Research On Instruction Recognition For Air Traffic Control Communication

With the rapid development of China’s civil aviation transport enterprises, the air traffic flow hasincreased dramatically which requires more and more high-level air traffic controllers. At present, itdepends on the air traffic control (ATC, for short) training simulator to finish the freshmen’s trainingwhere the training of ATC communication in English is considered as a very significant part. As weknow, the current ATC training mainly depends on the setting of special pilot seat. During the training,the ATC learners complete their ATC communication training under the guidance of the instructorworking as the pilots. However, due to the limited number of coaches, the ATC learners take turns tobe the pilots and cooperate with the training of other learners. In this situation, both the training timeand effects are very hard to be guaranteed, and there is a big unnecessary redundancy existing in thecomputing system.In this thesis, we analyze voice instruction recognition in ATC training simulator deeply andfocus on some key issues of voice instruction recognition in ATC communication, the main contentsare as follows：(1) According to the Radiotherapy Communications for Air Traffic Services given by the CivilAviation Administration of China (CAAC), which gives a very detailed analysis of all the terms, wesummarize its grammatical features, design the basic instruction format, and give different languagemodels (grammatically strict) of approach control terms, tower control terms and area control termsduring the different flight phrases. Through the construction of the special word pronunciationsequences, we tackle the problems of pronunciation recognition of the special terms in ATC, such asnumbers, letters, and some specific terms, etc.(2) Since the domestic ATC trainers’ English pronunciation characteristics are different fromabroad ones’s, in order to acquire high recognition accuracy in ATC, the current English acousticmodel should be adapted. So, a hidden markov model is established to train and update this model,especially, the numbers and letters, which tend to pronounce wrong in the process of the voiceinstruction recognition, are trained specifically.(3) Usually, the character string commands output from the speech recognition engine neithercontain any structure and semantic information, nor directly provide the required format processed bythe simulator server. Therefore, a parser tool, JavaCC, is used to replace parts of the instruction, anda formal description of XML of instruction is given. Besides, we design a kind of dynamic query matching method to increase the system fault tolerance and give the interface data standard andcommunication way to interact with the training simulator.(4) On the basis of the Sphinx-4, an AIRS (ATC Instruction Recognizer System) is designedand realized. The experimental results show that the accuracy rate of voice Instruction recognition hasbeen largely improved by the acoustic model adapting and the dynamic matching fault tolerance ofinstruction. AIRS system can reach ATC training simulator requirements and lay the foundation forthe realization of the replacement of the special pilot by the automatic pilot in the ATC trainingsimulator.