Airport San Francisco Intl (KSFO) San Francisco, California
Procedure VOICE APPROACH RWY 28L Ground · Tower · Approach
Issued AMDT 1 · 14 NOV 2026 First public release · MIT

Project 01 · Personal · May – Nov 2026

SkyControl

Voice air-traffic control at SFO, running on small models I trained. No LLM, no GPU at run time.

You are ground, tower or approach. Key the mic, speak a clearance; pilots read it back and fly it.

An airliner on short final over a line of runway approach lights at dusk.
FIG. 1 Short final, runway lights in sight.
Speech
ATC-ASR 110M
Command model
S1-2M · 2M params
Pilot voice
Paradee-8M
Run time
CPU + browser · no GPU
Licence
MIT
Role
Everything, end to end
TWR · RADIO CHECK
  1. YOUUnited two one four, climb and maintain five thousand.
  2. S1-2MUAL214 C 5000
  3. UAL214Climb and maintain five thousand, United two one four.

Notes

An ATC game that actually listens.

  1. 1

    Radio is clipped, fast and full of callsigns. I wanted a game where you talk like a controller and the aircraft simply do what you said.

  2. 2

    A large language model in the loop is slow, needs a GPU somewhere, and can pick the wrong aircraft. That last one is the error that matters.

  3. 3

    So I trained small, specialised models and put rules and a safety gate around them. The model only ever writes a typed command line.

Plan view · how it works

Mic to readback, fix by fix.

A stylised approach chart. The flight path visits nine fixes, one per pipeline stage, from the microphone to the pilot's spoken readback, and lands on runway 28L. SAN FRANCISCO BAY 37°40′N37°35′N 122°25′W122°20′W D5D10D15 N 28L SAY AGAIN
  1. MIC

    Key the mic

    Push to talk. You are ground, tower or approach at SFO.

  2. ASR

    ATC-ASR 110M

    Speech to text, trained for radio. 0.6% WER on real ATC recordings, about 0.2 s on CPU.

  3. RULES

    Clean-up rules

    Spoken numbers, callsigns and phraseology are normalised before anything else runs.

  4. PARSE

    Command parser

    A deterministic grammar goes first. The model only steps in when it fails.

  5. S1-2M

    Command model, in the browser

    A 2M-param transformer I trained from scratch writes one command line. int8 in a Web Worker, 0.1–0.2 s.

  6. GATE

    Safety gate

    Send, confirm, or “say again”. Wrong auto-sends fell from 19% to 0.9%.

  7. ENGINE

    Game engine

    The aircraft flies the clearance. The same engine was the reward during RL.

  8. PHRASE

    Phraseology

    The typed command is turned back into a proper readback.

  9. VOICE

    Paradee-8M pilot voice

    8.5 MB of TTS at 6× real time in WASM. The pilot reads it back, in the browser.

Detail A · the only thing the model writes

One word, one token. One line out.

I built a custom 1,349-token BPE where every ATC word is a single token. The model reads the situation and the transcript, then writes a typed command. Nothing else can reach the game.

Transcript in

  1. united
  2. two
  3. one
  4. four
  5. climb
  6. and
  7. maintain
  8. five
  9. thousand

Command out

UAL214 C 5000

Profile view · what runs where

Every model runs on your machine.

Profile view. Speech recognition runs on CPU in about 0.2 seconds; the command model runs int8 in a browser Web Worker in 0.1 to 0.2 seconds; the pilot voice runs in WebAssembly at six times real time. ASR S1-2M GATE VOICE CPUATC-ASR 110M · ~0.2 s WEB WORKER · INT8S1-2M · 0.1–0.2 s RULESsend / confirm / say again WASMParadee-8M · 6× real time RWY

Minimums · measured results

Small models, measured.

SkyControl measured results
ComponentValueConditions
Speech · ATC-ASR 110M0.6% WERReal ATC recordings · ~0.2 s on CPU
Command · S1-2M2M paramsGPT-style, trained from scratch on 205,000 simulated situations across 11 airports
Tokeniser1,349 tokensCustom BPE · every ATC word is one token
Decoding4.5× · 7×Hand-written KV cache · prefix caching
RL · GRPO79.6 → 83.8%Exact match, game engine as reward · SFT control 82.0%, p = 0.013
Wrong aircraft74 → 44Errors, before and after GRPO
Safety gate19% → 0.9%Wrong auto-sends
Pilot voice · Paradee-8M8.5 MB6× real time in WASM
Phone app · Flutter62 msOn-device streaming speech
INSET B Exact match, %

Axis from 75%. p = 0.013 vs SFT control.

INSET C Wrong auto-sends, %
INSET D Wrong-aircraft errors

Tower view · reel

Cleared to land.

FIG. 8 Reel. Gate to touchdown.

Ramp · screens

Ground, tower, approach.

An airliner parked at a gate at dusk, jet bridge attached.
FIG. 2 Ground: pushback and taxi.
An airliner seen head-on on the apron.
FIG. 3 Hold short.
An airliner landing at night, landing lights blazing head-on.
FIG. 4 Tower: cleared to land.
A wing above clouds at sunrise.
FIG. 5 Approach: vectors to final.
Earth at night from orbit, city lights glowing.
FIG. 6 Eleven airports of training data.
A business jet silhouetted against a sunset.
FIG. 7 Phone app, on-device speech.

Communications · stack

Built with

Web
TypeScript · Next.js · three.js · Mapbox GL
Training
PyTorch · HF TRL / PEFT · vLLM
Inference
ONNX Runtime Web · sherpa-onnx
Mobile
Flutter
Testing
Playwright

Next slice in time

Stratify

Aug 2026 – presentAn MCP server for honest options backtestingOpen

All work