Usable speech detection based on empirical mode decomposition

Abstract

Recently, usable speech criteria have been proposed to extract minimally corrupted speech for speaker identification in co-channel speech. Proposed is a new usable speech extraction method based on the pitch information obtained from a multi-resolution analysis by empirical mode decomposition. The idea is to retain the speech segments that have only one pitch detected and remove the others. Evaluation of this method is performed on the TIMIT database referring to the target to interferer ratio measure. Co-channel speech is constructed by mixing all possible gender speakers. Results do not show much difference for different mixtures. For the overall mixtures 94.97% of usable speech is correctly detected with false alarms of 16.52%.