Research On Deep Neural Networks Model Of Radiotelephony Communication Speech Recognition

The number of domestic and international flights has been increasing in recent years,which is an opportunity for civil aviation development,but it is a challenge for civil aviation safety.The radiotelephony communication is crucial for flight safety and is the primary medium of ATC instructions transmission between air traffic controllers and pilots.However,the ATC command misunderstandings event occur occasionally when the fatigue and different accents speakers is working in radiotelephony communication.The use of auto speech recognition technology to convert radiotelephony speech into text is significant for the air pilots to correctly understand the air traffic control command and reduce workload of the controller.Therefore,the deep neural networks model in radiotelephony communication for the recognition speech is studied in this paper.The main following works are:Firstly,in the construction of acoustic models of radiotelephony communication,Deep Neural Networks(DNN),Convolutional Neural Networks(CNN),Long Short-Term Memory(LSTM),Bidirectional Long Short-Term Memory(BiLSTM)networks are combined with HMM for realizing the acoustic modeling of radiotelephony communication,the influence of different neural networks on acoustic model capability of radiotelephony communication is studied.The different discriminative training strategies are applied for model optimization training.The experimental results show that the speech recognition result of radiotelephony communication based on the BiLSTM-HMM model is the best and the discriminative training can effectively improve the performance of acoustic models.However,there is a training complex shortcoming in speech recognition of radiotelephony communication based on HMM.Secondly,in terms of the training model issues,a speech recognition method based on BiLSTM-CTC model is proposed in this paper to achieve end-to-end acoustic modeling in radiotelephony communication.Therefore,the transfer learning and data augmentation are used to train the BiLSTM-CTC acoustic model with the limited radiotelephony communication datasets.Experimental results show that the proposed speech recognition method based BiLSTM-CTC is more convenient than the HMM hybrid model.The transfer learning can construct an acoustic model that suitable for speech recognition in radiotelephony communication.Data augmentation and transfer learning can effectively improve the performance of speech recognition.