Spoken language interfaces

The intrinsic properties of speech communication (e.g., the presence-of malformed utterances) and the characteristics of current recognition technology (inaccurate recognition) pose special problems for the design of a speech interface. We are interested in understanding these problems and in identifying an interface structure that allows speech to be a useful form of computer input. Ultimately, our goal is to understand how to turn speech into a conventional input modality, well integrated into a multimodal interface that includes keyboard and mouse. To fully exploit the advantages of spoken communication, a spoken language system must afford the user the following forms of flexibility: natural production, natural language, and a natural flow of interaction. The Carnegie Mellon Spoken Language Shell (CMSLS) attempts to provide such flexibility through the use of speaker-independent continuous-speech recognition, natural language processing, as well as rudiment ary “conversational skill” heuristics.