ESTIMATION OF DIRECTION OF ARRIVAL USING MATCHING PURSUIT AND ITS APPLICATION TO SOURCE SEPARATION
Yoshio Yamasaki · 2003
In this paper, we describe a new blind source separation (BSS) method that uses spatial information derived from the direction of arrival (DOA) estimates of each direct and reflected sound. The method we proposed has the following steps: (1) each DOA is es- timated using matching pursuit and re-optimized after each new DOA is estimated, (2) using these DOA estimates, the mixing ma- trix is also estimated and the inverse of the mixing matrix is used to separate the mixture signals. Our experiments yielded a better signal separation with the new method than with the conventional frequency domain independent component analysis (ICA) based BSS method. In this paper, we propose a BSS method that uses spatial infor- mation derived from the results of direction of arrival (DOA) esti- mates for direct and early reflected sounds. We need to find many DOAs to estimate the mixing system. However it is impossible to find true DOAs using conventional beam forming techniques when the number of sources exceeds that of microphones. We suggest a new DOA estimates technique, which is using a matching pursuit algorithm, and it is possible to find true DOAs even if the num- ber of sources exceeds that of microphones. The basic outline of our algorithm is as follows. We first find the normalized power of the array output, , as a function of the DOA, . Then to es- timate the DOA of direct and indirect (reflected) signals we apply a matching pursuit algorithm, which includes a re-optimization of the DOAs at each iteration step. The sounds coming from differ- ent DOAs are then classified into a small set of sources. We then form estimates of the impulse responses for each source and mi- crophone combination from these classified DOAs. The separated source signals are obtained by filtering the observations with the inverse of the mixing matrix estimate. We compared our method with the conventional frequency do- main ICA based BSS method (5) using two sources, two micro- phones, and a convolved mixture. Our experiments yielded better signal separation for the new method than that for conventional frequency domain ICA based BSS.