Automatic Detection and Removal of Disfluencies from Spontaneous Speech

Unlike rehearsed and prepared speech, spontaneous speech contains high occurrence of disfluencies, like repetitions, filled pauses, and hesitations. Disfluencies can seriously hamper the word recognition accuracy of an Automatic Speech Recogniser (ASR), by increasing word insertion and deletion and rejection rates. In this paper we introduce signal processing algorithms to automatically identify and remove repetitions and filled pauses from spontaneous speech before passing it to an ASR for transcription. The algorithms are tested with Dragon NaturallySpeaking Speech Recogniser and show significant improvements in the word recognition accuracy, and ensuing reductions in substitution and deletion and insertion errors.