JP2603065B2 - Prediction filter - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a prediction filter that performs successive adaptive prediction.

(Prior Art) FIG. 4 is a conceptual diagram of a general predictor used in a compression encoding circuit. All signals are discrete time series that have been sampled, and the subscript _(k) means the sample value at time k. In FIG. 4, an input signal S _(k) is input to a prediction filter circuit 42, and the prediction filter circuit 42 outputs a prediction signal _(k) to an adder 44 based on a linear combination coefficient a _i described later. The adder 44 includes a prediction signal _(k) and an input signal S _(k).
, And outputs a residual signal e _(k) .

That is, e _(k) = S _(k) - _(k) (1) At this time, the prediction signal _(k) is an output of the prediction filter circuit 42 calculated by a linear combination of the input signal S _(k) . Here, _{(k) where ai} is a linear combination coefficient (that is, a filter coefficient ₎ is
It can be expressed by the following equation.

n is equal to the order of the filter. In general, the larger the value of n, the more accurate the prediction. This filter coefficient a
_i is a fixed coefficient optimal for the signal sequence S _(k) when the property of the input signal S _(k) does not fluctuate so much, that is, when the spectrum of S _(k) does not fluctuate, but S _(k) If the spectrum fluctuates with time, the filter coefficients a _i are not fixedly set, but are sequentially selected as optimal ones corresponding to the fluctuations. I can't make good predictions. Various methods of controlling the filter coefficient a _{i in} such a case are conceivable, but a method based on the following equation (3) for controlling the power of the residual signal e _(k) to be minimum is general.

a _{i (k)} = a _{i (k-1)} + d · sgn (e _(k) · S _(k) ) (3) where d is the acceleration coefficient, sgn () is the polarity judgment, and A function that takes the sign of a variable.

However, the prediction filter circuit that uses a _i updated as a filter coefficient by the equation (3) does not guarantee the stability, and its guarantee method has been limited to the second order conventionally. Therefore, even if an adaptive type predictor is used, an accurate prediction signal cannot be obtained due to the restriction of the order (second order). Therefore, a configuration method of cascading the fixed predictor 57 and the adaptive predictor 58 as shown in FIG. 5 has been considered. As shown in FIG. 5, the fixed prediction filter circuit 52 and the adaptive prediction filter circuit
When the cascade connection 54 is simply connected, first, the fixed prediction filter circuit 52 calculates the prediction signal _{p1 (k)} for the input signal S _(k) , and outputs the prediction signal _{p1 (k)} to the adder 53. Output. The adder 53 outputs a residual signal e1 _(k) based on the prediction signal _{p1 (k)} and the input signal S _(k) . Subsequently, the residual signal e _{1 (k)} is input to the adaptive prediction filter circuit 54, and the adaptive prediction filter circuit 54 outputs a second prediction signal _{p2 (k)} based on the residual signal e _{1 (k).} Together with its prediction signal _{p2 (k)}
Is output to the adder 55. The adder 55 calculates a residual signal e _(k) on the basis of the residual signal e _{1 (k)} and the prediction signal _{p2 (k).}

(Problem to be Solved by the Invention) In the above-described conventional configuration shown in FIG. 5, the filter coefficient of the fixed prediction filter circuit 52 is set so that it can be optimally predicted for an input signal (data signal) having a small variation in frequency spectrum. Is set. Therefore, in many cases, the fixed prediction filter circuit 52 does not perform optimal prediction for an input signal (for example, an audio signal) other than the set input signal. Thus, according to the prior art, the fixed prediction filter circuit 52 regardless of the frequency spectrum of the input signal S _(k), and outputs a prediction signal _{p1 (k)} with respect to the input signal S _(k). Therefore, the frequency spectrum of the residual signal e _{1 (k)} obtained in this case is an unspecified one having a larger variation than the audio signal. That is, although the frequency spectrum of an audio signal fluctuates with time, the residual signal consisting of the difference between the amplitude of the audio signal and the prediction signal of the fixed prediction filter circuit 52 for the audio signal is larger than the fluctuation of the frequency spectrum of the audio signal. It has a fluctuating frequency spectrum.

As a result, the adaptive prediction filter circuit 54 assuming the fluctuation of the frequency spectrum of the audio signal generates the residual signal e _{1 (k)}
Cannot generate a prediction signal having a high correlation with

 Hereinafter, the problems of the related art will be described in detail.

A single adaptive prediction filter circuit can use various input signals, for example, a signal having a small variation in frequency spectrum regardless of a change in time (hereinafter, referred to as a data signal) or a signal whose frequency spectrum changes with time. In the case of designing so as to output a prediction signal having a high correlation with respect to two kinds of signals (hereinafter, referred to as an audio signal), the order of the filter is increased to cope with the problem. Is limited by the order of the filter.
It is difficult to generate a prediction signal having a high correlation with the type of signal.

Therefore, in general, it is conceivable to use a fixed prediction filter circuit 52 and an adaptive prediction filter circuit 54 that can generate prediction signals having high correlation with the two types of signals. However, simply connecting in cascade as shown in FIG. 5 causes the following problems.

That is, in the conventional configuration shown in FIG. 5, the fixed prediction filter circuit 52 has a fixed (preset) filter coefficient so that a prediction signal having a high correlation with the data signal can be output. Further, the adaptive prediction filter circuit 54 sequentially updates its own filter coefficient based on the above-described equation (3) with respect to the audio signal, and generates a prediction signal having a high correlation with a signal whose frequency spectrum varies with time. It is configured to output. However, the adaptive prediction filter circuit 54 is supposed to successively update the filter coefficients within the variation range of the frequency spectrum with respect to the audio signal and output a prediction signal having a high correlation with the audio signal due to the limitation of the filter order. .

In the conventional configuration shown in FIG. 5, regardless of whether the input signal is a data signal or a voice signal, first, a fixed prediction filter circuit 52 generates a first prediction signal for the input signal, and an adder 53 generates a first prediction signal for the input signal. The difference between the amplitude of the input signal and the amplitude of the first prediction signal is output as a first residual signal.
Next, a second prediction signal corresponding to the first residual signal is generated in the adaptive prediction filter circuit 54 using the first residual signal as an input signal, and the first residual signal and the second
Is output as a second residual signal (final residual signal).

Accordingly, when the input signal is a data signal, the first prediction signal having high correlation can be generated in the fixed prediction filter circuit 52, and thus the difference is the difference between the amplitude of the data signal and the amplitude of the first prediction signal. The amplitude of the first residual signal is minimized. That is, the first residual signal is the optimal residual signal for the data signal. However, the first residual signal is directly input to the adaptive prediction filter circuit 54. First
The adaptive prediction filter circuit 54 cannot generate a second prediction signal having a high correlation with the first residual signal, because the residual signal has a different frequency spectrum from that of the speech signal. Accordingly, the second residual signal composed of the difference between the amplitudes of the first residual signal and the second prediction signal is a signal having a larger amplitude than the first residual signal, and the final output (residual signal) ) Is inappropriate.

On the other hand, when the input signal is a speech signal, the adaptive prediction filter circuit 54 generates a prediction signal, and a residual signal including a difference between the speech signal and the prediction signal of the adaptive prediction filter circuit 54 is output as a final output. However, this cannot be done with the conventional configuration of FIG. 5, and the audio signal is input to the fixed prediction filter circuit 52. As described above, since the frequency spectrum of an audio signal fluctuates with time, the fixed prediction filter circuit 52 in which the filter coefficient is set for a data signal cannot generate a prediction signal having a high correlation with the audio signal. Therefore, the first residual signal which is the difference between the amplitude of the audio signal and the first prediction signal of the fixed prediction filter circuit 52 is not only an optimal residual signal for the audio signal, but also the frequency spectrum of the audio signal. The signal is out of the fluctuation range. For this reason, even if the adaptive prediction filter circuit 54 generates a prediction signal for the first residual signal using the first residual signal as an input signal, it is not possible to generate a highly correlated prediction signal. Therefore, the second residual signal comprising the difference between the amplitude of the first residual signal and the amplitude of the second predicted signal of the adaptive prediction filter circuit 54 also does not become the optimal residual signal for the audio signal.

As described above, if the fixed prediction filter circuit 52 and the adaptive prediction filter circuit 54 are simply cascaded, an optimum residual signal cannot be output with respect to the input of the data signal and the audio signal.

As described above, each of the conventional techniques has a problem, and accurate prediction cannot be performed for various input signals.

The present invention eliminates the above-described problems, generates a predicted signal having high correlation with various input signals (data and voice), and calculates the difference between the input signal and the predicted signal (the amplitude difference between the two signals). The purpose is to reduce the amount of transmission data by transmitting.

(Means for Solving the Problems) A prediction filter of the present invention derives a first prediction signal from an input signal using a fixed filter coefficient set for an arbitrary signal sequence of a discrete input signal. A fixed predictor; a first adder for deriving a first residual signal from the input signal and the first predicted signal; and a filter coefficient based on a past input time series in the first residual signal. An adaptive predictor for updating and deriving a second prediction signal from the first residual signal, and a second predictor for deriving a second residual signal from the second prediction signal and the first residual signal. An adder, which is connected to the output side of the fixed predictor, and based on the polarity of a correlation function consisting of a product of the first prediction signal and the first residual signal, The amplitude of the predicted signal of A first adaptive amplitude control circuit for increasing and decreasing and outputting to the first adder, connected to an output side of the adaptive predictor, and configured to calculate a product of the second prediction signal and the second residual signal. And a second adaptive amplitude control circuit for increasing and decreasing the amplitude of the second prediction signal by a predetermined relaxation factor based on the polarity of the correlation function and outputting the result to the second adder. It is characterized by the following.

(Operation) The prediction filter of the present invention includes a fixed predictor and an adaptive predictor, and an adaptive amplitude control circuit is connected to the output side of each predictor. The adaptive amplitude control circuit controls the amplitude of the prediction signal output from each predictor based on the polarity of the correlation function that is the product of the residual signal and the prediction signal. That is, control to attenuate the amplitude and control to restore the amplitude are performed.

Therefore, the prediction filter of the present invention can perform accurate prediction on various input signals including data and voice.

(Embodiment) The present invention provides an amplitude control circuit that connects a stable sequential adaptive predictor of any order and a fixed predictor in cascade and independently controls the amplitude of a prediction signal output from each predictor. It is provided.

FIG. 1 is a block diagram showing a configuration of such a successive adaptive predictor. In FIG. 1, a fixed predictor 12 sets a fixed filter coefficient, and outputs a prediction signal that is close (highly correlated ₎ to an input signal S _(k) having a small variation in frequency spectrum. It is designed to be. Further, the adaptive predictor 15 is designed to selectively update the filter coefficient so as to sufficiently follow a signal whose frequency spectrum fluctuates rapidly. in this way,
Fixed predictor 12 and adaptive predictor independently designed
In the configuration in which 15 are connected in cascade, adaptive amplitude control circuits 13 and 16 are connected to the output side of each predictor, respectively. The adaptive amplitude control circuit 13 controls to attenuate the amplitude of the prediction signal S ′ _{p1 (k)} output from the fixed predictor 12 and restores the amplitude to the original value (closer to the amplitude of the input signal) based on Expression (4) described later. Perform control. The adaptive amplitude control circuit 16
Based on the expression (10) described later, control is performed to attenuate the amplitude of the prediction signal ' _{p2 (k)} output from the adaptive predictor 15 and to restore the amplitude (to approach the amplitude of the input residual signal).

FIG. 2 is a circuit diagram showing the configuration of FIG. 1 in detail.
The configuration and operation of the prediction filter will be described in detail with reference to FIG.

In Figure 2, the composed fixed predictor 201 transversal filter 22, fixed is a set of filter coefficients, the approximation of the (correlated to variations in the frequency spectrum is small input signal S _(k) Signal)
It is designed to output _{p1 (k)} . The fixed predictor 201, the prediction signal to the input signal _{S (k) 'p1 (k} )
Is output to the adaptive amplitude controller 202.

The adaptive amplitude controller 202 performs control to attenuate the amplitude of the prediction signal ' _{p (k)} and control to restore the amplitude to the original value based on the following equation (4). Here, these controls of the adaptive amplitude controller 202 will be described in detail. The coefficient g ₁ of the adaptive amplitude controller 202 is
In consideration of the followability and stability of the prediction filter, the prediction filter is controlled by the relationship shown in the following equation (4).

_{g 1 (k) = δ g1} · {g 1 (k-1) -G 1} + G 1 + d g1 · sgn (e 1 (k) · 'p1 (k)) ... (4) where [delta] _g1: g Relaxation coefficient for smoothing the change of _{1 (k)} , G ₁ : g when the predicted signal has low correlation with the input signal
_{1 (k)} is an arbitrary constant that converges, d _g1 : acceleration coefficient corresponding to the amount of change from the previous value g _{1 (k-1)} based on the result of polarity stability, sgn: polarity determination, parentheses Is a function that takes the sign of the variables in.

 This equation (4) represents the following three actions.

The first effect is that d _g1 · sgn is based on the assumption that the predictor outputs a predicted signal highly correlated with the input signal.
The term (e _{1 (k)} · ′ _{p1 (k)} ) brings the amplitude of the predicted signal closer to the input signal based on the polarity judgment of the correlation function consisting of the product of e _{1 (k)} and ' _{p1 (k)} It is to increase or decrease as follows. In the case of only this condition, the expression (4) is expressed as g1 _(k) = g1 _(k-1)
+ D _g1 · sgn (e _{1 (k)} · ′ _{p1 (k)} ).
This means that the coefficient g1 _(k) is increased or decreased at a rate of the acceleration coefficient _{dg1 with} respect to the coefficient _{g1 (k-1)} one sample period ago. That is, the difference between the amplitude of the prediction signal and the amplitude of the input signal in units of one sample is compensated.

Specifically, when the amplitude of the prediction signal is smaller than the amplitude of the input signal, the polarity of sgn becomes positive, and acts to increase the amplitude of the prediction signal at the rate of the acceleration coefficient d _g1 .
Conversely, when the amplitude of the prediction signal is larger than the amplitude of the input signal, the polarity of sgn becomes negative, which acts to reduce the amplitude of the prediction signal at the rate of the acceleration coefficient d _g1 .

The second effect is to control the amplitude of the predicted signal when the predictor cannot output a predicted signal highly correlated with the input signal by the relaxation coefficient δ _g1 (experimentally 0.99). In this case, equation (4) gives g _{1 (k)} = δ _g1 · g
_{1 (k-1)} + _dg1 · sgn (e1 _(k) · ' _{p1 (k)} ).

This does not compensate for the amplitude for each sample unit described above, but reduces the amplitude of the predicted signal output from the predictor when a predicted signal having a high correlation with the input signal is not output (finally, Acts as “0”). Specifically, assuming a coefficient after n samples in the above equation, it is composed of a term of δ _{g1 to} the nth power and a term including sgn (e _{1 (k)} · ′ _{p1 (k)} ). Become. During this time, if the predictor cannot output a predicted signal having high correlation with the input signal, the magnitude relationship between the amplitude of the input signal and the predicted signal is randomly changed for each sample, so that sgn (e _{1 (k)} · ′
The polarity of _{p1 (k)} ) is also exchanged. Therefore, sgn (e
_The term including _{1 (k)} · ' _{p1 (k)} cancels out due to the difference in polarity, and finally converges to 0 by the term of δ _g1 (0.99) to the nth power. Therefore, if a predicted signal having a high correlation with the predictor input signal cannot be output, control is performed to reduce the amplitude of the predicted signal.

The third effect is the expression (4) is a point relates arbitrary constants G _1. Although the equation includes an arbitrary constant G ₁ , the arbitrary constant G ₁ is basically set to 0, and
Two actions can be achieved. However, when manufacturing an actual device, it is not always optimal to converge the amplitude of the predicted signal to “0” when the predictor cannot output a predicted signal having high correlation with the input signal. . Therefore, the arbitrary constants G ₁ is intended to be set in consideration of various conditions in the actual apparatus.

As can be seen from the term d _g1 · sgn (e _{1 (k)} · ′ _{p1 (k)} ) in the equation (4), the adaptive amplitude controller 202 determines whether the residual signal e _{1 (k)} and the prediction signal ′ _{p1 (} The amplitude of the prediction signal is increased or decreased so as to approach the input signal by judging the polarity of the correlation function formed by the product of the correlation function _k) . That is, when the predicted signal ' _{p1 (k)} and the input signal S _(k) are close to each other (high correlation), that is, when accurate prediction is performed, the predicted signal'
_When the amplitude of _{p1 (k)} is larger than the amplitude of the input signal, the polarity is determined to be positive (+). At this time, the adaptive amplitude controller
In step 202, control is performed to gradually reduce the amplitude of the input prediction signal ' _{p1 (k)} and output the signal. That is, the prediction signal _{p1 (k)}
At the rate of the acceleration coefficient d _g1 . When the amplitude of the prediction signal ' _{p1 (k)} is smaller than the amplitude of the input signal, the polarity is determined to be negative (-). At this time, the adaptive amplitude controller 202 gradually increases the amplitude of the input predicted signal ' _{p1 (k)} at the rate of the acceleration coefficient d _g1 and outputs the predicted signal _{p1 (k).}
Is output.

Further, (4) the _{δ g1 · {g 1 (k} -1) -G 1} terms and d _g1
As can be seen from the term of sgn (e _{1 (k)} · ' _{p1 (k)} ), when the predicted signal' _{p1 (k)} and the input signal S _(k) are not approximate (low correlation), That is, when prediction with low accuracy is being performed, the adaptive amplitude controller 202 performs control to attenuate the amplitude of the input prediction signal ' _{p1 (k)} .

As shown in FIG. 6, the internal structure of the adaptive amplitude controller 202 is obtained by expanding equation (4) into a circuit form.

The prediction signal _{p1 (k)} is input to the adder 24, and the adder 24 calculates a residual signal e1 _(k) from the input signal S _(k) and the prediction signal _{p1 (k)} . Then, the residual signal e _{1 (k)} is input to the adaptive predictor 203.

For the adaptive predictor 203, see ICC '84 IEEE Internati
onal Conference on Communications, Proceedings, Volu
It is described in detail in me 3 (pp.1500 to 1503), and it stably constructs an iterative adaptive predictor of any order.
Here, only the specific configuration will be described.

The adaptive predictor 203 that inputs the residual signal e _{1 (k)} and outputs the prediction signal ' _{p2 (k)} is a transversal filter 25.
, A coefficient control circuit 26, and a coefficient conversion circuit 27. Transversal filter 25 based on the filter coefficient output from the coefficient control circuit 26 and the coefficient conversion circuit 27, the prediction signal based on the residual signal e _{1 (k) 'p2 (k)} to the adaptive amplitude controller 204 Output. The coefficient control circuit 26 receives the residual signal e _{1 (k)} and the residual signal e _(k) and outputs filter coefficients (control signals) ω ₁ , ω ₂ ,..., Ω _m to the coefficient conversion circuit 27. . To explain in more detail, the coefficient control circuit 26
Inputs the adaptive predictor 203 residual signal is input to the e _{1 (k)} and the residual signal e outputted from the adder 29 _(k), on the frequency axis filter coefficients omega _i follows (5) Update according to the formula.

The coefficient ω _i is given by the following (6) for security assurance.
Satisfies the size relation of the expression.

0 <ω ₁ <ω ₂ <... <Ω _n <π (6) The coefficient conversion circuit 27 calculates the filter coefficient ω _i on the frequency axis controlled as in the equations (5) and (6) as follows. (7) of
Converts the filter coefficient h _i on the time axis defined by the identity equation, the filter coefficients (control signal) h _1, h _2,
..., and outputs the h _m in the transversal filter 25.

Here, C (Z) and D (Z) are represented by a pair of Chebyshev polynomials as shown in equations (8) and (9).

The adaptive amplitude controller 204 performs attenuation control and control to restore the amplitude of the prediction signal ' _{p2 (k)} based on the following equation (10). Here, these controls of the adaptive amplitude controller 204 will be described in detail. Coefficient g ₂ of the adaptive amplitude controller 204, like the coefficient g ₁ of the adaptive amplitude controller 202 described above, in consideration of the follow-up and stability of the predictive filter, controlled by the relationship shown in the following equation (10) Have been. That is, [delta] _g2 is the value of g _{2 (k)} converges when low correlation relaxation factor for smoothing the variation, with respect to G ₂ is the prediction signal is an input signal g _{2 (k)} any D _g2 is the previous value g based on the result of the polarity determination.
Acceleration coefficient corresponding to the value that increments for _{2 (k-1)} , sgn
Is a polarity determination and is a function that takes the sign of the variable in parentheses.

_{g 2 (k) = δ g2} · (g 2 (k-1) -G 2) + G 2 + d g2 · sgn (e (k) · 'p2 (k)) ... (10) (10) formula d _g2 As can be seen from the term sgn (e _(k) · ' _{p2 (k)} ), the adaptive amplitude controller 204 calculates the correlation function consisting of the product of the residual signal e _(k) and the predicted signal' _{p2 (k).} And the amplitude of the prediction signal is increased or decreased so as to approach the input signal. Predicted signal ' _{p2 (k)} and input residual signal e
_{1 (k)} is approximate (high correlation), that is, when accurate prediction is performed, and the prediction signal '
_When the amplitude of _{p2 (k)} is larger than the amplitude of the input residual signal e1 _(k) , the polarity is determined to be positive (+). At this time,
The adaptive amplitude controller 204 performs control of gradually attenuating the amplitude of the input prediction signal ' _{p2 (k)} and outputting the signal. That is, the amplitude of the prediction signal _{p2 (k)} is gradually reduced at the rate of the acceleration coefficient _dg2 . When the amplitude of the prediction signal ' _{p2 (k)} is smaller than the amplitude of the input residual signal e1 _(k) , the polarity is determined to be negative (-). At this time, the adaptive amplitude controller 204
Is the amplitude of the input prediction signal ' _{p2 (k)}
Control is performed such that the prediction signal _{p2 (k)} is output while gradually increasing at the rate of _g2 .

Furthermore, terms and d _g2 of (10) of _{δ g2 · {g 2 (k} -1) -G 2}
As can be seen from the term of sgn (e _(k) · ' _{p2 (k)} ), the predicted signal' _{p2 (k)} and the input residual signal e1 _(k) are not approximate (the correlation is low). ), That is, when prediction with low accuracy is being performed, the adaptive amplitude controller 204 performs control to attenuate the amplitude of the input prediction signal ' _{p1 (k)} .

As shown in FIG. 7, the internal configuration of the adaptive amplitude controller 204 is obtained by expanding equation (10) into a circuit form.

Prediction signal _{p2 (k)} are input to the adder 29, the adder 29 calculates a residual signal e _{1 (k)} and the prediction signal _{p2 (k)} from the residual signal e _(k), and the calculated remaining The difference signal e _(k) is output.

Subsequently, the operation when a specific input signal (data signal and audio signal) is input will be described.

For example, when the input signal S _(k) is a data signal having a small variation in frequency spectrum, the fixed predictor 201 outputs a prediction signal ' _{p1 (k)} highly correlated with the input signal S _(k) . I do. In this case, the adaptive amplitude controller 202 calculates the prediction signal ' _{p1 (k)} and the residual signal e1 _(k) as shown in Expression (4 _).
Based on the polarity determination of the correlation function consisting of the product of the above, control is performed to bring the amplitude of the prediction signal _{p1 (k)} closer to the amplitude of the input signal S _(k) . As a result, the residual signal e _{1 (k)} becomes a signal with a very small amplitude, that is, a signal with good prediction accuracy.

Subsequently, the residual signal e _{1 (k)} is input to the adaptive predictor 203, the adaptive predictor 203, because it is designed to be easy to follow the speech signal, the residual signal e _{1 (k )} near the prediction signal _{'p2 (k)} it is often not output. Therefore, in this case, the adaptive amplitude controller 204 determines the polarity of the correlation function consisting of the product of the prediction signal ' _{p2 (k)} and the residual signal e _(k) as shown in Expression (10). Based on the predicted signal '
Control is performed to attenuate the amplitude of _{p2 (k)} . as a result,
Although the summed prediction signal _{p2 (k)} and the residual signal e _(k) and in the adder 29, the residual signal e _(k) is not affected by the prediction signal _{p2 (k),} almost residual signal Output as e _{1 (k)} .
That is, since the prediction signal ' _{p2 (k)} from the adaptive predictor 15 does not affect the residual signal e1 _(k) , the residual signal e1 _{(k )} Is output as the final residual signal e _(k) .

On the other hand, when the input signal S _(k) is a speech signal whose frequency spectrum varies with time, the fixed predictor 201 often cannot output a prediction signal _{p1 (k)} having a high correlation with the speech signal. Therefore, unlike the case of the data signal described above, the adaptive amplitude controller 202 calculates the product of the prediction signal ' _{p1 (k)} and the residual signal e1 _(k) as shown in the equation (4). Based on the polarity determination of the correlation function, control is performed to attenuate the amplitude of the prediction signal ' _{p1 (k)} . Therefore, the residual signal e between the input signal S _(k) and the prediction signal _{p1 (k)} in the adder 24
_{1 (k)} is a signal that is approximately (highly correlated ₎ approximate to the input signal S _(k) .

Subsequently, when a (mutually high) residual signal e _{1 (k)} (近似 speech signal) approximate to the input signal S _(k) is input to the adaptive predictor 203, the adaptive predictor 203 outputs the residual signal e _The prediction signal ' _{p2 (k)} having a high correlation with _{1 (k)} is output. Therefore, the adaptive amplitude control circuit 1
6, as shown in equation (10), based on the polarity judgment of the correlation function consisting of the product of the prediction signal ' _{p2 (k)} and the residual signal e _(k) , the prediction signal _{p2 (k)} Control is performed so that the amplitude approaches the amplitude of the residual signal e _{1 (k)} input to the adaptive predictor 203.
As a result, in the adder 29, an accurate prediction signal
_{p2 (k)} and the residual signal e _{1 (k)} (≒ input signal S _(k) ) are added, and the residual signal e _(k) having the minimum amplitude is output.

FIG. 3 is a circuit diagram showing a second embodiment according to the present invention. In this example, a zero-type adaptive predictor is added to the predictor of the first embodiment.
306 is zero which is cascaded to the input of the zero shape adaptive predictor 306, in front of the front differential signal e _{2 (k),} to control the filter coefficient b _i with the e _{2 (k)} This is a coefficient control circuit 301. Update of the coefficient b _i is performed using the following equation (11).

b _{i (k)} = δ _b · (b _{i (k-1)} −B _i ) + B _i + d _b · sgn (e _(k) · e _(k-1) … (11) The coefficient updated in this way Bi is used as a filter coefficient of the transversal filter 302, and a prediction signal _{z (k)} is calculated.

Further, an example in which the cascade relation as described in the first embodiment is exchanged and the adaptive predictor is connected to the preceding stage and the fixed predictor is connected to the subsequent stage can be easily inferred and belong to the present invention.

(Effects of the Invention) As described above, according to the present invention, the adaptive amplitude controller performs control to attenuate the amplitude of the prediction signal based on the polarity determination of the correlation function including the product of the prediction signal and the residual signal. (Or control to return to the amplitude of the input signal). Therefore, a residual signal can be derived based on a highly accurate prediction signal output from an optimal predictor (fixed predictor or adaptive predictor). Also, a prediction signal output from a predictor that is not suitable for the properties of the input signal does not affect the residual signal.

Therefore, the prediction filter of the present invention widely follows various signals having different properties, and enables accurate prediction. Further, since the prediction accuracy is improved, it can be applied to a band compression transmission device such as an ADPCMCODEC or an echo canceller.

[Brief description of the drawings]

FIG. 1 is a block diagram of a predictor according to the present invention, FIG. 2 is a first embodiment of the present invention, FIG. 3 is a second embodiment of the present invention,
FIG. 4 is a conceptual diagram of a predictor, FIG. 5 is a diagram showing an example of a combination of a fixed predictor and an adaptive predictor, FIG. 6 is a circuit diagram of an adaptive amplitude controller 202 in the embodiment of FIG. The figure is a circuit diagram of the adaptive amplitude controller 204 in the embodiment of FIG. 12, fixed predictor; 13, 16 adaptive amplitude control circuit; 14, 17 adder; 15 adaptive predictor

 ──────────────────────────────────────────────────続 き Continuing on the front page (72) Inventor Shinji Kawaguchi 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. (72) Inventor Atsushi Shinbo 1-7-112 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd.

Claims (2)

(57) [Claims] 1. A fixed predictor that derives a first prediction signal from an input signal using a fixed filter coefficient set for an arbitrary signal sequence of a discrete input signal; A first adder that derives a first residual signal from one predicted signal, and a filter coefficient that is updated based on a past input time series in the first residual signal, and the first residual signal An adaptive predictor that derives a second prediction signal from a second predictor, and a second adder that derives a second residual signal from the second prediction signal and the first residual signal. An output of the predictor, which is connected to the output of the first prediction signal and the first residual signal, based on the polarity of a correlation function consisting of the product of the first residual signal, the amplitude of the first prediction signal of a predetermined relaxation coefficient Increase / decrease by the ratio and output to the first adder. The first to
An adaptive amplitude control circuit, which is connected to an output side of the adaptive predictor, based on a polarity of a correlation function composed of a product of the second predicted signal and the second residual signal. The amplitude of which is increased or decreased by a predetermined relaxation coefficient ratio and output to the second adder
A predictive filter, comprising: an adaptive amplitude control circuit according to (1), and (2). 2. The prediction filter according to claim 1, wherein said adaptive predictor has a transfer function comprising a pair of Chebyshev polynomials using coefficients arranged on a unit circle on a Z plane. A predictive filter characterized by using an adaptive predictor that sequentially updates the power of the residual signal so as to minimize the power of the residual signal and controls the updated coefficients to be always separated.