WO2010029247A1

WO2010029247A1 - Low-distortion noise cancellation

Info

Publication number: WO2010029247A1
Application number: PCT/FR2009/051670
Authority: WO
Inventors: Ludovick Lepauloux; Claude Marro; Pascal Scalart
Original assignee: France Telecom
Priority date: 2008-09-12
Filing date: 2009-09-04
Publication date: 2010-03-18
Also published as: FR2936116A1

Abstract

The invention relates to a signal processing, in a backward structure including at least first (W₂₁) and second (W₂₁) filters arranged between a main path (V1) conveying a first signal (X_p(z)) and a reference path (V2) conveying a second signal (X_p(z)), for reducing in the first signal (X_p(z)) the noise estimated in the second signal (X_r(z)). At least a first adaptive filtering calculation (W1) is applied based on a difference estimated from the first and second signals taken upstream from the first and second filters in said main and reference paths. The first filtering calculation is then imposed to one (W₂₁) of the first and second filters for updating.

Description

Low distortion noise cancellation

The present invention relates to the processing of a sound signal, especially after its acquisition.

A typical problem of sound recording concerns the acoustic environment in which a sound pick-up microphone is placed which, for example, can be integrated into a telephone. Due to the impossibility of completely controlling this environment, a disturbing signal (referred to as "noise") is present within the observation signal, including the really useful signal. In order to improve the quality of the signal possibly transmitted to the remote party, to increase its intelligibility and to reduce the correspondent's fatigue, it is important to develop noise reduction processes whose purpose is to extract useful information. by performing a treatment on the noisy observation signal. In addition to speech communication applications, the improvement of the speech signal quality is also necessary for voice recognition, the performance of which is greatly altered when the user is in a noisy environment.

In order to deal with this problem, solutions have already been studied and applied, such as adaptive algorithms (for example "stochastic gradient", LMS (least squares) or RMS (least recursive squares), or other). In these latter algorithms, it is sought to produce an estimate of the impulse responses connecting the sources of noise and speech to the sound pickup means (consisting of several sensors) in order to better subtract the disturbing signals from the useful speech signal. To do this, adaptive filters driven by adaptation steps, most often variable, are implemented. The control of these steps is difficult because it is necessary to allow the adaptation of the filter with a high learning rate only during the phases of noise alone, and reduce it, see cancel, as soon as the local speaker is active, in order to limit the distortions introduced during the convergence of the filters. However, this phenomenon is not the most critical. Since the development of adaptive filtering techniques, the major challenge for achieving tangible efficiency in a real environment lies in the ability to provide the sound acquisition means with a valid noise reference. The validity of the noise reference is evaluated in the degree of correlation that exists between the disturbing noise of the speech signal and that which is taken for image by means of a second microphone that advantageously comprises the aforementioned sound recording means. It is easy to understand that the similarity between the reference noise and the disturbing noise will be all the greater as the microphones will be placed sufficiently close to one another. This proximity constraint of the sensors on the acoustic front end gives rise to the phenomenon of leakage of the useful signal, called "crosstalk", at the origin of the main deteriorations on the signal produced at the output of a noise reduction system. In addition, this (necessary) proximity is also made necessary to limit the length of the adaptive filters calculated over time, in order to reduce the adjustment problems producing in particular echo effects.

Crosstalk-resistant noise reduction techniques in the context described above have already been proposed. The implementation of these existing techniques described later in the description generally leads to the addition of an additional filter for learning the signal contaminating the noise reference. They may also consist of introducing a block allowing the calculation of the adaptation step in order to better adjust the learning rate, which amounts more or less to developing a voice activity detection (or "DA V" below). ).

Some techniques use, in particular and separately, a detection DAV coupled to a stage of determination of the pitch. These basic architectures are divided into two main classes: the so-called "forward" and "backward" techniques. These techniques are used for the development of blind source separation algorithms. It is proposed below to rely on a backward type technique shown in Figure 2 and described below. Nevertheless, it is then necessary to solve a problem of the coupling existing between the filters and in particular also the noise problem of the gradient introduced by the standard least squares algorithm generally used, as explained hereinafter. The state of the known art and the problems it poses are now described in detail.

In general, FIG. 1 can correspond to a conventional adaptive noise canceller scheme using two microphones C1 and C2 which both pick up a useful signal S (z) (for example of speech) and a disturbing signal N ( z) (ambient noise for example), where z denotes the variable of a transform in Z. It will be understood that the position of the sensors C1 and C2 is judiciously chosen so that one more captures the ambient noise (C2 in the example shown in Figure 1) while the other captures more the useful signal (C1 in Figure 1). In Figure 1, the reference:

- X _p (z) denotes the signal from the first microphone C1 on the primary or "main" channel below (the index p designating this main channel) comprising the useful signal to which the noise is added,

X _r (z) is the signal from the second microphone C2 on a reference channel (the index r denoting this reference channel) comprising the noise to which is added the useful signal,

Hj _j (z) are acoustic channels of unknown lengths, variable in time,

W (z) is an identification filter of length L, and

S '(z) is a useful speech signal to be estimated.

Figure 1 then illustrates the problem below. The two-sensor sound recording system actually records two signals X (z) and X _r (z). So, we will understand that each of these signals X _p (z) and X _r (z) is the result of the sum of the convolutions of the signals coming from the sources: S (z) for the useful signal and N (z) for the noise, with the channel associated with the path connecting it to each of the microphones, ie: Xp (z) = H ₁ i (z) * S (z) + H ₂ , (z) * Ν (z) and X _r (z) = H ₁₂ (z) * S (z) + H ₂₂ (z) * N (z)

W (z) is a filter of length /,, whose coefficients | w (/)} _. _ are adapted in time, and whose purpose is to generate a noise estimate N \ z). This estimate requires, according to Wiener's theory, that the filter W (z) "learns" the impulse response H ₂₁ (z). Unfortunately, in the structure of Figure 1, nothing is done to estimate the crosstalk from the convolution between S (z) and H ₁₂ (z). Indeed, the output signal S '(z) is finally expressed only by an expression of the type: S' (z) = X _p (z) - W (z) X _r (z).

However, the presence of this "signal" of crosstalk provides an optimal solution different from that of the classical case considering H ₁₂ (z) as zero. Indeed, by assuming that H _n (z) = H ₂₂ (z) = 1, we show that the optimal solution taking into account the leakage of the speech signal through H _u (z) becomes: w _{oPl {z) (1)}

where Φ _m (z) and Φ _≈ (z) are the spectra of the N (z) and S (z) signals in the Z transform domain.

Thus, the cancellation of the noise can not be done completely. In addition, we can consider that the crosstalk then dominates equation (1) above, which boils down to:

In this case, the signal at the output of the system E (z) resulting from the subtraction between X _p (z) and W '"" (z) X _r (z) can be written as:

E (z) = X _p (z) - W ^op '(z) X _r (z)

≈ S (z) + H ₂ Az) N (Z) ¹ - - N (z) - ^^ - S (z) (3)

H ₁₂ (z) H ₁₂ (z)

≈ (H _1x [Z) - ^ -) N (z)

H ₁₂ (z)

The optimal solution therefore leads to complete cancellation of the signal, which is not satisfactory. A simple way to overcome this problem is to stop the adaptation when the useful signal is present, or to design structures explicitly taking into account the leakage of the useful signal (or "crosstalk signal") in providing a second adaptive filter, such as the "backward" structures described hereinafter.

Adaptive filtering consists of giving an expression of evolution of the coefficients of the filter over time that satisfies a convergence criterion. Several algorithms are used in noise cancellation (LMS, NLMS (standardized least squares), RLS, and others). The adaptation equation of the simplest of them (the LMS) is given hereafter by way of example, for which the criterion retained is the minimization of the power of the residual error and we have:

W (n) = W (n-1) + μ-e (n) - X _r (n), where: (3)

- W (n) = Fw ₀ (W ^ w ₁ (n), - - -, w _L _ _{ («) 1 is the vector of the L coefficients of the adaptive filter at time n, - X _r (n) =

is the vector of the last L samples present on the reference microphone, and

- μ is a factor called "no adaptation" which controls the speed of convergence. The bigger the step, the faster the convergence will be. Therefore, the more the signal contains useful information for the adaptation of the filter, the more it is necessary to accelerate the convergence and more then the chosen step will be large. It will be understood that the role of the pitch is important in controlling the stability of the filter. In cases where only noise is present, the filter can be adapted to converge rapidly. In the absence of noise, the adaptation of the coefficients is not desirable because it can lead to the maladaptation of the adaptive filter, and finally to noticeable echo feedback and the birth of distortion. It is therefore very often desired to freeze the filter in this case.

In the opposite case, the filter tries to suppress the useful word and becomes unadapted. In addition to the risk of divergence of the filter, this leads to significant degradation of the useful signal, or even their amplification. This is also observed for moments when the speech predominates, that is to say when the signal is slightly noisy and the filters are close to their convergence value.

Among the backward crosstalk management algorithms, those using dual filters are already known. Here, the addition of the filter W _v (z) makes it possible to subtract the crosstalk signal from the noise reference created by the other filter W _2X (z) and thus to reduce the distortions on the estimated useful signal S "(z) The objective of this second filter is therefore to estimate the response of the acoustic channel H _n (z) adaptively, in exactly the same way as for the noise through the filter W _n (z). symmetrical and the respective positions of the sources of noise and speech may not be known a priori, there is shown in Figure 2 such dual filters, with estimation of a reference channel for each type of signal, with the same notations than those in Figure 1.

To define an expression that satisfies the stability conditions of the adaptive filters, the power received on each of the channels is used to determine a normalized value of the adaptation step. Thus, an expression not μ _t that satisfies the desired behavior can be defined by a pair of constants (α ,, ^), noting σ, ² (") the signal strength considered at time n on the channel /, as follows:

where σf (n) = (l - a) σf (n - 1) + aε _j ^J (n) (6) and corresponds to an exponential smoothing. The term ε (n) represents a power of the signal on the channel y and entering the filter which processes the signal of the channel /. The index j evolves symmetrically with the index /, that is to say that the calculation of the power on the channel i = p associated with the signal X _p (z) involves the error calculated at the same time on the channel / = r associated with the signal X ₁ . (z), and vice versa.

Approaches have been dedicated to improving stewardship. It is not, strictly speaking, crosstalk management algorithms but rather to establish the expression of a variable adaptation step that allows to manage the adaptation of filters in any configuration. However, the basic architecture remains similar to that of Figure 2.

It was then proposed in the document Ikeda et al:

"An adaptive noise canceller With low signal-distortion in the presence of crossta \ K" by S. Ikeda and A. Sugiyama, in "IEICE Transactions on Fundamentals", Volume E82-A, No. 8, August 1999, pages 1517- 1525, a technique designed to change the pitch μ in an interval [μ _mm , μ _max ] as a function of the energy levels present on each of the microphones and the filter with which they are associated. In the opposite way, for one, it is to learn the noise and for the other, the signal of crosstalk.

For the step associated with the filter W _2x (z), the behavior is as follows: Μ (n) tends to a maximum μ _max in noise period alone, in order to converge the filter W ₂₁ , and

Μ (n) tends to a minimum μ _min in the useful speech phase alone, in order to stabilize the estimate of the response H _2i (z) made by the filter JF ₂₁ and thus not to disturb the learning of the noise by the speech signal leaking into the noise reference.

• Otherwise, μ (n) e [μ _mm , μ _max ] in noisy speech phase. It is necessary that μ (n) tends to μ _mm if the useful speech is preponderant in front of the noise and that μ (n) tends to μ _maJl in the opposite case.

The inverse reasoning applies to the step associated with the filter W _i2 (z). The threshold values of these steps are taken for example in the interval [0, 0.1]. They are assigned according to values of the signal-to-noise ratios (SNR) measured (or calculated from the measured energies) on each of the channels, these ratios being themselves thresholded. Finally, a simple way to comply with this behavior is to use, for example, an affine function that describes the evolution of the step (then linear evolution), in this interval using the values of the RSB ratios calculated over time. as well as the thresholds associated with them.

It has been proposed in US-6,700,976 (Zhang et al) to reduce the complexity of the structure implemented by Ikeda et al for the calculation of the pitch, using a calculation-based voice activity detection (DAV). the power and the correlation of the signals, received and output filters, to control a decision system for selecting, via a counter, the optimum value of predefined steps, depending on the context. In particular, the implementation of this technique is to vary the step discretely and not more continuously outside the threshold areas. In the document, six different values are used, which are indexed by the two types of DAVs artwork. The adaptive filters necessary for the calculation of the SNRs of the technique of Ikeda et al are thus removed.

However, these techniques have disadvantages.

The first of these methods (Ikeda et al) suffers from a major disadvantage: the noise introduced by the gradient processing algorithm which, because of the existing interdependence on the convergence of the filters, is amplified and thus prevents good convergence of filters W ₂₁ (z) and W _n (z). To overcome this problem, the document: "Improved adaptive noise cancellation in the presence of signal leakage on the noise reference channef \ MJ Al-Kindi and J. Dunlop, Signal Processing, Vol 17, No. 3, pages 241-250 (1989) proposes, in particular, to introduce, where possible, prior knowledge that comes down to considering that one of the filters is known in advance, and it is proposed to use a DAV to prevent joint adaptation of the two filters.

The other known methods, as for them, suffer, besides this problem stated above:

on the one hand the complexity they introduce and,

- on the other hand, a free choice difficult on the value of the thresholds to be imposed for the steps.

However, these can not be chosen at random. If the technique proposed by Zhang et al reduces the complexity in terms of calculation cost compared to the method of Ikeda et al, there are however many thresholds to be fixed, and this, for each of the blocks used (four DAV and those to be defined for the system of decision fixing the values of the steps to be attributed). In addition, in a very adverse context, the operation of the VAD is far from optimal and the use of a VAD often becomes a very disadvantageous handicap because the false detections cause very marked distortions on the output signal.

Finally, it is necessary to have, at the beginning, at least one relatively favorable reference path, that is to say a path with a very high signal-to-noise ratio or, if not very low, so that one filters can be initialized with precision, allowing the adaptation of the other, otherwise the adaptation remains difficult.

The present invention improves the situation.

For this purpose, it first proposes a signal processing method, in a structure comprising at least a first and a second filter, arranged between a main channel carrying a first signal and a reference channel carrying a second signal, and in a backward configuration to reduce, in the first signal, the estimated noise in the second signal.

The term "backward configuration" is understood to mean that the structure has an architecture similar to that illustrated in FIG. 2 and previously described. However, within the meaning of the invention, at least one of the first and second filters mentioned above is not a simple adaptive filter within the meaning of the prior art described above with reference to FIG. rather a first filtering calculation, adaptive as a function of an estimated difference as a function of the first and second signals taken upstream of the first and second filters in the main and reference channels, this first filtering calculation then being imposed on one first and second filters for updating this filter.

An advantage of the invention is that, if an adaptation becomes erroneous in one of the first and second filters, it does not taint the filtering calculation that is imposed for the other of the first and second filters. In particular, this filtering calculation is preferentially adapted as a function of the subtraction of one of the first and second signals, by the other of the first and second signals and to which the calculated filtering is applied.

Advantageously, it is possible to choose an adaptation step of the filtering calculation (to accelerate or not the convergence of the filtering calculation) as a function of a signal-to-noise ratio estimated on at least one of the main and reference channels. It will therefore be possible to choose the adaptation steps of the one and the other of the first and second filters independently, simply as a function of the adaptation conditions of the filters and in particular as a function of the signal-to-noise ratios on the main channels and reference. It is therefore possible to provide a measurement of the signal-to-noise ratio on at least one of the main and reference channels, or to estimate this ratio from measurements of other parameters such as energy or power, as will be seen later. .

For example, the adaptation for noise filtering can be fixed, without updating the corresponding filter, if the measured current noise is low compared to the useful signal (which corresponds to an estimated signal-to-noise ratio which is high) . This implementation makes it possible not to taint an error adaptation of this filter which is not feasible under optimal conditions.

Thus, in an advantageous embodiment, the filtering calculation and the updating of the first filter are implemented only if a signal-to-noise ratio, estimated on one of the main and reference channels, has a value in a chosen range, for example greater than a chosen threshold.

It is possible to provide a single filtering calculation imposed on one of the first and second filters. However, in a more sophisticated embodiment, the method of the invention may comprise:

a first filtering computation, adaptive as a function of a first estimated difference as a function of the first and second signals taken upstream of the first and second filters in the said main and reference channels, and a second filtering calculation, which is adaptive as a function of a second difference estimated as a function of the first and second signals taken upstream of the first and second filters in the said main and reference channels,

said first filtering computation being imposed on one of the first and second filters and said second filtering computation being imposed on the other of the first and second filters.

Thus, in the context of the treatment of the signal leak (or "crosstalk") in a system with at least two sensors, noise reduction, the implementation of the invention makes it possible to avoid the effects of a strong existing coupling between the outputs of conventional adaptive filters and brings a certain improvement in the adaptation. Indeed, tests have shown that the implementation of the invention makes it possible to provide crosstalk management in a number of adverse situations, in particular:

Different signal-to-noise ratios (SNRs) on each of the channels, which assumes that the reference sensor is isolated and therefore relatively well protected from the crosstalk signal,

• or identical RSB reports on each channel,

• or strong noise conditions,

• or a very spectrally colored channel with many reflections, room reverb effects, or others.

Other features and advantages of the invention will appear on examining the detailed description below, and the accompanying drawings in which, in addition to Figures 1 and 2 described above and relating to the prior art:

FIG. 3 illustrates the architecture of a device in the sense of the invention which, in the example shown but not limiting, comprises two sensors, being understood that the invention can be implemented in a device with more than two sensors,

FIG. 4 is a flowchart illustrating, by way of example, the steps for calculating the steps of the adaptive filters W 1 and W 1 in FIG. 3,

FIG. 5 illustrates the architecture of a device according to an embodiment of the invention, less sophisticated than that represented in FIG. 3, and

FIG. 6 illustrates, by way of example, the material elements that may comprise a device in the sense of the invention.

With reference to FIG. 3, a device within the meaning of the invention comprises:

at least two sensors C1 and C2, connected to two respective channels V1 and V2,

a first adaptive filtering W _t (z), as represented, for generating a noise estimate, preferably according to an adaptive algorithm (LMS, NLMS, APA, or other, as described below),

a second adaptive filtering W ₂ (Σ), as shown, dedicated to learning the crosstalk signal, preferably according to an adaptive algorithm (LMS, NLMS, APA, or other),

and in particular, a copy of these adaptive filters W _x (z) and W ₂ (z) (dotted lines of FIG. 3) in respective homologous filters W ₂ Φ) and W _n (z), fixed (c '). that is, non-adaptive) and placed in a backward configuration.

Thus, in more generic terms, the structure of FIG. 3 comprises at least:

a first sensor C1 and a second sensor C2, respectively capturing the first signal Xp (z) and the second signal Xr (z), and respectively connected to the main channel Vl and the reference channel V2, at least one first filter W21 of the second signal Xr (z), whose output is intended to be removed from the first signal Xp (z), and

a second filter W1 2 of the first signal Xp (z), the output of which is to be removed from the second signal Xr (z).

In the embodiment of the invention illustrated in FIG. 3 by way of nonlimiting example, upstream of the first and second filters in the main channels V1 and reference V2 are carried out:

a first filtering calculation W1, which is adaptive as a function of the difference between the first signal Xp (z) and the second signal Xr (z) to which the first computed filtering W1 is applied, and

a second filtering calculation W2, which is adaptive as a function of the difference between the second signal Xr (z) and the first signal Xp (z) to which the second calculated filtering W2 is applied.

The first filtering calculation W _\ is then imposed on the first filter

to update its parameters and the second filtering calculation W ₂ is imposed on the second filter W] ₂ to update its parameters.

Advantageously, the adaptation steps of the first and second filtering calculations are distinctly chosen. In particular, as will be seen in detail below with reference to FIG. 4, in the example shown in FIG.

a step μW2 of adaptation of the second filtering calculation varying like the signal-to-noise ratio estimated on the reference channel V2 (hereinafter referred to as RSBr), and

a step μW1 of adaptation of the first filtering calculation varying as the inverse of the estimated signal-to-noise ratio on the main channel Vl (denoted hereinafter RSBp). As will also be seen with reference to FIG. 4, the computation of the first filter W _\ and the update of the first filter Wι _\ are implemented only if the signal-to-noise ratio RSB _r , estimated on the reference channel V2, has a value less than a threshold value (denoted RSBw _Σ min).

It will then be understood that the algorithmic constraint is imposed here so as not to adapt one of the filtering calculations W ₁ (Z) or W ₂ (z) if the conditions are not optimal for this purpose. For example, when the decision to adapt the filter W ₂ (z) for learning the crosstalk is taken (more useful signal than noise), then the adaptation of the filter W _x (z) can be frozen ( not enough noise to develop his learning).

Moreover, since the allocation of the values of the adaptation steps is a function of the measured signal-to-noise ratios (SNRs), an appropriate adjustment of the comparison thresholds of the signal-to-noise ratios is provided. Advantageously, this adjustment is made so that the non-stationarity of the signal can be followed. The objective here is to make the behavior of the pitch associated with the filter W _x (z) as sensitive as that of a voice activity detector. In generic terms, a possible implementation of this embodiment may consist, as will be seen in detail with reference to Figure 4 below, to provide: - an adaptation step (noted μW2 further) of the second filtering calculation , for the crosstalk, constant for a signal-to-noise ratio lower than a first threshold value (denoted RSBW2min), then varying as the signal-to-noise ratio for a signal-to-noise ratio greater than this first threshold value, then constant for a signal ratio with noise greater than a second threshold value (denoted RSBW2max and greater than the first threshold value denoted RSBW2min), and

the adaptation step (denoted μWl below) of the first filtering calculation, for the noise, constant for a signal-to-noise ratio lower than a first threshold value (noted later RSBWlmin), then varying as the inverse the signal-to-noise ratio for a signal-to-noise ratio higher than the first threshold value, then constant for a signal-to-noise ratio greater than a second threshold value (denoted RSBW lmax and greater than the first threshold value RSBW 1min).

These threshold values mentioned above will then be chosen judiciously for optimum operation of the device within the meaning of the invention.

The aforementioned calculations of the adaptive filtering W _x (z) and W ₂ (z) are detailed below.

A purpose of the first adaptive filtering W _x (z) is to identify the channel and track its variations over time to provide a noise estimate. Its implementation can be based on an adaptive algorithm of the type:

"Least squares" or LMS for "Least Mean Square", or

"Least standard squares" or NLMS for "Normalized LMS",

"Affine projection" or APA for "Affine Projection Algorithm",

- or others.

The management of the adaptation of this filter can involve a law as indicated above and according to which, in a situation of leakage of the useful signal, the adaptation of the filtering W _x (z) is blocked (which amounts to imposing μ = 0 in equation (3) given previously for an LMS algorithm). Of course, other adaptation management laws are also possible.

The implementation of the invention makes it possible to overcome the problems related to the prior art.

Indeed, once the filtering W _x {z) calculated, it is then copied into a homologous filter W _2x (z) which itself is fixed (non-adaptive). So, every moment: W ₂₁ [Z) = W _x (Z) (5)

By realizing the adaptation of the filterings in this way and in particular according to the architecture of FIG. 3, the coupling problem indicated previously disappears. To be convinced of this, the output S '(z) of FIG. 2 in the sense of the prior art was expressed, at a time n, by:

^S '( ⁿ ) ^{ZZ X} _P ( ⁿ ) - 2l ^W 2 ^ε r ( ⁿ - ^{' i} ) ( ⁶ ) ι = 0

L ₂ -I with the term e _r [n) = x _r (n) - ^ w _n s' [n - i) which shows the interdependence between

/ -1 filters.

In FIG. 3 in the sense of the invention, the output S '(z) is expressed by:

_S '(n) = x _p [n) -Σ _Wι x _r (ni) (?)

Interdependence is removed and decoupling is achieved. As a reminder, the index / "above is relative to the primary path V] and the index r is relative to the reference path V ₂ of FIG.

Thus, assuming that only one filter is updated (eg W _2l (z) from

W _t (z)), the dual filtering (W ₂ (z)) is not active and a possible bad learning of one of the filters (W ₁ (Z)) does not affect its dual (W ₂ (z)) since, advantageously, the adaptation is made from signals which are independent, by construction of the architecture illustrated in FIG.

By the symmetry of the structure, the same reasoning can be held for obtaining an estimate of the crosstalk. In order to ensure that any adjustment faults in the noise learning filtering do not influence the convergence towards an optimal solution, the invention proposes to place the filtering W ₁ (Z) as shown in FIG. 3. As for the filtering W _x (z), it may be advantageous to base the implementation of the filtering W ₂ (z) on an adaptive algorithm of the LMS, NLMS, APA or other type.

The management of the adaptation of this filtering can involve a law as indicated previously: in situation of noise only, the adaptation of the filtering W ₂ (z) must be blocked (which amounts to imposing μ - 0 in the equation (3) given by way of example for an adaptive algorithm of the LMS type).

Once the computation of the filtering W ₂ (z) has been carried out, this latter is then copied into the homologous filter W _n (z), which itself is fixed. So, every moment:

W _u (n) = W ₂ (H) (8)

It should be noted, however, that the second filtering W ₂ (z) has been represented by way of example in FIG. 3: the use of two filtering calculations W _x (z) and W ₂ (z) in the same architecture although advantageous, it is not essential. It has been found that, depending on the noise conditions and any complexity constraints, the use of a single decoupling filter calculation W ₁ (Z) already allowed a clear improvement over the existing techniques.

With regard to the implementation of an effective algorithmic constraint, since strong signal components cause a significant gradient noise, the step is advantageously managed accordingly: in the example described, the value calculated at one instant is tested n for the steps μ (n) so as to decide whether the adaptation of the filtering W ₁ (z), for example, must be fixed or not. Typically, if the computed value of a step μ, (n) associated with the filtering W ₂ (z) dedicated to the learning and management of the crosstalk is different from its minimum value, it is then possible to force the pitch μ (n) associated with the filtering W _x (z) dedicated to the learning of the noise to have a null value, which freezes adaptation. This algorithmic constraint will be detailed below with reference to FIG. 4. Thus, the filters may not adapt together and, in particular in the example described, can not fit together if the crosstalk signal is significant. The importance of the crosstalk signal can be judged by the appropriately set RSB thresholds.

This calculation of the adaptation step is preferably carried out continuously, as will be seen below. The power received on each channel is advantageously calculated to determine the signal-to-noise ratios RSB. The thresholds set for the RSB ratios that make it possible to decide on the values that the different steps must take are advantageously chosen so that:

the filters follow the non-stationary signals, and

the step associated with the filter dedicated to learning the noise behaves like a noise detector.

Tests have shown that an advantageous solution for most of the target noise configurations is to choose values in the following proportions:

the minimum threshold of the ratio RSB _w associated with the first filtering W _x (z) may be close to 5 dB, while the maximum threshold may be close to 15 dB;

- Moreover, the minimum threshold RSB _W associated with the second filter W ₂ (z) may be close to -8 dB, while the maximum threshold may be close to 14 dB.

A process is described below with reference to the flowchart of FIG. 4. A first step 40 consists of measuring the signal-to-noise ratios RSB _P and RSB _r (or more precisely to calculate them from energies or powers). on each channel respectively Vl and V2. The first test 41 preferably relates to the RSB ratio _r and aims to compare it with a minimum threshold RSBw ₂ min. If it is near or below (output o of the test 41), then the step μwi of the computation of the adaptive filtering Wi is minimum (step 42) and the convergence is slow for a low noise useful signal ratio. It is possible to benefit from a low signal-to-noise ratio (relatively a lot of noise) to achieve the adaptation of the filtering calculation W _\ for the noise (steps 47 to 51), as will be seen below.

On the other hand, if the signal-to-noise ratio on the reference channel is greater than the threshold RSBw ₂ min (arrow n at the output of the test 41), the adaptation of the filtering W _\ is stopped by freezing its convergence and fixing its pitch μwi at 0 (step 43), then a suitable value of the pitch μw ₂ of the filtering W _% is calculated for the crosstalk signal. For this purpose, if the signal-to-noise ratio on the reference channel is greater than or equal to the maximum threshold RSBw _Σ max (arrow o at the output of test 44), then the step μw _{2 is set} to a maximum value μ _W2 max ( step 45) for fast convergence under good adaptation conditions (high signal-to-noise ratio). In the intermediate case where the signal-to-noise ratio RSB _r is of moderate value (arrow n at the output of the test 44), it is necessary to finely estimate the pitch (step 46) and, for this purpose, a function f, for example affine, can be considered, with:

uW ₂ = f (SNR _r) = a _r + SNR b, a and b being such that: μ = a RSBw inin _W2 _Σ min + b, and ₂ mW max = a + b RSBw _2max

It should be remembered that the adaptation of the filtering W _\ is carried out, together with that of the filtering W _% , only if μ _W2 ⁼ μw ₂ min (output o of the test 41). In this case, we now compare the signal-to-noise ratio RSB _P estimated on the main channel Vl to a maximum threshold RSBwimax (test 47) and if it is greater than this threshold (arrow o at the output of the test 47), which means that the signal-to-noise ratio is relatively high and that the convergence for the noise filtering W _\ must be as slow as possible, the pitch μwi is then set to the minimum μwimin (step 48).

Otherwise (arrow n at the output of the test 47), a suitable value of the pitch μwi of the filtering W _\ is calculated for the noise. For this purpose, if the signal-to-noise ratio RSB _P is less than or equal to the minimum threshold RSBw i min (arrow o at the output of the test 49), which corresponds to optimal conditions of adaptation of the filtering W \ for the noise, so we set the pace μwi at a maximum value μwimax (step 50) for fast convergence under good adaptation conditions (low signal-to-noise ratio). In the intermediate case where the signal-to-noise ratio is of moderate value (arrow n at the output of the test 49), it is necessary to finely estimate the pitch (step 51) and another affine function g can be envisaged, with:

μwi = f (RSB _p ) = a 'RSB _P + b', a 'and b' being such that μ _W imin = a 'RSBwimax + b \ and μwimax = a' RSBwimin + b '

It will be noted again that if μw ₂ ⁼ μw ₂ min, the pitch μwi is not set to 0 and a joint adaptation is possible of the two filters at once W \ and JF ₂ . However, it is possible, in a variant, to set the minimum μw ₂ min at 0 so as not to disturb the adaptation of the filtering JF ₂ under unfavorable conditions with little useful signal and a lot of noise and, in such an embodiment, it is forbidden in all cases a double joint adaptation.

Above, it is indicated that the RSB ratios are deduced from the power or energy measurements on the primary channels V1 and / or V2, but, of course, any other variant technique allowing the estimation of the RSB ratios on each ways can be considered.

The present invention also aims at a device for implementing the method described above and comprising at least a first and a second filter, arranged between a main channel carrying a first signal and a reference channel carrying a second signal, and according to a backward configuration to reduce, in the first signal, the estimated noise in the second signal.

In the device within the meaning of the invention, at least one of the first and second filters is updated by a first adaptive filtering calculation as a function of an estimated difference as a function of the first and second signals taken upstream of the first and second filters. and second filters in said main and reference channels. It will thus be understood that the parameters of this filter are regularly updated as the first filtering calculation is made.

FIG. 6 shows the material elements that can comprise a device of this type with in particular:

an input interface INT1 connected to the sensors C1 and C2 for receiving the signals X _p (z) and X _r (z),

a PROC processor (to execute a computer program in the sense of the invention as will be seen later),

a working memory MEM (which can store instruction codes of the aforementioned computer program),

and an output interface INT2 for outputting the useful signal S '(z).

An exemplary device capable of implementing the invention may for example be a sound recording equipment equipped with two or more microphones. The treatment described above can be programmed and integrated into the digital signal processing processor embedded in such equipment.

The device can also be a sophisticated computer or telephone terminal on which at least two microphones are connected to audio inputs. In this case, the computer program, an algorithm of which can be represented by the flowchart of FIG. 4, can be programmed in a language of the computer or of the terminal and integrated in memory of the latter.

The present invention is also directed to a computer program comprising instructions for implementing the method described above, when this program is executed by a processor.

Such a processor can realize in particular the first filtering calculation mentioned above, for example in a device within the meaning of the invention. By executing the program in the sense of the invention, the processor can perform other operations, such as in particular the dynamic determination of the adaptation steps, or the decision to freeze the adaptation, as has been seen previously with reference to FIG. may represent a flow chart part of the general algorithm of the computer program within the meaning of the invention.

Of course, the present invention is not limited to the embodiment described above by way of example; it extends to other variants.

For example, it is conceivable to dispense with one of the adaptive filters W ₁ or JF ₂ of FIG. 3 (for example the filtering W ₂ ). In this case, the homologous filter (the filter Wn {ι) in this example) becomes adaptive as in a classic backward structure. On the other hand, the dual filter (the JF ₂ i (z) filter in the same example) remains fixed (non-adaptive) and updated by the single head filtering calculation (W _\ in the same example). The present invention also aims at such a simplified embodiment, which is illustrated in FIG.

In the example shown in FIG. 5, the filtering calculation W _\ has been retained for the filter

Nevertheless, it could be planned to remove this filtering calculation and make the first filter

completely adaptive as in a conventional backward structure, but, on the other hand, to retain the filtering calculation W ₂ of FIG. 3 for the second filter W] ₂ .

In either case, it would be possible to set different values of signal-to-noise ratio thresholds for the determination of the adaptation steps according to the fact of providing only one or two filtering calculations upstream of the W ₂ filters. \ and Wn-

On the other hand, a fixed adaptation of the filter W _\ has been described above if the signal-to-noise ratio is not minimal. As a variant, it is possible to adapt the filtering W \ even if the signal-to-noise ratio is not minimal, but at regular time intervals over small time ranges below these time intervals. More generally, an implementation based on time ranges in selected time intervals can be envisaged at the same time for the adaptation of the filtering W _\ but also of the filtering W ₂ , the duration of the ranges relative to the duration of the intervals, for each adaptation, which can be chosen according to the estimates of signal-to-noise ratio on the channels Vl and V2.

The invention may be in the field of sound taking using at least two sensors, for example for applications according to the following non-exhaustive list:

Teleconference and videoconference in a noisy environment (in a dedicated room and / or from a multimedia computer, etc.);

Telephony, on conventional or mobile terminals, or others;

Telephony on hands-free terminals (desktop, mobile, or on-board mobile phones, or others);

Sound recording in public places (station, trains, airport, or other); Robust speech recognition to the acoustic environment;

Sound recording for the cinema and the media (radio, television, especially for journalism of sports events or concerts, etc.). However, more generally, the invention is also applicable to all the fields where a multisensor system is (or could be) used, in particular for:

Reception on a radio wave antenna (radiocommunications, atmospheric remote sensing, or others);

Sonar antenna reception (underwater imaging, underwater remote sensing, or other);

Multi-sensor treatment applied to biomedical signals (electroencephalogram, electrocardiogram, medical imaging, biomedical or other).

Claims

Signal processing method, in a structure comprising at least a first and a second (Wu) filter, arranged between a main channel (V1) conveying a first signal (X _p (z)) and a reference channel (V2) ) conveying a second signal (X _r (z)), and in a backward configuration for reducing, in the first signal (X _p (z)), estimated noise in the second signal (X ₁ - (Z)),

characterized in that the method comprises a first filtering calculation (Wi), adaptive as a function of an estimated difference as a function of the first and second signals taken upstream of the first and second filters in said main and reference channels, said first calculation filtering being imposed on one (Wn) of the first and second filters for updating.

2. Method according to claim 1, characterized in that the filtering calculation (Wi) is adapted as a function of the subtraction of one (X _p (z)) of the first and second signals, by the other (X _r (z)) of the first and second signals and to which the calculated filtering (W] is applied).

3. Method according to claim 1, characterized in that the method comprises:

a first filtering calculation (W {), which is adaptive as a function of a first estimated difference as a function of the first and second signals taken upstream of the first and second filters in the said main and reference channels, and

a second filtering calculation (Wi), which is adaptive as a function of a second estimated difference as a function of the first and second signals taken upstream of the first and second filters in the said main and reference channels, said first filtering computation being imposed on one (JF ₂ )) of the first and second filters and said second filtering computation being imposed on the other (JF ₁₂ ) of the first and second filters.

4. Method according to claim 2, in a structure comprising at least:

a first (C1) and a second (C2) sensor, respectively capturing the first (X _p (z)) and second (X _r (z)) signals respectively connected to the main channel (Vl) and the reference channel; (V2),

at least one first filter (W ₂ ) of the second signal (X _r (z)) whose output is to be removed from the first signal (X _p (z)), and

a second filter (JF) ₂ ) of the first signal (X _p (z)), whose output is intended to be removed from the second signal (X _r (z)),

characterized in that, upstream of the first and second filters in said main and reference channels are carried out:

a first filtering calculation (JFi), which is adaptive as a function of the difference between the first signal (X _p (z)) and the second signal (X _r (z)) to which the first calculated filtering (JF ₁ ) is applied, and

a second filtering calculation (Wi), which is adaptive as a function of the difference between the second signal (X _r (z)) and the first signal (X _p (z)) to which the second calculated filtering (JF ₂ ) is applied,

the first filtering calculation (JF)) being imposed on the first filter (JF ₂ )) and the second filtering calculation (JF ₂ ) being imposed on the second filter (Wn).

5. Method according to claim 1, characterized in that a step (μwi; μw ₂ ) of adaptation of the filtering calculation is chosen according to a signal-to-noise ratio estimated on at least one of the main channels. and reference.

6. Method according to claim 5, taken in combination with claim 4, characterized in that one chooses:

a step (μw ₂ ) of adaptation of the second filtering calculation varying like the signal-to-noise ratio estimated on the reference channel, and

a step (μwi) of adaptation of the first filtering computation varying as the inverse of the estimated signal-to-noise ratio on the main channel.

7. Method according to claim 6, characterized in that:

the adaptation step of the second filtering calculation (μw ₂ ) is constant for a signal-to-noise ratio lower than a first threshold value (RSB _W2 min), varies as the signal-to-noise ratio for a signal-to-noise ratio higher than the first threshold value, and is constant for a signal-to-noise ratio greater than a second threshold value (RSBw ₂ max), greater than said first threshold value, and

the adaptation step of the first filtering calculation (μwi) is constant (μwimax) for a signal-to-noise ratio lower than a first threshold value (RSBwimin), varies as the inverse of the signal-to-noise ratio for a signal-to-noise ratio; noise greater than the first threshold value, and is constant (μ _W imin) for a signal-to-noise ratio greater than a second threshold value (RSBw imax), greater than said first threshold value (RSBwimin).

8. Method according to claim 1, characterized in that the filtering calculation (W) and the updating of the first filter (W _{2 \} ) are implemented only if a signal to noise ratio, estimated on one of the main and reference channels, has a value in a chosen range.

9. The method of claim 8, taken in combination with claim 4, characterized in that the calculation of the first filtering (W) and the update of the first filter (W ₂ )) are implemented only in provided that the signal-to-noise ratio (RSB ₁ -), estimated on the reference channel, has a value lower than a threshold value (RSB _W2 min).

Device comprising at least a first (W ₂ ) and a second (W] ₂ ) filter, arranged between a main channel (V1) carrying a first signal (X _p (z)) and a reference channel (V2) conveying a second signal (X _r (z)), and in a backward configuration for reducing, in the first signal (X _p (z)), estimated noise in the second signal (X _r (z)),

characterized in that at least one of the first and second filters (W ₂ ) is updated by a first adaptive filtering computation (W] based on an estimated difference as a function of the first and second signals taken into consideration. upstream of the first and second filters in said main and reference channels.

1. A computer program comprising instructions for implementing the method according to claim 1, when this program is executed by a processor.