JPH01147932A - Linear predictive coding method - Google Patents

Abstract

PURPOSE:To improve the accuracy of prediction coefficient by applying blocking processing to a sample and multiplying a prescribed value with the sample for each block. CONSTITUTION:The multiplication of weighting by a time window function W(n) is applied to an input data Xt at a terminal 11 by a time window 21, the obtained data is compared with a maximum value of a preceding data to extract a maximum value of data Xt.W(n). Similar processing is applied to the data Xt of one block succeedingly. Then a maximum value Tmax after weighting in one block is extracted. Then the value Tmax is shifted left one by one bit sequentially, doubled, compared with a value TREF being a comparison reference and the processing is proceeded to obtain the result of Tmin>=TREF. Thus, the data Xt.W(n) after weighting is multiplied in the unit of blocks in the range that the maximum value Tmax is not subject to overflow. Even with a limit of word length for succeeding processing, the accuracy of prediction coefficient alphai is improved.

Description

[Detailed description of the invention] The explanation will be given in the following order.

A. Industrial application field B. Summary of the invention C Conventional technology D. Problem that the invention aims to solve E. Means to solve the problem (Figure 1) F. Effect G Example G1 first embodiment (Figure 1) G2 other examples Effect of H invention A. Industrial application field The present invention relates to a linear predictive coding method.

B. Summary of the Invention The present invention is a linear predictive coding method in which samples are divided into blocks and a predetermined multiplication is performed on the samples for each block, thereby increasing the precision of prediction coefficients.

C. Conventional technology For example, in 8 mm video, an optional function is to digitize the audio signal into a PCM signal during recording, record this PCM signal in the overscan section of the tape, and reverse the process during playback. It is accepted that the original audio signal can be obtained by performing

In this case, if the sampling frequency and number of quantization bits of the PCM signal are increased, it is possible to record and play back the audio signal with better characteristics. It ends up.

Therefore, when recording, the number of bits of the PCM signal is compressed.
It has been proposed to expand the number of bits during playback so that excellent recording and playback characteristics can be obtained even if the number of bits on the tape is small.

AD is a method of bit compression/expansion.
There is a method called PCM.

Figure 2 shows an example of a transmission system using this ADPCH. In this example, each 64 consecutive samples of input data is set as one block, and the prediction coefficients of the prediction filter are optimized for each block. This is the case when controlling the value. At this time, the main data bit-compressed for each sample of the input data is output, and auxiliary data related to the bit compression is output for each block.

That is, in FIG. 2, (10) is the encoder, (
30) is a signal transmission system, and (40) is a decoder. For example, if it is used for a PCM audio system in 8 mm video, an encoder (10) is provided in a recording system,
A decoder (40) is provided on the recycled yarn, and a transmission system (30) includes an error correction processing circuit, a nine-rotation magnetic head, and the like.

Then, in the encoder (lO), the digital data Xt is transmitted in parallel for each sample from the input terminal (11) through the delay circuits (12) and (13) to the subtraction circuit (
14). In this case, the input data Xt is a PCM signal that is linearly A/D converted from an analog audio signal, and for example, the sampling frequency is 48
kHz, and the number of quantization bits is 16 bits. Also,
As shown in FIG. 3, it is assumed that the data Xt is expressed as a fixed decimal point with −1≦Xt<1 and as a two's complement (the same applies to other values).

Furthermore, the delay circuits (12) and (13) are for synchronizing the timing of the main data and the auxiliary data, and each has a delay time of one block period (
Therefore, strictly speaking, if the input value of the terminal (11) is Xt, the output of the delay circuit (13) will be Xt-12@, but for the sake of complexity, we will write Xt on the car).

Further, a predicted value ×[ for data Xt is extracted from the prediction filter (19), and this value ×t is sent to the subtraction circuit (14).
The subtraction circuit (14) extracts the difference DtD1xt-xt between the value Xt and the value t. This value Di is the predicted value 5i! for the input value Xt! is the error (prediction residual) of t, therefore,
Ideally, the value D1 is Di-0, which is a small value even for boat fishing, so even if the word length of the value Dt is, for example, 16 bits (because it is expressed as a fixed decimal point), for example, Diko"o, ooo...011011',
All the significant bits on the MSH side become "O" (except for the sign bit), and the remaining few bits on the LSB side become 0" or "1" based on the difference between the values Xt and xt. Furthermore, when the value Dt becomes a large value, the lower bits can be ignored.

Therefore, this value DL is supplied to the gain control circuit (15) and multiplied by 6 (G≧1), resulting in a normalized value D
t-G, and this value G-1)t is the requantization circuit (16
) and requantized, for example, into a 4-bit value 15t-c.

Furthermore, this value 6t-c is supplied to the gain control circuit (17) and multiplied by 1/G, thus giving a non-normalized value tSt of the same order as the value Dt, and this value 15
t is supplied to the adder circuit (18), and the predicted value N from the filter (19) is supplied to the adder circuit (18). t father t=nu t+Qt is taken out and this value 51i:t is fed to the filter (19).

In this case, the value t is the predicted value for the value Xt, and the value t5t is the value obtained by truncating or rounding the lower bits of the error Dt at the time of prediction, so the sum of these values 5zt and 6t is A certain value t is approximately equal to the input value Xt, and since this value t is supplied to the filter (19), the value 5 < which is the filter output.
t can be a value that predicts the input value Xt◆1 at the next sample time.

Then, the value 15t-c from the requantization circuit (16) is
The signal is supplied to a decoder (40) through a transmission system (30).

In this decoder (40), the value 5t-c is multiplied by 1/G by the gain control circuit (41) to obtain the value I5t, and this value tbt is supplied to the adder circuit (42), and the added output is output from the output terminal. (44), and the prediction filter (43) configured similarly to the filter (19).
), and its filter output is supplied to an adder circuit (42).

Therefore, the output of the filter (43) becomes the value Xt, and the terminal (44) receives the data father t, in which the lower pits of the input data Xt are rounded, that is, the input data
Digital data father t approximately equal to t is retrieved.

Further, in order to set the prediction coefficients in the filters (19) and (43) to optimal values for each block, the following circuit is provided.

That is, the prediction filters (19) and (43) are
For example, it is considered to be a 4th order filter, and its 1st order ~
The fourth-order coefficients a1 to a4 can be changed to arbitrary values.

Input data Xt from the terminal (11) is supplied to a time window circuit (21), subjected to predetermined weighting, and then supplied to an autocorrelation circuit (22) to calculate a correlation coefficient. The coefficients are supplied to the prediction coefficient circuit (23), and the prediction coefficients α1 to α of the fourth order are calculated for each block of data Xt.
4 is calculated.

Further, data Xt from the delay circuit (12) is supplied to a prediction error filter (24), and the filter output is supplied to an intra-block maximum value detection circuit (25).

In this case, the filter (24) is the prediction filter (19)
a fourth-order prediction filter (241) configured in the same manner as
a subtraction circuit (242), and a coefficient circuit (242).
The prediction coefficients α1 to α1 from 3) are supplied to the filter (241), and the predicted value (
The prediction error) t5t is generated for each sample. Further, the detection circuit (25) detects 1 of the human power data Xt.
For each block, the predetermined error i within that block is
5t (there are 64 of them), the absolute value i5taax of the prediction error with the largest absolute value is detected.

Then, this maximum value 15Illaχ is supplied to the normalization gain calculation circuit (28), and the data of the gain G at the time of normalization, G
"b/f5 rtrax b is a safety factor of Q<b<1, and is converted to, for example, b-0,9, and this data G is used in the gain control circuit (15).

(17) and is also supplied to the gain control circuit @(41) through the latch (52). In this case, the value 5
Since i l1ax is the maximum value of the 64 values f5t, the value Dt-G is normalized to -1≦Dt-G<1.

Furthermore, considering the amount of data transmitted from the encoder (10) to the decoder (40) through the transmission system (30), the main data t5t-c is, for example, 4 bits and 1
The prediction coefficient α is transmitted for each sample and is auxiliary data.
1 to α and data G are transmitted for each block in 8 bits and 16 bits, respectively, so the amount of data in one block period is 4 bits x 64 samples + 8 bits XAffi + 16
It becomes 304 bits. 1 in the case of no data compression
The amount of data in a block period is 16 bits x 64 samples - 1024 bits. Therefore, the amount of data was compressed and transmitted to 304 bits/1024 bits ζ29.7%.

In this way, according to this system, data compression of digital audio data can be performed, but in this case,
In particular, according to this system, even if there is a limit to the word length of the coefficients and calculations, the prediction coefficients of the prediction filters (19) and (43) are controlled to optimal values according to the input data Xt, so that the decoded Errors caused by compression of data 5t can be minimized.

In addition, when transmitting the prediction residual Di, this residual Di is requantized to reduce the number of bits and normalized before the requantization, so the transmitted data 5t-G is , resulting in data with fewer bits and fewer errors.

Literature: "Fundamentals of Speech Information Processing", patent application published by Ohmsha, 1986
-299285 Specification and Drawing D Problems to be Solved by the Invention However, in the above system, the prediction coefficients α1 to
The accuracy of α2 cannot be made very high.

That is, the time window function (weighting function) W (n) in the time window circuit (21) is generally a Hamming window W (n) =0.54-0.46cos (2rc
n/ (N −1) ) N: Number of data in l block (
N = 64) n: Data number 0≦n S N -1 in l block is used.

Then, the data Xt fixed in this way is supplied to the autocorrelation circuit (22), and the autocorrelation coefficient vj is as follows: xn: Data Xt at t=n after weighting j=0. 1. 2. ..., p P: Calculated based on the order of the prediction filter (p = 4).

Furthermore, the processing of the prediction coefficient circuit (23) is performed using an algorithm as shown in FIG. 4, for example, to obtain prediction coefficients. However, in the same figure, v, ): Q-order autocorrelation coefficient vl: 1st-order I αi: forward linear prediction coefficient, αi = αiWn: cross-correlation between forward prediction residual and backward prediction residual un: forward prediction Root mean square of residuals kn◆i: Partial autocorrelation coefficient.

Therefore, the accuracy of the prediction coefficient αi is calculated in the following order: accuracy of weighted data XL (−Xn) → accuracy of the autocorrelation coefficient vj and cross-correlation wn → accuracy of the partial autocorrelation coefficient kl order 1 − accuracy of the prediction coefficient αi. is decided.

On the other hand, if the time window function of the time window circuit (21) is a Hamming window as described above, the weight W ((N-1)/2) for the data Xt at the center of the block is W ((N -1)/2)-1, but the weights W (0) and W (N-1) for the data Xt at the beginning and end of the block are W (0)
−0,08 w (N −1) −0,08, and the data Xt near t=o and t=N−i after direct alignment becomes a small value. That is, if expressed as a binary value, the focus will be shifted to the right by 3 to 4 points.

At this time, since the word length of the data Xt is limited, the lower bits shifted to the right are discarded, and as a result, the data Xt after stopping becomes less accurate.

As mentioned above, the accuracy of the weighted data Xt directly affects the accuracy of the prediction coefficient αi, so
As the accuracy of the data Xt after stopping decreases,
The accuracy of the prediction coefficient αl also decreases.

This invention attempts to solve these problems.

E Means for solving the problem As mentioned above and as shown in Figure 4, the prediction coefficient α
The precision of i is determined by the precision of the coefficient kn◆1, but this coefficient ■1 is calculated by division with the value wn(!:un. Then, at the time of this division, the dividend and divisor (numerator and denominator) ) are multiplied in advance by constants that are equal to each other, there will be no problem with the coefficient kn, which is the quotient of the division.

This invention focuses on these points, and the weighted data xt has a long word length corresponding to its weight W(n), and the weighted data Xt is divided into blocks for each block. Multiply by a coefficient of a predetermined size (≧1),
The prediction coefficient αi is determined using the multiplication result. In addition,
The coefficient to be multiplied is such that the multiplication result is, for example, data Xt
The size may be within the range that can be expressed in the data format.

F Prediction coefficients are calculated with high accuracy even when weighting is performed using a time window function on the human force data.

G Embodiment G1 First Embodiment In the time window circuit (21), the process shown in the flowchart shown in FIG. 1 is performed.

That is, in step (61), the input data Xi (for one sample) of the terminal (11) is weighted and multiplied by the time window function W(n), and the data Xt
-W(n), and then in step (62), the data xt-W(n) and the previous data Xt-W(n)
The maximum value (absolute value) of data Xt-W(n) is extracted by comparing the magnitude with the maximum value (absolute value) of step (61), (6
It is checked whether or not the process 2) has been performed. If it has not been performed, the process returns to step (61), and if it has been performed, the process advances to step (71).

Therefore, by steps (61) to (63), the time window function W(n) is calculated for all input data Xt of one block.
Weighting is performed, and the maximum value (absolute value) Tmax after weighting within that one block is extracted. It should be noted that the above processing is performed, for example, on data Xt
is 16 bits, function W(n) is 8 bits, data X t
-W(n), T+sax is performed using 24 bits, and data Xt-W(n).

The accuracy required for Tmax is ensured.

Then, in step (71), the bit shift counter BSFT is reset to rOJ, and then in step (72), the maximum value Tmax is compared in magnitude with a predetermined constant value TREF. In this case, the value TRIP
is a value that is several bits smaller than the maximum value that can be expressed with 24 bits, which is the word length of data Ru.

As a result of the size comparison, '1' wax < TRt
! l', the process proceeds to step (73), and in this step (73), the counter BSFT is rl.
The counter HSFT is incremented by J, and then in step (74) the counter HSFT is compared in magnitude with a constant predetermined value BLIM. This size comparison is to prevent the process from entering an infinite loop, as will be described later.
(2) Corresponding to the word length of ax, for example, BLIM=24.

Then, as a result of this size comparison, when BSFT<BLIF, the process proceeds to step (75), and in this step (75), the data Tl1aχ is doubled by, for example, being shifted to the left by 1 focus, and then the process is Return to step (72).

Therefore, by repeating steps (72) to (75), the value Tmax is sequentially shifted to the left by 1 pint and multiplied by 2, and the number of times it is shifted to the left is counted by the counter BSFT. In addition,
When Tmax-0, no matter how many times you shift left, Tn+a
Since it becomes an infinite loop with x=0, step (74
), the counter BSFT is checked and BSFT
When ≧BLIM, the process proceeds to step (74)
The routine then exits to step (76) and ends.

Then, by repeating steps (72) to (75), the value 'rtsax is sequentially shifted to the left by 1 bit and multiplied by 2. When Tslax≧TRIP, the process proceeds from step (72) to step (81). ), in this case the value ``rtaax'' is ``B'' whose original value is ``2''.
This means that the SFT has been multiplied, but step (72)
The value TREI', which is a comparison standard, is the word length (
For example, the maximum value that can be expressed with 24 bits (in this example, 1)
/22 times the value, the value 'raIax will not overflow.

Then, in step (81), the counter BSFT
is checked, and when BSFT>O, the process proceeds to step (82), and in this step (82), all data Xt - W (n) of one block are each shifted to the left by 1 bit and multiplied by 2, Then step (
At step 83), the counter BSFT is decremented by rlJ, and then the process returns to step (81).

Therefore, by steps (81) to (83), the total data Xt −W (n) of 64 samples in one block is
Each bit is shifted to the left by 1 bit and multiplied by 2.
When each data Xt -W (n) is all shifted to the left by BSl'T bits, it becomes BSFT-0.

Then, when BSFT-0 is reached, the process proceeds to step (81
), the routine proceeds to step (84) and ends this routine.

Therefore, the weighted data Xt -W(n) is shifted to the left by BSFT bits and supplied to the next stage autocorrelation circuit (22), so in subsequent processing, the LSB side is truncated due to tone restrictions. Even if the coefficient kl order 1 is
The prediction coefficient αi can be determined with high accuracy.

Note that at this time, the data Xt −W (n) is BSF
Only the T pit has been shifted to the left, so the BSF of "2"
It is multiplied by T, but the maximum value of data Xt - W (n) in each block is "2" B of ysax
Since the value BSFT is calculated so that the value multiplied by SFT does not overflow, the remaining data Xt-W (
Even if n) is multiplied by "2" to the BSFT power, it will not overflow.

Thus, according to the present invention, the weighted data Xt
-W(n) is multiplied in units of blocks within a range in which the maximum value Tsax does not overflow, so even if there is a word length limit in subsequent processing, the accuracy of the prediction coefficient αi can be increased.

G2 Other embodiments In the above, steps (61) to (33) are processed in the time window circuit (21), and step (71)
- (84) may be processed by a circuit newly provided between circuits (21) and (22).Also, although the above explanation was made using a flowchart, step (61)
) to (84) can be configured with discrete hardware. Furthermore, partial autocorrelation coefficient (PARCOR coefficient) k
i can also be used as the prediction coefficient αl.

Effects of the invention H According to this invention, the weighted data Xt·W(n)
is multiplied in block units within the range where its maximum value T wax does not overflow, so i
Even with the &4 restriction, the accuracy of the prediction coefficient αi can be increased.

[Brief explanation of the drawing]

FIG. 1 is a flowchart of an example of the present invention, and FIGS. 2 to 4 are diagrams for explaining the same. (10) is an encoder, (30) is a signal transmission system, (40
) is a decoder.

Claims (1)

[Claims] An input signal converted into a digital signal at a predetermined sampling frequency is set as one block for each predetermined number of samples, and the input signal is encoded using prediction residuals for each block. In the predictive coding method, the input signal is given a predetermined weight, the maximum value of each block is detected, the maximum value is multiplied so that it exceeds a predetermined value, and the block is assigned a predetermined weight according to this multiplication amount. A linear predictive coding method that calculates prediction coefficients by multiplying other samples within.