Speaker identification enhancement under cochannel conditions using sinusoidal model-based usable speech detection

The accuracy of present day speaker identification systems (SID) is degraded in adverse acoustical environments. The idea of usable speech is to identify and extract those portions of degraded speech which are considered useful for SID. Recently, a usable speech extraction system was proposed to classify cochannel speech as usable speech and unusable speech for SID. Speech segments can be declared usable based upon a target-to-interferer energy ratio (TIR). By considering only usable speech for SID instead of the corrupted cochannel speech, it is seen that there is an increase in the accuracy. A novel usable speech detection measure using the sinusoidal model of speech and ESPRIT (estimation of signal parameters via rotational invariance technique), spectral estimation is proposed and investigated, which resulted in 82% correct detection of usable speech segments based on TIR. The usable speech frames extracted using ESPRIT when tested with the SID system, resulted in 84% accuracy in detecting speaker identity as compared to using entire co-channel speech which resulted in only 45% accuracy.