A Voice Wake-up Circuit Based On Extremely Low Power Feature Extraction And Neural Network

Voice wake-up is a keyword spotting(KWS)technology,which usually applied in a small keyword retrieval system as an interface of complex voice interactive devices.With the development of the Internet of Things technology,the always-on voice wake-up circuit is widely used in mobile and wearable devices with limited battery,which shows strict requirements on power consumption.An extremely low-power voice wake-up circuit is proposed in this thesis,mainly including feature extraction circuit,neural network circuit and post-processing circuit.The feature extraction circuit is based on the improved Mel-frequency Cepstral Coefficients(MFCC)algorithm,which reduces the data dimension from 256 to 128 by adjusting the algorithm execution order.It also greatly reduces the calculation and storage of serial Fast Fourier Transform(FFT).Secondly,fine-grained hierarchical quantization is conducted on the feature extraction circuit,which compresses the data width and reduces the implementation onverhead.Finally,according to the output result of the feature extraction circuit,a gating circuit is designed to turn off the clock of the neural network and the posterior-handling circuit.The throughput of each part of feature extraction circuit is ensured,so that the circuit operates at 8KHz,which is consistent with the sampling frequency.The neural network circuit is designed using a binarized depthwise separable convolutional neural network(DSCNN).The frame-by-frame data multiplexing mode reduces the amount of calculation to 5.7% and storage to 16.5%.Secondly,a configurable voice window is introduced in the neural network circuit,which is configured according to different keywords,improving the flexibility of hardware design.In addition,the processing element(PE)is applicable to both multi-bit and single-bit calculation.The true-code based calculation method reduces the switching power of single-bit data operation.The Re Lu function is used instead of the traditional exponential function in post-processing circuit,so that the storage and calculation amount are compressed without accuracy loss.The proposed near-threshold voice wake-up circuit is implemented in TSMC 28 nm process.The memory array with low-leakage register files is customized,which greatly reduces the overall power consumption of the circuit.The simulation results show that the voice wake-up circuit has an accuracy rate of 97.1% for two keywords,and the power consumption of the overall circuit is 0.47-0.93?W,which meets the requirement of extremely low power consumption.