An Energy-efficient Reconfigurable Hybrid DNN Architecture for Speech Recognition with Approximate Computing

Abstract

This paper proposes an hybrid deep neural network (DNN) for speech recognition and an energy-efficient reconfigurable architecture with approximate computing for accelerating the DNN. The hybrid DNN consists of two network models: a binary weight network (BWN) for twenty key words recognition; a recurrent neural network (RNN) for processing acoustic model of high precision common words recognition. To accelerate the hybrid DNN and reduce the energy cost, we propose a digital-analog mixed reconfigurable architecture with approximate computing units, including: a BWN accelerator with analog multi-chain delay-addition units for bit-wise approximate computing, and a RNN accelerator with approximate multiplication units for different calculation accuracy requirements. Implementation and simulation with TSMC 28nm HPC+ process technology, the energy efficiency of proposed architecture can achieves 163.8TOPS/W for twenty key words recognition and 3.3TOPS/W for common words recognition. Comparing with State-of-the-Art architectures, this work achieves over 1.7X better in energy efficiency with approximate computing.