ASTRA: A Scalable Next-Generation ATCO Training Simulator with Autonomous Simpilots

ASTRA is an AI-powered air traffic control training simulator that automates the role of simpilots (human trainers) through advanced speech recognition and response generation systems. The system reduces speech recognition error rates from 107.80% to 23.45% for Singaporean-accented aviation speech and incorporates AI-assisted performance evaluation, addressing a critical training capacity bottleneck in aviation safety infrastructure.

ASTRA represents a meaningful advancement in aviation training infrastructure by automating roles traditionally filled by highly specialized human trainers. The core innovation addresses a real operational constraint: air traffic control operator training requires simpilots who can simultaneously role-play multiple pilot and controller voices, limiting training capacity and increasing instructor workload. This bottleneck directly impacts aviation safety readiness in regions like Singapore where Western-centric AI models fail dramatically, achieving word error rates exceeding 100%.

The system's technical foundation builds on open-source components (DSPy and Unsloth), making it reproducible and adaptable across different operational contexts and language variants. The fine-tuned ASR pipeline achieving 23.45% WER demonstrates the value of localized model adaptation over generic commercial solutions. Beyond transcription, ASTRA's end-to-end pipeline interprets controller instructions and generates contextually appropriate responses, creating a closed-loop training environment that reduces dependency on human trainers.

The AI-assisted evaluation framework adds secondary value by standardizing performance assessment across radiotelephony communications (accuracy at 91.7%, brevity at 88.2%, completeness at 86.9%), reducing subjective bias in trainee evaluation. This scalable assessment approach could enable more frequent, consistent training cycles and improved training outcomes across aviation authorities with resource constraints.

Market implications extend beyond Singapore to any region where aviation training infrastructure is constrained by trainer availability. Adoption could reduce training costs while improving standardization and safety outcomes. The open-source approach positions this as a potential template for other specialized training domains facing similar instructor bottlenecks.

• →ASTRA reduces speech recognition error rates from 107.80% to 23.45% for Singaporean aviation speech through localized model fine-tuning.
• →The system automates simpilot roles, directly addressing training capacity constraints that limit air traffic control operator certification programs.
• →AI-assisted performance evaluation standardizes radiotelephony assessment across accuracy, brevity, and completeness metrics with 86-91% post-optimization scores.
• →Built on open-source foundations, the approach enables scalable deployment across different aviation authorities and language contexts.
• →Reduced instructor workload and standardized training could improve safety outcomes while decreasing training program operational costs.