A speech separation system that is robust to reverberation

Abstract

A speech separation system has been built up that is self-supporting (i.e. does not require prior knowledge of pitch tracks obtained from the speech signals on their own as in the work of Stubbs and Summerfield, 1990), and is also robust to reverberation. A two-dimensional harmonic sieve that tracks harmonics in both amplitude and frequency has been developed. With the help sometimes of directional information in order to attribute the resulting pitch measurements to the appropriate speakers at times of pitch crossings or following breaks in voicing, reliable pitch tracking and accurate pitch extraction have been achieved in the present system. The present system can successfully extract one speech from mixed speech even when there are more than two overlapping speech signals. >