klangfarbeRequest the whitepaper

EU AI Act Article 50 and voice agents: marking is not identity

From August 2026, synthetic audio has to be machine-detectable. That tells a listener the voice is artificial. It does not tell them whose agent it is.

· su.engineering

Article 50 of the EU AI Act sets transparency duties for AI systems. Among them, providers of systems that generate synthetic audio, images, video or text must make sure the outputs are marked in a machine-readable format and detectable as artificially generated, subject to the regulation's scope and exceptions. These duties apply from 2 August 2026.

For voice agents, that pushes the industry toward watermarking generated speech. It is a real step forward, and it answers one question: is this voice artificial?

The question it leaves open

A contact center receiving a call from an AI agent, or a person picking up the phone, usually needs more than that:

  • Which organization authorized this agent?
  • What is it authorized to do on this call?
  • Is this a live, valid session, or a recording played back?
  • Did a different agent take over halfway through?

A synthetic-media mark cannot answer those. Two agents from two different companies, built on the same voice platform, can carry the same kind of mark.

Marking plus identity

Klangfarbe is designed as the identity layer above marking. Its binding is machine-readable and lives in the speech, and it points to signed evidence: the organization's Agent Authorization Credential and a per-call Session Attestation. A verifier can then show Acme Corporation, Claims Assistant, claims intake, live session instead of only "synthetic".

Klangfarbe is not a compliance product on its own and does not replace a provider's Article 50 assessment. It does make interoperability with watermarking ecosystems a priority, because regulation is creating the same machine-readable channel that identity can ride on.

What to do now

If you deploy voice agents in the EU, your Article 50 work will bring a marking mechanism into your audio pipeline. That is the right moment to ask whether the same channel should also carry who you are. Read the FAQ, see how the protocol works, or request the whitepaper.