Research On Speech Recognition For Civil Aviation Radiotelephony

The civil aviation radiotelephony(RTF),which is the communication between controllers and pilots,is the core part of the flight guidance process.However,due to the heavy workload of controllers and the noisy environment,misunderstandings that can lead to accidents are likely taken place.Therefore,it's an important problem to ensure the correct understanding of the communication between controllers and pilots,while preventing the occurrence of accidents in the civil aviation area.Automatic speech recognition is an important way to ensure the correct understanding of the civil aviation RTF and reduce accidents because it can be used to process the civil aviation RTF and monitor the behavior of controllers,providing early warning of the dangers caused by wrong instructions.But due to the specialization in pronunciation and intonation of the civil aviation RTF,the common speech recognition technologies cannot be used in the civil aviation RTF without change.At the same time,the data of civil aviation RTF is difficult to collect and mark,increasing the difficulty of the speech recognition.To solve the above problems,this thesis makes an in-depth study on the application of automatic speech recognition technology in the field of civil aviation and explores how to increase the recognition accuracy.In this thesis,some commonly used speech recognition technologies are improved to make them suitable for the recognition of civil aviation RTF.Moreover,to reduce the recognition error rate,this thesis uses technologies such as deep neural network(DNN),transfer learning and language identification to solve key issues of the civil aviation RTF.The detailed work of this thesis is as follows:1.It has studied language identification between Chinese and English civil aviation RTF.In this paper,the necessity of language recognition is analyzed firstly.Then,the corpus construction method is introduced.Based on the analysis and comparison of the models and methods of language identification,the appropriate models and methods are selected,and the language recognition suitable for Chinese and English civil aviation RTF is proposed.Finally,a language identification system has been built using Chinese and English civil aviation RTF data.The experiments are also conducted.2.It has studied the construction of speech recognition system for Chinese-English civil aviation RTF based on GMM-HMM.Firstly,the data of Chinese-English civil aviation RTF are analyzed.And,based on the CMUDICT dictionary and CE-DICT dictionary of Carnegie Mellon University,the dictionaries for Chinese-English civil aviation RTF are constructed semi-automatically.Then,the acoustic model for civil aviation RTF is trained based on GMM-HMM.Finally,the Chinese-English civil aviation RTF system was built,and the correct rate of the speech recognition system was tested.3.It has studied the construction of the speech recognition system for Chinese-English civil aviation RTF based on deep learning.To further improve the correct rate of speech recognition system,the acoustic modeling method of DNN-HMM,which is based on deep learning,is studied.And then,a DNN network structure suitable for Chinese-English civil aviation RTF is proposed.Finally,experiments are conducted and results are analyzed.4.It has studied the construction of cross-lingual deep neural network acoustic model based on transfer learning.It also studies the construction of civil aviation RTF speech recognition system based on depth learning.In view of the deficiency of data sparse problem in Chinese-English civil aviation RTF dataset,the commonly used method of transfer learning was analyzed,and a new transfer learning method was proposed.The new transfer learning method can be used to transfer knowledge from rich resources domain into the civil aviation RTF speech recognition system with light resources.Finally,experiments have been conducted in both Chinese and English civil aviation RTF speech recognition systems,and the result has been analyzed to compare the recognition rates of both systems.