An energy-efficient voice activity detector using deep neural networks and approximate computing

This paper proposed an energy-efficient reconfigurable DNN accelerator architecture for voice activity detection (VAD) based on deep neural networks and fabricated in 28-nm technology. To reduce the power consumption and achieve high energy efficiency, two optimization techniques are proposed. First, the processing elements contained in the DNN accelerator support digital-analog mixed approximate computing, including multi-step quantized multiplication units and time-delay based addition units. Second, the proposed approximate computing units can be dynamically reconfigured to adapt to different computing accuracy requirements. The proposed approximate computing can significantly reduce the power consumption by 76% similar to 88% compared to standard digital computing units. Implemented under TSMC 28 nm HPC + process technology, the layout size of the prototype system is 0.52 mm(2), and the estimated power is 6 similar to 12 mu W. The energy efficiency of our work achieves 33.33 similar to 66.67 TOPS/W, which is over 6.5X better than the state-of-the-art architecture.