Reconfigurable Computing for Speech Recognition: Preliminary Findings

摘要

Continuous real-time speech recognition is a highly computationally-demanding task, but one which can take good advantage of a parallel processing system. To this end, we describe proposals for, and preliminary findings of, research in implementing in programmable logic the decoder part of a speech recognition system. Recognition via Viterbi decoding of Hidden Markov Models is outlined, along with details of current implementations, which aim to exploit properties of the algorithm that could make it well-suited for devices such as FPGAs. The question of how to deal with limited resources, by reconfiguration or otherwise, is also addressed.