(An) area-efficient reconfigurable CNN-LSTM architecture for automatic speech recognition system

Abstract

Recently, deep neural networks replace all stages of state-of-the-art automatic speech recognition (ASR) algorithms. Also, current segregated modules for voice activity detection (VAD) and speech recognition (SR) are highly area-inefficient. Therefore, we propose the first end-to-end neural network reconfigurable real-time ASR hardware architecture using a CNN-LSTM dual core system. The aggregation of separated modules and adaptation of the Winograd algorithm to CNN core both drastically reduce the overall area overhead. We further improve the energy efficiency by frame packaging scheme and partial power/clock gating onto the core. The proposed architecture achieves 61.2% area reduction and 93.6% energy reduction of always-on process.