A Natural-Language Speech Interface Constructed Entirely as a Set of Executable Specifications

Abstract

SpeechNet is a collection of speech-accessible hyper-linked objects called sihlos. Sihlos are deployed over the Internet and are accessed by remote speech browsers. When a speech browser accesses a sihlo, it begins by download-ing a grammar file which is used to configure the browser in order to achieve high recognition accuracy. Sihlos are hy-perlinked in a manner that is similar to the linking of html pages. SpeechNet provides non-visual access to knowledge which is analogous to visual access provided by the web. One of the sihlos can answer thousands of spoken pseudo-natural-language questions about the solar system. This sihlo has been constructed entirely as a set of executable specifications of the language that it can process. SpeechNet The world-wide web is primarily based on visual brows-ing of hyperlinked text and graphical objects. The huge knowledge base available on the web is largely inacces-sible to visually-challenged users. The SpeechNet project at the University of Windsor (Frost 1999b) overcomes this problem by augmenting the web with a network of speech-accessible hyperlinked objects called sihlos. Each sihlo consists of a language processor and a gram-mar defining the syntax of the input language. When a remote speech browser contacts a sihlo, it downloads the grammar in order to configure the speech recognizer. This alleviates a major difficulty in building accurate speech interfaces to large knowledge bases. Sihlos enable the knowledge to be divided into small modules each having a “small ” language associated with it. This allows highly accurate user-independent continuous-speech recognition. Hyperlinking sihlos allows the user to navigate through a complex interrelated body of knowledge using spoken prompts in a way that is analogous to the way in which the web is navigated using visual prompts.