Encoding method, decoding method, encoding device, decoding device, program, and recording medium

The present invention relates to a technology for encoding or decoding signal sequences of acoustic signals, video signals, and other signals, such as voice and music, by vector quantization.

In an encoding device disclosed in Patent Literature 1, an input signal is first divided by a normalization value to perform normalization. The normalization value is quantized, and a quantization index is generated. The normalized input signal is vector-quantized, and an index of a representative quantization vector is generated. The generated quantization index and the generated representative quantization vector are output to a decoding device. The decoding device decodes the quantization index and generates a normalization value. The index of the representative quantization vector is also decoded, and a sample sequence is generated. A sequence of the values obtained by multiplying each sample in the generated sample sequence by the normalization value serves as a decoded signal sample sequence.

On the other hand, as highly efficient vector quantization methods that generate little quantization noise, the spherical vector quantization (SVQ) method (refer to Non-Patent Literature 1, for example) and other vector quantization methods that quantize a plurality of input signals together within a predetermined number of quantization bits are widely used.

In the SVQ method, samples of input signals such as modified discrete cosine transform (MDCT) coefficients are normalized by using a quantized normalization value, and the normalized samples are quantized together in units of sub-bands. Here, the number of bits (quantization bits) are dynamically assigned to a code corresponding to each sub-band in accordance with perceptual importance of each sub-band. Assuming that the input signals are sparse, the SVQ method quantizes the main elements of the input signals preferentially. Therefore, input signals having sparse energy in the frequency domain (sparse signals), such as harmonics signals and vowels, can be quantized with high precision.

However, the SVQ method increases the frequency that a frequency component included in the input signals is not included in decoded signals decoded from the quantized values (the decoded signals lack the frequency component) when the samples are quantized for input signals having energy in many frequencies. When the decoded signals lack a frequency component, the presence or absence of the frequency component in the decoded signals varies discontinuously over time at a high frequency. Humans are sensitive to those temporally discontinuous variations in the presence or absence of a frequency component. If the input signals are acoustic signals, these variations may be perceived as noise which is known as musical noise. If the input signals are video signals, block noise, which is equivalent to musical noise in the acoustic signals, may occur. Musical noise and block noise will be referred to as “musical noise and the like” below.

An algebraic vector quantization (AVQ) method (refer to Non-Patent Literature 2, for example) is a vector quantization method in which the decoded signals lack a frequency component at a lower frequency than with the SVQ method. Like the SVQ method, the AVQ method assumes that the signals are sparse, but the AVQ method can provide quantized values with which more frequency components can be restored than with the SVQ method.

Patent Literature 1: Japanese Patent Application Laid Open No. H7-261800 Non-Patent Literature 1: Recommendation ITU-T G.729.1, SERIES G: TRANSMISSION SYSTEMS AND MEDIA, DIGITAL SYSTEMS AND NETWORKS, Digital terminal equipments—Coding of analogue signals by methods other than PCM, G729-based embedded variable bit-rate coder: An 8-32 kbit/s scalable wideband coder bitstream interoperable with G729. Non-Patent Literature 2: Recommendation ITU-T G.718, SERIES G: TRANSMISSION SYSTEMS AND MEDIA, DIGITAL SYSTEMS AND NETWORKS, Digital terminal equipments—Coding of voice and audio signals, Frame error robust narrow-band and wideband embedded variable bit-rate coding of speech and audio from 8-32 kbit/s.

Problems to be Solved by the Invention

The amplitude quantization precision of the AVQ method is lower than that of the SVQ method, however. Even if the decoded signals lack a frequency component at a low frequency, low amplitude quantization precision could cause musical noise and the like. This problem is not limited to the AVQ method, and is common when musical noise and the like occur depending on the quantization precision. This problem can occur not only when the input signals are frequency-domain signals but also when the input signals are time-domain signals.

The present invention provides a technology for reducing musical noise and the like that can occur depending on the quantization precision.

Means to Solve the Problems

In encoding, a plurality of samples are collectively vector quantized to obtain a vector quantization index and the quantized value of each of the plurality of samples; bits are assigned in a predetermined order of priority based on auditory perceptual characteristics to one or more sets of sample positions among a plurality of sets of sample positions, each set having a plurality of sample positions and being given an order of priority based on the auditory perceptual characteristics, the number of bits not being larger than the number of bits obtained by subtracting the number of bits used for a code corresponding to the vector quantization index from the number of bits assigned for the code corresponding to the vector quantization index; and index information indicating a group of coefficients that minimizes the sum of the error between the value of each sample included in each of the sets of sample positions to which the bits are assigned and the value obtained by multiplying the quantized value of each sample included in the set of sample positions by a coefficient corresponding to the position of the sample, of all the sample positions included in the set of sample positions, among a plurality of groups of predetermined coefficients corresponding to the positions of the samples, is output.

In decoding, a plurality of values corresponding to an input vector quantization index are obtained as decoded values corresponding respectively to a plurality of sample positions; and, with a set of sample positions to which bits are assigned in a predetermined order of priority based on auditory perceptual characteristics being considered as a target of processing among a plurality of sets of sample positions that constitute the plurality of sample positions, and with the use of a group of predetermined coefficients corresponding to the plurality of sample positions and indicated by input index information, the values obtained by multiplying the decoded values and the coefficients, corresponding to the sample positions included in the set of sample positions considered as the target of processing, are output.

Effects of the Invention

In encoding, since index information indicating a group that includes a plurality of gradient coefficients is output to adjust the quantized values of a plurality of samples, the quantization error in decoding can be reduced, and as a result musical noise and the like can be reduced.

In decoding, since a plurality of decoded values are adjusted by using a group that includes a plurality of gradient coefficients indicated by index information, the quantization error can be reduced, and as a result musical noise and the like can be reduced.

FIG. 1 is a functional block diagram of an encoding device and a decoding device;

FIG. 2 is a flowchart illustrating an encoding method;

FIG. 3 is a flowchart illustrating an example of step E 4 ;

FIG. 4 is a flowchart illustrating the example of step E 4 ;

FIG. 5 is a flowchart illustrating a decoding method;

FIG. 6 is a flowchart illustrating an example of step D 3 ;

FIG. 7 is a flowchart illustrating the example of step D 3 ;

FIG. 8 illustrates an example of the relationship among input signals, quantized values, and gradient coefficients (tilt correction gains).

An embodiment of the present invention will now be described in detail.

(Configuration)

As shown in FIG. 1 , an encoding device 11 in this embodiment includes a normalization value calculator 112 , a normalization value quantizer 113 , a vector quantizer 115 , and a gradient calculator 116 , for example. The gradient calculator 116 corresponds to a bit assignment unit and a coefficient group selection unit.

As shown in FIG. 1 , a decoding device 12 in this embodiment includes a normalization value decoder 121 , a vector decoder 122 , and a gradient adjusting unit 124 , for example. The encoding device 11 may include a frequency-domain converter 111 , for example, when necessary. The decoding device 12 may include a time-domain converter 125 and a smoothing unit 126 , for example.

(Encoding)

The encoding device 11 executes the steps of an encoding method illustrated in FIG. 2 .

Input signals X(k) are input to the normalization value calculator 112 , the vector quantizer 115 , and the gradient calculator 116 . The input signals X(k) here are frequency-domain signals that can be obtained by transforming time-domain signals x(n), which are time-series signals such as acoustic signals, into the frequency domain. The input signals X(k) in the frequency domain may be input directly to the encoding device 11 . Alternatively, the frequency-domain converter 111 may transform the input signals x(n) in the time domain into the frequency domain to generate the input signals X(k) in the frequency domain. When the frequency-domain converter 111 generates the input signals X(k) in the frequency domain, the frequency-domain converter 111 transforms the input signals x(n) in the time domain to the input signals X(k) in the frequency domain by modified discrete cosine transform (MDCT), for example. Here, n indicates the number (discrete time number) of the signals in the time domain, and k indicates the number (discrete frequency number) of the signals (samples) in the frequency domain. A larger n value corresponds to a later time. A larger k value corresponds to a higher frequency. When a single frame includes L samples, the time-domain signals x(n) are transformed into the frequency domain in units of frames, and the input signals X(k) (k=0, 1, . . . , L−1) forming L frequency components in the frequency domain are generated. Here, L is a given positive integer greater than 1, such as 64 or 80. When MDCT is used, the input time-series signals are transformed into the frequency domain in units of frames each including L samples, and the frame to be transformed shifts by ½ frame, or L/2 samples at a time.

The normalization value calculator 112 calculates a normalization value τ X 0 − , which is a value representing a predetermined C 0 samples in L samples of the input signals X(k), in each frame (step E 1 ). Here, τ X 0 − is the character τ X 0 with an overbar, where τ is a unique integer not smaller than 0, assigned to each sub-band formed of the predetermined C 0 samples out of L samples in a single frame.

C 0 is L or a common divisor of L other than 1 or L. Setting C 0 to a common divisor of L other than 1 or L means that the group of L samples is divided into sub-bands (corresponding to sets of sample positions), and a normalization value is obtained for each group of C 0 samples constituting each sub-band. For example, when L=64 and when eight frequency components constitute a sub-band, eight sub-bands are formed, and a normalization value is calculated for each sub-band. When C 0 is L, τ=0, and the normalization value τ X 0 − represents L samples. In other words, when C 0 is L, a single normalization value τ X 0 − is calculated for each frame. When C 0 is a common divisor of L other than 1 or L, τ is an integer τ=0, . . . , (L/C 0 )−1 corresponding to each sub-band in the single frame, and the normalization value τ X 0 − is a value representing C 0 samples included in the sub-frame corresponding to τ. That is, when C 0 is a common divisor of L other than 1 or L, L/C 0 normalization values τ X 0 − (τ=0, . . . , (L/C 0 )−1) are calculated for each frame. Independently of the value of C 0 , k=τ·C 0 , . . . , (τ+1)·C 0 −1. The value τ X 0 − calculated by the normalization value calculator 112 is sent to the normalization value quantizer 113 .

[Examples of Normalization Value τ X 0 − ]

The normalization value τ X 0 − is a representative value of C 0 samples. In other words, the normalization value τ X 0 − is a value that corresponds to C 0 samples. An example of the normalization value τ X 0 − is the following square root to a power average value of the C 0 samples.

Another example of the normalization value τ X 0 − is the following value, which is obtained by dividing, by C 0 , the square root to a total power value of the C 0 samples.

Still another example of the normalization value τ X 0 − is the following average amplitude value of the C 0 samples.

The normalization value τ X 0 − is not limited to the examples given above (the description of [Examples of normalization value τ X 0 − ] ends here).

The normalization value quantizer 113 quantizes the normalization value τ X − to obtain a quantized normalization value τ X − and obtains a normalization-value quantization index corresponding to the quantized normalization value τ X − (step E 2 ). Here, τ X − is the character τ X − with an overbar. The quantized normalization value τ X − is sent to the vector quantizer 115 , and a code (bit stream) corresponding to the normalization-value quantization index is sent to the decoding device 12 .

The vector quantizer 115 generates a vector quantization index by collectively vector-quantizing a plurality of samples X(k) out of L samples of the input signals X(k) in each frame. The vector quantization index is an index indicating a representative quantization vector. The vector quantizer 115 here normalizes a plurality of X(k)′s by using the quantized normalization value τ X − and obtains a plurality of normalized samples X(k)′. For example, the vector quantizer 115 obtains X(k)′ by dividing X(k) by τ X − or by multiplying X(k) by the reciprocal of τ X − . The vector quantizer 115 performs vector quantization by selecting a representative quantization vector closest to the vector composed of the plurality of samples X(k)′, out of a plurality of representative quantization vectors stored in a vector codebook storage, which is not shown in the drawings, and outputting a vector quantization index indicating the selected representative quantization vector, for example. The vector quantizer 115 vector-quantizes C 0 samples X(k)′ together, for example. The vector quantizer 115 performs vector quantization by using a vector quantization method such as the AVQ method (refer to Non-Patent Literature 2, for example), but any other vector quantization method may be used. When C 0 is the number of samples in the sub-band, vector quantization may be unperformed on a sub-band with a low priority given in accordance with human auditory perceptual characteristics. A sub-band corresponding to a frequency that is easier to be perceived by humans is given a higher priority, for example. A sub-band having a greater quantized normalization value τ X − is given a higher priority, for example.

The bit number of a code obtained by the vector quantization varies depending on the input signals. For some input signals, the bit number of a code (the vector quantization index or the like) obtained by the vector quantization may be less than a bit number assigned for the vector quantization, and part of bits assigned for the vector quantization may remain unused. The “bits assigned for the vector quantization” mean bits assigned for a code (a code corresponding to the vector quantization index) obtained by the vector quantization, among codes sent from the encoding device 11 to the decoding device 12 . The “bit number assigned for the vector quantization” means the bit number of the bits assigned for the vector quantization. The “bit number assigned for the vector quantization” may be determined for each frame, or may be determined for each sub-band. In addition, the “bit number assigned for the vector quantization” may vary depending on the input signal, or may be constant irrespective of the input signal. The vector quantizer 115 calculates the number of bits that are not used in actual vector quantization, among the bits assigned for vector quantization, as the number of unused bits, U. In this embodiment, the number of unused bits, U, is calculated in each frame (in units of L samples). For example, the vector quantizer 115 obtains the number of unused bits, U, by subtracting, from the number of bits assigned for vector quantization in a target frame to be processed, the total number of bits of the vector quantization index obtained by vector quantization of L samples included actually in the frame. Here, U is an integer not smaller than 0.

The vector quantizer 115 further obtains a plurality of quantized values X^(k), which are local-decoded values of the vector quantization index, and outputs them. For example, the vector quantizer 115 outputs the values obtained by denormalizing the components X(k)′ of the representative quantization vector indicated by the vector quantization index, by using the quantized normalization value τ X −1 , as the quantized values X^(k) of X(k). For example, the vector quantizer 115 outputs the products of X(K)′ and τ X − as quantized values X^(k). Here, the quantized values X^(k) equal the decoded values X^(k) obtained by the decoding device 12 . The quantized values X^(k) of a sub-band that is not subjected to vector quantization become 0. Here, X^ indicates X with a superscript caret immediately above it.

The vector quantizer 115 sends the vector quantization index, the number of unused bits, U, and the quantized values X^(k) to the gradient calculator 116 (step E 3 ).

The gradient calculator 116 holds M MAX groups of C 0 gradient coefficients (tilt correction gains), for example, in a storage, which is not shown in the drawings. Here, M MAX is an integer not smaller than 2. For example, the gradient calculator 116 holds a gradient matrix γ given by Equation (1) where a gradient coefficient vector γ m =[γ m (0), . . . γ m (C 0 −1)] (a group of a plurality of gradient coefficients) composed of C 0 gradient coefficients (tilt correction gains) γ m (k′) (k′=0, . . . , C 0 −1) is provided as a row vector in the m-th row (m=0, . . . , M MAX −1).

The gradient calculator 116 obtains the row number m′ of the gradient coefficient vector that minimizes the error between a first vector composed of values corresponding to C 0 input samples X(k) (k=τ· τ C 0 , . . . , (τ+1)·C 0 −1) and a second vector composed of values corresponding to C 0 adjusted values obtained by multiplying the input C 0 quantized values X^(k) (k=τ· τ C 0 , . . . , (τ+1)·C 0 −1) by the elements γ m (k′) (k′=0, . . . , C 0 −1) of the gradient coefficient vector γ m , and writes index information idx indicating the row number m′ in the region of bits that are not used (referred to as an unused bit region), out of the bits assigned for vector quantization.

In other words, the gradient calculator 116 finds, out of a gradient matrix γ having gradient coefficient vectors γ m each composed of a plurality of gradient coefficients γ m (k′) as row vectors, the gradient coefficient vector that minimizes the error between a first vector composed of values corresponding to a plurality of samples X(k) and a second vector composed of values corresponding to the adjusted values obtained by multiplying the plurality of quantized values X^(k) by the elements of the gradient coefficient vector γ m , outputs index information idx indicating the row number m′ of that gradient coefficient vector, and places it in the unused bit region of the code (bit stream) corresponding to the vector quantization index (step E 4 ).

With this step, the encoding device 11 can send information for adjusting the quantization error of the amplitude to the decoding device 12 , using the unused bit region effectively, and can reduce musical noise and the like generated in accordance with the quantization precision.

The C 0 gradient coefficients γ m (0), . . . , γ m (C 0 −1) constituting the gradient coefficient vector γ m are correlated with one another. In other words, each gradient coefficient vector γ m is a vector composed of a plurality of mutually correlated gradient coefficients γ m (0), . . . , γ m (C 0 −1). It is the frequent case that X(τ·C 0 ), . . . , X((τ+1)·C 0 −1) are distributed lopsidedly on a straight line or a curve in a (k, X(k)) plane having k on its first axis and X(k) on its second axis. By using the gradient coefficient vector γ m composed of the gradient coefficients γ m (0), . . . , γ m (C 0 −1) with such characteristics of X(τ·C 0 ), . . . , X((τ+1)·C 0 −1) being taken into consideration, the quantization error can be adjusted with high precision. For example, it is assumed that the gradient coefficients γ m (0), . . . , γ m (C 0 −1) corresponding to the same row number m are distributed lopsidedly on a straight line or a specific curve in a (k′, γ m (k′)) plane having k′ (value corresponding to a frequency corresponding to the gradient coefficient γ m (k′)) on its first axis and γ m (k′) (value of gradient coefficient) on its second axis. In other words, it is assumed, for example, that the gradient coefficient vector γ m is a vector composed of a plurality of gradient coefficients γ m (0), . . . , γ m (C 0 −1) distributed lopsidedly on a straight line or a specific curve in the (k′, γ m (k′)) plane having the value k′ corresponding to the order of the element in the vector on its first axis and the value γ m (k′) of a vector element on its second axis. More specifically, it is assumed, for example, that the gradient coefficients γ m (0), . . . , γ m (C 0 −1) corresponding to the same row number m are placed on a straight line or a specific curve on the (k′, γ m (10) plane. In other words, it is assumed that a vector composed of the gradient coefficients γ m (0), . . . , γ m (C 0 −1) on a straight line or a specific curve in the (k′, γ m (10) plane is provided as the gradient coefficient vector γ m . The straight lines or specific curves in the (k′, γ m (k′)) plane are different depending on the row numbers m, for example. An example of the gradient matrix γ is shown below. The example shown is characterized by C 0 =8 and M MAX =3. In the example, the gradient coefficients γ m (0), . . . , γ m (7) are placed on a straight line given for each row number m (m=0, 1, 2).

γ

=

[

1.35

1.25

1.15

1.05

0.95

0.85

0.75

0.65

1.175

1.125

1.075

1.025

0.975

0.925

0.875

0.825

0.65

0.75

0.855

0.95

1.05

1.15

1.25

1.35

]

Examples of the first vector include a vector composed of C 0 X(k)′s, a vector composed of the magnitude |X(k)| of C 0 X(k)'s, and a vector composed of C 0 X(k)'s or |X(k)|'s multiplied by a constant or a variable. Examples of the second vector include a vector composed of C 0 adjusted values, a vector composed of the magnitude of the C 0 adjusted values, and a vector composed of C 0 adjusted values or their magnitude multiplied by a constant or a variable. Examples of adjusted values include the product of X^(k) and γ m (k′), the product of the magnitude |X^(k)| of X^(k) and γ m (k′), the magnitude of the product of X^(k) and γ m (k′), a value indicating the magnitude of the product of X^(k) and γ m (k′), and a value corresponding to the product of X^(k) and γ m (k′).

An example of the error between the first vector and the second vector is the distance between the first vector and the second vector. The distance is not especially defined and can be the Manhattan distance, the Euclidean distance, variations of those distances, and the like. Examples of the gradient coefficient vector that minimizes the error between the first vector and the second vector include a gradient coefficient vector that minimizes the error between the first vector and the second vector and a gradient coefficient vector that minimizes the error between the first vector and the second vector under given search conditions or within a given search range.

The unused bit region can be identified by the reference position (first address, for example) of a determined unused bit region and the input number of unused bits, U. The upper limit of the number of bits of the index information idx that can be written in the unused bit region is the number of unused bits, U. Therefore, the index information idx corresponding to all the row numbers cannot always be written in the unused bit region. So, the gradient calculator 116 specifies the range of row numbers that can be identified by the index information idx that can be written in the unused bit region as the search range. In other words, the gradient calculator 116 selects just a row number indicated by index information idx that can be written in the unused bit region. For example, the gradient calculator 116 identifies a row number m′ as given below, among the m MAX row numbers m=0, . . . , m MAX −1 that can be identified by index information idx that can be written in the unused bit region, and writes index information idx corresponding to the row number m′ in the unused bit region.

m ′ = arg ⁢ ⁢ min m ⁢  χ - A m · χ ^ 

The symbol ∥•∥ indicates the norm of; argmin m ∥•∥ means that m minimizing ∥•∥ becomes m′; argmin m means argmin with subscript m; and χ=[X(τ·C 0 ), . . . , X((τ+1)·C 0 −1), χ^=[X^(τ·C 0 ), . . . , X^((τ+1)·C 0 −1)]; and A m means a diagonal matrix having gradient coefficient vectors γ m =[γ m (0), . . . , γ m (C 0 −1)] (m=0, . . . , m MAX −1) corresponding to the row number m as its diagonal elements, as shown below.

When C 0 is the number of samples in a sub-band (τ>0), the gradient calculator 116 places index information idx indicating the row number m′ of the gradient coefficient vector in the unused bit region for each sub-band. In this case, in order to store index information idx corresponding to each sub-band, the unused bit region must be split for the sub-bands. The gradient calculator 116 determines the bit region in which index information idx of each sub-band is placed in the unused bit region in a predetermined order of priority based on human auditory perceptual characteristics, for example.

The gradient calculator 116 determines the size of the bit region where index information idx corresponding to each sub-band (corresponding to a set of sample positions) is placed in accordance with the priority level given to the sub-band, for example. A sub-band corresponding to a frequency that is easier to be perceived by humans is given a higher priority, for example. The gradient calculator 116 preferentially places index information idx of a sub-band with a higher priority in the unused bit region, for example. Specifically, the gradient calculator 116 makes the bit region where index information idx corresponding to a sub-band with higher priority is placed larger than the bit region where index information idx corresponding to a sub-band with lower priority is placed. In other words, the number of bits assigned to a first sub-band of the sub-bands is not smaller than the number of bits assigned to a second sub-band having a lower priority than the first sub-band. For example, the gradient calculator 116 does not assign a bit region for placing index information idx corresponding to a sub-band that has not been vector-quantized in the vector quantizer 115 . For example, since X^(k) of a sub-band that has not been vector-quantized are all 0, the gradient calculator 116 does not assign a bit region for placing index information idx of a sub-band if the sum of the amplitudes or energies of the included X^(k) elements is 0. For the sub-band that has not been vector-quantized, processing to identify index information idx, for example, is not necessary either.

As described above, the gradient calculator 116 assigns at least some of the bits that are not actually used for the code corresponding to the vector quantization index, of the bits assigned for the code corresponding to the vector quantization index, to a plurality of samples in a predetermined order of priority based on the auditory perceptual characteristics, and outputs index information idx having the number of bits assigned to the plurality of samples.

In other words, the gradient calculator 116 assigns bits the number of which is equal to or smaller than the number of bits obtained by subtracting the number of bits used for a code corresponding to the vector quantization index from the number of bits assigned for the code corresponding to the vector quantization index, to one or more sets of sample positions out of a plurality of sets of sample positions, each set including a plurality of sample positions and being given a priority level based on auditory perceptual characteristics, in a predetermined order of priority based on the auditory perceptual characteristics, and outputs index information idx indicating a group of coefficients that minimizes the sum of the error between the value of each sample belonging to each of the set of sample positions to which the bits are assigned, and the product of the quantized value of the sample belonging to the set of sample positions and a coefficient corresponding to the position of the sample, of all the sample positions belonging to the set of sample positions, out of a plurality of groups of predetermined coefficients corresponding to the positions of the samples. The positions of samples in this embodiment are positions on the frequency axis, corresponding to discrete frequency numbers k.

Now, index information idx can be transmitted just by using the unused bit region.

11 : Encoding device

111 : Frequency-domain converter

112 : Normalization value calculator

113 : Normalization value quantizer

115 : Vector quantizer

116 : Gradient calculator

12 : Decoding device

121 : Normalization value decoder

122 : Vector decoder

124 : Gradient adjusting unit

125 : Time-domain converter

126 : Smoothing unit