DOA estimation of multiple speech sources based on the single-source point detection using an FOA microphone

Abstract

This paper presents a method for direction of arrival (DOA) estimation of multiple speech sources based on the temporal correlation and local-frequency stationarity of speech signals. The distribution analysis of single-source points (SSPs) in a recorded signal shows that in the time–frequency (T-F) domain, the SSPs are distributed in the form of a small cluster. According to this distribution, a method for DOA estimation of multiple sound sources is developed based on the continuity between adjacent T-F points. In addition, low-reverberation single-source (LRSS) points are detected based on the phase consistency and used as guidance to detect whether adjacent T-F points are SSPs. The direction deviations between adjacent frequency points and between adjacent frames are used as the SSP detection criteria considering the temporal correlation and local-frequency stationarity. The kernel density estimation and peak search are performed to obtain the dynamic DOA estimation range of each source. Finally, DOA estimates of each source are obtained by statistical weighting-based fine localization. Experiments under both simulated and real conditions show that the proposed method can achieve better localization performance than several existing methods.