Design Of Ultra Low Power Intelligent Keyword Spotting Chip Based On Quantitative Feature Extraction

Keyword spotting system is the interface of human-computer interaction to control the complex system,which needs to keep normally on all the time.Therefore,low power design of keyword spotting chip is very important.The existing research usually has fixed data bit width and redundant hardware computing.To solve these problems,an ultra-low power intelligent keyword spotting chip based on quantitative feature extraction is designed,which consists of feature extraction circuit,neural network circuit and post-processing circuit.A custom Mel frequency cepstral coefficient(MFCC)algorithm is proposed for feature extraction circuit.Firstly,an intelligent quantization model based on genetic algorithm is designed by analyzing the bit width variation of internal data of MFCC algorithm,which can balance hardware cost and accuracy.Secondly,Fourier transform circuit in real sequence is optimized,and uses the custom memory to reduce the power consumption.In addition,by embedding adaptive speech detection module,keyword spotting system can be closed in the no-speech environment.For the design of neural network,a Temporal Depthwise Separable Convolutional Neural Network(TDSCNN)is proposed.And the neural network uses Process Element(PE)based on parallel adder tree to reduce the delay of the system recognition of keywords.In PE design,a kind of parallel accumulation method is put forward to reuse 8bit and 1bit calculations.And the near memory architecture is proposed to meet the high data bandwidth of parallel computing.In the design of the post-processing circuit,the post-processing algorithm based on hardmax function replaces the original softmax function,which further reduces the hardware implementation complexity of the post-processing circuit.Finally,in TSMC 28 nm process,a near-threshold keyword spotting chip is designed,and the accuracy of keyword spotting chip for one keyword can reach 98.2%,and 96.5% for two keywords.When the keyword spotting chip is gated by the speech detection module,the power consumption of the overall circuit is 401 n W.When the chip normally works,the power consumption is 492 n W.Therefore,the chip can meet the requirements of very low power consumption.