Research On Speech Recognition Method Of Chinese-English Cross-lingual Civil Aviation's Radiotelephony Communication

With the economic globalization,the number of domestic and international flights has increased year by year.The field of civil aviation has received great attention.Ensuring and improving the safety of civil aviation has become a top priority.From a domestic perspective,a controller needs to communicate with domestic and international flights within the same time frame,which involves frequent switching between English and Chinese.That will increase the probability of errors in radiotelephony communication.Therefore,the application of speech recognition technology in the field of radiotelephony communication will help reduce the semantic expression errors of radiotelephony communication,and it is of great significance to improve the work efficiency of controllers.This paper studies the speech recognition methods for Chinese-English civil aviation's radiotelephony communication(CARC).The specific research contents are as follows:Firstly,the characteristics of Chinese ang English radiotelephony communication languages were studied to build a database of Chinese-English radiotelephony communication.Acoustic models of Chinese-English radiotelephony communication were constructed using deep neural network(DNN),long-short-term memory network(LSTM),convolutional neural network(CNN)combined with hidden Markov model(HMM),and comparative experiments were performed on the database.Through the comparison of word error rates,it is proved that the model of DNN-HMM performs well in the field of Chinese-English CARC.Secondly,a convolution depth neural network-hidden Markov model(CDNN-HMM)for Chinese-English CARC is proposed,which resolves the differences in language rules and diversity issues of Chinese and English radiotelephony communication.When phonemes are labeled,CMU standard English phonemes are mapped to TIMIT standard English phonemes,and the reconstructed language model is used for recognition.A low frame rate is added to the acoustic feature extraction process to speed up decoding and training.A comparative analysis of the word error rate shows that the application of the acoustic model of CDNN to solve the speech recognition problem of CARC has obvious advantages.The method of phoneme mapping can further improve the recognition performance,and the low frame rate added effectively reduces the complexity of model training and decoding.