TL;DR: I’ve been beta testing Archer Aviator — a first-of-its-kind app that transcribes ATC radio calls in real time, entirely on-device, using GPS and ADS-B to add context most transcription tools can’t. It’s built by Archer Aviation, the eVTOL company behind the Midnight air taxi, not a traditional avionics shop. On a short VFR hop from KAWO to KPAE, it correctly tagged callsigns, clearances, and taxi instructions — including inferring a full frequency from shorthand ATC never fully stated — while missing a couple of airport names, a solvable problem given the position data the app already has. The beta is open to the public now.

The Gap Aviator Is Filling

Modernizing the National Airspace System is a decades-long, multi-billion-dollar undertaking — I made that case in The Future of ATC: most of what a controller reads you over the radio is structured data or the output of a computable rule, routed through a human voice channel because nothing else was ever built. CPDLC proves the alternative works. It’s also been rolling out since 2016 and still isn’t the default anywhere outside a handful of towers and oceanic tracks.

That’s the top-down path, and it moves at the pace of federal procurement and rulemaking. There’s a bottom-up path too, and it doesn’t need the FAA’s permission at all: pilots running better tools in the seat next to them. Archer Aviator is squarely in that category. It doesn’t change what ATC says or how the radio works — it sits inside the cockpit, listens to the same frequency you already have selected, and turns the raw audio into something more usable, in real time, entirely on the device in your yoke mount. No infrastructure investment, no rulemaking, no waiting for the NAS. Just better tools for the pilot who already has an iPad running.

What Aviator Actually Does

Aviator is built around a simple idea: the radio call you just heard has structure buried in it — a callsign, an instruction, sometimes a frequency or a runway — and most of that structure never makes it past your short-term memory. Aviator’s job is to capture it.

Before a flight, a pre-flight checklist confirms three things are ready: an ADS-B receiver is connected (Aviator fuses ADS-B traffic with radio calls for the transcription context), an audio source is detected from the panel, and the on-device “Archer AI Model” has finished loading. None of it depends on an internet connection once the model is downloaded — a deliberate design choice, since a cockpit at pattern altitude is exactly the kind of place you can’t count on connectivity.

Once you hit Start Recording, every transmission on frequency gets transcribed and displayed as a scrolling feed of tagged bubbles — attributed to whichever aircraft or ATC facility is speaking, not just your own. Where there’s something actionable in the call, Aviator pulls it out and tags it separately: a takeoff clearance shows up as a labeled CLR TAKEOFF block, taxi instructions get parsed into a route string, a frequency change gets flagged even when the controller only gave you the last two digits. A radar-style traffic view, built from the same ADS-B feed, sits alongside the transcript.

After the flight, Debrief mode plays the whole thing back — transcript, audio, and GPS track, all synced to a timeline you can scrub. It’s built to be shareable, too: recent updates let you send a debrief link to another pilot without asking them to install anything first.

Who’s Building It, and Why

Archer Aviation is not an avionics company. It’s an eVTOL manufacturer — founded in 2018 by Brett Adcock and Adam Goldstein, headquartered in San Jose, California, and traded on the NYSE as ACHR — building Midnight, a pilot-plus-four passenger electric vertical-takeoff-and-landing aircraft aimed at short urban air-taxi routes. Archer has been moving through FAA type certification, closing Phase 3 of that process earlier this year, and is participating in the White House’s eVTOL Integration Pilot Program — real-world test operations across multiple states ahead of scaled Advanced Air Mobility rules. It’s also pursuing a parallel certification path for Midnight in the UAE.

None of that is GA avionics. But it explains why a company like this would build Aviator at all. In January, Archer announced a partnership with NVIDIA, integrating the IGX Thor compute platform into its aircraft for onboard AI processing. A cockpit that listens to ATC, understands what’s being said in context, and extracts structured meaning from it in real time is a capability that matters a great deal for an eventual autonomous or reduced-crew air-taxi operation — arguably more than it matters for a GA renter pilot flying VFR patterns. Aviator, running on an iPad in a Sling TSi at KAWO, looks a lot like Archer testing that underlying capability against real radio traffic, real accents, and real airspace, years before it ever needs to fly on Midnight.

That’s speculation on my part about motive, worth stating as such — Archer’s public materials describe Aviator as a safety and situational-awareness tool for pilots and instructors, full stop. But the timing and the compute investment line up too well to ignore.

The Hard Problem: Real-Time, On-Device, No Internet

Running a transcription model on a tablet isn’t new; dictating a text message has worked offline for years. What Aviator does is a different problem. It has to keep up with a live audio stream indefinitely, not a ten-second voice memo — and it has to do it while sharing the iPad with whatever EFB you’re actually flying with. In this cockpit, that’s Garmin Pilot, which is also pulling GPS, rendering a moving map, and syncing traffic — all competing for the same CPU, GPU, and battery budget as Aviator’s model. Apple’s own on-device speech stack only recently got good enough for this class of app: SpeechAnalyzer, introduced with iOS 26, replaced the old SFSpeechRecognizer and its one-minute session cap with a framework built for exactly this kind of continuous, long-form audio. That’s a recent enough shift that an app doing this well in 2026 is genuinely riding the leading edge of what’s possible on a tablet, not falling behind a cloud-based competitor by choice.

“Intelligent Communications” implies a second model sitting on top of raw transcription — something reasoning over the text to decide that “point eight” means 121.8 on ground, or that a string of words is a takeoff clearance worth tagging. That’s a small language model, not just an ASR model, and it’s a newer trick to run well on a tablet. The pattern behind it — a few-billion-parameter model, quantized down to 4-bit weights to fit in an iPad’s memory and run on its neural engine rather than its CPU — is exactly what’s made local extraction and classification tasks like this practical only in the last year or so. It’s a meaningfully harder deployment problem than transcription alone: a second model, running after the first, adding its own latency and memory footprint on hardware that’s already busy.

The airport-name misses in the field notes below are a symptom of that constraint, not a separate issue. A general-purpose model has no reason to favor “Arlington” or “Paine” over any other plausible word it might be hearing — but Aviator already has the GPS position of every transmission it processes, which means biasing recognition toward airports, waypoints, and fixes near the aircraft’s actual location is a tractable, well-scoped improvement rather than a fundamental limit of on-device transcription.

Field Notes: KAWO to KPAE

A short VFR hop to Paine Field on August 25 for a ground VOR check turned out to be a good showcase for what Aviator gets right this early in beta — accurate callsign tracking, correct extraction of clearances and taxi instructions, and one genuinely clever bit of inference. A couple of airport-name misses are worth noting too, mostly because they point at an easy win the team is well-positioned to make.

Departing Arlington, the app caught the full self-announce call, tagged it to N117ZS, and correctly parsed the runway crossing and taxiway callout — “crossing three four at Alfa two.” “Arlington” came through as “altimeter,” a reasonable miss for a beta model working from a name with no built-in reason to expect it.

Same story on frequency change: “Paine tower” was heard as “janes tower,” but the callsign (including “Experimental”), ATIS code, and Langley waypoint were all correct in the same transmission. For a first-of-its-kind app parsing live ATC audio with zero connectivity, getting the substance of the call right while missing one proper noun is a solid result.

The app isn’t just listening for its own callsign. A takeoff clearance issued to N6091T, a different aircraft on the same frequency, was captured and correctly tagged CLR TAKEOFF: RWY 16L.

This one’s worth calling out on its own. Tower simply said “ground point,” and the readback was “point eight, ground, thank you” — no explicit frequency spoken anywhere. Aviator tagged it 121.8 (GROUND), applying the convention that ground frequencies run in the 121.X range to fill in what was only implied. That’s not transcription, that’s the app doing the kind of contextual reasoning the “Intelligent Communications” name promises — and it worked on the first real-world test I threw at it.

Taxi instructions for a Skyhawk on the ground came through clean — TAXI: A → J → D — even though “Paine” was misheard again here (“Millionaire ground”), interestingly after having transcribed the same airport’s own self-identification correctly moments earlier. That inconsistency is a useful data point in itself: the miss isn’t that the model doesn’t know the word, it’s audio-quality-dependent.

Taken together, the airport-name issue is the one clear pattern across the flight, and it’s a good one to have as your main note this early — Aviator already has your GPS position on every transmission, which means airport and waypoint names near your own location are exactly the kind of thing a context-aware model should be able to bias toward correctly. That’s a tractable fix, not a fundamental limitation, and it matters more as a target for IFR use, where a misheard fix or approach name carries more weight than a misheard airport name on a VFR self-announce.

Accent handling (I have a solid Greek accent), for what it’s worth, was a non-issue the entire flight.

Debrief Mode

Every one of the field-note screenshots above is actually Debrief mode, not the live in-flight view — which says something about how useful it is after the fact. The header carries the basics for the leg (tail number, route, date, max altitude, max ground speed, total flight time, distance), and the transcript, GPS track, and a synced altitude/speed trace all scrub together on one timeline. Tap anywhere in the transcript and the map, the audio, and the trace all jump to that moment together. It reads less like a flight log and more like a cockpit voice recorder with the trace data built in.

For IFR debrief specifically, this is the feature that matters most. Reviewing a transcript against the actual assigned altitudes, headings, and clearances after a training flight beats trying to reconstruct it from memory — and having it synced to the ground track makes it obvious exactly where in the approach or the pattern something happened. It’s also the one part of the app that doesn’t need to work in real time, which sidesteps the on-device constraints described above entirely.

Archer added one more piece worth mentioning: as of late August, debriefs are shareable by link, and the recipient doesn’t need Aviator installed to view one. That’s a meaningful shift for a beta that was fully closed a few weeks earlier — it’s now possible to send a CFI or another pilot an actual flight without asking them to install anything first, which is a reasonable way to have the app spread inside a flight school or a training relationship.

Closing Thoughts

Aviator is beta software, and it behaves like it — the transcription is good, not perfect, and Archer says so directly: no guarantee of accuracy at this stage, and you fly with your own training and judgment regardless of what the transcript says. That’s not a caveat particular to this app. It applies to every tool on the panel, including the ones that have been certified for decades. Aviator doesn’t replace a checklist, a chart, or a controller — it gives you a second, persistent record of what was actually said, and in a cockpit that’s most of the value.

That’s also the point of the bottom-up path. The NAS will get CPDLC everywhere eventually, on its own institutional timeline. In the meantime, an app running entirely on the iPad already mounted in the panel — no infrastructure, no rulemaking, no waiting — is doing real work today: catching a frequency a controller only gave in shorthand, keeping a synced record of an IFR debrief, giving a CTAF self-announce the same structure a controller’s readback gets. It’s the same problem from the other end.

The Aviator beta opened to public sign-up on August 21, 2026. If you want to try it in your own cockpit: