| With the increased number of flights in recent years,China’s civil aviation sector has entered a stage of fast expansion,and civil aviation safety has also become a major problem at this point.ATC voice land-to-air communication is critical to civil aviation safety since it is the primary mode of information exchange between controllers and pilots.However,due to the influence of channel interference,irregular language,mishearing and misrepresentation,and heavy workload,inconsistent understanding of instructions and "errors,forgetting,and omissions" between controllers and pilots may occur during the process of land-air communication,posing certain safety risks.The use of ATC automated voice recognition technology to translate the voice signal of ATC land-air communications into text information is critical to pilots’ proper interpretation of control orders,enhancing controller efficiency,and lowering controller effort.As a result,the following study is being carried out on the topic of ATC speech recognition,as well as the development and use of ATC command voice automated recognition technology based on deep learning.1.An end-to-end voice recognition method based on generative adversarial networks is suggested.Deep neural network-based feature mapping is a popular way for improving speech.This study combines such a mapping approach with an adversarial learning network,introduces a discriminator,and advances the distribution of improved features towards the distribution of clean features via adversarial multi-task training.The acoustic network model is trained with the feature mapping and discriminator networks to improve performance in the ATC voice recognition challenge.The experimental findings on the self-built dataset demonstrate that the algorithm’s average word error rate drops by 16.92% when compared to the baseline model,and the WER decreases to 17.73%.2.To encode and decode the network end-to-end ATC speech recognition,a hybrid CTC/Attention mechanism combining unsupervised pre-training is presented.A substantial amount of training data is required in the ATC voice recognition process in order to train the ATC speech recognition model.Because the labeled finished ATC voice data is limited,this article presents an unsupervised pre-training model to train the unlabeled model,followed by supervised training with the labeled speech data to produce a better acoustic representation.This paper proposes a hybrid CTC/Attention mechanism to fuse unsupervised pre-training to encode and decode the network’s end-to-end ATC speech recognition model,and two different fusion methods are proposed in order to use CTC to align the output sequences and eliminate unnecessary alignment assumptions,while using the attention mechanism to model the input sequences and improve recognition accuracy.When trials are done on self-built ATC voice data,the end-to-end network model of the fused unsupervised pre-trained CTC/Attention mechanism coding and decoding network reduces from 9.48% WER to the baseline model.3.We present a recurrent neural network-based approach for processing ATC audio commands.Relevant research for natural language understanding was conducted,which included analyzing the types and grammatical characteristics of air traffic control instructions,proposing a joint SLU-LSTM model,and conducting a final comparison test in a self-built ATC speech database,with the F1 value of the joint SLU-LSTM model improving by 1.49%compared to the RNN model and by 0.25% compared to the LSTM model. |
