Abstract

Simulations – especially human-in-the-loop real-time simulations – are important in the air traffic control (ATC) domain to train controllers and to test new features for controller working positions. One important reason for such simulations is the measurement of human workload. Verbal communication of aviation operators – contributing to this workload – is a central mean for safety and efficiency of air traffic. Speech recognition and understanding (ASRU) has reached pre-industry level, is about to enter operations, and therefore will become a vital part in training. The technology affects working procedures and reduces controller workload by roughly 20%. Thus, ASRU must be considered in simulations.
This paper describes a process model to integrate ASRU in ATC simulations. The model consists of three steps for efficient integration and adaption of ASRU: (1) collection of in-domain speech data for tuning of acoustic and language models, (2) compilation of configuration files and adaptation of speech understanding algorithms, and (3) manual checking of automatic transcriptions and extracted, semantic meanings of speech utterances.
We evaluate the process using a multiple remote tower environment case study. In this study, recognition error rates for words and callsigns were reduced by a factor of three compared to first simulations and command recognition rates increased from 81% to 92%.
Similarly feasible results are expected for other new ATC simulations with ASRU using the proposed process model.