Usable Speech Processing: A Filterless Approach in the Presence of Interference

Abstract

A speech signal is commonly corrupted by non-stationary interference in different environments. The interference distributes itself nonuniformly over the temporal regions of the signal. This makes certain segments more speech dominant and hence, “usable” for speech processing applications like speaker identification and speech recognition. The main part of this paper examines the detection of usable segments in a co-channel environment having two interfering speakers. A method based on determining the Target-to-Interferer Ratio to detect the usable portions in which one speaker is very dominant over the other is discussed. A short discussion of additive noise distortion follows. Experimental results are presented.