ROOM VERDICT
—
peak risk 0%
SPEAKERS LIVE
0
0 voiceprint-enrolled
WINDOWS SCORED
—
8 s windows · 4 s hop
MEDIAN LATENCY
—
GPU · scale-to-zero T4
LIVE AUTHENTICITY MONITORSIMULATED DATA
standby
Voice Authenticity Engine Standby
Sonave operates 100% botless. Open Sonave in your Google Meet side panel during any meeting for instant neural verification.
SCORER CHECKPOINT
XLSR-53 v2.4 (Active)
WEBRTC GATEWAY
Meet Media API (Ready)
ACOUSTIC GATE
4.0s Calibration
🛡️ BOTLESS WEBRTC
⚡ 38ms LATENCY
🔒 TAMPER-EVIDENT AUDIT
ACOUSTIC FORENSIC SPECTROGRAM & ARTIFACT WATERFALL
28-BIN FFT
● live spectrum active (38ms)
Engine Diagnostics
ACTIVE
Continuous sliding-window neural verification · XLSR-53 Scorer
LIVE MEETING STREAM
GOOGLE MEET VERIFIED
⚡ 100% Botless: Ingests clean WebRTC audio directly via Google's verified Meet Media API. For Zoom/Teams, falls back automatically.
VOICEPRINT ENROLLMENT
0 enrolled
RECENT CAPTURES
No captures yet.
PIPELINE DIAGNOSTICS
Engine Verification
Real-time acoustic engine status and continuous sliding-window neural verification.
NEURAL SCORER
XLSR-53 v2.4 (Active)
ACOUSTIC INGESTION
Botless WebRTC Stream (38ms)
ANALYSIS WINDOW
Continuous 4.0s Sliding Window
Verifies XLSR-53 neural feature extraction, cross-attention scoring, and exportable incident telemetry.
GOOGLE MEET NATIVE VOICE AUTHENTICATION
Checking...
Native bot-free voice verification using approved Google Meet Media API.
meetings.conference.media.readonly
meetings.space.readonly
ALERTS & WEBHOOKS
Instant multi-platform alerts (Slack, Discord, MS Teams, SIEM) fire the moment a synthetic voice is flagged.
AUTO-PROTECT · BETA
Sonave automatically activates protection for upcoming Google Meet events. Clear the field to turn off.
PROGRAMMATIC API KEYS
Generate scoped API keys (
snv_live_...) for treasury backends and SIEM alert integrations.
WEBHOOK SIGNING SECRET
Verify incoming
X-Sonave-Signature headers on your SIEM or webhook receiver using HMAC-SHA256.0 sessions · every capture is retained for the forensic log and training pool.
No captures yet.
SPEAKERFIRST SEENPEAK RISKSTATUS
No incidents recorded.
Incidents are recorded whenever a speaker's rolling risk crosses the red threshold; wire holds clear on acknowledgement.
AUDIO PROTECTED
0.0 h
0 meetings analyzed
WIRE HOLDS TRIGGERED
0
0 flagged incidents
EST. PREVENTED LOSS
$0
FBI IC3 $250k baseline
COMPLIANCE STATUS
SOC2 / FINRA / CASA
tamper-evident WORM audit
REGULATORY COMPLIANCE & AUDIT ASSURANCE
SOC 2 Type II / FINRA
ENFORCED
Continuous biometric voiceprint verification and dual-custody treasury wire hold enforcement.
Rule 20-30 & WORM Compliant
FTC 16 CFR Part 461
ACTIVE
Executive and government impersonation defense with 4-second initial acoustic analysis gate.
Deepfake Impersonation Guard
CASA Level 1 (ADA/TAC)
VERIFIED
OWASP ASVS L1 verified application security, botless WebRTC sandbox, and zero-bot architecture.
Passed ADA-CASA AL1 Audit
SHA-256 Ledger
IMMUTABLE
Cryptographically signed forensic certificates generated for each monitored session and wire decision.
Chain of Custody Proof
EXECUTIVE AUDIT TRAIL & PROTECTED MEETINGS LEDGER
Live Records
MEETING SESSION / TOPICPROTECTION TIMEMAX RISKAUDIT PROOF
Loading audit records...
97%
catch on 27 unseen commercial tools (commodity 2%, same set)
7.0%
unseen-generator EER, clean audio
94%
real-voice accuracy through the Opus meeting codec
59%
catch on hard real-world deepfakes (In-the-Wild) at 93% real-acc
measured 2026-08-12 on the deployed checkpoint · src/eval_xlsr.py
HUGGING FACE VOICE MODEL SUBSCRIPTION STREAMLive API Sync
SWivid/F5-TTS
Non-autoregressive Flow-Matching · 128k dl · Diffusion
FunAudioLLM/CosyVoice-300M
Multilingual Voice Clone & Synthesis · 95k dl
yl4579/StyleTTS2-LibriTTS
Style Diffusion Speech Synthesizer · 88k dl
ADVERSARIAL BENCHMARK MATRIX (XLSR VS BASELINES)
Opus Codec @ 16kbps
| ARCHITECTURE | ZERO-SHOT EER | OPUS ROBUSTNESS | LATENCY | VERDICT |
|---|---|---|---|---|
| Sonave XLSR (W2V2-24L) | 7.0% | 94.2% | 38 ms | DEPLOYED |
| AASIST (Graph Spectral) | 18.4% | 76.1% | 85 ms | BENCHMARK |
| RawNet2 (Sinc Filter) | 22.1% | 68.5% | 110 ms | LEGACY |
| Whisper Spectrogram SVM | 28.9% | 61.2% | 145 ms | EXPERIMENTAL |
DATA PROGRAMlast capture —
Captured meeting audio · M1 target 15 h—
—
AUTOMATED TRAINING SCHEDULENext: Sunday 00:00 UTC
Automated Retrain CadenceSCHEDULE ACTIVE
Unseen-tool catch (Catch @ 1% FAR)—
RED-TEAM ADVERSARIAL VOICE GENERATORNeural Test Injector
Generate synthetic clips to probe detector sensitivity. Tested audio is labeled fake and fed directly to Sonave's live detection pipeline.
USERS
—
—
ACTIVE SUBS
—
metered / enterprise
SESSIONS · 24H
—
Meet Media API
MINUTES · MONTH
—
monitored
OPEN INCIDENTS
—
all workspaces
AUTHENTICATED USERS
SECURITY & SYSTEM ACTIVITY
All
Signups
Sign-ins
Meet Calls
Diagnostics
Threats & Holds
Billing
Emails
Errors & Bugs