Demonstration abstract: Automatic speech recognition for resource-constrained embedded systems

Abstract

We demonstrate the design and implementation of a prototype hardware/software architecture for automatic single word speech recognition on resource-constrained embedded de vices. Designed as a voice-activated extension of an existing wireless nurse call system, our prototype device continually listens for a pre-recorded keyword, and uses speech recognition techniques to trigger an alert upon detecting a match. Preliminary experiments show that our prototype achieves a high average detection rate of 96%, while only dissipating 28.5 mW for continuous audio sampling and duty-cycled speech recognition.