New: catching cloned voices inside Teams calls
Synthetic media can resemble anyone. We help you spot it.
Resemble is an independent guide to deepfake detection and voice AI — how cloned audio, video and images get made, how real-time models catch them, and how to build with voice responsibly.
Detection that works wherever fakes show up.
Warn agents mid-call through carrier paths.
Get startedFlag cloned voices and faces in live meetings.
Start nowTriage reports without exposing anyone pictured.
Get startedHand cases a scored, explainable verdict.
Start nowConfirm new applicants are who they claim.
Get startedCatch edited invoices and IDs before payment.
Start nowScreen uploads before fakes reach the feed.
Get startedSpot spoofed leaders in calls and ads.
Create yours
Ship detection with one short request.
pip install requests
# score any clip in seconds import requests r = requests.post(API_URL, headers=AUTH, files={"file": open("clip.wav", "rb")}) print(r.json()["verdict"])
Detection plugs into stacks like these
Detection models
Generators keep getting more lifelike. Detection has to keep pace.
Resemble follows in-house detectors for audio, image, video and text, and explains what each does well.
DETECT-World
Resemble AI's newest detector for audio, video and images, using a world model that notices broken physics.
OUTPUT: Score, verdict and a readable summary
ACCURACY: Audio 99.5% | Video 98.2% | Image 95.8%
Three detector generations
AUDIO
IMAGE
VIDEO
Image and video arrived in DETECT-3B Omni. Audio scores: Podonos; image and video: internal tests.
Intelligence
Readable explanations attached to every detection score.
OUTPUT: Auditable, structured findings
Signal
Alerts users when audio matches a known fraud pattern, without recording.
OUTPUT: Compliant alerts and notices
PerTh Multimodal
Watermarks no one can see or hear, across media types.
OUTPUT: EU AI Act-ready files plus a verdict
Identity
Enroll a voice or likeness, then match new media to it.
OUTPUT: Match score with a clear verdict
Benchmarks
Independent scores, not self-grading
Public leaderboards are where claims get tested. Here is how Resemble AI's detectors rank, with source and setup noted.
See the table99.5%
Audio accuracy
95.8%
Image accuracy
98.2%
Video accuracy
Latest result: audio deepfake detection
Accuracy against real-time factor. Top-left is best.
Resemble AI
Commercial
Research baseline
Source: Podonos audio deepfake benchmark, August 2026.
| # | System | RTF (s) | Accuracy |
|---|---|---|---|
| 1 | Resemble DETECT-World | 0.12 | 99.47 |
| 2 | Resemble AI | 0.40 | 98.05 |
| 3 | Whispeak | 0.39 | 97.70 |
| 4 | Aurigin AI | 0.33 | 96.75 |
| 5 | Pella Research | 0.021 | 95.82 |
| 6 | Pindrop | 0.076 | 95.05 |
| 7 | Corsound AI | 0.035 | 87.79 |
| 8 | Hive | 0.34 | 83.53 |
| 9 | Reality Defender | 1.52 | 71.27 |
| 10 | Wav2Vec2 (2019 LA) | 0.14 | 62.89 |
| 11 | AST (ASVspoof 5) | 0.0017 | 56.83 |
| 12 | Wav2Vec2 (2024 mix) | 0.056 | 55.55 |
Top 12 of 18 systems shown
Illustrative team personas
Resemble finally made a detection score mean something to our call floor.
Fraud operations lead
We cared about speed. The model notes saved us weeks of vendor calls.
Head of trust engineering
The Resemble incident log made the risk real for our board in one sitting.
Security chief
Plain watermarking notes helped us plan for EU labelling rules early.
Compliance manager
Resemble's benchmark summary is the tab our analysts never close.
Forensic audio analyst
Our moderators now know which uploads to escalate and which to clear.
Safety operations lead
Security checklist
What to demand from any detection vendor
Resemble's checklist for teams where security can't bend — from air-gapped networks to cross-border data rules.
Personal data handled only on a lawful basis.
A certified information security management system.
Live detection within a day of signing.
One REST API per modality, ready for SIEM.
Audited controls for availability and confidentiality.
Health data handled in line with HIPAA.
Containerized on-prem install, zero outbound traffic.
Media deleted right after each scan.
Public incidents
Deepfake incident log
A rolling sample of cases that went public. Resemble logs them so patterns are easier to see.
Read the guidesRestaurant caller cloned the owner, then a staffer
A worker says a voice clone "resigned" for them.
Reddit thread points to a Meta apology over Modi fakes
Shared coverage says doctored clips spread widely.
Faked industrialist video used to solicit money
Rajkot scammers reportedly borrowed his likeness.
Hedge funds reportedly hit by an AI voice scam
Attackers allegedly cloned voices to reach big firms.
WhatsApp voice notes cost a Hong Kong man $1.27M
Fraudsters reportedly mimicked his father's voice.
Fake Trump Super Bowl clip ends in a $500K transfer
An investor's funds reportedly moved anyway.
Resources
New to Resemble? Start with these three.
How RAD-S1 tells human visitors from AI agents
It reads movement, typing and clicks to separate people from lifelike bots.
View section
Inside the H1 2026 deepfake threat picture
Resemble's summary of synthetic attacks, fraud losses and enterprise defenses.
View section
EU AI Act Article 50: a working checklist
What voice AI providers and deployers must disclose, and when.
View sectionSee what's genuine — and what's a genuine risk.
Build with voice AI on solid ground. Resemble keeps the detection side clear while you create.
Create yours