The problem
Even as automation grows, air taxis and other aircraft will continue to rely on voice communication with air traffic control. Modern text-to-speech models can produce convincing speech from short voice samples. A small change to an altitude, heading, or frequency in an instruction could alter its operational meaning entirely.
Our approach
GWU developed an evaluation framework that measures synthetic controller speech along several dimensions: recovery of the instruction content, resemblance to the original speaker, spectral characteristics, and procedural cadence. The study found that instruction content can be recovered very accurately even when pitch and timing still differ from real controllers.
The study deliberately does not test pilot acceptance. Its purpose is to provide acoustic evidence for assessing risk, and to inform defenses.
Looking ahead
Pilot readback and controller hearback already give opportunities to catch an unauthorized instruction. The next priority is studying how synthetic speech interacts with these procedural defenses.