Enhancing Air Traffic Simulation with Automated Speech Recognition and Natural Language Understanding : Developing Autonomous Pseudo-Pilot Systems in NARSIM

Sammanfattning

During the education of Air Traffic Controllers (ATCs), students are required to undergo simulator training together with simulator pilots. These so-called pseudo-pilots have the task of controlling the aircraft in the simulator and responding to commands from the ATC students. This report explores the implementation of an Artificial Intelligence (AI) system used for simulating the pseudo-pilot in the ATC simulator NARSIM. Such a system could greatly improve cost efficiency by reducing the dependency on human pseudo-pilots; availability, since the system can run continuously; reliability, as the AI system could reduce mistakes; and realism, as each simulated pilot can have a distinct voice. This system is of interest to Luftfartsverket (LFV), who rely on simulation and pseudo-pilots to conduct student exercises. The performance of Natural Language Understanding (NLU) and Automatic Speech Recognition (ASR) AI models were tested to determine the feasibility of a real-world implementation. A custom audio dataset was created with the purpose of validating results and fine-tuning the AI models. The study found that within the limited context of this work, an efficient AI model could be developed to act as a pseudo-pilot, with both high accuracy and low latency. While the results were promising, it is worth noting that the fine-tuning and testing data were similar, with the same sentence structure and the same people recording the audio clips. This raises uncertainty about the performance in real-world, dynamic scenarios.