sofi233/truesight-voice-deception

trusight-voice-deception (v1.0.0)

Pretrained acoustic classifier for the TRUSIGHT voice analysis service.

Model Summary

• Architecture: RandomForestClassifier(n_estimators=200, class_weight='balanced', random_state=42)
• Framework: scikit-learn 1.9.0
• Input Dimension: 42 acoustic features (pitch, jitter, shimmer, tempo, pauses, spectral, MFCCs)
• Output Classes: 0 = Truthful, 1 = Deceptive
• Decision Threshold: 0.5

Performance Metrics

• 5-Fold Cross-Validation Accuracy: 0.643 ± 0.075
• Held-out Test Accuracy: 0.588
• Held-out Test Precision: 0.588
• Held-out Test Recall: 0.588
• Held-out Test F1: 0.588
• Held-out ROC AUC: 0.6955

Automatic Download & Usage

This model is automatically downloaded and cached by the TRUSIGHT voice service upon first run:

uvicorn app.main:app --reload

To load manually with Python:

import joblib
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="sofi233/truesight-voice-deception",
    filename="deception_model.pkl",
    revision="v1.0.0"
)
model = joblib.load(model_path)

Intended Use & Ethical Disclaimers

This model estimates deception likelihood from vocal acoustic cues only as part of an experimental multimodal truth verification system. It does NOT make definitive lie-detection determinations on its own. Indian-English and non-native accents have not been validated for deception detection.