Correlation-Based TDOA-Estimation for Multiple Sources in Reverberant Environments

Estimation of time difference of arrival (TDOA) using a microphone array is an important task in many speech related applications such as speaker localization or near-field beamforming for noise reduction. Basically, two different approaches for TDOA estimation have been used: search for the extrema in cross-correlation of microphone signals [7, 10] or blindly estimate the room impulse responses [3]. Both approaches have been approved in many single source scenarios. However, little research work has been done so far on simultaneous TDOA estimation of multiple sources in reverberant environments.

The main idea to extend single source room impulse response techniques to the multiple source case is to split the multi-input multi-output (MIMO) system into several independent single-input multi-output (SIMO) systems. Usually, this is achieved either under the ideal assumption that each speaker is active exclusively during some time intervals [8, 9, 13] or by blind source separation [1, 5].

The cross-correlation technique is both conceptually and computationally simple. Under ideal conditions of a single white source signal in additive white noise without multipath propagation, the TDOA estimate derived from the cross-correlation reaches the Cramer-Rao lower bound and is efficient [7]. In order to combat spurious peaks in the cross-correlation caused by a colored source signal spectrum or even nearly periodic signals like voiced speech segments, different generalized cross-correlations have been proposed [10].