Optimization And Implementation Of Recognition Efficiency Of Chinese Acoustics Model Based On Convolution Neural Network

In recent years,Artificial Intelligence(AI)technology is developing rapidly.As one o f the key technologies of AI,Automatic Speech Recognition(ASR)receives much attention.With the rise of neural network' research boom,the researchers use CNN to participate in the acoustic model modeling and decode the posterior probabilities of the individual HMM states,which form the CNN-HMM acoustic model.In the following part,it is referred to as the CNN acoustic model.Compared with the traditional acoustic model,the CNN acoustic model achieves better recognition rate.However,the Chinese CNN acoustic model,which is based on the CPU(Central Processing Unit)processor,faces the problem of low recognition efficiency when it is deployed in the speech recognition system and continuously recognizes massive audio data.Therefore,it is urgent to solve the problem that how to improve the recognition efficiency for the Chinese CNN acoustic model in the field of speech recognition.This research is based on the project of IFLYTEK Company,named “Efficienc y optimization of the CNN acoustic model”.Through the analysis and comparison,this thesis optimizes the recognition efficiency of Chinese CNN acoustic model on CPU processor from the following three aspects: using the AVX2 instruction set system to accelerate the decod ing speed of acoustic model,proposing an 8-bit fixed point scheme to solve the problem of slow operation of floating-point speech data,optimizing convolution algorithm to improve the operational efficiency of CNN acoustic model.Finally,the Chinese CNN acoustic model is subjected to the above three optimizations,which can jointly improve the recognition efficiency.In this thesis,both module test and system test are conducted for the above optimizatio n schemes and the results before and after optimization are respectively compared to verify the effectiveness and feasibility of the proposed schemes.The model test results show that the optimized Chinese CNN model can improve recognition efficiency obviously.In the system test,the Chinese CNN acoustic models before and after optimization are respectively deployed to the same speech recognition system to test the recognition rate and recognition efficiency.The results show that the average recognition efficiency of the speech recognition system is relatively improved by 77.58% by using the optimized Chinese CNN acoustic model,while the average recognition accuracy of the speech recognition system after the optimization is reduced within 1%,which meets the requirements of the project's expected indicators.Finally,the stability of the optimized speech recognition system is tested for more than 10 hours.The results show that the optimized CNN acoustic model can work stably for a long time,and the memory usage is normal.In summary,the proposed efficiency optimization schemes in this thesis are effective and feasible for Chinese CNN acoustic model without affecting the recognition rate and the operation stability of the speech recognition system.At present,the above-mentioned optimization schemes have been applied to the in-vehicle speech recognition system of IFLYTEK Company.