Third International Conference on Spoken Language Processing (ICSLP 94)
We have proposed Sentence Spotting as one of the most suitable ways to handle spontaneous speech. In this paper, a novel sentence spotting algorithm called "Continuous Structuring Method"(CSM) is presented. This algorithm is able to spot word sequences that satisfy higher constraints such as syntax from the stream of input speech. This algorithm was evaluated in two experiments. In the first experiment, we used 110 sentences uttered continuously by 10 open male speakers. The task had a 113-word vocabulary. We used a phrase pair grammar for the syntactic constraint, with a phrase branching factor of 7.8. A sentence recognition rate of 70.0% and a phrase recognition rate of 89.0% were obtained. The second experiment was for simulated spontaneous speech containing a non-intensional phrase and a long pause and so on, spoken by 5 open male speakers. The syntax and vocabulary were the same as in the first one. We confirmed that CSM shows good performance throughout various utterance styles.
Bibliographic reference. Kiyama, Jiro / Itoh, Yoshiaki / Oka, Ryuichi (1994): "Sentence spotting using continuous structuring method", In ICSLP-1994, 807-810.