How VaaniX Detects AI Voice Clones in 10 Seconds
VaaniX runs a dual-engine architecture combining deep acoustic feature verification with semantic scam transcript analysis. All calculations are completely transparent — no black boxes.
1. The 10-Second Listening Window
Listening to full phone calls is intrusive, unnecessary, and raises severe privacy concerns. VaaniX activates strictly during seconds 5 to 15 of a connected call.
Ringing & Greetings
Initial phone connection, ringback tones, and routine hellos are intentionally skipped to avoid false alarms.
10-Second Capture
Captures a single 10-second monophonic 16kHz audio sample — the exact duration where callers state their identity and intent.
Immediate Evaluation
Recording immediately ceases. Audio is streamed into parallel acoustic and speech-to-text pipelines.
2. The Dual-Engine Pipeline
A scam cannot be judged by acoustic timbre alone, nor by vocabulary alone. Both pipelines run concurrently to prevent false positives.
Engine A: Acoustic Deepfake Model (Nes2Net-X)
Nes2Net-X evaluates subtle vocal artifacts that humans cannot perceive: spectral anomalies, glottal pulse discontinuities, and vocoder phase jitter typical of voice cloning software.
Engine B: Transcript & Intent Extractor
Using faster-whisper on CPU with VAD (Voice Activity Detection), speech is transcribed and normalized across Hindi script, Hinglish, and English to detect scam keywords and impersonation claims.
3. The Transparent Score Formula
Every verdict on VaaniX comes from an explicit mathematical formula, never an opaque proprietary black box.
Verdict Decision Rules:
- AI_CLONE:Triggered when
risk_score ≥ 60ANDvoice_score ≥ 50. (The voice model must agree; scam words alone cannot trigger a deepfake alert). - REAL:Triggered when
risk_score < 40. - INCONCLUSIVE:Scores between 40 and 59. Displays a caution badge; does not fire automated emergency Telegram messages.
Scam Keyword Weights:
Matched group weights are summed once and capped at 100. If no speech is heard, text score is 0.
4. Model Accuracy & Benchmark Integrity
95% accuracy in our internal tests (test set details coming soon)
We uphold strict academic integrity. While our internal evaluations with synthetic voice clones demonstrated a 95% detection accuracy, independent third-party evaluation datasets and comprehensive ROC curves will be published openly upon peer review. Processing duration averages ~10 seconds per check.
5. Privacy & Zero-Retention Architecture
We built VaaniX with the fundamental belief that privacy and fraud security must coexist.
No Audio Stored on Clean Calls
If the call is verified as natural or inconclusive, the audio buffer is wiped from memory instantly.
SHA-256 Forensic Evidence
Only calls flagged as AI_CLONE are securely retained (30 days maximum) with an immutable cryptographic digest for police complaints.
No Radar Tracking
Scam Radar stores aggregated caller phone reputation only. No victim or reporter identity is ever stored or displayed.
Your voice is your identity.
VaaniX protects it.
Real-time acoustic AI voice-clone verification and family impersonation shield.