Word error rate · real ATC recordings
0.6%
110M ASR · ~0.2 s on CPU
SFO · ground · tower · approach
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.
UAL214 C 5000
Filed as flight progress strips — the paper ATC still trusts.
Every clearance is a radio call. A game about it should listen, not offer buttons.
A general LLM needs a server and a GPU, and it can invent an aircraft. I wanted every turn to run in a browser tab.
A 110M speech model, a 2M command model and an 8.5 MB pilot voice. The model only ever writes a typed command line, like UAL214 C 5000.
Mic to pilot voice, all in the browser except speech-to-text.
S1-2M is a GPT-style transformer I trained from scratch on 205,000 simulated situations across 11 airports. It runs int8 in a Web Worker and answers in 0.1–0.2 s.
I wrote a custom 1,349-token BPE so every ATC word is one token. A hand-written KV cache made decoding 4.5× faster; prefix caching made it 7×. Then GRPO, with the game engine itself as the reward.
tokenizer 1,349 tokens · 1 ATC word = 1 token
train 205,000 situations · 11 airports
decode KV cache 4.5× · prefix cache 7×
GRPO · reward = game engine
exact match · p = 0.013 vs SFT control
Measured, not estimated.
Word error rate · real ATC recordings
0.6%
110M ASR · ~0.2 s on CPU
Wrong auto-sends
19%0.9%
after the safety / verifier gate
Wrong-aircraft errors
7444
with GRPO on the game engine
Command model
2M params
0.1–0.2 s · int8 · Web Worker
Pilot voices · Paradee-8M
6× real time
8.5 MB · WASM
Phone app · on-device speech
62ms
Flutter · streaming
Custom BPE vocabulary
1,349tokens
every ATC word is one token
Drag, or scroll on.






