V are banded matrices, the amount of calculation for actual conversion can be greatly reduced by the conversion in which the positions of non-zero samples are specified in advance, and many elements that are zero components in learning can also be obtained. Since learning is not necessary, learning can be performed at low cost. In addition, the amount of conversion and learning can be further adjusted by adjusting the bandwidth when setting the initial value.

(Decoding of the first embodiment)
A configuration example of a decoding apparatus corresponding to the encoding apparatus of the first embodiment is shown in FIG. As shown in FIG. 6, the decoding device according to the first embodiment includes an expansion / contraction linear prediction coefficient decoding unit 31, an expansion / contraction power spectrum envelope sequence generation unit 32, an inverse expansion / conversion conversion unit 33, a decoding unit 34, and an envelope inverse normality. For example, a conversion unit 35 and a time domain conversion unit 36 are provided. An example of each process of the decoding method of the first embodiment realized by this decoding apparatus is shown in FIG.

  In the decoding device, MDCT coefficients are reconstructed by processing in the reverse order to the encoding processing by the encoding device.

The decoding device receives at least the code corresponding to the normalized frequency domain sample sequence and the stretched linear prediction coefficient code output from the encoding device. Hereinafter, as codes corresponding to normalized frequency domain sample sequences, codes corresponding to normalized MDCT coefficient sequences X _N (0), X _N (1),..., X _N (N-1) are input. A case will be described as an example.

  Hereinafter, each part of FIG. 6 will be described.

<Extensible Linear Prediction Coefficient Decoding Unit 31>
The stretchable linear prediction coefficient decoding unit 31 receives the stretchable linear prediction coefficient code output from the encoding device.

The expansion / contraction linear prediction coefficient decoding unit 31 decodes the input expansion / contraction linear prediction coefficient code by, for example, a conventional decoding technique for each frame to quantize the expansion / contraction linear prediction coefficients ^ β ₁ , ^ β ₂ ,. _p is generated (step D1).

The generated quantized stretch linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p are output to the stretch power spectrum envelope sequence generation unit 32.

  Here, the conventional decoding technique is to obtain a quantized linear prediction coefficient by decoding the linear prediction coefficient code when the linear prediction coefficient code is a code corresponding to the quantized linear prediction coefficient, for example. A technique for obtaining a quantized LSP parameter by decoding a linear prediction coefficient code when the linear prediction coefficient code is a code corresponding to a quantized LSP parameter. In addition, the quantized linear prediction coefficient and the quantized LSP parameter can be converted to each other, and the conversion process may be performed according to the input prediction coefficient code and information necessary for the subsequent processing. Is well known. From the above, what includes the decoding process of the linear prediction coefficient code and the conversion process performed as necessary is “decoding by a conventional decoding technique”. Here, the input linear prediction coefficient code is a stretched linear prediction coefficient code, but the process is the same as the conventional decoding process.

<Elastic Power Spectrum Envelope Sequence Generation Unit 32>
The expansion / contraction power spectrum envelope sequence generation unit 32 receives the quantized expansion / contraction linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p generated by the expansion / contraction linear prediction coefficient decoding unit 31.

The expansion / contraction power spectrum envelope sequence generation unit 32 uses the quantized expansion / contraction linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p and performs the same processing as the expansion power spectrum envelope sequence generation unit 24 of the encoder. Then, the expansion / contraction power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W (N-1) expressed based on the nonlinear discretized sample point sequence is generated (step D2).

  The generated expansion / contraction power spectrum envelope sequences ~ W (0), ~ W (1), ..., ~ W (N-1) are output to the inverse expansion / conversion conversion unit 33.

<Reverse expansion / contraction conversion unit 33>
The inverse expansion / contraction conversion unit 33 receives the expansion / contraction power spectrum envelope sequences ~ W (0), ~ W (1), ..., ~ W (N-1) generated by the expansion / contraction power spectrum envelope series generation unit 32.

  The inverse expansion / conversion conversion unit 33 performs the same processing as the inverse expansion / contraction conversion unit 25 of the encoding device, and the expansion / contraction power spectrum envelope sequence ~ W (0), ~ W (1) expressed based on the non-linear discretization sample point sequence , ..., ~ W (N-1) are converted into power spectrum envelope sequences W (0), W (1), ..., W (N-1) expressed based on the linear discretized sample point sequence (steps) D3).

  The converted power spectrum envelope sequences W (0), W (1),..., W (N−1) are output to the envelope denormalization unit 35.

<Decoding unit 34>
A code corresponding to the normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1) output from the encoding device is input to the decoding unit 34.

The decoding unit 34 decodes a code corresponding to the input normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1) for each frame to normalize the MDCT coefficient. Columns X _N (0), X _N (1),..., X _N (N−1) are generated (step D4).

The generated normalized MDCT coefficient sequences X _N (0), X _N (1),..., X _N (N−1) are output to the envelope denormalization unit 35.

  For example, when Rice coding is used in the encoding device, the decoding unit 34 decodes the code by a decoding process corresponding to Rice coding.

When a gain code and an integer signal code are input as codes corresponding to the normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1), the decoding unit 34 By multiplying the quantized normalized coefficient sequence X _Q (0), X _Q (1), ..., X _Q (N-1) obtained by decoding the integer signal code by the gain specified by the gain code A normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1) is generated.

<Envelope inverse normalization unit 35>
The envelope denormalization unit 35 includes power spectrum envelope sequences W (0), W (1),..., W (N-1) converted by the inverse expansion / conversion conversion unit 33, and normalized MDCT coefficients generated by the decoding unit 34. The columns X _N (0), X _N (1), ..., X _N (N-1) are input.

The envelope inverse normalization unit 35 uses a power spectrum envelope sequence W (0), W (1),..., W (N-1) to normalize a MDCT coefficient sequence that is a sample sequence in the frequency domain. By denormalizing X _N (0), X _N (1), ..., X _N (N-1), MDCT coefficient sequence X (0), X (1), ..., X (N-1) Is generated (step D5).

  The generated MDCT coefficient sequences X (0), X (1),..., X (N−1) are output to the time domain conversion unit 36.

For example, the envelope inverse normalization unit 35, i = 0, 1, ..., a N-1, normalized MDCT coefficients _{X N (0), X N} (1), ..., X N of (N-1) MDCT coefficient sequence by multiplying each coefficient by the square root of each envelope value W _γ (i) of the power spectrum envelope sequence W (0), W (1), ..., W (N-1) to X _N (i) Generate X (0), X (1), ..., X (N-1). That is, as i = 0, 1,..., N−1, X (i) = X _N (i) * sqrt (W (i)). Here, sqrt (x) represents the square root of x, where x is a real number.

  The power spectrum envelope sequences W (0), W (1),..., W (N-1) are expanded power spectrum envelope sequences ~ W (0), ~ W (1), ..., ~ W (N- It is derived from 1).

Therefore, the envelope denormalization unit 35 normalizes MDCT coefficients which are frequency domain sample sequences based on the stretched power spectrum envelope sequences ~ W (0), ~ W (1), ..., ~ W (N-1). It can be said that the frequency domain sample sequence is generated by denormalizing the sequences X _N (0), X _N (1),..., X _N (N−1).

<Time domain conversion unit 36>
MDCT coefficient sequences X (0), X (1),..., X (N−1) generated by the envelope denormalization unit 35 are input to the time domain conversion unit 36.

  For each frame, the time domain conversion unit 36 converts the “MDCT coefficient sequence” obtained by the envelope denormalization unit into the time domain to obtain a sound signal (decoded sound signal) in units of frames (step D6).

[Second Embodiment]
(Encoding of the second embodiment)
The configuration example of the encoding device of the second embodiment is the same as the configuration example of the encoding device of the first embodiment shown in FIG.

  Hereinafter, a description will be given centering on differences from the first embodiment. Description of the same parts as those in the first embodiment is omitted.

  The encoding apparatus of the second embodiment differs in the power spectrum envelope sequence used when normalizing the MDCT coefficient sequence X (0), X (1),..., X (N−1). That is, the smoothing method for smoothing the expansion / contraction power spectrum envelope corresponding to the non-linear discretized sample point sequence is different. In other words, the generation method of the expansion / contraction power spectrum envelope series ~ W (0), ~ W (1), ..., ~ W (N-1) by the extension / smoothing power spectrum envelope series generation unit 24 is different.

In the encoding device according to the first embodiment, the stretched power spectrum envelope sequence ~ _Wo (0), ~ _Wo (1), ..., ~ _Wo (N-1) is an example of the stretched and smoothed power spectrum envelope series. ~ W _γ (0), ~ W _γ (1), ..., ~ W _γ (N-1) are the same as the conventional method of smoothing the power spectrum envelope sequence corresponding to the linear discretized sample point sequence, It is generated by smoothing a stretched pseudo power spectrum envelope sequence corresponding to a non-linear discretized sample point sequence.

  When the expansion / contraction power spectrum envelope sequence corresponding to the non-linear discretization sample point sequence generated in this way is inversely stretched and converted to the power spectrum envelope sequence corresponding to the linear discretization sample point sequence, the resolution in the linear discretization sample point sequence is calculated. In the high frequency region, the effect of smoothing applied to the stretched power spectrum envelope sequence is offset, and a large peak shape remains when inverse stretch transform is performed to the power spectrum envelope sequence corresponding to the linear discretized sample point sequence. May end up. As a result, a normalized MDCT coefficient sequence is obtained and encoded using a power spectrum envelope sequence in which a large peak shape remains, so that a sufficient smoothing effect cannot be obtained, resulting in a decrease in encoding efficiency. May end up.

  For this reason, the expansion / contraction power spectrum envelope sequence generation unit 24 of the second embodiment smoothes the power spectrum envelope in which the power spectrum envelope sequence after inverse expansion / conversion conversion is expressed with a linear discrete sampled point sequence with a uniform resolution. The smoothing effect is corrected in accordance with the degree of expansion / contraction g (k) of the non-linear discretized sample point sequence so as to approximate the smoothing power spectrum at that time (conventional smoothing power spectrum).

Specifically, the stretched and smoothed power spectrum envelope sequence generation unit 24 of the second embodiment performs stretched power spectrum envelope sequences ~ _Wo (0), ~ _Wo (1), ..., ~ _Wo (N-1 ), The expansion / contraction power spectrum envelope sequence ~ _Wγ (0), ~ _Wγ (1), ..., ~ _Wγ (N-1) defined by the equation (3 ') is generated (step E4). The generated stretched smoothed power spectrum envelope sequence ~ _Wγ (0), ~ _Wγ (1), ..., ~ _Wγ (N-1) is the stretched power spectrum envelope series ~ _Wo (0), ~ W (1) _o, ..., it is output to the inverse scale transformation unit 25 as _{~ W o (N-1)} .

The stretched and smoothed power spectrum envelope sequence ~ W _γ (0), ~ W _γ (1), ..., ~ W _γ (N-1) defined by equation (3 ') is expressed as g (k) (k = 0 , 1,..., N−1), the stretched and smoothed power spectrum envelope sequence defined by equation (3 ′) ~ W _γ (0), ~ W _γ (1), ..., ~ W _γ (N-1) is the corrected stretched smoothed power spectrum envelope sequence ~ W _γ (0), ~ W _γ (1), ..., ~ W _γ (N -1) Also called.

Comparing equation (3) and equation (3 ′), the difference is that the correction coefficient γ ⁿ is multiplied by g (k). g (k) is a value corresponding to the degree of expansion and contraction of the nonlinear discretized sample point sequence from the linear discretized sample point sequence, and is defined as follows, for example.

  f (k) represents the relative frequency position in the linear discretization sample point sequence of the frequency of the sample point corresponding to the kth index in the non-linear discretization sample point sequence. Therefore, the expression (3 ′) means that the correction value γ is equivalently reduced at the frequency at which the interval is shortened in the inverse expansion / contraction conversion. If the matrix that realizes the inverse transformation from the nonlinear discrete sample point sequence to the linear discrete sample point sequence is the inverse transformation matrix V, f (k) (k = 0,1, ..., N-1) is For example, it is expressed as follows.

  The inverse transformation matrix V can be expressed by a band matrix like the transformation matrix U. Similarly to the transformation matrix U, the inverse transformation matrix V can be obtained in advance by learning or the like from the correlation between the non-linear discretized sample point sequence and the linear discretized sample point sequence. The method of obtaining is the same as that described in the first embodiment. Note that since the nonlinear conversion of the frequency of a discrete signal is an irreversible operation, V does not necessarily have an inverse matrix relationship of U.

(Decoding of the second embodiment)
The configuration example of the decoding device of the second embodiment is the same as the configuration example of the decoding device of the first embodiment shown in FIG.

In the decoding apparatus according to the second embodiment, the expansion / contraction power spectrum envelope sequence generation unit 32 performs the expansion / contraction power spectrum envelope sequence ~ _Wo (0), ~ _Wo (1), ..., ~ _Wo (N- As 1), the corrected stretched and smoothed power spectrum envelope sequence ~ W _γ (0), ~ W _γ (1), ..., ~ W _γ (N-1) defined by Equation (3 ') is generated. This part is different from the decoding device of the first embodiment.

  The decoding device of the second embodiment is the same as the decoding device of the first embodiment with respect to other parts.

[Third embodiment]
(Encoding of the third embodiment)
An example of the configuration of the encoding apparatus according to the third embodiment is shown in FIG. As shown in FIG. 8, the encoding device according to the third embodiment includes a frequency domain transform unit 21, a stretched pseudo power spectrum sequence generation unit 22, a linear prediction analysis unit 23, and a stretch smoothing power spectrum envelope sequence generation unit. 213, a second inverse expansion / conversion conversion unit 28, an expansion / contraction unsmoothed power spectrum envelope series generation unit 29, a first inverse expansion / conversion conversion unit 210, an envelope normalization unit 26, and an encoding unit 27, for example. Yes. An example of each process of the encoding method according to the third embodiment realized by this encoding apparatus is shown in FIG.

  Hereinafter, a description will be given centering on differences from the first embodiment and the second embodiment. Description of the same parts as those in the first embodiment and the second embodiment is omitted.

  In the encoding device of the first embodiment and the second embodiment, when the MDCT coefficient sequence X (0), X (1),..., X (N-1) is encoded, the smoothed power spectrum envelope Only the information of was used. In contrast, the encoding device of the third embodiment uses the square root of the smoothed power spectrum envelope to divide the MDCT coefficient when encoding the MDCT coefficient sequence, and in addition to that, The ratio information for each sample of the power spectrum envelope in the frequency domain and the smoothed power spectrum envelope for each sample is used as auxiliary information for estimating the amplitude of the MDCT coefficient to be quantized.

<Extension / Smoothing Power Spectrum Envelope Sequence Generation Unit 213>
The stretched smoothed power spectrum envelope sequence generation unit 213 receives the quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p generated by the linear prediction analysis unit 23.

The stretched / smoothed power spectrum envelope sequence generation unit 213 performs the same process as the stretched / smoothed power spectrum envelope sequence generation unit 24 of the first embodiment or the second embodiment, thereby performing a line quantization stretched linear prediction coefficient ^ β ₁ , ^ Using β ₂ ,…, ^ β _p , the stretched and smoothed power spectrum envelope sequence ~ W _γ (0), ~ W _γ (1),…, ~ W _γ (N- 1) is generated (step E4).

The generated stretched and smoothed power spectrum envelope sequences ~ _Wγ (0), ~ _Wγ (1), ..., ~ _Wγ (N-1) are output to the second inverse stretch transform unit 28.

The stretched and smoothed power spectrum envelope sequence ~ W _γ (0), ~ W _γ (1), ..., ~ W _γ (N-1) is expressed by Expression (3) or Expression (3 ').

<Second Reverse Stretch Conversion Unit 28>
The second inverse expansion / conversion conversion unit 28 includes the expansion / contraction smoothed power spectrum envelope sequence generated by the expansion / contraction smoothed power spectrum envelope generation unit 24 ~ W _γ (0), ~ W _γ (1), ..., W _γ ( N-1) is input.

Then, the second inverse expansion / conversion conversion unit 28 performs the same processing as the inverse expansion / contraction conversion unit 25 of the first embodiment and the second embodiment, and the expansion / contraction smoothed power spectrum envelope sequence ~ W corresponding to the non-linear discretized sample point sequence. _γ (0), ~ W _γ (1), ..., ~ W _γ (N-1) are smoothed power spectrum envelope sequences W _γ (0), W _γ (1) corresponding to the linear discretized sample point sequence ,..., W _γ (N−1) (step E8).

The converted smoothed power spectrum envelope sequences W _γ (0), W _γ (1),..., W _γ (N−1) are output to the envelope normalization unit 26 and the encoding unit 27.

<Expandable Unsmoothed Power Spectrum Envelope Sequence Generation Unit 29>
The stretched non-smoothed power spectrum envelope sequence generation unit 29 receives the quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p generated by the linear prediction analysis unit 23.

Stretch textured power spectrum envelope sequence generating unit 29, the quantization telescopic linear prediction coefficient _{^ β 1, ^ β 2,} ..., ^ β with _p, before smoothing is defined by equation (2) stretching the non Smoothed power spectrum envelope sequences ~ _Wo (0), ~ _Wo (1), ..., _Wo (N-1) are generated (step E9).

The generated stretch / unsmoothed power spectrum envelope sequences ~ _Wo (0), ~ _Wo (1), ..., _Wo (N-1) are output to the first inverse stretch transform unit 210.

<First Reverse Stretch Conversion Unit 210>
The first inverse expansion / conversion conversion unit 210 includes expansion / contraction non-smoothed power spectrum envelope sequence generated by the expansion / contraction non-smoothed power spectrum envelope sequence generation unit 29 ~ W _o (0), ~ W _o (1), ..., W _o. (N-1) is input.

The first inverse expansion / conversion transform unit 210 converts the expansion / de-smoothing power spectrum envelope sequence corresponding to the non-linear discretized sample point sequence to W _o (0), to W _o (1), ..., W _o (N-1). Based on this, a non-smoothed power spectrum envelope sequence W (0), W (1),..., W (N-1) corresponding to the linear discretized sample point sequence is generated (step E10).

The generated non-smoothed power spectrum envelope sequences W _o (0), W _o (1),..., W _o (N−1) are output to the encoding unit 27.

The first inverse expansion / conversion conversion unit 210 performs non-smoothed power spectrum envelope sequences W _o (0), W _o (1),... By interpolation or linear conversion in the same manner as the inverse expansion / contraction conversion unit 25 of the first embodiment. Generate W _o (N-1).

When the non-smoothed power spectrum envelope sequence W _o (0), W _o (1),..., W _o (N−1) is generated by linear transformation, the first inverse expansion / conversion transformation unit 210 performs, for example, the first implementation. Using the inverse transformation matrix V described in the embodiment, non-smoothed power spectrum envelope sequences W _o (0), W _o (1),..., W _o (N−1) are generated by the following equations.

When generating non-smoothed power spectrum envelope sequences W _o (0), W _o (1),..., W _o (N−1) by interpolation, the first inverse expansion / conversion conversion unit 210 performs, for example, the following processing. Similar to the expansion / contraction pseudo power spectrum sequence generation unit 22, the inverse expansion / conversion conversion unit 25 interpolates the expansion / contraction power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W (N-1) by the sinc function. Obtained curve (envelope that smoothly connects the expansion and contraction power spectrum envelope). Then, a series of power spectrum envelope values of frequencies corresponding to the discrete sample points of the linear discretized sample point sequence on the curve are converted into non-smoothed power spectrum envelope sequences W _o (0), W _o (1),. , W _o (N-1).

<Envelope normalization unit 26>
The envelope normalization unit 26 includes the MDCT coefficient sequence X (0), X (1),..., X (N-1) converted by the frequency domain conversion unit 21 and the smoothing converted by the second inverse expansion / conversion conversion unit 28. Power spectrum envelope sequences W _γ (0), W _γ (1),..., W _γ (N−1) are input.

The envelope normalization unit 26 performs smoothing power spectrum envelope sequences W _γ (0), W _γ (1),..., W _γ (in the same manner as the envelope normalization unit 26 of the first and second embodiments. Frequency domain sample sequence normalized by normalizing MDCT coefficient sequence X (0), X (1), ..., X (N-1), which is a frequency domain sample sequence based on (N-1) Is generated (step E6). In this example, the normalized frequency domain sample sequence is a normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1). For example, assuming that i = 0, 1,..., N−1, X _N (i) = X (i) / sqrt (W _γ (i)).

  The generated normalized frequency domain sample sequence is output to the encoding unit 27.

<Encoding unit 27>
The encoding unit 27 includes the normalized frequency domain sample sequence generated by the envelope normalization unit 26 and the smoothed power spectrum envelope sequences W _γ (0), W _γ ( 1), ..., W _γ (N-1) and the non-smoothed power spectrum envelope sequence W _o (0), W _o (1), ..., W _o (N- 1) is entered. In this example, the normalized frequency domain sample sequence is a normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1).

  The encoding unit 27 encodes the normalized frequency domain sample sequence and generates a code corresponding to the normalized frequency domain sample sequence (step E7).

  The generated code is output to the decoding device.

The encoding unit 27 first divides each coefficient of the normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1) by a gain (global gain) g, and the result The code obtained by encoding the quantized normalized coefficient sequence X _Q (0), X _Q (1), ..., X _Q (N-1), which is a sequence of integer values obtained by quantizing To do. Quantized normalized coefficient series X _Q (0), X _Q (1),..., X _Q (N−1) are, for example, variable length encoded. The encoding unit 27 determines a gain g such that the number of bits of the integer signal code is equal to or smaller than the allocated bit number B, which is the number of bits allocated in advance, and as large as possible. Then, the encoding unit 27 generates a gain code corresponding to the determined gain g and an integer signal code corresponding to the determined gain g. In this case, the gain code and the integer signal code are codes corresponding to the normalized frequency domain sample sequence.

The encoding unit 27 is a value before smoothing when the quantized normalized coefficient series X _Q (0), X _Q (1),..., X _Q (N−1) is variable length encoded. Changing the information distribution of variable length coding according to the estimated value of the amplitude of the MDCT coefficient to be quantized obtained from the unsmoothed power spectrum envelope sequence and the smoothed power spectrum envelope sequence that is the value after smoothing In order to improve the compression efficiency.

Here, the amplitude of the MDCT coefficient X _Q (k) to be quantized is the square root of the value obtained by dividing the unsmoothed power spectrum envelope coefficient W _o (k) by the smoothed power spectrum envelope coefficient W _γ (k). Can be estimated. This is used as an estimated value of the amplitude of the MDCT coefficient X _Q (k) to be quantized.

For example, when using Rice coding as the variable-length coding method, a value proportional to the logarithm of the amplitude estimate of the MDCT coefficient X _Q (k) to be quantized is expressed as a quantized normalized coefficient sequence X _Q ( Rice coding is performed using k) as a rice parameter when performing rice coding. That is, based on the ratio of the unsmoothed frequency-domain power spectrum envelope to the smoothed power spectrum envelope for each sample, the Rice parameter for the Rician coding of the quantized normalized coefficient sequence X _Q (k) Select.

Instead of switching the rice parameter for each sample k, encoding may be performed by switching the rice parameter for each of a plurality of samples. In this case, for example, a value proportional to the logarithm of the average value of the amplitude estimation values of the MDCT coefficients X _Q (k) to be quantized of a plurality of samples is used as the quantized normalized coefficient series X _Q (k). Used as a rice parameter when encoding.

(Decoding of the third embodiment)
A configuration example of the decoding device according to the third embodiment is shown in FIG. As shown in FIG. 10, the decoding device of the third embodiment includes an expansion / contraction linear prediction coefficient decoding unit 31, an expansion / smoothing power spectrum envelope sequence generation unit 312, a second inverse expansion / conversion conversion unit 39, and expansion / contraction non-smoothing. For example, a power spectrum envelope series generation unit 37, a first inverse expansion / conversion conversion unit 38, a decoding unit 34, an envelope inverse normalization unit 35, and a time domain conversion unit 36 are provided. An example of each process of the decoding method according to the third embodiment realized by this decoding apparatus is shown in FIG.

  The following description will focus on the parts different from the decoding / decoding device of the first embodiment, and the description of the same parts as in the first embodiment will be omitted.

<Expandable Unsmoothed Power Spectrum Envelope Sequence Generation Unit 37>
The stretched non-smoothed power spectrum envelope sequence generation unit 37 receives the quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p generated by the stretched linear prediction coefficient decoding unit 31.

The stretched non-smoothed power spectrum envelope sequence generation unit 37 uses the quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p and uses the stretched non-smoothed power spectrum defined by equation (2). Envelope sequences ~ _Wo (0), ~ _Wo (1), ..., _Wo (N-1) are generated (step D7).

The generated stretch / unsmoothed power spectrum envelope sequences ~ _Wo (0), ~ _Wo (1), ..., _Wo (N-1) are output to the first inverse stretch transform unit 38.

<First Reverse Stretch Conversion Unit 38>
The first inverse expansion / conversion conversion unit 38 includes expansion / contraction non-smoothed power spectrum envelope sequence generated by the expansion / de-smoothing power spectrum envelope sequence generation unit 37 ~ W _o (0), ~ W _o (1), ..., W _o. (N-1) is input.

The first inverse expansion / conversion conversion unit 38 performs the same processing as the first inverse expansion / contraction conversion unit 210 of the encoding device according to the third embodiment, and the expansion / contraction non-smoothing power spectrum envelope sequence corresponding to the nonlinear discretized sample point sequence ~ W _{Based on o} (0), ~ W _o (1), ..., W _o (N-1), the unsmoothed power spectrum envelope sequence W _o (0), W _o ( 1),..., W _o (N−1) are generated (step D8).

The generated non-smoothed power spectrum envelope sequences W _o (0), W _o (1),..., W _o (N−1) are output to the decoding unit.

<Extension / Smoothing Power Spectrum Envelope Sequence Generation Unit 312>
The expansion / contraction smoothed power spectrum envelope sequence generation unit 312 receives the quantized expansion / contraction linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p generated by the expansion / contraction linear prediction coefficient decoding unit 31.

The stretched / smoothed power spectrum envelope sequence generating unit 32 uses the quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p and the stretched / smoothed power spectrum envelope sequence generating unit 24 of the encoder. The same processing generates the stretched and smoothed power spectrum envelope sequence ~ W _γ (0), ~ W _γ (1), ..., ~ W _γ (N-1) expressed based on the nonlinear discretized sample point sequence (Step D2).

The generated stretched and smoothed power spectrum envelope sequences ~ _Wγ (0), ~ _Wγ (1), ..., ~ _Wγ (N-1) are output to the inverse stretch conversion unit 33.

The stretched and smoothed power spectrum envelope sequence ~ W _γ (0), ~ W _γ (1), ..., ~ W _γ (N-1) is expressed by Expression (3) or Expression (3 ').

<Second Reverse Stretch Conversion Unit 39>
The second reverse expansion / conversion conversion unit 39 is the same as the reverse expansion / conversion conversion unit 25 of the first embodiment and the second embodiment and the second reverse expansion / contraction conversion unit 28 of the third embodiment.

The second inverse expansion / conversion conversion unit 39 includes the expansion / contraction smoothed power spectrum envelope sequence generated by the expansion / contraction smoothed power spectrum envelope generation unit 32 ~ _Wγ (0), ~ _Wγ (1), ..., ~ _Wγ ( N-1) is input.

The second inverse expansion / conversion conversion unit 39 corresponds to the non-linear discretization sample point sequence in the same manner as the inverse expansion / contraction conversion unit 25 of the first embodiment and the second embodiment and the second inverse expansion / contraction conversion unit 28 of the third embodiment. The smoothed power spectrum envelope sequence ~ W _γ (0), ~ W _γ (1), ..., ~ W _γ (N-1) to be smoothed power spectrum envelope series W corresponding to the linear discretized sample point sequence Convert to _γ (0), W _γ (1),..., W _γ (N−1) (step D9).

The converted smoothed power spectrum envelope sequences W _γ (0), W _γ (1),..., W _γ (N−1) are output to the decoding unit 34 and the envelope denormalization unit 35.

<Decoding unit 34>
The decoding unit 34 includes a code corresponding to the normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1) output from the encoding device, and a first inverse expansion / conversion conversion unit. , W _o (0), W _o (1),..., W _o (N−1) converted by 38 and the smoothed power spectrum envelope sequence converted by the second inverse expansion / conversion transform unit 39. W _γ (0), W _γ (1),..., W _γ (N−1) are input.

The decoding unit 34 decodes a code corresponding to the input normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1) for each frame to normalize the MDCT coefficient. Columns X _N (0), X _N (1),..., X _N (N−1) are generated (step D4).

The generated normalized MDCT coefficient sequences X _N (0), X _N (1),..., X _N (N−1) are output to the envelope denormalization unit 54.

When a gain code and an integer signal code are input as codes corresponding to the normalized MDCT coefficient sequence X _N (0), X _N (1),..., X _N (N−1), the decoding unit 34 First, the integer signal code is decoded by a decoding process corresponding to the encoding process used when the encoding unit 27 of the encoding device obtains the integer signal code, and a “normalized MDCT coefficient sequence normalized by gain” is obtained. . For example, when Rice coding is used in the encoding device, the decoding unit 34 decodes the integer signal code by a decoding process corresponding to Rice coding.

At this time, the decoding unit 34 selects a rice parameter based on the same criteria as the encoding unit 27 of the encoding device of the third embodiment. That is, rice decoding is performed using a value proportional to the logarithm of the amplitude estimation value of the MDCT coefficient X _Q (k) to be quantized as the rice parameter of the quantized normalized coefficient series X _Q (k). When obtaining the estimated value of the amplitude of the MDCT coefficient X _Q (k) to be quantized, similarly to the encoding unit 27 of the encoding device, the non-smooth generated by the first inverse expansion / conversion conversion unit 38 of the decoding device. The normalized power spectrum envelope coefficient W _o (k) is estimated by the square root of the value divided by the smoothed power spectrum envelope coefficient W _γ (k) generated by the second inverse expansion / conversion transform unit of the decoding device.

Next, the decoding unit 34 multiplies the obtained “normalized MDCT coefficient sequence normalized by gain” by the gain specified by the gain code, thereby obtaining a normalized MDCT coefficient sequence X _N (0), X _N (1), ..., X _N (N-1) is generated.

  Since the subsequent processing by the decoding apparatus, that is, the processing of the envelope denormalization unit 35 and the time domain conversion unit 36 is the same as that of the first embodiment and the second embodiment, description thereof will be omitted.

[Fourth embodiment]
(Encoding of the fourth embodiment)
An example of the configuration of the encoding apparatus according to the fourth embodiment is shown in FIG. As shown in FIG. 12, the encoding apparatus according to the fourth embodiment includes a frequency domain transform unit 21, a stretched pseudo power spectrum sequence generation unit 22, a linear prediction analysis unit 23, and a stretched unsmoothed power spectrum envelope sequence generation. For example, a unit 29, a first inverse expansion / conversion conversion unit 210, a high-order linear prediction analysis unit 211, a smoothed power spectrum envelope sequence generation unit 212, an envelope normalization unit 26, and an encoding unit 27. . An example of each process of the encoding method of the fourth embodiment realized by this encoding apparatus is shown in FIG.

  The encoding apparatus according to the second embodiment performs expansion / contraction smoothing corrected by a value g (k) corresponding to the degree of expansion / contraction from the linear discretization sample point sequence of the nonlinear discretization sample point sequence as the expansion / contraction power spectrum envelope sequence. A power spectrum envelope sequence was generated.

  In the fourth embodiment, after obtaining a power spectrum envelope corresponding to a linear discretized sample point sequence as an alternative method of correction of the second embodiment, higher-order linear prediction analysis is performed only for a specific partial frequency region. By obtaining a fine power spectrum envelope and smoothing it, an effect equivalent to that of the second embodiment is obtained.

  Hereinafter, a description will be given centering on differences from the third embodiment. Description of the same parts as those of the third embodiment is omitted.

  In the following, an example is given in which the same effect as that obtained by smoothing the expansion / contraction pseudo power spectrum envelope corresponding to the non-linear discretized sample point sequence in which the resolution is higher in the low frequency region than in the high frequency region is given as an example. I will explain.

  The processes of the frequency domain conversion unit 21, the expansion / contraction pseudo power spectrum sequence generation unit 22, the linear prediction analysis unit 23, the expansion / contraction unsmoothed power spectrum envelope sequence generation unit 29, and the first inverse expansion / conversion conversion unit 210 are the same as in the third embodiment. Therefore, the description thereof is omitted.

<Higher-order linear prediction analysis unit 211>
The higher-order linear prediction analysis unit 211 includes the non-smoothed power spectrum envelope sequence W _o (0), W _o (1) _,. A power spectrum envelope sequence corresponding to a predetermined frequency region for which the resolution is to be increased is input. When the predetermined frequency region whose resolution is to be increased is the low frequency region, the high-order linear prediction analysis unit 211 receives the unsmoothed power spectrum envelope sequence W _o (0), W _o (1),. _o (I _L ) is entered. _IL is an integer of 1 to N-2.

The higher-order linear prediction analysis unit 211 performs unsmoothed power corresponding to the low frequency region of the unsmoothed power spectrum envelope sequence W _o (0), W _o (1),..., W _o (N−1). Using the spectral envelope sequence W _o (0), W _o (1),..., W _o (I _L ), a higher-order linear prediction analysis, that is, a prediction order p higher than the prediction order p of the linear prediction analysis unit 23. The linear prediction coefficients ˜α ₁ , ˜α ₂ ,..., ˜α _{p ′} are generated by performing the linear prediction analysis at “Step E11”. p ′ is an integer larger than p. That is, p ′> p.

The generated linear prediction coefficients ˜α ₁ , ˜α ₂ ,..., ˜α _{p ′} are output to the smoothed power spectrum envelope sequence generation unit 212.

Note that the linear prediction coefficients ˜α ₁ , ˜α ₂ ,..., ˜α _{p ′} are used only for smoothing the spectral envelope in the low frequency region. The linear prediction coefficients ˜α ₁ , ˜α ₂ ,..., ˜α _{p ′} do not have to be quantized as parameters and output to the decoding device.

In this example, since it is premised on nonlinear stretch transformation that increases the resolution in the low frequency region, the high-order linear prediction analysis unit 211 includes the non-smoothed power spectrum envelope sequence corresponding to the linear discretized sample point sequence. High-order linear prediction analysis is performed using non-smoothed power spectrum envelope sequences W _o (0), W _o (1), ..., W _o (I _L ) in the low frequency region.

If an effect equivalent to a nonlinear expansion / contraction transformation that increases the resolution in the intermediate frequency domain is obtained, the higher-order linear prediction analysis unit 211 performs the non-smoothed power spectrum envelope corresponding to the linear discretized sample point sequence. Unsmoothed power spectrum envelope sequence W _o (I _ML ), W _o (I _ML +1),…, W _o (I _MH −1), W _o (I _MH ) Higher-order linear prediction analysis may be performed using (1 <I _ML <I _MH <N-2).

  In short, the high-order linear prediction analysis unit 211 performs high-order linear prediction analysis using a non-smoothed power spectrum envelope sequence corresponding to a predetermined frequency region of the non-smoothed power spectrum envelope sequence to be improved in resolution. Just do it.

<Smoothing power spectrum envelope sequence generation unit 212>
The smoothed power spectrum envelope sequence generation unit 212 converts the linear prediction coefficients ~ α ₁ , ~ α ₂ , ..., ~ α _{p '} generated by the high-order linear prediction analysis unit 211 and the first inverse expansion / contraction conversion unit 210. Non-smoothing corresponding to a frequency region other than a predetermined frequency region in which the resolution is desired to be increased in the non-smoothed power spectrum envelope sequence W _o (0), W _o (1), ..., W _o (N-1) A power spectrum envelope sequence is input. In this example, non-smoothed power spectrum envelope sequences corresponding to frequency regions other than a predetermined frequency region for which resolution is to be improved are W _o (I _L +1), W _o (I _L +2),. _o (N-1).

Smoothed power spectral envelope sequence generation unit 212, the linear prediction coefficient ~ alpha ₁ for sample point k included in a predetermined frequency region to increase the resolution, ~ alpha _2, ..., ~ using the α _{p 'W} γ ₍ k) is calculated, and W _γ (k), which is a value obtained by smoothing W _o (k), is calculated for sample points k not included in the predetermined frequency region for which resolution is to be increased, thereby _obtaining a smoothed power spectrum. Envelope sequences W _γ (0), W _γ (1),..., W _γ (N-1) are generated (step E12).

The generated smoothed per-spectrum envelope sequences W _γ (0), W _γ (1),..., W _γ (N−1) are output to the envelope normalization unit 26 and the encoding unit 27.

Specifically, the smoothed power spectrum envelope sequence generation unit 212 performs linear prediction coefficients ~ α ₁ , ~ α ₂ , ..., ~ α _{p '} for sample points k included in a predetermined frequency region where resolution is desired to be increased. Is used to calculate the smoothed power spectrum envelope W _γ (k) by the following equation. Examples of sample points k included in a predetermined frequency region to increase the resolution, k = 0,1, ..., a I _L.

  Specifically, the smoothed power spectrum envelope sequence generation unit 212 corrects the non-smoothed power spectrum envelope sequence corresponding to the frequency region other than the predetermined frequency region for which the resolution is to be improved, in the same manner as the conventional smoothing method. A smoothed power spectrum envelope sequence smoothed by the coefficient γ is generated.

An example of a non-smoothed power spectrum envelope sequence corresponding to a frequency domain other than a predetermined frequency domain for which resolution is to be increased is W _o (I _L +1), W _o (I _L +2), ..., W _o (N -1). In this case, the generated smoothed power spectrum envelope sequences are W _γ (I _L +1), W _γ (I _L +2),..., W _γ (N−1).

  Note that the smoothed power spectrum envelope sequence corresponding to a frequency region other than the predetermined frequency region whose resolution is to be increased is processed by the expansion / contraction smoothing power spectrum envelope sequence generation unit 213 and the second inverse expansion / conversion conversion unit 28 of the third embodiment. It may be generated by performing the same process.

That is, first, using the quantized linear prediction coefficients ^ β ₁ , ^ β ₂ ,..., ^ Β _p , the stretched smoothed power spectrum coefficient sequence ~ W _γ (0), ~ W _γ ( 1), ..., ~ W _γ (N-1) is generated. Then, the generated stretched and smoothed power spectrum coefficient sequence is converted into a smoothed power spectrum envelope sequence corresponding to the linear discretized sample point sequence. A portion of the converted smoothed power spectrum envelope sequence corresponding to a frequency region other than the predetermined frequency region whose resolution is to be increased corresponds to a smoothed power spectrum corresponding to a frequency region other than the predetermined frequency region whose resolution is to be increased. Envelope series.

  The encoding device of the fourth embodiment may include a stretch / smoothing power spectrum envelope sequence generation unit 213 and a second inverse stretch / conversion transform unit 28 for performing this process.

  Subsequent processing by the encoding device, that is, processing of the envelope normalization unit 26 and the encoding unit 27 is the same as that of the third embodiment, and thus description thereof is omitted.

(Decoding of the fourth embodiment)
FIG. 14 shows a configuration example of the decoding device according to the fourth embodiment. As shown in FIG. 14, the decoding device of the fourth embodiment includes a stretched linear prediction coefficient decoding unit 31, a stretched / unsmoothed power spectrum envelope sequence generating unit 37, a first inverse stretch transforming unit 38, and a high-order linearity. For example, a prediction analysis unit 310, a smoothed power spectrum envelope sequence generation unit 311, a decoding unit 34, an envelope denormalization unit 35, and a time domain conversion unit 36 are provided. An example of each process of the decoding method of the fourth embodiment realized by this decoding apparatus is shown in FIG.

  The following description will focus on the parts different from the decoding / decoding device of the third embodiment, and the description of the same parts as in the third embodiment will be omitted.

  The processes of the expansion / contraction linear prediction coefficient decoding unit 31, the expansion / contraction unsmoothed power spectrum envelope sequence generation unit 37, the first inverse expansion / conversion conversion unit 38, the decoding unit 34, the envelope inverse normalization unit 35, and the time domain conversion unit 36 This is the same as the embodiment.

<Higher-order linear prediction analysis unit 310>
The higher-order linear prediction analysis unit 310 includes the non-smoothed power spectrum envelope sequences W _o (0), W _o (1) _,. A non-smoothed power spectrum envelope sequence corresponding to a predetermined frequency region for which the resolution is to be increased is input.

The high-order linear prediction analysis unit 310 uses a non-smoothed power spectrum envelope sequence corresponding to a predetermined frequency region whose resolution is to be increased by the same processing as the high-order linear prediction analysis unit 211 of the encoding device, Linear prediction analysis, that is, linear prediction analysis at a prediction order p ′ higher than the prediction order p of the linear prediction analysis unit 23, to generate linear prediction coefficients ~ α ₁ , ~ α ₂ , ..., ~ α _{p '} (Step D10).

The generated linear prediction coefficients ˜α ₁ , ˜α ₂ ,..., ˜α _{p ′} are output to the smoothed power spectrum envelope sequence generation unit 311.

When the predetermined frequency region whose resolution is to be increased is a low frequency region, the non-smoothed power spectrum envelope sequence corresponding to the predetermined frequency region whose resolution is to be increased is, for example, W _o (0), W _o (1) , ..., W _o (I _L ).

<Smoothing power spectrum envelope sequence generation unit 311>
The smoothed power spectrum envelope sequence generation unit 311 converts the linear prediction coefficients ~ α ₁ , ~ α ₂ , ..., ~ α _{p '} generated by the high-order linear prediction analysis unit 310 and the first inverse expansion / contraction conversion unit 38. Non-smoothing corresponding to a frequency region other than a predetermined frequency region in which the resolution is desired to be increased in the non-smoothed power spectrum envelope sequence W _o (0), W _o (1), ..., W _o (N-1) A power spectrum envelope sequence is input.

The smoothed power spectrum envelope sequence generation unit 311 performs linear prediction coefficients ~ α ₁ , ~ α ₂ , ..., ~ α _{p '} and the resolution in the same manner as the smoothed power spectrum envelope sequence generation unit 212 of the encoding device. Using the non-smoothed power spectrum envelope sequence corresponding to the frequency domain other than the predetermined frequency domain to be enhanced, the smoothed per-spectrum envelope series W _γ (0), W _γ (1), ..., W _γ (N- 1) is generated (step D11).

The generated smoothed per-spectrum envelope sequences W _γ (0), W _γ (1),..., W _γ (N−1) are output to the decoding unit 34 and the envelope denormalization unit 35.

  Subsequent processing by the decoding device, that is, processing of the decoding unit 34, the envelope denormalization unit 35, and the time domain conversion unit 36 is the same as that of the third embodiment, and thus description thereof is omitted.

[Sample sequence generation apparatus and method]
The above-described expansion / contraction pseudo power spectrum sequence generation unit 22 generates a frequency sequence sample sequence ξ (0), ξ (1),..., Ξ (N−1) derived from the sound signal by a predetermined M × N. Sample sequence generation including a sample sequence generation unit for generating sample sequences ~ ξ (0), ~ ξ (1), ..., ~ ξ (M-1) defined by the following formula using the non-negative matrix U of It can be said that it is a device. In other words, the expansion / contraction pseudo power spectrum sequence generation step (step E2) by the expansion / contraction pseudo power spectrum sequence generation unit 22 described above is performed in the frequency domain sample sequence ξ (0), ξ (1),. ξ (N-1) is a sample sequence ~ ξ (0), ~ ξ (1), ..., ~ ξ (M defined by the following equation using a predetermined M × N non-negative matrix U It can be said that this is a sample sequence generation method including a sample sequence generation step for generating -1). For example, N = M.

  Here, the non-negative matrix U corresponds to the transformation matrix U described in the section <Expanded pseudo power spectrum sequence generation unit 22>. As described above, since the non-negative matrix U is often a sparse matrix, U may be a sparse matrix here. The conditions for the sparse matrix U are the same as the conditions for the transformation matrix U described in the section <Expanded pseudo power spectrum sequence generation unit 22>, and thus the description thereof is omitted here.

  The sample sequence ξ (0), ξ (1),..., Ξ (N-1) may be any sample sequence as long as it is a frequency-domain sample sequence derived from a sound signal. For example, sample sequences ξ (0), ξ (1), ..., ξ (N-1) are [1] MDCT coefficient sequences X (0), X (1), ..., X (N-1). Or [2] a sequence of absolute values of MDCT coefficients | X (0) |, | X (1) |, ..., | X (N-1) |, or [3] sound It may be a sequence X (0) ^ q, X (1) ^ q,..., X (N-1) ^ q of the absolute value of the power spectrum MDCT coefficient of the signal. Further, the MDCT coefficient X (i) (i = 0, 1, ..., N-1) is replaced with a frequency domain feature amount other than the MDCT coefficient, and X '(i) (i = 0, 1, ..., N-1), the sample sequence ξ (0), ξ (1), ..., ξ (N-1) is [4] X '(0), X' (1), ..., X '(N- 1), [5] | X '(0) |, | X' (1) |, ..., | X '(N-1) |, or [6] X It may be '(0) ^ q, X' (1) ^ q, ..., X '(N-1) ^ q. Further, the sample sequence ξ (0), ξ (1),..., Ξ (N-1) may be a power spectrum sequence of [7] sound signal or [8] power spectrum envelope sequence of sound signal It may be. Furthermore, the sample sequence ξ (0), ξ (1), ..., ξ (N-1) smoothes the correction values [9] [1] to [7], in other words, [1] to [7] It may be a series. Here, smoothing means a process of dulling the unevenness of the magnitude (amplitude) of the values of the series while maintaining the magnitude relationship between the values of the series.

  Here, when the sample sequence is a series consisting of only non-negative values such as [2] to [9], the approximation accuracy of the second sample sequence described later is increased.

  Similarly, the sample sequence ~ ξ (0), ~ ξ (1), ..., ~ ξ (M-1) can be any sample string as long as it is a frequency domain sample string derived from a sound signal. Good. An example of the frequency domain sample sequence derived from the sound signal is the same as described above.

  In addition, the reverse expansion / conversion conversion unit 25, the reverse expansion / conversion conversion unit 33, the first reverse expansion / contraction conversion unit 210, the second reverse expansion / contraction conversion unit 28, the first reverse expansion / contraction conversion unit 38, and the second reverse expansion / contraction conversion unit 39 described above are The frequency domain sample sequence derived from the sound signal ~ η (0), ~ η (1), ..., ~ η (M-1), using a predetermined N × M non-negative matrix U, It can be said that this is a sample sequence generation apparatus including a sample sequence generation unit that generates sample sequences η (0), η (1),..., Η (N−1) defined by the following formula. In other words, the reverse expansion / conversion conversion step (step E5) by the reverse expansion / conversion conversion unit 25, the reverse expansion / conversion conversion step (step D3) by the reverse expansion / conversion conversion unit 33, and the reverse expansion / conversion conversion step by the first reverse expansion / conversion conversion unit 210 (described above) Step E10), reverse expansion / conversion conversion step (step E8) by the second reverse expansion / conversion conversion unit 28, reverse expansion / conversion conversion step (step D8) by the first reverse expansion / conversion conversion unit 38, and reverse expansion / conversion conversion step by the second reverse expansion / conversion conversion unit 39 (Step D9) is a predetermined M × N non-negative matrix of frequency domain sample sequences ~ η (0), ~ η (1), ..., ~ η (M-1) derived from sound signals. It can be said that this is a sample sequence generation method including a sample sequence generation step for generating sample sequences η (0), η (1),..., Η (N−1) defined by the following formula using U. For example, N = M.

  Here, the non-negative matrix V corresponds to the transformation matrix V described in the section of <Inverse expansion / contraction transformation unit 25>. As described above, since the non-negative matrix V is often a sparse matrix, V may be a sparse matrix here. The conditions for the sparse matrix V are the same as the conditions for the conversion matrix V described in the section <Inverse expansion / conversion conversion unit 25>, and thus the description thereof is omitted here.

  Sample sequence ~ η (0), ~ η (1), ..., ~ η (M-1) and sample series η (0), η (1), ..., η (N-1) are derived from sound signals Any sample string may be used as long as it is a frequency-domain sample string. An example of the frequency domain sample sequence derived from the sound signal is the same as described above.

  In signal processing such as encoding and decoding of a sound signal, a sample sequence (also referred to as a first sample sequence) expressed with a resolution in a certain frequency domain is used as a sample sequence (second sample) expressed with a resolution different from the resolution. It may be necessary to convert it to a sample string.

  An example of the first sample sequence is a sample sequence expressed based on a linear discrete sample point sequence, and an example of the second sample sequence is a sample sequence expressed based on a nonlinear discrete sample point sequence.

  In this example, when all elements of U are non-zero, the calculation using U increases the approximation accuracy of the second sample sequence expressed based on the non-linear discrete sample point sequence, but the calculation amount increases. . On the other hand, by using U as a sparse matrix, the second sample sequence can be approximated with a small amount of computation even if the second sample sequence is essentially non-linear while maintaining a reasonable approximation accuracy.

  In addition, another example of the first sample sequence is a sample sequence expressed based on a nonlinear discrete sample point sequence, and another example of the second sample sequence is a sample expressed based on a linear discrete sample point sequence. Is a column.

  In this other example, when all elements of V are non-zero, the calculation using V increases the approximation accuracy of the second sample sequence expressed based on the linear discrete sample point sequence, but the amount of calculation is growing. On the other hand, by using V as a sparse matrix, the second sample sequence can be approximated with a small amount of computation even if the first sample sequence is essentially non-linear while maintaining a reasonable approximation accuracy.

  Note that the calculation using the sparse matrices U and V by the sample sequence generation apparatus and method can be used for signal processing other than encoding and decoding of a sound signal.

[Modifications, etc.]
The sample sequence generation method, the encoding method, the decoding method, and the processes described in these apparatuses are not only executed in time series in the order described, but also in parallel according to the processing capability of the apparatus that executes the processes or as necessary. Or may be performed individually.

  In addition, when each step of the sample string generation method is realized by a computer, the processing content of each step of the sample string generation method is described by a program. Each step of the sample sequence generation method is realized on the computer by executing this program on the computer.

  Similarly, when each step of the encoding method is realized by a computer, the processing content of each step of the encoding method is described by a program. Each step of the encoding method is realized on the computer by executing this program on the computer.

  Similarly, when each step of the decoding method is realized by a computer, the processing content of each step of the decoding method is described by a program. Each step of the decoding method is realized on the computer by executing this program on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  In addition, each step of the encoding method and the decoding method may be configured by causing a computer to execute a predetermined program, or at least a part of these processing contents may be realized by hardware. Also good.

  Needless to say, other modifications are possible without departing from the spirit of the present invention.

DESCRIPTION OF SYMBOLS 11 Frequency domain conversion part 12 Linear prediction analysis part 13 Power spectrum envelope sequence generation part 14 Envelope normalization part 15 Encoding part 21 Frequency domain conversion part 22 Expansion / contraction pseudo power spectrum series generation part 23 Linear prediction analysis part 24 Expansion / contraction power spectrum envelope series Generation unit 25 Inverse expansion / contraction conversion unit 26 Envelope normalization unit 27 Encoding unit 28 Second inverse expansion / contraction conversion unit 29 Expansion / non-smoothing power spectrum envelope sequence generation unit 210 First inverse expansion / conversion conversion unit 211 High-order linear prediction analysis unit 212 Smooth Power spectrum envelope sequence generation unit 31 expansion / contraction linear prediction coefficient decoding unit 32 expansion / contraction power spectrum envelope sequence generation unit 33 inverse expansion / conversion conversion unit 34 decoding unit 35 envelope inverse normalization unit 36 time domain conversion unit 37 expansion / contraction unsmoothed power spectrum envelope sequence Generation unit 38 First inverse expansion / conversion conversion unit 39 Second inverse expansion / contraction conversion unit 310 High-order linear prediction analysis unit 11 smoothed power spectral envelope sequence generator

Claims (15)

Using a predetermined N × N sparse matrix U, where ξ (0), ξ (1), ..., ξ (N-1) are sample sequences in the frequency domain derived from sound signals in a predetermined time interval A sample sequence generation step for generating ~ ξ (0), ~ ξ (1), ..., ~ ξ (N-1) defined by the following equation,

The sparse matrix U is a matrix that includes non-zero values only in the neighboring components of the diagonal component, and the number of nonzero elements included in the upper triangular matrix is greater than the number of nonzero elements included in the lower triangular matrix. Less
The sample point sequence corresponding to ξ (0), ξ (1),..., Ξ (N-1) is a linear discretized sample point sequence in which the frequency intervals of adjacent sample points are equal.
Sample column generation method. The sample sequence generation method according to claim 1,
The sample point sequence corresponding to ~ ξ (0), ~ ξ (1), ..., ~ ξ (N-1) is a non-linear discretized sample point sequence in which the frequency intervals of adjacent sample points are uneven. Yes,
Sample points of the linear discretization sample point sequence are linear discretization sample points, sample points of the nonlinear discretization sample point sequence are nonlinear discretization sample points,
Each row of the sparse matrix U corresponds to each nonlinear discretized sample point of the nonlinear discretized sample point sequence, and each column of the sparse matrix U corresponds to each linear discretized sample point of the linear discretized sample point sequence As to
Each row of the sparse matrix U is a non-zero value only in the neighboring elements of the column corresponding to the linear discretization sample point of the frequency closest to the frequency corresponding to the nonlinear discretization sample point corresponding to each row, and the other elements Is a sparse matrix such as 0,
Sample column generation method. A frequency sequence sample sequence derived from the sound signal for each predetermined time interval is ξ (0), ξ (1),..., Ξ (N-1), and a predetermined non-negative matrix U of M × N is defined as A sample sequence generation step for generating ~ ξ (0), ~ ξ (1), ..., ~ ξ (M-1) defined by the following equation:

The non-negative matrix U has a component of the i-th row and k-th column as U [i, k], and for each i = 1, 2,.

Where g ₁ <g ₂ <… <g _N is satisfied,
The sample point sequence corresponding to ξ (0), ξ (1), ..., ξ (N-1) or ~ ξ (0), ~ ξ (1), ..., ~ ξ (M-1) has a frequency Sample point sequence equally spaced in the direction,
Sample column generation method. The sample string generation method according to claim 1 or 2,
All elements of the sparse matrix U are non-negative,
Sample column generation method. A predetermined N × N sparse matrix V, where ~ η (0), ~ η (1), ..., ~ η (N-1) are sample sequences in the frequency domain derived from sound signals in a predetermined time interval A sample sequence generation step for generating η (0), η (1),..., Η (N-1) defined by the following equation:

The sparse matrix V is a matrix including non-zero values only in the neighboring components of the diagonal component, and the number of nonzero elements included in the upper triangular matrix is greater than the number of nonzero elements included in the lower triangular matrix. Many
The sample point sequence corresponding to η (0), η (1),..., Η (N-1) is a linear discretized sample point sequence in which the frequency intervals of adjacent sample points are equal.
Sample column generation method. The sample sequence generation method according to claim 5, comprising:
The sample point sequence corresponding to ~ η (0), ~ η (1), ..., ~ η (N-1) is a non-linear discretized sample point sequence in which the frequency intervals of adjacent sample points are uneven. Yes,
Sample points of the linear discretization sample point sequence are linear discretization sample points, sample points of the nonlinear discretization sample point sequence are nonlinear discretization sample points,
Each column of the sparse matrix V corresponds to each nonlinear discretized sample point of the nonlinear discretized sample point sequence, and each row of the sparse matrix V corresponds to each linear discretized sample point of the linear discretized sample point sequence As to
Each row of the sparse matrix V has a non-zero value only in the neighboring elements of the column corresponding to the non-linear discretization sample point of the frequency closest to the frequency of the linear discretization sample point corresponding to the row, and the other elements are 0. Is a sparse matrix such that
Sample column generation method. The method of generating a sample string according to claim 5 or 6,
All elements of the sparse matrix V are non-negative,
Sample column generation method. A sequence corresponding to the power of the sample sequence obtained by converting the sound signal of the predetermined time interval into the frequency domain by the sample sequence generation method according to claim 1 or 2 is the ξ (0), ξ (1 ), ..., ξ (N-1) and ~ ξ (0), ~ ξ (1), ..., ~ ξ (N-1) are expanded and contracted pseudo power spectrum series ~ Y (0), ~ Y (1 ), ..., ~ Y (N-1) to generate a stretching pseudo power spectrum sequence generation step;
p is a positive integer, and the stretched pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1) is a series converted to the time domain ~ X (0), ~ A linear prediction analysis step for generating a quantized stretch linear prediction coefficient ^ β ₁ , ^ β ₂ ,…, ^ β _p by linearly predicting X (1), ..., ~ X (N-1),
Stretched power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W, which is a sequence in the frequency domain corresponding to the quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ , ..., ^ β _p A stretching power spectrum envelope sequence generation step for generating (N-1);
8. The stretched power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W (N-1) is converted into ~ η (0), ... by the sample string generation method according to any one of claims 5 to 7. ~ η (1), ..., ~ η (N-1), and the above η (0), η (1), ..., η (N-1) are used as power spectrum envelope sequences W (0), W (1) , ..., inverse stretch conversion step generated as W (N-1),
The frequency domain sample sequence X (0), X (1),..., X (N-1) using the power spectrum envelope sequence W (0), W (1),. An envelope normalization step to generate a normalized frequency domain sample sequence by normalizing
An encoding step of encoding the normalized frequency domain sample sequence to generate a code corresponding to the normalized frequency domain sample sequence;
An encoding method including: The encoding method according to claim 8, comprising:
The frequency interval between adjacent sample points in the low frequency region of the sample point sequence corresponding to the stretchable power spectrum envelope sequence is greater than the frequency interval between adjacent sample points in the high frequency region of the stretchable power spectrum envelope sequence. narrow,
An encoding method characterized by the above. Stretching linear prediction coefficient decoding step for generating quantized stretching linear prediction coefficients ^ β ₁ , ^ β ₂ ,…, ^ β _p by decoding the input stretching linear prediction coefficient code, where p is a positive integer,
Stretched power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W, which is a sequence in the frequency domain corresponding to the quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ , ..., ^ β _p A stretching power spectrum envelope sequence generation step for generating (N-1);
8. The stretched power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W (N-1) is converted into ~ η (0), ... by the sample string generation method according to any one of claims 5 to 7. ~ η (1), ..., ~ η (N-1), and the above η (0), η (1), ..., η (N-1) are used as power spectrum envelope sequences W (0), W (1) , ..., inverse stretch conversion step generated as W (N-1),
The code corresponding to the input sampled frequency domain sample sequence X _N (0), X _N (1), ..., X _N (N-1) is decoded, and the normalized frequency domain sample string is decoded. A decoding step for generating sample sequences X _N (0), X _N (1),..., X _N (N−1);
.., W (N-1) using the power spectrum envelope sequence W (0), W (1),..., W (N-1), the frequency domain sample sequence X _N (0), X _N (1),. by inverse normalize X _{N (N-1),} sample sequence X in the frequency domain (0), X (1) , ..., and envelope denormalization step of generating X (N-1),
A decoding method including: A predetermined N × N sparse matrix U is used, where ξ (0), ξ (1),..., Ξ (N-1) are sample sequences in the frequency domain derived from sound signals for each predetermined time interval. Including a sample sequence generation unit for generating ~ ξ (0), ~ ξ (1), ..., ~ ξ (N-1) defined by the following formulas,

The sparse matrix U is a matrix having only non-zero values in the vicinity of diagonal components, and the number of non-zero elements included in the upper triangular matrix is greater than the number of non-zero elements included in the lower triangular matrix. Less
The sample point sequence corresponding to ξ (0), ξ (1),..., Ξ (N-1) is a linear discretized sample point sequence in which the frequency intervals of adjacent sample points are equal.
Sample sequence generator. Predetermined N × N sparse matrix with ~ η (0), ~ η (1), ..., ~ η (N-1) as sample sequences in the frequency domain derived from the sound signal for each predetermined time interval A sample sequence generation unit that generates η (0), η (1),..., Η (N-1) defined by the following equation using V,

The sparse matrix V is a matrix including non-zero values only in the neighboring components of the diagonal component, and the number of nonzero elements included in the upper triangular matrix is greater than the number of nonzero elements included in the lower triangular matrix. Many
The sample point sequence corresponding to η (0), η (1),..., Η (N-1) is a linear discretized sample point sequence in which the frequency intervals of adjacent sample points are equal.
Sample sequence generator. A sequence corresponding to the power of the sample sequence obtained by converting the sound signal for each predetermined time interval into the frequency domain is ξ (0), ξ (1), ..., ξ (N-1), and Generate ξ (0), ~ ξ (1), ..., ~ ξ (N-1) as the stretched pseudo power spectrum series ~ Y (0), ~ Y (1), ..., ~ Y (N-1) An expansion / contraction pseudo power spectrum sequence generation unit which is the sample string generation device of claim 11;
p is a positive integer, and the stretched pseudo power spectrum sequence ~ Y (0), ~ Y (1), ..., ~ Y (N-1) is a series converted to the time domain ~ X (0), ~ A linear prediction analysis unit that generates a quantized stretch linear prediction coefficient ^ β ₁ , ^ β ₂ ,…, ^ β _p by linearly predicting X (1), ..., ~ X (N-1),
Stretched power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W, which is a sequence in the frequency domain corresponding to the quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ , ..., ^ β _p A stretchable power spectrum envelope generation unit for generating (N-1),
The stretched power spectrum envelope series ~ W (0), ~ W (1), ..., ~ W (N-1) is changed to ~ η (0), ~ η (1), ..., ~ η (N-1). The η (0), η (1),..., Η (N-1) is generated as a power spectrum envelope sequence W (0), W (1),. A reverse expansion / conversion conversion unit that is a sample string generation device of
The frequency domain sample sequence X (0), X (1),..., X (N-1) using the power spectrum envelope sequence W (0), W (1),. An envelope normalization unit that generates a normalized frequency domain sample sequence by normalizing
An encoding unit that encodes the normalized frequency domain sample sequence and generates a code corresponding to the normalized frequency domain sample sequence;
An encoding device including: A stretched linear prediction coefficient decoding unit that decodes an input stretched linear prediction coefficient code and generates quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ ,…, ^ β _p , where p is a positive integer,
Stretched power spectrum envelope sequence ~ W (0), ~ W (1), ..., ~ W, which is a sequence in the frequency domain corresponding to the quantized stretched linear prediction coefficients ^ β ₁ , ^ β ₂ , ..., ^ β _p A stretchable power spectrum envelope generation unit for generating (N-1),
The stretched power spectrum envelope series ~ W (0), ~ W (1), ..., ~ W (N-1) is changed to ~ η (0), ~ η (1), ..., ~ η (N-1). The η (0), η (1),..., Η (N-1) is generated as a power spectrum envelope sequence W (0), W (1),. A reverse expansion / conversion conversion unit that is a sample string generation device of
The code corresponding to the input sampled frequency domain sample sequence X _N (0), X _N (1), ..., X _N (N-1) is decoded, and the normalized frequency domain sample string is decoded. A decoding unit for generating sample sequences X _N (0), X _N (1),..., X _N (N−1);
.., W (N-1) using the power spectrum envelope sequence W (0), W (1),..., W (N-1), the frequency domain sample sequence X _N (0), X _N (1),. By denormalizing X _N (N-1), an envelope denormalization unit for generating frequency domain sample sequences X (0), X (1), ..., X (N-1);
A decoding device.   The program for making a computer perform each step of the sample row | line production | generation method in any one of Claim 1 to 7.

Images