A sub-word level matching strategy in a speech understanding system.

Abstract

This paper describes the design of our speech understanding system and reports preliminary experimental results which support an underlying idea. Our speech understanding system is based on the idea that since we can locate some phones very accurately in the incoming speech, an interval bounded by such robust phones should be taken as the unit of matching. This eliminates the problem of ambiguous word boundaries in searching the most likely sentence, as such an interval have little correspondence to word boundaries. We introduced a new processing level called partial lattice hypothesis level to realize the above idea in our hierarchical SUS. The typical word pronunciation and their varieties are precompiled a lattice form, while modifications at word boundaries are performed by applying the phonological rules during the linguistic decoding. In order to observe how the unit of matching affects the search space of a SUS, we carried out a preliminary experiment of spotting two sets of short phonetic strings against a phoneme lattice-a set containing those strings bounded by robust phones, and the other set for the counterpart. The results verified the validity of our idea.