Locating a moving acoustic source

Abstract

The invention relates to processing sound signals acquired by at least one microphone, in order to locate at least one sound source, emitting from a plurality of discrete positions at respective discrete points in time (k, k'), in a space comprising at least one reflective, planar, wall, comprising: - in order to obtain at least: * a first vector (I), determining a direction (DoA) of a direct first acoustic path between the source and the microphone, * at least a second vector (II) representing a second acoustic path from at least one specular reflection and arriving at the microphone, * at least one delay (III) of the second path at the microphone relative to the direct path, - exploiting at least one property of the specular reflection according to which a Euclidian distance between two positions of the source at two discrete points in time is equal to a Euclidean distance between two respective image positions of the source and deriving from one or a plurality of same reflections, respectively, at the two discrete points in time, in order to determine at least one position (d(k), d(k')) of the source relative to the microphone, respectively, at the plurality of discrete points in time (k, k'), on the basis of, for each point in time k: * the first vector (I) to determine a direction (DoA) of the direct path, and * both the delay (III) and the second vector (II) to associate a distance d(k) between the source and the microphone with said direction (DoA) of the direct path.

Description

Description Title: Localization of a moving acoustic source

[0001] This description relates to the field of acoustic source localization, in particular for the estimation of acoustic direction or “DoA” (Direction of Arrival) by a compact microphone system (for example a microphone capable of capturing sounds in “ambiophonic” or “ambisonic” representation hereinafter).

[0002] One possible application is, for example, beamforming, which involves spatial separation of audio sources, in particular to improve speech recognition (for example, for a virtual assistant using voice interaction). Such processing can also be used in 3D audio coding (pre-analysis of a sound scene to code the main signals individually), or to enable spatial editing of immersive sound content, possibly in audiovisual applications (for artistic, radio, cinema, etc.). It also enables speaker tracking in teleconferencing or the detection of sound events (with or without associated video).

[0003] Document WO-2021 / 074502 has proposed an approach using the sound velocity vector to obtain in particular the direction of arrival of the sound, its delay (therefore the distance to the source), as well as the delays linked to possible reflections on the walls of a room and the determination of the positions of such walls (possibly partitions such as walls, the floor, the ceiling, but also reflective walls such as tables, screens, etc.). Such an implementation makes it possible to model the interference between the direct wave and at least one indirect wave (resulting from reflection) and to exploit the manifestations of this model on the entire velocity vector (on its imaginary part as well as on its real part).

[0004] An improvement of this approach was proposed in document FR2011874 using a modified velocity vector, called “generalized”, and constructed from the classic velocity vector which is generally expressed as a function of a component in the denominator which is omnidirectional. The generalized velocity vector then replaces the classic velocity vector in the sense of document WO- 2021 / 074502, with then a component in the denominator which is different from an omnidirectional component. This different component can in fact be more “selective” towards the direction of arrival of the sound.

[0005] In an embodiment presented in these documents, it is possible to obtain (from an ambisonic sensor for example) a succession of peaks characterizing an intensity or an acoustic energy, and each linked to a reflection on at least one wall, in addition to a peak linked to an arrival of the sound along the direct path (DoA) of the sound from the source.

[0006] However, in certain application cases where the sound source may be mobile, a robust method is sought for determining the distance between the source and the microphone, as the source moves, particularly when the precise orientation of the wall(s) causing the reflection(s) at a given time is not known a priori.

[0007] This description improves this situation.

[0008] To this end, it proposes to rely in particular on the reflections of the walls, at different distinct moments.

[0009] It then aims at a method for processing sound signals acquired by at least one microphone, for locating at least one sound source, emitting from a plurality of distinct positions at respective distinct times (k, k'), in a space comprising at least one plane reflective wall, the method comprising: - obtain at least, for each instant k: * a first vector determining a direction (DoA) of a first path acoustic, direct between the source and the microphone, * at least one second vector representing a second acoustic path originating from at least one specular reflection, and arriving at the microphone, * at least one delay that the second path has, at the microphone, compared to the direct route, - exploit at least one property of the reflection, specular, according to which a Euclidean distance between two positions of the source at two distinct times equals a Euclidean distance between two positions of respective images of the source and resulting from one or more same reflections, respectively at said two distinct times, to determine at least one position (d (k) , d ( k') ) of the source relative to the microphone respectively at said plurality of distinct instants (k, k'), as a function, for each instant k: * of the first vector to determine a direction (DoA) of the direct path, and * both late and the second vector, to associate a distance d (k) between the source and the microphone at this direction (DoA) of the direct path.

[0010] “At least one sound source, emitting from a plurality of distinct positions at respective distinct times” means a source that may be mobile and thus occupy these distinct positions at these respective times. Alternatively, it may be several sources having these respective distinct positions.

[0011] By "at least one wall" we mean possibly a set of walls parallel or forming any angle between them (two by two). Thus, "said at least one reflection" may possibly concern a plurality of successive reflections on the walls of this set.

[0012] It is then demonstrated below that, if the acoustic reflections in play can be considered as specular and if the walls considered are planar, then the aforementioned property of conservation of Euclidean distances (illustrated in figure 3a) makes it possible to obtain the distance from the source to the microphone at different times k and k', and this from the observations: - directions of arrival of the source at the microphone at these different instants (S1, S2 in Figure 3a), as well as - directions of arrival at the microphone of the images (S1 (w2) , S2 (w2) by example in figure 3a) resulting from the same reflections (on the wall (w2) for example in figure 3a) and respectively associated with the positions (S1, S2) of the source at different times k and k', and the arrival delays respectively associated and

[0013] Obtaining these observations for different instants k, k', etc., and possibly exploiting several reflections for the same instant k (for example individual reflections on different walls or successively on several walls), allows for example, as presented later in an embodiment, to obtain a system with several equations whose solutions are the distances d (k) , d (k') ... between each position of the source at a time k, k'... and the microphone.

[0014] Thus, it is possible to collect a number of observations at these different times, sufficient to resolve such a system.

[0015] In one embodiment, the method may further comprise: - exploit, in addition to said property of specular reflection, a second geometric property according to which a projection on a chosen axis of said Euclidean distance between two positions of the source at two distinct times corresponds to a projection on the same chosen axis of the Euclidean distance between two positions of respective images of the source and resulting from one or more of the same reflections, respectively at said two distinct times.

[0016] This conservation of the projection is illustrated in Figure 3b. Combined with the property of conservation of Euclidean distances, it allows us to obtain even more equations and thus perfect the determination of the distances d (k) , d (k') ...

[0017] However, it can impose geometric conditions which are not really restrictive in practice.

[0018] For example, the aforementioned chosen axis is parallel or perpendicular to said at least one wall.

[0019] For example, the microphone is also of the ambisonic type, and preferably arranged so that the height z axis of the microphone is parallel to the chosen axis.

[0020] These geometric conditions simply amount to considering that the microphone is placed on a wall such as a table for example (therefore a wall horizontal, perpendicular to the microphone's z-axis), in a space surrounded by walls such as walls parallel to the z-axis (but not necessarily parallel to each other), and with typically a floor and a ceiling as other walls, which are then perpendicular to the z-axis.

[0021] As previously indicated, the exploitation of the specular reflection property, combined with the exploitation of the second geometric property (of projection on the z axis), can generate a system of equations in which the positions of the source relative to the microphone, for different instants k, k', are the unknowns. In particular, this system of equations can, in general, be overdetermined (with therefore more equations than unknowns).

[0022] Regarding the consideration of the different instants k, k', etc., the sound signals can be acquired according to a succession in time of frames, and the first vector, the second vector and the delay can be obtained for a plurality of frames corresponding respectively to distinct instants (k, k').

[0023] In particular, it is possible to isolate "the good frames", the most useful for obtaining these parameters and determining, for example, a movement of the source between the different instants corresponding to these frames.

[0024] To obtain these parameters, different embodiments can be provided. Of course, the expression of the velocity vector can be used (as described in the documents presented above). However, other techniques can be used, for example the one presented in: Sakari Tervo, Jukka Pätynen, Antti Kuusinen, Tapio Lokki: “Spatial Decomposition Method for Room Impulse Response”, Journal of the Audio Engineering Society, Vol.61, No.1 / 2, 2013.

[0025] In this document, the parameters are derived from room impulse responses (RIRs), recorded by a antenna of simply co-located microphones (without even using ambiophonics here). It will thus be understood that a specifically ambisonic microphone is not necessary for the capture of sounds, on the one hand, and that the present description is not limited to the exploitation of the velocity vector to obtain the aforementioned parameters.

[0026] However, in an implementation where a velocity vector is used (and more particularly a generalized velocity vector within the meaning of document FR2011874 for better results in general), at least one parameter among the first vector , the second vector and the delay can be obtained from the expression of this velocity vector (generalized), the process then comprising: - apply a time-frequency transform to the acquired signals, - from the acquired signals, express in the frequency domain a generalized velocity vector, for a plurality of distinct instants (k, k'), each generalized velocity vector for a given instant k characterizing a composition between: * the first acoustic path, direct between the source and the microphone, represented by the first vector, and taking a delay between the emission of a sound by the source and reception of this sound by the microphone, and * at least the second acoustic path, represented by the second vector and having the delay at the microphone, compared to the direct path.

[0027] Typically, the DoA of the source (i.e., the first vector) can be obtained by a technique other than that using the velocity vector. To obtain the delays, it is nevertheless more comfortable to use the expression in the time domain of the velocity vector as follows.

[0028] For this purpose, the process may include: - further apply an inverse transform, from frequencies to time, to the (generalized) velocity vector to obtain, in the time domain, at least one peak linked to one or more reflections on one or more walls, in addition to a peak linked to an arrival of the sound along the direct path (DoA), the peak linked to one or more reflections being offset by the delay relative to the peak linked to the arrival of the sound according to the path direct.

[0029] The peaks in Figure 2 illustrate these time shifts (τ1, τ2, etc.) relative to the deadline between the emission of a sound by the source and the reception of this sound by the microphone.

[0030] Examples of equation formalisms that the present method proposes to solve in embodiments are presented below. We consider below: - a vector between the source and the microphone, at an instant k, written in function of the first vector, where d (k) is the distance Euclidean at time k between the source and the microphone, - a vector between an image of the source and the microphone, at an instant k, written as a function of the second vector : , with . where c is the speed of sound.

[0031] Then the aforementioned property of specular reflection translates, for two distinct instants k and k', by an expression of the type:

[0032] This expression can be developed into: with : where the notation<x,y> denotes the scalar product between two vectors x and y.

[0033] Then, the second geometric property mentioned above (of conservation of the projection on the z axis) is translated by an expression of type; where is a unit vector parallel to the aforementioned chosen axis, and this expression expands to: Or : and denotes a scalar product of type .

[0034] Then, the respective developments of the expressions can generate a system of bi-affine equations of type: whose variable d is a column vector having coefficients corresponding to the distances between the source and the microphone at different times 1, 2, ..., K: and where the vtriu dd operator T extracts coefficients from a diagonal and above the diagonal of the matrix dd T by concatenating them into a column vector.

[0035] This system of bi-affine equations can be solved by nonlinear minimization of a cost function l(.), given by: knowing that where Ib and ub are lower and upper limits given to the distances d (k) .

[0036] A regularization term λr(d) can be added to the term ℓ(.) to express the cost function globally as follows: Such an expression with the term λr(d) advantageously allows to adjust at least one smoothing structure applied to the coordinates of the vector d (one can thus "smooth" the form of displacement of the source between two points, or on the contrary wish to preserve a form of jerky movement for example).

[0037] Furthermore, a diagonal weighting matrix diag(ψ ) can be applied to the cost function as follows: , which amounts to weighting the different equations of the system Mf + q, for example to favor the weight of observations at a given time compared to other observations at another time.

[0038] The present invention also relates to a computer program comprising instructions for implementing the above method, when these instructions are executed by a processor of a processing circuit. It also relates to a non-transitory recording medium, readable by a computer, on which such a program is recorded.

[0039] It also relates to a computer device comprising a processing circuit configured to implement the above method.

[0040] Other features, details and advantages will become apparent upon reading the detailed description below, and upon analyzing the attached drawings, in which:

[0041] Figure 1 [Fig. 1] shows, for illustrative purposes, different parameters that can be used to estimate the distance from a source to the microphone.

[0042] Figure 2 [Fig. 2] shows as an illustration the different successive peaks that the temporal expression of a (generalized) velocity vector presents after its inverse transform from frequencies to time.

[0043] Figure 3a [Fig. 3a] illustrates the aforementioned Euclidean distance conservation property.

[0044] Figure 3b [Fig. 3b] illustrates the second property, mentioned above, of conservation of the projection on the chosen axis.

[0045] Figure 4 [Fig. 4] schematically shows a device for implementing the above method, according to one embodiment.

[0046] As briefly presented above, the present description proposes to estimate the distance and the direction of arrival (DoA) of a source at different times, for example for a moving source. The processing relating to the determination of the DoA is not impacted by the mobility of the source. However, it is more difficult to determine by conventional methods of the state of the art the distance from the source to the microphone for a moving source (without knowing a priori the orientation of the partitions which surround it).

[0047] It is proposed in the detailed implementation set out below to use the velocity vector as in the approach of the prior art documents presented above, and we first estimate the footprint of the reflections on the generalized velocity vector, hereinafter noted “GTVV” (for “Generalized Time Domain Velocity Vector”), using the processing described in particular in document FR2011874. A slightly different weighting can be carried out than that described in FR2011874, in the time-frequency domain before calculating the GTVV vector, as detailed later in a possible optional example but the principle described in this document remains the same.

[0048] The DoA of the source is estimated by observing the GTW vector at time t=0 (as in the aforementioned document). The DoAs and the relative delays of some acoustic reflections are then detected by selecting only a part of the peaks on a sequence from the GTW vector (for example, its norm as a function of the delay, or possibly this norm multiplied by the sign of the omnidirectional component). Of course, this implementation is simple but can admit more sophisticated variants to carry out the inference of the parameters. We note here that it is then generally possible to have a set of estimates of the DoAs over time and their associated delays, for example per signal frame (possibly not for all the frames, but for some, at least, which allow the continuation of the processing).

[0049] The identification of peaks in the succession of peaks can be carried out in practice using two different processes, but which can be combined. A first process searches for the DoA of the sound source itself. This process is the easiest, because it does not require identifying the reflection peaks, their number and the walls at the origin of these reflections. The second process offers a "complete" analysis of the peaks resulting from the reflections. The second process is then dedicated to the tracking of multiple image sources, based purely on the observed DoAs, and adapted to also record and use the observed relative delays. Here, "image sources" are understood to be the virtual sources generated by the reflections on the walls. The result of this second process gives the time sequences of the pairs (DoA, delay), with labels corresponding to the estimated individual reflections (i.e. the "trajectories" of the reflections).By exploiting the delays linked to reflection, the implementation of the second processing allows us to subsequently estimate a distance d between the source and the microphone, to be associated with the DoA of the source, and this at different distinct instants k, k', etc. For this purpose, we exploit geometric properties presented later. We can. finally construct a position vector of the source relative to the microphone from the DoA of the source and the distance separating the source from the microphone.

[0050] In one embodiment, it is indeed possible to construct a system of bi-affine equations (whose variable is the vector of distances), thanks to the estimated trajectories of the reflections and the source at different times. This system is generated by applying acoustic principles relative to geometric conditions of sound propagation in space.

[0051] Then, we minimize a cost function based on this system of equations, such as the sum of squared (or absolute) residuals. This is a non-linear and non-convex minimization: we can use known methods but adapted to the use case (for example, an accelerated (sub-)gradient descent, or others).

[0052] The GTW vector representation can thus be leveraged to estimate at least the DoA and the source-to-microphone distance for a moving sound source without a priori knowledge of the wall orientations. The trajectories of a moving source and the corresponding reflections are spatially and temporally related, which can be used to infer the absolute delay of the propagating source signal and, therefore, approximate the microphone-to-source distance.

[0053] The following objective is to exploit the GTW vector footprint to estimate the 3D position of a moving sound source, without a priori assumptions about the orientations of the reflecting walls. Since the source is mobile, the inference must be made, for example, for each frame of a succession of frames. By "frames" here, we mean packets of sound data acquired by a microphone (e.g., ambisonic), at distinct times. Thus, each frame acquired at a time k gives a sound image from which the current position of the moving source at this time k can be derived. Several frames acquired at distinct times k, k, etc., should make it possible to determine the movement of the source during these times k, k', etc.

[0054] The mobile source thus provides a "spatial diversity" which is then exploited to cope with the unknown geometry of the acoustic environment. Of course, this approach can be taken in a similar way in the case where the source is fixed while the microphone is mobile, due to the symmetry of the acoustic wave equations.

[0055] We recall here equation 39 of document WO-2021 / 074502, giving the expression of the velocity vector noted depending on the aforementioned peaks, linked to reflections and marked by Dirac functions positioned at delays τ n (let δ(t - kτ n )) relative to the first peak of abscissa τ0 corresponding to the arrival of the sound along the direct path. The other SARC terms are specific to reverberations and cross reflections and are not considered:

[0056] Eq.39

[0057] Figure 1 recalls the orientation of the vectors (associated with the source) and the vector associated with the reflection on a reflective wall such as the ground in the illustrated example. It should be noted that the vector can be more particularly associated with an image source (in thin lines, relative to the original source illustrated in strong lines) which is arranged on the “other side” of the reflecting wall. The original source (the child above the ground) and the image source (the child below) are represented as symmetrical with respect to the wall (horizontal in the illustrated example). This property of symmetry of the sources is linked to the conservation of Euclidean distances (between for example two original sources, on the one hand, and between their two respective image sources, on the other hand). This property remains true in the acoustic approach considering reflections as specular (which is generally the case) on flat walls, including for multiple reflections. This property is used and explained in more detail later, with reference to Figure 3a.

[0058] At each new peak in the succession of equation 39 of document WO-2021 / 074502, equation 40 of the same document gives a new delay τ new determined in relation to previous delays:

[0059] Eq.40

[0060] The generalized velocity vector in the sense of document FR2011874 is written in a similar way:

[0061] , and thus appears, as shown in the example in Figure 2: - a first peak in τ0, associated with the direct path, the vector U0 being obtained by normalization, - as many time series as there are reflections, each associated with the interference between a reflection and the direct sound, and abscissas τ n + τ0, - and the combined delay series noted SARC.

[0062] Regarding the β parameters n(noted BETAn below), starting from equation Eq.B6 of this document FR2011874 at the end of the appendix, we simply retain a particular relationship between two successive vectors of a series, in particular between the first two vectors V'(TAUn) and V'(2.TAUn), the most salient.

[0063] The representation of the GTW vector therefore makes it possible to directly determine the direct component (without reflection) indicating the DoA (and specific to the vector U0), by simply evaluating the sequence at v(t = 0). This is given by the first peak at τ0 en starting from the left in the example illustrated in Figure 2. Then, the gaps between peaks allow to determine some important reflections by identification and selection of peaks, because the peak index corresponds to the difference in arrival time (TDoA) of a given reflection, noted:

[0064] τ n (1)

[0065] It is therefore possible to obtain the DoA in the form of a unit vector, given for a frame of index k:

[0066] (2),

[0067] as well as a collection of pairs:

[0068] (3)

[0069] corresponding to the detected reflections and their associated TDoA time differences.

[0070] The position of the source relative to the microphone array (which may correspond to an ambisonic microphone) is given by the vector:

[0071]

[0072] and, similarly, the position of the nth image source is given by

[0073] vV

[0074] with , where c is the speed of sound.

[0075] Similarly, if the same reflection is detected in another frame k', the equivalent expressions for the position vectors and would be obtained.

[0076] Here we consider that the reflections detected, estimated and used in the model are a priori specular. It is then possible to use simple geometric arguments to arrive at the estimation of a distance d(k), to the frame indexed by the variable k. Indeed, the source images are obtained from the positions of the original sources by applying certain rigid transformations: reflections, translations or rotations (depending on the order and arrangement of the reflecting surfaces). Rigid transformations in principle preserve Euclidean distances.

[0077] In Figure 3a, two successive positions of a source in S1, then S2, as well as the position of the microphone M (assumed to be fixed in this example) are illustrated. The respective positions of the successive image sources relative to a first wall w1 in S1 are also illustrated. (w1) and S2 (w1) and with respect to a second wall w2 in S1 (w2) and S2 (w2) . As seen in particular for images S1 (w2) and S2 (w2) in this figure, the distance between the respective positions of images S1 (w2) - S2 (w2) associated with the same wall w2 is equal to the distance between the respective source positions S1-S2. This figure then illustrates the conservation of Euclidean distances between source positions and image positions.

[0078] This distance preservation property translates to:

[0079]

[0080] Expanding this expression using (4) and (5), we arrive at:

[0081]

[0082] with:

[0083] Or<x,y> corresponds to the scalar product between the vectors x and y.

[0084] An attractive quality of the distance-preserving hypothesis is that it does not require any specific assumptions about the geometry of the environment. A somewhat more restrictive, but in practice still very plausible, assumption is to consider all reflective surfaces as either horizontal (floors, ceilings, tables, etc.) or vertical (walls, windows, etc.). Such an assumption should nevertheless allow vertical surfaces to form arbitrary angles with each other, such as that of a door relative to the wall that supports it.

[0085] In order to exploit this assumption, the z-axis of the local coordinate system of the surround microphone should preferably be aligned with the z-axis of the global (room) coordinate system. This is generally true, since the microphone is most often placed on a horizontal surface (e.g., a table) or it is mounted on a vertical stand. If this is the case, the projection of the magnitude of the displacement vector on the z axis, for any index n, is equivalent to the same projection of the magnitude of the displacement vector of the corresponding source, i.e.:

[0086]

[0087] with :

[0088] Referring now to Figure 3b, this other geometric property (besides the conservation of Euclidean distances, but deriving from it nonetheless) translates as follows: the projection of the distance between source positions S1-S2 onto the z axis (parallel to the axis of the microphone M) is equal to the same projection of the distance between image positions S1 (w2) - S2 (w2) associated with the same wall w2. More precisely, the projection of the “magnitude” (positive amplitude) of the displacement vector S1 -S2 is equal to the same projection of the vector between S1 (w2) and S2 (w2) The term “magnitude of the vector” here designates a distance (therefore positive).

[0089] This property is geometrically verified if the walls considered are parallel or perpendicular to the z axis. Indeed, if a wall is vertical (parallel to the z axis) or horizontal (perpendicular to the z axis), the equations given here do not change because the projected magnitude (illustrated by the reference Pz in strong line in figure 3b) is always preserved in both cases.

[0090] By expanding equation (8) with equations (4) and (5), we obtain:

[0091]

[0092] with : and denotes the scalar product

[0093] It should be noted that distance preservation is not guaranteed under equation (8). Therefore, instead of being used independently, this condition instead complements the one given in equation (6). Expressions (7) and (9) can be written compactly as follows:

[0094]

[0095] Or , And

[0096] for the model based on equation (7); or

[0097] , And , for the model based on equation (8).

[0098] The operator extracts the coefficients of the diagonal as well as those which are above the diagonal of the matrix dd T and concatenates its inputs into a column vector.

[0099] Finally, the vector contains the distances estimated between the source and the microphone (k)for each frame k belonging to a set of frames indexed from 1 to K. It is specified here that the frames are not necessarily successive, that is to say that they do not necessarily follow each other immediately in time. For example, these may be frames of duration T such that the first indexed k=1 is emitted at time t, the second k=2 is emitted at t+4T, the one indexed k=3, emitted at t+5T, the one indexed k=4, emitted at t+7T, etc.

[0100] It is possible to obtain the position of the source relative to the microphone M at different times of its movement, these times being reflected here by the frame index k.

[0101] By assembling at least two frames at different times k and / ('belonging to [1, K] X [1, K] and where the same reflection was observed on the same wall, we arrive at a system of equations which is generally overdetermined (more equations than solutions), of the type:

[0102]

[0103] We note below: - M DP , q DP , the systems corresponding to the property of conservation of Euclidean distances, and - M HV , q HV , the systems corresponding to the hypothesis of walls parallel or perpendicular to the z axis of the microphone.

[0104] Thus, it is possible to present both geometric conditions simultaneously by assembling the two systems into a single combined system:

[0105]

[0106] Finally, the estimation of the distance vector d boils down to a regression problem based on the nonlinear system (12), as follows:

[0107] knowing that where 0 < Ib < ub denote the lower, respectively upper, limits of the distance estimate. The so-called "fidelity" term of the data ℓ(-) is generally a type of norm (squared or not), such as for example the sum of squares (ℓ = ℓ2 2 ), or absolute values ​​(ℓ= ℓ1). The first solution leads to a smooth optimization problem, but the advantage of the ℓ1 norm is that it can be more robust with respect to possible errors in the system parameters (the extracted DoAs and relative delays). An alternative could be the use of so-called "structured" norms (e.g. ℓ 1,2 ) if the parameters related to certain reflections are significantly more erroneous than the others. The term regularization term λr(d) can be optionally added to induce additional structure in the d vector, for example smoothing the source trajectory. Alternatively, adjusting this term can favor the detection of a displacement of a source moving by "jumps" from one position to another, this term r(d) encouraging for example a d vector which is only constant for sections of the path (and therefore not smoothed due to these "jumps").

[0108] In the matrix expression above: it is possible to relatively weight one of the systems M DP or M HV compared to the other M HV or M DP in order, for example, to give more weight to one of the geometric properties than the other, depending for example on the conditions of sound acquisition.

[0109] It is further possible to provide a linear combination of M DP and M HV , for example: M = MDP - M HV and q = q DP - q HV

[0110] Furthermore, it is possible to apply further weighting in equation (13) as follows: which amounts to adding a diagonal matrix ψ aimed at weighting the different equations of the system Mf + q. This weighting can be produced by applying confidence criteria on the extraction of delay parameters, DoA, etc., for example, for certain frames or peaks identified in these frames. For example, it can favor frames in which a sound attack is detected (to exploit direct sound and early reflections for example).

[0111] Problem (13) is non-convex, and a local solution can be found by applying a non-linear optimization method. In particular, a Fast Adaptive Shrinkage / Thresholding (FASTA) algorithm can be used, providing it with the appropriate (sub-)gradient of a cost function.

[0112] An advantage of the temporal representation of the velocity vector exploited with the processing presented here is that it is possible to group the reflections individual over time. In other words, in addition to tracking the source DoA, it is proposed here to further rely on the determination of reflections to strengthen the determination of the source / microphone distance over time.

[0113] Thus, the tracking algorithm described in the above-mentioned prior art documents FR2011874 and WO-2021 / 074502 can be modified so that it can process the observations in the form of DoA and the associated relative delays. A simple modification of this processing can consist of providing the measurements in the form of scaled vectors which allows the tracking processing to discriminate between reflections of very similar DoAs (e.g., a source near a wall), possibly adding a certain "depth" to the observations. In practice, two instances of tracking processing can be implemented: - the first instance follows the source itself, and is therefore responsible for estimating the trajectory of a single target using the DoA obtained from v(t = 0) (position of the first peak relative to the initial time t=0); - the second instance of processing performs multi-target tracking of reflections using the remaining observations (directions and relative delays obtained from the remaining peaks of the GTW velocity vector sequence).

[0114] Finally, it should be noted that, when defining the two aforementioned geometric conditions, it is possible to consider a pair of reflections (provided that they are both detected at two distinct times k and k), instead of coupling the parameters of the source and a single reflection. For example, from at least two reflections, it would be possible to determine the parameters related to these reflections and to deduce the source-microphone distance. However, tracking reflections is generally less stable and less accurate than directly tracking the source DoA. For example, some reflections may appear and disappear depending on the current position of the source, as they may become "invisible" to the microphone.

[0115] Furthermore, the case of a source that may be mobile and thus occupy distinct positions at respective times has been described above. However, the treatment proposed here can also be adapted to the case of several sources having these respective distinct positions at these different times. The exploitation of the reflections, mentioned above, can then be applied by obtaining information making it possible to distinguish the reflections corresponding to each source (for example by spectral analysis if the sources emit in different fundamental frequencies, or others).

[0116] In order to preferentially exploit the first reflections and avoid multiple reflections, which are more difficult to exploit, it is possible to weight the received signals by favoring sound attacks, as presented with reference to figure 5a of the documents cited above.

[0117] Figure 4 illustrates an example of a sound signal processing device, comprising a processing circuit arranged for implementing the above method. Such a processing circuit may then comprise: - an IN input interface to receive SIG signals acquired by the microphone (which may include several piezoelectric pellets to compose these signals, for example in an ambisonic context), - a PROC processor cooperating with a working memory MEM to process these signals, in particular to develop the expression of the generalized velocity vector in order to derive the desired parameters d0, U0, etc., the values ​​of which can be used to determine the distances d (1) , d (2) , ... , d (K) , between the source and the microphone to be delivered also by the OUT output interface.

[0118] Such a device may take the form of a module for locating a sound source in a 3D environment, this module being connected to a microphone (sound antenna type, or other). Conversely, it may be a sound rendering engine based on a given position of a source in a virtual space (comprising one or more walls) in augmented reality.

[0119] More generally, the object of this description can be used in numerous applications such as: - spatial audio coding, - the immersive experience (with six degrees of freedom) in augmented reality, - improving the separation of sources in space, - monitoring of multiple sources (active simultaneously or not), - assistance with robot navigation and instant mapping or “SLAM” (for “Simultaneous Localization and Mapping”), with a possible extension of the above principles to non-acoustic signals (for example radio waves).

Claims

DEMANDS

1. A method for processing sound signals acquired by at least one microphone, for the localization of at least one sound source, emitting from a plurality of distinct positions at distinct respective times (k, k'), in a space comprising at least one flat reflective wall, the method comprising: - to obtain at least, for each instant k: * a first vector determining a direction (DoA) of a first path acoustic, direct connection between the source and the microphone, * at least a second vector representing a second acoustic path stemming from at least one specular reflection, and arriving at the microphone, * at least one delay that the second path has, at the microphone, compared to the direct route, - exploit at least one property of reflection, specular reflection, according to which a Euclidean distance between two positions of the source at two distinct instants equals a Euclidean distance between two positions of respective images of the source and resulting from one or more of the same reflections, respectively at said two distinct instants, to determine at least one position (d (k) , d (k') ) of the source relative to the microphone respectively at said plurality of distinct instants (k, k'), as a function, for each instant k: * of the first vector to determine a direction (DoA) of the direct path, and * both the delay and the second vector to associate a distance d (k) between the source and the microphone in this direction (DoA) of the direct path.

2. A method according to claim 1, further comprising: - to exploit, in addition to the aforementioned property of specular reflection, a second geometric property according to which a projection onto a chosen axis of the said Euclidean distance between two positions of the source at two distinct instants corresponds to a projection onto the same chosen axis of the Euclidean distance between two respective image positions of the source and originating from one or more of the same reflections, respectively at the two said distinct moments.

3. Method according to claim 2, wherein the chosen axis is parallel or perpendicular to said at least one wall.

4. A method according to any one of claims 2 and 3, wherein the microphone is of the ambisonic type, and arranged so that the z-axis of height of the microphone is parallel to the chosen axis.

5. A method according to any one of claims 2 to 4, wherein the exploitation of said property of specular reflection, combined with the exploitation of the second geometric property, generates an overdetermined system of equations in which the positions of the source relative to the microphone, for different instants k, k', are the unknowns.

6. A method according to any one of the preceding claims, wherein the sound signals are acquired according to a time-sequenced sequence of frames, and wherein the first vector the second vector and the delay are obtained for a plurality of frames corresponding respectively to distinct instants ( , k') .

7. A method according to any one of the preceding claims, wherein at least one parameter among the first vector, the second vector and the delay is obtained from the expression of a velocity vector generalized, the process includes: - apply a time-frequency transform to the acquired signals, - From the acquired signals, express in the frequency domain a generalized velocity vector, for a plurality of distinct instants (k, k'), each generalized velocity vector for a given instant k characterizing a composition between: * the first acoustic path, direct between the source and the microphone, represented by the first vector, and taking a delay between the emission of a sound by the the source and reception of that sound by the microphone, and * at least the second acoustic path, represented by the second vector, and having the delay at the microphone, compared to the direct path.

8. A method according to claim 7, comprising: - furthermore, apply an inverse transform, from frequencies to time, to the generalized velocity vector to obtain, in the time domain, at least one peak linked to one or more reflections on one or more walls, in addition to a peak linked to the arrival of the sound along said direct path (DoA), the peak linked to one or more reflections being shifted by the delay relative to the peak linked to the arrival of the sound along the path direct.

9. A method according to any one of the preceding claims, wherein: - a vector between the source and the microphone, at a time k, written in function of the first vector, where d (k) is the distance Euclidean curve at time k between the source and the microphone, - a vector between an image of the source and the microphone, at time k, which can be written in terms of the second vector: , where c is the speed of sound, said specular reflection property translates, for two distinct instants k and k', into an expression of the type:

10. A method according to claim 9, wherein the expression is expanded to: with : where the notation<x,y> denotes the dot product between two vectors x and y.

11. A method according to claims 3 and 4, wherein said second geometric property is expressed by an expression of the type ; , where: - u z is a unit vector parallel to the chosen axis, - the notation<x,y> denotes the dot product between two vectors x and y, - is a vector between the source and the microphone, at a time k, written as a function of the first vector, where d (k)is the distance Euclidean curve at time k between the source and the microphone, - is a vector between an image of the source and the microphone, at a time k, written as a function of the second vector : . where c is the speed of sound, said expression developing into: Or : and denotes the dot product

12. A method according to claims 10 and 11, wherein the respective developments of the expressions and generate a system of bi- equations affine types: where the variable d is a column vector having coefficients corresponding to the distances between the source and the microphone at different times 1, 2, ..., K: and where the operator extracting the coefficients of a diagonal and above the diagonal of the matrix dd T by concatenating them into a column vector.

13. A method according to claim 12, comprising solving said system of bi-affine equations by non-linear minimization of a cost function ℓ(.) , given by: knowing that where Ib and ub are lower and upper limits given to the distances

14. A method according to claim 13, wherein a regularization term λr(d) is added to ℓ(.) to express the overall cost function as follows: the term λr(d) allowing to adjust at least one smoothing structure applied to the coordinates of the vector d.

15. A method according to any one of claims 13 and 14, wherein a diagonal weighting matrix diag(ψ) is applied to the cost function as follows:

16. Computer program comprising instructions for carrying out the method according to any one of the preceding claims, when said instructions are executed by a processor of a processing circuit.

17. Computer device comprising a processing circuit configured to implement the method according to any one of claims 1 to 15.

Images