An AI training environment that teaches agents to detect deepfake voices — the kind
used in phone scams, CEO impersonation, and identity fraud at scale.
Live on HuggingFace Spaces
5 investigation tasks
6-component scoring
Partial observability
48-dimensional features
$25B+
lost to voice fraud globally every year
banking · insurance · telecom · government
60s
to clone any voice with modern AI tools
real-time phone scams now possible at scale
0
existing benchmarks train agents this way
this environment fills that gap
1
How the AI agent investigates a voice
Most AI classifiers make a snap judgment from full information. This environment
forces the agent to work like a detective — starting with zero information, then gathering evidence step
by step before committing to a verdict. Skipping steps is penalised.
1
Check for natural vocal tremors
Real human voices have tiny involuntary wobbles in pitch and volume. AI
voices are unnaturally steady and smooth. The agent measures these micro-tremors first —
they are the strongest early signal.
Jitter · Shimmer · HNR
+0.10
2
Read the voice's sonic fingerprint
Every voice has a unique tonal texture — like acoustic DNA. The agent
examines 48 characteristics covering brightness, depth, and spectral variation across the
full frequency range.
48-dimensional feature vector
+0.10
3
Compare against a database of known voices
The agent measures how similar this sample is to confirmed real and
AI-generated voices — like comparing a fingerprint to a known database. It gets both a
similarity score and a distance measurement.
Cosine similarity · Centroid distance
+0.05
4
Weigh all evidence before deciding
Before giving a verdict, the agent reviews everything it found, looks
for contradictions between the signals, and works out how confident it should be. This step
is required — skipping it is penalised.
Evidence synthesis
+0.05
5
Give a verdict with honest confidence
The agent declares the voice real or fake — and states how confident it
is. Being 90% confident when wrong is penalised far more than saying "I'm only 50% sure."
This mirrors how real fraud decisions should work.
Final classification · Confidence calibration
scored
2
Five tests of increasing difficulty
Each test uses a different type of audio
degradation. Click any row to see the full score breakdown.
Clean studio recordingeasiest
74% scoreAgent detects reliably
Compressed / low-quality audiomedium
62% scoreCompression hides the clues
AI voice engineered to sound humanhard
55% scoreAdversarial AI fools the agent
Noisy live-stream conditionsvery hard
30% scoreBackground noise overwhelms signal
Real phone call conditionsextreme
22% scoreNear the limit of detection
3
How the score is calculated
Getting the right answer is only part of the score. Click any test
on the left to see its breakdown across all six components.
Got the right answer
Real or fake — was the verdict correct?
—
Honest about uncertainty
Overconfidence when wrong is penalised heavily
—
Followed investigation steps
Gathered evidence before deciding
—
Used all available clues
Checked both tremors and tonal texture
—
Explanation matched decision
Reasoning actually supported the verdict given
—
Did things in logical order
Gather first, analyse second, decide last
—
4
Watch it work — live investigation
Press the button to simulate an agent investigating a voice sample in real time. Each
line is an actual log message emitted by the agent as it works through the five steps.
inference.py — voice-authenticity-openenv
Ready. Press the button below to begin an investigation.