Live in-browser measurements evaluating acoustic room physics, biometric verification, and synthetic spoof liveness with active automated mitigations.
| Subsystem Metric |
Live Measurement |
Practical Relevance |
Active In-Browser Mitigation & Safeguard |
| Signal-to-Interference (SINR) |
-- dB |
Speech volume vs. wall bleed & room noise. |
Soft-knee tanh AGC + Dynamic 1.5s Auto-Calibration |
| Reverberation Decay (T60) |
-- ms |
Echo duration before decaying 60dB. |
Dynamic Hold Expansion (256ms → 450ms) |
| Biometric Similarity (Sim) |
-- |
Cosine distance to enrolled vocal centroid. |
L2-Norm Normalization + 5% Rolling EMA Centroid |
| Bispectral Liveness (ΔΦ) |
-- |
Robotic phase linearity vs. human vocal cords. |
AASIST 32k Padding + Forensic Tagging Router |
| Pitch Elevation (ΔF0) |
-- Hz |
User vocal strain under background noise. |
3-Slice Pitch-Invariant Enrollment Wizard |
| Spectral Crest Factor (PAPR) |
-- dB |
Transients (clicks/coughs) vs. sustained speech. |
Transient Gate (<180ms rejection) |
| Attribution Purity (API) |
98.5% |
Stream routing sorting accuracy. |
3-Frame Majority-Vote Hysteresis Smoothing |
| Inference Latency |
-- ms |
Time-to-first-token sequence decoding. |
WebGPU Pipeline + Microtask Yielding Queue |