A 6 mW, 5,000-Word Real-Time Speech Recognizer Using WFST Models

Abstract

We describe an IC that provides a local speech recognition capability for a variety of electronic devices. We start with a generic speech decoder architecture that is programmable with industry-standard WFST and GMM speech models. Algorithm and architectural enhancements are incorporated in order to achieve real-time performance amid system-level constraints on internal memory size and external memory bandwidth. A 2.5 × 2.5 mm test chip implementing this architecture was fabricated using a 65 nm process. The chip performs a 5,000 word recognition task in real-time with 13.0% word error rate, 6.0 mW core power consumption, and a search efficiency of approximately 16 nJ per hypothesis.