Aurigin’s real-time audio deepfake detection defends against phone, video call fraud

Companies, governments and individuals have been struggling to defend themselves from AI-generated voice fraud in which scammers use voice deepfakes to trick their victims. One of the biggest challenges for synthetic speech detection firms has been developing a solution that can detect fake voices in real time during phone calls, conferencing and streaming.

Swiss-based Aurigin.ai says it has solved this issue with a new audio deepfake detection model that promises a latency of less than 50ms.

The model, named Apollo, delivers a 97.7 percent accuracy, according to the firm. The figure is derived from a 2.3 percent weighted Equal Error Rate at 3 seconds of audio.

“We’ve managed to cut the error rate in half compared to the best commercial alternatives, while also outperforming all the open source solutions,” Luzi Sennhauser, the founder and CEO of Aurigin.ai, said during a Biometric Update webinar last week.

The webinar presented findings of the “2025 Deepfake Detection Market Report & Buyer’s Guide” along with examples of the deepfake protection technologies available today.

Aurigin.ai says it has evaluated the model across a wide range of audio spoofing and deepfake benchmarks, including ASVspoof variants, in‑the‑wild samples and Fake‑or‑Real.

Its weighted aggregate metrics include an F1 score of 97.45 percent and an Equal Error Rate (EER) of 2.76 percent overall. The False Positive Rate (FPR) was measured to be 2.13 percent, while the False Negative Rate (FNR) is 3.47 percent. The company says it computed the dataset‑level weights and aggregated metrics accordingly.

The Apollo model works regardless of phone‑call quality, re‑recordings and background noise and is tested across more than 40 languages.

“We’re fortunate to have some of the best audio experts helping us tackle this challenge,” adds Sennhauser.

Article Topics

AI fraud  |  Aurigin.ai  |  biometrics  |  deepfake detection  |  deepfakes  |  synthetic voice  |  voice biometrics