AI voice agent identity: questions and answers
The questions identity, telecom and contact-center teams ask first, answered in a paragraph each.
How can you verify which organization's AI voice agent is on a call?
With Klangfarbe, the organization issues the agent an authorization credential, the agent signs a short-lived attestation for each call, and a compact binding to that attestation is carried inside the speech. A verifier recovers the binding from the audio, fetches the attestation, and checks the credential and the organization's trust chain. The result names the organization, the agent and the authorized purpose.
Is an audio watermark enough to identify an AI agent?
No. A watermark can show that speech is machine-generated or came from a particular voice system. It does not say which organization authorized the agent, for what purpose, or whether the session is valid right now. Klangfarbe uses the audio only as a carrier for a binding to signed, verifiable session evidence. Compare Klangfarbe with audio watermarking.
How is Klangfarbe different from STIR/SHAKEN?
STIR/SHAKEN signs the calling telephone number in SIP signaling and says how well the originating carrier knows the caller. Klangfarbe identifies the AI agent speaking in the audio, the organization that authorized it and its purpose, and keeps checking for the whole call. They are complementary: a STIR/SHAKEN result can be one input to channel binding. Read the full comparison.
If no Klangfarbe binding is found, does that mean the caller is human?
No. Presence can prove; absence cannot disprove. Missing or damaged evidence is reported as unverified, never as human or safe. Klangfarbe is a way for an authorized agent to prove its identity, not a deepfake detector, and it never labels unmarked audio.
Does the binding survive phone networks and G.711?
That is the design goal and the first thing the working PoC measures. The binding is built to sit inside the 300 to 3400 Hz telephone band, and the MVP is tested across Opus, G.711 A-law and μ-law, resampling and recordings. The targets are a verified identity within 3 seconds of speech on WebRTC and within 5 seconds over telephony. These are targets, not measured results.
Can someone replay a recording of a verified agent?
A replayed recording carries a real binding from an old session. Short session lifetimes, channel binding to the SIP Call-ID or WebRTC session, verifier challenges and rolling sequence information make the verifier report it as an invalid session rather than a verified one. Relay into a live call is countered by channel and challenge binding.
Does verification need a su.engineering account or a central registry?
No. Verification is designed to be free and local, using open-source software. Organizations can be trusted through their own PKI, a federation, a sector trusted list or a Klangfarbe public trust service. su.engineering may run hosted resolvers and trust services, but the protocol never depends on them.
What does Klangfarbe reveal about the person receiving the call?
Nothing. Session handles in the audio are random and non-enumerable, and public session attestations contain only identity and authorization evidence. They never contain the recipient's phone number, customer names, case numbers or transcripts. Verification does not require a transcript.
Which standards does Klangfarbe build on?
W3C Verifiable Credentials 2.0 for the authorization credential, OpenID4VCI and OpenID4VP for issuance and presentation, IETF WIMSE and AIMS with SPIFFE for agent workload identity, OpenID Federation and EUDI-style trusted lists for trust, and C2PA with CAWG for recordings that outlive the call.
Does it help with the EU AI Act's rules on synthetic audio?
Article 50 asks providers of systems that generate synthetic audio to mark outputs in a machine-readable way. Klangfarbe's binding is machine-readable, but its goal is stronger: identifying the authorized organization and agent. It is not a compliance product on its own. Read the note on Article 50.
Is Klangfarbe available today?
Not yet. The whitepaper v1.0 was published for review on 21 September 2026, and a working proof of concept is being built now, over the next few months: a live WebRTC call, SIP over G.711, recorded-file verification and an attack suite. Design partners can shape what it proves first. Join as a design partner.