VoiceGuard listens to a live call and analyses it in 1.5-second windows, streaming back an explainable 0–100 risk score before the caller finishes asking you to move money, reset a password, or trust an instruction.
No signs of voice cloning. Stay alert to unusual requests.
No recording, no upload-then-wait. Each window is analysed and scored as it arrives.
Mic audio streams over WebSocket in ~150ms frames, re-windowed server-side into 1.5s / 0.5s-overlap chunks.
Each window is resampled, VAD-trimmed, and turned into a 64-bin log-mel spectrogram.
A CNN-LSTM estimates spoof probability; exponential smoothing turns it into a stable 0–100 score.
LOW / MEDIUM / HIGH pushes back instantly, with the reasoning behind the number.
Real-time, explainable, and designed so a false accusation is as costly as a missed one.
Not record-then-analyse — every window is scored the moment it arrives, with results back on-screen in ~1–2s.
Attention-weighted time segments and frequency bands show exactly what drove the score.
Raw audio is never persisted. Only scores, timings and explanation metadata leave the pipeline.
LOW is informational, MEDIUM asks for a second channel of verification, HIGH says stop and call back.
Mel-spectrogram features, no ASR or lexicon — built to extend across English, Tamil and Hindi.
The model exports to ONNX and is verified bit-for-bit against the PyTorch original — ready to run offline.
No signs of voice cloning detected. Stay alert to unusual requests, nothing more to do.
Possible synthetic voice. Confirm the caller through another channel before acting on anything.
High risk of a cloned voice. Hang up and call back on a number you already trust.
The same four steps run whether audio arrives over a live WebSocket or an uploaded file.
A real-time, explainable, privacy-preserving defence against voice-cloning fraud — from a live socket down to the risk engine.
Back to the live demo