High Robust Low Power Voice Activity Detection Design Based On DNN

In recent years,more and more voice recognition intelligent applications have entered everyone's daily life.Voice Activity Detection(VAD)is one of the key technologies in the field of speech recognition.Its accuracy and power consumption play an important role in the whole speech recognition system.Starting from the problems of low accuracy and high power consumption under the non-stationary noise faced by traditional VAD accelerators,this thesis designs and implements a high robust low power VAD based on Deep Neural Network(DNN).Based on the algorithm characteristics of VAD,this thesis makes algorithm optimization and hardware implementation for feature extraction and speech classification.Firstly,the feature extraction algorithm based on Mel Frequency Cepstrum Coefficient(MFCC)is optimized,and fast fourier transform,discrete cosine transform and speech classification are realized by high robust DNN,which improves VAD performance in Low SNR and non-stationary noise and achieves high robustness of VAD.Secondly,the digital-analog hybrid approximation calculation is introduced,and the multi-order quantization shared multiplier is used to realize the dynamic precision configurable operation of VAD,which reduces the computational complexity and circuit complexity,and realizes the low power consumption and high energy efficiency of VAD.Based on the TSMC28nm,the system area of VAD is 0.52mm~2,the working frequency is 1.6MHz,the power consumption is about 6~12?W,and the energy efficiency can reach33.33~66.67TOPS/W.The experimental results show that the proposed algorithm optimizes the VAD algorithm and hardware,compared with Price,the accuracy is increased by 9%,and the energy efficiency is about 6.5times that of Thinker,which solves the bottleneck of consumption and precision faced by current VAD.This thesis provides a new solution for VAD design.