Harsh33t/Sunniva-kavach

Kavach (कवच) — Real-Time Voice Impersonation Defense & Telephony Forensic Weights

Kavach is an open-source voice authenticity and anti-spoofing defense engine engineered specifically for real-time telephony, WebRTC live audio streams, and financial vishing prevention.

Model Summary

• Architecture: Frozen Wav2Vec2 (facebook/wav2vec2-xls-r-300m, layer 14) + AASIST Graph Neural Network backend
• Sampling Rate: 16 kHz mono Float32 PCM
• Telephony Robustness: Trained with simulated acoustic impairments, handset frequency response, AGC, and noise suppression filters (modern_phone series)
• Calibration: Multi-signal isotonic & logistic fusion (chunk_clean / chunk_fused_telephony) combining acoustic artifact detection with prosody dynamics

Checkpoints Included

• modern_phone/best.pt: Top-performing telephony checkpoint (Epoch 23, Validation EER 0.48% / 99.5% accuracy)
• modern_phone/last.pt: Final training checkpoint with optimization history
• codec_augmented/best.pt: Baseline wideband codec-augmented checkpoint
• calibrators/chunk_clean.joblib: Clean in-domain calibrated logistic regressor
• calibrators/chunk_fused_telephony.joblib: Multi-signal fusion calibrator
• signals/prosody.joblib: Secondary pitch/energy forensic scorer

Quick Setup for Team Members

Clone the repository and pull the models in one command:

git clone https://github.com/xOAviOx/kavach.git
cd kavach
uv sync
uv run python scripts/download_from_hf.py

Run serving:

uv run uvicorn kavach.serving.app:app --host 0.0.0.0 --port 8100