Real-time AI scam-call shield
Verity fuses deepfake detection, social-engineering analysis, and speaker verification into a single trust score -- in under two seconds.
<2s
Detection latency
5
ML models
0
Models trained
8
Languages
The Problem
A 30-second voice sample from a WhatsApp note or social media video is enough to clone anyone's voice. Scammers use these clones to impersonate family members, demanding money while the victim hears their loved one's voice begging for help.
Typical scam call transcript
"Mom, please, I've been arrested and I need ten thousand rupees right now. Don't tell dad. The police won't let me go until I pay this. Please hurry."
Architecture
Each layer runs independently. The Trust Engine fuses all outputs into a single 0-100 score. If any signal fires, the score drops. Defense in depth -- one layer failing doesn't matter.
Wav2Vec2 + AASIST3 neural networks analyze raw audio to detect synthetic speech, voice cloning, and TTS artifacts. Gemini 2.5 provides noise-robust analysis.
Gemini 2.5 Flash examines the live transcript for social engineering -- urgency, financial requests, authority threats, impersonation, and secrecy pressure.
ECAPA-TDNN compares the caller's voiceprint against your Family Voice Vault. If they claim to be your son, the voice must match.
Live Output
The Trust Score starts at 100. Each red flag subtracts points: synthetic voice detected, financial request identified, voiceprint mismatch confirmed. You see exactly why a call is flagged.
100
Safe
Real Scenarios
Four real-world scenarios showing how the three detection layers work together. Try them yourself in the dashboard's Demo Scenarios panel.
A scammer clones your grandson's voice from a WhatsApp note and calls demanding bail money.
Detection
Deepfake detector flags synthetic voice at 90%+. Scam classifier catches urgency and financial demand. Voice vault shows no match to enrolled grandson.
Automated call threatens arrest for unpaid taxes, demands payment via gift cards.
Detection
LLM detects authority threats (95%) and financial requests (90%). Trigger phrases flagged: 'warrant', 'arrest', 'gift cards'.
Fake bank rep claims your account is compromised, asks you to transfer money to a 'safe account'.
Detection
Financial request (92%) and impersonation (80%) flagged. Secrecy pressure detected: 'don't share this information'.
Your actual son calls to chat about weekend plans.
Detection
All scam signals at 0%. Deepfake probability 1%. Voice vault match: 95%. Trust Score: 98.
Platform
Not just detection -- a complete protection platform with forensic reporting, education, threat intelligence, and family safety tools.
Stream audio from your microphone for real-time call monitoring
Upload WAV, MP3, M4A, OGG files for instant analysis
Enroll family voiceprints. Verify callers against trusted contacts
Full-screen emergency warning with action buttons when scam detected
One-click export of HTML reports and raw JSON analysis data
Gemini explains why a voice sounds synthetic or a call seems suspicious
4 pre-built scenarios to demo the system without a backend
English, Hindi, Spanish, Tamil, French, German, Japanese, Chinese
No audio leaves your device. Voice Vault stays local. Zero cloud storage
Community feed showing active scam campaigns and trending attack types
Interactive library of scam playbooks with a quiz to test your knowledge
Quick-call buttons for trusted family members when a scam is detected
Voice Vault
Aarav
Son
Priya
Daughter
Mom
Mother
Protection
Enroll a 30-second voice sample of each family member. When someone calls claiming to be your son, Verity checks the voiceprint against the vault -- exposing clones that sound identical to the human ear.
The vault uses ECAPA-TDNN to create 192-dimensional voiceprint embeddings stored locally on your device. No biometric data ever leaves your phone.
Open Voice VaultLandscape
Scam-call protection exists -- for telecom networks, banks, and browsers. Nobody protects the consumer from AI voice clones. Until now.
| Solution | Protects | Gap |
|---|---|---|
| Hiya | Telecom networks | B2B only, no deepfake detection |
| Pindrop | Banks & call centers | Enterprise pricing, no consumer app |
| McAfee | Browser scam alerts | No real-time voice analysis |
| Verity | Your family's phone | All three signals, free, on-device |
Under the Hood
Every ML component is a pretrained downloadable artifact. We built the fusion and the product, not the models.
Audio Forensics
Gemini 2.5
Noise-robust deepfake + scam + transcript in 1 call
Deepfake Detection
Wav2Vec2
XLS-R 300M fine-tuned on ElevenLabs
Anti-Spoofing
AASIST3
KAN + GAT for voice conversion attacks
Speaker Verify
ECAPA-TDNN
192-dim voiceprint embeddings
Transcription
Whisper
distil-large-v3, 8 languages
FAQ
We use three layers: a Wav2Vec2 model trained specifically on ElevenLabs and modern TTS, AASIST3 for older voice conversion attacks, and Gemini 2.5 Flash for noise-robust audio forensics. If any layer detects synthesis, the trust score drops.
Yes. Even when the voice is real, our LLM analyzes the transcript for social engineering patterns -- urgency, financial demands, authority threats, secrecy pressure. Real human scam calls are caught through content analysis.
You enroll 30-second voice samples of trusted contacts (family members). When someone calls claiming to be your son, Verity's ECAPA-TDNN model compares the caller's voiceprint against the enrolled sample. A clone may sound perfect to your ear, but the voiceprint won't match.
No. Audio analysis happens on-device or via encrypted API calls. No audio is stored on our servers. The Voice Vault is local to your device.
Transcription supports English, Hindi, Spanish, Tamil, French, German, Japanese, and Chinese. The scam classifier works across all languages since it analyzes translated content.
Under 2 seconds for real-time streaming. Uploaded files take 10-30 seconds depending on length. The Gemini combined analyzer does deepfake detection, transcription, and scam classification in a single API call.
AI voice scams stole over a billion dollars from families globally in 2025. The voice is real. The crime is invisible. Verity makes it visible.