EP2515300A1 - Method and System for noise reduction - Google Patents

Abstract

The method involves converting noisy acoustic signal in time domain into noisy signal in frequency domain, and estimating a noise component in the signal from a square modulus by a minima controlled recursive averaging algorithm. Frequential band is cut into frequential subbands based on mutiband decomposition of the modulus and the component. A square modulus of a noiseless component of each subband is estimated. Noiseless output signal is determined from the estimated modulus using a determination module (36). The noiseless signal is converted into noiseless voice signal in time domain. The conversion of the noiseless signal is carried out by computing sampled output signal for time frames by applying inverse Fourier transform of the noiseless output signal and performing temporal reconstruction of noiseless output voice signal from the sampled signal. An independent claim is also included for a system for reducing noise in noisy acoustic signal obtained from a microphone operating in noisy environment.

Description

The present invention relates to a noise reduction method and an associated noise reduction system.

It relates more particularly to a method and a system for reducing noise on a noisy acoustic signal y (t) from a microphone operating in a noisy environment.

The present invention finds particular application in full-duplex, wireless and mono-sensor audio communication systems, in other words mono-microphone, which make it possible to establish audio communication between several users, in an autonomous manner (ie ie without connection to a transmission base or to a network) and which is easy to use (that is to say not requiring any intervention of a technician to establish the communication).

Such communication systems are generally used in a noisy environment, such as a marine environment or a theater, or extremely noisy, such as a construction site or a hall or stadium hosting a sporting event.

Thus, it is necessary, if not essential, to provide a method and / or system for noise reduction (or speech enhancement) in the signal processing chain in order to improve the audio quality so that the conversation is audible between the speakers. users of communication systems, in other words that the communication is understandable.

It is known to estimate the noise component contained in the noisy signal by an algorithm for estimating the power spectral density of the noise component according to a recursive averaging method of the controlled minima known as "MCRA" for "Minima Controlled Recursive" Averaging ".

• « Speech enhancement for non-stationary noise environments », par I. Cohen et B. Berdugo, Signal Processing, 2001, vol.81 , pp. 2403-2418 ;
• « Noise Estimation by Minima Controlled Recursive Averaging for Robust Speech Enhancement », par I. Cohen et B. Berdugo, IEEE Signal Processing Letters, Janvier 2002, vol. 9, No.1, pp. 12-15 ;

This so-called "MCRA" noise estimation method is known from the scientific literature, including the following articles:

• "Speech enhancement for non-stationary noise environments", by I. Cohen and B. Berdugo, Signal Processing, 2001, vol.81, pp. 2403-2418 ;
• "Noise Estimation by Minima Controlled Recursive Averaging for Robust Speech Enhancement", by I. Cohen and B. Berdugo, IEEE Signal Processing Letters, January 2002, vol. 9, No.1, pp. 12-15 ;

"A Modified Spectral Subtraction Method for Speech Enhancement Based on the Masking Property of Human Auditory System," by X. Bing-Yin et al, Wireless Communications & Signal Processing, International Conference on IEEE, November 13, 2009, pp. 1-5 , XP031594664.

The so-called "MCRA" noise estimation method is thus particularly well suited in environments where the noise signal is highly non-stationary and evolves relatively rapidly over time.

The state of the art can also be illustrated by teaching the article "Multi-band Spectral Subraction for Enhancing Speech Corrupted by Colored Noise", 2002 IEEE International Conference on Acoustics, Speech and Signal Processing. Proceedings (Car NO.02CH37334), vol.4, 2002 . This article discloses a multiband spectral subtraction method.

The present invention aims to provide a method and a noise reduction system adapted to highly noisy environments and where the noise component is highly non-stationary and evolves relatively quickly over time.

1. a) conversion du signal acoustique bruité y(t) dans le domaine temporel en un signal bruité Y_k,l dans le domaine fréquentiel, par découpage temporel du signal acoustique bruité y(t) en signaux échantillonnés y_l dans des trames temporelles I successives, fenêtrage des signaux échantillonnés y_l par application d'une fenêtre de pondération, et application d'une transformée de Fourier discrète, avec extraction du module carré |Y_k,l|², et éventuellement de la phase θ_k,l, du signal bruité Y_k,l ;
2. b) estimation d'une composante de bruit D̂_k,l contenue dans le signal bruité Y_k,l à partir du module carré |Y_k,l|², par un algorithme d'estimation de la densité spectrale de puissance de la composante de bruit selon une méthode de moyennage récursif des minima contrôlés dite « MCRA » ;
   ledit procédé étant remarquable en ce qu'il comprend en outre, après l'étape b), les étapes successives suivantes :
3. c) découpage de la bande fréquentielle en plusieurs sous-bandes fréquentielles SB_i=[e_i, b_i], suivie d'une décomposition multi-bandes du module carré |Y_k,l|² et de la composante de bruit D̂_k,l, consistant à décomposer le module carré |Y_k,l|² et la composante de bruit D̂_k,l en respectivement plusieurs modules carré de sous-bande |Y_k,l,i|² et plusieurs composantes de bruit de sous-bande D̂_k,l,i propres à chacune des sous-bandes SB_i ;
4. d) estimation, pour chacune des sous-bandes SB_i, du module carré |X̂_k,l,i|² d'une composante débruitée de sous-bande X̂_k,l,i propre à chaque sous-bande SB_i d'un signal débruité X̂_k,l, par un algorithme de soustraction spectrale multi-bandes dit « SSMB » à partir des modules carrés de sous-bande |Y_k,l,i|² et des composantes de bruit de sous-bande D̂_k,l,i ;
5. e) détermination d'un signal débruité de sortie SD_k,l à partir des modules carré |X̂_k,l,i|² issus de l'étape d), et éventuellement des phases θ_k,l extraites lors de l'étape a) ;
6. f) conversion du signal débruité de sortie SD_k,l en un signal vocal débruité de sortie sd(t) dans le domaine temporel, par une étape f.1) de calcul d'un signal échantillonné de sortie Sd_l propre à chaque trame temporelle I pair application d'une transformée de Fourier inverse du signal débruité de sortie SD_k,l, suivie d'une étape f.2) de reconstruction temporelle du signal vocal débruité de sortie sd(t) à partir des signaux échantillonnés de sortie Sd_l.

For this purpose, it proposes a method for reducing noise on a noisy acoustic signal y (t) originating from a microphone, preferably a single microphone, operating in a noisy medium, comprising the following successive steps:

1. a) converting the noisy acoustic signal y (t) in the time domain into a noisy signal Y _{k, 1} in the frequency domain, by time division of the noisy acoustic signal y (t) into sampled signals y ₁ in successive time frames I , windowing of sampled signals y _l by applying a weighting window, and applying a discrete Fourier transform, with extraction of the squared modulus | Y _{k, l} | ² , and possibly of the phase θ _{k, 1} , of the noisy signal Y _{k, 1} ;
2. b) estimating a noise component D _{k, 1} contained in the noisy signal Y _{k, 1} from the square module | Y _{k, l} | ² , by an algorithm for estimating the spectral power density of the noise component according to a method of recursive averaging of the controlled minima known as "MCRA";
   said method being remarkable in that it further comprises, after step b), the following successive steps:
3. c) division of the frequency band into several frequency sub-bands SB _i = [e _i , b _i ], followed by a multi-band decomposition of the module square | Y _{k, l} | ² and the noise component D _{k, 1} , consisting in breaking down the square module | Y _{k, l} | ² and the noise component D _{k, l} respectively at several square subband modules | Y _{k, l, i} | ² and a plurality of subband noise components D _{k, l, i} specific to each of the subbands SB _i ;
4. d) estimating, for each of the subbands SB _i , the square module | X _{k, l, i} | ² of a denoised component of subband X _{k, l, i} specific to each subband SB _i of a denoised signal X _{k, 1} , by a multiband spectral subtraction algorithm called "SSMB" from square subband modules | Y _{k, l, i} | ² and subband noise components D _{k, l, i} ;
5. e) determining an output denoised signal SD _{k, l} from the square modules | X _{k, l, i} | ² from step d), and optionally phases θ _{k, l} extracted during step a);
6. f) converting the output denoised signal SD _{k, l} into an output speech output signal sd (t) in the time domain, by a step f.1) of calculating a sampled output signal Sd _l specific to each frame temporal I pair application of an inverse Fourier transform of the output denoised signal SD _{k, l} , followed by a step f.2) of temporal reconstruction of the output unfracked speech signal sd (t) from the output sampled signals Sd _l .

Thus, the invention proposes to implement a multi-band spectral subtraction algorithm "SSMB" which consists of sharing the entire spectral band in sub-bands and to adapt in each subband the subtraction calculation between the noisy signal Y _{k, 1} and the noise component D _{k, 1} for extracting an output denoised signal SD _{k, l} ; applying the spectral subtraction operation in each sub-band to improve the sensitivity of the noise reduction method.

Simulations with this method that combines the "MCRA" and "SSMB" algorithms demonstrate a marked improvement in the audio quality of the processed noisy signal. For a wide range of noises, the result is far superior to conventional methods.

During step a), the extraction of the phase θ _{k, 1} , of the noisy signal Y _{k, 1} is optional. It is indeed possible, during step e), to determine the output denoised signal SD _{k, l} from the square modules | X _{k, l, i} | ² and the noisy signal Y _{k, l} without having to calculate its phase beforehand.

• b.1) calcul d'une composante bruitée filtrée S_k,l répondant à l'équation : $${\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\mathrm{\alpha }}_{\mathrm{s}}{\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{s}}\right){\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}$$
  
  
  où α_s est une constante prédéterminée caractéristique d'un filtre passe-bas ;
• b.2) calcul d'une densité de probabilité de présence de parole p̃_k,l par la mise en oeuvre du calcul progressif suivant :
  1. (i) calcul d'une composante minimale spectrale Smin_k,l avec :
     • si rem(k,1) = 0
       alors Smin_k,l = min (Smin_k,l-1 ; S_k,l) et Stmp_k,l = S_k,l
     • si rem(k,1) ≠ 0
       alors Smin_k,l = min (Stmp_k,l-1 ; S_k,l) et
       Stmp_k,l = min (Stmp_k,l-1 ; S_k,l)
       où rem(k,l) est le reste de la division entière de k par I, puis
  2. (ii) calcul d'un rapport spectral Sr_k,l répondant à l'équation : $${\mathrm{Sr}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\frac{{\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}{{\mathrm{Smin}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}$$
     
  3. (iii) calcul d'une variable indicatrice I_k,l avec :
     • si Sr_k,l > δ_TH alors I_k,l = 1
     • si Sr_k,l ≤ δ_TH alors I_k,l = 0
       où δ_TH est un paramètre prédéterminé de seuil fixe de détection de parole ;
  4. (iv) calcul de la densité de probabilité de présence de parole p̃_k.l avec : $${\tilde{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\mathrm{\alpha }}_{\mathrm{p}}{\tilde{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{p}}\right){\mathrm{I}}_{\mathrm{k}\mathrm{,}\mathrm{l}}$$
     
     
     où α_p est une constante prédéterminée ;
• b.3) calcul d'un coefficient α̃_k,l répondant à l'équation suivante : $${\tilde{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\mathrm{\alpha }\mathrm{+}\left(\mathrm{1}\mathrm{-}\mathrm{\alpha }\right){\tilde{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{,}$$
  
  
  où α est une constante prédéterminée ;
• b.4) calcul de la composante de bruit D̂_k,l répondant à l'équation suivante : $${\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\tilde{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\tilde{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right){\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{.}$$
  

According to one characteristic, the algorithm for estimating the power spectral density of the noise component according to the method of recursive averaging of the controlled minima known as "MCRA" during step b) implements the following calculation phases:

• b.1) calculating a filtered noise component S _{k, l} corresponding to the equation: $${\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\mathrm{\alpha }}_{\mathrm{s}}{\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{s}}\right){\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}$$
  
  
  where α _s is a predetermined constant characteristic of a low-pass filter;
• b.2) calculating a probability density of presence of speech p _{k, 1} by the implementation of the following progressive calculation:
  1. (i) calculating a minimum spectral component Smin _{k, 1} with:
     • if rem (k, 1) = 0
       then Smin _{k, l} = min (Smin _{k, l-1} ; S _{k, l} ) and Stmp _{k, l} = S _{k, l}
     • if rem (k, 1) ≠ 0
       then Smin _{k, l} = min (Stmp _{k, l-1} ; S _{k, l} ) and
       Stmp _{k, l} = min (Stmp _{k, l-1} ; S _{k, l} )
       where rem (k, l) is the remainder of the integer division of k by I, then
  2. (ii) calculating a spectral ratio Sr _{k, l} corresponding to the equation: $${\mathrm{Sr}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\frac{{\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}{{\mathrm{Smin}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}$$
     
  3. (iii) calculating an indicator variable I _{k, l} with:
     • if Sr _{k, l} > δ _TH then I _{k, l} = 1
     • if Sr _{k, l} ≤ δ _TH then I _{k, l} = 0
       where δ _TH is a predetermined parameter of fixed threshold of speech detection;
  4. (iv) calculating the probability density of speech presence p _kl with: $${\stackrel{\mathrm{~}}{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\mathrm{\alpha }}_{\mathrm{p}}{\stackrel{\mathrm{~}}{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{p}}\right){\mathrm{I}}_{\mathrm{k}\mathrm{,}\mathrm{l}}$$
     
     
     where α _p is a predetermined constant;
• b.3) calculating a coefficient α _{k, l} corresponding to the following equation: $${\stackrel{\mathrm{~}}{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\mathrm{\alpha }\mathrm{+}\left(\mathrm{1}\mathrm{-}\mathrm{\alpha }\right){\stackrel{\mathrm{~}}{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{,}$$
  
  
  where α is a predetermined constant;
• b.4) calculating the noise component D _{k, l} corresponding to the following equation: $${\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\stackrel{\mathrm{~}}{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\stackrel{\mathrm{~}}{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right){\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{.}$$
  

The so-called "MCRA" method uses the following principle: on a finite horizon window and for a given frequency, the minimum of the power spectral density (DSP) of the noisy signal Y _{k, l} corresponds to the value of the spectral density of power (DSP) of the noise component D _{k, l} . Thus, an estimation of the different minimum values of the spectrum on a sliding window makes it possible to obtain an estimate of the power spectral density (DSP) of the noise component D _{k, l} .

• d.1) calcul d'un rapport signal à bruit SNR_k,l,i propre à chaque sous-bande SB_i répondant à l'équation : $${\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{=}\mathrm{10.}{\log }_{\mathrm{10}}\left(\frac{\mathrm{\sum }_{\mathrm{k}\mathrm{=}\mathrm{ei}}^{\mathrm{bi}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}}{\mathrm{\sum }_{\mathrm{k}\mathrm{=}\mathrm{ei}}^{\mathrm{bi}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}\right)$$
  
• d.2) calcul du module carré |X̂_k,l,i|² de la composante débruitée de sous-bande X̂_k,l,i propre à chaque sous-bande SB_i, selon l'équation : $${\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{=}\mathrm{\{}\begin{array}{ll}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|& \mathrm{si}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{>}{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|\\ \mathrm{\beta }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}& \mathrm{si}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{\le }{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|\end{array}$$
  
  
  où - δ_i est un paramètre variable en fonction de la sous-bande SB_i correspondante, prenant des valeurs distinctes d'une sous-bande à l'autre ;
  • α_i est un paramètre variable qui dépend de la valeur du rapport signal à bruit SNR_k,l,i calculée dans la sous-bande SB_i correspondante ; et
  • β est une constante.

According to another characteristic, the multi-band spectral subtraction algorithm called "SSMB" of step d) implements the following calculation phases, for each of the sub-bands SB _i :

• d.1) calculating a signal-to-noise ratio SNR _{k, l, i} specific to each sub-band SB _i corresponding to the equation: $${\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{=}\mathrm{10.}{\log }_{\mathrm{10}}\left(\frac{\underset{\mathrm{k}\mathrm{=}\mathrm{ei}}{\overset{\mathrm{bi}}{\mathrm{\Sigma }}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}}{\underset{\mathrm{k}\mathrm{=}\mathrm{ei}}{\overset{\mathrm{bi}}{\mathrm{\Sigma }}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}\right)$$
  
• d.2) calculation of the square module | X _{k, l, i} | ² of the denoised component of sub-band X _{k, l, i} specific to each sub-band SB _i , according to the equation: $${\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{=}\mathrm{\{}\begin{array}{ll}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|& \mathrm{if}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{>}{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|\\ \mathrm{\beta }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}& \mathrm{if}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{\le }{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|\end{array}$$
  
  
  where - δ _i is a variable parameter as a function of the corresponding sub-band SB _i , taking values that are distinct from one subband to another;
  • α _i is a variable parameter which depends on the value of the signal-to-noise ratio SNR _{k, l, i} calculated in the corresponding sub-band SB _i ; and
  • β is a constant.

Such an algorithm is particularly advantageous since it takes into account the value of the signal-to-noise ratio SNR _{k, l, i} specific to each sub-band SB _i to perform the spectral subtraction in each sub-band.

Since the noise component is not distributed uniformly along the frequencies, the implementation of this multi-band spectral subtraction algorithm "SSMB" allows processing with different parameters depending on the sub-band where the 'We are. As a result, the noise component can be reduced more significantly in a sub-band if it is more dominant, unlike a full-band subtraction algorithm that will reduce noise equally all the sub-bands of the spectrum and will therefore be less precise and therefore less effective.

où α_c1, α_c2, α_c3 et α_c4 sont des constantes prédéterminées, et SNR₁ et SNR₂ sont des seuils prédéterminés.In a particular embodiment, the parameters α _i respond to the following equations: $${\mathrm{\alpha }}_{\mathrm{i}}\mathrm{=}\mathrm{\{}\begin{array}{ccc}{\mathrm{\alpha }}_{\mathrm{vs}\mathrm{1}}\hfill & \mathrm{if}& {\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{<}{\mathrm{<figure-callout\; id="1"\; label="SNR"\; filenames="imgf0001.png"\; state="\{\{state\}\}">SNR</figure-callout>}}_{\mathrm{1}}\\ {\mathrm{\alpha }}_{\mathrm{vs}\mathrm{2}}\mathrm{+}{\mathrm{\alpha }}_{\mathrm{vs}3}{\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\hfill & \mathrm{if}& {\mathrm{<figure-callout\; id="1"\; label="SNR"\; filenames="imgf0001.png"\; state="\{\{state\}\}">SNR</figure-callout>}}_{\mathrm{1}}\mathrm{\le }{\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{\le }{\mathrm{<figure-callout\; id="2"\; label="SNR"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">SNR</figure-callout>}}_{\mathrm{2}}\\ {\mathrm{\alpha }}_{\mathrm{vs}\mathrm{4}}\hfill & \mathrm{if}& {\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{>}{\mathrm{<figure-callout\; id="2"\; label="SNR"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">SNR</figure-callout>}}_{\mathrm{2}}\end{array}$$

where α _c1 , α _c2 , α _c3 and α _c4 are predetermined constants, and SNR ₁ and SNR ₂ are predetermined thresholds.

et $${\mathrm{\alpha }}_{\mathrm{c}\mathrm{1}}\mathrm{=}\mathrm{5}\mathrm{,}{\mathrm{\alpha }}_{\mathrm{c}\mathrm{2}}\mathrm{=}\mathrm{4}\mathrm{,}{\mathrm{\alpha }}_{\mathrm{c}\mathrm{3}}\mathrm{=}\mathrm{-}\mathrm{0.15}\mathrm{,}{\mathrm{\alpha }}_{\mathrm{c}\mathrm{4}}\mathrm{=}\mathrm{1}$$

The values α _c1 , α _c2 , α _c3 and α _c4 and SNR ₁ and SNR ₂ are chosen experimentally, in particular by numerical simulation. By way of non-limiting example, the following values have given good results in numerical simulation: $${\mathrm{<figure-callout\; id="1"\; label="SNR"\; filenames="imgf0001.png"\; state="\{\{state\}\}">SNR</figure-callout>}}_{\mathrm{1}}=-5\mathrm{dB},{\mathrm{<figure-callout\; id="2"\; label="SNR"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">SNR</figure-callout>}}_2=20\mathrm{dB};$$

and $${\mathrm{\alpha }}_{\mathrm{vs}\mathrm{1}}\mathrm{=}\mathrm{5}\mathrm{,}{\mathrm{\alpha }}_{\mathrm{vs}\mathrm{2}}\mathrm{=}\mathrm{4}\mathrm{,}{\mathrm{\alpha }}_{\mathrm{vs}\mathrm{3}}\mathrm{=}\mathrm{-}\mathrm{0.15}\mathrm{,}{\mathrm{\alpha }}_{\mathrm{vs}\mathrm{4}}\mathrm{=}\mathrm{1}$$

According to a first possibility of the invention, step e) consists of determining the denoised signal X _{k, 1} from the square modules | X _{k, l, i} | ² debruitées components of sub-band X _{k, l, i} , and optionally phases θ _{k, l} extracted in step a), so that the output unedited signal SD _{k, l} corresponds to the denoised signal X _{k , l} , ie SD _{k, l} = X _{k, l} .

Thus, it is considered that the output unedited signal SD _{k, 1} corresponds to the denoised signal X _{k, l} whose square modules | X _{k, l, i} | ² are directly derived from step d) implementing the multi-band spectral subtraction algorithm "SSMB".

• déterminer, pour chacune des sous-bandes SB_i, le module carré
  
  d'une composante débruitée combinée de sous-bande
  
  propre à chaque sous-bande SBi d'un signal débruité combiné
  
  répondant à l'équation correspondante :
  
  où γ est un coefficient d'amplification prédéterminé, préférentiellement compris entre 0,01 et 0,1 ;
• déterminer un signal débruité combiné
  
  _k,l à partir des modules carrés |
  
  _k,l,i|² des composantes débruitées combinées de sous-bande
  
  _k,l,i, et éventuellement des phases θ_k,l extraites lors de l'étape a), de sorte que le signal débruité de sortie SD_k,l corresponde au signal débruité combiné
  
  _k,l,soit SD_k,l =
  
  _k,l.

According to a second possibility of the invention, as a variant of the first possibility, step e) consists in:

• determining, for each of the subbands SB _i , the square module
  
  a combined denoised subband component
  
  specific to each sub-band SBi of a combined de-noise signal
  
  responding to the corresponding equation:
  
  where γ is a predetermined amplification coefficient, preferably between 0.01 and 0.1;
• determine a combined debris signal
  
  _{k, l} from the square modules |
  
  _{k, l, i} | ² combined debris components of subband
  
  _{k, l, i} , and possibly phases θ _{k, l} extracted during step a), so that the output denoised signal SD _{k, 1} corresponds to the combined de-noise signal
  
  _{k, l} , ie SD _{k, l} =
  
  _{k, l} .

The processed signal, i.e., the denoised signal X _{k, l} , may suffer in terms of the quality and intelligibility of a distortion problem. This distortion, whose origin is usually the noise reduction process, depends on the parameters of noise reduction algorithms ("MCRA" and "SSMB" algorithms) but also on the level and type of noise component to be reduced.

Thus, in the context of this second possibility of the invention, after step d) implementing the multi-band spectral subtraction algorithm "SSMB", a step of reinjection of the noisy signal Y _{k is carried out} , _l from the microphone in the denoised signal X _{k, l} ; this reinjection being controlled by the amplification coefficient (otherwise called attenuation parameter) chosen very low, of the order of a few percent. The reinjected signal corresponds to the signal γ | Y _{k, l, i} | ² and it is not distorted because it has not been processed by the "MCRA" and "SSMB" algorithms.

This principle of reinjection of a very small part of the noisy signal Y _{k, l} from the microphone makes it possible to remedy at least part of this problem of distortion.

In a first embodiment, the step f.2) involves reconstructing the voice signal output denoised sd (t) only from the Sd output signals from _the the step f.1), these output signals Sd _l corresponding to the inverse Fourier transforms of the output denoised signal SD _{k, l} specific to each time frame I.

Thus, in this first embodiment, the speech output speech signal sd (t) is reconstructed from the only output signals Sd _I from step f.1).

• g) calculer un rapport moyen signal à bruit r_l propre à la trame temporelle I à partir du module carré |Y_k,l|² et de la composante de bruit D̂_k,l ;
• h) comparer le rapport moyen signal à bruit r_l avec un seuil Ψ_TH prédéterminé ;
• i) reconstruire le signal vocal débruité de sortie sd(t) à partir en considérant que :
  • si le rapport moyen signal à bruit r_l est inférieur audit seuil Ψ_TH pour la trame temporelle I, alors le signal considéré avant reconstruction temporelle pour cette trame temporelle I correspond au signal de sortie Sd_l issu de l'étape f.1) ;
  • si le rapport moyen signal à bruit r_l est supérieur audit seuil Ψ_TH pour la trame temporelle I, alors le signal considéré avant reconstruction temporelle pour cette trame temporelle I correspond au signal échantillonné y_l issu de l'étape de découpage de l'étape a).

In a second embodiment, as a variant of the first embodiment, step f.2) consists of, for each time frame I:

• g) calculating a signal-to-noise average ratio r _I specific to the time frame I from the square module | Y _{k, l} | ² and the noise component D _{k, l} ;
• h) comparing the average signal-to-noise ratio r ₁ with a predetermined threshold Ψ _TH ;
• i) reconstructing the output speechless signal sd (t) from considering that:
  • if the average signal-to-noise ratio r ₁ is less than said threshold Ψ _TH for the time frame I, then the signal considered before time reconstruction for this time frame I corresponds to the output signal Sd ₁ resulting from the step f.1);
  • if the average signal-to-noise ratio r _l is greater than said threshold Ψ _TH for the time frame I, then the signal considered before time reconstruction for this time frame I corresponds to the sampled signal y ₁ resulting from the step of cutting step at).

• lors de l'étape g), une détection de bruit par calcul d'un rapport moyen signal à bruit r_l ;
• lors de l'étape h), une comparaison de ce rapport moyen signal à bruit r_l avec un seuil Ψ_TH pour établir si le bruit est présent (r_l < Ψ_TH) ou si le bruit est absent ou du moins est extrémement faible (r_l > Ψ_TH) ; et
• lors de l'étape i), on fait en sorte, pour chaque trame I, que :
  • si le bruit est présent (r_l < Ψ_TH), alors on prend en compte le signal traité numériquement, c'est-à-dire le signal de sortie Sd_l pour la reconstruction temporelle ;
  • si le bruit est absent ou extrémement faible (r_l > Ψ_TH), alors on ne prend pas en compte le signal de sortie Sd_l mais on prend en compte directement le signal échantillonné y_l pour la reconstruction temporelle, ce qui revient à ignorer les traitements de réduction du bruit (MCRA, SSMB) pour cette trame I, avec l'avantage d'éviter des distorsions inutiles lorsque le niveau de bruit est tel qu'un traitement de réduction du bruit n'est pas nécessaire.

Thus, in this second embodiment, we implement:

• in step g), a noise detection by calculating a signal-to-noise average ratio r _l ;
• in step h), a comparison of this signal-to-noise ratio r _l with a threshold Ψ _TH to establish if the noise is present (r _l <Ψ _TH ) or if the noise is absent or at least is extremely weak (r _l > Ψ _TH ); and
• during step i), it is made, for each frame I, that:
  • if the noise is present (r _l <Ψ _TH ), then we take into account the digitally processed signal, that is to say the output signal Sd _l for time reconstruction;
  • if the noise is absent or extremely weak (r _l > Ψ _TH ), then we do not take into account the output signal Sd _l but we take into account directly the sampled signal y _l for the time reconstruction, which amounts to ignoring the noise reduction processing (MCRA, SSMB) for this frame I, with the advantage of avoiding unnecessary distortions when the noise level is such that a noise reduction treatment is not necessary.

This second embodiment allows somehow the deactivation of steps b), c) and d) noise reduction when the noise is not present. As a result, the distortions that can be brought about by the noise reduction processing, in this situation of absence or near absence of noise, will be eliminated.

To make the process more robust, it would be advantageous for the same decision (presence or absence of noise) to be taken on a succession of time frames I.

• g.1) calcul d'une composante de bruit moyenne D̅₁ à partir de la composante de bruit D̂_k,l estimée lors de l'étape b) et répondant à l'équation : $${\stackrel{\mathrm{‾}}{\mathrm{D}}}_{\mathrm{1}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\mathrm{\sum }_{\mathrm{k}\mathrm{=}\mathrm{0}}^{\mathrm{M}\mathrm{-}\mathrm{1}}{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}$$
  
  
  où M est une constante prédéterminée, de préférence égale à N ou à [1+N/2], N étant le nombre de points de la transformée de Fourier ;
• g.2) calcul d'un module carré moyen |Y_k,l|² du signal bruité Y_k,l répondant à l'équation : $$\stackrel{\mathrm{‾}}{{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\mathrm{\sum }_{\mathrm{k}\mathrm{=}\mathrm{0}}^{\mathrm{M}\mathrm{-}\mathrm{1}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}$$
  
• g.3) calcul d'une composante filtrée P_l du module carré moyen |Y_k,l|² répondant à l'équation : $${\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="imgf0001.png"\; state="\{\{state\}\}">P</figure-callout>}}_{\mathrm{1}}\mathrm{=}\mathrm{\lambda }{\mathrm{P}}_{\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}\mathrm{\lambda }\right)\stackrel{\mathrm{‾}}{{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}}$$
  
  
  où - λ est une constante prédéterminée caractéristique d'un filtre passe-bas,
  de préférence compris entre 0,80 et 0,99 ;
  $\mathrm{-}{\mathrm{P}}_{\mathrm{0}}\mathrm{=}{\stackrel{\mathrm{‾}}{\mathrm{D}}}_{\mathrm{0}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\mathrm{\sum }_{\mathrm{k}\mathrm{=}\mathrm{0}}^{\mathrm{M}\mathrm{-}\mathrm{1}}{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{0}}$
  
  pour initialiser l'algorithme.
• g.4) calcul du rapport moyen signal à bruit r_l répondant à l'équation :
  • si D̅_l > 0 alors ${\mathrm{r}}_{\mathrm{l}}\mathrm{=}\frac{{\mathrm{P}}_{\mathrm{l}}}{{\stackrel{\mathrm{‾}}{\mathrm{D}}}_{\mathrm{l}}}\mathrm{,}$
    
  • si D̅_l ≤ 0 alors r_l = 0.

In a particular embodiment, step g) implements the following calculation algorithm, for each time frame I:

• g.1) calculating a mean noise component D̅ ₁ from the noise component D _{k, l} estimated in step b) and corresponding to the equation: $${\stackrel{\mathrm{~}}{\mathrm{D}}}_{\mathrm{1}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\underset{\mathrm{k}\mathrm{=}\mathrm{0}}{\overset{\mathrm{M}\mathrm{-}\mathrm{1}}{\mathrm{\Sigma }}}{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}$$
  
  
  where M is a predetermined constant, preferably equal to N or [1 + N / 2], where N is the number of points of the Fourier transform;
• g.2) calculating a mean square module | Y _{k, l} | ² of the noisy signal Y _{k, l} corresponding to the equation: $$\stackrel{\mathrm{~}}{{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\underset{\mathrm{k}\mathrm{=}\mathrm{0}}{\overset{\mathrm{M}\mathrm{-}\mathrm{1}}{\mathrm{\Sigma }}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}$$
  
• g.3) calculating a filtered component P _l of the average square module | Y _{k, l} | ² answering the equation: $${\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="imgf0001.png"\; state="\{\{state\}\}">P</figure-callout>}}_{\mathrm{1}}\mathrm{=}\mathrm{\lambda }{\mathrm{P}}_{\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}\mathrm{\lambda }\right)\stackrel{\mathrm{~}}{{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}}$$
  
  
  where - λ is a predetermined constant characteristic of a low-pass filter,
  preferably between 0.80 and 0.99;
  $\mathrm{-}{\mathrm{P}}_{\mathrm{0}}\mathrm{=}{\stackrel{\mathrm{~}}{\mathrm{D}}}_{\mathrm{0}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\underset{\mathrm{k}\mathrm{=}\mathrm{0}}{\overset{\mathrm{M}\mathrm{-}\mathrm{1}}{\mathrm{\Sigma }}}{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{0}}$
  
  to initialize the algorithm.
• g.4) calculation of the average signal to noise ratio r _l corresponding to the equation:
  • if D̅ _l > 0 then ${\mathrm{r}}_{\mathrm{l}}\mathrm{=}\frac{{\mathrm{P}}_{\mathrm{l}}}{{\stackrel{\mathrm{~}}{\mathrm{D}}}_{\mathrm{l}}}\mathrm{,}$
    
  • if D̅ _l ≤ 0 then r _l = 0.

• pour l'étape a), un découpage du signal acoustique bruité y(t) en trames temporelles avec un recouvrement entre les trames temporelles successives ;
• pour l'étape f.2), la reconstruction du signal vocal débruité de sortie sd(t) est réalisée par les additions successives des parties en recouvrement des signaux de deux trames temporelles successives.

Advantageously, the conversion steps a) and f) implement an OLA recovery and addition method, with:

• for step a), cutting the noisy acoustic signal y (t) into time frames with an overlap between the successive time frames;
• for step f.2), the reconstruction of the output defracked speech signal sd (t) is performed by the successive additions of the parts in overlap of the signals of two successive time frames.

This OLA overlap and add method is a typical time reconstruction method that uses juxtaposed weighting windows (ie overlapping or overlapping windows). then who Adds the output signals taking into account the overlap of the time frames.

• une unité de conversion du signal acoustique bruité y(t) dans le domaine temporel en un signal bruité Y_k,l dans le domaine fréquentiel, comportant :
  • un module de découpage du signal acoustique bruité y(t) en signaux échantillonnés y_l dans des trames temporelles I successives ;
  • en sortie du module de découpage, un module de fenêtrage des signaux échantillonnés y_l par application d'une fenêtre de pondération ;
  • en sortie du module de fenêtrage, un module de calcul d'une transformée de Fourier discrète qui délivre en sortie le signal bruité Y_k,l ;
• une unité de traitement numérique dans le domaine fréquentiel comportant, en sortie de l'unité de conversion :
  • un premier module d'extraction du module carré |Y_k,l|² du signal bruité Y_k,l ; et éventuellement un deuxième module d'extraction de la phase θ_k,l du signal bruité Y_k,l ;
  • en sortie du premier module d'extraction, un module d'estimation, dit « MCRA », d'une composante de bruit D̃_k,l contenue dans le signal bruité Y_k,l à partir du module carré |Y_k,l|² issu du premier module d'extraction, par un algorithme d'estimation de la densité spectrale de puissance de la composante de bruit selon une méthode de moyennage récursif des minima contrôlés dite « MCRA » ;
  • en sortie du premier module d'extraction, un module de découpage de la bande fréquentielle en plusieurs sous-bandes fréquentielles SB_i=[e_i, b_i] ;
  • en sortie du module d'estimation « MCRA » et du module de découpage de la bande fréquentielle, un module d'estimation, dit « SSMB », du module carré |X̂_k,l,i|² d'une composante débruitée de sous-bande X̂_k,l,i propre à chaque sous-bande SB_i d'un signal débruité X̂_k,l, par un algorithme de soustraction spectrale multi-bandes à partir de modules carrés de sous-bande |Y_k,l,i|² et de composantes de bruit de sous-bande D̂_k,l,i ;
  • en sortie du module d'estimation « SSMB », et éventuellement du deuxième module d'extraction, un module de détermination d'un signal débruité de sortie SD_k,l à partir des modules carré |X̂_k,l,i|², et éventuellement des phases θ_k,l ;
• une unité de conversion dans le domaine temporel comportant, en sortie de l'unité de traitement numérique :
  • un module de calcul d'un signal de sortie Sd_l propre à chaque trame temporelle I par application d'une transformée de Fourier inverse du signal débruité de sortie SD_k,l ; et
  • un module de reconstruction d'un signal vocal débruité de sortie sd(t) dans le domaine temporel à partir desdits signaux de sortie Sd_l.

The invention also relates to a system for reducing noise on a noisy acoustic signal y (t) originating from a microphone operating in a noisy medium, comprising:

• a noisy acoustic signal conversion unit y (t) in the time domain into a noisy signal Y _{k, 1} in the frequency domain, comprising:
  • a module for cutting the noisy acoustic signal y (t) into sampled signals y ₁ in successive time frames I;
  • the output of the cutting module, a windowing module of the sampled signals y _l by applying a weighting window;
  • at the output of the windowing module, a module for calculating a discrete Fourier transform which outputs the noisy signal Y _{k, l} ;
• a digital processing unit in the frequency domain comprising, at the output of the conversion unit:
  • a first extraction module of the square module | Y _{k, l} | ² of the noisy signal Y _{k, l} ; and possibly a second extraction module of the phase θ _{k, 1} of the noisy signal Y _{k, 1} ;
  • at the output of the first extraction module, an estimation module, called "MCRA", of a noise component D _{k, 1} contained in the noisy signal Y _{k, 1} from the square module | Y _{k, l} | ² from the first extraction module, by an algorithm for estimating the spectral power density of the noise component according to a recursive averaging method of the controlled minima known as "MCRA";
  • at the output of the first extraction module, a frequency band division module in several frequency subbands SB _i = [e _i , b _i ];
  • at the output of the estimation module "MCRA" and the frequency band division module, an estimation module, called "SSMB", of the square module | X _{k, l, i} | ² of a denoised component of subband X _{k, l, i} specific to each subband SB _i of a denoised signal X _{k, 1} , by a multiband spectral subtraction algorithm from square subband modules | Y _{k, l, i} | ² and subband noise components D _{k, l, i} ;
  • at the output of the estimating module "SSMB", and possibly of the second extraction module, a module for determining an output denoised signal SD _{k, l} from the square modules | X _{k, l, i} | ² , and optionally phases θ _{k, l} ;
• a time domain conversion unit comprising, at the output of the digital processing unit:
  • a module for calculating an output signal Sd ₁ specific to each time frame I by applying an inverse Fourier transform of the output denoised signal SD _{k, l} ; and
  • a module for reconstructing a speech output speech signal sd (t) in the time domain from said output signals Sd _l .

• un sous-module racine-carré pour calculer le module |X̂_k,l,i| des composantes débruitées de sous-bande X̂_k,l,i, et
• un sous-module de recombinaison des composantes débruitées de sous-bande X̂_k,l,i pour obtenir le signal débruité X̂_k,l à partir des modules |X̂_k,l,i|, et éventuellement des phases θ_k,l, de sorte que le signal débruité de sortie SD_k,l corresponde au signal débruité X̂_k,l, soit SD_k,l = X̂_k,l

According to a first possibility of the invention, the module for determining the output denoised signal SD _{k, 1} comprises:

• a root-square submodule to compute the module | X _{k, l, i} | debrueted components of sub-band X _{k, l, i} , and
• a recombination sub-module of the de-banded components of sub-band X _{k, l, i} to obtain the denoised signal X _{k, l} from the modules | X _{k, l, i} |, and possibly phases θ _{k, l} , so that the output unedited signal SD _{k, 1} corresponds to the denoised signal X _{k, 1} , ie SD _{k, l} = X _{k, l}

• en sortie du premier module d'extraction, un sous-module d'amplification selon un coefficient d'amplification γ, préférentiellement compris entre 0,01 et 0,1, afin de délivrer un signal amplifié γ |Y_k,l|² ;
• en sortie du module d'estimation « SSMB », un sous-module additionneur propre à additionner le signal amplifié γ |Y_k,l|² et les modules carré |X̂_k,l,i|², afin de délivrer en sortie les module carré |
  
  _k,l,i|² de composantes débruitées combinées de sous-bande
  
  _k,l,i propres à chaque sous-bande SBi d'un signal débruité combiné
  
  _k,l, répondant à l'équation correspondante : $${\left|{\mathrm{X}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{=}{\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{+}\mathrm{\gamma }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}};$$
  
• un sous-module racine-carré pour calculer le module |
  
  _k,l,i| des composantes débruitées combinées de sous-bande
  
  _k,l,i des et
• un sous-module de recombinaison des composantes combinées de sous-bande
  
  _k,l,i pour obtenir le signal débruité combiné
  
  _k,l à partir des modules |
  
  _k,l,i|, et éventuellement des phases θ_k,l, de sorte que le signal débruité de sortie SD_k,l correspond au signal débruité combiné
  
  _k,l, soit SD_k,l =
  
  _k,l,

According to a second possibility of the invention, as a variant of the first possibility, the module for determining an output denoised signal SD _{k, 1} comprises:

• at the output of the first extraction module, an amplification sub-module according to an amplification coefficient γ, preferably between 0.01 and 0.1, in order to deliver an amplified signal γ | Y _{k, l} | ² ;
• at the output of the estimating module "SSMB", an addering sub-module capable of adding the amplified signal γ | Y _{k, l} | ² and the square modules | X _{k, l, i} | ² , in order to output the square module |
  
  _{k, l, i} | ² combined debrueted components of sub-band
  
  _{k, l, i} specific to each sub-band SBi of a combined de-noise signal
  
  _{k, l} , corresponding to the equation: $${\left|{\mathrm{X}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{=}{\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{+}\mathrm{\gamma }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}};$$
  
• a root-square submodule to compute the module |
  
  _{k, l, i} | combined debrueted subband components
  
  _{k, l, i} and
• a recombination sub-module of the combined subband components
  
  _{k, l, i} to obtain the combined noise signal
  
  _{k, l} from the modules |
  
  _{k, l, i} |, and possibly phases θ _{k, 1} , so that the output unedited signal SD _{k, 1} corresponds to the combined de-noise signal
  
  _{k, l} , ie SD _{k, l} =
  
  _{k, l} ,

In this second possibility, the system implements a step of reinjection of a very small part of the noisy signal Y _{k, 1} coming from the microphone into the denoised signal X _{k, 1} to remedy at least in part the distortion problems induced by the estimation modules "MCRA" and "SSMB".

• un module de calcul d'un rapport moyen signal à bruit r_l propre à chaque trame temporelle I à partir du module carré |Y_k,l|² et de la composante de bruit D̂_k,l ;
• un module de comparaison du rapport moyen signal à bruit r_l propre à chaque trame temporelle I avec un seuil Ψ_TH prédéterminé ;
• un module de contrôle du module de reconstruction du signal vocal débruité de sortie sd(t) qui est conçu pour que :
  • si le rapport moyen signal à bruit r_l est inférieur audit seuil Ψ_TH pour la trame temporelle I, alors le signal considéré avant reconstruction pour cette trame temporelle I correspond au signal de sortie Sd_l issu du module de calcul dudit signal de sortie Sd_l ;
  • si le rapport moyen signal à bruit r_l est supérieur audit seuil Ψ_TH pour la trame temporelle I, alors le signal considéré avant reconstruction pour cette trame temporelle I correspond au signal échantillonné y_l issu du module de découpage du signal acoustique bruité y(t).

In an advantageous embodiment, the digital processing unit further comprises, at the output of the estimation module "MCRA", a noise detection module comprising

• a module for calculating a signal-to-noise average ratio r ₁ specific to each time frame I from the square module | Y _{k, l} | ² and the noise component D _{k, l} ;
• a module for comparing the average signal-to-noise ratio r ₁ for each time frame I with a predetermined threshold Ψ _TH ;
• a module for controlling the output defracked speech signal reconstruction module sd (t) which is designed so that:
  • if the average signal-to-noise ratio r ₁ is less than said threshold Ψ _TH for the time frame I, then the signal considered before reconstruction for this time frame I corresponds to the output signal Sd ₁ from the calculation module of said output signal Sd _l ;
  • if the average signal-to-noise ratio r _l is greater than said threshold Ψ _TH for the time frame I, then the signal considered before reconstruction for this time frame I corresponds to the sampled signal y ₁ from the noisy acoustic signal cutting module y (t ).

• la figure 1 est une vue schématique d'un premier système de réduction du bruit conforme à l'invention ;
• la figure 2 est une vue schématique d'un second système de réduction du bruit conforme à l'invention ;
• la figure 3 est une vue schématique d'une variante du second système de la figure 2.

Other features and advantages of the present invention will appear on reading the detailed description below, of an example of non-limiting implementation, with reference to the appended figures in which:

• the figure 1 is a schematic view of a first noise reduction system according to the invention;
• the figure 2 is a schematic view of a second noise reduction system according to the invention;
• the figure 3 is a schematic view of a variant of the second system of the figure 2 .

The description of a system 1 for reducing the noise on a noisy acoustic signal y (t) originating from a single microphone operating in a noisy medium, and the method of reducing the associated noise, is made with reference to the Figures 1 to 3 .

In the three embodiments, the system 1 comprises a conversion unit 2 of the noisy acoustic signal y (t) in the time domain into a noisy signal Y _{k, 1} in the frequency domain.

This conversion unit 2 comprises a switching module 21 of the noisy acoustic signal y (t) in sampled signals y ₁ in successive time frames I.

In this clipping module 21, the noisy acoustic signal y (t) is cut into frames of 240 samples, which at a sampling frequency of 8 kHz corresponds to time frames of 30 milliseconds.

It is also conceivable to have a frame division of 256 samples, which at a sampling frequency of 8 kHz corresponds to time frames of 32 milliseconds.

In this clipping module 21, the successive time frames overlap or overlap. For example, successive time frames overlap on 120 samples, which corresponds to fifty percent (50%) of overlap.

This overlap of the time frames is intended to allow the implementation of a method of recovery and addition called "OLA", which allows the initial temporal division of the noisy acoustic signal y (t) then the final restitution at the output of the system 1 in the time domain.

This conversion unit 2 comprises, at the output of the chopper module 21, a windowing module 22 of the sampled signals y ₁ by application of a weighting window, in particular of the Hanning window or Hamming window type, in order to output weighted sampled signals {y ₁ }.

Thus, the time frames are then apodized with a weighting window, before applying a Fourier transform, in order to to minimize the edge effects due to the blanking cut performed by the cutting module 21.

This conversion unit 2 comprises, at the output of the windowing module 22, a calculation module 23 of a discrete Fourier transform which outputs the noisy signal Y _{k, l} .

avec x(t) le signal de parole utile et d(t) la composante de bruit.From a mathematical point of view, we note: $$\mathrm{there}\left(\mathrm{t}\right)\mathrm{=}\mathrm{x}\left(\mathrm{t}\right)\mathrm{+}\mathrm{d}\left(\mathrm{t}\right)\mathrm{,}$$

with x (t) the useful speech signal and d (t) the noise component.

The switching module 21 receives as input the noisy acoustic signal y (t) and outputs the sampled signal y ₁ , where I is the time index (or index of the time frame).

avec ω_l, le signal représentatif de la fenêtre de pondération.Cutting module 21 receives as input the sampled signal y _l and outputs the weighted sampled signal {y _t}, with: $$\left\{{\mathrm{there}}_{\mathrm{l}}\right\}\mathrm{=}{\mathrm{\omega }}_{\mathrm{l}}\mathrm{.}{\mathrm{there}}_{\mathrm{l}}$$

with ω _l , the signal representative of the weighting window.

The calculation module 23 receives as input the weighted sampled signal {y ₁ } and outputs the noisy signal Y _{k, 1} which corresponds to the Discrete Fourier transform of {y (1)}, where k represents the index of Frequency: $${\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\mathrm{DFT}\left(\left\{{\mathrm{there}}_{\mathrm{l}}\right\}\right)\mathrm{=}{\mathrm{r}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{+}\mathrm{j}{\mathrm{i}}_{\mathrm{k}\mathrm{,}\mathrm{l}}$$

The calculation of the Discrete Fourier Transform (DFT) is for example carried out by a Fast Fourier Transform (FFT) calculation with a size N which can be equal to 256 (N corresponds to the number of points of the Fourier transform).

The system 1 comprises, at the output of the conversion unit 2, a digital processing unit 3 in the frequency domain which carries out denoising or speech enhancement processing on the noisy signal Y _{k, l} .

This digital processing unit 3 comprises a first extraction module 31 of the square module | Y _{k, l} | ² of the noisy signal Y _{k, l} .

From a mathematical point of view, the first extraction module 31 performs the following calculation: $${\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{=}{\mathrm{r}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}^{\mathrm{2}}\mathrm{+}{\mathrm{i}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}^{\mathrm{2}}\mathrm{.}$$

In the embodiments of figures 1 and 2 this digital processing unit 3 comprises a second extraction module 32 of the phase θ _{k, 1} of the noisy signal Y _{k, l} . As detailed later, it is also conceivable to dispense with this second extraction module 32.

avec $$\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|\mathrm{=}\sqrt{{\mathrm{r}}_{\mathrm{k}\mathrm{,}\mathrm{l}}^{\mathrm{2}}\mathrm{+}{\mathrm{i}}_{\mathrm{k}\mathrm{,}\mathrm{l}}^{\mathrm{2}}}$$

From a mathematical point of view, the second extraction module 32 performs the following calculation: $$\cos {\mathrm{\theta }}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\frac{{\mathrm{r}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|}\mathrm{,}\mathrm{and\; sin}{\mathrm{\theta }}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\frac{{\mathrm{i}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|}$$

with $$\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|\mathrm{=}\sqrt{{\mathrm{r}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}^{\mathrm{2}}\mathrm{+}{\mathrm{i}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}^{\mathrm{2}}}$$

The digital processing unit 3 comprises, at the output of the first extraction module 31, an estimation module 33, called "MCRA", of a noise component D _{k, 1} contained in the noisy signal Y _{k, l} from the square module | Y _{k, l} | ² from the first extraction module 31, by an algorithm for estimating the power spectral density of the noise component according to a recursive averaging method of the controlled minima known as "MCRA".

• b.1) calcul d'une composante bruitée filtrée S_k,l répondant à l'équation : $${\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\mathrm{\alpha }}_{\mathrm{s}}{\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{s}}\right){\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}$$
  
  
  où α_s est une constante prédéterminée caractéristique d'un filtre passe-bas ;
• b.2) calcul d'une densité de probabilité de présence de parole p̃_k,l par la mise en oeuvre du calcul progressif suivant :
  1. (i) calcul d'une composante minimale spectrale Smin_k,l avec :
     • si rem(k,1) = 0
     alors Smin_k,l = min (Smin_k,l-1 ; S_k,l) et Stmp_k,l = S_k,l
     • si rem(k,l) ≠ 0
     alors Smin_k,l = min (Stmp_k,l-1 ; S_k,l) et
     Stmp_k,l = min (Stmp_k,l-1 ; S_k,l)
     où rem(k,l) est le reste de la division entière de k par I, puis
  2. (ii) calcul d'un rapport spectral Sr_k,l répondant à l'équation : $${\mathrm{Sr}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\frac{{\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}{{\mathrm{Smin}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}$$
     
  3. (iii) calcul d'une variable indicatrice I_k,l avec :
     • si Sr_k,l > δ_TH alors I_k,l = 1
     • si Sr_k,l ≤ δ_TH alors I_k,l = 0
       où δ_TH est un paramètre prédéterminé de seuil fixe de détection de parole ;
  4. (iv) calcul de la densité de probabilité de présence de parole p̃_k,l avec : $${\tilde{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\mathrm{\alpha }}_{\mathrm{p}}{\tilde{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{p}}\right){\mathrm{I}}_{\mathrm{k}\mathrm{,}\mathrm{l}}$$
     
     
     où α_p est une constante prédéterminée ;
• b.3) calcul d'un coefficient α̃_k,l répondant à l'équation suivante : $${\tilde{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\mathrm{\alpha }\mathrm{+}\left(\mathrm{1}\mathrm{-}\mathrm{\alpha }\right){\tilde{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{,}$$
  
  
  où α est une constante prédéterminée ;
• b.4) calcul de la composante de bruit D̂_k,l répondant à l'équation suivante : $${\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\tilde{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\tilde{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right){\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{.}$$
  

From a mathematical point of view, the algorithm for estimating the power spectral density of the noise component according to the recursive averaging method of the controlled minima known as "MCRA" implements the following calculation phases:

• b.1) calculating a filtered noise component S _{k, l} corresponding to the equation: $${\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\mathrm{\alpha }}_{\mathrm{s}}{\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{s}}\right){\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}$$
  
  
  where α _s is a predetermined constant characteristic of a low-pass filter;
• b.2) calculating a probability density of presence of speech p _{k, 1} by the implementation of the following progressive calculation:
  1. (i) calculating a minimum spectral component Smin _{k, 1} with:
     • if rem (k, 1) = 0
     then Smin _{k, l} = min (Smin _{k, l-1} ; S _{k, l} ) and Stmp _{k, l} = S _{k, l}
     • if rem (k, l) ≠ 0
     then Smin _{k, l} = min (Stmp _{k, l-1} ; S _{k, l} ) and
     Stmp _{k, l} = min (Stmp _{k, l-1} ; S _{k, l} )
     where rem (k, l) is the remainder of the integer division of k by I, then
  2. (ii) calculating a spectral ratio Sr _{k, l} corresponding to the equation: $${\mathrm{Sr}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\frac{{\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}{{\mathrm{Smin}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}$$
     
  3. (iii) calculating an indicator variable I _{k, l} with:
     • if Sr _{k, l} > δ _TH then I _{k, l} = 1
     • if Sr _{k, l} ≤ δ _TH then I _{k, l} = 0
       where δ _TH is a predetermined parameter of fixed threshold of speech detection;
  4. (iv) calculating the probability density of speech presence p _{k, 1} with: $${\stackrel{\mathrm{~}}{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\mathrm{\alpha }}_{\mathrm{p}}{\stackrel{\mathrm{~}}{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{p}}\right){\mathrm{I}}_{\mathrm{k}\mathrm{,}\mathrm{l}}$$
     
     
     where α _p is a predetermined constant;
• b.3) calculating a coefficient α _{k, l} corresponding to the following equation: $${\stackrel{\mathrm{~}}{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\mathrm{\alpha }\mathrm{+}\left(\mathrm{1}\mathrm{-}\mathrm{\alpha }\right){\stackrel{\mathrm{~}}{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{,}$$
  
  
  where α is a predetermined constant;
• b.4) calculating the noise component D _{k, l} corresponding to the following equation: $${\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\stackrel{\mathrm{~}}{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\stackrel{\mathrm{~}}{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right){\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{.}$$
  

The estimation module "MCRA" 33 thus outputs the noise component D _{k, l} .

The digital processing unit 3 comprises, at the output of the first extraction module 31, a switching module 34 of the frequency band in several frequency subbands SB _i = [e _i , b _i ].

The frequency band can for example be divided into three frequency sub-bands, namely SB ₁ for f _i <1000 Hz, SB ₂ for 1000 Hz ≤ f _i ≤ 2000 Hz and finally SB ₃ for f _i > 2000 Hz, where f _i is the subband frequency.

The digital processing unit 3 comprises, at the output of the estimation module "MCRA" 33 and the module of division 34 of the frequency band, an estimation module 35, called "SSMB", of the square module | X _{k, l, i} | ² of a denoised component of sub-band X _k , _l , _i specific to each sub-band SB _i of a denoised signal X _{k, l} , by a multi-band spectral subtraction algorithm from square modules of sub-band | Y _{k, l, i} | ² and subband noise components D _{k, l, i}

avec µ_i un coefficient prédéterminé.The principle of the multi-band spectral subtraction algorithm (SSMB) proceeds from a generalization of the spectral subtraction algorithm which consists in subtracting from the spectral power density of the noisy signal Y _{k, l} coming from the microphone a portion of the spectral power density of the noise component estimated by the "MCRA" method. The multi-band spectral subtraction algorithm applies the same principle by splitting the spectral space into several frequency sub-bands SB _i and then, in each sub-band SB _i , the spectral subtraction operation is applied, as follows : $${\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{=}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}-{\mathrm{\mu }}_{\mathrm{i}}{\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}$$

with μ _i a predetermined coefficient.

This multi-band spectral subtraction algorithm is combined with the so-called OLA overlay and addition method.

It is of course conceivable to refine the multi-band spectral subtraction relationship given above, as described below.

• d.1) calcul d'un rapport signal à bruit SNR_k,l,i propre à chaque sous-bande SB_i répondant à l'équation : $${\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{=}\mathrm{10.}{\log }_{\mathrm{10}}\left(\frac{\mathrm{\sum }_{\mathrm{k}\mathrm{=}\mathrm{ei}}^{\mathrm{bi}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}}{\mathrm{\sum }_{\mathrm{k}\mathrm{=}\mathrm{ei}}^{\mathrm{bi}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}\right)$$
  
• d.2) calcul du module carré |X̂_k,l,i|² de la composante débruitée de sous-bande X̂_k,l,i propre à chaque sous-bande SB_i, selon l'équation : $${\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{=}\mathrm{\{}\begin{array}{ll}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|& \mathrm{si}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{>}{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|\\ \mathrm{\beta }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}& \mathrm{si}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{\le }{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|\end{array}$$
  
  
  où - δ_i est un paramètre variable en fonction de la sous-bande SB_i
  correspondante, prenant des valeurs distinctes d'une sous-bande à l'autre ;
  • α_i est un paramètre variable qui dépend de la valeur du rapport signal à bruit SNR_k,l,i calculée dans la sous-bande SB_i correspondante ; et
  • β est une constante.

From a mathematical point of view, the multi-band spectral subtraction algorithm "SSMB" implemented by the estimation module "SSMB" 35, performs the following calculation phases, for each of the sub-bands SB _i :

• d.1) calculating a signal-to-noise ratio SNR _{k, l, i} specific to each sub-band SB _i corresponding to the equation: $${\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{=}\mathrm{10.}{\log }_{\mathrm{10}}\left(\frac{\underset{\mathrm{k}\mathrm{=}\mathrm{ei}}{\overset{\mathrm{bi}}{\mathrm{\Sigma }}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}}{\underset{\mathrm{k}\mathrm{=}\mathrm{ei}}{\overset{\mathrm{bi}}{\mathrm{\Sigma }}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}\right)$$
  
• d.2) calculation of the square module | X _{k, l, i} | ² of the denoised component of sub-band X _{k, l, i} specific to each sub-band SB _i , according to the equation: $${\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{=}\mathrm{\{}\begin{array}{ll}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|& \mathrm{if}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{>}{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|\\ \mathrm{\beta }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}& \mathrm{if}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{\le }{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|\end{array}$$
  
  
  where - δ _i is a variable parameter depending on the sub-band SB _i
  corresponding, taking values distinct from one subband to another;
  • α _i is a variable parameter which depends on the value of the signal-to-noise ratio SNR _{k, l, i} calculated in the corresponding sub-band SB _i ; and
  • β is a constant.

où α_c1, α_c2, α_c3 et α_c4 sont des constantes prédéterminées, et SNR₁ et SNR₂ sont les seuils prédéterminés.It is conceivable to compare the value of the signal-to-noise ratio SNR _{k, l, i} with two thresholds SNR ₁ and SNR ₂ to establish the value of the parameters α _i . Thus, the parameters α _i respond to the following equations: $${\mathrm{\alpha }}_{\mathrm{i}}\mathrm{=}\mathrm{\{}\begin{array}{ccc}{\mathrm{\alpha }}_{\mathrm{vs}\mathrm{1}}\hfill & \mathrm{if}& {\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{<}{\mathrm{<figure-callout\; id="1"\; label="SNR"\; filenames="imgf0001.png"\; state="\{\{state\}\}">SNR</figure-callout>}}_{\mathrm{1}}\\ {\mathrm{\alpha }}_{\mathrm{vs}\mathrm{2}}\mathrm{+}{\mathrm{\alpha }}_{\mathrm{vs}3}{\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\hfill & \mathrm{if}& {\mathrm{<figure-callout\; id="1"\; label="SNR"\; filenames="imgf0001.png"\; state="\{\{state\}\}">SNR</figure-callout>}}_{\mathrm{1}}\mathrm{\le }{\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{\le }{\mathrm{<figure-callout\; id="2"\; label="SNR"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">SNR</figure-callout>}}_{\mathrm{2}}\\ {\mathrm{\alpha }}_{\mathrm{vs}\mathrm{4}}\hfill & \mathrm{if}& {\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{>}{\mathrm{<figure-callout\; id="2"\; label="SNR"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">SNR</figure-callout>}}_{\mathrm{2}}\end{array}$$

where α _c1 , α _c2 , α _c3 and α _c4 are predetermined constants, and SNR ₁ and SNR ₂ are the predetermined thresholds.

For example, it is conceivable to have the following relations for the parameters α _i : $${\mathrm{\alpha }}_{\mathrm{i}}\mathrm{=}\mathrm{\{}\begin{array}{ccc}5& \mathrm{if}& {\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{<}-5\\ 4-\frac{3}{20}{\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\hfill & \mathrm{if}& -5\mathrm{\le }{\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{\le }20\\ 1& \mathrm{if}& {\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{>}20\end{array}$$

Concerning the variable parameters δ _i , it is conceivable to have the following relations, in the case of the division into three frequency subbands described above: $${\mathrm{\delta }}_{\mathrm{i}}\mathrm{=}\mathrm{\{}\begin{array}{ccc}1& \mathrm{if}& \hfill {\mathrm{f}}_{\mathrm{i}}<1000\left(\mathrm{first\; sub}-\mathrm{bandaged}{\mathrm{<figure-callout\; id="1"\; label="SB"\; filenames="imgf0001.png"\; state="\{\{state\}\}">SB</figure-callout>}}_1\right)\\ 2.75\hfill & \mathrm{if}& \hfill 1000\le {\mathrm{f}}_{\mathrm{i}}\le 2000\left(\mathrm{second\; sub}-\mathrm{bandaged}{\mathrm{<figure-callout\; id="2"\; label="SB"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">SB</figure-callout>}}_2\right)\\ 1.75& \mathrm{if}& \hfill {\mathrm{f}}_{\mathrm{i}}\mathrm{>}2000\left(\mathrm{third\; sub}-\mathrm{bandaged}{\mathrm{<figure-callout\; id="3"\; label="SB"\; filenames="imgf0001.png,imgf0003.png"\; state="\{\{state\}\}">SB</figure-callout>}}_3\right)\end{array}$$

Regarding the constant β, it is possible to have values of the order of 0.002 or 0.0015, for example β = 0.002 or β = 0.0015.

The digital processing unit 3 also comprises, at the output of the estimation module "SSMB" 35 and the second extraction module 32, a module 36 for determining an output denoised signal SD _{k, l} from the modules square | X _{k, l, i} | ² and phases θ _{k, l} .

• un sous-module racine-carré 361 pour calculer le module |X̂_k,l,i| des composantes débruitées de sous-bande X̂_k,l,i, et
• un sous-module de recombinaison 362 des composantes débruitées de sous-bande X̂_k,l,i pour obtenir le signal débruité X̂_k,l à partir des modules |X̂_k,l,i| et des phases θ_k,l, de sorte que le signal débruité de sortie SD_k,l corresponde au signal débruité X̂_k,l, soit SD_k,l = X̂_k,l.

In the first embodiment illustrated on the figure 1 , the determination module 36 of the output denoised signal SD _{k, 1} comprises:

• a root-square sub-module 361 for calculating the module | X _{k, l, i} | debrueted components of sub-band X _{k, l, i} , and
• a recombination sub-module 362 of the subband de-waving components X _{k, l, i} to obtain the denoised signal X _{k, l} from the modules | X _{k, l, i} | and phases θ _{k, 1} , such that the output unedited signal SD _{k, 1} corresponds to the denoised signal X _{k, 1} , ie SD _{k, l} = X _{k, l} .

• le sous-module racine-carré 361 réalise le calcul : $\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|\mathrm{=}\sqrt{{\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}}\mathrm{;}$
  
  et
• le sous-module de recombinaison 362 effectue la réinjection de la phase, comme suit : SD_k,l = X̂_k,l = |X̂_k,l| cos θ_k,l + j |X̂_k,l| sin θ_k,l.

From a mathematical point of view, we have:

• the root-square sub-module 361 performs the calculation: $\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|\mathrm{=}\sqrt{{\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}}\mathrm{;}$
  
  and
• the recombination submodule 362 performs the reinjection of the phase, as follows: SD _{k, l} = X _{k, l} = | X _{k, l} | cos θ _{k, l} + j | X _{k, l} | sin θ _{k, l} .

• en sortie du premier module d'extraction 31, un sous-module d'amplification 363 selon un coefficient d'amplification γ, préférentiellement compris entre 0,01 et 0,1, afin de délivrer un signal amplifié γ |Y_k,l|² ;
• en sortie du module d'estimation « SSMB » 35, un sous-module additionneur 364 propre à additionner le signal amplifié γ |Y_k,l|² et les modules carré |X̂_k,l,i|², afin de délivrer en sortie les module carré |X̂_k,l,i|2 de composantes débruitées combinées de sous-bande X̂_k,l,i propres à chaque sous-bande SBi d'un signal débruité combiné X̂_k,l, répondant à l'équation correspondante : $${\left|{\mathrm{X}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{=}{\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{+}\mathrm{\gamma }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}};$$
  
• un sous-module racine-carré 361 pour calculer le module |X̂_k,l,i| des composantes débruitées combinées de sous-bande X̂_k,l,i ; et
• un sous-module de recombinaison 362 des composantes combinées de sous-bande X̂_k,l,i pour obtenir le signal débruité combiné X̂_k,l, à partir des modules |X̂_k,l,i| et des phases θ_k,l, de sorte que le signal débruité de sortie SD_k,l correspond au signal débruité combiné X̂_k,l, soit SD_{k,l =} X̂_k,l.

In the second embodiment illustrated on the figure 2 , the determination module 36 of the output denoised signal SD _{k, 1} comprises:

• at the output of the first extraction module 31, an amplification sub-module 363 according to an amplification coefficient γ, preferably between 0.01 and 0.1, in order to deliver an amplified signal γ | Y _{k, l} | ² ;
• at the output of the estimation module "SSMB" 35, an adder sub-module 364 capable of adding the amplified signal γ | Y _{k, l} | ² and the square modules | X _{k, l, i} | ² , in order to output the square modules | x _{k, l, i} | 2 of combined de-banded components of sub-band X _{k, l, i} specific to each sub-band SBi of a combined de-signaled signal X _{k, l} , responding to the corresponding equation: $${\left|{\mathrm{X}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{=}{\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{+}\mathrm{\gamma }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}};$$
  
• a root-square sub-module 361 for calculating the module | X _{k, l, i} | combined de-banded components of sub-band X _{k, l, i} ; and
• a recombination sub-module 362 of the combined sub-band components X _{k, l, i} to obtain the combined de-signaled signal X _{k, l} , from the modules | X _{k, l, i} | and phases θ _{k, 1} , so that the output unedited signal SD _{k, 1} corresponds to the combined de-noise signal X _{k, 1} , ie SD _{k, l =} X _{k, l} .

• le sous-module d'amplification 363 délivre en sortie le signal amplifié $$\mathrm{\gamma }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}};$$
  
• le sous-module additionneur 364 réalise le calcul : $${\left|{\mathrm{X}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{=}{\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{+}\mathrm{\gamma }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}$$
  
• le sous-module racine-carré 361 réalise le calcul :
  
  et
• le sous-module de recombinaison 362 effectue la réinjection de la phase, comme suit : SD_{k,l =} X̂_k,l = |X̂_k,l| cos θ_k,l + j |X̂_k,l| sin θ_k,l

From a mathematical point of view, we have:

• the amplification sub-module 363 outputs the amplified signal $$\mathrm{\gamma }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}};$$
  
• the adder submodule 364 performs the calculation: $${\left|{\mathrm{X}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{=}{\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}\mathrm{+}\mathrm{\gamma }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">i</figure-callout>}}\right|}^{\mathrm{2}}$$
  
• the root-square sub-module 361 performs the calculation:
  
  and
• the recombination submodule 362 performs the reinjection of the phase, as follows: SD _{k, l =} X _{k, l} = | X _{k, l} | cos θ _{k, l} + j | X _{k, l} | sin θ _{k, l}

Thus, in this second embodiment, the system 1 implements a step of reinjection of a very small part of the noisy signal Y _k , ₁ , coming from the microphone, into the denoised signal X _{k, l} ; the reinjected signal corresponding to the amplified signal γ | Y _{k, l} | ² . In this way, at least some of the distortion problems introduced by the "MCRA" 33 and "SSMB" estimation modules 35 are remedied.

The system 1 comprises, at the output of the digital processing unit 3, a conversion unit 4 in the time domain.

This conversion unit 4 comprises a calculation module 41 of an output signal Sd ₁ specific to each time frame I by the application of an inverse Fourier transform of the output denoised signal SD _{k, l} .

où IDFT correspond à la fonction de transformée de Fourier discrète inverse, qui peut être du type transformée de Fourier rapide inverse (IFFT).From a mathematical point of view, this calculation module 41 implements the calculation of the inverse fast Fourier Transform (IFFT) with a size N equal to 256 (N corresponding to the number of points of the Fourier transform), with the following relation: $${\mathrm{<figure-callout\; id="1"\; label="sd"\; filenames="imgf0001.png"\; state="\{\{state\}\}">sd</figure-callout>}}_{\mathrm{1}}\mathrm{=}\mathrm{IDFT}\left({\mathrm{SD}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right)$$

where IDFT corresponds to the inverse discrete Fourier transform function, which can be of the inverse fast Fourier transform (IFFT) type.

In addition, due to the symmetry of the Fourier transform (DFT or FFT) amplitude of the real signals, the noise reduction and reconstruction processing according to the "OLA" method will be done only on the first [N / 2 + 1] first sampling points, ie on the first 129 sampling points for N equal to 256, knowing that we have the following Hermitian symmetry relation: $${\mathrm{SD}}_{\mathrm{-}\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\mathrm{SD}}_{\mathrm{k}\mathrm{,}\mathrm{l}}^{\mathrm{*}}$$

In the case of the first embodiment illustrated on the figure 1 we have the following relation: Sd _l = x _l = IFFT (X _{k, l} ).

In the case of the second embodiment illustrated on the figure 2 we have the following relation: Sd _l = x _l = IFFT (X _{k, l} ).

This conversion unit 4 comprises, at the output of the calculation module 41, a reconstruction module 42 of a speech output signal sd (t) in the time domain from the output signals Sd ₁ .

qui se traduit dans le premier mode de réalisation par : $\left\{{\widetilde{\mathrm{sd}}}_{\mathrm{l}}\right\}\mathrm{=}\left\{{\hat{\mathrm{x}}}_{\mathrm{l}\mathrm{-}\mathrm{1}}\right\}+\left\{{\hat{\mathrm{x}}}_{\mathrm{l}}\right\}\mathrm{,}$

et qui se traduit dans le second mode de réalisation par :

The temporal signal is restored according to the "OLA" overlay and addition method, the reconstruction of the output defracked speech signal sd (t) being performed by the successive additions of the parts in recovery of the signals of two successive time frames, according to the principle: $$\left\{{\stackrel{\mathrm{~}}{\mathrm{sd}}}_{\mathrm{l}}\right\}\mathrm{=}\left\{{\mathrm{sd}}_{\mathrm{l}\mathrm{-}\mathrm{1}}\right\}\mathrm{+}\left\{{\mathrm{sd}}_{\mathrm{l}}\right\}\mathrm{,}$$

which is translated in the first embodiment by: $\left\{{\stackrel{\mathrm{~}}{\mathrm{sd}}}_{\mathrm{l}}\right\}\mathrm{=}\left\{{\hat{\mathrm{x}}}_{\mathrm{l}\mathrm{-}\mathrm{1}}\right\}+\left\{{\hat{\mathrm{x}}}_{\mathrm{l}}\right\}\mathrm{,}$

and which is translated in the second embodiment by:

Whenever a time frame of the output denoised signal SD _k , _l is delivered in the frequency domain and its inverse Fourier transform Sd _l is computed, the first N / 2 sampling points will be added together with the last N / 2 sampling points of the previous processed frame. The last N / 2 sampling points of the current processed frame will be stored in memory for use in processing the next frame.

In other words, after processing in the spectral domain, the time frame of the output denoised signal SD _{k, 1} passes through the calculation module 41 of the inverse Fourier transform, then its first half (N / 2 first sampling points). is added with the second half (N / 2 last sampling points) saved from the previous frame, while its second half (N / 2 first sampling points) is saved for the next block.

With N equal to 256, we have N / 2 which is equal to 128, being reminded that the recovery rate of two successive frames is fifty percent (50%).

• mise en oeuvre d'un calcul d'un rapport moyen signal à bruit r_l pour effectuer une détection de bruit ;
• comparaison de ce rapport moyen signal à bruit r_l avec un seuil ψ_TH pour établir si le bruit est présent (r_l < ψ_TH) ou si le bruit est absent ou du moins extrémement faible (r_l > ψ_TH) ; et
• pilotage du module de reconstruction 42 selon les règles suivantes :
• si le bruit est présent (r_l < ψ_TH), alors on prend en compte le signal traité numériquement, c'est-à-dire le signal de sortie Sd_l pour la reconstruction temporelle ;
• si le bruit est absent ou extrémement faible (r_l > ψ_TH), alors on ne prend pas en compte le signal de sortie Sd_l mais on prend en compte directement le signal échantillonné y_l pour la reconstruction temporelle, ce qui revient à ignorer les traitements de réduction du bruit (« MCRA », « SSMB ») pour cette trame I, avec l'avantage d'éviter des distorsions inutiles lorsque le niveau de bruit est tel qu'un traitement de réduction du bruit n'est pas nécessaire.

In an optional and advantageous manner, the digital processing unit 3 further comprises, at the output of the estimation module "MCRA" 33, a noise detection module 37 which drives the reconstruction module 42 according to the following principle:

• implementing a calculation of a signal-to-noise average ratio r ₁ for performing noise detection;
• comparing this signal to noise ratio r ₁ with a threshold ψ _TH to establish whether the noise is present (r _l <ψ _TH ) or if the noise is absent or at least extremely low (r _l > ψ _TH ); and
• driving the reconstruction module 42 according to the following rules:
• if the noise is present (r _l <ψ _TH ), then we take into account the digitally processed signal, that is to say the output signal Sd _l for time reconstruction;
• if the noise is absent or extremely weak (r _l > ψ _TH ), then we do not take into account the output signal Sd _l but we take into account directly the sampled signal y _l for the time reconstruction, which amounts to ignoring noise reduction processing ("MCRA", "SSMB") for this frame I, with the advantage of avoiding unnecessary distortions when the noise level is such that a noise reduction treatment is not necessary .

Therefore, in a situation of absence or near absence of noise, the distortions that can be made by the noise reduction treatment will be eliminated.

• un module de calcul d'un rapport moyen signal à bruit r_l propre à chaque trame temporelle 1 à partir du module carré |Y_k,l|² et de la composante de bruit D_k,l;
• un module de comparaison du rapport moyen signal à bruit r_l propre à chaque trame temporelle I avec un seuil ψ_TH prédéterminé ;
• un module de contrôle du module de reconstruction 42 du signal vocal débruité de sortie sd(t) qui est conçu pour que :
  • si le rapport moyen signal à bruit r_l est inférieur audit seuil ψ_TH pour la trame temporelle I, alors le signal considéré avant reconstruction pour cette trame temporelle I correspond au signal de sortie Sd_l issu du module de calcul dudit signal de sortie Sd_l ;
  • si le rapport moyen signal à bruit r_l est supérieur audit seuil ψ_TH pour la trame temporelle I, alors le signal considéré avant reconstruction pour cette trame temporelle I correspond au signal échantillonné y_l issu du module de découpage du signal acoustique bruité y(t).

From a structural point of view, the noise detection module 37 comprises:

• a module for calculating a signal-to-noise average ratio r ₁ specific to each time frame 1 from the square module | Y _{k, l} | ² and the noise component D _{k, l} ;
• a module for comparing the average signal-to-noise ratio r ₁ for each time frame I with a predetermined threshold ψ _TH ;
• a module of control of the reconstruction module 42 of the output speech signal noiseless output sd (t) which is designed so that:
  • if the average signal-to-noise ratio r ₁ is less than said threshold ψ _TH for the time frame I, then the signal considered before reconstruction for this time frame I corresponds to the output signal Sd ₁ from the calculation module of said output signal Sd _l ;
  • if the average signal-to-noise ratio r _l is greater than said threshold ψ _TH for the time frame I, then the signal considered before reconstruction for this time frame I corresponds to the sampled signal y ₁ from the noisy acoustic signal cutting module y (t ).

• g.1) calcul d'une composante de bruit moyenne D_l à partir de la composante de bruit D̂_k,l estimée par le module d'estimation « MCRA » 33 et répondant à l'équation : $${\stackrel{\mathrm{‾}}{\mathrm{D}}}_{\mathrm{1}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\mathrm{\sum }_{\mathrm{k}\mathrm{=}\mathrm{0}}^{\mathrm{M}\mathrm{-}\mathrm{1}}{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}$$
  
  
  où M est une constante prédéterminée égale à N/2, N étant pour rappel le nombre de points de la transformée de Fourier ;
• g.2) calcul d'un module carré moyen |Y_k,l|² du signal bruité Y_k,l répondant à l'équation : $$\stackrel{\mathrm{‾}}{{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\mathrm{\sum }_{\mathrm{k}\mathrm{=}\mathrm{0}}^{\mathrm{M}\mathrm{-}\mathrm{1}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}$$
  
• g.3) calcul d'une composante filtrée P_l du module carré moyen |Y_k,l|² répondant à l'équation : $${\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="imgf0001.png"\; state="\{\{state\}\}">P</figure-callout>}}_{\mathrm{1}}\mathrm{=}\mathrm{\lambda }{\mathrm{P}}_{\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}\mathrm{\lambda }\right)\stackrel{\mathrm{‾}}{{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}}$$
  
  
  où - λ est une constante prédéterminée caractéristique d'un filtre passe-bas, de préférence compris entre 0,80 et 0,99 ;
  • ${\mathrm{P}}_{\mathrm{0}}\mathrm{=}{\stackrel{\mathrm{‾}}{\mathrm{D}}}_{\mathrm{0}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\mathrm{\sum }_{\mathrm{k}\mathrm{=}\mathrm{0}}^{\mathrm{M}\mathrm{-}\mathrm{1}}$
    
    D̂_k,0 pour initialiser l'algorithme.
• g.4) calcul du rapport moyen signal à bruit r_l répondant à l'équation :
  • si D _l > 0 alors ${\mathrm{r}}_{\mathrm{l}}\mathrm{=}\frac{{\mathrm{P}}_{\mathrm{l}}}{{\stackrel{\mathrm{‾}}{\mathrm{D}}}_{\mathrm{l}}}\mathrm{,}$
    
  • si D _l < 0 alors r_l = 0.

From a mathematical point of view, the noise detection module 37 implements the following calculation algorithm, for each time frame 1:

• g.1) calculation of an average noise component D _l from the noise component D _{k, l} estimated by the estimation module "MCRA" 33 and corresponding to the equation: $${\stackrel{\mathrm{~}}{\mathrm{D}}}_{\mathrm{1}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\underset{\mathrm{k}\mathrm{=}\mathrm{0}}{\overset{\mathrm{M}\mathrm{-}\mathrm{1}}{\mathrm{\Sigma }}}{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}$$
  
  
  where M is a predetermined constant equal to N / 2, N being for reference the number of points of the Fourier transform;
• g.2) calculating a mean square module | Y _{k, l} | ² of the noisy signal Y _{k, l} corresponding to the equation: $$\stackrel{\mathrm{~}}{{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\underset{\mathrm{k}\mathrm{=}\mathrm{0}}{\overset{\mathrm{M}\mathrm{-}\mathrm{1}}{\mathrm{\Sigma }}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}$$
  
• g.3) calculating a filtered component P _l of the average square module | Y _{k, l} | ² answering the equation: $${\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="imgf0001.png"\; state="\{\{state\}\}">P</figure-callout>}}_{\mathrm{1}}\mathrm{=}\mathrm{\lambda }{\mathrm{P}}_{\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}\mathrm{\lambda }\right)\stackrel{\mathrm{~}}{{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">l</figure-callout>}}\right|}^{\mathrm{2}}}$$
  
  
  where - λ is a predetermined constant characteristic of a low-pass filter, preferably between 0.80 and 0.99;
  • ${\mathrm{P}}_{\mathrm{0}}\mathrm{=}{\stackrel{\mathrm{~}}{\mathrm{D}}}_{\mathrm{0}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\underset{\mathrm{k}\mathrm{=}\mathrm{0}}{\overset{\mathrm{M}\mathrm{-}\mathrm{1}}{\mathrm{\Sigma }}}$
    
    D _{k, 0} to initialize the algorithm.
• g.4) calculation of the average signal to noise ratio r _l corresponding to the equation:
  • if D _l > 0 then ${\mathrm{r}}_{\mathrm{l}}\mathrm{=}\frac{{\mathrm{P}}_{\mathrm{l}}}{{\stackrel{\mathrm{~}}{\mathrm{D}}}_{\mathrm{l}}}\mathrm{,}$
    
  • if D _l <0 then r _l = 0.

Of course the implementation example mentioned above is not limiting and other improvements and details can be made to the reduction system according to the invention, without departing from the scope of the invention.

Thus, it is conceivable to dispense with the extraction of the phase of the phase θ _{k, 1,} from the noisy signal Y _{k, l} as illustrated in FIG. figure 3 where the system 1 does not include the second extraction module 32 of the phase θ _{k, 1} of the noisy signal Y _{k, l} . In this case, the recombination sub-module 362 of the de-banded components of sub-band X _{k, l, i} calculates the denoised signal X _{k, l} from the modules | X _{k, l, i} | and the noisy signal Y _{k, l} , Indeed, the reinjection of the phase can be carried out directly from the noisy signal Y _{k, l} which intrinsically integrates this phase.

soit X̂_k,l = G_k,l Y_k,l.For this purpose, the recombination sub-module 362 performs the following calculation: $${\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|\frac{{\mathrm{r}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|}\mathrm{+}\mathrm{j}\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|\frac{{\mathrm{i}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|}\mathrm{=}{\mathrm{BOY\; WUT}}_{\mathrm{k}\mathrm{,}\mathrm{l}}{\mathrm{r}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{+}\mathrm{j}{\mathrm{BOY\; WUT}}_{\mathrm{k}\mathrm{,}\mathrm{l}}{\mathrm{i}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{,}$$

let X _{k, l} = G _{k, l} Y _{k, l} .

With G _{k, l} the gain of the noise reduction algorithm. With this calculation, it is no longer necessary to calculate, store and reinject the phase.

System 1 of the figure 3 is a variant of the system 1 of the figure 2 without the second extraction module 32, but it is of course also conceivable to provide for the removal of the second extraction module 32 in the system 1 of the figure 1 .

Claims (14)

A method of reducing noise on a noisy acoustic signal y (t) from a microphone operating in a noisy medium, comprising the following successive steps: a) converting the noisy acoustic signal y (t) in the time domain into a noisy signal Y _{k, 1} in the frequency domain, by time division of the noisy acoustic signal y (t) into sampled signals y ₁ in successive time frames I , windowing of sampled signals y _l by applying a weighting window, and applying a discrete Fourier transform, with extraction of the squared modulus | Y _{k, l} | ² , and possibly of the phase θ _{k, 1} , of the noisy signal Y _{k, 1} ; b) estimating a noise component D _{k, 1} contained in the noisy signal Y _{k, 1} from the square module | Y _{k, l} | ² , by an algorithm for estimating the spectral power density of the noise component according to a method of recursive averaging of the controlled minima known as "MCRA";
characterized in that it further comprises, after step b), the following successive steps: c) division of the frequency band into several frequency sub-bands SB _i = [e _i , b _i ], followed by a multi-band decomposition of the square module | Y _{k, l} | ² and the noise component D _{k, 1} , consisting in breaking down the square module | Y _{k, l} | ² and the noise component D _{k, l} respectively at several square subband modules | Y _{k, l, i} | ² and a plurality of subband noise components D _{k, l, i} specific to each of the subbands SB _i ; d) estimating, for each of the subbands SB _i , the square module | X _{k, l, i} | ² of a denoised component of subband X _{k, l, i} specific to each subband SB _i of a denoised signal X _{k, 1} , by a multiband spectral subtraction algorithm called "SSMB" from square subband modules | Y _{k, l, i} | ² and subband noise components D _{k, l, i} ; e) determining an output denoised signal SD _{k, l} from the square modules | X _{k, l, i} | ² from step d), and optionally phases θ _{k, l} extracted during step a); f) converting the output denoised signal SD _{k, l} into an output speech output signal sd (t) in the time domain, by a step f.1) of calculating a sampled output signal Sd _l specific to each frame 1 by applying an inverse Fourier transform of the output denoised signal SD _{k, l} , followed by a step f.2) of temporally reconstructing the output defracked speech signal sd (t) from the output sampled signals Sd _l . The method according to claim 1, wherein the algorithm for estimating the power spectral density of the noise component according to the method of recursive averaging of the controlled minima known as "MCRA" in step b) implements the phases following calculations: b.1) calculating a filtered noise component S _{k, l} corresponding to the equation: $${\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\mathrm{\alpha }}_{\mathrm{s}}{\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{s}}\right){\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|}^{\mathrm{2}}$$

where α _s is a predetermined constant characteristic of a low-pass filter; b.2) calculating a probability density of presence of speech p _{k, 1} by the implementation of the following progressive calculation: (i) calculating a minimum spectral component Smin _{k, 1} with: - if rem (k, l) = 0
then Smin _{k, l} = min (Smin _{k, l-1} ; S _{k, l} ) and Stmp _{k, l} = S _{k, l} - if rem (k, l) ≠ 0
then Smin _{k, l} = min (Stmp _{k, l-1} ; S _{k, l} ) and
Stmp _{k, l} = min (Stmp _{k, l-1} ; S _{k, l} )
where rem (k, l) is the remainder of the integer division of k by 1, then (ii) calculating a spectral ratio Sr _{k, l} corresponding to the equation: $${\mathrm{Sr}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\frac{{\mathrm{S}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}{{\mathrm{Smin}}_{\mathrm{k}\mathrm{,}\mathrm{l}}}$$

(iii) calculating an indicator variable I _{k, l} with: if Sr _{k, l} > δ _TH then I _{k, l} = 1 if Sr _{k, l} ≤ δ _TH then I _{k, l} = 0
where δ _TH is a predetermined parameter of fixed threshold of speech detection; (iv) calculating the probability density of speech presence p _{k, 1} with: $${\stackrel{\mathrm{~}}{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\mathrm{\alpha }}_{\mathrm{p}}{\stackrel{\mathrm{~}}{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{p}}\right){\mathrm{I}}_{\mathrm{k}\mathrm{,}\mathrm{l}}$$

where α _p is a predetermined constant; b.3) calculating a coefficient α _{k, l} corresponding to the following equation: $${\stackrel{\mathrm{~}}{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}\mathrm{\alpha }\mathrm{+}\left(\mathrm{1}\mathrm{-}\mathrm{\alpha }\right){\stackrel{\mathrm{~}}{\mathrm{p}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{,}$$

where α is a predetermined constant; b.4) calculating the noise component D _{k, l} corresponding to the following equation: $${\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\mathrm{=}{\stackrel{\mathrm{~}}{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}{\stackrel{\mathrm{~}}{\mathrm{\alpha }}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right){\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|}^{\mathrm{2}}\mathrm{.}$$

Method according to any one of claims 1 and 2, wherein the multi-band spectral subtraction algorithm called "SSMB" of step d) implements the following calculation phases, for each of the sub-bands SB _i : d.1) calculating a signal-to-noise ratio SNR _{k, l, i} specific to each sub-band SB _i corresponding to the equation: $${\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{=}\mathrm{10.}{\log }_{\mathrm{10}}\left(\frac{\underset{\mathrm{k}\mathrm{=}\mathrm{ei}}{\overset{\mathrm{bi}}{\mathrm{\Sigma }}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}}{\underset{\mathrm{k}\mathrm{=}\mathrm{ei}}{\overset{\mathrm{bi}}{\mathrm{\Sigma }}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}\right)$$

d.2) calculation of the square module | X _{k, l, i} | ² of the denoised component of sub-band X _{k, l, i} specific to each sub-band SB _i , according to the equation: $${\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}\mathrm{=}\mathrm{\{}\begin{array}{ll}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}\mathrm{-}{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|& \mathrm{if}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}\mathrm{>}{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|\\ \mathrm{\beta }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}& \mathrm{if}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}\mathrm{\le }{\mathrm{\alpha }}_{\mathrm{i}}{\mathrm{\delta }}_{\mathrm{i}}\left|{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|\end{array}$$

where - δ _i is a variable parameter depending on the sub-band SB _i
corresponding, taking values distinct from one subband to another; α _i is a variable parameter which depends on the value of the signal-to-noise ratio SNR _{k, l, i} calculated in the corresponding sub-band SB _i ; and - β is a constant. The method of claim 3, wherein the parameters α _i respond to the following equations: $${\mathrm{\alpha }}_{\mathrm{i}}\mathrm{=}\mathrm{\{}\begin{array}{ccc}{\mathrm{\alpha }}_{\mathrm{vs}\mathrm{1}}\hfill & \mathrm{if}& {\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{<}{\mathrm{SNR}}_{\mathrm{1}}\\ {\mathrm{\alpha }}_{\mathrm{vs}\mathrm{2}}\mathrm{+}{\mathrm{\alpha }}_{\mathrm{vs}3}{\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\hfill & \mathrm{if}& {\mathrm{SNR}}_{\mathrm{1}}\mathrm{\le }{\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{\le }{\mathrm{SNR}}_{\mathrm{2}}\\ {\mathrm{\alpha }}_{\mathrm{vs}\mathrm{4}}\hfill & \mathrm{if}& {\mathrm{SNR}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\mathrm{>}{\mathrm{SNR}}_{\mathrm{2}}\end{array}$$

where α _c1 , α _c2 , α _c3 and α _c4 are predetermined constants, and SNR ₁ and SNR ₂ are predetermined thresholds. A method as claimed in any one of claims 1 to 4, wherein step e) comprises determining the denoised signal X _{k, 1} from the square modules | X _{k, l, i} | ² debruitées components of sub-band X _{k, l, i} , and optionally phases θ _{k, l} extracted in step a), so that the output unedited signal SD _{k, l} corresponds to the denoised signal X _{k , l} , ie SD _{k, l} = X _{k, l} . The method of any one of claims 1 to 4 wherein step e) comprises: determining, for each of the sub-bands SB _i , the square module | X _{k, l, i} | ² of a combined denoised component of sub-band X _{k, l, i} specific to each sub-band SBi of a combined de-noise signal X _{k, l} , corresponding to the corresponding equation: $${\left|{\mathrm{X}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}\mathrm{=}{\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}\mathrm{+}\mathrm{\gamma }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}$$

where γ is a predetermined amplification coefficient, preferably between 0.01 and 0.1; determining a combined noise signal X _{k, l} , from the square modules | x _{k, l, i} | ² of the combined de-current components of sub-band X _{k, l, i} and possibly phases θ _{k, l} extracted during step a), so that the output de-let signal SD _{k, 1} corresponds to the combined de-noise signal X _{k, l} , ie SD _{k, l} = X _{k, l} . A method as claimed in any one of claims 1 to 6, wherein step f.2) is to reconstruct the output speech output signal sd (t) only from the output signals Sd ₁ from step f. 1), said output signals Sd ₁ corresponding to the inverse Fourier transforms of the output de-let signal SD _{k, 1} belonging to each time frame I. The method of any one of claims 1 to 6, wherein step f.2) comprises, for each time frame I: g) calculating a signal-to-noise average ratio r _I specific to the time frame I from the square module | Y _{k, l} | ² and the noise component D _{k, l} ; h) comparing the average signal-to-noise ratio r ₁ with a predetermined threshold ψ _TH ; i) reconstructing the output speechless signal sd (t) from considering that: if the average signal-to-noise ratio r ₁ is less than said threshold ψ _TH for the time frame I, then the signal considered before time reconstruction for this time frame I corresponds to the output signal Sd ₁ resulting from the step f.1) ; if the average signal-to-noise ratio r ₁ is greater than said threshold ψ _TH for the time frame I, then the signal considered before time reconstruction for this time frame I corresponds to the sampled signal y ₁ resulting from the step of splitting the step a). The method of claim 8, wherein step g) implements the following calculation algorithm, for each time frame I: g.1) calculation of an average noise component D _l from the noise component D _{k, l} estimated in step b) and corresponding to the equation: $${\stackrel{\mathrm{~}}{\mathrm{D}}}_{\mathrm{1}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\underset{\mathrm{k}\mathrm{=}\mathrm{0}}{\overset{\mathrm{M}\mathrm{-}\mathrm{1}}{\mathrm{\Sigma }}}{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{l}}$$

where M is a predetermined constant, preferably N or N / 2, where N is the number of sample points of the Fourier transform; g.2) calculating a mean square module | Y _{k, l} | ² of the noisy signal Y _{k, l} corresponding to the equation: $$\stackrel{\mathrm{~}}{{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|}^{\mathrm{2}}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\underset{\mathrm{k}\mathrm{=}\mathrm{0}}{\overset{\mathrm{M}\mathrm{-}\mathrm{1}}{\mathrm{\Sigma }}}{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|}^{\mathrm{2}}$$

g.3) calculating a filtered component P _l of the average square module | Y _{k, l} | ² answering the equation: $${\mathrm{P}}_{\mathrm{1}}\mathrm{=}\mathrm{\lambda }{\mathrm{P}}_{\mathrm{l}\mathrm{-}\mathrm{1}}\mathrm{+}\left(\mathrm{1}\mathrm{-}\mathrm{\lambda }\right)\stackrel{\mathrm{~}}{{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}}\right|}^{\mathrm{2}}}$$

where - λ is a predetermined constant characteristic of a low-pass filter, preferably between 0.80 and 0.99; - ${\mathrm{P}}_{\mathrm{0}}\mathrm{=}{\stackrel{\mathrm{~}}{\mathrm{D}}}_{\mathrm{0}}\mathrm{=}\frac{\mathrm{1}}{\mathrm{M}}\underset{\mathrm{k}\mathrm{=}\mathrm{0}}{\overset{\mathrm{M}\mathrm{-}\mathrm{1}}{\mathrm{\Sigma }}}{\hat{\mathrm{D}}}_{\mathrm{k}\mathrm{,}\mathrm{0}}$

to initialize the algorithm. g.4) calculation of the average signal to noise ratio r _l corresponding to the equation: - if D _l > 0 then ${\mathrm{r}}_{\mathrm{l}}\mathrm{=}\frac{{\mathrm{P}}_{\mathrm{l}}}{{\stackrel{\mathrm{~}}{\mathrm{D}}}_{\mathrm{l}}}\mathrm{,}$

- if D _l <0 then r _l = 0. Method according to any one of the preceding claims, wherein the conversion steps a) and f) implement an OLA recovery and addition method, with: for step a), cutting the noisy acoustic signal y (t) into time frames with an overlap between the successive time frames; for step f.2), the reconstruction of the output defracked speech signal sd (t) is performed by the successive additions of the parts in overlap of the signals of two successive time frames. System (1) for reducing noise on a noisy acoustic signal y (t) from a microphone operating in a noisy environment, comprising: a conversion unit (2) of the noisy acoustic signal y (t) in the time domain into a noisy signal Y _{k, 1} in the frequency domain, comprising: a switching module (21) of the noisy acoustic signal y (t) in sampled signals y ₁ in successive time frames I; - at the output of the chopper module (21), a windowing module (22) sampled signals y _l by applying a weighting window; - at the output of the windowing module (22), a calculation module (23) of a discrete Fourier transform which outputs the noisy signal Y _{k, l} ; - a digital processing unit (3) in the frequency domain comprising, at the output of the conversion unit: a first extraction module (31) of the square module | Y _{k, l} | ² of the noisy signal Y _{k, l} ; and possibly a second extraction module (32) of the phase θ _{k, 1} of the noisy signal Y _{k, 1} ; at the output of the first extraction module (31), an estimation module (33), called "MCRA", of a noise component D _{k, 1} contained in the noisy signal Y _{k, 1} from the module square | Y _{k, l} | ² from the first extraction module (31), by an algorithm for estimating the spectral power density of the noise component according to a recursive averaging method of the controlled minima known as "MCRA"; at the output of the first extraction module (31), a module for cutting (34) the frequency band into a plurality of frequency sub-bands SB _i = [e _i , b _i ], in particular of the filter bank type; at the output of the estimation module "MCRA" (33) and the module for cutting (34) the frequency band, an estimation module (35), called "SSMB", of the square module | X _{k, 1, i} | ² of a denoised component of sub-band X _{k, l, i} specific to each sub-band SB _i of a denoised signal X _{k, 1} , by a multi-band spectral subtraction algorithm from square modules of sub -band | Y _{k, l, i} | ² and subband noise components D _{k, l, i} ; at the output of the estimation module "SSMB" (35), and possibly of the second extraction module (32), a module for determining (36) a denoised output signal SD _{k, 1} from the square modules | X _{k, l, i} | ² , and optionally phases θ _{k, l} ; a conversion unit (4) in the time domain comprising, at the output of the digital processing unit (3): a calculation module (41) for an output signal Sd ₁ specific to each time frame I by applying an inverse Fourier transform of the output denoised signal SD _{k, l} ; and a module for reconstructing (42) an output speech squelch signal sd (t) in the time domain from said output signals Sd ₁ . System (1) according to claim 11, wherein the determination module (36) of the output denoised signal SD _{k, 1} comprises: a root-square submodule (361) for calculating the module | X _{k, l, i} | debrueted components of sub-band X _{k, l, i} , and a recombination sub-module (362) of the de-banded components of sub-band X _{k, l, i} to obtain the denoised signal X _{k, l} from the modules | X _{k, l, i} |, and possibly phases θ _{k, l} , so that the output denoised signal SD _{k, 1} corresponds to the denoised signal X _{k, 1} , ie SD _{k, l} = X _{k, l} . System (1) according to claim 11, wherein the determination module (36) of an output denoised signal SD _{k, l} comprises: at the output of the first extraction module (31), an amplification sub-module (363) according to an amplification coefficient γ, preferably between 0.01 and 0.1, in order to deliver an amplified signal γ | Y _{k, l} | ² ; at the output of the estimation module "SSMB" (35), an adder sub-module (364) able to add the amplified signal γ | Y _{k, l} | ² and the square modules | X _{k, l, i} | ² , in order to output the square modules | X _{k, l, i} | ² of combined debranched components of sub-band X _{k, l, i} specific to each sub-band SBi of a combined de-signaled signal X _{k, l} , corresponding to the corresponding equation: $${\left|{\mathrm{X}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}\mathrm{=}{\left|{\hat{\mathrm{X}}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}}\mathrm{+}\mathrm{\gamma }{\left|{\mathrm{Y}}_{\mathrm{k}\mathrm{,}\mathrm{l}\mathrm{,}\mathrm{i}}\right|}^{\mathrm{2}};$$

a root-square submodule (361) for calculating the module | X _{k, l, i} | ² combined de-banded components of sub-band X _{k, l, i} ; and a recombination sub-module (362) of the combined sub-band components X _{k, l, i} to obtain the combined de-signaled signal X _{k, 1} , from the modules | X _{k, l, i} | ² and possibly phases θ _{k, 1} , such that the output unedited signal SD _{k, 1} corresponds to the combined de-signaled signal X _{k, 1} , ie SD _{k, l} X _{k, l} . System (1) according to any one of claims 11 to 13, wherein the digital processing unit (3) further comprises, at the output of the estimation module "MCRA" (33), a noise detection module (37) including a module for calculating a signal-to-noise average ratio r ₁ specific to each time frame I from the square module | Y _{k, l} | ² and the noise component D _{k, l} ; a module for comparing the average signal-to-noise ratio r ₁ specific to each time frame I with a predetermined threshold ψ _TH ; a module for controlling the reconstruction module (42) of the output speech-output signal sd (t) which is designed so that: if the average signal-to-noise ratio r ₁ is less than said threshold ψ _TH for the time frame I, then the signal considered before reconstruction for this time frame I corresponds to the output signal Sd ₁ from the calculation module of said output signal Sd _l ; if the average signal-to-noise ratio r ₁ is greater than said threshold ψ _TH for the time frame I, then the signal considered before reconstruction for this time frame I corresponds to the sampled signal y ₁ originating from the noisy acoustic signal cutting module y ( t).

Images