Behind the app

Building an ATC speech engine that sounds real and understands everyone

July 2026 · 6 min read

Anyone can build an app that plays a recording of a controller and asks you to tap a button. What almost nobody has built is a system that actually talks to you like a real tower, listens to you talk back, and knows whether you got it right, on a phone, in real time, no matter who you are or where you learned to speak. That is the thing Tower Drill set out to do, and it turned out to be far harder than it looks. This is the story of why.

Two opposite problems, one impossible standard

A radio trainer that feels real has to solve two challenges that pull in completely opposite directions. The controller in your ear has to sound like a real controller: not a soothing narrator, but the fast, clipped, band-limited voice that comes crackling out of a real radio and makes a new pilot's stomach drop. And at the same time, the app has to understand you, reading it back, under stress, in your own voice and your own accent, and know instantly whether you nailed it or blew the runway assignment.

Neither of those is what phones are built for. Everyday speech technology is tuned for calm, clear, conversational English: "text mom I'm running late." Aviation radio is the opposite of that in every way. It is a foreign dialect where three is "tree," nine is "niner," numbers run together with no pauses, and a single transmission packs a runway, a heading, a frequency, and a squawk code into one breathless line. Point an ordinary voice assistant at it and it falls apart: it hears a frequency as a phone number and a clearance as gibberish. Getting it right meant building something that does not exist anywhere else.

Making a machine sound like a real controller

The easy, wrong instinct is to pick the nicest, clearest voice a phone offers and call it done. But a real controller never comes through clean. The voice is thin and forward, squeezed through the narrow band of a radio, with the low warmth and airy top stripped away, and it arrives fast, because real controllers are busy. If your trainer sounds like a meditation app, the first transmission at a busy Class C airport will still flatten you, because you practiced against something that does not exist in the real world.

So we did the harder thing. Every transmission is shaped to sound like it actually came over the air, and each controller position, Ground, Tower, Approach, sounds like a different person, the way it does at a real field, so your brain learns to shift gears on a handoff instead of hearing one flat robotic voice throughout. You can even push the delivery faster, closer to the rapid-fire pace of a major airport, so the speed stops being the thing that ambushes you on your first solo. Dialing all of this in was not a single clever trick. It was months of listening, tuning, comparing against live ATC recordings, and adjusting until it stopped sounding like an app and started sounding like a controller having a busy afternoon.

Understanding every pilot, in every voice

This is the harder half by far, and it is where most of the work went. Aviation is global. The pilot reading back a clearance might have grown up speaking English in Texas, or Tagalog in Manila, or Portuguese in São Paulo, and every one of them says "niner" a little differently. The system has to understand all of them, saying words that are not even real English, having never heard your particular voice before, and it has to decide in a heartbeat whether your readback was correct. A single missed word in aviation is not a typo: it is the difference between the runway you were cleared onto and the one with traffic on it.

Teaching a phone to do that is not something you write once and ship. It is a long, grinding loop of feeding the system the strange vocabulary of the radio, testing it against voice after voice, watching exactly where it stumbles, and refining. We drilled it on the way pilots actually talk, not the way a textbook says they should: the ums, the half-corrected numbers, the students who mix "tree" and "three" in the same sentence without noticing. Every mishearing we found became a fix. Then we tested again, and found the next one. That cycle, repeated far more times than we care to count, is the difference between a demo that works when the founder speaks slowly into it and an app that understands a nervous student pilot on their worst day.

And it does not just transcribe you, it judges you fairly. It knows the difference between the words you must get exactly right, the runway, the hold-short, the squawk, and the small human stumbles that a real controller would never fail you for. It gives you credit for a correct readback delivered imperfectly, and catches the genuine error hiding inside a confident one. That judgment, the part that makes the practice actually mean something, is the hardest-won piece of all, and the part you will not find in an app that just plays you a clip and moves on.

All of it, on your phone, in real time

The shortcut would have been to send your voice off to a server somewhere and let a giant model in a data center do the thinking. We refused to. Every bit of this runs on your device, and that choice cost us enormous effort, because a phone is a far tighter space to pull it off in. But it is the only way to deliver what actually matters to a pilot:

• It works anywhere. On the ramp, in a metal hangar, at a remote field with no signal, in airplane mode. Your practice never depends on a connection.
• It answers instantly. There is no lag between keying up and being understood, exactly like a real radio.
• Your voice never leaves your phone. Nothing you say is uploaded, stored, or fed into anyone's cloud. What you say in the cockpit stays in the cockpit.

Why it was worth doing the hard way

It would have been easy to ship a polished-sounding trainer with a pretty voice and a stock recognizer, and it would demo beautifully. It would also train you against a controller who is not real, understand you a beat too slowly, misjudge your readbacks, and quit the moment you lose signal. We built the version that is genuinely hard to build, because a pilot's confidence on the radio is not something you fake with a recording. Getting a phone to sound like a real controller on a busy afternoon, and to understand a nervous student from anywhere in the world reading it back, is not a feature. It is the whole reason the app exists, and as far as we know, nobody else has put in the miles to get it this close to the real thing.