EP3384688B1 - Successive decompositions of audio filters - Google Patents

Description

The present invention relates to the field of the restitution of sound data.

It finds applications, in particular, but not exclusively, in the context of telecommunications services offering spatialized sound reproduction, such as for example in the case of an audioconference between several speakers, of a cinema trailer broadcast or broadcasting of any type of multichannel audio content. The invention also applies in the case of telecommunications terminals, in particular mobile, for which it is envisaged a sound reproduction with a stereophonic listening system (a headset for example) allowing the listener to position the sound sources in space.

To this end, the invention uses invariant and stationary linear systems that can be characterized by a set of filters depending on a direction between the sound source and one of the listener's auditory canals.

This set of filters represents the directivity of the system. Filters can be represented in their time (as an impulse response) or frequency (as a transfer function) form.

By way of example, an individual or an artificial head with a microphone at the entrance to each auditory canal are particular cases of such an invariant and stationary linear system. In this case, the system can be characterized by its transfer functions, specific to each individual.

The transfer functions define the spatial hearing characteristics of the individual by taking into account in particular the reflections linked to his morphology.

The transfer functions are classically called transfer functions of the HRTF type for “ Head Related Transfer Function ”, when the filters are given in the frequency domain, and HRIR for “ Head Related Impulse Response ”, when the filters are given in the time domain. . It is possible to go from one representation to another by a Fourier transform.

HRTF transfer functions are therefore a set of complex values. It is possible to return to real values by taking their respective moduli: we thus obtain the moduli of the HRTF.

The division of each module by the spatial average of the modules for a given frequency makes it possible to obtain what is commonly called in the literature “ Directional Transfer Function”.

The invention can be generalized to the directivities of systems having different shapes and / or numbers of sensors (for example a mobile telephone with 3 microphones). Without prejudicing the generalization of the invention to any linear system that can be characterized by ORTFs, and in order to facilitate understanding of the invention, the particular case of DTF transfer functions is considered hereinafter. In fact, it is possible to pass from the DTF transfer functions to the HRTF transfer functions by calculating minimum phase filters associated with the DTF transfer functions and by adding a delay to them modeling the propagation delays between the capsules (inter-aural delay by a human ). The personalization of these delays is obtained by other techniques well known and not described here.

One technique using HRTF-type transfer functions is binaural synthesis. This technique is based on the use of so-called “binaural” filters, which reproduce the acoustic transfer functions between the sound source (s) and the auditory canals of the listener. These filters are used to simulate auditory localization cues that allow a listener to locate sound sources in a real listening situation.

The techniques linked to binaural synthesis are therefore based on a pair of binaural signals which feeds a reproduction system. The two binaural signals can be obtained by signal processing, by filtering a monophonic signal by binaural filters which reproduce the properties of the acoustic propagation between the source placed at a given position and each of the listener's ear canals.

Binaural synthesis can be used for different reproductions such as, for example, a reproduction by means of a headset with two earphones, or by means of two loudspeakers. The goal is reconstruction a sound field at the level of the listener's ears practically identical to that induced by real sources in space.

Binaural filters take into account all the acoustic phenomena which modify the acoustic waves in their path between the source and the auditory canals of the listener. Acoustic phenomena include in particular diffraction from the listener's head and reflections on the auditory pinna and upper torso of the user.

These acoustic phenomena vary according to the position of the sound source relative to the listener and the variations allow the listener to locate the source in space. Indeed, these variations determine a form of acoustic coding of the position of the source. The hearing system of an individual knows, by learning, how to interpret this coding to locate the sound source (s).

Nevertheless, the acoustic diffraction / reflection phenomena strongly depend on the morphology of the listener. A quality binaural synthesis therefore relies on binaural filters which best reproduce the acoustic coding that the listener's body naturally produces, taking into account the individual specificities of his morphology.

When these conditions are not respected, a degradation in the performance of binaural rendering is induced, which is reflected in particular by an intracranial perception of the sources and confusion between the front and rear locations.

Thus, the binaural filters represent the acoustic transfer functions or HRTF type transfer functions which model the transformations generated by the listener's torso, head and pinna on the acoustic signal coming from a sound source. Each sound source position is associated with a pair of HRTF functions, one for each ear. In addition, these HRTF transfer functions carry the acoustic imprint of the morphology of the individual on which they were measured.

In a well-known manner, the HRTF transfer functions are obtained during a measurement phase. A selection of directions which cover more or less finely the whole of the space surrounding the listener is fixed. For each direction, the left and right HRTF transfer functions are measured using microphones inserted at the entrance to the listener's ear canals. Usually, a listener-centered sphere is defined in this way.

For a good quality measurement, the measurement must be carried out in an anechoic chamber, or an anechoic chamber, so that only the reflections and acoustic phenomena linked to the listener are taken into account. In the end, if M directions are measured, we obtain, for a given listener, a database of 2M HRTF type transfer functions (because two right and left auditory channels) representing, for each auditory canal, each of the positions of the sources. . These techniques therefore require carrying out the measurements on the listener directly. The duration of such a measuring operation is very long because it is necessary to measure a large number of directions.

Some individuals thus spend long hours in the laboratory in order to have the acoustic signature associated with their physiognomy analyzed in detail, as well as their ability to perceive sound space in three dimensions. These individuals then benefit from binaural listening shaped from the analysis results, offering comfort and a high quality sound impression.

In order to provide this quality and comfort to a wider set of listeners, particularly in the context of services intended for the general public, it is necessary to have personalized filters for each of the listeners.

However, it is hardly conceivable to measure all the customers of a service in deaf rooms (which are rare and expensive). In addition, the duration and arduousness of the measures are difficult to bear for the general public.

It is thus desirable to have solutions making it possible to obtain the acoustic signatures of individuals in a rapid, reliable and not very intrusive manner in order to be able to generalize the results obtained in an anechoic chamber on a small number of subjects, to a very large population.

A practical solution starting to emerge is to offer the user to measure his own HRTF transfer functions in his usual listening place in order to emulate his listening experience on a headset. in the studio or in the living room. The drawbacks associated with this type of solution are linked to the fact of measuring only a small number of fixed positions and of making it difficult to separate the information linked to the broadcasting device itself and the listening place. Various studies have been devoted to the development of methods making it possible to reduce certain practical constraints such as dynamic measurement ( “Dynamic measurement of room impulse responses using a moving microphone”, Ajdler, Sbaiz, Vetterli, 2007 ) or reciprocal measurement, in which the roles of the microphone and the speaker are reversed (“ Fast head-related transfer function measurement via reciprocity”, Zotkin, Duraiswami, Grassi, Gumerov, 2006). The applications of this solution are limited to professional mixing studios or "home-cinema" installations.

Different avenues offering alternative solutions are explored. A first approach consists in calculating the filters on the basis of the acquisition of the listener's morphology and in particular of his flag. I Wen Zhang et al. thus propose to decompose an HRTF transfer function on a basis of independent functions (" Efficient Continuous HRTF Model Using Data Independent Basis Functions: Experimentally Guided Approach ", IEEE Transactions on Audio, Speech and Language Processing, vol. 17, no. 4, May 1, 2009, pages 819-829 ). Personalization can also be based on the transformation of non-individual HRTF transfer functions extracted from a database including the morphologies associated with HRTF transfer functions ( "Individualization of spectral indices for binaural synthesis: research and exploitation of interindividual similarities for the adaptation or reconstruction of HRTF", Guillon, P, PhD Thesis, University of Maine, Le Mans, France, 2009 ; see also the patent FR 2958825 ).

The transformation of the HRTF transfer functions to adapt them to a given individual is then controlled by the comparison of the morphologies of the original flag from the database and the target flag of the given individual. This comparison is based on a technique of matching the three-dimensional meshes of the pavilions. Another method consists in using morphological parameters to create or deform a three-dimensional mesh, which will then be used for a detailed calculation and a numerical simulation of the HRTF transfer functions of the individual, by finite elements of border for example. It is also possible, from the morphological parameters of a given individual, to search in a database a third individual with similar morphological parameters.

Some works propose to use as input a three-dimensional mesh of the morphology of the subject and more particularly of his pinna, and others of the measurements of morphological parameters of the users. One method of acquiring pinna morphology is to use a three-dimensional scan, but this method is sometimes problematic as it requires both specific hardware and implementation.

Alternative solutions are developed either by deriving three-dimensional scans from a set of photographs ( “Reconstructing head models from photographs for individualized 3D-audio processing”, Dellepiane, Pietroni, Tsingos, Asselot, Scopigno, 2008 ), or by using methods resulting from image processing making it possible to obtain three-dimensional meshes from a camera and reconstruction techniques ( "shape from shading", "shape from structured light ") or even from Kinect ™ type sensors associated with depth analysis techniques.

Other works attempt to set up learning methods that combine two opposing approaches.

The first approach consists in studying the capacity of the auditors to appropriate generic HRTF transfer functions which are not initially adapted to them. The second approach, on the contrary, suggests learning by a computer of the reactions of a user participating in an interactive game or responding to an interactive questionnaire. The computer iteratively reconstructs the set of HRTF transfer functions suitable for the user from the observation of his localization performance and / or his responses.

However, storing the sets of transfer functions, transmitting and loading them is complicated due to the size of the data representing each set of transfer functions.

In addition, the solutions necessary for the personalization of a set of transfer functions, to adapt it to a given listener, other than the measurements in an anechoic room do not yet exist. However, as explained above, measurements in anechoic rooms are complex and costly in terms of hardware and software resources as well as time, and are therefore not transposable to a large population.

The present invention improves the situation.

To this end, a first aspect of the invention relates to a method for processing individualized data representative of the directivity of an individualized audio system, according to claim 1.

The successive decomposition in a first base of N independent components common to all the individuals of the first set, then in a second base of P independent components advantageously makes it possible to compress the stored data. To this end, the numbers N and P of independent components can be chosen as a function of criteria linked to the size of the data stored and to the precision desired for the filter sets. The second basis of P independent components can be a basis of spherical harmonics of order P and the second set of weighting coefficients is a set of spherical coefficients.

• obtention de données morphologiques courantes d'un nouvel individu ;
• sélection d'un individu parmi l'ensemble initial par comparaison entre les données morphologiques courantes et les ensembles de données morphologiques des individus de l'ensemble initial ;
• application d'une transformation au deuxième jeu de coefficients de pondération associé à l'individu sélectionné en vue d'obtenir un deuxième jeu de coefficients de pondération transformé, la transformation étant déterminée à partir des données morphologiques courantes ;
• stockage du deuxième jeu de coefficients de pondération transformé en association avec un identifiant du nouvel individu.

The decomposition in a base of spherical harmonics advantageously makes it possible to have sets of spherical coefficients which can be easily transformed by application of transformations comprising a rotation. Each individual of the initial set of individuals is further associated with a set of morphological data, and the method further comprises the following steps:

• obtaining current morphological data of a new individual;
• selecting an individual from the initial set by comparison between the current morphological data and the morphological data sets of the individuals of the initial set;
• applying a transformation to the second set of weighting coefficients associated with the selected individual in order to obtain a second set of transformed weighting coefficients, the transformation being determined from current morphological data;
• storing the second set of weighting coefficients transformed in association with an identifier of the new individual.

The storage of the sets of filters in the form of morphological data advantageously makes it possible to easily apply transformations in order to adapt the second set of weighting coefficients of an individual from the initial set. The initial set can thus be used as a starting point for a quick and non-binding determination of sets of filters for users other than the users of the initial set.

In addition, the transformation can comprise at least the application of a rotation matrix to the set of spherical coefficients associated with the selected individual.

The application of a rotation matrix to a set of coefficients in a base of spherical coordinates makes it possible to easily apply a rotation to the directivities of the sets of filters represented, and thus makes it possible to easily adapt the sets of filters of the individuals of the 'initial set.

• application d'une homothétie aux N composantes indépendantes, l'homothétie étant déterminée à partir des données morphologiques courantes, afin d'obtenir N composantes indépendantes transformées ;
• multiplication du jeu de coefficients sphériques transformé par une matrice formée par les N composantes indépendantes transformées, en vue d'obtenir un jeu de filtres modifié en association avec l'identifiant du nouvel individu.

In addition, the method can further comprise the following steps:

• application of a homothety to the N independent components, the homothety being determined from current morphological data, in order to obtain N transformed independent components;
• multiplication of the set of spherical coefficients transformed by a matrix formed by the N transformed independent components, with a view to obtaining a set of filters modified in association with the identifier of the new individual.

• multiplication du jeu de coefficients sphériques transformé par une matrice formée par les N composantes indépendantes, en vue d'obtenir un nouveau jeu de filtres ;
• application d'une homothétie par ré-échantillonnage temporel du nouveau jeu de filtres en vue d'obtenir un jeu de filtres modifié en association avec l'identifiant du nouvel individu.

As a variant, the method may further comprise the following steps:

• multiplication of the set of spherical coefficients transformed by a matrix formed by the N independent components, with a view to obtaining a new set of filters;
• application of a scaling by temporal resampling of the new set of filters with a view to obtaining a set of filters modified in association with the identifier of the new individual.

In addition, when the new set of filters is in the frequency domain, the method comprises the application of an inverse Fourier transform to the new set of filters prior to the temporal resampling.

According to one embodiment, the morphological data relate at least to the auditory pinna of the user.

Thus, the morphological data having the most influence on the filter set associated with an individual are taken into account when determining a new set for a new individual.

According to embodiments of the invention, the filters can be transfer functions in the frequency domain (or the modules of these transfer functions), each independent component can be a function having a non-zero spectrum in a frequency band. given, and the given frequency bands may be distinct.

In addition, the independent components can be expressed in logarithmic scale of frequencies.

The use of a logarithmic scale makes it possible to translate more precisely the perceptual functioning of the human ear (more sensitive in the high frequencies than the low frequencies).

According to one embodiment, for each set of filters obtained, the modules of the set of filters can be deconvolved by a spatial average of the modules of the set of filters and the N independent components can be determined from the deconvolved modules.

This embodiment makes it possible to reduce the variance of the filters by eliminating the part common to all the filters and makes it possible to work on real values rather than complex (DTF).

A second aspect of the invention relates to a computer program product comprising program code instructions recorded on a medium readable by a computer, for the execution of the steps of the method according to the first aspect of the invention.

A third aspect relates to a device for processing individualized data representative of the directivity of an audio system, according to claim 10.

• la figure 1 est un diagramme représentant les étapes d'un procédé de traitement de données selon un mode de réalisation de l'invention;
• la figure 2 représente une décomposition d'un jeu de filtres dans une base de composantes indépendantes selon un mode de réalisation de l'invention;
• la figure 3 représente des composantes indépendantes obtenues à partir d'un ensemble initial de jeux de filtres selon un mode de réalisation de l'invention ;
• la figure 4 illustre des figures de directivité pour une même composante indépendante pour huit individus distincts, selon un mode de réalisation de l'invention ;
• la figure 5 illustre un dispositif selon un mode de réalisation de l'invention.

Other characteristics and advantages of the invention will become apparent on examination of the detailed description below, and of the appended drawings in which:

• the figure 1 is a diagram representing the stages of a data processing method according to one embodiment of the invention;
• the figure 2 represents a decomposition of a set of filters in a base of independent components according to one embodiment of the invention;
• the figure 3 represents independent components obtained from an initial set of sets of filters according to one embodiment of the invention;
• the figure 4 illustrates directivity figures for the same independent component for eight distinct individuals, according to one embodiment of the invention;
• the figure 5 illustrates a device according to an embodiment of the invention.

The figure 1 is a diagram illustrating the general steps of a data processing method according to one embodiment of the invention.

In a step 101, for each individual of an initial set of individuals, a personalized set of filters is obtained. The initial set of individuals is a restricted set of individuals for which the solutions of the state of the art could be applied in order to obtain a set of filters personalized for each of the individuals.

For example, each individual was tested in an anechoic chamber in order to obtain at least one personalized set of filters. Typically, two sets of custom filters are obtained for each individual, one for each ear canal.

No restriction is however attached to the way in which the sets of filters were acquired in step 101. The sets of filters of the initial set of individuals are stored in a step 102, for example in a memory of a device implementing the method according to the invention.

Filter sets can be expressed as the coefficients of a matrix. As detailed previously, the example of HRTF transfer functions in the frequency domain is considered without limitation as sets of filters.

In a step 103, N independent components common to the sets of filters obtained are determined. For example, the decomposition into independent components disclosed in the document “Independent conduct analysis”, Stone JV, 2004, John Wiley & Sons , can be applied to the modules of the filters of a set (HRTF transfer functions), the modules being optionally deconvolved (frequency division) by the spatial mean of the set of filters. Such an operation is equivalent to removing from the HRTF transfer functions the frequency components common to all the filters. Such deconvolved modules are called DTF hereafter. Modules can be optionally smoothed in order to keep only the perceptually relevant frequency variations.

Thus, any HRTF (or DTF) transfer function modulus of the initial set of individuals can be reconstructed by a linear combination of independent components weighted by weighting coefficients, as shown in Figure figure 2 .

A first matrix 200 of coefficients w _{i, j} , i varying between 1 and M (2 * M being the total number of directions measured, M filters corresponding to one of the two ears of the listener) and j varying between 1 and N represents the weighting coefficients obtained after decomposition of the filters corresponding to one of the ears of a set on a basis formed of the N independent components.

A second matrix 201 of coefficients c _{n, f} , with n varying between 1 and N and f varying between 1 and F represents the coefficients of the N independent components, each row corresponding to one of the N independent components.

A third matrix 202 represents a set of filters (the deconvolved modules of the HRTF transfer functions in the preceding example) for an individual, for an ear, obtained in step 101, and comprises coefficients d _{m, f} , m varying between 1 and M and f varying between 1 and F. Each row m of the third matrix 202 represents a filter for a given direction of space, and each column corresponds to a frequency (or a band of frequencies more precisely), translating thus the spectrum of HRTF transfer functions. The HRTF transfer function modules can be in logarithmic or linear scale, abscissa or ordinate, which results in four distinct configurations (linear, linear), (logarithmic, linear), (linear, logarithmic) and (logarithmic, logarithmic). A logarithmic scale on the abscissa amounts to resampling the spectrum of a transfer function (a row of the matrix 202) with a logarithmic and non-linear frequency step, which more precisely reflects the perceptual functioning of the human ear. (more sensitive in high frequencies than low frequencies). A logarithmic scale on the ordinate amounts to considering 20 * log ₁₀ (abs (HRTF)), abs (HRTF) representing the modules of the HRTF transfer functions.

As indicated above, each row of the second matrix 201 represents an independent component, each coefficient of the row corresponding to the energy of the independent component in a given frequency band.

The figure 3 presents a set of spectra for N = 20 independent components, according to one embodiment of the invention. These N independent components can be determined in step 103 described above, from the set of third matrices 202 of the individuals of the initial set. As shown on the figure 3 , each independent component can correspond to a band of the spectrum in which the energy is non-zero, the independent components having disjoint spectrum supports.

The first matrix 200 depends on the azimuth and the elevation (on the ordinate) and on the weights assigned to each independent component (on the abscissa). The set of coefficients w _{m, n} for a given column n, represents, for an individual, the directivity for an independent component for the component n. Each index m corresponding to a measurement for a direction (azimuth (m), elevation (m)). The first matrix 200 is determined in a step 104, by decomposing each of the sets of filters obtained in step 101, into the base formed from the N independent components. The coefficients of a column of the first matrix 200 represent the values of the weights for an independent component for the different measurement directions. They thus represent a figure of spatial directivity.

The figure 4 illustrates such figures of spatial directivity for eight individuals of the initial set of individuals, according to one embodiment of the invention. We see on the figure 4 that the spatial directivities look the same from one individual to another and that rotations can be applied to bring these spatial directivities closer together.

The figure 4 presents in particular the weighting coefficients of the first matrix 200, for each individual, applied to the third independent component (third row of the second matrix 201) for eight different individuals. These are therefore the respective third columns of the first matrices 200 for the eight individuals. The columns are re-cut by the same elevation and represented in a three-dimensional way. The abscissa corresponds to the azimuth expressed in degrees, and the ordinate corresponds to the elevation in degrees. The third dimension is represented by color variations (in shades of gray on the figure 4 ). The shades of gray represent the values of the weighting coefficients. The figure 4 can thus be interpreted as a set of directivity figures for the third independent components of eight individuals of the initial set.

In order to search for such rotations and to further reduce the quantity of information describing an individual, the invention provides for breaking down in a step 105, each set of weighting coefficients (each first matrix 200 of an individual from the initial set ) in a base of P independent functions in the mathematical sense, for example in a base of spherical harmonics of order P-1, in order to obtain a set of spherical coefficients. The choice of the base of spherical harmonics allows the easy application of rotations to the sets of spherical coefficients in order to recalculate a new set of spherical coefficients following a rotation of the measurement reference frame, which is not the case with a basis of independent components in two dimensions.

dans lequel HS_m,i est un vecteur de la taille du nombre de mesures et dont la valeur vaut la valeur de l'harmonique sphérique i pour la direction de mesure correspondant à l'indice m (azimut(m), elevation(m))The determination of a set of spherical coefficients amounts to carrying out a spatial Fourier transform of the directivities (of a first matrix 200 therefore). The decomposition of the directivities cw _{ic, p} for the component independent ic to the order P in spherical harmonics is expressed as follows: $${\displaystyle \sum _{p=0}^{P-1}{\mathit{cw}}_{\mathit{ic},p}}\times {\mathit{HS}}_{m,p}=w_{m,\mathit{ic}}$$

in which HS _{m, i} is a vector of the size of the number of measurements and whose value is the value of the spherical harmonic i for the measurement direction corresponding to the index m (azimuth (m), elevation (m) )

From each set of filters of an individual of the initial set, we thus obtain a set of spherical coefficients. In a step 106, each set of spherical coefficients obtained in association with an identifier of the individual to which it corresponds.

The decomposition into spherical harmonics thus makes it possible to fully characterize a set of filters corresponding to the directivity of the ear canal of one of the individuals by means of spherical coefficients cw _{ic, p} which are of dimension P * N, where P-1 is the order of the decomposition into spherical harmonics and N the number of independent components.

The basis of spherical harmonics and the N independent components are common to all individuals, and therefore to all HRTF or DTF filter sets.

The successive application of a decomposition on a basis of N independent components then on a basis of spherical harmonics allows a first advantage which is to reduce the quantity of information to be analyzed and makes it possible to compress the HRTF filter sets.

By way of example, current solutions provide for the acquisition of HRTF filter sets comprising 1680 directions, for two ears of an individual, for a filter size of 512 points at 48 kHz sampling frequency, i.e. 1680 * 2 * 512 = 1720320 floats.

A decomposition on N = 64 independent components, then on a basis of spherical harmonics of order 20 allows an almost perfect reconstruction by storing only 2 * 64 * (20 + 1) = 2688, that is to say a compression factor of 640. It also suitable to store the 64 components independent (64 * 512 = 32768). As for the values of the basis of spherical harmonics, they can be calculated or stored in tables. The latter and the N independent components are common to all the sets of filters of individuals.

The values N and P can thus be chosen as a function of a compromise between the level of compression and storage constraints, in order to ensure that the complexity of the HRTF transfer functions is reduced after successive decompositions.

A second advantage arising from the successive application of a decomposition on a basis of N independent components then on a basis of spherical harmonics is linked to the customization of the transfer functions HRTF or DTF. In fact, the steps 101 to 106 detailed previously have been applied to an initial set of individuals, the set comprising a limited number of individuals (around fifty for example) due to the complexity linked to the acquisition of the functions of HRTF transfer to step 101. However, the sets of spherical coefficients determined for this restricted number of individuals can also be used to rapidly determine a set of filters for a new individual, not belonging to the initial set. The advantage of decomposition on a basis of spherical harmonics is that, to perform a rotation of the whole of a set of HRTF or DTF filters (rotation of the measurement reference frame), it suffices to apply a rotation matrix à to the corresponding set of spherical coefficients cw (see in this regard the thesis "Representation of acoustic fields, application to the transmission and reproduction of complex sound scenes in a multimedia context", Daniel J, University of Paris 6, 2000 ).

To this end, in a step 107, the method according to the invention can comprise obtaining current morphological data of a new individual. Indeed, it is possible to switch from a morphology of an individual to a morphology of a new individual by applying a transformation which can include a simple rotation defined by three axes of rotation (θ, ϕ, ρ). In addition a homothety To can be applied. The transformation parameters can be obtained by comparison between two three-dimensional meshes of two individuals, and more generally by comparison between data morphological data of the individuals of the initial set and the current morphological data of the new individual.

The parameters θ, ϕ, ρ and λ can also depend on a factor f representing a frequency band or a set of frequency bands.

To this end, morphological data of the individuals of the first set can also be obtained in step 101 described above and then stored in step 102. These morphological data can describe the geometry of the linear system whose directivity is characterized by the set of associated filters.

No restriction is attached to the means used to obtain the morphological data of the individuals of the initial set as well as the current morphological data. For example, they can be obtained by direct measurements on the individual, from photographs or even using the three-dimensional scanner of the Kinect ™ type for example. The morphological data related to the flag of the individual can be particularly taken into account in determining the transformation parameters. Indeed, the bell is the most influencing factor in the information of HRTF filter sets.

Thus, at a step 108, the current morphological data is compared with the set of morphological data of the individuals of the initial set, with a view to selecting, at a step 109, an individual from the initial set. For example, the individual of the initial set having the parameters closest to the current parameters is selected. By way of example by considering the morphological data to be stored and compared as being 3D meshes of the pavilions, one can search in the base for the 3D mesh which, after a rotation and a homothety, will be closest to the current 3D mesh. No restriction is attached to the criterion used to characterize the proximity of the morphological parameters.

In a step 110, a transformation to be applied to the set of spherical coefficients associated with the selected individual is determined from the current morphological data. The transformation is determined by determining the first parameters that allow data to be passed current morphological data to the morphological data of the individual selected from the initial set. In the example above, the values of the rotation found in the previous step are used. From these first parameters, are deduced the transformation parameters making it possible to transform the set of filters of the selected individual into a new set of filters.

In general, such a method amounts to determining a transformation model and its parameters on the sets of filters characterizing the directivities of the systems from a signal point of view, another transformation model and its parameters describing the geometries, shapes or morphologies of the systems, and also to determine a function to match these two models.

The transformation is then applied to the set of spherical coefficients associated with the selected individual in order to obtain a set of spherical coefficients transformed in a step 111.

The transformed set of spherical coefficients is stored in association with an identifier of the new individual at a step 112.

• dilatation fréquentielle des N composantes indépendantes de la deuxième matrice 201 d'un facteur À, puis application de la multiplication matricielle de la figure 2 pour obtenir un nouveau jeu de filtres HRTF ou DTF. Dans ce calcul, la première matrice 200 est obtenue à partir du jeu de coefficients sphériques transformé ;
• application de la multiplication matricielle de la figure 2, la première matrice 200 étant obtenue à partir du jeu de coefficients sphériques transformé, puis application d'une transformée de Fourier inverse pour revenir dans le domaine temporel, et application de l'homothétie λ par un ré-échantillonnage temporel. La transformée de Fourier inverse est ainsi appliquée dans l'exemple particulier des jeux de filtres HRTF ou DTF. Toutefois, elle n'est pas nécessaire dans le cas où l'invention est appliquée à des jeux de filtres dans le domaine temporel.

In addition, the λ homothety can be applied in different ways:

• frequency expansion of the N independent components of the second matrix 201 by a factor λ, then application of the matrix multiplication of the figure 2 to get a new set of HRTF or DTF filters. In this calculation, the first matrix 200 is obtained from the transformed set of spherical coefficients;
• application of the matrix multiplication of the figure 2 , the first matrix 200 being obtained from the set of transformed spherical coefficients, then application of an inverse Fourier transform to return to the time domain, and application of the homothety λ by time resampling. The inverse Fourier transform is thus applied in the particular example of the HRTF or DTF filter sets. However, it is not necessary in the case where the invention is applied to sets of filters in the time domain.

The figure 5 represents device 500 according to one embodiment of the invention.

The device 500 comprises a random access memory 503 and a processor 502 for storing instructions allowing the implementation of steps 101 to 112 of the method described above with reference to figure 1 . The device also comprises a database 504 for storing data intended to be kept after the application of the method, in particular the sets of spherical coefficients, the independent components, and optionally the base of spherical harmonics. The device 500 further comprises an input interface 501 intended to receive the sets of filters of the initial set of individuals, and optionally the morphological parameters of the individuals of the initial set and the current morphological parameters. The device 500 further comprises an output interface 505 for the transmission of data resulting from the application of the method according to the invention. For example, the output interface can transmit the modified filter set or the transformed spherical coefficient set obtained for the new user.

The present invention is not limited to the embodiments described above by way of examples; it extends to other variants, included within the scope of the protection defined by the following claims. Thus, the present invention makes it possible to improve the quality of immersive audio rendering in binaural systems, since it makes it possible to easily obtain a set of filters personalized for an individual from morphological data, without requiring long and expensive measurements on each of the individuals. The invention thus applies to communications services including audio conferencing and content distribution services or applications (music, films, games, user interfaces, etc.). In addition, the present invention allows the compression of the sets of filters (HRTF or DTF for example), which facilitates the storage, exchange or loading thereof.

Claims (10)

1. Method for processing individualized data representative of the directivity of an individualized audio system, said method comprising the following steps:

   - obtaining (101), for each individual of an initial set of individuals, at least one personalized set of binaural filters;

   - determining (103) N independent components common to the sets of filters obtained;

   - decomposing (104) each of the sets obtained into a first base formed from the N independent components with a view to obtaining, for each set of filters, a first set of weighting coefficients;

   - decomposing (105) each first set of weighting coefficients into a second base of P independent components, in order to obtain a second set of weighting coefficients, the second base of P independent components being a base of spherical harmonics of order P-1 and the second set of weighting coefficients being a set of spherical coefficients;

   - storing (106) each second set of weighting coefficients obtained in association with an individual identifier from among the initial set of individuals;

   wherein each individual of the initial set of individuals is further associated with a set of morphological data, said method further comprising the following steps:

   - obtaining (107) current morphological data of a new individual;

   - selecting (109) an individual from among the initial set by comparison between the current morphological data and the sets of morphological data of the individuals of the initial set;

   - applying (111) a transformation to the second set of weighting coefficients that are associated with the individual selected with a view to obtaining a transformed second set of weighting coefficients, said transformed second set of weighting coefficients being a transformed set of spherical coefficients, the transformation being determined on the basis of the current morphological data with a view to obtaining a modified set of filters;

   - storing (112) the transformed second set of weighting coefficients in association with an identifier of the new individual.
2. Method according to Claim 1, wherein the transformation comprises at least the application of a rotation matrix to the set of spherical coefficients that is associated with the individual selected.
3. Method according to Claim 2, wherein the method further comprises the following steps with a view to obtaining the modified set of filters on the basis of the transformed set of spherical coefficients:

   - applying a homothety, by frequency dilation, to the N independent components, said homothety being determined on the basis of the current morphological data, in order to obtain N transformed independent components;

   - multiplying the transformed set of spherical coefficients by a matrix formed by the N transformed independent components, with a view to obtaining the modified set of filters in association with the identifier of the new individual.
4. Method according to Claim 2, wherein the method further comprises the following steps with a view to obtaining the modified set of filters on the basis of the transformed set of spherical coefficients:

   - multiplying the transformed set of spherical coefficients by a matrix formed by the N independent components, with a view to obtaining a new set of filters;

   - applying a homothety by temporally resampling the new set of filters with a view to obtaining the modified set of filters in association with the identifier of the new individual.
5. Method according to Claim 4, wherein when the new set of filters is in the frequency domain, the method comprises the application of an inverse Fourier transform to the new set of filters prior to the temporal resampling.
6. Method according to one of the preceding claims, wherein the morphological data relate at least to the ear auricle of the user.
7. Method according to one of the preceding claims, wherein the binaural filters are transfer functions in the frequency domain, wherein each independent component is a function having a non-zero spectrum in a given frequency band, and wherein the data frequency bands are distinct.
8. Method according to Claim 7, wherein the independent components are expressed on a logarithmic scale of frequencies.
9. Computer program product comprising program code instructions stored on a medium that is readable by a computer, for the execution of the steps of the method according to any one of Claims 1 to 8.
10. Device for processing individualized data representative of the directivity of an audio system, said device (500) comprising a processor configured to:

    - obtain, via an input interface (501) of the device, for each individual of an initial set of individuals, at least one personalized set of binaural filters;

    - determine N independent components common to the sets of filters obtained;

    - decompose each of the sets obtained into a first base formed from the N independent components with a view to obtaining, for each set of filters, a first set of weighting coefficients;

    - decompose each first set of weighting coefficients into a second base of P independent components, in order to obtain a second set of weighting coefficients, the second base of P independent components being a base of spherical harmonics of order P-1 and the second set of weighting coefficients being a set of spherical coefficients;

    - storing, in a memory (504) of the device, each second set of weighting coefficients obtained in association with an individual identifier from among the initial set of individuals;

    wherein each individual of the initial set of individuals is further associated with a set of morphological data, said processor further being configured to:

    - obtain current morphological data of a new individual;

    - select an individual from among the initial set by comparison between the current morphological data and the sets of morphological data of the individuals of the initial set;

    - apply a transformation to the second set of weighting coefficients that are associated with the individual selected with a view to obtaining a transformed second set of weighting coefficients, said transformed second set of weighting coefficients being a transformed set of spherical coefficients, the transformation being determined on the basis of the current morphological data with a view to obtaining a modified set of filters; and

    - store the transformed second set of weighting coefficients in association with an identifier of the new individual.

Images