EP2772916B1 - Method for suppressing noise in an audio signal by an algorithm with variable spectral gain with dynamically adaptive strength - Google Patents

Description

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

It relates in particular to the processing of speech signals picked up by "hands-free" telephony devices intended to be used in a noisy environment.

These devices include one or more microphones 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 unintelligible the words of the speaker. 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" for motor vehicles, whether in-vehicle equipment or accessories in the form of removable housing incorporating all components and functions signal processing for telephone communication.

Indeed, the large distance between the microphone (placed at the dashboard or in an upper corner of the roof of the cockpit) and the speaker (whose distance is constrained by the driving position) leads to the capture of a relatively low level of speech compared to the ambient noise, which makes it difficult to extract the useful signal embedded in the noise. In addition to this permanent stationary rolling noise component, the very noisy environment typical of the automotive environment has non-stationary spectral characteristics, that is to say, which evolve unpredictably according to the driving conditions: passage over deformed or paved roads, car radio in operation, etc.

Similar difficulties arise in the case where the device is a headset type microphone / headset combined used for communication functions such as "hands-free" telephony functions, in addition to listening to a source audio (music for example) from a device to which the headphones are connected.

In this case, it is a question of ensuring sufficient intelligibility of the signal picked up by the microphone, that is to say the speech signal of the close speaker (the helmet wearer). Now, the headset can be used in a noisy environment (Metro, busy street, train, etc.), so that the microphone will not only pick up the words of the wearer of the helmet, but also the surrounding noise. The wearer is certainly protected from this noise by the helmet, especially if it is a model with closed earphones isolating the ear from the outside, and even more if the headset is provided with an "active control of noise". On the other hand, the distant speaker (the one at the other end of the communication channel) will suffer from the noise picked up by the microphone and being superimposed and interfere with the speech signal of the near speaker (the helmet wearer). In particular, certain speech formers essential to the understanding of the voice are often embedded in noise components commonly encountered in the usual environments.

The invention relates more particularly to single-channel selective denoising techniques, that is to say operating on a single signal (as opposed to techniques using several microphones whose signals are combined judiciously and are subject to a spatial or spectral coherence analysis, for example by beamforming or other techniques). However, it will apply with the same relevance to a signal recomposed from several microphones by a beamforming technique , insofar as the invention presented here applies to a scalar signal.

In this case, it is to operate the selective denoising of a noisy audio signal, generally obtained after scanning the signal collected by a single microphone of the telephony equipment.

The invention aims more particularly at an improvement made to the noise reduction algorithms based on signal processing in the frequency domain (thus after application of an FFT Fourier transformation) of applying a calculated spectral gain according to several estimators of probability of presence of speech.

More specifically, the signal y from the microphone is cut into frames of fixed length, overlapping or not, and each frame of index k is transposed into the frequency domain by FFT. The resulting frequency signal Y ( k, l ) , which is also discrete, is then described by a set of frequency "bins" (frequency bands) of index l , typically 128 bins of positive frequencies.

For each signal frame, a number of estimators are updated to determine a frequency probability of speech presence p ( k, l ) . If the probability is large, the signal will be considered as useful signal (speech) and thus preserved with a spectral gain G ( k, l ) = 1 for the considered bin. In the opposite case, if the probability is low, the signal will be assimilated to noise and thus reduced or even eliminated by applying a spectral attenuation gain that is much smaller than 1.

• Cette technique peut être notamment implémentée par un algorithme de type OM-LSA (Optimally Modified - Log Spectral Amplitude) telle que ceux décrits par :
  1. [1] I. Cohen et B. Berdugo, "Speech Enhancement for Non-Stationary Noise Environments", Signal Processing, Vol. 81, No 11, pp. 2403-2418, Nov. 2001 ; et
  2. [2] I. Cohen, "Optimal Speech Enhancement Under Signal Presence Uncertainty Using Log-Spectral Amplitude Estimator", IEEE Signal Processing Letters, Vol. 9, No 4, pp. 113-116, Apr. 2002 .

In other words, the principle of this algorithm consists in calculating and applying to the useful signal a "frequency mask" which preserves the useful information of the speech signal and eliminates the noise noise signal:

• This technique can be implemented in particular by an OM-LSA type algorithm ( Optimally Modified - Log Spectral Amplitude) such as those described by:
  1. [1] I. Cohen and B. Berdugo, "Speech Enhancement for Non-Stationary Noise Environments," Signal Processing, Vol. 81, No. 11, pp. 2403-2418, Nov. 2001 ; and
  2. [2] I. Cohen, "Optimal Speech Enhancement Under Signal Presence Uncertainty Using Log-Spectral Amplitude Estimator," IEEE Signal Processing Letters, Vol. 9, No. 4, pp. 113-116, Apr. 2002 .

The US 7,454,010 B1 also describes a comparable algorithm taking into account, for the calculation of the spectral gains, information of presence or absence of the voice in a current time segment.

We can also refer to WO 2007/099222 A1 (Parrot ), which describes a denoising technique implementing a probability calculation of presence of speech.

The effectiveness of such a technique lies, of course, in the model of the speech probability estimator which must discriminate between speech and noise.

In practice, the implementation of such an algorithm encounters a number of defects, the two main ones being the "musical noise" and the appearance of a "robotic voice".

"Music noise" is characterized by a non-uniform residual background noise layer, favoring certain specific frequencies. The tone of the noise is then no longer natural, which makes listening disturbing. This phenomenon results from the fact that the frequency treatment of denoising is operated without dependence between neighboring frequencies during frequency discrimination between speech and noise, because the treatment does not include a mechanism to prevent two very different neighboring spectral gains. However, in periods of noise alone, ideally a uniform attenuation gain is needed to preserve the tone of the noise; but in practice, if the spectral gains are not homogeneous, the residual noise becomes "musical" with the appearance of frequency notes at less attenuated frequencies, corresponding to bins falsely detected as containing useful signal. It should be noted that this phenomenon is all the more marked as it allows the application of significant attenuation gains.

The phenomenon of "robotic voice" or "metallic voice", meanwhile, occurs when one chooses to operate a very aggressive noise reduction, with significant spectral attenuation gains. In the presence of speech, frequencies corresponding to speech but which are falsely detected as noise will be strongly attenuated, making the voice less natural or even totally artificial ("robotization" of the voice).

The parameterization of such an algorithm therefore consists in finding a compromise on the aggressiveness of the denoising, so as to remove a maximum of noise without the undesirable effects of the application of too large attenuation spectral gains becoming too perceptible. This last criterion proves however extremely subjective, and on a comparatively large control group of users it is difficult to find a compromise setting that can be unanimous.

To minimize these defects, inherent to a denoising technique by applying a spectral gain, the "OM-LSA" model provides for setting a lower bound G _min for the attenuation gain (expressed on a logarithmic scale, this gain of attenuation therefore corresponds in the remainder of this document to a negative value) applied to the zones identified as noise, so as to prevent too much denoising to limit the appearance of the defects mentioned above. This solution is however not optimal: it certainly helps to eliminate the undesirable effects of excessive noise reduction, but at the same time it limits the performance of denoising.

The problem of the invention is to overcome this limitation, by making the noise reduction system more efficient by applying a spectral gain (typically according to an OM-LSA model), while respecting the constraints mentioned above, namely effectively reduce noise without altering the natural appearance of speech (in the presence of speech) or of noise (in the presence of noise). In other words, the undesirable effects of the algorithmic processing must be made imperceptible by the remote speaker, while at the same time attenuating the noise significantly.

The basic idea of the invention is to modulate the calculation of the spectral gain G _OMLSA - calculated in the frequency domain for each bin - by a global indicator observed at the time frame and no longer at a single level. frequency bin.

This modulation will be performed by a direct transformation of the lower limit G _min of the attenuation gain - a terminal which is a scalar commonly referred to as "denoising hardness" - into a time function whose value will be determined according to a temporal descriptor ( or "global variable") reflected by the state of the various estimators of the algorithm. These will be chosen according to their relevance to describe known situations for which it is known that the choice of denoising hardness G _min can be optimized.

Subsequently and depending on the case, the temporal modulation applied to this logarithmic G _min attenuation gain may correspond to either an increment or a decrement: a decrement will be associated with a greater noise reduction hardness (gain logarithmic greater in absolute value), conversely an increment of this negative logarithmic gain will be associated with a smaller absolute value, hence a lower noise reduction hardness.

Indeed, it can be seen that a frame-scale observation can often correct certain defects in the algorithm, particularly in very noisy areas where it can sometimes falsely detect a noise frequency as a frequency of speech: thus, if a frame of only noise is detected (at the level of the frame), one will be able to denoise in a more aggressive way without introducing musical noise, thanks to a more homogeneous denoising.

• dans le premier mode de réalisation, l'optimisation consistera à moduler dans le sens adéquat la valeur de la dureté de débruitage G_min pour mieux réduire le bruit en phase de bruit seul, et mieux préserver la voix en phase de parole ;

Conversely, over a period of noisy speech, we can allow less denoising in order to perfectly preserve the voice while ensuring that the energy variation of the residual background noise is not noticeable. We thus have a double lever (hardness and homogeneity) to modulate the importance of denoising according to the case considered - noise phase alone or speech phase -, the discrimination between one or the other case resulting from an observation on the scale of the time frame:

• in the first embodiment, the optimization will consist in modulating in the appropriate direction the value of the denoising hardness G _min to better reduce noise in the noise phase alone, and better preserve the voice in the speech phase;

1. a) génération de trames temporelles successives du signal audio bruité numérisé ;
2. b) application d'une transformation de Fourier aux trames générées à l'étape a), de manière à produire pour chaque trame temporelle de signal un spectre de signal avec une pluralité de bandes de fréquences prédéterminées ;
3. c) dans le domaine fréquentiel :
   • c1) estimation, pour chaque bande de fréquences de chaque trame temporelle courante, d'une probabilité de présence de parole ;
   • c3) calcul d'un gain spectral, propre à chaque bande de fréquence de chaque trame temporelle courante, en fonction de : i) une estimation de l'énergie du bruit dans chaque bande de fréquences, ii) la probabilité de présence de parole estimée à l'étape c1), et iii) une valeur scalaire de gain minimal représentative d'un paramètre de dureté du débruitage ;
   • c4) réduction sélective de bruit par application à chaque bande de fréquences du gain calculé à l'étape c3) ;
4. d) application d'une transformation de Fourier inverse au spectre de signal constitué des bandes de fréquences produites à l'étape c4), de manière à délivrer pour chaque spectre une trame temporelle de signal débruité ; et
5. e) reconstitution d'un signal audio débruité à partir des trames temporelles délivrées à l'étape d).

More specifically, the invention proposes a method of denoise of an audio signal by applying a variable spectral gain algorithm according to a probability of presence of speech, comprising in a manner known per se the following successive steps:

1. a) generating successive time frames of the digitized noisy audio signal;
2. b) applying a Fourier transform to the frames generated in step a), so as to produce for each signal time frame a signal spectrum with a plurality of predetermined frequency bands;
3. c) in the frequency domain:
   • c1) estimating, for each frequency band of each current time frame, a probability of presence of speech;
   • c3) calculating a spectral gain, specific to each frequency band of each current time frame, as a function of: i) an estimate of the noise energy in each frequency band, ii) the probability of presence of estimated speech in step c1), and iii) a minimum gain scalar value representative of a denoising hardness parameter;
   • c4) selective reduction of noise by applying to each frequency band the gain calculated in step c3);
4. d) applying an inverse Fourier transform to the signal spectrum consisting of the frequency bands produced in step c4), providing for each spectrum a time frame of denoised signal; and
5. e) reconstituting an audio signal denoised from the time frames delivered in step d).

• ladite valeur scalaire de gain minimal est une valeur modulable de manière dynamique à chaque trame temporelle successive ; et
• le procédé comporte en outre, préalablement à l'étape c3) de calcul du gain spectral, une étape de :
  • c2) calcul, pour la trame temporelle courante, de ladite valeur modulable en fonction d'une variable globale observée au niveau de la trame temporelle courante pour toutes les bandes de fréquences ; et
• ledit calcul de l'étape c2) comprend l'application, pour la trame temporelle courante, d'un incrément/décrément apporté à une valeur paramétrée nominale dudit gain minimal.

Characteristically, the invention

• said minimum gain scalar value is a dynamically scalable value at each successive time frame; and
• the method further comprises, prior to step c3) of calculating the spectral gain, a step of:
  • c2) calculating, for the current time frame, said scalable value as a function of a global variable observed at the current time frame for all the frequency bands; and
• said calculation of step c2) comprises the application, for the current time frame, of an increment / decrement brought to a nominal parameterized value of said minimum gain.

In a first implementation of the invention, the global variable is a signal-to-noise ratio of the current time frame evaluated in the time domain.

• k étant l'indice de la trame temporelle courante,
• G_min (k) étant le gain minimal à appliquer à la trame temporelle courante,
• G_min étant ladite valeur nominale paramétrée du gain minimal,
• ΔG_min (k) étant ledit incrément/décrément apporté à G_min , et
• SNR_y (k) étant le rapport signal sur bruit de la trame temporelle courante.

The scalar value of minimum gain can notably be calculated in step c2) by applying the relation: $${\mathrm{BOY\; WUT}}_{\mathit{min}}\left(k\right)={\mathrm{BOY\; WUT}}_{\mathit{min}}+\mathrm{\Delta }{\mathrm{BOY\; WUT}}_{\mathit{min}}\left({\mathit{SNR}}_{\mathrm{there}}\left(k\right)\right)$$

• where k is the index of the current time frame,
• G _min ( k ) being the minimum gain to be applied to the current time frame,
• G _min being said parameterized nominal value of the minimum gain,
• Δ G _min (k) being said increment / decrement brought to G _min , and
• SNR _y (k) being the signal-to-noise ratio of the current time frame.

In a second implementation of the invention, the global variable is an average probability of speech, evaluated at the level of the current time frame.

• k étant l'indice de la trame temporelle courante,
• G_min (k) étant le gain minimal à appliquer à la trame temporelle courante,
• G_min étant ladite valeur nominale paramétrée du gain minimal,
• P_speech (k) étant la probabilité moyenne de parole évaluée au niveau de la trame temporelle courante,
• Δ₁ G_min étant ledit incrément/décrément, apporté à G_min en phase de bruit, et
• Δ₂ G_min étant ledit incrément/décrément, apporté à G_min en phase de parole.

The scalar value of minimum gain can notably be calculated in step c2) by applying the relation: $${\mathrm{BOY\; WUT}}_{\mathit{min}}\left(k\right)={\mathrm{BOY\; WUT}}_{\mathit{min}}+\left(P_{\mathit{speech}}\left(k\right)-1\right)\mathrm{.}{\mathrm{\Delta }}_1{\mathrm{BOY\; WUT}}_{\mathit{min}}+P_{\mathit{speech}}\left(k\right)\mathrm{.}{\mathrm{\Delta }}_2{\mathrm{BOY\; WUT}}_{\mathit{min}}$$

• where k is the index of the current time frame,
• G _min ( k ) being the minimum gain to be applied to the current time frame,
• G _min being said parameterized nominal value of the minimum gain,
• P _speech ( k ) being the average probability of speech evaluated at the level of the current time frame,
• Δ ₁ G _min being said increment / decrement, brought to G _min in the noise phase, and
• Δ ₂ G _min being said increment / decrement, brought to G _min in speech phase.

• l étant l'indice de la bande de fréquences,
• N étant le nombre de bandes de fréquences dans le spectre, et
• p(k,l) étant la probabilité de présence de parole de la bande de fréquences d'indice l de la trame temporelle courante.

The average probability of speech can notably be evaluated at the level of the current time frame by applying the relation: $$P_{\mathit{speech}}\left(k\right)=\frac{1}{\mathrm{NOT}}{\displaystyle \underset{l}{\overset{\mathrm{NOT}}{\Sigma }}}p\left(k,l\right)$$

• where l is the index of the frequency band,
• N being the number of frequency bands in the spectrum, and
• p ( k , l ) being the probability of presence of speech of the index frequency band I of the current time frame.

In a third implementation of the invention, the global variable is a Boolean voice activity detection signal for the current time frame, evaluated in the time domain by analysis of the time frame and / or by means of an external detector.

• k étant l'indice de la trame temporelle courante,
• G_min (k) étant le gain minimal à appliquer à la trame temporelle courante,
• G_min étant ladite valeur nominale paramétrée du gain minimal,
• VAD (k) étant la valeur du signal booléen de détection d'activité vocale pour la trame temporelle courante, et
• ΔG_min étant ledit incrément/décrément apporté à G_min .

The scalar value of minimum gain can notably be calculated in step c2) by applying the relation: $${\mathrm{BOY\; WUT}}_{\mathit{min}}\left(k\right)={\mathrm{BOY\; WUT}}_{\mathit{min}}+\mathit{VAD}\left(k\right)\mathrm{.}\mathrm{\Delta }{\mathrm{BOY\; WUT}}_{\mathit{min}}$$

• where k is the index of the current time frame,
• G _min ( k ) being the minimum gain to be applied to the current time frame,
• G _min being said parameterized nominal value of the minimum gain,
• VAD (k) being the value of the Boolean voice activity detection signal for the current time frame, and
• Δ G _min being said increment / decrement brought to G _min .

• La Figure 1 illustre de façon schématique, sous forme de blocs fonctionnels, la manière dont est réalisé un traitement de débruitage de type OM-LSA selon l'état de la technique.
• La Figure 2 illustre le perfectionnement apporté par l'invention à la technique de débruitage de la Figure 1.

An embodiment of the device of the invention will now be described with reference to the appended drawings in which the same reference numerals designate identical or functionally similar elements from one figure to another.

• The Figure 1 schematically illustrates, in the form of functional blocks, the manner in which an OM-LSA type denoising treatment according to the state of the art is carried out.
• The Figure 2 illustrates the improvement brought by the invention to the technique of denoising the Figure 1 .

The process of the invention is implemented by software means, schematized in the figures by a number of functional blocks corresponding to appropriate algorithms executed by a microcontroller or a digital signal processor. Although, for the sake of clarity, the various functions are presented as separate modules, they implement common elements and correspond in practice to a plurality of functions globally executed by the same software.

OM-LSA denoising algorithm according to the state of the art

The Figure 1 schematically illustrates, in the form of functional blocks, the manner in which an OM-LSA type denoising treatment according to the state of the art is carried out.

The digitized signal y ( n ) = x (n) + d ( n ) comprising a speech component x ( n ) and a noise component d (n) ( n being the rank of the sample) is cut out (block 10 ) in segments or time frames y ( k ) (k being the frame index) of fixed length, overlapping or not, usually frames of 256 samples for a signal sampled at 8 kHz ( narrowband telephone standard).

Each time frame of index k is then transposed into the frequency domain by a fast Fourier transform FFT (block 12): the resulting obtained signal or spectrum Y ( k , l ), also discrete, is then described by a set of Frequency bands or "bins" (where l is the bin index), for example 128 bins of positive frequencies. A spectral gain G = G _OMLSA ( k, l), specific to each bin, is applied (block 14) to the frequency signal Y ( k , l ), to give a signal X ( k, l ): $$\hat{X}\left(k,l\right)={\mathrm{BOY\; WUT}}_{\mathit{OMLSA}}\left(k,l\right)\mathrm{.}Y\left(k,l\right)$$

The spectral gain G _OMLSA (k, l) is calculated (block 16) as a function of a part of a probability of presence of speech p ( k , l ), which is an estimated frequency probability (block 18) for each bin and on the other hand a parameter G _min , which is a scalar value of minimum gain, commonly referred to as "denoising hardness". This parameter G _min sets a lower limit to the attenuation gain applied to the zones identified as noise, in order to avoid that the phenomena of musical noise and robotic voice become too marked due to the application of spectral gains of too much and / or heterogeneous attenuation.

The spectral gain G _OMLSA (k, l) calculated is of the form: $${\mathrm{BOY\; WUT}}_{\mathit{OMLSA}}\left(k,l\right)={\left\{\mathrm{BOY\; WUT}\left(k,l\right)\right\}}^{p\left(k,l\right)}\mathrm{.}{\mathrm{BOY\; WUT}}_{\mathit{min}}^{1-p\left(k,l\right)}$$

• [2] I. Cohen, "Optimal Speech Enhancement Under Signal Presence Uncertainty Using Log-Spectral Amplitude Estimator", IEEE Signal Processing Letters, Vol. 9, No 4, pp. 113-116, Apr. 2002 .

The calculation of the spectral gain and that of the probability of presence of speech are therefore advantageously implemented in the form of an OM-LSA (Optimally Modified Log Spectral Amplitude) type algorithm such as that described in the article (cited above):

• [2] I. Cohen, "Optimal Speech Enhancement Under Signal Presence Uncertainty Using Log-Spectral Amplitude Estimator," IEEE Signal Processing Letters, Vol. 9, No. 4, pp. 113-116, Apr. 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 criterion is adapted because the distance chosen is in better adequacy with the behavior of the human ear and thus gives qualitatively better results.

In all cases, it is a question of reducing the energy of the highly parasitized frequency components by applying them a weak gain, while leaving intact (by the application of a gain equal to 1) those which are it little or not at all.

The algorithm "OM-LSA" ( Optimally-Modified LSA ) improves the computation of the LSA gain by weighting it by the conditional probability p ( k , l ) of presence of speech or SPP ( Speech Presence Probability ), for the computation of the gain final: the noise reduction applied is all the more important (that is to say that the applied gain is even lower) that the probability of presence of speech is low.

• [3] I. Cohen et B. Berdugo, "Two-Channel Signal Detection and Speech Enhancement Based on the Transient Beam-to-Reference Ratio", IEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2003, Hong-Kong, pp. 233-236, Apr. 2003 .

The probability of presence of speech p ( k , l ) is a parameter that can take several different values between 0 and 100%. This parameter is calculated according to a technique in itself known, examples of which are in particular exposed in:

• [3] I. Cohen and B. Berdugo, "Two-Channel Signal Detection and Speech Enhancement Based on the Transient Beam-to-Reference Ratio," IEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2003, Hong Kong, pp. 233-236, Apr. 2003 .

As is frequently the case in this field, the method described is not intended to identify precisely on which frequency components of which frames the speech is absent, but rather to give a confidence index between 0 and 1, a value 1 indicating that the speech is absent for sure (according to the algorithm) while a value 0 declares the opposite. By its nature, this index is likened to the probability of absence of speech a priori, that is to say the probability that speech is absent on a given frequency component of the frame considered. This is of course a non-rigorous assimilation, in the sense that even if the presence of speech is probabilistic ex ante, the signal picked up by the microphone presents at each moment only one of two distinct states: at the moment considered, it can either include speech or not contain it. In practice, however, this assimilation gives good results, which justifies its use.

We can also refer to WO 2007/099222 A1 (Parrot ), which describes in detail a denoising technique derived from this principle, implementing a probability calculation of presence of speech.

The resulting signal X ( k , l ) = G _OMLSA ( k, l ). Y ( k , l ), that is to say the useful signal Y ( k , l ) to which the frequency mask G _OMLSA ( k , l ) has been applied, is then subjected to an inverse Fourier transformation. iFFT (block 20), to go back from the frequency domain to the time domain. The resulting time frames are then collected (block 22) to give a digitized denoised signal x ( n ).

OM-LSA denoising algorithm according to the invention

The Figure 2 illustrates the changes made to the algorithm just described. Blocks bearing the same numerical references correspond to functions that are identical or similar to those described above, as well as the references of the various processed signals.

In the known implementation of the Figure 1 , the scalar value G _min of the minimal gain representative of the denoising hardness was chosen more or less empirically, so that the degradation of the voice remains low audible, while ensuring an acceptable attenuation of the noise. As stated in the introduction, however, it is desirable to denoise more aggressively in the noise phase alone, but without introducing musical noise; conversely, over a period of noisy speech, we can allow less denoising to perfectly preserve the voice while ensuring that the energy variation of the residual background noise is not noticeable.

Depending on the case (noise phase alone or speech phase), there may be a double interest in modulating the hardness of denoising: this will be modulated by dynamically varying the scalar value of G _min , in the appropriate direction that will reduce noise in the noise phase alone and will better preserve voice in the speech phase.

To do this, the scalar value G _min , initially constant, is transformed (block 24) into a time function G _min ( k ) whose value will be determined according to a global variable (also called "temporal descriptor"). that is, a variable considered globally at the level of the frame and not the frequency bin. This global variable can be reflected by the state of one or more different estimators already calculated by the algorithm, which will be chosen according to the case according to their relevance.

These estimators can be: i) a signal-to-noise ratio, ii) an average probability of presence of speech and / or iii) a voice activity detection. In all these examples, the denoising hardness G _min becomes a time function G _min ( k ) defined by the estimators, themselves temporal, making it possible to describe known situations for which it is desired to modulate the value of G _{min in} order to influence on noise reduction by dynamically modifying the signal denoise / degradation compromise.

Incidentally, in order for this dynamic modulation of the hardness not to be perceptible by the listener, it is necessary to provide a mechanism for preventing abrupt variations of G _min ( k ), for example by a conventional temporal smoothing technique. It will thus be avoided that abrupt temporal variations in the hardness G _min ( k ) are audible on the residual noise, which is very often stationary in the case for example of a motorist in rolling condition.

Time descriptor: signal-to-noise ratio

The starting point of this first implementation is the observation that a speech signal picked up in a quiet environment has little or no need to be de-noised, and that an energetic denoising applied to such a signal would lead quickly to audible artifacts, without the comfort of listening improved from the point of view of residual noise.

Conversely, an excessively noisy signal can quickly become unintelligible or cause progressive listening fatigue; in such a case the benefit of a large denoising will be indisputable, even at the cost of an audible (however reasonable and controlled) degradation of speech.

In other words, the noise reduction will be all the more beneficial for the understanding of the useful signal that the untreated signal is noisy.

• G_min (k) étant le gain minimal à appliquer à la trame temporelle courante,
• G_min étant une valeur nominale paramétrée de ce gain minimal,
• ΔG_min (k) étant l'incrément/décrément apporté à la valeur G_min , et
• SNR_y (k) étant le rapport signal sur bruit de la trame courante, évalué dans le domaine temporel (bloc 26), correspondant à la variable appliquée sur l'entrée n° ① du bloc 24 (ces "entrées" étant symboliques et n'ayant qu'une valeur illustrative des différentes possibilités alternatives de mise en oeuvre de l'invention).

This can be taken into account by modulating the hardness parameter G _min as a function of the signal-to-noise ratio a priori or the noise level of the processed signal: $${\mathrm{BOY\; WUT}}_{\mathit{min}}\left(k\right)={\mathrm{BOY\; WUT}}_{\mathit{min}}+\mathrm{\Delta }{\mathrm{BOY\; WUT}}_{\mathit{min}}\left({\mathit{SNR}}_{\mathrm{there}}\left(k\right)\right)$$

• G _min ( k ) being the minimum gain to be applied to the current time frame,
• G _min being a parameterized nominal value of this minimum gain,
• Δ G _min (k) being the increment / decrement brought to the value G _min , and
• SNR _y ( k ) being the signal-to-noise ratio of the current frame, evaluated in the time domain (block 26), corresponding to the variable applied on the input n ° ① of the block 24 (these "entries" being symbolic and n having an illustrative value of the different alternative possibilities of implementing the invention).

Temporal descriptor: mean probability of speech

Another relevant criterion for modulating the hardness of the reduction may be the presence of speech for the time frame considered.

With the conventional algorithm, when one tries to increase the denoising hardness G _min , the phenomenon of "robotic voice" appears before that of "musical noise". Thus, it seems possible and interesting to apply a greater denoising hardness in a noise phase alone, by simply modulating the denoising hardness parameter by a global indicator of presence of speech: in noise-only period, the residual noise - at the origin of the listening fatigue - will be reduced by application of a greater hardness, and this without counterpart since the hardness in speech phase can remain unchanged.

• P_speech (k) étant la probabilité moyenne de parole évaluée au niveau de la trame temporelle courante,
• N étant le nombre de bins du spectre, et
• p(k,l) étant la probabilité de présence de parole du bin d'indice l de la trame temporelle courante.

Since the noise reduction algorithm is based on a frequency presence probability calculation, it is easy to obtain a mean index of the presence of speech at the frame scale from the different frequency probabilities, so as to differentiate the frames mainly made of noise from those containing useful speech. For example, we can use the classical estimator: $$P_{\mathit{speech}}\left(k\right)=\frac{1}{\mathrm{NOT}}{\displaystyle \underset{l}{\overset{\mathrm{NOT}}{\Sigma }}}p\left(k,l\right)$$

• P _speech ( k ) being the average probability of speech evaluated at the level of the current time frame,
• N being the number of bins in the spectrum, and
• p ( k , l ) being the probability of presence of speech of the bin of index l of the current time frame.

• G_min (k) étant le gain minimal à appliquer à la trame temporelle courante,
• G_min étant une valeur nominale paramétrée de ce gain minimal, et
• Δ₁ G_min étant un incrément/décrément apporté à G_min en phase de bruit, et
• Δ₂ G_min étant un incrément/décrément apporté à G_min en phase de parole.

This variable P _speech ( k ) is calculated by the block 28 and applied to the input No. ② of the block 24, which calculates the denoising hardness to be applied for a given frame: $${\mathrm{BOY\; WUT}}_{\mathit{min}}\left(k\right)={\mathrm{BOY\; WUT}}_{\mathit{min}}+\left(P_{\mathit{speech}}\left(k\right)-1\right)\mathrm{.}{\mathrm{\Delta }}_1{\mathrm{BOY\; WUT}}_{\mathit{min}}+P_{\mathit{speech}}\left(k\right)\mathrm{.}{\mathrm{\Delta }}_2{\mathrm{BOY\; WUT}}_{\mathit{min}}$$

• G _min ( k ) being the minimum gain to be applied to the current time frame,
• G _min being a parameterized nominal value of this minimum gain, and
• Δ ₁ G _min being an increment / decrement brought to G _min in the noise phase, and
• Δ ₂ G _min being an increment / decrement brought to G _min in speech phase.

• l'augmentation de la dureté de la réduction de bruit d'un facteur Δ₁ G_min en phase de bruit afin de réduire le bruit résiduel, typiquement Δ₁ > 0, par exemple Δ₁ = +6 dB ; et
• la diminution de la dureté de la réduction de bruit d'un facteur Δ₂ G_min en phase de parole afin de mieux préserver la voix, typiquement Δ₂ < 0, par exemple Δ₂ = -3 dB.

The above expression clearly highlights the two complementary effects of the presented optimization, namely:

• increasing the hardness of the noise reduction by a factor Δ ₁ G _min in the noise phase in order to reduce the residual noise, typically Δ ₁ > 0, for example Δ ₁ = + 6 dB; and
• reducing the hardness of the noise reduction by a factor Δ ₂ G _min in the speech phase in order to better preserve the voice, typically Δ ₂ <0, for example Δ ₂ = -3 dB.

Time descriptor: voice activity detector

In this third implementation, a voice activity detector or VAD (block 30) is used to perform the same type of hardness modulation as in the previous example. Such a "perfect" detector delivers a binary signal (absence vs. presence of speech), and is distinguished from systems delivering only a probability of presence of variable speech between 0 and 100% continuously or in successive steps, which can introduce false important detections in noisy environments.

• G_min (k) étant le gain minimal à appliquer à la trame temporelle courante,
• G_min étant une valeur nominale paramétrée dudit gain minimal,
• VAD (k) étant la valeur du signal booléen de détection d'activité vocale pour la trame temporelle courante, évalué dans le domaine temporel (bloc 30) et appliqué à l'entrée n° ③ du bloc 24, et
• ΔG_min étant l'incrément/décrément apporté à la valeur G_min.

Since the voice activity detection module takes only two distinct values '0' or '1', the modulation of the denoising hardness will be discrete: $${\mathrm{BOY\; WUT}}_{\mathit{min}}\left(k\right)={\mathrm{BOY\; WUT}}_{\mathit{min}}+\mathit{VAD}\left(k\right)\mathrm{.}\mathrm{\Delta }{\mathrm{BOY\; WUT}}_{\mathit{min}}$$

• G _min ( k ) being the minimum gain to be applied to the current time frame,
• G _min being a parameterized nominal value of said minimum gain,
• VAD ( k ) being the value of the Boolean voice activity detection signal for the current time frame, evaluated in the time domain (block 30) and applied to the input No. ③ of the block 24, and
• Δ G _min being the increment / decrement brought to the value G _min .

The voice activity detector 30 may be implemented in different ways, of which three examples of implementation will be given below.

In a first example, the detection is performed from the signal y ( k ), intrinsically to the signal collected by the microphone; an analysis of the more or less harmonic nature of this signal makes it possible to determine the presence of a vocal activity, because a signal having a strong harmonicity can be considered, with a small margin of error, as being a voice signal, therefore corresponding to a presence of speech.

In a second example, the voice activity detector 30 operates in response to the signal produced by a camera, installed for example in the passenger compartment of a motor vehicle and oriented so that its angle of view encompasses in all circumstances the head of the vehicle. driver, considered to be the close speaker. The signal delivered by the camera is analyzed to determine from the movement of the mouth and lips whether the speaker speaks or not, as described inter alia in the EP 2 530 672 A1 (Parrot SA) , which can be referred to for further explanation. The advantage of this image analysis technique is to have complementary information completely independent of the acoustic noise environment.

A third example of a sensor that can be used for voice activity detection is a physiological sensor capable of detecting certain vocal vibrations of the speaker that are not or only slightly corrupted by the surrounding noise. Such a sensor may especially consist of an accelerometer or a piezoelectric sensor applied against the cheek or the temple of the speaker. It can be in particular incorporated in the earpad pad of a combined microphone / headset assembly, as described in the EP 2 518 724 A1 (Parrot SA), which can be referred to for more details.

Indeed, when a person makes a voiced sound (that is, a speech component whose production is accompanied by a vibration of the vocal cords), a vibration propagates from the vocal cords to the pharynx and to the bucco-nasal cavity, where it is modulated, amplified and articulated. The mouth, the soft palate, the pharynx, the sinuses and the nasal fossae then serve as a sounding board for this voiced sound and, their wall being elastic, they vibrate in turn and these vibrations are transmitted by internal bone conduction and are perceptible at the cheek and temple.

These vibrations in the cheek and temple have the characteristic of being, by nature, very little corrupted by the surrounding noise. Indeed, in the presence of external noise, even important, the tissues of the cheek and the temple do not vibrate almost, and this regardless of the spectral composition of the outside noise. A physiological sensor which collects these noise-free vocal vibrations gives a signal representative of the presence or absence of voiced sounds emitted by the speaker, thus making it possible to discriminate very clearly the speech phases and the speaker's silence phases.

Implementation variant of the OM-LSA denoising algorithm

As a variant or in addition to the above, the spectral gain G _OMLSA - calculated in the frequency domain for each bin - can be modulated indirectly, by weighting the probability of presence of frequency speech p ( k , l ) by an indicator global time observed at the level of the frame (and no longer a single particular frequency bin).

In this case, if a single noise frame is detected, it can advantageously be considered that each frequency probability of speech should be zero, and the local frequency probability can be weighted by a global datum, this global datum making it possible to make a deduction on the real case encountered at the frame scale (speech / phase phase noise alone), that the only data in the frequency domain does not allow to formulate; in the presence of noise alone, we can reduce ourselves to a uniform denoising, avoiding any musicality of the noise, which will keep its "grain" of origin.

In other words, the probability of presence of speech initially frequency will be weighted by a probability of global presence of speech at the scale of the frame: one will then strive to denoise in a homogeneous way the whole of the frame in a absence of speech (denoise uniformly when speech is absent).

en pondérant la probabilité fréquentielle de présence de parole par une donnée globale p_glob (k) de présence de parole évaluée au niveau de la trame : $$G_{\mathit{OMLSA}}\left(k,l\right)={\left\{G\left(k,l\right)\right\}}^{p\left(k,l\right)\mathrm{.}{p}_{\mathit{glob}}\left(k\right)}\mathrm{.}{G}_{\mathit{min}}^{1-p\left(k,l\right)\mathrm{.}{p}_{\mathit{glob}}\left(k\right)}$$

• G_Oivf_LSA (k,l) étant le gain spectral à appliquer au bin d'indice l de la trame temporelle courante,
• G (k,l) étant un gain de débruitage sous-optimal à appliquer au bin d'indice l,
• p(k,l) étant la probabilité de présence de parole du bin d'indice l de la trame temporelle courante,
• p_glob (k) étant la probabilité globale et seuillée de parole, évaluée au niveau de la trame temporelle courante, et
• G_min étant une valeur nominale paramétrée du gain spectral.

Indeed, as explained above, the presence of speech P _speech ( k ) (calculated as the arithmetic mean of the frequency probabilities of presence of speech) is a rather reliable indicator of the presence of speech on the scale of speech. the frame. One can then consider modifying the conventional expression of the OM-LSA gain calculation, namely: $${\mathrm{BOY\; WUT}}_{\mathit{OMLSA}}\left(k,l\right)={\left\{\mathrm{BOY\; WUT}\left(k,l\right)\right\}}^{p\left(k,l\right)}\mathrm{.}{\mathrm{BOY\; WUT}}_{\mathit{min}}^{1-p\left(k,l\right)}$$

by weighting the frequency probability of presence of speech by a global data item p _glob ( k ) of speech presence evaluated at the level of the frame: $${\mathrm{BOY\; WUT}}_{\mathit{OMLSA}}\left(k,l\right)={\left\{\mathrm{BOY\; WUT}\left(k,l\right)\right\}}^{p\left(k,l\right)\mathrm{.}{p}_{\mathit{glob}}\left(k\right)}\mathrm{.}{\mathrm{BOY\; WUT}}_{\mathit{min}}^{1-p\left(k,l\right)\mathrm{.}{p}_{\mathit{glob}}\left(k\right)}$$

• G _Oi vf _LSA ( k , l ) being the spectral gain to be applied to the bin of index l of the current time frame,
• G (k, l) being a suboptimal denoising gain to be applied to bin of index l,
• p ( k , l ) being the probability of presence of speech of the bin of index l of the current time frame,
• p _glob ( k ) being the global and thresholded probability of speech, evaluated at the level of the current time frame, and
• G _min being a parameterized nominal value of the spectral gain.

$$P_{\mathit{speech}}\left(k\right)=\frac{1}{N}{\displaystyle \sum _l^N}p\left(k,l\right)$$

• P_seuil étant une valeur de seuil de la probabilité globale de parole, et
• N étant le nombre de bins dans le spectre.

The global datum p _glob ( k ) at the time frame can be evaluated in particular as follows: $$p_{\mathit{glob}}\left(k\right)=\frac{1}{P_{\mathit{threshold}}}\mathrm{.}\mathit{max}\left\{P_{\mathit{speech}}\left(k\right);P_{\mathit{threshold}}\right\}$$

$$P_{\mathit{speech}}\left(k\right)=\frac{1}{\mathrm{NOT}}{\displaystyle \underset{l}{\overset{\mathrm{NOT}}{\Sigma }}}p\left(k,l\right)$$

• P _threshold being a _threshold value of the overall probability of speech, and
• N being the number of bins in the spectrum.

$$p_{\mathit{combinée}}\left(k,l\right)=p\left(k,l\right)\mathrm{.}{p}_{\mathit{glob}}\left(k\right)$$

This amounts to replacing in the conventional expression the frequency probability p ( k , l ) by a combined _combined probability p ( k , l ) which integrates a weighting by the global data p _{glo b} (k ), non-frequency, evaluated at the level of the time frame in the presence of speech: $${\mathrm{BOY\; WUT}}_{\mathit{OMLSA}}\left(k,l\right)={\left\{\mathrm{BOY\; WUT}\left(k,l\right)\right\}}^{p_{\mathit{combined}}\left(k,l\right)}\mathrm{.}{\mathrm{BOY\; WUT}}_{\mathit{min}}^{1-p_{\mathit{combined}}\left(k,l\right)}$$

$$p_{\mathit{combined}}\left(k,l\right)=p\left(k,l\right)\mathrm{.}{p}_{\mathit{glob}}\left(k\right)$$

• en présence de parole au niveau de la trame, c'est-à-dire si P_speech (k) > P_seuil , l'expression conventionnelle du calcul du gain OM-LSA reste inchangée ;
• en l'absence de parole au niveau de la trame, c'est-à-dire si P_speech (k) < P _seui/, les probabilités fréquentielles p(k,l) seront en revanche pondérées par la probabilité globale p_glob (k) faible, ce qui aura pour impact d'uniformiser les probabilités en diminuant leurs valeurs ;
• dans le cas asymptotique particulier P_speech (k) = 0, toutes les probabilités seront nulles et le débruitage sera totalement uniforme.

In other words:

• in the presence of speech at the level of the frame, that is to say if P _speech ( k )> P _threshold , the conventional expression of the calculation of the gain OM-LSA remains unchanged;
• in the absence of speech at the level of the frame, that is to say if P _speech ( k ) < P _{threshold /} , the frequency probabilities p ( k , l ) will instead be weighted by the global probability p _glob ( k ) low, which will have the effect of standardizing probabilities by decreasing their values;
• in the particular asymptotic case P _speech ( k ) = 0, all the probabilities will be zero and denoising will be totally uniform.

The evaluation of the global data p _glob ( k ) is schematized on the Figure 2 by the block 32, which receives as input the data P _threshold (parameterizable threshold value) and P _speech ( k , l ) (value itself calculated by the block 28, as described above), and outputs the value p _glob ( k ) which is applied to the input ④ of the block 24.

Here again, a global data item calculated at the level of the frame is used to refine the calculation of the frequency gain of denoising, and this as a function of the case encountered (absence / presence of speech). In particular, the global data makes it possible to estimate the actual situation encountered at the of the frame (speech phase vs. noise phase alone), which the only frequency data would not allow to formulate. And in the presence of noise alone, we can reduce to a uniform denoising, ideal solution because the perceived residual noise will never be musical.

Results obtained by the algorithm of the invention

As has just been explained, the invention is based on the demonstration that the noise denoising / degradation compromise is based on a spectral gain calculation (a function of a minimum gain scalar parameter and a probability of presence of speech) whose model is suboptimal, and proposes a formula involving a temporal modulation of these elements of calculation of the spectral gain, which become a function of relevant temporal descriptors of the noisy speech signal.

The invention is based on the exploitation of a global datum in order to process each frequency band in a more relevant and adapted manner, the denoising hardness being made variable as a function of the presence of speech on a frame (the noise is no longer disconnected when the risk to have a counterpart is weak).

In the conventional OM-LSA algorithm, each frequency band is treated independently, and for a given frequency is not integrated a priori knowledge of other bands. However, a broader analysis that observes the entire frame to calculate a global indicator characteristic of the frame (here, a speech presence indicator capable of discriminating even roughly noise phase alone and speech phase) is a useful means and effective in refining the processing at the frequency band scale.

Specifically, in a conventional OM-LSA algorithm, the denoising gain is generally adjusted to a compromise value, typically of the order of 14 dB.

The implementation of the invention makes it possible to adjust this gain dynamically to a value varying between 8 dB (in the presence of speech) and 17 dB (in the presence of noise alone). The noise reduction is thus much more energetic, and makes the noise virtually imperceptible (and in any case non-musical) in the absence of speech in most situations commonly encountered. And even in the presence of speech, denoising does not change the tone of the voice, whose rendering remains natural.

Claims (8)

1. A method for denoising an audio signal by applying an algorithm with a variable spectral gain, which is function of a speech presence probability, including the following successive steps:

   a) generating (10) successive time frames (y(k)) of the digitized noisy audio signal (y(n)) ;

   b) applying a Fourier transform (12) to the frames generated at step a), so as to produce for each signal time frame a signal spectrum (Y(k, l)) with a plurality of predetermined frequency bands;

   c) in the frequency domain:

   c1) estimating (18), for each frequency band of each current time frame, a speech presence probability (p(k, l)) ;

   c3) calculating (16) a spectral gain (G_omLsA(k,l)), specific to each frequency band of each current time frame, as a function of: i) an estimation of the noise energy in each frequency band, ii) the speech presence probability estimated at step c1), and iii) a scalar value of minimum gain (G_min) representative of a parameter of hardness of the denoising;

   c4) performing (14) a selective noise reduction by applying to each frequency band of the gain calculated at step c3);

   d) applying an inverse Fourier transform (20) to the signal spectrum (X̂(k,l)) consisted of the frequency bands produced at step c4), so as to deliver for each spectrum a time frame of denoised signal; and

   e) reconstructing (22) a denoised audio signal from the time frames delivered at step d),

   the method being characterized in that:

   - said scalar value of minimum gain (G_min) is a value (G_min(k)) that is dynamically modulatable at each successive time frame (y(k)) ; and

   - the method further includes, previously to step c3) of calculating the spectral gain, a step of:

   c2) calculating (24), for the current time frame (y(k)), said modulatable value (G_min(k)), as a function of a global variable (SNR_y(k)); (P_speech(k)); (VAD(k)) observed at the current time frame for all the frequency bands; and

   - said calculation of step c2) comprises apply, for the current time frame, an increment/decrement (ΔG_min(k); Δ ₁ G_min ,Δ₂ G_min; ΔG_min) added to a parameterized nominal value (G_min) of said minimum gain.
2. The method of claim 1, wherein said global variable is a signal-to-noise ratio (SNR_y(k)) of the current time frame, evaluated (26) in the time domain.
3. The method of claim 2, wherein the scalar value of minimum gain is calculated at step c2) by application of the relation: $${\mathit{G}}_{\mathit{min}}\left(\mathit{k}\right)=G_{\mathit{min}}+\mathrm{\Delta }{G}_{\mathit{min}}\left({\mathit{SNR}}_{\mathit{y}}\left(\mathit{k}\right)\right)$$

   

   k being the index of the current time frame,

   G_min(k) being the minimum gain to be applied to the current time frame,

   G_min being said parameterized nominal value of the minimum gain,

   ΔG_min(k) being said increment/decrement added to G_min , and

   SNR_y(k) being the signal-to-noise ratio of the current time frame.
4. The method of claim 1, wherein said global variable is an average speech probability (P_speech(k)), evaluated (28) at the current time frame.
5. The method of claim 4, wherein the scalar value of minimum gain is calculated at step c2) by application of the relation: $${\mathit{G}}_{\mathit{min}}\left(\mathit{k}\right)=G_{\mathit{min}}+\left({\mathit{P}}_{\mathit{speech}}\left(\mathit{k}\right)-1\right)\mathrm{.}{\mathrm{\Delta }}_1{G}_{\mathit{min}}+{\mathit{P}}_{\mathit{speech}}\left(\mathit{k}\right)\mathrm{.}{\mathrm{\Delta }}_2{G}_{\mathit{min}}$$

   

   k being the index of the current time frame,

   G_min(k) being the minimum gain to be applied to the current time frame,

   G_min being said parameterized nominal value of the minimum gain,

   P_speech(k) being the average speech probability evaluated at the current time frame,

   Δ₁ G_min being said increment/decrement added to G_min in phase of noise, and

   Δ ₂ G_min being said increment/decrement added to G_min in phase of speech.
6. The method of claim 4, wherein the average speech probability is evaluated at the current time frame by application of the relation: $$P_{\mathit{speech}}\left(k\right)=\frac{1}{N}{\displaystyle \sum _l^N}\mathit{p}\left(\mathit{k},\mathit{l}\right)$$

   

   l being the index of the frequency band,

   N being the number of frequency bands in the spectrum, and p(k,l) being the speech presence probability of the frequency band of index l of the current time frame.
7. The method of claim 1, wherein said global variable is a Boolean signal of voice activity detection (VAD(k)) for the current time frame, evaluated (30) in the time domain by analysis of the time frame and/or by means of an external detector.
8. The method of claim 7, wherein the scalar value of minimum gain is calculated at step c2) by application of the relation: $${\mathit{G}}_{\mathit{min}}\left(\mathit{k}\right)=G_{\mathit{min}}+\mathit{VAD}\left(\mathit{k}\right)\mathrm{.}\mathrm{\Delta }{G}_{\mathit{min}}$$

   

   k being the index of the current time frame,

   G_min(k) being the minimum gain to be applied to the current time frame,

   G_min being said parameterized nominal value of the minimum gain,

   VAD(k) being the value of the Boolean signal of voice activity detection for the current time frame, and

   ΔG_min being said increment/decrement added to G_min .

Images