Binaural synthesis is a sound spatialization technology, which is the closest to na-tural hearing. It allows the spatialization of a monophonic sound source with only twofilters for a given position. The filters are defined by the HRTFs (Head Related TransferFunction) corresponding to the left and right ear of the listener. The major drawback ofbinaural synthesis is that the HRTF, which are related to the listener's morphology, arestrongly individual. Listening with non-individual HRTF could lead to audible artifacts.Therefore binaural rendering of high quality requires individualized HRTF. This thesistackles the problem of the individualization of binaural synthesis in the framework ofits implementation as a pure delay, the interaural time di®erence (ITD), and a minimalphase filter determined by the magnitude of the HRTF. The work conducted on the ITDvalidates the implementation chosen even for the positions where the HRTF are poorlyminimum phase filters. In addition the ITD calculation methods which are close to per-ception are pointed out. An experimental study is also undertaken to investigate theresolution of the ITD with the elevation angle along the cones of confusion. Perceptualresults indicate that the ITD variation with the elevation angle needs to be reproduced.In order to account for this variation, a new formula is proposed on the basis of thespherical head model. Optimization of the parameters of this formula for a whole ITDdatabase provides an average formulation which is appropriate for a large number of sub-jects and for many applications. Concerning the modeling of the spectral cues (HRTFmagnitude), the Boundary Element Method (BEM) has been examined. It is concludedthat BEM methods are useful in combination with measurement for the modeling ofthe low frequency part. A new approach, which involves statistical learning technique, isproposed for the HRTF prediction. A neural network is built to compute HRTF in anydirection from a limited set of measured HRTF. Preliminary assessment of this modelingshows that the neural network succeeds well in individualizing spectral cues. This resultsuggests a simplified protocol of HRTF measurement : HRTF are measured for only afew directions and the HRTF for the other locations are obtained by the neural network.
