Masahito Togami(Line Corporation) and Robin Scheibler(LINE)
We propose a direction-of-arrival (DOA) estimation
technique which assumes that speech sources are sufficiently sparse and
there is only one active speech source at each
time-frequency (T-F) point. The proposed method estimates the DOA of the active speech source
at each T-F point.
A typical way for DOA estimation is based on grid-searching for all possible directions. However,
computational cost of grid-searching is proportional to the resolution of search area.
Instead of accurate grid-searching, the proposed method adopts
rough grid-searching followed by an iterative parameter optimization based on
Majorization-Minimization (MM) algorithm.
We propose a parameter optimization method which guarantees a monotonical increase
of the objective function.
Experimental results show that the proposed method estimates DOAs of speech sources more accurately than
conventional DOA estimation methods when computational cost of each method is almost the same.