EP3025342B1 - Method for suppressing the late reverberation of an audible signal - Google Patents

Description

TECHNICAL AREA

The invention relates to a method for suppressing the late reverberation of a sound signal. The invention is more particularly, but not exclusively, adapted to the field of the treatment of reverberation in a closed space.

STATE OF THE ART

The figure 1 shows an omnidirectional sound source 100 positioned in a closed space 110, such as a motor vehicle or a room, and a microphone 120. A sound signal emitted by the omnidirectional sound source 100 propagates in all directions. Thus, the signal observed at the microphone is formed by the superposition of several delayed and attenuated versions of the sound signal emitted by the omnidirectional sound source 100. Indeed, the microphone 120 first captures the source signal 130, also called signal direct 130, but also reflected signals 140 on the walls of the closed space 110. The various reflected signals 140 have traveled acoustic paths of different lengths and have been attenuated by the absorption of the walls of the closed space 110, the phase and the amplitude of the reflected signals 140 picked up by the microphone 120 are therefore different.

Two types of reflections exist, early reflections and late reverberation. The microphone 120 picks up the early reflection signals with a small delay compared to the source signal 130, of the order of zero milliseconds to fifty milliseconds. Said early reflection signals are temporally and spatially separated from the source signal 130, but the human ear does not perceive these early reflection signals and the source signal 130 separately by virtue of an effect called "precedence effect". In the case where the sound signal emitted by the omnidirectional sound source 100 is a speech signal, the temporal integration of the early reflection signals by the human ear makes it possible to highlight certain characteristics of the speech, which favors the speech signal. intelligibility of the sound signal.

Depending on the size of the room, the boundary between early reflections and late reverberation is between fifty milliseconds and eighty milliseconds. The late reverberation includes many signals reflected close together in time and therefore impossible to separate. The set of these reflected signals is therefore considered in a probabilistic framework as a random distribution whose density increases with time. In the case where the sound signal emitted by the omnidirectional sound source 100 is a speech signal, the late reverberation degrades the quality of said sound signal and its intelligibility. Said late reverberation also affects the performance of speech recognition and sound source separation systems.

According to the prior art, a first method called "inverse filtering" seeks to identify the impulse response of the closed space 110 to then build an inverse filter to compensate for the effects of reverberation at the sound signal.

This type of process is described for example in the following scientific publications: " BWGillespie, HS Malvar, and DAF Florencio, Speech dereverberation via maximum-kurtosis subband adaptive filtering, Proc. International Conference on Acoustics, Speech and Signal Processing, Volume 6 of ICASSP '01, pages 3701-3704. IEEE, 2001 » , " Mr. Wu and DL Wang. A two-stage algorithm for one-microphone reverberant speech enhancement, Audio, Speech, and Language Processing, IEEE Transactions on, 14 (3): 774-784, 2006 »,« Saeed Mosayyebpour, Abolghasem Sayyadiyan, Mohsen Zareian, and Ali Shahbazi, Single Channel Inverse Filtering of Impulse Response Room by Maximizing Skewness of LP Residual . ".

This method exploits in the time domain distortions introduced by the reverberation on parameters of a model of linear prediction of the sound signal. Starting from the observation that the reverberation mainly modifies the residual of the linear prediction model of the sound signal, a filter maximizing the higher order moments of said residual is constructed. This method is suitable for short pulse responses and is mainly used to compensate for early reflection signals.

However, this method assumes that the impulse response of the closed space 110 is invariant in time. In addition, this method does not model the late reverberation. This method must thus be combined with another method dealing with late reverberation. These two combined processes require numerous iterations before obtaining convergence, so that said methods can not be implemented for a real-time application. In addition, inverse filtering introduces artifacts such as pre-echoes, which must then be compensated.

A second method called "cepstral" aims to separate the effect of the closed space 110 and the sound signal in the cepstral domain. Indeed, the reverberation modifies the average and the variance of the cepstres of the signals reflected with respect to the average and the variance of the cepstres of the source signal 130. Thus, when the mean and the variance of the cepstres are normalized, the reverberation is attenuated.

This type of process is described, for example, in the following scientific publication: D Bees, M Blostein, and P Kabal, Reverberant speech enhancement using cepstral processing, ICASSP '91 Proceedings of the Acoustics, Speech, and Signal Processing, 1991 ".

This method is particularly useful for speech recognition problems since the reference databases of the recognition systems can also be normalized to approach the signals picked up by the microphone 120. However, the effects of the closed space 110 and the signal sound are not completely separable in the cepstral domain. The implementation of the method thus causes a distortion of the timbre of the sound signal emitted by the omnidirectional sound source 100. In addition, this method processes the early reflections rather than the late reverberation.

A third method called "estimating the spectral power density of late reverberation" allows to establish a parametric model of late reverberation.

This type of process is for example described in the following scientific publications: EAP Habets, Single- and Multi-Microphone Speech Dereverberation using Spectral Enhancement, PhD thesis, Technische Universiteit Eindhoven, 2007 »,« T. Yoshioka, Speech Enhancement, Reverberant Environments, PhD Thesis, 2010 ".

According to this third method, an estimate of the spectral density of late reverberation power makes it possible to construct a spectral subtraction filter for the dereverberation. Spectral subtraction introduces artifacts, such as musical noise, but said artifacts can be limited by applying more complex filtering schemes used by denoising methods.

However, an important parameter for estimating the power spectral density of late reverberation in this third method is the reverberation time. However, the reverberation time is a difficult parameter to estimate accurately. The reverberation time estimate is distorted by background noise and other interfering sound signals. In addition, this estimate of the reverberation time is time consuming and therefore lengthens the execution time.

A fourth method exploits the parsimony of the speech signals in the time / frequency plane.

This type of process is described for example in the following scientific publication " T. Yoshioka, Speech Enhancement in Reverberant Environments, PhD Thesis, 2010 ".

In this publication, late reverberation is modeled as a delayed and attenuated version of the current observation whose attenuation factor is determined by solving a maximum likelihood problem, with a parsimony constraint.

This type of process is further described in the following scientific publication " H Kameoka, T Nakatani, and T Yoshioka, Proceedings of the 2009 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP '09, pp 45-48. IEEE Computer Society, 2009 ".

Dereverberation is approached in this publication as a problem of deconvolution by factorization in non-negative matrices, which makes it possible to separate the response of the closed space 110 and the sound signal. However, this process introduces a lot of noise and distortions. In addition, said method depends on the initialization of the matrices for the factorization.

In addition, the methods mentioned require a plurality of microphones to accurately process the reverberation.

STATEMENT OF THE INVENTION

The invention is intended in particular to solve all or part of the aforementioned problems.

• captation d'un signal d'entrée formé par la superposition de plusieurs versions retardées et atténuées du signal sonore,
• application d'une transformation temps-fréquence au signal d'entrée afin d'obtenir une transformée temps-fréquence complexe du signal d'entrée,
• calcul d'une pluralité de vecteurs de prédiction,
• création d'une pluralité de vecteurs d'observation à partir du module de la transformée temps-fréquence complexe du signal d'entrée,
• construction d'une pluralité de dictionnaires de synthèse à partir de la pluralité de vecteurs d'observations,
• estimation d'un spectre de réverbération tardive à partir de la pluralité de dictionnaires de synthèse et de la pluralité de vecteurs de prédiction,
• filtrage de la pluralité de vecteurs d'observations afin d'éliminer le spectre de réverbération tardive et d'obtenir un module de signal déréverbéré.

To this end, the invention relates to a method for suppressing the late reverberation of a sound signal characterized in that it comprises the following steps:

• capture of an input signal formed by the superposition of several delayed and attenuated versions of the sound signal,
• applying a time-frequency transformation to the input signal to obtain a complex time-frequency transform of the input signal,
• calculating a plurality of prediction vectors,
• creating a plurality of observation vectors from the module of the complex time-frequency transform of the input signal,
• constructing a plurality of synthetic dictionaries from the plurality of observation vectors,
• estimating a late reverberation spectrum from the plurality of synthetic dictionaries and the plurality of prediction vectors,
• filtering the plurality of observation vectors to eliminate the late reverberation spectrum and obtain a dereverberated signal module.

Thus, the method which is the subject of the invention is rapid and has a reduced complexity. This method is therefore usable in real time. In addition, this method does not introduce artifacts and is robust to background noise. In addition, said method reduces background noise and is compatible with noise reduction methods.

The invention can be implemented according to the embodiments advantageous below, which can be considered individually or in any combination technically operative.

• création d'un module sous échantillonné en fréquence à partir du module de la transformée temps-fréquence complexe du signal d'entrée,
• création d'une pluralité de vecteurs d'observation sous échantillonnés à partir dudit module sous échantillonné en fréquence,
• construction d'une pluralité de dictionnaires d'analyse à partir de la pluralité de vecteurs d'observation sous échantillonnés,
• calcul de la pluralité de vecteurs de prédiction à partir de la pluralité de vecteurs d'observation sous échantillonnés et de la pluralité de dictionnaires d'analyse.

Advantageously, the method further comprises the following steps:

• creating a subsampled frequency module from the module of the complex time-frequency transform of the input signal,
• creating a plurality of subsampled observation vectors from said subsampled frequency module,
• constructing a plurality of analysis dictionaries from the plurality of subsampled observation vectors,
• calculating the plurality of prediction vectors from the plurality of subsampled observation vectors and the plurality of analysis dictionaries.

Advantageously, the step of calculating the plurality of prediction vectors is performed by minimizing, for each prediction vector, the expression ∥ Xv-D ^a α ∥ ₂ , which is the Euclidean norm of the difference between the vector of subsampled observation associated with said prediction vector and the analysis dictionary associated with said prediction vector multiplied by said prediction vector, taking into account the stress ∥ α ∥ ₁ ≤ λ , according to which the norm 1 of said prediction vector is lower or equal to a maximum intensity parameter of the late reverberation.

Advantageously, the value of the maximum intensity parameter of the late reverberation is between 0 and 1.

• création d'un signal complexe déréverbéré à partir du module de signal déréverbéré et de la phase de la transformée temps-fréquence complexe du signal d'entrée.

Advantageously, the method further comprises the following step:

• creating a dereverberated complex signal from the dereverberated signal module and the phase of the complex time-frequency transform of the input signal.

• application d'une transformation fréquence-temps au signal complexe déréverbéré afin d'obtenir un signal temporel déréverbéré.

Advantageously, the method further comprises the following step:

• applying a frequency-time transformation to the complex dereverberated signal in order to obtain a dereverberated temporal signal.

où ξ est le rapport signal à bruit a priori, et où la borne d'intégration υ est calculée selon le modèle $$\upsilon =\gamma \frac{\xi }{1+\xi }$$

où γ est le rapport signal à bruit a postériori.Advantageously, the method further comprises a step of constructing a dereverberation filter according to the model $$\mathrm{BOY\; WUT}=\frac{\mathrm{<figure-callout\; id="1"\; label="\xi "\; filenames="imgf0002.png"\; state="\{\{state\}\}">\xi </figure-callout>}}{1+\xi }\exp \left({\int }_{\upsilon }^{\infty }\frac{e^{-t}}{t}\mathit{dt}\right),$$

where ξ is the signal-to-noise ratio a priori, and where the integration bound υ is calculated according to the model $$\upsilon =\gamma \frac{\xi }{1+\xi }$$

where γ is the signal-to-noise ratio a posteriori.

• capter un signal d'entrée formé par la superposition de plusieurs versions retardées et atténuées du signal sonore,
• appliquer une transformation temps-fréquence au signal d'entrée afin d'obtenir une transformée temps-fréquence complexe du signal d'entrée,
• calculer une pluralité de vecteurs de prédiction,
• créer une pluralité de vecteurs d'observation à partir du module de la transformée temps-fréquence complexe du signal d'entrée,
• construire une pluralité de dictionnaires de synthèse à partir de la pluralité de vecteurs d'observations,
• estimer un spectre de réverbération tardive à partir de la pluralité de dictionnaires de synthèse et de la pluralité de vecteurs de prédiction,
• filtrer la pluralité de vecteurs d'observations afin d'éliminer le spectre de réverbération tardive et d'obtenir un module de signal déréverbéré.

The invention also relates to a device for suppressing the late reverberation of a sound signal characterized in that it comprises means for:

• capture an input signal formed by the superposition of several delayed and attenuated versions of the sound signal,
• applying a time-frequency transformation to the input signal to obtain a complex time-frequency transform of the input signal,
• calculate a plurality of prediction vectors,
• creating a plurality of observation vectors from the module of the complex time-frequency transform of the input signal,
• constructing a plurality of synthetic dictionaries from the plurality of observation vectors,
• estimating a late reverberation spectrum from the plurality of synthetic dictionaries and the plurality of prediction vectors,
• filtering the plurality of observation vectors to eliminate the late reverberation spectrum and obtain a dereverberated signal module.

PRESENTATION OF FIGURES

• Figure 1 (déjà décrite) : une représentation schématique d'une source sonore omnidirectionnelle et d'un microphone positionnés dans un espace fermé selon un exemple de réalisation de l'invention ;
• Figure 2 : une représentation schématique d'un dispositif de déréverbération d'un signal sonore selon un exemple de réalisation de l'invention ;
• Figure 3 : une représentation schématique d'une unité de déréverbération d'un dispositif de déréverbération d'un signal sonore selon un exemple de réalisation de l'invention ;
• Figure 4 : une représentation schématique d'une unité d'estimation de la réverbération tardive d'un dispositif de déréverbération d'un signal sonore selon un exemple de réalisation de l'invention ;
• Figure 5 : une représentation schématique d'un regroupement en sous bandes d'un module d'une transformée temps-fréquence complexe d'un signal d'entrée selon un exemple de réalisation de l'invention ;
• Figure 6 : une représentation schématique d'une unité de calcul de vecteurs de prédiction d'un dispositif de déréverbération d'un signal sonore selon un exemple de réalisation de l'invention ;
• Figure 7 : une représentation schématique d'une unité de calcul de vecteurs de prédiction d'un dispositif de déréverbération d'un signal sonore selon un exemple de réalisation de l'invention ;
• Figure 8 : une représentation schématique d'une unité d'évaluation de la réverbération d'un dispositif de déréverbération d'un signal sonore selon un exemple de réalisation de l'invention ;
• Figure 9 : un diagramme fonctionnel montrant différentes étapes du procédé selon un exemple de réalisation de l'invention.

The invention will be better understood on reading the following description, given by way of non-limiting example, and with reference to the figures which represent:

• Figure 1 (already described): a schematic representation of an omnidirectional sound source and a microphone positioned in a closed space according to an exemplary embodiment of the invention;
• Figure 2 : a schematic representation of a device for dereverberation of a sound signal according to an exemplary embodiment of the invention;
• Figure 3 : a schematic representation of a dereverberation unit of a device for dereverbating a sound signal according to an exemplary embodiment of the invention;
• Figure 4 : a schematic representation of a unit for estimating the late reverberation of a device for the dereverberation of a sound signal according to an exemplary embodiment of the invention;
• Figure 5 : a schematic representation of a grouping in sub-bands of a module of a complex time-frequency transform of an input signal according to an exemplary embodiment of the invention;
• Figure 6 : a schematic representation of a prediction vector calculation unit of a sound signal dereverberation device according to an exemplary embodiment of the invention;
• Figure 7 : a schematic representation of a prediction vector calculation unit of a sound signal dereverberation device according to an exemplary embodiment of the invention;
• Figure 8 : a schematic representation of a unit for evaluating the reverberation of a device for the dereverberation of a sound signal according to an exemplary embodiment of the invention;
• Figure 9 : a functional diagram showing different steps of the method according to an exemplary embodiment of the invention.

In these figures, identical references from one figure to another designate identical or similar elements. For the sake of clarity, the elements shown are not to scale unless otherwise stated.

DETAILED DESCRIPTION OF EMBODIMENTS

The invention uses a device for the dereverberation of a sound signal emitted by an omnidirectional sound source 100 positioned in a closed space 110, such as a motor vehicle or a room, and picked up by a microphone 120. Said device for dereverberation is inserted into the audio processing chain of a device such as a telephone. This dereverberation device comprises a unit for applying a time-frequency transform 200, a dereverberation unit 210 and a unit for applying a frequency-time transform 220 (cf. figure 2 ). The dereverberation unit 210 comprises a unit for estimating the late reverberation 300 and a filtering unit 310 (cf. figure 3 ). The late reverberation estimation unit 300 comprises a subband consolidation unit 400, a prediction vector calculation unit 410 and a reverberation evaluation unit 420 (see FIG. figure 4 ). The prediction vector calculation unit 410 comprises an observation construction unit 700, an analysis dictionary construction unit 710 and a resolution unit of the LASSO 720 (cf. figure 7 ). The reverberation evaluation unit 420 comprises a synthesis dictionary construction unit 800 (cf. figure 8 ).

In a step 900, a microphone 120 captures an input signal x ( t ) formed by the superposition of several delayed and attenuated versions of the sound signal emitted by the omnidirectional sound source 100. In fact, the microphone 120 first captures the source signal 130, also called direct signal 130, but also reflected signals 140 on the walls of the closed space 110. The various reflected signals 140 have traveled acoustic paths of different lengths and have been attenuated by the absorption of the walls of the closed space 110, the phase and the amplitude of the reflected signals 140 picked up by the microphone 120 are therefore different.

Two types of reflections exist, early reflections and late reverberation. The microphone 120 captures the reflection signals early with a small delay compared to the source signal 130, of the order of zero milliseconds to fifty milliseconds. Said early reflection signals are temporally and spatially separated from the source signal 130 but the human ear does not perceive these early reflection signals and the source signal 130 separately by virtue of an effect called "precedence effect". In the case where the sound signal emitted by the omnidirectional sound source 100 is a speech signal, the temporal integration of the early reflection signals by the human ear makes it possible to highlight certain characteristics of the speech, which favors the speech signal. intelligibility of the sound signal.

The microphone 120 captures the late reverberation between fifty milliseconds and eighty milliseconds after the arrival of the source signal 130. The late reverb includes many signals reflected close together over time and thus impossible to separate. The set of these reflected signals is therefore considered in a probabilistic framework as a random distribution whose density increases with time. In the case where the sound signal emitted by the omnidirectional sound source 100 is a speech signal, the late reverberation degrades the quality of said sound signal and its intelligibility. Said late reverberation also affects the performance of speech recognition and sound source separation systems.

The input signal x ( t ) is sampled at a sampling frequency f _s . The input signal x ( t ) is thus subdivided into samples. In order to suppress the late reverberation of said input signal x (t), the power spectral density of the late reverberation is estimated then a dereverberation filter is constructed by the dereverberation unit 210. The estimate of the spectral density of power of the late reverberation, the construction of the dereverberation filter and the application of said dereverberation filter are performed in the frequency domain. Thus, in a step 901, a time-frequency transformation is applied to the input signal x ( t ) by the application unit of the Short-Fourier Transform 200 to obtain a complex time-frequency transform of the input signal x ( t ) denoted X ^C (cf. figure 2 ). In one example, the transformation Time-frequency is a short-term Fourier Transformation.

de la transformée temps-fréquence complexe X^C est calculé de la façon suivante : $$X_{k,n}^C={\displaystyle \sum _{m=0}^{M-1}}x\left(m+\mathit{nR}\right)w\left(m\right)e^{-\frac{2\mathit{j\pi km}}{M}}$$

où k est un indice fréquentiel d'échantillonnage de valeur comprise entre 1 et un nombre K, n est un indice temporel de valeur comprise entre 1 et un nombre N, w(m) est une fenêtre glissante d'analyse, m est l'indice des éléments appartenant à une trame, M est la longueur d'une trame, c'est-à-dire le nombre d'échantillons d'une trame et R est le pas d'avancement de la transformation temps-fréquence.Each element $X_{k,\mathrm{not}}^{\mathrm{VS}}$

of the complex time-frequency transform X ^C is calculated as follows: $$X_{k,\mathrm{not}}^{\mathrm{VS}}={\displaystyle \underset{m=0}{\overset{M-1}{\Sigma }}}x\left(m+\mathit{n\; R}\right)w\left(m\right)e^{-\frac{2\mathit{j\pi km}}{M}}$$

where k is a frequency sampling index of value between 1 and a number K, n is a time index of value between 1 and a number N, w (m) is a sliding window of analysis, m is the index of the elements belonging to a frame, M is the length of a frame, that is to say the number of samples of a frame and R is the step of advancement of the time-frequency transformation.

défini par $$X_{k,n}^C=\left|X_{k,n}\right|e^{-j\angle X_{k,n}},$$

où |X_k,n | est le module du signal complexe $X_{k,n}^C$

et ∠X_k,n est la phase du signal complexe $X_{k,n}^C\mathrm{.}$

The input signal x ( t ) is analyzed by frames of length M with a pitch R equal to M / 4 samples. For each frame of the input signal x ( t ) in the time domain a discrete time-frequency transform of sampling frequency index k and of time index n is thus calculated by means of the time-frequency transformation algorithm to get a complex signal $X_{k,\mathrm{not}}^{\mathrm{VS}}$

defined by $$X_{k,\mathrm{not}}^{\mathrm{VS}}=\left|X_{k,\mathrm{not}}\right|e^{-j\angle X_{k,\mathrm{not}}},$$

where | X _{k, n} | is the module of the complex signal $X_{k,\mathrm{not}}^{\mathrm{VS}}$

and ∠X _{k, n} is the phase of the complex signal $X_{k,\mathrm{not}}^{\mathrm{VS}}\mathrm{.}$

The estimate of the power spectral density of the late reverberation is performed on the complex time-frequency transform module of the input signal X ^C , denoted X. The phase of the complex time-frequency transform X ^C , denoted ∠X is kept in memory and is used to reconstruct a dereverberated signal in the time domain after application of the dereverberation filter.

The module X of the complex time-frequency transform of the input signal X ^C is then grouped into sub-bands. More precisely, said module X comprises the number K of spectral lines denoted X _k . The term "spectral line" here designates all the samples of the module X of the complex time-frequency transform of the input signal X ^C for the sampling frequency index k and all the time indices n. In a step 903, the sub-banding unit 400 groups the K spectral lines X _k into a number J of sub-bands, in order to obtain a sub-sampled module denoted X having a number J of spectral lines denoted X _j , where j is a subsampling frequency index between 1 and the number J. The number J is less than the number K. Each subband thus comprises a plurality of spectral lines X _k , the frequency index k belonging to an interval having a lower bound b _j and an upper bound e _j . In one example, each subband corresponds to an octave to take into account the sound perception model of the human ear. Then, in a step 904, the subband grouping unit 400 calculates, for each subband, a mean Mean of the spectral lines X _k of said subband in order to obtain the J spectral lines X _i of the sub-sampled module. frequency X (cf. figure 5 ).

Then, the prediction vector calculation unit 410 calculates for each spectral line X _j of the sub-sampled module at frequency X and for each temporal index n a prediction vector α _{j, n} (cf. figure 6 ). More precisely, in a step 905, the observation construction unit 700 constructs, for each temporal index n and sub-sampling frequency index j, a sub-sampled observation vector Xv _{j, n} from the set of samples X _{j , n 1: n} belonging to the j-th spectral line X _j of the module sub-sampled at frequency X and between the instants n ₁ = nN +1 and n, where n is the index of the current instant and nn ₁ is the size of the memory of the dereverberation device. Each subsampled observation vector Xv _{j, n} is defined by $$\tilde{X}{v}_{j,\mathrm{not}}:={\left[{\tilde{X}}_{j,\mathrm{not}}...{\tilde{X}}_{j,\mathrm{not}-\mathrm{NOT}+1}\right]}^T,$$

Each observation vector Xv _{j, n} is of size N × 1, where the number N is the length of the observation. The length of observation N is the number of the frames of the time-frequency transformation necessary for the estimation of the late reverberation. The length of the observation N makes it possible to define the temporal resolution of the estimate. As the length of observation N increases, the complexity of the system decreases. The subsampling of the X module of the complex time-frequency transform of the input signal X ^C makes it possible, among other things, to apply the method in real time.

est construit en concaténant un nombre L de vecteurs d'observations passées déterminés à l'étape 905. Le dictionnaire d'analyse $D_{j,n}^a$

se définit ainsi comme la matrice $$D_{j,n}^a:=\left[\begin{array}{cccc}{\tilde{X}}_{j,n-\delta }& {\tilde{X}}_{j,n-\delta -1}& \cdots & {\tilde{X}}_{j,n-\delta -L+1}\\ {\tilde{X}}_{j,n-\delta -1}& {\tilde{X}}_{j,n-\delta -2}& \cdots & {\tilde{X}}_{j,n-\delta -L}\\ ⋮& ⋮& ⋮& ⋮\\ {\tilde{X}}_{j,n-\delta -N+1}& {\tilde{X}}_{j,n-\delta -N}& \cdots & {\tilde{X}}_{j,n-\delta -L-N+2}\end{array}\right]$$

où L est le nombre de vecteurs d'observations passées et donc la taille du dictionnaire d'analyse $D_{j,n}^a,$

et $\mathrm{\delta }\in \mathbb{R}*$

est le retard du dictionnaire d'analyse $D_{j,n}^a\mathrm{.}$

Plus précisément, le retard δ est le retard de trames entre le vecteur d'observation courante sous échantillonné X̃v_j,n et les autres vecteurs d'observations sous échantillonnés appartenant au dictionnaire d'analyse $D_{j,n}^a\mathrm{.}$

Ledit retard δ permet de réduire les distorsions introduites par le procédé. Ce retard δ permet en outre de d'améliorer la séparation de la réverbération tardive et des réflexions précoces. Pour calculer le vecteur d'observation courante X̃v_j,n et le dictionnaire d'analyse $D_{j,n}^a$

et donc le vecteur de prédiction α_j,n pour chaque ligne spectrale X̃_j et pour chaque indice temporel n, un nombre L+N+δ de trames doit être gardé en mémoire.In a step 906, the analysis dictionaries building unit 710 builds analysis dictionaries D ^a . More precisely, for each temporal index n and sub-sampling frequency index j, an analysis dictionary $D_{j,\mathrm{not}}^{\mathrm{at}}$

is constructed by concatenating a number L of past observation vectors determined in step 905. The analysis dictionary $D_{j,\mathrm{not}}^{\mathrm{at}}$

is defined as the matrix $$D_{j,\mathrm{not}}^{\mathrm{at}}:=\left[\begin{array}{cccc}{\tilde{X}}_{j,\mathrm{not}-\delta }& {\tilde{X}}_{j,\mathrm{not}-\delta -1}& \cdots & {\tilde{X}}_{j,\mathrm{not}-\delta -\mathrm{The}+1}\\ {\tilde{X}}_{j,\mathrm{not}-\delta -1}& {\tilde{X}}_{j,\mathrm{not}-\delta -2}& \cdots & {\tilde{X}}_{j,\mathrm{not}-\delta -\mathrm{The}}\\ ⋮& ⋮& ⋮& ⋮\\ {\tilde{X}}_{j,\mathrm{not}-\delta -\mathrm{NOT}+1}& {\tilde{X}}_{j,\mathrm{not}-\delta -\mathrm{NOT}}& \cdots & {\tilde{X}}_{j,\mathrm{not}-\delta -\mathrm{The}-\mathrm{NOT}+2}\end{array}\right]$$

where L is the number of past observation vectors and therefore the size of the analysis dictionary $D_{j,\mathrm{not}}^{\mathrm{at}},$

and $\mathrm{\delta }\in \mathbb{R}*$

is the delay of the analysis dictionary $D_{j,\mathrm{not}}^{\mathrm{at}}\mathrm{.}$

More precisely, the delay δ is the frame delay between the subsampled current observation vector Xv _{j, n} and the other subsampled observation vectors belonging to the analysis dictionary. $D_{j,\mathrm{not}}^{\mathrm{at}}\mathrm{.}$

Said delay δ makes it possible to reduce the distortions introduced by the method. This delay δ also makes it possible to improve the separation of the late reverberation and the early reflections. To calculate the current observation vector Xv _{j, n} and the analysis dictionary $D_{j,\mathrm{not}}^{\mathrm{at}}$

and therefore the prediction vector α _{j, n} for each spectral line X _j and for each temporal index n, a number L + N + δ of frames must be kept in memory.

en tenant compte de la contrainte ∥α_j,n ∥ ≤ λ où λ est un paramètre d'intensité maximale. Pour résoudre ledit problème, la meilleure combinaison linéaire des L vecteurs du dictionnaire permettant d'approcher l'observation courante doit être trouvée. Dans un exemple, un procédé connu, appelé LARS, selon l'acronyme anglo-saxon de "Least Angle Regression" permet de résoudre ledit problème. La contrainte ∥α_j,n ∥₁ ≤ λ permet de privilégier les solutions ayant peu d'éléments non nuls, c'est-à-dire les solutions parcimonieuses. Le paramètre d'intensité maximale λ permet de régler l'intensité maximale estimée de la réverbération tardive. Ce paramètre d'intensité maximale λ dépend a priori de l'environnement acoustique, c'est-à-dire dans un exemple de l'espace fermé 110. Pour chaque espace fermé 110, une valeur optimale du paramètre d'intensité maximale λ existe. Cependant, des essais ont montré que ledit paramètre d'intensité maximale λ peut être fixé à une valeur identique pour tous les espaces fermés 110, sans que ladite valeur introduise de dégradations par rapport à la valeur optimale. Ainsi le procédé fonctionne dans une grande variété d'espaces fermés 110 sans nécessiter de réglage particulier, ce qui permet de s'affranchir des erreurs d'estimation du temps de réverbération de l'espace fermé 110. En outre, le procédé selon l'invention ne nécessite pas de paramètre devant être estimé, ce qui permet l'application dudit procédé en temps réel. La valeur du paramètre d'intensité maximale λ est comprise entre 0 et 1. Dans un exemple, la valeur du paramètre d'intensité maximale λ est égale à 0,5, ce qui est un bon compromis entre la réduction de la réverbération et la qualité globale du procédé.In a step 907, the resolution unit of LASSO 720 solves a problem called "LASSO" which is to minimize the Euclidean norm ${\Vert \tilde{X}{v}_{j,\mathrm{not}}-D_{j,\mathrm{not}}^{\mathrm{at}}{\alpha }_{j,\mathrm{not}}\Vert }_2$

taking into account the stress ∥ α _{j, n} ∥ ≤ λ where λ is a maximum intensity setting. To solve this problem, the best linear combination of the L vectors of the dictionary allowing to approach the current observation must be found. In one example, a known method, called LARS, according to the English acronym of "Least Angle Regression" solves the problem. The constraint ∥ α _{j, n} ∥ ₁ ≤ λ makes it possible to favor solutions with few non-zero elements, that is, parsimonious solutions. The maximum intensity parameter λ adjusts the estimated maximum intensity of the late reverberation. This parameter of maximum intensity λ depends a priori on the acoustic environment, that is to say in an example of the closed space 110. For each closed space 110, an optimum value of the maximum intensity parameter λ exists . However, tests have shown that said maximum intensity parameter λ can be set to an identical value for all the closed spaces 110, without said value introducing degradations with respect to the optimal value. Thus, the method operates in a wide variety of closed spaces 110 without requiring any particular adjustment, which makes it possible to avoid errors in estimating the reverberation time of the closed space 110. In addition, the method according to The invention does not require a parameter to be estimated, which allows the application of said method in real time. The value of the maximum intensity parameter λ is between 0 and 1. In one example, the value of the maximum intensity parameter λ is equal to 0.5, which is a good compromise between the reduction of the reverberation and the overall quality of the process.

In a step 908, for each time index n and each sampling frequency index k, a current observation vector Xv _{k, n} is created from the set of samples belonging to the k-th spectral line X _k of module X of the complex time-frequency transform and between the instants n ₁ and n, denoted by X _{k, n1 : n} where n is the current-time index and n - n ₁ is the size of the memory of the device of dereverberation. Each observation vector Xv _{k, n} is defined by the formula Xv _{k, n} : = [ X _{k, n} ... X _{k, nN +1} ] ^T , and is of size N × 1, where N is the length of observation.

est construit en concaténant un nombre L de vecteurs d'observations passées déterminés à l'étape 908. Le dictionnaire de synthèse $D_{k,n}^s$

se définit ainsi comme la matrice $$D_{k,n}^s:=\left[\begin{array}{cccc}{X}_{k,n-\delta }& X_{k,n-\delta -1}& \cdots & X_{k,n-\delta -L+1}\\ X_{k,n-\delta -1}& X_{k,n-\delta -2}& \cdots & X_{k,n-\delta -L}\\ ⋮& ⋮& ⋮& ⋮\\ X_{k,n-\delta -N+1}& X_{k,n-\delta -N}& \cdots & X_{k,n-\delta -L-N+2}\end{array}\right]$$

où L et δ sont les même paramètres que pour le dictionnaire d'analyse $D_{j,n}^a\mathrm{.}$

In a step 909, the construction unit of a synthesis dictionary 800 constructs a synthesis dictionary D ^s . More precisely, for each temporal index n and each sampling frequency index k, the summary dictionary $D_{k,\mathrm{not}}^s$

is constructed by concatenating a number L of past observation vectors determined at step 908. The summary dictionary $D_{k,\mathrm{not}}^s$

is defined as the matrix $$D_{k,\mathrm{not}}^s:=\left[\begin{array}{cccc}{X}_{k,\mathrm{not}-\delta }& X_{k,\mathrm{not}-\delta -1}& \cdots & X_{k,\mathrm{not}-\delta -\mathrm{The}+1}\\ X_{k,\mathrm{not}-\delta -1}& X_{k,\mathrm{not}-\delta -2}& \cdots & X_{k,\mathrm{not}-\delta -\mathrm{The}}\\ ⋮& ⋮& ⋮& ⋮\\ X_{k,\mathrm{not}-\delta -\mathrm{NOT}+1}& X_{k,\mathrm{not}-\delta -\mathrm{NOT}}& \cdots & X_{k,\mathrm{not}-\delta -\mathrm{The}-\mathrm{NOT}+2}\end{array}\right]$$

where L and δ are the same parameters as for the analysis dictionary $D_{j,\mathrm{not}}^{\mathrm{at}}\mathrm{.}$

est construit par multiplication du dictionnaire de synthèse $D_{k,n}^s$

avec le vecteur de prédiction α_j,n selon la formule $$X_{k,n}^{\ell }=D_{k,n}^s{\alpha }_{j,n}\forall k\in \lfloor b_j,e_j\rfloor ,\mathit{j}=1,\dots ,J$$

In a step 910, for each time index n and each sampling frequency index k, an estimate of the power spectral density of the late reverberation or the late reverberation spectrum $X_{k,\mathrm{not}}^{\ell }$

is built by multiplication of the synthetic dictionary $D_{k,\mathrm{not}}^s$

with the prediction vector α _{j, n} according to the formula $$X_{k,\mathrm{not}}^{\ell }=D_{k,\mathrm{not}}^s{\alpha }_{j,\mathrm{not}}\forall k\in \lfloor b_j,e_j\rfloor ,\mathit{j}=1,...,J$$

The prediction vector α _{j, n} therefore indicates the columns of the synthesis dictionary that have been selected for estimating the reverberation, as well as the contribution of each of them to the reverberation. The spectrum of late reverberation X ^ℓ is considered in the rest of the process as a noise signal to be eliminated.

où ξ_k,n est le rapport signal à bruit a priori, calculé de la façon suivante $${\xi }_{k,n}=\beta G_{k,n-1}^2{\gamma }_{k,n-1}+\left(1-\beta \right)\max \left\{{\gamma }_{k,n}-1,0\right\}$$

et où la borne d'intégration ν_k,n est calculée de la façon suivante $${\upsilon }_{k,n}={\gamma }_{k,n}\frac{{\xi }_{k,n}}{1+{\xi }_{k,n}}$$

où γ_k,n est le rapport signal à bruit a postériori, calculé selon la formule $${\gamma }_{k,n}=\frac{{\left|X_{k,n}\right|}^2}{{\left|R_{k,n}\right|}^2}$$

où R_k,n est la réverbération tardive lissée calculée de la façon suivante $$R_{k,n}={\mathit{\alpha R}}_{k,n-1}+\left(1-\alpha \right)\left|X_{k,n}^{\ell }\right|$$

où α est une première constante de lissage et β est une seconde constante de lissage. Dans un exemple, la première constante de lissage α vaut 0.77 et la seconde constante de lissage β vaut 0.98.For this purpose, a filtering of the reverberation is carried out by the filtering unit 310. More specifically, in a step 911, for each time index n and each sampling frequency index k, a dereverberation filter G _{k, n} is built according to the formula $${\mathrm{BOY\; WUT}}_{k,\mathrm{not}}=\frac{{\xi }_{k,\mathrm{not}}}{1+{\xi }_{k,\mathrm{not}}}\exp \left({\int }_{{\upsilon }_{k,\mathrm{not}}}^{\infty }\frac{e^{-t}}{t}\mathit{dt}\right)$$

where ξ _{k, n} is the signal to noise ratio a priori, calculated as follows $${\xi }_{k,\mathrm{not}}=\beta {\mathrm{BOY\; WUT}}_{k,\mathrm{not}-1}^2{\gamma }_{k,\mathrm{not}-1}+\left(1-\beta \right)\max \left\{{\gamma }_{k,\mathrm{not}}-1,0\right\}$$

and where the integration bound ν _{k, n} is computed as follows $${\upsilon }_{k,\mathrm{not}}={\gamma }_{k,\mathrm{not}}\frac{{\xi }_{k,\mathrm{not}}}{1+{\xi }_{k,\mathrm{not}}}$$

where γ _{k, n} is the signal-to-noise ratio a posteriori, calculated according to the formula $${\gamma }_{k,\mathrm{not}}=\frac{{\left|X_{k,\mathrm{not}}\right|}^2}{{\left|R_{k,\mathrm{not}}\right|}^2}$$

where R _{k, n} is the smoothed late reverberation calculated as follows $$R_{k,\mathrm{not}}={\mathit{\alpha R}}_{k,\mathrm{not}-1}+\left(1-\alpha \right)\left|X_{k,\mathrm{not}}^{\ell }\right|$$

where α is a first smoothing constant and β is a second smoothing constant. In one example, the first smoothing constant α is 0.77 and the second smoothing constant β is 0.98.

Indeed, the estimated reverberation is non-stationary in the long term because the sound signal emitted by the omnidirectional sound source 100, which causes said estimated reverberation is not stationary in the long term. Excessive variations in the estimated reverb can introduce annoying artifacts during filtering. To limit these effects, a recursive smoothing is performed to calculate the power spectral density of the late reverberation.

In a step 912, for each time index n and each sampling frequency index k, the observation vectors Xv _{k, n} are filtered by the dereverberation filter G _{k, n} calculated in step 911 in order to obtain a dverbvered signal module Y _{k, n} calculated as follows $$Y_{k,\mathrm{not}}={\mathrm{BOY\; WUT}}_{k,\mathrm{not}}{X}_{k,\mathrm{not}}\mathrm{.}$$

The filter constructed in step 911 strongly attenuates certain observation vectors Xv _{k, n} which generates artifacts harmful to the quality of the dereverberated signal. To limit said artifacts, a lower bound is imposed on the attenuation of the filter. Thus, for each sampling frequency index k and for each time index n, if the dereverberation filter G _{k, n} is less than or equal to a minimum value of the dereverberation filter Gmin, then said dereverberation filter G _{k, n} is equal to the said minimum value of the filter of Gmin dereverberation.

sont multipliés afin de créer un signal complexe déréverbéré Y^C. In a step 913, for each sampling frequency index k and each time index n, the dereverberated signal module Y _{k, n} and the phase ∠X _{k, n} of the complex signal $X_{k,\mathrm{not}}^{\mathrm{VS}}$

are multiplied to create a complex dereverberated signal Y ^C.

afin d'obtenir un signal temporel déréverbéré y(t) dans le domaine temporel. Dans un exemple, la transformation fréquence-temps est une Transformation de Fourier Inverse à Court Terme.In a step 914, a frequency-time transformation is applied by the unit for applying a frequency-time transformation 220 to the dereverberated complex signal. $Y_{k,\mathrm{not}}^{\mathrm{VS}}$

to obtain a temporal signal dereverbere y ( t ) in the time domain. In one example, the frequency-time transformation is a Short Term Inverse Fourier Transformation.

In one implementation, the value of the number of observation vectors L is equal to 10, the value of the number of observation length N is equal to 8, the value of the delay δ is equal to 5, the value of the parameter of maximum intensity λ is equal to 0.5, the value of the number K is equal to 257, the value of the number J is equal to 10, the value of the length of a frame M is equal to 512 and the minimum value of the filter Gmin dereverberation is equal to -12 decibels. This choice of parameters allows the application of the process in real time.

The method of suppressing the late reverberation of a sound signal according to the invention is rapid and has a reduced complexity. This method is therefore usable in real time. In addition, this method does not introduce artifacts and is robust to background noise. In addition, said method reduces background noise and is compatible with noise reduction methods.

The method of suppressing the late reverberation of a sound signal according to the invention requires a single microphone to accurately process the reverberation.

Claims (6)

1. Method for suppressing the late reverberation of a sound signal, characterized in that it includes the following steps:

   • capturing (900) an input signal (x) formed by the superimposition of a multiplicity of delayed and attenuated versions of the sound signal,

   • applying (901) a time-frequency transformation to the input signal (x) in order to obtain a complex time-frequency transform (X^C ) of the input signal (x),

   • generating a frequency-subsampled modulus (X̃) from the modulus of the complex time-frequency transform (X^C ) of the input signal (x),

   • generating (905) a plurality of subsampled observation vectors from said frequency-subsampled modulus (X̃),

   • constructing (906) a plurality of analysis dictionaries (D^a ) from the plurality of subsampled observation vectors,

   • calculating (907) a plurality of prediction vectors (α) from the plurality of subsampled observation vectors and from the plurality of analysis dictionaries (D^a ), by minimizing, for each prediction vector (α), the expression ∥X̃v - D^aα∥₂, which is the Euclidean norm of the difference between the subsampled observation vector associated with said prediction vector (α) and of the analysis dictionary (D^a ) associated with said prediction vector (α) multiplied by said prediction vector (α), while taking into consideration the constraint ∥α∥₁ ≤ λ, according to which the norm 1 of said prediction vector (α) is less than or equal to a maximum intensity parameter of the late reverberation (λ),

   • generating (908) a plurality of observation vectors from the modulus of the complex time-frequency transform (X^C ) of the input signal (x),

   • constructing (909) a plurality of synthesis dictionaries (D^s ) from the concatenation of the plurality of observation vectors,

   • estimating (910) a late reverberation spectrum (X^ℓ ) from the multiplication of the plurality of synthesis dictionaries (D^s ) by the plurality of prediction vectors (α),

   • filtering (912) the plurality of observation vectors in order to eliminate the late reverberation spectrum (X ^ℓ) and to obtain a dereverberated signal modulus (Y).
2. Method according to Claim 1, characterized in that the value of the maximum intensity parameter of the late reverberation (A) is between 0 and 1.
3. Method according to either of Claims 1 and 2, characterized in that it furthermore includes the following step:

   • generating (913) a dereverberated complex signal (Y^C ) from the dereverberated signal modulus (Y) and from the phase (∠X) of the complex time-frequency transform (X^C ) of the input signal (x).
4. Method according to Claim 3, characterized in that it furthermore includes the following step:

   • applying (914) a frequency-time transformation to the dereverberated complex signal (Y^C ) in order to obtain a dereverberated time signal (y).
5. Method according to one of Claims 1 to 4, characterized in that it furthermore includes a step of constructing a dereverberation filter in accordance with the model $$G=\frac{\xi }{1+\xi }\exp \left({\int }_{\upsilon }^{\infty }\frac{e^{-t}}{t}\mathit{dt}\right),$$

   

   where ξ is the a priori signal-to-noise ratio, and where the bound of integration v is calculated in accordance with the model $$\upsilon =\gamma \frac{\xi }{1+\xi }$$

   

   where γ is the a posteriori signal-to-noise ratio.
6. Device for suppressing the late reverberation of a sound signal, characterized in that it includes means for:

   • capturing an input signal (x) formed by the superimposition of a multiplicity of delayed and attenuated versions of the sound signal,

   • applying a time-frequency transformation to the input signal (x) in order to obtain a complex time-frequency transform (X^C ) of the input signal (x),

   • generating a frequency-subsampled modulus (X̃) from the modulus of the complex time-frequency transform (X^C ) of the input signal (x),

   • generating a plurality of subsampled observation vectors from said frequency-subsampled modulus (X̃),

   • constructing a plurality of analysis dictionaries (D ^a) from the plurality of subsampled observation vectors,

   • calculating a plurality of prediction vectors (α) from the plurality of subsampled observation vectors and from the plurality of analysis dictionaries (D ^a), by minimizing, for each prediction vector (α), the expression ∥X̃v - D^aα∥₂, which is the Euclidean norm of the difference between the subsampled observation vector associated with said prediction vector (α) and of the analysis dictionary (D^a ) associated with said prediction vector (α) multiplied by said prediction vector (α), while taking into consideration the constraint ∥α∥₁ ≤ λ, according to which the norm 1 of said prediction vector (α) is less than or equal to a maximum intensity parameter of the late reverberation (λ),

   • generating a plurality of observation vectors from the modulus of the complex time-frequency transform (X^C ) of the input signal (x),

   • constructing a plurality of synthesis dictionaries (D^s ) from the concatenation of the plurality of observation vectors,

   • estimating a late reverberation spectrum (X ^ℓ) from the multiplication of the plurality of synthesis dictionaries (D^s ) by the plurality of prediction vectors (α),

   • filtering the plurality of observation vectors in order to eliminate the late reverberation spectrum (X^ℓ ) and to obtain a dereverberated signal modulus (Y).

Images