1 October 2025
desk@falconatc.com

Air traffic control voice recognition and inference

Voice recognition software in air traffic control simulators today are unforgiving in phraseology yet inaccurate in voice detection. The rapid development of large language models (LLMs) have opened up many possibilities for voice recognition ATC simulators to accept a wide variety of nonstandard speech.

The most basic layer of text-based (as transcribed by voice recognition) instructions to air traffic control simulators is like a line of code. Consider a traditional simulator that taking a climb command:

Delta eight niner five climb one seven thousand

(computer: Amending DAL895 target altitude to 17000 ft)

We can separate a simple sentence into two components: an aircraft identifier, which is the aircraft we are commanding, and the instruction. However, things can get a bit messy.

Delta eight niner five climb uhhh one seven thousand

(computer: “uhhh one seven thousand” is not a valid altitude!)
(user: WTF?! This simulator is so stupid!)

Delta ate nice five climb one seven thousand

(computer: Delta ate nice five is not an aircraft I can find!)
(user: I’m ragequitting.)

The fix

The traditional method of fixing this was hardcoding error detection. For example, automatically replacing “ate” with “eight” and deleting any uhhhs. While this can solve 50% of the most common errors, this still leaves a wide failure surface that is enough to frustrate a user to leave. This is where we add the LLM layer.

LLMs can infer what the user wants to do and communicate it to the simulator in a way the simulator understands without needing complicated hardcoding. It can also identify and correct errors in the voice recognition layer.

Delta ate nice five climb uhh one seven thousand →Delta eight nine five climb one seven thousand

(computer: Amending DAL895 target altitude to 17000 ft)

Drive-by benefits

Aviation uses standardized phraseology. This allows controllers and pilots to communicate in a predictable way even without having ever met. Adherence to this standardized phraseology is taught and graded in pilot and ATC schools. LLMs can now not only understand nonstandard phraseology, but recognize and note its use.