How it works

From METAR to spoken broadcast, in five steps

  1. 1
    Live observation
    Authoritative METAR
  2. 2
    Deterministic parse
    Structured fields
  3. 3
    Phraseology engine
    Standard ATIS/AWOS phraseology
  4. 4
    Per-language renderer
    EN / ES / FR / PT
  5. 5
    Synthetic voice
    Broadcast audio

AeroVox changes how aviation weather is delivered without touching the data, the chain of authority, or the phraseology pilots already expect. The AI is the voice, never the author.

The single most important property: every spoken word is produced by deterministic rules operating on the authoritative observation and the airport's configured runway and approach data, never improvised by a generative model. The voice is the only AI component. That is why it is auditable and defensible to a regulator.

Step 1 · Live observation

Same data the official broadcast uses

AeroVox pulls the airport's live METAR, the standardized weather observation already produced by the existing ASOS/AWOS/AWS infrastructure. No new sensors, no parallel weather source, no second source of truth. If the FAA trusts the observation, AeroVox does too.

METAR KDFW 311053Z 19008KT 10SM CLR 24/21 A2986 RMK AO2 SLP101

For the live flagships, AeroVox re-checks the observation on a polling cycle (default 30 minutes, matching the standard ATIS reissue cadence) and only re-voices a broadcast when the weather actually changes. Stable weather costs nothing.

Steps 2 and 3 · Deterministic parse and phraseology

Words come from rules, not from a model

The METAR is parsed into structured fields (wind direction and speed, visibility, sky layers, temperature/dewpoint, altimeter, weather phenomena), handling both U.S. format (statute miles / inHg) and international format (meters / hPa / CAVOK / TEMPO).

A fixed phraseology engine then assembles the broadcast script by rule, aligned to FAA Order JO 7110.65 conventions: aviation digits ("niner"), phonetic information letters, standard order, the data-group cadence pilots are trained to copy. The same observation always produces the same script.

"Fort Worth International Airport, information Bravo. Wind two three zero at eight. Visibility one zero. Sky clear. Temperature two four, dew point two one. Altimeter two niner eight six."

Crucially, the engine refuses to invent. If a weather field is not in the observation, it doesn't appear in the broadcast. Every weather phrase is traceable to a specific field in the METAR; runway and approach phrases come from the airport's configuration. Easy to audit, with no generative step that could invent content.

Step 4 · Per-language renderer

Native, not translated

Each supported language has its own rule-based renderer, with its own digit words, lexicon, cardinal-number grammar, unit conventions, and word order. Translation is never done by a model. The same parsed observation flows into separate native renderers for English, Spanish, French, Portuguese.

  • EN · "Scattered clouds at two thousand two hundred"
  • ES · "Nubes dispersas a dos mil doscientos pies"
  • FR · "Nuages épars à deux mille deux cents pieds"
  • PT · "Nuvens dispersas a dois mil e duzentos pés"

Adding a new language is a new renderer subclass, not a model retrain.

Step 5 · Natural voice

Natural voice, ATIS cadence

The finished script is voiced by a neural TTS engine, at a measured ATIS pace deliberately chosen so a pilot can copy the broadcast without missing data: the same group-by-group rhythm a controller uses on the radio. Information letters advance Alfa, Bravo, Charlie on each weather change, just like a real ATIS reissue.

The voice layer can be swapped (a different broadcaster voice, a per-language voice, or a customer's preferred voice) without touching the engine. The audio is the only piece a buyer ever hears change.

Why this approach is defensible

Side by side: a generative-AI broadcast vs. AeroVox

Generative AI approach

Asks a model to "describe the weather"

The model can hallucinate, paraphrase, drop fields, or invent phrases under stress. Two runs of the same input can produce different broadcasts. There is no defensible audit trail back to the original observation.

AeroVox approach

Renders broadcast text by rule

Words come from a deterministic engine reading authoritative weather data. Same observation in, same broadcast out, every time. Every phrase traces to a specific field. The voice cannot say something the data did not contain.

Compliance and operational fit

Designed to slot into existing operations

  • No change to the data. AeroVox is downstream of ASOS/AWOS/SWIM: it modernizes how the existing observation is delivered, not what it says.
  • No change to authority. The official observation and the official frequency assignment remain unchanged. AeroVox is the voice layer, not a chain-of-authority change.
  • Phraseology-aligned. Output is modeled on FAA Order JO 7110.65 / ICAO Annex 10 phraseology; full per-region alignment is completed during pilot integration.
  • Auditable by design. Every broadcast has a deterministic mapping back to a METAR plus a timestamp, both replayable and independently verifiable.
  • Software-defined. Changes ship as configuration and software updates rather than a hardware program.
  • Dual-use. The same engine serves military expeditionary, civil non-towered fields, towered ATIS modernization, and international markets.

Hear it live

It's already running on real weather.

Pick any airport in the live demo and you'll hear today's robotic broadcast, then the same observation in the AeroVox voice, in up to four languages.

Hear the live demo Request a briefing