EP0948237A2 - Method for noise suppression in a microphone signal - Google Patents

Abstract

The noise reduction method uses subtraction of a noise compensation signal from the microphone signal (y), the noise compensation signal provided by simulation of the noise signal via an adaptive filter (H) using a least mean square algorithm, supplied with a reference signal, e.g. a loudspeaker signal, with transformation of the microphone signal, the compensation signal and the output signal into the frequency range.

Description

The invention relates to a method for eliminating interference Microphone signal.

Such methods are particularly useful for voice input of commands and / or for hands-free phones increasingly of importance, especially the situation in is an important application for a vehicle.

This often creates a special situation in vehicles given that a playback device such as e.g. a radio, a Cassette or CD player through a loudspeaker Generated noise environment, which as a noise signal from one Microphone recorded voice signal, for example for voice recognition or telephone transmission is superimposed. For Recognition of voice inputs in a speech recognizer or for understandable voice transmission over the phone Microphone signal as far as possible from interference signal components to free.

That from a source of interference, especially a loudspeaker outgoing interference signal does not only arrive at the shortest direct Way to the microphone, but also occurs via numerous Reflections as an overlay of a plurality of Echoes with different delay times in the microphone signal in Appearance. The total exposure to the interference signal from the Interference source on the microphone signal can by a priori unknown transfer function of the room, for example of the passenger compartment of a motor vehicle are described. The transfer function changes depending on the occupation of the Vehicle and according to the position of the individual. By Simulation of this transfer function and filtering a Reference signal from the interference source with this replica a compensation signal can be generated, which by Subtraction from the microphone signal is freed from the interference signal Signal, for example a pure voice signal. In the real case, the replica mentioned represents a more or less good approximation to the unknown transfer function and the malfunction cannot be completely eliminated become.

The object of the present invention is a method to provide interference from a microphone signal that at reasonable signal processing effort good properties with regard to interference suppression.

The invention is described in claim 1. The subclaims contain advantageous configurations and Developments of the invention.

It is essential in the method according to the invention that the Compensation of the interference signal component in the microphone signal by means of one from the reference signal via the simulation of the Transfer function generated compensation signal in the frequency domain is made so that microphone signal, compensation signal and output signal in the frequency domain, i.e. are in the form of spectra. The signal processing required in this step in the frequency domain a spectral transformation of the microphone signal, but takes into account that the simulation of the transfer function is more advantageous in the frequency domain and provides for an advantageous subsequent additional noise reduction of the output signal, which is also typically in the frequency domain is already a special one suitable waveform ready.

By simple approximations when replacing a processing step with a time window can be made by moving to a Convolution in the frequency domain a significant reduction in the Processing effort can be reduced.

For long impulse responses of the transfer function or its Replica sees an advantageous development of Invention a division of the replica filter into several Sub-filter for time-shifted segments of the segmented Reference signal before, its coefficient update can be staggered in time, reducing the signal processing effort can be kept low.

Trouble shooting proves to be particularly advantageous of a speech signal based on a setting of the Replica filters in a previous language break won and saved.

The division of the replica filter into several sub-filters and the interference clearance based on one in one Speech pause filter settings are also independent of interference signal compensation in the frequency domain independently for the interference suppression of a microphone signal feasible and advantageous.

Fig. 1

ein Prinzip der Kompensation eines Radiosignals

Fig. 2a

ein Blockschaltbild zu Fig. 1

Fig. 2b

ein Blockschaltbild zur Filternachbildung

Fig. 3

ein detailliertes Beispiel zu Fig. 2b

Fig. 4

eine Erweiterung auf mehrere Teilfilter

Fig. 5

einen Übergang zur Kompensation im Frequenzbereich

Fig. 6

ein detailliertes Beispiel zu Fig. 5b

Fig. 7

ein Ausführungsbeispiel mit mehreren Teilfiltern

Fig. 8

ein Ausführungsbeispiel mit Speicherung der Filtereinstellungen

Fig. 9

Signale einer synthetischen Beispielsszene

Fig. 10

Impulsantwort und Übertragungsfunktion zu Fig. 9

Fig. 11

Signal einer ersten Meßszene

Fig. 12

Impulsantwort und Übertragungsfunktion zu Fig. 11

Fig. 13

das Beispiel nach Fig. 11 mit Speicherung der Filtereinstellungen

Fig. 14

eine Sprachpausendetektion zu Fig. 13

Fig. 15

Impulsantworten und Übertragungsfunktionen zu Fig. 11 und Fig. 13

Fig. 16

Übergang von einem Zeitfenster zu einer Faltung im Frequenzbereich

Fig. 17

ein Rechteck-Zeitfenster mit Linienspektrum

Fig. 18

ein Hamming-Zeitfenster mit Linienspektrum

Fig. 19

Staffelung von Signalblöcken bei der Filterberechnung

Fig. 20

Signale einer zweiten Meßszene

Fig. 21

eine Sprachpausendetektion zu Fig. 20

Fig. 22

Impulsantworten und Übertragungsfunktionen zu Fig. 20 und Fig. 21

Fig. 23

Signale einer dritten Meßszene

Fig. 24

eine Sprachpausendetektion zu Fig. 23

Fig. 25

Impulsantworten und Übertragungsfunktionen zu Fig. 23 und Fig. 24

Fig. 26

Signale einer vierten Meßszene

Fig. 27

eine Sprachpausendetektion zu Fig. 26

Fig. 28

Impulsantworten und Übertragungsfunktionen zu Fig. 26 und Fig. 27.

The invention is illustrated in more detail below on the basis of preferred exemplary embodiments with reference to the figures. It shows:

Fig. 1

a principle of compensation for a radio signal

Fig. 2a

2 shows a block diagram of FIG. 1

Fig. 2b

a block diagram for filter replication

Fig. 3

a detailed example of Fig. 2b

Fig. 4

an extension to several sub-filters

Fig. 5

a transition to compensation in the frequency domain

Fig. 6

a detailed example of Fig. 5b

Fig. 7

an embodiment with several sub-filters

Fig. 8

an embodiment with storage of the filter settings

Fig. 9

Signals from a synthetic example scene

Fig. 10

Impulse response and transfer function for Fig. 9

Fig. 11

Signal from a first measurement scene

Fig. 12

Impulse response and transfer function for Fig. 11

Fig. 13

the example of FIG. 11 with storage of the filter settings

Fig. 14

a speech pause detection for FIG. 13

Fig. 15

Pulse responses and transfer functions for FIGS. 11 and 13

Fig. 16

Transition from a time window to a convolution in the frequency domain

Fig. 17

a rectangular time window with a line spectrum

Fig. 18

a Hamming time window with a line spectrum

Fig. 19

Staggering of signal blocks in the filter calculation

Fig. 20

Signals from a second measurement scene

Fig. 21

a speech pause detection for FIG. 20

Fig. 22

Impulse responses and transfer functions for FIGS. 20 and 21

Fig. 23

Signals from a third measurement scene

Fig. 24

a speech pause detection for FIG. 23

Fig. 25

Pulse responses and transfer functions to FIGS. 23 and 24

Fig. 26

Signals from a fourth measurement scene

Fig. 27

a speech pause detection for FIG. 26

Fig. 28

Impulse responses and transfer functions to FIGS. 26 and 27.

1 shows the principle of a device for (single-channel) Radio signal compensation. That from the speaker emitted acoustic signal arrives directly, but also about numerous reflections in the vehicle interior, on the microphone of the voice input system. Under the Assumption that the transmission path G is accordingly Transversal filter with a weighted sum of time delays Represents echoes, can be a filter replica H find, which in the ideal case H = G a complete compensation of the radio signal.

The loudspeaker signal x is filtered by the a priori unknown transfer function G of the vehicle interior. The interference component r arises, which is added to the microphone signal y with the speech signal s. In order to compensate for the interference component r, an estimate r ^ is generated from the loudspeaker signal x using the filter simulation H. The output of the circuit provides the estimate for the speech signal: s ^ = s + r - r ^ = s + E

The speech signal s is therefore still the error signal at the output of the circuit E = rr ^ superimposed, which should be kept as small as possible in practice. The voice signal can still contain interference in the form of, for example, engine noise or external noise, but these are not dealt with explicitly in this context.

H is an adaptive filter and works according to one in the Literature known standard method, the LMS algorithm (least mean squares). In addition to the input signal x the error signal E needed to adapt the coefficient to accomplish in filter H. For this is the output signal s ^ fed to the determination of the filter coefficients.

2a shows the arrangement again in another representation 1 as radio signal compensation. The adaptive system H can e.g. in the time domain as FIR filter (finite impulse response filter) will be realized. With long impulse response lengths, as they often occur in practice however, this requires a very high computing effort. Different advantages over a time domain solution offers the implementation of the LMS algorithm in the frequency domain (FLMS). Because of the block processing of Data realized in the discrete Fourier transforms spectral transformations and filter implementation in the frequency domain by multiplication this becomes Process particularly economical in terms of computing time.

2b shows a block diagram of the FLMS algorithm. The associated theory is known per se and is therefore not dealt with in detail here. F is a spectral transformation FFT of a time signal into the frequency domain and F ^{-1 is} the inverse IFFT. The processing steps designated as projections P1, P2 and P3 are used for the correct segmentation of the data by the block-wise use with the FFT or IFFT and are explained in more detail later. The filter works by multiplying the reference spectrum X by the filter coefficient vector H. The spectrum of the filter output R ^ is transformed back into the time domain via F ^-1 . After applying the projection P2 to the real part of the compensation signal thus obtained, the signal r ^ is available. The difference of the signals s ^ = y - r ^ = s + r - r ^ = s + E represents the actual outcome, an estimate of the speech input.

An essential part of the adaptive filter is the coefficient adaptation in block K, that in FIG. 2b by the renewal equation H '= H' + ΔH ' is described. The projection P1, which is particularly complex here with two spectral transformations, calculates from H 'the coefficient vector H required for the filtering. In addition to the reference spectrum X, the spectrum S ^ of the output signal evaluated with P ₃ is used to calculate the correction vector ΔH' s + r - r ^ needed.

A detailed block diagram of the FLMS algorithm shown in FIG. 2b is shown in FIG. 3. The sample values of a signal and the reference points of the FFT are commonly referred to as samples. All spectral transformations and their inverses are to be segmented as 256-point FFTs, each overlapping by 128 samples. It should be noted that the output signal s ^ is made up of 128 sample blocks in the time domain. It arises from the difference between the second block halves (that is, samples 129 to 256) of the microphone signal and the filtered compensation signal r ^. The projection P1 is complex, which requires 2 FFTs and converts the vector H 'into the vector H. The first half (samples 1 to 128) is cut out of the complex 256-point result vector of the backward transformation from the frequency to the time domain (IFFT) and the second half (samples 129 to 256) is set to zero. After using this rectangular window in the time domain, the transformation into the frequency domain is carried out again using FFT. The projection P2 is simple. It consists of the sectioning of the last 128 samples already described above, which again results in non-overlapping 128-sample blocks from overlapping 256-sample blocks. Finally, the projection P3 is also very simple, which in turn provides overlapping 256-sample blocks from non-overlapping 128-sample blocks of the output signal by prepending 128 zero values. The adaptation of the filter coefficients H ' _{L + 1} for a cycle L + 1 consists of the addition of a renewal vector ΔH' _L to the old coefficient vector H ' _L. This renewal is calculated from the product between the spectrum S ^ _{L of} the output signal and the conjugate complex spectrum X * _{L of} the reference signal - weighted with a spectral power standardization 2µ _L , ΔH ' L = 2µ L · X * L · S ^ L . For the purpose of this power standardization, the reciprocal of the smoothed reference power spectrum S _{xx, L} multiplied by a constant 2α must be calculated 2µ L = 2α / S xx, L , for which a recursive filter of the 1st order with a constant β is used S xx, L = β · | X L | 2nd + (1-β) · S xx, L-1 .

The operation of the LMS algorithm is significantly different from that Adaptation constant α and the smoothing constant β influenced. Buffer in recursion loops are with Sp designated.

The arrangement of the FLMS algorithm described so far allows Filter replicas with a maximum impulse response length of half a FFT length, so in the example 128 samples. Should longer impulse responses be compensated is the already known FLMS algorithm for one Extend sub-filter (Fig. 4a) to n sub-filters. A 3-part filter solution with an impulse response length of 3 * 128 = 384 samples has proven itself in radio signal suppression in the car proven with a voice input system (Fig. 4b). The one in 4a block designated B with the input signals X. and S ^ and the compensation spectrum R ^ as an output to be replaced by the extension shown in FIG. 4b. The spectrum X of the reference signal is stored in a buffer D delayed by 1 or 2 block lengths and the undelayed X1 and the two delayed spectra X2, X3 separately in with an extended projection P1 multiplied certain coefficient vectors H1, H2, H3. The coefficient vectors are formed analogously to Case of only a partial filter, whereby in K1, K2, K3 each associated reference spectrum with the spectrum S ^ of the output signal is linked. The effort is mainly by tripling the P1 projection considerably elevated. Additional space requirements will be necessary the spectra of the older one by 1 or 2 block lengths To provide reference signals X.

• Bei der zeitgleichen Spektralanalyse der Signale x und y ist nur eine einzige 256-Punkte-FFT mit geringem Zusatzaufwand für eine spektrale Separation notwendig. Man erzielt eine Einsparung von 1 FFT.
• Die hier mit P4 gekennzeichnete und neu definierte Projektion ist bis auf das verwendete Zeitfenster formal identisch mit der Projektion P1. Wie später gezeigt wird, läßt sich P4 durch eine relativ einfache Faltungsoperation im Frequenzbereich ersetzten, ohne daß eine merkliche Einbuße an Qualität in Kauf genommen werden muß. Man erzielt eine Einsparung von 2 FFT's.

In the case of the example of the task of suppressing the radio signal when voice input is given in the motor vehicle, it is advantageous not to output the output data in the time domain but in the frequency domain, since this enables an improved adaptation to a downstream noise suppression to be achieved. According to FIG. 5a, the FLMS algorithm with a partial filter already presented requires a total of 5 FFTs for an output signal in the time domain. If an FFT is connected downstream of the output, the effort for a frequency range output signal increases to 6 FFT's. The same number of FFTs initially results from an equivalent solution according to FIG. 5b. However, this variant has the following advantages:

• With the simultaneous spectral analysis of the signals x and y, only a single 256-point FFT is required with little additional effort for spectral separation. A saving of 1 FFT is achieved.
• The projection identified and redefined here with P4 is formally identical to projection P1 except for the time window used. As will be shown later, P4 can be replaced by a relatively simple convolution operation in the frequency domain without having to accept a noticeable loss in quality. A saving of 2 FFT's is achieved.

Figure 6 provides a more detailed block diagram of the FLMS algorithm with frequency domain output signal and allows a comparison with FIG. 3 again (time domain output). The filter adaptation has remained unchanged consisting of smoothing the spectral power, power normalization and coefficient renewal. They are new FFT in the microphone channel, the difference Y-R ^ in the frequency instead of in the time domain for the output formation, and finally the newly defined projection P4, which is only through the complementary time window of the projection P1 differs.

As a precursor to a preferred one described below See FIG. 7 for execution. The FLMS algorithm is shown with 3 sub-filters (384-sample impulse response), which has a sufficient suppression of the radio signal in the microphone channel of the speech input system. The Projections P1 and P4 are shown in simplified form. It is the additional effort already known from FIG. 4b in the form the memory P and the tripling of the projection P1 evident. In contrast to the 1-part filter solution according to Fig. 6 becomes the sum W of the current and the two in time previous reference power spectra on the Given the input of the recursive filter. The fact that at the filter output now practically 3 times the spectral smoothed Performance is available after the reciprocal by multiplying by the constant 6α. After the spectral performance standardization of the modified in P4 The filter adaptation is now the output spectrum S ^ for the 3 coefficient vectors of the 3 sub-filters separately carried out.

An example Z0 for the operation of the invention according to Figure 7 shows Figure 9. The input data has been synthesized generated. The reference signal X represents 100,000 samples a white Gaussian noise at a sample rate of fs = 12 kHz. The microphone signal Y was created by Convolution of this noise signal with a likewise constructed one 384-sample impulse response and the addition of one extremely weak speech signals. When listening to this in 9 signal y recorded above are the 10 spoken Digits just in color (because filtered) Recognizing noise. That transformed back into the time domain Output signal of the estimator frees up after a approx. 1 second (12000 samples) settling process very effective the speech input from the noise and delivers an undistorted but slightly reverberated speech signal S ^ (Fig. 9 below). The two parameters used were α = 0.05 and β = 0.5, values that are also in the later presented Examples have proven successful.

From the 129 sample-long partial coefficient vectors H1, H2, H3 of the 3 sub-filters according to Fig. 7 can now be used the resulting 3 * 128-sample impulse response at any time or the associated filter transfer function to calculate. 10 shows the 384 sample impulse response, how they look at the very end of the scene the digit "0" was spoken - results. It is a very exact image of the impulse response that is used for folding with white Gaussian noise and thus for synthetic generation of the signal was used micro. The associated Amount transfer function (Fig. 10 below) in the range between the frequencies 0 and fs / 2 = 6 kHz provide one with numerous narrow-band resonance peaks Low pass frequency response.

White noise as a reference input signal and filtered "colored" noise as a microphone input signal in the Meaning of the task, a replica of this filter to find the simplest case. Because the reference signal contains all frequency components by definition, the Filter adaptation is the fastest here. The additional additive Voice input in the microphone input signal - that is actual useful signal of the voice input system - provides for the (F) LMS algorithm represents a disturbance which the correct adaptation of the filter coefficients hindered. Different in other words: the system is only able to do this during pauses in speech Position the room acoustics of the vehicle interior (route Radio speakers to the microphone) and correctly thereby to compensate for the radio reproduction. In the example shown in FIG. 9 demonstrated above, this succeeds very good, since the microphone input is essentially off Noise and only a very small part Voice input exists.

On the other hand, this came from real measurements in the vehicle Reference signal tapped at the radio speaker terminals radio and that recorded by the microphone of the voice input system Signal micro of scene Z1. This microphone signal is shown in Fig. 11 above, consists of 100000 samples and therefore has a sampling frequency of 12 kHz a duration of approximately 8.3 seconds. It is about to fluent and relatively fast spoken language vehicle occupants sitting in the rear right while at the same time music with normal volume from the car radio speaker sounds. After applying the interference suppression measure 7 and conversion into the time range the output signal shown in Fig. 11 below. Of the Hearing test shows a clear elaboration of the language component or a remarkable one especially in the short language breaks Music suppression. Noticeable and disadvantageous is that the desired radio signal suppression in depends to a large extent on whether one is currently speaking or Not. The 384-sample impulse response determined again at the end of the scene with the associated transfer function is off Fig. 12 can be seen. A correct impulse response is to the to recognize typical zero samples (dead time) at the beginning, which depends on the duration of the direct sound from the radio speaker come to the microphone. From the existing ones strong disturbances at the beginning and at the end of the impulse response can therefore be concluded that the filter adaptation at this point extremely because of existing voice input is insufficient.

The embodiment described below with reference to FIG. 8 is based on the following basic idea: a suitable one Feature serves as an indicator along with a threshold for voice input. Falls below the characteristic the threshold, so this is a sign of missing Voice input. In this case - as stated above - A largely undisturbed filter adaptation. With voice input, the filter coefficient set is now used resorted to the immediately before the Threshold crossing - i.e. at the end of the previous one Speech pause - was saved. This saved Coefficients H10, H20, H30 usually provide a clear better radio signal compensation than that under the disturbing influence of voice input is constantly changing current coefficients H, H2, H3.

Fig. 8 shows an embodiment with a further improved FLMS processing with 3 partial filters. In addition to the already in Fig. 7 existing current filter coefficient vectors H1, H2, H3, which were continuously adopted to form the Output signal y-R were required, there is now an additional one Output signal (y-Ro) that is stored using Coefficients H10, H20, H30 is formed. The current coefficient sets H1, H2, H3 only provide missing speech input in the steady state usable compensation filter in the frequency domain, on the other hand provide inadequate filter properties for voice input, because the adaptation process in the control loop is constantly disturbed. If there is no voice input, i.e. The three switches are closed and of high filter quality the current coefficient sets are stored in the coefficient memory M1, M2, M3 written: H10 = H1, H20 = H2, H30 = H3. The outputs (y-Ro) and (y-Ra) are identical. Inserting Voice inputs open the 3 switches, whereby the last ones in the memories M1, M2, M3 Coefficients H10, H20, H30 no longer overwritten will and remain unchanged. This state in which the outputs (Y-Ro) and (Y-Ra) differ hold until a speech pause is detected again and the switches are closed.

The smoothed sum has become the speech pause feature fea all absolute values of the coefficient correction vectors ΔH1 ', ΔH2 ', ΔH3' proven (Fig. 8a). This size is zero or has small numerical values if there is none or only there is little need to change the coefficients. This is the case during breaks in speech, the control loop is practical steady. Disorders such as those caused by voice input - but also by movements of the vehicle occupants - have an increased need for readjustment result, which is characterized by correspondingly large numerical values noticeable with ΔH1 ', ΔH2', ΔH3 'and thus with the characteristic fea makes. A smoothing filter, for example, a recursive one 1st order low pass with the feat input on his Output the smoothed speech pause feature fea is available, which after comparison with a threshold value th the Switch for coefficient acceptance controls.

How the improved FLMS algorithm works Figure 8 demonstrates Figure 13. Above is the recorded one Signal y of scene Z1 (see FIG. 11 above) shown below the output signal obtained. Already the visual comparison of the output signals of FIGS. 13 and 11 the improved elaboration of the language passages. Of the comparative hearing test confirms this: also during the Voice input is much better at music rejection. The course of the speech pause feature and the constant Threshold over time (scaled here in FFT blocks) shows Fig. 14 above. In the fall below the threshold detected speech pauses (Fig. 14 below) takes place continuously the transfer of the coefficients into the memory as described instead of being there during voice input stored coefficients are available. The 384 sample impulse response measured at the end of the scene in FIG with associated amount transfer function 15 as the current impulse response (a) or current Transfer function (b) shown. In contrast to this due to speech input, the estimate from the current coefficient H1, H2, H3 is from the saved Coefficients H10, H20, H30 an impulse response (c) and a transfer function (d) of high quality can be calculated. The impulse response from the stored coefficients points the typical zero samples at the beginning, which are indicated by the Running time of the direct sound from the radio speaker to Voice input microphone. From the example readable dead time of approx. 40 samples the distance between the speaker and the microphone determine.

As already indicated above, the complex Projection P4 (IFFT, window on the right in the time domain, FFT) without noticeable loss of quality due to a relatively simple Replacing convolution in the frequency domain, resulting in 2 FFT's be saved. Consider Fig. 16. In one The first step is the "right-hand" 128-sample rectangular window in the time domain (Fig. 16a) with the ideal projection replaced by a 128-sample Hamming window (Fig. 16b). This has the advantage over the rectangular window of a significantly narrower spectrum. Like Fig. 17 shows, the real part of the spectrum exists in the rectangular window from a single line (DC component), while the middle antisymmetric imaginary part spectrum from many lines slowly descending towards the outside with alternating lines Zeros exist. In contrast, it is limited the complex spectrum of the Hamming window (Fig. 18) a total of 7 lines, of which only in the symmetrical real part 3 and in the antisymmetric imaginary part only 4 values of Are zero different. All parts outside are negligible. This special property the Hamming window advantageously makes it possible to replace the multiplication in the time domain (Fig. 16b) by convolution with the associated 7-sample spectrum in the Frequency range and thus to save an IFFT and an FFT (Fig. 16c).

In principle, the projection P1 can of course also be used (IFFT - left-sided rectangular window - FFT) replace with a corresponding convolution operation in the frequency domain the conjugate complex 7-line spectrum. Experiments have shown, however, savings at this point be bought with a significant deterioration in the Transient response. Effortless solutions can be nevertheless achieve by following in the LMS algorithm 8 the 3 projections P1 not simultaneously in one 256-sample input data block must be processed. The with 128-sample overlapping input data blocks of length 256 are numbered starting at "1" sketched in Fig. 19a. So it is e.g. possible at modulo-3-counting of the input data blocks the 3 sub-filter projections not in parallel (Fig. 19b) but sequentially to be calculated in successive blocks Fig. 19. Thereby are not 6 per data block with ideal projection P1 but only 2 FFT's necessary. It has shown, that the compensation of the radio signal is also sufficient works if the distances between those to be calculated Sub-filter projections can be chosen even larger. If you count the blocks e.g. modulo 6, that's just To calculate a projection in every second block (Fig. 19d). Even a reduction of four Blocks between two consecutive P1 calculations using modulo-12 counting still leads to useful results (Fig. 19e).

The performance of the FLMS algorithm with 3 sub-filters according to block circuit Fig. 8 and a sequential Calculation of the ideal projection P1 according to the time grid 19e and the projection P2 by means of convolution in the frequency domain (Fig. 16c) with a complex 7-line spectrum (Fig. 18) will be demonstrated using 3 measurement scenes.

The first of these scenes Z2 includes voice input from Digits, the radio speaker almost white Noise emits at a relatively high volume. The associated 100000 sample microphone signal is in Fig. 20 above, the extracted output signal is shown in Fig. 20 below. A clear release of noise from the output signal compared to the microphone input is made by listening comparison firmly. The time course of the speech pause feature is up along with the constant threshold th Fig. 21 mapped and the derived language breaks or the assigned switch positions in Fig. 21 below. Finally, FIG. 22 shows the in an analogous manner to FIG impulse response (a) and transfer function found at the end of the scene (b) based on the current coefficients and the corresponding sizes (c), (d) based on the Speech pause setting. It is clearly recognizable that the current impulse response found at the end of the scene Speech input represents disturbed result while the out the impulse response from the last speech pause stored coefficient sets has a high quality.

The first 100000 samples of a measuring scene Z3 with POP music on the radio and fluent to quickly spoken language of the The person sitting on the right rear is in the form of a microphone signal y recorded in Fig. 23 above. After about 10,000 samples (0.83 s) the radio signal is suppressed usably (Fig. 23 below). Even in the last third of this POP music suppression remains when voice input begins effectively preserved, making speech intelligibility noticeably improved here compared to the microphone signal becomes. After a long pause in speech, it comes because of the subsequent one non-stop voice input to one Falling below threshold (Fig. 24). This is why the impulse response recorded at the bottom of the scene in Fig. 25 based on the stored coefficients Relatively obsolete in time because it was already approx was current (215 blocks * 10.7 ms). Again, the current impulse response (Fig. 25 above) strong from the Interference arising from voice input. Like a comparison with the similar scene Z1 according to FIGS. 11 to 15, is the quality despite the greatly reduced computing effort the interference exemption remains high.

The last scene Z4 according to FIG. 26 was without speech input finally creates and is supposed to again the music suppression properties of the FLMS algorithm described to demonstrate. After about 18000 samples or 1.5 s - how seen from Fig. 26 below - the music effectively suppressed. This property is unchanged until the end of the scene Maintain quality. Fig. 27 shows that the speech pause size fea predominantly below the threshold th remains. The times when the saved Coefficients are therefore used only very much short. Impulse response and transfer function from current Coefficients are therefore essentially the same Course from speech pause coefficients identical.

Claims (12)

Method for the elimination of interference of a microphone signal from Share a source signal that as a reference signal (x) is present and after passing through a transmission link with a priori unknown transfer function (G) in the microphone signal as a disturbance signal (r) a speech signal (s) superimposed, by adaptive simulation of the interference signal and Compensation of the actual and the simulated interference signal in an output signal, the microphone signal also transformed into the frequency domain, the compensation made on signals in the frequency domain and the output signal in the frequency domain for adaptation the replica with that in the frequency domain Reference signal is linked. A method according to claim 1, characterized in that transformed the output signal spectrum into the time domain, the time signal by prefixing zero values brought twice the length, transformed back into the frequency domain and the simulation of the transfer function is taken as a basis. A method according to claim 1, characterized in that the output signal spectrum with the spectrum of a Hamming time window is folded and the replica of the transfer function is taken as a basis. Method according to one of claims 1 to 3, characterized in that that to simulate the interference signal component adaptive filter function of a replica filter the reference signal is applied. A method according to claim 4, characterized in that the filter function is specified by a coefficient vector whose coefficients are adjusted adaptively. A method according to claim 4 or 5, characterized in that that the appearance of a speech signal component in the microphone signal is detected and when a speech signal occurs the one set before the speech signal appeared Maintain filter function to generate the output signal becomes. A method according to claim 6, characterized in that adaptive tracking even when a speech signal is detected a current filter function in addition to education of the output signal is continued. A method according to claim 7, characterized in that the appearance of a speech signal from a change in current filter function is detected. A method according to claim 8, characterized in that the change of the current filter function for the detection the occurrence of a speech signal smoothed in time becomes. Method according to one of claims 4 to 9, characterized characterized in that the filter function into several sub-filter functions to successive sections of an overall impulse response all sub-filters are split and on reference signal spectra at time-staggered time segments of the segmented reference time signal applied becomes. A method according to claim 10, characterized in that the adaptation of the filter function for the partial filters in parallel is carried out. A method according to claim 10, characterized in that the adaptation of the filter function for the individual sub-filters is carried out sequentially.

Images