Low-power Circuit Design Of Convolutional Recurrent Neural Network For High-noise Keyword Recognition

Speech keyword recognition is an important part of human-computer interaction,which requires real-time,high precision,low power consumption,and high robustness.Once the noise environment becomes complicated,the current keyword recognition system can hardly meet the requirements in terms of performance.Therefore,this paper has realized a set of keyword recognition system solutions that can be applied to a low SNR(signal-to-noise ratio)environment.Algorithmically,the speech keyword recognition algorithm of convolutional recurrent neural network is realized.First,a set of suitable lightweight hybrid neural networks for speech keyword recognition with low SNR is designed.Second,a fixed-point quantitative training method is proposed,and then the weight is converted into 10bit,and the data is directly fixed-point into 16bit.When the SNR is-5d B,the recognition accuracy of 4 keywords can reach 86.4%,and the accuracy can reach more than 97%in a clean environment.In terms of hardware,this thesis has realized the circuit optimization design of the above-mentioned network.First,intra-frame calculations is proposed for real-time speech recognition,which disperses centralized neural network calculations into different time periods.The real-time recognition delay is reduced by 80%.Second,the reconfigurable computing array is designed to implement CNN,RNN,and FC operations,so as to make full use of computing resources and save60%of the computing circuit area.Third,the multiplier in the calculation array is replaced with a multi-bit wide multiplier,which provides the possibility of hierarchical quantization.The circuit design of this thesis adopts TSMC 22nm craft,the area after the layout is 0.76mm~2.When the working frequency is 250KHz,real-time voice processing can be realized,and the power consumption is about 6.89?W.