SwanOpen source, AGPL-3.0

Deepfake voice detection through phonation physics.

Swan measures the biological physics of speech to tell a human voice from a synthetic one. It never needs retraining for the next voice-cloning tool.

G.711
Telephony and HD audio
Real-time
Verdict per interaction
AGPL-3.0
Self-hostable
01How it works

Measure physics, not fingerprints.

Swan reads the acoustic behavior that only biological phonation produces, then integrates it into a stable verdict.

01

Measure

Jitter, spectral flatness, cross-band divergence, phase dispersion, and onset and offset transition morphology, frame by frame.

02

Integrate

A leaky-bucket integrator accumulates risk over time, so one noisy frame cannot flip the verdict.

03

Adjudicate

A forensic-style determination: human, synthetic, or inconclusive, with a stated behavioral driver.

04

Report

A JSON report plus readable dashboard and audit reports for every interaction.

02Self-host

Run it in minutes.

Ships with no API keys. You bring your own. The engine is local; the only optional external call is LLM adjudication, off by default.

# clone, install, run
git clone https://github.com/sriramnatrajhen/humsana-swan.git
cd humsana-swan
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env      # set your own keys
python run_bioauth.py

Decision support, not identity proof.

Swan assesses whether a voice is consistent with human phonation. It does not confirm who is speaking.