SPEECHACTS - A SPOKEN-LANGUAGE FRAMEWORK

Speech Acts is a prototype testbed for developing spoken natural langu age applications. It lets software developers without special expertis e in speech or natural language create applications with which users c an speak naturally, as if they were conversing with a personal assista nt. We believe we have achieved a degree of conversational naturalness similar to that of the outstanding Air Traffic information Systems di alogues, and we have done so with simpler natural language techniques. Currently, SpeechActs supports a handful of speech recognizers: BBN's Hark, Texas Instruments' Dagger, and Nuance Communications' recognize rs. These recognizers are all continuous-they accept normally spoken s peech with no artificial pauses between words-and speaker-independent- they require no training by individual users. For output, the framewor k provides text-to-speech support for Centigram's TruVoice and AT&T's TrueTalk. As an existing set of applications, SpeechActs is both a pro of of concept and an effective system that about a dozen people now de pend upon when they travel. Powerful enough to be useful, it is easy t o use with little training. As a framework for building speech applica tions, SpeechActs' contributions include a Unified Grammar to create s ynchronized grammars for speech recognition and semantic parsing, reus able plug-in speech components, and the Swiftus natural language proce ssor. SpeechActs also includes important discourse management techniqu es. Both the discourse stack and a simple context queue in SpeechActs model the current state of the discourse so that SpeechActs can respon d naturally. These simple, straightforward components combine to make SpeechActs a powerful framework in which to design speech applications .