The Science of Detecting Synthetic Voice
Generative AI and voice cloners leave micro-acoustic artifacts invisible to the human ear. VoiceShield isolates these digital fingerprints in milliseconds.
Encrypted Stream Ingestion
Audio chunks (16-bit PCM, 16kHz or 8kHz narrowband) stream over authenticated WSS with client verification.
WebSocket RFC 6455 • TLS 1.3 • Ring buffer memory management
DSP Signal Preprocessing & VAD
Voice Activity Detection (VAD) separates conversational speech from ambient silence, line noise, and DTMF tones.
Zero-crossing density • Energy thresholding • 100ms analysis frames
Spectral & Biometric Extraction
Spectrogram analysis reveals phase discontinuities, robotic harmonic pitch artifacts, and absence of natural human jitter.
FFT Multi-band flux • Pitch period tracking • Glottal cycle consistency
Ensemble Inference & Kalman Smoothing
Scores from local acoustic models feed into a Kalman filter that prevents false alerts caused by fleeting network glitches.
State-space estimation • Dynamic covariance • Calibrated synthetic risk [0.0 - 1.0]
Forensic Chain-of-Custody & Alerting
Incidents generate structured cryptographic evidence packages including Section 65B certificates for formal dispute filing.
SHA-256 fingerprinting • Append-only audit logs • Twilio SMS & Webhook dispatch