ASTRA: AI-Powered Air Traffic Control Training Simulator Reduces Speech Recognition Errors in Singapore

Researchers have introduced ASTRA, a next-generation training simulator designed to automate the role of simpilots—specialized human trainers who role-play as both pilots and air traffic control operators during training exercises. The system addresses a critical bottleneck in Air Traffic Control Operator (ATCO) training capacity by leveraging artificial intelligence to replace manual instruction roles while maintaining high-quality performance evaluation.

A key innovation of ASTRA is its adaptation to regional speech patterns. Existing automated training solutions rely on Western-centric speech recognition models that perform poorly in non-Western operational contexts. Off-the-shelf systems exhibited Word Error Rates (WER) of up to 107.80% when processing Singaporean-accented aviation speech—rendering them effectively unusable in that setting. ASTRA's fine-tuned Automatic Speech Recognition (ASR) pipeline dramatically reduces this error rate to 23.45%, making it substantially more effective for regional deployment.

Beyond speech recognition, ASTRA incorporates an AI-assisted performance evaluation framework that assesses trainee radiotelephony communications across three critical dimensions. The system evaluates accuracy, brevity, and completeness of communications, achieving post-optimization assessment scores of 91.7%, 88.2%, and 86.9% respectively. This automated evaluation standardizes training assessment while reducing the workload on human instructors.

The system is built on open-source foundations including DSPy and Unsloth, enabling scalability and accessibility. By automating simpilot roles, ASTRA addresses training capacity constraints while maintaining training quality through rigorous AI-powered assessment. The approach represents a shift toward standardized, technology-enabled ATCO training that can be adapted to regional speech and operational contexts.