The Design And FPGA Verification Of Speech Command Recognition Algorithm For Assisted Driving

With the help of deep learning technology,the accuracy of speech keywords spotting algorithm has been greatly improved,and the algorithm has high value in applications of smart home,vehicle system and other fields.However,compared with the traditional algorithm,the amount of parameter and operations of the speech keywords spotting algorithm based on neural network are huge.The deployment of speech keywords spotting recognition system with high accuracy on the vehicle terminal equipment is faced with the problems of insufficient storage resources and computing power.In order to meet the requirements of high real-time,high accuracy and miniaturization of the speech keywords spotting system for vehicle terminal equipment,it has certain research significance to realize the efficient speech keywords spotting algorithm based on software and hardware collaborative design.A speech keywords spotting algorithm for vehicle terminal equipment is designed in this thesis.By using depthwise separable convolution and attention mechanism,the recognition accuracy is improved and the model is compressed from the perspective of convolution neural network structure.The size of the model is further compressed by the layered quantization of the model,and the loss of accuracy caused by quantization is reduced at the same time.The dataset is preprocessed and the training strategy is adjusted to improve the performance and robustness of the model.In addition,a neural network accelerator is designed based on speech keywords spotting algorithm and FPGA,including convolution operation module,state control module,data cache module and function unit module,to accelerate the forward reasoning process of speech keywords spotting.Finally,the verification system of speech keywords spotting algorithm is built based on FPGA to realize speech keywords spotting.The experimental results show that the recognition accuracy of 10 keywords is 95.16% on the Speech Commands dataset,and the recognition accuracy of hierarchical 8-bit quantization is 94.56%.On the self-made dataset,the recognition accuracy of 5 keywords is 98.94%.The network model size of speech keyword spotting is 148.125 KB,and the hierarchical quantization model size is 37 KB.The reasoning time of speech keywords spotting algorithm on neural network accelerator is only 8.52 ms,and it takes a toal of 40.145 ms for a recognition time.The research content of this thesis has a certain reference significance for the terminal implementation of speech keywords spotting system in the future.