How PulmoForge
detects synthetic voices.

Most detectors look for AI artifacts. We also look for what makes a voice biologically human — and that distinction is what makes detection resilient.

680%
rise in voice deepfakes year-over-year in 2024
11s
to clone any voice with modern AI
$25M
lost in a single deepfake call — Hong Kong, 2024
4
detection layers combined into one risk score
Detection approach
A different question changes everything.
Traditional detection
"Does this sound like AI?"

Searches for artifacts left behind by synthetic generation. As voice cloning improves, those artifacts disappear — and the detector degrades with them. A purely artifact-based approach puts defenders in a perpetual catch-up position.

PulmoForge
"Does this sound like a real human?"

Also analyzes biological signals present in genuine speech — micro-variations in breath, tension, and vocal fold behavior that synthetic systems struggle to replicate consistently in real time.

Vocal instability
Breath & plosives
Phase continuity
Spectral behavior
Speech dynamics
Multi-layer analysis
Four layers. One risk score.
01
Biological Liveness
Measures whether speech contains characteristics expected from a real human speaker — including the micro-variations in breath, tension, and vocal fold behavior that synthetic systems struggle to reproduce at scale.
02
Acoustic Analysis
Examines pitch, energy distribution, and spectral patterns across the speech signal. Looks for the consistency and variation profiles characteristic of natural versus generated audio.
03
Signal Integrity
Detects inconsistencies caused by synthetic generation, post-processing, or signal manipulation — including artifacts introduced by real-time codec compression such as Zoom's Opus pipeline.
04
Ensemble Scoring
Combines signals from all three layers into a single continuous risk score. No single layer is relied upon in isolation — the ensemble is more robust to individual layer evasion.
Adversarial testing
Tested against a changing threat.

Synthetic voice technology evolves rapidly. An attacker-minded testing approach helps improve resilience as the threat landscape changes. PulmoForge is continuously evaluated against a range of synthetic speech systems.

Codec-compressed audio — specifically Zoom's Opus 32kbps pipeline — is a core training condition, not an afterthought.

Test category Coverage
Commercial voice cloning systems Continuously tested
Open-source speech generators Continuously tested
Emerging synthetic speech models Monitored and evaluated
Codec-compressed audio (Opus 32kbps) Core training condition
Why this matters
Trust in real-time communication is at risk.

AI-generated voices are becoming increasingly difficult to identify by ear. The gap between synthetic and natural speech is narrowing at a pace that outpaces human perception.

Organizations need tools that can provide additional confidence when important decisions depend on the identity of the speaker. PulmoForge is building technology designed to help verify that trust — in real time, during the call, before any damage is done.

Early access
Be first when we launch.

Free beta access goes to the waitlist first — no spam, just a single email when it's ready.