Audio-Deepfakes & Voice-Phishing: Wenn die Stimme am Telefon lügt

Three seconds of voice recording are enough today to clone a voice with deceptive realism. The tools for this are free, require no specialist knowledge, and are improving weekly. What began as a technical curiosity has become a tool for fraud: specialized providers of telephone fraud detection report an increase in deepfake attacks by over 1,300 % within a short period of time.

The Scam: Trust via Voice

Audio deepfakes target the oldest human security system: the recognition of a familiar voice. The typical attacks:

• CEO Fraud via phone call: The "management" calls the accounting department and orders an urgent transfer. Sound, tone, manner of speaking: everything fits.
• Grandparent Scam 2.0: Relatives are pressured using cloned voices of children or grandchildren: accidents, emergencies, immediate money.
• Attack on voice authentication: Where systems are unlocked by voice ("My voice is my password"), the clone becomes a master key.
• Manipulated recordings: Fake voice messages and recordings are appearing in disputes, extortion cases, and disinformation campaigns.

Why the ear is no longer enough

Modern speech synthesis reproduces not only timbre, but also breathing, speaking rhythm, and emotion. In tests, humans can hardly distinguish current clones from real recordings, especially over the phone, where audio quality is already reduced. The consequence: Humans alone are overwhelmed as a testing authority. Technical detection is required that looks for synthesis process artifacts that are inaudible to the human ear.

What the EU AI Act requires

Since August 2026, the following applies Article 50 of the EU AI ActAnyone who creates or distributes AI-generated or manipulated audio, image, or video content must label it as such. For companies, this means two things:

• Anyone using speech AI (assistants, announcements, avatars) must ensure labeling both organizationally and technically.
• Those who process incoming communication (contact centers, voice systems, transcription) benefit from detection technology: it provides proof that synthetic content has been detected and handled.

In addition, ISO/IEC 30107-3, the international testing framework for the detection of presentation attacks on biometric systems, against which professional solutions must be measured.

How can you recognize effective protection?

The market for "deepfake detection" is growing rapidly, and with it the number of promises. Three criteria separate substance from marketing:

• Public benchmarks instead of self-promotion. Reputable systems can be measured using independent test datasets (such as the Speech-DF-Arena benchmark, which compares detection systems across more than a dozen datasets). The key metric is the Equal Error Rate (EER): the lower, the better. Top systems are in the range of a few percent.
• Generalization to unknown methods. The crucial point is not to recognize known fakes, but new ones, which were not present at all during training. Reputable providers provide exactly this proof.
• Fail-Closed Principle When in doubt, reject; never let through. A detection system that grants access during uncertainty is mere decoration.

What companies should do now

• Hardening processes: Payment authorizations and critical instructions should never be accepted based on a verbal request alone, but always confirmed via a second, independent channel (callback to a known number, four-eyes principle).
• Raising employee awareness: Those who know that voices can be faked are more likely to question unusual calls. Code words for emergencies also help in private life.
• Check technical detection: Where voice communication is business-critical (contact centers, voice systems, approval processes), a detection layer must be integrated into the architecture, capable of being operated locally and auditable.

Insight-IT is developing its own Voice-Security-Gatea pre-positioned verification instance that checks incoming voices for authenticity before they are converted into a transcript, access, or authorization, on-premise and according to the fail-closed principle. In independent benchmark measurements, the system reaches the level of the best open detection systems. If you want to know what this means for your telephony or voice processes: we would be happy to demonstrate it using your own use cases.

Note: This article serves as an initial classification and does not replace legal advice, particularly regarding the application of the EU AI Act to your specific case.

← all guides