Sound source localization based on binaural signals

摘要

The proposed method is directed a sound source localization using a binaural detector configuration. It comprises the steps of: Preprocessing the two signals using a filter bank to gain (2D) time-frequency spectrum for each signal. Jointly calculating ITD and ILD measures using a common matching procedure working on 2D patches of the time-frequency spectrum. This results in two frequency vs. displacement matrices (one for the ITD and one for the ILD measurement) that jointly characterize the sound source. These matrices are used as a whole (all displacements and all frequencies) in the following steps. Learning the averaged frequency vs. displacement matrices for each sound source location. This is done separately for the ITD and the ILD measurements. Comparing a measured frequency vs. displacement matrix with the learned matrices to gain a probability distribution for the sound source location (for each frequency channel). This is again done both for the ITD and the ILD measurements. Combining the ITD and ILD probability distribution matrices to gain a single joint probability distribution for the sound source localization. Applying information-theoretic procedures to extract the best estimation for the sound source location.