27.2 A 6mW 5K-Word real-time speech recognizer using WFST models

Hardware-accelerated speech recognition is needed to supplement today's cloud-based systems in power- and bandwidth-constrained scenarios such as wearable electronics. With efficient hardware speech decoders, client devices can seamlessly transition between cloud-based and local tasks depending on the availability of power and networking. Most previous efforts in hardware speech decoding [1-2] focused primarily on faster decoding rather than low-power devices operating at real-time speed. More recently, [3] demonstrated real-time decoding using 54mW and 82MB/s memory bandwidth, though their architectural optimizations are not easily generalized to the weighted finite-state transducer (WFST) models used by state-of-the-art software decoders. This paper presents a 6mW speech recognition ASIC that uses WFST search networks and performs end-to-end decoding from audio input to text output.