Improved localization of sound sources

Abstract

The invention relates to processing a sound signal acquired by at least one microphone, for example a microphone of the ambisonic type, in order to localize at least one sound source in a space including at least one wall. A time-frequency transform is applied to the acquired signal and, from the acquired signal, a complex velocity vector V(f) having a real part and an imaginary part is expressed in the frequency domain, wherein the velocity vector characterizes the synthesis between: * a direct first acoustic path between the source and the microphone, represented by a first vector U0, and * at least one second acoustic path resulting from reflections on the wall and represented by a second vector U1, the second path having a first delay TAU1 at the microphone relative to the direct path. From the delay TAU1, the first vector U0 and the second vector U1, at least one parameter is determined among the direction of the direct path (DoA), the distance d0 from the source to the microphone, and the distance z0 from the source to the wall.

Description

Technical Field

[0001] The present invention relates to the field of sound source localization and in particular to estimating the acoustic direction or "DOA" (direction of arrival) by means of a compact microphone system (eg microphones capable of picking up sounds hereinafter indicated as "ambiophonic" or "ambisonic"). Background Art

[0002] Possible applications are for example the formation or "beamforming" of channels which then undergo a spatial separation of audio sources in order to improve speech recognition in particular (e.g. for virtual assistants interacting by voice). Such processing can also be used for 3D audio encoding (pre-analysis of the sound scene to encode the main signal separately), or even allow the spatial editing of immersive sound content, possibly in an audiovisual way (for purposes of art, radio, cinema, etc.). It also allows monitoring speakers in teleconferences or detecting sound events (with or without associated video).

[0003] In the prior art involving ambisonic type coding (or equivalent), most methods are based on spatial components resulting from frequency analysis (usually a time-frequency representation resulting from processing via a Short Term Fourier Transform or "STFT", or a representation of a narrowband time signal produced by a bank of filters).

[0004] The ambisonic signal of order 1 is collected in vector form according to equation Eq.1 given in the Appendix below. The coding convention of Eq.1 is presented here for convenience, but is not limiting, as conversion to other conventions can also be achieved. Therefore, when the field is comparable to a single plane wave coming from the direction described by the unit vector U1 (hence the directional DOA of the source) and carrying the transmitted signal s1(t), it can be written according to Eq.2 (Appendix).

[0005] In practice, the signal is analyzed frame by frame (in French: “trame”) in the frequency domain, which gives the expression Eq.3 (Appendix) and, for the case of a single wave, in the form of Eq.4 and, by extension, for N waves, in the form of Eq.5.

[0006] One class of methods relies on the analysis of the velocity vector V(f) or the intensity vector I(f) (the first being an alternative version of the second, normalized by the power of the omnidirectional reference component), as expressed in Eq. 6 and Eq. 7.

[0007] The method using complex frequency samples essentially bases the location estimate on the information contained in the real part of this vector (related to the properties of the wave propagation and the effective intensity, since it is directly related to the phase field gradient).

[0008] On the other hand, the imaginary part (the reactive part associated with the energy gradient) is considered to be characteristic of steady-state acoustic phenomena.

[0009] In practice it can be seen that within the framework of a single plane wave, the velocity vector can be summarized as V=U1.

[0010] The known method (called "DirAC") operates on time samples filtered into subbands (in which case they are real and the intensity vector is also real), or on complex frequency samples (in which case the real part of the vector intensity is used only to specify the direction of the origin (or more precisely, its inverse). Furthermore, the calculation of the so-called "diffusivity" coefficient associated with the ratio between the norm of the vector and the energy of the sound field makes it possible to decide whether the available information at the considered frequency is a directional component (whereby the vector direction determines the position) or a characteristic of the "ambience" (the field resulting from a mixture of diffuse reverberation and / or undifferentiated secondary sound sources).

[0011] Another method, denoted "VVM" hereinafter, is based on statistics of the velocity vector and the angular direction of its real part, weighted by certain factors associated with the ratio between the real and imaginary parts and their norm. A spherical mapping (2D histogram, e.g. equidistant bar graph) is built by collecting the values ​​over all frequency samples and over a certain number of time frames. The estimation is therefore essentially based on the maximum probability and is subject to a certain delay.

[0012] Another class of so-called "covariance" methods (sometimes presented as an extension of the first approach) involves computing the covariance matrix (sometimes also called the power spectral density matrix or "PSD") of the spatial components by frequency subband. Here again, the imaginary parts are sometimes completely ignored. It should be noted that when the spatial components are of the ambisonic type, the first row (or first column) of this matrix is ​​equivalent to an intensity vector. Many of these methods involve "subspace" methods and algorithms that are sometimes expensive, especially when they work on a large number of frequency subbands and when they exploit higher spatial resolution.

[0013] These "vector-based" or "matrix-based" methods try Figure 1 A "directional" component associated with a localizable sound source or path is discerned from the ambient component on the one hand, and a "directional" component associated with the ambient component is discerned on the other hand.

[0014] Among the observed limitations of this approach, even in the case of a single simultaneously active sound source, these are hampered by the interference of direct sound (which indicates the direction of the sound source) with reflections. For example, beyond a certain level of presence of room effects, they do not give reasonable estimates often enough and / or the estimates are too often biased. When the object comprising the capture and sound localization device (e.g., an ambisonic microphone) is placed, for example, on a table or near a wall (and / or if this is the case near the sound source), such reflective surfaces tend to cause angular biases in the system.

[0015] In practice, localization is often biased by the systematic interference of direct sound and reflections associated with the same sound source. When it is based on velocity vectors, mainly the real part of the velocity vector is considered, while the imaginary part is usually ignored (or at least not fully used). Acoustic reflections, which are considered to be disturbing, are not included in the estimation problem. Therefore, they remain neglected components, not modeled, without considering the specific structures causing the interference.

[0016] Therefore, for applications of the type described above, the acoustic localization is usually estimated only in angular terms. Furthermore, there seems to be no efficient way to propose an estimate of the distance from a single capture point (which is considered unique for coincident or more generally "compact" microphone systems, i.e. contained in a volume of small dimensions compared to the distance from the sound source, typically about ten centimeters for ambisonic microphones).

[0017] However, some application contexts require additional information besides its orientation (and thus 3D positioning in XYZ), namely the distance from the source. These are for example:

[0018] - virtual navigation in a real environment captured in 3D (since the appropriate modification of the angle and intensity of a source depends on the relative XYZ translation between that object and the microphone),

[0019] - localization of sources to identify the speaker (especially for connected speakers or the like),

[0020] - Monitoring and alarm equipment in home or industrial environments,

[0021] or something else. Summary of the invention

[0022] The present invention improves this situation.

[0023] A method for processing sound signals acquired by at least one microphone for localizing at least one sound source in a space comprising at least one wall is proposed, wherein:

[0024] - applying a time-frequency transform to the acquired signal,

[0025] - From the acquired signal, express in the frequency domain a velocity vector V(f), which is a complex number with a real part and an imaginary part, wherein the velocity vector characterizes the composition between:

[0026] * a first acoustic path directly between the source and the microphone, the first acoustic path being represented by a first vector U0, and

[0027] * at least one second acoustic path resulting from reflections on the wall and represented by a second vector U1,

[0028] The second path has a first delay TAU1 at the microphone relative to the direct path,

[0029] - Determine at least one parameter from the following based on the delay TAU1, the first vector U0 and the second vector U1:

[0030] * Direction of direct path (DoA),

[0031] * the distance d0 from the source to the microphone,

[0032] * The distance z0 from the source to the wall.

[0033] It is therefore proposed to model the interference between the direct wave and at least one indirect wave (generated by the reflections) and to exploit the behavior of this model on the entire velocity vector (on its imaginary as well as its real part).

[0034] In an embodiment, the sound signal is acquired by an ambisonic microphone, and the velocity vector V(f) in the frequency domain is represented by an order 1 ambisonic component in the form of the following type:

[0035] V(f)=1 / W(f)[X(f),Y(f),Z(f)] T .

[0036] The order considered here is 1, which makes it possible to express the components of the velocity vector in a three-dimensional reference space, but other embodiments are possible, in particular with higher ambisonic orders.

[0037] In one embodiment, the direction of the direct path can be determined by a first vector U0 based on the average value over a set of frequencies of the real part of the velocity vector V(f) expressed in the frequency domain (Eq. 24 in the Appendix presented at the end of this specification).

[0038] Therefore, even the expression of the velocity vector in the frequency domain already makes it possible to give the vector U0.

[0039] However, in a more advanced embodiment:

[0040] - applying an inverse transformation from frequency to time to said velocity vector to express said velocity vector in the time domain V(t),

[0041] - after the duration of the direct path, finding at least the maximum value V(t)max in the expression of the velocity vector as a function of time, and

[0042] - Thereby deriving said first delay TAU1 corresponding to the time at which said maximum value V(t)max occurs.

[0043] In this example:

[0044] - Estimation of the second vector U1 from the first vector U0 and from the maximum value V(t)max by a relationship of the following type:

[0045] U1=U0–2(U0·V(t)max)V(t)max / ||V(t)max||2

[0046] Then:

[0047] - The first vector U0 and the second vector U1 can be determined relative to the wall ( Figure 1 ) with their respective angles PHI0 and PHI1 such that:

[0048] PHI0=arcsin(U0·n R ) and PHI1=arcsin(U1·n R ), where n R is a unit vector and is perpendicular to the wall, and

[0049] The distance d0 between the source and the microphone is determined from the first delay TAU1 by a relationship of the following type:

[0050] d0 = (TAU1 x C) / ((cosPHI0 / cosPHI1)-1), where C is the speed of sound.

[0051] Furthermore, the distance z0 from the source to the wall can be determined by a relationship of the following type:

[0052] z0=d0(sinPHI0–sinPHI1) / 2

[0053] Thus, here in the presence of a single wall, all parameters relevant to the localization of the source (e.g., from Figure 1 ), although the model can be generalized to the presence of several walls.

[0054] Thus, in embodiments where the space includes multiple walls:

[0055] - Apply the inverse transform from frequency to time to the velocity vector to represent it as a series of peaks in the time domain V(t) (Eq. 39, Figure 2 ),

[0056] - identifying, among the series of peaks, peaks associated with reflections on walls of the plurality of walls, each identified peak having a time abscissa that is a function of a first delay TAUn of the acoustic path resulting from the reflection on the corresponding wall n relative to the direct path,

[0057] - from each first delay TAUn, the first vector U0 and each second vector Un representing the acoustic path resulting from the reflection on the wall n, at least one parameter is determined from:

[0058] * the direction of the direct path (DoA),

[0059] * the distance d0 from the source to the microphone, and

[0060] * At least the distance zn from the source to the wall n.

[0061] For example, as in Figure 5B As can be observed in the figure, the expression of the velocity vector (frequency versus time) after the inverse transformation presents a series of peaks, which is also Figure 2 , where for multiple values ​​of the above-mentioned delays (TAU1, 2TAU1, etc.; TAU2, 2TAU2, etc.) between the direct path and the path resulting from at least one reflection on the wall, and for combinations of these delays (TAU1+TAU2, 2TAU1+TAU2, TAU1+2TAU2, etc.), the maximum value is reached.

[0062] These peaks can then be used to identify in particular at least the peaks associated with the reflection on the wall n and which therefore have a plurality of time abscissas ( x1 , x2 , x3 , etc. ) of the delay TAUn associated with this wall n.

[0063] Since combinations of different delays may complicate the identification of simple delays (TAU1, TAU2, TAU3, etc.) and the presence of the associated wall, the first part of the peak at the minimum positive time abscissa can be preselected so that each peak associated with a reflection on the wall is identified in this part (so there are no combinations of different delays TAU1+TAU2, 2TAU1+TAU2, TAU1+2TAU2, etc. that may appear after the first peak).

[0064] In one embodiment, the acquired signal is delivered in the form of consecutive frames of samples:

[0065] - for each frame, it is possible to estimate the fraction of said frame in which a sound onset is present, and

[0066] - Frames having a score above a threshold may be selected for determining the at least one parameter.

[0067] In fact, in the method of finding the first instantaneous reflection of the wall, it may be preferred to find the instantaneous reaction of the room wall to the sound onset (ie the beginning of the sound emission, expressed in French as "attaque sonore").

[0068] In one embodiment, the signal acquired by the microphone is in the form of a series of samples to which a weighting window (in French: " de pondération”), the weighting window has an exponential variation that decreases over time ( Figure 5A ).

[0069] Furthermore, possibly, the window can be placed at the very beginning of the sound onset (or even just before the onset). This avoids the discomfort of multiple reflections.

[0070] The application of such a weighting window allows less biased estimates of the parameters U0, d0, etc. derived from the expression using the velocity vector in the time domain. In fact, in some cases where the cumulative amplitude of the reflections is greater than the cumulative amplitude of the direct sound, the estimates of the above parameters may be biased. When the negative time abscissa in the time expression of the velocity vector ( Figure 5B These cases can be detected when peaks are observed at the top of the curve). The application of a weighted window of the above type makes it possible to bring these peaks back to the positive abscissa, such as Figure 5B The bottom curve shows this and gives a less biased estimate.

[0071] In one embodiment, the weights q(f) each associated with a frequency band f may also be iteratively applied to the velocity vector V(f) in the frequency domain according to an expression of the type (Eq. 49):

[0072] q(f)=exp(-|Im(V(f))·m| / (||Im(V(f))||)),

[0073] where Im(V(f)) is the imaginary part of the complex velocity vector, and m is the perpendicular to the line consisting of vector U0 and the wall normal (z, Figure 1 ) (usually commented on in detail later Figure 1 The unit vector of the plane defined by the z-axis.

[0074] Such an embodiment enables the selection of the most available frequency band for determining the above mentioned parameters.

[0075] The present invention also relates to a device for processing a sound signal, comprising a processing circuit for implementing the method according to the above.

[0076] As an illustration, Figure 4 Such a processing circuit is schematically shown, which may then include:

[0077] an input interface IN for receiving a signal SIG acquired by a microphone (which may comprise several piezoelectric pads to compose these signals, for example in the context of ambisonics),

[0078] - A processor PROC which cooperates with the working memory MEM to process these signals and in particular to build up a representation of the velocity vector in order to extract therefrom the desired parameters d0, U0, etc., the values ​​of which can be delivered by the output interface OUT.

[0079] Such a device may take the form of a module for localizing sound sources in a 3D environment, connected to a microphone (sound antenna type or other type). Alternatively, it may be a sound rendering engine based on a given position of a source in a virtual space (including one or more walls) in augmented reality.

[0080] The 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.

[0081] Figure 3A and 3B A flow chart example of an algorithm for such a procedure is shown.

[0082] In another aspect, there is provided a non-transitory computer-readable recording medium having such a program recorded thereon. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Other features, details and advantages will become apparent upon reading the following detailed description and analyzing the accompanying drawings, in which:

[0084] Figure 1 Various parameters involved in locating a sound source according to an embodiment are illustrated in a diagrammatic manner.

[0085] Figure 2 The various successive peaks exhibited by a time representation of a velocity vector after an inverse frequency to time transform ("IDFT") of the velocity vector are graphically shown.

[0086] Figure 3A shows the initial steps of the algorithmic processing for determining the relevant parameters U0, d0, etc.,

[0087] as well as Figure 3B Shows Figure 3A Continuation of the processing steps.

[0088] Figure 4 A device within the meaning of the invention according to an embodiment is schematically shown.

[0089] Figure 5A An exponentially decreasing weighting window over time of samples of an acquired signal is shown according to one embodiment.

[0090] Figure 5B Compare the time representation of the velocity vector after IDFT:

[0091] - without preprocessing the samples by weighting window (top curve),

[0092] - and has a window treatment (bottom curve). DETAILED DESCRIPTION

[0093] The velocity vector can be calculated in a manner known per se. However, some specific settings can be recommended in order to improve the final result obtained.

[0094] Typically, the spectrum B(f) of the ambisonic reverberation signal is first obtained, typically by a short-term Fourier transform (or STFT) for a series of time frames b(t), which are typically overlapped (e.g., with addition / superposition). Here, the order of the ambisonic reverberation components may be m=1 for four components (however, the calculation can be adapted to higher orders without loss of generality).

[0095] Then, for each time frame, the velocity vector is then calculated as the ratio of the directional components X(f), Y(f), and Z(f) to the omnidirectional component W(f) for all frequency samples (Eq. 6 in the Appendix). Embodiments that also introduce temporal smoothing or merging by weighted sums can be considered, as described below.

[0096] Using such ratios (X(f) / W(f), Y(f) / W(f), Z(f) / W(f)), if the spectral content of the audio signal excites a large number of useful frequencies (for example, over a wide frequency band), the characteristics of the source signal are essentially removed to highlight the characteristics of the acoustic channel.

[0097] In the application presented previously, one can consider the case where a sound source with stable properties (in position and radiation, at least over several consecutive frames) emits a signal s(t) in a stable acoustic environment (reflections, possible diffractions off walls and objects etc., and thus responsible for what is usually called the "room effect", even though it may be outside the "room"). These signals are received by an ambisonic microphone. The ambisonic signal b(t) results from the combined spatial encoding of different versions of the signal s(t) along its direct and indirect paths, the so-called "acoustic channel effects". This results in the convolution of the signal with a spatial impulse response h(t), each channel (or dimension) of which is associated with an ambisonic component, as expressed in Eq. 8 of the Appendix.

[0098] This impulse response is called the Spatial Room Impulse Response, SRIR, and is usually represented as a series of time peaks:

[0099] - the first peak at time t=TAU0 (flight time), corresponding to the direct sound,

[0100] - a second peak at t=TAU1, corresponding to the first reflection,

[0101] etc.

[0102] Therefore, it should be possible to read the direction of origin of these wavefronts in these peaks, where the vector u given in Eq. 9-1 is n As a first approximation, the expression for is . In practice, the spatial impulse response is unknown data, but here it is revealed how to indirectly get back to some of its characteristics via the velocity vector calculated based on the ambisonic reverberation signal b(t).

[0103] To emphasize this point, the connection between the impulse response h(t), the transmitted signal s(t) and the ambisonic reverberation signal b(t) (Eq. 9-2) is first described over a selected observation time interval. Specifically, this expression assumes the absence of any measurement noise and the absence of other sound sources whose signals will be captured directly or indirectly over the considered time interval. Therefore, all direct and indirect source signals are captured over this time interval.

[0104] It is shown that by performing a Fourier transform over the entire time interval, the resulting velocity vector is the only characteristic of the spatial impulse response. This so-called LT transform (because it is "longer term" than the STFT) converts b(t), s(t), and h(t) into B(f), S(f), and H(f) according to Eq. 10. The time support may correspond to a time window extending over several consecutive signal frames.

[0105] From the expression of the convolution in the frequency domain, we then derive the convolution expression of the velocity vector calculated with equation Eq. 11. This expression Eq. 11 becomes a characteristic of the acoustic channel (in other words, the room effect) and no longer of the transmitted signal, as long as the latter has a non-zero energy (in practice, detectable) for each frequency f during the time period considered.

[0106] In practice and as already mentioned, a common approach is to perform a time-frequency analysis per frame, where each short-term Fourier transform is applied to a time windowing (in French: "fenêtrage") which does not verify a priori that the observed signal results entirely and exclusively from the convolution product of Eq. 9. This means that, strictly speaking, the velocity vector cannot be written in a form that characterizes only the acoustic channel (as in the right-hand side of Eq. 11). However, it is approximated here as much as possible in the context of this specification (Eq. 20 detailed later), while taking advantage of the short-term analysis presented below.

[0107] At a later stage, a search is made for a sequence of energy peaks which characterize on the one hand the direct path of the signal emitted from the source and received by the microphone, and then the first reflections on one or more walls, as long as these reflections are identifiable. It is then possible to focus on the characteristics of what is the beginning of the spatial impulse response, i.e. first the first time peak from which we derive the direction of the direct sound, and possibly subsequent time peaks from which we derive the characteristics of the first reflection.

[0108] To this end, the influence of the interference between the direct sound and at least one reflection on the expression of the complex velocity vector is examined in order to estimate the relevant parameters for defining the position of the sound source.

[0109] For the beginning of the impulse response, a simplified model of the direct path (n=0) combined with N mirror reflections (n=1, ..., N) is introduced as shown in Eq. 12, where g n , TAU n and u n are respectively the attenuation, delay and origin direction of the wave indexed n (nth reflection) arriving at the microphone system. In the following, for simplicity and without limiting generality, the delay and attenuation associated with the direct sound are considered, which is equivalent to setting the terms of Eq. 13 for n=0.

[0110] The corresponding frequency expression is given in Eq. 14, where for the special case of gamma0 = 1 for direct sound. Of course, the variable gamma n (regardless of whether n is greater than 0) is a function of frequency f.

[0111] If the latter part is ignored, the frequency expression of the ambisonic reverberation field is as follows, given by expression Eq.16.

[0112] The short-term velocity vector is then expressed by Eq. 17 or even by Eq. 18 according to a regularized version with a non-zero ε(EPSILON) term in order to avoid (quasi) infinite values ​​when W is (almost) zero.

[0113] Short-term analysis allows the observation of the frequency footprint (hereinafter denoted as "FDVV"), which is a characteristic of a submix of the wavefront within the spatial impulse response, over time and according to the dynamic evolution of the source signal. The characteristic submix of a given observation ("submix" smx) is modeled in the time and frequency domains according to Eq. 19.

[0114] In the method described below, one strives to characterize the frequency footprint FDVV as represented by the implicit sub-mixture model H by the approximation expressed in Eq. 20 (which is usually not exactly equivalent except in special cases). smx produce.

[0115] Especially at the start time of the signal, the implicit model h smx (t) appears to be similar to h at least in terms of relative wavefront direction and delay early (t) The beginning of the spatial impulse response. Due to the relative gain g n The implicit parameters are affected by the time windowing and dynamics of the signal, so they do not necessarily appear to correspond to those of the impulse response. Here we focus mainly on the direct wave (which provides the DOA) and one or a few early reflections, essentially considering the case where the observation is characteristic.

[0116] In particular, for illustrative purposes, an example is described below considering only a single reflection with an estimate in the frequency domain. The case of a simple interference is treated here (basically between the direct sound and the first reflection) and it is shown how the parameters sought can be determined by highlighting a specific space-frequency structure by looking not only at the real part of the velocity vector but also at the imaginary part of the velocity vector. In practice, the ambisonic field is described according to Eq. 21 and the velocity vector is derived from it according to Eq. 22. From this expression it follows that, when the frequency travels through the audio spectrum considered, the real and imaginary parts travel parallel segments in 3D space (affine and linear, respectively), as shown in Eq. 23. The affine segment (real part) is on the line containing the unit vectors U0 and U1 pointing to the direct wave and the indirect wave, respectively, and the two segments are orthogonal to the plane between these two vectors (and therefore the imaginary part of the vector itself is always the case, since it is on a linear segment). Furthermore, assuming a uniform distribution of the phase shifts between the waves (and therefore a representative scan of the frequencies), according to a statistical calculation, the mean value of the real part of the velocity vector is equal to the vector U0, as shown in Eq.24, and the maximum probability is the mean value of U0 and U1 weighted by the corresponding amplitudes of the waves, as shown in Eq.25. Therefore, the DOA detection based on the maximum probability is contaminated by the systematic angular deviation, giving an intermediate direction between the direct sound and its direction. Equation Eq.23 shows that this spatial scan is done with a frequency periodicity equal to the inverse of the delay TAU1 between the two waves. Therefore, when such a space-frequency structure can be observed, the directions U0 and U1 and the delay TAU1 can be extracted from the observations. Another embodiment for estimating these parameters in the time domain (in combination with Figure 2 description).

[0117] By having a priori about the orientation of the reflecting surface relative to the microphone's reference frame, it is then possible to derive information about the absolute distance d of the source relative to the microphone from the estimates of U0, U1, TAU1, and possibly the height of both. Figure 1 As shown, by recording the distance d0 from the source S0 to the microphone M and the distance d1 of its mirror image S1 relative to the reflecting surface R, the surface R being orthogonal to the plane formed by the vectors U0 and U1. These three points (M, S0, S1) are in the same plane orthogonal to the surface R. It is then necessary to define the parameters to be determined in order to define the orientation (or inclination) of the reflecting plane. In the case of a reflection from the floor or ceiling (and therefore detected because U1 is pointing towards the floor or ceiling), the following assumption can be used: this is horizontal and parallel to the plane XY of the frame of the ambisonic microphone. The distances d0 and d1 are then related by the relation 26, which also directly gives the distance from the microphone M to the axis (S0, S1), PHI0 and PHI1 being the respective elevation angles of the vectors U0 and U1.

[0118] An estimate of the delay TAU1 of the reflected sound relative to the direct sound is also obtained, which allows access to another relationship between the distances Eq. 27, since their difference converts the acoustic path delay by a factor c, which is the speed of sound.

[0119] By expressing d1 as a function of d0, this last quantity becomes the only unknown quantity, which can be estimated according to Eq. 28. Also obtained is the distance from the source to the reflecting plane, ie its height or latitude z0 relative to the ground according to Eq. 29, and the height or latitude of the microphone in Eq. 30.

[0120] exist Figure 1 The various parameters U0, U1, PHI0, PHI1, d1, d0, etc. are shown in the example of reflections on a floor. Of course, similar parameters can be derived for reflections on a ceiling. In the same way, similar parameters can be derived for reflections on any other reflecting surface R whose orientation relative to the reference frame of the microphone is known, as determined by the normal n R (unit vector orthogonal to the surface R). It is sufficient to redefine the angles PHI0 and PHI1 relative to the reflecting surface R, usually PHI0 = arcsin (U0·n R ) and PHI1=arcsin(U1·n R ). Therefore, the respective positions of these obstacles can be determined by the vector U1 associated with each reflection case for applications in augmented reality or robotics for positioning estimation by acoustic detection.

[0121] The position n on the reflecting surface R If it is not known a priori, it can be fully estimated if the wavefront parameters associated with at least two source positions for which reflections were detected by the same reflection plane have been estimated by observations at different times. Thus, there is a first set of parameters (U0, U1, TAU1) and at least a second set of parameters (U0', U1', TAU1'). Since U0 and U1 define a plane orthogonal to the plane R, their vector product defines the axis of this plane R, and the same applies to the vector product obtained from U0' and U'1.

[0122] These corresponding vector products (non-collinear) together define the orientation of plane R.

[0123] However, this model, which is limited to two interfering waves (direct sound and reflected wave), is likely to have difficulty distinguishing different first reflections on the partition. Moreover, when additional reflections are introduced, the space-frequency behavior of the velocity vector quickly becomes more complex. In practice, the routes of the real and imaginary parts are then combined, and in a non-trivial way, along several axes:

[0124] - In the parallel planes for the direct wave and two reflections,

[0125] - Or generally in the whole space.

[0126] When several reflecting surfaces are to be considered, these complex spatio - frequency distributions make the determination of the model parameters too cumbersome.

[0127] The solution to this problem is to operate a time - frequency analysis that is more selective in time (i.e., has a shorter time window) to have the opportunity to see a simpler acoustic mixing occur during the amplitude onset (transient, signal rise), i.e., to reduce the number of reflections of the direct sound in the mixing present in the interfering frames. However, in some cases, the delays associated with successive reflections may be too close to each other to isolate the effect of the first reflection in their interference from the direct sound.

[0128] Then a treatment is proposed below that allows the effects of multiple interferences to be easily separated and characterized. The first step consists of transforming the footprint of the velocity vector in the time domain (or "TDVV", "Time - Domain Velocity Vector") by means of an inverse Fourier transform as presented in Eq.31. This has the effect of compressing the frequency cycles associated with certain axes, and the effects of these frequency cycles manifest in the complex drift of the velocity vector, manifesting in less data and thus being easier to analyze. In fact, this transformation results in a series of peaks occurring at regular time intervals, where the most important peaks are easily detectable and extractable (see for example Figure 5B ).

[0129] A remarkable property is that, by construction (due to the inverse Fourier transform), the vector at t = 0 is equal to the average of the velocity vector in the frequency domain (the average of its real part if we only consider the positive frequencies of the half - spectrum). Such an observation is related to the estimation of the main DoA U0.

[0130] Starting from the frequency model of the velocity vector for two interfering waves (direct sound and one reflection), the denominator can be usefully reformulated due to the Taylor expansion in Eq.32. Using the conditions on x and gamma1 given in Eq.32, an expression for the velocity vector Eq.33 is obtained, and under the condition that the reflection has a smaller amplitude than the direct sound (g1 < g0 = 1, which is usually the case at the start of the sound onset), the inverse Fourier transform of this expression converges and is formulated as in Eq.34, where the first peak is identified at t = 0, which gives U0 (the direction of the direct sound), and then a series of peak characteristics of the interference between the reflection and the direct sound.

[0131] These peaks are placed at multiples of the delay TAU1 (non-zero integer k>0) and have exponentially decreasing amplitudes in norm (according to the gain g1). They are all associated with directions that are collinear with the sum and difference U0-U1, and therefore orthogonal to the middle plane between these two vectors, and with alternating directions (signs). The advantage of converting the velocity vector to the time domain is that it saves and almost immediately presents the parameters sought ( Figure 2 ).

[0132] Therefore, in addition to the main DoA U0, it is possible to determine:

[0133] -TAU1 delay, possibly for several different walls,

[0134] - Then, the vector collinear with U0-U1, which is normalized to the unit vector n, can be used with, for example, Eq. 41:

[0135] - derive U1 as a vector symmetric to U0 with respect to the mid-plane with U0, and

[0136] - Optionally, an attenuation parameter g1 (this may be modified by the time-frequency analysis parameters, in particular by the shape of the analysis window and by its temporal placement relative to the observed acoustic events. Estimation of this parameter is therefore of less practical use in the context of the applications mentioned herein).

[0137] Observation of the following time peaks makes it possible to check whether they correspond essentially to the same series (multiple delays TAU1, multiple delays TAU2, etc.) and therefore to the characteristics of the same interference, otherwise it would be necessary, for example, to determine the presence of multiple reflections.

[0138] In the following, the case with "favorable conditions" is highlighted, where in the case of N reflections, the Taylor expansion is applied to give the velocity vector according to Eq. 35, provided that the sum of the N gammas in Eq. 35 remains below 1. The Taylor series that transforms the denominator in the initial expression can be rewritten using the polynomial law of Eq. 36, which makes it possible to reorganize the expression of the velocity vector V model as the sum of several series, where the "cross series" is represented by the term SC of Eq. 37.

[0139] Under the conditions of Eq.38, for any frequency f, the following time series Eq.39 is derived by means of the inverse Fourier transform, where the series has a combined delay SARC. The first peak at t=0 is identified, which gives U0 (the direction of the direct sound), and then for each reflection, a series of peak characteristics of the interference of that reflection with the direct sound are identified. For example, at Figure 2In the graph, these peaks are placed at consecutive positive time abscissas TAU, 2TAU, 3TAU, etc., which are multiples of the delay TAU between the reflection on the wall and the direct path.

[0140] There then appears a characteristic series of interferences between several reflections on several walls and the direct sound (for larger time abscissas), whose delays are other combinations of their different delays (with positive integer factors).

[0141] In fact, Figure 2 Such a series is shown in the simplified case of two reflections interfering with the direct sound. Each mark (circle, cross, diamond, respectively) indicates by its ordinate the contribution of the vectors U0, U1, U2 (characteristics of the direct sound, the first reflection and the second reflection, respectively) to the time footprint TDVV as a function of the time abscissa. It can thus be seen that the reception of the direct sound is characterized by a first peak at time zero and with an amplitude of 1, shown by a circle. The interference of the first reflection (delay TAU1) with the direct path results in a first series of peaks in TAU1, 2xTAU1, 3xTAU1, etc., which are marked in this article by a cross at one end and a circle (up-down) at the other end. The interference of the second reflection (TAU2 delay) with the direct path results in a second series of peaks in TAU2, 2xTAU2, 3xTAU2, etc., which are marked in this article by a diamond at one end and a circle at the other end. Then there are the elements of the "cross series", i.e. the interference between the reflections (first delay: TAU1+TAU2, then 2TAU1+TAU2, then TAU1+2TAU2, etc.). For the sake of brevity, these cross-series, whose expressions are available but are generally very long to write, are not explained in this paper, especially because it is not necessary to utilize them to estimate the relevant parameters in the treatment presented in this paper.

[0142] The analysis of the time footprint by sequential estimation of parameters is described in the following.

[0143] The estimation of the parameters of the model from the calculated time series is done in a manner similar to the case of a single reflection previously described. First, in the most general case (excluding the specific cases treated later), corresponds to the favorable case when the delays do not "overlap": the above series then do not show any temporal coincidence, that is, any identifiable peak belongs to only one of them. Therefore, by recording the time peaks by increasing the delay from t=0, any new peak detected with a delay TAUnew can be attributed to an already identified series, or define the beginning of a new series. In fact, considering a set of delay characteristics of already identified reflections, the first case is detected if there is a positive integer k or partly zero that gives TAUnew according to Eq.40, otherwise the second case applies and the group of identified reflections is increased by introducing a new delay TAUN+1 associated with a direction that can be estimated in the manner described in the case of a single reflection.

[0144] In practice, there may not be a need to seek to explain many temporal peaks. Restriction is placed on the first peaks observed, in particular because they are most easily detected due to their larger amplitude (or magnitude (absolute value)) than subsequent peaks. Thus, the case where the delays have a common multiple but high (or non-low) rank Ki; Kj can be analyzed as a function of amplitude by the above process.

[0145] As long as the sum of the implicit gain blocks gn(n>0) is less than 1 (Eq. 38), the inverse Fourier transform (Eq. 31) gives a unidirectional time footprint evolving in positive time.

[0146] On the other hand, if the sum of the implicit gain blocks gn(n>0) is greater than 1, then the inverse Fourier transform gives a "bidirectional" TDVV time footprint, where the series generally develops towards positive and negative times ( Figure 5B The top curve is used for illustration). For example, when the amplitude of the direct wave is smaller than the sum of the amplitudes of the waves resulting from reflections on one or more partitions, a situation may be encountered in which one or more reflection gains will be greater than 1. In this "unfavorable situation", the main peak at time zero no longer strictly corresponds to the vector u0, but to a mixture of the vector u0 and a more or less significant proportion of the vector indicating the direction of reflection. This leads to a positioning bias (positioning bias of the "estimated DOA"). Another symptom is that the main peak has a norm that is generally different from 1, and more usually lower than 1. Depending on the relative importance of the problem, the extent to which the U0 vector provides reasonable (weakly biased) DOA estimates can be assessed, thereby providing a confidence factor about the estimate and making it possible to give priority to the estimates made on some frames. When the risk of estimation bias proves to be too great, the frames that are least exposed to the problem can be selected, as described below with reference to Figure 3 (shown by Figure 3A and 3B composition).

[0147] Thus, the first peaks of a given room can be observed continuously by frequency analysis of time subframes. The frames where the signal onset is localized (energy rise, transient, etc.) are those that allow observation of the acoustic mixture involving only the earliest wavefronts: the direct sound and one or more reflections (so that the above-mentioned "gamma sum" remains below 1 according to Eq. 38).

[0148] With respect to the frame containing the onset of the signal, the time window used for the frequency analysis can be adjusted (possibly dynamically), for example by giving it an asymmetric and globally decreasing shape, so that the "bump" of the window gives more weight to the signal rise (onset, transients) and thus gives progressively smaller weights to the direct sound (for example, but not necessarily, in an approximately exponential manner). Thus, the amplitude of the later wavefronts is artificially reduced relative to the earlier wavefronts, and the convergence condition is approached, whereby the Taylor series ensures a unidirectional time development.

[0149] In the following an example of reduced exponential time windowing is presented to be applied to the analyzed signal in order to bring the analysis of the resulting time footprint back to a favorable situation without substantially biasing the estimate of the direction of arrival of the wave. The set operation, for convenience, is valid from time t0 designated as time 0 and preferably corresponds to the instant of signal onset before silence, as shown in equation Eq. 42, where ALPHA>0, and the form of equation Eq. 43 is found by reintegrating the convolution form involving s(t) and h(t).

[0150] Equation Eq.44 then enables the properties of the exponent whose choice is justified so as to obtain the form given in Eq.45, which is equivalent to establishing equation Eq.46.

[0151] Therefore, if the impulse response is modeled by a set of specular reflections added to the direct sound, Eq. 47 is obtained.

[0152] Therefore, if the sum of the gammas is greater than or equal to 1 (possibility of "bidirectional progression"), it is always possible to determine the attenuation factor ALPHA such that the sum of the gains thus "adapted" (Eq. 48) becomes lower than 1.

[0153] It is then observed that the temporal footprint is unidirectional in nature, as evidenced by the peaks occurring only in positive times after the reduced exponential window is applied ( Figure 5BIt is also observed that, in practice, the energy of the observed signal decreases exponentially very quickly, and the numerical impact of the truncation of said signal on the estimate becomes quite negligible after a relatively short truncation time. In other words, the advantages of a long-term analysis covering both the entire exciter signal and its reverberation are obtained in a shorter time. In fact, the observed "TDVV" conforms to the interference model without errors due to the dynamics of the signal. This is therefore a dual property of weighting by such a window, which ideally makes it possible to obtain a usable time footprint.

[0154] In practice, without knowing the amplitude of the reflections in advance, it is appropriate to determine the attenuation ALPHA, preferably finding a compromise between a value low enough to ensure the unidirectionality of the temporal footprint and a value not too low to avoid reducing the chances of detecting and estimating indirect waves. For example, this value can be based on the attenuation factor a EXP By duration t EXP To determine, the duration t EXP Physically represents the observed phenomenon (usually 5ms), so that ALPHA = -(log a EXP ) / t EXP .

[0155] An iterative process (e.g., by bisection) may be implemented to adjust the attenuation value. Depending on the threshold attenuation value, when the obtained time footprint is detected as bidirectional and therefore has a priori a deviation vector U0, the analysis is repeated with a stronger attenuation, otherwise at least an estimate of U0 is adopted, and if subsequent peaks are barely discernible (because they decrease with attenuation), the analysis is repeated with an intermediate attenuation between two previous peaks, if necessary, and so on, until the vector U1 can be estimated.

[0156] However, the method with exponentially decreasing window may be sensitive to interference, especially at the beginning of windowing, where the interference is significantly amplified at the beginning of windowing. Interference other than noise can simply be the reverberation of the source itself at the beginning of windowing, if it was activated shortly before. De-noising can then be introduced to reduce this interference.

[0157] Often time windows of different shapes and / or sizes, or even overlaps and superpositions between windows, may be provided in order to maximize the chances of obtaining a "favorable footprint".

[0158] An initial DFT size is chosen that is typically larger than the analysis window.

[0159] Of course, we place ourselves in the context of processing digital audio signals in the form of successive blocks (or "frames") of samples sampled at a given sampling frequency.

[0160] It is also possible to optionally provide for pre-processing the detection of onsets, transients etc. and then for time-frequency denoising, for example by defining a mask (a time-frequency filter, possibly a binary filter) in order to avoid introducing elements from other ambient and / or diffuse field sources into the interference footprint. It is appropriate to calculate the impulse response of the mask (the result of the inverse transformation) to control the influence of the mask on the peak analysis. It can alternatively be integrated into the frequency weighting of the footprint of the frame considered to be stored in order to then calculate a weighted average of the a priori frequency footprint corresponding to a similar interference mixture (usually on the signal onset, by checking that the relevant source has not moved, which can be guessed by an estimate of the delay).

[0161] So then extraction is done and peaks are observed, e.g. according to the norm |V(t)|: largest peak, then the next peak, giving TAU1 (usually) etc.

[0162] The time footprint is then diagnosed by testing (according to {tau_n} and V(sum(k_n·tau_n))):

[0163] - whether there is time recirculation (a kind of loop "aliasing") due to choosing the FFT on too short a time support,

[0164] - whether there is a progressive unidirectional series or, conversely, a bidirectional series,

[0165] Or even whether there are special cases of series without significant decay (when the sum of the gains sum(gn) remains close to 1) or even retrograde series (when at least one implicit gain g_n>1).

[0166] Then you can:

[0167] - assign a score of "good frame" or "good footprint" (allowing for reliable estimation, a priori and without bias of DOA, since it is unidirectional), and memorize it,

[0168] - perform the estimation (u_n), and

[0169] - If necessary, adjust upstream analysis by selecting appropriate time windows.

[0170] The analysis of time footprints has been described above, but the frequency analysis can be performed more simply as follows.

[0171] It is easy to show mathematically that the peak at time zero is by construction equal to the average of the velocity vector over its entire spectrum (the real parts cancel out by Hermitian symmetry), or even to its real part (if only positive frequencies are considered). It can be estimated that if only direct sound is of interest, it is useless to calculate the inverse transform of the FDVV to have an estimate of the DOA. However, the temporal check of the TDVV makes it possible to detect whether this DOA is reliable (development criteria towards positive and increasing time).

[0172] This favorable situation is more reasonably observed during the onset of the source signal, when the mixture is still not very complex. Usually, it is sufficient to estimate these moments.

[0173] Furthermore, in practice, the frequency and time footprint of the VV is not always identifiable with an ideal model of a mixture of interfering waves. It is possible that, due to a lack of transmitted power, the source signal does not sufficiently or not always excites a significant range of frequencies at the critical moment, possibly taking into account competition from other components of the captured sound field (insufficient SNR or SIR). This can be associated with more or less diffuse background sounds (other sound sources) or microphone noise.

[0174] Then at least one of these processing operations or several of these processing operations in combination may be performed:

[0175] -Selection of time-frequency samples with onset detection based on advanced algorithms,

[0176] - possibly smoothing the velocity vector over several frames via the intensity vector (e.g. by |W(f)| of the relevant frames 2 and can be the average value of V(f) weighted by the forgetting factor of the dynamic factor),

[0177] -The selection of the starting frame is performed by |W(f)| 2 Weighted averaging of V(f) (if the extracted delays are the same) to complete the frequency footprint and merge the time footprint.

[0178] To simplify the calculations, it may also be advisable to perform the calculations upstream of the TDVV or even the FDVV only for frames that are detected as more consistent in terms of information: for example, if the case is a start frame detectable by simple processing, in this case it may even be advantageous to position the analysis window at the rise of the signal.

[0179] To properly estimate non-integer delays (fractional delays and their multiples in a time series), one may consider peak estimation via inter-sample interpolation and / or local frequency analysis (by isolating peaks in a narrow neighborhood in time) and fine-tuning the delay based on the phase response.

[0180] The prior selection of the time peak may be performed based on a current estimate of the characteristic delay of the series.

[0181] Therefore, it can be summarized as follows Figure 3A and 3B Steps implemented in an exemplary possible embodiment shown. At step S1, the Fourier transform (from time to frequency) of the ambisonic signal is calculated, which may be in the form of a series of consecutive "frames" (blocks of consecutive samples). For each transformed frame k (step S2), a dynamic mask may be applied to some of the frequency bands for which the signal-to-noise ratio is below a threshold (some frequency bands may indeed be noisy, for example by microphone or other inherent noise, such that the utilization of the signal captured in this frequency band is impaired). In particular, a search for noise in each frequency band is performed preferentially on the "omni" component W at step S3, and frequency bands altered by noise (exceeding a threshold, e.g. SNR<0 dB) are masked (i.e. set to zero) at step S4.

[0182] Then, in step S5, the velocity vector V(f) is calculated in the frequency domain, for example by means of equation Eq. 6 (or even in the form of Eq. 11, Eq. 18 or Eq. 20).

[0183] In an exemplary embodiment, a weight q(f) calculated as described below is applied to give more or less importance to a frequency band f. Such an embodiment makes it possible to express the velocity vector V(f) in a frequency band whose evolution is important. To this end, the optimal weight is iteratively calculated from U0 and V(f). Thus, returning to Figure 3A In step S6, various weights q(f) are set to 1. In step S7, the weighted q(f) applied to V(f) of each frequency band is applied so that Vbar(f)=q(f)V(f). In step S8, U0 is determined for each frame k so that:

[0184] U0(k)=E(Re(Vbar(f))), where E(x) here as an example is the expected value of x, which is therefore similar to the average over all frequencies of the real part of the estimated velocity vector Vbar(f).

[0185] Of course, the first estimate of U0(k) is rough. It is iteratively refined by calculating the previously determined weights for U0(k) based on the imaginary part of the vector V(f) using Eq. 49, and where the vector m is a unit vector, perpendicular to the plane defined by the vector U0 and perpendicular to the wall (e.g., Figure 1 In step S9, the vector m is also iteratively estimated from U0, and then the weights are calculated by Eq. 49 in step S10. The found weights are applied in step S7 and the estimate of U0 is refined until convergence at the output of test S11. At this stage, U0(k) has been estimated for different frames.

[0186] U1 can be derived from it by a relationship of the type described above in Eq. 41. In the variant described here, U1 is determined by Eq. 50 to Eq. 52, which has previously been applied in step S12 to the vector Vbar(f) found in step S7 to obtain the time representation V(t) of the velocity vector. As mentioned above with reference to Figure 2 As can be seen, such an embodiment makes it possible to identify various delays TAU1, TAU2, etc. in the case of different reflecting surfaces. The first delay TAU1 is identified because it is the first peak of V(t) in time after the reception instant of the direct path. Therefore, in equation Eq. 51, tmax(k) is the moment that maximizes the modulus of V(t)k calculated for frame k.

[0187] In test S13, it is verified for each frame that the modulus of V(t=0) is much greater than the modulus of V(t) for t>0. In step S14, the frames that do not meet this condition are discarded. The different delays TAU1, then TAU2, etc. are then determined in step S15 (by removing modules from V(t)k, as compared with those corresponding to the delay TAU1 in Eq.51). Considering that the time t and tmax(k) are first expressed in sample indices (taking time zero as reference for the zero index), the delay TAUm is given by the component tmax found in each iteration m divided by the sampling frequency fs according to Eq.52. The vectors U1, U2, etc. can then also be calculated using equation Eq.50.

[0188] Other parameters may also be determined, notably d0 given by equation Eq. 28 in step S16 (and then checked in test S17 for consistency with regular room data such as d0min=0 and d0max=5m, otherwise the frame comprises errors and may be rejected at step S14).

[0189] Step S18 may comprise further selecting a "good" frame representing the onset of sound with the first reflection. The criterion D(k) for selecting such a frame may be illustrated by way of example by equation Eq. 53, where C(f) i (k) Specifies the amplitude (amplitude in absolute value) detected on ambisonic channel i at the time-frequency sample (t, f) resulting from the first transform (time to frequency) of frame k. ε specifies a non-zero positive value to avoid zero in the denominator in the absence of signal. F specifies the total number of frequency subbands used.

[0190] Therefore, in step S22, only frames whose criterion D(k) is not less than 90% of the maximum value Dmax found in step S21 can be selected from the criteria of all frames D(k) calculated from Eq.53.

[0191] Therefore, in step S18, the value D(k) of all frames is calculated, and then in step S19, U0(k), d0(k), D(k) of different frames are processed. In step S20, the value D(k) is collected to identify the highest value in step S21, and frames with a value D(k) less than 0.9Dmax are eliminated in step S22.

[0192] Finally, in step S23, the selected vector U0 is preferably the median (not the average) vector among the vectors U0 of the various selected frames. The selected distance d0 is also the median among the distances d0 of the various selected frames.

[0193] Of course, the invention is not limited to the embodiment described above by way of example; it extends to other variants.

[0194] The application of the processing to an ambisonic signal (FOA) of order 1 has been described above. For enrichment in terms of spatial resolution, the order may be higher.

[0195] In fact, an order 1 ambisonic representation has been described above, but higher orders can be provided. In this case, the velocity vector calculation is increased by the ratio of the higher order directional components on the component W(f), and the vector Un is implicitly increased by the same number of dimensions. Increasing the dimensions (more than 3) and therefore increasing the spatial resolution makes it possible to better distinguish the vectors U0, U1, ..., Un from each other, and to more easily detect the peak value V(k*TAUn) proportional to (U0-Un) in the time footprint, even when the vectors U0 and Un are close in angle (which occurs in the case of grazing reflections (e.g., when the source is far away from and / or close to the ground)). Therefore, this makes it possible to estimate the parameters U0, U1, d0, etc. that are being sought more finely.

[0196] More generally, the velocity vectors can be replaced by ratios between components of a spatial acoustic representation of the "coincidence" type in the frequency domain, and it is possible to work in the coordinate system characteristic of said spatial representation.

[0197] To overcome situations with several sources for example, the computation of TDVV can be exploited more generally in association with artificial intelligence methods (including neural networks). Some training strategies envisioned (e.g., on footprints from a model or windowed SRIR, and not necessarily from the original signal) can allow the network to learn to exploit consecutive frames to improve detection and estimation compared to a given room situation.

[0198] appendix

[0199]

[0200]

[0201] Eq.13 g0=1;τ0=0

[0202]

[0203] Eq.15 γ0=1

[0204]

[0205]

[0206] Eq.27 d1-d0=τ1c

[0207]

[0208] Eq.44 e -αt =e -α(t-τ) .e -ατ

[0209]

[0210]

Claims

1. A method for processing sound signals acquired by at least one microphone for localizing at least one sound source in a space comprising at least one wall, wherein: - applying a time-frequency transform to the acquired signal, - From the acquired signal, express in the frequency domain a velocity vector V(f), which is a complex number with a real part and an imaginary part, which characterizes the synthesis between: * a first acoustic path directly between the source and the microphone, the first acoustic path being represented by a first vector U0, and * at least one second acoustic path resulting from reflections on said wall and represented by a second vector U1, The second acoustic path has a first delay TAU1 at the microphone relative to the direct path, - and, as a function of the delay TAU1, the first vector U0 and the second vector U1, determining at least one parameter from: * the direction of the direct path, * the distance d0 from the source to the microphone, * the distance z0 from the source to the wall; Therein, the direction of the direct path is determined by the first vector U0 which is an average value over a set of frequencies of the real part of the velocity vector V(f) expressed in the frequency domain.

2. The method according to claim 1, wherein the sound signal is acquired by an ambisonic microphone, and the velocity vector V(f) in the frequency domain is represented by an ambisonic component of order 1 in the form of: V(f)=1 / W(f)[X(f),Y(f),Z(f)] T 。 3. The method according to claim 1, wherein: - applying an inverse transformation from frequency to time to said velocity vector to express said velocity vector in the time domain V(t), - after the duration of the direct path, finding at least the maximum value V(t)max in the expression of the velocity vector as a function of time, and - Thereby deriving said first delay TAU1 corresponding to the time at which said maximum value V(t)max occurs.

4. The method according to claim 3, wherein: - Estimation of the second vector U1 from the first vector U0 and from the maximum value V(t)max by a relationship of the following type: U1=U0–2(U0·V(t)max)V(t)max / ||V(t)max||².

5. The method according to claim 4, wherein: - determining the respective angles PHI0 and PHI1 of the first vector U0 and the second vector U1 relative to the wall, such that: PHI0=arcsin(U0·n R ) and PHI1=arcsin(U1·n R ), where n R is a unit vector and is perpendicular to the wall, and The distance d0 between the source and the microphone is determined from the first delay TAU1 by a relationship of the following type: d0 = (TAU1 x C) / ((cosPHI0 / cosPHI1)-1), where C is the speed of sound.

6. The method according to claim 5, wherein: The distance z0 from the source to the wall is determined by a relationship of the following type: z0=d0(sinPHI0–sinPHI1) / 2.

7. The method according to claim 1, wherein: The space includes a plurality of walls, and: - applying the inverse transformation from frequency to time to said velocity vector to represent it in the time domain V(t) as a series of peaks, - identifying, among the series of peaks, peaks associated with reflections on walls of the plurality of walls, each identified peak having a time abscissa that is a function of a first delay TAUn of the acoustic path resulting from the reflection on the corresponding wall n relative to the direct path, - from each first delay TAUn, the first vector U0 and each second vector Un representing the acoustic path resulting from the reflection on the wall n, at least one parameter is determined from: * the direction of the direct path, * the distance d0 from the source to the microphone, and * At least the distance zn from the source to the wall n.

8. The method according to claim 7, wherein: A peak value associated with a reflection on a wall n has a time abscissa which is a multiple of the delay TAUn associated with this wall n.

9. The method according to claim 7, wherein: A first portion of peaks at the minimum positive time abscissa is preselected in order to identify in said portion each peak associated with a reflection on a wall.

10. The method according to claim 1, wherein: The acquired signal is delivered in the form of successive sample frames, and wherein: - for each frame, estimate the fraction of the frame in which a sound onset is present, and - Selecting frames having a score above a threshold for determining the at least one parameter.

11. The method according to claim 1, wherein: The signal acquired by the microphone is in the form of a series of samples to which a weighting window is applied with an exponential variation that decreases over time.

12. The method according to one of the preceding claims, wherein: The weights q(f), each associated to a frequency band f, are applied iteratively to the velocity vector V(f) in the frequency domain according to an expression of the type: q(f)=exp(-|Im(V(f))·m| / (||Im(V(f))||)), where Im(V(f)) is the imaginary part of the complex velocity vector and m is a unit vector perpendicular to the plane defined by the vector U0 and the normal to the wall.

13. A device for processing sound signals, comprising a processing circuit for implementing the method according to one of the preceding claims.

14. A non-transitory computer storage medium storing a computer program comprising instructions for implementing the method according to one of claims 1 to 12 when the instructions are executed by a processor of a processing circuit.

Images