SIH26104 · Team Xenthron

Know if the voice on the
call is real — before you act.

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.

How it works Explore features
0
LOW
Waiting for audio…

🛡️ Voice appears genuine

No signs of voice cloning. Stay alert to unusual requests.

1.5s windowsLive WebSocketAttention explainabilityNo audio stored
Illustrative playback of the Live Check UI — cycling through LOW → MEDIUM → HIGH.
How it works

Four steps, running continuously during the call

No recording, no upload-then-wait. Each window is analysed and scored as it arrives.

01

Listen

Mic audio streams over WebSocket in ~150ms frames, re-windowed server-side into 1.5s / 0.5s-overlap chunks.

02

Analyse

Each window is resampled, VAD-trimmed, and turned into a 64-bin log-mel spectrogram.

03

Score

A CNN-LSTM estimates spoof probability; exponential smoothing turns it into a stable 0–100 score.

04

Alert

LOW / MEDIUM / HIGH pushes back instantly, with the reasoning behind the number.

Why VoiceGuard

Built for the moment it actually matters

Real-time, explainable, and designed so a false accusation is as costly as a missed one.

Real-time streaming

Not record-then-analyse — every window is scored the moment it arrives, with results back on-screen in ~1–2s.

Explainable, not a black box

Attention-weighted time segments and frequency bands show exactly what drove the score.

Privacy-preserving by default

Raw audio is never persisted. Only scores, timings and explanation metadata leave the pipeline.

Graded, actionable alerts

LOW is informational, MEDIUM asks for a second channel of verification, HIGH says stop and call back.

Language-agnostic

Mel-spectrogram features, no ASR or lexicon — built to extend across English, Tamil and Hindi.

On-device ready

The model exports to ONNX and is verified bit-for-bit against the PyTorch original — ready to run offline.

The graded alert

A score is only useful if it tells you what to do

LOW · Informational

No signs of voice cloning detected. Stay alert to unusual requests, nothing more to do.

MEDIUM · Verify

Possible synthetic voice. Confirm the caller through another channel before acting on anything.

HIGH · Stop

High risk of a cloned voice. Hang up and call back on a number you already trust.

Under the hood

One pipeline, live socket and offline clip alike

The same four steps run whether audio arrives over a live WebSocket or an uploaded file.

Mic / CallPCM16 @16kHz
PreprocessVAD · log-mel
CNN-LSTMSpoof probability
Risk EngineScore + band
Tech stack

Fixed, boring choices where it counts

FastAPI · async backend
PyTorch · CNN-LSTM
ONNX Runtime · edge inference
PostgreSQL · durable storage
WebSockets · live streaming
Flutter · Android + iOS
librosa / torchaudio · audio DSP
nginx + pm2 · production
1.5sanalysis window
<2sperceived latency
0raw audio stored
3languages targeted
VoiceGuard produces a probabilistic risk assessment, not a forensic determination. Scores are advisory and should be corroborated before any enforcement or financial action.

Built for the call that can't wait.

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