Fast Spoken Term Detection using pre-retrieval results of syllable bigrams

We propose a method of the Spoken Term Detection (STD) based on a priori retrieval results in which plural syllables are used as query terms. In the proposed method, all N-syllable combinations such as syllable bigrams are searched for in spoken documents. In the first step of the method, the retrieval results are prepared a priori, where pre-retrieval results include candidates with scores matching those of each N-syllable sequence. Given a query, the syllable sequence of the query is divided into plural syllable sequences whose lengths are the same as those of the pre-retrieval results. In the second step, the candidate sections are filtered by using the scores of query's syllable combinations. This reduction in the number of candidate sections for detailed matching leads to a large reduction of the retrieval time. In the third step, these candidates sections are rescored by performing detailed matching. Experimental results show that the proposed method reduces the retrieval time by 93% with a performance degradation of less than 2 points.