DK153266B - PROCEDURE TO REMOVE A NOISE SIGNAL FROM A COMPOSED SIGNAL AND ANIMAL EQUIPMENT TO EXERCISE THE PROCEDURE - Google Patents

Description

DK 153266 B

in

The invention relates to a method for removing a noise signal from a composite signal which, in addition to the noise signal, comprises a desired signal. The invention also relates to a device for practicing this method.

5 Various methods and equipment for removing noise are known. Over the last few years, significant advances have been made in the field of echo deletion in two-wire full-duplex transmission lines. An example of such a technique is described in the article: "A new Digital 10 Echo Canceller for Two-Wire Full-Duplex Data Transmission", by K.H. Mueller, in IEEE Transactions on Communications, Bind Com. 24, No. 9, pp. 956-962, September 1976, in the disclosure of U.S. Patent No. 4,087,644 (KH Mueller), in the disclosure to German Patent No. 3,009,450, and in Article 15 "Digital Echo Cancellation for Baseband Data Transmission "by Uerhoeckx, van den Elzen, Snijders, and Gerwen in the IEEE, Volume ASSP-27, No. 6, pp. 768-781, December 1979.

From the above references, it is known in the past to use a transversal filter to remove noise from a known noise source. The problem discussed in the first article relates to the removal of noise, or to put it another way, from reflections of the signal (s) transmitted in a non-ideal hybrid circuit.

As can be seen from the explanations in these references, it is known in the past to use a stochastic sign algorithm to adjust the tapping coefficients of a transversal filter included in such noise canceling equipment. When such an algorithm is used, the tapping coefficients are adjusted at each sampling with a fixed size or a fixed step of value V. The polarity of Ύ is determined by a multiplication of samples and the desired signal s and the noise signal e. This stochastic sign algorithm is defined by the following terms: 2

DK 153266B

aC | <= y. sgn (å + é) · (é).

Here, Δο ^ is the change in the filter coefficient c ^, r is the fixed value of the step change, 5 words shorten for sign or signal, & is a sampling of the desired signal s, and one sampling of the noise signal e.

It should be mentioned that the signal έ + é, which depends both on the desired signal s and on the noise signal e, is preferably already corrected with a roughly estimated estimate correction. Additional noise can also be added as explained below.

This algorithm represents a slow approach to obtaining the desired value. To avoid overcorrection, unstable control conditions with fluctuations, and, at worst, divergence, the value of "V" must be relatively small.

However, this is an unfortunate feature if large inequalities are detected between the desired and the current value, since only a slow adjustment to the desired value is obtained when r is small.

Therefore, using this simple stochastic sign algorithm, the value of Y must be carefully chosen. In reality, it must be chosen as a compromise between two requirements. i) The Ύ value must be so small that large fluctuations in the c ^ values will not occur. 2) The T-value must be so large that it will not take too long before an approximate correct value of is built up.

Because of these two contradictory claims, a really good estimate for Y will not be possible to find, instead a compromise must be satisfied which will detract from both requirements.

3

. .. DK 153266 B

This is a general disadvantage of the current noise cancellation circuits.

From the latter reference (the article by Verhoeckx et al.) It is even suggested that one can accumulate 5 the correction values over several bits before the final correction is made. This is implied indirectly by Equation 13 in the article and by the comments after that equation. Here it is stated that k can assume different values such as k = K, k = 2K, k = 3K, .... and 10 that is to say, accumulating over a similar number of bits.

Although some form of accumulation is made here, this will cause the same disadvantages as mentioned above. Fluctuations will occur or a slow adjustment will occur.

The method of the present invention, on the other hand, is to have an accumulation which, simply put, varies with the actual deviation.

The main object of the invention is to provide a new noise cancellation method by which both rapid adjustment of the filter coefficients is achieved when there is a large distance between the desired value and the existing value, and a stable and correct value for the filter coefficients when the inequality is small. . This is achieved by using a method and / or equipment according to the characterizing part of claim 1 or claim 11, respectively.

It can be said that the main idea of the present invention is based on the use of a sign correlation principle which is such that not only the polarity, but also the magnitude of the adjustments to the tapping coefficient will be made automatically. The size of the change- 4

DK 153266B

The adjustment of the value to the tapping coefficient should be estimated by adjustment in such a way that neither excessive overcompensation nor large undercompensation is obtained. Previously, no method has been known which has provided such characteristics.

BRIEF DESCRIPTION OF THE DRAWINGS The invention will now be described in more detail with reference to the accompanying drawings, in which: FIG. 1 shows the principle of a system using a noise cancellation method according to the invention; FIG. 2 shows how this method can be used in a transmission system; FIG. 3 shows how the probability density of a signal can be measured; FIG. 4 shows another method for determining the probable heat seal; FIG. 5 is a block diagram of a noise canceling device according to the invention; FIG. 6 shows a mathematical model for the principle of the present invention; and FIG. 7 shows the probability density of g and G ,.

I 'K

DK 153266B

5 In FIG. 1, the basic principle of the present invention is shown. Here, the signal generator 1 (S) produces a signal s.

Although it is primarily assumed that this signal is an information-carrying signal of electrical or electronic nature, it may as well be a signal based on other physical value. Thus, e.g. be low frequency noise produced by a variable magnetic field in a power transformer, where it is precisely the low frequency noise which is the undesired signal that one will suppress.

Thus, in such a case, the signals in question do not include any information, and the desired signal would then be silence (absence of acoustic noise).

A separate signal generator N1 (2) produces an error signal or a noise signal e. This error signal e may be subjected to various kinds of changes due to the nature of the surroundings. This is shown in FIG. 1 is indicated by a block 3 which represents a function c which is dependent on external parameters such as time (t), space or place (x), temperature (T), etc. 20 or the noise signal e are completely uncorrected signals, i.e. they cannot be derived from each other.

Now, the two signals s and e are added in an adding device 4 and the composite signal s + e is obtained. This composite signal s + e may also be subject to changes in a similar manner to signal e. All of these signal procedures are known in the past and are in fact normal signal processes in most physical systems.

Let us now consider the noise relief portion 8 of the circuit of FIG.

First

DK 153266B

6

The filter 5 represents a transfer function c, which is a close approximation of the transfer function c for block 3.

Since the noise signal e is applied to the filter 5, the output of the filter will be a signal é which is strongly approximated to the signal 5 e. In the subtraction unit 7 the signal e is subtracted from the composite signal s + e = s ^ s. The closer e is to e, the closer will be the desired signal s.

For the best approximation, filter 5 is assigned 10 different coefficients c (at least one). It is assumed that the number of coefficients is n, and they are denoted by the terms c ^, ... c ^, .... cn, where is one of any of the coefficients. The crucial point now is to select values for all c., So that the function c becomes as close as k 'to the function c as possible.

The following procedures are now followed:

A comparison circuit 10 or correlator receives samples of the noise signal e and the corrected signal s. In the correlator 10, these samples are compared or multiplied. The product of this multiplication is stored with its associated sign in accumulator 11. The next two samples from e and s are treated in the same way, and the new product is added (taking the sign into account) the contents of accumulator 11. This accumulation-25 process is continued, either for a predetermined period of time or until a predetermined accumulated value is reached. When this is done, the accumulated value (one for each filter coefficient) and the current correlation time (the period during which the accumulation is made) are read out to the arithmetic unit 12 and a new accumulation process is started.

The correlation time and the accumulated value, also called the correlation value, are both used to determine

DK 153266 B

7 the new coefficients for the filter 5, e.g. by determining the changes in the preceding values stored in the coefficient register 6. It can be assumed that the change in the coefficient with order number k is defined by m Δο, = γ. 2_ sgn (å + é) .. é.

i = l il 5 where the new symbols are Ύ, which is a factor (variable or constant) m is the number of accumulation steps made.

(It must always be remembered here that the έ + é signal is preferably corrected in advance by means of a first approximate 10 estimated correction, as assumed in Fig. 1 and in Claim 1).

Here, the index "i" can vary between 1 and m and represent the number of samples over which the correlation is accumulated, while the index k can vary between 1 and n representing the number of filter coefficients. In a transverse filter, the number of coefficients will correspond to the number of filter pins.

Whatever signals e and s + e are, it can be assumed that the sampling of these signals is done at intervals, and it is then these sampling values of the 20 signals that are passed to the correlator. The signals themselves can be analog or digital without affecting the process. And they can be coded according to any known coding rule. Therefore, the correlator 10 receives pair values of simultaneously generated instantaneous values of the signals.

In this way, therefore, a variable value of the coefficient c1 is obtained, depending on the accumulated value of the sign correlation. Therefore, if the samplings vary steadily in one direction during the entire correlation period, the change in the coefficient value with order number k, i.e. Ac ^, get a large positive or negative value. In this way, 8

DK 153266 B

also random fluctuations are suppressed safely.

The value of the accumulated sum m y— average (έ + é). * é ΙΞΪ is an integer of value between -m and + m and is easily found in 5 in a practical embodiment. The two signals é and (å + é) can be easily multiplied in an exclusive or circuit, and the sign for these products can be accumulated over m samples. The result can be stored in a register.

In a simple embodiment, the noise-canceling equipment 8 may merely consist of the units 5, 6, 10, 11 and 12 of FIG. First

However, a signal evaluation circuit 9 may be added to evaluate the instantaneous values taken by the signal taking into account the applicable code and other conditions. This signal evaluator also estimates, in a preferred embodiment, the probability density of the corrected signal s as explained below.

And if the desired signal s has a distribution equal to 0 (or near 0) at 0 amplitude, an additional noise with a controlled e.g. rectangular distribution, be added.

This is indicated by the dashed noise source (N2, 13) in FIG. First

The need and behavior of this additional noise will be explained below.

In all cases, sampling of the noise signal e to be removed and the corrected signal (s + e-e) are multiplied or otherwise compared in the correlator or comparison circuit 10 to give the relative sign of the samples. The correlation values, which may be the same (+1 or -1) or which may also have other values if more than one bit is considered, are accumulated in

DK 153266B

9 The accumulator 11. And the accumulated values, one for each filter coefficient, are read out after a predetermined period, or at the time when a predetermined accumulated value is reached. finally, the accumulation time, the accumulated values, and optionally the density, are converted into optimal filter coefficients by the arithmetic unit 12.

Let us now calculate the coefficient error or deviation in a binary system. Let's look at the system 10 of FIG. 2, which represents a transfer system with a n-tab transversal filter 18 and with the transfer function cii of the erasure equipment. Here, a ^, ··, a ^ ... represent an order of binary symbols to be transmitted, and together they represent the noise signal e; while s is the desired remote end signal with added noise. In addition, q is controlled by internal noise, and its significance will be described later. Finally, d the composite signal s + e is corrected with the erase error e and possibly with added controlled noise q.

20 If the entire system is now considered discrete and if you look at the leak e only at the sampling points, e can be expressed by: 03 e = a. C.

kio k k where c (= c ^, ···) represents the impulse response to the real leakage path.

25 e is the deletion signal given by n é => a, 6, k k where n is the number of pins in the transverse filter and c ^ is the coefficient of the pin k.

The embodiment of FIG. 1 illustrates the method of the invention used in a generally applicable system. Thus, 3Q, it can e.g. used to listen to details in the fetus-

DK 153266B

10 the sound. Then the sounds emanating from the mother will be noise to be erased from the composite signal that is obtained when listening to the fetus. The principle can generally be used to detect and study any weak signal blurred by strong noise signals uncorrelated with the desired signal.

Here, however, we will concentrate on showing the principle in an electronic transmission system. Such an embodiment is shown in FIG. 2. In this figure, one transmitter / receiver station 10, called station west, is shown, and it is connected, via transmission line 15, to a similar transmitter / receiver station east, which is not shown in the figure. The weak signal s is received on line 15 and here represents the desired signal, while the strong signal e from 15 transmitter / receiver west (16) represents the noise to be deleted because the signal e over the hybrid circuit 17 is superimposed on the desired signal s. and the composite signal s + e becomes the received result. In the figure, the reference numeral 18 represents a transversal filter with approximately the same transfer function c as the hybrid circuit 17. It is assumed that the subtraction unit 20 is analogous, and therefore a digital / analog converter 19 is disposed between the transversal filter 18 and the subtraction unit 20 *. the process is very similar to that shown in FIG. 1. However, it should be mentioned that 25 only one tap is included in the calculation of the figure, namely tap k.

It should also be noted that fig. 6 shows the mathematical model for the noise relief system, and in the following reference is made to figures 2 and 6.

30 For the sake of simplicity only the conditions at a single tap k are seen. See fig. 6. The composite error signal d is multiplied by the transmitted signal a ^. The product g ^ can be split into a deterministic part g ^ e caused by the coefficient error, and a stochastic or arbitrary part g ^ which mainly

DK 153266 B

n is uncorrelated with the emitted symbol a ^. The part includes the contribution of the other pins, ie. all bins other than the order number k. Assuming a binary system, ak is either +1 or -1, and it is found that g ^ is 5 a v-random variable with the expectation value and with a probability distribution pgk (x) as shown. in FIG. 7a. The peak value of the hit to g ^ is equal to r. The sign of g ^ is also any variable G ^ whose value can be +1 or -1.

10 The following terms are used: G, = measured, accumulated value for G. measured over m-steps.

_km

Gkm = expectation value to G ^ measured over m-steps.

pgk = probability distribution for g ^.

PGk = probability distribution for G ^.

15 If the added noise from the noise generator 21 (N2) now has a substantially rectangular distribution and is so strong that Δο ^ is less than r, pgk (x) will be practically constant up to x = 4 ° ^ · The probability distribution is shown in FIG. 7a and b for pgk and pGK.

20 If also assumes a Gaussian distribution, which is allowed when m is a large number, we get:

Gkm

Ack “m-pgk (o) and to justify the use of the measured value of Gkm instead of the expectation value G ^^ in this formula, m must satisfy the following expression:

Ack * pgk (o) m -—-—> i (l-2-.Ackpgk (o)) 2 25 We have now found not only a formula for calculating the coefficient deviation but also a formula by which the accuracy for the tap error calculation can be judged. The accuracy will usually not be good

DK 153266B

12 enough to give the correct tapping coefficients in one step, but the required number of repetitions will be quite low.

We now need a value for Pg | <(0) which can be inserted into the formula above. But first, let's look at the probability distribution Pg ^ (g ^). If the deletion error Ac is small and the internal noise q is not added, g ^ would be dominated by the remote end signal s. In a digital transmission system, both ρ ^ (0) = 0 and = 0 due to the signal pattern, and we can not at all obtain any 10 estimated value for Ac ^. Therefore, an uncorrected but controlled noise q must be added, which will fill the gaps in the distribution of the desired signal.

It is possible to measure ρ ^ (0) by sampling g ^, and detect how often the value falls between two threshold values selected symmetrically around g = 0.

A circuit for measuring the probability density function of this method is shown in FIG. Third

The composite signal after deletion (s + e-e) is fed to a pair of comparison circuits (D and E) with threshold levels 20 defined by resistance chain A, B, C. If the input voltage lies between the two threshold values -f-V ^. and -V ^., the output of AND gate F will be active and activate counter G. The counter is clock-controlled at a suitable frequency and reset after j count periods. If the counter output is just 25 before reset, we have:

L

pgk 2jvt *

This is the average value for pgk between the two threshold levels.

However, it may be more convenient to measure the peak value r and use the approximation:

DK 153266B

13

PG (0> “fe”

Another circuit for estimating the probability density relative to this latter method is shown in FIG. 4th

The composite signal s + ee is compared with a reference level from digital / analog converter I in comparator circuit 5 K. When the amplitude of the composite signal is greater than r, an impulse will be given to the controller H. The output r 'from H is a digital representation. of r. By making a successive approximation, the correct value of r is found. The probability density p ^ is then determined by the equation r, - L_ pgk 2r * 10 This is a good estimate if the internal applied noise q has a rectangular distribution and is the dominant amplitude of g. The amplitude of q must be chosen in such a way that the probability distribution is quite smooth without giving an extremely large value for the peak value r.

A more detailed description of a transmission system utilizing the present invention is shown in FIG. 5. Transfer data is shifted into register 30. The multiplexer 31 can select one of the last n transmitted bits, where n is a fixed number, e.g. 8. These n bits are multiplied by one by 20 a coefficient c ^ stored in Random Access Memory 32 (ARAM). The multiplication is performed by means of the exclusive-or-gate 33. The coefficient c ^ is the coefficient with the number i of a total number of n-coefficients. The result is accumulated in 34 so that the final erase value upon receipt of the last of the required n periods can be stored in the buffer storage 35 for controlling the digital / analog converter 36. Then new n periods occur in which the n-correlation values are increased or decreased. The correlation values k .. are retrieved from ARAM 32, the exclusive-or-gate is blocked, 30, and k ^ is increased or decreased in addition circuit 37 with the result 14

DK 1S3266B

the log multiplication between the error log and the current data bit is retrieved from register 30. The result is returned to ARAM 32 through buffer layer 38. When one of the correlation values k ^ exceeds a certain limit, processor 39 is interrupted, the correlation process stops and the processor reads all correlation values. k ^ through a two-way buffer buffer 40. The time frames previously used for correlation are now made available to processor 39 (MUX 49) for arbitrary access 10 to ARAM 32. Based on these correlation values, the time elapsed and the sign correlation, the processor now calculates new sets of coefficients c ^ which are written into ARAM 32 at address locations previously used for correlation values. Thereafter, the addressing for ARAM 32 15 is changed so that the new set of coefficients is used, while the range of the previous coefficients is deleted and used for correlation purposes.

The circuit 42 is a multiplexer, 43 a counter.

The accuracy depends on the accuracy with which 20 Pg ^ (0) (= il / 2r) and G ^ m can be determined. There is no reason to improve the estimate of G ^ m if the estimate of p ^^ (0) is the actual limiting factor.

The echo deletion circuit according to the present principle has been simulated on a computer. The calculation of the echo signal e and the far-end signal s has been carried out in the frequency range by means of existing computer programs (PSIM), where the cable characteristics have been remotely processed according to a known program model (BALCAB).

This is important to get a picture of the desired number of pins in the transverse filter. The remaining part of the system has been modulated in the time zone.

Several variants of the system have been simulated. The one chosen for the implementation of the equipment had the consequence-

DK 153266B

15 the characteristics: - bit rate 12 kb / s - line code: biphase - calculation of a new set of coefficients as soon as the correlation register has reached the value - 128.

- The length of the transversal filter comprises 8 symbols (bits) with a resolution of 4 samples per unit. bits.

- The line length is a maximum of 15 km, which corresponds to a attenuation of 50 dB at 12 kHz. The cable is made up of two 10 sections with different characteristics to give the effect of a remote echo.

- The controlled noise has a rectangular distribution and a given peak amplitude corresponding to the peak amplitude of the desired signal s.

15 It has been found that after 10 repetitions, or 15000 bits, the maximum value of the signal-to-error ratio is obtained. In order to suppress fluctuations, 30 m limiting the maximum achievable signal to noise ratio, it was necessary to reduce the Ac values obtained by the described algorithm as the tapping coefficients approach their final value.

When this algorithm is adapted for use in a practical system, various possibilities exist. The system tested in practice makes use of selecting a fixed 25 value for Gm optimized for the last repetitions in the tapping coefficient calculation. Each loss has its own correlation register. Each time a new one of the registers gets crowded, new tapping coefficients are calculated and put into use.

30 Signal processing systems tend to be costly when high speed is combined with complicated algorithms. Therefore, the arithmetic circuit has been divided into two parts. A fast, real-time arithmetic circuit generates deletion samples by the addition / subtraction of pins

DK 153266 B

16 coefficients, and increases / decreases in the correlation values Gm for all pins.

The calculation of new sets of tapping coefficients includes division and multiplication. This is done by a standard 5 microprocessor. The time required to perform the calculation is not critical as the system will run on the old set of coefficients until the new ones are ready for use.

The fast arithmetic circuit and the relatively slow microprocessor share a common coefficient storage. The memory is 10 divided into two parts a and b. The tapping coefficients are stored in one part a and the correlation values G ^ are accumulated in the other part b. When one of the correlation registers becomes overcrowded, the processor is interrupted and reads out all correlation values. Once the new tapping coefficients have been calculated, they are read into section b of the memory. Then the use of the two parts of the coefficient storage is exchanged. The arithmetic circuit reads out the tapping coefficients from section b and accumulates the correlation in section a.

20 The mathematical model of FIG. 7 includes a true time multiplication given by the following expression:

Gk = sgn (ak * (d)).

In the equipment realized, this was replaced by Gk = ak * sgn (d), which is carried out using an exclusive-or-gate.

Solutions according to the invention may also be of interest in noise canceling equipment, although the desired signal and the main signal are mutually correlated to each other. The principle described above can then also be used to detect an undesirable degree of correlation between the far-end signal s representing the desired information and the

DK 153266B

17 local transmitters emit signal e representing the noise. Because the correlation values accumulate before they are used, it is possible to allow a detected high degree of correlation between these two signals to affect an omission circuit which passivates the otherwise active coefficient update and utilizes the already last stored coefficients until uncorrelated signals again is obtained.

Claims (11)

A method of removing a noise signal (s) from a composite signal (s + s) comprising, in addition to the noise signal (s), a desired signal (s), wherein the noise signal (s) and the desired signal are mutually uncorrelated, wherein the noise signal 5 (e) or a signal which is a function of the same (e0) above is monitored and the desired signal (s) may be much weaker than the noise signal (s), the removal of the noise signal (s) being carried out by first transmitting the noise signal (e or e0) through a transverse filter circuit (5, 18) having a transfer function (e) producing an estimated value (s) of the continuous effect of the noise signal on the composite signal (e + s), so that a resultant signal (e + se) arises that sign information obtained from a comparison between the noise signal and the resulting signal in a correlation circuit (10) is used to generate the correction values (Ac ^) for adjustable coefficients (c ^) in the transverse filter circuit (5, 18), whereby they obtained new corrections coefficient values are used to derive from the filter the subsequent estimated value (s), whereby the resultant signal (e + se) is obtained a gradual approximation to the desired signal (s) and the sign information obtained from a pure sign comparison between instantaneous values of the monitored noise signal (s) and the composite signal (s + s), consisting solely of + 1's and 25's, is accumulated to form an accumulated sign correlation value (which may have a absolute value '1), characterized in that the accumulation of the sign-correlated values runs continuously over a variable period, the duration of which is determined by the continuous magnitude of the correlation value (Ac ^), the larger the time the smaller the correlation value is; the final accumulated sign correlation value that exists at the end of the current period, together with the duration of the variable accumulation period, is used in an arithmetic unit (12) to determine the subsequent filter coefficient (s) (c ^ ... ... c) or the subsequent adjustment value (40 ^ ... 40 ^ ... 4cn) for the same and that this process is repeated continuously throughout the course of the signal. Method according to claim 1, characterized in that the accumulation of the sign correlated values runs continuously until a predetermined accumulated correlation value is obtained. The method of claim 1 or 2, wherein at least two filter coefficients are used, characterized in that the accumulation, which is preferably a pure summation of sign correlation values, is made in parallel for all filter coefficients and the accumulation is continued until one of the accumulated values. reaches a predetermined value, after which all values are reset and a new accumulation process is initiated to determine new filter coefficients or new adjustment values for the same, while the last obtained accumulated log values divided by the 20 accumulation time are used as a factor in the discharge of new filter coefficients. Method according to claim 1, 2 or 3, characterized in that the probability density (pgk) of the composite signal (s + e) is measured or evaluated and used as a scaling factor in determining the filter coefficients. Method according to claim 4, characterized in that the adjustment values for the filter coefficients (Ac Ac) are calculated to be proportional to the accumulated sign correlation value (γ - - average) and inversely proportional to the correlation value (represented by m) and with the probability density (pgk) of the composite signal (s + e) for amplitudes less than the maximum noise amplitude. DK 153266 B 20 The method of claim 1, 2, 3, 4 or 5, wherein the probability density (pgk) of said amplitude range is either very small or not constant (non-uniform) and wherein controlled noise (q) with an approximately rectangular for - 5 division and with a maximum amplitude (s) at least equal to the maximum signal amplitude is added to the corrected signal (s + ee), characterized in that the probability density is added to the corrected signal with controlled noise (q) (s + e) -e + q) is used to calculate the change in filter coefficients. Method according to any one of the preceding claims, characterized in that the noise signal (s) is a digital signal or a linear derivation of such a signal. Method according to claim 5, 6 or 7, characterized in that the value for the probability density (pgk) is estimated by measuring the relative time in which the composite signal (s + e-e + q) has a value between two predetermined thresholds. Method according to claim 5, 6 or 7, characterized in that the value for the probability density (pgk) is estimated by inversion of the maximum amplitude (2r) for the composite signal (s + e-e + q). Method according to any one of claims 1-9, characterized in that the filter coefficients (c 1) and the accumulated correlation values between the composite signal and signals from the noise source are stored at different memory locations or in completely different memories, that the current coefficients are used. i · fast real-time periods with a non-destructive read / write operation, that the correlation values (and duration of the correlation time) are read out slowly in a local time period to calculate new corrected filter coefficient values, while the new corrected coefficients are loaded. in the address locations where the correlation factors were previously stored and the connections are only switched to memory 5 upon completion of a complete set of new corrected filter coefficients, so that the memory portion that was last used as coefficient storage is now used to store new accumulated correlation values , while the new set of coefficients t ages in use for calculating the next generation of deletion values. 11. Noise canceling circuit (8) to reduce the noise (s) in a composite signal (s + e), which consists partly of a desired signal (s) from a signal generator (1) and partly of a noise signal (s) which is coming from a noise generator (2) and transformed by a noise transfer function (c) present in outer parameters (3), said noise cancellation circuit (8) comprising a correlator (10), an accumulator (11) and a transverse filter (5) ) with multiple (1 ... k ... n) pins (6), each having an adjustable filter coefficient (c ^) and the output of the transverse filter (s) coupled to the composite signal (s) + e) to the input side sf a subtraction named (7), characterized in that the output (s) of the subtraction unit (7) and the output of the noise generator (2) are coupled to the correlator (10) and are compared where the result of this comparison is accumulated in the accumulator (11) with variable accumulation time, that signals representing accumulated values and accumulation time are passed to a arithmetic unit (12) and passing the output of this arithmetic unit (12) to the pins (6) of the transverse filter (5) for correction of previous filter coefficients.