On the Clustering of Head-Related Transfer Functions Used for 3-D Sound Localization
Chin‐Shyurng Fahn, Yueh-Chuan Lo · 2003
Head-related transfer functions (HRTFs) serve the increasingly dominant role of implementation 3-D audio systems, which have been realized in some commercial ap-plications. However, the cost of a 3-D audio system cannot be brought down because the efficiency of computation, the size of memory, and the synthesis of unmeasured HRTFs remain to be made better. This paper presents a way to moderate the memory require-ment and computational complexity in order to reduce the cost of a 3-D audio system. We employ the library of HRTF measurements called Knowles Electronics Mannequin for Acoustic Research (KEMAR) as the original data [8]. First of all, each HRTF meas-urement has to be approximated in the minimum phase, and the length of the HRTF is limited by use of a modified Hamming window function, if necessary. Second, we pro-pose an improved LBG-based clustering algorithm to lower the huge number of HRTFs. During the clustering, each HRTF is represented by its power cepstrum. Only portions of the HRTFs are reserved, and the others are neglected on condition that the minimal av-erage mismatch distance between measured and synthesized HRTFs is achieved. Before applying localization of 3-D sounds, both unmeasured and removed HRTFs can be syn-thesized by linear interpolation and interaural time difference insertion. Experimental results reveal that the average and the maximum mismatch distances deriving from our improved LBG-based clustering method are less than those from the uniform clustering and LBG-based clustering methods [13].