EP1251493B1 - Method for noise reduction with self-adjusting spurious frequency - Google Patents

Abstract

The signals are processed together in pairs. Only one of the processed signals is subjected to a spectral subtraction, and combined with the other signal to form a difference signal. The primary signal may be connected as a differential array of two channels (1,2), or as a sum and difference signal of two channels.

Description

The invention relates to a method for noise reduction according to the preamble of patent claim 1.
A frequently used method for noise reduction of a disturbed useful signal, eg a speech signal, music signal etc. is the spectral subtraction. Advantage of the spectral subtraction is the low complexity and that the disturbed useful signal is needed only in one variant (only 1 channel). Disadvantage is the signal delay (due to the block processing in the spectral range), the limited maximum achievable noise reduction and the difficulty to compensate for transient noises. Stationary noise can be reduced, for example, by 12dB with good voice quality.

If a higher noise reduction or better voice quality is required, several recording channels are required. For example, microphone arrays are used. Of the various microphone arrays, such are for many practical applications particularly interesting, which make do with small geometrical dimensions for the microphone arrangement. Small differential microphone arrays (also called super-directive arrays) are formed and an adaptive variant of this microphone arrangement, whereby the LMS (least mean square) algorithm is used for the adaptation. In the adaptive form of this array, two microphones are time-compensated subtracted in two ways such that a virtual microphone with a cardioid polar pattern is turned away from the speaker and a "virtual" microphone with a kidney-shaped characteristic faces away from the speaker. The delay compensation corresponds to the time that the sound needs for the distance between the two microphones, eg 1.5cm. This results in a "back-to-back" kidney-shaped directional characteristic. The speaker-directed microphone is the primary signal for the adaptive filter, and the oppositely-directed microphone is the reference signal of the interference.

Figure 1 shows an adaptive arrangement for a beamformer. The runtime compensation with an all-pass ALL is realized by shifting by whole samples. The combination of two single microphones with omnidirectional characteristic described above results in a cardioid polar pattern characteristic for the speaker and an oppositely directed cardioid polar pattern characteristic as interference reference. The adaptive filter H 1 is adapted in the time domain using the LMS (least mean square) algorithm. A low-pass filter TP at the system output raises low frequency components, which are attenuated during the formation of the cardioid polar pattern.
The arrangement of the microphones M in succession according to FIG. 1 is referred to as an "end fire array", in contrast the arrangement of the microphones is designated side by side with a "broadside array".

Figure 2 shows an arrangement for a "broad side array" of two microphones in the distance, wherein the two microphone signals are preprocessed with the aid of spectral subtraction (SPS). A runtime compensation with the all-pass all between both channels is carried out and serves to compensate for movements of the speaker. The sum of the two pre-processed microphone signals forms the primary input and the difference the reference input for an adaptive filter H 1. The adaptive filter in this arrangement with sum and difference input is also referred to as 'generalized sidelobe canceller'. The adaptation takes place with the LMS algorithm, whereby the implementation of the LMS in the frequency domain takes place. A post-processing of the microphone signals is performed with a modified cross-correlation function in the frequency domain. The basic structure with spectral preprocessing by means of SPS, beam shaping and post-processing is described in the patent EP 0615226B1, wherein a precise specification of the beamformer has not been made.

Figure 3 shows an overview of circuitry of microphones for forming the directional characteristics for two microphones. The two individual microphones themselves can already have a kidney-shaped characteristic or the so-called spherical characteristics. "ALL" refers to a passpipe all-pass. 'Gain' is a gain equalization between both channels which is required in practice to equalize the sensitivity of the microphone capsules.

The Einsprechrichtung in the polar diagrams of the directional characteristics is 90 °. The first 3 arrangements a, b and c are suitable as a voice channel, since there is a maximum at 90 ° and an attenuation is present for the other directions. Arrangement a and b lead to the same directional characteristic. The arrangements a, b are referred to as a sum or difference array and arrangement c as a differential array.
The arrangements d and e have a zero at 90 ° in the polar diagram and are therefore suitable as a fault reference. The zero point at 90 ° in the polar diagram is necessary so that no speech components get into the reference channel. Speech components in the reference channel lead to partial compensation of the language.
Under ideal conditions, according to arrangements d and e, for the disturbance reference a zero will be set towards the speaker. However, in practical applications this will not be the case. The consequence is that speech components are treated as interference signals and thus removed from the actual speech signal.

Beamformers are usually adapted only in the speech pauses, in order to allow no adaptation to speech components. Nevertheless, even in this case existing speech components in the reference are compensated since they are always superimposed on the noise.

Another approach is to equalize the gain of channels so that, ideally, they are subtracted to produce a zero. This is necessary because Microphones from mass production identify tolerances. In the arrangements of Figure 3, this is taken into account with the function block 'Gain', which compensates for different microphone sensitivities.

In applications, despite the sensitivity compensation with 'Gain', no zero point is set for the speech signal in the reference. Only under the condition that the microphone is operated in the acoustic free field (without reflections), the voice components can be fully compensated. Real applications have due to reflections a certain amount of sound from different directions, which does not give rise to a zero for the speech signal. In arrangements according to FIG. 1 or FIG. 2, there will always be a certain proportion of speech in the reference signal of the beamformer, which leads to speech distortions.

The present invention is therefore based on the object to provide a method for noise reduction, with the crosstalk of the useful signal is minmiert in the interference reference signal.

The invention is specified in claim 1. Advantageous embodiments and further developments can be found in the dependent claims.

The invention has the advantage that significantly fewer useful signal components, e.g. Voice components are present in the interference reference signal than with the previous methods. The elimination of the disturbing speech components is thus under real conditions with reflections of the speech signal in real spaces such as e.g. in the vehicle possible.

The invention assumes that a one-sided spectral subtraction is performed to form the interference reference signal. It is essential that the spectral subtraction takes place to form a reference signal only on one channel, which is referred to as 'one-sided'. The one channel thus contains useful and interference signals, the second channel after the spectral subtraction contains only useful signals. During the subsequent subtraction of the two channels, the useful part is subtracted and the fault remains. This difference is the disturbance reference signal.

If e.g. Microphones used to receive speech signals, the speech signals are processed so that the interference reference signal has a zero to the speaker in the form of a kidney-shaped or an eight-shaped characteristic. The unilateral spectral subtraction leads to a self-regulating control of the characteristic, such that the zero occurs only in voice activity. In speech pauses, the one-sided spectral subtraction results in nothing or only a small signal being subtracted and thus approximately the characteristic of the single microphone (e.g., kidney or bullet) available for the perturbation.

The ideal zero for the speech signal in the reference is only achieved with an ideal spectral subtraction in the acoustic free field. An ideal spectral subtraction gives the undisturbed speech signal as an output signal and would then make any further processing unnecessary. The spectral subtraction in practice gives only a good approximation of the speech signal with noise residues in the speech pauses. Since the one-sided spectral subtraction is used in addition to the microphone zero point, the speech components of the reference reduce significantly.

und $$\mathrm{Y}\left(\mathrm{i}\right)=\mathrm{W}\left(\mathrm{i}\right)\cdot \mathrm{X}\left(\mathrm{i}\right);$$

The residual noise of the spectral subtraction in speech pauses is set with a parameter, the "spectral floor". The spectral floor b is the minimum value of a filter coefficient W of the spectral subtraction at each frequency index i. The output signal Y (i) is obtained by multiplying the filter coefficients W (i) by the input value X (i): $$\mathrm{W}\left(\mathrm{i}\right):=\mathrm{Max}\left(\mathrm{W}\left(\mathrm{i}\right).\mathrm{b}\right);$$

and $$\mathrm{Y}\left(\mathrm{i}\right)=\mathrm{W}\left(\mathrm{i}\right)\cdot \mathrm{X}\left(\mathrm{i}\right);$$

The maximum value for W is 1 (output = input). If b = 1 is selected, the spectral subtraction is practically switched off. With b = 0, the spectral subtraction achieves the maximum effectiveness. In practice, b = 0 results in poor voice quality.

With the parameter b results for the present invention, the ability to adjust the one-sided spectral subtraction continuously in their effectiveness. With a value of e.g. b = 0.25 achieves a noise suppression of about 12dB and a good voice quality.

The invention is explained below with reference to exemplary embodiments with reference to schematic drawings.

FIG. 4 shows three block diagrams with one-sided spectral subtraction for the reference input. In FIG. 4 a, the primary useful signal P of the beamformer (eg voice signal) is connected as a differential array DA for the channels 1, 2 (arrangement c in FIG. 3). Figure 4b, 4c shows a circuit of the primary signal P as a sum and difference array SD (arrangement a and b in Figure 3).
The interference reference input processes the reference signal R with the additional extension of the one-sided spectral subtraction in differential form according to the arrangement d and e in FIG. 3. The difference between useful signal in channel 2 and interference-canceled useful signal from channel 1 is applied to the adaptive filter H 1. The adaptive filter H1 is adapted in the time domain or in an equivalent form in the frequency domain using the LMS algorithm. The filtered interference reference signal R is then subtracted from the primary useful signal P.

A further embodiment of the invention according to Figure 5 is that the one-sided spectral subtraction, PLC ₁ 'is performed once on the channel 1 for the useful signal to form together with the useful signal in channel 2, a first reference signal R1. A second time, the unilateral spectral subtraction, SPS ₂ 'performed on the useful signal of the channel 2, to form together with the useful signal in channel 1, a second reference signal R2. The result is a system with 2 reference signals, which are subtracted from the primary signal P. In the case of speech signals, the interference is detected in each case with the characteristic of the individual microphones during speech pauses, and a zero point for the speech signal is generated during speech activity.

4, the modification with 2 reference inputs for 'end fire' microphone arrangement or 'broad side' arrangement is used. Figure 5 shows the block diagram for the 'end fire' arrangement. The beamformer consists of the channel 1 for the speech signal and two reference channels 2, 3. Each reference input is filtered by an adaptive filter 'H ₁ ', or 'H ₂ '. Filter balancing is performed with a multi-channel LMS algorithm.

If more than 2 input signals are available, a one-sided spectral subtraction is performed by combining two inputs each in the manner described in order to obtain a reference signal. If, for example, a 'broad side array' with 3 microphones is assumed, this results in 6 combinations for pairing. Taking into account that for each pair the one-sided spectral subtraction is optionally performed on one or the other channel, so doubles the number of combinations and thus the number of Referenzkänale. An array of multiple microphones uses a limited number of possible combinations.
The invention is not limited to the recording of the useful signals by microphones, but receiving systems such as antennas can be used. Useful signals can be any kind of acoustic and electrical signals.

Claims (11)

1. A method of producing an interference reference signal R for noise reduction of a primary useful signal, which is produced by the combination of the signals from at least two channels, particularly voice channels, wherein the signals are processed together in pairs and wherein only one of each of the signals processed in pairs is subjected to spectral subtraction and is used to form a difference with the other signal so that an interference reference signal R is produced as the result which contains substantially only the interference signal itself as the reference.
2. A method as claimed in Claim 1, characterised in that the primary useful signal is connected in the form of a differential array (DA) of two channels (1, 2).
3. A method as claimed in Claim 1, characterised in that the primary useful signal is connected in the form of a sum and difference array (SD) of two channels (1, 2).
4. A method as claimed in one of the preceding claims, characterised in that the interference reference signal is produced with the additional broadening of the one-sided spectral subtraction in differential form so that the difference between the screened useful signal from (1) channel and the useful signal from a further channel 2 is supplied to an adaptive filter (H1) and that the filtered interference reference signal (R) is then subtracted from the primary useful signal (P).
5. A method as claimed in one of Claims 1 to 3, characterised in that a spectral subtraction (SPS₁) is performed at a first channel (1) for the useful signal and is supplied together with the useful signal in a second channel (2) to an adaptive filter (H1) and a first reference signal (R1) is formed, that a further spectral subtraction (SPS₂) is performed on the useful signal of the second channel (2) and is supplied together with the useful signal from the first channel (1) to an adaptive filter (H2) in a further channel (3) and a second reference signal (R2) is formed and that the two reference signals (R1, R2) are subtracted from the primary useful signal (P).
6. A method as claimed in one of the preceding claims, characterised in that the filters (H1, H2) are adapted in the time domain or in the frequency domain with the LMS algorithm.
7. A method as claimed in one of the preceding claims, characterised in that the useful signal is recorded by microphones.
8. A method as claimed in one of the preceding claims, characterised in that a voice signal is used as the useful signal.
9. A method as claimed in one of the preceding claims, characterised in that the spectral subtraction is continually adjusted with a parameter in its effectiveness.
10. A method as claimed in Claim 9, characterised in that the parameter is formed as a minimal value of a filter coefficient of the spectral subtraction at each frequency index.
11. A method as claimed in one of the preceding claims, characterised in that, with more than two input signals, a spectral subtraction is performed by a combination of two inputs to produce a reference signal.

Images