Relative transfer function (RTF) estimation utilising peaks in time-domain RTF

Abstract

An effective algorithm for a more accurate estimation of relative transfer function (RTF) to improve the directional-blind source separation-based RTF estimation is presented. The peaks in time-domain RTF (PTDR) as a feature of time-domain RTF to improve the conventional algorithm are utilised. The PTDRs contain important characteristics of RTF; however, they are smoothed due to the geometric constraint in the conventional algorithm. In the proposed algorithm, the PTDRs are recovered by employing the semi-non-negative matrix factorisation algorithm to effectively find the PTDRs from other peaks. A series of experiments using interfere and various noises demonstrates effectiveness of the proposed RTF estimation algorithm utilising the PTDR.