Effectiveness of an expanded dictionary in information retrieval system by keyword spotting

摘要

An information retrieval system by voice input for broadcast news is investigated. Our purpose is to design the system in which robust speech recognition is possible to inquiries from a user with comparatively high flexibility. The transcription result of a news speech is used to contract a user's keyword dictionary for inquiry. In order to solve a problem of time difference between training data and evaluation data, the N-gram language model was created from the news story in the newest Web data, and it was adapted to evaluation data. In addition, widely used vector space model and the LSI method were investigated to deal with the words out of vocabulary. Through the experiments, some effect was verified. The simulation experiment of keyword potting was conducted and the system validity was shown.