SYSTEM ARCHITECTURE & DETECTION PIPELINE
V-SHIELD is a real-time, privacy-preserving audio verification engine designed for Indian telephony and VoIP ecosystems. Built to decide in under 269ms, survive 8 kHz G.711 narrowband compression, and operate with zero raw-audio persistence under India's DPDP Act 2023.
FIVE COUPLED SUBSYSTEMS
INGESTION & PREPROCESSING
WebRTC and SIP/RTP media streams ingested into single-channel 16-bit linear PCM at 16 kHz. Silero Voice Activity Detection (VAD) discards background silence, comfort noise, and hold music so only active speech enters the classifier.
MULTI-DOMAIN DSP EXTRACTION
Extracts dual-domain acoustic markers: Linear Frequency Cepstral Coefficients (LFCC) across 0–8 kHz preserving high-frequency vocoder residuals, Bispectral matrix B(f1, f2) measuring quadratic phase non-linearities, and F0 pitch jitter with cycle-to-cycle variance.
QUANTIZED NEURAL CLASSIFIER
Dual-stream neural backbone: Modified RawNet2 with parameterized sinc-convolutions for raw waveform inspection combined with a lightweight 2D ResNet with Squeeze-and-Excitation for spectral envelopes. Quantized to INT8 ONNX for sub-40ms CPU inference.
KALMAN SMOOTHING & TRI-STATE
Rolling 1-D Kalman filter over 1.5-second observation window suppresses isolated false-positive spikes from cellular packet loss or talker overlap. Emits a deterministic tri-state policy verdict: GREEN (<0.35), AMBER (0.35–0.75), or RED (≥0.75).
ACTIVE MITIGATION & CHALLENGE
When score enters the Amber band in a sensitive transaction, the engine injects an unpredictable multilingual phonemic verification prompt (EN/HI/TA/TE). Tests caller response latency: humans answer under 800ms while neural vocoders require 800–2500ms pipeline latency.
FAIL-CLOSED FOR PAYMENTS · FAIL-OPEN FOR SOS
Integrators bind policy actions (Allow, Challenge, Terminate, Escalate) to transaction value and caller roles. High-value RTGS/NEFT transfers are fail-closed; emergency lines remain open with operator alerts.
269.0 MS TOTAL DECISION ROUND-TRIP
Sum of ingestion window, DSP feature compute, INT8 neural inference, Kalman logic, and network dispatch remains strictly inside a natural conversational pause.
ZERO RAW-AUDIO PERSISTENCE BY DESIGN
- 1.Volatile RAM Only: Audio frames reside strictly in circular volatile memory queues, overwritten immediately after feature extraction.
- 2.Non-Reconstructible Vectors: Extracted LFCC and bispectral features cannot be inverse-synthesized into intelligible human speech.
- 3.Append-Only Audit: Telemetry logs store session ID, C(t) risk trace, codec tag, and timestamps only — never waveform bytes.
ROBUST AGAINST INDIAN NARROWBAND PATHS
- 1.G.711 / AMR Adaptation: Trained and validated on band-limited 300–3400 Hz telephony audio with comfort-noise insertion.
- 2.Target EER ≤ 5.4%: Maintained on held-out telephony-transcoded evaluation split, avoiding optimistic broadband-only lab results.
- 3.Accent Resilience: Trained on Indian English, Hindi, Tamil, and Telugu to prevent regional phonetics from triggering false challenges.