Seiichi Nakagawa

Japanese spoken document retrieval considering OOV keywords using LVCSR system with OOV detection processing

This paper describes a Japanese spoken document retrieval system that is robust for Out-of-Vocabulary (OOV) words. A standard approach to spoken document retrieval is to automatically transcribe spoken documents into word sequences, which can be directly matched against queries. In this approach, the documents including OOV words and words mis-recognized as other words can not be retrieved. To avoid this problem, we propose a novel method of spoken document retrieval considering OOV keywords. One of our approach is to create a index from multiple recognizers outputs to deal with transcribed documents including mis-recognized words. The index becomes better to use multiple recognizers which have different characteristics to one another. The other is to use both word based indexing for in-vocabulary keywords and syllable based indexing for OOV keywords, and to switch them according to in-vocabulary/OOV keywords in the query. Evaluation results clearly show that this approach benefits from advantages of both indexing methods and that the proposed technique is quite effective in robustly retrieving spoken documents.