EP1755110A2 - Method and device for adaptive reduction of noise signals and background signals in a speech processing system - Google Patents

Abstract

The method involves filtering an audio-input signal under application of adaptive filters (F1, F2), preferably of FIR configuration, for producing predicted-output signals with reduced noise. The filter is implemented under application of several coefficients for formation of several prediction errors and for the formation of the errors from several of prediction errors (sv1-sv4). The amount of coefficients is continuously reduced by several reduction parameters. An independent claim is also included for a device for reduction of noise signals and background signals in a speech processing system.

Description

The invention relates to a method for reducing noise and background signals in a speech processing system with the preamble features of claim 1 and to an apparatus for performing such a method with the preamble features of claim 18.

In speech processing systems, e.g. Systems for speech recognition, speech detection or speech compression, noise such as noise and non-speech background noise reduce the quality of speech processing, e.g. in terms of the detection or compression of the speech or speech signal components contained in an input signal. These disturbing background signals should be eliminated with as little computational effort as possible.

For reducing noise and background signals in speech processing systems, filter devices are employed which include at least one audio input, an audio output, a memory and a processor or a field programmable device or an application-specific integrated circuit (ASIC) Perform filtering procedure.

By means of a complex process using a spectral subtraction is in EP 1080465 and in US 6,820,053 For the reduction of noise and background signals, a spectrum of an audio signal is calculated using the Fourier transform and z. B. deducted a slowly increasing proportion. By inverse transformation into the time domain, a noise-reduced output signal is subsequently obtained. The computational effort is disadvantageous in this process is disadvantageous. In addition, the storage space consumption is very high. In addition, the parameters used in the spectral subtraction can be very poorly adapted to other sample rates.

To reduce noise and background signals, there are other methods, such as center clipping, in which an autocorrelation of the signal is formed and used as information of the noise content of the input signal US 5,583,968 or US 6,820,053 work with neural networks, which must be trained consuming or procedures, according to eg US 5,500,903 working with multiple microphones to separate noise and speech signals. At least, however, an estimate of the noise amplitudes is performed. The computational effort of a Fourier transform (FFT) is O (n log (n)), that of an autocorrelation O (n ² ), that of the entire method presented here is O (n).

mit µ << 1, z.B. µ = 0,01 als einer Lernrate, s(t) als einem Audio-Eingangssignal zur Zeit t, e = s(t) - sv(t) als einem Fehler aus einer Differenz aller einzelner Vorhersagefehler vom Audio-Eingangssignal, sv(t) als Ausgangssignal aus einer Summe der Terme c_i(t-1) . s(t-i), d.h. der einzelnen Vorhersagefehler über alle i von 1 bis N, N als Anzahl der Koeffizienten und c_i(t) als einem individuellen Koeffizienten mit einem Parameter i zur Zeit t.Commonly known is the use of an FIR filter (Finite Impulse Response / Finite impulse response), which is trained to predict the input signal from eg speech and noise as well as possible from the past n values, this using an LPC (Linear Predictive coding / linear prediction coding). The output values of the filter are these predicted values. The magnitudes of coefficients c (i) of such a filter increase on average more slowly in the case of noise signals than on speech signals, the coefficients being calculated in accordance with FIG $${\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\mathrm{+}\mathrm{1}\right)\mathrm{=}{\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\right)\mathrm{+}\mathrm{\mu }\mathrm{\cdot }\mathrm{e}\mathrm{\cdot }\mathrm{s}\left(\mathrm{t}\mathrm{-}\mathrm{i}\right)$$

with μ << 1, eg μ = 0.01 as a learning rate, s (t) as an audio input signal at time t, e = s (t) - sv (t) as an error from a difference of all individual prediction errors from Audio input signal, sv (t) as output from a Sum of the terms c _i (t-1). s (ti), ie the individual prediction error over all i from 1 to N, N as the number of coefficients and c _i (t) as an individual coefficient with a parameter i at time t.

The object of the invention is to improve a method for the reduction of noise and background signals in a voice-processing system or a device for carrying out such a method in terms of applicability, in particular to make it more flexible.

This object is achieved by a method for reducing noise and background signals in a voice-processing system having the features of patent claim 1 or to an apparatus for carrying out such a method having the features of patent claim 18. Advantageous embodiments are the subject of dependent claims.

Accordingly, a method of reducing noise and background signals in a voice processing system in which an audio input signal is filtered by means of filtering using an adaptive filter to generate a predicted output signal with reduced noise, wherein the filtering is performed using, is accordingly preferred a plurality of coefficients for forming a plurality of prediction errors and for forming one of the plurality of prediction errors, wherein by means of a plurality of reduction parameters, the amounts of the coefficients are continuously reduced.

In particular, a method is preferred in which the continuous reduction of the coefficients is produced by multiplying the coefficients by a factor smaller than 1, in particular by multiplying by a factor between 0.8 and 1.0.

mit

• k mit 0 > k << 1, insbesondere k <= 0,0001, als einem Reduktionsparameter,
• µ << 1, insbesondere µ <= 0,01, als einer Lernrate,
• s(t) als einem Audio-Eingangssignal zu einer Zeit t,
• e als einem Fehler aus einer Differenz aller einzelner Vorhersagefehler (sv1 - sv4) vom Audio-Eingangssignal s(t),
• sv(t) als dem Vorhersage-Ausgangssignal aus einer Summe aller einzelnen Vorhersagefehler, mit N als Anzahl der Koeffizienten c_i(t) und
• c_i(t) als individuellem Koeffizient mit einem Index i zur Zeit t.

In particular, a method in which the coefficients c _i (t) are calculated is preferred $${\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\mathrm{+}\mathrm{1}\right)\mathrm{=}{\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\right)\mathrm{+}\mathrm{\mu }\mathrm{\cdot }\mathrm{e}\mathrm{\cdot }\mathrm{s}\left(\mathrm{t}\mathrm{-}\mathrm{i}\right)-\mathrm{k}{\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\right)$$

With

• k with 0> k << 1, in particular k <= 0.0001, as a reduction parameter,
• μ << 1, in particular μ <= 0.01, as a learning rate,
• s (t) as an audio input signal at a time t,
• e as an error from a difference of all individual prediction errors (sv1 - sv4) from the audio input signal s (t),
• sv (t) as the prediction output from a sum of all the individual prediction errors, with N as the number of coefficients c _i (t) and
• c _i (t) as an individual coefficient with an index i at time t.

mit

• 
  und
• 

In particular, such a method is preferred in which the coefficients are calculated in accordance with $${\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\mathrm{+}\mathrm{1}\right)\mathrm{=}{\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\right)\mathrm{+}\mathrm{\mu }\mathrm{\cdot }\mathrm{e}\mathrm{\cdot }\mathrm{s}\left(\mathrm{t}\mathrm{-}\mathrm{i}\right)-\mathrm{k}{\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\right)$$

With

• 
  and
• 

In particular, a method is preferred in which the predicted output signal is used as a prediction of the reduced noise audio input signal as an input to a subsequent second filtering to produce a second prediction. In particular, such a method is preferred in which the second filtering is carried out by means of predictive filtering with a second, known per se filtering with a set of second coefficients, wherein a learning rate for adjusting the coefficients by a few powers of ten is chosen to be less than a learning rate of first filtering.

In particular, a method is preferred in which the second prediction is then subtracted from the prediction output signal in order to eliminate long-lasting background noise.

In particular, a method is preferred in which a learning rule for determining the further coefficients is asymmetrical, so that the magnitude of the further coefficients falls more sharply in magnitude than can rise and drop rapidly to zero, but increases only with a small slope.

In particular, a method is preferred in which, instead of the audio input signal for determining individual prediction errors, only its sign is used so as not to disadvantage small signals.

Particularly preferred is a method in which the coefficients are limited to avoid drifting of the coefficients, in particular to a range of e.g. -4 ... 4, if the audio input signal is standardized from -1 ... 1.

In particular, a method is preferred in which a maximum of a speech signal component of the audio input signal is detected and the output signal is normalized, in particular sluggishly, to this maximum.

In particular, a method is preferred in which the output signal of the first and / or the second filtering in relation to their input signal is used in particular simultaneously as a measure of the presence of speech in the input signal.

In particular, a method is preferred in which a filter is used for the first and / or the second filtering, which performs an error prediction by means of an LMS adaptation (Least Mean Squares Adaption). In particular, it is preferred a method in which an FIR filter is used for the first and / or the second filtering.

In particular, a method is preferred in which a sigmoid function is multiplied by the predicted output signal to avoid overdriving the signal in the case of a bad prediction.

Particularly preferred is a method in which the predicted output signal as the original signal, the audio input signal is mixed to produce a more natural sound.

In particular, a method is preferred in which a field-programmable component or an ASIC (Application-Specified-Integrated-Circuit) is correspondingly programmed for carrying out the method.

Accordingly, what is preferred is an apparatus, in particular apparatus for carrying out a method for reducing noise and background signals in a speech processing system, having an audio input for inputting an audio input signal to an adaptive filter for filtering audio input signal to produce a prediction output signal with reduced noise, with a memory for storing a plurality of coefficients for the filter, wherein the filter is configured or switched to form a plurality of prediction errors and to form an error of the plurality of prediction errors, wherein a coefficient providing arrangement is formed or switched by means of at least one reduction parameter to reduce the amounts of the coefficients continuously.

In particular, a device is preferred in which the coefficient providing arrangement for multiplying the coefficients by the reduction parameter as a factor less than 1, in particular with a factor between 0.8 and 1.0 is formed or connected.

In particular, a device is preferred in which a first filter stage with the filter as the first filter is followed by a second filter stage with a second filter for supplying the predicted output signal as a prediction of the audio input signal with reduced noise as an input signal for the second filter for generating a second prediction.

Particularly preferred is a device having a subtraction circuit for subtracting a sum of error predictions of the second filtering from the prediction output signal to produce the prediction.

In particular, a device is preferred in which the second filter is designed or switched by an LMS adaptation filter for carrying out an error prediction.

In particular, a device is preferred in which the first filter and / or the second filter is formed or switched by an FIR filter for performing a signal prediction.

In particular, a device which is formed by a field-programmable component or an ASIC is preferred.

In particular, a device is provided with a multiplier for weighting the optionally time delayed audio input signal or for weighting the prediction output signal with a weighting factor less than one, in particular about 0.1 and an adder for adding the weighted signal to the prediction output signal or the Prediction for generating a noise-reduced audio output signal.

Across from EP 1080465 and US 6,820,053 is the computational effort in the preferred method here many times lower. In addition, the storage space consumption is many times lower. In addition, the problem of very poor adaptation of the parameters used to other sample rates as in spectral subtraction is eliminated.

Compared to the various known methods, the computational effort is much lower. Whereas in a Fourier transformation the computation outlay is O (n log (n)) and the computational cost of an autocorrelation is O (n ² ), the computational effort of the preferred method of the entire method of both filter stages is only O (n), where n a number of sampled samples (nodes) of the input signal and O is a general function of the filter overhead.

The particularly preferred filter arrangement results in a large number of advantages. A speech signal is only delayed by a single sample. An adaptation is for noises instantaneously and for long-lasting background noises, the adaptation is preferably delayed about 0.2 s to 5.0 s.

The method is much less computationally expensive than conventional methods. Especially with only four coefficients one obtains respectable results, so that only four multiplications and four additions have to be calculated for the prediction of one sample and only four to five further operations for the adaptation of the filter coefficients are required.

In addition, there is a lower memory consumption than for conventional methods, such as the spectral subtraction.
A simple adjustment of the parameters is also possible with different sample rates. In addition, the amount of filtering for noise and for long-lasting background signals can be set separately.

Fig. 1

eine bevorzugte Filteranordnung sur Reduktion von Rausch- und Hintergrundsignalen in einem sprachverarbeitenden System mit zwei hintereinander geschalteten Filterstufen,

Fig. 2

vergrößert dargestellt die erste der biden Filterstufen und

Fig. 3

vergrößert dargestellt die zweite der beiden Filterstufen.

An embodiment will be explained in more detail with reference to the drawing. Show it:

Fig. 1

a preferred filter arrangement sur reduction of noise and background signals in a voice processing system with two filter stages connected in series,

Fig. 2

enlarged the first of the biden filter stages and

Fig. 3

shown enlarged the second of the two filter stages.

As can be seen from FIG. 1, the particularly preferred method consists of two adaptive filters F1, F2, which are connected in series as a first and a second filter stage. Autonomously advantageous, however, is already the use of only the first filter stage.

In the particularly preferred circuit arrangement, an audio input signal s (t) is input via an audio input 1. The audio input signal is applied to a group of delay elements 2, which z. B. are formed as a buffer and delay the respective applied value of the audio input signal s (t) by one clock. In addition, the audio input signal s (t) is supplied to a first adder 3. The values s (t-1) -s (t-4) delayed by means of the delay elements 2 are output from the respective delay element 2 to the next of the delay elements 2 and respectively two corresponding multipliers of two groups of multipliers 4 created. The group of second multipliers 5 is applied to a further multiplication input in each case a coefficient c1 - c4 as a filter coefficient of an adaptive filter. The multiplication results of the group of second multipliers 5 are output to a second adder 6 as individual prediction errors sv1-sv4. A time sequence of the addition values of the second adder 6 forms a prediction output sv (t).

The sequence of values of the predicted output signal sv (t) are directly output according to a first advantageous embodiment to form an output signal o (t) (FIG. 2).

The sequence of values of the predicted output sv (t) are also applied to the first adder 3 formed as a subtraction circuit at a subtraction input to subtract these values from the current later value of the audio input signal s (t). The subtraction result of the first adder 3 forms an error e from a corresponding sequence of individual error values. This error e is applied to a third multiplier 8, at the second multiplication input of which a value of a learning rate μ with preferably μ≈0.01 is applied. The multiplication result is applied to the inputs of the group of first multipliers 4 for multiplication by the delayed values s (t-1) -s (t-4).

The multiplication results of the group of first multipliers 4 are supplied to a group of third adders 10, which form an input of a coefficient providing arrangement 9. The output values of the group of third adders 10 form the coefficients c1 - c3, which are applied to the corresponding multipliers 5 of the group of second multipliers 5. In addition, these coefficients c1-c4 are each applied to an adder 11 of a group of fourth adders 11 and in each case to a multiplier 12 of a group of fourth multipliers 12. To the group of fourth multipliers 12 For example, a reduction parameter k is applied to a multiplication input, wherein the value of the reduction parameter k is, for example, 0.0001. By means of the reduction parameter k, the respective value of the coefficients c1-c4 is correspondingly reduced by this factor. The corresponding multiplication result of the fourth multipliers 12 is applied to the respective one of the fourth adder 11 formed as a subtraction circuit, to which the corresponding coefficient c1-c4 has previously been applied, at a subtraction input. The output value of the respective adders 11 of the fourth group adder 11 is applied to another input of the corresponding third adder of the group of third adders 10. In this case, the group of third adders 10 adds the respective addition value of the group of fourth adders 11 to the respectively applied and delayed audio signal input value s (t-1) -s (t-4) in order to learn the coefficients.

By means of an adder 7, the prediction output signal sv (t) for forming the output signal o (t) can optionally be added with a weighted value formed directly from the instantaneous or optionally from a correspondingly delayed value of the audio input signal s (t). The weighted value is provided by a weighting multiplier 15, which multiplies the input signal s (t) by a factor η <1, in particular η ≈ 0.1.

Preferably, the predicted output sv (t) or the output o (t) is not output as a final output but provided as an input to a second filter stage with the second filter F2.

As shown in Fig. 3, the second filter F2 is again an adaptive filter arrangement, the construction of which is preferably substantially equal to the structure of the first filter stage. Therefore, only differences to the first filter stage will be described below. The respective components and signals or values are with a Star for distinguishing the corresponding components and signals or values of the first filter stage characterized.

Different is the generation of the coefficients c * 1 - c * 4 in a comparison with the first filter stage modified coefficient providing means 9 *. The coefficients c * 1 - c * 4 are in a known manner of a z. B. adaptive FIR filter formed without a multiplication with a reduction parameter k. Another difference compared to both the first filter stage of the first filter F1 and a conventional FIR filter is that the value of a learning rate μ * for the second filter F2 smaller, especially much smaller than the value of the learning rate μ of the first filter F1 is selected becomes.

The output result of the second filter F2 is correspondingly provided by a second adder 6 * of the second filter F2 and added to the input signal or the corresponding input value of the input signal sv (t) of the second filter F2 by means of a fifth adder 13 * or preferably subtracted therefrom Case of an adder 6 *, preferably designed as a subtraction circuit. The output of the fifth adder 13 * forms a second prediction sv * (t) as a second prediction output. Preferably, the values of the prediction sv * (t) are added to the optionally time-delayed and weighted audio input signal s (t) or sv (t) by means of a sixth adder 14 * to produce a noise-reduced audio output signal o * (t). For the purpose of weighting, a multiplication of the audio input signal s (t) with a weighting factor η * <1, in particular η ≈ 0.1 in a multiplier 15 *, which precedes the sixth adder 14 *. In order to control the method steps, the arrangement has further components in a conventional manner or is connected to further components such as a processor for control functions and a clock generator for providing a clock signal. To store the coefficients c1 - c4, c * 1 - c * 4 and, if necessary, other values, the arrangement has a memory or can access a memory.

The first filter F1 reduces the noise over the entire perceived frequency range. Here, a modified adaptive FIR filter is trained to predict the audio input signal s (t) containing eg speech and noise as well as possible from the past n values. The output is the predicted values as the prediction output sv (t). The amounts of the general coefficients c _i (t) according to FIG. 1 with an index i = 1, 2, 3, 4 and corresponding to the coefficients C 1 -C 4 of such a first filter F1 increase more slowly with noise signals than with speech signals.

mit $$\mathrm{e}\mathrm{=}\mathrm{S}\left(\mathrm{t}\right)\mathrm{-}\mathrm{sv}\left(\mathrm{t}\right)\mathrm{,}$$

$$\mathrm{sv}\left(\mathrm{t}\right)\mathrm{=}\mathrm{\sum }_{\mathrm{i}\mathrm{=}\mathrm{1}\mathrm{\dots }\mathrm{N}}{\mathrm{c}}_{\mathrm{i}}\mathrm{(}\mathrm{t}\mathrm{-}\mathrm{1}\mathrm{)}\mathrm{\cdot }\mathrm{s}\left(\mathrm{t}\mathrm{-}\mathrm{i}\right)\mathrm{und}$$

und mit k mit 0 > k << 1, z.B. k = 0,0001, als einem Reduktionsparameter, mit µ << 1, z.B. µ = 0,01, als einer Lernrate, mit s(t) als einem Audio-Eingangssignal zur Zeit t, mit e als einem Fehler aus einer Differenz aller einzelner Vorhersagefehler vom Audio-Eingangssignal, mit sv(t) als einem Vorhersage-Ausgangssignal aus einer Summe der Koeffizienten multipliziert mit den zugehörigen verzögerten Signalen, mit N als Anzahl der Koeffizienten c_i(t) und mit c_i(t) als individuellem Koeffizient mit einem Parameter bzw. Index i zur Zeit t.The filtering is done in analogy to the LPC. Instead of a delta rule or an LMS learning step according to the prior art, a modified filtering method is now used in which the coefficients c _i (t) are generally calculated according to a new learning rule $${\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\mathrm{+}\mathrm{1}\right)\mathrm{=}{\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\right)\mathrm{+}\mathrm{\mu }\mathrm{\cdot }\mathrm{e}\mathrm{\cdot }\mathrm{s}\left(\mathrm{t}\mathrm{-}\mathrm{i}\right)-\mathrm{k}{\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\right)$$

With $$\mathrm{e}\mathrm{=}\mathrm{S}\left(\mathrm{t}\right)\mathrm{-}\mathrm{sv}\left(\mathrm{t}\right)\mathrm{.}$$

$$\mathrm{sv}\left(\mathrm{t}\right)\mathrm{=}\underset{\mathrm{i}\mathrm{=}\mathrm{1}\mathrm{...}\mathrm{N}}{\mathrm{\Sigma }}{\mathrm{c}}_{\mathrm{i}}\mathrm{(}\mathrm{t}\mathrm{-}\mathrm{1}\mathrm{)}\mathrm{\cdot }\mathrm{s}\left(\mathrm{t}\mathrm{-}\mathrm{i}\right)\mathrm{and}$$

and with k with 0> k << 1, eg k = 0.0001, as a reduction parameter, with μ << 1, eg μ = 0.01, as a learning rate, with s (t) as an audio input signal to Time t, with e as an error of a difference of all individual prediction errors from the audio input signal, with sv (t) as a prediction output signal of a sum of the coefficients multiplied by the associated delayed signals, with N as the number of coefficients c _i ( t) and with c _i (t) as an individual coefficient with a parameter or index i at time t.

According to the learning rule using the reduction parameter k, the amounts of the coefficients c _i (t) are continuously reduced, resulting in smaller predicted amplitudes in noise signals than in speech signals. It is determined with the reduction parameter k, how much the noise should be suppressed.

The second filter F2 reduces long-lasting background noise. It exploits the fact that the energy of speech signal components in the audio input signal s (t) repeatedly falls to zero in individual frequency bands, whereas long-lasting tones tend to have a constant energy in the frequency band. An adaptive FIR filter with extremely low learning rate of e.g. μ = 0.000001 is now adapted so slowly for prediction by means of LPC in particular that the speech signal component in the audio input signal s (t) is predicted with much lower amplitude than long-lasting signals. Finally, the prediction sv * (t) thus obtained in the second filter F2 is subtracted from the input signal s (t), so that the long-lasting signals from the input signal s (t) are eliminated or at least greatly reduced.

The first and second filters F1, F2 are particularly efficient when executed one after the other on the input signal s (t), as shown in FIG. In this case, first the first filter F1 is executed and its output or predicted output signal sv (t) is passed as an input signal to the second filter F2 for further additional filtering.

1 schematically shows an amplitude curve a over the time t of an exemplary input signal s (t) in the time domain before and after the filtering by the first filter F1 for noise suppression. While the input signal s (t) includes speech and noise, the prediction output sv (t) of the first filter F1 includes speech and a reduced noise.

Fig. 2 shows schematically an amplitude curve a over the time t of an exemplary input signal s (t) and the prediction output signal sv (t) in the frequency range before and after the filtering by the second filter F2 for the suppression of long-lasting background noise. In this case, the x-axis corresponds to the time t, the y-axis corresponds to a frequency f and a brightness corresponds to an amplitude. Recognizable is a spectrum of a prominent 2 kHz tone in the background in front of the second filter F2 compared to a spectrum with a reduced 2 kHz tone after the second filter F2.

Instead of a continuous reduction of the coefficients C1-C4 according to formula (2), the reduction of the coefficients c _i (t) can alternatively or additionally also be produced by the coefficients c _i (t) having a fixed or variable factor between in particular 0, 8 and 1.0 are multiplied.

Advantageously, a method and apparatus, respectively, in which a sigmoidal function, e. G. (T), is applied after the first filter F1 has been used with its predicted output sv (t). a hyperangular hyperbaric, which avoids overdriving the signal in case of a bad prediction.

A method and a device are advantageous if the audio input signal (s (t)) is added to the prediction output signal (sv (t)) as the original signal in order to produce a more natural sound.

Instead of a single reduction parameter k for all coefficients c1-c4, a plurality of reduction parameters for the different coefficients c1-c4 can also be determined or determined individually. In particular, the reduction parameter or k can also be dependent on e.g. be varied according to the received audio input signal.

Claims (27)

Method for reducing noise and background signals in a speech processing system, in which an audio input signal (s (t)) is filtered by means of filtering using an adaptive filter to produce a predicted output signal (sv (t)) with reduced noise, wherein the filtering is performed using a plurality of coefficients (c _i (t); c1-c4) to form a plurality of prediction errors (sv1-sv4) and to form an error (e) from the plurality of prediction errors (sv1-sv4) SV4), characterized in that - By means of a plurality of reduction parameters (k), the amounts of the coefficients (c _i (t); c1 - c4) are continuously reduced. Method according to Claim 1, in which the continuous reduction of the coefficients (c _i (t)) is produced by multiplying the coefficients (c _i (t)) by a factor of less than 1, in particular by a factor between 0.8 and 1.0 multiplied. Method according to Claim 1 or 2, in which the coefficients (c _i (t)) are calculated in accordance with $${\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\mathrm{+}\mathrm{1}\right)\mathrm{=}{\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\right)\mathrm{+}\mathrm{\mu }\mathrm{\cdot }\mathrm{e}\mathrm{\cdot }\mathrm{s}\left(\mathrm{t}\mathrm{-}\mathrm{i}\right)-\mathrm{k}{\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\right)$$

With k with 0> k << 1, in particular k <= 0.0001, as a reduction parameter, p << 1, in particular μ <= 0.01, as a learning rate, s (t) as an audio input signal at a time t, e as an error from a difference of all individual prediction errors (sv1-sv4) from the audio input signal s (t), sv (t) as the prediction output from a sum of all single prediction error, with N as the number of coefficients c _i (t) and - c _i (t) as an individual coefficient with an index i at time t. The method of claim 3, wherein the coefficients (c _i (t)) are calculated according to $${\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\mathrm{+}\mathrm{1}\right)\mathrm{=}{\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\right)\mathrm{+}\mathrm{\mu }\mathrm{\cdot }\mathrm{e}\mathrm{\cdot }\mathrm{s}\left(\mathrm{t}\mathrm{-}\mathrm{i}\right)-\mathrm{k}{\mathrm{c}}_{\mathrm{i}}\left(\mathrm{t}\right)$$

With

and

A method according to any preceding claim, wherein the predicted output signal (sv (t)) is used as a prediction of the reduced noise audio input signal as an input to a subsequent second filtering (F2) to produce a second prediction (sv * (t )). Method according to Claim 5, in which the second filtering (F2) is carried out by means of prediction filtering with a second, in particular per se known, filtering with a set of second coefficients (c _i (t), c * 1-c * 4)), wherein a learning rate (μ *) for adjusting the coefficients by a few powers of ten is chosen smaller than a learning rate (μ) of the first filtering (F1). The method of claim 5 or 6, wherein the second prediction (sv * (t)) is then subtracted from the predicted output signal (sv (t)). Method according to one of Claims 5 to 7, in which a learning rule for determining the further coefficients (c _i * (t); c * 1 - c * 4) is made asymmetrical, such that the magnitude of the further coefficients (c _i * ( t); c * 1 - c * 4) fall more in the amount than rise and fall rapidly to zero but can only increase with a small slope. Method according to any preceding claim, wherein instead of the audio input signal (S (t)) for determining individual prediction errors (sv1 - sv4) only its sign is used. A method as claimed in any preceding claim, wherein the coefficients (c _i (t); c1-c4) are limited to avoid drifting the coefficients, in particular -4 ... 4, when the audio input signal is -1 ... 1 is normalized. Method according to one of the preceding claims, wherein a maximum of a speech signal component of the audio input signal (s (t)) is detected and the output signal (o (t)) is normalized again to this maximum. A method as claimed in any preceding claim, wherein the output signal (sv (t); sv * (t)) of the first and / or second filtering in relation to its input signal (s (t); sv (t)) is a measure of the presence of speech is used in the input signal. Method according to one of the preceding claims, in which a filter which performs an error prediction by means of a Least Mean Square Adaptation (LMS) adaptation is used for the first and / or the second filtering. Method according to any preceding claim, wherein an FIR filter is used for the first and / or the second filtering. A method as claimed in any preceding claim, wherein a sigmoidal function is multiplied by the predicted output signal (sv (t)) to avoid overdriving the signal in the case of a bad prediction. A method according to any preceding claim, wherein the audio input signal (s (t)) is mixed in with the predicted output signal (sv (t)). A method according to any preceding claim, wherein a field programmable device or an ASIC (Application-Specified-Integrated-Circuit) is programmed accordingly to perform the method. Device for reducing noise and background signals in a speech processing system, with an audio input (1) for inputting an audio input signal (s (t)), an adaptive filter (F1) for filtering audio input signal (s (t)) for generating a predicted output signal (sv (t)) with reduced noise, a memory for storing a plurality of coefficients (c _i (t); C1-C4) for the filter (F1), - wherein the filter (F1) is designed or switched to form a plurality of prediction errors (sv1 - sv4) and to form an error (e) from the plurality of prediction errors (sv1 - sv4), characterized in that - A coefficient-providing arrangement (9) is formed or connected by means of at least one reduction parameter (k) to reduce the amounts of the coefficients (c _i (t); C1 - C4) continuously. Apparatus according to claim 18, wherein the coefficient providing means (9) for multiplying the coefficients (c _i (t)) by the reduction parameter (k) as a factor k is less than 1, in particular with a factor between 0.8 and 1.0 is formed or switched. Apparatus according to claim 18 or 19, in which a first filter stage with the filter as the first filter (F1) is followed by a second filter stage with a second filter (F2) is for supplying the predicted output signal (sv (t)) as a prediction of the reduced noise audio input signal (s (t)) as an input to the second filter (F2) to produce a second prediction (sv * (t)) , Apparatus according to claim 20, further comprising an adder (13) for adding a sum of error predictions (sv * 1 -sv * 4) of the second filter (F2) from the prediction output signal (sv (t)) of the first filter (F1) to Generating the prediction (sv * (t)). Apparatus according to any one of claims 18 to 21, wherein the second filter (F2) is formed or switched by an LMS adaptation filter to perform error prediction. Apparatus according to any one of claims 18 to 21, wherein the first filter (F1) and / or the second filter (F2) is formed or switched by a FIR filter for performing signal prediction. Device according to one of claims 18 to 22, which is formed by a field programmable device or an ASIC. Apparatus according to any one of claims 18 to 24, including a subtraction circuit (14) for subtracting the values of the prediction (sv * (t)) of values of the audio input signal (s (t)) to produce a noise-reduced audio output signal (o * ( t)). Device according to one of claims 18 to 25 with - a multiplier (15; 15 *) for weighting the optional time-delayed audio input signal (s (t)) or weighting the prediction output signal (sv (t)) with a weighting factor (η; η *) less than one, in particular about 0.1 and an adder (7; 14 *) for adding the weighted signal to the predicted output signal (sv (t)) or the prediction (sv * (t)) for generating a noise-reduced audio output signal (o (t); o * (t)). Apparatus according to any one of claims 18-26, adapted to carry out a method according to any one of claims 1-17.

Images