Cluster Analysis On Individualized Head-related Transfer Functions In Wavelet Domain And Application To Binaural5.1Channel Surround Sound Reproduction

In the binaural reproduction of5.1channel surround sound, the5.1channel signals arefiltered with Head-related Transfer Functions (HRTFs) and then rendered via headphone tocreate spatial information correctly. Individualized HRTFs are preferred in binaural filtering orsynthesis, because non-individualized HRTFs often result in directional errors in perceivedvirtual sources, such as elevation error and reversal error, or poor externalization.Experimental measurement is a relatively accurate method to obtain individualized HRTFs.However, measurement is complicated and time-consuming, thus impractical for uses.Accordingly, simply and approximately methods for obtaining individualized or matchedHRTF are need. Given that the similarities or differences in individualized HRTFs depend onfrequency, moreover，individualized HRTFs in different frequency ranges contribute unequallyto localization, the present thesis analyzes the similarities and cluster of HRTFs in differentfrequency ranges using the wavelet decomposition and proposes a scheme to customizematched HRTFs. The matched HRTFs are used for binaural5.1channel surround soundreproduction.Firstly, three levels wavelet decomposition is applied to the individualized head-relatedimpulse response (HRIR) in a baseline database with52subjects. Preliminary analyses on thesimilarities of individualized HRIR in different sub-band show that, two levels waveletdecomposition is sufficient for representing the similarities of individualized HRTFs atdifferent frequency bands. Based on this conclusion, a further analysis on the similarity of twolevels wavelet decomposition of individualized HRTFs indicates that HRTFs of differentindividuals are highly similar within0to6kHz but vary considerably above6kHz.Nevertheless, HRTFs above12kHz contribute less to localization. Based on this result, clusteranalysis is applied to the individualized HRTFs within6kHz to12kHz to obtain8clustercenters. The cluster centers are combined with the general HRTFs at sub-bands [0,6kHz] and[12kHz,24kHz] to reconstruct the customized or matched HRTFs.The virtual source localization experiment indicates that, compared withnon-individualized HRTFs of KEMAR, the matched HRTFs partly improve the localization performance in the binaural5.1channel reproduction.