A system for retrieving broadcast news speech documents using voice input keywords and similarity between words

This paper describes a robust speech documents retrieval system that uses voice input keywords. To solve the inevitable problems which arise when the input to the system is speech, i.e. misrecognition, a novel method was developed, where, before the retrieval processing, unproductive keyword candidates are discarded by a grouping processing using the similarity between words and the recognition score of keywords. In retrieval experiments, we used the proposed method to retrieve Japanese broadcast news documents through voice keywords input to the system and showed its e ectiveness.