On detecting single-talk intervals in two-input speech separation systems

摘要

Considering mixtures of recorded speech sources, this paper presents a method for detection the time intervals where only one input source is active. Information theoretical criteria which are used for a more general class of sources, can give wrong results in some situations. We propose a different approach to solve the problem. Like the theoretical criteria specify, first, we compute the eigenvalues of the covariance matrix of the mixtures. Then, the eigenvalues ratios are used to find a threshold which allows us to establish if a time interval contains only one speech source. Experimental results, which use two or three mixtures of two speech sources and prove the quality of the proposed method, are also presented.