EP0615226A2 - Method for noise reduction in disturbed voice drannels - Google Patents

Abstract

The invention relates to a method which can be used not only for removing noise, for example in automatic speech recognition, but also for improving the speech quality for humans, for example hands-free speech on the car phone. The noise reduction is carried out in a two-channel or multiple-channel manner, in such a way that the temporal and spatial-acoustic signal properties of speech and noise are used systematically in a stepwise fashion. <IMAGE>

Description

The invention relates to a method according to the preamble of patent claim 1.

Such a method is used in automatic speech recognition or in hands-free systems to improve speech quality, e.g. in offices or in a motor vehicle.

Disrupted speech is easier to grasp if it is recorded with two or more channels. Language and interference should be present in each channel. The multi-channel signals are processed with digital signal processing.

In the case of multi-channel systems, the transit time difference of the useful signal in the individual channels must first be determined. It will later be possible to merge the individual channels into one channel in the correct phase.

Systems with 2 channels are of particular interest because they can already be used to resolve a spatial sound field in individual directions, but the computational effort remains tolerable.

If the direction from which the sound event of interest arrives is known, an acoustic directional lobe is set for this event.

The noise reduction is first carried out in each individual channel. Since the noise reduction does not work correctly, distortions and artificial insertions (e.g. "musical tones") can occur. When merging the individual processed channels, these errors are averaged and thus reduced.

The sum signal is then post-processed using the cross-correlation of the signals in the individual channels. It is assumed that interference or reverberation is less correlated than the useful signal of the channels.

A method for merging 2 disturbed speech channels is described in the publication "Multimicrophone signal-processing technique to remove room reverberation from speech signals" by Allen, Berkley and Blauert (J: Acoust. Soc. Am., Vol. 62, No. 4, October 1977) and out "Noise Suppression Signal Processing Using 2-Point Received Signals" by Kaneda and Tohyame (Electronics and Communication in Japan, Vol. 67-A, No. 12, 1984). The first method is intended for the dehalling of speech signals and does not use a real phase compensation of the useful signal and the dehalling with noise reduction is only carried out in a post-processing stage. The second method uses a simple linear phase compensation of the channels, but here too the noise is only reduced in the post-processing stage.

The invention is therefore based on the object of specifying a method for noise reduction in which the noise reduction is carried out in several stages and a significant improvement in the speech quality is achieved.

The object is achieved by the features specified in the characterizing part of patent claim 1. Advantageous refinements and / or further developments can be found in the subclaims.

• 1.) räumliche Eigenschaft der Schallfelder:
  • a) Dämpfung von Punktstörquellen
    Mit digitalen Richtungsfiltern am Eingang der Kanäle wird zusammen mit der Phasenschätzung eine akustische Richtkeule auf den Sprecher ausgerichtet. Für die Phasenschätzung wird das in der unveröffentlichten deutschen Patentanmeldung P 42 43 831 beschriebene Verfahren verwendet. Es ist robust gegenüber Störungen und benötigt nur einen geringen Rechenaufwand. Die Richtungsfilter sind fest eingestellt. Es wird angenommen, daß der Sprecher sich relativ nahe an den Mikrofonen befindet (Abstand ≦ 1m) und sich nur in einem beschränkten Bereich bewegt. Instationäre und stationäre Punkt-Störquellen werden durch diese räumliche Auswertung gedämpft.
  • b) Dämpfung von diffusen Störquellen
    In der Nachverarbeitung werden mit Hilfe der Kreuzkorrelation die diffusen Stör- und Hallanteile gedämpft.
• 2.) zeitliche Signaleigenschaften:
  Die spektrale Subtraktion schätzt die Störung in den Sprachpausen und führt eine betragsmäßige Subtraktion im Spektralbereich durch. Hier werden die zeitlich stationären Störanteile gedämpft.
• 3.) Mittelung der Kanäle (Addition):
  Durch die räumliche Trennung der Aufnahmekanäle (Mikrofone in einem bestimmten Abstand) treten Fehler der spektralen Subtraktion (Verzerrung und "musical tones") in den einzelnen Kanälen z.T. zeitlich zufällig auf. Eine Mittelung der Kanäle vermindert diesen Fehler.

With the method according to the invention, the spatial and the temporal properties of the useful signal and the disturbance are used systematically:

• 1.) spatial property of the sound fields:
  • a) Attenuation of point sources of interference
    With digital directional filters at the entrance of the channels, an acoustic directional lobe is aligned with the speaker together with the phase estimation. The method described in the unpublished German patent application P 42 43 831 is used for the phase estimation. It is robust against interference and requires little computing effort. The directional filters are fixed. It is assumed that the speaker is relatively close to the microphones (distance ≦ 1m) and only moves in a limited area. Transient and stationary point sources of interference are dampened by this spatial evaluation.
  • b) Attenuation of diffuse sources of interference
    In post-processing, the cross-correlation dampens the diffuse interference and Hall components.
• 2.) Temporal signal properties:
  The spectral subtraction estimates the disturbance in the speech pauses and carries out an amount-based subtraction in the spectral range. Here the temporally stationary disturbance components are damped.
• 3.) Averaging the channels (addition):
  Due to the spatial separation of the recording channels (microphones at a certain distance), spectral subtraction errors (distortion and "musical tones") in the individual channels sometimes occur at random in time. Averaging the channels reduces this error.

FIG. 1

zeigt ein Blockdiagramm des gesamten Verfahrens.

FIG. 2

zeigt einen Vergleich der gemittelten Ausgangsleistungen Z verschiedener Verfahren mit der Leistung des Original-Geräuschsignals (Beispiel: Mikrofonabstand 12cm, Fahrzeug mit 140km/h). Es wird die zunehmende Geräuschreduktion gezeigt wenn die Verarbeitung mit einem Kanal, mit zwei Kanälen und mit zwei Kanälen mit Nachverarbeitung durchgeführt wird.

The invention is explained in more detail using exemplary embodiments and reference to schematic drawings.

FIG. 1

shows a block diagram of the entire method.

FIG. 2nd

shows a comparison of the average output powers Z of different methods with the power the original sound signal (example: microphone distance 12cm, vehicle at 140km / h). The increasing noise reduction is shown when the processing is carried out with one channel, with two channels and with two channels with post-processing.

The microphone signals x and y are transformed into the frequency range (FFT, Fast Fourier Transformation). The segments are half overlapped and weighted with a Hanning window. The segments are each N values long and are expanded by an additional N zeros. The transformation length is chosen to be 2N = 512, for example. The transformed segments X _l (i) and Y _l (i) result. The output signal z results after inverse transformation and the overlap of the segments. l denotes the block index of the segments, i the discrete frequency (i = 0,1,2 ..., 2N-1). The sampling rate of the signals x and y is, for example, 12 kHz.

$${\mathit{\text{β}}}_{\mathit{\text{l}}}\text{=}\mathit{\text{gβ₀}}$$

   mit $${\mathit{\text{g}}}_{\mathit{\text{l}}}\text{=}\frac{\text{2}{\mathit{\text{L}}}_{\mathit{\text{l}}\text{-1}}}{{\mathit{\text{L}}}_{\mathit{\text{l}}\text{-1}}\text{+}{\mathit{\text{K}}}_{\mathit{\text{l}}}}$$

$${\mathit{\text{L}}}_{\mathit{\text{l}}}\text{= (1 -}{\mathit{\text{β}}}_{\mathit{\text{l}}}\text{)}{\mathit{\text{L}}}_{\mathit{\text{l}}\text{-1}}\text{+}{\mathit{\text{β}}}_{\mathit{\text{l}}}{\mathit{\text{K}}}_{\mathit{\text{l}}}$$

$${\widehat{\mathit{\text{S}}}}_{\mathit{\text{nn,l}}}\text{(}\mathit{\text{i}}\text{) = (1 -}{\mathit{\text{β}}}_{\mathit{\text{l}}}\text{)}{\widehat{\mathit{\text{S}}}}_{\mathit{\text{nn,l}}}\text{₋₁(}\mathit{\text{i}}\text{) +}{\mathit{\text{β}}}_{\mathit{\text{l}}}\text{|}{\mathit{\text{X}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)|²}$$

In the frequency domain, the long-term mean of the magnitude spectrum is subtracted (spectral subtraction H _SPS ). The short-term average K and the long-term average L are used to calculate a first adaptive smoothing constant β. The interference spectrum S _nn (i) is estimated with β. This adaptive smoothing constant replaces the otherwise common speech pause detector. l denotes the block index, i the discrete frequency. For example, β _o = 0.03 is used as the smoothing constant β _o .

$${\mathit{\text{β}}}_{\mathit{\text{l}}}\text{=}\mathit{\text{gβ₀}}$$

With $${\mathit{\text{G}}}_{\mathit{\text{l}}}\text{=}\frac{\text{2nd}{\mathit{\text{L}}}_{\mathit{\text{l}}\text{-1}}}{{\mathit{\text{L}}}_{\mathit{\text{l}}\text{-1}}\text{+}{\mathit{\text{K}}}_{\mathit{\text{l}}}}$$

$${\mathit{\text{L}}}_{\mathit{\text{l}}}\text{= (1 -}{\mathit{\text{β}}}_{\mathit{\text{l}}}\text{)}{\mathit{\text{L}}}_{\mathit{\text{l}}\text{-1}}\text{+}{\mathit{\text{β}}}_{\mathit{\text{l}}}{\mathit{\text{K}}}_{\mathit{\text{l}}}$$

$${\widehat{\mathit{\text{S}}}}_{\mathit{\text{nn, l}}}\text{(}\mathit{\text{i}}\text{) = (1 -}{\mathit{\text{β}}}_{\mathit{\text{l}}}\text{)}{\widehat{\mathit{\text{S}}}}_{\mathit{\text{nn, l}}}\text{₋₁ (}\mathit{\text{i}}\text{) +}{\mathit{\text{β}}}_{\mathit{\text{l}}}\text{|}{\mathit{\text{X}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) | ²}$$

$${\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) = (1 -}\frac{{\widehat{\mathit{\text{S}}}}_{\mathit{\text{nn,l}}}\text{(}\mathit{\text{i}}\text{)}}{\text{|}{\mathit{\text{X}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)|²}}\text{)}{\mathit{\text{X}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)}$$

The interference spectrum is normalized and subtracted. $$\text{|}{\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) | = |}{\mathit{\text{X}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) | -}\frac{{\widehat{\mathit{\text{S}}}}_{\mathit{\text{nn, l}}}\text{(}\mathit{\text{i}}\text{)}}{\text{|}{\mathit{\text{X}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) |}}$$

$${\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) = (1 -}\frac{{\widehat{\mathit{\text{S}}}}_{\mathit{\text{nn, l}}}\text{(}\mathit{\text{i}}\text{)}}{\text{|}{\mathit{\text{X}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) | ²}}\text{)}{\mathit{\text{X}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)}$$

$${\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) = ƒ₀}{\mathit{\text{S}}}_{\mathit{\text{nn,l}}}\text{(}\mathit{\text{i}}\text{) ; sonst}$$

A modified form results with: $${\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) = (1 - α}\sqrt{\frac{{\widehat{\mathit{\text{S}}}}_{\mathit{\text{nn, l}}}\text{(}\mathit{\text{i}}\text{)}}{{\mathit{\text{S}}}_{\mathit{\text{xx, l}}}\text{(}\mathit{\text{i}}\text{)}}}\text{)}{\mathit{\text{X}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{); for (1 - α}\sqrt{\frac{{\widehat{\mathit{\text{S}}}}_{\mathit{\text{nn, l}}}\text{(}\mathit{\text{i}}\text{)}}{{\mathit{\text{S}}}_{\mathit{\text{xx, l}}}\text{(}\mathit{\text{i}}\text{)}}}\text{)}{\mathit{\text{X}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) <ƒ₀}{\mathit{\text{S}}}_{\mathit{\text{nn, l}}}\text{(}\mathit{\text{i}}\text{)}$$

$${\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) = ƒ₀}{\mathit{\text{S}}}_{\mathit{\text{nn, l}}}\text{(}\mathit{\text{i}}\text{); otherwise}$$

$${\text{α}}_{\mathit{\text{l}}}\text{= 2 -}{\mathit{\text{g}}}_{\mathit{\text{l}}}\text{; für 0.5 < 2 -}{\mathit{\text{g}}}_{\mathit{\text{l}}}\text{< 2.0}$$

$${\text{α}}_{\mathit{\text{l}}}\text{= 0.5 ; für 0.5 > 2 -}{\mathit{\text{g}}}_{\mathit{\text{l}}}$$

$${\text{α}}_{\mathit{\text{l}}}\text{= 2 ; für 2 < 2 -}{\mathit{\text{g}}}_{\mathit{\text{l}}}$$

f_o wird als "spectral floor" bezeichnet. Es wird ein Teil des Hintergrundgeräuschs zugelassen, um einen natürlich Höreindruck zu erzeugen und um einen Teil der "musical tones" zu maskieren. α ist ein Überschätzfaktor für das Geräusch und dient der weiteren Reduzierung des Restgeräuschs. Für diese Werte kann z.B. f_o = 0.2 und α = 1.5 gewählt werden.The following applies to the power density S _{xx, l of} a channel: $${\widehat{\mathit{\text{S}}}}_{\mathit{\text{xx, l}}}\text{(}\mathit{\text{i}}{\text{) = (1 - α}}_{\mathit{\text{l}}}\text{)}{\widehat{\mathit{\text{S}}}}_{\mathit{\text{xx, l}}\text{-}}\text{₁ (}\mathit{\text{i}}{\text{) + α}}_{\text{l}}\text{|}{\mathit{\text{X}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) | ²}$$

$${\text{α}}_{\mathit{\text{l}}}\text{= 2 -}{\mathit{\text{G}}}_{\mathit{\text{l}}}\text{; for 0.5 <2 -}{\mathit{\text{G}}}_{\mathit{\text{l}}}\text{<2.0}$$

$${\text{α}}_{\mathit{\text{l}}}\text{= 0.5; for 0.5> 2 -}{\mathit{\text{G}}}_{\mathit{\text{l}}}$$

$${\text{α}}_{\mathit{\text{l}}}\text{= 2; for 2 <2 -}{\mathit{\text{G}}}_{\mathit{\text{l}}}$$

f _o is referred to as "spectral floor". Part of the background noise is allowed to create a natural auditory impression and to mask part of the "musical tones". α is an overestimation factor for the noise and serves to further reduce the residual noise. For these values, for example, f _o = 0.2 and α = 1.5 can be selected.

In contrast to the known forms of spectral subtraction, a second adaptive smoothing with α is used to reduce a further part of the "musical tones" by smoothing the power density S _xx little during speech and strongly smoothing during pause.

The corresponding equations apply to the second channel Y.

$${\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) :=}{\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)(cos(}\mathit{\text{i}}\text{* φ) +}\mathit{\text{j}}\text{sin(}\mathit{\text{i}}\text{* φ))}$$

The method specified in the unpublished patent application P 42 43 831 is used to calculate the linear phase shift between useful parts in the channels. This method fits seamlessly into the noise reduction method according to the invention. The phase shift is estimated from a selected number of the maximums of the cross power density and the phase correction is achieved by multiplication in the frequency domain with the all-pass function H _ALLP . $${\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{): =}{\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}{\text{)H}}_{\mathit{\text{ALLP, l}}}$$

$${\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{): =}{\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) (cos (}\mathit{\text{i}}\text{* φ) +}\mathit{\text{j}}\text{sin (}\mathit{\text{i}}\text{* φ))}$$

If there are more than two channels, the phase correction is carried out for the other channel. The first channel serves as a reference.

The directional filters for the channels are calculated using a "beamforming process". Various cases can be considered as noise. Different directional filters H _R result depending on the noise _situation . A set of these filters is selected, however, if the system status is known in later operation, it is possible to switch to a specific set or the filters can be continuously adapted. For example, the "Beamforming method" is the Frost gradient method ("An Algorithm for Linearly Constrained Adaptive Array Processing" Proc. IEEE, Vol. 60, No. 8, 1972) or according to Sondhi and Elko ("Adaptive Optimization of Microphone Arrays under a Nonlinear Contraint" Int. Conf. on ASSP, Tokyo, 1096, pp. 981-984).

The multiplication for directional filtering results in the frequency domain: $${\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{): =}{\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)}{\mathit{\text{H}}}_{\mathit{\text{R}}}\text{(}\mathit{\text{i}}\text{)}$$

The addition of the channels with the directional filters results in the overall directional characteristic and the output signal $${\mathit{\text{Z.}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) =}{\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{) +}{\widehat{\mathit{\text{Y}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)}$$

In addition, the addition of the channels leads to an averaging and thus reduction of the statistical errors of the spectral subtraction.

The cross power density of the two channels is then calculated using a smoothing constant (eg γ = 0.3). $${\mathit{\text{S}}}_{\mathit{\text{xy, l}}}\text{(}\mathit{\text{i}}\text{) = (1 - γ)}{\mathit{\text{S}}}_{\mathit{\text{xy, l}}}\text{₋₁ (}\mathit{\text{i}}\text{) + γ}{\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)}{\widehat{\mathit{\text{Y}}}}_{\mathit{\text{l}}}\text{* (}\mathit{\text{i}}\text{)}$$

$${\mathit{\text{H}}}_{\mathit{\text{KKF,l}}}\text{(}\mathit{\text{i}}\text{) = 0.3; sonst}$$

mit $${\mathit{\text{S}}}_{\mathit{\text{xx,l}}}\text{(}\mathit{\text{i}}\text{) = (1 - γ)}{\mathit{\text{S}}}_{\mathit{\text{xx,l}}}\text{₋₁(}\mathit{\text{i}}\text{) + γ}{\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)}{\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}^{\text{*}}\text{(}\mathit{\text{i}}\text{)}$$

$${\mathit{\text{S}}}_{\mathit{\text{yy,l}}}\text{(}\mathit{\text{i}}\text{) = (1 - γ)}{\mathit{\text{S}}}_{\mathit{\text{yy,l}}\text{-1}}\text{(}\mathit{\text{i}}\text{) + γ}{\widehat{\mathit{\text{Y}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)}{\widehat{\mathit{\text{Y}}}}_{\mathit{\text{l}}}^{\text{*}}\text{(}\mathit{\text{i}}\text{)}$$

The cross power density S _xy is standardized with the sum of the power densities S _xx, S _{yy of} the individual channels. The result is a modified coherence function: $${\mathit{\text{H}}}_{\mathit{\text{KKF, l}}}\text{(}\mathit{\text{i}}\text{) =}\frac{{\mathit{\text{S}}}_{\mathit{\text{xy, l}}}\text{(}\mathit{\text{i}}\text{)}}{{\mathit{\text{S}}}_{\mathit{\text{xx, l}}}\text{(}\mathit{\text{i}}\text{) +}{\mathit{\text{S}}}_{\mathit{\text{yy, l}}}\text{(}\mathit{\text{i}}\text{)}}\text{; For}\frac{{\mathit{\text{S}}}_{\mathit{\text{xy, l}}}\text{(}\mathit{\text{i}}\text{)}}{{\mathit{\text{S}}}_{\mathit{\text{xx, l}}}\text{(}\mathit{\text{i}}\text{) +}{\mathit{\text{S}}}_{\mathit{\text{yy, l}}}\text{(}\mathit{\text{i}}\text{)}}\text{> 0.3}$$

$${\mathit{\text{H}}}_{\mathit{\text{KKF, l}}}\text{(}\mathit{\text{i}}\text{) = 0.3; otherwise}$$

With $${\mathit{\text{S}}}_{\mathit{\text{xx, l}}}\text{(}\mathit{\text{i}}\text{) = (1 - γ)}{\mathit{\text{S}}}_{\mathit{\text{xx, l}}}\text{₋₁ (}\mathit{\text{i}}\text{) + γ}{\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)}{\widehat{\mathit{\text{X}}}}_{\mathit{\text{l}}}^{\text{*}}\text{(}\mathit{\text{i}}\text{)}$$

$${\mathit{\text{S}}}_{\mathit{\text{yy, l}}}\text{(}\mathit{\text{i}}\text{) = (1 - γ)}{\mathit{\text{S}}}_{\mathit{\text{yy, l}}\text{-1}}\text{(}\mathit{\text{i}}\text{) + γ}{\widehat{\mathit{\text{Y}}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)}{\widehat{\mathit{\text{Y}}}}_{\mathit{\text{l}}}^{\text{*}}\text{(}\mathit{\text{i}}\text{)}$$

The following applies to the output signal Z: $${\mathit{\text{Z.}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{): =}{\mathit{\text{Z.}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)}{\mathit{\text{H}}}_{\mathit{\text{KKF, l}}}\text{(}\mathit{\text{i}}\text{)}$$

If directional filters are used according to the Sondhi and Elko method, an inverse filter for frequency response correction is required. This filter is used to raise the lower frequencies because the frequency response of the directional filters (for the desired direction, direction of the speaker) leads to a reduction in these frequencies. This filter H _INV can be approximated in a simple manner from the calculated frequency response. $${\mathit{\text{Z.}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{): =}{\mathit{\text{Z.}}}_{\mathit{\text{l}}}\text{(}\mathit{\text{i}}\text{)}{\mathit{\text{H}}}_{\mathit{\text{INV, l}}}\text{(}\mathit{\text{i}}\text{)}$$

If the adaptation is carried out using the Frost method, an inverse filter is not necessary because the Frequency response in the direction of the speaker has a constant value of 1.

The method according to the invention is not limited to systems with two channels, but can be used for multi-channel systems (3 and more channels).

Claims (4)

Method for noise reduction of at least two disrupted voice channels, the disrupted voice channels being combined to form an output channel, characterized in that that a swiveling, acoustic directional lobe is generated for the individual channels by means of digital directional filters and a linear phase estimate, which follows the speaker movement and thereby dampens the spatial interference sources, - that the interference is estimated in the individual channels during the pauses in speech and that the temporally stationary sources of interference are damped by spectral subtraction, - That the individual speech channels are then added, and thereby the statistical disturbances of the spectral subtraction are averaged, and - That the sum signal is post-processed with a modified coherence function and thereby the diffuse interference and Hall components are damped. A method according to claim 1, characterized in that the spectral subtraction is carried out with two adaptive smoothing constants α, β, - That the interference spectrum S _{nn is} estimated with the first adaptive smoothing constant β, and - That with the second adaptive smoothing constant α, the power density S _{xx of} the individual channels is greatly smoothed during speech pauses and little smoothed during speech. Method according to claim 1, characterized in that the linear phase shift of at least two signals is determined over a certain number of maxima of the cross power density in the frequency domain. Method according to Claim 1, characterized in that the phase correction, the directional filtering and any necessary inverse filtering are carried out in the frequency domain.

Images