JPH05145378A - Adaptive digital filter having adaptive input signal white level processing - Google Patents

Abstract

PURPOSE:To improve the convergence of the adaptive digital filter by providing a linear FIR digital filter whose filter coefficient is variable to a proper location of the adaptive digital filter and controlling the filter adaptively. CONSTITUTION:A DLY is an element delaying a signal by a time and outputting the result and a COF is an element multiplying a coefficient with the signal and outputting the result. A block 1a is an adaptive linear FIR digital filter and a block 1h is an arithmetic operation section deciding a characteristic of white level processing of the block 1a. Moreover, a signal x(n) is an input signal, a d(n) is an object signal, e(n) is an error output signal, U(n) is a signal resulting from applying white level processing to the X(n) and the v(n) is a signal resulting from applying white level processing to the d(n). The adaptive digital filter inputs the signal X(n) to an M-degree of FIR digital filter whose filter coefficient is variable, its output y(n) is used to correct filter coefficients h0(n)-hM(n) sequentially based on the input signal d(n) with correlation to as to be made close to the input signal d(n).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an adaptive filter application apparatus which requires a real-time processing, for example,
The present invention relates to an adaptive digital filter added with an adaptive input signal whitening process used in an adaptive noise canceller, an eco-canceler, or the like.

[0002]

2. Description of the Related Art Conventionally, in an application device of an adaptive digital filter that requires real-time processing, a complicated processing method cannot be used due to the limitation of the processing performance of a processor at present, and basically, a procedure is required. Zing, I-A-Y-A, Volume 63, 1975, Page 16921-1716 (Proc. IEEE, Vol. 63, (1975), pp1692-171.
The processing method of the adaptive noise canceller discussed in 6) is adopted. This is a simplification of the calculation of the internal matrix assuming that the input signal is white noise, and for signals other than white noise, the speed and accuracy of obtaining the optimum solution, so-called convergence, were poor. In recent years, since this convergence is related to the degree of autocorrelation of the input signal, many studies have been conducted to reduce the autocorrelation of the signal and bring it closer to white noise to improve the convergence.

Also, with the need to improve the quality of acoustic signals,
Improvement in performance is also desired in applied equipment.

[0004]

From the above, the conventional adaptive filter has a problem in that its performance is limited when it is applied to an adaptive noise canceller or an eco-canceler that handles actual voice or the like. On the other hand, making the process too complicated for this improvement is a problem in real-time processing.

It is an object of the present invention to bring an input signal close to white noise by the simplest possible processing, improve the convergence of the adaptive digital filter within a range where real-time processing is possible, and improve the performance of the adaptive operation of the applied device. It is what

[0006]

To achieve the above object, the present invention provides a first-order FIR (Finite Impulse
Response: FiniteImpulse Response) By installing a digital filter and adaptively controlling these,
Whitening is applied to an input signal whose properties are unknown, and the convergence of the adaptive digital filter is improved.
In addition, the processing is simplified by performing the above-mentioned control by focusing on only the component of the autocorrelation of the signal that is delayed by one time.

In order to achieve the above whitening treatment,
The first-order FIR digital filter is used to remove the autocorrelation component delayed by one time, and the filter coefficient is sequentially adjusted to an optimum value by a relatively simple arithmetic process based on the gradient method. It was done.

[0008]

The adaptive first-order FIR digital filter for whitening the signal in the previous period works to adaptively remove the autocorrelation component of the one-time portion of the signal, and the process for adjusting the filter coefficient is the above-mentioned filter. By updating the filter coefficient so that the average of the product of the output signal of 1) and the input signal of one hour before becomes closer to 0, the autocorrelation component delayed by one time of the input signal acts so as to be the smallest.

[0009]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of an adaptive digital filter to which an adaptive input signal whitening process of the present invention is added. Newly added to the conventional adaptive digital filter by the present invention are blocks 1a and 1b. FIG. 2 shows a block configuration of the block 1a. In the figure, DLY is an element that delays the signal by one time and outputs it, and COF is an element that outputs the signal after multiplying it by a coefficient. A block 1a is an adaptive first-order FIR digital filter that performs whitening processing on a signal, and a block 1b is an arithmetic unit that determines characteristics of the signal whitening processing of the block 1a. Further, the signal x (n) is the input signal of the adaptive digital filter, d (n) is the target input signal, and e
(n) is the error output signal, u (n) is the first tap of the FIR digital filter that constitutes the adaptive digital filter,
A signal obtained by whitening the signal x (n) by the block 1a, v
(n) is a signal obtained by whitening d (n).

The adaptive digital filter inputs the signal x (n) to an M-th order FIR digital filter whose filter coefficient is variable, and outputs this output y (n) as another correlated input signal d (n). As a target, filter coefficients h ₀ (n) to h
_It is a method of sequentially correcting _M (n) and bringing them closer to each other. If the order M is sufficiently large and the signal d (n) can be obtained from x (n) by a linear conversion as shown in Expression 1, the signal y
(n) can be made completely equal to d (n), and the error output signal e (n) can be zero.

[0011]

## EQU1 ## d (n) = w ₀ x (n) + w ₁ x (n-1) + ... + w _M x (nM) At this time, the M-th order FIR digital filter becomes the optimal Wiener filter and each filter coefficient h _It shows that ₀ (n) to h _M (n) converged to the optimal Wiener solution w _{0 to} w _M. this is,
This means that the echo canceller identified the echo transmission path, and the adaptive noise canceller identified the noise propagation path. However, in reality, the order M must be cut off within a practical range, and the signal d (n) involves not only a linear conversion from x (n) but also some disturbance, so that it is necessary to converge to the optimal Wiener solution. It is difficult, and this approximation is obtained. However,
The conventional adaptive digital filter converges to the fastest and best approximate solution when the input signal x (n) is white noise with no autocorrelation, and even for an input signal with strong autocorrelation such as speech. In order to bring out this performance, the functional blocks 1a and 1b of the present invention are added.

The operation of the adaptive digital fill using the configuration of the present invention will be described below with reference to FIGS. 1 and 2. First,
The input signals x (n) and d (n) are defined. x (n) is an approximation of a signal with strong autocorrelation, such as voice, and is represented by the following expression, which is a signal in which the autocorrelation of a part delayed by one time is added to white noise.

[0013]

## EQU00002 ## x (n) =. Epsilon. (N) +. Alpha.x (n-1) .epsilon. (N) is white noise, and .alpha. Is a parameter indicating the degree of autocorrelation. The target input signal d (n) is calculated from the equations 1 and 2 as follows.

[0014]

[Equation 3]

These signals are processed by the block 1a as follows.

[0016]

## EQU00004 ## u (n) = x (n)-. Beta. (N) x (n-1) =. Epsilon. (N) +. Alpha.x (n-1)-. Beta. (N) x (n-1)

[0017]

[Equation 5]

Β (n) is an estimated value of α. At this time, the output signal y (n) of the FIR digital filter forming the adaptive digital filter is as follows by the product-sum operation with the filter coefficients h ₀ (n) to h _M (n).

[0019]

[Equation 6]

Here, if β (n) is the optimum value α, Equations 5 and 6 become Equations 7 and 8, respectively.

[0021]

[Equation 7]

[0022]

[Equation 8]

The error output signal at this time is e (n) = v (n)-
Since y (n), the filter coefficients h ₀ (n) to h _M (n) of the Mth-order FIR digital filter are estimated only by processing time-series signals of white noise ε (n). This improves the convergence to the optimal Wiener solution and improves the performance of the adaptive digital filter.

In parallel with this, the operation of the block 1b at this time determines the coefficient β (n) for performing optimum signal whitening. This is because the output signal u (n) when the signal x (n) is applied to the block 1a is closest to white noise and has the smallest variance when the coefficient β (n) is the optimum value. , This minimum point is searched by a calculation based on the gradient method. This arithmetic expression is shown in Expression 9.

[0025]

## EQU9 ## β (n + 1) = β (n) + ηu (n) × (n-1) η is called a step size parameter and determines the size of update of β (n). Adjust in advance to a suitable value that converges quickly without diverging. Input signal x
When the amplitude of (n) is normalized to -1 to 1, it is roughly
It takes a value of 0.1 ≦ η <1. The arithmetic algorithm for estimating the filter coefficient of the M-th order FIR digital filter forming the adaptive digital filter is not particularly mentioned, but the most used LMS (Least Mean Square) at present. ) Method and learning identification method, which are relatively easy to operate.

[0026]

The adaptive digital filter of the present invention, to which the adaptive input signal whitening process is added, exhibits better convergence than a conventional adaptive digital filter even when a signal having a strong autocorrelation is input, and the processing amount is increased. Since it is also small, when applied to an adaptive noise canceller or an eco-canceller that handles voice and the like and requires real-time processing, the time required to attenuate noise and eco-can be shortened, and the amount of attenuation also increases. The effect is obtained.

[Brief description of drawings]

FIG. 1 is a block diagram of an adaptive digital filter to which an adaptive input signal whitening process is added according to the present invention.

FIG. 2 is a block configuration diagram of an adaptive first-order FIR digital filter that performs whitening processing of an input signal, which corresponds to block 1a in FIG.

[Explanation of symbols]

DLY: delay element, COF: coefficient element, 1a: signal whitening adaptive first-order FIR digital filter, 1b: signal whitening filter coefficient adjusting unit. x (n): input signal, d
(n): target input signal, u (n): whitened input signal, v
(n): Whitened target input signal, e (n): Error output signal.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Akira Osawa 216 Totsuka-cho, Totsuka-ku, Yokohama Hitachi, Ltd. Information & Communication Division (72) Inventor Yoshinote Ota 292 Yoshida-cho, Totsuka-ku, Yokohama Hitachi, Ltd. Video Media Research Institute

Claims (2)

[Claims] 1. A first-order FIR digital filter having a zero-order filter coefficient of 1 and a variable first-order filter coefficient is formed by the input position of a signal as a target of the adaptive digital filter and the adaptive digital filter. An FIR digital filter which is provided at the output position of each tap of the FIR digital filter so that all the primary filter coefficients of the variable FIR digital filter have the same value, and this value constitutes the adaptive digital filter. In the configuration in which the input signal is adjusted to approach white noise by the input signal, the whitening of the input signal shortens the time required to estimate the filter coefficient of the adaptive digital filter and improves the accuracy of the estimation. The adaptive digital filter to which the adaptive input signal whitening process is added is characterized by Data. 2. A first-order FIR digital filter in which the 0th-order filter coefficient is 1 and the first-order filter coefficient is variable, and the product is added to the product of the input signal of the filter one time before and the output signal of the filter. A processing method for signal whitening, characterized in that a first-order filter coefficient of the filter is adjusted so that an output signal always approaches white noise in a configuration in which processing is performed every time (sampling period).