Machine-Actionable ATC Radio Understanding with Generative Audio–Text Models

Short description of the innovation study

Air traffic control (ATC) radio is how controllers issue taxi clearances, hold-short instructions and runway crossings, yet this information exists mostly as speech. General-purpose speech recognition struggles with noisy radio audio, accents and aviation phraseology, and a transcript alone does not tell a system which instruction was given, to which aircraft, or for which runway or taxiway. In this study, NoamAI uses European HPC systems to train a generative audio-text model that turns ATC audio directly into a transcript, the intent of the instruction and the aviation entities it refers to, using airport and traffic context to keep outputs consistent. The model is evaluated for accuracy and for robustness to noise and accents against standard speech recognition and pipeline approaches. The goal is a validated model that makes ATC communications machine-actionable for airport safety monitoring and analysis.

Organisations involved: 

Main SME: NoamAI