Klangfarbe and audio watermarking: synthetic, or authorized
A watermark says the voice is synthetic. Klangfarbe says which organization's agent is speaking, for what, and whether the session is valid now.
Audio watermarking and Klangfarbe both put information into speech. What that information can prove is very different.
What a watermark proves
Systems such as Google SynthID and Meta's open-source AudioSeal embed an imperceptible mark in generated audio. Detecting the mark shows that the audio came from a particular generator, or that it is machine-generated. AudioSeal, for example, supports an optional 16-bit message and streaming detection.
That is valuable, and regulators increasingly expect it. But a mark by itself does not say which organization authorized the agent, what it was allowed to do, whether the call is happening now or was recorded last week, or whether the mark was copied onto other speech.
What Klangfarbe adds
Klangfarbe uses the audio only as a carrier. The payload is a compact, opaque handle and a rolling authenticator derived from a per-call session key. They point to a Session Attestation signed by the agent, which points to an Agent Authorization Credential issued by the organization, which a verifier checks against a trust framework it already accepts.
The result is an identity, not a label: Acme Corporation, Claims Assistant, claims intake, live session, verified.
Side by side
| Audio watermark | Klangfarbe | |
|---|---|---|
| Proves | Audio is generated, or came from a given system | An authorized agent of a named organization is speaking, for a stated purpose |
| Payload | A mark, sometimes a few bits | A session handle plus a rolling authenticator |
| Trust comes from | The watermark vendor | The organization's credential and a trust framework you choose |
| Replay of old audio | Usually still detected as marked | Reported as an invalid session |
| Copying the mark onto other speech | Hard to tell apart | Fails the rolling authenticator |
| Mark removed | Undetected | Reported as unverified, never as human |
Built to use watermarks, not to replace them
The carrier in Klangfarbe is pluggable. A better watermark should slot in without changing the credential model, and AudioSeal is a reasonable research baseline. What Klangfarbe defines is the identity layer above the carrier and the verification states around it. The open question it measures first is how many authenticated bits survive three seconds of speech after Opus and G.711.
The limit both share
Any mark in audio can be stripped or lost through re-synthesis. Klangfarbe does not pretend otherwise. Its rule is that presence can prove and absence cannot disprove, so a missing binding is always unverified, and the product never markets absence as authenticity. Read the protocol overview.