Lie to Me — Deception Detection Model
Trained on the Real-Life Trial Dataset (121 courtroom clips).
Modalities: facial emotion, voice (MFCCs, pitch), body language, linguistic features.
Performance
- AUC: 0.94
- Accuracy: 84% on held-out test set
Features
Emotion (DeepFace), Voice (librosa), Body (MediaPipe), Language (rule-based LIWC-style)
Usage
import joblib, json
model = joblib.load("model.pkl")
scaler = joblib.load("scaler.pkl")
feat_cols = json.load(open("feature_cols.json"))
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support