Always-On Speech Recognition Terminals: Designs based on approximate computing methods

Abstract

Always-on speech recognition terminals (ASRTs), which detect a user’s speech all the time and convert it into text for the speech interaction system, have broad prospects. However, the conventional implementations of ASRTs, which are always based on accurate computing design, suffer from redundant power consumption, high processing latency, and extensive memory access. Since the processing of the algorithms used for ASRTs has an error-tolerance property in its nature, this article adopts analog and digital approximate computing techniques to solve these challenges.