EP3427260B1 - Optimized coding and decoding of spatialization information for the parametric coding and decoding of a multichannel audio signal - Google Patents

Description

The present invention relates to the field of encoding / decoding digital signals.

The coding and decoding according to the invention is suitable in particular for the transmission and / or storage of digital signals such as audio-frequency signals (speech, music or others).

More particularly, the present invention relates to the parametric multichannel encoding and decoding of multichannel audio signals.

The invention is therefore concerned with multichannel signals, and in particular with binaural signals which are sound signals recorded with microphones placed at the entrance to the duct of each ear (of a person or of a mannequin) or else synthesized. artificially through filters known as HRIR (Head-Related Impulse Response) filters in the time domain or HRTF (Head-Related Transfer Function) in the frequency domain, which are a function of the direction and distance of the sound source and morphology of the subject. Binaural signals are associated with listening typically through headphones or earphones and have the advantage of representing a spatial image giving the illusion of being naturally in the middle of a sound scene; it is therefore a reproduction of the soundstage in 3D with only 2 channels. It will be noted that it is possible to listen to binaural sound on loudspeakers by means of complex processing operations to invert the HRIR / HRTF filters and to reconstitute binaural signals.

A distinction is made here between binaural signals and stereo signals. A stereo signal is also made up of two channels, but in general it does not allow perfect reproduction of the sound scene in 3D. For example, a stereo signal can be built by taking a given signal on the left channel and a zero signal on the right channel, listening such a signal will give a sound source location on the left but in a natural environment this artifice is not possible because the signal to the right ear is a filtered version (including a time shift and attenuation) of the signal to the left ear depending on the body type of the person.

Parametric multichannel coding is based on the extraction and coding of parameters of spatial information so that in decoding these spatial characteristics can be used to recreate the same spatial image as in the original signal. Examples of codecs based on this principle can be found in the 3GPP e-AAC + or MPEG Surround standards.

We consider here by way of example the case of parametric stereo coding with N = 2 channels, insofar as its description is simpler than in the case of N> 2 channels.

A parametric stereo encoding / decoding technique is for example described in the document of J. Breebaart, S. van de Par, A. Kohlrausch, E. Schuijers, entitled "Parametric Coding of Stereo Audio" in EURASIP Journal on Applied Signal Processing 2005: 9, pp. 1305-1322 . This example is repeated with reference to figures 1 and 2 describing respectively an encoder and a parametric stereo decoder.

So the figure 1 describes a stereo encoder receiving two audio channels, a left channel (denoted L for Left in English) and a right channel (denoted R for Right in English).

The time signals L ( n ) and R ( n ), where n is the integer index of the samples, are processed by the blocks 101, 102, 103 and 104 which perform a short-term Fourier analysis. The transformed signals L [ k ] and R [ k ] , where k is the integer index of the frequency coefficients, are thus obtained.

• Le "downmix" passif qui correspond à un matriçage direct des canaux stéréo pour les combiner en un seul signal - les coefficients de la matrice de downmix sont en général réels et de valeurs prédéterminées (fixes);
• Le "downmix" actif (adaptatif) qui inclut un contrôle de l'énergie et/ou de la phase en plus de la combinaison des deux canaux stéréo.

The block 105 performs a channel reduction or “downmix” processing in order to obtain in the frequency domain from the left and right signals, a monophonic signal hereinafter referred to as a mono signal. Several techniques have been developed for channel reduction or stereo to mono downmix processing. This "downmix" can be carried out in the time or frequency domain. In general, we distinguish:

• The passive "downmix" which corresponds to a direct matrixing of the stereo channels to combine them into a single signal - the coefficients of the downmix matrix are generally real and of predetermined (fixed) values;
• The active (adaptive) downmix which includes energy and / or phase control in addition to the combination of the two stereo channels.

Spatial information parameter extraction is also performed in block 105. The extracted parameters are as follows.

où L[k] et R[k] correspondent aux coefficients spectraux (complexes) des canaux L et R, chaque bande de fréquence d'indice b = 0, ..., B - 1 comprend les raies fréquentielles dans l'intervalle [k_b, k _b+1 - 1], le symbole * indique le conjugué complexe et B est le nombre de sous-bandes.The parameters ICLD or ILD or CLD (for " InterChannel / Channel Level Difference " in English), also called inter-channel intensity differences, characterize the energy ratios by frequency sub-band between the left and right channels. These parameters are used to position sound sources in the horizontal stereo plane by "panning". They are defined in dB by the following formula: $$\mathit{ICLD}\left[b\right]=10.{\mathit{log}}_{10}\left\{\frac{{\displaystyle {\sum }_{k=k_b}^{k_{b+1}-1}L\left[k\right].L^{\ast }\left[k\right]}}{{\displaystyle {\sum }_{k=k_b}^{k_{b+1}-1}R\left[k\right].R^{\ast }\left[k\right]}}\right\}\mathrm{dB}$$

where L [ k ] and R [ k ] correspond to the (complex) spectral coefficients of the L and R channels, each frequency band with index b = 0, ..., B - 1 includes the frequency lines in the interval [ k _b , k _{b +1} - 1], the symbol * indicates the complex conjugate and B is the number of subbands.

où ∠ indique l'argument (la phase) de l'opérande complexe.The parameters ICPD or IPD (for " InterChannel Phase Difference " in English), also called phase differences, are defined according to the following relation: $$\mathit{ICPD}\left[b\right]=\angle \left({\displaystyle {\sum }_{k=k_b}^{k_{b+1}-1}L\left[k\right].R^{\ast }\left[k\right]}\right)$$

where ∠ indicates the argument (phase) of the complex operand.

où d définit l'intervalle de recherche du maximum. On notera que la corrélation à l'équation (3) peut être normalisée.It is also possible to define in an equivalent manner to the ICPD, an inter-channel time shift called ICTD or ITD (for “ InterChannel Time Difference ”). The ITD can be measured for example as the delay maximizing the cross-correlation between L and R: $$\mathit{ITD}={\max }_{-d\le \tau \le d}{\displaystyle {\sum }_{\mathrm{not}=0}^{\mathrm{NOT}-\tau -1}L\left(\mathrm{not}+\tau \right).R\left(\mathrm{not}\right)}$$

where d defines the maximum search interval. Note that the correlation to equation (3) can be normalized.

où la corrélation peut être normalisée comme pour l'éq. (3).Unlike the ICLD, ICPD and ICTD parameters which are location parameters, the ICC parameter (for " InterChannel Coherence " in English) represents the level of inter-channel correlation (or coherence) and is associated with the spatial width of a sound source; ICC can be defined as: $$\mathit{ICC}={\max }_{-d\le \tau \le d}\left|{\displaystyle {\sum }_{\mathrm{not}=0}^{\mathrm{NOT}-\tau -1}L\left(\mathrm{not}+\tau \right).R\left(\mathrm{not}\right)}\right|$$

where the correlation can be normalized as for eq. (3).

It is noted in the article by Breebart et al. that the ICC parameters are not necessary in the sub-bands reduced to a single frequency coefficient - indeed the amplitude and phase differences completely describe the spatialization in this "degenerate" case.

The ICLD and ICPD parameters are extracted by analysis of the stereo signals, by block 105. The ICTD or ICC parameters can also be extracted by sub-band from the spectra L [ k ] and R [ k ]; however, their extraction is generally simplified by assuming an identical inter-channel time shift for each sub-band and in this case a parameter can be extracted from the time channels L ( n ) and R ( n ).

The mono signal M [k] is transformed in the time domain (blocks 106 to 108) after short-term Fourier synthesis (inverse FFT, windowing and addition-overlap known as OverLap-Add or OLA in English) and a mono coding (block 109) is then carried out. In parallel, the stereo parameters are quantized and coded in block 110.

In general, the spectrum of the signals ( L [ k ], R [ k ]) is divided according to a non-linear frequency scale of the ERB ( Equivalent Rectangular Bandwidth ) or Bark type. The parameters (ICLD, ICPD, ICC, ITD) are coded by scalar quantization possibly followed by entropy coding and / or differential coding. For example, in the article cited above, the ICLD is encoded by a non-uniform quantizer (ranging from -50 to +50 dB) with differential entropy coding. The non-uniform quantization step exploits the that the greater the value of the ICLD, the lower the hearing sensitivity to variations in this parameter.

For the coding of the mono signal (block 109), several quantization techniques with or without memory are possible, for example coding with "Coded Pulse Modulation" (PCM), its version with adaptive prediction called "Modulation by Coded Pulse Adaptive Differential "(ADPCM) or more advanced techniques such as perceptual transform coding or" Code Excited Linear Prediction "(CELP) coding or multi-mode coding.

We are more particularly interested here in the 3GPP EVS standard (for “Enhanced Voice Services”) which uses multi-mode coding. Algorithmic details of the EVS codec are provided in 3GPP TS 26.441 to 26.451 specifications and are therefore not repeated here. Hereafter, these specifications will be referred to by the name EVS.

• o "EVS Primary":
  • o débits fixes: 7.2, 8, 9.6, 13.2, 16.4, 24.4, 32, 48, 64, 96, 128
  • o mode à débit variable (VBR) avec un débit moyen proche de 5.9 kbit/s pour la parole active
  • o mode "channel-aware" à 13.2 en WB et SWB uniquement
• o "EVS AMR-WB IO" dont les débits sont identiques au codec 3GPP AMR-WB (9 modes)

The input signal of the EVS codec (mono) is sampled at the frequency of 8, 16, 32 or 48 kHz and the codec can represent telephone audio bands (narrowband, NB), widened (wideband, WB), super-widened (super-wideband, SWB) or full band (fullband, FB). The bit rates of the EVS codec are divided into two modes:

• o "EVS Primary":
  • o fixed flow rates: 7.2, 8, 9.6, 13.2, 16.4, 24.4, 32, 48, 64, 96, 128
  • o Variable bit rate mode (VBR) with an average bit rate close to 5.9 kbit / s for active speech
  • o "channel-aware" mode at 13.2 in WB and SWB only
• o "EVS AMR-WB IO" whose bit rates are identical to the 3GPP AMR-WB codec (9 modes)

Added to this is the discontinuous transmission mode (DTX) in which the frames detected as inactive are replaced by SID frames (SID Primary or SID AMR-WB IO) which are transmitted intermittently, approximately once every 8 frames .

At the decoder 200, with reference to the figure 2 , the mono signal is decoded (block 201), a de-correlator is used (block 202) to produce two versions M̂ ( n ) and M̂ ' ( n ) of the decoded mono signal. This decorrelation, necessary only when the ICC parameter is used, makes it possible to increase the spatial width of the mono source M̂ ( n ). These two signals M̂ ( n ) and M̂ ' ( n ) are passed into the frequency domain (blocks 203 to 206) and the decoded stereo parameters (block 207) are used by the stereo synthesis (or shaping) (block 208) to reconstruct the left and right channels in the frequency domain. These channels are finally reconstructed in the time domain (blocks 209 to 214).

An example of parametric stereo coding seeking to represent binaural signals (without respecting the nature of HRTF filters) is described in the article by Pasi Ojala, Mikko Tammi, Miikka Vilermo, titled "Parametric binaural audio coding", in Proc. ICASSP, 2010, pp. 393-396 . Two parameters are coded to restore a spatial image with a localization close to a binaural image: the ICLD and the ITD. In addition, an ALC (for “Ambiance Level Control”) parameter similar to the ICC is also coded, making it possible to control the level of the “ambience” associated with the use of decorrelated channels. This codec is described for super-wideband signals with 20 ms frames and a bit rate of 20 or 32 kbit / s to encode the mono signal to which is added a bit rate of 5 kbit / s to encode the spatial parameters.

Another example of a parametric stereo codec developed with a specific mode for encoding binaural signals is given by the G.722 Annex D standard, in particular in the R1ws wideband stereo encoding mode at 56 + 8 kbit / s. This codec works with "short" frames of 5 ms according to 2 modes: a "transient" mode where ICLDs are coded on 38 bits and a "normal" mode where ICLDs are coded on 24 bits with a full band ITD / IPD on 5 bits. The details of estimating the ITD, coding of the ICLD and ITD parameters are not included here. It will be noted that the ICLDs are coded by “decimation” by distributing the coding of the ICLDs over several successive frames, by coding only a subset of the parameters of a given frame.

In both examples it is important to note that these are not binaural codecs, but stereo codecs seeking to reproduce a spatial image similar to a binaural signal.

Note that the case of parametric multichannel coding with N> 2 follows the same principle of the N = 2 case, however in general the downmix may not be mono but stereo and the inter-channel parameters must cover more than 2 channels. An exemplary embodiment is given in the MPEG Surround standard where parameters ICLD, ICTD and ICC are coded. It will also be noted that the MPEG Surround decoder includes binaural restitution, parameterized by HRTFs filters.

où ${\sigma }_L^2\left[b\right]$

et ${\sigma }_R^2\left[b\right]$

représentent respectivement l'énergie du canal gauche (L[k]) et du canal droit (R[k]): $$\{\begin{array}{l}{\sigma }_L^2\left[b\right]={\displaystyle {\sum }_{k=k_b}^{k_{b+1}-1}L\left[k\right].L^{\ast }\left[k\right]}\\ {\sigma }_R^2\left[b\right]={\displaystyle {\sum }_{k=k_b}^{k_{b+1}-1}R\left[k\right].R^{\ast }\left[k\right]}\end{array}$$

Let us now consider the case of a stereo encoding and decoding of ICLD type parameters as described in figures 1 and 2 and take the case of a wideband signal, sampled at 16 kHz and analyzed with 20 ms frames and sinusoidal windowing covering 40 ms (including 20 ms of "lookahead"). For the extraction of the ICLD parameters (block 105), the spectra L [ k ] and R [ k ] can for example be split into B frequency sub-bands according to the ERB scale. For each frame, the ICLD of the sub-band b = 0, ..., 34 is calculated according to the equation: $$\mathit{ICLD}\left[b\right]=10.{\mathit{log}}_{10}\left\{\frac{{\sigma }_L^2\left[b\right]}{{\sigma }_R^2\left[b\right]}\right\}$$

or ${\sigma }_L^2\left[b\right]$

and ${\sigma }_R^2\left[b\right]$

respectively represent the energy of the left channel (L [k]) and of the right channel (R [ k ]): $$\{\begin{array}{l}{\sigma }_L^2\left[b\right]={\displaystyle {\sum }_{k=k_b}^{k_{b+1}-1}L\left[k\right].L^{\ast }\left[k\right]}\\ {\sigma }_R^2\left[b\right]={\displaystyle {\sum }_{k=k_b}^{k_{b+1}-1}R\left[k\right].R^{\ast }\left[k\right]}\end{array}$$

• 5 bits pour le premier paramètre ICLD (codé en absolu),
• 4 bits pour les 32 paramètres ICLD suivants (codés en différentiel),
• 3 bits pour les 2 derniers paramètres ICLD (codés en différentiel).

ce qui donne un total de 5 + 32x4 + 2x3 = 139 bits / trame, soit un débit proche de 7 kbit/s dans le cas de trames de 20 ms. Ce débit ne comprend pas les autres paramètres.According to the state of the art, the coding of a block of 35 ICLDs of a given frame can be carried out for example with:

• 5 bits for the first ICLD parameter (absolute coded),
• 4 bits for the following 32 ICLD parameters (differential coded),
• 3 bits for the last 2 ICLD parameters (differential coded).

which gives a total of 5 + 32x4 + 2x3 = 139 bits / frame, that is to say a rate close to 7 kbit / s in the case of frames of 20 ms. This rate does not include the other parameters.

This rate of approximately 7 kbit / s can be reduced on average by using variable rate entropy coding, for example Huffman coding; however, the flow reduction cannot be drastic in most cases.

To divide the encoding rate of the ICLD parameters by 2, it would be possible to use the alternate encoding approach described previously in the case of G.722 stereo encoding. However, the associated bit rate remains high for coding with 35 sub-bands and 20 ms of frame; moreover, the temporal resolution of the coding would be reduced, which can be problematic in the case of non-stationary signals. Another approach would be to reduce the number of subbands from 35 to for example 20 subbands. This would reduce the throughput associated with the ICLD parameters, but would generally degrade the fidelity of the synthesized spatial image.

If we assume that the encoder of the figure 1 is a stereo encoder operating for example at bit rates of 16.4, 24.4, 32, 48, 64, 96, 128 kbit / s and that it is based on a downmix encoded by a mono EVS codec, then for the lowest bit rates , for example 16.4 kbit / s in stereo, if the downmix is encoded with the 13.2 kbit / s mono EVS codec, only 3.2 kbit / s remains to encode all the spatial parameters in order to faithfully represent a spatial image. If it is necessary to encode not only ICLD parameters, but also other spatial parameters, it will be understood that the encoding of the ICLD parameters described above requires too much bit rate.

L Gao, R Hu, Y Yang, X Wang, W Tu, T Wu, Azimuthal perceptual resolution model based adaptive 3D spatial parameter coding, multimedia modeling (Springer International Publishing, Sydney, 2015 ) proposes to quantify the ICLD parameters of the MPEG Surround codec after having transformed them into azimuthal angles.

There is therefore a need to represent the spatial parameters of a multichannel signal efficiently, at a bit rate as low as possible and with an acceptable quality.

The invention improves the state of the art.

• extraction d'une pluralité d'informations de spatialisation du signal multicanal d'au moins deux types ;
• obtention d'au moins un modèle de représentation des informations de spatialisation extraites ;
• détermination d'au moins un paramètre d'angle d'un modèle obtenu ;
• codage du au moins un paramètre d'angle déterminé pour coder les aux moins deux types d'informations de spatialisation extraites lors du codage d'informations de spatialisation.

To this end, it proposes a method of parametric coding of a multichannel digital audio signal comprising a step of coding a signal resulting from a channel reduction processing applied to the multichannel signal and coding of spatialization information of the multichannel signal. . The process is such that it comprises the following steps:

• extracting a plurality of spatialization information from the multichannel signal of at least two types;
• obtaining at least one representation model of the extracted spatialization information;
• determination of at least one angle parameter of a model obtained;
• encoding of the at least one angle parameter determined to encode the at least two types of spatialization information extracted during the encoding of spatialization information.

The method of encoding spatialization information is based on a model-based approach which makes it possible to approximate the spatial information. Thus, the coding of a plurality of spatial information is reduced to the coding of an angle parameter, which considerably reduces the coding rate compared with the direct coding of the spatial information. The bit rate required for coding this parameter is therefore reduced.

In a particular sub-band embodiment, the spatialization information is defined by frequency sub-bands of the multichannel audio signal and at least one angle parameter per sub-band is determined and coded.

In a particular embodiment, the method further comprises the steps of calculating reference spatialization information and coding this reference spatialization information.

Thus, the encoding of reference information can improve the quality of decoding. The coding rate of this reference information does not require too high a rate.

This method is particularly well suited to the coding of spatial information of the inter-channel time shift (ITD) type and / or of the inter-channel intensity difference (ILD) type.

• estimation d'une information de différence d'intensité intercanale à partir du modèle obtenu et du paramètre d'angle déterminé ;
• codage de la différence entre l'information de différence d'intensité intercanale extraite et estimée.

To further improve the decoding quality of ILD type information, the method further comprises the following steps:

• estimation of inter-channel intensity difference information on the basis of the model obtained and the determined angle parameter;
• coding of the difference between the extracted and estimated inter-channel intensity difference information.

The coding of this residue requires an additional coding rate but this method always brings a gain in rate compared to the direct coding of the spatialization information ILD.

In a particular embodiment, a model of representation by spatialization information is obtained. It can be fixed and stored in memory.

This fixed and registered model is, for example, a sine-shaped model. This type of model is adapted to the form of the information ITD or ILD according to the position of the source.

In an alternative embodiment, obtaining a model of representation of the spatialization information is performed by the selection from a table of models defined for different values of the spatialization information.

Several models can be selectable depending on the characteristics of the multichannel signal. This makes it possible to best adapt the spatialization information model to the signal.

The index of the chosen model can then be, in one embodiment, encoded and transmitted.

In an alternative embodiment, a representation model common to several spatialization information items is obtained.

This makes it possible to pool the selection of a model with several spatialization information, which reduces the processing operations to be carried out.

• réception et décodage d'au moins un paramètre d'angle codé ;
• obtention d'au moins un modèle de représentation d'informations de spatialisation ;
• détermination d'une pluralité d'informations de spatialisation d'au moins deux types du signal multicanal à partir du au moins un modèle obtenu et du au moins un paramètre d'angle décodé.

The invention also relates to a method for parametric decoding of a multichannel digital audio signal comprising a step of decoding a signal originating from a channel reduction processing applied to the multichannel and coded signal and of decoding spatialization information. multichannel signal. The method is such that it comprises the following steps for decoding at least one piece of spatialization information:

• receiving and decoding at least one encoded angle parameter;
• obtaining at least one spatialization information representation model;
• determination of a plurality of spatialization information of at least two types of the multichannel signal on the basis of the at least one model obtained and of the at least one decoded angle parameter.

In the same way as for coding, this method based on the use of a model of representation of spatialization information makes it possible to find the information with good quality without it being necessary to have too high a bit rate. . At reduced bit rate, a plurality of spatialization information is recovered by the decoding of a simple angle parameter.

In a particular embodiment, the method comprises a step of receiving and decoding a model table index and obtaining at least one model of representation of the spatialization information to be decoded from the decoded index.

Thus, it is possible to adapt the model to be used according to the characteristics of the multichannel signal.

• un module d'extraction d'une pluralité d'informations de spatialisation du signal multicanal d'au moins deux types ;
• un module d'obtention d'au moins un modèle de représentation des informations de spatialisation extraites ;
• un module de détermination d'au moins un paramètre d'angle d'un modèle obtenu ;
• un module de codage du au moins un paramètre d'angle déterminé pour coder les au moins deux types d'informations de spatialisation extraites lors du codage d'informations de spatialisation.

The invention relates to a parametric encoder of a multichannel digital audio signal comprising a module for encoding a signal originating from a channel reduction processing module applied to the multichannel signal and modules for encoding spatialization information. multichannel signal. The encoder is as it comprises:

• a module for extracting a plurality of spatialization information from the multichannel signal of at least two types;
• a module for obtaining at least one model of representation of the extracted spatialization information;
• a module for determining at least one angle parameter of a model obtained;
• a module for encoding the at least one angle parameter determined to encode the at least two types of spatialization information extracted during the encoding of spatialization information.

The encoder has the same advantages as the method which it implements.

• un module de réception et décodage d'au moins un paramètre d'angle codé ;
• un module d'obtention d'au moins un modèle de représentation des informations de spatialisation ;
• un module de détermination d'une pluralité d'informations de spatialisation d'au moins deux types du signal multicanal à partir du au moins un modèle obtenu et du au moins un paramètre d'angle décodé.

The invention relates to a parametric decoder of a multichannel digital audio signal comprising a module for decoding a signal resulting from a channel reduction processing applied to the multichannel signal and coded and a module for decoding spatialization information. multichannel signal. The decoder is as it includes:

• a module for receiving and decoding at least one coded angle parameter;
• a module for obtaining at least one model of representation of spatialization information;
• a module for determining a plurality of spatialization information of at least two types of the multichannel signal from the at least one model obtained and from the at least one decoded angle parameter.

The decoder has the same advantages as the method which it implements.

Finally, the invention relates to a computer program comprising code instructions for implementing the steps of a coding method according to the invention, when these instructions are executed by a processor, to a computer program comprising instructions code for implementing the steps of a decoding method according to the invention, when these instructions are executed by a processor.

The invention finally relates to a storage medium readable by a processor on which is recorded a computer program comprising code instructions for the execution of the steps of the encoding method as described and / or of the decoding method as described.

• la figure 1 illustre un codeur mettant en œuvre un codage paramétrique connu de l'état de l'art et précédemment décrit;
• la figure 2 illustre un décodeur mettant en œuvre un décodage paramétrique connu de l'état de l'art et précédemment décrit;
• la figure 3 illustre un codeur paramétrique selon un mode de réalisation de l'invention;
• les figures 4a, 4b et 4c illustrent les étapes du procédé de codage selon différents modes de réalisation de l'invention par une illustration détaillée des blocs de codage d'informations spatiales;
• les figures 5a, 5b illustrent les notions de perception sonore en 3D et 2D et la figure 5c illustre une représentation schématique de coordonnées polaires (distance, azimuth) d'une source audio dans le plan horizontal par rapport à un auditeur, dans le cas binaural ;
• la figure 6a illustre des représentations de modèles d'énergie totale de HRTFs adaptés à représenter des informations spatiales de type ILD ;
• la figure 6b illustre une configuration de microphones stéréo de type ORTF captant un exemple de signal à deux canaux à coder selon un mode de réalisation du procédé de codage de l'invention ;
• les figures 6c à 6g illustrent des représentations d'un modèle d'information M_ILD(m,t) (pour m =0 et t correspondant à un azimuth de 0 à 360°) de spatialisation de type ILD par sous-bandes dans une découpe en 1/3 d'octave, en fonction de l'angle d'azimuth ;

la figure 7 illustre un décodeur paramétrique ainsi que le procédé de décodage selon un mode de réalisation de l'invention ;

• la figure 8 illustre une variante de réalisation d'un codeur paramétrique selon l'invention;
• la figure 9 illustre une variante de réalisation d'un décodeur paramétrique selon l'invention ; et
• la figure 10 illustre un exemple matériel d'un équipement incorporant un codeur apte à mettre en œuvre le procédé de codage selon un mode de réalisation de l'invention ou un décodeur apte à mettre en œuvre le procédé de décodage selon un mode de réalisation de l'invention.

Other characteristics and advantages of the invention will emerge more clearly on reading the following description, given solely by way of non-limiting example, and made with reference to the appended drawings, in which:

• the figure 1 illustrates an encoder implementing a parametric encoding known from the state of the art and described above;
• the figure 2 illustrates a decoder implementing a parametric decoding known from the state of the art and described above;
• the figure 3 illustrates a parametric encoder according to one embodiment of the invention;
• the figures 4a , 4b and 4c illustrate the steps of the coding method according to various embodiments of the invention with a detailed illustration of the spatial information coding blocks;
• the figures 5a, 5b illustrate the concepts of sound perception in 3D and 2D and the figure 5c illustrates a schematic representation of polar coordinates (distance, azimuth) of an audio source in the horizontal plane relative to a listener, in the binaural case;
• the figure 6a illustrates representations of total energy models of HRTFs adapted to represent spatial information of ILD type;
• the figure 6b illustrates a configuration of ORTF type stereo microphones picking up an example of a two-channel signal to be encoded according to an embodiment of the encoding method of the invention;
• the figures 6c to 6g illustrate representations of an information model M _ILD ( m, t ) (for m = 0 and t corresponding to an azimuth of 0 to 360 °) of spatialization of the ILD type by sub-bands in a cut in 1/3 octave, depending on the azimuth angle;

the figure 7 illustrates a parametric decoder as well as the decoding method according to one embodiment of the invention;

• the figure 8 illustrates an alternative embodiment of a parametric encoder according to the invention;
• the figure 9 illustrates an alternative embodiment of a parametric decoder according to the invention; and
• the figure 10 illustrates a hardware example of equipment incorporating an encoder capable of implementing the coding method according to an embodiment of the invention or a decoder capable of implementing the decoding method according to an embodiment of the invention .

With reference to the figure 3 , a two-channel parametric signal encoder according to one embodiment of the invention, delivering both a mono binary stream and spatial information parameters of the input signal is now described. This figure presents both the entities, hardware or software modules controlled by a processor of the coding device and the steps implemented by the coding method according to one embodiment of the invention.

The case of a two-channel signal is described here. The invention also applies to the case of a multichannel signal with a number of channels greater than 2.

To avoid overloading the text, the encoder described in figure 3 will be called a "stereo encoder" even if it allows the encoding of binaural signals. Likewise the parameters ICLD, ICTD, ICPD will be respectively denoted ILD, ITD, IPD even if the signal is not binaural.

This parametric stereo encoder as illustrated uses EVS mono encoding according to 3GPP TS 26.442 (fixed point source code) or TS 26.443 (floating point source code) specifications, it works with stereo or multichannel signals sampled at frequency d. F _s sampling of 8, 16, 32 and 48 kHz, with frames of 20 ms. Subsequently, without loss of generality, the description is mainly given for the case F _s = 16 kHz and for the case N = 2 channels.

It should be noted that the choice of a frame length of 20 ms is in no way restrictive in the invention which applies similarly in variants of the embodiment where the frame length is different, for example from 5 or 10 ms, with a codec other than EVS.

Moreover, the invention applies equally to other types of mono coding (eg: IETF OPUS, UIT-T G.722) operating at identical sampling frequencies or not.

Each time channel (L (n) and R (n)) sampled at 16 kHz is first pre-filtered by a high pass filter (HPF for High Pass Filter English) typically eliminating components below 50 Hz ( blocks 301 and 302). This pre-filtering is optional, but it can be used to avoid DC bias in estimating parameters such as ICTD or ICC.

The channels L ' ( n ) and R' ( n ) coming from the pre-filtering blocks are analyzed in frequencies by discrete Fourier transform with sinusoidal windowing with 50% overlap of length 40 ms or 640 samples (blocks 303 to 306) . For each frame, the signal ( L ' ( n ), R' ( n )) is therefore weighted by a symmetrical analysis window covering 2 frames of 20 ms, ie 40 ms (ie 640 samples for F _s = 16 kHz). The 40 ms analysis window covers the current frame and the future frame. The future frame corresponds to a “future” signal segment commonly called a “lookahead” of 20 ms. In variants of the invention, other windows could be used, for example an asymmetric window with low delay called “ALDO” in the EVS codec. In addition, in variants, the analysis windowing could be made adaptive according to the current frame, in order to use an analysis with a long window on stationary segments and an analysis with short windows on transient / non-transient segments. stationary, possibly with transition windows between long and short windows.

For the current frame of 320 samples (20 ms at F _s = 16 kHz), the spectra obtained, L [k] and R [k] (k = 0 ... 320), include 321 complex coefficients, with a resolution of 25 Hz by frequency coefficient. The coefficient of index k = 0 corresponds to the continuous component (0 Hz), it is real. The coefficient of index k = 320 corresponds to the Nyquist frequency (8000 Hz for F _s = 16 kHz), it is also real. The coefficients of index 0 < k <160 are complex and correspond to a sub-band of width 25 Hz centered on the frequency of k.

The L [ k ] and R [ k ] spectra are combined in block 307 to obtain a mono signal (downmix) M [k] in the frequency domain. This signal is time-converted by inverse FFT and windowing-overlap with the "lookahead" part of the previous frame (blocks 308 to 310).

An example of a frequency "downmix" technique is described in the document entitled " A stereo to mono downmixing scheme for MPEG-4 parametric stereo encoder "by Samsudin, E. Kurniawati, N. Boon Poh, F. Sattar, S. George, in Proc. ICASSP, 2006 . In this document, the L and R channels are aligned in phase before performing the channel reduction treatment.

où R'[k] est le canal R aligné, k est l'indice d'un coefficient dans la b ^ième sous-bande fréquentielle, ICPD[b] est la différence de phase inter-canal dans la b ^ième sous-bande fréquentielle donnée par l'équation (2).More precisely, the phase of the L channel for each frequency sub-band is chosen as the reference phase, the R channel is aligned according to the phase of the L channel for each sub-band by the following formula: $$R\prime \left[k\right]=e^{j.\mathit{ICPD}\left[b\right]}R\left[k\right]$$

where R ' [ k ] is the aligned R channel, k is the index of a coefficient in the b ^th frequency sub-band, ICPD [ b ] is the inter-channel phase difference in the b ^th frequency sub-band given by equation (2).

Note that when the sub-band of index b is reduced to a frequency coefficient, we find: $$R\prime \left[k\right]=\left|R\left[k\right]\right|.e^{j\angle L\left[k\right]}$$

Finally, the mono signal obtained by the “downmix” of the document by Samsudin et al. cited above is calculated by averaging the channel L and the channel R 'aligned, according to the following equation: $$M\left[k\right]=\frac{L\left[k\right]+R\prime \left[\mathrm{<figure-callout\; id="2"\; label="k"\; filenames="imgf0011.png"\; state="\{\{state\}\}">k</figure-callout>}\right]}{2}$$

avec w ₁ = 0,5 et $w_2=\frac{e^{J.\mathit{ICPD}\left[b\right]}}{2}$

dans le cas où la sous-bande d'indice b ne comporte qu'une valeur fréquentielle d'indice k.Alignment in phase therefore makes it possible to conserve energy and avoid attenuation problems by eliminating the influence of phase. This “downmix” corresponds to the “downmix” described in the document by Breebart et al. or: $$M\left[k\right]=w_{1}L\left[k\right]+w_{2}R\left[k\right]$$

with w ₁ = 0.5 and ${\mathrm{<figure-callout\; id="2"\; label="w"\; filenames="imgf0011.png"\; state="\{\{state\}\}">w</figure-callout>}}_2=\frac{e^{J.\mathit{<figure-callout\; id="2"\; label="ICPD\; b"\; filenames="imgf0011.png"\; state="\{\{state\}\}">ICPD</figure-callout>}\left[b\right]\left[\right]}}{2}$

in the case where the sub-band of index b only comprises a frequency value of index k .

Other “downmix” methods can of course be chosen without modifying the scope of the invention.

The algorithmic delay of the EVS codec is 30.9375 ms at F _s = 8 kHz and 32 ms for the other frequencies F _s = 16, 32 or 48 kHz. This delay includes the current 20 ms frame, the additional delay compared to the frame length is therefore 10.9375 ms at F _s = 8 kHz and 12 ms for the other frequencies ( i.e. 192 samples at F _s = 16 kHz), the mono signal is delayed (block 311) of T = 320-192 = 128 samples so that the delay accumulated between the mono signal decoded by EVS and the original stereo channels becomes a multiple of the frame length (320 samples). Consequently, to synchronize the extraction of stereo parameters (block 314) and the spatial synthesis from the mono signal performed at the decoder, the lookahead for the calculation of the mono signal (20 ms) and the mono encoding / decoding delay at which is added the delay T to align the mono synthesis (20 ms) correspond to an additional delay of 2 frames (40 ms) compared to the current frame. This delay of 2 frames is specific to the implementation detailed here, in particular it is related to the symmetrical sinusoidal windows of 20 ms. This delay could be different. In an alternative embodiment, one could obtain a delay of one frame with an optimized window with a lower overlap between adjacent windows with a block 311 not introducing any delay ( T = 0).

The shifted mono signal is then encoded (block 312) by the mono EVS encoder for example at a rate of 13.2, 16.4 or 24.4 kbit / s. In variants, the coding could be carried out directly on the non-shifted signal; in this case the shift can be performed after decoding.

Consideration is given in a particular embodiment of the invention, illustrated here at figure 3 , that block 313 introduces a delay of two frames on the spectra L [ k ] , R [ k ] and M [k] in order to obtain the spectra L _buf [ k ], R _buf [ k ] and M _buf [ k ] .

In terms of the quantity of data to be stored, one could more advantageously shift the outputs of the block 314 for extracting the parameters or the outputs of the quantization blocks 318, 316 and 319. This shift could also be introduced at the decoder at the same time. reception of the bit stream from the stereo encoder.

In parallel with the mono coding, the coding of the spatial information is implemented in the blocks 315 to 319 according to a coding method of the invention. Furthermore, the coding comprises an optional step of classifying the input signal in block 321.

This classification block, depending on the multichannel signal to be coded, can make it possible to switch from one coding mode to another. One of the coding modes being that implementing the invention for coding spatialization information. The other coding modes are not detailed here, but conventional stereo or multichannel coding techniques can be used, including parametric coding techniques with ILD, ITD, IPD, ICC parameters. The classification is indicated here with the time signals L and R at the input, possibly the signals in the frequency domain and the stereo or multichannel parameters can also be used for the classification. The classification can also be used to apply the invention to a given spatial parameter (for example to encode the ITD or the ILD), in other words, to switch the type of coding of spatial parameters with a possible choice between a coding method according to a model as in the invention or an alternative coding method of the state of the art.

The spatial parameters are extracted (block 314) from the spectra L [ k ], R [ k ] and M [k] shifted by two frames: L _buf [ k ] , R _buf [ k ] and M _buf [ k ] and coded (blocks 315 to 319) according to a coding method described with reference to figures 4a to 4c and detailing blocks 315 and 317.

For the extraction of the ILD parameters (block 314), the spectra L _buf [ k ] and R _buf [ k ] are for example split into frequency sub-bands.

In one embodiment, a cut into 1/3 octave sub-bands defined in Table 1 below will be taken:

This table covers all cases of sampling frequency, for example for an encoder with a sampling frequency at 16kHz, only the B = 20 first sub-bands will be retained. Thus, we can define the table: $$k_{b=0..20}=\left[\begin{array}{ccccccccccccccccccccc}0& 4& 6& 7& 9& 11& 14& 18& 22& 28& 36& 45& 57& 71& 90& 113& 143& 180& 226& 285& 320\end{array}\right]$$

The above table defines (in Fourier line index) the frequency sub-bands with index b = 0 to B-1 for the case F _s = 16 kHz. Each sub-band of index b comprises the coefficients k _b = 0 to k _{b +1} - 1. The frequency line of index k = 320 which corresponds to the Nyquist frequency is not taken into account here.

In variants, it is possible to use another cut into sub-bands, for example according to the ERB scale; in this case, we can use B = 35 sub-bands, these are defined by the following borders in the case where the input signal is sampled at 16 kHz: $$\begin{array}{l}{k}_{b=0..35}=[\begin{array}{cccccccccccccccccc}0& 1& 2& 3& 5& 6& 8& 10& 12& 14& 17& 20& 23& 27& 31& 35& 40& 46\end{array}\\ \begin{array}{ccccccccccccccccc}52& 58& 66& 74& 83& 93& 104& 117& 130& 145& 162& 181& 201& 224& 249& 277& 307\end{array}\\ 320]\end{array}$$

The above table defines (in Fourier line index) the frequency sub-bands with index b = 0 to B-1. For example the first sub-band ( b = 0) goes from the coefficient k _b = 0 to k _{b +1} - 1 = 0; it is therefore reduced to a single coefficient which represents 25 Hz. Similarly, the last sub-band ( k = 34) goes from the coefficient k _b = 307 to k _{b +1} - 1 = 319, it includes 12 coefficients (300 Hz ). The frequency line of index k = 320 which corresponds to the Nyquist frequency is not taken into account here.

où ${\sigma }_L^2\left[b\right]$

et ${\sigma }_R^2\left[b\right]$

représentent respectivement l'énergie du canal gauche (L_buf [k]) et du canal droit (R_buf [k]): $$\{\begin{array}{l}{\sigma }_L^2\left[b\right]={\displaystyle {\sum }_{k=k_b}^{k_{b+1}-1}L\left[k\right].L^{\ast }\left[k\right]}\\ {\sigma }_R^2\left[b\right]={\displaystyle {\sum }_{k=k_b}^{k_{b+1}-1}R\left[k\right].R^{\ast }\left[k\right]}\end{array}$$

For each frame, the DLI of the sub-band b = 0, ..., B-1 is calculated according to equations (5) and (6) repeated here: $$\mathit{HE\; D}\left[b\right]=10.{\mathit{log}}_{10}\left\{\frac{{\sigma }_L^2\left[b\right]}{{\sigma }_R^2\left[b\right]}\right\}$$

or ${\sigma }_L^2\left[b\right]$

and ${\sigma }_R^2\left[b\right]$

respectively represent the energy of the left channel ( L _buf [ k ]) and of the right channel ( R _buf [ k ]): $$\{\begin{array}{l}{\sigma }_L^2\left[b\right]={\displaystyle {\sum }_{k=k_b}^{k_{b+1}-1}L\left[k\right].L^{\ast }\left[k\right]}\\ {\sigma }_R^2\left[b\right]={\displaystyle {\sum }_{k=k_b}^{k_{b+1}-1}R\left[k\right].R^{\ast }\left[k\right]}\end{array}$$

According to a particular embodiment, the ITD and ICC parameters are extracted in the time domain (block 320). In variants of the invention, these parameters can be extracted in the frequency domain (block 314), which is not shown on the diagram. figure 3 so as not to weigh down the figure. An example of how to estimate the ITD in the frequency domain is given in standard ITU-T G.722 Annex D from the smoothed product L [ k ] .R ^∗ [ k ].

avec par exemple d= 630µs x F_s , soit 10 échantillons à 16 kHz. Cette valeur de 630µs s'obtient pour le cas binaural, à partir de la loi de Woodworth définie ci-après, avec une approximation sphérique de la tête (avec un rayon moyen a=8,5cm) et un azimuth θ = π/2.In one embodiment, the ITD and ICC parameters are estimated as follows. The ITD is sought by intercorrelation according to equation (3) shown here: $$\mathit{ITD}={\max }_{-d\le \tau \le d}{\displaystyle {\sum }_{\mathrm{not}=0}^{\mathrm{NOT}-\tau -1}L\left(\mathrm{not}+\tau \right).R\left(\mathrm{not}\right)}$$

with for example d = 630µs x F _s , ie 10 samples at 16 kHz. This value of 630µs is obtained for the binaural case, from Woodworth's law defined below, with a spherical approximation of the head (with an average radius a = 8.5cm) and an azimuth θ = π / 2 .

The ITD obtained according to equation (3) is then smoothed to attenuate its temporal variations. The advantage of smoothing is to attenuate the fluctuations of the instantaneous ITD which can degrade the quality of the spatial synthesis at the decoder. The smoothing method adopted goes beyond the scope of the invention and is not detailed here.

When calculating the ITD, the ICC is also calculated according to equation (4) defined above.

The parameters or spatial information ILD and ITD are coded according to a method forming the subject of the invention and described with reference to figures 4a to 4c which detail blocks 315 and 317 of the figure 3 according to different embodiments of the invention.

These blocks 315 and 317 implement methods based on models of respective representations of the information ITD and ILD.

Certain parameters of the respective models obtained at the output of blocks 315 and 317 are then coded in 316 and 318, for example according to a scalar quantization method.

All the spatialization information thus coded is multiplexed by the multiplexer 322 before being transmitted.

We remind figures 5a and 5b some important notions on sound perception. To the figure 5a is illustrated a median plane M, a frontal plane F and a horizontal plane H, relative to the head of a listener. Sound perception allows a 3D localization of a sound source, this localization is typically identified by spherical coordinates (r, θ, ϕ ) according to the figure 5b ; in the case of a stereo signal, the perception takes place on a horizontal plane and in this case polar coordinates ( r , θ ) are sufficient to locate the source in 2D. It is also recalled that a stereo signal only allows reproduction on a line between 2 loudspeakers on the horizontal plane, whereas a binaural signal normally allows 3D perception.

In one embodiment, the signal is considered to include a sound source located in the horizontal plane.

In the case of a binaural signal, it may be useful to define the position of a virtual source associated with the multichannel signal to be encoded. As shown in figure 5c , if we consider only the case of a sound source 510 located in the horizontal plane (2D) around the person represented by a head approximated by a sphere at 540, the position of the source is specified by the polar coordinates ( r , θ ).

The angle θ is defined between the front axis 530 of the listener and the source axis 520. The two ears of the listener are represented in 550R for the right ear and in 550L for the left ear. The time offset information between the two channels of a binaural signal is associated with the interaural time difference, that is to say the time difference that a sound takes to reach both ears. If the source is directly in front of the listener, the wave reaches both ears at the same time and the ITD information is zero.

où θ est l'azimuth dans le plan horizontal, a est le rayon d'une approximation sphérique de la tête et c la vitesse du son (en m.s^-1) qui peut être définie comme c=343 m.s^-1. Cette loi est indépendante de la fréquence, et elle est connue pour donner de bons résultats en termes de localisation spatiale.The interaural time difference (ITD) can be simplified by using a geometric approximation in the form of the following sine law: $$\mathit{ITD}\left(\theta \right)=\mathit{asin}\left(\theta \right)/\mathrm{vs}$$

where θ is the azimuth in the horizontal plane, a is the radius of a spherical approximation of the head and c is the speed of sound (in ms ^-1 ) which can be defined as c = 343 ms ^-1 . This law is independent of the frequency, and it is known to give good results in terms of spatial localization.

où $${\mathit{ITD}}_{\mathit{max}}=a/c$$

A virtual sound source can therefore be located with an angle θ and the ITD information can be deduced by the following formula: $$\mathit{ITD}\left(\theta \right)={\mathit{ITD}}_{\mathit{max}}\mathit{sin}\left(\theta \right)$$

or $${\mathit{ITD}}_{\mathit{max}}=\mathrm{at}/\mathrm{vs}$$

The value given to ITD _max can for example correspond to 630 μs, which is the limit of perceptual separation between two pulses. For larger ITD values the subject will hear two different sounds and will not be able to interpret the sounds as a single sound source.

qui est valable pour un champ lointain (typiquement une source à une distance d'au moins 10. a). En reprenant le principe d'une normalisation par une valeur maximale ITD_max comme à l'équation (15), le modèle d'ITD selon la loi de Woodworth peut être écrit sous la forme : $$\mathit{ITD}\left(\theta \right)=\frac{{\mathit{ITD}}_{\mathit{max}}\left(\mathit{sin}\left(\theta \right)+\theta \right)}{1+\pi /2}$$

où $${\mathit{ITD}}_{\mathit{max}}=a\left(1+\pi /2\right)/c$$

In variants of the invention, the sine law may be replaced by Woodworth's ITD model defined in the work of RS Woodworth, Experimental Psychology (Holt, New York), 1938, pp. 520-523 , by the following equation: $$\mathit{ITD}\left(\theta \right)=\mathrm{at}\left(\mathit{sin}\left(\theta \right)+\theta \right)/\mathrm{vs}$$

which is valid for a far field (typically a source at a distance of at least 10. a). By taking the principle of a normalization by a maximum value ITD _max as in equation (15), the ITD model according to Woodworth's law can be written in the form: $$\mathit{ITD}\left(\theta \right)=\frac{{\mathit{ITD}}_{\mathit{max}}\left(\mathit{sin}\left(\theta \right)+\mathrm{<figure-callout\; id="1"\; label="\theta "\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">\theta </figure-callout>}\right)}{1+\pi /2}$$

or $${\mathit{ITD}}_{\mathit{max}}=\mathrm{at}\left(1+\pi /2\right)/\mathrm{vs}$$

où $${\mathit{ITD}}_{\mathit{max}}=a/c$$

In variants, it would be possible to define a multiplying factor which does not represent the maximum value of the ITD but a proportional value, for example the factor a / c. The invention also applies in this case. For example, to simplify the expression of Woodworth's law it is possible to write: $$\mathit{ITD}\left(\theta \right)={\mathit{ITD}}_{\mathit{max}}\left(\mathit{sin}\left(\theta \right)+\theta \right)$$

or $${\mathit{ITD}}_{\mathit{max}}=\mathrm{at}/\mathrm{vs}$$

In this case the value of ITD _max does not represent the maximum value of ITD. Subsequently, this “rating gap” will be used.

Thus, with reference to the figure 4a , the block 315 which receives inter-channel time shift information (ITD) by the extraction module 320, comprises a module 410 for obtaining a model of representation of the inter-channel time shift information.

This model is for example the model as defined above in equation (15) with a value ITD _max = 630 μs predefined in the model or the model of equation (20).

In variants, the ITD _max value can be made flexible by coding either this value directly, or by coding the difference between this value and a predetermined value. This approach indeed makes it possible to extend the application of the ITD model to more general cases, but it has the drawback of requiring an additional throughput. To indicate that the explicit coding of the ITD _max value is optional, block 412 appears in dotted lines at the end of the figure 4a .

A module 411 for determining the angle θ as defined above is implemented to obtain the angle defined by the sound source. More precisely, this module searches for the azimuth parameter θ which makes it possible to get as close as possible to the extracted ITD. When the law is known as in equation (15), this angle can be obtained analytically: $$\theta =\mathit{asin}\left(\mathit{ITD}/{\mathit{ITD}}_{\mathit{max}}\right)$$

In variants, the asin function can be approximated.

An equivalent approach for determining azimuth can be implemented in block 411. According to this approach, the determination of the angle θ for the sine law involves a search using the ITD model, the closest value according to the possible azimuth values: $$\theta ={\mathit{argmin}}_{\theta \in T}{\left(\mathit{ITD}-{\mathit{ITD}}_{\mathit{<figure-callout\; id="2"\; label="max\; sin\; \theta "\; filenames="imgf0011.png"\; state="\{\{state\}\}">max</figure-callout>}}\mathit{sin}\left(\theta \right)\left(\right)\right)}^2$$

This search can be performed by pre-storing the different candidate values of ITD _max .sin ( θ ) from the ITD model in an M _ITD table for a search interval which can be T = [- π / 2, π / 2] assuming ITD is symmetrical when the source is in front of or behind the subject. In this case, the values of θ are discretized, for example with a step of 1 ° over the search interval.

In the case of Woodworth's law, we can also follow the same approach as above for the sine law. The analytical expression of the inverse function of sin ( θ ) + θ not being trivial, we may prefer the search: $$\theta ={\mathit{argmin}}_{\theta \in T}{\left(\mathit{ITD}-{\mathit{ITD}}_{\mathit{max}}\left(\mathit{sin}\left(\theta \right)+\mathrm{<figure-callout\; id="2"\; label="\theta "\; filenames="imgf0011.png"\; state="\{\{state\}\}">\theta </figure-callout>}\right)\right)}^2$$

où la table est donnée pour le cas d'une quantification scalaire uniforme sur 4 bits $$Q_{\theta }=\left\{-\pi ,-\frac{7\pi }{8},\dots ,0,\frac{\pi }{8},\dots ,\frac{7\pi }{8}\right\}$$

The angle parameter θ determined in block 411 is then coded according to a conventional coding method, for example by scalar quantization on 4 bits by block 316. This block carries out a search for the quantization index $$i={\mathit{argmin}}_{j=0,...,15}{\left(\theta -Q_{\theta }\left[\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="imgf0011.png"\; state="\{\{state\}\}">j</figure-callout>}\right]\right)}^2$$

where the table is given for the case of a uniform scalar quantization on 4 bits $$Q_{\theta }=\left\{-\pi ,-\frac{7\pi }{8},...,0,\frac{\pi }{8},...,\frac{7\pi }{8}\right\}$$

In variants, the number of bits allocated to the coding of the azimuth could be different, and the quantization levels could be non-uniform to take account of the perceptual limits of the location of a sound source according to the azimuth.

It is the encoding of this parameter which makes it possible to encode the time shift information ITD, possibly with the encoding of ITD _max (block 412) as additional information if the value predetermined by the ITD model must be adapted. The spatialization information will therefore be found on decoding by decoding the angle parameter, possibly by decoding ITD _max , and by applying the same model of representation of the ITD. The bit rate required for encoding this angle parameter is low (for example 4 bits per frame) when no correction of the ITD _max value pre-defined in the model is encoded. Thus, the coding of this spatialization information (ITD) consumes little bit rate.

At very low bit rate, the encoding of a single angle θ can be implemented to encode the spatialization information of a binaural signal.

In an alternative embodiment, it is possible to estimate an ITD per frequency band, for example by taking a cut in B sub-bands defined above. In this case, an angle θ per frequency band is coded and transmitted to the decoder, which for the example of B sub-bands gives B angles to be transmitted.

In another variant, it is possible to ignore the estimate of the ITD for certain high frequency bands for which the phase differences are not perceptible. Likewise, the ITD estimate can be omitted for very low frequencies. For example, the ITD could not be estimated for the bands greater than 1 kHz, and for a sub-band cut as defined previously, the bands b = 0 to 11 could be retained in the embodiment using the 1/3 octave and 1 to 16 in the variants using the ERB scale (the first band b = 0 being omitted in the latter case because these are frequencies below 25 Hz). In variants of the invention, a sub-band cutting with a resolution other than 25 Hz could be used; one will thus be able to regroup certain sub-bands because the cut in 1/3 octave or the ERB scale can be too fine for the coding of the ITD. This avoids coding too many angles per frame. For each frequency band, the ITD is then converted to an angle as in the case of a single angle described above with a bit allocation which can be either fixed or variable depending on the importance of the sub- bandaged. In all these variants where several angles are determined and coded, vector quantization can be implemented in block 316.

The figure 4b represents an alternative embodiment of the invention which can replace the mode described in figure 4a . The principle of this variant is to combine in particular the blocks 411 and 316 into a block 432.

In this variant embodiment, the definition of several “concurrent” models for coding the ITD is considered, knowing that the invention also applies when a single ITD model is defined.

où N _M est le nombre de modèles dans la table de modèles ITD, N_θ (m) est le nombre d'angles d'azimuth considérés pour le m-ième modèle et M_ITD(m,t) correspond à une valeur précise de l'information ITD.Thus, the model as defined for the inter-channel time shift (ITD) information may not be fixed and be configurable. Each model defines a set of ITD values as a function of an angle parameter: the sine law and Woodworth's law are two examples of models. In this variant, for coding, from an ITD model table obtained in 430, a model index and an angle index (also called an angle parameter) to be encoded according to block 432 are determined in block 432. the following equation: $$\left(m_{\mathit{opt}},t_{\mathit{opt}}\right)={\mathit{argmin}}_{\begin{array}{c}m=0,...,{\mathrm{NOT}}_{\mathrm{M}}-1\\ t=0,...,{\mathrm{NOT}}_{\theta }\left(m\right)-1\end{array}}{\left(\mathit{ITD}-{\mathrm{M}}_{\mathrm{<figure-callout\; id="2"\; label="ITD\; m\; t"\; filenames="imgf0011.png"\; state="\{\{state\}\}">ITD</figure-callout>}}\left(m,t\right)\left(\right)\right)}^2$$

where N _M is the number of models in the ITD model table, N _θ ( m ) is the number of azimuth angles considered for the m- th model and M _ITD ( m , t ) corresponds to a precise value of ITD information.

où chaque valeur est en ms. L'indice d'angle t correspond en fait à un angle θ couvrant l'intervalle $\left]-\frac{\pi }{2},\frac{\pi }{2}\right]$

avec un pas de $\frac{\pi }{8}.$

An example of a model M _ITD ( m , t ) is given below in the case of a model of index m = 0 according to a Woodworth law as in equation 20 with ITD _max = 0.2551 ms: $${\mathrm{M}}_{\mathrm{ITD}}\left(m=1,t=0...7\right)=\left[\begin{array}{cccccccc}-0.5362& -0.3807& -0.1978& 0& 0.1978& 0.3807& 0.5362& 0.6558\end{array}\right]$$

where each value is in ms. The angle index t corresponds in fact to an angle θ covering the interval $\left]-\frac{\mathrm{<figure-callout\; id="2"\; label="\pi "\; filenames="imgf0011.png"\; state="\{\{state\}\}">\pi </figure-callout>}}{2},\frac{\pi }{2}\right]$

with a step of $\frac{\pi }{8}.$

This table can also be reduced to samples, for example in the case of sampling at 16 kHz, we obtain in an equivalent way: $${\mathrm{M}}_{\mathrm{ITD}}\left(m=1,t=0...7\right)=\left[\begin{array}{cccccccc}-8.5795& -6.0919& -3.1648& 0& 3.1648& 6.0919& 8.5795& 10.4930\end{array}\right]$$

In this case, N _θ ( m ) = 8 and N _M = 1. It is therefore possible to encode the ITD information on 3 bits with this unique model.

Note that for a given model index m, the model M _ITD ( m , t ) is implicitly a function of the azimuth angle, insofar as the index t in fact represents a quantization index of the angle θ . Thus, the M _ITD ( m , t ) model is an efficient way to combine the relationship between ITD and θ , and the quantization of θ on N _θ ( m ) levels, and potentially use several models (at least one), indexed by m _opt when more than one model is used.

• m=0 : Un modèle binaural défini précédemment avec la loi de Woodworth avec ITD(θ) = ITD_max (sin(θ) + θ) et ITD_max = 10 (échantillons à 16 kHz) m=1 : Un modèle selon une loi en sinus comme à l'équation (15) mais pour un micro A-B (2 microphones omnidirectionnels séparés d'une distance a). La loi en sinus s'applique ici aussi, seul le paramètre a dépend de la distance entre les microphones : ITD(θ) = ITD_maxsin(θ) et ITD_max = 30 (échantillons à 16 kHz)

For example, in one embodiment, the case of two different models is considered:

• m = 0: A binaural model defined previously with Woodworth's law with ITD ( θ ) = ITD _max ( sin ( θ ) + θ ) and ITD _max = 10 (samples at 16 kHz) m = 1: A model according to a law in sine as in equation (15) but for a microphone AB (2 omnidirectional microphones separated by a distance a). The sine law also applies here, only the parameter a depends on the distance between the microphones: ITD ( θ ) = ITD _max sin ( θ ) and ITD _max = 30 (samples at 16 kHz)

Note that the size N _θ ( m ) can be identical for all models, but in the general case it is possible that different sizes are used. For example, we can define N _θ ( m ) = 16 and N _M = 2. It is therefore possible to encode the ITD information on 4 + 1 = 5 bits.

An index of the selected law m _opt is then encoded on ┌log ₂ N _M ┐ bits and transmitted to the decoder in addition to the azimuth angle t _opt encoded on ┌ log ₂ N _θ ┐ bits. In the example taken above, we can code m _opt on 1 bit, and t _opt on 4 bits.

In a variant, the model m = 0 could be replaced by an ITD table as a function of the azimuth resulting from real measurements of HRTFs, without a parametric law, but with ITD values estimated on the real data; in this case, the size N _θ ( m ) may depend on the angular resolution used to measure HRTFs (assuming that no angular interpolation has been applied).

As at the figure 4a , the coding of a correction information of the value ITD _max is optional, thus the block 312 is indicated in dotted lines. When the budget of bits allocated to the coding of ITD _max is zero, the value of ITD _max pre-defined in the representation model of the ITD will therefore be taken.

In a variant of the invention, the representation model of the ITD could be generalized to be reduced only to the horizontal plane but also to include the elevation. In this case, two angles are determined, the azimuth angle θ and the elevation angle ϕ.

avec N_ϕ (m) le nombre d'angles d'élévation considérés pour le m-ième modèle et p_opt représentant l'angle d'élévation à coder.The search for the two angles can be done according to the following equation: $$\left(m_{\mathit{opt}},t_{\mathit{opt}},p_{\mathit{opt}}\right)={\mathit{argmin}}_{\begin{array}{c}m=0,...,{\mathrm{NOT}}_{\mathrm{M}}-1\\ t=0,...,{\mathrm{NOT}}_{\theta }\left(m\right)-1\\ p=0,...,{\mathrm{NOT}}_{\phi }\left(m\right)-1\end{array}}{\left(\mathit{ITD}-{\mathrm{M}}_{\mathrm{ITD}}\left(\mathrm{<figure-callout\; id="2"\; label="m\; t\; p"\; filenames="imgf0011.png"\; state="\{\{state\}\}">m</figure-callout>},tp,\right)\right)}^2$$

with N _ϕ ( m ) the number of elevation angles considered for the m- th model and p _opt representing the elevation angle to be coded.

In the invention, it is also sought to reduce the coding rate of other spatialization information than the ITD, such as the inter-channel intensity difference (ILD) spatialization information. Note that block 316 of figure 4b can encode and multiplex in different ways with fixed or variable bit rate coding m _opt , t _opt , p _opt and ITD _{max information} only when these must be transmitted.

où f est la fréquence, r la distance avec la source sonore et c la vitesse du son.Thus, in the same way as for the ITD one can resort to a parameterization of the ILD. In the binaural case, according to the thesis of Jérôme Daniel, entitled "Representation of acoustic fields, application to the transmission and reproduction of complex sound scenes in a multimedia context", University Paris 6, July 2011 , the DLI can also be approximated according to the following law: $$\mathit{HE\; D}\left(r,\theta \right)=\frac{80\pi \mathit{Fr}\mathit{<figure-callout\; id="10"\; label="sin\; \theta \; cln"\; filenames="imgf0012.png,imgf0013.png"\; state="\{\{state\}\}">sin</figure-callout>}\left(\theta \right)}{\mathit{cln}}$$

where f is the frequency, r the distance from the sound source and c the speed of sound.

By defining a relative ILD ILD _max it is possible under certain conditions to reduce this approximation to the equation: $${\mathit{HE\; D}}_{\mathit{glob}}\left(\theta \right)={\mathit{HE\; D}}_{\mathit{max}}\mathit{sin}\left(\theta \right)$$

The above law is only an approximation corresponding to the global level of HRTFs at a given azimuth; it does not make it possible to completely characterize the spectral coloration given by the HRTFs but it only characterizes their overall level.

The reference ILD can be defined - in deferred time, during the definition of the ILD model, by taking a base of normalized signals or a base of HRTFs filters - by taking the maximum of the total ILD of a signal binaural. It is considered in the invention that this sine law applies not only to the total (or global) ILD but also to the ILD by sub-bands; in this case, the parameter ILD _max depends on the index of the sub-band and the model becomes: $$\mathit{HE\; D}\left[b\right]\left(\theta \right)={\mathit{HE\; D}}_{\mathit{max}}\left[b\right]\mathit{sin}\left(\theta \right)$$

Experimentally, we can verify that if we calculate the energy of the HRTFs filters (illustrated with reference to figure 6a for several values of elevation ϕ ), it appears that the approximation of the global DLI (in the sense of global level difference between channels) follows a sine law for the elevations represented ϕ = 0 °, 15 ° and 30 °, as a function of the azimuth θ .

Note that even if the symmetry of the front half-plane (azimuth in [0, 180] degrees) and the half-plane behind the head (azimuth in [180,360] degrees) is generally not fully valid , this sine law is used in the invention to encode and decode the ILD.

As for the case of the ITD where an ITD _max value has been defined, we can therefore either transmit the ILD _max parameter, or use a predetermined and stored ILD _max value, to derive from it an ILD _glob ( θ ) value according to equation (30) and thus apply a global ILD, valid over the entire signal spectrum to obtain a rudimentary (global) location.

Another sample model is based on the ORTF stereo microphone setup shown in figure 6b .

avec $$L\left(\theta \right)=a\left(1+\mathit{cos}\left(\theta -{\theta }_0\right)\right)$$

$$R\left(\theta \right)=a\left(1+\mathit{cos}\left(\theta +{\theta }_0\right)\right)$$

où θ₀ (en radians) correspond à 55°.In this example, the sub-band ILD model could be defined in relation to an ORTF microphone configuration as follows: $$\mathit{HE\; D}\left(\theta \right)=L\left(\theta \right)-R\left(\theta \right)=\mathrm{at}\left(\mathit{cos}\left(\theta -{\theta }_0\right)-\mathit{cos}\left(\theta +{\mathrm{<figure-callout\; id="0"\; label="\theta "\; filenames="imgf0007.png"\; state="\{\{state\}\}">\theta </figure-callout>}}_0\right)\right)$$

with $$L\left(\theta \right)=\mathrm{at}\left(1+\mathit{cos}\left(\theta -{\theta }_0\right)\right)$$

$$R\left(\theta \right)=\mathrm{at}\left(1+\mathit{cos}\left(\theta +{\mathrm{<figure-callout\; id="0"\; label="\theta "\; filenames="imgf0007.png"\; state="\{\{state\}\}">\theta </figure-callout>}}_0\right)\right)$$

where θ ₀ (in radians) corresponds to 55 °.

It is possible to write this model also in the form: $$\mathit{HE\; D}\left(\theta \right)=L\left(\theta \right)-R\left(\theta \right)=\mathrm{at}\left(\mathit{cos}\left(\theta \right)\mathit{cos}\left({\theta }_0\right)+\mathit{sin}\left(\theta \right)\mathrm{<figure-callout\; id="0"\; label="sin\; \theta "\; filenames="imgf0007.png"\; state="\{\{state\}\}">sin</figure-callout>}\left({\theta }_{}\right)\left({}_0\right)\right)$$

Here again, we can define a _max ILD value which corresponds to: $${\mathit{HE\; D}}_{\mathit{max}}=\mathrm{at}$$

Here again, it is assumed that the model defined in equation 35 applies not only to the case of a total (or global) ILD but also to the sub-band ILD; in this case the parameter ILD _max (or a proportional version) will be dependent on the sub-band in the form ILD [ b ] _max .

Thus, with reference to the figure 4a , in the same way as for the information ITD, the block 317 which receives an information of inter-channel intensity difference (ILD) by the extraction module 314, comprises a module 420 for obtaining a model of representation of the inter-channel intensity difference (ILD) information.

This model is for example the model as defined above in equation (30) or with other models described in this document.

The angle parameter θ already defined in 411 is re-used at the decoder to find the global ILD or the ILD in sub-bands as defined by equation (30), (31) or (35); this makes it possible to “mutualize” the coding of the ITD and the ILD. If the ILD _max value is not fixed, it is determined in 423 and coded.

In a particular embodiment, a module 421 for estimating inter-channel intensity difference information is implemented on the one hand from the angle parameter obtained by block 411 to encode the offset information. temporal (ITD) and on the other hand the representation model of equation (30), (31) or (35). Optionally, the module 422 calculates a residue of the ILD information, that is to say the difference between the actual inter-channel intensity difference (ILD) information extracted at 314 and the difference information of Inter-channel intensity (ILD) estimated at 421 from the ILD model.

This residue can be encoded at 318 for example by a conventional scalar quantization method. However, unlike the coding of a direct ILD, the quantization table can for example be limited to a dynamic range of +/- 12 dB with a step of 3 dB.

This ILD residue makes it possible to improve the quality of decoding of the ILD information in the case where the ILD model is too specific and only applies to the signal to be encoded in the current frame; it is recalled that a classification can optionally be used at the encoder to avoid such cases, however in the general case it may be useful to code an ILD residue.

Thus, the coding of these parameters as well as that of the angle of the ITD makes it possible to find at the decoder the information on the inter-channel intensity difference (ILD) of the binaural audio signal with good quality.

In the same way as for the ITD, the spatialization information (global or by sub-bands) will therefore be found on decoding by applying the same representation model and by decoding, if necessary, the residual and ILD parameters of reference. The bit rate required for encoding these parameters is lower than if the ILD information itself were encoded, in particular when the ILD residue does not have to be transmitted and when using the ILD _max parameter (s) pre-defined in the ILD model (s). Thus, the coding of this spatialization information (ILD) can be bit rate consuming.

This model of ILD using only one value of total ILD is however very simplistic because in general the ILD is defined on several sub-bands.

In the encoder described above, B sub-bands following a 1/3 octave cut or according to the ERB scale have been defined. To make it possible to represent more than one parameter of total (or global) ILD, the model of representation of the ILD is therefore extended to several sub-bands. This extension applies to the invention described in figure 4a , however the associated description is given below in the context of figure 4b to avoid too much redundancy. The model is a function of the angle θ and possibly of the elevation; this model can be the same in all the sub-bands, or vary according to the sub-bands.

où N _M est le nombre de modèles dans la table de modèles ILD, N_θ (m) est le nombre d'angles d'azimuth considérés pour le m - ième modèle, M_ILD(m,t) correspond à une valeur précise de l'information ILD et dist(.,.) est un critère de distance entre vecteurs d'ILD.Consider the variant embodiment described in figure 4b for the coding of the DLI. As for the ITD, in this variant we define representation models of the ILD. The model as defined for the inter-channel intensity difference (ILD) information is not fixed but can be configured. The model is defined by a _max ILD value and an angle parameter. In the general case, from a table of ILD models obtained in 440, a model index m _opt and an angle index to be coded in 442 are determined according to the following equation: $$\left(m_{\mathit{opt}},t_{\mathit{opt}}\right)={\mathit{argmin}}_{\begin{array}{c}m=0,...,{\mathrm{NOT}}_{\mathrm{M}}-1\\ t=0,...,{\mathrm{NOT}}_{\theta }\left(m\right)-1\end{array}}\mathit{dist}\left(\mathit{HE\; D},{\mathrm{M}}_{\mathrm{HE\; D}}\left(m,t\right)\right)$$

where N _M is the number of models in the ILD model table, N _θ ( m ) is the number of azimuth angles considered for the m - th model, M _ILD ( m , t ) corresponds to a precise value of the information ILD and dist ( .,. ) is a criterion of distance between vectors of ILD.

However, in an alternative embodiment, this search could be simplified by using the angle information already obtained in block 432 for the ITD model. Note that the values t = 0, ..., N _θ ( m ) - 1 for the DLI model do not necessarily correspond to the same set of values as for the ITD model, however it is advantageous to harmonize these sets to have consistency between representation models for the DLI and the ITD.

où q = 1 ou 2.For example, we can take as possible distance criteria: $$\mathit{dist}\left(X,Y\right)={\left|{\displaystyle {\sum }_{b=0}^{B-1}X\left[b\right]}-{\displaystyle {\sum }_{b=0}^{B-1}Y\left[b\right]}\right|}^q$$

where q = 1 or 2.

An example of a DLI model is shown in figures 6c to 6g for several frequency bands. The corresponding values (in dB) are not given here in the form of tables so as not to overload the text, approximate values can be taken from the graphs of the figures 6c to 6g . This figure considers the case of a 1/3 octave cut already defined previously. Thus each figure represents the ILD for the frequency band defined by the third octave number defined in Table 1 above with a central frequency fc depending on the band. Each point marked with a circle on each sub-figure corresponds to a value M _ILD ( m , t ); in addition to defining the table of ILD associated with the model, we have also shown the sine law scaled by a pre-defined parameter ILD _max and depending on the sub-band.

avec N_ϕ (m) le nombre d'angles d'élévation considérés pour le m-ième modèle et p_opt représentant l'angle d'élévation à coder.In a variant of the invention, the representation model of the DLI could be generalized so as not to be reduced only to the horizontal plane but also to include the elevation. In this case, the search for two angles becomes: $$\left(m_{\mathit{opt}},t_{\mathit{opt}},p_{\mathit{opt}}\right)={\mathit{argmin}}_{\begin{array}{c}m=0,...,{\mathrm{NOT}}_{\mathrm{M}}-1\\ t=0,...,{\mathrm{NOT}}_{\theta }\left(m\right)-1\\ p=0,...,{\mathrm{NOT}}_{\phi }\left(m\right)-1\end{array}}\mathit{dist}\left(\mathit{HE\; D},{\mathrm{M}}_{\mathrm{HE\; D}}\left(m,t,p\right)\right)$$

with N _ϕ ( m ) the number of elevation angles considered for the m-th model and p _opt representing the elevation angle to be coded.

• calculer les ILDs par sous-bande entre canaux gauche et droit par sous-bande
• éventuellement normaliser les ILDs
• stocker les IlDs et déterminer la valeur de ILD_max dans chaque sous-bande pour ajuster un facteur de dilatation des ILD

In a variant, an example of a model M _ILD ( m , t, p) can be obtained from a set of HRTFs as follows. Given the HRTFs filters for θ and ϕ , we can:

• calculate ILDs per sub-band between left and right channels per sub-band
• possibly standardize ILDs
• store the IlDs and determine the value of ILD _max in each sub-band to adjust an expansion factor of the ILDs

The multidimensional table M _ILD ( m , t, p) can be seen as a directivity model brought back to the domain of the ILD.

An index of the selected law m _opt is then encoded and transmitted to the decoder at 318.

In the same way as for the figure 4a , an ILD residue can be calculated (blocks 421 and 422) and coded.

So far separate models have been considered for ITD and ILD, although it has been noted that the angle determination can be "mutualized". For example, azimuth can be determined using the ITD model and this same angle is directly used for the ILD model. We now consider another variant embodiment using an “integrated model” (joint). This variant is described in figure 4c .

In this variant, rather than having separate models for the ITD and the ILD (M _ITD ( m , t, p ) and M _ILD ( m , t, p )) we can define a joint model in the block 450: M _{ITD, ILD} ( m , t, p ) whose entries include candidate values of ITD and ILD; thus, for different discrete values representing θ and ϕ “vectors” (ITD, ILD) are defined. In this case, the distance measure used for the search must combine the distance on the ITD and the distance on the ILD, however it is still possible to perform a separate search.

Thus, an index of the selected law m _opt , of the azimuth angle t _opt and of the elevation angle p _opt determined in 453, are coded in 331 and transmitted to the decoder, in the same way for the figures 4a and 4b , the parameters ITD _max , ILD _max and the residual ILD can be determined and coded.

A variant of the encoder shown in figure 3 implementing the joint model of figure 4c is illustrated in the figure 8 . It will be noted that in this variant of the encoder, the ITD and ICC parameters are estimated in the block 314. In addition, we consider here the general case where IPD parameters are also extracted and coded in the block 332. The blocks 330 and 331 correspond to the blocks. indicated and detailed in figure 4c .

With reference to the figure 7 a decoder according to one embodiment of the invention is now described.

This decoder comprises a demultiplexer 701 in which the coded mono signal is extracted to be decoded at 702 by a mono EVS decoder (according to the 3GPP TS 26.442 or TS 26.443 specifications) in this example. The part of the binary train corresponding to the mono EVS encoder is decoded according to the bit rate used at the encoder. It is assumed here that there is no loss of frames or of binary errors on the binary train to simplify the description, however known techniques for correcting the loss of frames can obviously be implemented in the decoder.

The decoded mono signal corresponds to M̂ ( n ) in the absence of channel errors. A short-term discrete Fourier transform analysis with the same windowing as in the encoder is performed on M̂ ( n ) (blocks 703 and 704) to obtain the spectrum M̂ [ k ]. We consider here that a decorrelation in the frequency domain (block 720) is also applied. This decorrelation could also be applied in the time domain.

The details of the implementation of the block 708 for the synthesis of the stereo signal are not presented here because they go beyond the scope of the invention, but the conventional synthesis techniques known from the state of the art could be used.

$$\widehat{R}\left[k\right]=c_2\widehat{M}\left[k\right]e^{-j2\mathit{\pi kITD}/\mathit{NFFT}}$$

où c = 10 ^ILD[b]/10 (avec b l'indice de la sous-bande contenant la raie d'indice k), $$c_1=\sqrt{\frac{2c}{1+c}},$$

et $$c_2=\sqrt{\frac{2}{1+c}},$$

ITD est l'ITD décodé pour la raie k (si un seul ITD est codé, cette valeur est identique pour les différentes raies d'indice k) et NFFT est la longueur de la FFT et de la FFT inverse (blocs 704, 709, 712).In synthesis block 708, it is for example possible to reconstruct a two-channel signal with the following processing on the decoded mono signal and transformed into frequencies: $$\widehat{L}\left[k\right]={\mathrm{vs}}_1\widehat{M}\left[k\right]$$

$$\widehat{R}\left[k\right]={\mathrm{vs}}_2\widehat{M}\left[k\right]e^{-\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="imgf0011.png"\; state="\{\{state\}\}">j</figure-callout>}2\mathit{\pi kITD}/\mathit{NFFT}}$$

where c = 10 ^{ILD [ b ] / 10} (with b the index of the sub-band containing the line of index k ), $${\mathrm{vs}}_1=\sqrt{\frac{2\mathrm{<figure-callout\; id="1"\; label="vs"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">vs</figure-callout>}}{1+\mathrm{vs}}},$$

and $${\mathrm{vs}}_2=\sqrt{\frac{2}{1+\mathrm{vs}}},$$

ITD is the decoded ITD for line k (if only one ITD is coded, this value is identical for the different lines of index k) and NFFT is the length of the FFT and of the inverse FFT (blocks 704, 709, 712).

The ICC parameter decoded in 718 can also be taken into account to recreate a non-localized sound environment (background noise) in order to improve the quality.

The spectra L̂ [ k ] and R̂ [ k ] are thus calculated and then converted into the time domain by inverse FFT, windowing, addition and overlap (blocks 709 to 714) to obtain the synthesized channels L̂ ( n ) and R̂ ( n ) .

The parameters which have been encoded to obtain the spatialization information are decoded in 705, 715 and 718.

In 718, it is the ICC ^{information q} [ b ] which is decoded if however they have been coded.

In 705, it is the angle parameter θ which is decoded with possibly a value ITD _max . From this parameter, the module 706 for obtaining a model of representation of an inter-channel time shift information item is implemented to obtain this model. As for the encoder, this model can be defined by equation (15) defined above. So, from this model and the decoded angle parameter, it is possible for the module 707 to determine the inter-channel time shift (ITD) information of the multi-channel signal.

If at the decoder an angle by frequency or by frequency band is coded, then these different angles by frequency or frequency bands are decoded to define the ITD information by frequency or frequency bands.

Likewise, in the case where parameters making it possible to encode the inter-channel intensity difference (ILD) information are encoded, they are decoded by the module for decoding these parameters at 715, at the decoder.

Thus, the parameters of residue (Resid. ILD) and of reference ILD _{(ILD max} ) are decoded in 715.

On the basis of these parameters, the module 716 for obtaining a model for representing information on the difference in inter-channel intensity is implemented in order to obtain this model. As for the encoder, this model can be defined by equation (30) defined above.

Thus, from this model, residual parameters of ILD (i.e. the difference between the actual inter-channel intensity difference (ILD) information and the inter-channel intensity difference information ( ILD) estimated with the model), the reference ILD parameter ( ILD _max ) and the angle parameter decoded in 705 for the ITD information, it is possible for the module 717 to determine the intensity difference information inter-channel (ILD) of the multi-channel signal.

If at the encoder the coding parameters of the ILD have been broken down by frequency band, then these various parameters by frequency bands are decoded to define the ILD information by frequency or frequency bands.

Note that the decoder of the figure 7 is related to the encoder of the figure 4a . It will be understood that if the coding according to the invention is carried out according to the figures 4b or 4c , the decoder will be modified accordingly to decode in particular model and angle indices in the form m _opt , t _opt , p _opt and reconstruct the values of ITD and ILD according to the model used and the associated indices to reconstruction values

In a variant of the invention, the decoder of the figure 7 is thus modified as illustrated in figure 9 . In this variant, the decoded ILD and ITD parameters are not directly reconstructed. Stereo synthesis (block 708) is replaced by binaural synthesis (block 920). Thus, the decoding of the ILD and ITD information is reduced to a decoding (block 910) of the angular coordinates. By using a predefined base of HRTFs (block 930) it is therefore possible to decode a binaural signal and not a stereo signal. In variants, the HRTFs filters can be applied in the time domain.

The encoder presented with reference to figure 3 and the decoder presented with reference to figure 7 have been described in the particular case of stereo coding and decoding. The invention has been described on the basis of a decomposition of the stereo channels by discrete Fourier transform. The invention also applies to other complex representations, such as for example the MCLT decomposition (Modulated Complex Lapped Transform) combining a modified discrete cosine transform (MDCT) and discrete modified sine transform (MDST), as well as the case of Pseudo-Quadrature Mirror Filter (PQMF) type filter banks. Thus, the term “frequency line” used in the detailed description can be extended to the concept of “sub-band” or “frequency band”, without changing the nature of the invention.

Encoders and decoders as described with reference to figures 3 and 7 can be integrated into multimedia equipment such as living room decoder, "set top box" or audio or video content player. They can also be integrated into communication equipment of the mobile phone or communication gateway type.

The figure 10 represents an exemplary embodiment of such equipment in which an encoder as described with reference to figures 3 , 8 and 4a to 4c or a decoder as described with reference to figure 7 or 9 , according to the invention is integrated. This device comprises a processor PROC cooperating with a memory block BM comprising a storage and / or working memory MEM.

In the case of an encoder, the memory block can advantageously comprise a computer program comprising code instructions for implementing the steps of the encoding method within the meaning of the invention, when these instructions are executed by the processor PROC, and in particular the steps of extracting a plurality of spatialization information from the multichannel signal, of obtaining at least one representation model of the extracted spatialization information, of determining at least one angle parameter. a model obtained and encoding of at least one angle parameter determined to encode the spatialization information extracted during the encoding of spatialization information.

In the case of a decoder, the memory block can advantageously comprise a computer program comprising code instructions for the implementation of the steps of the decoding method within the meaning of the invention, when these instructions are executed by the processor PROC, and in particular the steps of receiving and decoding at least one coded angle parameter, of obtaining at least one model of representation of spatialization information and of determining a plurality of spatialization information of the multichannel signal from the at least one model obtained and from the at least one decoded angle parameter.

The memory MEM can store the model or models of representation of various spatialization information which are used in the encoding and decoding methods according to the invention.

Typically, descriptions of figures 3 , 4 on the one hand and 7 on the other hand repeat the steps of an algorithm of such a computer program respectively for the encoder and for the decoder. The computer program can also be stored on a memory medium readable by a reader of the device or equipment or downloadable in the memory space thereof.

Such equipment as an encoder comprises an input module capable of receiving a multichannel signal, for example a binaural signal comprising the R and L channels for right and left, either by a communication network, or by reading stored content. on a storage medium. This multimedia equipment can also include means for capturing such a binaural signal.

The device as an encoder comprises an output module capable of transmitting a mono signal M resulting from a channel reduction processing and at least an angle parameter θ making it possible to apply a model for the representation of information of spatialization to find this spatial information. If necessary, other parameters such as the residual parameters of ILD, ILD or reference ITD (ILDmax or ITDmax) are also transmitted via the output module.

Such equipment as a decoder comprises an input module capable of receiving a mono signal M coming from a channel reduction processing and at least one angle parameter θ making it possible to apply an information representation model. spatialization to find this spatial information. If necessary, to find the spatialization information, other parameters such as the residual parameters of ILD, ILD or reference ITD (ILDmax or ITDmax) are also received via the input module E.

The device as a decoder comprises an output module capable of transmitting a multichannel signal, for example a binaural signal comprising the R and L channels for right and left.

Claims (15)

1. Method of parametric coding of a multichannel digital audio signal comprising a step (312) of coding a signal (M) arising from a channels reduction processing (307) applied to the multichannel signal and of coding spatialization cues in respect of the multichannel signal, characterized in that it comprises the following steps:

   - extraction (314, 320) of a plurality of spatialization cues in respect of the multichannel signal of at least two types;

   - obtaining (315, 317) of at least one representation model of the spatialization cues extracted;

   - determination (315, 411) of at least one angle parameter of a model obtained;

   - coding (316, 318) of the at least one determined angle parameter so as to code the at least two types of spatialization cues extracted during the coding of spatialization cues.
2. Coding method according to Claim 1, characterized in that the spatialization cues are defined by frequency sub-bands of the multichannel audio signal and in that at least one angle parameter per sub-band is determined and coded.
3. Method according to one of Claims 1 and 2, characterized in that it furthermore comprises the steps of calculating a reference spatialization cue and of coding this reference spatialization cue.
4. Coding method according to one of the preceding claims, characterized in that one of the spatialization cues of a first type is an interchannel time shift (ITD) cue.
5. Coding method according to one of the preceding claims, characterized in that one of the spatialization cues of a second type is an interchannel intensity difference (ILD) cue.
6. Method according to Claim 5, characterized in that it furthermore comprises the following steps for coding an interchannel intensity difference cue:

   - estimation of an interchannel intensity difference cue on the basis of the model obtained and of the angle parameter determined;

   - coding of the difference between the interchannel intensity difference cue extracted and estimated.
7. Method according to one of the preceding claims, characterized in that a spatialization-cue-based representation model is obtained.
8. Method according to one of the preceding claims, characterized in that a representation model common to several spatialization cues is obtained.
9. Coding method according to one of Claims 1 to 8, characterized in that the obtaining of a representation model of the spatialization cues is performed by selecting from a table of models defined for various values of the spatialization cues.
10. Method according to Claim 9, characterized in that an index of the table corresponding to the selected model is coded.
11. Method of parametric decoding of a multichannel digital audio signal comprising a step (702) of decoding a coded signal arising from a channels reduction processing applied to the multichannel signal and of decoding (705, 715) spatialization cues in respect of the multichannel signal, characterized in that it comprises the following steps for decoding at least one spatialization cue:

    - reception and decoding (705, 715) of at least one coded angle parameter;

    - obtaining (706, 716) of at least one representation model of spatialization cues;

    - determination (707, 717) of a plurality of spatialization cues of at least two types in respect of the multichannel signal on the basis of the at least one model obtained and of the at least one decoded angle parameter.
12. Decoding method according to Claim 11, characterized in that it comprises a step of receiving and decoding an index of table of models and of obtaining the at least one representation model of the spatialization cues to be decoded on the basis of the decoded index.
13. Parametric coder of a multichannel digital audio signal comprising a module (312) for coding a signal (M) arising from a module for channels reduction processing (307) applied to the multichannel signal and modules (315, 317, 316, 318) for coding spatialization cues in respect of the multichannel signal, characterized in that it comprises:

    - a module (314, 320) for extracting a plurality of spatialization cues of at least two types in respect of the multichannel signal;

    - a module (315, 317) for obtaining at least one representation model of the spatialization cues extracted;

    - a module (411) for determining at least one angle parameter of a model obtained;

    - a module (316, 318) for coding the at least one angle parameter determined so as to code the at least two types of spatialization cues extracted during the coding of spatialization cues.
14. Parametric decoder of a multichannel digital audio signal comprising a module (702) for decoding a coded signal arising from a channels reduction processing applied to the multichannel signal and a module (705, 7015) for decoding spatialization cues in respect of the multichannel signal, characterized in that it comprises:

    - a module (705, 715) for receiving and decoding at least one coded angle parameter;

    - a module (706, 716) for obtaining at least one representation model of the spatialization cues;

    - a module (707, 717) for determining a plurality of spatialization cues of at least two types in respect of the multichannel signal on the basis of the at least one model obtained and of the at least one decoded angle parameter.
15. Storage medium readable by a processor on which is recorded a computer program comprising code instructions for the execution of the steps of the coding method according to one of Claims 1 to 10 and/or of the decoding method according to one of Claims 11 and 12.

Images