JP3010637B2 - Quantization device and quantization method - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a quantization device and a quantization method for performing high-efficiency coding of an input signal.

[Summary of the Invention]

The present invention provides a quantization apparatus and a quantization method that improve quantization efficiency by performing quantization in units of a matrix including a two-dimensional data array of a time axis and an axis orthogonal to the time axis for an input signal. To provide.

[Conventional technology]

As one method of high-efficiency signal encoding, while dividing the input signal into multiple channels on the time axis or frequency axis,
There is an encoding technique based on bit allocation (bit allocation) that adaptively allocates the number of bits for each channel. For example, a coding technique based on the above-mentioned bit allocation of a voice signal includes band division coding (sub-band coding) that divides a voice signal on a time axis into a plurality of frequency bands and codes them.
So-called adaptive transform coding (ATC), which converts a signal on the time axis into a signal on the frequency axis (orthogonal transform), divides the signal into multiple frequency bands, and adaptively codes each band. Alternatively, the SBC is combined with so-called adaptive prediction coding (APC), and a signal on the time axis is divided into bands, and each band signal is converted into a baseband (low band), and then a multi-order linear prediction analysis is performed. Coding techniques such as so-called adaptive bit allocation (APC-AB) for predictive coding.

Here, the above-mentioned band division coding is performed by a signal transmission device as shown in FIG. In FIG. 10, a digital audio signal supplied to an input terminal 110 is first converted into a frequency division filter (for example, QMF: quadrature
Bandpass and low-frequency conversion are performed by mirror filter groups 131 _{1 to} 131 _n such as mirror filters. That is, in the frequency division filter groups 131 _{1 to} 131 _n , after being band-divided by a band-pass filter (band-pass filter: BPF), each signal is passed through a low-pass filter and passed through the center frequency of the pass band. , And these signals are down-sampled at quantizers 134 _{1 to} 134 _n at an appropriate sampling frequency. Each signal whose data has been compressed by the requantization in this way is output from the terminal 138 via the multiplexer 136, transmitted to the terminal 148 of the decoder 140 via the transmission path, and
8 through the demultiplexer 149 to the inverse quantizers 144 _{1 to} 14
After being decoded by 4 _n, it is converted into a signal of each band on the time axis by frequency converters 142 _{1 to} 142 _n , and is output from terminal 150 as a decoded audio signal through adder 146.

Here, in the data compression processing of the signal by the encoder 130, the quality is improved by adaptively allocating quantization bits to each band so as to reduce the influence of noise generated by the data compression on the decoded speech signal. Is being planned. The decoder 140 also obtains bit allocation information by some method and performs decoding.

Conventionally, in order to obtain the above-mentioned bit allocation information, a method of transmitting energy value information of each band as auxiliary information (side information) separately from a signal of each band has been adopted. That is, in this method, the encoder
The energy values of the signals in each band are calculated by the energy detection means 133 _{1 to} 133 _n from the signals that have been band-divided by the 130 frequency division filter groups 131 _{1 to} 131 _n , and the quantization coefficients are calculated based on the calculated values. Means 135 performs optimal bit allocation and quantization step calculation as optimal quantization coefficients for quantization of signals in each band, and uses the results to re-create signals in each band in quantizers 134 _{1 to} 134 _n. Quantized. Further, the output signal of the quantization coefficient calculation means 135, that is,
Is transmitted to the quantization coefficient calculation means 145, and the information from the quantization coefficient calculation means 145 is transmitted to the inverse quantizers 144 _{1 to} 144 _N , where the information is inverse to the above-described quantizers 134 _{1 to} 134 _N. Is performed to decode the signal.

In such band division coding, noise shaping or the like can be considered in accordance with human auditory characteristics, and a large contribution is made to subjective quality such as a band in which voice energy is unbalanced and clarity. More information can be allocated to the band. The quantization and dequantization of the signals in the respective bands are performed with the allocated number of quantization bits, whereby the degree of auditory disturbance of quantization noise can be reduced, and the number of bits can be reduced as a whole. Also, by performing the band division coding, quantization noise occurs only in the divided band, and does not affect other bands. The method of transmitting the energy value information as auxiliary information as described above has an advantage that the energy value of the signal of each band is simultaneously used as the quantization step width (normalization factor) of the signal of each band.

[Problems to be solved by the invention]

In coding by bit allocation, for example, in the above-described band division coding (SBC), a time waveform of each band divided by a frequency band of an input signal is quantized by an adaptive number of allocated bits independently for each channel. The quantization at this time is performed by a vector quantization or the like that divides a sequence of sample values (a series of a plurality of words) into an appropriate size and collectively quantizes them as one vector. ing.

In addition, for example, in transform coding such as the above-described adaptive transform coding (ATC), a signal on the time axis is converted into, for example, a discrete cosine transform (DC
T), Fast Fourier Transform (FFT) or Discrete Fourier Transform (DFT), etc., and then perform band division. The DCT coefficient, FFT coefficient, DFT coefficient, etc. of each of these divided bands are calculated as described above. It was quantized by adaptive bit allocation.
As the quantization processing at this time, each coefficient is quantized by the scalar quantization for quantizing each independent sample value of the signal or the above-described vector quantization, and is processed.
Here, the scalar quantization as described above can remove the correlation between words, but when each word is quantized at a low bit rate of, for example, about 2 bits, quantization distortion increases. In contrast, the vector quantization is
Even with a small number of bits, the quantization distortion is small and the correlation between words can be largely removed, so that the quantization efficiency is high.

However, even in this vector quantization, the quantization efficiency cannot be said to be sufficiently high yet, and a higher efficiency quantization is desired.

The present invention has been made to solve such a problem, and an object of the present invention is to provide a quantization device and a quantization method that can perform quantization with higher efficiency.

[Means for solving the problem]

A quantization device according to the present invention has been proposed to achieve the above object, and generates and outputs a matrix composed of a two-dimensional data array of a time axis and an axis orthogonal to the time axis for an input signal. Means and a code matrix most similar to a matrix consisting of a two-dimensional data array of a time axis and an axis orthogonal to the time axis generated from a code book provided with a plurality of code matrices and identification codes of the code matrices. And a matrix quantizer for outputting the identification code.

Further, the quantization method of the present invention generates a matrix consisting of a two-dimensional data array of a time axis and an axis orthogonal to the time axis for an input signal, and is provided with a plurality of code matrices and identification codes of these code matrices. A code matrix most similar to a matrix composed of a two-dimensional data array of the generated time axis and an axis orthogonal to the time axis is selected from the generated codebook.

The frequency axis means an axis orthogonal to the time axis in a broader sense, and includes, for example, a sequence axis by Hadamard transform.

[Action]

According to the present invention, an identification code of a code matrix most similar to a matrix composed of a two-dimensional data array of a time axis and a frequency axis for an input signal is obtained. Therefore, it is possible to remove the correlation of the input signal on the time axis and the frequency axis.

〔Example〕

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 shows a schematic configuration of an apparatus according to an embodiment of the present invention. The quantizing apparatus shown in FIG. ₁ is preferably applied to, for example, each of the quantizers 134 _{1 to} 134 _n of the apparatus for performing the band division coding shown in FIG.

In the apparatus shown in FIG. 1, an input signal such as sound from the input terminal 1 is, for example, QMF (quadrature mirror filter).
r), the input signal is supplied to a filter bank 10 which is a matrix output means for outputting a matrix composed of a two-dimensional data array of a time axis and an axis orthogonal to the time axis for the input signal. That is, the filter bank 10 divides the input signal into a plurality of (n-channel) frequency bands, and outputs a time waveform of these band-divided signals and a time waveform (axis) orthogonal to the time axis of the time waveform. Time waveform of each channel at the same time in adjacent bands)
It is output as a matrix of a dimensional data array.
This output matrix is supplied to the matrix quantizer 13 after being subjected to a floating process described later. The matrix quantizer 13 has a code book in which a plurality of code matrices and identification codes of these code matrices are provided in advance, and selects a code matrix most similar to the output matrix from the filter bank 10 to select The identification code is output. The identification code selected in this way is output from the output terminal 2.

In the present embodiment, the input signal is sampled at, for example, a sampling frequency fs = 32 kHz, and the filter bank 10 has, for example, n = 4 channels, that is, a division number of 4
It is assumed that a bandwidth of 0 to 15 kHz is divided. In this case, the frequency range of each band is 0 to 4 kHz, 4 to 8 kHz, 8 to 12 kHz.
z, 12 to 15 kHz.

The output matrix at this time is a k-dimensional column vector (the number of vector elements is k) in which the sample sequence of k samples in the time axis direction for each channel of the waveform of each band of four channels as shown in FIG. , For example, k = 4), and these are combined for four channels to produce a (k × 4) matrix. When k is 4, a (4 × 4) matrix as shown in FIG. 3 is obtained. This can be regarded as data of k rows (for example, 4 rows) of the four-dimensional row vector data at the same time of each channel forming each band. That is, in the example of FIG. 3, the sampling data S ₁₁ to S ₄₁ were collected each sampling data S 4 pieces of time waveform of each channel of the second view in the channel direction, S ₁₂ to S _42, S ₁₃ ~S _43, S ₁₄ ~S ₄₄
4 rows of each row vector or each sampling data
_{S 11 ~S 14, S 21 ~S} 24, S 31 ~S 34, S 41 2 -dimensional data sequence composed of 4 rows of column vector to S ₄₄ (matrix, matrix) has become.

Next, floating processing is performed on each data set as the output matrix as described above. That is, each data divided into each band of the four channels are respectively transmitted to the effective value calculating circuit 11 ₁ to 11 _4, the these effective value calculating circuit 11 ₁ to 11 _4, the effective value calculating i.e. general RMS A calculation is made.

The effective value obtained by such an RMS calculation becomes a floating coefficient, and the floating coefficient is sent to the floating coefficient quantizer.

In the floating coefficient quantizer 14, a plurality of floating coefficients are defined as one block, and the one block is defined as one block.
Vector quantization for quantizing as one vector is performed. Here, the floating coefficient quantizer 14 includes:
There is a code book in a memory in which a code vector (representative vector) associated with an index (identification code of vector quantization) is recorded. The vector of the floating coefficient and the stored code vector are, for example, distance calculation or the like. , The closest (similar) code vector is found, and the index corresponding to that code vector is read. The index and the code vector corresponding to the index are output from the floating coefficient quantizer 14, and the index is output from the terminal 3 as auxiliary information as shown in FIG.
Further, the code vector corresponding to the vector of the floating coefficients is fed to the divider 12 ₁ to 12 _n.

The divider 12 ₁ to 12 _n, the vector of the divided and data of each channel is supplied with, for each channel of at該除adder 12 ₁ to 12 _n data and the floating coefficient from the filter bank 10 Is performed with each data of the corresponding code vector, thereby performing floating processing of the data of each channel. The matrix composed of the data of each channel subjected to the floating process in this way is sent to the matrix quantizer 13.

As described above, the matrix quantizer 13 has a code book provided with a plurality of code matrices and identification codes. The matrix quantizer 13 compares the input matrix which has been subjected to the floating processing with the input matrix and the code matrix, for example, by calculating a distance, thereby obtaining a code closest to (similar to) the input matrix. A matrix is selected, and an identification code corresponding to the code matrix is output. At this time, matrix quantization may be performed so that adaptive bit allocation as described later is performed.

That is, with respect to the signal of each channel in FIG. 1, the bit allocation (bit allocation) is changed for each channel according to the RMS value obtained for each of the k samples, that is, adaptive bit allocation is performed.
In this case, the bit allocation for each column vector in the matrix of FIG. 3 changes for each matrix. A specific example of such a quantization of a matrix composed of vectors whose bit allocation can change will be described in detail in the description of a second embodiment described later.

From the above, in the present embodiment,
Since the quantization of the matrix consisting of the time axis of the input signal and the two-dimensional data array on the axis orthogonal to the time axis is performed, the dependency on the change on the time axis of the input signal can be removed, and each channel can be removed. Efficient quantization that can also remove the correlation between them becomes possible. Further, the quantization efficiency can be further increased by performing adaptive bit allocation.

Next, as a second embodiment of the present invention, the quantization apparatus according to the present invention performs transform coding, that is, coefficients (FFT) obtained by performing orthogonal transform as in the above-described adaptive transform coding.
A description will be given of a case where the present invention is applied to quantization of coefficients, DCT coefficients, and the like.

That is, in the second embodiment, the input signal on the time axis is, for example, FFT processed by the fast Fourier transform (FFT).
A plurality of FT coefficients are grouped in the time axis direction and arranged two-dimensionally to form a matrix, and this matrix is subjected to the same matrix quantization processing as described above.

Here, the matrix of the two-dimensional data array in the second embodiment will be described with reference to FIGS. 4 to 6.

An audio signal or the like on the time axis as shown in FIG. 4 is divided into, for example, periods A, B, and C, and the audio signal in each of the periods A, B, and C is divided into frequency axes (an axis that is generally orthogonal to the time axis). FIG. 5 shows the frequency spectrum data of each of the periods A, B, and C together with the time axis. In the present embodiment, for example, the sixth
As shown in the figure, four data groups a _{1 to} a ₄ , b _{1 to} b ₄ , each of which is surrounded by a curve on the frequency axis for periods A, B, and C in FIG.
Each of c _{1 to} c _{4 and the} like is defined as a row vector, and these row vectors are grouped together in the time axis direction, for example, three (3 rows), and (4 ×
3) constitutes a two-dimensional data array matrix. In other words, from a different perspective, the data groups a _{1 to} c ₁ , a _{2 to} c ₂ , a _{3 to} c ₃ , and a _{4 to} c ₄ in the time axis direction are column vectors, and this column vector is 4 pieces (4 rows)
Collectively, a two-dimensional data array matrix (4 × 3 data) is obtained.

Hereinafter, the operation of the matrix quantization process when the frequency axis of the matrix of the two-dimensional data array, that is, the axis that is generally orthogonal to the time axis is, for example, an FFT coefficient as described above will be described with specific examples. Will be explained.

Input signals such as audio signals are converted to sampling frequency fs
= Frequency Fourier Transform (FFT) processing is performed for every predetermined number of sample data obtained by sampling at 32 kHz, for example, for every 1024 points, and 512 points of FFT coefficients are obtained between 0 and 16 kHz. Can be Here, the FFT coefficient (complex number)
Is converted into amplitude and phase data using the following equations (2) and (3). That is, Amp = (Re ² + Im ² ) ^1/2 ‥‥‥‥‥ (2) θ = tan ⁻¹ (Im / Re) ‥‥‥‥‥ (3) where Amp is amplitude, Re is a real number, Im indicates an imaginary number and θ indicates a phase. As described above, the FFT coefficient is calculated by the second equation and the third equation.
By converting to amplitude and phase data by the formula, 0 to 16 kHz
By then, spectra at 31.25 Hz intervals are required. The spectrum obtained in this way is divided into a plurality of bands. For example, as shown in FIG. 7, in consideration of human auditory characteristics, 0 to 16 kHz is divided into 24 bands, each of which is close to a so-called critical bandwidth (critical band). Each piece of data thus divided by the critical band is formed into a matrix of a two-dimensional data array and quantization is performed.

At this time, floating processing is performed on the data of each of the divided bands, for example, as shown in FIG. The floating coefficient of this floating process is determined in consideration of, for example, the power of the input signal, or the presence of a strong signal, which is one of the human auditory characteristics, causes weak noise in the vicinity of the signal to be weak. It is required in consideration of so-called masking or the like, which is masked and the interference effect is reduced. That is, for example, FF
The T-spectrum envelope or the quantized average level of each band is used as a floating coefficient, and the above-described floating coefficient is used to perform the above-described floating processing for interrupting the FFT coefficient of each band.

Next, as a matrix quantization process, the FFT coefficient of each band subjected to the floating process as described above,
A two-dimensional data array matrix is arranged in the time axis direction, and this matrix is subjected to matrix quantization in the same manner as in FIG. That is, as shown in FIG. 8, for example, a row vector composed of the FFT coefficients at the intervals of 31.25 Hz, for example, F ₁₁ , F ₂₁ , F ₃₁ , and F ₄₁ is 1024 × (1 / 32k) = 32 [ms] Are obtained every 32 ms, and three rows of these row vectors, for example,
FFT coefficients F ₁₁ to F row vector L ₁ made of _41, F ₁₂ to F row vector L ₂ of _42, F ₁₃ row vector L ₃ of to F ₄₃ is summarized as a matrix (matrix) of (4 × 3) Is done. This matrix is input to a matrix quantizer, and the similarity with a plurality of code matrices (representative matrices) stored in a code book in advance, for example, the distance, is calculated. An index (identification code) corresponding to the matrix is output. Note that the critical band described above has a wide bandwidth in a high frequency range, and thus it is necessary to form a plurality of matrices in this high frequency range.

Here, an example in which the above-described adaptive bit allocation is applied to the matrix on the time axis and the frequency axis as described above will be described.

In the case of performing this adaptive bit allocation, the number of bits allocated to each band obtained by dividing a plurality of bands in the frequency axis direction as shown in FIG. We decide by taking the average. At this time, the interval between each row vector is
Since the width is as large as 32 ms, when the change of the input signal is large, the number of bits assigned to each band may change, and the number of bits assigned to each row vector may differ even in the same matrix.

For example, in the matrix shown in FIG. 8 consisting of the FFT coefficients F _{11 to} F ₄₃ , that is, a matrix in which three rows of row vectors each consisting of 4-word coefficients are grouped, for example, each coefficient is represented by 1
As a specific example of the number of bits adaptively allocated to each row vector when performing the above-described matrix quantization, _one word is used for each word of the row vector L1.
Bit, 2 bits to each word of the row vector L ₂ is a 3-bit to each word of the row vector L ₃ is assigned.

Matrix quantization according to such bit allocation can be realized with a circuit configuration as shown in FIG. In FIG. 9, an input matrix consisting of 12 words (4 × 3 words) of the 16-bit FFT coefficient is input to an input terminal 41, and a first matrix quantization (M
Q) Supplied to the circuit 42. The matrix quantizer 42 performs matrix quantization such that the input matrix (12 words) is represented by one bit per word, that is, a 12-bit index is output.
Thus, from the matrix quantizer 42, the identification code of 12 bits corresponding to the most similar code matrix and the input matrix is obtained and outputted from the output terminal 47 as an output code C _A. Further, the code matrix from the matrix quantizer 42 and the input matrix are subjected to subtraction processing by the subtractor 43 to obtain 12 (= 4 × 3).
A word quantization error is obtained. Of the quantization error, the row quantization error vector corresponding to the vector L ₁ is removed, the respective middle and lower row vector L ₂ and L ₃ corresponding 8 (= 4 × 2) words quantization The error matrix em is sent to the matrix quantizer 44 and the subtractor 45, respectively. In the matrix quantizer 44, the input matrix (4 × 2 = 8 words) is subjected to matrix quantization such that it is expressed by one bit per word, that is, by eight bits. 8-bit identification code is obtained from the identification code is outputted from the terminal 48 as an output code C _B. The code matrix from the matrix quantizer 44 is sent to a subtractor 45, where the quantization error matrix em
And a code matrix are subtracted to obtain a quantization error. The quantization error is sent to a vector quantization (VQ) unit 46. However, this time, from the quantization error quantization error vector corresponding to the middle row vector L ₂ is removed, the above-mentioned vector quantizer 46 quantization error vector data ev corresponding to the row vector L ₃ is Sent. Therefore, the vector data ev
Index of codevector vector quantization corresponds to the decorated with with vector data ev is obtained, this index is output from an output terminal 49 as an output code C _C.

By performing the matrix quantization processing as described above, it is possible to perform adaptive bit allocation to each row vector. That is, the respective output code C _A -C _C The matrix quantization process described above by performing reverse processing, 1 bit for each word of the input matrix row vector L ₁ is assigned, a row vector L ₂ each word of 2 bits, 3 bits to each word of the row vector L ₃ is, to be allocated respectively.

This can be similarly performed when the matrix composed of a plurality of columns of the column vector is vector-quantized by changing the bit allocation for each column as in the first embodiment.

As described above, if the quantization apparatus according to the present embodiment is used for adaptive transform coding or the like that performs orthogonal transform, for example, it is possible to not only remove the correlation between adjacent spectra on the frequency axis, but also change the time axis. (Non-linear) dependence can be removed, so that efficient quantization of data can be performed. In addition, since adaptive bit allocation can be performed, efficiency can be further improved.

When the present embodiment is applied to the above-described band division coding and adaptive transform coding, it is preferable to perform the following analysis window variable length processing in order to efficiently and adaptively allocate bits. That is, at the time of matrix quantization processing, in addition to the above-described adaptive bit allocation for each word in the matrix, by performing processing of changing the window (analysis window) from which the data of the matrix is extracted to a variable length. , More efficient quantization can be performed.

For example, calculate the energy for each block of the input signal (signal block taken out in the analysis window), and when the energy difference between adjacent blocks is large, use a separate analysis block as a transient part, and when the energy difference is small, Treats, for example, two blocks together as one analysis block as a stationary part. Here, as the analysis window length which is a variable length,
By using two or more types of analysis windows (time windows) and separating / combining the analysis windows, the analysis window length (time window length) can be made variable. For example, in the stationary part, the time resolution is reduced or the frequency resolution is increased. In the transition section, the time resolution is increased or the frequency resolution is decreased.
As a result, analysis, that is, bit allocation, which is dynamically adapted to signal characteristics can be performed.

For example, in band division coding, in the transient part of a signal, the output of each divided band (sub-band) of multiple channels is synthesized, that is, the bandwidth is widened, and the time response is shortened (quickly), so that efficient bit allocation is achieved. It becomes possible. Also, in adaptive transform coding, by assigning a shorter time window in the transient part of the signal and increasing the time window in the stationary part, it becomes possible to allocate bits in accordance with the characteristics of the signal.

〔The invention's effect〕

In the quantization device and the quantization method of the present invention, since the matrix composed of the two-dimensional data array on the time axis and the axis orthogonal to the time axis for the input data is quantized, the time variation of the signal on the time axis is Not only can the dependency be removed, but also the correlation of the axis orthogonal to the time axis such as the frequency axis can be removed, and highly efficient quantization can be performed.

Therefore, by applying the quantization apparatus and the quantization method to various high-efficiency coding techniques, it is possible to realize higher-efficiency coding.

[Brief description of the drawings]

FIG. 1 is a block circuit diagram showing a schematic configuration of a quantization device according to one embodiment of the present invention, FIG. 2 is a waveform diagram showing an input signal waveform for each channel, and FIG. 3 is a two-dimensional data array of one embodiment. FIG. 4 is a waveform diagram showing an audio signal waveform, FIG. 5 is a diagram showing a frequency spectrum, FIG. 6 is a diagram showing a matrix of a two-dimensional data array of the example of FIG. 5, and FIG. FIG. 8 is a diagram showing a critical band, FIG. 8 is a diagram showing a matrix of a two-dimensional data array of another embodiment, FIG. 9 is a block circuit diagram showing a configuration for adaptive bit allocation of another embodiment, FIG. 10 is a block circuit diagram showing a schematic configuration of an apparatus for performing band division coding. 10 ... Filter bank 11 _{1 to} 11 ₄ ... Effective value calculation circuit 12 _{1 to} 12 ₄ ... Divider 13,42,44 ... Matrix quantizer 14 ... Floating coefficient quantizer 46 ... Vector quantization Unit 43,45 …… Subtractor

Continuation of the front page (72) Inventor Kenzo Akagiri 6-7-35 Kita-Shinagawa, Shinagawa-ku, Tokyo Sony Corporation (72) Inventor Yoshihito Fujiwara 6-35, Kita-Shinagawa, Shinagawa-ku, Tokyo Sony Stock In-house (56) References JP-A-62-139089 (JP, A) JP-A-1-158496 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) H03M 7/30

Claims (2)

(57) [Claims] 1. A matrix output means for generating and outputting a matrix composed of a two-dimensional data array of a time axis and an axis orthogonal to the time axis for an input signal, a plurality of code matrices and identification codes of these code matrices. And a matrix quantizer for selecting a code matrix most similar to a matrix formed of a two-dimensional data array of the generated time axis and an axis orthogonal to the time axis from the generated code book and outputting the identification code. A quantization device characterized by comprising: 2. A matrix comprising a two-dimensional data array of a time axis and an axis orthogonal to the time axis for an input signal is generated. From a code book provided with a plurality of code matrices and identification codes of these code matrices, A quantization method, comprising: selecting a code matrix most similar to a matrix composed of the two-dimensional data array of the generated time axis and an axis orthogonal to the time axis.