Design Of Low Voltage Computing Circuit For Keyword Spotting

With the development of computer technology,Keyword Spotting(KWS)has been widely used in small portable devices with human-computer interaction,so low power consumption is an important consideration for KWS systems.The introduction of approximate computing can effectively reduce the power consumption of KWS system,and based on the characteristics of approximate computing,the introduction of low voltage design can reduce power consumption without loss of precision.Based on the above analysis,this thesis designs a low-voltage-based low-power KWS system.Based on a typical KWS system,this thesis makes detailed optimization from two aspects of approximate computing and low voltage design to reduce the power consumption of the KWS system as much as possible.The main research work of this thesis includes:1)According to the introduction of approximate computing in the circuit,the margin of the critical path delay will increase,and a dual-rail voltage standard cell library design based on TSMC 22nm process standard cell library is proposed.And a low-voltage approximate adder and a low-voltage approximate multiplier based on the dual-rail voltage standard cell library are proposed,which can effectively reduce power consumption while ensuring accuracy;2)The designed low-voltage-based approximate computing unit is applied to the Fast Fourier Transform(FFT)unit in the feature extraction module,which greatly reduces the power consumption of the FFT unit;3)For the calculation unit of the Binarized Weight Network(BWN),a 4-layer addition tree with reconfigurable precision is designed.In addition,the pipeline design is adopted to improve the computing efficiency,and the convolutional layer and the fully connected layer are processed through task scheduling and time-division multiplexing.This thesis completes the low-voltage-based KWS architecture and circuit design based on the TSMC 22nm process.The simulation results show that:in the typical mode(2.5MHz),the high power supply voltage is 0.55V,the low power supply voltage is 0.45V,the power consumption of the KWS system is 37.9μW,and the delay is 1.175ms;in the low power consumption mode(250k Hz),the power supply voltage is 0.45V,the power consumption of the KWS system is 7.8μW,and the delay time is 11.75ms.The layout design area is 0.25mm~2,and the KWS system recognition accuracy can reach 86.3%in the near-microphone scene,and the power consumption can be reduced by up to29.8%.