2026
The Silent Co-Driver
Reads driver stress from F1 team radio and lines it up against real lap times — 3rd place, AI Race Month
A pit wall watches tyre temperatures, fuel and sector deltas — dozens of streams, all numeric. The one channel nobody processes in real time is the driver's own voice, and it carries frustration and exhaustion before either shows up in a lap time.
This listens to Formula 1 team radio, scores how stressed the driver sounds from 0 to 100, and lines that score up against his real lap times.
Three models from the Hugging Face Hub, chained: faster-whisper large-v3 for transcription with a timestamp on every word, HuBERT reading emotion from the sound, and DistilRoBERTa reading emotion from the words. The blend is the part we wrote — tone weighted 65/35 over word choice, because a driver saying "I'm fine" through gritted teeth is not fine. Vocal energy separates the two ways of not being okay: stress is loud, fatigue is flat. Both thresholds are the 90th percentile measured across 63 real radio clips, not numbers that felt right.
278 laps, 72 radio calls, 4 drivers from the 2019 Brazilian Grand Prix. The radio comes from a Hugging Face dataset and the lap times from the F1 timing feed via fastf1 — two unrelated sources, joined only on a timestamp. They agree: the call that says "outlap critical" lands on a genuine out-lap, and the safety car talk lands on the laps that took two minutes instead of 73 seconds.
The hardest bug was invisible. Whisper was silently dropping whichever speaker talked first, so we were keeping the engineer's calm reply and throwing away the driver's question — on a project built to read the driver's voice. Caught it because the loudest part of a waveform had no words against it.
FastAPI backend, no framework on the frontend, every model running locally so it works with the wifi off. 3rd place at AI Race Month · GrandPrix, built in a day by a team of two.
Next: Café de' Mercanti → Building something like it? Get in touch →