EP0918317A1 - Frequency filtering method using a Wiener filter applied to noise reduction of audio signals - Google Patents

Abstract

The noisy signal is converted into the frequency domain by Fourier analysis (1). A model is created of the noise (2) and estimates made of the spectral density, energy level and the conversion coefficient for statistical dispersion. In parallel to this similar estimates are made of the noisy signal (3). The two results are then recombined in a Wiener filter (4) which compensates for energy levels and applies noise overestimation. The signal is then reconstructed (5).

Description

The present invention relates to a method of frequency filtering using a Wiener filter.

It applies in particular, although not exclusively for the denoising of sound signals containing speech captured in noisy environments and more general to denoising of all sound signals.

The main areas concern telephone or radiotelephone communications, the voice recognition, sound recording on board aircraft civil or military, and more generally of all noisy vehicles, on-board intercoms, etc.

By way of nonlimiting example, in the case of a aircraft, noises result from engines, air conditioning, ventilation of on-board equipment or aerodynamic noises. All these noises are picked up, at less partially, by the microphone in which the pilot or another member of the crew. In addition, for this type of application in particular, one of the characteristics noises is to be very variable over time. In indeed, they are very dependent on the operating regime engines (take-off phase, stabilized speed, etc.). Useful signals, i.e. signals representing the conversations, also present particularities: they are most often short-lived.

Finally, whatever the application envisaged, if we are interested in "voicing", we can highlight certain peculiarities. As is known, the voicing relates to elementary characteristics of pieces of speech, and more specifically concerns vowels, as well only part of the consonants: "b", "d", "g", "j", etc. These letters are characterized by an audio signal from pseudo-periodic structure.

In speech processing, it is common to consider that stationary regimes, in particular the mentioned above, are established on durations included between 10 and 20 ms. This time interval is characteristic of the elementary phenomena of production of the floor and will be referred to as the frame below.

Also, it is usual for denoising processes take into account this important characteristic of audio signals including speech.

These processes generally include the steps following main ones: a division into frames of the signal to denoise, the processing of these frames by a Fourier transform operation (or a similar transform) to move into the field frequency, the proper denoising treatment by digital filtering, and dual processing of the first, by an inverse Fourier transform, to return to the time domain. The last step is a signal reconstruction. This reconstruction can be obtained by multiplying each of the frames by a window of weighting.

One of the most used digital filters for this type of application is the Wiener filter, in particular a so-called optimal Wiener filter. This one presents the advantage of differentiating the frames successive.

In other words, and more generally, The Wiener's optimal filtering is at the center of the methods signal processing, based on second order statistical characteristics and therefore of the notion of correlation.

Wiener filtering allows the separation of decorrelation signals. Its importance is linked to the simplicity of theoretical calculations. In addition, it can apply to a multitude of specific processes, and in particular, with regard to the preferred application targeted by the invention, the extraction of noise polluting a signal of speech.

However, in the known art, a classic problem encountered during denoising by Wiener filtering is the presence of a noise, called musical noise, which degrades the perception of the denoised signal. This musical noise is due to fluctuations in the spectral densities of the noise present in the input signal. For some frames, in fact, the noise spectral density is higher, at least on a frequency channel, to that of the noise model that we used in these techniques. In this case, the mechanisms specific to Wiener filtering cause the appearance of a residual noise on the denoised signal. It is particularly unpleasant from a perceptual point of view of share its instability. Indeed, when listening to a signal speech, there are residual noises in the form of "gurgling", which resembles distortions that one can attribute to a large variability of the polluting noise the denoised speech signal or "useful" signal.

The invention therefore sets itself the aim of overcoming the disadvantages of filtering methods of the known art, especially the main drawback which has just been recalled: the presence of a parasitic residual noise in the denoised signal, called "musical noise". The invention aims to more generally, to increase the intelligibility of the speech, in its main application.

• la probabilité de bruit musical est d'autant plus forte que l'estimée des densités spectrales du bruit est instable d'une trame à l'autre ;
• la probabilité de présence de bruit musical est d'autant plus forte que l'estimée de la densité spectrale du bruit est faible par rapport à sa densité spectrale réelle.

In order to greatly reduce the effects of musical noise, the invention takes advantage of the following two experimental observations:

• the probability of musical noise is all the greater when the estimate of the spectral densities of the noise is unstable from one frame to another;
• the probability of the presence of musical noise is higher when the estimate of the spectral density of the noise is low compared to its real spectral density.

According to a main characteristic of the invention, the Wiener filter used for digital filtering is modified in an optimized way by introducing a term of energy compensation aimed at overestimating the level of noise. In addition, this compensation term is adaptive.

• élaboration à partir desdits signaux bruités d'un modèle de bruit sur un nombre N déterminé desdites trames, N étant compris entre des bornes minimale et maximale prédéterminées ;
• application d'une transformée de Fourier auxdites N trames ;
• estimation, pour chaque trame dudit modèle, de la densité spectrale de cette trame ;
• estimation de la densité spectrale moyenne dudit modèle de bruit ;
• calcul, à partir de ces deux estimations, d'un coefficient de surestimation statistique, ledit coefficient statistique étant égal au rapport maximal, pour lesdites N trames du modèle de bruit, entre le maximum de la densité spectrale d'une trame considérée dudit modèle de bruit, et le maximum de la densité spectrale estimée du modèle de bruit ;
• estimation, pour chaque trame desdits signaux à débruiter, de sa densité spectrale ; et
• modification, pour chaque trame desdits signaux à débruiter, des coefficients dudit filtre de Wiener pour que la relation suivante soit vérifiée : W(ν) = β1 - α · maxi · γ x ν γ u ν ,    relation dans laquelle α et β sont des coefficients fixes prédéterminés, dits coefficient statique de compensation énergétique et coefficient d'atténuation exponentielle, respectivement, ν décrit l'ensemble des canaux fréquentiels de ladite transformée de Fourier, γ _u (ν) étant l'estimée de la densité spectrale de la trame à débruiter, γ _x (ν) est ladite densité spectrale du modèle de bruit, et maxi ledit coefficient de surestimation statistique, modifiant le coefficient statique de compensation énergétique α.

The subject of the invention is therefore a method of frequency filtering for the denoising of noisy sound signals consisting of so-called useful sound signals mixed with noise signals, the method comprising at least one step of splitting said sound signals into a series of identical frames. of a determined length and a frequency filtering step using a Wiener filter, characterized in that it further comprises the following steps:

• elaboration from said noisy signals of a noise model on a determined number N of said frames, N being between predetermined minimum and maximum limits;
• applying a Fourier transform to said N frames;
• estimation, for each frame of said model, of the spectral density of this frame;
• estimation of the average spectral density of said noise model;
• calculation, from these two estimates, of a statistical overestimation coefficient, said statistical coefficient being equal to the maximum ratio, for said N frames of the noise model, between the maximum of the spectral density of a considered frame of said model noise, and the maximum of the estimated spectral density of the noise model;
• estimation, for each frame of said signals to be denoised, of its spectral density; and
• modification, for each frame of said signals to be denoised, of the coefficients of said Wiener filter so that the following relation is verified: W (ν) = β 1 - α · max · γ x ν γ u ν , relation in which α and β are predetermined fixed coefficients, called static energy compensation coefficient and exponential attenuation coefficient, respectively, ν describes the set of frequency channels of said Fourier transform, γ _u (ν) being the estimated of the spectral density of the frame to be denoised, γ _x (ν) is said spectral density of the noise model, and maxi said statistical overestimation coefficient, modifying the static energy compensation coefficient α.

• la figure 1 illustre, sous forme de bloc diagramme, les principales étapes du procédé selon l'invention ;
• la figure 2 illustre schématiquement un filtre de Wiener de l'art connu ;
• la figure 3 est un diagramme illustrant la densité spectrale d'un modèle de bruit et les densités spectrales γ _u de chaque trame de ce modèle de bruit ;
• les figures 4a et 4b sont des diagrammes comparatifs illustrant ces mêmes paramètres avec surestimation de la densité spectrale du modèle de bruit ;
• la figure 5 est un diagramme illustrant ces mêmes paramètres avec surestimation adaptative de la densité spectrale du modèle de bruit ;
• la figure 6 représente un exemple typique de signal issu d'une prise de son bruitée ;
• la figure 7 est un organigramme représentant les étapes d'un procédé particulier de recherche d'un modèle de bruit ;
• et la figure 8 est un organigramme détaillé représentant les étapes du procédé de filtrage numérique selon un mode de réalisation préféré de l'invention.

The invention will be better understood and other characteristics and advantages will appear on reading the description which follows with reference to the appended figures, among which:

• FIG. 1 illustrates, in the form of a block diagram, the main steps of the method according to the invention;
• FIG. 2 schematically illustrates a Wiener filter of the known art;
• FIG. 3 is a diagram illustrating the spectral density of a noise model and the spectral densities γ _u of each frame of this noise model;
• FIGS. 4a and 4b are comparative diagrams illustrating these same parameters with overestimation of the spectral density of the noise model;
• FIG. 5 is a diagram illustrating these same parameters with adaptive overestimation of the spectral density of the noise model;
• FIG. 6 represents a typical example of a signal coming from a noisy sound recording;
• FIG. 7 is a flowchart showing the steps of a particular method of searching for a noise model;
• and FIG. 8 is a detailed flow diagram representing the steps of the digital filtering method according to a preferred embodiment of the invention.

The main phases and stages of the process according to the invention will now be described with reference to block diagram of figure 1. Each block, referenced 0 to 5, represents a phase of the process, which itself can be subdivided into elementary steps.

In what follows, to fix the ideas and without this in any way limits the scope of the invention, we will be placed, within the framework of speech processing noisy. As noted earlier, it is common to consider that stationary regimes, in particular the voiced, established over durations between 10 and 20 ms, time interval characteristic of the phenomena of speech production and that will be referred to as the frame below.

As in the known art, the method of the invention, includes a step of splitting the signal into frames audiophonic to denois (block 0).

In practice, digital techniques are used. Also, the frame signals are not "continuously evolving" signals, but discrete signals obtained by sampling. It is assumed that the signals are sampled at period T _e , before digital processing. It is common to then consider 2 ^p samples for a signal frame, choosing p so that the value 2 ^p T _e is of the order of magnitude of the duration D of a frame. By way of example, for a sampling frequency of 10 kHz, frames of 12.8 ms are often chosen, so that 128 points are available for each frame, which constitutes a power of two. The number of samples corresponding to a frame will be noted below LGtrame. The following relation: D = LGframe × T _e is therefore satisfied. The frame cutting step, as indicated in FIG. 1, is therefore preceded by a sampling digitization step.

By convention, the input signal will be noted u (t) , the useful signal s (t) and the disturbing noise x (t) so that: u (t) = s (t) + x (t) in continuous time u (( kT e ) = s (( kT e ) + x (( kT e ) in discrete time

The stages of digitization and cutting into frames (block 0) are common to known art. The samples numeric thus created are stored in a buffer memory circulating type "FIFO" (ie type "first in - first out ") to be read as frames successive.

The successively read frames then undergo a series of autonomous processing steps, in two ways that we can qualify as "parallels".

The operations carried out in block 1, consists in identify segments of the signal to be denoised that do not contain only noise. The output of this block consists of a series of digital samples representative of noise only. In other words, a noise model is developed at from noisy signals, or more precisely from frames successively read (block 0). Many processes can be implemented and an example of a method of Noise model research will be explained below.

• l'estimation de la densité spectrale moyenne du bruit (par exemple par spectre moyen et corrélogramme lissé) ;
• la détermination de l'énergie moyenne du modèle de bruit ;
• et la détermination d'un coefficient traduisant la dispersion statistique du bruit.

In block 2, three steps are carried out and consist, from the samples provided by block 1, of performing:

• estimation of the average spectral density of the noise (for example by average spectrum and smoothed correlogram);
• determining the average energy of the noise model;
• and determining a coefficient reflecting the statistical dispersion of the noise.

The above steps, including the last step which is one of the main characteristics of the invention will be detailed below.

In the "parallel" branch, block 3 has a step of estimating the spectral density of the frame signal current and its energy calculation.

In block 4, according to another characteristic essential of the invention, the coefficients of the filter frequency denoising the signal are determined in the manner which will be detailed below. As it has been indicated, the method of the invention is based on a energy compensation and noise overestimation.

Finally, in block 5, the denoised time signal is rebuilt, ensuring the best continuity possible between frames. In applications other than the main application targeted by the invention the signals can be used as is by various methods such as automatic speech recognition. In itself, this phase of the process is common to the known art, and there is no no need to detail the reconstruction method or for processing the signals at the output of block 4.

According to the main characteristic of the invention, the process makes it possible to modify and optimize the coefficients of the Wiener filter used for the phase of denoising proper (block 4), so as to eliminate or, to say the least, strongly attenuate the so-called parasitic noises "musical".

a/ la probabilité de bruit musical est d'autant plus forte que l'estimée des densités spectrales du bruit est instable d'une trame à l'autre ;

b/ la probabilité de présence de bruit musical est d'autant plus forte que l'estimée de la densité spectrale du bruit est faible par rapport à la densité spectrale réelle du bruit.

As mentioned, these noises are due to two main causes:

a / the probability of musical noise is all the greater when the estimate of the spectral densities of the noise is unstable from one frame to another;

b / the probability of the presence of musical noise is higher the lower the estimated spectral density of the noise compared to the real spectral density of the noise.

According to the invention, in relation to the cause a /, the dispersion is quantified by a coefficient from the analysis carried out in block 2, from the model of noise developed in block 1.

Similarly, in relation to the cause b /, to reduce the influence of the spectral density of the noise, in particular when it is low, the method according to the invention overestimates this spectral density, in y introducing a degree of adaptability in order to optimize the perception of the denoised signal.

Before describing in more detail the process of the invention, it is useful to briefly recall the characteristics of a Wiener filter according to known art.

FIG. 2 very schematically illustrates a Wiener filter used to denoise a noisy signal U (n) .

• Yves THOMAS : "Signaux et systèmes linéaires", éditions MASSON (1994) ; et :
• François MICHAUT : "Méthodes adaptatives pour le signal", édition HERMES (1992).

By way of nonlimiting examples, Wiener filters are described in the following books, to which one can profitably refer:

• Yves THOMAS: "Signals and linear systems", MASSON editions (1994); and:
• François MICHAUT: "Adaptive methods for the signal", HERMES edition (1992).

• U(n) : transformée de Fourier discrète du processus aléatoire observé, soit le signal bruité ;
• S(n) : transformée de Fourier discrète du processus "désiré", à estimer par filtrage linéaire de U(n) ;
• X(n) : transformée de Fourier discrète du bruit additif polluant le signal utile ;
• S and(n) : estimation de S(n) exprimée dans le domaine de

Fourier, avec ε=S and-S= erreur d'estimation (S étant le signal débruité réel) ; et

• W(z) : filtre d'estimation exprimé dans le domaine fréquentiel.

In Figure 2 the following conventions have been adopted:

• U (n) : discrete Fourier transform of the observed random process, ie the noisy signal;
• S (n) : discrete Fourier transform of the "desired" process, to be estimated by linear filtering of U (n);
• X (n) : discrete Fourier transform of the additive noise polluting the useful signal;
• S and (n) : estimate of S (n) expressed in the domain of

Fourier, with ε = S and-S = estimation error ( S being the real denoised signal); and

• W (z) : estimation filter expressed in the frequency domain.

The optimal Wiener filter minimizes the distance between the random variables S (n) and S and (n) measured by the mean square error J: J = E [(S (n) - S (not) ) 2 ]

The minimization of this criterion amounts to rendering the estimation error orthogonal to the observed signal, which results in the principle of orthogonality: E [ε (n) .U * (n)] = 0

γ_S

la densité spectrale du signal utile, et

γ_X

la densité spectrale du bruit parasite,

   le filtre de Wiener est décrit par la relation suivante : W(n) = γ S(n) γ S(n) + γ X(n) Noting:

γ _S

the spectral density of the useful signal, and

γ _X

the spectral density of the parasitic noise,

the Wiener filter is described by the following relation: W (n) = γ S (not) γ S (not) + γ X (not)

By taking into account the independence of S (n) and X (n) , we obtain the relation below: γ U = γ S + γ X relation in which γ _U representing the spectral density of the observed signal.

The relation describing the Wiener filter therefore finally becomes: W (n) = γ S (not) γ S (not) + γ X (not) = 1 - γ X (not) γ U (not)

In practice, it is this second formulation of the Wiener filter which is used, since it does not intervene only directly accessible terms, that is to say, on the one hand, the noisy signal received from block 3 and, on the other hand, the noise, previously determined by the calculation of the noise model (block 1).

It should be noted that the coefficients W (n) of the Wiener filter are always positive. If computational artifacts cause a negative value for a coefficient, that coefficient is made equal to zero.

According to known art, the elimination of additive noise by a spectral subtraction method, as it is produced by a Wiener filter, leads to the creation of so-called "musical" noises. To avoid the appearance of these unpleasant noises when listening and harmful to speech intelligibility, or at the very least prevent maximum their appearance, according to a characteristic essential of the invention, the coefficients of the filter Wiener are modified using parameters determined in blocks 2 and 3, in the way that will now be detailed.

When the input signal contains only noise, additional "musical noise" is present because, in practice, the estimate of the ratio γ _s / γ _u fluctuates at each frequency, although in theory this ratio should be equal to unity whatever the frequencies. It is these estimation errors which produce attenuating filters whose variations in the coefficients are random, depending on the frequencies and over time.

To fix the ideas, we consider the example of the denoising of a single noise, sampled at 44 kHz. The spectral density γ _x of a noise model chosen using this signal is determined and the spectral densities γ _u of each frame (of length LG frame ) of this noise.

The variation of these two parameters has been shown in the form of curves in the diagram of FIG. 3, as a function of the number of FFT Fourier transform channels. To plot the curves, it was assumed that the frame length was 128 samples, i.e. LGtrame = 128.

This diagram clearly shows that the paces of the two curves γ _x and γ _u are similar, but the two estimates have a clear difference in amplitude. The main peak of γ _u , which is located at the frequency 2.75 kHz (64 FFT channels corresponding to 22 kHz, i.e. the sampling half-frequency) has an amplitude approximately seven times greater than that of γ _x located at the same frequency . This is the main reason for the presence of "musical" noises. When, for certain frequencies referenced ν, γ _u (ν) is much greater than γ _x (ν), this means, in theory, that the frame contains not only noise but another part of the signal. In this case, Wiener filtering according to known art denoising the corresponding frame as if it contained useful speech signal, which causes the presence of noise residues.

To avoid this parasitic effect, the process according to the invention optimally modifies the coefficients of the Wiener filter and introduces a compensation term energy, artificially overestimating the level of noise, with different levels of adaptivity of this compensation.

The coefficients of the modified Wiener filter obey the following relation: W (ν) = β 1 - α · max · E x E u · γ x ν γ u ν

β : coefficient d'atténuation exponentielle ;

α : coefficient statique de compensation énergétique ;

E _x / E _u : rapport de pondération énergétique ; et

maxi : coefficient de surestimation statistique issu de l'analyse statistique du bruit, ce à partir d'un modèle de bruit établi lors de la phase du procédé correspondant au bloc 1.

Referring again to relation (7), it is easy to see that four new terms have been introduced, namely:

β: exponential attenuation coefficient;

α: static coefficient of energy compensation;

E _x / E _u : energy weighting ratio; and

max : statistical overestimation coefficient from the statistical noise analysis, based on a noise model established during the process phase corresponding to block 1.

Each of these terms will now be explained.

The exponential attenuation coefficient β is a term commonly used in the literature on digital filtering and, more specifically, denoising. A typical value for this parameter is 0.5.

As a nonlimiting example, we can see article by L. Arslan, A. Mc Cree and V. Viswana- Than, entitled: "New Methods for adaptive noise deletion ", IEEEE, May 1995, pages 812-815.

The static energy compensation coefficient ^α makes it possible to overestimate the noise and is particularly relevant in the case of noise suppression alone. Indeed, a typical value of α = 10 applied to the example of FIG. 3 increases the estimate of the average noise spectrum γ _x by approximately +10 dB, which then makes it possible to reduce the level of residual noise, since the coefficients of the Wiener filter cannot be negative. Otherwise, they are then forced to zero.

However, if this modification is very effective in eliminating noise alone, it in turn poses problems when the frames to be denoised contain useful signal. If this useful signal is much more energetic than the noise, this multiplier coefficient α has no effect on the degradation of this signal. But, in the opposite case, there may exist frequencies ν for which a useful signal frame has a non negligible energy but close to that of noise for the same frequencies. In this case, the multiplication by α of γ _x (ν) imposes zero Wiener coefficients W ( ν ) zero and therefore causes the signal energy to disappear for these frequencies.

γ _u : : densité spectrale de la trame de signal considérée (trame de signal faiblement énergétique devant le bruit) ; et

γ _x : densité spectrale du modèle de bruit choisi (bloc 1).

This problem is illustrated in Figures 4a and 4b. In these figures the following conventions have been adopted.

γ _u :: spectral density of the signal frame considered (low energy signal frame in front of noise); and

γ _x : spectral density of the noise model chosen (block 1).

The curve in FIG. 4a shows that the energy of the signal in the frequency band Δν, represented by the spectral density γ _x , is not negligible.

Referring to Figure 4b, we can see that the multiplication of γ _x by the parameter α = 10 makes α. γ _x greater than γ _u in the Δν band. It follows that the Wiener gain is zero for this frequency band which no longer appears in the denoised frame.

The energy weighting report described below reduces this distortion in the signal noisy.

As previously mentioned, noise denoising alone is okay but it can be too brutal within parts of the useful signal.

In a preferred variant of the invention, remedy this drawback by varying the coefficient α, depending on the presence or not of part of useful signal in the signal to denois. In a way advantageous, α remains close to a typical equal value to 10, when the noisy signal contains only noise, and varies between 0 and 10, when a useful signal is present in the noisy signal. A degree is therefore advantageously introduced adaptivity.

It is the function which is assigned to the ratio E _x / E _u which comes to multiply ^α in the relation (8), ratio in which E _x is the average energy of the noise model and E _u the energy of the current frame . This therefore allows the coefficients of the Wiener filter to change at each frame in a differentiated manner according to the greater or lesser presence (in terms of energy) of the speech signal.

If E _x ≅ E _u , then α≅10 and the frame is considered as noise alone. It is properly denoised.

If on the contrary E _x << E _u , this means that the frame considered is very energetic in front of the noise and that it is necessary to attenuate this signal part as much as possible.

This third modification is illustrated by FIG. 5. In this figure, the signal frame considered is the same as that used for FIGS. 4a and 4b, α = 10 and E x E u = 0.2.

Thanks to this weighting of the coefficient α by E _xx / E _uu , the frequency band Δν 'in which the useful signal is eliminated, (i.e. the frequencies for which the coefficients of γ _x are greater than those of γ _u ) is much less important than during the modification by multiplication of the only coefficient α = 10.

This type of filter therefore has good efficiency. in terms of eliminating degraded signal segments in which speech is absent and decrease in distortions inflicted on the wanted speech signal.

The probability of generating "musical noise" is also related, as noted, to the variance of estimates of the spectral density of noise over all frames.

Indeed, the higher the estimated spectral densities of the noise vary from frame to frame, plus the formation of the "musical" noise is likely.

According to another important aspect of the invention, the value of the overestimation coefficient is made dependent on the statistical properties of the noise. To do this, a coefficient, called maximum below, is introduced proportional to the dispersion of the spectral density values of the noise.

relation dans laquelle :

• N est le nombre de trames du modèle de bruit ;
• ν décrit l'ensemble des canaux fréquentiels, soit LGtrame/2 canaux ;
• γ _i(ν) est la densité spectrale de la i ème trame du modèle de bruit dans le canal ν ; et
• γ _x(ν) est la densité spectrale du modèle de bruit.

The overestimation coefficient then becomes:
α = α * max , with max satisfying the following relation:

relationship in which:

• N is the number of frames of the noise model;
• ν describes all the frequency channels, ie LGframe / 2 channels;
• γ _i ( ν ) is the spectral density of the i th frame of the noise model in the channel ν; and
• γ _x ( ν ) is the spectral density of the noise model.

The maximum coefficient is equal to the maximum ratio, for all the frames of the noise model, between the maximum of the spectral density of the frame of the noise model considered, and the maximum of the estimated spectral density of the noise model.

In other words, this coefficient characterizes the maximum noise disparity for frequency channels carrying significant energy. Multiplied by the coefficient α, it provides proportional additional attenuation to this disparity.

To elaborate part of the parameters entering changing the coefficients of the Wiener filter it is necessary to have a noise model (block 1 of the figure 1).

Developing a signal noise model noisy is a classic operation in itself. However, the specific method implemented for this operation can to be a method of known art, but also a method original.

We will describe below, with reference to Figures 6 and 7, a method for developing a noise model, particularly suitable for the main targeted applications by the process of the invention, in particular the denoising of noisy speech signals.

The method is based on a permanent and automatic search for a noise model. This research is done on the signal samples u (t) digitized and stored in an input buffer memory. This memory is capable of simultaneously memorizing all the samples of several frames of the input signal (at least 2 frames and, in the general case, N frames).

The noise model sought consists of a succession of several frames including energy stability and the relative energy level suggest that it is ambient noise and not a speech signal or other disturbing noise. We will see later how this automatic search.

When a noise model is found, all the samples of the N successive frames representing this noise model are kept in memory, so that the spectrum of this noise can be analyzed and can be used for denoising. But the automatic noise search continues from the input signal u (t) to possibly find a newer and more suitable model, either because it better represents the ambient noise, or because the ambient noise has evolved. The more recent noise model is stored in place of the previous one, if the comparison with the previous one shows that it is more representative of the ambient noise.

• le bruit qu'on veut éliminer est le bruit de fond ambiant,
• le bruit ambiant a une énergie relativement stable à court terme,
• la parole est le plus souvent précédée d'un bruit de respiration du pilote qu'il ne faut pas confondre avec le bruit ambiant; mais ce bruit de respiration s'éteint quelques centaines de millisecondes avant la première émission de parole proprement dite, de sorte qu'on ne retrouve que le bruit ambiant juste avant l'émission de parole,
• et enfin, les bruits et la parole se superposent en termes d'énergie de signal, de sorte qu'un signal contenant de la parole ou un bruit perturbateur, y compris la respiration dans le microphone, contient forcément plus d'énergie qu'un signal de bruit ambiant.

The starting postulates for the automatic development of a noise model are as follows:

• the noise that we want to eliminate is the ambient background noise,
• ambient noise has a relatively stable energy in the short term,
• speech is most often preceded by a breathing noise from the pilot which should not be confused with ambient noise; but this breathing noise goes out a few hundred milliseconds before the first actual speech emission, so that we only find the ambient noise just before the speaking emission,
• and finally, noise and speech are superimposed in terms of signal energy, so that a signal containing speech or disturbing noise, including breathing in the microphone, necessarily contains more energy than a ambient noise signal.

It follows that we will make the simple assumption next: ambient noise is a signal with a minimum energy stable in the short term. In the short term, it must hear a few frames, and we will see in the example practice given below that the number of frames intended for assess the noise stability is from 5 to 20. The energy must be stable on several frames, otherwise we must assume that the signal contains speech or a noise other than ambient noise. It must be minimal, otherwise, the signal is considered to contain breathing or phonetic elements of speech resembling noise but superimposed on ambient noise.

Figure 6 shows a typical configuration of time evolution of the energy of a signal microphone at the start of a broadcast, speech, with a breathing noise phase, which goes out during a few tens to hundreds of milliseconds to make room for ambient noise alone, after which an energy level high indicates the presence of speech, to finally return to the ambient noise.

The automatic search for ambient noise then consists in finding at least N1 successive frames (for example N1 = 5) whose energies are close to each other, that is to say that the ratio between the signal energy contained in a frame and the signal energy contained in the or, preferably, the preceding frames is located within a determined range of values (for example between 1/3 and 3). When such a succession of relatively stable energy frames has been found, the digital values of all the samples of these N frames are stored. This set of NxP samples constitutes the current noise model. It is used in denoising. Analysis of the following frames continues. If we find another succession of at least N1 successive frames meeting the same conditions of energy stability (frame energy ratios in a determined range), then we compare the average energy of this new succession of frames to l average energy of the stored model, and the latter is replaced by the new succession if the ratio between the average energy of the new succession and the average energy of the stored model is less than a determined replacement threshold which may be 1, 5 for example.

From this replacement of a noise model by a more recent model which is less energetic or not much more energetic, it follows that the noise model is generally wedged on permanent ambient noise. Even before speaking, preceded by breathing, there is a phase where ambient noise alone is present for a sufficient time to be taken into account as an active noise model. This phase of ambient noise alone, after breathing, is brief. The number N1 is chosen to be relatively low, so that there is time to readjust the noise model to the ambient noise after the breathing phase.

If the ambient noise changes slowly, the change will be taken into account that the comparison threshold with the stored model is greater than 1. If it evolves more rapidly in the increasing direction, evolution risks not not be taken into account, so it is better to schedule a reset from time to time looking for a noise pattern. For example, on an airplane on the ground when stationary, the ambient noise will be relatively low, and it should not be that during the takeoff phase the noise model remains frozen on what it was at standstill the fact that a noise model is only replaced by a less energetic or not much more energetic. The reset methods will be explained later. considered.

Figure 7 shows a flowchart of automatic noise pattern search operations ambient.

The input signal u (t) , sampled at the frequency F _e = 1 / T _e and digitized by an analog-digital converter, is stored in a buffer memory capable of storing all the samples of at least 2 frames.

The number of the current frame in a search operation for a noise model is designated by n and is counted by a counter as the search is carried out. At the initialization of the search, n is set to 1. This number n will be incremented as a model of several successive frames is developed. When analyzing the current frame n , the model already includes by hypothesis n -1 successive frames meeting the conditions imposed to be part of a model.

We first consider that this is a first model development, no other previous model having been built. We will then see what happens for further elaborations.

The signal energy of the frame is calculated by summation of the squares of the numerical values of the samples of the frame. It is kept in memory.

We then read the next frame of rank n = 2, and its energy is calculated in the same way. It is also kept in memory.

The ratio between the energies of the two frames is calculated. If this ratio is between two thresholds S and S 'one of which is greater than 1 and the other of which is less than 1, it is considered that the energies of the two frames are close and that the two frames can be part of a noise model. The thresholds S and S ' are preferably inverse to each other ( S ' = 1 / S ) so that it suffices to define one to have the other. For example, a typical value is S = 3, S '= 1/3. If the frames can be part of the same noise model, the samples that compose them are stored to start building the model, and the search continues by iteration by incrementing n by one.

If the ratio between the energies of the first two frames leaves the imposed interval, the frames are declared incompatible and the search is reset by resetting n to 1.

In the case where the search continues, the rank n of the current frame is incremented, and in an iterative procedure loop, the energy of the next frame is calculated and a comparison with the energy of the previous frame or previous frames, using thresholds S and S ' .

It will be noted in this connection that two types of comparison are possible for adding a frame to n -1 previous frames which have already been considered as energy homogeneous: the first type of comparison consists in comparing only the energy of the frame n to l energy of the frame n -1. The second type consists in comparing the energy of frame n with each of frames 1 to n -1. The second way leads to a greater homogeneity of the model but it has the disadvantage of not taking into account sufficiently well the cases where the noise level increases or decreases quickly.

Thus, the energy of the frame of rank n is compared with the energy of the frame of rank n -1 and possibly of other previous frames (not necessarily all for that matter).

• ou bien n est inférieur ou égal à un nombre minimal N1 en dessous duquel le modèle ne peut pas être considéré comme significatif du bruit ambiant parce que la durée d'homogénéité est trop courte; par exemple N1 = 5; dans ce cas on abandonne le modèle en cours d'élaboration, et on réinitialise la recherche au début en remettant n à 1 ;
• ou bien n est supérieur au nombre minimal N1. Dans ce cas, puisqu'on trouve maintenant un manque d'homogénéité, on considère qu'il y a peut-être un début de parole après une phase de bruit homogène, et on conserve à titre de modèle de bruit tous les échantillons des n-1 trames de bruit homogènes qui ont précédé le manque d'homogénéité. Ce modèle reste stocké jusqu'à ce qu'on trouve un modèle plus récent qui semble également représenter du bruit ambiant. La recherche est réinitialisée de toute façon en remettant n à 1.

If the comparison indicates that there is no homogeneity with the previous frames, because the energy ratio is not between 1 / S and S , two cases are possible:

• or n is less than or equal to a minimum number N1 below which the model cannot be considered as significant of the ambient noise because the duration of homogeneity is too short; for example N1 = 5; in this case we abandon the model being developed, and we reset the search at the beginning by giving n to 1;
• or n is greater than the minimum number N1. In this case, since we now find a lack of homogeneity, we consider that there may be a start of speech after a homogeneous noise phase, and we keep as a noise model all the samples of the n -1 homogeneous noise frames which preceded the lack of homogeneity. This model remains stored until a newer model is found which also appears to represent ambient noise. The search is reset anyway by resetting n to 1.

But the comparison of the frame n with the previous ones could still have resulted in the observation of a frame still homogeneous in energy with the previous one (s). In this case, either n is less than a second number N2 (for example N2 = 20) which represents the maximum length desired for the noise model, or else n has become equal to this number N2. The number N2 is chosen so as to limit the computation time in the subsequent operations for estimating the spectral noise density.

If n is less than N2 , the homogeneous frame is added to the previous ones to help build the noise model, n is incremented and the next frame is analyzed.

If n is equal to N2 , the frame is also added to the previous n -1 homogeneous frames and the model of n homogeneous frames is stored for use in eliminating noise. The search for a model is also reset by setting n to 1.

The previous steps relate to the first model search. But once a model has been stored, it can be replaced at any time by a model more recent.

The replacement condition is still a energy condition but this time it's about the average energy of the model and no longer on the energy of each frame.

Consequently, if a possible model has just been found, with N frames where N1 < N < N2, we calculate the average energy of this model which is the sum of the energies of the N frames, divided by N , and we compare it at the average energy of the N 'frames of the previously stored model.

If the ratio between the average energy of the new possible model and the average energy of the current model in force is less than a replacement threshold SR, the new model is considered to be better and it is stored in place of the previous one. Otherwise, the new model is rejected and the old one remains in force.

The threshold SR is preferably slightly greater than 1.

If the SR threshold were less than or equal to 1, we would store the least energetic homogeneous frames each time, which corresponds well to the fact that we consider that ambient noise is the energy level below which we never descend . But, we would eliminate any possibility of evolution of the model if the ambient noise started to increase.

If the SR threshold was too high above 1, there is a risk of making a poor distinction between ambient noise and other disturbing noises (breathing), or even certain phonemes that sound like noise (whistling or hissing consonants for example). Eliminating noise from a model of noise calibrated on respiration or on whistling or hissing consonants would risk damaging the intelligibility of the denoised signal.

In a preferred example the threshold SR is approximately 1.5. Above this threshold we will keep the old model; below this threshold we will replace the old model with the new one. In both cases, the search will be reinitialized by recommencing the reading of a first frame of the input signal u (t) , and setting n to 1.

To make the development of the noise model more reliable, we can predict that the search for a model is inhibited if speech emission is detected in the useful signal. Digital signal processing commonly used in speech detection allow identify the presence of words based on the characteristic spectra of periodicity of certain phonemes, in particular phonemes corresponding to vowels or voiced consonants.

The purpose of this inhibition is to prevent certain sounds are taken for noise, whereas they are useful phonemes, that a noise model based on these sounds be stored and that noise suppression after the development of the model then tends to remove all similar sounds.

Furthermore, it is desirable to provide from time to time for a reset of the search for the model to allow updating of the model when the increases in ambient noise have not been taken into account since SR is not much greater than 1.

Ambient noise can indeed increase significantly and rapidly, for example during the acceleration phase of the engines of an airplane or other vehicle, air, land or sea. However, the threshold SR requires that the previous noise model be kept when the average noise energy increases too quickly.

If we want to remedy this situation, we can do it in different ways but the most simple is to reset the model periodically to looking for a new model and imposing it as a model active regardless of the comparison between this model and the previously stored model. Periodicity can be based on the average duration of speech in the application considered; for example the speaking times are in average of a few seconds for the crew of an airplane, and resetting can take place with a frequency of a few seconds.

The implementation of the method of developing a noise model (Figure 1: block 1) and, more general of the process according to the invention, can be done at from non-specialized computers, provided with necessary calculation programs and receiving the digitized signal samples as supplied by an analog-digital converter, via a port adapted.

This implementation can also be done from a specialized computer based on signal processors digital, which enables faster processing large number of digital signals.

The computers are associated, as is well known, with different types of memories, static and dynamic, for recording the programs and the intermediate data, as well as with circulating memories of the "FIFO" type. Finally, the system includes an analog-to-digital converter, for digitizing u (t) signals , and a digital-to-analog converter, as needed, if the denoised signals are to be used in analog form.

In conclusion, and to describe more detailed the process of the invention, we can cut the steps differently than what has been described in reference to Figure 1 (which illustrates the process so more synthetic). Figure 8 is a summary diagram all the stages of the filtering process according to the invention, in a preferred embodiment.

These stages are divided into a first subset steps to determine the parameters depending on the noise model, and a second subset steps to determine the dependent parameters only from the current frame of the signal to be denoised.

The first step of the first subset, includes an initial step of selecting a suitable noise model to the specific application, advantageously a model of noise determined by the method described above, in reference to Figures 6 and 7.

This first subset of steps includes two branches.

In the first branch, the energy of the frame is calculated for each frame of the noise model (in the time domain), then the average energy of the frames of the model is calculated, which makes it possible to estimate the average energy of the model, i.e. the parameter E _x .

In the second branch, a Fourier transform is applied to the frames of the noise model, so as to pass into the frequency domain. Then we determine successively the spectral density of the frame i (with i = 1 .. N ) of the noise model in the frequency channel ν , ie γ _i ( ν ) , and the spectral density of the noise model in the frequency channel ν , let γ _x ( ν ) . From these two parameters, the maximum statistical coefficient is determined so that it verifies the relation (9). The parameter γ _x ( ν ) is also used for the calculation of one of the other coefficients of the Wiener filter.

The second subset of steps also includes two branches.

In the first branch, the energy of the current frame, E _u , is determined, and in the second branch, the spectral density of the current frame γ _{u is} estimated.

From these two parameters and the parameters γ _x and E _x , determined previously, we obtain the coefficients [E _x / E _u ] and [γ _x ( ν ) / γ _u ( ν ) ].

All coefficients of the Wiener filter, conforming to relation (8), are therefore determined at the end of these steps. The coefficients α and β are fixed coefficients predetermined, typically equal to 10 and 0.5, respectively.

On reading the above, we can easily see that the invention does achieve the goals it has set for itself.

It should be clear, however, that the invention is not not limited to only examples of achievements explicitly described, in particular in relation to FIGS. 1 to 8.

In particular, the numerical examples have only been given only to better specify the invention but are essentially related to the specific application envisaged. Therefore, they participate in a simple technological choice within the reach of the skilled person.

Furthermore, as will be recalled, the invention is not not reduced to the only domain of filtering of signals containing noisy speech, even if this domain constitutes one of the favorite apps.

Claims (9)

Frequency filtering method for denoising noisy sound signals ( u (t) ) consisting of so-called useful sound signals mixed with noise signals, the method comprising at least one step of cutting (0) said sound signals into a series of frames identical of a determined length and a frequency filtering step (4) using a Wiener filter, characterized in that it further comprises the following steps: elaboration from said noisy signals ( u (t) ) of a noise model (1) over a determined number N of said frames, N being comprised between predetermined minimum and maximum limits; applying a Fourier transform to said N frames; estimation (2), for each frame of said model, of the spectral density of this frame; estimation (2) of the average spectral density of said noise model; calculation (2), from these two estimates, of a statistical overestimation coefficient, said statistical coefficient being equal to the maximum ratio, for said N frames of the noise model, between the maximum of the spectral density of a frame considered said noise model, and the maximum of the estimated spectral density of the noise model; estimation (3), for each frame of said signals to be denoised ( u (t) ), of its spectral density; and: modification (4), for each frame of said signals to be denoised ( u (t) ), of the coefficients of said Wiener filter so that the following relation is verified: W (ν) = β 1 - α · max · γ x ν γ u ν , relation in which α and β are predetermined fixed coefficients, called static energy compensation coefficient and exponential attenuation coefficient, respectively, ν describes the set of frequency channels of said Fourier transform, γ _u ( ν ) being the estimated of the spectral density of the frame to be denoised, γ _x ( ν ) is said spectral density of the noise model, and maxi said statistical overestimation coefficient, modifying the static energy compensation coefficient α. Method according to claim 1, characterized in that said maximum statistical coefficient satisfies the following relationship:

Method according to one of claims 1 or 2, characterized in that it comprises the following additional steps: calculating the average energy of said noise model E _x ; calculation, for each frame of said signals to be denoised ( u (t) ), of the energy of the current frame E _u ; and multiplication of said static energy compensation coefficient α by an energy weighting coefficient equal to the ratio E _x / E _u , so as to selectively modify these coefficients for each frame of said signals to be denoised ( u (t) ) by applying a coefficient continuously variable between an extrema and a minima, the extrema being substantially equal to unity when said useful signals are absent from said signals to be denoised ( u (t) ) and substantially equal to zero when the energy of said useful signals is much greater than the energy of said noise signals, and that said coefficients of the Wiener filter satisfy the following relation: W (ν) = β 1 - α · max · E x E u · γ x ν γ u ν Method according to any one of previous claims, characterized in that said static energy compensation coefficient α is equal to 10. Method according to any one of previous claims, characterized in that said exponential attenuation coefficient β is 0.5. Method according to any one of the preceding claims, characterized in that it comprises an initial step (0) consisting in digitizing said signals to be denoised ( u (t) ) by sampling, each frame comprising p samples. Method according to Claim 6, characterized in that the said noise model (1) is obtained by a repetitive search carried out continuously in the said signals to be denoised ( u (t) ), by searching for N successive frames, of p samples each, having the expected characteristics of a noise, by storing the corresponding NxP samples to constitute said noise model, and by repeating the search to find a new noise model and to store the new model to replace the previous one or to keep the previous model according to the characteristics respective of the two models. Application of the method according to any one of the preceding claims to the denoising of noisy speech signals ( u (t) ). Application of the method according to claim 8, characterized in that the duration of said frames is in the range 10 to 20 ms.

Images