A new architecture of a speech understanding system - A hybrid of a hierarchical and a network models

This paper describes the design of our speech understanding system, One of the disadvantages of the hierarchical model is the inadequate treatment of coarticulation effects across word boundaries. In analyzing continuous speech in Japanese, word-by-word matching fails easily because our language has many important functional words which decide semantic cases. The network model suggests one solution to the coarticulation problem by compiling all the knowledge sources in a single network and finding the optimal path in it. However, a network model places strong constraints on the grammar of the language to be accepted. It is unacceptable since our language has much more flexible structures than English. Our speech understanding system is based on the idea that we can locate some phones very accurately in the incoming speech and that an interval bounded by such robust phones should be taken as the unit of matching. Such an interval have little correspondence to word boundaries, We add to our hierarchical speech understanding system a new partial lattice hypothesis level in which the above idea is realized.