Computer-aided method for stable processing of an audio signal using an adapted LMS algorithm

Description

[0001] The invention relates to a computer-aided method for processing an audio signal using an adapted LMS algorithm according to the preamble of claim 1.

[0002] Least mean squares (LMS) algorithms, the collective term for algorithms that minimize the squared error, have long been used in electroacoustics, particularly for adaptive filters, in numerous variations and for a wide range of applications. The most commonly used variants are known by the following abbreviations: LMS, NLMS, sign-LMS, variable step LMS (VSS-LMS), correlation factor LMS, and, for specific applications, variants such as filtered-x LMS or filtered-e LMS (see International Journal of Electrical and Computer Engineering Vol. 7, No. 5, October 2017, LMS Adaptive Filters for Noise Cancellation: A Review, pp. 2520ff).

[0003] The application of LMS algorithms, for example in feedback suppression, is known from the prior art. In particular, sign-based LMS can be used in this context.

[0004] US patent 2008 / 0063230 demonstrates the use of a classic LMS algorithm in conjunction with an Adaptive Digital Filter (ADF). The ADF is controlled by the LMS algorithm.

[0005] Another example of its use is US 9,807,503, which employs an equalizer to optimize the playback of a digital acoustic source (for example, Bluetooth audio). The equalizer is driven by an adaptive filter, whose filter coefficients are obtained, for example, by applying LMS and then smoothed using linear interpolation and / or the exponential method.

[0006] WO0019605 shows a standard implementation of an adaptive feedback canceller using an NLMS algorithm.

[0007] US 2017 / 0201276 shows an LMS (Standard System Estimation) application in which the input signals of the LMS algorithm are delayed differently.

[0008] Furthermore, the NPL document "Design the adaptive noise canceller based on an improved LMS algorithm and realize it by DSP" by Xu Yanhong and Zhang Ze shows the possibility of using an improved sign-LMS as an adaptive system estimator.

[0009] A disadvantage of all solutions is the potential instability of the system. Exponential smoothing uses the same coefficient for both the attack and release cases. Due to this property, under certain unfavorable conditions, the filter coefficients of adaptive filters can converge to infinitely high values. Such an overshoot of the filter coefficients is prevented by the word width of the processor used. This property is called clipping. A typical example of such clipping is acoustic feedback suppression: If feedback occurs (the microphone picks up the loudspeaker), it is perceived as a very loud sine wave ("howling"). In this case, an adaptive filter is configured to filter the loudspeaker output and subtract it from the microphone output. Ideally, it filters exactly the feedback sine wave and subtracts it from the microphone output, thus canceling out the sine wave.However, if the system is confronted with a sinusoidal tone from an external source (e.g., music), the adaptive filter attempts to cancel it out. Since the tone doesn't weaken despite the subtraction (because the source is no longer the loudspeaker but external), the filter coefficients swell towards infinity. The result is a distorted sound. To avoid this erroneous convergence, it is known in the prior art to monitor the coefficients using an external algorithm and limit them at a specific threshold. While this yields usable results, it increases complexity and computational effort.

[0010] A problem arises, for example, with in-ear hearing systems such as headphones or hearing aids when the device is removed from the ear. Due to the loss of passive noise reduction, a method using a classic LMS algorithm attempts to increase the amplification of the ANC filters indefinitely, since no ANC power is present. Therefore, additional sensors must be used to detect whether the device is in the ear or not.

[0011] There is therefore a need for an LMS algorithm that requires less complexity of the acoustic system and less computing power, and can be used not only in ANC systems, but anywhere adaptive filters are used.

[0012] According to the invention, these problems are solved by using a novel PEAK-LMS algorithm that has the features specified in the characterizing part of claim 1; in other words, in which a signum function is calculated from an input signal sequence and a filter output sequence, which, together with a sequence of initial filter values ​​and a sequence of time constants for the attack and release cases (which need not be the same), is used to calculate a sequence of adapted filter values. In one embodiment, upper and lower limits for the coefficients are set within the method / algorithm itself, so that the risk of upward convergence is intrinsically excluded, thereby also eliminating the need for costly external monitoring.

[0013] The invention is illustrated below by examples. These examples show the Fig. 1the schematic structure of a feedback suppression system for an acoustic system, which Fig. 2 stable values ​​for the filter coefficients, which are determined according to the in Fig. 1 described methods were obtained using the PEAK-LMS algorithm Fig. 3 Filter coefficients of an LMS algorithm for a system with sinusoidal external tones, which Fig. 4 Filter coefficients of a PEAK-LMS algorithm for a system with sinusoidal external tones, which Fig. 5 the application of the invention to the control of an amplifier, the Fig. 6 a block diagram of an example where an LMS algorithm is used in its known form, which Fig. 7 the unit step response of the filter for the attack case, which Fig. 8 the unit step response of the filter for the attack and release cases, which Figures 9a and 9bin two further variants the adaptation rate and the residual error of the PEAK filter, as used by the method according to the invention, which Fig. 10 the convergence of a filter coefficient of an advantageous embodiment of the invention, in which this method is also applied to the adaptation rate, which Fig. 11 For example, the application of feedback suppression, which Fig. 12 For example, consider the use case of echo cancellation, which Fig. 13 For example, the use case of adaptive ANC, which Fig. 14 For example, one variant of the use case of adaptive beamforming, which Fig. 15 For example, one variant of the use case of an adaptive equalizer and the Fig. 16 Here is an example of a second use case for an adaptive equalizer.

[0014] Fig. 1Figure 1 shows the schematic setup of a feedback suppression system for an acoustic system, as known from the prior art, but which can also be used for the method according to the invention. The microphone 1 records the output of the driver (loudspeaker 2), thereby creating so-called acoustic feedback 5 in the form of the feedback path. h ( n ) is generated. In the unfiltered case, this can be audible as a very loud sine wave. The function block for adapting filter 4 (adaptation block for short) calculates this from the reference signal. d ( n ) (Playback) and the error signal e ( n ) (Result of the subtraction) Filter coefficients for an adaptive filter 3, which are then used as filter output y( n) are output. The adaptation block 4 thus attempts to estimate the transfer function of the feedback path 5 and apply it using adaptive filter 3. In the example shown, the filtered output of the driver is subtracted at the microphone input of the circuit (downstream of the physical microphone). If the transfer function is estimated correctly, the filter "cuts out" the sine wave and subtracts it from itself at the input, thereby suppressing the feedback. Fig. 1 The following algorithms are listed in adaptation block 4: LMS, SignLMS, and PEAK-LMS. These represent variants of the adaptation algorithm. These three variants are compared below under identical conditions: In a simulation, white noise (duration 1 second, sampling rate 48 kHz) is fed to the microphone. The same learning rate is used for all three algorithms. For PEAK-LMS, the individual filter coefficients are defined by the upper limit. λ pos= 1 and the lower limit λ neg The attack time is limited to -1. For this application, the attack time is equal to the release time. The time constants for the attack and release times must be chosen in a manner known to an expert from the prior art, depending on the sampling rate. For the sampling rate of 48 kHz specified here, for example, it would be 300 ms.

[0015] The learning rate is a factor by which adaptation occurs. LMS is a gradient desaturation method where the gradient is based on the square of the error. The learning rate thus defines a step size by which the coefficients of the LMS algorithm are adjusted. For a learning rate of 0.00005, the LMS algorithm does not produce a valid result for the filter coefficients because the adaptation is too slow for a learning rate of 0.00005. A Sign-LMS algorithm exhibits similar behavior and becomes unstable for a learning rate that is too high, while it adapts too slowly for a learning rate that is too low. In this example, a learning rate of 0.00001 is close to the optimum for an LMS or Sign-LMS algorithm. The PEAK-LMS algorithm shows a significantly higher tolerance: For 0.00005, the suppression performance is reduced, but the algorithm maintains stable values ​​for the coefficients.

[0016] Fig. 2 displays values ​​for coefficients for a filter function c (n ), which after the in Fig. 1 The described procedure using the PEAK-LMS algorithm yielded results for an FIR filter at a learning rate of 0.00005. It is clearly evident that the coefficient values ​​remain stable. The abscissa represents the number of the filter coefficient, and the ordinate represents the coefficient value.

[0017] Fig. 3This figure shows an example of the filter coefficients of a state-of-the-art LMS algorithm for a system with sinusoidal external tones. The abscissa represents the number of the filter coefficient, and the ordinate represents its value. A typical problem for feedback suppression is the case of an "external" sinusoidal tone, i.e., one not introduced into the system by feedback. In this case, the adaptive filter detects the sinusoidal tones and attempts to suppress them, as they are mistakenly interpreted as feedback. For a tone whose source is not feedback, but rather, for example, music, the problem arises with the adaptive filter that the subtraction at the microphone does not reduce the perceived feedback, and the filter is misadjusted (this effect is called "entrainment").For an LMS algorithm, this leads to the filter coefficients tending towards infinity, since the gradient – ​​due to the constant error – always remains the same. The sharp increase in the filter coefficients is particularly noticeable for the first two coefficients. The range of -4000 to +2000, in which the coefficients lie, is partially beyond the capabilities of processors, and the filtering result is unusable because digital-to-analog converters (DACs) in acoustics are limited to a range of ±1. In the example above, the filter coefficients reach ±∞ after only a short time, and the simulation terminates because the coefficients can no longer be calculated.

[0018] Fig. 4 shows, as an example, the filter coefficients for a filter function. c ( n), determined using the PEAK-LMS algorithm according to the invention, for a system with sinusoidal external tones, identical to the one in Fig. 3 The abscissa shows the number of the filter coefficient and the ordinate the value of the coefficient. By limiting the filter coefficients to λ pos = 1 and λ neg By setting the instabilities to -1, the instabilities that occur in a classical LMS algorithm can be avoided (see Fig. 3In addition to the advantage of the filter's general stability, the value range of ±1 can be easily represented, for example, in typical fixed-point processors for audio (e.g., 24-bit). While the invention is not limited to fixed-point processors, these are widely used in audio applications. In the illustrated example, PEAK-LMS prevents the coefficients from increasing too much. In extreme cases, the PEAK-LMS algorithm would limit the filter coefficients to ±1, but in the illustrated case, it generally prevents incorrect settings. It is also immediately apparent that a method according to the invention based on a PEAK-LMS algorithm can operate more efficiently than one that uses, for example, a classic NLMS algorithm, resulting in savings in the required computing power, which can either be used for other processes or contribute to extending battery life.

[0019] Fig. 5This example illustrates the control of an amplifier based on estimating the gain of a transfer function. In the case of an in-ear ANC (Active Noise Cancelling) hearing system, the gain of a transfer function must be controlled in real time. In this context, real time means with the lowest possible latency and time invariance; the processing cannot occur independently of the signal's timing. Processing a previously recorded audio file would therefore not be real time, as the calculation is theoretically not subject to any time constraints. For ANC, however, incoming microphone data must be processed as quickly as possible. Depending on the wearing situation (insertion depth; individual ear canal shape), the gain of the primary path (external microphone) and / or the gain of the secondary path (internal microphone) must be adapted to optimize the performance of the ANC system.

[0020] In this example as well, conventional LMS variants will always attempt to converge towards a theoretical optimum. A problem arises, for instance, when the device is removed from the ear: there is no longer any passive attenuation, making the system acoustically open. In this case, the LMS algorithm would attempt to increase the gain of the ANC filters to infinity, since no ANC performance is present. Therefore, additional sensors must be used to detect whether the device is in the ear or not. The PEAK LMS algorithm's automatic limitation at λ pos = 1 and λ neg Setting it to -1 prevents clipping and the resulting artifacts at the driver (speaker).

[0021] The inventive method, carried out using the PEAK-LMS algorithm, can be used not only in ANC systems, but wherever adaptive filters are used, for example in echo suppression, feedback suppression, system estimation (detection of an unknown transfer function), channel equalization (also applies to RF technology), adaptive inverse control, hum suppression (filtering 50Hz hum from the mains for e.g. ECG sensors), vector voltmeters, separation of signals with different correlations, interference suppression in audiological measurement systems (e.g. measurement of otoacoustic emissions) and many others.

[0022] The exact design of the PEAK-LMS algorithm for use in the method according to the invention is explained below.

[0023] Fig. 6 shows a block diagram of an example where an LMS algorithm is used in its known form. x ( n) describes the input signal, h ( n ) describes the transfer function to be estimated (transfer function of the acoustic system), y ( n ) describes the output of the filter, c ( n ) describes the coefficients of the adaptive filter, d ( n ) describes the reference signal for the LMS algorithm and e ( n ) the error signal that results from the difference between the obtained value y and the given value x, in each nth step. Furthermore, it shows Fig. 6 the LMS block, which makes the necessary adaptations to the filter c ( n ) certainly.

[0024] The LMS algorithm in its known form is (Formulas I): e n = x n − y n c n + 1 = c n + μ ∗ e n ∗ x n

[0025] As is customary in this context, the vectors are printed in bold, the scalars in normal font. x ( n ) is, for example, the element of the vector x (n ) at time n.

[0026] Specifically, c ( n + 1) for the result of the nth adaptation of c , c ( n ) for the value obtained in the n-1-th step of c ; µ (also mu) for the adaptation rate; e ( n ) for the error signal in step n, x ( n ) for the input signal sequence that represents the input data of the audio signal up to time n and y ( n ) for the output of the filter in the nth step.

[0027] The in Fig. 6 The basic form shown, which is from the prior art, is, as already explained, sensitive to level fluctuations of the input signal. x ( n) in that it negatively affects the convergence of the algorithm. Therefore, as already mentioned, various adaptations of the LMS method have been proposed, which, under the names NLMS and sign-LMS, are part of the state of the art and reduce this sensitivity. However, NLMS is unsuitable for fixed-point processors because division is necessary for normalization. Divisions increase computational overhead and are subject to quantization effects on fixed-point systems. Sign-LMS, on the other hand, does not require division, but this comes at the cost of reduced convergence speed and the occurrence of a larger residual error of the adaptation, which can lead to poorer filtering properties.

[0028] For a sign-LMS algorithm, the formulas derived from formulas I are: e n = x n − y n c n + 1 = c n + μ ∗ sgn e n ∗ x n

[0029] The meaning of the vectors and scalars is analogous to the case of the classical LMS algorithm (see above). As can be seen directly, a small step corresponds to a coefficient ( µ * sgn ( e ( n ) * x ( n )) ) added or subtracted. In case of an incorrect setting (as explained above using feedback suppression, filter coefficients can converge to infinity under unfavorable conditions), this change can theoretically continue to infinity, which causes problems in a finite system.

[0030] This problem can be prevented by the so-called PEAK algorithm, which is shown in the following formulas III: C α β = x n > y n − 1 : α x n ≤ y n − 1 : β y n = y n − 1 + C α β ∗ x n − y n − 1

[0031] The time constants α and β These represent the attack and release cases and thus determine the rise and fall times of the algorithm. α and βare purely numerical; their temporal meaning depends on the chosen sampling rate. Both move within the interval [0;1]. .

[0032] A filter that uses a sign-LMS algorithm with a PEAK filter is a single-pole, recursive filter per coefficient, possessing an attack and a release time constant. If the input signal is larger than the filter's output, the attack time constant is applied; otherwise, the release time constant is used (see Formulas III).

[0033] Fig. 7The figure shows the unit step response for a single coefficient of the filter, which is approximated according to the inventive method (the value towards which the filter tends for an input signal that jumps from 0 to 1 and then remains constant). The influence of the attack time constant is also evident here. The abscissa shows the approximation steps relative to its discrete time axis, and the ordinate shows the value of the filter coefficient.

[0034] By combining the sign-LMS with the PEAK filter, the calculation of each individual coefficient is no longer performed by adding a step size, but rather via the step response of the PEAK filter. This results in the following formulas IV: w n = sgn e n ∗ x n c n + 1 = c n + C α β ∗ w n − c n

[0035] This is w ( n ) a signum function that represents the summary of the signum term of formulas II. The inputs of the algorithm e ( n ) and x ( nThe signals could be filtered before being fed into the algorithm. This has the advantage that the adaptation can be focused on a specific frequency range, usually with the help of a suitable pre-filter (typically a high-pass, low-pass, or band-pass filter, or a combination thereof). The desired frequency range is application-specific and can be individually defined by a specialist according to the requirements. Due to the exponential function of the PEAK filter, the LMS method can no longer converge to infinity, even if, for example, the step size (adaptation rate) for a coefficient is kept constant due to an incorrect setting (see the feedback cancellation example above). µ is added.

[0036] Fig. 8Figure 1 shows the unit step response of a filter for a single coefficient, approximated according to the inventive method using the PEAK-LMS algorithm, for both the attack and release cases. The abscissa represents the approximation steps relative to its discrete time axis, and the ordinate represents the value of the filter coefficient. The attack and release time constants in Formula IV have been equated. Therefore, in this case, the following applies: C ( α, β ) = α = β. This means that the algorithm (for example, rectangular pulse) converges at the same speed in both directions.

[0037] However, it is also possible to use different values ​​for the attack and release time constants in the PEAK-LMS algorithm ( α ≠ β ), which allows different convergence times to be obtained for positive and negative coefficients. An adaptation of formulas IV leads to formulas V: w n = sgn e n ∗ x n C α β = w n > c n : α w n ≤ c n : β c n + 1 = c n + C α β ∗ w n − c n

[0038] These and other modifications open up possibilities and areas of application that are not accessible with the various LMS variants of the state of the art.

[0039] One example of this is the introduction of the filter coefficient. λ The signum operator returns 1, 0, or -1; by introducing the factor λ The limit of the coefficient can be controlled, as shown by Formulas VI, which are an extension of Formulas V: w n = λ ∗ sgn e n ∗ x n C α β = w n > c n : α w n ≤ c n : β c n + 1 = c n + C α β ∗ w n − c n

[0040] The factor λ The interval is basically freely definable and application-specific. However, for audio applications, the interval [0,1] is useful, while for a filter, [-1,1] might also be suitable.

[0041] This also makes it possible to set different limits for the coefficients in the negative and positive ranges, and even individual limits for each coefficient. An example of this is Formulas VII, which are an extension of Formulas V: w n = sgn e n ∗ x n λ = w n > 0 : λ + w n = 0 : 0 oder w n w n < 0 : λ − C α β = w n > c n : α w n ≤ c n : β c n + 1 = c n + C α β ∗ λ − c n λ ± have fixed, predefined values ​​that can be adapted by the specialist to the desired use case.

[0042] By using different attack and release time constants, the adaptation of the filter can be handled extremely flexibly, while retaining the advantages of sign-LMS over other variants, such as NLMS, since no divisions are required during the calculation process, which is advantageous when implementing it on a fixed-point DSP.

[0043] An exemplary implementation of the above principles is adaptive gain control. Typically, the gain should adjust adaptively within the range of ±6 dBFS. By setting an upper limit for λ above = 2 and a lower limit of λ below At a value of 0.5, the PEAK-LMS algorithm automatically adjusts for ± 6 dBFS without requiring additional calculations, such as converting linear values ​​to decibels and vice versa. The elegance of the solution is therefore obvious.

[0044] The ±6 dBFS interval is used here specifically for adaptive gain control in ANC headphones. Other applications include feedback suppression, ANC, echo cancellation, adaptive beamforming, adaptive gain control, and similar applications. The gain control range depends on the specific product and the calculated ANC filters. However, a practical interval range is generally ±10 dBFS. Depending on the wearing situation of, for example, in-ear ANC headphones and the wearer's ear geometry, the volume of the pressure chamber created by the ear canal and the in-ear headphone varies. This, in turn, must be taken into account and compensated for by an adaptive ANC system.

[0045] Fig. 9The graph shows the adaptation rate and residual error of the PEAK filter as used by the method according to the invention. The abscissa represents the discrete time axis and the ordinate the value of the error magnitude. As briefly mentioned earlier, sign-LMS leaves a residual convergence error, the magnitude of which depends on the step size. µ depends: A larger one µ This leads to rapid convergence, but also to a larger residual error (the algorithm frequently "oscillates" around an optimum without reaching it); a smaller µ On the other hand, this leads to slower convergence with a smaller residual error. The variant used in the invention shows similar behavior. Specifically, the variant in Fig. 9a a PEAK filter with a time constant C ( α, β ) of 0.001 and the variant in Fig. 9b a PEAK filter with a time constant C ( α, β ) of 0.0025.

[0046] It is generally known in LMS algorithms that the result can be improved by adaptively selecting the step size depending on the error magnitude. This is generally known as variable step-size LMS. By adapting formulas I, formulas VIII can be specified: μ n + 1 = a ∗ μ n + b ∗ e 2 n e n = x n − y n c n + 1 = c n + μ n ∗ e n ∗ x n

[0047] This can be used for µ An upper and lower limit are introduced again. The step size is chosen to be large if the error is large, and small if the error is small, thus achieving a compromise between convergence rate and residual error. a and b are coefficients that ensure an adaptive step size and a weighting between µ ( n ) and e 2< ( n ) allow. For example, the ratio a = 1 - b chosen. Typical values ​​here would be 1 > a ≥ 0.8.

[0048] Fig. 10 Figure 1 shows the adaptation rate and residual error of the PEAK filter as used in an advantageous embodiment of the method according to the invention. The abscissa represents the discrete time axis and the ordinate the value of the error magnitude. For the case of equal attack and release time constants, formulas IX yield: w n = λ ∗ sgn e n ∗ x n c n + 1 = c n + C α , β n ∗ w n − c n C α , β n + 1 = a ∗ C α , β n + b ∗ e 2 n

[0049] Formula IX is a simplified version of Formula VII for the case C α,β ( n ) = α = β In other words, this is a case where α and β are the same, but can be controlled over time. For the sake of simplicity, the formulas were therefore set in a time-based context. A significant advantage of this variant lies in the lower requirement for computing power. The improved convergence of Fig. 10 opposite Fig. 9a and bThis is clearly evident from the exponential convergence behavior. If, in the embodiment of the invention presented, different time constants are additionally used for this method, formula X is obtained as an extension of formula IX: w n = λ ∗ sgn e n ∗ x n c n + 1 = c n + C n ∗ w n − c n C n + 1 = a ∗ C n + b ∗ e 2 n

[0050] In this case, there are no predefined attack / release times, as the formula in the bottom row recalculates the time constants in each iteration.

[0051] The invention is not limited to the example and its variants, but can be used wherever LMS algorithms are used and where external conditions pose a risk of instability or a rapid response is desired.

[0052] The terms "step response" and "jump response" are used synonymously in this application.

[0053] The algorithm was always defined in the time domain in the formulas given. An expert can easily define it in the frequency domain using appropriate methods, such as the Fourier transform, with the same result.

[0054] In summary, this is a computer-aided method for the stable processing of an input audio signal from an acoustic system, comprising at least one audio signal input (e.g., a microphone, an external music playback device, input of a digital interface (USB, Bluetooth, etc.), reading a stored audio file, or any other audio signal source), at least one audio signal output (e.g., a loudspeaker, a recording device, output of a digital interface (USB, Bluetooth, etc.), writing a stored audio file, or any other audio signal sink), a data connection between the audio signal input and the audio signal output influenced by an adaptive filter (3), and an LMS algorithm (4) controlling the adaptive filter (3), wherein a) the LMS algorithm (4) based on a reference signal d ( n ) and an error signal e ( n) a transfer function h ( n ) (this is the target value of the LMS algorithm, which can be approximated as closely as possible but never fully achieved), b) the estimation of the transfer function is applied as an adaptive filter (3) and is used there as an initial filter function c ( n ) is available (the estimate is therefore passed to the filter block and is located there as a mathematical filter function) c ( n ) in the form of filter coefficients), c) the output of a filter output sequence as y ( n ) (This is a sequence of samples that are as close as possible to the output sequence of h ( n ) should be located) is carried out, d) a comparison of the output of the transfer function h ( n ) and the filter output sequence y ( n ) the initial filter function c ( n) occurs, whereby the error signal is derived from the difference e ( n ) results in, e) characterized by the fact that the LMS algorithm used to adapt the adaptive filter (3) is a PEAK-LMS algorithm, where f) is a signum function for the algorithm w ( n ) is used, which is via i. an input signal sequence representing the input data of the input audio signal x ( n ) (This is the actually desired audio input signal, h ( n ) in contrast, an unwanted environmental influence such as feedback, echo, noise, etc., which undesirably alters x(n) in the application), ii. the filter output sequence representing the output of the filter y ( n ), iii. the difference between an input signal sequence altered by unwanted influences (e.g. environmental influences, feedback, echo, noise, etc.). x ( n ) and the filter output sequence y ( n ) through e ( n ) = x ( n ) - y ( n ) defined error signal e ( n ) (The task of the PEAK-LMS algorithm is to minimize this error signal), iv. in the form w ( n ) = sign ( e ( n ) * x ( n )) is calculated, g) which together with the initial filter function c ( n ) and time constants C ( α, β ), where α the time constants of the attack and β The time constants of the release case are described and can be used to determine the sequence of adapted filter values. c ( n + 1) using the equation c n + 1 = c n + C α β ∗ w n − c n to calculate.

[0055] The invention is therefore a computer-aided method for the stable processing of an input audio signal of an acoustic system, comprising at least one audio signal input, at least one audio signal output, a data connection between the audio signal input and the audio signal output influenced by an adaptive filter, and an LMS algorithm controlling the adaptive filter, wherein the controlling algorithm is a novel combination of a classical LMS algorithm and a PEAK filter, which ensures the stable processing of the input audio signal of the acoustic system.

[0056] To improve processing, the input signal sequence can be x ( n ) and the error signal e ( n The signal undergoes pre-filtering before being fed to the algorithm. Commonly used filters include high-pass, low-pass, or band-pass filters, as well as combinations thereof.

[0057] For better understanding, the following are application examples for acoustic systems in which the inventive method with the PEAK-LMS algorithm can be applied. For simplification, ADC (analog-to-digital converter) and DAC (digital-to-analog converter) blocks are omitted in the graphics, and acoustic paths are represented discretely. H ( z ) instead of analog H ( s ) interpreted as corresponding to the preceding figures h ( n ) . Reference signal d ( n ) , Error signal e ( n ), Filter output sequence y ( n ) and input signal sequence x ( n ) are, for the sake of simplicity, listed without the sample index ( n ). Furthermore, Ĥ ( z ) instead of the filter function c ( n) is used to clarify that the goal is either to create a complementary function or to estimate the objective function.

[0058] Fig. 11 shows the general case of feedback cancellation (applicable to all types of headphones, hearing aids, ANC systems, etc.) analogous to Fig. 1 A sound source is additively mixed with the sound from loudspeaker 2 and recorded by microphone 1 of the system. The task of the PEAK-LMS algorithm is to determine the acoustic feedback path. H ( z The system estimates the loudspeaker signal, filters it accordingly, and subtracts it from the microphone signal. The error signal e (which must be minimized) is the result of the subtraction, and the reference d is the result for the output (usually taken directly from the DAC buffer). M ( z) denotes the transfer function of the system (e.g., equalizer, multiband compression, etc.). Ideally, H ( z ) optimal in Ĥ ( z ) approximates ( Ĥ ( z ) = H ( z )), i.e., the sound from loudspeaker 2, which is mixed in before microphone 1, is subtracted after the ADC. This results in infinite cancellation of the feedback sine wave. An example of a suitable processor for this is the "Onsemi Ezairo 7100" for hearing aids. This processor has a block-floating point in an auxiliary processor.

[0059] Fig. 12 This illustrates the use case of echo cancellation. Similar to feedback suppression, a PEAK-LMS algorithm must here analyze the acoustic path of the echo ( H ( z)) estimate. This acoustic path mixes with the speaker in front of microphone 1 and must be subtracted after it. Unlike feedback suppression, the microphone here does not play to loudspeaker 2, but to a receiving device (examples: wireless microphones and monitoring boxes, headsets in mobile communications and intercom systems). The labels are analogous to Fig. 11 .

[0060] Fig. 13 This illustrates the use case of adaptive ANC. With adaptive ANC, two paths can be estimated: feedforward and feedback. Depending on the system, there may be only one of these paths or both. H(z) here denotes the passive attenuation of the system, and a PEAK-LMS algorithm must calculate this path from feedforward microphone 6 and feedback microphone 7. Ĥ ( z ) estimate. This path will be inverted and output to speaker 2. Ĥ ( z ) and H(z) cancel out. Another path is J ( z) - this refers to the acoustics inside the ear cup of the ANC headphones. A PEAK-LMS algorithm is also required here. J ( z ) in Ĵ ( z ) estimate in order to achieve annihilation. J ( z ) in Ĵ ( z ) are analogous to Ĥ ( z The paths shown are H(z) and H(z), and are named differently because they represent different paths. An example of a suitable processor is the "ADAU 1860". This is a fixed-point DSP without floating-point capability.

[0061] Fig. 14 This shows a variant of the adaptive beamforming use case. Two microphones 1 are operated here as an end-fire array, meaning one signal is delayed and subtracted from the other (the delay is determined by the block). z - n(< shown). This is done for two paths, resulting in two opposing cardioid polar patterns. A PEAK-LMS algorithm then estimates the transfer function between one cardioid and the other, filters the former, and subtracts its signal from the latter. The result is adaptive beamforming, where the zeros of the final polar pattern are controlled by the adaptive filter. The loudspeaker is not shown in this simplified circuit diagram. An example of a suitable processor is the Onsemi BelaSigna 300, a fixed-point processor with a block-floating-point unit.

[0062] Fig. 15 This shows a variant of the use case of an adaptive equalizer. In this application, the block of a PEAK-LMS algorithm is arranged so that it corresponds to noise (noise disturbance). H ( z)) removed from a signal. If a useful signal (source) is disturbed with additive noise, the PEAK-LMS algorithm controls a filter in the channel and equalizes the channel according to the difference between the source and the disturbed transmission. The loudspeaker is not shown in this simplified circuit diagram. The block z -n< represents a delay element.

[0063] Fig. 16 Channel equalization is shown as a second variant of the use case of an adaptive equalizer. In this form, a filter I ( z ) via convolution (in the time domain) with a transfer function H ( z ) linked. The reference R ( z ) is placed before the PEAK-LMS algorithm block (which is also evident from the channel designation d), so the adaptation will equalize the channel in the sense of H ( z ) * I ( z ) = R ( z ) .

[0064] As can be seen from the exemplary, non-exhaustive list of applications, the potential field of application is broad. Based on the descriptions, a specialist will easily be able to adapt the method according to the invention to their own tasks.

Claims

1. Computer-aided method for processing an input audio signal of an acoustic system, comprising at least one audio signal input, at least one audio signal output, a data link between the audio signal input and the audio signal output that is influenced by an adaptive filter (3), and an LMS algorithm (4) that controls the adaptive filter (3), wherein a) the LMS algorithm (4) estimates a transfer function h(n) based on a reference signal d(n) and an error signal e(n), b) the estimation of the transfer function is applied as an adaptive filter (3) and is present there as an initial filter function c(n), c) the output of the adaptive filter (3) is provided as a filter output sequence y(n), d) a comparison is carried out between an input signal sequence x(n) representing the input data of the input audio signal and the filter output sequence y(n) of the initial filter function c(n), whereby the error signal e(n) results from the difference, e) characterised in that the LMS algorithm used to adapt the adaptive filter (3) is a PEAK-LMS algorithm, wherein f) a sign function w(n) is used for the algorithm, which is calculated based on i. the input signal sequence x(n), ii. the filter output sequence y(n), iii. the error signal e(n), which defines the difference between the input signal sequence x(n) and the filter output sequence y(n) as e(n) = x(n) - y(n), iv. in the form of w(n) = sign(e(n) * x(n)) g) which, together with the initial filter function c(n) and time constants C(α, β), wherein α describes the time constants of an attack case and β the time constants of a release case of the adaptive filter (3), wherein C(α, β) = α if w(n) > c(n) and C(α, β) = β if w(n) ≤ c(n), is used to calculate the sequence of adapted filter values c(n + 1) using the equation c n + 1 = c n + C α β ∗ w n − c n .

2. Computer-aided method for the stable processing of an input audio signal of an acoustic system according to claim 1, characterised in that the time constant α of the attack case and the time constant β of the release case are equal.

3. Computer-aided method for the stable processing of an input audio signal of an acoustic system according to one of claims 1 or 2, characterised in that the input signal sequence x(n) is subjected to prefiltering before being fed to the algorithm.

4. Computer-aided method for the stable processing of an input audio signal of an acoustic system according to one of claims 1 to 3, characterised in that the error signal e(n) is subjected to prefiltering before being fed to the algorithm.

5. Computer-aided method for the stable processing of an input audio signal of an acoustic system according to one of claims 3 or 4, characterised in that high-pass, low-pass, or band-pass filters, as well as combinations thereof, are used for prefiltering.

6. Computer-aided method for the stable processing of an input audio signal of an acoustic system according to one of claims 1 to 5, characterised in that the sign function w(n) is extended by a filter coefficient λ such that the following holds: w(n) = λ * sgn(e(n) * x(n)).

7. Computer-aided method for the stable processing of an input audio signal of an acoustic system according to claim 6, characterised in that the filter coefficient λ lies in the interval [0,1].

8. Computer-aided method for the stable processing of an input audio signal of an acoustic system according to one of claims 6 or 7, characterised in that the filter coefficient λ may vary depending on the sign of the sign function w(n) according to λ = w n > 0 : λ + w n = 0 : 0 oder w n w n < 0 : λ _ .

9. Computer-aided method for the stable processing of an input audio signal of an acoustic system according to one of claims 1 to 8, characterised in that the step size µ of the PEAK-LMS algorithm is selected adaptively as a function of the error magnitude according to the system of equations μ n + 1 = a ∗ μ n + b ∗ e 2 n e n = x n − y n c n + 1 = c n + μ n ∗ e n ∗ x n wherein a and b are coefficients that ensure an adaptive step size and allow for a weighting between µ(n) and e2(n), and are linked via a = 1 - b.

10. Computer-aided method for the stable processing of an input audio signal of an acoustic system according to one of claims 1 to 9, characterised in that an adjustment of the adaptation rate is applied as a function of the error magnitude, according to the formulas: w n = λ ∗ sgn e n ∗ x n c n + 1 = c n + C α , β n ∗ w n − c n C α , β n + 1 = a ∗ C α , β n + b ∗ e 2 n wherein Cα,β(n) = α = β.

11. Computer-aided method for the stable processing of an input audio signal of an acoustic system according to claim 10, characterised in that the time constant α of the attack case and the time constant β of the release case are different.

12. Computer-aided method for the stable processing of an input audio signal of an acoustic system according to one of claims 1 to 9, characterised in that the algorithm is applied to the frequency domain.