Machine Learning of Air Traffic Controller Command Extraction Models for Speech Recognition Applications

Increasing digitization and automation is a widely accepted method to cope with the challenges of constantly increasing air traffic. The analogue communication of air traffic controllers (ATCo) to pilots has been excluded so far from the digitization process. However, the content of this communication is of decisive importance for various automation systems. Although Assistant Based Speech Recognition (ABSR) has recently significantly improved the recognition performance and, therefore, enables the digitization of ATCo-pilot-communication, its adaptation to other airports is a critical and costly process, This is even more important, if ATCos tend to deviate from the published ICAO phraseology: “start reducing to two fifty” instead of “reduce two five zero knots” is just an example. User acceptance requires that these deviations are also correctly recognized. Therefore, this paper presents an approach, which automatically learns a so-called Command Extraction Model from labelled controller utterances. The initial Command Extraction Model without learning only covers 60% of the commands, whereas the automatically learned Command Extraction Model covers more than 98%. With just six hours of training data we could achieve 94