검색 상세

Reverberation-robust multiple sound source localization based independent component analysis with imposition of the disjoint condition and a sparse prior

서로소 조건 및 성김 특성을 적용한 반향에 강인한 독립성분분석 기반 다음원 국지화 방법

초록/요약

This thesis presents a method to estimate the time delay for multiple sound source localization in reverberant environments. A method has been proposed to accomplish time delay estimation (TDE) for speaker localization, and its experimental results demonstrated that it could efficiently estimate a desired time delay even in highly reverberant environments by imposing sparse priors on estimated channel impulse responses. Although the method can estimate an exact speaker location successfully, it can be used for the localization of a single dominant speaker at a time. Because the conventional methods based on the cross-relation between channels are not applicable to multiple sound source localization, this thesis extends the joint maximum likelihood estimation of the independent components and the mixing matrix to deal with acoustic observations modeled by convolutive mixtures, and derives a frequency-domain approach for efficient computations. A sparse prior is imposed in the process of mixing channel estimation to help detection of direct paths from temporally correlated signals such as speech. Furthermore, assuming sparsity of sound sources in the time-frequency domain, disjoint condition of sources is exploited to improve the performance of TDE for closely located sources. Experimental results show that the presented method can efficiently estimate the time delay for multiple sound source localization in reverberant environments.

more