Fast subword-based approach for open vocabulary spoken term detection

This paper describes an efficient two-stage approach using sub-phonetic segment N-gram index and shift continuous dynamic programming for open vocabulary spoken term detection. With this two-stage search, we attempt to improve performance in both retrieval accuracy and process time. In the speech recognition process, a more sophisticated subword that is shorter than phonemes is used to minimize the effect of recognition error. Then, in the indexing and search process, N-gram and block addressing techniques are adopted to improve the search speed. In addition, in order to reduce missed errors in indexing, the N-best hypotheses are directly added to the inverted index. We investigate the properties of each method and examine their usefulness for the open vocabulary spoken term detection task.