EP2309499A1 - Method for optimised filtering of non-stationary interference captured by a multi-microphone audio device, in particular a hands-free telephone device for an automobile. - Google Patents

Abstract

The method involves combining signals picked up by two microphones to make a noisy combined signal, and calculating a probability that speech is absent from the noisy combined signal on the basis of respective spectral energy levels of the noisy combined signal and of a referent noise signal. Noise is selectively reduced by applying variable gain that is specific to each frequency band and to each time frame on the basis of the probability that speech is absent and on the basis of the noisy combined signal.

Description

The invention relates to the treatment of speech in a noisy environment.

It concerns in particular, but without limitation, the processing of speech signals received by telephone devices for motor vehicles.

These devices include a microphone ("microphone") sensitive sensing not only the voice of the user, but also the surrounding noise, noise that is a disruptive element that can go, in some cases, to make incomprehensible the speaker's words. It is the same if one wants to implement speech recognition techniques, because it is very difficult to perform a form recognition on words embedded in a high noise level.

This difficulty related to surrounding noise is particularly restrictive in the case of devices "hands-free". In particular, the large distance between the microphone and the speaker causes a high relative level of noise that makes it difficult to extract the useful signal embedded in the noise. In addition, the highly noisy environment typical of the automotive environment has non-stationary spectral characteristics, that is to say that evolve unpredictably depending on the driving conditions: passage on deformed or paved roads, car radio operating etc.

Some of these devices provide for the use of several microphones, usually two microphones, and use the average of the signals picked up, or other more complex operations, to obtain a signal with a lower level of interference. In particular, a so-called beamforming technique makes it possible to create by software means a directivity which improves the signal / noise ratio, but the performances of this technique are very limited when only two microphones are used (concretely, it is considered that such a method provides good results only if you have a network of at least eight microphones).

Moreover, conventional techniques are especially adapted to the filtering of diffuse noise, stationary, coming from the surroundings of the device and found at comparable levels in the signals picked up by the two microphones.

On the other hand, a nonstationary noise or "transient", that is to say a noise evolving unpredictably as a function of time, will not be discriminated from the speech and will not be attenuated.

However, in an automotive environment these nonstationary and directive noises are very common: blow of horn, passage of a scooter, overtaking by a car, etc.

A difficulty in filtering these nonstationary noises lies in the fact that their temporal and spatial characteristics are very close to those of speech, hence the difficulty on the one hand, of estimating the presence of a speech (because the speaker do not speak all the time) and on the other hand to extract the useful speech signal in a very noisy environment such as a passenger compartment of a motor vehicle.

• de distinguer de façon efficace les bruits non stationnaires de la parole ; et
• d'adapter le débruitage à la présence et aux caractéristiques des bruits non stationnaires détectés, sans altérer la parole éventuellement présente, afin de traiter le signal bruité de la manière la plus efficace.

One of the aims of the present invention is to propose a multi-microphone hands-free device, in particular a system which implements only two microphones, allowing:

• to effectively distinguish nonstationary noises from speech; and
• to adapt the denoising to the presence and characteristics of the non-stationary noises detected, without altering the speech possibly present, in order to treat the noisy signal in the most effective manner.

• la parole présente généralement une cohérence spatiale supérieure au bruit ; et par ailleurs que
• la direction d'incidence de la parole est généralement bien définie, et peut être supposée connue (dans le cas d'un véhicule automobile, elle est définie par la position du conducteur, vers lequel est tourné le micro).

The starting point of the invention consists in associating (i) a spatial coherence analysis of the signal picked up by the two microphones, to (ii) an analysis of the direction of incidence of these signals. The invention is based on two observations, namely that:

• speech generally has a higher spatial coherence than noise; and besides that
• the direction of incidence of speech is generally well defined, and can be assumed to be known (in the case of a motor vehicle, it is defined by the position of the driver, to which the microphone is turned).

• une première référence de bruit calculée en fonction de la cohérence spatiale des signaux captés - une telle référence sera intéressante dans la mesure où elle intègre les bruits non stationnaires peu directifs (accrocs dans le ronronnement du moteur, etc.) ; et
• une seconde référence de bruit calculée en fonction de la direction principale d'incidence des signaux - cette caractéristique est en effet déterminable lorsque l'on utilise un réseau de plusieurs micros (au moins deux), conduisant à une référence de bruit intégrant surtout les bruits non stationnaires directifs (coups de klaxon, passage d'un scooter, dépassement par une voiture, etc.).

These two properties will be used to calculate two noise references according to different methods:

• a first noise reference calculated according to the spatial coherence of the signals captured - such a reference will be interesting insofar as it incorporates non-stationary non-directional noises (snags in the purring of the engine, etc.); and
• a second noise reference calculated as a function of the main direction of incidence of the signals - this characteristic is indeed determinable when using a network of several microphones (at least two), leading to a noise reference integrating especially the noises non-stationary directional (honking, passing a scooter, overtaking by a car, etc.).

• de manière générale, la première référence de bruit (celle calculée par cohérence spatiale) sera utilisée par défaut ;
• en revanche, lorsque la direction principale d'incidence du signal sera éloignée de celle du signal utile (la direction du locuteur, supposée connue a priori) - c'est-à-dire en présence d'un bruit directif assez puissant - la seconde référence de bruit sera utilisée de façon à introduire majoritairement dans cette dernière les bruits non stationnaires directifs et puissants.

These two noise references will be used alternately according to the nature of the noise present, as a function of the direction of incidence of the signals:

• in general, the first noise reference (that calculated by spatial coherence) will be used by default;
• on the other hand, when the main direction of incidence of the signal will be distant from that of the useful signal (the direction of the speaker, assumed to be known a priori ) - that is to say in the presence of a rather powerful directional noise - the second Noise reference will be used to introduce predominantly directional and powerful non-stationary noise into the latter.

Once the noise reference thus selected, this reference will be used to, firstly, calculate a probability of absence / presence of speech and, on the other hand, denoise the signal picked up by the microphones.

More specifically, the invention aims, in general, a method of denoising a noisy acoustic signal picked up by two microphones of a multi-microphone audio device operating in a noisy environment, including a hands-free telephone device for a vehicle automobile. The noisy acoustic signal includes a speech component derived from a directional speech source and a noise noise component, said noise component itself including a directional non-stationary side noise component.

Such a method is for example disclosed by I. Cohen and B. Berdugo, Two-Channel Signal Detection and Speech Enhancement Based on the Transient Beam-to-Reference Ratio, Proc. ICASSP 2003, Hong Kong, pp. 233-236, Apr. 2003 .

1. a) calcul d'une première référence de bruit par analyse de cohérence spatiale des signaux captés les deux microphones, ce calcul comprenant un filtrage linéaire prédictif appliqué aux signaux captés par les deux microphones et comprenant une soustraction avec compensation du déphasage entre le signal capté et le signal de sortie du filtre prédictif ;
2. b) calcul d'une seconde référence de bruit par analyse des directions d'incidence des signaux captés par les deux microphones, ce calcul comprenant le blocage spatial des composantes des signaux captés dont la direction d'incidence est située à l'intérieur d'un cône de référence défini de part et d'autre d'une direction prédéterminée d'incidence du signal utile ;
3. c) estimation d'une direction principale d'incidence des signaux captés par les deux microphones ;
4. d) sélection comme signal de bruit référent de l'une ou l'autre des références de bruit calculées aux étapes a) et b), en fonction de la direction principale estimée à l'étape c) ;
5. e) combinaison des signaux captés par les deux microphones en un signal combiné bruité ;
6. f) calcul d'une probabilité d'absence de parole dans le signal combiné bruité, à partir des niveaux respectifs d'énergie spectrale du signal combiné bruité et du signal de bruit référent ;
7. g) à partir de la probabilité d'absence de parole calculée à l'étape f) et du signal combiné bruité, réduction sélective du bruit par application d'un gain variable propre à chaque bande de fréquences et à chaque trame temporelle.

In a characteristic manner of the invention, this method comprises, in the frequency domain for a plurality of defined frequency bands. for successive time frames of signal, the following signal processing steps:

1. a) calculating a first noise reference by spatial coherence analysis of the signals picked up by the two microphones, this calculation comprising a predictive linear filtering applied to the signals picked up by the two microphones and comprising a subtraction with compensation of the phase shift between the signal picked up and the output signal of the predictive filter;
2. b) calculating a second noise reference by analyzing the directions of incidence of the signals picked up by the two microphones, this calculation comprising the spatial blocking of the components of the signals picked up whose direction of incidence is located inside a reference cone defined on either side of a predetermined direction of incidence of the useful signal;
3. c) estimating a principal direction of incidence of the signals picked up by the two microphones;
4. d) selecting as a reference noise signal from one or other of the noise references calculated in steps a) and b), depending on the principal direction estimated in step c);
5. e) combining the signals picked up by the two microphones into a noisy combined signal;
6. f) calculating a probability of absence of speech in the noisy combined signal from the respective spectral energy levels of the noisy combined signal and the reference noise signal;
7. g) from the probability of absence of speech calculated in step f) and the noisy combined signal, selective noise reduction by applying a variable gain specific to each frequency band and each time frame.

• le filtrage linéaire prédictif comprend l'application d'un algorithme de prédiction linéaire de type moindres carrés moyens LMS ;
• l'estimation de la direction principale d'incidence de l'étape c) comprend les sous-étapes successives suivantes : c1) partition de l'espace en une pluralité de secteurs angulaires ; c2) pour chaque secteur, évaluation d'un estimateur de direction d'incidence à partir des signaux captés par les deux microphones ; et c3) à partir des valeurs d'estimateurs calculées à l'étape c2), estimation de ladite direction principale d'incidence ;
• la sélection de l'étape d) est une sélection de la seconde référence de bruit comme signal de bruit référent si la direction principale estimée à l'étape c) est située hors d'un cône de référence défini de part et d'autre d'une direction prédéterminée d'incidence du signal utile ;
• la combinaison de l'étape e) comprend un préfiltrage de type fixed beamforming ;
• le calcul de probabilité d'absence de parole de l'étape f) comprend l'estimation de composantes de bruit pseudo-stationnaire respectives contenues dans le signal combiné bruité et dans le signal de bruit référent, la probabilité d'absence de parole étant calculée à partir également de ces composantes de bruit pseudo-stationnaire respectives ;
• la réduction sélective du bruit de l'étape g) est un traitement par application d'un gain à amplitude log-spectrale modifié optimisé OM-LSA.

According to various advantageous subsidiary features:

• predictive linear filtering comprises the application of a LMS mean least squares linear prediction algorithm;
• the estimate of the main direction of incidence of step c) comprises the following successive sub-steps: c1) partitioning the space into a plurality of angular sectors; c2) for each sector, evaluation of an incidence direction estimator from the signals captured by the two microphones; and c3) from the estimator values calculated in step c2), estimating said main direction of incidence;
• the selection of step d) is a selection of the second noise reference as a reference noise signal if the main direction estimated in step c) is located outside a reference cone defined on either side of a predetermined direction of incidence of the useful signal;
• the combination of step e) comprises prefiltering of fixed beamforming type ;
• the speech absence probability calculation of step f) comprises the estimation of respective pseudo-stationary noise components contained in the noisy combined signal and in the reference noise signal, the probability of absence of speech being calculated from also these respective pseudo-stationary noise components;
• the selective noise reduction of step g) is a processing by applying an OM-LSA optimized modified log-spectral amplitude gain.

An example embodiment of the method of the invention will now be described with reference to the appended figure.

The Figure 1 is a block diagram showing the different modules and functions implemented by the method of the invention as well as their interactions.

The method of the invention is implemented by software means, which can be broken down and schematized by a number of illustrated blocks 10 to 36 Figure 1 .

These processes are implemented in the form of appropriate algorithms executed by a microcontroller or a digital signal processor. Although, for the sake of clarity, these various treatments are presented as separate modules, they implement common elements and correspond in practice to a plurality of functions globally executed by the same software.

The signal that one wishes to denoise comes from a plurality of signals picked up by a network of microphones (which, in the minimum configuration, can be simply a network of two microphones, as in the example shown) arranged in a configuration predetermined. In practice, these two microphones can for example be installed on the ceiling of a car interior, about 5 cm from each other; and have the main lobe of their driver oriented directivity diagram. This direction, considered a priori known, will be designated the direction of incidence of the useful signal.

"Lateral noise" shall be termed a directional non-stationary noise whose direction of incidence is remote from that of the wanted signal, and the "preferred cone" shall be called the direction or angular sector of the space where the useful signal source is located ( the speech of the speaker) in relation to the network of microphones. When a sound source will manifest outside the preferred cone, it will be a lateral noise, which we seek to mitigate. As illustrated on the Figure 1 , the noisy signals picked up by the two microphones x ₁ ( n ) and x ₂ ( n ) are transposed in the frequency domain (blocks 10) by a short-term Fourier Transform (FFT) calculation of which the result is denoted respectively X ₁ ( k, l ) and X ₂ ( k, l ), where k is the index of the frequency band and l is the index of the time frame. The signals from the two microphones are also applied to a module 12 implementing a predictive LMS algorithm schematized by the block 14 and giving, after calculation of a short-term Fourrier transform (block 16) a signal Y ( k, l ) which will be used to calculate a first noise reference Ref ₁ ( k, l ) executed by a block 18, essentially on a spatial coherence criterion.

Another noise reference Ref ₂ ( k, l ) is calculated by a block 20, essentially on an angular blocking criterion), from the signals X ₁ ( k, 1 ) and X ₂ ( k, 1 ) obtained directly, in the frequency domain, from the signals x ₁ ( n ) and x ₂ ( n ) .

A block 22 effects the selection of one or the other of the noise references Ref ₁ ( k, l ) or Ref ₂ ( k, l ) as a function of the result of a calculation of the angle of incidence of the signals operated by the block 24 from the signals X ₁ ( k, l ) and X ₂ ( k, l ). The chosen noise reference, Ref ( k, l ), is used as referent noise channel of a block 26 for calculating a probability of speech absence. operated on a noisy signal X ( k, l ) resulting from a combination, performed by the block 28, of the two signals X ₁ ( k, 1 ) and X ₂ ( k, 1 ). Block 26 also takes into account the respective pseudo-stationary noise components of the reference noise channel and the noisy signal, components estimated by blocks 30 and 32.

The result q ( k, l ) of the speech absence probability calculation and the noisy signal X ( k, l ) are inputted to an OM-LSA gain control algorithm (block 34) whose result I Ŝ ( k, l ) is subjected (block 36) to an inverse Fourier transformation (iFFT) to obtain in the time domain an estimate ŝ ( t ) of the denoised speech signal.

We will now describe in detail each step of the treatment.

Fourier transform of the signals picked up by the microphones (blocks 10)

avec : $$d_n\left(k\right)=e^{-\mathit{i}\mathrm{2}{\mathit{\pi f}}_k{\mathit{\tau }}_n}$$

l

étant l'indice de la trame temporelle,

k

étant l'indice de la bande de fréquences, et

f_k

étant la fréquence centrale de la bande de fréquences indicée par k,

S(k,l)

désignant la source de signal utile,

α _n et τ _n

désignant l'atténuation et le délai subis par le signal utile capté au niveau du micro n, et

V_n (k,l)

désignant le bruit capté par le micro n.

The signal in the time domain x _n ( t ) from each of the N micros ( N = 1.2 in the example shown) is digitized, divided into frames of T time points, window temporally by a Hanning type window, then the Fast Fourier Transform FFT (short-term transform) X _n ( k, l ) is calculated for each of these signals: $$X_{\mathrm{not}}\left(kl\right)={\mathrm{at}}_{\mathrm{not}}\mathrm{.}{d}_{\mathrm{not}}\left(k\right)\times S\left(kl\right)+V_{\mathrm{not}}\left(kl\right)$$

with: $$d_{\mathrm{not}}\left(k\right)=e^{-\mathit{i}\mathrm{2}{\mathit{\pi f}}_k{\mathit{\tau }}_{\mathrm{not}}}$$

l

being the index of the time frame,

k

being the index of the frequency band, and

f _k

being the center frequency of the frequency band indexed by k ,

S ( k, l )

designating the useful signal source,

α _n and τ _n

designating the attenuation and the delay experienced by the useful signal picked up at the level of the microphone n , and

V _n ( k, l )

designating the noise picked up by the microphone n .

Calculation of a first noise reference by spatial coherence (block 12)

The fundamental idea on which the invention is based is that, in a telecommunications environment, speech is a signal emitted by a well localized source, relatively close to microphones and almost entirely captured in direct path. Conversely, stationary and nonstationary noises, which come mainly from the user's surroundings, can be associated with distant sources, in large numbers and having a statistical correlation lower than the speech between the two microphones.

In a telecommunications environment, speech is therefore more spatially coherent than noise.

From this principle, it is possible to exploit the spatial coherence property to build a richer and more suitable reference noise channel than with a beamformer. The system provides for this purpose to use a predictive filter 14 type LMS ( Least Mean Squares, Least Mean Squares ) having for inputs the signals x ₁ ( n ) and x ₂ ( n ) captured by the pair of microphones. Note y ( n ) the output of the LMS and e (n) the prediction error.

This predictive filter is used to predict from x ₂ ( n ) the speech component in x ₁ ( n ). Indeed, being more coherent spatially, speech will be better predicted by the adaptive filter than noise.

E(k,l), X ₁(k,l) et Y(k,l) étant les transformées de Fourier à court terme (TFCT) respectives de e(k,l), x₁ (k,l) et y(k,l).A first possibility consists in taking the Fourier transform of the prediction error for the reference noise channel: $$E\left(kl\right)=X_1\left(kl\right)-Y\left(kl\right)$$

E ( k, l ), X ₁ ( k, l ) and Y ( k, l ) being the short-term Fourier transforms (TFCTs) of e ( k, l ), x ₁ ( k, l ) and y ( k, l ).

However, in practice there is a certain phase shift between X ₁ ( k, l ) and Y ( k, l ) due to an imperfect convergence of the LMS algorithm, which prevents good discrimination between speech and noise.

To overcome this defect, it is possible to define the first referent noise signal Ref ₁ ( k, l ) by: $${\mathit{Ref}}_1\left(kl\right)=X_1\left(kl\right)-X_1\left(kl\right)\frac{\left|Y\left(kl\right)\right|}{\left|X_1\left(kl\right)\right|}$$

Unlike many conventional noise estimation methods, no stationarity hypothesis is used on noise to compute this first reference noise channel Ref ₁ (k, l ). One of the advantages is therefore that this noise channel integrates a part of nonstationary noises, in particular those which have a low statistical correlation and which are not predictable between the two microphones.

Calculation of a second noise reference by spatial blocking (block 20)

In a telecommunications environment, it is possible to encounter noises whose source is well localized and relatively close to the microphones. In general, it is about occasional noise rather powerful (passage of a scooter, overtaking by a car, etc.), and which can be troublesome.

The assumptions used for calculating the first reference noise channel are not checked on this type of noise; on the other hand, these noises have the peculiarity of having a direction of incidence well defined and distinct from the direction of incidence of the speech.

To exploit this property, it will be assumed that the angle of incidence θ _S of the speech is known, for example defined as being the angle between the mediator of the pair of microphones and the reference direction corresponding to the source of useful speech.

More precisely, we perform a partition of the space into angular sectors that describe space, and each of which corresponds to a direction defined by an angle θ _j , j ∈ [1, M ], with for example M = 19, giving the collection of angles {-90 °, -80 ° ..., 0 °, ... + 80 °, + 90 °}. Note that there is no link between the number of microphones N and the number M angles tested: for example, it is quite possible to test M = 19 angles with a single pair of microphones ( N = 2). ).

We get the partition { A , I } of angles θ _j which are respectively "allowed" and "forbidden", the angles θ _a ∈ A being "authorized" in that they correspond to signals coming from a cone privileged centered on θ _S , while the angles θ _i ∈ I are "prohibited" in that they correspond to unwanted side noises.

X ₁(k,l)

étant la TFCT du signal enregistré par le micro d'indice 1,

X ₂(k,l)

étant la TFCT du signal enregistré par le micro d'indice 2,

f_k

étant la fréquence centrale de la bande de fréquences k,

l

étant la trame,

d

étant la distance entre les deux micros,

c

étant la célérité du son, et

|A|

étant le nombre d'angles "autorisés" du cône privilégié.

The second reference noise channel Ref ₂ ( k, l ) is defined as follows: $${\mathit{Ref}}_2\left(kl\right)=\frac{1}{\left|\mathrm{AT}\right|}{\displaystyle \underset{{\theta }_{\mathrm{at}}\in \mathrm{AT}}{\overset{}{\Sigma }}}\left(X_1\left(kl\right)-X_2\left(kl\right)\times e^{\frac{i2\pi \mathrm{.}{f}_k\mathrm{.}d\mathrm{.}\sin {\mathit{\theta }}_{\mathrm{at}}}{\mathrm{vs}}}\right)$$

X ₁ ( k, l )

being the TFCT of the signal recorded by the micro of index 1,

X ₂ ( k, l )

being the TFCT of the signal recorded by the micro of index 2,

f _k

being the center frequency of the frequency band k ,

l

being the frame,

d

being the distance between the two pickups,

vs

being the speed of sound, and

| A |

being the number of "allowed" angles of the privileged cone.

In each term of this sum, we subtract from the signal of the micro of index 1 the signal of the micro of index 2 out of phase by an angle θ _α which belongs to A (sub-collection of "authorized" angles). Thus, in each term, signals having a propagation direction θ _α "allowed" are spatially blocked. This spatial blocking is performed for all authorized angles.

In this second referent noise channel Ref ₂ ( k, l ), any lateral noises (non-stationary directional noises) are thus allowed to pass, by spatially blocking the speech signal.

Choice of the noise reference according to the direction of incidence of the signals (blocks 22 and 24)

This selection involves an estimation of the angle of incidence θ ( k, l ) of the signals.

avec : $$P_{1,2}\left({\theta }_{j}kl\right)=E\left(X_1\left(kl\right)\mathrm{.}{\bar{X}}_2\left(kl\right)\mathrm{.}{e}^{-i2\pi f_k{\tau }_j}\right)$$

et $${\tau }_j=\frac{d}{c}\sin {\theta }_j$$

This estimator (block 24) can for example be based on an intercorrelation calculation, taking as the direction of incidence the angle that maximizes the modulus of the estimator, that is: $$\hat{\theta }\left(kl\right)=\underset{{\theta }_j,j\in \left[1M\right]}{\mathrm{arg\; max}}\Vert {\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="imgf0001.png"\; state="\{\{state\}\}"><figure-callout\; id="2"\; label="P"\; filenames="imgf0001.png"\; state="\{\{state\}\}">P</figure-callout></figure-callout>}}_{1,2}\left({\theta }_{j}kl\right)\Vert$$

with: $${\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="imgf0001.png"\; state="\{\{state\}\}"><figure-callout\; id="2"\; label="P"\; filenames="imgf0001.png"\; state="\{\{state\}\}">P</figure-callout></figure-callout>}}_{1,2}\left({\theta }_{j}kl\right)=E\left(X_1\left(kl\right)\mathrm{.}{\tilde{X}}_2\left(kl\right)\mathrm{.}{e}^{-i2\pi f_k{\tau }_j}\right)$$

and $${\tau }_j=\frac{d}{\mathrm{vs}}\sin {\theta }_j$$

• si θ̂(k,l) est "autorisé" (θ̂(k,l) ∈ A), alors Ref(k,l) = Ref ₁(k,l)
• si θ̂(k,l) est "interdit" (θ̂(k,l) ∈ I), alors Ref(k,l) = Ref ₂(k,l)
• si θ̂(k,l) n'est pas défini, alors Ref(k,l) = Ref ₁(k,l)

The selected reference noise channel Ref ( k, l ) will depend on the detection of an "allowed" or "forbidden" angle for the frame l and the frequency band k :

• if θ ( k, l ) is "allowed" (θ ( k, l ) ∈ A ), then Ref ( k, l ) = Ref ₁ ( k, l )
• if θ ( k, l ) is "forbidden" ( θ ( k, l ) ∈ I ), then Ref ( k, l ) = Ref ₂ ( k, l )
• if θ ( k, l ) is not defined, then Ref ( k, l ) = Ref ₁ ( k, l )

Thus, in the case of a detected "allowed" angle, or in the absence of directional signals at the input of the microphones, the reference noise channel Ref ( k, l ) is calculated by spatial coherence, which allows Integrate non-stationary non-directional noise.

On the other hand if a "forbidden" angle is detected, it means that a directional noise and rather powerful is present. In this case, the reference noise channel Ref ( k, l ) is calculated according to a different method, by spatial blocking, so as to efficiently introduce in this channel non-stationary directional and powerful noises.

Constitution of a combined signal partially de-denuded (block 28)

avec : $$d_2\left(k\right)=e^{i2\pi f_k{\tau }_s}\mathrm{avec}{\tau }_s=\frac{d}{c}\sin {\theta }_s$$

The signals X _n ( k, l ) (the TFCTs of the signals picked up by the microphones) can be combined with one another by a simple pre-filtering technique. beamforming of the Delay and Sum type , which is applied to obtain a partially denoised combined signal X ( k, l ): $$X\left(kl\right)=\frac{1}{2}\left[X_1\left(kl\right)+\widetilde{d_2\left(k\right)}\mathrm{.}{X}_2\left(kl\right)\right]$$

with: $$d_2\left(k\right)=e^{i2\pi f_k{\tau }_s}\mathrm{with}{\tau }_s=\frac{d}{\mathrm{vs}}\sin {\theta }_s$$

When the system considered comprises, as in the present example, two microphones whose mediator cuts the source, the angle θ _S is zero and it is a simple average that is made on the two microphones. It should also be noted that, specifically, the number of microphones being limited, this treatment provides only a slight improvement in the signal / noise ratio, of the order of 1 dB only.

Estimation of pseudo-stationary noise (blocks 30 and 32)

The purpose of this step is to calculate an estimate of the pseudo-stationary noise component present in the noise reference Ref ( k, l ) (block 30) and, in the same way, the pseudo-stationary noise component present in the signal to denoise X ( k, l ) (block 32).

There are many publications on this subject, the estimate of the pseudo-stationary noise component being indeed a fairly well resolved conventional problem. Different methods are effective and usable for this purpose, in particular an algorithm for estimating the energy of the pseudo-stationary minimum-recursive averaging noise component (MCRA) as described by I. Cohen and B. Berdugo, Noise Estimation by Minima Controlled Recursive Averaging for Robust Speech Enhancement, IEEE Signal Processing Letters, Vol. 9, No. 1, pp. 12-15, Jan. 2002 .

Calculation of the probability of absence of speech (block 26)

An effective and recognized method for estimating the probability of no speech in a noisy environment is that of the ratio of transients, described by I. Cohen and B. Berdugo, Two-Channel Signal Detection and Speech Enhancement Based on the Transient Beam-to-Reference Ratio, Proc. ICASSP 2003, Hong Kong, pp. 233-236, Apr. 2003 .

X(k,l)

étant le signal combiné partiellement débruité,

Ref(k,l)

étant le canal de bruit référent calculé dans la partie précédente,

k

étant la bande de fréquences, et

l

étant la trame

The ratio of transients is defined as follows: $$\mathrm{\Omega }\left(kl\right)=\frac{S\left[X\left(kl\right)\right]-M\left[X\left(kl\right)\right]}{S\left[\mathit{Ref}\left(kl\right)\right]-M\left[\mathit{Ref}\left(kl\right)\right]}$$

X ( k, l )

being the combined signal partially denoised,

Ref ( k, l )

being the reference noise channel calculated in the previous part,

k

being the frequency band, and

l

being the weft

The operator S is an estimate of the instantaneous energy, and the operator M is an estimate of the pseudo-stationary energy (estimate made by the blocks 30 and 32). S - M provides an estimate of the transient parts of the analyzed signal, also called transients.

The two signals analyzed here are the combined noisy signal X ( k, l ) and the signal of the reference noise channel Ref ( k, l ). The numerator will therefore highlight the speech and noise transients, while the denominator will extract only the transients of noise in the reference noise channel.

Thus, in the presence of speech but in the absence of non-stationary noise, the ratio Ω ( k, l ) will tend towards a high limit Q _max ( k ), whereas conversely, in the absence of speech but in the presence of nonstationary noise, this ratio will approach the low limit Ω _min ( k ), where k is the frequency band. This will make it possible to differentiate between speech and nonstationary noises.

In the general case, we have: $${\mathrm{\Omega }}_{\min }\left(k\right)\le \mathrm{\Omega }\left(kl\right)\le {\mathrm{\Omega }}_{\max }\left(k\right)$$

The probability of no speech, noted here q ( k, l ), will be calculated as follows.

For each frame I and each frequency band k:

1. i) Calculation of S [ X ( k, l )], S [ Ref ( k, l )], M [ X ( k, l )] and M [ Ref ( k, l )];
2. ii) If S [ X ( k, l )] ≥ α _x M [ X ( k, l )], the speech is likely to be present, the analysis is continued in step (iii),
   in the opposite case, speech is absent: then q (k, l ) = 1;
3. (iii) If S [ Ref ( k, l )] ≥ α _Ref M [ Ref ( k, l )], transient noise is likely to be present, the analysis is continued in step (iv),
   in the opposite case, this means the transients found in X ( k, l ) are all speech transients: then q ( k, l ) = 0 ;
4. iv) Calculation of the ratio
5. v) Determination of the probability of speech absence: $$q\left(kl\right)=\max \left(\min \left(\frac{{\mathrm{\Omega }}_{\max }\left(kl\right)-\mathrm{\Omega }\left(kl\right)}{{\mathrm{\Omega }}_{\max }\left(kl\right)-{\mathrm{\Omega }}_{\min }\left(kl\right)}1\right),0\right)\mathrm{.}$$
   

The constants α _X and α _Ref used in this algorithm are in fact transient detection thresholds. The parameters α _X , α _Ref and Ω _min ( k ) and Ω _max ( k ) are all chosen to correspond to typical situations, close to reality.

Noise reduction by applying an OM-LSA gain (block 34)

The probability q ( k, l ) of absence of speech calculated at block 26 will be used as an input parameter in a technique (in itself known) of denoising. It has the advantage of making it possible to identify the periods of absence of speech even in the presence of a nonstationary, non-directive or directive noise. The probability of speechlessness is a crucial estimator for the proper functioning of a denoising structure such as we will use, because it underlies the good estimate of noise and the calculation of an efficient denoising gain.

It is advantageous to use an OM-LSA ( Optimally Modified Log Spectral Amplitude ) denoising method such as that described by: I. Cohen, Optimal Speech Enhancement Under Signal Presence Uncertainty Using Log-Spectral Amplitude Estimator, IEEE Signal Processing Letters, Vol. 9, No 4, April 2002 .

Essentially, the application of a gain called Log-Spectral Amplitude (LSA) is used to minimize the mean squared distance between the logarithm of the amplitude of the estimated signal and the logarithm of the amplitude of the original speech signal. This second criterion is superior to the first because the distance chosen is in better adequacy with the behavior of the human ear and thus gives qualitatively better results. In all cases, the essential idea is to reduce the energy of the highly parasitized frequency components by applying a low gain, while leaving intact (by the application of a gain equal to 1) those that are little or not at all.

The "OM-LSA" ( Optimally-Modified Log-Spectral Amplitude ) algorithm improves the calculation of the LSA gain to be applied by weighting it by the conditional probability of presence of speech.

avec : $${\alpha }_{\mathit{Bruit}}\left(kl\right)={\alpha }_{\mathit{B}}+\left(1-{\alpha }_{\mathit{B}}\right)\mathrm{.}{p}_{\mathit{spa}}\left(kl\right)$$

In this method, the probability of absence of speech occurs at two important moments, for the estimation of the noise energy and for the calculation of the final gain, and the probability q ( k, l ) will be used at these two levels. . If λ _Noise ( k, l ) is the estimate of the spectral power density of the noise, this estimate is given by: $${\hat{\mathit{\lambda }}}_{\mathit{Noise}}\left(kl\right)={\alpha }_{\mathit{Noise}}\left(kl\right)\mathrm{.}{\hat{\mathit{\lambda }}}_{\mathit{Noise}}\left(k,l-1\right)+\left[1-{\alpha }_{\mathit{Noise}}\left(kl\right)\right]\mathrm{.}{\left|X\left(k\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png"\; state="\{\{state\}\}">l</figure-callout>}\right)\right|}^2$$

with: $${\alpha }_{\mathit{Noise}}\left(kl\right)={\alpha }_{\mathit{B}}+\left(1-{\alpha }_{\mathit{B}}\right)\mathrm{.}{p}_{\mathit{spa}}\left(kl\right)$$

It can be noted here that the probability q ( k, l ) modulates the forgetfulness factor in the estimation of the noise, which is updated more rapidly on the signal noisy X ( k, l ) when the probability of absence of speech is high, this mechanism completely conditioning the quality of λ _Noise ( k, l ).

The denoising gain G _OM-LSA ( k, l ) is given by: $${\mathrm{BOY\; WUT}}_{\mathit{OM}-\mathit{LSA}}\left(kl\right)={\left\{{\mathrm{BOY\; WUT}}_{H1}\left(kl\right)\right\}}^{1-q\left(kl\right)}\mathrm{.}{\mathrm{BOY\; WUT}}_{\min }^{q\left(kl\right)}$$

G _{H 1} ( k, l ) being a denoising gain (whose calculation depends on the noise estimate λ _Noise ) described in the aforementioned article by Cohen, and G _min being a constant corresponding to the denoising applied when speech is considered absent.

It should be noted that the probability q ( k, l ) plays a large role here in determining the gain G _OM-LSA ( k, l ). In particular, when this probability is zero, the gain is equal to G _min and a maximum noise reduction is applied: if, for example, a value of 20 dB is chosen for G _min , the non-stationary noises previously detected are attenuated by 20 dB.

The denoised signal Ŝ ( k, l ) at the output of block 34 is given by: $$\hat{S}\left(kl\right)={\mathrm{BOY\; WUT}}_{\mathit{OM}-\mathit{LSA}}\left(kl\right)\mathrm{.}X\left(kl\right)$$

It will be noted that usually such a denoising structure produces an unnatural and aggressive result on nonstationary noises, which are confused with useful speech. One of the major advantages of the invention is, on the contrary, to effectively eliminate these nonstationary noises.

On the other hand, in an advantageous variant, it is possible to use in the expressions given above a probability of absence of hybrid q _hybrid speech ( k, l ), which will be calculated using q ( k, l ) and another probability of absence of speech q _std ( k, l ), for example evaluated according to the method described in the WO 2007/099222 A1 (Parrot SA). We then have: $$q_{\mathit{hybrid}}\left(kl\right)=\max \left(q\left(kl\right),q_{\mathit{std}}\left(kl\right)\right)$$

Time reconstitution of the signal (block 36)

The last step consists in applying to the signal Ŝ ( k, l ) a fast inverse Fourier transform iFFT to obtain in the time domain the denoised speech signal ŝ ( t ) sought.

Claims (7)

A method of denoising a noisy acoustic signal picked up by two microphones of a multi-microphone audio device operating in a noisy medium, in particular a "hands-free" telephone device for a motor vehicle,
the noisy acoustic signal comprising a useful speech component derived from a directional speech source and a noise noise component, this noise component itself including a directional non-stationary lateral noise component,
characterized in that it comprises, in the frequency domain for a plurality of defined frequency bands for successive time frames of signal, the following signal processing steps: a) calculation (18) of a first noise reference by spatial coherence analysis of the signals picked up by the two microphones, this calculation comprising a predictive linear filtering applied to the signals picked up by the two microphones and comprising a subtraction with compensation of the phase shift between the captured signal and the output signal of the predictive filter; b) calculating (20) a second noise reference by analyzing the directions of incidence of the signals picked up by the two microphones, this calculation comprising the spatial blocking of the components of the signals picked up whose incidence direction is located at the inside a reference cone defined on either side of a predetermined direction of incidence of the useful signal; c) estimating (24) a principal direction of incidence (θ ( k, l )) of the signals picked up by the two microphones; d) selecting (22) as a reference noise signal ( Ref ( k, l )) from one or other of the noise references calculated in steps a) and b), according to the main direction estimated at step c); e) combining (28) the signals picked up by the two microphones into a noisy combined signal ( X ( k, 1 )); f) calculating (26) an absence of speech probability ( q ( k, l )) in the noisy combined signal from the respective spectral energy levels of the noisy combined signal ( X ( k, l )) and the reference noise signal ( Ref ( k, l )); g) from the probability of no speech ( q ( k, l )) calculated in step f) and the noisy combined signal ( X ( k, l )), selective noise reduction (34) per application a variable gain specific to each frequency band and each time frame. The method of claim 1, wherein the predictive filtering comprises applying a LMS mean least squares linear prediction algorithm. The method of claim 1, wherein estimating (24) the main direction of incidence of step c) comprises the following successive substeps: c1) partitioning the space into a plurality of angular sectors; c2) for each sector, evaluating an incidence direction estimator from the two signals picked up by the two corresponding microphones; and c3) from the estimator values calculated in step c2), estimating said main direction of incidence. The method of claim 1, wherein the selection (22) of step d) is a selection of the second noise reference as a reference noise signal if the principal direction estimated in step c) is located out of a reference cone defined on either side of a predetermined direction of incidence of the useful signal. The method of claim 1, wherein the combination (28) of step e) comprises prefixing of fixed beamforming type . The method of claim 1, wherein the speech absence probability calculation (26) of step f) comprises estimating (30, 32) respective pseudo-stationary noise components contained in the noisy combined signal. and in the reference noise signal, the probability of no speech ( q ( k, l )) being calculated from these respective pseudo-stationary noise components. The method of claim 1, wherein the selective noise reduction (34) of step g) is a processing by applying an OM-LSA optimized modified log-spectral amplitude gain.

Images