Bachelor Thesis of Bechtel, Phillip
To present people a spatial sound through headphones, it is necessary to determine the head related transfer function HRTF, for each desired position of the virtual sound source. These HRTF include information on the position of sound source, as well as the geometrical dimensions of the perSOll, for example shoulder, head and ear dimensions. As it is too expensive to carry out individual measurements for each person, an average, human head, a so-called dummy head was developed. With this dummy head an HRTF-Set is measured and fitted. Fitting is the approximation of a transfer function with poles and zeros. Fitting makes a HRTF -Set compact and reduces the memory overhead. The aim of fitting is to generate the best possible approximation of the measured HRTF so that no characteristics get lost. Shifting the poles and zeros of a fitted HRTF-Set, set the changes. This shifting allows individualization of a HRTF-Set. Consequently the number of poles and zeros should be kept as low as possible. Otherwise too many free parameters have to be determined for a potential individualization. In this paper an algorithm has been developed which divides the considered frequency band in smaller areas. These areas are fitted and joined with band pass back together. This results in smaller, independent solution vectors. To get the same result with a conventional fitting algorithm, the same number of poles and zeros are necessary. Generally the number of required poles and zeros depends on the desired quality of the fit and on the resolution of the measured HRTF. To evaluate the developed algorithm the case with a divisioll into four frequency bands was compared with a conventional algorithm. For comparison, the interaural time difference (ITD), interaural level difference (ILD), monaural cnes and the spectral difference were considered. The results of this comparison show that both algorithms provide similar results with an identical number of poles and zeros. The fitted HRTFs of the conventional algorithm is a bit better in the lower frequency ranges and the newly developed algorithm performs better at higher frequencies. Thus the phase is reconstructed more accurately at higher frequencies.