Research On Ship Command Speech Recognition Technology Based On Deep Learning

Now is in an era of rapid development of artificial intelligence,voice recognition technology is increasingly used in all walks of life.At present,when naval personnel use the naval training simulator for training,the instructor verbally utters the naval training instructions.After the participants hear the instructions,the naval training simulator is operated accordingly.When the simulator judges the operations of the students,The simulator does not know the specific content of the training instructions,and it is necessary to manually enter the instructions,which will affect the speed of the judgment.When using speech recognition technology to recognize the training instructions,when the instructor speaks the instructions of the ship,the corresponding instructions can be automatically input into the simulator,and the simulator can judge the operation of the students according to the content of the recognized instructions,which can greatly improve the speed of judgement,and then improve the speed and effect of the training ship students.In this paper,based on the fully sequenced convolutional neural network(DFCNN)proposed by HKUST,and on this basis,the network structure is optimized and related parameters are adjusted,and an improved fully connected full sequence volume is proposed.The integrated neural network(enhanced deep fully convolutional neural network,E-DFCNN),combined with the connection time series classification as the loss function during neural network training,is used in the acoustic model part of the ship command speech recognition system in this paper.This part is in the model After the training is completed,the ship commands can be converted into Chinese Pinyin.The language model uses the Transformer language model proposed by Google.Through model building and training,the Chinese Pinyin generated by E-DFCNN can be converted into Chinese characters.Collect and record the relevant ship command audio as the training set and test set of the experiment.Compare E-DFCNN with the original DFCNN and other mainstream models in speech recognition.Experimental results of the acoustic model show that EDFCNN has the same recognition speed The recognition speed of DFCNN is comparable,and the error rate of recognized words is reduced from 16.77% to 9.15%,the error rate of the entire sentence is reduced from 28% to 16%,and the recognition speed and accuracy of these two models are better than other mainstream models.The feasibility and superiority of E-DFCNN combined with Transformer’s speech recognition framework proposed in this paper are verified.