Official Google Meet Add-on · Google Workspace Marketplace ↗
Catch cloned voices on the calls that move money.
Real-time deepfake-voice detection built for compressed meeting audio. Every speaker receives an immediate REAL / SUSPECT / FAKE verdict, and an automated wire hold fires before fraudulent transfers clear.
1 hr free trial
•
+4 bonus hours with review
•
No credit card required
meet.google.com/wire-approval-q3
LIVE VERDICT
SJ
Sarah Jenkins
Chief Financial Officer
!
Unknown Caller
Phone Dial-in · +1 (415) •••-••92
97%
Sonave Catch Rate
Catches 27 out of 27 commercial voice-clone tools through live meeting codecs.
● Benchmark: 27 clone families
2%
Commodity Detectors
Generic detectors collapse from 99% down to 2% once audio passes through meeting compression.
● 98% collapse on Opus codec
94%
Real-Voice Accuracy
Authentic voices confirmed without false alarms, preventing disruption to legitimate calls.
● Production model: sonave_xlsr_meet
<1s
Time-to-Verdict
From spoken word to on-screen verdict, powered by ~98ms GPU inference per window.
● 4.0s rolling analysis
SEE IT IN ACTION
The moment the wire gets held.
Inspect high-definition captures of in-meeting speaker verification, synthetic voice alerts, and the executive forensic console.
https://meet.google.com/wire-approval-q3 · Sonave Active
Enlarge
ZOOMED IN-CALL INSPECTION
Live Side Panel
Sarah Jenkins (CFO)
REAL · 99.2%
Acoustic envelope and harmonic phase match verified human vocal tract.
Unknown Caller (+1 415...)
FAKE · 94.0%
High-frequency neural artifacts detected. Synthetic clone signature confirmed.
Enforcement Webhook:
WIRE_HOLD_ARMED
HOW IT WORKS
Every voice is scored live. Suspicion holds the wire.
Zero installs for meeting participants. Sonave connects at the audio ingestion layer to evaluate voices in real time.
01 · INGESTION
Capture
Sonave ingests live meeting audio directly in Google Meet over encrypted WebRTC/PCM. Zero browser extensions, zero participant installs, and zero attendee friction.
● 16kHz PCM · TLS 1.3 Encrypted
02 · INFERENCE
Detect
An XLS-R neural detector evaluates 4-second rolling windows entirely in volatile GPU RAM. Specifically trained on compressed Opus meeting audio. Tensors are never written to disk or permanent storage.
● 98ms TensorRT · Zero Disk Writes
03 · ENFORCEMENT
Act
Live verdicts trigger real-time on-screen badges, automated alerts, and an instant webhook that pauses wire approvals in your treasury ERP before money leaves the bank. The audio buffer is immediately flushed.
● Instant ERP Hold · Immutable Audit Log
THE REAL-WORLD DIFFERENCE
Lab accuracy is not meeting accuracy.
Even deepfake incumbents now admit: "Post-call detection comes too late." If funds already cleared, a post-call scan is just an autopsy. Here is why Sonave stops fraud in the room.
Capability / Vector
Commodity Detectors
Sonave Authenticity Engine
Real-World Outcome
Meeting Codec Audio
Google Meet & Zoom (Opus)
2% catch rate
Collapses under Opus compression; trained only on clean studio WAVs.
97% catch rate
Trained directly on codec-compressed meeting audio and packet jitter.
Fraud caught on call
Detection Timing
Continuous vs post-call
Post-Mortem / Delayed
Post-call file uploads or carrier APIs reporting fraud after funds already cleared.
<4s Live Verdicts
Continuous 4.0s rolling windows while participants are still talking.
Before wire approval
Action & Enforcement
What happens on red alert
Passive Dashboard
Displays a score on a dashboard; zero integration with treasury or ERP.
Automated Wire Hold
Instant webhook pauses treasury approvals + exportable IC3 forensic report.
Funds frozen in ERP
Audio Privacy
Memory vs persistent disk
Permanent Server Storage
Audio recorded, uploaded, and stored indefinitely on vendor servers.
Zero-Retention Volatile RAM
Audio scored in GPU memory and purged immediately after verdict generation.
Zero leak liability
We publish our numbers — including the hard ones. See the full benchmark dataset →
CRITICAL INCIDENT DEFENSE
When a deepfaked CEO asks for $2M, the verdict is on screen before the approval.
Every scored window lands in an immutable forensic log — score, verdict, timestamp, audio snippet — with a one-click exportable report per meeting. Voiceprint enrollment adds a second factor: confirming whether the voice truly belongs to your authorized executive.
Automated webhook triggers freeze in treasury & ERP systems
One-click IC3 and insurance legal certificate of authenticity
Voiceprint enrollment verifies biometric identity and liveness
!
Unknown caller
Joined by phone · +1 (415) •••-••92
SYNTHETIC RISK SCORE
94%
Wire hold triggered — re-authentication required
TRANSPARENT PRICING
Start free. Pay for what you monitor.
No seat licenses, no enterprise dark patterns. Monitor your calls with confidence.
FREE TIER
Free
1 hour included free (+ 4 bonus hours unlocked with a Marketplace review). Sign in with Google, monitor meetings in real time, and receive live verdicts. No credit card required.
- 1 hour free trial (extendable to 5h with review)
- Google Meet Add-on & Live Media
- Live REAL / SUSPECT / FAKE verdicts
- 4-second continuous sliding window
- Zero browser extension installs
RECOMMENDED FOR TEAMS
$8
/ monitored hour
Billed by the minute (~$0.13/min). Pay only while actively monitoring a call. Zero monthly seat licenses.
- Everything in Free, no monthly hour limits
- Instant wire-hold treasury webhooks
- Real-time SMS & Slack red alerts
- Exportable PDF forensic certificates
ENTERPRISE
Custom
Dedicated tenant, custom ERP / Treasury wire-hold webhooks, executive voiceprint enrollment, and 99.9% uptime SLA.
- High-volume discounted rates
- Dedicated GPU tenant & VPC peering
- SOC 2 compliance vault
- Dedicated incident engineer support
QUESTIONS
Straight answers.
Everything you need to know about how Sonave intercepts synthetic voice attacks.
Can a deepfake voice be detected on a live Google Meet, Zoom or Teams call?
Yes. Sonave connects directly via the Google Meet Add-on and Live Media stream to inspect each speaker's audio with neural detection models trained specifically on meeting-codec acoustics. Every speaker gets a REAL / SUSPECT / FAKE verdict about 4 seconds after they start talking, re-scored every 4 seconds for the rest of the call.
How accurate is deepfake voice detection, really?
On 27 unseen commercial voice-clone tools played through meeting audio, Sonave catches 97% where a commodity detector catches 2%. On the hardest public real-world deepfakes (In-the-Wild), it catches about 57% at a 94%+ real-voice accuracy — which is why verdicts are designed as a second factor next to callbacks, not a replacement. Full methodology at usesonave.com/benchmarks.
Does everyone on the call need to install something?
No. Only the meeting participant protecting the call opens the Sonave Google Meet Add-on from the side panel. Nobody else installs or configures anything on their laptop or phone.
What is a wire hold?
When a voice scores in the red band for three consecutive windows, Sonave fires an automated webhook into your payment-approval flow to pause a wire transfer pending secondary out-of-band re-authentication. It also produces an exportable forensic report containing scores, timestamps, and audio evidence.
What happens to my meeting audio?
Audio is processed in volatile memory to produce verdicts. Voiceprints are stored solely as mathematical vector embeddings that you can delete in one click. We never sell data or share meeting audio between customers — full details at usesonave.com/privacy.
How much does Sonave cost?
1 hour of call monitoring included free upon install. Users can unlock an additional 4 hours free by leaving a quick review on the Google Workspace Marketplace. After that it's $8 per monitored hour, billed by the minute, only while actively monitoring a call. No seats, no tiers.
Sonave Screenshot Preview