EP1438870A1

EP1438870A1 - Interference cancelling method and system for multisensor antenna

Info

Publication number: EP1438870A1
Application number: EP02790538A
Authority: EP
Inventors: Claude Marro; Jean-Philippe Thomas
Original assignee: France Telecom SA
Current assignee: Orange SA
Priority date: 2001-10-25
Filing date: 2002-10-23
Publication date: 2004-07-21
Also published as: WO2003037033A1; FR2831717A1; US20040264610A1

Abstract

The invention concerns a method for reducing interference for a reception system using a multisensor antenna and at least a channel formatter supplying an antenna signal from signals received by the different sensors of said antenna. The method consists in estimating, from a reference signal for regenerating said interference, the transfer function of a first filter, called first-post-filter, and in filtering the antenna signal with said first post-filter.

Description

Interference elimination method and system for multisensor antenna

The present invention relates generally to an interference reduction method in a multi-sensor reception system as well as a multi-sensor reception system implementing said interference reduction method.

The invention more particularly finds application in the field of acoustic echo cancellation for sound recording and in the field of hands-free telephony.

Prior to the presentation of the state of the art, we will make certain terminological conventions and specify certain notations. These conventions and details are valid for the entire description. Any device capable of transforming a physical quantity into a measurement signal, for example an antenna or an acoustic transducer, will be called a "sensor". A reception system using a plurality of such sensors will be called a “multi-sensor reception system”. An antenna consisting of a plurality of such sensors, for example an array of sensors, will be called a “multi-sensor” antenna.

The term “gain” will not be understood in the strict sense and will subsequently cover both amplification (gain greater than 1) and attenuation (gain less than 1). The notation A (t, f), represents a quantity a at time t and at the frequency / This makes it possible to describe a quantity in the frequency domain but which varies over time. As we know, the passage from the time domain to the frequency domain requires an observation of the quantity a in a time window. In this sense, we understand that A (t, f) is a known value at time t, but that its calculation may have required an observation of the quantity a for a certain duration. A (t, f) can be obtained from a by means of a spectral estimation in a time window. Ait, f) can be a signal, a spectral quantity, or the transfer function of a filter varying over time. Subsequently, the notations A t, f) and a will be used interchangeably to designate a quantity a whose spectrum varies over time.

The notation a (t) is reserved for a quantity a at time t, varying in time. The notation A (f), meanwhile, represents a quantity a at the frequency f, remaining invariant over time (for example the transfer characteristic of an invariant filter).

The notation " ^Λ " denotes in the following the estimate of the quantity to which it applies.

Fig. 1 schematically represents a multisensor reception system known from the state of the art. Such a system typically comprises a plurality of sensors 100 _1? .., 100N transforming a physical quantity into received signals X _x {t, f), .., X _N (t, f), respectively. These signals are then used to perform beamforming using the filters 1 10 ₁₅ .., 110 _N and the adder 120. More precisely, the signals V _x (t, f), .., V _N (t, f) at filter output

HO. _^ l 10 _N are summed in the summator 120 to provide a signal, called the antenna signal, noted Y (t, f). The filters in question introduce delays and / or phase shifts as well as a weighting of the signals received. It is clear that, in the general case, the transfer functions of these filters can depend on both time and frequency. In practice, the filters 110I, .., 1 10N simply perform a multiplication using complex weighting coefficients. These coefficients are determined according to the desired reception pattern, so that the main lobe of the reception pattern points in the direction of a useful signal source and has a given opening angle. The choice of coefficients is also dictated by the desired level of attenuation of the secondary lobes. Likewise, these coefficients can be determined so as to introduce one or zeros in the reception diagram for directions in which there are interfering sources. In addition, the coefficients can be calculated adaptively in order to follow a mobile useful signal source or to reduce interference arriving at a variable angle of incidence. However, this interference reduction technique by means of the reception diagram does not make it possible to eliminate interference signals received by the main lobe of the diagram, which lobe can be large if the number of sensors is small.

Another technique for eliminating interfering signals involves knowing or estimating these signals. It is notably used for echo cancellation. It is recalled that an echo is generated in a transmission system when a transmitted signal is reflected towards the transmitter after having propagated in the transmission channel, the reflection possibly being due to characteristics of the channel (lack of adaptation impedance in particular) or a coupling between a receiver and a transmitter at the near end (far end) or far (far end) of the channel. Fig. 15 illustrates a simple echo generation situation. There is shown a multi-media terminal 1500 comprising a monitor 1510, a pair of speakers 1520 and a multi-sensor acoustic antenna 1530 made up of microphones. The transmission channel is characterized by the direct propagation paths between the loudspeakers and the user (placed in front of the multi-media terminal). The arrows 1540 illustrate the phenomenon of echo generation, attributable here to the acoustic coupling between the loudspeakers and the acoustic antenna. The coupling is due to the propagation paths between the loudspeakers and the antenna as well as to the reflections of the signals emitted by the loudspeakers on the environment of the terminal (people, walls, objects etc.). Fig. 2 schematically reproduces a known system of echo cancellation reception, as described for example in the article by W. Kellermann entitled “Strategies for combining acoustic echo cancellation and adaptive beamforming microphone arrays” published in April 1997 in Proc. ICASSP-1997, Vol. 1, pages 219-222. We find in Fig. 2, the multi-sensor system of FIG. 1 consisting of sensors 200 _I , .., 200N, filters 210 ₁₃ .. ₃ 210N and summing device 220. The system further comprises an element 201 symbolizing a sensor or a socket making it possible to obtain a reference signal e of the signal interfere, also noted E (t, f) in frequency representation. In general, we will call the reference signal the interfering signal, a picked up or picked up signal, allowing the signal to be reconstructed. interfere. For example, in the system of FIG. 15, the reference signal could be the signal transmitted to a loudspeaker and the element 201 could be a socket on the loudspeaker control circuit. The reference signal e is filtered by an echo canceller filter 205 of transfer function H _e (t, f). This filter actually models the echo propagation channel. The filter output is subtracted from the antenna signal 230 to provide a signal Z (t, f), ideally rid of the echo. In the context of audio or speech signals, this echo cancellation technique is also called adaptive acoustic echo cancellation. Indeed, the coefficients of the echo canceller filter 205 are calculated adaptively in the calculation module 206, so as to minimize the mean square value of the error signal Z (t, f).

If several interfering sources are present, it is difficult to implement the technique of adaptive acoustic echo cancellation. For example, in the case of a multi-media terminal, if the sound is spatialized (several speakers then reproducing different spatial components) taking into account all the echoes is extremely complex. Furthermore, when the antenna is multi-channel, that is to say if several reception diagrams are generated by a plurality of sets of different channel filters, the problem of echo cancellation for the different channels is almost insoluble.

Furthermore, the performance of acoustic echo cancellers is limited by the size of the room in which the sound recording system is used. Indeed, in the case of large rooms, the echo canceller must identify an acoustic impulse response (transfer function H _e (t, f)) which can reach several seconds, which leads to a considerable increase in digital complexity of the adaptation algorithm. In addition, the duration of the impulse response to be identified has a direct influence on the performance of the algorithm: the convergence time (i.e. the speed of adaptation of the filter) as well as the mismatch (i.e. i.e. the difference between the effective reduction of the echo level and the theoretical maximum reduction) increases with the duration of the impulse response. Thus, if the impulse response changes suddenly, the new echo may be filtered in error for several seconds, which can be particularly annoying to the ear.

The general objective of the present invention is to propose a method of eliminating interfering signals for a multi-sensor reception system, which does not have the drawbacks of the state of the art. In particular, one of the goals of the present invention is to provide a method of eliminating interfering signals, in particular of an echo, having a higher adaptation speed than in the prior art. A secondary object of the present invention is to provide a simple method of eliminating interfering signals capable of taking into account a plurality of interfering sources, including when the reception system is multi-channel.

The problem is solved by the invention defined by an interference reduction method for a reception system using a multisensor antenna and at least one channel trainer supplying an antenna signal from the signals received by the various sensors of said receiver. antenna. According to this method, the transfer function of a first filter, called the first post-filter, is estimated from a reference signal making it possible to regenerate said interference, and said antenna signal is filtered by said first postfilter.

Advantageously, said transfer function is obtained from a short-term estimate and a long-term estimate of the spectral density of said reference signal, for example from a ratio between these two estimates.

The short-term estimation and the long-term estimation of the spectral density are preferably obtained by low-pass filtering of a spectrum of the reference signal. For example, the short-term estimate Φ ^ (t, f) of the spectral density is obtained by recursive filtering of the type: where E (t, f) is a spectral component of the reference signal at frequency / and at time t, a is a coefficient between 0 and 1, δt is the delay in the loop of the recursive filtering and. denotes the conjugation operation and the long-term estimate φ (t, f) of the spectral density is obtained by recursive filtering of the type:

where a and α ₂ are coefficients such that 0 <α ₂ <o.ι <l.

According to a second embodiment, the signals received by the various sensors being filtered by at least one set of channel filters before being summed for supplying said antenna signal, the transfer function of a second filter, called second post-filter, is estimated from said received signals, before or after filtering by said channel filters, and the signal is filtered. antenna by said second post-filter. Advantageously, the transfer function of said second post-filter is estimated from an average of the power spectral densities and an average of the interspectral power densities of said received signals, after filtering by said channel filters. For example, the transfer function W _s (t, j) of said second post-filter is estimated by:

where Φ _vv (/, /) and Φ., _v (t, f) are respectively estimates of the spectral densities and interspectral power densities of the signals received after channel filtering, b _t (f) are the transfer functions of the different channel filters free of phase-shifting terms, N is a number of antenna sensors and γ (.) denotes the actual value or the module. According to another variant of the second embodiment, the signals received by the various sensors being filtered by at least one set of channel filters before being summed to provide said antenna signal, the transfer function of an second filter, said second post-filter, from said signals received after filtering by said channel filters, as well as from the antenna signal, and the antenna signal is filtered by said second post-filter. Advantageously, the transfer function of said second post-filter is estimated from an average of the interspectral power densities of said received signals, after filtering by said channel filters, and from an estimation of the spectral density of the antenna signal. . For example, the function of estimated by:

where Φ _{V Vι} (t, f) and Φ _w (t, /) are respectively the spectral and interspectral power densities of the signals received after channel filtering, b, (/) are the transfer functions of the different channel filters removed in terms of re-phasing, Nest a number of antenna sensors and γ (.) denotes the actual value or the module. Advantageously, the filtering of the antenna signal by the first post-filter and that of the second post-filter are applied in a combined manner, by filtering the antenna signal by means of a post-filter, called the first combined post-filter. having the transfer function a combination of the transfer functions of said first and second post-filters.

According to a third embodiment, a statistical analysis is carried out of spectral components of the transfer function of the second post-filter and / or of the transfer function of the first combined post-filter and that an indication of presence is deduced therefrom. or lack of a useful signal. According to a variant of the third embodiment, the statistical analysis is also carried out on spectral components of the transfer function of the first post-filter. Advantageously, said statistical analysis uses a criterion of spectral occupancy rate and / or variance of said spectral components. According to the third embodiment, a switching signal is generated from said indication of presence or absence of a useful signal and the antenna signal is filtered by means of the first combined post-filter when the switching signal is in a first state and is filtered by means of a second combined post filter when the switching signal is in a second state, the transfer function of the second combined post filter being a combination of the transfer function of the first post -filter and a predetermined attenuation. According to a fourth embodiment, the received signals being filtered by a plurality of sets of channel filters to form a plurality of channel signals, a statistical analysis is carried out of spectral components of the transfer functions of the second post-filters associated with the different channel filter sets and that the channel with the highest probability of the presence of a useful signal is deduced therefrom. According to a varainte of the fourth embodiment, the statistical analysis is also carried out on spectral components of the transfer function of the first post-filter. Advantageously, said statistical analysis uses a criterion of spectral occupancy rate and / or variance of said spectral components. The antenna signal is then obtained from the channel signals relating to the channel having the highest probability of presence of useful signal.

The invention is also defined by a reception system comprising a multisensor antenna, at least one channel trainer and interference reduction means, said interference reduction means being adapted to implement said method of reduction of interference. The characteristics of the invention mentioned above, as well as others, will appear more clearly on reading the following description of exemplary embodiments, said description being made in relation to the accompanying drawings.

Fig. 1 schematically represents a multi-sensor reception system known from the state of the art;

Fig. 2 schematically represents a multi-sensor reception system with echo cancellation, as known from the state of the art;

Fig. 3A schematically represents a multi-sensor reception system according to a first embodiment of the invention; Fig. 3B schematically represents a multi-sensor reception system according to a variant of the first embodiment of the invention;

Fig. 4 schematically represents a multi-sensor reception system according to a second embodiment of the invention;

Fig. 5 schematically represents a multi-sensor reception system according to a variant of the second embodiment of the invention;

Fig. 6 schematically represents a multi-sensor reception system with useful signal detection;

Fig. 7 schematically represents a multi-sensor reception system according to a third embodiment of the invention; Fig. 8 schematically represents a module of the multisensor reception system shown in FIG. 7 according to a first alternative embodiment;

Fig. 9 schematically represents a module of the multisensor reception system shown in FIG. 7 according to a second alternative embodiment;

Fig. 10 schematically represents another module of the multi-sensor reception system shown in FIG. 7;

Fig. 11 schematically represents a multi-sensor reception system according to a fourth embodiment of the invention;

Fig. 12 shows an example of a multi-sensor reception system according to the third embodiment of the invention; Fig. 13 schematically represents a module of the multi-sensor reception system shown in FIG. 12;

Fig. 14 schematically represents another module of the multi-sensor reception system shown in FIG. 12; Fig. 15 schematically represents a context of application of a multi-sensor reception system according to the invention.

We again refer to the context of a multi-sensor reception system and to the case where there is a reference to the interfering signal. Fig. 3A illustrates a first embodiment of the invention. We find there the reception system comprising a plurality of sensors 300J, .., 300N, filters 310 ₁₅ .., 310N and a summator 320 performing the channel formation. An element 301, sensor or socket, provides a reference signal e of the interfering signal. An estimation module 351, which will be detailed below, calculates from said reference signal the transfer function W _e (t, f) of a filter 350 at the antenna output, for this reason called "post-filter"".

The post-filter 350 receives from the estimation module the information allowing it to perform said transfer function. The channel (or antenna) signal Y (t, f) filtered by the post-filter 350 is denoted Z (t, f).

In general, the transfer function W _e (t, f) of the post-filter is determined so that in the absence of an interfering signal, it is worth 1 (the filter

350 is "transparent") and that in the event of the emergence of an interfering signal, it attenuates the frequency components of the antenna signal corresponding to the energetic components of the interfering signal.

More precisely, the estimation module 351 advantageously calculates the transfer function W _e (t, f) of the post-filter in the following manner:

where Φ ^ T (- ^* , /) is the power spectral density (dsp) of the reference signal e, estimated in the short term and where Φ (t, f) is the power spectral density of the reference signal e, estimated long-term.

The long-term estimation makes it possible to follow the evolution of the background noise and that in the short term the signal emergences interfere. In the presence of background noise alone, in other words in the absence of an interfering signal, the transfer function W _e {t, f) is close to 1. On the other hand, in the presence of an interfering signal, the transfer function W _e (tJ) represents the inverse of the spectral power of this signal, reduced to the power of the background noise. It is clear that expressions of W _e t, f) other than that given in (1) could also be suitable insofar as they would make it possible to attenuate the frequency components corresponding to the energy frequency components of the signal interfere. Short-term and long-term spectral densities are estimated from an averaging of a periodogram of the reference signal. Advantageously, this averaging is performed recursively:

Φ ^ {t) ^^ (t-δt (l-ά) E (t, f) E t, β (2) and

Φ ^L t aΦ ^L ^ -δt, fY {\ - ~ a ^ E (t, f) E \ t, f) si Φ _e {t, f) <Φ ^c _e (t, f) (3)

if Ô ^r (, Φ _e ^c : (r,) (4)

where α is a constant, called short-term time constant, between 0 and 1 and where a _\ and α ₂ are constants such that 0 <o: ₂ <α! <l, a \ representing a time constant at long term and α ₂ a long term time constant. The calculations of expressions (2) to (4) are carried out simply by means of a first order recursive filtering, δt being the delay present in the filter loop. The constants, is chosen very close to 1 so as to take into account the past over a long period while the values of and ΩΓ ₂ are chosen lower so as to react more quickly to variations in spectral density of the reference signal.

Other types of low-pass filtering can of course be envisaged for the averaging of the periodogram of the reference signal.

Although the above-mentioned long-term and short-term spectral density estimation method operates in the frequency domain, it is however clear that spectral estimation methods in the time domain can also be used, for example from an autocorrelation of the reference signal e. Fig. 3B illustrates a variant of the first embodiment of the invention.

This variant applies in the presence of a plurality of interfering signals. Let e \, .., ep be the respective reference signals relating to these interfering signals, as picked up or received in 301 _1? .., 301p. The signals e _l ., E ^ are subject to short-term and long-term spectral estimation in the respective estimation modules. 35 li, .., 351p. Each of these modules also calculates a transfer function of a post-filter according to 1. The elementary transfer functions W _e (t, f), .., W _e (t, f) are then combined in a combination module 352 to provide a global transfer function W _e t, f). The term “combination” must be taken here in the broad sense: it can especially be a linear combination or, preferably, a multiplication of elementary transfer functions. In all cases, the module

352 transmits information to the post-filter 350 enabling said global transfer function to be carried out. The antenna signal Y (t, f) filtered by the post-filter 350 is ideally free of the P interfering signals. Fig. 4 illustrates a second embodiment of the invention. There are a plurality of sensors 400I, .., 400N, channel filters 410I, .., 410N respectively filtering the signals received by the different sensors and a summator 420 allowing channel formation to be carried out. In 401, the reference signal e of an interference signal is received or sampled. The estimation module 451 determines, as before, the transfer function W _e (t, f) post-filter 450. This embodiment differs from the previous one in that there is provision for additional post-filtering by means of a post-filter 440 placed upstream or downstream of the post-filter 450. The post-filter 440 has a transfer function Wft, f) provided by an estimation module

441 which will be detailed later. According to a variant embodiment not shown, the filters 440 and 450 can be replaced by a single filter with transfer function W (t, f) obtained by combination (in the broad sense) of the transfer functions W _e (t, f) and

W {t, f). For example, we could choose:

W (t, f) = W _s {t, f) W _e {t, f) (5)

or, more generally:

W <, f ^' ) = W f) W% f) (6)

where μ, v are strictly positive integers.

Post-filtering in 440 aims to improve the signal-to-noise ratio (SNR) at the antenna output. By “noise” here is meant all of the disturbing signals for which there is no reference. If you have a signal reference disruptive, the latter can be taken into account in the expression of W _e (t, f) as we have seen in Fig. 3B.

The post-filtering technique in order to improve the signal-to-noise ratio at the output of a multi-sensor antenna was notably described in the article by C. Marro et al. entitled "Analysis of noise reduction and deverberation techniques based on microphone arrays with postfiltering" published in May 1998 in IEEE Trans. on Speech and Audio Processing, Vol. 6, No. 3, pages 240-259. The main results are recalled below.

We consider the case where a source emits a useful signal s. This signal is received by the various 400 _1. , 400 _N sensors after propagating to the antenna. The signals received by the various sensors are disturbed by noise. We will note n _{ the noise at the sensor 400 ,. We will make the following assumptions:

• the signal x, at the level of the sensor i is modeled by the sum of the useful signal received after propagation to the sensor and of the noise n,; • for each sensor, the noise n _t and the useful signal received are decorrelated;

• the noise power spectral densities n, are identical on each sensor;

• the noises are decorrelated between sensors 400, and 400, in other words the interspectral power densities are zero for i ≠ j; • the input signals x, are perfectly put back in phase with respect to s by the filters 410 _I , .., 410 _N. The optimal post-filter W _s is that which minimizes the mean square error between the desired signal s and the signal at the output of the filter. As shown in the article by KU Simmer et al. in the article “Time delay compensation for adaptive multichannel speech enhancement Systems” published in Proc. ISSE-92 in September 1992, the expression of this optimal filter can be written from the useful signal s and the average noise n at the antenna output:

where Φjyt, f) and Φ _m (t, f) are the spectral power densities of the wanted signal and of the noise at the output of the channel formation.

The spectral densities Φ _ss (t, f) and Φ __ (t, /), necessary for the calculation of W _s (t, f), are a priori unknown and the difficulty lies in their estimation. It has been proposed to estimate them from the signals received on the various sensors. Indeed, if we respectively note Φ _vv (t, /) and Φ „„ (*, /) the spectral and interspectral power densities of the signals x, re-phased, that is to say signals v „We have, under the above assumptions:

^φ _{V /} fe / fe /). ° ^{i ≠} J ⁽ 9 ⁾

One means of estimating W _s (t, f) then consists in using an average of the spectral densities Φ _vv (t,) and interspectral Φ _{v ι} , (t, f) respectively in the denominator and in the numerator of (7), that is to say :

where γ (.) = Re (.) or γ (.) = |. |

The use of the operator module or real part γ (.) Is justified by the fact that the quantity in the numerator must be real and positive.

The module 441 in FIG. 4 performs the estimation of W _s (t, f) according to expression (10). For this, it uses either the signals x, and puts them back in phase, or directly the signals v ,. The signals x, and v, have been represented in FIG. 4 as vectors V fJ)) where Xit, f) and V, {t, f, are the frequency notations of x, and v ,. Thus, the estimation module 441 uses either the vector X (t, f) or the vector N (t, /). The two possibilities are shown in Fig. 4. Advantageously, an estimate of W _s (t, f) close to the expression will be used

(10) but which has the advantage of being standardized:

where b, (f) is equal to the coefficient a _l (f) of the filter 410 ,, rid of the resetting phase, that is to say if we denote r, the propagation time of the signal between the useful signal source and the sensor 400,, b _l (f) = a _t (f) s ^β ', and where a, is the attenuation coefficient between the useful signal source and the sensor 400 ,. If the distances between sensors are small compared to the distance separating the antenna from the useful source, the attenuation coefficients are close and the expression (11) can then be reduced to:

It should be noted that the estimation according to (10), (11) or (12) may relate only to a subset of the signals x, or v „of cardinal M <N, the mean of the spectral and interspectral densities n 'then being taken only on this subset. Finally, said estimation can be carried out in the frequency domain or in the time domain.

Fig. 5 illustrates a variant of the second embodiment. The elements 500 ₁₅ .., 500N, 510 _I , .., 510N, 520, 501, 550, 551 are respectively identical to the elements 400 ₁ , .., 400 _N , 410ι, .., 410 _N , 420, 401, 450, 451 of FIG. 4.

According to this variant, an estimate of the transfer function of the post-filter is used, as proposed by UK Simmer et al. in the article entitled "Adaptive microphone arrays for noise suppression in the frequency domain" published in Proc. of the Sec. Cost 229 Work on Adaptive Algorithms in Communications, pages 185-194, Bordeaux, France, 1992. This estimate uses the signals x, re-phased, ie signals v, as well as the antenna signal JΛ More precisely W _s {t, f) is estimated by:

where Φ _yy (t, f) is the spectral density of the antenna signal;. Again, as in the previous case, we can advantageously adopt a standardized estimate, namely:

with the same rating conventions.

Finally, the estimate (13) or (14) may relate only to a subset of the signals x, or v „, the mean of the spectral and interspectral densities then being taken only on this subset. Said estimation can be carried out in the frequency domain or in the time domain.

The module 541 of FIG. 5 performs the estimation of W _s (t, f) according to expression (11). For this, it uses on the one hand the antenna signal y (Y (t, f) in frequency representation) and, on the other hand, either the signals x, (vector X (t,)) after having delivered them in phase, ie directly the signals v, (vector N (t, /)). The multi-sensor reception systems shown in Figs. 4 and 5 allow, as we have seen, to significantly improve the signal to noise ratio at the antenna output. However, when the useful signal is absent, the post-filters 440 and 450 or 540 and 550 tend to distort the noise, which can prove to be annoying, in particular in sound recording applications. To remedy this drawback, it was proposed in patent application FROO 05601 filed on 28.4.2000 by the applicant and included here by reference, not to apply the post-filter permanently but only when the useful signal is present, a constant gain being applied otherwise. A switching signal, denoted below K (t), ensures the switching of the antenna output either to a constant gain filter or to the transfer function filter W _s (t, f).

Fig. 6 illustrates a multi-sensor reception system with switchable post-filter as set out in the above-mentioned application.

The elements 600 _1. , 600Ν, 610I, .., 610Ν, 620, 641 are respectively identical to the elements 500ι, .., 500 _Ν , 510 _1; .., 510 _N , 520 and 541 (or 441). The arrow in dashed lines carrying the reference Y (tf) indicates that the module 641 can carry out an estimate of W _s (t) according to (10), (11), (12) or else according to (13) or (14). Components of the transfer function W _s (t, f) (or more precisely its estimated) are the subject of a statistical analysis in the analysis module 661. The latter performs a set of operations which will be detailed more far and which aim to identify by means of analysis of variance and / or spectral occupancy rate an index of presence of useful signal in the spectrum of W _s (t, f). The module 661 provides a binary indicator for the presence / absence of a useful signal denoted P_A (t) which is temporally smoothed in the low-pass filtering module 662 to give a gain value G (t). This gain value is compared with a threshold value ST in the comparator 663, the result of the comparison, K (t) controlling the switching of the antenna output, by means of a switch 630, ie towards the post- filter 640 or to an attenuator 645 with constant gain G _SA - The respective outputs of the post-filter and the attenuator are both connected to the input of an amplifier 660 with gain controlled by the gain value G (t) .

This reception system makes it possible to avoid distortion of the residual disturbance when the useful signal is absent. To ensure continuity during switching, a temporal gain smoothing is provided. The temporal smoothing of gain carried out in the low-pass filter 662 is intended to ensure a smooth transition when switching at the antenna output.

Fig. 7 illustrates a multi-sensor reception system according to a third embodiment of the invention. This embodiment uses post-filtering and switching of presence / absence of useful signal in FIG. 6. The elements 700 ₁ , .., 700 _N , 710 ₁ , .., 710 _N , 720, 730, 740, 741, 760, 762, 763 are respectively identical to the elements 600 _1; .., 600 _N , 610 _l5 .., 610 _N , 620, 630, 640, 641, 660, 662, 663 of the latter. The system of FIG. 7 differs from that of FIG. 6 in that it implements an elimination of interfering signals according to the principle of the invention.

Indeed, 701 denotes an element making it possible to receive or take a reference signal from an interfering signal and 751 denotes an estimation module identical to the module 351 of FIG. 3A. The estimation modules 741 and 751 respectively estimate the transfer function W _s (t, f) of a first post-filter and the transfer function W _e (t, f) of a second post-filter, the first allowing to increase the signal to noise ratio subject to the aforementioned hypotheses and the second allowing to eliminate interference (or interference as in Fig. 3B) for which there is a reference. The transfer functions W _s (t, f) and W _e (t, f) (or more precisely their estimates) are provided, as well as their product W (t, f) obtained by the multiplier 725, to a statistical analysis module 761 similar to module 661 and the function of which will be detailed below. This module provides a binary indicator of presence / absence of useful signal P_A (t) which is smoothed temporally to provide the gain G (t). In addition, the output W {t, f) = W _s (t, f) .W _e (t, f) of the multiplier is transmitted to the filter 740. Likewise, the estimation module 751 transmits the indication W _e (t) to the filter 750, which allows it to carry out the transfer function G _S A- W _e (tj).

According to a first variant, the filters 740 and 750 respectively have the transfer functions W {t, f) = W _s ^μ {t, f) W _e ^v (t, f) and G _SA W _e ^v (t, fi where μ , v are strictly positive integers According to a second variant, the attenuation G _S A depends on the frequency while remaining constant over time, which can in particular make it possible to systematically reject certain parts of the spectrum, independently of the interference.

Thus when a useful signal is present, the antenna output is filtered by the combination of the two post-filters as in FIG. 5 whereas, when the useful signal is absent, it is simply subject to filtering by W _e (t, f) (or W _e ^v {t, f)) and attenuation.

The functional diagram of the statistical analysis module 761, according to a first exemplary embodiment, is shown in FIG. 8. The transfer functions W (t, f), W _s (t, f), W _e (t, f) undergoing analogous processing, we will limit ourselves to that of the function W _s (t, f). We extract from W _s (t, f), by extraction means 811, a set of frequencies F _ocp set by the user. The spectrum thus obtained possibly undergoes a non-linear transformation (not shown) to give more relevant information. If necessary, a logarithmic transformation (in decibels) will be advantageously used:

The spectral components reduced to the set of frequencies F _ocp are then compared to a threshold SOC ^s in the comparator 812. We then determine in module 813 the rate τ ^ (t) of frequencies for which W _s (t, F _0Cp ) exceeds a predetermined SOC ^s threshold, i.e. w _t ytaiUe of the spectrum in F _ocp such that W _s (t, F _ocp )> SOC -.w ⁿ . ∞ ^> size of the spectrum of F

The occupancy rate τ r O ₀ 'c ^* p _p (t) is then compared with a predetermined threshold value TOC "*. In the comparator 814. This comparison provides a binary signal p_a'. The binary signal p_a 'indicates a useful signal is present when the occupancy rate τ '(t) is greater than the TOC ^* threshold. Similarly, the processing chain consisting of the extraction module 821 (resp. 831), of the comparator 822 (resp. 832), the occupancy rate calculation module 823 (resp. 833) and the comparator 824 (resp. 834) provides a binary signal p_a ^w (resp. p_a '). The binary signal p_a ^* gives a indication of the presence of the useful signal without taking the interference signal into account, on the other hand, the binary signal p_a 'gives an indication of the presence of the interference signal. The binary signal p_a ^w gives an indication of the presence of useful signal from global information taking into account both the useful signal and the inter signal The three binary signals p_a ^s , p_a ^w , p_a 'are combined in the combination module 850 to provide a binary indicator of presence / absence of useful signal P_A (f). The combination function used may in particular depend on the gain value G (t). We may indeed wish to favor more or less the indicator p_ a ^w depending on whether or not we are already in the presence of a useful signal. According to a first simplified variant, the statistical analysis module only has the processing chains of W _s (t, f), W _e (t, j) and the combination at 850 relates only to the indicators p_a * and p ^e . According to a second simplified variant, the statistical analysis module only comprises the processing chain of W (t, f) and, consequently, no combination is carried out. The functional diagram of the statistical analysis module 761, according to a second embodiment, is shown in FIG. 9. This example uses a variance criterion instead of an occupancy rate criterion.

The transfer functions W (t, f), W _s (t, j), W _e (t, f) undergoing similar processing, we will limit ourselves to that of the function W _s (t, f). After extraction of the spectral components at a set of frequencies F _var set by the user in 911 and possibly non-linear transformation, a thresholding of the spectrum reduced to these frequencies is carried out in 912 by means of a predetermined threshold value SVAR ^W ' , only the components above the threshold being retained. The variance va ^s of the spectrum thus obtained is then calculated in 913 and then compared with a threshold value VAR 'in a comparator 914 to provide provides a binary signal p_α *. The binary signal p_α "indicates that a useful signal is present if the variance va 'is less than the threshold value VAR ^s . In the same way, the processing chain consisting of the extraction module 921 (resp. 931), comparator 922 (resp. 932), variance calculation module 923 (resp. 933) and comparator 924 (resp. 934) provide a binary signal p_a ^w (resp. p_a ^e ). The three binary signals are combined in 950 to provide the binary indicator P_A (t) The remarks made for the first example in Fig. 8 also apply here mutatis mutandis.

According to another alternative embodiment, it is possible to combine the binary signals or indicators obtained according to the occupancy rate criterion and the variance criterion to form a synthetic binary indicator P_A (f).

The low-pass filter 762 is shown schematically in FIG. 10. We will denote P the state of presence and A the state of absence of the useful signal. The function of the filter 762 is to continuously decrease the gain G (t) to a value S _mm when passing from state P to state A and to increase the gain G (t) to value S _max during the pass in reverse.

The low-pass filter receives the binary indicator P_A (t) which varies over time. This indicator controls the switching between a minimum gain value S _mm and a maximum gain value S _max by means of a first switch 1010. When the useful signal is present, the maximum value S _max (generally fixed at 1) is injected at l 'common input of two low-pass filters 1020 and 1030. When the useful signal is absent, it is the minimum value S _mm which supplies the common input. To ensure continuous growth and then maintaining G (f) at the value S _max during the transitions from state A to state P, the input signal is filtered by the low-pass filter 1020 of time constant tp. The choice of this time constant conditions the rise time of the signal G (t). Similarly, to continuously decrease and then maintain from G (t) to the value S _mn during the transitions from state P to state A, the input signal is filtered by the low-pass filter 1030 of time constant TA, which conditions the descent time of G (t). The outputs of the two low-pass filters are connected to the inputs of a second switch 1040 which selects, using the binary indicator P_A (t), the output of the low-pass filter with time constant T _A if the useful signal is absent and the output of the low-pass filter with time constant τ if the useful signal is present. The common output of switch 1040 provides the smoothed gain signal Gif).

Fig. 11 illustrates a multi-sensor reception system according to a fourth embodiment of the invention. Unlike the previous mode, a plurality K of channels are formed using R. separate sets of channel filters 1110 ^, .., ll lOi, k = \, .., K, operating in parallel, each clearance to form a path in a particular direction of reception. The 1141 estimation module estimates the transfer functions , k = \, .., K, post-filters associated with these K games. To do this, it uses either the K signal sets at the output of the K sets of channel filters, ie the signals x, - after compensation by K sets of phase shift. In any event, the transfer functions are analyzed on the basis of a variance or spectral occupancy rate criterion by a first statistical analysis module 1145 and the channel ko having the highest probability of receiving a useful signal is selected. The index ko generated by the module 1145 selects by means of the multiplexer 1115 the signals V _! ⁰ , .., ^ ⁰ which provides them to the summator 1120 to form the channel in question. The transfer function of the post-filter associated with channel kb is transmitted to a second statistical analysis module 1161 identical to module 761 in FIG. 7.

According to a variant embodiment, the first statistical analysis module 1145 also receives the transfer function W _e t, f) from the estimation module 1151 and takes it into account for the selection of the channel kb. The first statistical analysis module 1145 then performs all of the statistical analysis operations and directly supplies the binary indicator P_A (f). The elements 1130, 1140, 1150, 1160, 1162 and 1163 are in both cases identical to the elements 730, 740, 750, 760, 762 and 763 in FIG. 7.

Fig. 12 gives an example of the third embodiment of the invention in an application to taking hands-free sound for interactive communication contexts (teleconferencing, communicating personal computers, etc.). In such a context, the acoustic echo, that is to say the signal of the distant speaker emitted by the loudspeaker constitutes the interfering signal. The disturbance signal e is taken directly from the loudspeaker.

We will present this example in the context of digital signal processing in the frequency domain. The temporal samples are indexed by n representing the index temprel at discrete time. The representation in the frequency domain is ensured by a discrete Fourier transform inside a sliding time window. The signals x, (") received by the microphones 1200 _l ., 1200 _N are subjected to a weighting in the time window by means of the filters 1205 _l ., 1205 _N then to a discrete Fourier transform (TFD) in the short term in 1207ι, .., 1207 _N - At the output of the TFD modules, there is a temporal representation X /, ω _q ) with:

^ U ^) = Σ (- "K ^ +" to "T ^o = q, ..., Ml (17)

'= 0

where the frequency axis has been uniformly discretized: ω _q = 2τvqjM, q - 0, ...,

MA with length of the analysis window (in samples); where the h _a (n) are the weighting coefficients within the analysis window, R is the window shift step (in samples) and p is the frame index; and where ^ = e ^ ^M = \ Conversely, at the output of the reception system, the signal Z ₂ (p, <) is transformed into a temporal representation by means of a module 1270 of inverse discrete Fourier transform (TFDI), that is to say :

where h _s (ή) are the weighting coefficients within the summary window.

The reference signal e taken from the loudspeaker is subjected like the microphone signals to a short-term Fourier transform in 1208 after weighting in 1206. The representation of e in the frequency domain, as obtained at the output of 1208 , is written with the previous notations:

E {p, ω _q ) = ∑h _a (-n) e (pR + n) W- '^>"q = Q, ..., MA (19) n = 0

The signal at the antenna output is obtained by performing the sum in 1220 of the input signals previously filtered by the channel filters 1210, -, i = l, .., N, that is:

The filtering of this signal by the post-filter depends on the state of the binary signal K (p). If K (p) = 1, i.e. if the useful signal is present, the output signal is expressed, in frequency representation:

Otherwise, K (p) = 0, i.e. if the useful signal is absent, the output signal is expressed:

where Gs _A (3 _q ) is the inverse of the frequency response of the coupling between the loudspeaker and the acoustic antenna. Indeed, for some applications, the sound pickup system and the speaker are located at fixed locations. The coupling between the two transducers is then obtained by a prior measurement. Thus the gain Gs _A (a> _q ) constitutes a fixed compensation filter which removes part of the disturbance.

The estimation of the transfer functions of the post-filters W _e (p, ω _g ) and W _s (p, ω _q ) is carried out respectively by the estimation modules 1251 and 1241. If the estimates are made by (1) and (12), we have:

The values of W _e (p, ω _q ) thus obtained are always positive and bounded between 1 and a value typically less than 1 predetermined and which corresponds to the maximum attenuation of the interfering signal. W (p, ω _q ) is calculated as a product (in 1225) of W _e (p, ω _q ) and W _s (p, ω _q ). To limit the effect of estimation errors and to avoid unwanted amplifications, W (p, ω _q ) is in practice clipped so as to belong to the interval [-1; 1].

For the estimate Φ ^c _e (t, f) in (23), we use the recursive equation (2), that is to say:

? (p, »= αΦΞ ^' (pl,»,) + (l-α) E (p ^ ^, (p, ^) (25)

where a is a short-term time constant. For the estimate Φ (?, /), We use, as we saw in (3) and (4) a long-term time constant ai or short term αr ₂ , depending on whether Φ _∞ (t, f) is less than or greater than :

If Φ ^c > Φ ^L _e î then ) + (l - _αι ) E (p, _ω E ^* ( _? , _ω

(26)

(27)

It should be noted that a _\ is more much closer to 1 than α. Concretely, it is advantageous to choose for the proposed realization a _\ - 0.9999 and a% = 0.9.

In the present example, the detection of the useful signal is carried out on the basis of a statistical analysis of the frequency values of the filter W (p, ω _g ) alone. The functional diagram of the statistical analysis module 1261 is given in FIG. 13.

We extract in 1321 from W (p, ω _q ) a set of frequencies F _cp fixed by the user. Typically, a frequency band is chosen where the signal to be detected is well represented, so as to have optimal detection and avoid costly calculation over the entire band. For speech signal, E ₀ ^ - = [lkHz; 3 kHz]). The components at the frequencies of F ^ _p are thresholded in 1322 by means of a predetermined threshold SOC ^w . Among the set of frequencies F ^ _p thus retained in for which wlp, F λ exceeds a SOC ^w threshold, that is:

The occupancy rate τ _o ^w _cp (jA) is then compared to an occupancy threshold STOC ^w in the comparator 1324. The comparator delivers a binary indicator p_μ ^W (p) of presence of useful signal which here is none other that the global indicator P_A (p) since here only the transfer function W (p, ω _q ) is used for the detection of useful signal. In summary :

P_A (p) = l if τ _o ^w _cp (p)> STOC ^w (29)

P_A (p) -0 if τ _p (p) ≤ STOC ^w (30)

The diagram of the gain smoothing filter 762 is given in FIG. 14. As we have seen, the function of this filter consists in decreasing continuously towards a predetermined value S _min the gain G (p) during the transition from state P (P_A (p) = 1) to state A (P_A (p) - 0) and to make it grow continuously towards a predetermined value S _max (fixed here at 1 to ensure gain transparency in the presence of the useful signal) during the passage in the opposite direction. In this example, the gain smoothing filter is digital. The gain G (p) is smoothed by a recursive filtering conditioned by the state of P_A (p):

G {p) = β _P G (p - \) + (l -β _P ) S _max if P_A (p) = \ (31)

G {p) = β _A G {- \) + (\ -β _A ) S _min if PjL {p) = Q (32)

The quantities βp and β_, are time constants which respectively determine the rates of ascent and descent of G (p). More precisely, the predetermined gains S _min and S _max are switched by the binary indicator P_A (p) to the common output of the switch 1410. This output is connected on the one hand to the input of the recursive filter 1420 of time constant βj. and at the input of the recursive filter 1430 of time constant β. The outputs of the two filters are switched by switch 1440, controlled by the binary indicator P_A (p). Returning to FIG. 12, the antenna output switching is provided by the switch 1230 controlled by the signal K (p). It makes it possible to apply the antenna output signal Y (p, ω _q ) either to the post-filter 1240 or to the post-filter 1245. The switching signal K (p) is obtained by comparison of the smoothed gain G (p ) at a predetermined threshold ST, that is:

Kp) ≈ l if G (p)> ST (33)

K (p) = 0 if G (p) ≤ST (34)

Although the invention has been essentially described, for reasons of clarity of presentation, in the form of functional modules each performing a function, those skilled in the art understand that in practice, these modules can be produced by means of a processor single executing these different functions or by means of a plurality of processors, specialized or not, executing one or more of said functions.

In addition, it is important to emphasize that the invention is not limited to the elimination of acoustic echo in a multi-sensor sound pickup system, but generally applies to any multi-sensor reception system having a reference of one or more interference signals.

Claims

1) Interference reduction method for a reception system using a multisensor antenna and at least one channel trainer supplying an antenna signal from the signals received by the various sensors of said antenna, characterized in that estimates, from a reference signal making it possible to regenerate said interference, the transfer function of a first filter, called the first post-filter, and that said antenna signal is filtered by said first post-filter.

2) Interference reduction method according to claim 1, characterized in that said transfer function is obtained from a short-term estimate and a long-term estimate of the spectral density of said reference signal.

3) Interference reduction method according to claim 2, characterized in that said transfer function is obtained from a ratio between the long-term estimate and the short-term estimate of said spectral density.

4) Interference reduction method according to claim 2 or 3, characterized in that the short-term estimate and the long-term estimate of the spectral density are obtained by low-pass filtering of a spectrum of the signal of reference.

5) Interference reduction method according to claim 4, characterized in that the short-term estimate Φ ^ (t, f) of the spectral density is obtained by recursive filtering of the type: where E (t, f) is a spectral component of the reference signal at frequency and time t, a is a coefficient between 0 and 1, δt is the delay in the loop of the recursive filtering and. denotes the conjugation operation.

6) Interference reduction method according to claim 5, characterized in that the long-term estimate Φ ^ ^r (t, /) of the spectral density is obtained by a recursive filtering of the type:

Φ f) = a _x Φ ^ {t-δt, fy (\ - a) E (t, f) E \ t ^ si Φ? _e {t, f) <Φ ^c _e (t, f) and where i and or ₂ are coefficients such that 0 <α ₂ <"ι <l.

7) Interference reduction method according to one of the preceding claims, the signals received by the various sensors being filtered by at least one set of channel filters before being summed to provide said antenna signal, characterized in that the transfer function of a second filter, known as the second post-filter, is estimated from said signals received, before or after filtering by said channel filters, and that the antenna signal is filtered by said second post-filter.

8) Interference reduction method according to claim 7, characterized in that the transfer function of said second post-filter is estimated from an average of the spectral power densities and from an average of the interspectral power densities of said signals received, after filtering by said channel filters.

9) Interference reduction method according to claim 8, characterized in that the transfer function W _s (t, f) of said second post-filter is estimated by:

where Φ _{V (V /} (t, /) and Φ _vv (/, /) are respectively estimates of the spectral densities and interspectral power densities of the signals received after channel filtering, b, (f) are the transfer functions of the different channel filters removed from the phase-back terms, N is a number of antenna sensors and γ (.) denotes the real value or the module.

10) Interference reduction method according to one of claims 1 to 6, the signals received by the various sensors being filtered by at least one set of channel filters before being summed to provide said antenna signal, characterized in that the transfer function of a second filter, known as the second post-filter, is estimated from said signals received after filtering by said channel filters, as well as from the antenna signal and that the the antenna signal is filtered by said second post-filter.

11) Interference reduction method according to claim 10, characterized in that the transfer function of said second post-filter is estimated from an average of the interspectral power densities of said received signals, after filtering by said channel filters , and an estimate of the spectral density of the antenna signal.

12) Interference reduction method according to claim 11, characterized in that the transfer function W _s (t, f) of said second post-filter is estimated by:

where Φ _vv (, /) and Φ (t, /) are respectively the spectral and interspectral power densities of the signals received after channel filtering, bι (f) are the transfer functions of the different channel filters rid of discount terms in phase, N is a number of antenna sensors and γ (.) denotes the real value or the module.

13) Interference reduction method according to one of claims 7 to 12, characterized in that the filtering of the antenna signal by the first post-filter and that of the second post-filter are applied in a combined manner, by filtering the antenna signal by means of a post-filter, said first combined post-filter having for transfer function a combination of the transfer functions of said first and second post-filters. 14) Interference reduction method according to claim 13, characterized in that a statistical analysis is carried out of spectral components of the transfer function of the second post-filter and / or of the transfer function of the first post-filter combined and that an indication of the presence or absence of a useful signal is deduced therefrom.

15) Interference reduction method according to claim 14, characterized in that the statistical analysis is also carried out on spectral components of the transfer function of the first post-filter.

16) Interference reduction method according to claim 14 or 15, characterized in that said statistical analysis uses a criterion of spectral occupancy rate and / or variance of said spectral components.

17) Interference reduction method according to one of claims 14 to 16, characterized in that a switching signal is generated from said indication of presence or absence of useful signal and that the antenna signal by means of the first combined post-filter when the switching signal is in a first state and when it is filtered by means of a second combined post filter when the switching signal is in a second state, the function of transfer of the second combined post-filter being a combination of the transfer function of the first post-filter and of a predetermined attenuation.

18) Interference reduction method according to one of claims 7 to 12, the received signals being filtered by a plurality of sets of channel filters to form a plurality of channel signals, characterized in that a statistical analysis of spectral components of the transfer functions of the second post-filters associated with the different sets of channel filters and that the channel with the highest probability of the presence of a useful signal is deduced therefrom.

19) Interference reduction method according to claim 18, characterized in that the statistical analysis is also carried out on spectral components of the transfer function of the first post-filter. 20) Interference reduction method according to claim 18 or 19, characterized in that said statistical analysis uses a criterion of spectral occupancy rate and / or variance of said spectral components.

21) Interference reduction method according to one of claims 18 to 20, characterized in that the antenna signal is obtained from the channel signals relating to the channel having the highest probability of presence of useful signal.

22) Reception system comprising a multisensor antenna, at least one channel trainer and interference reduction means, characterized in that said interference reduction means comprise a filter at the output of the channel trainer, known as the first post- filter and means for estimating from a reference signal making it possible to generate said interference the transfer function of said first post-filter.

23) Reception system comprising a multi-sensor antenna, at least one channel trainer and interference reduction means, characterized in that said interference reduction means are suitable for implementing the interference reduction method according to one of claims 1 to 21.