Blind Relative Interchannel Delay Estimation

Abstract

Incorporation of localization information improves the convergence and accuracy of many popular multichannel speech separation, extraction, or enhancement algorithms in noisy, reverberant scenarios with overlapping speech of multiple speakers. This letter addresses the problem of multi-source estimation of time differences of arrival (TDOAs), for a microphone array of unknown geometry. The proposed blind relative interchannel delay estimation (BRIDE) algorithm estimates the TDOAs at the microphones and assigns them to each of the concurrent sources. The processing involves a selection of candidate TDOAs based on PHAT-normalized beamforming, unique assignment of chosen time delays to sources based on Pearson cross-correlation scores on broadband beamformer outputs, and final refinement that accounts for symmetries in the source-microphone setups; the latter two steps constitute the main novelty. Experimental evaluation shows that the BRIDE algorithm outperforms clustering-based techniques as well as narrowband methods based on eigenvalue decomposition (EVD), even with ground-truth time-frequency masks, in terms of accuracy of TDOA estimation and effectiveness when applied as spatial prior in guided source separation.