Third International Conference on Spoken Language Processing (ICSLP 94)
This study aims to examine how to combine reliable bottom-up information such as pauses in a top-down approach. As the fir&t step, we select pause information as reliable bottom-up information and study an efficient recognition method that accepts syntactically well-formed spoken sentences that contain natural pauses at free positions. In order to reduce the computational costs that increase exponentially in the recognition process, we examine two ways to use pause information. One way is to merge/pack candidates of the same syntactic categories when detecting pauses in the input speech. The other way is to reduce the phone verification range by using the pause information. We have conducted preliminary experiments using a conference registration task (Japanese speech). The results suggest that this method becomes effective more as the input speech becomes longer.
Bibliographic reference. Takezawa, Toshiyuki / Morimoto, Tsuyoshi (1994): "An efficient predictive LR parser using pause information for continuously spoken sentence recognition", In ICSLP-1994, 1-4.