A Robust EM-Based Method for Underdetermined Convolutive Blind Source Separation of Speech

摘要

Underdetermined Blind Source Separation is a technology to estimate the individual source components from the mixed signals where there are few sensors than the sources. This paper presents a novel method to separate the convolutive mixed signals. This method is divided into two steps. The first step separates the signals at each frequency bin using the sparse property of speech signal by EM algorithm. The second step solves the permutation problem via the time difference of arrival (TDOA) and the result of EM algorithm in the first step. Experimental result with real reverberant records shows a good performance of the proposed method.