Task-Oriented Approaches to Automatic Speech Recognition

We describe several computer-implemented systems for automatic speech recognition. The systems are designed for specific communication tasks in which the human talker and machine interact in a disciplined dialog. One speaker-dependent system recognizes individual spoken words and provides information on airline flight schedules. The same system, combined with a programmed syntax analyzer, recognizes whole sentences that are chosen from a sub-set of natural English. Another system provides automated telephone directory assistance by recognizing voiced-spelled names and speaking back the requested telephone number. Still another system recognizes digits spoken by any speaker, and provides the capability for automatic voice dialing. All the speech recognition systems utilize dynamic programming to match spoken input with stored reference templates. The reference templates for each vocabulary word are constructed from linear predictor coefficients (LPC) which are measured over word utterances. We give performance data for the systems when they are operated over conventional dialed-up telephone connections.