README.md

metadata

title: AInterviewer SER Analyzer
emoji: 🎙️
colorFrom: blue
colorTo: indigo
sdk: docker
app_port: 7860
pinned: false

AInterviewer SER Analyzer

Speech Emotion Recognition (SER) inference Space for the AInterviewer project.

Endpoints

• GET /health — Returns {"status": "ok", "svm_loaded": bool}
• GET /version — Returns {"version": "ser_colombian_v1"} when SVM loaded, "emotion2vec_plus_base" in fallback mode
• POST /analyze — Accepts WAV file, returns {"dominant": str, "distribution": {...}, "model": str}

Architecture

Two-stage inference:

1. emotion2vec_plus_large extracts 768-dim audio embeddings
2. SVM Pipeline (StandardScaler + SVC) classifies into 4 emotions: angry, happy, neutral, sad

Falls back to emotion2vec+ direct classification if the Colombian SVM artifact is unavailable. That fallback is an operational safety net, not the final Colombian model deployment.