Low Power Keyword Recognition Accelerator based on Approximate Calculation of Deep-Shift Neural Network

In this paper, an ultra-low power always-on keyword spotting (KWS) accelerator is implemented based on an optimized convolutional neural network (CNN). To reduce the power consumption while maintaining the system recognition accuracy, we proposed a deep-shift neural network (DSNN) to reduce the hardware resources requirements without recognition accuracy loss. This new form of conventional CNNs and fully connected networks can almost waive multiplication operations and greatly reduce the power consumption. Implementation results show that this DSNN based accelerator can support four keywords real-time high accuracy recognition. Compared with the binary weight neural networks(BWN), the proposed DSNN reduces the multiplication-accumulation (MAC) operation by 50%. Our work proves that DSNN may be a better option than BWN in some applications.