PIT WALL — model backend

PIT WALL — the model backend

Upload or record a few seconds of speech. This runs the same four-stage pipeline the PIT WALL app uses: Whisper for the words, a wav2vec2 dimensional-affect model for the voice, a RoBERTa sentiment model for the text, and a calibration layer that places the result against 2,042 real F1 team-radio messages.

This Space is the backend. It exists so the app's Live Analysis works for anyone, without running anything locally. The app itself is the place to look — this page is the raw endpoint, and is also callable as an API at /gradio_api/call/analyze.

Two things it will tell you honestly: the valence axis of the affect model scores at chance against gold labels, so a state's calm/stressed direction is much less reliable than its high/low activation; and the index is calibrated against F1 radio, so a clip of ordinary speech will be scored as unusually calm.

Source: github/pitwall · Dataset: pitwall-f1-radio-analysis