Deep neural network-based automatic speech recognition for ATC-pilot audio transcription

Artificial Intelligence (AI) has demonstrated ability to
\n manage complex processes highly effectively, and thus is widely
\n seen as a key component in future airport ATM systems. Future AI
\n tools for ATM will rely on digital data, such as surveillance, radar,
\n weather, flight plans, for their operation. However, the foundational
\n Air Traffic Control Officer (ATCo)-pilot communication medium
\n is voice, which is a vital source of situational data. Controller Pilot
\n Data Link Communications (CPDLC) has been developed as an
\n alternative, text-based communication delivery method, however
\n ATCo-pilot communications will not be completed transitioned to
\n this framework in the near-term future. Moreover, as CPDLC is
\n a one-to-one communication paradigm, the additional situational
\n awareness of other traffic provided by traditional party-line VHF
\n communications is potentially lost. Therefore, an automated speech
\nto-text translation tool can be seen as a missing link, enabling
\n traditional ATCo-pilot voice communications to be automatically
\n translated and input into a datalink system such as CPDLC.
\n To this end this paper presents a Machine Learning (ML)
\nbased Automatic Speech Recognition (ASR) framework that is
\n able to accurately translate ATCo-pilot speech communication to
\n text, achieving a Word Error Rate of only 6.13%. Moreover, the
\n presented model is able to extract seven entities with an accuracy
\n and F1-score of 91.8% and 84.4% respectively, which is similar
\n to previously presented models but can only capable of extracting
\n three. A detailed design of the framework is provided to enable its
\n replication by the wider research community.