JPH0244176B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a waveform distortion removal method for acoustic signals, etc., and in particular, for a partially distorted signal waveform such as a peak-clipped signal waveform.
This invention relates to a signal restoration method that attempts to remove the distortion by restoring the distorted part by estimating it from the non-distorted part.

Conventionally, a signal waveform that has undergone nonlinear distortion could be restored by passing it through a circuit that has input/output characteristics with the inverse characteristic of the nonlinear characteristic, if the inverse characteristic of the nonlinear characteristic can be realized. Conventionally, it has been impossible to restore signal waveforms that have undergone nonlinear distortion and do not have inverse characteristics, such as clips.

SUMMARY OF THE INVENTION It is therefore an object of the present invention to provide a signal restoration method that eliminates the above-mentioned drawbacks when linearly predicting a signal waveform with partial signal loss from a portion without signal loss.

In order to achieve such an object, the present invention provides a signal restoration method for restoring a signal waveform with partial signal loss from a part without signal loss by linear prediction. A forward estimated waveform is formed from the waveform by forward linear prediction, a backward estimated waveform is formed by backward linear prediction from the signal waveform after the end of the signal loss, and a weighted average waveform of the forward estimated waveform and the backward estimated waveform is formed. The method is characterized in that the weighted average waveform is arranged as a restored signal in the signal missing portion.

The present invention will be described in detail below with reference to the drawings.

First, a forward linear prediction method, which is the basis of the signal restoration method of the present invention, will be explained. Acoustic signals (including the human voice) most often originate from vibrating systems that are described by linear differential equations, and signals generated from such vibrating systems can be linearized with good accuracy. Can be predicted. That is,
If the sampled signals are ..., x _l-2 , x _l-1 , x _l , x _l+1 , x _l+2 , ..., then the forward linear predicted value x^ _l = _P 〓 ^k=1 α _k x _lk (1) is a good forward estimate of x _l . Here, α ₁ , α ₂ , ..., α _p are the simultaneous linear equations where R(m) is the m-th order autocorrelation coefficient of the signal. This is the solution. However, p needs to be appropriately determined depending on the characteristics of the signal.

Using this fact, one section of the signal can be estimated from its adjacent sections. That is, the sampled signals ..., x _i-2 , x _i-1 , x _i , x _i+1 , ..., x _j , x _j+1 ,
Let's assume that the interval of x _i , x _i+1 , ..., x _j is unknown for some reason in ..., and consider estimating it. First, calculate the autocorrelation coefficients R(0), R(1), ..., R(p) from the data in the known interval, and (2)
Solve the equation to find α ₁ , ..., α _p . Next, by repeatedly using equation (1), the forward linear predicted values x _{^ i} , x^ _i+1 , ..., x^ _j of x _i , x _i+1 , ..., x _j are calculated. Find them sequentially. In this way, x _i , ..., x _j
The estimated value of is obtained, and the original waveform can be approximately restored. However, if the unknown signal section becomes long, this alone may cause problems. In other words, as k increases, the error between the forward linear predicted value x^ _k and the true value x _k increases, and the forward linear predicted value x^ _j corresponding to the last sampled signal of the unknown interval and the unknown interval A step may occur in the waveform at the connection with the first sampled signal value x _j+1 after completion.

Therefore, the present invention aims to eliminate such drawbacks, and uses the distorted section of the signal as an unknown section, and calculates the level difference from the forward linear predicted value and backward linear predicted value from the undistorted section. Enables restoration of signals that do not occur.

More specifically, the backward linear predicted value x _l = _P 〓 ^k=1 α _k x _l+k (4) is also a good backward estimate of x _l , so (4)
Using the formula repeatedly, the backward linear predicted values x _i , ..., x _j of x _i , ..., x _j are Find them sequentially by Then, the weighted average of the forward predicted value and the backward predicted value, for example, _k = 1/j-i ((j-k) x^ _k + (k-i) x _k ) (6) (k = i, ..., j) Let be the estimated values of x _i , ..., x _j . _k is the estimated value of x k in which the part close to x _i-1 is weighted with the forward linear predicted value, and the part close to j+ ₁ is weighted with the backward linear predicted value, _i = x^ _i , _j = x _j Since this holds true, the above-mentioned problem of the difference in level at the connecting portion does not occur.

FIG. 1 shows an example of the configuration of an apparatus for removing signal waveform distortion using the signal restoration method of the present invention. In FIG. 1, a signal supplied to a terminal A passes through a sampling filter 1, is converted into a digital signal by an A/D converter 2, and is temporarily stored in a digital input buffer 3. Of the signals stored in the digital input buffer 3, a portion to be processed is extracted using a time window and stored in the processing buffer 4. Reference numeral 5 denotes a distortion-free section detection unit that detects an undistorted section of the signal waveform among the signals stored in the processing buffer 4. Taking clip distortion as an example, it is as shown in FIG. In the peak clipped waveform, the part where the amplitude is smaller than the clip level (solid line part in Figure 2)
is detected as an undistorted section.

6 is the signal stored in the processing buffer,
This is an autocorrelation coefficient calculation unit that calculates autocorrelation coefficients R(0), R(1), ..., R(p) from data in undistorted signal sections, and the autocorrelation coefficients calculated here are The relationship coefficients are led to the simultaneous linear equation solving section 7, where the solutions α ₁ , α ₂ , . . . , α _p of equation (2) are found.

Here, in the case of audio signals, the autocorrelation coefficient does not change rapidly, so the autocorrelation coefficient obtained from the sampled signal near the signal missing part is forward-looking,
Since it can be used for either backward linear prediction, the result from the solver 7 is transferred to the terminal B of the forward predictor 8 and the terminal D of the backward predictor 9. The forward predictor 8 is a part that calculates a forward linear predicted value, and a detailed example thereof is shown in FIG. Here, the signal...
..., x _i-2 , x _i-1 , x _i , x _i+1 , ..., x _j , x _j+1 , ...
Let x _i , ..., x _j be the distorted sections. In FIG. 3, 21 is a shift register, and shift register stages 21-1, 21-2, . . .
..., 21-p, and each initial value of the forward linear prediction from the processing buffer 4 is stored in these shift register stages 21-1, 21-2, ..., 21-p.
It is assumed that x _i-1 , x _i-2 , . . . , x _ip are set via the initial value setter 22. The solutions α ₁ , α ₂ , _.
-2, . . . , 23-p, respectively.
Values α ₁ x i-1 , α _{2 x i-2 ,} α _{p x} ip obtained by passing the outputs from each register stage 21-1 to 21-p through coefficient units 23-1 to 23-p, respectively is supplied to the adder 24, and the added output is taken out from the output terminal 25. As a result, for every clock supplied to the shift register 21, the forward linear predicted values x _{^ i} , x _{^ i+1} , ..., x _{j ^} of x i , x _i+1 , ..., x _j are Output terminal 25
can be obtained sequentially from

The backward predictor 9 that calculates backward linear predicted values can be constructed in the same way as the forward predictor ₈ , and in _{FIG .} 1 to 23-p, and set x _j+1 +x _j+2 , ..., x _j+p from the processing buffer 4 as the initial value of each stage 21-1 to 21-p of the shift register 21. If you set x _j , x _j-1 ,
..., backward linear predicted value of x _i x _j , x _j-1 , ..., x _i
is obtained.

Forward linear predicted value x _i ^ from forward predictor 8,
x _i+1 , . . . , x _j ^ and backward linear predicted values x _j , x _{j-1 ,} . In this weighted average unit 10, the forward linear predicted value x _i
Weighted averages i, ..., _j of _^ , ..., x _j ^ and backward linear predicted values _x i , ..., _{x j} are calculated, and the results are led to the waveform connector 11. The signal from the processing buffer 4 is also supplied to this waveform connection unit 11, of which the distorted signal sections x _i , x _i+1 , ..., x _j have their estimated values _i , x _{i +1} , ..., replaced by _i .
The signal section for which processing has been completed in this manner is outputted in the form of a digital signal via the output buffer 12, or is supplied to the D/A converter 13, where it is D/A converted, and then sent to the sampling filter 14. , and is taken out from output terminal H as an audio signal.

As is clear from the above, according to the present invention, it is possible to reduce the distortion of a signal that is partially distorted and can detect the distorted section. It is applied to remove distortion from acoustic signals, and has the effect of reducing the auditory sense of distortion.

As explained above, the present invention is based on the principle of estimating the unknown part of a signal from the known part. It can also be effectively applied to restoring waveforms.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an example of the configuration of a signal restoration device according to the present invention, FIG. 2 is a waveform diagram showing an example of a peak clip waveform, and FIG. 3 is a block diagram showing an example of the configuration of a linear predictor. . 1... Sampling filter, 2... A/D converter, 3...
Digital input buffer, 4... Processing buffer, 5
...Distortion-free interval detection unit, 6...Autocorrelation coefficient calculation unit, 7
...Simultaneous linear equation solving unit, 8...Forward predictor, 9
... Backwards predictor, 10... Weighted average section, 11... Waveform connection section, 12... Output buffer, 13... D/A converter, 14... Sampling filter, 21... Shift register, 21-1 to 21-p... Shift register stage, 2
2... Initial value setter, 23-1 to 23-p... Coefficient unit, 24... Adder, 25... Output terminal.

Claims (1)

[Claims] 1. In a signal restoration method in which a signal waveform with partial signal loss is restored by linear prediction from a portion without signal loss, the signal waveform of a predetermined section immediately before the signal loss occurs. A forward estimated waveform is formed by forward linear prediction, a backward estimated waveform is formed by backward linear prediction from the signal waveform of a predetermined section immediately after the end of the signal loss, and a weighted average waveform of the forward estimated waveform and the backward estimated waveform is calculated. The following formula _k = 1/j-i ((j-k) x^ _k + (k-i) x _k ) (k = i,..., j) [ _k : Weighted average x^ _k : Forward linear predicted value x _k : backward linear predicted value] and arranges the weighted average waveform as a restored signal in the signal missing portion.