Towards a Speech - to - Speech Translation System ( 1 )

We describe our Universal Parser Architecture and its use in a speech translation system, based on the Machine Translation system under development at the Center for Machine Translation at Carnegie Mellon. To "understand" natural language, a system must have syntactic knowledge of the language and semantic knowledge of the domain. The Universal Parser Architecture allows grammar writers to develop these kinds of knowledge separately in a declarative manner, and the compiler/interpreter integrates these two knowledge bases dynamically in order to parse input sentences. We recently integrated our system with a speech recognition system to accept spoken sentences (rather than typed sentences) by extending our runtime parser component to handle "noisy" phoneme sequences (of spoken utterances) that possibly include recognition errors. The nature of this modification is explained and examples are presented. We find our architecture very suitable for speech translation, as it combines the use of domain semantic knowledge with a highly efficient runtime parsing algorithm, thus accommodating the increased search space necessary for parsing speech input.