Facilitating speech recognition with atuomatically generated ATN's, theme identifiers, and syntax analyzers (abstract only)
Thomas M. English, Lois Boggess · 1985
We are developing a system to recognize the utterances of persons with speech handicaps. Central to the system is a component that builds “from scratch” an augmented transition network (ATN) describing the user's inputs. Information in the network is used to predict the user's subsequent inputs, thus facilitating speech recognition. When presented with novel inputs the system adds paths to the ATN. During periods when the user does not require use of the system, the ATN is optimized. Our current focus is upon predictive mechanisms for use when the user's sentence diverges from established paths in the ATN. We have developed procedures to perform rigorous statistical analysis of frequencies of particular words, both globally and in particular contexts. Planned is a module to discover clusters of low frequency words that are related by theme. We are also developing data-driven procedures that “learn” over time which of the system's multiple heuristics is most useful in a given context. These procedures are facilitated by an innovative method of abstracting syntactic categories from user input.