EP3987822B1 - Sound pickup device with improved microphone network - Google Patents

Description

The invention relates to acoustic capture equipment intended to be integrated into a building, for domestic use (domestic context - connected house) or professional use (company context).

For example, this equipment aims to capture the sounds present in a room in order to feed an ambient intelligence system composed of a set of sensors and actuators making it possible to control the parameters (for example temperature, light, or others) and the corresponding building devices (connected objects in particular such as a connected heating installation, connected lamps, etc.).

Capturing ambient sounds in this context raises several issues.

The sounds to be captured can be located anywhere in a room. It is not possible to know their position a priori and to position the sound recording equipment accordingly. It is therefore necessary to have capture equipment capable of covering the entire space in a homogeneous manner.

However, for reasons of cost and bulk, it is not possible to line the surfaces of the room with microphones. It is therefore also necessary to seek to minimize the total number of sensors.

The visual appearance of the part can also be a constraint parameter. The aesthetics of the room should not be degraded by a multitude of capture equipment. It is therefore necessary to favor discreet and compact capture equipment.

Today, acoustic capture solutions do not meet all of these constraints. This is an audio ambient intelligence issue.

Concerning connected objects in general typically equipped with audiovisual tracking (or “monitoring”) devices embedding a camera and microphones, the number of sensors is insufficient to provide wide coverage of acoustic pick-up. They are limited to nearby sound sources. At least for distant sources, the signal to noise ratio (due to ambient noise and reverberation) is unfavorable and does not allow a reliable analysis of the received signals.

There are also known voice assistants that today have good voice recognition performance in order to improve the quality of interaction with the user. They have a microphone array (often circular) so that they can focus the pickup on the source of interest (i.e. the user) by applying antenna processing (typically training methods of tracks or “beamforming”). This makes it possible to improve the quality of the signals captured, and to overcome interactions with surrounding noise and the room effect.

This type of solution is not satisfactory because it is optimized for a specific category of sources: voice signals, sources limited to a portion of space. It is not suitable for picking up wideband signals (or outside the voice bandwidth). In addition, voice assistants are generally placed at head height (on a table, typically) and their capture is degraded by the presence of noise sources in their vicinity (television, radio, etc.) and by furniture that impede the propagation of sound.

More generally, the microphone arrays that can be designed for the audio ambient intelligence context are typically of the linear or spherical type. The linear geometry is not optimal because it requires a large number of sensors for effective capture. In addition, this type of geometry (linear or spherical) requires placing the antenna in the middle of the room to take advantage of its omnidirectional coverage, which is incompatible with the discretion constraint of the equipment. On the other hand, by placing the acoustic antenna close to a wall, the geometry is sub-optimal in the sense that the microphones pointed towards the wall are useless, and can even be a source of disturbance (capture of unwanted reflections by example).

The invention improves the situation.

• une pluralité de capsules de microphone (par exemple des capsules électrostatiques ou piézoélectriques, des électrets ou des MEMS), réparties sur une portion P de sphère S circonscrite entre deux ou trois plans perpendiculaires entre eux, les trois plans se coupant entre eux en un point correspondant au centre de la sphère S, et les deux plans se coupant en une droite passant par le centre de la sphère S, et la portion P de sphère étant telle que P = n S/8, avec n=1,2,
• une unité de traitement reliée aux capsules pour recevoir les signaux captés par les capsules, ladite unité de traitement étant agencée pour :
  • * matricer les signaux selon une représentation ambisonique dans laquelle seules sont retenues les composantes ambisoniques associées à des harmoniques sphériques symétriques par rapport à au moins deux des plans précités, et
  • * traiter une matrice ainsi obtenue pour identifier au moins une source sonore dans un espace environnant la portion de sphère, et interpréter un signal sonore issu de cette source.

A sound recording device is proposed, comprising at least:

• a plurality of microphone capsules (for example electrostatic or piezoelectric capsules, electrets or MEMS), distributed over a portion P of sphere S circumscribed between two or three mutually perpendicular planes, the three planes intersecting with each other at a point corresponding to the center of the sphere S, and the two planes intersecting in a straight line passing through the center of the sphere S, and the portion P of the sphere being such that P = n S/8, with n=1.2,
• a processing unit connected to the capsules to receive the signals picked up by the capsules, said processing unit being arranged to:
  • * matrix the signals according to an ambisonic representation in which only the ambisonic components associated with spherical harmonics symmetrical with respect to at least two of the aforementioned planes are retained, and
  • * process a matrix thus obtained to identify at least one sound source in a space surrounding the sphere portion, and interpret a sound signal from this source.

Thus, such a device can be inserted for example in an upper corner of a room or between a wall and a ceiling, discreetly. In addition, one advantage of such an embodiment is that the number of capsules to be provided can be reduced, compared with what is usually required for an embodiment based on a solid sphere. In particular, the reflections of the ceiling and of the wall(s) are used here to limit the number of spherical harmonics to be taken into account and thus retain a limited number of ambisonic components. Indeed, the supposedly rigid walls induce a large number of zero components. Only the harmonics respecting the symmetry can be exploited.

In an embodiment where n=1 and the capsules are then distributed over an eighth of a sphere, the ambisonic components retained are associated with spherical harmonics symmetrical with respect to each of the three perpendicular planes and intersecting with each other at the center of the sphere S.

It is thus possible to select only the harmonics presenting such symmetries.

In such an embodiment, the device may further comprise a fixing support suitable for fixing the device in an upper corner of the room defined by two perpendicular walls and a ceiling overhanging the walls, the walls and the ceiling coinciding with the three perpendicular planes mentioned above and acting as reflective walls of sound waves.

As will be seen later with reference to the picture 3 , these reflections make it possible to consider virtual sources, mirrors of real sources, which can contribute to increasing the fineness of detection of a source for example. There are thus both virtual sources and virtual microphones which complement the real microphones and then constitute a complete sphere.

With an eighth of a sphere to consider, the ambisonic components retained are associated with spherical harmonics having a degree l and an order m (couples {l,m} of the picture 3 described later), such as:
l and m are even AND m greater than or equal to 0.

In such an embodiment, the number of ambisonic components retained is equal to (A+1)(A+2)/2 where A is the integer part of half of a maximum degree L of the spherical harmonics with which the ambisonic components are associated retained.

As will be seen in the embodiments presented below, the aforementioned maximum degree L is greater than 4 and preferably greater than 6.

In the embodiment where n=2 and therefore the capsules are distributed over a quarter of a sphere, the ambisonic components selected are associated with spherical harmonics symmetrical with respect to two perpendicular planes and intersecting each other in a straight line passing through the center of the sphere S.

In such an embodiment, the device may further comprise a fixing support suitable for fixing the device in a corner of a room defined by a wall and a ceiling, perpendicular to each other, the wall and the ceiling coinciding with said two perpendicular planes and acting as reflective walls of sound waves.

In either of the aforementioned embodiments (n=1 or 2), the capsules can be positioned on a spherical Gauss-Legendre mesh, and in this case, the device preferably comprises a number N of capsules given by:
N=2n/8 (L+1) ² (or N=n/4 (L+1) ² ), where L is a maximum degree of the spherical harmonics associated with the ambisonic components retained.

• b est une matrice vecteur contenant les composantes ambisoniques retenues,
• C est une constante réelle (par exemple C=8 dans le cas d'un huitième de sphère présenté plus loin),
• E est une matrice diagonale contenant des filtres d'égalisation radiale de chaque capsule,
• Y est une matrice contenant les harmoniques sphériques auxquels sont associées les composantes ambisoniques retenues, et
• G est une matrice diagonale contenant des poids d'intégration d'un maillage de Gauss-Legendre pour chacune des capsules,

s étant un vecteur contenant des signaux issus des capsules.In such an embodiment, the processing unit can be configured to decompose signals from the microphone capsules, on the spherical harmonics associated with the retained ambisonic components, using a matrixing of the type:
b = C EYGs, where:

• b is a vector matrix containing the retained ambisonic components,
• C is a real constant (for example C=8 in the case of an eighth of a sphere presented later),
• E is a diagonal matrix containing radial equalization filters of each capsule,
• Y is a matrix containing the spherical harmonics with which the retained ambisonic components are associated, and
• G is a diagonal matrix containing integration weights of a Gauss-Legendre mesh for each of the capsules,

s being a vector containing signals from the capsules.

In such an embodiment, the processing unit can also be configured to then weight the vector b by a deflection vector given in azimuth and elevation with respect to a reference frame defined by the center of the sphere S and the three intersections between the three planes. For example, it is possible to provide a sweep of this angle of the deflection vector to probe the various sources of a room.

In one embodiment, the device may comprise a plurality of portions of spheres P=n S/8, with n=1.2 (compact or separated by forming an installation, for example with several shells of sphere portions), each comprising a plurality of microphone capsules, distributed over each portion P of sphere S, and the processing unit is further arranged to process the signals coming from the capsules of each portion of sphere separately by matrixing, and to refine by cross-checking on the matrices thus obtained l identification of at least one sound source in a space surrounding the sphere portions.

Indeed, such an embodiment based on several sphere portions makes it possible to increase the signal-to-noise ratio by cross-checking the various processed signals originating from the capsules of these sphere portions. It is then typically possible to refine a source detection for example, or even to remove ambiguities, or to be able to take advantage of a better point of view (more exactly “listening point”) on the target source.

• les signaux captés par les capsules sont matricés selon une représentation ambisonique dans laquelle seules sont retenues les composantes ambisoniques associées à des harmoniques sphériques, symétriques par rapport à au moins deux des plans précités, et

* la matrice ainsi obtenue (typiquement un vecteur des composantes ambisoniques par exemple) est traitée pour identifier au moins une source sonore dans un espace environnant la portion de sphère, et interpréter un signal sonore issu de cette source. On peut ainsi focaliser par exemple l'écoute dans une direction donnée.The invention also relates to a method implemented by a processing unit of a device of the above type, in which:

• the signals picked up by the capsules are matrixed according to an ambisonic representation in which only the ambisonic components associated with spherical harmonics, symmetrical with respect to at least two of the aforementioned planes, are retained, and

* the matrix thus obtained (typically a vector of ambisonic components for example) is processed to identify at least one sound source in a space surrounding the sphere portion, and to interpret a sound signal from this source. It is thus possible, for example, to focus listening in a given direction.

Such a realization can be illustrated by way of example by the flowchart of the figure 6 , on which, following the obtaining of the signals from the capsules in step sa, these signals are matrixed to obtain the aforementioned vector b of the ambisonic components in step S1. This vector b can be weighted in step S2 by a deflection vector as presented above. Optionally, it is possible to provide in step S3 a processing of signals originating from several portions of sphere P to produce the weighted vectors b(A), b(B), etc. specific to each portion A, B, etc. Such an embodiment makes it possible to refine the detection of source(s) in step S4 for a better interpretation of the sound signal SIG coming from this (or these) source(s). Thus, it is possible for example in an embodiment where the device is used as a voice assistant to distinctly recognize a COM command in step S5.

The present invention also relates to a computer program comprising instructions for implementing the above method when this program is executed by a processor.

• une interface d'entrée IN pour recevoir les signaux issus des capsules,
• une mémoire MEM stockant au moins les données d'instructions d'un tel programme informatique au sens de l'invention,
• le processeur PROC apte à coopérer avec la mémoire MEM pour lire ces instructions et exécuter ainsi le procédé illustré à titre d'exemple sur la figure 6,
• et une interface de sortie OUT apte à délivrer par exemple le signal de commande COM interprété (ou alternativement le signal sonore issu de la source détectée, ou alternativement encore des signaux ambisoniques traités permettant d'identifier une source sonore générant le signal SIG).

Alternativement, la sortie OUT peut délivrer l'interprétation de(s) l'événement(s) sonore(s) (alarme, aboiement de chien, chute de personne, etc., ou toute autre situation caractérisée par les sons identifiés), et toute information associée à cet événement (localisation temporelle et/ou spatiale).This may typically be the processor PROC of a processing unit UT as illustrated by way of example on the figure 7 further comprising:

• an IN input interface to receive the signals from the capsules,
• a memory MEM storing at least the instruction data of such a computer program within the meaning of the invention,
• the processor PROC capable of cooperating with the memory MEM to read these instructions and thus execute the method illustrated by way of example on the figure 6 ,
• and an output interface OUT able to deliver for example the interpreted command signal COM (or alternatively the sound signal coming from the detected source, or alternatively also processed ambisonic signals making it possible to identify a sound source generating the signal SIG).

Alternatively, the OUT output can deliver the interpretation of the sound event(s) (alarm, dog barking, person falling, etc., or any other situation characterized by the identified sounds), and any information associated with this event (temporal and/or spatial location).

The present invention also relates to a non-transitory recording medium readable by a computer on which is recorded a program for the implementation of the above method when this program is executed by a processor.

As indicated previously, it may be the aforementioned memory MEM.

Brief description of the drawings

• Fig. 1
  [Fig. 1] montre des exemples de réalisation de portions de sphère.
• Fig. 2
  [Fig. 2] montre les directivités des harmoniques sphériques jusqu'au degré maximum L= 5, les deux nuances de couleur représentent les valeurs positives et négatives respectivement.
• Fig. 3
  [Fig. 3] illustre le principe d'une source et d'un microphone image dans le cas d'une réflexion acoustique (sur une paroi telle qu'un mur, un plafond).
• Fig. 4
  [Fig. 4] illustre un réseau de microphones réels sur une fraction de 1/8 sphère et des microphones images (grisés) générés grâce aux réflexions sur parois rigides.
• Fig. 5
  [Fig. 5] montre un exemple de formation de voies utilisant les harmoniques sphériques.
• Fig. 6
  [Fig. 6] montre un exemple d'ordinogramme définissant une succession d'étapes d'un procédé selon un mode de réalisation.
• Fig. 7
  [Fig. 7] montre un exemple de structure d'une unité de traitement UT d'un dispositif selon un mode de réalisation.

Other characteristics, details and advantages will appear on reading the detailed description below, and on analyzing the appended drawings, in which:

• Fig. 1
  [ Fig. 1 ] shows exemplary embodiments of sphere portions.
• Fig. 2
  [ Fig. 2 ] shows the directivities of the spherical harmonics up to the maximum degree L= 5, the two shades of color represent the positive and negative values respectively.
• Fig. 3
  [ Fig. 3 ] illustrates the principle of a source and an image microphone in the case of an acoustic reflection (on a wall such as a wall, a ceiling).
• Fig. 4
  [ Fig. 4 ] illustrates an array of real microphones on a fraction of 1/8 sphere and image microphones (shaded) generated thanks to reflections on rigid walls.
• Fig. 5
  [ Fig. 5 ] shows an example of channel formation using spherical harmonics.
• Fig. 6
  [ Fig. 6 ] shows an example of a flowchart defining a succession of steps of a method according to one embodiment.
• Fig. 7
  [ Fig. 7 ] shows an example of the structure of a processing unit UT of a device according to one embodiment.

Description of embodiments

Reference is now made to the figure 1 on which a device within the meaning of the DIS invention is in the form of a quarter of a sphere (upper part of the figure 1 ) or in the form of an eighth of a sphere (lower part of the figure 1 ). The surface of these sphere portions is meshed (in a chosen way which may correspond to the Gauss-Legendre spherical mesh as described later) and MIC microphone capsules are placed on this mesh in a number which can also be determined by the aforementioned Gauss-Legendre mesh. These MIC capsules are connected to a processing unit UT (visible on the upper part of the figure 1 ) to receive the sound signals picked up and process them by matrixing in Ambisonic representation as described in detail later.

• dans un coin supérieur d'une pièce (entre deux murs perpendiculaires et un plafond) pour un huitième de sphère tel qu'illustré au bas de la figure 1, ou encore
• sur une arête entre un mur et le plafond pour un quart de sphère tel qu'illustré en haut de la figure 1.

Moreover, as can also be seen on the figure 1 , the device DIS may also comprise a mounting support SUP to be fixed, for example:

• in an upper corner of a room (between two perpendicular walls and a ceiling) for an eighth of a sphere as shown at the bottom of the figure 1 , or
• on an edge between a wall and the ceiling for a quarter of a sphere as illustrated at the top of the figure 1 .

The invention thus proposes a capture device consisting of one or more elementary networks of PCM capsules which can be distributed for example in a building room. The geometry of an elementary network is a fraction of a sphere (1/8 or 1/4) which is naturally inserted into the upper corners of a room so as to match the architecture, or even on an edge of intersection between a ceiling and a wall, in order to take advantage of reflections on such walls. The set of capture systems obtained is thus very discreet, makes it possible to considerably reduce the number of microphones while maintaining high directivity, and offers wide coverage of ambient sounds in the room. In fact, the microphones being located high up, they benefit from a privileged pick-up point over the entire room without being bothered by furniture or nearby users.

Even if the positioning in height improves the coverage of the part, it is necessary to envisage that only one network cannot cover all the part. In particular if the latter has a complex geometry (presence of nooks, areas of acoustic shadow without direct wave), it is preferable to have several networks. One embodiment then relates to a processing jointly exploiting the information coming from the various networks of sensors to acquire a reliable and complete representation of the sound scene captured. Obtaining several results of the presence of possible sound source(s) makes it possible to cross-check this information and thus ultimately improve a signal-to-noise ratio of source(s) detection.

Moreover, the choice of a spherical geometry is advantageous in the sense that it makes it possible to obtain (by associating the microphones with an appropriate processing of antenna signals) a high directivity with a small number of sensors. Indeed, in the case of a spherical geometry, the processing of the antenna signals uses spherical harmonic functions in a so-called “ambisonic” context. In the case where one is limited to a fraction of a sphere, the conventional harmonic functions cannot be applied directly and they should be adapted to the geometry chosen for the array of microphones, according to one embodiment.

In addition, the choice of the positions of the microphones on the sphere fraction is to be optimized. The optimal mesh must satisfy the best compromise between the number of sensors (to be minimized) and the quality of the information captured (which imposes a minimum number of sensors). This is a spatial sampling problem to be fitted to a fraction of a sphere.

The family of spherical harmonics forms a base. Each spherical harmonic is described by its degree l and its order m. At degree l, there is (2l+1) spherical harmonics. Up to the maximum degree L, there are (L+1) ² harmonics. In an Ambisonic context, a spherical array of microphones is usually used to decompose a sound pressure field on the basis of spherical harmonics, a representation of which is then illustrated on the figure 2 . Each line of the picture 2 is relative to a degree l and the representation up to degree L which includes all the components up to this degree. Thus, for the degree l=0 we have only one component. For degree l=1, we have 1 (first line) + 3 (second line) = 4 ambisonic components. For the degree l=2, we have 9 components, etc.

As a general rule, if the network is designed to perform a decomposition up to the maximum degree L of the Ambisonic components), it must be able to estimate Q = (L+1) ² components. For an accurate decomposition, the number of microphones, N, must be greater than or equal to the number Q of components to be estimated.

There figure 2 presents the directivities of the spherical harmonics up to the maximum degree L = 5. They are arranged in a pyramid in ascending order of degree l and order m: {l; m}.

For the implementation of the embodiment described here, only the components of the harmonics having a symmetry with respect to a plane of reflection of the sound wave (a wall or the ceiling) are retained. We note these different planes Oxy (the ceiling), Oxz (a wall) and Oyz (another wall in the case where 1/8th of a sphere is used, rather than a quarter of a sphere).

• la pression rayonnée par la source sans la paroi, et
• la pression issue de la réflexion sur la paroi rigide.

The reason for this selection of components is explained as follows, with reference to the picture 3 . The situation to the left of the picture 3 where a source (for example an HP loudspeaker) and a sensor (a MIC microphone) are placed close to an acoustically rigid wall (referenced MUR on the picture 3 ). The sound pressure at the sensor is the sum of:

• the pressure radiated by the source without the wall, and
• the pressure resulting from the reflection on the rigid wall.

It is also possible to mathematically solve the equations related to this configuration by removing the wall and adding a source and a microphone image, symmetrical with respect to the wall, as shown on the right side of the picture 3 . These are then “acoustic images”, the wall acting as a “mirror” of the sound wave.

The pressure received by the image sensor is assumed to be the same as that received by the real sensor without the wall.

• m supérieur ou égal à 0 ET m est pair, OU
• m < 0 ET m est impair
• (et présentant donc une symétrie par rapport au plan Oyz) soient déjà une première sélection des harmoniques dont les composantes sont retenues.

Symmetry with respect to the Oyz plane (a wall typically) dictates that spherical harmonics of degree l and order m such that:

• m greater than or equal to 0 AND m is even, OR
• m < 0 AND m is odd
• (and therefore presenting a symmetry with respect to the Oyz plane) are already a first selection of the harmonics whose components are retained.

• la somme l + m est paire
• (et présentant donc une symétrie par rapport au plan Oxy) soient ensuite une deuxième sélection des harmoniques dont les composantes sont à retenir.

Moreover, the symmetry with respect to the plane Oxy (the ceiling typically) imposes that the spherical harmonics of degree l and order m such that:

• the sum l + m is even
• (and therefore having a symmetry with respect to the Oxy plane) are then a second selection of the harmonics whose components are to be retained.

• m supérieur ou égal à 0 ET m est pair OU m < 0 ET m est impair
• ET (l + m) est pair.

Thus, for a quarter of a sphere (inserted in an intersection between two planes), the conditions can be:

• m greater than or equal to 0 AND m is even OR m < 0 AND m is odd
• AND (l + m) is even.

Of course, this is an embodiment where the device is fixed between a wall and the ceiling, such as for example the Oxy and Oyz planes. It can also be fixed between two walls Oyz and Oxz and the symmetry condition m greater than or equal to 0, which is specific to Oxz, should be added to the previous condition relating to Oyz (m greater than or equal to 0 AND m is even, OR m < 0 AND m is odd),
which ultimately amounts to m greater than or equal to 0 AND m is even.

In any case, we find the same number of spherical harmonics to retain, whatever the two planes of symmetry chosen.

• m supérieur ou égal à 0
• (et présentant donc une symétrie par rapport au plan Oxz) soient, avec les conditions ci-dessus, les harmoniques dont les composantes sont retenues.

For an eighth of a sphere, it is possible to take into account in addition the symmetry with respect to the Oxz plane (another wall typically), which imposes that the spherical harmonics of degree l and order m such that:

• m greater than or equal to 0
• (and therefore having a symmetry with respect to the Oxz plane) are, with the above conditions, the harmonics whose components are retained.

These conditions for an eighth of a sphere can be summarized, in fine, as follows:
l is even AND m greater than or equal to 0 AND m is even.

For a maximum degree fixed and noted L, the total number of harmonics respecting the symmetries with respect to the planes Oxy, Oxz, Oyz jointly is given by: $$\tilde{Q}=\frac{\left(\lfloor \frac{I}{2}\rfloor +1\right)\left(\lfloor \frac{I}{2}\rfloor +2\right)}{2}$$

[L/2] designating the integer part of L/2.

Thus, following a reasoning of acoustic images (as seen previously with reference to the picture 3 ), it is possible to use a fraction of a sphere 1/8 or 1/4 (or even possibly 1/2 but without real interest for an application in a building as presented above), and to place acoustically rigid walls in the appropriate plans to generate image microphones. It is then possible to use the resulting spherical network of microphones to decompose on the basis of the spherical harmonics still represented in this configuration, that is to say those respecting the conditions stated above on l and m. Furthermore, the image microphones receive the same pressure as the corresponding real microphones. Consequently, during the projection, the components on the spherical harmonics which do not respect the symmetries above (conditions on l and m) are considered null. For example, on the figure 2 , up to the maximum degree L=5, there are only six spherical harmonics which respect these conditions and which are symmetrical with respect to the planes Oxy, Oxz, Oyz jointly and it would then suffice at least to N = 6 microphones on 1/8 of a sphere (in baffle) to be able to estimate the components of the acoustic field on these harmonics.

In the context of portions of sphere with reflections, it is chosen in particular to carry out a mesh as illustrated on the figure 4 , known as the “Gauss-Legendre spherical mesh”, which gives the number and position of the microphones on a sphere to estimate the decomposition up to a chosen maximum degree L. By choosing L odd, the resulting mesh respects the symmetries with respect to the plans Oxy, Oxz, Oyz jointly. For example, the figure 4 shows a mesh with N = 72 microphones, capable of performing a precise decomposition up to the maximum degree L = 5 (with N=2(L+1) ² to respect the aforementioned Gauss-Legendre mesh which imposes twice the number of capsules, minimal, required (L+1) ² ).

Here, using only the nine microphones (nine points shown in a different shade on the figure 4 ) and using the shaded walls in the figure, it is possible to generate sixty-three image microphones. Because of the symmetries, only six components are non-zero here.

• b est un vecteur contenant les composantes ambisoniques associées aux harmoniques sphériques respectant les symétries précitées,
• E est une matrice diagonale (carrée) contenant des filtres d'égalisation radiale de chaque microphone,
• Y est une matrice (non carrée car traitant plus de signaux issus des capsules que de composantes ambisoniques en sortie) contenant les harmoniques sphériques respectant les symétries précitées évaluées aux différentes directions des microphones, et
• G est une matrice diagonale (carrée) contenant des poids d'intégration de la quadrature de Gauss-Legendre pour chacun des microphones du huitième de sphère,

s étant un vecteur contenant les signaux issus des microphones.As illustrated on the figure 5 , the signals of the microphones S1, S2, ..., SN, are decomposed (for example in the frequency domain) on the spherical harmonics, using an equation of the type:
b = 8 EYGs, where:

• b is a vector containing the ambisonic components associated with the spherical harmonics respecting the aforementioned symmetries,
• E is a diagonal (square) matrix containing radial equalization filters of each microphone,
• Y is a matrix (not square because processing more signals from the capsules than ambisonic components at the output) containing the spherical harmonics respecting the aforementioned symmetries evaluated at the different directions of the microphones, and
• G is a diagonal (square) matrix containing Gauss-Legendre quadrature integration weights for each of the eighth-sphere microphones,

s being a vector containing the signals from the microphones.

Such a realization amounts to applying a spherical Fourier transform (referenced SFT on the figure 5 ).

For channel formation (or "beamforming") in the field of spherical harmonics in order to identify one or more sound sources in a space surrounding the sphere portion, and thus interpret a sound signal from this source, the spherical harmonic components are first estimated using the matrix equation above. The obtained vector b is then weighted by a steering vector (or “steering vector”) which makes it possible to describe listening in a steering direction. Finally, the weighted components are summed to obtain the output signal.

One can predict weights w _lm for a regular directivity function given by the following equation: $$w_{\mathit{lm}}=Y_{\mathit{lm}}\left(\mathrm{<figure-callout\; id="0"\; label="theta"\; filenames="imgf0004.png,imgf0005.png"\; state="\{\{state\}\}">theta</figure-callout>}0,\mathrm{<figure-callout\; id="0"\; label="phi"\; filenames="imgf0004.png,imgf0005.png"\; state="\{\{state\}\}">phi</figure-callout>}0\right)$$

An example steering angle can be such that theta0 and phi0 are 45 and 135° respectively (pointing in this example towards the inside of the room). These respective azimuth and elevation coordinates are given relative to the base formed by the intersections of the three planes Oxy, Oxz, Oyz.

For the example of the eighth of a sphere, the directivity function obtained is the superposition of eight directivity functions of a complete sphere pointing in directions symmetrical with respect to the planes Oxy, Oxz, Oyz jointly. This superposition can however present itself as a disadvantage for low degrees L (L<6) and L=7 can present itself as a good compromise between the number of capsules on the one hand and the quality of the decomposition on spherical harmonics.

In this case, provision is usually made for at least N=(L+1) ² capsules for correct capture quality, ie N=64. Nevertheless, on an eighth of a sphere only, it is advisable to divide by 8 this number, that is to say the useful number N=8.

Nevertheless, to respect the aforementioned spherical Gauss-Legendre mesh, it is necessary to multiply this number N by 2, so that in the example embodiment mentioned above at L=7, provision may preferably be made for N=16 capsules or more.

In this case, as indicated above, the number of Ambisonic components retained is Q=(3+1) (3+2)/2 = 10.

• une prise de son homogène sur l'ensemble d'une pièce,
• la capacité d'extraire une source sonore dans une direction donnée grâce au traitement de signaux d'antenne (débruitage et déréverbération pour améliorer le rapport signal-utile-à-bruit),
• un dispositif résultant de cette conception qui se présente compact et discret, intégré et s'adaptant à la configuration d'une pièce classique.

Thus the invention combines the following advantages:

• a homogeneous sound recording over the whole of a room,
• the ability to extract a sound source in a given direction thanks to the processing of antenna signals (denoising and dereverberation to improve the signal-useful-to-noise ratio),
• a device resulting from this design which is compact and discreet, integrated and adapting to the configuration of a classic room.

• la domotique utilisant des objectés connectés notamment pour un système d'intelligence ambiante audio permettant à partir de l'analyse et reconnaissance des sons ambiants d'inférer des actions et de proposer des services aux habitants d'une maison ou de personnes d'une entreprise (potentiellement applicable à n'importe quel lieu de vie) ;
• les assistants vocaux avec un dispositif de captation de sons ambiants possiblement utilisé pour capter la voix des utilisateurs et ainsi alimenter un assistant vocal ;
• les systèmes de surveillance audio pour la détection d'intrusions (bris de glace), des alarmes, les bruits de chute de personnes, ou autres.

The invention finds numerous applications, in particular in:

• home automation using connected objects in particular for an audio ambient intelligence system allowing from the analysis and recognition of ambient sounds to infer actions and offer services to the inhabitants of a house or people of a company (potentially applicable to any place of life);
• voice assistants with an ambient sound capture device possibly used to pick up the voice of users and thus feed a voice assistant;
• audio surveillance systems for detecting intrusions (glass breakage), alarms, the noise of people falling, or others.

Claims (15)

1. Sound pickup device, comprising at least:

   - a plurality of microphone capsules, distributed over a portion P of a sphere S circumscribed between two or three planes perpendicular to one another, the three planes intersecting one another at a point corresponding to the centre of the sphere S, and the two planes intersecting at a straight line passing through the centre of the sphere S, and the sphere portion P being such that P = n S/8, with n=1,2,

   - a processing unit connected to the capsules in order to receive the signals picked up by the capsules, said processing unit being designed to:

   * matrix the signals according to an ambisonic representation in which only ambisonic components associated with spherical harmonics symmetrical about at least two of the abovementioned planes are selected, and

   * process a matrix thus obtained in order to identify at least one sound source in a space surrounding the sphere portion, and interpret a sound signal originating from this source.
2. Device according to Claim 1, wherein, for n=1, the capsules being distributed over an eighth of a sphere, the selected ambisonic components are associated with spherical harmonics symmetrical about each of the three perpendicular planes intersecting one another at the centre of the sphere S.
3. Device according to Claim 2, furthermore comprising a fastening support designed to fasten the device in an upper corner of a room defined by two perpendicular walls and a ceiling overhanging the walls, the walls and the ceiling coinciding with said three perpendicular planes and acting as sound wave-reflecting walls.
4. Device according to either of Claims 2 and 3, wherein the selected ambisonic components are associated with spherical harmonics having a degree 1 and an order m such that:
   1 and m are even AND m is greater than or equal to 0.
5. Device according to Claim 4, wherein the number of selected ambisonic components is greater than or equal to (A+1) (A+2) /2, where A is the integer part of half a maximum degree L of the spherical harmonics with which the selected ambisonic components are associated.
6. Device according to Claim 5, wherein the maximum degree L is greater than 4, and preferably greater than 6.
7. Device according to Claim 1, wherein, for n=2, the capsules being distributed over a quarter of a sphere, the selected ambisonic components are associated with spherical harmonics symmetrical about two perpendicular planes intersecting one another at a straight line passing through the centre of the sphere S.
8. Device according to Claim 7, furthermore comprising a fastening support designed to fasten the device in a corner of a room defined by a wall and a ceiling that are perpendicular to one another, the wall and the ceiling coinciding with said two perpendicular planes and acting as sound wave-reflecting walls.
9. Device according to one of the preceding claims, wherein the capsules are positioned on a Gauss-Legendre spherical mesh, and the device comprises a number N of capsules given by N = 2n/8 (L+1)², wherein L is a maximum degree of the spherical harmonics associated with the selected ambisonic components.
10. Device according to Claim 9, wherein the processing unit is configured to decompose signals originating from the microphone capsules, on the spherical harmonics associated with the selected ambisonic components, using matrixing of the type:
    B = C EYGs, where:

    - b is a vector matrix containing the selected ambisonic components,

    - C is a real constant,

    - E is a diagonal matrix containing radial equalization filters of each capsule,

    - Y is a matrix containing the spherical harmonics with which the selected ambisonic components are associated, and

    - G is a diagonal matrix containing integration weights of a Gauss-Legendre mesh for each of the capsules,

    s being a vector containing signals originating from the capsules.
11. Device according to Claim 10 taken in combination with Claim 6, wherein the processing unit is furthermore configured to then weight the vector b with a steering vector given in azimuth and in elevation with respect to a reference system defined by the centre of the sphere S and the three intersections between the three planes.
12. Device according to one of the preceding claims, comprising a plurality of sphere portions P = n S/8, with n=1,2, each comprising a plurality of microphone capsules, distributed over each portion P of a sphere S, and wherein the processing unit is furthermore designed to process the signals originating from the capsules of each sphere portion separately through matrixing, and to refine, by cross-checking the matrices thus obtained, the identification of at least one sound source in a space surrounding the sphere portions.
13. Method implemented by a processing unit of a device according to one of the preceding claims, wherein:

    - the signals picked up by the capsules are matrixed according to an ambisonic representation in which only ambisonic components associated with spherical harmonics symmetrical about at least two of the abovementioned planes are selected, and

    * the matrix thus obtained is processed in order to identify at least one sound source in a space surrounding the sphere portion, and interpret a sound signal originating from this source.
14. Computer program comprising instructions for implementing the method according to Claim 13 when this program is executed by a processing unit of a device according to one of Claims 1 to 12.
15. Non-transient computer-readable recording medium on which there is recorded a program for implementing the method according to Claim 13 when this program is executed by a processing unit of a device according to one of Claims 1 to 12.

Images