High-performance speech recognition is extremely computationally expensive, limiting its use in the mobile domain. We therefore propose a low-power hardware speech recognition architecture for mobile applications, exploiting the orders-of-magnitude efficiency improvements dedicated hardware can offer. Our system is based on the Sphinx 3.0 software recognizer developed at Carnegie Mellon University, capable of large-vocabulary, speaker-independent, continuous, real-time speech recognition. We show through cycle-accurate simulation that our hardware, targeting the backend search stage of recognition, is capable of recognizing speech from a 5,000 word vocabulary 1.3 times faster than real-time, within an approximately 200mW power budget.
Bibliographic reference. Bourke, Patrick J. / Rutenbar, Rob A. (2008): "A low-power hardware search architecture for speech recognition", In INTERSPEECH-2008, 2102-2105.