Design Of Low-power Keyword Recognition Feature Extraction Module For High Noise Scence

In recent years,more and more speech recognition applications have entered everyone's daily life.Among them,keyword recognition is one of the indispensable key technologies in the field of speech recognition applications.However,these application devices have strict requirements for recognition accuracy in low-power consumption and noise scenes.Therefore,a special integrated circuit module for low-power keyword recognition feature extraction in high noise scenes will be proposed in this thesis.In this thesis,the design is mainly optimized in two aspects,namely algorithm and circuit.And the power consumption of keyword recognition system feature extraction is reduced as much as possible while ensuring recognition accuracy.The main work includes:(1)Based on the traditional Mel Frequency Cepstrum Coefficient(MFCC)feature extraction algorithm,a simplified rectangular Rectangular Mel Frequency Cepstrum Coefficient(SRMFCC)feature extraction algorithm architecture is designed,compared with the traditional MFCC,the work of this thesis can maintain good recognition accuracy in noise scenes with different signal-to-noise ratios.(2)The mode switching design with different signal-to-noise ratio noise is realized and the multi-order quantization shared multiplier in which the Fast Fourier Transform(FFT)module multiplication calculation can adapt to different calculation accuracy requirements,compared with the standard multiplication calculation,it can significantly reduce power consumption.Based on the TSMC 22nm process,in this thesis,it is completed that the design of a low-power keyword recognition feature extraction module in high-noise scenes.The experimental results show that SRMFCC algorithm reduces the overall data volume by 50%.Compared with the traditional MFCC,the multiplication calculation of SRMFCC is reduced by 77.8%.The verification system is binarized weight network(BWN).The total area after layout is 0.72mm~2.The area of feature extraction is 0.195mm~2.The main frequency of operation is 250k Hz and the delay time is 12ms.The power consumption is 2.81u W in low power mode and5.32u W in high precision mode,compared with the main stream feature extractions,power consumption is reduced by 74%,the recognition accuracy is 91.3%at-5dB SNR,95.47%at 5dB SNR.