EP1154674A2 - Circuit and method for adaptive noise suppression - Google Patents

Abstract

The circuit for adaptive noise suppression is part of a digital hearing aid consisting of two microphones (1, 2), two AD converters (3, 4), two compensation filters (5, 6), two delay elements (7, 8), two subtractors ( 9, 10), a processing unit (11), a DA converter (13), a receiver (15) and the two filters (17, 18). The method for adaptive noise suppression can be implemented with the specified circuit. Depending on their spatial arrangement or their directional characteristics and depending on the location of the acoustic signal sources, the two microphones (1, 2) deliver different electrical signals (d1 (t), d2 (t)) which are present in the AD converters (3, 4) digitized and pre-processed with the two fixed compensation filters (5, 6). This is followed by the filters (17, 18) arranged symmetrically crosswise in the forward direction with the adaptive filter coefficients (w1, w2). The filter coefficients (w1, w2) are calculated using a stochastic gradient method and updated in real time while minimizing a quadratic cost function consisting of cross-correlation terms. This selectively amplifies spectral differences in the input signals. With suitable placement of the microphones (1, 2) or selection of the directional characteristics, the signal-to-noise ratio of output signals (s1, s2) can thus be significantly increased compared to that of the individual microphone signals (d1 (t), d2 (t)). One of the improved output signals (s1, s2) is preferably subjected to the customary hearing-device-specific processing in one of the processing units (11, 12), sent to one of the DA converters (13, 14) and acoustically again via one of the listeners (15, 16) spent. In the present invention, four additional cross-link filters (19-22) perform a signal-dependent transformation of the input and output signals (y1, y2; s1, s2), and only the transformed signals are used to update the filter coefficients (w1, w2) . This enables the filter coefficients (w1, w2) to be quickly reacted and yet computationally efficient and, in contrast to other methods, causes only minimal audible distortion. <IMAGE>

Description

The present invention relates to a circuit and a method for adaptive Noise suppression according to the generic terms of the independent Claims. For example, it is used in digital hearing aids.

The healthy human hearing system allows you to relax during a conversation in a noise situation disturbed by noise towards a conversation partner focus. Many hearing aid users, however, suffer from a greatly reduced Speech intelligibility as soon as next to the desired speech signal There are noises.

Many methods of noise cancellation have been proposed. she can be divided into single-channel methods, which only have one input signal need, and in multi-channel processes, which by means of several acoustic Inputs that use spatial information in the acoustic signal.

No relevant improvement of the Speech intelligibility can be demonstrated. It will only improve the subjectively perceived signal quality achieved. In addition, these fail Procedure in the case of practical importance in which both the useful and the Interference signal is speech (so-called cocktail party situation). None of the single channel method is able to convert a single speech signal from one Highlight mixture selectively.

In the multi-channel method for noise suppression, the Assumption assumed that the acoustic source from which the Useful signal is sent out in front of the listener while the noise is off other directions. This simple assumption has proven itself in practice and accommodates the supportive lip reading. The multi-channel Processes can be further divided into fixed systems, which have a fixed one have predetermined directional characteristics, and adaptive systems, which adapt to the current sound situation.

The fixed systems work either using directional microphones, which have two acoustic inputs and one from the direction of incidence dependent output signal, or using several Microphones whose signals are processed electrically. Manual Switching may allow you to choose between different Polar patterns. Such systems are and will be available on the market increasingly also built into hearing aids.

One is hoping from the adaptive systems currently in development that that it optimally interferes with noise depending on the current sound situation suppress and thus surpass the fixed systems. An approach with an adaptive directional microphone was described in Gary W. Elko and Anh-Tho Nguyen Pong, "A Simple Adaptive First-Order Differential Microphone", 1995 IEEE ASSP Workshop on Applications of Signal Processing to Audio and Acoustics, New Paltz NY. The shape of the Polar pattern set depending on the signal. This allows a single one to the side incoming interference signal can be suppressed. By restricting yourself to one only adaptive parameters, the system works only in simple Acoustic situations with a single interference signal.

Numerous studies have been made using two Microphones that are placed on each ear. With these so-called adaptive Beamformem are sum and difference signal of the two microphones as Input used for an adaptive filter. The basics of this type of Processing was by L. J. Griffiths and C. W. Jim, "An Alternative Approach to Linearly Constrained Adaptive Beamforming ", IEEE Transactions on Antennas and Propagation, vol. AP-30 no. 1 pp. 27-34, Jan. 1982. This Griffiths-Jim beamformer can also work with more than two microphone inputs. This can successfully suppress noise. But there are problems the room echoes present in real rooms. In extreme cases this can do this cause the useful signal to be suppressed or distorted instead of the interference signals.

In recent years there has been great progress in the area of so-called blind signal separation has been made. A good compilation of the previous research results can be found in Te-Won Lee, "Independent Component Analysis, Theory and Applications ", Kluwer Academic Publishers, Boston, 1998. It starts from an approach where M is statistical independent source signals from N sensors in different Mixing ratios are included (M and N are natural numbers), where the transfer functions from the sources to the sensors are unknown. It is the goal of blind signal separation, from the known sensor signals to reconstruct statistically independent source signals. This is in principle possible if the number of sensors N is at least the number of sources M corresponds, d. H. N ≥ M. Many different algorithms have been proposed been, most not at all for efficient processing in Are suitable in real time.

The algorithms that can be considered as a subgroup are: instead of statistical independence, only uncorrelatedness of the reconstructed Request source signals. These approaches are from Henrik Sahlin, "Blind Signal Separation by Second Order Statistics, "Chalmers University of Technology Technical Report No. 345, Gothenburg, Sweden, 1998 been. He was able to demonstrate that uncorrelated output signals are required completely sufficient for acoustic signals. For example, minimization with a quadratic cost function consisting of cross correlation terms be carried out using a gradient method. Thereby filter coefficients gradually changed in the direction of the negative gradient. Such one The process is described in Henrik Sahlin and Holger Broman, "Separation of Real World Signals ", Signal Processing vol. 64 no. 1, pp. 103-113, Jan. 1998. There it is used for noise suppression on a mobile phone.

It is an object of the invention to provide a circuit and a method for adaptive Noise suppression specify which on the known systems build, but surpass them in essential properties. In particular, should optimal convergence behavior with as little effort as possible minimal, inaudible distortion and without additional signal delay can be achieved.

The problem is solved by the circuit and the method as described in the independent claims are defined.

The present invention belongs to the group of blind systems Signal separation using second order methods, i.e. H. with the aim of Achievement of uncorrelated output signals. Essentially, there will be two Microphone signals by means of blind signal separation into useful signal and interference signals Cut. Consistent behavior at the exit can be achieved if that Signal-to-noise ratio of a first microphone is always greater than that of a second microphone. This can be achieved either by: the first microphone is placed closer to the useful source than the second microphone or in that the first microphone as opposed to the second microphone has a directional characteristic geared to the source of use.

The decorrelated output signals are calculated with minimization a quadratic cost function consisting of cross correlation terms. For this purpose, a special stochastic gradient method is derived, in which Expected values of cross correlations are replaced by their instantaneous values become. This has a quick-reacting and computationally efficient update of the Filter coefficients result.

Another difference from the generally known method is that for the update of the filter coefficients transformed versions depending on the signal the input and output signals are used. The transformation by means of Cross-link filter performs spectral smoothing so that the signal powers be distributed more or less evenly over the frequency spectrum. This means that when the filter coefficients are updated, all spectral components are weighted evenly regardless of the current one Power distribution. This also does not allow for real acoustic signals neglecting autocorrelation functions a low-distortion Processing with satisfactory convergence behavior.

For an optimal functioning of the circuit according to the invention and the The method according to the invention can be used with the microphone inputs Compensation filter are compared. It will be one Standardization size used for updating all filter coefficients. she is calculated in such a way that only one of the two filters with maximum Speed is adapted, depending on whether currently the useful signal or Interference signals dominate. This approach enables correct convergence even in the singular case where only useful signal or only interference signals are present.

The present invention differs significantly from all of them so far published systems for noise suppression, in particular by the special stochastic gradient methods, the transformation of the signals for the update of the filter coefficients as well as the interplay of Compensation filter and standardization unit for controlling the Adaptation speed.

Overall, the system according to the invention has a very large range of signal-to-noise ratios consistent behavior, i. H. the signal-to-noise ratio is always improved and never deteriorated. So it can make an optimal contribution to better communication in difficult sound situations.

Fig. 1

ein allgemeines System zur adaptiven Geräuschunterdrückung mittels der Methode der blinden Signaltrennung gemäss Stand der Technik,

Fig. 2

das erfindungsgemässe System,

Fig. 3

eine Detailzeichnung eines Kompensationsfilters des erfindungsgemässen Systems,

Fig. 4

eine Detailzeichnung eines Verzögerungselements des erfindungsgemässen Systems,

Fig. 5

eine Detailzeichnung eines Filters des erfindungsgemässen Systems,

Fig. 6

eine Detailzeichnung eines Kreuzglied-Filters des erfindungsgemässen Systems,

Fig. 7

eine Detailzeichnung eines Kreuzkorrelators des erfindungsgemässen Systems,

Fig. 8

eine Detailzeichnung einer Vorberechnungseinheit vom Typ V des erfindungsgemässen Systems,

Fig. 9

eine Detailzeichnung einer Vorberechnungseinheit vom Typ B des erfindungsgemässen Systems,

Fig. 10

eine Detailzeichnung einer Aufdatierungseinheit des erfindungsgemässen Systems,

Fig. 11

eine Detailzeichnung eines Kreuzglied-Dekorrelators des erfindungsgemässen Systems,

Fig. 12

eine Detailzeichnung einer Glättungseinheit des erfindungsgemässen Systems und

Fig. 13

eine Detailzeichnung einer Normierungseinheit des erfindungsgemässen Systems.

The invention is described in detail below with reference to figures. The block diagrams show:

Fig. 1

a general system for adaptive noise suppression using the method of blind signal separation according to the prior art,

Fig. 2

the system according to the invention,

Fig. 3

1 shows a detailed drawing of a compensation filter of the system according to the invention,

Fig. 4

2 shows a detailed drawing of a delay element of the system according to the invention,

Fig. 5

2 shows a detailed drawing of a filter of the system according to the invention,

Fig. 6

2 shows a detailed drawing of a cross-link filter of the system according to the invention,

Fig. 7

2 shows a detailed drawing of a cross correlator of the system according to the invention,

Fig. 8

2 shows a detailed drawing of a type V precalculation unit of the system according to the invention,

Fig. 9

2 shows a detailed drawing of a type B precalculation unit of the system according to the invention,

Fig. 10

2 shows a detailed drawing of an update unit of the system according to the invention,

Fig. 11

2 shows a detailed drawing of a cross-link decorrelator of the system according to the invention,

Fig. 12

a detailed drawing of a smoothing unit of the system according to the invention and

Fig. 13

a detailed drawing of a standardization unit of the system according to the invention.

A general system for adaptive noise suppression using the method of blind signal separation, as is known from the prior art, is shown in FIG. 1 . Two microphones 1 and 2 deliver the electrical signals d ₁ (t) and d ₂ (t). The following AD converters 3 and 4 determine digital signals therefrom at the discrete times d ₁ (n * T) and d ₂ (n * T), in abbreviated form d ₁ (n) and d ₂ (n) or d ₁ and d ₂ . T = 1 / f _{s is} the sampling period, f _{s is} the sampling frequency and n is a continuous index. Compensation filters 5 and 6 follow which, depending on the application, can carry out a fixed frequency response correction on the individual microphone signals. The resulting input signals y ₁ and y ₂ are now led to delay elements 7 and 8 and to filters 17 and 18 according to FIG. 1. Subsequent subtractors 9 and 10 provide output signals s ₁ and s ₂ .

Processing units 11 and 12 follow which, depending on the application, carry out any linear or nonlinear postprocessing. Their output signals u ₁ and u ₂ can be converted into electrical signals u ₁ (t) and u ₂ (t) via DA converters 13 and 14 and made audible by loudspeakers or earphones 15 and 16.

The aim of the blind signal separation is to obtain output signals s ₁ and s ₂ which are as statistically independent as possible, starting from the input signals y ₁ and y ₂ and using the filters 17 and 18. For the only briefly stationary acoustic signals, the requirement of uncorrelated output signals s ₁ and s ₂ is sufficient. We will minimize a cost function for the calculation of the optimal filter coefficients w ₁ and w ₂ in filters 17 and 18. It is the following quadratic cost function J consisting of cross correlation terms. The operator ^* stands for conjugate complex in applications where we are dealing with complex signals.

The cross correlation terms can be expressed using the output signals s ₁ and s ₂ . The operator E [] stands for the expected value. R s 1 s 2 ( l ) + E [ s * 1 ( n ) · s 2 ( n + l )]

The output signals s ₁ and s ₂ can be expressed by the input signals y ₁ and y ₂ and by means of the filter coefficients w ₁ and w ₂ . W _1k denote the elements of the vector w ₁ and w _2k the elements of the vector w ₂ .

To minimize the cost function J by means of a gradient method, the derivatives with regard to the filter coefficients w ₁ and w _{2 must} be calculated. After a few transformations we get the following expressions.

To derive the stochastic gradient method according to the invention, the summation limits must now be replaced by limits dependent on the coefficient index. The following substitutions are required for this. L l = L 2 - D 2 + k L u = L 2 + D 2 - k L l = L 1 + D 1 - k L u = L 1 - D 1 + k

The derivatives can now be expressed using the modified summation limits.

In the transition from the normal gradient to the stochastic gradient, expected values are replaced by instantaneous values. This is carried out in the method according to the invention for the cross-correlation terms of the output signals s ₁ and s ₂ . The latest available instantaneous values are used according to the following relationship.

∂J ∂w 1k (n) = -2·[v 1(n)·s 2(n-k)+b 1(n-k)·s * 1(n)] ∂J ∂w 2k (n) = -2·[v 2(n)·s 1(n-k)+b 2(n-k)·s * 2(n)] Inserting the instantaneous values simplifies the calculation of the derivatives and we get the following relationships. The intermediate sizes v ₁ , b ₁ , v ₂ and b ₂ enable a simplified spelling and also a simplified calculation, since only one new value of each size has to be calculated at each discrete point in time. With this novel procedure, we achieve a considerable reduction in computing effort in the method according to the invention.

∂ J ∂ w 1 k (n) = -2 · [ v 1 ( n ) · s 2 ( n - k ) + b 1 ( nk ) · s * 1 ( n )] ∂ J ∂ w 2 k (n) = -2 · [ v 2 ( n ) · s 1 ( n - k ) + b 2 ( nk ) · s * 2 ( n )]

The filter coefficients w ₁ and w ₂ are now updated in the direction of the negative gradient. Μ is the step size. A relationship similar to the known LMS algorithm (Least Mean Square) is obtained. The two terms per coefficient are only necessary because we have used the latest estimates for the instantaneous value. This is useful if we want to react quickly. w 1 k ( n +1) = w 1 k ( n ) + µ · [ v 1 ( n ) · s 2 ( n - k ) + b 1 ( n - k ) · s * 1 ( n )] w 2 k ( n +1) = w 2 k ( n ) + µ · [ v 2 ( n ) · s 1 ( n - k ) + b 2 ( n - k ) · s * 2 ( n )]

In order to obtain a uniform behavior with varying signal powers, we formulate a standardized version for updating the filter coefficients w ₁ and w ₂ . The standardization quantity must be proportional to the square of a power quantity p ₁ or p ₂ . Β is the rate of adaptation. w 1 k ( n +1) = w 1 k ( n ) + β [ p 1 ( n )] 2 · [ v 1 ( n ) · s 2 ( n - k ) + b 1 ( n - k ) S * 1 ( n )] w 2 k ( n +1) = w 2 k ( n ) + β [ p 2 ( n )] 2 · [ v 2 ( n ) · s 1 ( n - k ) + b 2 ( n - k ) · s * 2 ( n )]

The system described so far for adaptive noise suppression using the method of blind signal separation is not yet sufficient due to the not negligible autocorrelation function of real acoustic signals in order to achieve low-distortion processing in a realistic environment with satisfactory convergence behavior. The system can be improved if the update of the filter coefficients w ₁ and w _{2 is} not based directly on the input signals y ₁ and y ₂ and the output signals s ₁ and s ₂ , but on transformed signals.

The system according to the invention according to FIG. 2 uses four cross-link filters 19, 20, 21 and 22 for the signal-dependent transformation of the input and output signals. The cross-link filter structures known from voice signal processing have proven to be particularly suitable for fast signal-dependent transformation. They are used there for linear prediction.

Two cross-link decorrelators 31 and 32 and a smoothing unit 33 are provided for determining the coefficients k of the cross-link filter. The cross-link decorrelators each determine a coefficient vector k ₁ and k ₂ based on the input signals y ₁ and y ₂ . In the smoothing unit, the two coefficient vectors are averaged and smoothed over time and passed on to the cross-section filter as coefficient vector k.

In contrast to the known system from FIG. 1, in the system according to the invention all calculations for updating the coefficients are based on the transformed input and output signals y _1M , y _2M , s _1M and s _2M . Two cross correlators 23 and 24 calculate the required cross correlation vectors r ₁ and r ₂ . The precalculation units 25, 26, 27 and 28 determine the intermediate variables v ₁ , v ₂ , b ₁ and b ₂ . The updating units 29 and 30 determine the modified filter coefficients w ₁ and w ₂ and make them available to the filters 17 and 18.

A standardization variable p, which is common for the update of the filter coefficients w ₁ and w _{2, is} calculated in the standardization unit 34. The optimal choice of the standardization variable p together with the correct setting of the compensation filters 5 and 6 ensure a clean and unambiguous convergence behavior of the method according to the invention.

A specific embodiment of the present invention is described in more detail below on the basis of FIG. 2. The microphones 1 and 2, the AD converters 3 and 4, the DA converters 13 and 14 and the listeners 15 and 16 are assumed to be ideal in the consideration. The characteristics of the real acoustic and electrical transducers can be taken into account in the compensation filters 5 and 6 or in the processing units 11 and 12 and, if need be, compensated for. The following relationships apply to AD converters 3 and 4 and DA converters 13 and 14. T and f _s denote the sampling period or sampling frequency and the index n the discrete point in time. d 1 ( n · T ) ⇒ d 1 ( n ) u 1 ( n ) ⇒ u 1 ( n · T ) d 2 ( n · T ) ⇒ d 2 ( n ) u 2 ( n ) ⇒ u 2 ( n · T ) T = 1 / f s f s = 16 kHz

K = 2 The compensation filters 5 and 6 are constructed according to FIG. 3 and the following relationships apply. The structure corresponds to a general recursive filter of order K. The coefficients b _1k , a _1k , b _2k and a _2k are set in such a way that the average frequency response of one input adjusts to the other input. It is preferably averaged over all possible locations of acoustic signal sources or over all possible directions of incidence.

K = 2

The delay elements 7 and 8 are constructed according to FIG. 4 and the following relationships apply. The necessary delay times D ₁ and D ₂ depend primarily on the distance between the two microphones and the preferred direction of sound incidence. Small delay times are desirable because it also reduces the overall system delay time. e.g. 1 ( n ) = y 1 ( n - D 1 ) e.g. 2 ( n ) = y 2 ( n - D 2 ) D 1 = D 2 = 1

The following relationships apply to the subtractors 9 and 10. s 1 ( n ) = e.g. 1 ( n ) - e 1 ( n ) s 2 ( n ) = e.g. 2 ( n ) - e 2 ( n )

The following relationships apply to the processing units 11 and 12. The functions f ₁ () and f ₂ () stand for any linear or non-linear functions of their arguments. They result from the usual processing of hearing aids. u 1 ( n ) = f 1 ( s 1 ( n ), s 1 ( n -1), s 1 ( n -2), ...) u 2 ( n ) = f 2 ( s 2 ( n ), s 2 ( n -1), s 2 ( n -2), ...)

N 1 = N 2 = 63 The filters 17 and 18 are constructed according to FIG. 5 and the following relationships apply. The filter orders N ₁ and N ₂ result from a compromise between the achievable effect and the computational effort.

N 1 = N 2 = 63

y 20(n) = y 2(n) x 20(n) = y 2(n)

s 10 (n) = s 1(n) x 30(n) = s 1(n)

s 20(n) = s 2(n) x 40(n) = s 2(n)

M = 2 The cross-link filters 19, 20, 21 and 22 are constructed in accordance with FIG. 6 and the following relationships apply. The filter order M can be chosen to be quite small. y 10 ( n ) = y 1 ( n ) x 10 ( n ) = y 1 ( n )

y 20th ( n ) = y 2 ( n ) x 20th ( n ) = y 2 ( n )

s 10 ( n ) = s 1 ( n ) x 30th ( n ) = s 1 ( n )

s 20th ( n ) = s 2 ( n ) x 40 ( n ) = s 2 ( n )

M = 2

g = 63/64 h = 1-g = 1/64 L 1 = L 2 = 31 The cross correlators 23 and 24 are constructed according to FIG. 7 and the following relationships apply. The constants g and h, which determine the time behavior of the averaged cross correlations, should be adapted to the filter orders N ₁ and N ₂ . The constants L ₁ and L ₂ determine how many cross-correlation terms are taken into account in the subsequent calculations.

G = 63/64 H = 1- G = 1/64 L 1 = L 2 = 31

Type V 25 and 26 precalculation units are constructed according to FIG. 8 and the following relationships apply. The standardization was chosen so that the intermediate variables v ₁ and v _{2 are} dimensionless.

Type B 27 and 28 precalculation units are constructed according to FIG. 9 and the following relationships apply. The standardization was chosen so that the intermediate sizes b ₁ and b _{2 are} dimensionless.

The update units 29 and 30 are constructed in accordance with FIG. 10 and the following relationships apply. The adaptation speed β can be chosen according to the desired convergence behavior. w 1 k ( n +1) = w 1 k ( n ) + β p ( n ) · [ v 1 ( n ) · s 2 M ( n - k ) + b 1 ( n - k ) · s 1 M ( n )] (0≤ k ≤ N 1 ) w 2 k ( n +1) = w 2 k ( n ) + β p ( n ) · [ v 2 ( n ) · s 1 M ( n - k ) + b 2 ( n - k ) · s 2 M ( n )] (0≤ k ≤ N 2 )

f 20(n) = y 2(n) b 20(n) = y 2(n)

The cross-link decorrelators 31 and 32 are constructed according to FIG. 11 and the following relationships apply. The cross-link decorrelators calculate the coefficient vectors k ₁ and k ₂ required for decorrelation of their input signals. f 10 ( n ) = y 1 ( n ) b 10 ( n ) = y 1 ( n )

f 20th ( n ) = y 2 ( n ) b 20th ( n ) = y 2 ( n )

f = 1.0 l = 0.25 The smoothing unit 33 is constructed in accordance with FIG. 12 and the following relationships apply. The constants f and I are chosen so that the averaged coefficients k get the desired smoothed course.

f = 1.0 l = 0.25

The standardization unit 34 is constructed according to FIG. 13 and the following relationships apply. First the four powers of y _1M , y _2M , s _1M and s _{2M are} calculated and from this the standardization variable p is determined. i 1 ( n ) = G · i 1 ( n- 1) + H· [ y 1 M ( n )] 2 O 1 ( n ) = G · O 1 ( n -1) + H · [ s 1 M ( n )] 2 i 2 ( n ) = G · i 2 ( n -1) + H · [ y 2 M ( n )] 2 O 2 ( n ) = g · o 2 ( n -1) + H · [ s 2 M ( n )] 2

The preferred embodiment can easily be on a commercial Signal processor programmed or implemented in an integrated circuit become. To do this, all variables must be appropriately quantized and the operations based on the existing architectural blocks are optimized. A special Attention is paid to the treatment of square sizes (services) and the division operations. Depending on the target system, there are optimized ones Procedures. But in and of themselves these are not the subject of present invention.

Claims (11)

Circuit for calculating two decorrelated digital output signals (s ₁ , s ₂ ) from two correlated digital input signals (y ₁ , y ₂ ), comprising two filters (17, 18) arranged symmetrically crosswise in the forward direction with adaptive filter coefficients ( w ₁ , w ₂ ), two delay elements (7, 8) and two subtractors (9, 10) for calculating the output signals (s ₁ , s ₂ ) in the time domain from the input signals (y ₁ , y ₂ ) while minimizing a quadratic cost function consisting of cross-correlation terms, characterized in that the circuit contains four cross-link filters (19-22) for signal-dependent transformation of the input and output signals (y ₁ , y ₂ ; s ₁ , s ₂ ) and that all computing units for updating the filter coefficients ( w ₁ , w ₂ ) downstream of the cross-section filters (19-22). Circuit according to claim 1, characterized by two cross-correlators (23, 24), four pre-calculation units (25-28) and two update units (29, 30) for quickly reacting and computationally efficient update of the filter coefficients ( w ₁ , w ₂ ). Circuit according to Claim 1 or 2, characterized by two cross-member decorrelators (31, 32) which follow the statistics of the two input signals (y ₁ , y ₂ ) and a smoothing unit (33) for calculating averaged and smoothed coefficients ( k ) for the cross-link filter (19-22). Circuit according to one of Claims 1-3, characterized by a standardization unit (34) which calculates an optimal standardization variable (p) for updating the filter coefficients ( w ₁ , w ₂ ). Device for adaptive noise suppression in acoustic input signals, comprising two microphones (1, 2) and two AD converters (3, 4) for converting the acoustic input signals into two digital input signals (y ₁ , y ₂ ), a circuit for processing the digital input signals (y ₁ , y ₂ ) into digital output signals (s ₁ , s ₂ ), at least one DA converter (13, 14) and at least one loudspeaker or receiver (15, 16) for converting the digital output signals (s ₁ , s ₂ ) in acoustic output signals, characterized in that the circuit for processing the digital input signals (y ₁ , y ₂ ) into digital output signals (s ₁ , s ₂ ) is a circuit according to one of claims 1-4. Apparatus according to claim 5, characterized by at least one compensation filter (5, 6) for matching the average frequency response of one microphone (1) to the average frequency response of the other microphone (2). Method for calculating two decorrelated digital output signals (s ₁ , s ₂ ) from two correlated digital input signals (y ₁ , y ₂ ), executable by means of a circuit according to one of claims 1-4, using two filters arranged symmetrically crosswise in the forward direction (17, 18) with adaptive filter coefficients ( w ₁ , w ₂ ), two delay elements (7, 8) and two subtractors (9, 10) the decorrelated output signals (s ₁ , s ₂ ) from the input signals (y ₁ , y ₂ ) while minimizing a quadratic cost function consisting of cross correlation terms in the time domain, characterized in that a signal-dependent transformation of the input and output signals (y ₁ , y ₂ ; s ₁ , s ₂ ) is carried out by means of four cross-link filters (19-22) and only the transformed signals (y _1M , y _2M ; s _1M , s _2M ) are used to update the filter coefficients ( w ₁ , w ₂ ). Method according to claim 7, characterized in that two cross-member decorrelators (31, 32) follow the statistics of the two input signals (y ₁ , y ₂ ) and a smoothing unit (33) the averaged and smoothed coefficients ( k ) for the cross-member filters (19-22) calculated. Method according to claim 7 or 8, characterized in that in a standardization unit (34) an optimal standardization quantity (p) is calculated for updating the filter coefficients ( w ₁ , w ₂ ). Method for adaptive noise suppression in acoustic input signals, the acoustic input signals being converted into digital input signals (y ₁ , y ₂ ), the digital input signals (y ₁ , y ₂ ) being processed into digital output signals (s ₁ , s ₂ ) and the digital ones Output signals (s ₁ , s ₂ ) are converted into acoustic output signals, characterized in that a method according to one of claims 7-9 is used for processing the digital input signals (y ₁ , y ₂ ) into digital output signals (s ₁ , s ₂ ) becomes. A method according to claim 10, characterized in that two microphones (1, 2) are used for converting the acoustic input signals and the average frequency response of a microphone (1) by means of at least one compensation filter (5, 6) the average frequency response of the other microphone (2) is adjusted.

Images