Determining the threshold for usable speech within co-channel speech with the SPHINX automated speech recognition system

Abstract

Much research has been and is continuing to be done in the area of separating the original utterances of two speakers from co-channel speech. This is very important in the area of automated speech recognition (ASR), where the current state of technology is not nearly as accurate as human listeners when the speech is co-channel. It is desired to determine what types of speech (voiced, unvoiced, and silence) and at what target to interference ratio (TIR) two speakers can speak at the same time and not reduce speech intelligibility of the target speaker (referred to as usable speech). Knowing which segments of co-channel speech are usable in ASR can be used to improve the reconstruction of single speaker speech. Tests were performed using the SPHINX ASR software and the TIDIGITS database. It was found that interfering voiced speech with a TIR of 6 dB or greater (on a per frame basis) did not significantly reduce the intelligibility of the target speaker in co-channel speech. It was further found that interfering unvoiced speech with a TIR of 18 dB or greater (on a per frame basis) did not significantly reduce the intelligibility of the target speaker in co-channel speech.