README.md

metadata

title: PIT WALL — Model Backend
emoji: 📻
colorFrom: red
colorTo: gray
sdk: gradio
sdk_version: 5.49.1
python_version: '3.12'
app_file: app.py
pinned: false
license: mit
short_description: Live radio analysis backend for the PIT WALL Space
preload_from_hub:
  - openai/whisper-small.en
  - audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim
  - cardiffnlp/twitter-roberta-base-sentiment-latest
startup_duration_timeout: 1h

PIT WALL — model backend

This Space is the backend for PIT WALL. Go there first; this page is the raw endpoint.

Why it is separate

PIT WALL is a Next.js frontend against a FastAPI backend, and the two run together in the project's Docker image. Hugging Face charges for Docker and Gradio Spaces on personal accounts, so the public deployment is a static Space — which serves the precomputed 12-race corpus perfectly, but has nowhere to run a model. Live Analysis was disabled in public as a result.

Free accounts in good standing may host two ZeroGPU Gradio Spaces, so the model backend lives here and the static frontend calls it cross-origin. Gradio accepts any origin when the host is not a localhost alias, which is the case on *.hf.space, so no proxy is involved.

What it runs

Stage	Model
Speech to text	openai/whisper-small.en
Voice affect	audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim
Text sentiment	cardiffnlp/twitter-roberta-base-sentiment-latest
Calibration	percentile map fitted on 2,042 real F1 radio messages

Every model call goes through the pipeline/ package copied verbatim from the main repository. This is not a second implementation — if it ever disagrees with the app, that is a build bug rather than a fork.

As an API

curl -X POST https://<this-space>.hf.space/gradio_api/call/analyze \
  -H "Content-Type: application/json" \
  -d '{"data": [{"path": "https://example.com/clip.mp3", "meta": {"_type": "gradio.FileData"}}]}'
# -> {"event_id": "..."}   then GET .../analyze/<event_id> for the SSE result

The response matches POST /api/analyze in the FastAPI backend field for field, so the frontend renders live and precomputed results with the same component.

Two honest caveats

The valence axis is at chance. Validated against CREMA-D gold labels, the affect model's arousal scores 79.4% against a 61.8% baseline, but valence scores 60.9% against 61.9% — no better than guessing. A state's high/low activation is reliable; its calm/stressed direction is much less so. This is measured and reported rather than smoothed over.

Calibration is F1-specific. Scores are percentiles against team radio, which is shouted over engine noise through a compressed channel. Ordinary speech will be placed as unusually calm because it is, relative to that reference.

Quota

ZeroGPU time is charged to the caller, not to this Space. Anonymous visitors get a couple of GPU-minutes a day, which is roughly a hundred clips. Signing in to Hugging Face raises it.