A fixed-point neural network for keyword detection on resource constrained hardware

摘要: Keyword detection is typically used as a front-end to trigger automatic speech recognition and spoken dialog systems. The engine needs be continuously listening, which has strong implications on power memory consumption. In this paper, we devise neural network architecture for keyword present set of techniques reducing the requirements in order make suitable resource constrained hardware. Specifically, fixed-point implementation considered; aggressively scaling down precision weights lowers compared naive floating-point implementation. For further optimization, node pruning technique proposed identify remove least active nodes network. Experiments are conducted over 10 keywords selected from Resource Management (RM) database. trade-off between performance assessed different weight representations. We show that with few 5 bits per yields marginal acceptable loss performance, while requiring only 200 kilobytes (KB) on-board latency 150 ms. A hardware using single multiplier consumption less than 10mW also presented.