A High-Performance Hardware Speech Recognition System for Mobile Applications

While commercial speech recognition systems remain limited in their capabilities, research systems are now relatively mature, although computationally expensive. Specialized hardware is one solution to this problem, and certainly the only solution at present for mobile applications. We present a queue-based hardware architecture for large-vocabulary, speaker-independent, continuous, real-time speech recognition in the mobile environment, demonstrating better than real-time per-formance. We base our results on simulation of approximately one hour of speech data for a 5,000 word vocabulary.