The Silent Co-Driver
The signal between
the radio calls.
An experimental audio pipeline that estimates vocal stress in Formula 1 team radio and aligns it with real lap times.
- 3rd place
- AI Race Month GrandPrix
- Built in a day, team of two
The question
A pit wall already watches numerical telemetry. What can the driver’s voice add? This prototype examines vocal and textual signals, then puts its estimated stress score beside lap timing. It is an experimental signal, not a clinical assessment.
Three models, one timeline
Whisper provides timestamped transcription, HuBERT reads emotion from audio, and DistilRoBERTa reads emotion from words. Our blend weights tone 65/35 over word choice. Vocal-energy thresholds were measured against 63 radio clips rather than chosen by intuition.
The invisible bug
Whisper was silently dropping whichever speaker talked first. That could leave the engineer’s calm reply while losing the driver’s question. The clue was a loud section of waveform with no corresponding words. Aligning audio, transcript, and timing made that failure visible.
The outcome
The analysis joined 278 laps and 72 radio calls from four drivers at the 2019 Brazilian Grand Prix. Built by two people in a day, the project placed third at AI Race Month. The FastAPI backend runs its models locally.
Care about what people see.
Care just as much about what they don’t.
