| Résumé |
Head-Related Transfer Functions (HRTF) vary upon individuals due to the specificity of one's head/ear/torso shapes. Perceptual artefacts entailed by the use of non-individual HRTF can be significantly reduced by scaling in frequency the non individual HRTF to best fit the listener's ones. This method, which was proposed by Middlebrooks ([1], [2]), consists in translating one head's spectra on the log-frequency axis to match the features of the second head's HRTF. As an extension, we first show that the efficiency of the scaling is improved when applied independently to two frequency bands above 1kHz. In order to automatically set the scaling factors, correlation to the dimension of morphological features is studied. In order to further simplify the evaluation of the scaling factors, it is shown that a satisfying estimate can be obtained that relies on a limited number of HRTF. They correspond to the "principal directions" given by a statistical analysis of HRTF. |
