Recognition And Processing System Of Air Traffic Control Speech Command

With the development of the aviation industry,the air traffic has become more and more busy,and the demand for air traffic control(ATC)has also increased.However,the training cost of air traffic control related personnel is high,and the complicated navigation arrangements require large human resources.It has become an urgent need for machines to complete part of the work of air traffic controllers.The recognition and processing of air traffic control speech commands is the basis for completing related tasks.Based on the characteristics of air traffic control commands,this paper studies the algorithms of automatic speech recognition(ASR)and information extraction and processing of ATC commands.The main contents are as follows:Firstly,the ASR of ATC commands is studied in the paper,the technologies of speech sampling quantization,preprocessing and feature extraction are specifically analyzed.The FBank feature extracted from speech frame is determined the input of the speech model.After analyzing and comparing the recognition effects of different end-to-end model frameworks and neural network encoders,the CTC framework based on the self-attention mechanism is determined.On this basis,the self-attention mechanism is improved according to the characteristics of speech commands of ATC,then we propose the speech model called CNN-WSA-CTC.Speech decoding technology can combine the speech model and 3-gram language model to realize the speech recognition function of converting speech into text.In terms of ATC commands information extraction and processing,the classification and the grammatical composition and characteristics of commands are analyzed in detail in the paper.An improved IOB notation method and Semantic framework suitable for ATC commands which uses semantic slots to extract information are proposed.The information extraction process is simplified to the sequence labeling process.In this paper,the SA-CRF model is proposed by combining neural network sequence modeling with CRF sequence labeling,and the validity of the model is verified.The process of information extraction and semantic slot filling based on sequence tagging and intention matching is designed.Then,the pronunciation characteristics and sentence characteristics of ATC commands are analyzed in this paper,and the generation method of ATC commands sentences is designed.Then 30 hours of speech data based on the ATC commands corpus is recorded.Based on the data set,the speech model,the language model and the sequence labeling model are trained and the optimal structure of the model is designed.In order to improve the recognition effect of the speech model,the speech data is processed by scaling and adding noise.Then 100 hours of speech data is obtained for speech model training.Finally,a word error rate of 1.44% is obtained on the test set.the sentence error rate of the sequence labeling model is 0.33%.In the final joint test of the model,the sentence accuracy rate is 93.9% on the noise-free test set and 90.3% on the noisy test set.Finally,the paper designs a simulation system for the recognition and processing of ATC speech commands.The ASR part and the information extraction and processing part are integrated.The python-based Tkinter framework is used for user interface design,and the Tensorflow framework is used to build and load the model.The process of ASR and information extraction is implemented and tested to verify the effectiveness of the model and system.