Towards Proactive Air Traffic Safety with Speech LLMs: Transcription, Attribute Tagging, and Readback Detection

Abstract

Air traffic communication (ATC) is a highly specialized domain where noisy acoustic conditions, rapid speech rates, and domain-specific terminology pose significant challenges for automatic speech recognition. We investigate the use of Speech Large Language Models (Speech LLMs) to address these challenges by combining accurate transcription with structured information extraction. We leverage publicly available ATC datasets and apply data augmentation together with metadata-enriched training, which improves multi-turn dialogue handling and reduces transcription errors. On the test set, our model achieved a word error rate of 16.70% and a role classification accuracy of 96.60%. Furthermore, we propose a system for extracting callsigns, commands, and values, enabling automated readback detection. Case studies show that the system can determine whether pilots correctly repeat critical instructions, thereby supporting safety monitoring. Taken together, these results demonstrate the potential of Speech LLMs to deliver reproducible transcription accuracy, reliable role attribution, and proactive safety monitoring in air traffic communication.