JP2000132194A - Signal encoding device and method therefor, and signal decoding device and method therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the calculation quantity for bit assignment in the quantization of coefficient data on an orthogonally transformed frequency axis by a bit assignment according to an input, and perform a stable quantization even to a signal largely changed in bit assignment between frames. SOLUTION: Coefficient data on frequency axis obtained by orthogonally transforming a time base input signal is inputted to an input terminal 1, and divided into some bands by a band dividing circuit 3. A weight calculation circuit 2 calculates the weight of each orthogonally transformed coefficient data on the basis of a parameter such as linear predicted coefficient according to the time base input signal, and this weight is also divided by the band dividing circuit 3. The band-divided weight and coefficient are transmitted to sort circuits 40-4L-1, and the coefficients within the band are sorted (rearranged) according to the weight for every band, and vector-quantized.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a signal encoding apparatus and method for quantizing an input signal by converting the time axis / frequency axis and a signal decoding apparatus and method, and more particularly to a high efficiency encoding of an audio signal. The present invention relates to a signal encoding device and a method suitable for the above-described case, and a signal decoding device and a method.

[0002]

2. Description of the Related Art Conventionally, there has been proposed an encoding method in which signal compression is performed by utilizing the statistical properties of an audio signal (including a voice signal and a music signal) in a time domain and a frequency domain and characteristics of human hearing. Various are known. This encoding method is roughly classified into encoding in the time domain, encoding in the frequency domain, and analysis-synthesis encoding.

[0003]

By the way, in transform coding for orthogonally transforming an input signal on the time axis into a signal on the frequency axis and performing coding, a dynamic bit corresponding to the input signal is used for the purpose of reducing the rate. It is proposed to perform allocation and quantize coefficient data on the frequency axis.
The calculation of this bit allocation is complicated, and especially when the coefficient data on the frequency axis is divided into several sub-vectors and vector-quantized, if the bit allocation for each coefficient changes, Calculation of bit allocation is complicated.

[0004] In addition, when the bit allocation is extremely changed for each frame which is a conversion unit of the orthogonal transform, there is a disadvantage that the reproduced sound is likely to be unstable.

[0005] The present invention has been made in view of such circumstances, and in coding with orthogonal transformation, while performing dynamic bit allocation according to an input signal,
It is an object of the present invention to provide a signal encoding device and method, and a signal decoding device and method that can easily calculate bit allocation and that does not make reproduced sound unstable even when bit allocation changes greatly between frames. .

[0006]

SUMMARY OF THE INVENTION In order to solve the above-mentioned problems, the present invention provides a method for performing weighting on an input signal on a time axis by using an orthogonal transform in accordance with the input signal. Is calculated, and the coefficient data obtained by the orthogonal transformation is ranked in accordance with the order of the weights, and highly accurate quantization is performed in accordance with the rank.

In the above-mentioned quantization, the higher the coefficient data in the order, the larger the number of bits allocated for quantization.

Further, the coefficient data obtained by the orthogonal transformation is divided into a plurality of bands on a frequency axis, and the coefficient data in each band is independently ranked for each band in accordance with the order of the weights. It is preferable to perform quantization.

[0009] It is preferable that the coefficient data is divided into a plurality of coefficient data from the coefficient data on the higher rank in the ranked order to obtain respective coefficient vectors, and these coefficient vectors are respectively vector-quantized.

[0010]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram illustrating a schematic configuration of a quantization circuit used in a signal encoding device according to an embodiment of the present invention.

In FIG. 1, coefficient data on a frequency axis obtained by orthogonally transforming a time axis signal is supplied to an input terminal 1, and a weight calculation circuit 2 is provided with, for example, LP.
Parameters such as a C coefficient, a pitch parameter, and a bark scale factor are input. In the weight calculation circuit 2,
The weight w is calculated based on such parameters.
Here, it is assumed that a coefficient for one frame of the orthogonal transformation is represented by a vector y and a weight for one frame is represented by a vector w.

The coefficient vector y and the weight vector w are divided into L (L ≧ 1) bands by sending them to the band dividing circuit 3 as necessary. The number of bands is, for example, about three bands of low band, middle band, and high band (L =
3), but the invention is not limited to this, and the band does not have to be divided. The coefficient for each band, for example, k-th
The coefficient of the th band is y _k , and the weight is w _k (0 ≦ k ≦
When L-1) _{and, y = (y 0, y} 1, ..., y L-1) w = (w 0, w 1, ..., a w _L-1). The number of bands in this band division and the number of coefficients for each band are fixed to preset numerical values.

Next, the coefficient vector y ₀ , y of each band
_1, ..., y _L-1 the sorting circuit 4 respectively _0, 4 _1, ..., 4
_L-1 and ranks the coefficients in each band according to the order of the weight for each band. this is,
The coefficients in each band may be rearranged (sorted) according to the order of the weights. However, only the index (index) indicating the position or order of each coefficient on the frequency axis is sorted in the order of the weights. The precision (quantity of allocated bits, etc.) at the time of quantization of each coefficient may be determined according to the index (index). In the case of sorting the coefficients themselves, the coefficients of the coefficient vector y _k are sorted in the order of weight for an arbitrary k-th band, and a coefficient vector y ′ _k sorted in the order of weight is obtained.

Next, the coefficient vectors y ′ ₀ , y ′ ₁ ,..., Y ′ _L−1 sorted for each band according to the order of the weights are vector quantizers 5 ₀ , 5 ₁ ,. ., 5 _L-1 to perform vector quantization.

Next, each vector quantizer 5 ₀ ,
The vectors c ₀ , c ₁ ,..., C _L-1 of the coefficient indexes for each band from 5 ₁ ,..., 5 _L−1 are combined into a vector c of the coefficient indexes of all the bands, and the terminal It is taken out from 6.

The more specific operation of the quantization circuit shown in FIG. 1 will be described later with reference to FIGS.

With such a configuration, even if the weight for each coefficient dynamically changes for each frame, the coefficients sorted in the order of the weight can be vector-quantized in order, and the bit allocation processing can be simplified. In addition, by fixing the number of allocated bits for each band, the number of bits allocated to each band is stable even for a signal whose weight changes significantly between frames, and a stable reproduced sound is obtained. can get.

Although the signal encoding apparatus shown in FIG. 1 is shown as a hardware configuration, it can be realized as software using a so-called DSP (digital signal processor) or the like.

Next, an audio signal encoding apparatus as a more specific configuration example using the quantization circuit of the above-described signal encoding apparatus according to the embodiment of the present invention will be described with reference to FIG.

In the audio signal encoding apparatus shown in FIG. 2, the supplied time axis signal is subjected to time axis / frequency axis conversion (T / F conversion) by, for example, MDCT (improved discrete cosine transform) in the orthogonal transform unit 25. The data is subjected to encoding by performing data encoding on the frequency axis (MDCT coefficients) and quantizing the data in the coefficient quantization unit 40. In this embodiment, the time axis signal before the orthogonal transform is performed. For L
The characteristics of the input signal waveform are extracted by PC analysis, pitch analysis, envelope extraction, and the like, and parameters representing these characteristics are separately quantized and extracted. The normalization (whitening) circuit unit 11 removes these characteristics. Alternatively, the coding efficiency is increased by removing the correlation of the signal to make the signal close to white noise, that is, a so-called noise-like signal.

Further, when deciding the bit allocation (bit allocation) at the time of quantizing the coefficient data after the orthogonal transformation,
The LPC coefficient obtained by the LPC analysis and the pitch parameter obtained by the pitch analysis are used. In addition, a bark scale factor may be used as a normalization factor by extracting a peak value, an rms value, and the like for each critical band (critical band) on the frequency axis. Based on these LPC coefficients, pitch parameters, and bark scale factors, weights at the time of quantization for orthogonal transform coefficient data such as MDCT coefficients are calculated, thereby determining bit allocation for all bands and performing coefficient quantization. When the weight determination at the time of quantization is made by a predefined parameter,
For example, when only the LPC coefficient, the pitch parameter, and the bark scale factor are used, only these parameters are transmitted to the decoder side, and the same bit allocation (bit allocation) as the encoder side is performed.
Is reproduced, there is no need to send additional information (side information) on the bit allocation itself.

Further, at the time of coefficient quantization, coefficient data is rearranged (sorted) in an order according to the weight at the time of quantization or the number of allocated bits, and quantization with high precision is performed in order. . In this quantization, it is preferable to divide the sorted coefficients into sub-vectors in order from the top and perform vector quantization for each. Sorting may be performed on coefficient data of all bands, or may be divided into several bands and sorted for each band. In this case as well, if the parameters used for bit allocation are specified in advance, the bit allocation, sorting order, etc. can be performed without directly transmitting the bit allocation information or the position information of the sorted coefficients by simply transmitting the parameters. Can be reproduced on the decoder side.

In FIG. 2, a digital audio signal obtained by A / D-converting a so-called wideband audio signal of, for example, about 0 to 8 kHz at a sampling frequency Fs = 16 kHz is supplied to an input terminal 10. This input signal is
The signal is sent to the LPC inverse filter 12 of the normalization (whitening) circuit unit 11 and cut out, for example, by 1024 samples, and sent to the LPC analysis / quantization unit 30. This L
The PC analysis / quantization unit 30 calculates an LPC coefficient of about the 20th order, that is, an α parameter after applying a Hamming window, and obtains an LPC residual by the LPC inverse filter 11. At the time of this LPC analysis, 1 which is a unit of analysis is used.
Some of the 1024 samples in the frame, for example, one half of 512 samples are overlapped with the next block, and the frame interval is 512 samples. This is to use the aliasing cancellation of MDCT (improved discrete cosine transform) adopted as the orthogonal transform at the subsequent stage. This LP
The C analysis / quantization unit 30 converts the α parameter, which is an LPC coefficient, to an LSP (line spectrum pair) parameter and transmits the result after quantization.

The α parameter from the LPC analysis circuit 32 is sent to the α → LSP conversion circuit 33 and is converted into a line spectrum pair (LSP) parameter. This converts the α parameter obtained as a direct type filter coefficient into, for example, 20, ie, 10 pairs of LSP parameters. The conversion is performed using, for example, the Newton-Raphson method. The conversion to the LSP parameter is because it has better interpolation characteristics than the α parameter.

The LSP parameter from the α → LSP conversion circuit 33 is vector-quantized or matrix-quantized by an LSP quantizer 34. At this time, vector quantization may be performed after obtaining the difference between frames, or matrix quantization may be performed on a plurality of frames at once.

The quantized output from the LSP quantizer 34, that is, the index of LSP vector quantization is taken out via a terminal 31, and the quantized LSP vector or the dequantized output is converted to an LSP interpolation circuit 36. And the LSP → α conversion circuit 38.

The LSP interpolation circuit 36 is an LSP quantizer 3
4 interpolates the set of the previous frame and the current frame of the vector of the LSP that has been vector-quantized for each frame and sets the rate required for the subsequent processing. Interpolated to rate.

In order to execute the inverse filtering of the input voice using the LSP vector on which the interpolation has been performed, the LSP → α conversion circuit 37 converts the LSP parameter into, for example, a coefficient of a direct type filter of about the 20th order. Convert to α parameter. The output from the LSP → α conversion circuit 37 is sent to an LPC inverse filter circuit 12 for obtaining the LPC residual, and the LPC inverse filter 12 performs an inverse filtering process using an α parameter updated at an eight-fold rate. To obtain a smooth output.

Further, the 1-times-rate LSP coefficient from the LSP quantization circuit 34 is sent to an LSP → α conversion circuit 38 to be converted into an α parameter, and a bit allocation calculation for performing the above-described bit allocation is performed. It is sent to a circuit (bit allocation determining circuit) 41. The bit allocation calculation circuit 41 also calculates a weight w (ω) used for quantizing MDCT coefficients, which will be described later, in addition to the allocated bits.

The output from the LPC inverse filter 12 of the normalization (whitening) circuit section 11 is sent to a pitch inverse filter 13 and a pitch analysis circuit 15 for pitch prediction as long-term prediction.

Next, the long-term prediction will be described. The long-term prediction is performed by subtracting the waveform shifted on the time axis by the pitch period or pitch lag obtained by the pitch analysis from the original waveform to obtain a pitch prediction residual. In this example, three-point pitch prediction is performed. Has gone by. The pitch lag refers to the number of samples corresponding to the pitch cycle of the sampled time axis data.

That is, the pitch analysis circuit 15 performs the pitch analysis once per frame, that is, the analysis length is one frame, and the pitch lag in the pitch analysis result is sent to the pitch inverse filter 13 and the bit allocation calculation circuit 41. The pitch gain is sent to a pitch gain quantizer 16. The pitch lag index from the pitch analysis circuit 15 is extracted from the terminal 52 and sent to the decoder.

In the pitch gain quantizer 16, the pitch gain at the three points corresponding to the three-point prediction is vector-quantized, a codebook index (pitch gain index) is extracted from the output terminal 53, and the representative value vector or The inverse quantization output is sent to the pitch inverse filter 13. The pitch inverse filter 13 outputs a pitch prediction residual whose three-point pitch is predicted based on the pitch analysis result. The pitch prediction residual is sent to the division circuit 14 and the envelope extraction circuit 17, respectively.

The pitch analysis will be further described. In this pitch analysis, a pitch parameter is extracted using the LPC residual. The pitch parameter is composed of a pitch lag and a pitch gain.

First, the pitch lag is determined. The above LPC
For example, 512 samples are cut out from the center of the residual, and x (n)
(N = 0 to 511), and is represented by x. Assuming that 512 samples in the past k samples from x are x _k , the pitch k is given as minimizing _{ x−gx _k } ² . _{That, g = (x, x k} ) as / ‖X _k ‖ ^2, by searching k that maximizes (x, x _k) a ² / ‖X _k ‖ ^2, can determine the optimal lag K. In the present embodiment, K is 12 ≦ K ≦ 240. This K may be used as it is, or a tracking result using a pitch lag of a past frame may be used. For K determined in this manner, 3
The optimum pitch gain at the point (K, K-1, K + 1) is obtained. That is, g ₋₁ , g ₀ , and g ₁ that minimize {x− (g ₋₁ × _{L + 1} + g ₀ × _L + g ₁ × _L−1 )} ² are obtained, and the three-point pitch gain for the optimal lag K is obtained. And The three-point pitch gain is sent to a pitch gain quantizer 16 and vector-quantized as a whole. Further, a pitch inverse filter 13 is formed using the quantized pitch gain and the optimal lag K. Find the residual. The obtained pitch residual is linked with the past pitch residual that has already been obtained, and is subjected to MDCT conversion as described later. At this time, time axis gain control may be performed before MDCT conversion.

Here, FIG. 3 shows the above L with respect to the input signal.
It shows the relationship between PC analysis processing and pitch analysis processing,
The analysis length of 1024 samples per frame FR is, for example,
It has a length corresponding to the MDCT transform block described later. Time t ₁ is the current new LPC analysis center (LSP
₁ ), and time t ₀ indicates the LPC analysis center (LSP ₀ ) one frame before. The second half of the current frame is new data ND, and the first half is previous data (previous
data) is PD, the LPC residuals a is obtained by interpolation of LSP ₀ and LSP ₁ in Figure, b is the LPC residual of the previous frame, c is the portion (the first half of the second half + a of b) A new pitch residual obtained from the target pitch analysis, and d indicates a past pitch residual.
When all of the new data ND in FIG. 3 has been input, data a can be obtained, and a new pitch residual c can be calculated using this a and b which has already been obtained. Of the frame F to be orthogonally transformed by connecting the pitch residual d
R can be created. The data of this one frame FR is MDC
Orthogonal transformation processing such as T can be performed.

Next, FIG. 4 shows an LPC based on LPC analysis.
FIG. 5 shows the change of the time axis signal through the inverse filter and the pitch inverse filter based on the pitch analysis.
The change on the frequency axis of the signal through the C inverse filter and the pitch inverse filter is shown, respectively. That is, FIG. 4A shows an input signal waveform, and FIG. 5A shows its frequency spectrum, and the characteristic of the waveform is extracted and removed by passing through an LPC inverse filter based on LPC analysis. Thus, a substantially periodic pulse-like time axis waveform (LPC residual waveform) as shown in FIG. 4B is obtained. The spectrum on the frequency corresponding to this LPC residual waveform is
The result is as shown in FIG. By passing the LPC residual through a pitch inverse filter based on the pitch analysis as described above, a pitch component is extracted and removed.
5C shows a time-axis signal close to white noise (noise-like), and its spectrum on the frequency axis is as shown in FIG.

Further, in the embodiment of the present invention,
In the normalization (whitening) circuit section 11, the gain of the data in the frame is smoothed. This corresponds to the time axis waveform in the frame (in this embodiment, the pitch inverse filter 13
From the residual), an envelope is extracted by an envelope extraction circuit 17, and the extracted envelope is sent to an envelope quantizer 20 via a switch 19, and the time axis waveform (pitch) is calculated based on the value of the quantized envelope. By dividing the residual (residual from the inverse filter 13) by the divider 14, a signal smoothed on the time axis is obtained. The signal from the divider 14 is normalized (whitened).
The output of the circuit section 11 is sent to the orthogonal transformation circuit section 25 in the next stage.

By this smoothing, so-called noise shaping can be realized in which the magnitude of a quantization error when the orthogonal transform coefficient after quantization is inversely transformed into a time signal follows the envelope of the original signal.

The extraction of the envelope by the envelope extraction circuit 17 will be described. The signal supplied to the envelope extracting circuit 17, that is, the residual signal normalized by the LPC inverse filter 12 and the pitch inverse filter 13 is x (n), n = 0 to N-1 (where N is Number of samples of frame FR, orthogonal transformation window length, for example, N
= 1024), the envelope is a rms (root mean square) value for each sub-block or sub-frame extracted by a length M shorter than the conversion window length N, for example, a window of M = N / 8. . That is, as the rms of each normalized sub-block (sub-frame),
rms _{i of the} i-th sub-block (i = 0 to M−1)
Is defined by the following equation (1).

[0041]

(Equation 1)

Scalar quantization is performed on each i of rms _i obtained by the above equation (1), or rms _i
Vector quantization can be performed with the whole as one vector. In the present embodiment, the envelope quantizer 2
At 0, vector quantization is performed, and the index is taken out from the terminal 21 as a parameter for time axis gain control, that is, an envelope index, and transmitted to the decoder side.

The rms _i of each of the sub-blocks (sub-frames) quantized in this way is defined as qrms _i, and the input residual signal x (n) is interrupted by the divider 14 based on this value, whereby time is obtained. Axis-smoothed signal x
_s (n) is obtained. However, the rms _i obtained in this manner is
Of the ratio of the largest to the smallest in the frame,
When the value is equal to or more than a certain value (for example, 4), the above-described gain control is performed, and a predetermined number of bits (for example, 7) is used for quantization of the parameter (the envelope index).
While assigned a bit), the normal processing is not performed gain control when the ratio of the rms _i for each sub-block (sub-frame) in the frame is less than the predetermined value, a bit for the gain control Is assigned to quantization of other parameters, for example, frequency axis parameters (orthogonal transform coefficient data). The determination as to whether or not to perform the gain control is performed by the gain control on / off determination circuit 18, and the determination output (gain control SW) is used as a switching control signal for the switch 19 on the input side of the envelope quantizer 20. The signal is sent to the coefficient quantization circuit 45 in the coefficient quantization unit 40, which will be described later, and is used to switch the number of allocated bits of the coefficient when the gain control is on and off. The gain control on / off discrimination output (gain control SW) is connected to terminal 2
2 and sent to the decoder side.

The signal x that has been gain-controlled (or gain-compressed) by the divider 41 and smoothed on the time axis
_s (n) is sent to the orthogonal transformation circuit unit 25 as an output of the normalization circuit unit 11, and is converted into a frequency axis parameter (coefficient data) by, for example, MDCT. The orthogonal transformation circuit unit 25 includes a windowing circuit 26 and an MDCT circuit 27. The windowing circuit 26 performs windowing by a window function such that the aliasing cancellation of MDCT by 1/2 frame overlap can be used.

At the time of decoding on the decoder side, inverse quantization is performed from the quantization index of the transmitted frequency axis parameter (for example, MDCT coefficient), and then the time is inversely orthogonally transformed as the frequency axis / time axis conversion. Using the time axis gain control parameter that has been returned to the axis signal and then inversely quantized, overlap addition and inverse processing of gain smoothing at the time of encoding (gain expansion or gain restoration processing) are performed. However, when gain smoothing is used, overlap addition assuming a window in which a normal symmetric and sum of squares of the window value at the superimposed position becomes a constant value cannot be used. Is required.

That is, FIG. 6 is a diagram showing a state of overlap addition and gain control on the decoder side. In FIG. 6, w (n), n = 0 to N-1 are analysis / synthesis windows. And g (n) is a time axis gain control parameter, that is, g (n) = qrms _j (j
Satisfies jM ≦ n ≦ (j + 1) M), g ₁ (n) is g (n) of the current frame FR ₁ , and g ₀ (n) is g (n) of the previous frame (previous frame FR ₀ ). g (n). FIG.
In this example, one frame is divided into eight sub blocks (sub frames) SB (M = 8).

For the latter half of the previous frame FR ₀ ,
On the encoder side, g ₀ (n + (N
/ 2)), the analysis window w ((N / 2) -1 for MDCT
-n), so after the inverse MDCT on the decoder side,
The signal obtained by applying the analysis window w ((N / 2) -1-n) again, that is, the sum P of the principal component and the aliasing component
(n) is expressed by the following equation (2).

[0048]

(Equation 2)

On the encoder side, the data in the first half of the current frame FR ₁ is divided by g ₀ (n) for gain control, and then the analysis window w (n) for MDCT.
Therefore, a signal obtained by applying the analysis window w (n) again after the inverse MDCT on the decoder side, that is, the sum Q (n) of the main component and the aliasing component is given by the following equation (3). become that way.

[0050]

(Equation 3)

Therefore, x (n) to be reproduced is obtained by the following equation (4).

[0052]

(Equation 4)

By performing such windowing and performing gain control using the rms of each sub-block (sub-frame) as an envelope, a sound that changes drastically over time, for example, a musical sound having a sharp attack or a pitch between pitch peaks can be obtained. For speech that attenuates relatively quickly, quantization noise such as pre-echo, which is easily heard, can be reduced.

Next, the MDCT coefficient data obtained by the MDCT processing in the MDCT circuit 27 of the orthogonal transformation circuit unit 25 is converted into the frame gain normalization circuit 43 of the coefficient quantization unit 40.
And the frame gain calculation / quantization circuit 47.
In the coefficient quantization unit 40 of the present embodiment, first, the MDC
After calculating the frame gain (block gain) of the entire coefficient of one frame, which is a T-transform block, and performing gain normalization, the critical band (critical band), which is a sub-band whose bandwidth is widened in a higher frequency range in accordance with hearing, is further increased. band)
Divided into scale factors for each band,
A so-called bark scale factor is calculated, and normalization is performed again by this. As the bark scale factor, for each band, a peak value of a coefficient in the band, a root mean square (rms), or the like can be used, and the bark scale factor of each band is collectively vector-quantized. .

That is, the frame gain calculation / quantization circuit 47 of the coefficient quantization unit 40 calculates and quantizes the gain for each frame, which is the above-mentioned MDCT transform block, and uses the codebook index (frame gain index) as a terminal. The frame gain of the quantized value is sent to the frame gain normalization circuit 43, and is normalized by dividing the input by the frame gain. The output normalized by the frame gain is sent to a bark scale factor calculation / quantization circuit 42 and a bark scale factor normalization circuit 44.

The bark scale factor calculation / quantization circuit 42 calculates and quantizes the bark scale factor for each critical band, and fetches a codebook index (bark scale factor index) via a terminal 54 and decodes the codebook index. Side, and the Bark scale factor of the quantized value is calculated by the bit allocation calculating circuit 4.
1 and a bark scale factor normalization circuit 44. The bark scale factor normalization circuit 44 normalizes the coefficient within the band for each of the above critical bands, and the coefficient normalized by the bark scale factor is used as the coefficient quantization circuit 4.
Sent to 5.

In the coefficient quantization circuit 45, quantization is performed by allocating the number of quantization bits to each coefficient according to the bit allocation information from the bit allocation calculation circuit 41.
The total number of allocated bits is switched according to the gain control SW information from the gain control on / off determination circuit 18 described above. For example, when performing vector quantization, two sets of codebooks, one for the above-described gain control and one for the off-state, are prepared.
What is necessary is just to switch these codebooks according to the gain control SW information.

Here, the bit allocation (bit allocation) in the bit allocation calculation circuit 41 will be described. The LPC coefficient, the pitch parameter, the bark scale factor, etc.
The quantization weight for the T coefficient is calculated, and thereby the bit allocation of the MDCT coefficient for the entire band is determined and quantization is performed. This weight can be considered as a noise shaping factor, and a desired noise shaping characteristic can be provided by changing each parameter. As an example, in the present embodiment, the weight W (ω) is calculated using only the LPC coefficient, the pitch parameter, and the bark scale factor as shown in the following equation.

[0059]

(Equation 5)

As described above, the weight determination at the time of quantization is performed by LP
Since only the C, pitch and bark scale factors are used, if only these three types of parameters are transmitted to the decoder, exactly the same bit allocation as that of the encoder will be reproduced. The rate of (auxiliary information) can be reduced.

Next, a specific example of quantization in the coefficient quantization circuit 45 will be described with reference to FIGS. 7, 8 and FIG.

FIG. 1 described above shows the coefficient quantization circuit 4 of FIG.
5 and the input terminal 1 in FIG.
The normalized coefficient data (for example, MDCT coefficients) y from the Bark scale factor normalizing circuit 44 is supplied. The weight calculation circuit 2 is a bit allocation calculation circuit 41 shown in FIG.
, But only the part for calculating the weight of each coefficient for allocating quantization bits is extracted. The weight calculation circuit 2 calculates the weight w based on parameters such as the above-described LPC coefficient, pitch parameter, and bark scale factor. Here, the coefficient for one frame is represented by a vector y and the weight for one frame is represented by a vector w.

FIG. 7 shows the state of this sort. FIG. 7A shows the weight vector w _k of the k-th band.
FIG. 7B shows a coefficient vector y _k of the k-th band. In the example of FIG. 7, the number of elements in the k-th band is, for example, eight, and the eight weights that are the respective elements of the weight vector w _k are w ₁ , w ₂ ,.
₈ , the eight coefficients that are each element of the coefficient vector y _k are _represented by y
₁ , y ₂ ,..., Y ₈ . (A) in FIG. 7, in the example of (B) is the largest weight w ₃ corresponding to the coefficient y _3, in the following order the _{weights, w 2, w 6, ...} , w
_{It is 4} . (C) in FIG. 7, the coefficient y _1, y ₂ in the order of the weights, ..., and sort the y ₈ (sorting), turn y _3,
The coefficient vector y ′ _k is _defined as y ₂ , y ₆ ,..., y ₄ .

Next, as described above, the coefficient vectors y ′ ₀ ,
_{y '1, ..., y'} L-1 , respectively vector quantizer 5 _0, 5 _1, ..., send the 5 _L-1, perform the respective vector quantization. Here, it is preferable that the number of allocated bits for each band is fixed in advance to prevent the allocation of the number of quantized bits to each band from fluctuating even if the energy for each band changes.

For the vector quantization for each band,
If the number of elements in one band is large, the band may be divided into several subvectors and vector quantization may be performed for each subvector. That is, the sorted coefficient vector y ′ _k of an arbitrary k-th band is divided into several sub-vectors according to a predetermined number of elements as shown in FIG.
For example, three subvectors y ′ _k1 , y ′ _k2 , y ′ _k3 may be used, and these may be respectively vector-quantized to obtain codebook indexes c _k1 , c _k2 , c _k3 .
The indexes _ck1 , _ck2 , _ck3 of the k-th band are collectively referred to as a coefficient index vector _ck . Here, in the quantization of the sub-vector, quantization is performed according to the weight by assigning a larger number of quantization bits to the head vector. For example, in FIG. 8, the vector y ′ _k1 is assigned 8 bits, the vector y ′ _k2 is assigned 6 bits, and the vector y ′ _k3 is assigned 8 bits, so that the number of assigned bits per coefficient is large. , So that bit allocation according to the weight can be realized.

Next, each vector quantizer 5 ₀ ,
The vectors c ₀ , c ₁ ,..., C _L-1 of the coefficient indexes for each band from 5 ₁ ,..., 5 _L−1 are combined into a vector c of the coefficient indexes of all the bands, and the terminal It is taken out from 6. This terminal 6 corresponds to the terminal 51 in FIG.

In the specific examples shown in FIGS. 1, 7 and 8, the orthogonally transformed coefficients on the frequency axis (for example, MDCT
The coefficient) itself is sorted according to the above-mentioned weights, and the number of allocated bits is reduced from a larger number to a smaller number according to the order of the sorted coefficients (more bits are allocated to higher-ranked coefficients in the sorted order). Only the index (index) indicating the position or order on the frequency axis of each coefficient obtained by the orthogonal transformation is sorted in the order of the weight,
The quantization precision (number of allocated bits and the like) of each coefficient may be determined in accordance with the sorted index (index). Further, in the above specific example, the vector quantization is used as the coefficient quantization, but the present invention may be applied to scalar quantization or quantization in which scalar quantization and vector quantization are used in combination. Easy.

Next, an example of the configuration of an audio signal decoding device (decoder side) corresponding to the audio signal encoding device (encoder side) as shown in FIG.
This will be described with reference to FIG.

In FIG. 9, each of the input terminals 60 to 67
2 is supplied with data from each output terminal of FIG. 2, and the input terminal 60 of FIG. 9 is connected to the output terminal 51 of FIG.
Are supplied with an index of orthogonal transform coefficients (for example, MDCT coefficients). The input terminal 61 is supplied with the LSP index from the output terminal 31 of FIG.
2 to 65 are supplied with data from the output terminals 52 to 55 in FIG. 2, that is, pitch lag index, pitch gain index, bark scale factor index, and frame gain index, respectively.
The envelope indexes and the gain control SW from the output terminals 21 and 22 in FIG.

The coefficient index from the input terminal 60 is inversely quantized by a coefficient inverse quantization circuit 71 and sent to an inverse orthogonal transform circuit 74 such as an IMDCT (inverse MDCT) via a multiplier 73.

The LSP index from the input terminal 61 is
The data is sent to the inverse quantizer 81 of the LPC parameter reproducing unit 80 and inversely quantized into LSP data.
2 and the LSP interpolation circuit 83. The α parameter (LPC coefficient) from the LSP → α conversion circuit 82 is sent to the bit allocation circuit 72. L from the LSP interpolation circuit 83
The SP data is converted into an α parameter (LPC coefficient) by an LSP → α conversion circuit 84, and the LPC synthesis circuit 7 described later.
7

The bit allocation circuit 72 has a pitch lag from the input terminal 62 and a pitch gain obtained from the input terminal 63 via the inverse quantizer 91 in addition to the LPC coefficient from the LSP → α conversion circuit 82. And the bark scale factor obtained via the inverse quantizer 92 from the input terminal 64, and the same bit allocation as that on the encoder side can be reproduced based only on these parameters. The bit allocation information from the bit allocation circuit 72 is
The coefficient is transmitted to the coefficient inverse quantizer 71 and is used to determine a quantization assignment bit of each coefficient.

The frame gain index from the input terminal 65 is sent to the frame gain inverse quantizer 86 and inversely quantized, and the obtained frame gain is sent to the multiplier 73.

The envelope index from the input terminal 66 is supplied via a switch 87 to the envelope inverse quantizer 8.
8, and the resulting envelope data is sent to an overlap adding circuit 75. The gain control SW information from the input terminal 67 is sent to the coefficient inverse quantizer 71 and the overlap addition circuit 75 and used as a control signal for the switch 87. The coefficient inverse quantizer 71 switches the total number of allocated bits in accordance with ON / OFF of the gain control as described above. In the case of inverse vector quantization, the codebook at the time of the gain control ON is used. And the off-state codebook may be switched.

The overlap addition circuit 75 has an IMDCT
The signal returned to the time axis for each frame from the inverse orthogonal transform circuit 74 is added while overlapping by 1/2 frame. When the gain control is on, the envelope inverse quantizer 88 is added. The overlap addition is performed while performing the gain control (gain expansion or gain restoration described above) using the envelope data from.

The time base signal from the overlap adding circuit 75 is sent to the pitch synthesizing circuit 76 to restore the pitch component. This corresponds to the inverse processing of the processing by the pitch inverse filter 13 in FIG. 2, and uses the pitch lag from the terminal 62 and the pitch gain from the inverse quantizer 91.

The output from the pitch synthesizing circuit 76 is LPC
The LPC inverse filter 12 shown in FIG.
An LPC synthesis process corresponding to the reverse process of the above process is performed and extracted from the output terminal 78.

Here, when the coefficient quantizing circuit 45 of the coefficient quantizing section 40 on the encoder side uses a vector quantizing coefficient sorted for each band according to the weight as shown in FIG. As the coefficient inverse quantization circuit 71, a configuration as shown in FIG. 10 can be used.

In FIG. 10, the input terminal 60 corresponds to the input terminal 60 in FIG. 9, and the coefficient index (M
A codebook index obtained by quantizing orthogonal transform coefficients such as DCT coefficients is supplied to the weight calculation circuit 79. The α parameter (LPC coefficient) from the LSP → α conversion circuit 82 in FIG. Pitch lag from input terminal 62, pitch gain from inverse quantizer 91, inverse quantizer 9
2, a Bark scale factor and the like are supplied.
The weight calculation circuit 79 is provided in the bit allocation circuit 72 in FIG.
The components up to calculating the weight of each coefficient obtained during the calculation of the quantization bit allocation are extracted. As described above, the weight calculation circuit 79 calculates the weight W (ω) using only the LPC coefficient, the pitch parameter (pitch lag and pitch gain), and the bark scale factor by the calculation of the above equation (5). are doing.
The input terminal 93 has an index (index) indicating the position or order of the coefficient on the frequency axis, that is, if there are N coefficient data in the entire band, the numerical value of 0 to N-1 (this is expressed as a vector I and To be supplied). Note that N weights for each of the N coefficients from the weight calculation circuit 79 are represented by a vector w.

The weight w from the weight calculation circuit 79 and the index I from the input terminal 93 are sent to the band division circuit 94, and are divided into L bands as in the encoder side. If the encoder side is divided into, for example, three bands (L = 3) of a low band, a middle band, and a high band, the decoder side is also divided into three bands. The indices and the weights for each of these band-divided bands are respectively stored in the sort circuits 95 ₀ , 95
₁ , ..., 95 Sent to _L-1 . For example, the index I _k and the weight w _k in the k-th band are sent to the k-th sorting circuit 95 _k . In the sorting circuit 95 _k , the indexes I _k in the k-th band are rearranged (sorted) according to the order of the weights w _k of the coefficients, and the sorted indexes I ′ _k are output. Each sort circuit 95 ₀ , 95 ₁ , ...,
The indexes I ₀ , I ₁ ,..., I _L−1 sorted for each band from 95 _L−1 are sent to the coefficient reconstructing circuit 97.

When the index of the orthogonal transform coefficient from the input terminal 60 in FIG. 9 is quantized on the encoder side, the band is divided into L bands as described with reference to FIGS. The coefficients sorted in the order of weight for each band are obtained by performing vector quantization for each sub-vector divided by the number based on a predetermined rule in one band. Specifically, for the L bands, a set of coefficient indexes for each band is a vector c ₀ , c ₁ ,..., C _L−1, and the coefficient index for each of these bands vector c _0, c ₁ of, ..., c _L-1, respectively inverse quantizer 96 _0, 95 _1, ..., it is transmitted to the 95 _L-1. The coefficient data obtained by the inverse quantization by these inverse quantizers 96 ₀ , 95 ₁ ,..., 95 _L−1 are sorted in the order of the weights in each band, 7 sort circuits 4
_0, 4 _1, ..., coefficient vector y _'0 from 4 _L-1,
y ′ ₁ ,..., y ′ _L−1 , and the arrangement order is different from the position on the frequency axis. Therefore, the index I representing the position of the coefficient on the time axis is sorted in advance according to the weight, and the sorted index is associated with the coefficient data obtained by the inverse quantization to obtain The function of the coefficient reconstruction circuit 97 is to return to the order on the time axis. That is, the coefficient reconstruction circuit 97, the inverse quantizer 96 _0, 95 _1, ..., from 95 _L-1, the coefficient data sorted in the weight order in each band, the sort circuit 9
_{5 0, 95 1, ...,} 95 L-1 are associated a sorted index for each band from, (reverse sort) Sort the inverse quantized coefficient data according to the sorted index As a result, the coefficient data y arranged in the original order on the time axis is obtained and extracted from the output terminal 98. The coefficient data from the output terminal 98 is output to the multiplier 7 in FIG.
Sent to 3.

The present invention is not limited to the above-described embodiment. For example, the input time axis signal may be an audio signal in a telephone band, a video signal, or the like, in addition to an audio signal including voice and music. . Further, the configuration of the normalization circuit unit 11 and the LPC analysis and the pitch analysis are not limited to these, and various configurations for extracting and removing the characteristics or correlation of the time-axis input waveform by linear prediction or non-linear prediction are available. Possible. In addition, in addition to vector quantization, each quantizer may use scalar quantization or a combination of scalar quantization and vector quantization.

[0083]

As is clear from the above description, according to the present invention, when encoding an input signal on the time axis using orthogonal transform, the weight is calculated according to the input signal. According to the order of the weights, the coefficient data obtained by the orthogonal transformation is ranked, and the quantization is performed with high precision according to the rank. Therefore, even if the bits are dynamically allocated according to the input signal, Calculation of the number of bits allocated to each coefficient is simple.

In particular, when the coefficient data is divided into several parts and vector-quantized as sub-vectors, even if the weight of each coefficient changes dynamically and the number of bits assigned to each coefficient changes, the order of the weights changes. By sub-vectoring after sorting, the number of bits allocated to each sub-vector can be determined only by weight calculation, and the amount of calculation can be reduced.

Further, the coefficient data of the entire band on the frequency axis is divided into several bands, and the number of bits allocated to each band is defined in advance, so that the weight of each coefficient changes extremely between frames. However, since the number of allocated bits for each band is stable, a rapid change in quantization distortion and the like can be prevented, and a stable reproduced sound can be obtained.

Further, by preliminarily defining parameters for calculating bit allocation and transmitting these parameters to the decoder side, it is not necessary to transmit the bit allocation information itself to the decoder side, and additional information (side information) is provided. And the bit rate can be reduced. In addition, by using an improved discrete cosine transform (MDCT) as the orthogonal transform, highly efficient encoding of an audio signal with good sound quality can be performed.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a schematic configuration of a quantization circuit used in an embodiment of the present invention.

FIG. 2 is a block diagram illustrating an audio signal encoding device that is a more specific configuration example of an embodiment of the present invention.

FIG. 3 is a diagram illustrating a relationship between an LPC analysis process and a pitch analysis process for an input signal.

FIG. 4 is a time-axis signal waveform diagram for explaining removal of correlation by a LPC analysis and a pitch analysis of a time-axis input signal.

FIG. 5 is a diagram showing frequency characteristics for explaining the removal of correlation by LPC analysis and pitch analysis of a time axis input signal.

FIG. 6 is a time axis signal waveform diagram for explaining overlap addition on the decoder side.

FIG. 7 is a diagram for explaining sorting according to the weight of a coefficient in one band divided into bands.

FIG. 8 is a diagram for explaining a process in which coefficients sorted according to weights in one band-divided band are divided into sub-vectors and vector-quantized.

9 is a block diagram illustrating an example of an audio signal decoding device as a decoding side configuration corresponding to the audio signal encoding device of FIG. 2;

FIG. 10 is a block diagram showing a specific example of an inverse quantization circuit of the audio signal decoding device in FIG. 9;

[Explanation of symbols]

11 normalization circuit section, 12 LPC inverse filter, 1
3 pitch inverse filter, 15 pitch analysis circuit, 1
6 pitch gain quantization circuit, 17 envelope extraction circuit, 18 gain control on / off determination circuit, 20 envelope quantization circuit, 25 orthogonal transformation circuit section, 26 windowing circuit, 27 MDCT circuit,
30 LPC analysis / quantization unit, 32 LPC analysis circuit, 33 α → LSP conversion circuit, 34 LSP quantization circuit, 36 LSP interpolation circuit, 37, 38 LS
P → α conversion circuit, 40 coefficient quantization circuit section, 41
Bit allocation calculation circuit, 42 bark scale factor calculation / quantization circuit, 43 frame gain normalization circuit,
44 bark scale factor normalization circuit, 45 coefficient quantization circuit, 47 frame gain calculation / quantization circuit

Claims (14)

[Claims] 1. A signal encoding apparatus for performing orthogonal transform on an input signal on a time axis by orthogonal transform means to perform encoding, wherein: a weight calculating means for calculating a weight according to the input signal; A signal encoding device, comprising: a coefficient ranking unit that ranks coefficient data from a conversion unit in accordance with the order of weights from the weight calculation unit, and performs quantization with higher precision in a higher rank. 2. The signal encoding apparatus according to claim 1, wherein in the quantization, the higher the coefficient data in the order, the larger the number of bits allocated for quantization. 3. The coefficient data from the orthogonal transform means is
2. The signal encoding method according to claim 1, wherein the signal is divided into a plurality of bands on a frequency axis, and coefficient data in each band is quantized independently in each band in the order of the weight. apparatus. 4. The method according to claim 1, wherein the coefficient data is divided into a plurality of coefficient data from the coefficient data on the upper side of the ranked order to obtain respective coefficient vectors, and these coefficient vectors are respectively vector-quantized. Signal encoding device. 5. The signal code according to claim 1, wherein the weight calculating means calculates the weight based on a parameter representing a statistical property of the input signal including a linear or non-linear prediction of the input signal. Device. 6. The signal encoding apparatus according to claim 1, wherein said orthogonal transform means transforms a time-axis signal input by an improved discrete cosine transform (MDCT) into coefficient data on a frequency axis. 7. An input side of said orthogonal transformation means, provided with a normalization means for removing a correlation of a signal waveform of said input signal and extracting a residual, said normalization means performing linear predictive coding on said input signal. (L
PC) an LPC inverse filter that outputs an LPC prediction residual of the input signal based on the LPC coefficient obtained by the analysis, and the LPC prediction residual based on a pitch parameter obtained by performing a pitch analysis on the LPC prediction residual 2. The signal code according to claim 1, further comprising: a pitch inverse filter that removes a correlation between the pitches, and wherein the weight calculation means calculates a weight based on the LPC coefficient and the pitch parameter. Device. 8. A signal encoding method for encoding an input signal on a time axis using orthogonal transform, comprising: a weight calculating step of calculating a weight according to the input signal; And a quantizing step of ranking the obtained coefficient data in accordance with the order of the calculated weights and performing highly accurate quantization in accordance with the rank. 9. The signal encoding method according to claim 8, wherein, in the quantization, the higher the coefficient data in the order, the larger the number of quantization allocation bits. 10. The coefficient data obtained by performing the orthogonal transformation is divided into a plurality of bands on a frequency axis, and the coefficient data in each band is independently ranked for each band according to the order of the weights. 9. The signal encoding method according to claim 8, wherein the signal is quantized. 11. The method according to claim 8, wherein the coefficient data is divided into a plurality of coefficient data from the coefficient data on the upper side of the ranked order to obtain respective coefficient vectors, and these coefficient vectors are respectively vector-quantized. Signal encoding method. 12. An orthogonal transform is applied to an input signal on a time axis, and coefficient data obtained by the orthogonal transform is subjected to quantization with a quantization accuracy according to a weight based on the input signal. A signal decoding device for decoding the encoded data, wherein a parameter for calculating the weight is input, and weight calculating means for calculating a weight at the time of the encoding based on the parameter; and the coefficient data is quantized. A signal decoding device, comprising: inverse quantization means for performing inverse quantization with quantization accuracy on the data obtained as a result of the calculation by the weight calculation means. 13. A weight calculated based on a parameter representing a statistical property of the input signal including a linear or non-linear prediction of the input signal is used as the weight, and the parameter is transmitted and supplied. 13. The signal decoding device according to claim 12, wherein the weight calculating means calculates the weight based on the supplied parameters. 14. An orthogonal transform is performed on an input signal on a time axis, and coefficient data obtained by performing the orthogonal transform is quantized by a quantization precision according to a weight based on the input signal. A signal decoding method for decoding the encoded data, wherein a parameter for calculating the weight is input, and a weight calculating step of calculating a weight at the time of the encoding based on the parameter; and wherein the coefficient data is quantized. A dequantization step of performing dequantization on the data obtained as a result with quantization precision in accordance with the weight calculated by the weight calculation means.